Academia.edu no longer supports Internet Explorer.
To browse Academia.edu and the wider internet faster and more securely, please take a few seconds to upgrade your browser.
ADITYA KUMARRelated papers
Monitoring of polymer extrusion and compounding processes
2020
This work focuses on monitoring of polymer extrusion and melt compounding using single or twin-screw extruders. Standard extruder instrumentation aims at detecting eventual instabilities in flow and heat transfer in the extruder. Strategies for in-process monitoring considering the actual characteristics of the material being processed are presented and discussed. On-line and in-line approaches are defined. The advantages of sampling along the extruder barrel are demonstrated. Examples of the development and application of in-process optical spectroscopy and rheometry techniques are presented and discussed
Low-cost process monitoring for polymer extrusion
Transactions of the Institute of Measurement and Control, 2013
Polymer extrusion is regarded as an energy-intensive production process, and the real-time monitoring of both energy consumption and melt quality has become necessary to meet new carbon regulations and survive in the highly competitive plastics market. The use of a power meter is a simple and easy way to monitor energy, but the cost can sometimes be high. On the other hand, viscosity is regarded as one of the key indicators of melt quality in the polymer extrusion process. Unfortunately, viscosity cannot be measured directly using current sensory technology. The employment of on-line, in-line or off-line rheometers is sometimes useful, but these instruments either involve signal delay or cause flow restrictions to the extrusion process, which is obviously not suitable for real-time monitoring and control in practice. In this paper, simple and accurate real-time energy monitoring methods are developed. This is achieved by looking inside the controller, and using control variables to ...
Thermal optimisation of polymer extrusion using in-process monitoring techniques
Applied Thermal Engineering, 2013
Workshop on New Measurement Technology for Polymer Processing - Gaithersburg, MD - December 3-4, 1990
Journal of Research of the National Institute of Standards and Technology, 1991
Industrial Polymers
Introduction, Classification, Thermosetting thermoplastic and other
EXTRUSION OF INTELLECTUAL POLYMER MATERIALS
Problem of obtain information about the characteristics of parts and components, in particular, their stress-strain state, in real time is extremely relevant in our time. Intrusion into the products made of polymer composite materials intelligent sensors allows for remote diagnostics and monitoring of these products directly during their operation. The article considers aspects of design of processing equipment and modeling of processing processes and non-destructive testing of sensors in polymeric material with the corresponding mathematical models. Also considered constructions of devices for obtaining information on the position of the sensors in the polymer melt based on the electrostatic method.
Polymer processing research in industry: Role of academic-industrial cooperation for fundamental research
Macromolecular Symposia, 1997
Engineering thermoplastics for additive manufacturing: a critical perspective with experimental evidence to support functional applications
Journal of applied biomaterials & functional materials, 2017
Among additive manufacturing techniques, the filament-based technique involves what is referred to as fused deposition modeling (FDM). FDM materials are currently limited to a selected number of polymers. The present study focused on investigating the potential of using high-end engineering polymers in FDM. In addition, a critical review of the materials available on the market compared with those studied here was completed. Different engineering thermoplastics, ranging from industrial grade polycarbonates to novel polyetheretherketones (PEEKs), were processed by FDM. Prior to this, for innovative filaments based on PEEK, extrusion processing was carried out. Mechanical properties (i.e., tensile and flexural) were investigated for each extruded material. An industrial-type FDM machine (Stratasys Fortus® 400 mc) was used to fully characterize the effect of printing parameters on the mechanical properties of polycarbonate. The obtained properties were compared with samples obtained by...
Barrel temperature control for quality of thermoplastic polymers in the extrusion process
Control and systems engineering, 2018
I nt e llige nt Ex t rude r for Polym e r Com pounding
Intelligent Extruder for Polymer Compounding
Final Technical Report
Prepared by
Alper Eker
Mark Giammattia
Paul Houpt (P-I)
Aditya Kumar
Oscar Montero
Minesh Shah
Norberto Silvi
Timothy Cribbs (GEIS)
GE Global Research
Automation and Controls Laboratory
Room KWD215, 1 Research Circle
Schenectady, NY 12309
Period of Performance
1/ 8/ 99-12/ 31/ 02
Prepared for
U.S. Department of Energy, Office of Industrial Technology
Contract DE-FC02-99-CH10972
DE-FC02-99-CH10972
1
I nt e llige nt Ex t rude r for Polym e r Com pounding
Table of Contents
Table of Contents ................................................................................................................ 2
Index of Figures .................................................................................................................. 4
Acknowledgements .............................................................................................................. 6
Disclaimer ........................................................................................................................... 6
1
Abstract ....................................................................................................................... 7
2
Introduction................................................................................................................. 7
2.1
Background ......................................................................................................... 7
2.2
Compounding basics ........................................................................................... 8
2.3
Problem and program objectives....................................................................... 10
2.4
Prior and related work ....................................................................................... 12
3
Overview of Tasks and Key Results .......................................................................... 13
3.1
Task 1: System Requirements ........................................................................... 13
3.2
Task 2: Process Models for Diagnostics and Controls...................................... 16
3.3
Task 3: Extruder Diagnostics ............................................................................ 16
3.4
Task 4: Extruder Inferential Estimation and Parameter Identification ............. 17
3.5
Task 5: Inferential Control System ................................................................... 18
3.6
Task 6: Control Platform and Experimental Extruder System.......................... 19
3.7
Task 7: Production Scale Demonstration and Validation ................................. 19
3.8
Task 8: Commercialization Plan ....................................................................... 20
3.9
Publications and Patents.................................................................................... 21
4
Experimental Extruder Setup for Demonstrations .................................................... 21
4.1
WP 25mm Lab Extruder Description................................................................ 21
4.2
Data Acquisition and Monitoring...................................................................... 22
4.3
D-Space Implementation for Closed-Loop Experiments .................................. 23
5
Extruder Modeling for Estimation, Diagnostics and Control................................... 25
5.1
Process Description & Modeling ..................................................................... 25
5.1.1
Process Description ................................................................................... 25
5.1.2
Physics-Based Lumped Model.................................................................. 26
5.1.3
Dynamic Model for Internal Holdup and Compositions........................... 27
5.1.4
Torque, Die Pressure and Viscosity Relations .......................................... 29
5.1.5
Dynamic Model Parameters ...................................................................... 33
5.1.6
Comparison of Model Predictions............................................................. 35
5.2
Extruder Modeling Summary............................................................................ 37
6
On-line Parameter Identification .............................................................................. 38
6.1
On-Line Parameter Identification Results......................................................... 40
6.2
On-line Parameter Identification Summary ...................................................... 41
7
Inferential Sensing..................................................................................................... 42
7.1
Viscosity Estimation Results............................................................................. 43
7.2
Viscosity Estimation Summary......................................................................... 47
8
Extruder Diagnostics................................................................................................. 48
8.1
Problem and Approach...................................................................................... 48
8.2
Fault Diagnostics – Noryl Case Study .............................................................. 52
8.2.1
Fault Signatures......................................................................................... 52
8.2.2
Fault Diagnostics with Noryl on 25mm Research Extruder ..................... 55
DE-FC02-99-CH10972
2
I nt e llige nt Ex t rude r for Polym e r Com pounding
8.2.3
Fault Identification Results with Noryl for Raw Material Changes.......... 57
8.2.4
Fault Identification of Feeder Bias............................................................ 59
8.3
Diagnostics Summary ....................................................................................... 60
9
Extruder Control ....................................................................................................... 61
9.1
Approach to control in the presence of upsets .................................................. 61
9.2
Viscosity Adaptive Control............................................................................... 62
9.3
Control implementation for research extruder .................................................. 64
10
Scale-Up for Production-Scale Extruders............................................................. 70
10.1 Scale of Dynamic Input-Output Model of Extruder.......................................... 70
10.2 Scale-Up of On-line Adaptation of Model Parameters ..................................... 73
10.3 Scale-Up of Viscosity Estimation ..................................................................... 74
10.4 Scale-Up Summary ........................................................................................... 75
11
Benefits .................................................................................................................. 78
11.1 Overview of benefits derivation........................................................................ 78
11.2 Approach to benefits quantification .................................................................. 80
12
Commercialization Plan........................................................................................ 83
12.1 Market Opportunity........................................................................................... 83
12.2 Commercialization Strategy.............................................................................. 86
12.2.1 Product – Advanced Process Control and Service .................................... 86
12.2.2 Remote Service Offering........................................................................... 87
12.2.3
Distribution – Utilizing CWP’s Distribution Channels............................. 87
12.2.4
Commercialization Sales Tools................................................................. 87
12.3 Commercialization Status ................................................................................. 88
13
Conclusions and Follow on Recommendations .................................................... 88
14
References ............................................................................................................. 89
APPENDIX-A Extruder Dynamic Models......................................................................... 91
A.1 Dynamic model for Hold up.................................................................................. 91
A.2 Dynamic model for composition........................................................................... 92
A.3 References for Appendix A.................................................................................... 97
APPENDIX B - Extruder Drive Torque Estimation using Electrical Variables............... 98
B.1 Summary................................................................................................................. 98
B.2 Requirements and Objectives ................................................................................ 98
B.3 Approaches to observer based torque estimation .................................................. 99
Current/Voltage model................................................................................................ 102
B.4 Saber Simulation.................................................................................................. 103
B.5 Conclusions from simulation results .................................................................... 106
B.6 References for Appendix B .................................................................................. 107
DE-FC02-99-CH10972
3
I nt e llige nt Ex t rude r for Polym e r Com pounding
Index of Figures
Figure 1 Extruder for Polymer Compounding............................................................... 8
Figure 2 Profile of Material State in Extruder.................................................................... 9
Figure 3 Sources of Finishing Variability........................................................................ 11
Figure 4: WP ZSK-25mm twin-screw extruder used at GE Global Research for
experiment runs ......................................................................................................... 22
Figure 5: Schematic representation of capillary and RDS rheometers ............................. 23
Figure 6: GUI interface for estimation and controls algorithms using D-Space....... 24
Figure 7: A typical extruder setup..................................................................................... 26
Figure 8: Schematic representation of extrusion process.................................................. 26
Figure 9: Schematic representation of mixing in partially and completely filled sections.
................................................................................................................................... 28
Figure 10: Input-output relation for torque at steady state................................................ 30
Figure 11: Comparison of measured viscosity variation with PPO fraction and model fit.
................................................................................................................................... 31
Figure 12: Comparison of Die Pressure measurement with model prediction ................. 34
Figure 13: Comparison of torque measurement with model prediction............................ 34
Figure 14: Inverse response in torque with respect to screw speed .................................. 36
Figure 15: Inverse response in die pressure with respect to screw speed ......................... 36
Figure 16: Inverse response in die pressure with respect to PS feed-rate ......................... 37
Figure 17: Schematic representation of on-line parameter identification approach. ........ 39
Figure 18: Comparison of model prediction after on-line adaptation vs measurement for
die pressure under composition change from nominal conditions............................ 40
Figure 19: Comparison of model prediction after on-line adaptation and measurement of
die pressure under raw material change from nominal conditions............................ 41
Figure 20: Approach for model-based estimation of viscosity ......................................... 43
Figure 21: Comparison of on-line viscosity estimation and off-line lab viscosity
measurements for experiment on 5/8/2001 with nominal (.46 IV) PPO blend......... 44
Figure 22: Comparison of on-line viscosity estimation and off-line lab viscosity
measurements for run on 6/25/2001 with nominal (.46 IV) PPO blend ................... 44
Figure 23: Comparison of on-line viscosity estimation and off-line lab viscosity
measurements for run on 7/25/2001 with nominal (.46 IV) PPO blend ................... 45
Figure 24: Comparison of on-line viscosity estimation and off-line lab viscosity
measurements for run on 8/27/2001 with low IV (.33 IV) PPO blend ..................... 45
Figure 25: Comparison of on-line viscosity estimation and off-line lab viscosity
measurements for run on 10/11/2001 medium IV (50/50 mixture of .46 IV and .33
IV) PPO blend ........................................................................................................... 46
Figure 26: Comparison of on-line viscosity estimation and off-line lab viscosity
measurements for run on 11/14/2001 with medium IV (50/50 mixture of 0.33IV and
0.46 IV) PPO blend ................................................................................................... 46
Figure 27: Schematic diagram of extrusion process and sources of product variability... 48
Figure 28: Schematic description of fault detection and correction with distinct phases –
(1) initial nominal operation, (2) fault occurrence and detection – mismatch between
model and measurement, (3) model update – on-line model adaptation and fault
DE-FC02-99-CH10972
4
I nt e llige nt Ex t rude r for Polym e r Com pounding
identification, (4) corrective control action for fault, (5) on-spec operation after fault
correction................................................................................................................... 49
Figure 29: Schematic of fault detection approach............................................................. 51
Figure 30: Fault signatures for changes in raw material ................................................... 52
Figure 31: Fault signatures for PPO/PS feeder bias.......................................................... 54
Figure 32: Fault occurrence, detection and identification using model parameter
identification for a representative run with raw material change.............................. 56
Figure 33: Comparison of measured and predicted die pressure for bias in PPO/PS feeder
using nominal parameters.......................................................................................... 60
Figure 34: Schematic diagram for closed-loop control of product viscosity .................... 62
Figure 35: Shematic block-diagram of the closed-loop control algorithm based on on-line
viscosity estimation ................................................................................................... 63
Figure 36: Schematic diagram of the final implementation using D-Space for closed-loop
control experiments ................................................................................................... 64
Figure 37: Closed-loop control of viscosity – set point tracking with nominal raw
materials on 12/14/2001 ............................................................................................ 66
Figure 38: Comparison of closed-loop control of viscosity –set point tracking using
nominal raw materials - with off-line viscosity measurement on 12/14/2001.......... 67
Figure 39: Experiment run on 12/14/2001 with nominal raw materials, accounting for
PPO and PS feeder bias............................................................................................. 68
Figure 40: Experiment run on 12/14/2001 with medium IV PPO, accounting for PPO and
PS feeder bias ............................................................................................................ 69
Figure 41: Comparison of model-predictions and measured value of torque on a 120mm
extruder during a DOE run at GEP Selkirk on Jan 18, 2001. ................................... 71
Figure 42: Comparison of model-predictions and measured value of die pressure on a
120mm extruder during a DOE run at GEP Selkirk on Jan 18, 2001. ...................... 72
Figure 43: Validation of model for torque using DOE data on a 120mm extruder at GEP
Selkirk on Nov 21, 2000. .......................................................................................... 72
Figure 44: Validation of model for die pressure using DOE data on a 120mm extruder at
GEP Selkirk on Nov 21, 2000................................................................................... 73
Figure 45: Comparison of measured and model predicted values of die pressure during
parameter adaptation with DOE data from 120mm extruder at GEP Selkirk on Jan
18, 2001..................................................................................................................... 74
Figure 46 : Steps for implementing developed model-based estimation, diagnostics and
control algorithm in an extruder application. ............................................................ 77
Figure 47 - How Intelligent Extruder Derives Benefits .................................................... 78
Figure 48 Process Timeline for Benefit Calculations ....................................................... 79
Figure 49 Process Spread Sheet used for Benefits Analysis with Customer .................... 81
Figure 50: Potential Extruder Market Segments ............................................................... 85
Figure 51: Detailed Opportunity Fishbone for Intelligent Extruder ................................. 86
Figure 52: Schematic representation of mixing in partially and completely filled sections.
................................................................................................................................... 92
Figure 53: Viscosity vs composition of PPO .................................................................... 94
DE-FC02-99-CH10972
5
I nt e llige nt Ex t rude r for Polym e r Com pounding
Acknowledgements
This material is based upon work supported by the U.S. Department of Energy under
Award No. DE-FC02-99-CH10972. The authors wish to acknowledge the guidance and
generous use of facilities of the Coperian Werner & Pfleiderer Corporation and GE
Plastics. At W&P in Ramsey, N.J., we are particularly indebted to John Curry and Rich
Taylor of their engineering team for their sustained in-kind support, ideas and guidance
throughout the program, and to Eberhard Dietrich in sales for perspective on markets and
applications. Many valuable contributions were provided by technology and
manufacturing staff at GE Plastics, Selkirk, NY and Mt. Vernon IN, including: Robert
Hossan, Ashish Kulkarni, Bo Liu, Eric Mortensen, Eric Gohr, Horst Oberst (retired), and
Dennilu Sosa. Prof. David Bigio, Dept. ME, U. Md, College Park, provided valuable
expertise in process modeling.
Disclaimer
Any opinions, findings, and conclusions and recommendations expressed in this material
are those of the authors and do not necessarily reflect the views of Department of Energy.
DE-FC02-99-CH10972
6
I nt e llige nt Ex t rude r for Polym e r Com pounding
1 Abstract
“Intelligent Extruder” described in this report is a software system and associated support
services for monitoring and control of compounding extruders to improve material
quality, reduce waste and energy use, with minimal addition of new sensors or changes to
the factory floor system components. Emphasis is on process improvements to the
mixing, melting and de-volitization of base resins, fillers, pigments, fire retardants and
other additives in the “finishing” stage of high value added engineering polymer
materials. While GE Plastics materials were used for experimental studies throughout the
program, the concepts and principles are broadly applicable to other manufacturers
materials. The project involved a joint collaboration among GE Global Research, GE
Industrial Systems and Coperion Werner & Pleiderer, USA, a major manufacturer of
compounding equipment. Scope of the program included development of algorithms for
monitoring process material viscosity without rheological sensors or generating waste
streams, a novel detection scheme for rapid detection of process upsets and an adaptive
feedback control system to compensate for process upsets where at line adjustments are
feasible. Software algorithms were implemented and tested on a laboratory scale extruder
(50 lb/hr) at GE Global Research and data from a production scale system (2000 lb/hr) at
GE Plastics was used to validate the monitoring and detection software. Although not
evaluated experimentally, a new concept for extruder process monitoring through
estimation of high frequency drive torque without strain gauges is developed and
demonstrated in simulation. A plan to commercialize the software system is outlined, but
commercialization has not been completed.
2 Introduction
2.1 Background
U.S. polymer resins and their compounds are produced at the rate of 90 billion lb/ yr,
one-third of which are engineering thermoplastics valued at more than $1/lb and used in
applications from small electrical connectors to medical products, optics to computer
cases, kitchen countertops, and automotive fenders and bumpers. Before reaching the
injection molder manufacturer, nearly every pound of resin passes through a final
“finishing” stage in a compounding line ( Figure 1), in which component materials are
blended in a 1000 to 15,000 hp extruder to achieve critical properties such as melt flow
behavior, color, mechanical strength, and fire resistance. Injection molders depend on the
values of these properties for their equipment to produce quality parts. Variations in resin
properties increase the initial setup time for injection molding and readjustments during a
run, resulting in more scrapped parts due to incomplete or excessive mold flow or
substandard properties such as color. The outcome, for both molders and resin makers, is
productivity loss, missed deliveries, additional landfill scrap, higher production costs, and
DE-FC02-99-CH10972
7
I nt e llige nt Ex t rude r for Polym e r Com pounding
dissatisfied customers. By improving the quality of material produced to meet molders’
expectations, the compounding industry and its customers benefit while energy and
waste generated in recycle is reduced.
Material Feeders
Drive-Motor
Die head
Finished
Pellets
Melt strand
Gear Box
Screw-Barrel
Water-Bath
Pelletizer
Figure 1 Extruder for Polymer Compounding
2.2 Compounding basics
Extruders are widely used, not only in polymer preparation, but throughout the petrochemical and food industries for mixing, blending, reacting, cooking, devolatilizing and
numerous other tasks, often at the end of the manufacturing chain where material quality
attributes are established and value has been added. Figure 1 shows the principle
components. Dry materials are conveyed from storage hoppers with loss-of-weight
feeders into various ports in the barrel. The screw conveys, recirculates, mixes and
kneads the materials toward the die head, imparting sufficient work to achieve the
required degree of mixing and temperature and pressure rise by the time it reaches the die
head that melted material emerges in continuous multiple strands. After cooling in a
water bath, the solid polymer strands are chopped into pellets and packed in boxes or
railcars for shipment to injection molding customers.
DE-FC02-99-CH10972
8
I nt e llige nt Ex t rude r for Polym e r Com pounding
Figure 2 Profile of Material State in Extruder
Figure 2 shows what the internal material state transitions might look like, moving from
the input to the output of the extruder barrel at the die head, recognizing that at any
position the material flow and melt state is highly complex and 3-dimensional. The
design of the screw is a complex science and art is not part of our investigation. For a
given screw, base resin and complementary components to create a desired alloy, there is
a desired profile the process engineer seeks. This cartoon example illustrates typical
complexities encountered, with dry ingredients added at two locations, an open devolatization vent, and changes in the solid melt ratio along the barrel, but ultimately
producing 100% melt at the die head with an appropriate melt temperature and pressure
to assure well mixed strand formation. Heaters and/or cooling jackets are located along
the barrel length to help maintain temperature conditions in the melt, but the bulk of the
energy that melts and blends the material comes from the screw drive.
DE-FC02-99-CH10972
9
I nt e llige nt Ex t rude r for Polym e r Com pounding
2.3 Problem and program objectives
The objective of the intelligent extruder program is to
develop and demonstrate a prototype software-based
monitoring, diagnostic, and control package that will reduce
production variability, energy use, and offgrade or waste
stream generation.
Both large manufacturers and small independent operations, have used extruders and
associated feeders, mixers, and pelletizers for polymer compounding for many years
without exotic automation equipment. But market forces are driving change in the
industry:
• Smaller lots of material, especially those made to order with short lead times (72
hr), put a premium on efficiency of setup and changeover, which today can
consume an entire shift.
• To improve the productivity of their equipment, injection molders are narrowing
acceptable quality limits on material properties (melt viscosity and color) from
their resin suppliers.
• Price deflation and cost pressures mandate increased productivity (dollars per
pound produced), while reducing energy used and landfill waste generation.
• For products in demand, every pound of recycle processed is a pound loss in
virgin material capacity that will lead to missed orders or expensive capital
equipment to raise plant capacity to compensate for low first-pass yield.
Some of the principle sources of variability in compounding are illustrated in .
Variability can be attributed to: (1) operator errors; (2) incoming material variations; (3)
equipment faults of various types; (4) process faults/upsets. Each listed fault types in the
figure can have an impact quality metrics of importance.
Developing a coherent and systems-based approach to these challenges is the objective of
this program, leading to products and services that the GE-Coperian Werner-Pfleiderer
Corp. team can sell, and that resin manufacturers can exploit for increased profitability.
A cross-functional team was formed to identify the most important needs in
compounding as identified in. Prioritized goals by the team included:
• Reduce operator errors. Continuous online monitoring could reduce or
eliminate human errors by detecting them quickly and allowing corrective
intervention
DE-FC02-99-CH10972
10
I nt e llige nt Ex t rude r for Polym e r Com pounding
Process Faults
Operator
Errors
• formulation
• setup
• QC errors
Incoming
Material
Variation
Feeders
Gears
Screw
Pelletizer
Resin MW
Resin color
Resin morphology
SPD pellet
Additive feed rates
Product
Variability
Heaters
Motor
Bath
Drive
•
•
•
•
•
• clogged feeders, die screen
• poor barrel temp regulation
• vent, die port blockage
• strand drop, fuse
• hot pellets
Equipment Faults
• MV
• Composition
• Color
• FPY
• COST / #
• feeder drive, motor , control
• dull pellet cutter
• screw shaft breakage
• screw wear
• gear box (bearing, gears)
Figure 3 Sources of Finishing Variability
•
•
•
Reduce the effects of incoming material variation. New inferential sensing
technology could be used to provide a continuous on line estimate of
property shifts, which can be used in closed loop or manually to initiate
trim corrections with secondary feeds or other machine adjustment.
Detect and correct process faults. Algorithms to detect and diagnose
process faults could be used to quickly divert product and/or initiate
operator intervention to make corrective action before upsets produce a
degraded product.
Anticipate and detect equipment faults. New diagnostic methods could
look for online data trends characteristic of impending faults (e.g., screw
wear).
The objective of the intelligent extruder program is to develop advanced diagnostics and
controls, responsive to the challenges described above, which show technical feasibility
to reduce product variability, increase first-pass yield, while reducing energy use and
waste generation in the compounding of polymer resin. The primary deliverables include:
demonstration of software and algorithms to provide process and equipment diagnostics
for a defined scope of fault coverage; provide an “inferential” predictor of output material
properties (melt viscosity) from available temperature, pressure, power, speed, and torque
DE-FC02-99-CH10972
11
I nt e llige nt Ex t rude r for Polym e r Com pounding
and internal extruder data derived from torque; provide closed-loop control based on
estimates from the inferential predictor; prove that the methods scale from laboratory to
full-scale production for at least one GE polymer material; and implement, test, and
integrate all software/controls for demonstration purposes on commercially available
industrial hardware.
