2017 IEEE 33rd International Conference on Data Engineering

Vibration Analysis for IoT Enabled Predictive

Maintenance

Deokwoo Jung Zhenjie Zhang Marianne Winslett

Big Data Technology Lab. Advanced Digital Sciences Center Department of Computer Science
SK Telecom Co. Ltd. Seoul Illinois at Singapore Pte. Ltd. University of Illinois at Urbana Champaign
deokwoo.jung@sk.com zhenjie@adsc.com.sg winslett@illinois.edu

Abstract—Vibration sensor is becoming an essential part of New generation of vibration sensor, using MEMS ac-
Internet of Things (IoT), fueled by the quickly evolving technology celerometer as measurement device, is reshaping the equip-
improving the measurement accuracy and lowering the hardware ment servicing and replacement workflow in last few years.
cost. Vibration sensors physically attach to core equipments in It overcomes the limitations of conventional vibration sensor,
control and manufacturing systems, e.g., motors and tubes, pro- based on piezoelectric accelerometer, which is bulky, energy
viding key insight into the running status of these devices. Massive
hungry and expensive. These new vibration sensors are now
readings from vibration sensors, however, pose new technical
challenges to the analytical system, due to the non-continuous sufficiently light for easy installation, sustainable for months
sampling strategy for sensor energy saving, as well as hardness under battery power and cheap to purchase and replace. For
of data interpretation. To maximize the utility and minimize the the first time, vibration analysis is now applicable for system
operational overhead of vibration sensors, we propose a new administrator to unveil a handful of key insights into the
analytical framework, especially designed for vibration analysis system. Firstly, vibration directly reflects the operational status
based on its unique characteristics. In particular, our data of the equipments, mostly generated by mechanical behavior
engine targets to support Remaining Usefulness Lifetime (RUL) of the equipments, such as the rotation of the motors and
estimation, known as one of the most important problems in the running flow in the pipes. Secondly, vibration is fairly
cyber-physical system maintenance, to optimize the replacement independent of other external factors, e.g., temperature and
scheduling over the equipments under monitoring. Our empirical
humidity. The readings of the vibrations over the target devices
evaluations on real manufacturing sites show that scalable and
accurate analysis over the vibration data enables to prolong the provide more reliable information, than conventional sensors
average lifetime of the tubes by 1.2x and reduce the replacement could offer. Thirdly, vibration behavior is believed to evolve
cost by 20%. quickly with respect to the ageing equipments. It is thus a per-
fect indicator of early signs on the failures and problems, way
before the actual problem occurs to the devices. With all these
I. I NTRODUCTION enticing features, vibration sensors are becoming an essential
part of IoT in CPS, especially to support predictive mainte-
We are now at the step of the door to a new industrial era nance over the replaceable components. Instead of replacing
with the adoption of Internet of Things (IoT) in almost every the equipment at a fixed period, vibration-based predictive
Cyber Physical System (CPS). The key vision of the new era is maintenance allows the system to much more closely monitor
the possibilities enabled by IoT on more accurate monitoring and evaluate the equipments in real-time, to alert the system
over and real-time optimal response to the physical devices in administrator with servicing or replacement suggestions only
the system. Device failure is one of the most important classes when the equipment is about to fail or degenerate.
of events CPS expects to prevent, detect and fix, resorting to
the sensor and data analysis techniques in IoT. In a typical Predictive maintenance with vibration sensor data, how-
semiconductor fabrication plant (FAB), for example, thousands ever, demands new processing and analysis techniques in the
of vacuum pumps endlessly evacuate air from chambers in existing Manufacturing Execution Systems (MES) [1], [3],
semiconductor wafer fabrication, to constantly meet the strin- [4], [2]. The challenges are in two-fold. Firstly, due to the
gent vacuum condition. To maintain a required yield rate of spectrum nature of vibration data, the readings from vibration
wafers, factory operators usually adopt a conservative and risk sensors span in a high-dimensional space, only part of which
averse strategy for its pump replacement, by which every pump is useful to the maintenance prediction objective. Such high
must be replaced after it reaches a maximal working period dimensionality requires effective feature selection tools based
even if the pump remains in perfect condition. As we will on the spectral properties of vibration behaviors, in order to
present in our empirical evaluations, although a fraction of identify key dimensions for accurate device failure prediction.
the pumps keep healthy for over 24 months, they are now Secondly, the bandwidth between vibration sensor and base
replaced within the time frame of 6 months, because of the station is limited, due to the hardware and energy consumption
large variance on the lifetime distribution. While most of constraints on the sensors. The back-end analytical engine is
the manufacturers are aware of the waste, such replacement expected to handle asynchronous and incomplete observations
strategy is the only applicable option to the decision makers, from multiple sensors for predictive modelling.
due to the limited analytical capability over the pumps. Any
minor problem with a pump in operation could render huge In this paper, we present our novel solution to vibration
cost of defected products and stoppage of the pipeline. analysis and lessons learnt from our practice. We build a

2375-026X/17 $31.00 © 2017 IEEE 1273

1259
1271
DOI 10.1109/ICDE.2017.170
Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on August 29,2023 at 19:25:27 UTC from IEEE Xplore. Restrictions apply.
 * " !) #  ! " ) !  ! %!) TABLE I: Comparison between two generations of vibration
&' &' sensors based on piezoelectric accelerometer (Piezo Sensor)
 !  ( (
 " !
#  "" $  # $%% and MEMS.
* " ! #  !   ! "