2.4 Prior and related work
A vast literature exists on modeling, diagnostics and adaptive control in general, and we
make no attempt at a comprehensive survey here. What follows are some of the past and
ongoing work of others that motivated this work. In the following, citation numbers in
bracket [] refer to references in Section 14 of this report.
In an effort to develop more comprehensive closed loop controls for composition and
MFI, the use of melt flow rheometers has been popular and is used widely, for example
as described by Gottfert [6] and Dealy [7], the latter providing a useful survey of known
techniques, but the cost, maintenance and generation of a waste stream of these sensors is
a major drawback. Moreover, to minimize the waste produced, low flow rates are
employed with the result that latency in obtaining a suitable measurement can be several
minutes. Sensors working in the near IR have been developed for extruder rheological
measurements, exemplified by the work of Hansen et al [10], but these costly devices are
at least today expensive and complex to use at the production level. A promising, noncontact transient IR spectroscopy being developed by McClelland et al is described in
[16]. This technique uses laser induced transient ultrasound spectroscopy to extract
information about composition and viscosity (among other properties). Following
correspondence and discussions by Prof. McClelland with the GE team, this method
looked very promising as an at-line system since it could work with hot strands as they
emerged from a die head, and we will continue to follow his developments. Like all
optical systems, it may be a challenge to use in a production environment due to
contamination, steam, suspended dust and other grime. Many other at-line sensors have
been studied for use in compounding, particularly for color. On-line colorimetry using
video based systems for pellet color, for example, are available for sale by Macbeth
instruments and others. But these systems suffer accuracy from the end “stress
whitening” that results from cutting strand into pellets. The GE team has developed and
demonstrated, but not commercialized, a video colorimetry system (see Campo et al[2])
that works directly on strand and uses off the shelf RGB video cameras and can measure
color to within 1.5 delta E (in the CIE L-a-b color coordinate system). While all of these
systems have shown promise for production operations, they all have limitations, and will
still require a QA lab for calibration and maintenance.
Prior work by the GE team on process control for extruders was focused on color control
by Houpt [2] which was enabled by a real-time video based colorimeter above. Process
control in extruders can go beyond the alloying or non-reactive blending addressed in this
DE-FC02-99-CH10972
12
I nt e llige nt Ex t rude r for Polym e r Com pounding
program, e.g. Pabedinskas [5] and Curry [13, 14]. In fact, the team believes reaction
control is a promising future direction for the Intelligent Extruder system development.
In this program, our goal has been to overcome deficiencies of controls that arise from
transient dynamics that confound the simplest steady-state relationships or transferfunctions between system inputs and responses. And while there is no shortage of solid
fundamental work in understanding process dynamics for purposes of process
optimization and screw design (see e.g. Curry [13-15], McKay [8]), such models are
typically unsuitable for control design owing to their complexity. Rather, the philosophy
is to extract the simplest possible model to capture the “important” dynamics (where
what constitutes importance is the outputs in response to inputs are correct), and further
to link these models that relate measurable process variables (speed, torque, feed rates,
temperature etc.) to un-measurable ones (viscosity etc). Purely heuristic models such as
derived from neural networks (e.g. Eerikainen et al [9] who have applied such methods to
food extrusion) have been extremely popular in many other fields, or taking a fuzzy
modeling and control approach (see Isermann [4] for an overview), and may yet provide
value for this effort. But “neuro-fuzzy” approaches almost always over-parameterize data
that models input/output behavior to match and require lots of (scarce) data. There is
usually little physical insight to what is going on that can be explained by the model fits.
Models that formed the basis of our work were first used by Gao and Bigio [11,12].
These models were originally aimed at studying residence time distributions for extruders
(how we have adapted and extended the models is described in Section 5.1 of this report).
Diagnostics and controls based on these models are described in detail in this report. The
underlying philosophy of model based detection and some of the potential techniques
considered for use in this effort are described in a survey by Gertler in [3]; the reader
interested in a comprehensive treatment of the subject can consult his excellent book [17]
which also provides numerous examples in the “art” of the methodology as applied here.
3 Overview of Tasks and Key Results
In this section we provide a summary of the major task objectives, actual
accomplishments and provide a road map to the relevant sections of the report where the
details are discussed.
Work was broken down into eight major tasks plus program management. The eight
major tasks are summarized below, with details provided in Sections 4-13 of the report.
The approach to and estimates of projected benefits is described in Section 11.
3.1 Task 1: System Requirements
Objective
Four major sub-systems were identified for development, including
• Extruder process diagnostics
• Extruder inferential estimation
• Extruder inferential control system
• Control implementation platform
DE-FC02-99-CH10972
13
I nt e llige nt Ex t rude r for Polym e r Com pounding
The goal for this task was to identify requirements for each of these systems applied to
polymer compounding applications, recognizing that our vision for Intelligent Extruder
included other products (see Commercialization Plan details in Section 12.1). Table 1
provides a brief summary of the purpose of each system to be developed, key functional
requirements, and where applicable, specific targets to be achieved. These requirements,
identified by a cross-functional team, address many of the upsets known to occur in
typical polymer processing operations. A more comprehensive list of “critical to quality”
(CTQ) parameters which is the source of Table 1, is provided in Table 2. The team
examined more than 60 potential needs, grouped by sub-system. After prioritizing based
on impact and likelihood, the proposed functionality was proposed.
System
Inferential Process Sensing /
Estimation: Obtain estimates
of process model constants
and dynamic variables from
measurable machine sensors
(drive torque, die pressure,
speed, etc) with minimum
material use and lab checks
Process diagnostics- Use
process understanding and
models to detect and classify
process upsets of interest
Inferential control – Building
on above systems, develop
control strategy to make
feeder or drive adjustments
on-line to keep process
viscosity within specs
Control implementation
platform
Key Requirements
Identify key system model
parameters (see 5.1.2) and
provide means to track
changes from process shifts
and grade changes
Specific targets
Model parameters to match
input/output dynamic response
Viscosity(MFI) +/- 10% or
better (vs. 5% 1-σ for typical
lab rheometry)
Estimate process viscosity
(or corresponding melt flow
index (MFI) from model
and machine variables
Detect and correctly classify
process upsets from raw
material variations, feeder
anomalies (blockage, drift,
sensor bias etc)
Regulate viscosity to no worse
the twice the measurement
error variance steady state;
provide adaptive framework to
simplify the modeling efforts
over material grade changes
Design all software to be
capable of running on typical
industrial DCS process control
equipment and/or PC-class
industrial controllers
95% Detection of feeder/resin
shift faults as defined in
Section 8.2, with 10% false
alarm rates
For correctable upsets,
maintain viscosity +/- 10% of
target value
Achieve continuous adaptation
for “neighboring” grades to
minimize transition waste
material
Target PC class control
computer interfaced to drive
PLC within scope of
demonstration program
Table 1: System Requirements for Intelligent Extruder
The main conclusion from examining Table 2, is that most upsets of importance derive
from problems external to the extruder (feeders and material variation). The drive,
gearbox and controls are comparatively reliable and have extensive diagnostics of their
own, or are already provided by 3rd parties, e.g. vibration based gear-box monitors.
Clearly, there are other important process defects that lead to rejected material, e.g. color,
surface defects and mechanical strength. But other means must be provided for tracking
DE-FC02-99-CH10972
14
I nt e llige nt Ex t rude r for Polym e r Com pounding
color shifts/upsets since these are rarely, except in extreme cases, detectable from
machine variables. But viscosity, important in its own right, is often leading indicator of
problems in composition which in turn is linked to material properties. Thus it was the
opinion of the team that considerable benefits derive from detecting and managing
viscosity related upsets, which became the primary focus of this effort.
System
FEEDERS
SCREW
DRIVE
Fault
Likelihood
(1=LOW;2=MED;3=HI)
1<LIK<4
CTQ Impacted (0=none; 1= weak; 3=moderate; 9=strong)
Capacity
Melt V
Composition
FP Yield
Total
Ult Yield Pellet Size Pellet App. Matl Cost Proc EHS
Other
Weighted
blocked main feed port
3
9
9
9
9
9
9
9
9
216
blocked chute
3
8
8
9
9
9
8
8
9
204
blocked sec feed port
3
8
7
8
8
8
3
5
9
168
resin l.o.w. feed controller
2
8
8
8
8
8
8
8
8
128
pigment line block
2
8
6
9
6
6
2
6
9
104
filler feeder drive
1
9
9
9
9
9
5
5
9
64
filller conveyer
1
9
9
9
9
9
5
5
9
64
transfer conveyer drive
1
9
9
9
9
9
0
0
9
54
pigment l.o.w. drive
1
8
6
9
6
6
2
6
9
52
18
drive computer
1
9
0
0
0
0
0
0
9
drive power electronics
1
9
0
0
0
0
0
0
9
18
drive setup error
1
9
0
0
0
0
0
0
9
18
drive / DCS comm error
1
9
0
0
0
0
0
0
9
18
gearbox lube pump
2
9
0
0
0
0
0
0
9
36
gearbox lube contamination
2
6
0
0
0
0
0
0
4
20
motor shaft
1
9
0
0
0
0
0
0
9
18
and
motor winding
1
9
0
0
0
0
0
0
9
18
GEARS
motor brushes
1
9
0
0
0
0
0
0
9
18
gearbox shaft
1
9
0
0
0
0
0
0
9
18
gearbox bearings
1
9
0
0
0
0
0
0
9
18
gearbox gears
1
9
0
0
0
0
0
0
9
18
motor bearing
1
6
0
0
0
0
0
0
6
12
feedport(s) blockage
4
8
7
7
7
7
4
6
8
216
feed hopper bridge
3
8
7
9
7
6
8
5
7
171
unmelt particle passage
4
6
4
0
5
6
3
5
4
132
feed hopper buildup & sloughing
3
4
6
6
5
4
6
3
4
114
vent port blockage
4
6
5
2
2
3
0
2
4
96
clogged die port(s)
4
3
2
0
1
1
6
6
5
96
strand fuse=fused pellets
4
0
0
0
0
0
9
9
5
92
feed cycling (bad tuning?)
2
5
6
7
6
5
5
3
6
86
blend separation in feeder
2
3
7
8
7
5
4
2
5
82
MOTOR
SCREW
worn screw elements
3
4
6
0
5
5
1
1
4
78
wrong screw makeup
3
5
5
0
5
5
0
2
4
78
clogged die screen
4
5
3
0
2
2
1
2
4
76
wrong KWH/# input
4
5
2
0
3
4
1
1
3
76
broken screw elements
2
7
8
0
7
7
1
1
6
74
feed flooding
2
6
5
5
4
5
4
4
4
74
vent vacuum failure
3
6
5
2
2
3
0
2
4
72
pigment aglom. via cold working
2
2
2
6
5
7
0
5
5
64
barrel heater(s) failure
3
4
1
0
1
1
2
1
1
33
barrel temp sens failure
3
4
1
0
1
1
2
1
1
33
strand drop
4
3
0
0
0
0
0
0
5
32
Table 2 "Complete" List of CTQs for Upset Coverage in Intelligent Extruder
DE-FC02-99-CH10972
15
I nt e llige nt Ex t rude r for Polym e r Com pounding
3.2 Task 2: Process Models for Diagnostics and Controls
Objective- Develop simplified extruder process dynamic models suitable for use in
advanced diagnostics and controls. Because of the complex, dynamic nature of
extruder physics, it was the team’s experience that many past efforts at monitoring and
diagnostics, e.g. based on simple statistical models, failed or were not robust. On the
other hand, detailed 3-D extruder flow models are too complex to exploit in a real-time
setting. The goal was to find a simple enough, but physically “correct”, model to use in
various aspects of the controls and diagnostic development.
Accomplishments – Motivated by work by Bigio and Gao [11,12], a lumped model is
derived and which despite its simplicity, is adequate to capture the nonlinear input-output
dynamics of the extruder. In response to changes in feeds and drive speed (inputs), the
model predicts die pressure/temperature and torques (outputs). Details on the model
derivation and its validation on both research scale and production scale extruders are
contained in Section 5 and 10 of this report respectively. The model contains certain
unknown parameters, which depend on the machine type and screw configuration, and on
the material being used and operating conditions (temperatures, pressures, production
rate). Explicit means to identify and adaptively track these parameters in a highly
efficient manner are developed as part of Task 4, and are described in Section 6 of the
report. The machine parameters need to be identified only once (for a given screw
configuration), whereas the material parameters vary for each grade or major change in
operating regime. Having an efficient means to identify and track model parameters is
key to making all the methods developed in this program practical with a minimum of
calibration/test and wasted material.
3.3 Task 3: Extruder Diagnostics
Objective—Develop and demonstrate algorithms for detection and classification of
key extruder compounding system faults, emphasizing material feed sub-systems
Accomplishments
The models from Task 2 and the identification methods from Task 4 form the basis of our
proposed algorithms for fault detection. As discussed above in the requirements task, our
investigation of process related faults suggested that feeder anomalies and material
variation were among the most important events to be able to detect rapidly and take
corrective action. In Section 8 of the report, the main method is derived. The derivation
shows that detection of fault events derives from comparing measured variables with
those predicted by a system model, assuming no faults exist. Differences that exceed a
threshold indicate a fault condition. A new approach is developed in which the pattern of
identified parameters following the fault occurrence can be used to isolate among the
possible causes. A demonstration case study illustrates the method using data from the
three-feeder research extruder with Noryl materials, in the presence of various property
shifts and feeder faults. Section 8 also outlines how the method can be extended to
materials with multiple constituents. It is shown that typical upset events can be detected
in time frames as short as a few seconds up to a few minutes of operation. The main
DE-FC02-99-CH10972
16
I nt e llige nt Ex t rude r for Polym e r Com pounding
limitation of the methodology is that faults are assumed to occur one at time. Multiple
concurrent faults can be handled, but the decision logic is more complex (see
Gertler[17]). Although concurrent extruder system faults are detectable, it may be
impossible to disambiguate the root cause in general, but at least manual intervention via
an alarm can be flagged.
3.4 Task 4: Extruder Inferential Estimation and Parameter
Identification
Objective – Develop and validate experimentally, algorithms to identify unknown
parameters in the extruder system models from Task 2. Using these models, derive
and demonstrate means to “infer” or estimate process states, emphasizing viscosity
or MFI, from machine variables.
Accomplishments
Model Parameter Identification
Simplified dynamic process models form the foundation for all the results in this
program, including estimation, diagnostics and controls. In Section 6 of the report, we
show that a total of six parameters must identified, four dependent on machine geometry
and two of which depend on the specific material (and process operating conditions).
Even though the two material parameters must be found for each grade of material (of
which there may be hundreds), the fact that only two material parameters suffice to
parameterize models for the estimation and diagnostics results that follow was surprising
and counter-intuitive. Moreover, it can be shown that material of similar grades can be
identified during the grade transition and lineout. That is, starting from a calibrated grade
for which the coefficients are known, the models can be continuously adapted to grades
with different MFI’s and material feeds that influence it. While we offer no definitive
“proof”, we believe this capability can reduce by an order of magnitude the complexity of
the data-base requirements that would otherwise be needed. Section 6 also illustrates the
recursive on-line method for parameter identification using classical techniques, and
provides a brief summary of experimental results on various grades of GE’s Noryl (PPO)
material.
Inferential Sensing and Estimation
Using the models and adaptive identification above, the goal of inferential sensing of
viscosity is shown in Section 7 to be a straightforward extension using a linearized
viscosity relation from Section 5. Using data from both our 25 mm research extruder and
a 120 mm production facility for Noryl, we showed that viscosity predictions within +/10% or better can be obtained. The significance of this approach in comparison to
traditional discrete lab quality control checks typically good for 2% accuracy, is that a
continuous audit of viscosity results. Task 3 on leverages this capability in developing
diagnostic algorithms.
DE-FC02-99-CH10972
17
I nt e llige nt Ex t rude r for Polym e r Com pounding
Fast Response Torque Estimation
The team had proposed using “high frequency” components of shaft reaction torque as a
measurement for use in estimation / diagnostics of screw condition, proper fill conditions,
surge behavior etc. High frequency in this context means torque variation at or faster than
the per-revolution rate of the screw. That such data would be informative and valuable in
monitoring and diagnostics was conjecture based on preliminary experimental studies by
one of the team members. Torque signals provided in the control electronics of most of
the installed base of extruders is heavily filtered due to the noise that would be present,
and because it is not required for proper drive operation. It was therefore believed that
this could not provide information of interest at or near the frequencies that result from
torque fluctuations on a per-revolution basis. Direct torque measurement with high
bandwidth was found to be technically possible using a variety of strain-gauge or in-line
shaft-strain devices, but not practical or cost effective for use in retrofit in a production
environment, so this sub-task was initially abandoned. Late in the program, a concept was
proposed using an adaptation of ideas used in conventional AC drive controls. A small
effort was devoted to show the potential of this technique, using an “observer” torque
estimation algorithm, optimized to provide high frequency torques. The results of a
design and detailed circuit simulation study (without the details of the screw load) are
provided in Appendix B. Observer based torque estimation requires measurement of
certain voltages and currents (or power) in the machine, but it is feasible and inexpensive
to add such sensors, and a CPU to do the required signal processing, even to existing
drives. Since AC drives are used in a large percentage of the installed base and most new
systems (particularly large machines), we believe this approach has merit for
investigation in future studies. Since we did not have time or resources to go beyond the
basic simulations studies, we can make no claim to benefits that would derive from this
estimation scheme, but believe it merits experimental validation in future studies.
3.5 Task 5: Inferential Control System
Objective: Develop and demonstrate a computer control system that integrates all
the diagnostic and estimation algorithms with a control system that allows
automatic on-line recovery to correctable faults.
Accomplishments
Section 9.1 summarizes the philosophy and approach to control design for inferential
adaptive control of viscosity. The methodology integrates all elements developed during
the program, i.e. to monitor viscosity with an estimator, to detect when an upset occurs,
to classify the root cause of the upset, and initiate corrective action, all in a timely manner
to minimize out of spec material. In Section 9.2, the details of design are presented with
consideration to key tradeoffs which must be selected by the designer. The resulting
control requires either sufficient natural process disturbance presence or introduction of
small set-point perturbations (well within spec limits) to be able to reliably isolate root
cause of the disturbance and to then take the right corrective action. A key benefit of this
approach is shown in Section 9.3 to be the ability to reliably change the operating set
DE-FC02-99-CH10972
18
I nt e llige nt Ex t rude r for Polym e r Com pounding
point. The significance of this is allowing a production operator to change grade on the
fly, with the adaptive control re-tuning the models so that both control performance (e.g.
stability) and detection algorithms follow the new set point automatically. This is one of
the key benefits of the integrated Intelligent Extruder approach when viewed from a
production point of view. Although our closed-loop feedback controller was only
demonstrated on the research extruder, our success in proving scalability with the
modeling and diagnostics (see Task 7 and the results in Section 10 ) gives us confidence
that adaptive control will scale similarly.
3.6 Task 6: Control Platform and Experimental Extruder System
Objective: Implement and demonstrate the diagnostic and control systems developed
in Tasks 1-5 on a research scale extruder.
Accomplishments
Section 4 of this report describes the 25-mm research extruder used throughout this
program. Using a 30HP drive, the Werner & Pfleiderer ZSK-25 is capable of producing
about 50-100 lb/hr compared to commercial production rates of 2000 lb/hr and up. Using
identical but scaled screw geometry, however, it is possible to correctly compound most
polymer materials in the same manner as they will produced on the big machines.
(Scaling is addressed in Task 7). This section of the report also identifies the lab test
procedures used for measuring the actual material properties in the intelligent extruder
validation. To prototype the various diagnostic and control strategies developed in the
program, all logic was first implemented in well known and widely Matlab and/or
Simulink, a high level analysis package by The Mathworks Inc. Using the “Real-Time
Workshop” capability, it was possible to compile the various algorithms into executable
software that would run real-time on a special target hardware system called D-Space.
The D-Space package offered extensive data acquisition and user interface prototyping
capability making interfacing and control of the extruder particularly easy. While this
platform offers far greater computer power than needed to implement these algorithms on
a commercial product, the software modules make rapid prototyping easy and all the code
is in a C-language format that can be readily ported to other platforms, e.g. high end
PLCs or plant distributed control systems that support C. Further details on the control
algorithms developed and how they were implemented on the extruder control platform
are described in Section 9 of this report. The software developed in this program is for
research purposes only and is not available as a commercial product. Researchers and
other potential users wishing to replicate the software should contact the principal
investigator or GE Industrial Systems for more information.
3.7 Task 7: Production Scale Demonstration and Validation
Objective: Demonstrate that Intelligent Extruder diagnostics and control concepts
developed on the lab scale system are extensible to production scale operations.
DE-FC02-99-CH10972
19
I nt e llige nt Ex t rude r for Polym e r Com pounding
Since the majority of the experimental work in this program was carried out on a small
laboratory research extruder, a major question addressed in this task is: do the concepts
work on extruders at more typical production rates which are 20-100x the production rate
of the ZSK-25 lab system? As described in Section 10 of the report, data was obtained
from a 120 mm extruder system at GE Plastics making the same Noryl grade used
throughout our lab testing. The system was operating at approximately 2000 lb/hr, a
typical production rate in the Selkirk, N.Y. facility. This was special test in which
perturbations in feed rates were allowed (still keeping product within spec). Results show
that the simplified modeling and adaptive parameter identification framework scales to
predict viscosity within +/- 4% on the production rate machine. While it was not
permitted to perturb the machine out of spec to produce faults, simulation with the model
suggests identical qualitative behavior and that therefore all the proposed diagnostics
should perform similar to the research extruder. The extruder line where tests were
conducted did not have a means to close the loop on extruder feeds so it was not possible
to demonstrate the closed loop control capability as planned within the original scope of
this task. However, given that the modeling and adaptive identification methodology
appeared to scale well, we are confident that production scale closed loop correction
should be feasible to accommodate feeder upsets and base resin property shifts.
3.8 Task 8: Commercialization Plan
Objective: Develop a commercialization plan to transition the research results of the
Intelligent Extruder program to a product and/or service offering for the polymer
compounding industry.
Accomplishments
Section 12 of this report provides an overview of the commercialization plan developed
by the team to market the Intelligent Extruder as a value add product/service offering to
potential customers in the polymer industry. The size of the market for potential retrofits
and new installations is segmented. Alternative sales channels are identified. Market
forces driving the need for such offerings are identified, leading to a summary
opportunity fishbone for both polymer and non-polymer applications where Intelligent
Extruder ideas are applicable. Based on the expected benefit (see Section 11) of
approximately 2% in first pass yield, it was anticipated that an attractive value
proposition could be presented to customers through a services offering for new and
retrofit markets. Over the course of the program, three potential customers were
approached who possessed large numbers of production extruders, and two proposals
were submitted which included various elements of the Intelligent Extruder system
concept. While the nature and details of these commercial discussions are proprietary, as
of the date of this report submission, no sales or implementations have been
consummated.
DE-FC02-99-CH10972
20
I nt e llige nt Ex t rude r for Polym e r Com pounding
3.9 Publications and Patents
This research has resulted in one accepted paper in a refereed conference, with two other
papers planned for submission in 2003.
Alper Eker, Aditya Kumar and Paul K. Houpt, “A Model Based approach for an
Intelligent Extruder,” 2003 IEEE Conference on Control Applications, June 2003.
One patent disclosure has been filed as part of the program,
“Model based estimation diagnostics and control of an extruder,” Docket filing RD30498 (GE Internal reference), April 12, 2002
4 Experimental Extruder Setup for Demonstrations
4.1 WP 25mm Lab Extruder Description
In this section, we describe the extruder setup used at GE Global Research to perform the
experiment runs during the various phases of inferential estimation, diagnostics and
control. The setup used consisted of a WP ZSK-25mm twin-screw extruder, two K-Tron
loss-of-weight feeders with K-Tron feeder controllers and a high-performance data
acquisition system from IOtech. Figure 4 shows the ZSK-25mm extruder used which was
capable of maximum screw speed of 1200rpm, maximum throughputs of 100 lb/hr with a
maximum torque of 164 Nm. The extruder was powered by a 30HP GE Innovation drive,
which provided a measurement of total screw torque (estimated from motor current and
voltage) and screw speed (measured using a high-resolution optical encoder). The
extruder had six thermocouples along the barrel length and heating elements to control
the temperatures in the corresponding barrel zones. In addition to these, the extruder had
a thermocouple and pressure probe in the die zone to measure the melt product
temperature and the die pressure. Finally, the K-Tron feeder controllers had provision to
remotely command a desired set-point for the feed-rates and provided the measurement
of the actual feed-rates estimated internally by monitoring the loss of weight in the feeder
hoppers.