    

Piezo Sensor MEMS Sensor
& $   #   
 ! "
Price US$ 300+ within US$ 10+
*  &"  
Power 27mW 3mW
Size 1.97×0.98×1 inch 0.2×0.2×0.05 inch
* + !! 
&'( Noise Density 700μg 4000μg
&!" !   
Resonance freq. 20 kHZ 22 kHz
 #  
&"! Accelerate range 10g 100g
".  !
/0"

  !% !$%%!%   

(///1234536 
,%-!"
" !"  + !  !%
+ ! major components, including a sensor mote component cap-

turing high-frequency vibration from the attached instrument,
a gateway component aggregating data from multiple sources,
Fig. 1: System Overview and an analysis component with data storage and processing
functionalities as well as a GUI for the end user. In this section,
completely new workflow, covering the whole analysis pro- we focus on the characteristics of modern vibration sensor and
cess, from data collection, pre-processing, analysis to result the our own experience with the sensor networks deployed
visualization. Our core analytical technique is an uncontrolled for industrial level vibration data collection. These features
approach. Different from existing methods which investigates inspire the design and implementation of our vibration analysis
the lifetime and patterns over a number of target devices since system, with details described by the following section.
its birth to its death, our new approach only passively monitors
the running status of existing devices in operation, therefore Most of the vibration sensors are now based on MEMS.
saving huge amount of observation time and cost on hardware. In Table I, we list the specifications of two different models
To evaluate the effectiveness of our proposal, we deploy our of sensors, including piezoelectric accelerometer sensor1 and
analytical engine in a world leading semiconductor plant, on MEMS sensor2 , on a number of key features. The new gen-
maintenance service over the pumps used in the plant. The eration of vibration sensors, based on MEMS, greatly reduces
preliminary results show that our analysis enables the plant the price, power consumption and size. MEMS also enables a
to prolong the lifetime of the pumps by 1.2x and reduces the wide operation range, due to its high resonance frequency and
replacement cost by 20% correspondingly. larger detectable acceleration range. These enticing features
are driving the growing popularity of MEMS sensors used
The major contributions of the paper is summarized as in real applications. On the other hand, due to the weakened
follows: accuracy of MEMS-based sensors, the analysis of the error-
prone vibration data poses huge challenges on data processing
1) We discuss the general framework of vibration sensor data and analytical softwares, which are expected to correct the
processing and analysis, covering the complete workflow errors and generate reliable results for the decision makers.
from data collection, cleaning, analysis and visualization.
2) We propose a novel analytical algorithm to enable ac- Each sensor mote (or equivalently sensor node) runs in
curate and scalable prediction on the failure of devices either active or sleep state. We design the sensor mote to
under vibration sensor monitoring. support active sensor operation at lower energy consumption,
3) We empirically evaluate the usefulness of our system on and ultra-lower power during sleep period. This follows the
real manufacturing sites and report the performance on common design principle of existing low-power senor motes,
both prediction accuracy and cost saving. e.g., Telos[9]. The sleeping schedule is carefully controlled
by the central server, in order to meet the expected sensor
The rest of the paper is organized as follows. Sec. II lifetime based on its battery power capacity. By using MEMS
introduces the background of vibration sensor network. Sec. sensor, the system is capable of capturing reliable vibration
III provides descriptions of the data used in the analysis data approximately from 150 Hz to 22 kHz of sampling
and the mathematical formulation of the analysis objective. rate, which is dynamically set by the software. Hence, the
Sec. IV covers the details of our approaches behind our frequency bandwidth of vibration observable by the sensor
analytical engine. Sec. V reports our empirical evaluations is up to 11kHz, half of the maximum sampling rate at 22
of our proposal on real world manufacturing site. Sec. VI kHz by Nyquist sampling theorem. For each measurement for
reviews existing studies on relevant topics, and Sec. VII finally vibration, the sensor collects the vibration in three dimension
concludes this work and addresses future research directions. independently by measuring the acceleration rate. Therefore,
it is straightforward to see that higher sampling rate renders
II. V IBRATION S ENSOR N ETWORK lower frequency resolution but wider frequency range in obser-
vation. In order to get higher frequency resolution, the sensor
In Figure 1, we illustrate the end-to-end system archi-
tecture, covering the whole workflow from the generation 1 http://swann-associates.com/pdf/probitech/TM0784A.pdf
of physical vibration to the analytical result visualization. 2 http://www.analog.com/media/en/technical-documentation/data-sheets/
Generally speaking, the system architecture consists of three ADXL001.pdf

1272
1260
1274

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on August 29,2023 at 19:25:27 UTC from IEEE Xplore. Restrictions apply.
 35 5
4
30 3

 2
   

Report Period Lower Bound (Hour)

25 1
 
0 2
10 10 3 10 4 10 5
20 Target Year= 1 year
     Target Year= 2 year
Target Year= 3 year
Fig. 2: Non-Intrusive Vibration Monitoring on Pump’s Suction 15 Target Year= 4 year
Connector using Wireless Vibration Sensor
       10
 

 
   
  5
$ 
# 

0 2
    10 10 3 10 4 10 5
   
!  "#
Sampling Frequency (Hz)
    

 %)&)' Fig. 5: Trade-off :sampling frequency, report periods, and node
Reliable Data Collection ! " % & ' lifetime
&)(
$ 

#  &)(

& (

put it together back to 6 kbyte of vibration data by Flush

Fig. 3: Illustration for sensor operation with reliable data in the sensor management server in Fig. 1. Fig. 3 and Fig. 4
collection [8] shows the overall sensor operation with Flush protocol. During
initialization, each sensor mote sends its boot-up notification
mote operates at lower sampling rate with narrow frequency to sensor management server and receives configuration on
range. its wakeup slot for next round of measurement. Sensor mote
transits into active mode at the specified wakeup time slot.
Fig. 2 illustrates the use of vibration sensor for monitoring There are two periods in the wakeup time slot; a round period
vibration of pump in service. On-site diagnostics for pump’s and a heartbeat period. The sensor collects the samples and
health condition can be directly done by analyzing real-time transmits the samples to the data server in the round period,
vibration data from the rotating motor inside pump. During its and updates its liveliness with the sensor management server in
operation, it however is not feasible to dissemble the pump to the heartbeat period. If the central server does not receive the
install vibration sensor. Instead, our IoT driven approach is to heartbeat signal from a particular sensor, the base station marks
continually monitor vibrations propagated from the motor to that server as dead. Once the Flush transaction is successfully
the exterior suction connector in a non-intrusive manner. completed, the sensor turns into sleep mode until the next
wakeup slot.
For each single measurement, the sensor collects 1024
samples, each of which consists of 2 byte readings on three
dimension, with 6 Kbyte data in total. Due to the limitation of
maximum packet size of low-power radio per transmission, the The key to the sensor management is the scheduling on
6Kbyte data is partitioned and transmitted to the gateway in their wakeup time slots, as is shown in Fig.4. The management
120 data packets. It is crucial to the system to reliably receive system attempts to maximize the information collected, under
all packets, in order to recover all 1024 samples. To accomplish the constraints of expected sensor battery lifetime, and the
that, we employ Flush, the reliable bulk transport protocol [8], sampling frequency for vibration. This is a unique challenge to
in which a bulk of packet stream is guaranteed to be delivered the data analytical system as well. Fig. 5 presents the tradeoff
to a base-station using NACK protocol. Those packets are among target node lifetime, sampling frequency, and report
period. It can be easily seen that the minimum report period
for targeted node lifetime grows and the data collected by the
   
  sensor network shrinks as decreasing the sampling frequency.

   Let us assume that the sampling rate of sensors is set to 150Hz
  

in Fig. 5. Then one option of the system is to collect 2,576

 
  
   
  
vibration measurements in three years of target lifetime (=
 3 year x 365 days x 24 hours / 10.2hours ), while another
    
        option is to collect 3,650 measurements for 2 years of target
 lifetime (= 2 year x 365 days x 24 hours / 5.2hours ). In our
  

IoT-based data analytic settings, we must acknowledge that
   
data is expensive and valuable resource. This is a significant
difference to other data analytical systems, in which it is
Fig. 4: Sensor management with round and heartbeat period always cheap to collect more data.