DE-FC02-99-CH10972
21
I nt e llige nt Ex t rude r for Polym e r Com pounding
Figure 4: WP ZSK-25mm twin-screw extruder used at GE Global Research for
experiment runs
4.2 Data Acquisition and Monitoring
The signals for the machine variables were recorded using an IOTech data acquisition
system capable of recording up to 24 channels. The torque, screw speed and die pressure
measurements were available as voltage signals, which could be directly interfaced with
the IOTech equipment. Also, the IOTech equipment had extensions to directly hook up
with thermocouple measurements. However, the feed-rate signals provided by the K-Tron
feeder controllers were in frequency, which had to be converted to voltage signals using
frequency to voltage converters before interfacing with the IOtech equipment. For our
experiments we recorded the main signals, i.e. torque, die pressure, two feed-rates, screw
speed and melt temperature, and for the most part they were recorded at a sampling rate
of 10Hz-1kHz and later sub-sampled through software as required. The collected data
was analyzed in Matlab and the algorithm development for modeling, inferential
estimation, diagnostics and control was carried out using Matlab/Simulink. For analyzing
the performance of the estimated viscosity using the measured extruder signals in
comparison to lab measurements, we collected samples of the product pellets at several
steady state operating conditions and measured their viscosity in the lab using two main
techniques shown in Figure 5.
DE-FC02-99-CH10972
22
I nt e llige nt Ex t rude r for Polym e r Com pounding
The capillary rheometer is commonly used to measure the static viscosity of the molten
polymer at medium to high shear rates. On the other hand, the RDS rheometer measures
the dynamic viscosity of the molten polymer subjected to oscillatory shear between two
discs at low shear rates. The initial experiment runs with polycarbonate mixtures used
both capillary and RDS rheometers, while in the later runs with Noryl (blend of
polyphenylene oxide (PPO) and polystyrene (PS) obtained from GEP Selkirk), we
exclusively used the capillary rheometer.
Pressure
Reservoir
Melt
Capillary
Die
Flow Rate
(a) Capillary rheometer
(b) RDS rheometer
Figure 5: Schematic representation of capillary and RDS rheometers
4.3 D-Space Implementation for Closed-Loop Experiments
The final implementation of the developed algorithms for estimation and closed-loop
control was done using D-Space, a high-end data acquisition and controls platform that
readily allows implementing algorithms developed in Matlab/Simulink. In addition to
DE-FC02-99-CH10972
23
I nt e llige nt Ex t rude r for Polym e r Com pounding
measuring the regular signals from the extruder, for the closed-loop control runs we also
needed the capability to automatically adjust the feed-rates. This was achieved using two
voltage to frequency converters that converted the voltage output signals from D-Space to
corresponding frequency signals used by the K-Tron feeder controllers to adjust the feedrate set-points. Figure 6 shows a screen shot of the GUI interface built in D-Space for online data acquisition, monitoring and closed-loop control. The interface allowed
monitoring measured values of raw material feed-rates, measured and predicted (by
model) values of die pressure and torque, estimated values of viscosity used in control
and other parameters used in diagnostics, all in real time.
Figure 6: GUI interface for estimation and controls algorithms using D-Space.
In addition to experiments performed at GE Global Research on the ZSK-25mm extruder,
we also obtained and analyzed data from production-scale extruders – 120mm, 2000lb/hr
throughput – at GEP Selkirk.
DE-FC02-99-CH10972
24
I nt e llige nt Ex t rude r for Polym e r Com pounding
5 Extruder Modeling for Estimation, Diagnostics and
Control
We adopted a model-based approach to achieve the estimation, diagnostics and control
objectives in a unified framework. There are several types of models that one can develop
using measured input-output data, e.g. static correlations for steady state relationships or
linear dynamic input/output dynamic models identified using standard identification
techniques. However, models obtained by these methods are often sensitive to the
particular data set and are not readily generalized. Moreover, they often lack any insight
into the physical process itself. On the other hand, first principles physics for melt flows
in extruders can be too complicated. A dynamic model which captures only enough of the
behavior to enable proposed methods is the goal.
The developed model is a representation of the physical process of extrusion that captures
the dynamic effect of common process inputs, e.g. raw material feed-rates and screw
speed, on measured process outputs, e.g. total screw torque and die pressure, without
getting into unnecessary details of the actual screw geometry and detailed material flow
characteristics. In particular, we used the work of Gao et al [11,12] on steady state
models for residence time distribution (RTD) in extruders as a starting point.
5.1 Process Description & Modeling
5.1.1 Process Description
Consider a typical extruder setup in Figure 7, which consists of the main drive, the
extruder barrel with one or two (co- or counter-rotating) screws and feeders (screw or
belt) for raw materials. Two or more raw materials (typically pre-blended with
appropriate additives) are fed to the extruder at controlled feed-rates and mixed and
melted in the extruder via the rotating screws with specifically designed conveying/
mixing/ kneading elements to produce the final molten product that is extruded at the end
as strands through holes in a die plate. The molten strands are then typically cooled and
solidified in a water bath and finally chopped into pellets for packaging and shipping as
final product. While typically some heat for the melting of the solid raw materials is
added to the extruder barrel externally through heating elements along the extruder barrel
length, most of the heat required for melting the raw materials is provided by friction
from the turning screws. This is especially true for the large industrial production-scale
extruders. Moreover, the extruder geometry, specifically the individual screw elements
and their sequence, varies from one application to another. In a typical plant, however,
the extruder geometry is often optimized and fixed for a wide grade of products, and
changed only occasionally for maintenance or when changing to very different product
grades.
DE-FC02-99-CH10972
25
I nt e llige nt Ex t rude r for Polym e r Com pounding
Figure 7: A typical extruder setup
5.1.2 Physics-Based Lumped Model
We start with a dynamic model that describes the dominant characteristics associated
with the mixing of the raw materials. To this end we develop a lumped two-section
dynamic mixing model that captures the effect of the inputs to the process (raw material
feed-rates and screw speed) on the measured process outputs ( total screw torque and die
pressure: the pressure developed prior to the die plate as the molten product is stranded
into the water bath).
Feed rates Q1 Q2
partially fille d
section 1
completely filled
section 2
Speed (N)
Die P (DP)
Torque (T)
M1 , x1
M2 , x2
Qo , , xo
viscosity (µ)
Figure 8: Schematic representation of extrusion process
Consider the schematic representation of a typical extrusion process shown in Figure 8.
For simplicity here, consider extruders with two key raw materials fed at rates Q1 and Q2.
In our experiments we worked with the NORYL resins from GE plastics which is
produced from polyphenylene oxide (PPO) and polystyrene (PS), but the methodology
can be generalized to multiple raw material feeds. The operating conditions of an
extruder are typically characterized by the combination of total throughput Q=Q1+Q2 and
screw speed N. Capacity of the machine depends on the screw design, material, and drive
torque power capability. As process outputs we measure the total shaft torque T and the
die pressure DP, which vary as a function of the operating conditions. Leveraging the
work of Gao et al (1999, 2000) [1,2] for steady state RTD, we consider the extruder with
two distinct sections during regular operation - a completely filled section (mixing,
kneading) and a partially filled section (conveying). Note that in an actual extruder there
are multiple conveying, mixing and kneading blocks and hence the partially/completely
filled sections may be interspersed. It will be shown that for the purpose of capturing the
overall dominant dynamics for use in estimation, diagnostics and control, it suffices to
assume two “equivalent partially” filled and completely filled zones into the respective
sections.
DE-FC02-99-CH10972
26
I nt e llige nt Ex t rude r for Polym e r Com pounding
5.1.3 Dynamic Model for Internal Holdup and Compositions
Under steady state operating conditions with a specific total feed-rate (throughput)
Q=Q1+Q2, and screw speed N, the total material holdup M1 and M2 in the partially and
completely filled sections, respectively, are given by
Eq 1
M1 = B
Q
,
N
M2 = A
where the ratio Q/N is referred to as the specific throughput and the parameters A, B are
related to the maximum capacities of the completely filled and partially filled sections,
respectively, depending on the specific screw design/geometry (see Gao et al (1999,
2000) [1,2] for more details). While the holdup M2 in the completely filled section is
constant, the holdup M1 in the partially filled section varies with the operating conditions,
specifically the ratio Q/N. In particular, the transient variation in the holdup M1 due to
changes in total feed-rate Q and screw speed N is described by the total material balance:
Eq 2
dM 1
= Q - Q1o
dt
In the above equation, the total inlet feed-rate to this section (from the feeders) is Q while
the total outlet mass flow rate, denoted by Q1o, varies with the operating conditions, in
particular the fill fraction φ (i.e. the fraction of the total void volume filled with the
material holdup) and the screw speed N. More specifically, the maximum flow capacity
of this section Q1fc corresponding to the maximum filled capacity M1fc (based on the void
volume from screw geometry) is proportional to the screw speed N, i.e., Q1fc = k N with
the proportionality constant k depending on the screw design/geometry. During regular
operation, when this section is only partially filled and the fill fraction is φ =M1 / M1fc ,
(0 < φ < 1), the total outlet mass flow rate is given by
Q 1o = k φ N
Eq 3
M 1
= k
M
1 fc
M 1N
=
B
N
where B=M1fc / k is a parameter that depends only on the screw design/geometry.
Combining equations Eq 2 and Eq 3 gives the dynamic mass balance relation for the
holdup M1:
Eq 4
M N
dM 1
=Q- 1
B
dt
Note that at steady state, the inlet and outlet mass flow rates are equal, i.e. Q=Q1o, and
the dynamic material balance in Eq 4 reduces to the steady state version: M1 = BQ/N. In
contrast with the partially filled section, the total holdup M2 in the filled section is
constant (since the void volume is filled to maximum capacity). Furthermore, the outlet
DE-FC02-99-CH10972
27
I nt e llige nt Ex t rude r for Polym e r Com pounding
flow rate from this filled section is always the same as the inlet flow rate, which in turn is
the same as the outlet flow rate from the partially filled section, i.e. Q1o.
Furthermore, in addition to the total material balance, we also need to capture variations
in composition of this material holdup to be able to predict the final product composition
as a function of the operating conditions (feed-rates & screw speed), which has a direct
bearing on the product properties, e.g. viscosity. To this end, we denote the weight
fraction of PPO in the material holdup in the partially and completely filled sections by x1
and x2, respectively. At steady state, these compositions are the same and are determined
solely by the feed rates Q1 and Q2, i.e.
x1 = x2 =
Eq 5
Q1
Q1 + Q2
Similar to the dynamic total mass balance, we also need to model the transient variations
in the compositions x1 and x2 in the partially and completely filled sections, respectively.
The composition in these sections changes due to mixing of the two raw materials in the
respective sections. More specifically, as the inlet feed-rates are changed thereby
changing the raw material composition at the extruder inlet, this change in composition at
the inlet propagates down the length of the extruder depending on the degree of mixing in
various sections. The detailed mixing mechanisms are governed by the screw design for
the various conveying, mixing, kneading sections and are too complex to model. For our
purposes, we seek a simple parameterized mixing model where the parameters can be fit
with measured input/output data to describe the overall effect of the mixing in the
partially and completely filled sections. This is achieved by modeling the level of mixing
in the individual sections through a combination of delay and recycle. Figure 9 shows
schematically this representation in the two sections.
Feedrates Q1 Q2
T
Speed N
partially filled
section 1
completely filled
section 2
M 1 , x1
Qo , xo
(µ)
M 2 , x2
M 1 , x1
R1*Q1o
Q
xi
DP
R2*Q2o
Q1o
x1
M 2 , x2
Q2o
x2
Figure 9: Schematic representation of mixing in partially and completely filled
sections.
For instance, the mixing in the first section that governs relation between the composition
x1 and the inlet composition xi, is governed by the delay depending on the ratio M1/Q1o
and the recycle ratio R1. The actual level of mixing can be captured by adjusting the
recycle ratio R1 between the extreme limits of 0 (no mixing) and infinite – in practice a
DE-FC02-99-CH10972
28
I nt e llige nt Ex t rude r for Polym e r Com pounding
large value (perfect mixing). Similarly, the level of mixing in the second section is
captured by the recycle ratio R2. The overall model for the transient behavior of the
compositions in the two sections is given in a compact form in the Laplace domain:
x1 ( s ) =
Qe − t d 1 s
x i ( s ), where
Q + R 1 * Q 1 (1 − e − t d 1 s )
x2 (s) =
e −td 2 s
x 1 ( s ),
1 + R 2 * (1 − e − t d 2 s )
Eq 6
wher e
M1
Q 1o
td1 =
td 2 =
M2
Q 2o
For a more detailed derivation of the above model, see APPENDIX-A Extruder Dynamic
Models.
5.1.4 Torque, Die Pressure and Viscosity Relations
The above equations (Eq 4 and Eq 6) describe the dynamics for the material holdup M1,
M2 (constant) and the compositions x1, x2. However, these internal state variables are not
measured on-line and need to be related to the output variables that are measured, namely
torque T and die pressure DP. The overall shaft torque arises from the combination of the
resistive torque in the individual conveying, mixing, kneading sections and a detailed
physics-based model involving the details of the screw design would be too complex. We
seek to develop a simple overall relationship for the total shaft torque in the following
general form:
Eq 7
T = α 0 + α 1 M 1 (α 2 + N ) x 1 + α 3 M 2 Nx 2
The above expression for torque has three key terms, the offset and the two contributions
from the partially and completely filled sections, respectively. The latter two terms
depend on the respective holdups and compositions and the screw speed. We tested the
validity of the above relationship for torque, using multiple measurements of torque at
various steady state operating conditions. At steady state, using the corresponding steady
state relations for the holdups, the above relation for torque reduces to the following
relation:
T = α 0 + α 1α 2 B
Eq 8
= c 0 + c1
Q
x i + α 3 ANx i + α 1 BQx i
N
Q
x i + c 2 Nx i + c 3 Qx i
N
We tested the validity of this equivalent steady state relation using measured input-output
data for feed-rates, screw speed and torque at various steady state conditions over
multiple days.
DE-FC02-99-CH10972
29
I nt e llige nt Ex t rude r for Polym e r Com pounding
Figure 10: Input-output relation for torque at steady state
Figure 10 shows the comparison of the measured torque and the model fit using Eq 8 for
45 steady state points obtained on four separate days. In particular, we fit the model
parameters ci, or equivalently αi, using the data from the first 25 points and tested its
validity against the last 20 points. Clearly, the model validates very well against the
measured torque data with an overall R2 value of 93.4%.
Similarly, a relation for die pressure DP as a function of the process variables is obtained
from physics using the laminar flow relation for pressure drop in a circular pipe (treating
the die plate holes as an effective short pipe) for the molten product with a viscosity µ
flowing through the die plate at a rate Qo (for a detailed description of die pressure model
see Appendix Section 2). It can be seen that the key parameters that affect the die
pressure DP are the product throughput and viscosity.
DE-FC02-99-CH10972
30
I nt e llige nt Ex t rude r for Polym e r Com pounding
Figure 11: Comparison of measured viscosity variation with PPO fraction and
model fit.
In general, the product viscosity at the die depends nonlinearly on the corresponding
product composition, temperature and shear rate. For the NORYL product we tested, the
viscosity of the product has a quadratic dependence on the weight fraction of PPO
content. In general, the dependence of viscosity on composition may be described by a
higher order polynomial. Figure 11 shows the plot of viscosity measured using a capillary
rheometer for 45 samples collected at steady state conditions over 4 days with a wide
composition range, and the fit obtained with a 2nd order equation using data from first two
days and validating against data from the last two days. Clearly, the 2nd order equation for
viscosity as a function of the product PPO weight fraction has a good fit with an R2 value
of 96.4% and maximum error between measured and fit values less than 8% (1σ − 3%).
The dependence of viscosity on shear rate, is typically given by a power law, while its
dependence on temperature is governed by an Arrhenius-type exponential function. For
DE-FC02-99-CH10972
31
I nt e llige nt Ex t rude r for Polym e r Com pounding
our purposes, we approximate the nonlinear dependence of viscosity on composition,
shear rate and temperature with a linearized relation:
Eq 9
µ = µ o + µ1 ( xo − xo ) − µ 2 (Qo − Qo ) − µ 3 (To − To )
The above relation describes the variation in the product viscosity around the nominal
operating conditions xo , Qo , To ; here To denotes the melt temperature measured at the die.
The above linearized relation for viscosity yields the following relationship for die
pressure:
Eq 10
DP = kQo [ µ o + µ 1 ( xo − xo ) − µ 2 (Qo − Qo ) − µ 3 (To − To )
= β 1Qo + β 2 Qo ∆xo − β 3 Qo ∆Qo − β 4 Qo ∆T
The dynamic material holdup and composition relations and the relations for torque and
die pressure comprise the overall dynamic model for the extrusion process, relating
changes in the process inputs (feed-rates and screw speed) to the measured output
variables (torque and die pressure)
Dynamic Process Model:
dM 1
= Q - Q1o
dt
M2 = A
Eq 11
x1 ( s ) =
(Q = Q1 + Q2 ,
Q1o =
Qe −td 1s
xi ( s ) ,
Q + R1 * Q1o (1 − e −td 1s )
( xi =
e −t d 2 s
x2 ( s) =
x1 ( s ),
1 + R2 * (1 − e −td 2 s )
( td 2 =
M1N
)
B
Q1
,
Q1 + Q2
t d1 =
M1
)
Q1o (1 + R1 )
M2
)
Q1o
T = α 0 + α 1 M 1 (α 2 + N ) x1 + α 3 M 2 Nx 2
DP = β 1Qo + β 2 Qo ∆xo − β 3Qo ∆Qo − β 4 Qo ∆T
The above model is a simple low-order lumped model in the so-called state-space form
with several input, state, output variables and parameters listed below.
Inputs
Q1 : feed-rate of PPO
Q2 : feed-rate of PS
N : screw speed
Qi : total feed-rate = Q1 + Q2
xi : weight fraction PPO at inlet = Q1 /Qi
States
M1 : mass holdup in partially filled section of screw
M2 : mass holdup in completely filled section of screw
x1 : wt. fraction PPO in partially filled section
x2 : wt. fraction PPO in completely filled section
DE-FC02-99-CH10972
32
I nt e llige nt Ex t rude r for Polym e r Com pounding
Outputs
Qo : outlet mass flow rate
xo : PPO fraction at outlet = x2
T : total torque
DP : die pressure
Parameters
A,B : extruder geometry-dependent parameters for holdups
R1, R2 : recycle ratios to capture mixing in partially/completely filled sections
α0-α3 : parameters in torque relation
β1-β4 : parameters in die pressure relation
5.1.5 Dynamic Model Parameters
The parameters in the dynamic model in Eq 11 will depend on the specific extruder
geometry and the product application and need to be identified from measured
input/output data. However, the task of identifying these parameters is simplified by
observing that the parameters can be categorized into two sets:
1. the first set consisting of the parameters A, B, R1 and R2 depend on the specific
screw geometry and will be invariant once the extruder screw geometry is fixed.
2. the second set consisting of the parameters αi and βi depend on the particular
process conditions and will change from one product grade (family) to another
and these parameters need to be identified depending on the operating conditions.
An initial value of all the parameters can be obtained through an off-line least squares fit
using the measured input-output data from an initial calibration experiment. Thereafter,
we need to update only the process-dependent parameters αi and βi, while the machinedependent parameters A, B, R1 and R2 are kept constant as long as the screw geometry
remains unchanged. We will address the on-line identification of the process-dependent
set of parameters αi and βi in a later on-line identification section. In this section, we
present the results obtained by fitting the parameters to match the model input-output
predictions with on-line measurements obtained during a calibration run.
Figure 12 and Figure 13 show the results of the initial off-line least squares fit comparing
the model predictions for the torque and die pressure (shown in red) with the on-line
measurements (shown in blue) obtained during an experiment run on 05/08/2001 on the
25mm extruder with NORYL polymer using nominal raw materials (PPO IV 0.46) and
nominal composition (PPO fraction 0.52). Clearly, the model predictions match very well
with the on-line measurements with R2 values of 91% and 93% for die pressure and
torque, respectively.
DE-FC02-99-CH10972
33
Speed (rpm)
(lb/hr)
Scaled die pressure
I nt e llige nt Ex t rude r for Polym e r Com pounding
error histogram
Speed (rpm)
(lb/hr)
Scaled torque
Figure 12: Comparison of Die Pressure measurement with model prediction
error histogram
Figure 13: Comparison of torque measurement with model prediction
DE-FC02-99-CH10972
34
I nt e llige nt Ex t rude r for Polym e r Com pounding
5.1.6 Comparison of Model Predictions
Despite the simplicity of the model it suffices for our purposes of predicting the
responses in key measured outputs like total screw torque and die pressure as a function
of the process inputs like feed-rates and screw speed and other process parameters. In
particular, the model captures both the steady state and transient characteristics for the
two measured outputs very well.
The transient response for total screw torque T shows an inverse response with respect to
screw speed changes (see Figure 14). This occurs due to the fact that starting at some
steady state as the screw speed is increased, the torque initially rises proportionally (see
Eq-A 9). However, due to the increased speed more material is withdrawn from the
partially filled section until the holdup M1 in this section reaches a new lower steady state
value – since total feed-rate Q is unchanged, the new steady state value has to be lower
such that the outlet flow rate Q1o =M1 N / B is the same as before the increase in screw
speed N. Consequently, as a result of reduced holdup M1, the total torque eventually
reduces after the initial increase.
The die pressure DP shows an inverse response with respect to screw speed N (see Figure
15). This occurs due to the fact that starting from a steady state, as the screw speed N is
increased, the outlet flow rate Q1o from section 1 and hence from section 2 increases
thereby leading to an initial increase in DP. But, again since the overall feed-rate has not
been changed, the outlet flow rate will reach a final steady state value same as before the
increase in the screw speed. However, the increased speed generates more heat due to
viscous dissipation thereby increasing the melt temperature, which reduces the product
viscosity and thus reducing the die pressure DP.
The die pressure DP shows an inverse response with respect to the PS feed-rate (see
Figure 16). Initially, starting from a steady state, as the PS feed-rate is increased, the total
throughput and hence the outlet product flow rate increases thereby increasing the die
pressure. However, the increased PS feed-rate reduces the PPO weight fraction xi, which
eventually leads to a reduced PPO weight fraction x2o at the outlet which leads to a
reduced product viscosity and hence reduced die pressure DP. The die pressure doesn’t
exhibit such an inverse response with respect to PPO feed-rate since an increase in the
PPO feed-rate leads to an increased throughput and an increase in x2o, both of which
contribute to an increased die pressure DP.
DE-FC02-99-CH10972
35
I nt e llige nt Ex t rude r for Polym e r Com pounding
Inverse response
in Torque
t (s)
Speed (rpm)
(lb/hr)
Increase in
screw speed
(psi)
Figure 14: Inverse response in torque with respect to screw speed
Inverse response
in Die Pressure
Increase in
screw speed
Speed (rpm)
(lb/hr)
t (s)
Figure 15: Inverse response in die pressure with respect to screw speed
DE-FC02-99-CH10972
36
(psi)
I nt e llige nt Ex t rude r for Polym e r Com pounding
Inverse response
in Die Pressure
Increase in
PS feedrate
Speed (rpm)
(lb/hr)
t (s)
Figure 16: Inverse response in die pressure with respect to PS feed-rate
5.2 Extruder Modeling Summary
We have developed a simple, physics-based dynamic model for a typical extrusion
process that can be applied to a wide variety of extrusion applications to describe their
transient input-output behavior. We have demonstrated the application of the model on a
lab-scale 25mm extruder to describe the dynamic input-output behavior, specifically the
response in on-line measurements like torque and die pressure (readily measured in most
extruder applications) as a function of variations in extruder operating conditions. Scaleup of the model to an industrial production-scale extruder is also demonstrated in Section
10.
The model has several unknown parameters, which need to be identified for specific
applications based on on-line input-output measurements from experiments. However,
the identification of these parameters is greatly simplified by grouping the parameters
into machine-dependent and process condition-dependent sets. The former set of
parameters need to be identified only once for a given extruder geometry, while the latter
set of parameters will, in general, vary depending on the process conditions. This
variation of the parameters will be addressed through on-line identification and used for
the fault diagnostics and inferential estimation as described in later sections.