1273
1261
1275

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on August 29,2023 at 19:25:27 UTC from IEEE Xplore. Restrictions apply.
 III. P ROBLEM D ESCRIPTION TABLE II: Table of notation descriptions.
In this section, we discuss the model of the data and for- Symbol Description
mulate the mathematical problem of vibration analysis based M number of vibration sensors
on the data representation. We also highlight the challenges N number of measurements
behind our data analysis problem, which motivate our analysis K number of samples in each measurement
methodologies introduced in subsequent sections in detail. m m-th vibration sensor
n n-th measurement
A. Problem Overview k k-th sample in a measurement
In Figure 6, we present an example analysis over a number amnk 3-dimensional vector in {x, y, z} space
of homogeneous equipments, say pumps in a semiconductor almn k-dimensional vector on readings over l ∈
factory, to illustrate the general workflow of vibration analysis. {x, y, z}
Generally speaking, given the periodic samples on the acceler- âlmn normalized vector of almn
ation reading data from the vibration sensors, the ultimate goal a L2-norm of the vector a
2
of our analysis is to estimate the Remaining Useful Lifetime rmn RMS feature of measurement n over equip-
(or RUL in short) of the equipments, based on the historical ment m
records in the database system. From operational perspective, smn feature vector of measurement n over equip-
it is more important to identify the equipments when they are ment m
close to end of their lifetime and about to fail in the near future. C set of all candidate labels
In the following, we list a number of basic assumptions on the Ci a particular label for every i ∈ {1, 2, 3, 4}
data and the equipments, in plain language, in order to justify (smn , qmn ) a training sample with label qmn ∈ C
our data and analysis model in the rest of the section. T s , Te starting and ending time of data used in
Based on the data collection mechanism introduced in analysis
Sec. II, we make a number of assumptions on the infor-
mation coming from the vibration sensors. Firstly, the data 1) Indirect Measurement: The acceleration rates in three
only provides indirect information over the vibrations, i.e., dimensional space is an indirect and approximate mea-
samples on the acceleration rate of the target equipment in surement over the vibration phenomenon of the target
three dimensions. Secondly, samples from different sensors equipment.
may cover various intervals on the time domain. In Fig. 4, 2) Noisy and Unaligned Observations: The readings of
for example, the samples from five equipments are collected acceleration data from different sensors cannot be well
at completely different time points. Therefore, it does not aligned and may contain huge amount of noise.
generate any meaningful results by directly comparing the 3) Variance on Initial Status: The initial status of the target
acceleration data from two vibration sensors. Thirdly, there equipment could be completely different from each other.
is always significant noise in the data. Although the technical 4) Diversity on Lifetime model: The usage and lifetime
advances on sensor technology has greatly reduced the error, model of the target equipment depends on a number of
the acceleration rate of the sensors may not accurately reflect unknown and external factors.
the vibration phenomenon of the physical equipment. The
analysis algorithm is supposed to be robust to the noise, to B. Data Representation
ensure the usefulness of the analysis outcomes.
In this part of the section, we discuss the data represen-
On the other hand, the unique characteristics of cyber- tation in the analysis system. There are three types of data
physical systems also bring up a few assumptions on the involved in the analysis process, including sensor data, human
equipments, which must be taken into consideration when de- label data and analysis meta data.
signing data and analysis model. Firstly, the equipments under
monitoring have different ages when the vibration sensors are Sensor Data We assume that there is one and only one
attached to them. Basically, it means that the initial status of vibration sensor attached to each equipment of interest. We
the equipment could be completely different. When plotting leave the extension from single sensor to multiple sensors to
the lifetime of the equipments in Fig. 6, for example, they our future research work. The notations used throughout the
are in very different positions in the time-feature visualization section are summarized in Tab. II. Following the conventions,
on the right part of the figure. Secondly, the lifetime of the we use bold font of lowercase to denote vectors and bold font
target equipment has huge variance, which depends on a of uppercase to denote matrices. Given a vector a, we use a
group of external factors. The expected lifetime of a pump, to denote the L2-norm of the vector a.
for example, depend on the manufacturing process where the
pump is installed during its operation. On the right part of Fig. Assume that we have M equipment in the cyber-physical
6, there are two general groups of equipments, which evolve on system, each of which is monitored by a vibration sensor.
two different directions in the feature space over usage time. We use m to indicate the id of the equipment as well as
The analysis model is supposed to distinguish between such the corresponding vibration sensor. Each vibration sensor has
equipment classes, in order to accomplish accurate predictions. taken N measurement, and each measurement contains K
samples of acceleration readings on three dimensions (x, y, z)
As a short summary, we design the analysis engine for the independently. Given the measurement id n and the sample id
vibration sensor network, by addressing the following major k, the readings from vibration sensor m is thus a 3-dimensional
challenges: vector amnk = (axmnk , aymnk , azmnk ).

1274
1262
1276

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on August 29,2023 at 19:25:27 UTC from IEEE Xplore. Restrictions apply.
        
  
 ! " #
 

$%& '  (

 
 
 

 
  
 


  

  $%& ' 
"  

 
$ #

  
 
 
 


 
  

    




   

Fig. 6: Problem Overview

Similarly, given l ∈ {x, y, z}, we use almn to denote the Human Label Data To accurately capture the running status
vector of length K, containing all measurement on direction l of the equipments, we also employ human experts to contribute
in lthe n-th measurement from vibration sensor m, i.e., almn = labels to the running status of the equipments. There are
amn1 , amn2 , . . . , almnK .
l
four options to the experts, when they are asked to label the
usability of the equipments. We use C to denote the complete
Due to the impact of gravity with earth, the original set of labels. In our problem setting, there are four labels
acceleration readings may contain bias irrelevant to the be- C = {C1 , . . . , C4 }, with details of the labels listed below:
havior of the target equipment. To eliminate the bias, we
apply normalization over the original readings, by subtracting • C1 : known as Zone A, Vibration of newly Commissioned
the average over all readings from one single measurement. machines.
Therefore, the normalized readings are calculated as âlmn = • C2 : known as Zone B, Acceptable for unrestricted long
K axmnk -term operation.
amn − 1 · k=1 K , in which 1 is a unit vector of length
l

K. • C3 : known as Zone C, Unsatisfactory for long term

continuous operation.
In our analysis system, we consider two types of features, • C4 : known as Zone D, Vibration of sufficient severity to
namely root mean square (or RMS in short) and power spectral cause damage to machine.
density (or PSD in short). RMS is commonly used to measure
the overall magnitude of vibrations. PSD allows us to better The experts label the data in two different ways. The
capture the subtle characteristics of the vibration behavior, by first way is data-driven, in which they read the acceleration
identifying detailed harmonic components behind RMS. These readings from vibration sensors and choose one out of the
features are mathematically formulated as follows. four labels to mark the status of the equipment, based on
their experience and intuition. The second way is physical-
checking, in which they physically check the equipment after
 1 it is replaced. Based on the labeling approaches above, each
2
rmn = (rmn
l
)2 , rmn
x
= √ axmn  equipment has at most one physical-checking label. Whether
l∈{x,y,z}
K the labels are collected in either of the approaches above, we
 1 have a record in the database, in form of a pair (smn , qmn ),
smn = slmn , sxmn = (ax WK )2 , in which smn is the feature vector for m-th measurement over
2K mn equipment n and qmn ∈ C is the class label evaluation from the
l∈{x,y,z}
domain expert over the equipment at the measurement time.
where WK is a K × K discrete cosine transform (DCT) Meta Data Given the observational data from sensor measure-
matrix which expresses the discrete time signal axmn by a ment database, we assume there are two timestamps, i.e. Ts to
weighted sum of cosine functions of different frequencies. Te , indicating the time interval covering all the observations
Note that rmsxmn is also simply a standard deviation
Kof vibra- (e.g. the beginning time could be Ts = 08/01/2016, and the
tion occurred in x- axis. Note that (rmsxmn )2 = k=1 sxmnk ending time could be Te =08/01/2017). This time interval is
holds by Parseval’s theorem, hence smn alone is sufficient to referred as analysis period . Note that the analysis period,
construct feature space. represented as a pair of (Ts , Te ), is automatically set by the