DE-FC02-99-CH10972
37
I nt e llige nt Ex t rude r for Polym e r Com pounding
6 On-line Parameter Identification
The dynamic process model described in the previous section involves unknown model
parameters that need to be identified from measured input-output data. These parameters
will vary for different extruders and process applications. The physics-based nature of the
model allows categorizing the parameters into two broad classes:
(i) parameters depending on the machine geometry (A, B, R1, R2), and
(ii) parameters depending on the material properties and process operating conditions
(αi, βi).
The unknown model parameters can be identified for a given extruder setup once using
experimental input-output data and a least-squares fit (see previous section). Thereafter,
the machine parameters are fixed for the specific extruder and screw geometry. In
contrast, the material parameters will in general vary from day to day, due to variations in
process conditions, raw materials etc., and need to be identified on-line.
For the on-line identification, the machine-dependent parameters (A, B, R1, R2) are fixed
at the values obtained by off-line identification, while the parameters αi in the torque
relation in Eq 7 and the parameters βi in the die pressure relation in Eq 10 are to be
identified on-line. In our experiment runs with NORYL PX5511 grade resin, we observed
that the nominal set of parameters αi (identified by off-line identification) captured the
transient variation in torque with process condition changes quite well, in-spite of
changes in raw material and composition. Moreover, the parameters αi are difficult to
interpret due to a lack of explicit relationship to physical parameters. On the other hand,
the parameters βi for die pressure have explicit relationship with physical parameters µi in
the viscosity relation (see Eq 9 and Eq 10), given by:
Eq 12
β 1 = kµ o
β 2 = kµ1
where k is a machine/product grade (family) dependent calibration parameter.
Furthermore, these parameters varied significantly under varying raw material and/or
composition variations as expected from the physics. So, we focused exclusively on the
on-line identification of βi to meet the objectives of diagnostics and estimation.
Equation Eq 10 has two important features for our purposes.
1. It relates process inputs and process output, die pressure, where the parameters βi
have a physical significance owing to their explicit relationship to the parameters
µi – we will exploit this to meet the objectives of estimation and diagnostics.
2. It is linear with respect to the parameters βi.
Having a relation linear with respect to parameters allows use of on-line recursive
adaptation techniques, with relatively low computational burden. To facilitate the online
identification of the parameters βi (or equivalently µi) we provided excitation to the
system via a pseudo-random binary sequence (PRBS) variation in the inputs (feed-rates
DE-FC02-99-CH10972
38
I nt e llige nt Ex t rude r for Polym e r Com pounding
and screw speed) and recorded the corresponding die pressure measurement. This
excitation of the process (referred to as persistent excitation) is necessary for correct
identification of the new parameters βi under changing raw material/ process conditions.
For the online identification of the parameters βi, we adopted the following well-known
recursive least-squares formulation in our approach, given by (see e.g. Ljung[18] Chapter
11):
θˆ(t) = θˆ(t − 1) +
Eq 13
P ( t − 1)
Eq 14
P(t − 2)φ(t − 1)
T
[
(t ) − φ(t − 1) θˆ(t − 1)]
y
1+ φ(t − 1)T P(t − 2)φ(t − 1)
P ( t − 2 )φ ( t − 1)φ ( t − 1) T P ( t − 2 )
= P (t − 2) +
1 + φ ( t − 1) T P ( t − 2 )φ ( t − 1)
where θ = [ β 0 m β 3 ]T denotes the parameters to be recursively identified,
φ (t ) = [1 Qo (t ) Qo (t ) xo (t ) Qo (t )∆T (t )]T denotes the coefficients of these parameters
in the die pressure relation in Eq 10 and P(t) is the parameter covariance matrix. The
covariance matrix P(t) is initialized with a pre-selected positive definite covariance
matrix Po reflecting the confidence in the initial estimates of the parameters.
Q1
Q2
Qo, , xo
(µ)
N
Inputs
(Q, N)
Extruder
Model
updated
β
predicted
outputs
Measured Outputs
(T, DP)
-
+
Parameter
Adaptation
Figure 17: Schematic representation of on-line parameter identification approach.
Figure 17 shows the schematic approach for the on-line parameter identification for the
extrusion process. In particular, the measured process inputs (feed-rates and screw speed)
are fed to the model and its prediction for the outputs (die pressure) are compared with
the on-line measurements to generate the residual error. Under nominal conditions, the
residual error will be normally distributed (due to noise) with a zero mean. However, if
the operating conditions change, e.g. change in raw material, feed composition, then the
model predictions with the nominal parameters βi will no longer match the measured
values, i.e. the residual error will no longer be zero mean or normally distributed. Under
DE-FC02-99-CH10972
39
I nt e llige nt Ex t rude r for Polym e r Com pounding
such a situation, the on-line identification of the unknown parameters βi will be started
with a pre-set sequence of persistently exciting variations in the inputs, PPO feed, PS
feed and screw speed that yield a well-conditioned information matrix formed with
vectors of φ (t ) = [1 Qo (t ) Qo (t ) xo (t ) Qo (t )∆T (t )]T in time (see [18] and Section 8
on fault diagnostics in this report for more detail).
6.1 On-Line Parameter Identification Results
Speed (rpm)
(lb/hr)
Scaled die pressure
We tested the on-line identification capability in multiple runs over several months for
variations in raw material and composition from nominal conditions. We show the results
of the model predictions compared with the on-line measurements for two representative
runs after the recursive parameter adaptation, one with change in raw materials and
another with a large change in composition (similar to a product grade change). In all
cases, the parameter adaptation started with parameter values initialized for nominal
conditions obtained by the least squares optimal fit obtained in Section 5.1.5.
error histogram
Figure 18: Comparison of model prediction after on-line adaptation vs
measurement for die pressure under composition change from nominal conditions
Figure 18 demonstrates the match between die pressure predictions of the extruder model
and measurements obtained in experiment run on 5/22/01 at GE GR for nominal raw
materials (0.46 IV PPO) but a significantly different composition (PPO fraction 0.35) –
this large change in composition corresponds to another product grade. During this run, a
PRBS variation in the inputs (PPO, PS feed-rates and screw speed) was used and the
parameters βi were adapted following the above recursive least squares method starting
from the nominal values obtained on 05/08/2001. Note that the die pressure during this
run (Figure 18) is distinctly lower than that in the nominal run on 05/08/2001 (Figure 12)
DE-FC02-99-CH10972
40
I nt e llige nt Ex t rude r for Polym e r Com pounding
Speed (rpm)
(lb/hr)
Scaled die pressure
- the die pressure is plotted after normalizing with respect to the nominal value of 450psi,
i.e. the normalized die pressure is 1.0 at nominal conditions. This is as expected, since the
lower PPO composition implies a lower product viscosity and hence lower die pressure.
Clearly, the parameters βi were adapted well to match the model predictions with the
measurements and capture the effect of composition change, with a modified R2 value of
90.5% and a normally distributed residual error between the measurement and the model
prediction.
error histogram
Figure 19: Comparison of model prediction after on-line adaptation and
measurement of die pressure under raw material change from nominal conditions
Similarly, we conducted another experiment on 8/27/01 at GE GR with the nominal
composition (PPO fraction 0.52) but with a different raw material. In particular, we used
a low IV PPO (0.33 IV) and conducted a similar PRBS experiment to facilitate the online identification of the parameters βi. Again, the lower IV PPO results in a product with
lower viscosity and hence lower die pressure than in the nominal run.
Figure 19 demonstrates the fit obtained after the on-line identification between die
pressure measurement and the model prediction with updated βi. Clearly, the parameters
were identified accurately to capture the effect of raw material change, with a modified
R2 value of 95.9% and a normally distributed residual error between the model prediction
and measurement.
6.2 On-line Parameter Identification Summary
Process dependent parameters αi and βi in the torque and die pressure relations will, in
general, change depending on changes in raw materials and feed composition. Under
DE-FC02-99-CH10972
41
I nt e llige nt Ex t rude r for Polym e r Com pounding
nominal conditions, the initial set of parameters obtained for a family of product grades
using the off-line least squares fit described in Section 5.1.5 will suffice for the model
predictions for torque and die pressure to match with the on-line measurements.
However, in the presence of deviations from nominal conditions, these parameters will
need to be adapted on-line. We applied recursive on-line adaptation techniques to track
changes in these parameters. For the NORYL grade polymers we studied, the nominal
parameters αi described the variations in torque despite changes in raw materials and
composition. So we focused exclusively on the adaptation of the parameters βi for die
pressure. The on-line identification of the parameters βi works very well in the presence
of changing process conditions. The on-line identification of the parameters βi to capture
effects of changes in raw material and/or composition will be exploited in the subsequent
sections for fault diagnostics and inferential sensing.
7 Inferential Sensing
In this section, methods are derived for the on-line inferential estimation of product
viscosity from on-line measurements of the inputs (feed-rates and screw speed) and the
outputs (die pressure). In a typical production environment, the product viscosity is
measured only off-line using samples from the finished product in a QA lab. Very often,
due to the limited resources and the large number of production lines and production
batches each day, only one sample is collected per batch for the QA lab analysis. This
leads to ineffective characterization of good/bad product batches: A batch of good
material that was out of spec only at the end of the run may be rejected, and bad material
that was good at the time of sampling may be passed with once per batch QA tests. Such
inefficient characterization of production quality leads to avoidable losses of material and
energy. The ability to monitor viscosity on-line is a significant leap that will enable quick
classification of good or bad product and enable corrective action, e.g., on-line closedloop control, to maintain products within specification limits and minimize waste
production. We will address the on-line estimation of product viscosity using the physicsbased modeling and adaptation framework of previous sections.
DE-FC02-99-CH10972
42
I nt e llige nt Ex t rude r for Polym e r Com pounding
Variability in composition / raw material
(unknown)
Variation in Torque / Die Pressure
(measured)
Measured Outputs
(T, DP)
Inputs
(Q, N)
variation in viscosity
Extruder
Model
predicted
outputs
-
+
updated
α, β
Parameter
Adaptation
Model adaptation and updated
online estimation of viscosity
Figure 20: Approach for model-based estimation of viscosity
Figure 20 shows the approach for online viscosity estimation from the measurements of
the process inputs and outputs. In particular, as the raw materials or feed compositions
change, thereby changing the viscosity, we obtain a continuous estimate for the viscosity
from on-line identification of the parameters β’s. Once the extruder model parameters β’s
are identified, they are used to estimate nominal viscosity µ0 and slope µ1 using a value
for k (see Eq 9, Eq 10 and Eq 12). The required value of k is obtained by comparing the
parameters β’s and the viscosity measured in the lab for samples collected during an
initial calibration run, and then fixed thereafter for all other runs. The product viscosity is
then estimated using the linearized viscosity relation (Eq 9) disregarding the shear rate
and temperature effects since these are fixed for the lab measurements and the product
quality specifications, i.e.
Eq 15
µ = µ o + µ1 ( xo − xo )
and the product composition at the extruder outlet xo=x2 obtained from the dynamic
model. The following figures depict the comparison of on-line viscosity estimates with
the off-line lab viscosity measurements of samples collected at multiple steady state
conditions with different compositions during each run. In all these comparison figures,
the magenta line represents the estimated value of viscosity and blue line represents the
off-line viscosity measurement in the lab.
7.1 Viscosity Estimation Results
Figure 21-Figure 26 that follow, show the viscosity estimation results for the multiple
runs with nominal raw materials and lower IV PPO over a large composition range
DE-FC02-99-CH10972
43
I nt e llige nt Ex t rude r for Polym e r Com pounding
spanning nominal compositions for several NORYL product grades at GEP Selkirk. The
estimation results shown by magenta lines are compared with the corresponding off-line
lab measurements of samples collected during steady state operating conditions in each
run.
Steady State viscosity comparison
% error
measurement
µ
estimate
Sample #
Sample #
Figure 21: Comparison of on-line viscosity estimation and off-line lab viscosity
measurements for experiment on 5/8/2001 with nominal (.46 IV) PPO blend
Steady State viscosity comparison
% error
measurement
µ
estimate
Sample #
Sample #
Figure 22: Comparison of on-line viscosity estimation and off-line lab viscosity
measurements for run on 6/25/2001 with nominal (.46 IV) PPO blend
DE-FC02-99-CH10972
44
I nt e llige nt Ex t rude r for Polym e r Com pounding
Steady State viscosity comparison
measurement
µ
estimate
Sample #
Sample #
Figure 23: Comparison of on-line viscosity estimation and off-line lab viscosity
measurements for run on 7/25/2001 with nominal (.46 IV) PPO blend
Steady State Viscosity Comparison
% error
estimate
µ
measurement
Sample #
Sample #
Figure 24: Comparison of on-line viscosity estimation and off-line lab viscosity
measurements for run on 8/27/2001 with low IV (.33 IV) PPO blend
DE-FC02-99-CH10972
45
I nt e llige nt Ex t rude r for Polym e r Com pounding
Steady State Viscosity Comparison
% error
estimate
µ
measurement
Sample #
Sample #
Figure 25: Comparison of on-line viscosity estimation and off-line lab viscosity
measurements for run on 10/11/2001 medium IV (50/50 mixture of .46 IV and .33
IV) PPO blend
Steady State Viscosity Comparison
% error
estimate
µ
measurement
Sample #
Figure 26: Comparison of on-line viscosity estimation and off-line lab viscosity
measurements for run on 11/14/2001 with medium IV (50/50 mixture of 0.33IV and
0.46 IV) PPO blend
DE-FC02-99-CH10972
46
I nt e llige nt Ex t rude r for Polym e r Com pounding
7.2 Viscosity Estimation Summary
Date of
Experiment
5/8/2001
No. of
samples
10
Raw Materials
used
Composition
(PPO fraction)
Estimation error
Nominal
0.42-0.61
-4 to +4%
6/25/2001
10
Nominal
0.43-0.62
-10 to +2%
7/25/2001
10
Nominal
0.43-0.62
-12 to +2%
8/27/2001
10
Low IV PPO
0.43-0.62
10/11/2001
4
0.43-0.6
11/14/2001
4
Medium IV
PPO
Medium IV
PPO
-6 to 0% (except 2
samples)
-6 to 0%
0.43-0.6
-8 to 0% (except 1
sample)
Table 3: Summary of viscosity estimation results
Table 3 above summarizes the results of model-based on-line viscosity estimation
compared to off-line lab measurements. Altogether, we collected about 50 samples
during steady state conditions in experiment runs conducted over several months with a
wide range of variations in product composition and raw materials spanning multiple
NORYL product grades. We compared the viscosity of these samples as measured in the
lab, using a capillary rheometer with those estimated by the model-based estimation. The
overall estimation results are very good with an error in the range of +/-10% of the
measured values, which is within the range of error of the off-line viscosity measurement
using capillary rheometer. The error is larger than 10% for specific samples (e.g. samples
4 & 8 on 8/27/2001 -Figure 24, and sample 2 on 11/14/2001 - Figure 26) that correspond
to very low PPO content and thus low viscosity. These samples correspond to the
maximum deviation from the nominal composition (PPO fraction 0.52) and the linear
approximation for viscosity as a function of composition used in the die pressure model
(see Eq 9 and Eq 10) becomes inaccurate under these extreme deviations, thus leading to
larger error. One way to alleviate these inaccuracies under these extreme composition
limits is to include higher order nonlinear terms in the dependence of viscosity on
composition, albeit at the expense of additional model parameters and increased
complexities in the recursive on-line parameter identification. A simpler approach would
be to linearize the viscosity model close to the extreme composition limits to minimize
the errors due to linearization.
DE-FC02-99-CH10972
47
I nt e llige nt Ex t rude r for Polym e r Com pounding
8 Extruder Diagnostics
In this section, we address the problem of fault diagnostics for the extruder using the
developed modeling and adaptation framework described in the previous sections. Figure
27 shows a schematic diagram of a typical extrusion process and various sources of
variability that cause product variability. The sources of variability include a wide array
of raw material, equipment and operator variations. Among these possible sources of
faults, the major ones that occur most frequently and are not readily detected by simple
practical means, and affect product quality are associated with variations in raw material
quality and feeder variations. Motivated by this, we will focus on the detection of raw
material and feeder variation as the main process faults, using available on-line
measurements.
• clogged feeders, die screen
• poor barrel temp regulation
• vent, die port blockage
Process Faults
Operator
Errors
• formulation
• setup
• QC errors
Raw
Material
Variation
• Resin MW
• Resin color
• Resin morphology
• SPD pellet
Feeders
Screw
Gear
Pelletizer
Product
Variability
Heater
Bath
Motor
Drive
Equipment Faults
•
•
•
•
MV
Composition
Color
mechanical
strength
• feeder drive, motor control
• dull pellet cutter
• screw shaft breakage
• screw wear
Figure 27: Schematic diagram of extrusion process and sources of product
variability
8.1 Problem and Approach
The problem of fault diagnostics entails the detection that a fault has occurred, and the
classification of the cause of the faults among multiple potential candidates. A standard
issues that arises in the detection and identification of faults is the trade-off between false
alarms, i.e. false declaration of faults in nominal conditions, and missed detects, i.e.
missing the occurrence of an actual fault. A tradeoff between is required between
acceptable limits of nuisance from false alarms and acceptable losses from missed fault
detection. Another important issue for fault detection is the latency or time elapsed from
DE-FC02-99-CH10972
48
I nt e llige nt Ex t rude r for Polym e r Com pounding
the occurrence of a fault to its detection, identification and correction to minimize the
waste production in this time interval.
Measurement, Model prediction
Figure 28 shows a schematic representation of the various phases during a typical fault
occurrence, detection, identification and correction. In phase 1, the process is initially
operating at nominal conditions and the model prediction using nominal model
parameters matches well with the on-line measurements (e.g. for die pressure). An
unknown fault then occurs (either abruptly or gradually over time) at the end of phase 1,
which leads to the on-line measurements deviating from the model prediction (using
nominal parameters). Depending on the type of fault, the difference between the
measurement and the model prediction (or the integral square of the difference—see e.g.
[17] for alternate approaches to detection and classification using comparison of model to
measurements) would exceed a threshold, indicating the occurrence of a fault (at the end
of phase 2). The source of fault is however not known at this stage. After the fault
detection, the process can be excited with changes in the inputs (feed-rates and screw
speed) and the model parameters adapted to match its prediction with the on-line
measurement, thereby obtaining a new set of model parameters reflecting the “faulted”
process conditions. At the end of this adaptation stage 3, the source of the fault is
uniquely classified and enables appropriate corrective control action in phase 4, which
may be either a physical action, e.g. recalibration of drifted feed-rate measurements, or an
automatic on-line closed-loop control correction using inferred viscosity based on the
adapted model parameters. Finally, the corrective control action brings the process back
within specification limits to achieve on-spec production in phase 5. The time duration of
the fault detection, fault identification and corrective control in phases 2, 3 and 4
determine the amount of off-spec waste produced during a fault. Note however, that the
detection of a fault at the end of phase 2 enables the possibility of physically diverting the
off-spec product during phases 3 and 4 to minimize the contamination of the finished
product.
On-line measurement
Model prediction
1
2
3
4
5
Time
Figure 28: Schematic description of fault detection and correction with distinct
phases – (1) initial nominal operation, (2) fault occurrence and detection – mismatch
between model and measurement, (3) model update – on-line model adaptation and
fault identification, (4) corrective control action for fault, (5) on-spec operation after
fault correction.
DE-FC02-99-CH10972
49
I nt e llige nt Ex t rude r for Polym e r Com pounding
The fault diagnostics clearly consists of two phases in the above description, the initial
detection phase where a fault (unknown) is declared to have occurred, and the second
classification phase, where a specific fault is identified, e.g. raw material change or
feeder bias.
For our purposes the two key on-line extruder output measurements are die pressure and
torque. During our experiments with NORYL, while both torque and die pressure were
sensitive to composition changes, only die pressure was sensitive to raw material
changes. Moreover, the changes in torque and die pressure were often correlated.
Motivated by these, we focused on die pressure for the diagnostics algorithm. Using die
pressure as the output signal, the detection of occurrence of a fault is addressed simply by
monitoring the difference between the measured die pressure and the value predicted by
the dynamic model as a function of measured inputs, i.e. feed-rates and screw speed.
Under nominal conditions this difference will be below a threshold – chosen based on the
noise characteristics of the signals ([17]). However, in the presence of a fault, i.e. feeder
bias or raw material change, the product viscosity will change, thereby changing the die
pressure. Thus, the measured die pressure will deviate from the model prediction using
the nominal parameters, and for a sufficiently large enough fault, the discrepancy
between the measured die pressure and the model prediction will exceed the threshold –
thereby declaring a fault occurrence. Clearly, the choice of such a threshold for the
mismatch between the model prediction and measured value depends on the signal to
noise ratio and affects the occurrence of false alarms and missed detects. Based on the
actual signal to noise characteristics and the normal operation spec-limits, an optimal
threshold can be chosen to optimize the trade-off between false alarms and missed
detects. Once a fault has been declared based on the discrepancy between the model
prediction and the measured die pressure, the second stage of fault classification, i.e. the
identification of the fault source will initiate.
There are several techniques available for fault identification, e.g. multi-model
hypothesis testing, data classification and correlation with fault/nominal signatures, as
summarized in Gertler for example [17].Multiple-model hypothesis testing approaches
the problem of fault detection by comparing the measured output with predictions from a
bank of models, one each for the nominal and all the fault conditions. With assumptions
of linear dynamic models and Gaussian noise, one can compute the a posteriori likelihood
that the data came from each model, and classification is as simple as picking the
maximum likelihood model. It relies critically in properly modeling all individual faults
and is computationally expensive. On the other hand, data-based classification and
correlation techniques rely on identification of distinct signatures in the measured data for
faults and nominal operation, which are often very problem-specific and difficult to
generalize.
The approach selected for fault classification derives from a variation on the classic
model based approach: to use the signature of the adaptively tracked process parameters
as they change over time.. A side benefit of continuous parameter adaptation, is the
ability to estimate the new viscosity after the occurrence of a fault, and enabling
automatic corrective action using closed-loop control where feasible (see Section 9 ).
DE-FC02-99-CH10972
50
I nt e llige nt Ex t rude r for Polym e r Com pounding
To facilitate the identification of the modified model parameters, the process is excited
through small test signals added to feed-rates and/or screw speeds. Recursive estimates of
the model parameters βi for die pressure are obtained using the methods in the previous
section. The adapted parameters βi, will in turn, enable the identification of the fault
source. Figure 29 depicts this approach to fault identification in more detail. The fault
identification depends on comparing the obtained set of parameters βi against a precalculated set of distinct signatures for nominal/fault conditions. Clearly, the generation
of generic fault signatures that are not problem specific is important for successful
application of this approach to a wide range of products/processes. We develop such
generic fault signatures that are not related to a specific process or product. While we
demonstrate the results of this approach as applied to the NORYL product, we emphasize
that the approach should extend easily to other products due to the generic physics-based
nature of the underlying model and fault signatures.
Q1
Q2
N
Unknown Variability
• raw material
• feeder bias
Qo, , xo
(µ)
Inputs
Extruder
Model
(Q, N)
Perform adaptation
when detect model
mismatch
updated
α, β
predicted
outputs
--
+
Measured Outputs
(T, DP)
Parameter
Adaptation
model parameter estimates
Identify source
of variability
Detection Algorithm
• match parameter change with fault signature
raw material variation
Nominal
feeder variation
Fault Signatures
• distinct parameter change (from nominal) for anomalous operation
• physics based - generically applicable
Figure 29: Schematic of fault detection approach
DE-FC02-99-CH10972
51
I nt e llige nt Ex t rude r for Polym e r Com pounding
8.2 Fault Diagnostics – Noryl Case Study
8.2.1 Fault Signatures
The underlying principle behind the developed approach for fault detection is that any
change observed in on-line identified parameters, β’s, would be due to changes in
operating conditions or in the raw material, assuming the screw geometry remains fixed
during the operation (see Eq-A 15). In particular, faults like raw material and feeder
variation affect the nominal viscosity µο and the slope µ1 of the viscosity with respect to
composition xo in distinct manner. We will exploit the relation of the identified β’s to the
physical parameters µ’s, in particular the relation of βο and β1 to µο and µ1, respectively
(see Eq 12) to identify and isolate these faults.
Note that from Eq 9, for a nominal shear rate or equivalently throughput Q o = Q o , and
nominal melt temperature T o = T o , the viscosity varies with the product composition xo
according to the linear relation in Eq 15 around the nominal composition x o . Any
changes in the raw materials PPO or PS or variations in PPO or PS feed-rates will affect
the viscosity of the product and consequently, the die pressure, which will be manifested
as distinct changes in the parameters µο and µ1. Figure 30 shows the plots for viscosity as
a function of composition under various raw material (PPO or PS for NORYL) changes
and the corresponding distinct combinations of increase/decrease in µο and µ1. These
variations in µo and µ1 form the basis for the distinct fault signatures.
• Change in PPO
∆µ1 < 0
∆µ0 > 0
viscosity
viscosity
∆µ0 < 0
Nominal
IV PPO
∆µ1 > 0
High IV
PPO
Low IV
PPO
PPO fraction
PPO fraction
• Change in PS
∆µ1 < 0
∆µ0 < 0
∆µ1 > 0
viscosity
viscosity
∆µ0 > 0
High IV
PS
Low IV
PS
PPO fraction
PPO fraction
Figure 30: Fault signatures for changes in raw material
DE-FC02-99-CH10972
52
I nt e llige nt Ex t rude r for Polym e r Com pounding
As seen in Figure 30 top left corner, if the raw material PPO changes from
nominal IV to lower IV, while the other raw material PS is nominal, the resulting curve
for variation in viscosity with respect to composition will be as shown in blue. In
particular, in comparison to the nominal case (shown in red) the viscosity will remain
unchanged at xo=0 (pure PS, which is nominal) while it will have the maximum reduction
at xo=1 (pure PPO). Thus, around the nominal composition xo , the new curve for
viscosity (blue) will have a reduced value µο as well as a reduced slope µ1, compared to
the values for original nominal curve (red). Similar logic applies to other cases of raw
material variations; for example, higher IV PPO, or lower IV PS, or higher IV PS as seen
in Figure 30, will yield a distinct set of signatures for changes in µο & µ1 corresponding
to raw material changes. On the other hand, for nominal raw materials, an unknown bias
in the PPO or PS feeders resulting in an increase/decrease of PPO fraction xo will also
affect the product viscosity and thus, die pressure.
Figure 31 shows the variation in µο & µ1 for “small” variations in PPO/PS feeders. For
instance as shown in the top left figure, a positive bias in the PPO feeder would imply
that the observed throughput Q will be greater than the actual throughput. Similarly, the
observed composition xo will also be greater than the actual. However, the observed die
pressure DP corresponds to the actual Q and xo. Thus, since both the actual throughput
and viscosity (as function of composition xo) are lower, the observed die pressure DP will
be lower than expected with respect to the observed Q and xo. This will be estimated as a
lower viscosity at nominal conditions and thus lead to the signature: ∆µo < 0. Note that
for a “small” bias, the local slope will be unaffected i.e. ∆µ1 = 0. Similarly, the signature
for negative bias in PPO feeder would correspond to the signature: ∆µo > 0, ∆µ1 = 0. In
contrast to PPO feeder bias, a positive bias in the PS feeder would imply that the
observed composition xo, will be less than the actual while the observed throughput Q
will be greater than the actual. These changes in the feed-rate and feed composition have
opposite effects on the expected change in die pressure (see Eq 10) and effectively reduce
the net effect on die pressure, thereby making the detection of PS feeder bias more
difficult. In our experiments, we observed that the effect of composition change (which
affects viscosity and hence die pressure) is more dominant than the effect of throughput
change, i.e. the net effect of positive PS feeder bias is an increase in the die pressure DP.
This leads to the signature ∆µo > 0, ∆µ1 = 0, which is similar to negative bias in PPO
feeder.
In short, a “small” bias in PPO or PS feeder will correspond to a change in the nominal
viscosity µo but no appreciable change in µ1, which is distinct from signatures for raw
material change. On the other hand if the feeder bias is large, then the slope µ1 will also
change from the nominal value and the resulting fault signatures will be similar to raw
material changes. We assume that such large feeder biases will be identified through
other independent means. For instance, in industrial applications the feeder hoppers are
typically refilled automatically if the weight of the raw material in the hopper falls below
a critical limit. For large gross errors in the feeders, the refill frequency will change
significantly from nominal rates, thus indicating a large positive or negative bias in the
DE-FC02-99-CH10972
53
I nt e llige nt Ex t rude r for Polym e r Com pounding
feeder. We will focus on the detection of “small” feeder biases, which cannot be readily
detected through other simple practical means.
Observed Q < Actual Q
Observed Q > Actual Q
vis
co
sity
vis
co
sity
∆µο < 0
∆µ1 = 0
Observed xo
Actual
∆µ1 = 0
Actual
Observed xo
PPO fraction
PPO fraction
• +ve bias in PPO feeder
• -ve bias in PPO feeder
Observed Q > Actual Q
Observed Q < Actual Q
vis
co
sity
∆µο > 0
vis
co
sity
∆µο < 0
∆µ1 = 0
Observed xo
Actual
PPO fraction
• -ve bias in PS feeder
∆µο > 0
∆µ1 = 0
Actual
Observed xo
PPO fraction
• +ve bias in PS feeder
Figure 31: Fault signatures for PPO/PS feeder bias
Condition/Fault
Nominal raw material, no feeder bias
High IV PPO
Low IV PPO
High IV PS
Low IV PS
Positive bias in PPO feeder
Negative bias in PPO feeder
Positive bias in PS feeder
Negative bias in PS feeder
∆µo
0
>0
<0
>0
<0
<0
>0
>0
<0
∆µ1
0
>0
<0
<0
>0
=0
=0
=0
=0
Table 4: Fault Signatures for raw material variation and feeder bias
Table 4 summarizes the fault signatures for raw material and feed-rate changes in terms
of combinations of changes in µo and µ1 from nominal values. Note that the signatures for
positive/negative bias in the PPO feed-rate are the same as signatures for
negative/positive bias in PS feed-rate, respectively. This is due to the fact the signatures
are governed mainly by the PPO fraction in the product, which may arise from a change
DE-FC02-99-CH10972
54
I nt e llige nt Ex t rude r for Polym e r Com pounding
in either feed-rate. Thus, the fault detection can identify a feed-rate bias distinctly from
raw material changes but not distinguish between the PPO and PS feed-rates.
8.2.2 Fault Diagnostics with Noryl on 25mm Research Extruder
The fault occurrence, detection, identification and correction will occur in successive
stages as shown in Figure 28. We ran multiple experiments with nominal and nonnominal raw materials over multiple days to test the performance of the fault diagnostics.
In particular, we used either nominal PPO (high IV = 0.46) or non-nominal PPO
(medium (0.4) / low (0.33) IV) while using nominal PS in all experiments. Furthermore,
each experiment run started with the nominal raw materials or the faulty (lower IV) PPO
from the beginning, focusing in particular on the fault identification stage, since the fault
detection is achieved simply by monitoring the residual error between measured die
pressure and the model prediction using nominal parameters against a threshold. Figure
32 shows the sequence of fault occurrence, detection and identification in a representative
run using data from the experiment on Nov 14, 2001 with raw material variation. In this
figure, we consider the scenario of starting with nominal operation in phase 1, and then a
fault – PPO changed from nominal (0.46IV) to medium IV – occurring at t = 2 min.
DE-FC02-99-CH10972
55
I nt e llige nt Ex t rude r for Polym e r Com pounding
Transition from nominal raw
materials to medium IV PPO
Detection of a fault
3
1
2
Model adaptation for fault identification
3
3
3
3
Figure 32: Fault occurrence, detection and identification using model parameter
identification for a representative run with raw material change.
Clearly, in phase 1 under nominal conditions, the measured value for die pressure
matches well with the model prediction using nominal parameters. However, after the
fault, i.e. the transition from nominal raw materials to a medium IV PPO, the product
viscosity, and thus, the die pressure drops significantly leading to a mismatch between
the measured die pressure and the model prediction. A fault is detected when this residual
error between the measured die pressure and the nominal model prediction crosses a
threshold. The choice of the threshold affects the trade-off between false alarms and
missed detects, and it depends on the signal to noise ratio. For the above change in raw
material, clearly the change in die pressure is significantly larger than the measurement
noise. Such a large signal to noise ratio allows an easy selection of a threshold. For
instance, even a conservative choice for threshold of +/- 30 psi, would trigger a fault
alarm in phase 2 within less than 2 minutes of the fault occurrence. The detection of a
fault in phase 2 will then initiate the fault identification in stage 3. In this identification
stage, the process is excited by a pre-defined PRBS input variation in the feed-rates and
screw speed and the model parameters are updated using the on-line recursive adaptation.
Finally, the new adapted parameters, specifically β1 and β2 (or equivalently µ0 and µ1
from Eq 12), are compared with the nominal values and the fault signatures in Table 4 to
DE-FC02-99-CH10972
56
I nt e llige nt Ex t rude r for Polym e r Com pounding
identify the specific fault. The lower four plots in Figure 32 show the plots of the 4
parameters β1-β4 during this adaptation phase 3. The parameters β1 and β2, or
equivalently µ0 and µ1, converge to a value lower than for nominal conditions by about
t=20 min., correctly matching the fault signature for lower IV PPO. Thus, the model
adaptation in phase 3 allows the identification of the correct fault.
Once a fault is correctly identified, then an appropriate corrective action can be taken to
bring the process back on-spec, e.g. fixing the raw material change, or using closed-loop
automatic control to correct for the raw material variation (see Section 9 for more
details).
8.2.3 Fault Identification Results with Noryl for Raw Material Changes
We summarize the results of the fault identification algorithm based on the modeling and
adaptation framework for the multiple runs with NORYL using nominal/non-nominal
raw materials in . The actual nominal/fault conditions imposed in the individual
experiments are listed in the second column and the nominal viscosity (at nominal PPO
fraction) measured in the lab using samples collected during steady state conditions are
listed in column 3. The estimated nominal viscosity µo and the slope µ1 of viscosity with
respect to PPO fraction obtained after the model adaptations in each run are listed in
columns 4 and 5, respectively. As mentioned before, the fault identification relies on
matching the deviations in µo and µ1 from the nominal values against the fault signatures
in Table 4. Due to the noise in the process measurements, the model parameters for
nominal conditions will lie in a normally distributed region. Thus, a threshold has to be
chosen to declare +ve or –ve changes in µo and/or µ1 before comparing with the fault
signatures. Again, these thresholds can be chosen statistically to optimize the trade-off
between false alarms and missed detects and their choice depends on the signal to noise
ratio. For instance, comparing the first three runs with nominal conditions, it is clear that
the parameters µo and µ1 are tightly clustered together and distinctly different from the
other runs with non-nominal raw materials. This distinct separation between the
parameter values for nominal and fault conditions facilitates a clear selection of the
thresholds. For instance, a choice of the thresholds for ∆µo = +/-50 and for ∆µ1 = +/-150
would suffice to correctly identify the nominal conditions in the first three runs and
declare the lower IV PPO in the last three runs (both ∆µo and ∆µ1 are –ve, matching the
signature for lower IV PPO). The fault identification results are summarized in the last
column of the table, which can be compared with the actual fault condition introduced
during each experiment, given in column 2. Clearly, the proposed fault identification
algorithm works very well to correctly identify the nominal runs and the specific faults
with raw material changes in each run.
DE-FC02-99-CH10972
57
Faulty raw
materials
Nominal raw materials
I nt e llige nt Ex t rude r for Polym e r Com pounding
Date of
Expt.
Fault
condition
(raw
material)
introduced
Viscosity
measured in
lab at
nominal
composition
Estimated
values of
nominal
Viscosity,
µ0 (∆µο)
Estimate of
Slope,
µ1 (∆µ1)
Fault
Identification
Result
6/25/2001
Nominal
(calibration
run)
Nominal
Nominal
Nominal
1152
1139 (-)
2944 (-)
Calibration
run
1129
1196
1167 (+28)
1174 (+35)
1155 (+16)
3083 (+139)
2891 (-53)
2844 (-100)
Nominal
Nominal
Nominal
Low IV PPO
Med IV PPO
Med IV PPO
920
1010
964
877 (-262)
1002 (-137)
940 (-199)
2537 (-407)
2550 (-394)
2647 (-297)
lowerIV PPO
lowerIV PPO
lowerIV PPO
5/22/2001
5/08/2001
7/25/2001
8/27/2001
10/11/2001
11/14/2001
Table 5: Diagnostics results for experiment runs with nominal/non-nominal raw
materials
In addition to the above six runs, we also analyzed later experiment runs conducted for
closed-loop control. Table 6 shows the results for the control runs and yields the correct
fault detection for the run on 12/06/2001. For the run on 12/04/2001 with medium IV
PPO, the calculated slope µ1 is close to the nominal value while the nominal viscosity µο
is lower than nominal – matching the signature for a small feeder bias. However, since
we deliberately changed the raw material, it would mean that two faults (raw material
change and feeder bias) occurred simultaneously, which cannot be isolated
unambiguously based on our fault detection method. The last run on 12/14/2001 during
closed-loop control yields µ0 close to values for nominal raw materials and a slope µ1 that
is distinctly lower than nominal runs. This change in µ0 and µ1 does not match any of the
raw material or feeder bias fault indicating possible multiple faults. Indeed, after thee
xperiment, we found that the feeders had a large bias, which together with the choice
of medium IV PPO led to a combination of faults in raw material and feed-rate. As
mentioned earlier, the fault diagnostics, in general, cannot correctly identify simultaneous
multiple faults since their effects are confounded. Note however that in all these three
runs, including the ones with multiple faults, the viscosity estimation in column 4 was
very good (within 10%) of the lab measurements given in column 3 – this will enable the
corrective closed-loop control based on the updated viscosity estimate.
DE-FC02-99-CH10972
58
I nt e llige nt Ex t rude r for Polym e r Com pounding
Date of
Expt.
12/04/2001
12/06/2001
12/14/2001
Fault
condition
(raw
material)
introduced
Med IV
PPO
Med IV
PPO
Med IV
PPO
Viscosity
measured in
lab at
nominal
composition
995
Estimate of
nominal
Viscosity,
µ0 (∆µο)
Estimate of
lope,
µ1 (∆µ1)
Fault
Identification
Result
930 (-209)
2920 (-24)
Feeder bias
980
968 (-171)
2594 (-350)
1095
1096 (-43)
1135* (-4)
2554 (-390)
2516* (-428)
Lower IV
PPO
Multiple
faults?
Table 6: Diagnostics results for control runs (*estimated after feeder bias correction)
8.2.4 Fault Identification of Feeder Bias
We also tested the fault detection for PPO/ PS feeder bias through specifically designed
off-line simulations using actual recorded input/output data from the nominal run on
06/25/2001. In particular, we added a specific positive/negative bias in the PPO or PS
feed-rates and performed the recursive parameter identification to test the performance of
the fault detection based on the fault signatures in Table 4. A positive bias in, for
example, the PPO feed-rate means that the observed feed-rate is higher than the actual
feed-rate. Figure 33 shows the comparison of measured die pressure with the die pressure
predicted using the observed feed-rates (with the bias introduced in PPO/PS feeder) and
nominal parameters. For instance, in the case of positive PPO feeder bias (upper left
plot), the observed throughput and PPO fraction are higher than the actual values and
thus the die pressure prediction is higher than actual. The fault detection results obtained
for the four cases using the parameter identification are summarized in Table 7. Clearly,
the calculated changes in µo, µ1 match with the expected fault signatures. However, while
the parameter µo changes quite significantly from the nominal value for a bias in the PPO
feeder and is easily detected, the change in this parameter for PS feeder bias is quite
small, i.e. detection of PS feeder bias is difficult as mentioned before due to the
competing effects of changes in feed-rate and composition in this case as mentioned
earlier.
DE-FC02-99-CH10972
59
predicted
Die P (psi)
Die P (psi)
I nt e llige nt Ex t rude r for Polym e r Com pounding
m easured
+10% PPO bias
Die P (psi)
Die P (psi)
-10% PPO bias
+10% PS bias
-10% PS bias
Figure 33: Comparison of measured and predicted die pressure for bias in PPO/PS
feeder using nominal parameters.
Condition/Fault
Nominal
+10% bias in PPO feeder
-10% bias in PPO feeder
+10% bias in PS feeder
-10% bias in PS feeder
µo (∆µο)
1139 (-)
1042 (-97)
1228 (+89)
1160 (+21)
1109 (-30)
µ1 (∆µ1)
2944 (-)
2922 (-22)
2926 (-18)
2914 (-30)
2945 (+1)
Detection result
Calibration
∆µo < 0, ∆µ1 = 0
∆µo > 0, ∆µ1 = 0
∆µo > 0? , ∆µ1 = 0
∆µo < 0? , ∆µ1 = 0
Table 7: Parameter identification and fault detection for feeder bias
8.3 Diagnostics Summary
In summary, the proposed modeling and adaptation provides a natural framework for
fault detection and identification using on-line process measurements and can be used as
the basis for taking corrective control action in the presence of faults. We have developed
a novel model-based fault detection methodology that is not tied to any specific process
or product grade and demonstrated its successful application on the 25mm research
extruder with NORYL polymer product. The key features of the developed diagnostics
algorithm are:
DE-FC02-99-CH10972
60
I nt e llige nt Ex t rude r for Polym e r Com pounding
(i) The approach is generically applicable to wide range of extruder applications
owing to the generic physics-based underlying model and physically motivated
fault signatures that are not problem specific.
(ii) As demonstrated by the multiple runs over a span of several months, the proposed
fault detection works very well to consistently identify raw material or feeder
faults, which are the main process faults that affect product quality. The distinct
fault signatures allow distinguishing between raw material changes and feed-rate
variations. However, owing to similar effects of the two raw material feed-rate
changes, it is not possible to distinguish between a bias in the PPO or the PS feedrate.
(iii)While we demonstrated the application of the developed fault detection approach
to a process with two main raw materials, it can be extended to the case of more
than two key raw materials – at the expense of increased number of model
parameters for viscosity and die pressure and thus, correspondingly increasing
complexity in the adaptation and fault identification. In particular, for multicomponent mixtures with N components, the relation for viscosity will involve N1 slopes µ1,i and the resulting fault signatures will involve unique combinations of
changes in µo, µ1,i making the fault detection for raw material changes more
complex.
The fault diagnostics algorithm assumes however that the faults occur one at a time. The
algorithm, in general, will not identify the correct fault in the case of multiple
simultaneous faults since their effects are confounded. When multiple faults are
occurring, the it is typical that severe upsets in machine operation have occurred and can
be detected by more conventional limit checks are gross anomalies like dropped strands.
The real power of the present technique is detecting small “drift” type faults that may
take a while for the effects to be seen. With the proposed scheme, such faults can be
detected, classified and dealt with before out of spec material is shipped.
9 Extruder Control
9.1 Approach to control in the presence of upsets
In this section, we present results on the closed-loop on-line control of viscosity. In the
absence of an on-line measurement/estimate of product viscosity, a common industrial
practice is to monitor on-line measurements like die pressure and torque to verify that
they are within a desired tolerance of nominal values. However, such practice is often
inadequate since the absolute magnitudes of these machine variables depend on both the
process conditions that affect product quality, e.g. feed composition, raw material
characteristics, and those that do not, e.g. throughput. In contrast, the online estimation of
product viscosity using the measurement of machine variables (feed-rates, die pressure
etc.) enables active closed-loop control of viscosity in the presence of disturbances like
changes in raw material (see Figure 34). A change in raw material properties can be
corrected by varying the incoming product composition to maintain the product viscosity
DE-FC02-99-CH10972
61
I nt e llige nt Ex t rude r for Polym e r Com pounding
on target. Thus, the on-line closed-loop control of viscosity entails the following two key
steps:
• On-line estimation (inferential sensing) of viscosity from measurement of
machine variables (feed-rates, die pressure)
• Corrective action to change feed-rates, and thus product composition, to
compensate for the effect of raw material changes and maintain product viscosity
at its nominal set-point.
Disturbances (raw
material change)
Feed composition
Online machine measurements
Control
Algorithm
Estimation
Inputs
(Q, N)
Extruder
Model
Measured Outputs
(T, DP)
predicted
outputs
-
+
updated
α, β
Parameter
Adaptation
-
Online viscosity estimate
+
Desired Viscosity
Setpoint
Figure 34: Schematic diagram for closed-loop control of product viscosity
9.2 Viscosity Adaptive Control
Figure 35 shows the schematic block diagram of the closed-loop control approach for
maintaining viscosity close to the desired set-point (for any product grade). The product
viscosity is affected by the main process disturbances like raw material and feeder
variations. Under nominal conditions, the model predictions with nominal parameters βi
matches the on-line measurements of die pressure, and thus, the on-line estimate of
product viscosity is obtained using the nominal parameters. In the presence of significant
disturbances, the model predictions for die pressure will differ from on-line
measurements leading to the model parameter adaptation. The newly adapted parameters
will reflect the effect of the disturbances and provide a corresponding updated estimate of
viscosity. The estimated viscosity is compared with the target set-point and the PI
DE-FC02-99-CH10972
62
I nt e llige nt Ex t rude r for Polym e r Com pounding
controller is used to manipulate the feed composition to maintain the estimated viscosity
at the set-point. The required feed composition is implemented by using ratio control
logic to vary the individual PPO and PS feed-rates while maintaining desired throughput.
These calculated PPO and PS feed-rates are finally provided as set-points for the lower
level feeder controllers.
Disturbances
• raw material
• feedrate
viscosity
set point
µSP
+
Control Algorithm
PI
Controller
_
PPO
feedrate
Ratio logic
composition
Extruder
PS feedrate
measured
die pressure
estimated viscosity
µ
Viscosity
Estimation
predicted product
composition
Estimation Algorithm
Dynamic
Model
_+
predicted die
pressure
Model
Parameter
Adaptation
Figure 35: Shematic block-diagram of the closed-loop control algorithm based on
on-line viscosity estimation
In order to tune the PI controller, the dynamics of the extruder are approximated by a first
order model. In particular, the product viscosity µ is given as a function of the production
composition xo (see nomenclature in Section 5), where the product composition xo = x2 is
given as a function of the feed composition xi by the model in Eq 11. The overall
response of the product composition, and thus viscosity µ, can be approximated by a
first-order response:
Eq 16
µ − µo =
µ1
( xi − x i )
τ ps +1
with a time constant τp. For our experiments on the lab-scale extruder we calculated the
time constant τp=30 seconds. The PI controller is then tuned for a closed-loop first order
response:
Eq 17
∆µ
1
=
∆µ SP τ f s + 1
where ∆µ and ∆µSP denote the deviation in estimated viscosity and desired set-point from
nominal values and τf is the desired closed-loop time constant for viscosity set-point
DE-FC02-99-CH10972
63
I nt e llige nt Ex t rude r for Polym e r Com pounding
tracking – we tuned the PI controller to achieve the closed-loop time constant τf = 60sec.
For more details on the controller tuning, see Appendix A to Section 5.
9.3 Control implementation for research extruder
The developed algorithms for estimation and control using Matlab/Simulink were
implemented on the extruder for closed-loop experiments using D-Space. Figure 36
shows the schematic representation of the various software/hardware elements of the
final implementation on the lab-scale extruder. In particular, the KTron feeder controllers
provided the internally calculated feed-rates as frequency signals, which were converted
to voltage signal using frequency to voltage conversion. In addition, the measured torque,
die pressure, melt temperature and screw speed were recorded as voltage signals. The
conversion of the raw voltage signals to physical units was handled by the software
algorithm in D-Space. The calculated feed-rate set-points from the control algorithm in
D-Space were in volts, which were converted to frequency signals using voltage to
frequency converters to interface with the KTron feeder controllers for automatic
adjustment of the feed-rate set-points. The software in D-Space was provided with a GUI
to allow monitoring the signals in real time, and provide the capability to enable/disable
the parameter adaptation and closed-loop automatic control (see Figure 6). In our
experiments, we manipulated the screw speed manually during the PRBS excitation for
parameter adaptation – screw speed was kept constant at nominal value during closedloop control. However, even the screw speed control on the extruder drive could be
interfaced with the D-Space implementation to allow a completely automated preprogrammed PRBS feed-rate, screw speed sequence during the parameter adaptation.