1275
1263
1277

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on August 29,2023 at 19:25:27 UTC from IEEE Xplore. Restrictions apply.
system to support periodical update, unless explicitly specified  
by the system administrator. For instance, the system’s analysis    
period could be set as long as an hour and refreshed when Te    !      #$  
is passed, i.e. Tsj = Tsj−1 and Tej + 1 hour, which forces   "  

the analytical engine to update the results in every hour. To   
simplify our notations, we assume that Ts and Te are pre-      
     
defined and unchanged, when the context is clear in the rest
of the paper. During the analysis period, the vibrations report   
their vibration measurements during their assigned time slot in         
     
every round as shown Fig.4. To simplify our problem without
loss generality, we assume that all sensors send that total N )  
     
number of measurements during the analysis period.        
 % &'(' 
C. Problem Formulation        
         
Given the abstract problem in III-A and the data model in
previous subsection, we provide a mathematical formulation *    +*& 
for the predictive maintenance problem.
) ,-   % 
We t = (t1 , t2 , t3 , t4 ) to denote the class label vector, in     / .
which tk (k = 1, 2, 3, 4) is a binary indicator to the class label .    
Ck . Since each equipment of interest4 is attached with exactly
one label at any time, we have k=1 tk = 1.
Let zmn denote an feature mapping of smn by a mapping Fig. 7: Analytical workflow overview.
function φ(s). Given the training data in the database, D =
{(smn , qmn )}, the goal of model training is to construct an where Θ = [θ1 , · · · , θL ]T , and evaluating (2) for Zpred =
evaluation of the probability over any class labels Ck ∈ C, XTpred ŴΘ̂ where Xpred = X + XRU L and XRU L is a
i.e., remaining useful lifetime matrix for m, n.
In the following section, we will discuss how we run data
P (qmn = Ck |zmn , D) (1) cleaning, feature extraction and model optimization, based on
the formulation above.
Given the probability evaluation function, our model pre-
dicts the class label for an unseen sample zmn by finding the IV. METHODOLOGY
estimated class label q̂ with the maximal likelihood, as
In this section, we describe our overall methodology and
algorithms developed for our predictive analytic. In Fig. 7,
q̂ = arg max P (qmn = Ck |zmn , D) (2) we present a layered architecture of the vibration analysis
Ck ∈C
engine. The data retrieval layer provides a common restful-type
In order to find the optimal function φ∗ for the class API for data transformation layer to retrieve data from factory
label prediction task, we aim to optimize with the following database and sensor database given a analysis period. The data
objective function: transformation layer performs various conversion functions to
 transform unitless raw data into various measurement data
φ∗ = arg max P (qmn |zmn , D) (3) such as acceleration in g(= m/s2 ), power spectral density
φ
(zmn ,qmn )∈D (g 2 /Hz), and etc. The data preprocessing layer performs a
number of tasks, including 1) an outlier detection to remove
Let xmn denote a service time of equipment m at nth invalid measurements and a moving average with user-defined
measurement since its installation. Then let X denote a M N × time window (1 day by default) to reduce noises in sensor
1 matrix of xmn sorted in an increasing order, and Z denote measurement, and 2) matrix construction for various variables
a M N × 1 matrix of zmn corresponding to X. in Table II while eliminating invalid measurements to prevent
unwanted computations followed in next layers. The feature
Let L denote the number of underlying equipment lifetime matrix extraction layer extracts a feature matrix from data
models. Then we assume that a linear model z = θi x holds matrix and stores them in data analytic server. Finally, the
for each lifetime model i = 1, · · · , L. Let w = (w1 , · · · , wL ) RUL model layer gathers aggregates pumps’ feature matrix
denote the label vector ofa binary indicator for lifetime model and build RUL model. More details on these processing steps
L
where wl ∈ {0, 1} and 1=1 wl = 1. Then let W denote a are elaborated in the following section.
M N × L matrix of w that assigns zmn to one of the lifetime
model l = 1, · · · , L. A. Data Preprocessing
Finally, our RUL estimation is done by first solving the The low-cost MEMS based sensor often suffers from a
following objective function: long-term offset drifts, by which a zero-offset of acceleration
measurement gradually increases or decreases over time. In our
Ŵ, Θ̂, L̂ = arg min XT WΘ − Z (4)
W,Θ,L problem setting, it does affect our analysis results, because