PLC 1
KTron
feeder 1
PLC 2
KTron
feeder 2
V/F converters
PPO feed
PS feed
F/V converters
measured feedrates (V)
DSPACE GUI
Software
die pressure (V/I)
I/O interface
melt temperature (TC) WP Extruder
torque (V/I)
screw speed (V/I)
GE Drive
Figure 36: Schematic diagram of the final implementation using D-Space for closedloop control experiments
DE-FC02-99-CH10972
64
I nt e llige nt Ex t rude r for Polym e r Com pounding
Experimental Control results with Noryl Material
Figure 37 depicts the set-point tracking experiment on 12/14/2001 with nominal raw
materials. As seen in the top figure, in the first part of the experiment, the system was
subjected to a PRBS input variation to facilitate the parameter identification. The bottom
plot shows the PRBS variation in the PPO and PS feed-rates during this adaptation phase.
Note that the estimated viscosity also varies during this phase, partly due to the
adaptation of the model parameters β, and partly because of the variation in the feed
composition during the PRBS excitation of the feed-rates. After this parameter adaptation
phase, in the second part of the experiment, the closed-loop controller was enabled and
the set-point for viscosity was increased from the initial nominal value of 1116 Pa-s to
1250 Pa-s and then decreased back to 1116 Pa-s. Clearly, the controller steered the
estimated viscosity of the system to the desired values by manipulating the PPO and PS
feed rates. In particular, the controller increased the PPO:PS feed ratio as expected to
achieve the higher viscosity when the set-point for viscosity was increased to 1250 Pa-s,
and then reduced it back to nominal after the set-point was reduced back to 1116 Pa-s.
DE-FC02-99-CH10972
65
I nt e llige nt Ex t rude r for Polym e r Com pounding
PRBS excitation
closed-loop
Figure 37: Closed-loop control of viscosity – set point tracking with nominal raw
materials on 12/14/2001
The blue plot in the figure represents the on-line viscosity estimate, which is dependent
on the composition determined by the measured feed-rates of the PPO blend and PS. If
the feed-rates were with out any bias, (feed-rates were implemented by the low level KTron controller exactly the same as set-points of the feeders: 16 lbs/hour PPO blend, 15
DE-FC02-99-CH10972
66
I nt e llige nt Ex t rude r for Polym e r Com pounding
lbs/hour PS rate for nominal composition fraction of 0.516) then the viscosity estimate
seen in
Figure 37 would be within measurement error range of the off-line viscosity
measurement. But the off-line lab measurements for two samples collected before and
after the set-point change and denoted with red line in Figure 38 (same Figure 37 as
with off-line measurement data) are higher than estimated values demonstrated in
Viscosity (Pas)
Figure 37. This means that a large feeder bias caused a higher PPO fraction (this high
viscosity was later found to correspond to PPO fraction of 0.571) in the mixture, thereby
increasing the viscosity and the die-pressure measurement.
lab measurements
estimated viscosity
t (s)
Figure 38: Comparison of closed-loop control of viscosity –set point tracking using
nominal raw materials - with off-line viscosity measurement on 12/14/2001
On the day of the experiment (12/14/01), we observed that even with nominal raw
materials and running at the nominal feed PPO composition of 0.571, the measured die
pressure was higher by 15% than similar nominal experiments done until that date.
Moreover, the identified parameters led to a higher µo and µ1, i.e. ∆µo > 0, ∆µ1 > 0, a
fault signature corresponding to high IV PPO. However, this was not possible since the
nominal PPO used in the experiments was already the highest IV (0.46IV) PPO available
for our experiments. In fact, this fault signature was caused by the large feeder bias
(possibly due to different set of feeders used on 12/14/01), which as mentioned in the
diagnostics section can yield a fault signature similar to raw material change - in this case
similar to a higher IV PPO. As mentioned in the diagnostics section, the detection of such
large feeder biases can be addressed separately through simple practical means (e.g. feed
hopper refill frequency). The developed diagnostics algorithm focuses only on “small”
feeder biases that have a distinct signature compared to nominal runs or runs with bad
raw materials.
Given the large feeder bias indicated by the off-line lab viscosity measurements for the
samples collected during the nominal run, which were about 20% higher than that
observed for similar nominal raw material and composition samples on prior days, we
DE-FC02-99-CH10972
67
I nt e llige nt Ex t rude r for Polym e r Com pounding
analyzed the data from the run after correcting the feed-rates with estimated feeder bias.
More specifically, the increased viscosity corresponded to an increased PPO fraction of
0.571 as opposed to the nominal composition of 0.516, indicating a bias in PPO and/or
PS feeders. Moreover, the increased viscosity and online die pressure measurements were
of the same order of magnitude indicating that the total throughput Q was close to
nominal and only the PPO fraction had changed. To verify this hypothesis, we reevaluated the online data from the run with an imposed positive bias correction in
measured PPO feed-rate (+1.7 lb/h), a negative bias correction in PS feed-rate (-1.7 lb/h)
and using the original calibrated value of k obtained from the nominal calibration run on
06/25/2001. The resulting viscosity estimation during the adaptation and closed-loop
control is shown in Figure 39. It shows the estimated viscosity (shown in blue in the
bottom plots) is close to the off-line measurements (shown in red), corroborating the
feeder bias hypothesis.
Viscosity (Pas)
PRBS excitation and adaptation
closed-loop
control
lab measurements
estimated viscosity
t (s)
Figure 39: Experiment run on 12/14/2001 with nominal raw materials, accounting
for PPO and PS feeder bias
DE-FC02-99-CH10972
68
I nt e llige nt Ex t rude r for Polym e r Com pounding
closed-loop
control
PRBS excitation and adaptation
t (s)
Viscosity (Pas)
estimated viscosity
lab measurements
PPO feed (lb/h)
t (s)
Feed-rate
variation by
controller
PRBS excitation
PS feed (lb/h)
t (s)
PRBS excitation
t (s)
Figure 40: Experiment run on 12/14/2001 with medium IV PPO, accounting for
PPO and PS feeder bias
On 12/14/2001 a further control test using medium IV PPO blend and nominal PS
mixture was conducted. Since the same unknown feeder bias was present for the same
day, we analyzed the data offline with the above-mentioned feeder bias corrections. The
results are shown in Figure 40. Again, in the beginning a PRBS excitation on the feedrates and screw speed was imposed to aid the adaptation. With the new adapted model
DE-FC02-99-CH10972
69
I nt e llige nt Ex t rude r for Polym e r Com pounding
parameters a new viscosity estimate was obtained. Finally, the closed-loop control was
enabled and the viscosity set-point was increased. Clearly, the controller yielded the
increased viscosity by increasing the PPO feed concentration, specifically by increasing
the PPO feed-rate and decreasing the PS feed-rate to maintain the nominal total feed-rate.
The values of viscosity measured in the lab using two samples collected at steady state,
one before the set-point increase and the other after the set-point change are compared
with the on-line estimate. Again, the on-line estimate of viscosity matches well with the
lab measurements and the closed-loop controller smoothly changed the product
composition and thus viscosity during the set-point increase.
As demonstrated by the above figures the closed-loop control scheme works well to
compensate for the effects of unknown raw material changes and track changes in desired
viscosity set-point. The effectiveness of the control scheme is clearly dependent on the
performance of the online viscosity estimation.
10 Scale-Up for Production-Scale Extruders
10.1 Scale of Dynamic Input-Output Model of Extruder
In this section, we address the scale-up of the developed approach and its application to
industrial production scale extruders. In particular, we study the application of the
approach developed above on the lab-scale 25mm extruder when scaled up to the
production-scale 120mm extruders at GEP Selkirk. We obtained and analyzed data
recorded at 0.2 Hz from the GEP Selkirk site on a specific extruder line that was
instrumented to measure and record the signals like torque, die pressure, feed-rates and
screw speed during DOE runs conducted on Nov 21, 2000 and Jan 18, 2001 while
making Noryl PX5511 product. In these two DOE runs, the process inputs, i.e. the screw
speed and the PPO, PS feed-rates were varied in a PRBS manner with small variations
from the nominal operating values so as to excite the system dynamics enough for
analysis while maintaining the product quality within desired specifications.
To test the scale-up of the model developed for the lab-scale extruder, we applied the
model to the data from the DOE run on Jan 18, 2001 and obtained the best-fit model
parameters using least squares off-line optimization. Thereafter, we used the model with
these best-fit parameters and compared its predictions of torque and die pressure using
the data from the DOE run on Nov 21, 2000, to test the validity of the model. Our initial
comparison of the measured torque and die pressure data for the two DOE runs showed
the nominal die pressure on Jan 18, 2001 to be significantly lower than on Nov 21, 2000
even though the operating conditions were exactly the same. This occurred due to the use
of a different screen-pack – used to filter out particulates in the molten product just after
the die pressure probe – which effectively changed the resistance to flow and thus the
pressure drop. We corrected for this difference by adding a constant bias value from the
die pressure measured on Jan 18, 2001 to make it consistent with the nominal value on
Nov 21, 2000. Furthermore, the melt temperature measurement on the extruder was
DE-FC02-99-CH10972
70
I nt e llige nt Ex t rude r for Polym e r Com pounding
faulty; it measured a constant temperature irrespective of any changes in throughput
and/or screw speed. In the absence of a good melt temperature measurement, we used a
simple first-order model to predict the changes in the melt temperature (we need only the
changes from the nominal value for fitting our model) as a function of the screw speed
and throughput (see Eq-A 21) and used it in the die pressure relation (see Eq-A 18).
Speed (rpm)
(lb/hr)
Torque (%)
Figure 41 shows the comparison of the measured torque and the model predictions for
torque during the DOE run on Jan 18, 2001. The model can be seen to fit the measured
torque variations during the DOE run very well. Figure 42 shows the comparison of the
model predictions and measured values of die pressure during the same DOE run. Again,
the model prediction for die pressure matches very well with the measured die pressure
variations. We tested the validity of the model with the obtained best-fit parameters by
comparing it against the data from the DOE run on Nov 21, 2000. Figure 43 and Figure
44 show the model predictions for torque and die pressure compared to measured values
on Nov 21, 2000, respectively. Clearly the model predictions match very well with the
measured values, i.e. the model validates very well with independent set of data. This
validation demonstrates the scale-up of the developed modeling approach and its
applicability to large industrial production-scale extruders.
error histogram
Figure 41: Comparison of model-predictions and measured value of torque on a
120mm extruder during a DOE run at GEP Selkirk on Jan 18, 2001.
DE-FC02-99-CH10972
71
Speed (rpm)
(lb/hr)
Die pressure (psi)
I nt e llige nt Ex t rude r for Polym e r Com pounding
error histogram
Speed (rpm)
(lb/hr)
Torque (%)
Figure 42: Comparison of model-predictions and measured value of die pressure on
a 120mm extruder during a DOE run at GEP Selkirk on Jan 18, 2001.
error histogram
Figure 43: Validation of model for torque using DOE data on a 120mm extruder at
GEP Selkirk on Nov 21, 2000.
DE-FC02-99-CH10972
72
Speed (rpm)
(lb/hr)
Die pressure (psi)
I nt e llige nt Ex t rude r for Polym e r Com pounding
error histogram
Figure 44: Validation of model for die pressure using DOE data on a 120mm
extruder at GEP Selkirk on Nov 21, 2000.
10.2 Scale-Up of On-line Adaptation of Model Parameters
For the DOE run on Jan 18, 2002, we tested the performance of the recursive adaptation
of the model parameters β, starting from slightly perturbed values to test the performance
of the adaptation. Table 8 summarizes the results of the parameter adaptation starting
from the initial perturbed values to yield the final adapted values, which are close to the
best-fit values obtained by the off-line least squares optimization. Figure 45 shows the
comparison of the measured die pressure and the model prediction during the adaptation
starting from the initial perturbed values of β. It can be seen that the model adaptation
works well to match the predicted values of die pressure with the measured values and
yield the correct values of the parameters β.
DE-FC02-99-CH10972
73
I nt e llige nt Ex t rude r for Polym e r Com pounding
Adaptation
Off-line
optimization
Initial
values
Final
values
Optimal values
β1
17.38
15.365
15.38
β2
19.02
22.46
23.0
β3
1.077
0.87
0.87
β4
3.43
3.23
3.13
Die pressure (psi)
Table 8: Scale-Up of parameter adaptation for die pressure for 120mm extruder
data from DOE run at GEP Selkirk on Jan 18, 2001.
Figure 45: Comparison of measured and model predicted values of die pressure
during parameter adaptation with DOE data from 120mm extruder at GEP Selkirk
on Jan 18, 2001
10.3 Scale-Up of Viscosity Estimation
In order to perform the viscosity estimation using the estimated parameters β1 and β2, and
the product composition xo, the calibration parameter k needs to be calculated to perform
the conversion between β1 and β2 and µο and µ1 (see Eq 12). We had QA lab
measurements of viscosity of 10 samples collected at various steady state operating
conditions during the DOE run on Nov 21, 2000. Since, this was the only viscosity data
we had for the DOE runs conducted at GEP Selkirk, we performed the calibration using a
few samples and compared the estimation against others. More specifically, a linear fit
between the measured viscosity and PPO weight fraction is performed to calculate the
DE-FC02-99-CH10972
74
I nt e llige nt Ex t rude r for Polym e r Com pounding
parameters µο and µ1 and thus, calculate the calibration parameter k. The linear fit yielded
the linear relation for viscosity:
Eq 18
µ = 25.49 + 272.6 * xo
= 178.15 + 272.6 * ( xo − xo )
with µο=178.15 and µ1=272.6 at the nominal mean PPO fraction xo = 0.56 . Thereafter,
the values of µο and µ1 and β1 and β2 were used to calculate the calibration parameter k
and estimate the product viscosity for other samples using the product composition xo.
The results are summarized in Table 9. Clearly, the estimated viscosity matches very
well with the measured values in the QA lab, with maximum error less than 4% even for
the samples not used in calculating the calibration parameter.
Sample no.
1*
2
3
4*
5
6*
7
8*
9
10
11*
12
PPO fraction
xo
0.58
0.58
0.58
0.58
0.54
0.56
0.54
0.54
0.58
0.56
0.58
0.54
Measured
viscosity (Pas)
186
186.6
182.7
184.6
171.4
175.4
168.6
174.2
186.9
178.8
183.7
171.9
Estimated
viscosity (Pas)
185.0
185.7
184.9
184.8
174.9
180.3
175.0
175.1
185.8
180.6
185.7
175.1
% error
-0.5376
-0.48
1.20
0.11
2.04
2.79
3.80
0.52
-0.59
1.01
1.09
1.86
Table 9: Comparison of measured and estimated values of viscosity for samples
collected during DOE run on Nov 21, 2000 at GEP Selkirk (* samples used in calculating
the calibration parameter)
10.4 Scale-Up Summary
The above-mentioned results clearly demonstrate the scale-up of the developed modelbased framework for estimation, diagnostics and control. In particular, the excellent fit
and validation of the model with data from a production-scale 120mm extruder
demonstrates the validity of the physics-based model and its application to any extruder
irrespective of size. Furthermore, the model-based adaptation also applies equally well to
the large production-scale extruder. Finally, the viscosity estimation using the model and
the model-based estimation works very well and the estimated viscosity was within 4%
of the off-line measurements in the QA lab, well within the accuracy of the capillary
rheometer. Due to limited resources, we addressed only the modeling, adaptation and
viscosity estimation on the 120 mm extruder and did not have the opportunity to
DE-FC02-99-CH10972
75
I nt e llige nt Ex t rude r for Polym e r Com pounding
investigate the fault diagnostics and closed-loop control through multiple experiments.
However, the extension of the modeling, adaptation and viscosity estimation to the
120mm extruder gives ample confidence in the scale-up of the diagnostics and closedloop control results as well.
We summarize the various steps in the developed model-based approach for estimation,
diagnostics and control to any extruder application in Figure 46.
DE-FC02-99-CH10972
76
I nt e llige nt Ex t rude r for Polym e r Com pounding
Diagnostics
• Generate fault signatures for main raw material and feeder variations for each product
grade/family
• Tune thresholds for fault detection and identification based on signal-to-noise ratio
and desired specifications on product quality – trade-off false alarms vs. missed
detects
Modeling
• Collect experimental data from extruder with a PRBS excitation of the inputs – if
normal daily operation has enough variability, may be possible to use that historical
data instead
• Obtain an initial value of the model parameters using off-line least-squares
optimization to get the best-fit between the model prediction and measured values of
torque and die pressure
o Identify these parameters under nominal conditions for product grades or
families of similar product grades (similar key raw materials and composition)
o Fix the machine-dependent parameters (one for each extruder for multiple
extruder lines) and use on-line adaptation for process-dependent parameters
Estimation
• Initial calibration:
o Obtain viscosity vs. composition (key raw materials) data for each product
grade/family - use historical data if it covers wide enough composition range,
else perform controlled experiments
o Fit linear/nonlinear correlation between viscosity and composition for each
grade/family – linearize around nominal compositions for each grade/family if
required
o Calibrate conversion parameter between model parameters β1, β2 and the
parameters µο, µ1 in linearized viscosity-composition relation
• Parameter adaptation and estimation:
o Initiate model with nominal parameters β for each grade/family
o Perform on-line recursive adaptation if there is mismatch between model
predictions and measured values of torque and die pressure
o Obtain on-line viscosity estimate with model parameters β and predicted product
composition
• Occasional verification of viscosity estimation with lab measurements of samples
collected during steady state operation
Control
• Specify target specifications and bounds for viscosity for each product grade
• Tune PI controller based on parameters for first-order approximation of process
dynamics and desired closed-loop response
Figure 46 : Steps for implementing developed model-based estimation, diagnostics
and control algorithm in an extruder application.
DE-FC02-99-CH10972
77
I nt e llige nt Ex t rude r for Polym e r Com pounding
11 Benefits
11.1 Overview of benefits derivation
Figure 47 - How Intelligent Extruder Derives Benefits
As shown in Figure 47, the Intelligent Extruder System provides benefits that include
reduced waste and energy use, improved quality, and by reducing recycled product,
increases system capacity to process virgin product, potentially delaying the need to add
costly compounding equipment.
The intelligent extruder monitor and diagnostic system provides benefits to a polymer
producer by alerting process control room operators and production line personnel when
out-of-spec material is being produced, and the closed loop adaptive control system (if
used) provides a means to bring production back into compliance automatically when
process corrections are feasible by changing the appropriate feed-stocks or other
manipulated variables. Specific benefits can be grouped as follows:
DE-FC02-99-CH10972
78
I nt e llige nt Ex t rude r for Polym e r Com pounding
•
•
•
•
•
•
Improved first pass yield by reducing the quantity of out of spec material that
must be recycled or sold as cheaper grade
Reducing the quantity of material that must scrapped to landfill
Reducing or eliminating material that escapes quality checks but is shipped to the
customer and returned, incurring transportation expense and unhappy buyers
Reducing the energy consumed in the net land-filled waste plus material recycled
(recyclable waste plus customer returns)
Reduced loading on the quality lab for process checks
A continuous quality audit assures material shipped meets customer expectations,
and reduces or eliminates the likelihood of returns: monitoring and controlling
consistency of product is often as import to end users as conformance to absolute
specifications
Figure 48 Process Timeline for Benefit Calculations
Figure 48 shows an event time line for a typical extruder based process contrasted with a
timeline with Intelligent Extruder in place. As suggested in the figure, benefits in waste
material reduction and recovered production capacity come simply from more rapid
detection of process upsets and corrections which get the process back on line.
DE-FC02-99-CH10972
79
I nt e llige nt Ex t rude r for Polym e r Com pounding
11.2 Approach to benefits quantification
To quantify the potential benefits, the following factors must be considered:
• Frequency of Process variation—The temporal behavior of process upsets, must
be quantified, either statistically from past QA data (including customer returns,
base resin shifts, etc) or based on first principles understanding. Understanding
this factor would be done as part of an initial process audit; for example, based on
a pareto of multiple factors as defined in Section 3.1 would be considered. Often
this data is hard to come by in our experience, either because the temporal
sampling resolution is not frequent enough, or determinable (e.g. customer
returns).
• Frequency of QC checks during a production run—Because of cost and
logistics, QA tests from a lab may only be done initially in the formulation for
short runs, or at low frequencies (e.g. each shift) for long campaigns. The lag
between when an upset occurs and QA (or the operator in extreme cases) detects
it is waste or off spec material.
• Time and method to correct the problem—Recovery time is dependent on the
problem and specifics of the material; for the 80% of quickly correctable
problems ( base resin shifts, feeder blockages etc), it is vital to know whether
operator intervention is required (e.g. feeder blockage or strand drops) or can be
handled through automation (adjustment to feed constituent rate set points).
• Time to restart and QC check the process
There is no easy way to get numbers for a given process. GE sales teams or the customer
would use a variety of calculation tools and information about the process capability to
assess the potential benefits of the Intelligent Extruder services offering. For example,
Figure 49 shows a typical spread sheet that was prototyped to aid in quantifying the
benefits in applying Intelligent Extruder at GE Plastics.
DE-FC02-99-CH10972
80
I nt e llige nt Ex t rude r for Polym e r Com pounding
GE Advanced Process Services
Compounding Extruder/Finishing Line Economics
Project: GEP Selkirk - Line No. 8
Analysis Period:
Product Type: Noryl PX5515
5
Noryl PX5515 Sales Price
Order size:
20,000 pounds
Revenues:
Feeder A:
Feeder B: Feeder C: Feeder D:
Feedstock Costs:
Type: S cre w
1 Vi b ra tory
2 Belt
3 Belt
3
Energy Costs:
Feeder Composition:
PPO
Chalk
PS
The rest
Recycle Costs:
Feed Ratio:
21%
29%
15%
35%
Total Operating Costs:
Feedstock Cost: $
0.75 $
0.75 $
0.25 $
0.25 Contribution Margin:
0.25
Sp. Energy (kw-hr/kg):
Production Rate: 2037.5
Finished Production Rate:
2014.3
Motor RPM's
Order Transition Time: 101.568
210.0
Machine Info
Machine Type:
Screw Size:
Screw Pitch:
Gear Ratio:
Screw Length (L/D):
ZS K -1 2 0
4.72
2.36
0.13503
24
8
inches
inches
speed out/in
inches/inches
Yield loss at transition (1st pass):
CTQ sampling time:
Yield loss at transition w/sampling:
% Recycle possible:
Recycle costs: $
24
$
$
$
$
hours
1.90 /pound
91,854
24,450
27%
1,619
2%
44
0%
26,113
28%
65,741
71.6%
lbs/hr
lbs/hr (includes losses)
seconds
0.3% (perfect setup)
5 minutes
1.1%
80%
0.10 /pound
Figure 49 Process Spread Sheet used for Benefits Analysis with Customer
Range of Expected Energy and Waste Reduction Benefits Results
The following data represents a composite over several engineering plastic polymers
similar to those used during this study. A summary of findings for the industry:
1. The range of first pass yield, the quantity of material that meets customer
requirements and can be shipped after one pass through the production
line, was approximately 88% for a poorly performing line to 98% for a
well performing line (98-99% is the perfect stoichiometric yield due to
volatile removal of about 1-2%).
2. Taking 94% as an average (actual data is confidential), about 6% of
material produced cannot be sold. Of this about 2% is non-pellet waste,
including volatiles, drool piles from startup, or strand waste from startup.
3. The remaining 4% is pellet waste that is out of specification due to
viscosity, mechanical properties, color or other defects.
4. The cost of land-fill is about the same as virgin polymer, and recycle in
high grade materials can degrade properties, so minimal recycle can be
assumed
5. Based on industry data and data from GE, about half or 2% is candidate
for improvement with our Intelligent Extruder system technology. Where
first pass yields are lower, the percentage will generally be even >2%, so
we will use this as a conservative bound. That is, if it can be assumed that
our technology can yield a 2% improvement in FPY from 94% to 96% (or
88 to 90% etc), the savings are significant.
DE-FC02-99-CH10972
81
I nt e llige nt Ex t rude r for Polym e r Com pounding
To put a 2% FPY improvement in context, GE produces approximately 7 billion pounds
of material per year. Table 10 shows a comparison of the potential benefits to GE to go
from 94% to 96% FPY. The benefit is a savings of $91MM, 140MM lb less solids to the
land fill, and about 146 MM Mwhr less energy use, and discharged 7MM lbs fewer
volatiles into the atmosphere. For simplicity we have combined the distillation and drying
energy to make the resin with specific energy per lb to process the resin in the extruder.
The reader can readily put his or her own numbers to run the calculation for their own
situation. To be fair, this projection assumes all extruders for all materials would get the
full 2% benefit, but the spread of benefits would also be offset by a spread in FPY in
which the upside could be greater for lines/materials with low FPY.
Assumptions
Pounds
Produced
7.00E+09
Base
94.00%
6.00%
Material
(lb)
4.20E+08
2% FPY Delta
Comparison
96.00%
4.00%
2.80E+08
First Pass Yield
Loss
Material
($)
2.10E+08
Energy
Energy
(kwhr)
($)
4.37E+08 2.18E+07
Finishing
($)
4.20E+07
1.40E+08
2.91E+08 1.46E+07
2.80E+07
$0.05
Average industrial contract price
resin mat'l cost ($)
$0.50
Base resin +pigments+fillers+other additives
spec energy (kwh/lb)
1.04
Includes energy (steam/elect) to produce resin + extruder energy
finishing cost ($/lb)
$0.10
Average non-energy compounding cost per pound
volatiles (lb-vol/lb-resin)
8.000E-05
Average for various engineering polymers
$183
Savings
$-MM
Land fill (MM-lb)
Energy (Mwhr)
Energy (Quad)
Volatiles (MM-lb)
Data
kwh cost ($)
Profile of Waste Produced
Cost
Solids
Energy
($-MM)
(MM-lb)
(Mwhr)
$274
420
436,800
280
291,200
$91
140
145,600
4.968E-04
7.0
Table 10 Impact of 2% FPY Improvement on GE Production
If this table is extrapolated to the whole engineering polymer industry which is about
30MM pounds per year, the results are shown in Table 11. Again this assumes 100%
penetration and that all production could benefit, but the interested application engineer
can readily work the numbers to their situation
Assumptions
Pounds
Produced
3.00E+10
Base
94.00%
6.00%
Material
(lb)
1.80E+09
2% FPY Delta
Comparison
96.00%
4.00%
1.20E+09
First Pass Yield
Loss
Material
($)
9.00E+08
Energy
Energy
(kwhr)
($)
1.87E+09 9.36E+07
Finishing
($)
1.80E+08
6.00E+08
1.25E+09 6.24E+07
1.20E+08
Data
kwh cost ($)
$0.05
Average industrial contract price
resin mat'l cost ($)
$0.50
Base resin +pigments+fillers+other additives
spec energy (kwh/lb)
1.04
Includes energy (steam/elect) to produce resin + extruder energy
finishing cost ($/lb)
$0.10
Average non-energy compounding cost per pound
volatiles (lb-vol/lb-resin)
8.000E-05
Average for various engineering polymers
Profile of Waste Produced
Cost
Solids
Energy
($-MM)
(MM-lb)
(Mwhr)
$1,174
1800
1,872,000
$782
Savings
$-MM
Land fill (MM-lb)
Energy (Mwhr)
Energy (Quad)
Volatiles (MM-lb)
1200
1,248,000
$391
600
624,000
2.129E-03
30.0
Table 11 Extrapolation of 2% FPY Improvement to Engineering Polymer Industry
DE-FC02-99-CH10972
82
I nt e llige nt Ex t rude r for Polym e r Com pounding
To more carefully quantify the benefits in a particular manufacturing context, it is
necessary to work up the details specific the process as outlined above and to establish
whether composition/viscosity properties and the root causes of variability can have the
same impact as suggested here. Potential users wishing to carry out a careful audit to
quantify potential benefits of the Intelligent Extruder system on their process should
contact GE Industrial Systems for assistance.
12 Commercialization Plan
12.1 Market Opportunity
The technology developed as part of this program presents sizable market opportunities
in the application of process control equipment and services primarily, but not limited to,
the plastics industry in which compounding screw extruders are used. To the end user, as
discussed in the Benefits discussion above, this technology presents a significant
opportunity to eliminate waste and increase yield, increase throughput, product quality,
and as a result, competitiveness. Because of the attractiveness of this market, GE
Industrial Systems (GEIS) has maintained ongoing discussion for an alliance with
Coperion Werner & Pfleiderer Inc (USA) (CWP) in order to expedite the introduction of
this technology to the market.
GEIS is a supplier of large AC and DC motors, adjustable speed drives, process control
equipment and related services. CWP is a supplier of state of the art compounding twinscrew extruders, related controls and services. Both parties have supplied equipment to
the plastics industry for over 15 years. Sales opportunities are anticipated from, in order
of priority, new CWP extruders, modernization of existing CWP extruders,
modernization of other existing extruders, and new and existing extruders in other
industries, as tabulated in Table 12 Market Segmentation .
In all of the markets in Table 12, it is the team’s opinion that service is the cornerstone of
the product offering as it provides the opportunity to optimize the process and to
recognize further opportunities for savings and reduction of wastes. Based upon historical
data, service revenues could exceed 20% of the overall project revenues (precise numbers
are proprietary to GEIS and CWP) and are attractive due to generation of an ongoing
revenue stream and an installed based for future upgrades as software innovations are
developed.
DE-FC02-99-CH10972
83
I nt e llige nt Ex t rude r for Polym e r Com pounding
Market
Anticipated Market
Opportunity (based on CWP data)
New CWP
40 units/year*
Extruders
Existing
CWP
Extruders
5000 units worldwide at
1% annual replacement
= 50 units/year*
Other
Existing
Extruders
Extruders
in other
Industries
10,000 units worldwide,
0.25% annual replacement
=25 units/year*
Food – market 10 times
plastics market
Metal casting– few
customers, but huge
savings opportunity
Technical, cost and competitive factors
“Value” price will far exceed cost, so
economic justification will be inherent to
the selling process. Customer alternative is
to buy traditional manual control
These customers would normally purchase
new control equipment only for reliability
reasons. Will need to overcome customer
reluctance to new technology (“selling up”)
through strong economic justification.
Customer alternative is to do nothing unless
reliability is an issue.
Selling up to advanced technology may be
difficult w/smaller users who have differing
requirements and limited capital resources.
Technology or CTQ’s may not be
applicable to these processes
Table 12 Market Segmentation linked to extruder sales
To examine the market opportunity more broadly, a cross-functional team from GE,
CWP and potential customers (GE Plastics and a small independent supplier) developed
data that identified other segments where Intelligent Extruder technologies might be
relevant. The results of these discussions are summarized in Figure 50: Potential Extruder
Market Segments. In addition to the polymer industry segment that is our focus, coatings
(insulation, toners, etc), fiber and film manufacture and injection molding all looked
attractive as follow on segments to pursue. These markets combined could exceed by 23X the value of the polymer manufacturing industry; though these are gross market
estimates only, it reinforces the potential synergy of the underlying polymer finishing
Intelligent Extruder technologies with related segments.
This team also drilled down into the polymer market opportunity to better understand the
factors driving the market. Findings are summarized in Figure 51: Detailed Opportunity
Fishbone for Intelligent Extruder. Overall, this sub-segment was believed to offer
approximately a $260MM market for the kinds of advanced automation and monitoring
products and services being developed in this program, initially targeted and achieving
resin quality and consistency for the target customers.
DE-FC02-99-CH10972
84
I nt e llige nt Ex t rude r for Polym e r Com pounding
Figure 50: Potential Extruder Market Segments
DE-FC02-99-CH10972
85
I nt e llige nt Ex t rude r for Polym e r Com pounding
Figure 51: Detailed Opportunity Fishbone for Intelligent Extruder
12.2 Commercialization Strategy
The commercialization of new process control technologies is not new to GE Industrial
Systems. GE has successfully supplied such breakthrough technologies to customers in a
number of industries since the early 1960’s when the on-line computer became
commercially viable. As stated above, to expedite the introduction of this technology to
the marketplace, GEIS and CWP have proposed forming a technology and market
development alliance. Throughout the definition of this alliance, joint marketing
discussions have included “what is the product and who will produce it?”, “how will
customers obtain this product?”, and “how will the product be sold?”.
12.2.1
Product – Advanced Process Control and Service
A typical extruder control system is expected to consist of an adjustable speed drive with
coordinated control, an operator interface with process displays and reporting, a general
purpose controller for miscellaneous control and sequencing functions, and a controller
for the monitoring, diagnostic and adaptive control functions described above.
The goal remains to have a standard platform that is simple, easy-to-use, and inexpensive
so that it may be sold on its own. Furthermore, the platform will be constructed so as to
DE-FC02-99-CH10972
86
I nt e llige nt Ex t rude r for Polym e r Com pounding
facilitate service opportunities including data collection to help field engineers to conduct
initial plant surveys locate additional opportunities for finishing operations
improvements, and remote diagnostics. To enhance market opportunities and
accommodate customer needs, modular software will be written so as to facilitate
implementation on alternate control platforms from other vendors.
12.2.2
Remote Service Offering
From the outset, it was the vision that service offerings could be provided remotely, after
initial setup and equipment validation, utilizing GEIS’ Onsite Center in Salem, Virginia.
Using telephone or secure web-enabled links, data would be downloaded through the
factory floor process control equipment, where data conditioning and processing
algorithms would be maintained. This facility already exists for remote monitoring of
drive equipment in the steel and paper industry, providing a scalable infrastructure to
support the needs of this program.
12.2.3
Distribution – Utilizing CWP’s Distribution Channels
The new control and services platform developed will be sold as an integral part of each
new mid and high end extruder sold by CWP wherever possible. Therefore, CWP’s direct
sales and distribution channels will be utilized. In the CWP extruder retrofit market,
where there are also opportunities for mechanical equipment upgrades, CWP sales and
distribution channels will be used, just as for new extruders. For other extruder retrofit
opportunities, sales and distribution will be directly by GE.
12.2.4
Commercialization Sales Tools
Through the alliance with CWP, it is planned, outside the scope of this program, that the
marketing objectives would be achieved by developing several sales tools. These are
expected to include the following:
Plant-wide Benefits/payback awareness presentation – a customer presentation that
will outline the benefits of the technology developed as it is implemented at the time
of the presentation and the project costs. An “experience list” will be developed as the
technology is installed in various facilities.
Standard technical specifications which describe the capabilities of the technology
and its associated control functionality.
Utilize full-scale technology pilot plant for future customer sales references.
Jointly-developed Advertising and Sales Promotion (A&SP) tools which will include
brochures, catalog pages, and Internet Web Pages.
Additional promotional events are anticipated, such as technical papers presented in
trade associations as well as KWP’s own Polymer Processing Conference held
annually in Ramsey, NJ.
DE-FC02-99-CH10972
87
I nt e llige nt Ex t rude r for Polym e r Com pounding
Each of these sales tools would have been distributed to the GE and CWP sales forces, as
appropriate, and each tool updated as the technology matured in actual target
implementations.
12.3 Commercialization Status
In spite of a promising market, and the potential for 2% increases in first pass yield, the
team has not yet succeeded in commercializing the Intelligent Extruder as system as of
the writing of this report. A key barrier may be the complexity of the system coupled
with the requirement to maintain a data-base for the models used based on material
grades. Although efforts have shown various means to streamline this data-base through
adaptive control, more work needs to be done to stream-line the effort required. In
simple terms, the team needs to invest more into developing the value story in the context
of specific market opportunities. Readers interested in using, adapting or extending any
part of this work should contact GE Industrial Systems or Coperian Werner & Pfleiderer
to initiate further discussion, including assessment of the applicability to their process.
13 Conclusions and Follow on Recommendations
Concepts developed in this program allow waste and energy reduction in polymer
compounding operations for high value engineering materials. With advanced diagnostic
and control software applied to existing extruder drive systems, benefits are obtained
from a continuous quality audit synthesized by inferring material properties from readily
measured machine variables. This allows rapid detection of out of spec material and
corrective action in contrast to the infrequent quality checks that are performed today. We
estimate that about 2% of first pass yield losses can conservatively be attributed to out of
spec material produced this way, and caught after manufacture or when received at the
customer. By increasing first pass yield by 2%, the 30 Billion pound per year engineering
plastics producers in North America would save some $391MM dollars in material and
energy costs, avoid land fill solids in excess of 600MM lbs, save 624,000 Mwhr of
electrical and equivalent steam energy, and remove 30MM lbs of volatiles from the
atmosphere. This also represents 600MM lbs of found capacity, equivalent to 51, 2000
lb/hr extruders operating two shifts, 5 days a week for a year.
All the algorithms developed can be run on readily available process control
instrumentation available off the shelf from multiple suppliers, using PCs, PLCs with
PC-class process cards, or plant level DCS systems. Maintenance and calibration of the
algorithms is required, but means are proposed to efficiently develop and maintain this
required data. Viscosity estimation within 10% was demonstrated on both 25-mm
research extruders and a 120-mm production system at GE plastics. A means to detect
and classify a number of common feeder and disturbance faults was developed, under the
restriction only one fault at time occurs (though this can be relaxed), and the system was
demonstrated on the research extruder under various operating conditions. Diagnosis is
based on certain characteristic process models. Parameters specific to machine geometry
are identified once, and parameters specific to a grade must be estimated. But means to
efficiently adapt the material parameters around a nominal grade is demonstrated,
DE-FC02-99-CH10972
88
I nt e llige nt Ex t rude r for Polym e r Com pounding
simplifying the required database that must be developed. While the methods developed
do not eliminate the need for a QA lab, the load for QA testing should be greatly reduced
In this program, the power of exploiting greatly simplified but dynamic process models
was demonstrated. It is believed that more information is contained in machine variable
data, particularly torque at frequencies at and above the rotation speed of the machine.
Though we did not have time or resource to conduct experiments with it, a new torque
observer algorithm was developed and studied in simulation. Because no extruder testing
was actually conducted, the results are relegated to Appendix B, which shows that “high”
frequency reaction torque components can be synthesized from measurements of voltage
and current (and/or power) in any AC drive. By eliminating troublesome strain gauges or
complex torque sensors, this algorithmic technique we believe can form the basis of new
diagnostics to complement those developed in this program, and should be implemented
and tested in a future research program.
Two program goals that were not achieved were the successful commercialization of the
system, and the feedback control system scale-up demonstration. The full scale
production system used for validating scale-up did not have provision for closing the loop
on extruders, and changes needed were not high enough priority to GE Plastics given
their commercial production constraints. There is no fundamental barrier to
demonstrating this capability in a follow up study when a willing partner can be
identified. As to commercialization, the GE Industrial Systems sales team has curtailed
further investment in marketing this technology until an appropriate customer /partner
can be identified. Parties interested in some or all aspects of Intelligent Extruder should
contact GE Industrial Systems through their regional office for further discussions.
14 References
1. Houpt, Paul K., Campo, Peter, U.S. Patent 5,559,173, System for Controlling the
Color of Compounded Polymers using In-Process Measurements, Sept. 24 1996
2. Campo, Peter J., Schneiter, John L., and Dixon, Walter V., U.S. Patent 5,526,285,
Imaging Color Sensor, Jun. 11, 1996
3. Gertler, Janos J., “Survey of Model Based Failure Detection and Isolation in Complex
Plants, IEEE Control Systems Magazine, December 1988.
4. Isermann, Rolf, “On Fuzzy Logic Applications for Automatic Control, Supervision
and Fault Diagnosis,” IEEE Transactions of Systems, Man and Cybernetics, Vol. 28,
No.2, March 1998
5. Pabedinskas, A., Cluett, W.R., “Controller Design and Performance Analysis for A
Reactive Extrusion Process,” Polymer Engineering and Science, Vol. 34, No. 7April
1994
6. Gottfert, A., “Real time viscosity control with capillary rheometer, Kunststoffe 76,
Vol. 12, 1986,p.1200
7. Dealy, D.M., Broadhead, T.O, “Process Rheometers for Molten Plastics: A Survey of
Existing Technology,” Polymer Engineering and Sci., Vol. 33, No. 23, Dec.1993.
DE-FC02-99-CH10972
89
I nt e llige nt Ex t rude r for Polym e r Com pounding
8. McKay, B. et al, “Extruder Modeling: A Comparison of Two Paradigms,”,
UKACC/IEE International Conference on Control, Conf. Publication No. 427, 1996
9. Eerikainen, T., Zhu Y-H, Linko, P., “Neural networks in extrusion process
identification and control,” Food Control Volume 5, No. 2, 1994, p. 111
10. Hansen, M.G., Vedula, S., “In-line Fiber-Optic Near-Infrared Spectroscopy:
Monitoring of Rheological Properties in an Extrusion Process.” Journal of Applied
Polymer Science, Vol 68, 1998, p.859
11. Gao, J., Walsh, G. C., Bigio, D., Briber, R. M., Wetzel, M. D., “A Residence Time
Distribution Model for Twin Screw Extruders,” AIChE J., vol 45, no. 12, pp 25412549, 1999.
12. Gao, J., Walsh, G. C., Bigio, D., Briber, R. M., Wetzel, M. D., “Polymer Eng. &
Sci.,” vol 40, no. 1, pp 227-237, 2000.
13. Curry, J, Anderson, P., “Controlling the Crosslink Density of Co-Reactive Polymers
in an Extruder Reactor,” SPE ANTEC Papers (1990b) 36 p.1938
14. Curry, J., Jackson, S., Stoehrer, B, van der Veen A., “Control Strategy for Free
Radical Degradation of Polypropylene via Reactive Extrusion”, Chem. Eng. Prog.
(1988) 84 p.43
15. Curry, J. , Kiani, A., “Experimental Identification of the Distribution of Fluid Stresses
in Continuous Melt Compounders, Part. II,” SPEC ANTEC Papers(1991) 114, p.37
16. Jones, R.W and McClelland, J.F, “On-line analysis of solids and viscous liquids by
transient infrared spectroscopy,” Process Control and Quality, 4 (1993), p. 253-260
(Elsevier, Amsterdam).
17. Gertler, Janos J, Fault Detection and Diagnosis in Engineering Systems, Marcel
Dekker, NY 1998
18. Ljung, Lennart, System Identification: Theory for the User, Prentice-Hall, N.J. 1987
DE-FC02-99-CH10972
90
I nt e llige nt Ex t rude r for Polym e r Com pounding
APPENDIX-A Extruder Dynamic Models
A.1 Dynamic model for Hold up
The transient variation in the holdup M1 due to changes in total feed-rate Q and screw
speed N is described simply by the total material balance:
Eq-A 1
dM 1
= Q - Q1o
dt
In the above equation, the total inlet feed-rate to this section (from the feeders) is Q while
the total outlet mass flow rate, denoted by Q1o , varies with the operating conditions, in
particular the fill fraction φ (i.e. the fraction of the total void volume filled with the
material holdup) and the screw speed N. More specifically, the maximum flow capacity
of this section Q1fc corresponding to the maximum filled capacity M1fc (based on the void
volume from screw geometry) is proportional to the screw speed N, i.e., Q1fc = k N with
the proportionality constant k depending on the screw design/geometry. During regular
operation, when this section is only partially filled and the fill fraction is φ =M1 / M1fc ,
(0 < φ < 1), the total outlet mass flow rate is given by
Q 1o = k φ N
Eq-A 2
M 1
= k
M
1 fc
M 1N
=
B
N
where B=M1fc / k is a parameter that depends only on the screw design/geometry.
Combining equations Eq-A 1 and Eq-A 2 gives the dynamic mass balance relation for the
holdup M1
Eq-A 3
dM 1
M N
=Q- 1
B
dt
Note that at steady state, the inlet and outlet mass flow rates are equal, i.e. Q=Q1o, and
the dynamic material balance in Eq-A 3 reduces to the steady state version: M1 = BQ/N.
In contrast with the partially filled section, the total holdup M2 in the filed section is
constant (since the void volume is filled to maximum capacity). Furthermore, the outlet
flow rate from this filled section is always the same as the inlet flow rate, which in turn is
the same as the outlet flow rate from the partially filled section, i.e. Q1o.
DE-FC02-99-CH10972
91
I nt e llige nt Ex t rude r for Polym e r Com pounding
A.2 Dynamic model for composition
partially filled
section 1
Feedrates Q1 Q2
T
Speed N
completely filled
section 2
R1*Q1o
M 1 , x1
Qo , xo
(µ)
M 2 , x2
M 1 , x1
Q
xi
DP
R2*Q2o
Q1o
x1
M 2 , x2
Q2o
x2
Figure 52: Schematic representation of mixing in partially and completely filled
sections.
A way to obtain a simple parameterized model that can capture varying degrees of
mixing, from no mixing at one extreme to perfect instantaneous mixing at the other, is to
use a combination of plug flow with recycle. More specifically, consider the schematic
representation in Figure 52, where raw material is fed to the partially filled section at total
flow rate Q = Q1+Q2 and composition xi = Q1 /(Q1+Q2), and exits from this section to the
completely filled section at a total flow rate Q1o and composition x1. The internal mixing
inside this section can be modeled through a recycle with a flow rate R1*Q1o , where R1
denotes the ratio of recycle to outlet flow rate. The combination of the recycle stream and
the feed stream yields a net inlet stream with a total flow rate Q1i = Q+R1*Q1o and
composition x1i = (Q*xi+R1*Q1o*x1)/(Q+R1*Q1o).
If we model the material flow through the section as pure plug flow (i.e., first in first out)
then its dynamics are described by a simple time delay, i.e.,
x1(t) = x1i(t-td1)
with a time delay
td1 = M1 / (Q1o*(1+R1))
governed by the total material holdup and the total outlet flow rate. The above relations in
time can be expressed equivalently in the Laplace domain:
Eq-A 4
x1i ( s) =
(Q * x i ( s) + R1 * Q1o * x1 ( s))
(Q + R1 * Q1o )
x1 ( s ) = x1i ( s )e −t d 1s
Solving for x1i(s) from the above relations yields the overall input-output relation
between the inlet composition xi(s) and the outlet composition x1(s):
Eq-A 5
Qe − t d 1 s
x1 ( s ) =
xi (s)
Q + R 1 * Q 1 (1 − e − t d 1 s )
DE-FC02-99-CH10972
92
I nt e llige nt Ex t rude r for Polym e r Com pounding
The above relation provides a general mixing relation for the partially filled section,
where the recycle ratio R1 is the parameter that can be fit to capture the actual degree of
mixing in a given screw design. In particular, the case R1=0, i.e. no recycle, corresponds
to no mixing and the input-output relation in Eq-A 5 reduces to a pure time delay
x1(s) = xi (s)e -td 1 s ,
Eq-A 6
where t d 1 =
M1
Q1o
On the other hand, the case with infinite recycle, i.e. R1= ∞ , corresponds to perfect
instantaneous mixing and the input-output relation reduces to a simple first order
response
x1 ( s ) =
Eq-A 7
M
1
xi ( s ), where τ = 1
τs + 1
Q
So, the parameter R1 can be fit with measured input-output data to capture the actual
degree of mixing in the section for a particular screw configuration. Similarly, the mixing
in the completely filled section can be captured through a combination of plug flow and
recycle to obtain the following input-output relation:
Eq-A 8
x 2 (s) =
e −td 2 s
x1 ( s ),
1 + R2 * (1 − e −t d 2 s )
where t d 2 =
M2
Q2 o
& Q2o = Q1o
Section 1: Torque relation
We develop the following relationship for the total torque
Eq-A 9
T = α 0 + α 1 M 1 (α 2 + N ) x 1 + α 3 M 2 Nx 2
which has three main components corresponding to the offset and the contributions from
the partially-filled and completely-filled sections. The above equation yields torque as an
instantaneous function of the total holdups (M1, M2), the compositions (x1, x2), and the
screw speed N – it is an algebraic output map relating the process inputs/states to the
output in a standard state-space description. The validity of such a relation can be tested
against steady state input-output data from the extruder. In particular, at steady state, the
relations in Eq 1 and Eq 5 can be used for the holdups and compositions to arrive at the
equivalent relation for steady state torque:
Eq-A 10
T = α 0 + α 1α 2 B
Q
x i + α 3 ANx i + α 1 BQxi
N
Section 2: Die pressure relation
DE-FC02-99-CH10972
93
I nt e llige nt Ex t rude r for Polym e r Com pounding
128L
DP = 4 Qo µ T
πD ρ
Eq-A 11
= kQo µ
In the above relation, L, D denote the length and diameter of the pipe, ρ is the density of
the material and they are lumped into an unknown parameter k. Furthermore, Qo is the
flow rate of the product through the die plate and µ denotes the viscosity of the product
at the prevailing temperature and shear rate at the die. The viscosity of the product
depends on the composition, i.e. the weight fraction xo of PPO, temperature To in the melt
zone and the shear rate on the molten product as it flows through the die plate holes.
While the melt temperature To is measured using a thermocouple in the melt pool just
before the die plate, the shear rate is considered to be proportional to the total material
flow rate Qo.
We will approximate the nonlinear dependence of the product viscosity on composition,
temperature and shear rate with a linear approximation that is valid in a local
neighborhood of the nominal point of operation. In particular, we will use the following
linear approximation for the product viscosity:
µ = µ o + µ1 ( xo − xo ) − µ 2 (Qo − Qo ) − µ3 (To − To )
Eq-A 12
Viscosity (Pas)
where ( ) denotes respective nominal steady state values at the nominal operating point.