1276
1264
1278

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on August 29,2023 at 19:25:27 UTC from IEEE Xplore. Restrictions apply.
 1.0 Cluster Label a harmonic peak feature by defining a feature space with a
x-axis acceleration group of pairs of significant peaks’ value and frequency in
0.5 y-axis acceleration
z-axis acceleration PSD. The harmonic peak feature of sn is formally defined by
0.0 pn = {(fnk , pnk )}k=1,··· ,np where pnk ,fnk , and np are peak
0.5 value, peak frequency, and the maximum number of peaks to
be searched, respectively.
1.0

1.5 Our algorithm searches the harmonic peaks following the

2.0 procedure below:
2.5
Jul 21 2016 Aug 04 2016 Aug 18 2016 Sep 01 2016 Sep 15 2016 Sep 29 2016 1) Smooth PSD over adjacent frequencies by convolutions
(a) Stable sensor measurements using Hann window wh (n) = 0.5(1 − cos( n2πn h −1
)) where
3
Cluster Label nh (=24 by default) is the window size.
2 x-axis acceleration 2) Find points where its first order differential changes from
y-axis acceleration
1 z-axis acceleration positive to negative
0
Together with np the Hann window size nh is an important
1
control parameter in the algorithm deciding the sensitivity
2 of the peaks. We obtain stable harmonic features by finding
3 the optimal values over (np , nh ), which essentially performs
4
a joint smoothing over time and frequency. Furthermore,we
define harmonic peak distance Dij to quantify a dissimilarity
5
Jul 25 2016 Aug 08 2016 Aug 22 2016 Sep 05 2016 Sep 19 2016 Oct 03 2016 between two harmonic peak features, pi and pj , as described
(b)Unstable sensor measurements in Algorithm1. The basic idea is to approximate their Euclidian
distance pi − pj .
Fig. 8: The average x,y,z accelerations of installed vibration sensors; Invalid
measurements marked with white rectangular boxes using outlier detection
algorithm in (b) Algorithm 1 Compute Peak Harmonic Feature Distance
Require: pmax ← max(pij ), fmax ← max(fij )∀i, j
we normalize any offset in every acceleration measurement Ensure: nh = 24 , np = 20
by subtracting its average. Nonetheless, the stability of zero- pij ← pij /pmax , fij ← fij /fmax ∀i, j
offset (i.e., the average of acceleration measurements) is an sumcur ← 0, cntcur ← 0
important indicator of integrity of acceleration measurements Queuei ← [(fi1 , pi1 ) · · · (fi20 , pi20 )]
itself. It is important to make sure the averages of acceleration Queuej ← [(fj1 , pj1 ) · · · (fj20 , pj20 )]
measurements are stable over time, used as a check on the while Queuei not empty do
correctness of acceleration measurement. Fig. 8 shows a trace (fi , pi ) ← Queuei .pop()
of the average of acceleration for two vibration sensors in- j∗ ← do binray search for fi in [fj1 · · · fj20 ]
stalled in pumps. Note that the acceleration measurements are if |fi − fj∗ | ∗ fmax < nh then
supposed to be a constant through installation, since vibration (fj∗ , pj∗ ) ← Queuej .pop(fj∗ )
sensors do not move in the operation after its installation. distcur ← distcur + (fi , pi ) − (fj∗ , pj∗ )
Compared to Fig.8(a), we can clearly observe considerable else
long-term drift on the acceleration measurements, as well as distcur ← (fi , pi )
abrupt changes in the average at the middle of plot. To tackle end if
this problem, we apply mean shift clustering algorithm [5] sumcur ← sumcur + distcur
to detect outlier measurements directly exclude these data cntcur ← cntcur + 1
from our analysis. In Fig. 8(b), we show how our outlier end while 
detection algorithm effectively detects these data, by running Dij ← (sumcur + pjk )/(cntcur + len(Queuej ))
a 3-dimensional clustering over the acceleration averages.

In Fig.9 we show an example to illustrate harmonic peak

B. Feature Extraction features extracted from PSD measurement and corresponding
In previous section, we discuss how to extract PSD features harmonic peak distance over different PSD samples. On top of
from acceleration readings from the sensors. However, it is the figure, there are 20 major peaks extracted by Algorithm1,
worthwhile to emphasize that there are two major drawbacks which are generally consistent with their actual peaks with
by using PSD directly for analysis. First, PSD is a high- high significance. Using the harmonic peak feature as a base-
dimensional feature (i.e., 1024 dimensions in our case) that line, we compute harmonic peak distance for three different
often generates singular matrix when calculating (sT s), thus PSD samples. We hereby emphasize that our peak harmonic
leading to difficulties for regression algorithms. Second, PSD distance measures give more distance penalty for disagreement
feature is unreliable due to a large random fluctuation in their in peaks at higher frequency regions. It is actually a desirable
amplitudes over frequency due to measurement noise inherent feature, as equipment in abnormal condition tends to give off
in MEMS sensor. To solve this potential problem, we develop high-frequency noises.

1277
1265
1279

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on August 29,2023 at 19:25:27 UTC from IEEE Xplore. Restrictions apply.
 6
Power Spectral Density
Hz )

5
4 Detected Harmonic Peaks [Baseline]
Amplitde (g/

3
2
1
00 500 1000 1500 2000 2500
9
8 Power Spectral Density
Hz )

Peak Harmonic Distance =0.116

7 Baseline Harmonic Peaks
6
Detected Harmonic Peaks
Amplitde (g/

5
4
3
2
1
00 500 1000 1500 2000 2500
20
Power Spectral Density
Hz )

Peak Harmonic Distance =0.097

15 Baseline Harmonic Peaks
Detected Harmonic Peaks
Amplitde (g/

10
5
00 500 1000 1500 2000 2500
35
Power Spectral Density
Hz )

30 Peak Harmonic Distance =0.232

25 Baseline Harmonic Peaks
20 Detected Harmonic Peaks
Amplitde (g/

15
10
5
00 500 1000 1500 2000 2500
Frequency, (Hz)
Fig. 9: Peak harmonic feature distance comparison: (Top) a PSD sample and its peak harmonic feature taken from a heathy condition, Zone A [baseline peak
harmonic feature] (from the Second to the Bottom) Comparison of peak harmonic distances from the baseline feature (on Top) for PSD samples randomly
chosen taken other Zones.

C. Learning Remaining-Useful-Time (RUL) Model A. Experiment Setup

To learn RUL model, our algorithm first classifies mea- We collected vibration data over several months from
surements from individual equipments into their corresponding dozens of vacuum pumps of Vacuum and Abatement System
heath condition category in C given their harmonic distance in a Fab of a major semiconductor manufacturer. At a fixed
from an exemplary measurement sample from healthy con- period of 10 minutes, our vibration sensor network system
dition, Zone A. Let Da denote the Peak Harmonic Distance collects 1024 samples of vibration measurement at 4 kHz
from Zone A. Da is a measure to quantify how dissimilar the sampling rate on three orthogonal directions. A Fab manager
harmonic peaks from equipment from normal harmonic peaks classifies equipments for 4 category zones from Zone A
trained for zone A. The Peak Harmonic Distance Classification to Zone D according to equipments’ maintenance needs as
algorithm constructs a classifier using training data to deter- well as their health condition. Zone A corresponds to the
mine equipment’s health condition category given the observed normal operating condition where no maintenance is required
measurement’s Da . for a long-term operation where only yearly inspection or
Da is expected to monotonically increase over equip- maintenance required. Zone B and Zone C denote health
ment’s service time unless maintenance actions are taken. conditions that require additional caution for monthly and
Note that in practice many maintenance actions and other weekly inspection, respectively. Zone D is a nearly hazard
operational events add outliers into the monotonic increasing condition, such that immediate equipment maintenance or
model of Da over service time. To overcome the problem, replacement must be taken. A Fab manager mainly performs
we use the well known random sample consensus (RANSAC) two types of maintenance, BM (Breakdown Maintenance) and
approach[6]. The Recursive RANSAC regression algorithm PM (Planned Maintenance). BM is performed when major
iteratively runs the RANSAC algorithm on outliers until no malfunctions or unstable operations are identified by either
more monotonically increasing linear model could be found visual inspection or alarms from factory information and con-
any more. Regarding RUL prediction, our algorithm first learns trol system (FICS). Meanwhile, PM is periodically performed
the threshold Da between zone C and zone D. This threshold based on maintenance and inspection schedule set by Fab
is chosen to minimize the error of wrongly classifying records manager. We obtain detailed information on BM and PM, used
in zone C and zone D. The algorithm then determines the in our analysis framework. Although FICS generates huge
equipment’s RUL by projecting the equipment’s features over amount of data in real-time, covering over 100 attributes of
the space and checking if the projection exceeds the threshold. main equipment systems, in which only a small fraction of the
attributes are related to peripheral equipment including Pumps.
V. E MPIRICAL E VALUATION In our empirical evaluations, we only use the temperature
information from FICS in our analysis.
In this section, we report our empirical evaluation over real
world manufacturing domain, with introduction to experimen- We select 12 vacuum pumps and corresponding 3 months
tal setup in Sec. V-A and results presented in Sec. V-B data for our experiment and performance evaluation by a

1278
1266
1280

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on August 29,2023 at 19:25:27 UTC from IEEE Xplore. Restrictions apply.
domain expert for Fab equipments and a Fab manager. To 100
Prob(Da ZoneA)
|
simplify the annotation process, we do not distinguish between Prob(Da ZoneBC)
|
Zone B and C. Therefore, we use Zone BC to indicate the 80 Prob(Da ZoneD)
|