In the above linear approximation for viscosity, µο denotes the nominal viscosity at the
nominal operating point, while µ1, µ2, µ3 denote the slope of viscosity with respect to
composition, shear rate and temperature, respectively at this operating point (see Figure
53 )
µο
slope = µ1
xo
product
composition xo
composition
Figure 53: Viscosity vs composition of PPO
Inserting the linear relation for viscosity from Eq-A 12 in
DE-FC02-99-CH10972
94
I nt e llige nt Ex t rude r for Polym e r Com pounding
Section 2: Die pressure relation
Eq-A 11,
we obtain the following model for die pressure:
DP = kQo [ µ o + µ1 ( xo − xo ) − µ 2 (Qo − Qo ) − µ 3 (To − To )]
Eq-A 13
The above equations together comprise a dynamic model for the extruder that relate the
changes in the process inputs (feed-rates, screwspeed) to the measured output variables
(torque, die pressure):
Eq-A 14
dM 1
= Q - Q1o
dt
M2 = A
x1 ( s ) =
M 1N
)
B
(Q = Q1 + Q2 ,
Q1o =
Qe− t d 1s
xi ( s ) ,
Q + R1 * Q1o (1 − e − t d 1s )
( xi =
Q1
,
Q1 + Q2
( td 2 =
M2
)
Q1o
e −t d 2 s
x2 ( s ) =
x1 ( s ),
1 + R2 * (1 − e −t d 2 s )
td 1 =
M1
)
Q1o (1 + R1 )
T = α 0 + α1M 1 (α 2 + N ) x1 + α 3M 2 Nx2
DP = kQ1o [ µ o + µ1 ( x2 − x2 ) − µ 2 (Q1o − Q1o ) − µ3 (To − To )
Section 3: Linearized DP equation and Nonlinear DP equation
Linear form:
Eq-A 15
_
DP= [DP− kQo µo + kQo 2 µ2 ] + [k(µo − µ1xo ) − kQo µ2 ] Qo + (kµ1 )Qo x0 − kµ3Qo ∆T
or written in a compact form
Eq-A 16
DP = β o + β 1 Q o + β 2 Q o x o − β 3 Q o ∆ T
The above relations for die pressure DP yield a simple relationship between the model
parametes βi and the physical parameters µi.
On the other hand, if we adopted a model structure for die pressure without any
linearization, i.e. retaining the same structure as equation, Eq-A 13:
DP = kQo [ µ o + µ1 ( xo − xo ) − µ 2 (Qo − Qo ) − µ 3 (To − To )] ,
Eq-A 17
then it can be written in the following compact form:
Eq-A 18
DP
= β 1Q
DE-FC02-99-CH10972
o
+ β 2Q o ∆ x o − β 3Q o ∆ Q
o
− β 4Q o ∆ T
95
I nt e llige nt Ex t rude r for Polym e r Com pounding
In this new nonlinear structure, comparing Eq-A 17 and Eq-A 18, we obtain another
simple relationship between µ0, µ1 and β1, β2 involving only the calibration parameter k
and independent of any other terms. More specifically,
Eq-A
β1 = kµ 0
β 2 = kµ 1
This simplification of calculation for µ0, µ1 from the identified parameters β’s results in
better estimates and improved diagnostics.
Section 4: Model for Melt Temperature variation
In the absence of good melt temperature measurement, we can use a simple first-order
dynamic model to predict the variations in the melt temperature from its nominal value.
In particular, assuming the heat input from the heating barrels to the extruder is constant
while the heat generated in the extruder due to friction in the screws varies as a function
of the operating conditions, we can obtain the model for melt temperature variation
through a simple dynamic heat balance. It should be mentioned that the ratio of heat input
from the barrels to the heat generated internally due to friction in the screws is smaller in
larger industrial-scale extruders compared to the smaller lab-scale extruders due to the
reduced surface area to extruder volume. The heat generated H by the screws is
proportional to the shaft work, i.e. H = k1*T*N, where T is the shaft torque, N is the
screw speed and k1 is a proportionality constant. This heat generated is used to melt the
solid raw materials and raise the temperature of the molten product to the outlet
temperature To from the inlet temperature Ti, i.e. at steady state,
H = k1 * (T * N ) = Qo (∆H m + c p (To − Ti ))
Eq-A 19
or
T *N
H
= k1 *
= ∆H m + c p (To − Ti )
Qo
Qo
where Qo is the throughput, ∆Hm is the heat of melting the raw materials, cp is the specific
heat capacity of the molten material, Ti is the inlet temperature (assumed constant) and To
is the outlet melt temperature. Since, we need only changes in the melt temperature from
the nominal steady state value Tos corresponding to the nominal values of Ts, Ns, Qos, we
can obtain the following relation for the change in melt temperature by subtracting the
above relation from a similar relation at the nominal steady state:
Eq-A 20
∆T = To − Tos = k 2 (
T *N
T * N Ts * N s
) = k2 * ∆
−
Qo
Qo
Qos
where k2 is another proportionality constant and ∆(T*N/Qo) denotes the deviation in the
specific energy generation, SE= (T*N/Qo), from the nominal steady state value. Finally,
since we are interested in the dynamic transient responses of the changes in melt
temperature, we get it by using a first-order transient response in conjunction with the
above steady state value (the first-order response can be obtained by performing a simple
lumped dynamic energy balance):
DE-FC02-99-CH10972
96
I nt e llige nt Ex t rude r for Polym e r Com pounding
∆T =
Eq-A 21
k2
T *N
*∆
τs + 1
Qo
where τ is the time constant for the transient response in the melt temperature. In the
absence of a good measurement of melt temperature, the above relation for ∆T can be
used in the die pressure relation (Eq-A 16 or Eq-A 18).
Section 5: Closed-loop Control
The process response between the product viscosity µ and the feed composition xi has the
following (linearized) first-order dynamics:
Eq-A 22
µ − µo =
µ1
( xi − x i )
τ ps +1
Since IMC tuning (see, e.g. Morari and Zafiriou, 1989 [3]) of a first order process such as
the one given above, for a corresponding first-order closed-loop response:
Eq-A 23
∆µ
1
=
∆µ SP τ f s + 1
yields a Proportional and Integral (PI) action controller, a PI controller is tuned for robust
control of this process with Internal Model Control (IMC) rulings. The resulting PI
controller in Laplace domain is in the following form:
Eq-A 24
C (s) =
τ ps +1
µ1 (τ f s )
=
τp
1
1 +
µ1τ f τ p s
in which the proportional gain and the integral time constant are apparent in terms of the
process parameters µ1 and τp, and the IMC tuning factors, i.e. the closed-loop time
constant τf.
A.3 References for Appendix A
1. Gao, J., Walsh, G. C., Bigio, D., Briber, R. M., Wetzel, M. D., “A Residence Time
Distribution Model for Twin Screw Extruders,” AIChE J., vol 45, no. 12, pp 25412549, 1999.
2. Gao, J., Walsh, G. C., Bigio, D., Briber, R. M., Wetzel, M. D., “Polymer Eng. &
Sci.,” vol 40, no. 1, pp 227-237, 2000.
3. Morari, M. and Zafiriou, E. Robust Process Control, Prentice-Hall, Inc., New
Jersey, 1989.
DE-FC02-99-CH10972
97
I nt e llige nt Ex t rude r for Polym e r Com pounding
APPENDIX B - Extruder Drive Torque Estimation using
Electrical Variables
B.1 Summary
Reaction torque from a loaded extruder is one potential indicator for process health. In
prior unpublished studies by Coperian Werner-Pfleiderer, strain gauges in line with the
drive shaft on special research extruders showed that there was potential information
contained in the shaft torque at frequencies at or above the shaft rotation rate which could
be correlated to proper filling, mixing performance and screw condition, for example.
Unfortunately, strain gauges are not easy to install, maintain and calibrate in a production
environment, so definitive work on interpreting relationships between dynamic torque
behavior and quality has not received much attention. Here we propose some alternative
methods for extracting high frequency torque dynamics which do not rely on strain
gauges or any other mechanical sensor in the drive line. The proposed methods instead
exploit readily available electrical sensors, which are commonly found in modern drives
used in extruders and which can often be retrofit to older machines. Software algorithms
are derived which use these electrical measurements to extract torque data that is usually
removed by filtering since it is not required for the functioning of the motor control.
Scope of this development is limited to preliminary simulations, since we did not have
time or resources to evaluate what quality attributes and/or diagnostics can be rigorously
linked to the torque data; we have included these results as a foundation for future
“Intelligent Extruder” developments.
B.2 Requirements and Objectives
An “observer” is a software algorithm which takes as inputs certain measured variables
and produces as outputs unmeasured variables of interest, such as torque. For this project
the requirements for the observer are as follows:
1) Obtain a method to determine the fundamental component and harmonic
components of the electrical torque produced by a synchronous or asynchronous
machine driving an extruder. Machine is assumed to operate at speed > 0 rpm.
2) Provide means that the observer can be implemented:
• As part of the drive controller.
• As a separate unit (e.g. PC or process equipment) able to interface with any
inverter.
3) The observer should have a frequency response reaching up to a couple of kHz to
be able to observe the extruder dynamics of interest
4) Motor constants are assumed available when used stand alone.
DE-FC02-99-CH10972
98
I nt e llige nt Ex t rude r for Polym e r Com pounding
5) The observer will be simulated using SABER with the best possible
approximation to the GE-Innovation drive control.
Requirement (1) assures that the methods will be applicable to the majority of production
scale extruders installed today, and normal operation is always at non-zero speed allows
simplification. Frequency response in (2) is a conjecture based on suggestions by
Coperian W-P, but the actual frequency range of requirement must be established by
experiment. Since we were not able to implement the torque observer, the SABER
simulation provided a well validated drive simulation we could could use to at least
provide a proof of concept.
B.3 Approaches to observer based torque estimation
Instantaneous electromagnetic torque computation can be derived by implementation of
the following expression:
−
−
t e = 32 P λ sg × i sg
Eq B- 1
−
Where P is the number of pole-pairs, λ sg is the space phasor of the stator flux linkages
−
expressed in a general reference frame and i sg is the space phasor of the stator currents in
the same general reference frame. Eq 1 can be further represented using two-axis
components for stator flux and stator currents as:
t e = 32 P(λ sd i sq − λ sq i sd )
Eq B-2
In general, there are two ways to obtain the flux value generated by a machine:
t
λ (t ) = ∫ v(t )dt
Eq B- 3
0
Or
λ (t ) = kL ⋅ i (t )
Use of Eq 3 leads to a method known as the Voltage model [1]. On the other hand, using
Eq 4 leads to a method known as Current model [2]. Both models are depicted in Figure
1.
EqB-4
iu
iα
T iv
iβ
γm
Lm
1+sTr
Lm
1+sTr
λαs
λ urr
-1
rr T
λv
γm
iβ
Current Model
X
λ βs X
iα
-
3 np
2
Telectric
Torque
Calculator
(a)
DE-FC02-99-CH10972
99
I nt e llige nt Ex t rude r for Polym e r Com pounding
Lσ
iα
Rs
uα
s
r
s
λ
λ
β
β
1
uβ
iβ
- 1 λ α s -λ α r
s
Rs
-
Lσ
Voltage Model
iβ
X
3 Lm n
2 L2 p
Telectric
X Torque
iα
Calculator
(b)
Figure 1: Block diagram: (a) Current model. (b) Voltage model
The Current model requires sensing of motor currents as well as motor parameters such
as mutual inductance and electric time constant and coupling inductances. Computation
of angle γm requires the actual rotor mechanical speed. But no motor voltages are required
at all. On the other hand, the Voltage model requires motor voltages and currents plus
some motor parameters such as stator resistance and inductances. Voltage model requires
an integrator as shown in Eq 3; the use of integrators at low speeds may lead to saturation
yielding in inaccurate torque calculation. Since both methods employ electric motor
parameters it is important to note that those parameters change under different
environmental and electrical conditions, e.g. stator resistance is affected by temperature
and inductance may saturate.
DE-FC02-99-CH10972
100
I nt e llige nt Ex t rude r for Polym e r Com pounding
Current model
• Parameter Sensitive (temperature,
inductance saturation)
• It works well at low frequencies.
• Requires the motor mechanical speed
• Not suitable for analog implementation
• Resulting torque has the full frequency
spectrum
• DSP implementation:
Resulting torque have limited frequency
spectrum due to sensors and sampling
•
•
•
•
•
•
Voltage model
Works well at high frequency.
Integrator has problems with sensors
offset
Does not work at zero freq.
Parameter sensitive (Temperature)
Analog implementation possible:
Complicated to implement in analog
circuitry
Integrators have to be replaced by 1st
order transfer function
Resulting torque have the full frequency
spectrum
DSP implementation:
Requires complicated voltage sensor
(VCO’s, ASIC)
Resulting torque have limited frequency
spectrum due to sensors
Table B-1: Salient characteristics for voltage model and current model for torque
estimation
A combination of methods mentioned above can be achieved as shown in Figure 2 [3].
iu
iα
T
iβ
iβ
Rs
uα
λ urr
Current Model
- 1 λ α s -λ α r
1
s
Rs
-
-
Ki
Kp +
s
∆ λαs
λ ur
Ki
Kp +
s
-1
λ vr T
λ vrr Field Corrector
γm
∆ λ βs
Corrected
Voltage Model
λu
λ βs - λ βr
γm
λ βs
rs
T λ vrs
Lσ
-
rs
λ vrs
s
uβ
λu
λ vrr
Lm
1+sTr
Lσ
iα
λ urr
1+sTr
iv
γm
∆ λ αs
∆ λ βs
Lm
iα
X
iβ
X λαs
3 np
2
Telectric
Torque
Calculator
Figure B-2: Current/voltage model
DE-FC02-99-CH10972
101
I nt e llige nt Ex t rude r for Polym e r Com pounding
Salient features of this method are presented in Table B-2
Current/Voltage model
•
The use of both models allows the extension of the operating region down to zero
frequency.
• Requires knowledge of motor parameters.
• Digital signal processing CPU required for implementation due to its complexity.
• Requires complicated voltage and current sensors.
Table B-2: Salient characteristics of Current/Voltage model
Another method, based on machine’s electric power, is derived from the following
expression:
Eq B-5
P = te ⋅ωe
Where P is the electric power and ωe is the machine angular speed. Figure 3 and Table 3
show the main features of this method.
v1
v2
v3
vα
va
3/2
U/V
vb
ia
ib
Power Model
vα
-
ωe
Rs
vβ
iα
X
iβ
iα
3/2
X
Rs
iβ
vβ
D
N
:
3 np
2
Telectric
Active
Power
-
Figure B- 3: Power model
Power model
• Less sensitive to parameters variation.
• It works only for frequencies not equal zero due to division.
• Easily implemented with analog circuits.
• Requires the applied frequency from the inverter or a two phase PLL for its
generation
• Can provide the full torque spectrum if implemented in analog fashion.
• The model is valid for synchronous and asynchronous motors.
DE-FC02-99-CH10972
102
I nt e llige nt Ex t rude r for Polym e r Com pounding
•
Two line-to-line voltages could also be used as long as they are isolated (LEM
voltage sensors).
• A small DSP could be used to set up gains and parameters from a PC serial interface.
• DSP could also include thermal models to correct for resistance variations.
Table B-3: Salient characteristics of Power model
B.4 Saber Simulation
Results regarding the Power model are presented in this section. Figure 4 shows the
schematic diagram for SABER used to simulate the motor drive and the torque observer.
Some basic assumptions for the simulation are:
• Detailed induction motor model considering only copper losses and saturation.
• Detailed model of a PWM inverter considering: dead-time, device losses and DC
voltage ripple.
• Sampling delays and quantization are included in the speed-current control for the
motor model.
The following parameters correspond to the values normally found in GE Innovation
Series controller. These are readily modified for other motors or controllers.
DC link voltage=900V
PWM frequency=2.5KHz
Sampling Time=250µs
Power=15KW
Pole pairs=2
rs=0.319Ω
rr=0.220Ω
ls=0.00269H
lr=0.004139H
lmo=0.077H
Table B-4: Parameters used in simulation
Figure B-4: Schematic diagram used for SABER simulation
A virtual torque sensor (ideal) in the simulation is attached to the motor to obtain the
“actual” torque. Information from this torque sensor is compared with results from the
torque observer for validation. Figure 5 shows the behavior of the motor during the
DE-FC02-99-CH10972
103
I nt e llige nt Ex t rude r for Polym e r Com pounding
starting process. Waveforms for motor speed in rad/s, motor current in Amps, observed
torque in Nm and motor torque in Nm are presented. Observed torque represents the
torque estimated using the Power model.
Starting with 20 Nm as a constant load
(-) : t(s)
100.0
Rotor Freq[Rad/s]
(-)
80.0
60.0
40.0
20.0
(A) : t(s)
(A)
20.0
Motor Current[A]
0.0
-20.0
(-) : t(s)
40.0
Torque Observer[Nm]
20.0
(-)
0.0
-20.0
-40.0
-60.0
(-) : t(s)
40.0
Motor Torque[Nm]
(-)
20.0
0.0
-20.0
0.0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
t(s)
Figure B-5: Behavior of motor at starting. From top to bottom: Rotor speed, Motor current,
Observed torque, Motor torque.
It can be noticed that the observed torque is following the actual torque very closely and
settles down at steady state.
Figure B-6 is a continuation in time of Figure B-5. It shows the motor running at 40 rad/s
under no-load conditions (T≈0Nm) when a step load of 20Nm is applied at t=75ms. In
consequence the torque developed by the motor jumps from 0Nm to 20Nm. This test
shows the ability of the observed torque to follow sudden load changes.
DE-FC02-99-CH10972
104
I nt e llige nt Ex t rude r for Polym e r Com pounding
Step response of 20Nm at 40Rad/s
(-) : t(s)
40.0
Observer_Torque[Nm]
(-)
20.0
0.0
-20.0
(-) : t(s)
40.0
torque(ind_mot1.ind_mot1)
(-)
20.0
0.0
-20.0
(-) : t(s)
60.0
Rotor_speed[Rad/s]
(-)
40.0
20.0
0.0
(A) : t(s)
(A)
20.0
Motor_Current[A]
0.0
-20.0
0.06
0.065
0.07
0.075
0.08
0.085
0.09
0.095
0.1
0.105
0.11
0.115
0.12
0.125
0.13
0.135
0.14
0.145
0.15
t(s)
Figure B-6: Behavior of motor during a step load. From top to bottom: Observed torque, Motor
torque, Rotor speed and Motor current.
Step response 0 to 20Nm to 0Nm
(-) : t(s)
100.0
Rotor_freq(Rad/s)
(-)
80.0
60.0
40.0
20.0
0.0
(-) : t(s)
40.0
Observer_torque(Nm)
(-)
20.0
0.0
-20.0
-40.0
(-) : t(s)
40.0
Motor_torque(Nm)
(-)
20.0
0.0
-20.0
-40.0
(A) : t(s)
20.0
Motor_current(A)
(A)
10.0
0.0
-10.0
-20.0
0.05
0.075
0.1
0.125
0.15
0.175
0.2
0.225
0.25
t(s)
Figure B-7: Behavior of motor during sudden application and demotion of load. From top to
bottom: Rotor speed, Observed torque, Motor torque and Motor current.
Figure B-7 shows the motor running at approx. 40 rad/s under driving a load of 20Nm
applied at t=75ms. Later, the load is removed at 0.15s. This figure shows the ability of
the observer to follow step loads when the motor is loaded or unloaded.
DE-FC02-99-CH10972
105
I nt e llige nt Ex t rude r for Polym e r Com pounding
FFT torque at 1200RPM
dB(-/Hz)
(343.03, 17.083)
dB(-/Hz)
40.0
: f(Hz)
fft(Observer)
20.0
0.0
-20.0
-40.0
-60.0
dB(-/Hz)
dB(-/Hz)
(343.44, 17.114)
40.0
: f(Hz)
fft(Motor)
20.0
0.0
-20.0
-40.0
0.0
0.5k
1.0k
1.5k
2.0k
2.5k
3.0k
3.5k
4.0k
4.5k
5.0k
5.5k
f(Hz)
(-) : t(s)
60.0
Observer_torque(Nm)
(-)
40.0
20.0
0.0
-20.0
-40.0
(-) : t(s)
60.0
Motor_torque(Nm)
(-)
40.0
20.0
0.0
-20.0
-40.0
0.1
0.105
0.11
0.115
0.12
0.125
0.13
0.135
0.14
0.145
0.15
t(s)
Figure B- 8: Torque during steady state. Top traces: FFT of Observed torque, FFT of Motor torque.
Bottom traces: expanded view of Observed torque and Motor torque respectively.
Bottom traces of Figure 8 shows the response of the torque with oscillation due to noncompensated dead-time at inverter side. The frequency of this oscillation is six times the
fundamental frequency. It can be noticed that even under irregular conditions the
observed torque is able to follow the actual motor torque. Top traces in same figure
shows the frequency spectrum for the aforementioned torque traces. The magnitude of
fundamental torque component is shown to illustrate the accuracy of the observed torque.
B.5 Conclusions from simulation results
• The observer using the Power model is able to detect the torque with a wide
bandwidth if implemented using analog components, allowing reaction torques at perrevolution resolution to be seen for use in advanced diagnostics
• The observer based on the Power model could be combined with a small
microprocessor or DSP and multiplying D/As to input the motor parameters and
correct them on line for temperature variations.
Since no attempt was made in the scope of this program to implement these concepts on a
functioning extruder, it remains to validate the value of high frequency information in the
estimated torque to diagnose process and equipment problems in extruders and drives.
DE-FC02-99-CH10972
106
I nt e llige nt Ex t rude r for Polym e r Com pounding
B.6 References for Appendix B
[1] P.L. Jansen, R.D. Lorenz and D.W. Novotny. Observer-based direct field
orientation: analysis and comparison of alternative methods. IEEE Transactions on
Industry Applications, vol. 30, no. 4, July/August 1994.
[2] C. Lascu and A.M. Trzynadlowski. A TMC320C243-based toque estimator for
induction motor drives. IEEE International Electric Machines and Drives Conference,
pp.733-735, 2001.
[3] S. Kühne and U. Riefenstahl. A new torque calculator for AC induction motor
drives that improves accuracy and dynamic behavior. In Proceedings of the IEEE
International Symposium on Industrial Electronics, pp. 498-503, 1999.
[4] “Sensorless vector control and direct torque control”: Peter Vas; Oxford University
Press, 1998.
[5] “Dynamic simulation of electric machinery”: Chee-Ming Ong; Prentice Hall PTR
1997.
DE-FC02-99-CH10972
107
Related papers
Azerbaycan-Türk Dilinin Lügat Terkibinin Zenginleşmesinde Morfoloji ve Sentaktik Usulün Rolü
DergiPark (Istanbul University), 2022
L’architecture par l’intérieur : concepts et imaginaires d’une discipline en devenir
Critique d’art, 2021
Remote Sensing and GIS Modelling of Roman Roads in South West Britain
Journal of Computer Applications in Archaeology, 2023
Perspectivas del concepto de poder en Byung-Chul Han
Zenodo (CERN European Organization for Nuclear Research), 2022
El amor según los médicos
ASEMEYA - Asociacion de medicos escritores y artistas, 2023
Social inequalities in maternal opinion of child development in southern Brazil
Acta Paediatrica, 2007
COMPARIson between capacitive and photo sensors in depth control of onion harvester
Journal of Engineering Studies and Research, 2011
Magnetic Resonance Imaging of the Brain in Diabetes
Diabetes, 2004
Reconstruction of Premature Atrial Contraction and Premature Ventricular Contraction on ECG Traces by Applying PLA as Segmentation Process
Journal of Image and Graphics, 2015
Atmospheric Pressure Gas-Phase H/D Exchange of Serine Octamers
Analytical Chemistry, 2003
Tarihsel Süreklilik Olarak Mimarlık: Mimari Tasarım Eğitimi’nde Tarihsel Nesne Olarak Kent ve Mimari Yapı
Her Dem Yeşil Yapraklı Bir Ağaç: Cevat Erder’e Armağan, 2003
The bubble frequency in a fluidized bed at elevated pressure
Powder Technology, 1990
Gastrostomy button: why complicate an office procedure?
Gastrointestinal Endoscopy, 1989
- Find new research papers in:
- Physics
- Chemistry
- Biology
- Health Sciences
- Ecology
- Earth Sciences
- Cognitive Science
- Mathematics
- Computer Science