combined label for both Zone B and Zone C. Note that

Number of samples
the human generated labels include physical inspection on
60
audial (almost noise only) and visual inspection for pumps
and event logs from FICS as well. We collect labels for 2800
measurements containing 700 labels on Zone A, 1400 labels 40
on Zone BC, and 700 labels on Zone D, respectively. In some
rare cases, the label could be invalid due to human mistakes. 20
We simply discard the label and corresponding measurement
from our data for model building and evaluation.
0
0.00 0.05 0.10 0.15 0.20 0.25 0.30
Note that all those selected pumps are an identical model Peak Harmonic Distance from Zone A, Da
and products from the same pump manufacturer. Their ex- Fig. 11: Probability density function estimates of Da given
pected lifetime follows the same distribution, before their Zone x∈ {A, BC, D}, a decision boundary of Da for Zone D
operation over the target equipments. However, their remaining is 0.21
useful lifetime (RUL) vary dramatically during their operation,
due to different environment and behavior of the target equip-
functions for each histogram are estimated using gaussian
ments. After 3 months of experiment, a domain expert and
kernel density as well. We can easily find that the posterior
Fab manager estimates RUL of these 12 pumps by extensive
probability P rob(Da |Zone x) is clearly separated for all these
diagnostics, and compares the diagnostic results against our
three types of labels. After simple calculation, we identify the
analysis outputs.
optimal boundary between zone D and other two zones at 0.21
in peak harmonic distance domain.
B. Performance
Zone A Zone BC
In this part of the section, we reports the results on 1.0 1.0
classification and prediction tasks in our analysis system. For
0.8 0.8
classification, we use only 50 labels for training and all other
labels for testing. For prediction, we learn RUL model on entire 0.6 0.6
precision

measurements without labels, and use classification model

0.4 0.4
learnt training data.
0.2 0.2
60 60 60
Zone A Zone BC Zone D 0.0 0.0
5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50
50 50 50
Zone D Average
1.0 1.0
Hz )

40 40 40
Amplitde (g/

30 30 30 0.8 0.8
20 20 20 0.6 0.6
10 10 10
0.4 0.4
Peak harmonic dist.
00 00 00
500 1000 1500 2000 2500 500 1000 1500 2000 2500 500 1000 1500 2000 2500 0.2 0.2 Euclidian dist.
Frequency, (Hz) Frequency, (Hz) Frequency, (Hz) Mahal dist.
0.0 0.0 Temp.
Fig. 10: 100 sample traces for Zone A, Zone BC, and Zone D 5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50
Number of training samples
1) Classification: In Fig. 10 we show PSD of 100 sample
measurements in frequency domain for Zone A, Zone BC, and Fig. 12: Precision
Zone D, respectively. From the results, it is clear that overall
amplitude, shape and peak location in frequency domain are Fig. 12, Fig. 13 and Fig. 14 present the precision, recall and
all different from zone to zone. Note that zone BC and zone accuracy of our classification algorithm using peak harmonic
D become nearly indistinguishable when the random noise feature over different number of training samples. We com-
grows to cover each frequency area in PSD measurement pare the performance against other feature metric, including
domain. Furthermore, from zone BC to zone D, the variance of Euclidian/Mahalanobis distance of vibration measurement and
PSD at each frequency area increases proportionally, and the temperature measurement from FICS. These results imply that
fluctuation of PSD grows accordingly in zone D. Following our metric is much more stable, leading to better classification
results show that harmonic peak distance allows us to capture performance than other metrics. In particular, it shows that
most distinguishing features under a presence of large signal temperature data does not work for classification at all. It
fluctuation. is because equipments’ temperature is greatly affected by
the factory control system rather than equipments’ inherent
In Fig. 11, we show histograms of Da , the peak harmonic condition. Tab. IV shows the confusion table given 15 training
distance from training samples assigned to zone A, for all 2800 samples. In the table, Euclidian and Mahalanobis distance
labelled measurements given pump’s condition in zone A , feature often misclassify measurements from zone D as Zone
zone BC and Zone D. The corresponding probability density BC. Such errors are mostly fatal to the Fab factory, and thus

1279
1267
1281

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on August 29,2023 at 19:25:27 UTC from IEEE Xplore. Restrictions apply.
 TABLE III: Confusion Table at number of training samples = 15

Peak Harmonic Euclidean Mahalanobis Temperature

Zone A Zone BC Zone D Zone A Zone BC Zone D Zone A Zone BC Zone D Zone A Zone BC Zone D
Zone A 690 0 0 690 0 0 690 0 0 486 62 151
Zone BC 18 1352 5 330 1045 0 219 1156 0 12 984 352
Zone D 0 0 689 0 522 167 0 224 465 548 151 0

Zone A Zone BC time. The threshold value of Da between the zone D and zone
1.0 1.0
BC are plotted at 0.21 over service times as well. The pump
0.8 0.8 with Da greater than 0.21 is very likely to operate under health
0.6 0.6
condition of zone D, hence whose RUL becomes 0. The figure
recall

also shows a significant fraction of outliers deviating from

0.4 0.4 both of the models. The RANSAC algorithm first discovers the
0.2 0.2 most significant linear model and effectively classifies outliers.
Recursive RANSAC then follows, until no linear model found
0.0 0.0 with the predefined positive slope threshold for these outlier
5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50
Zone D Average
points. These results also prove our observation on the waste
1.0 1.0 of resource, highlighted in the introduction section.
0.8 0.8 Fig. 16 shows the same vibration feature versus service
operation time for individual pumps. Compared to Fig. 16,
0.6 0.6
monitoring pump’s vibration independently would require a
0.4 0.4 significant amount of time until collecting enough data to
Peak harmonic dist.
0.2 0.2 Euclidian dist. build RUL model. In the figure, asterisk symbol represents
Mahal dist. the last day of experiment where RUL are predicted using
0.0 0.0 Temp. the most suitable RUL model selected for each pump. The
5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50
Number of training samples results are shown in Tab. IV and compared with diagnosed
RUL intensively evaluated by domain experts. Note pump 4, 5,
7, 8 are replaced during the experiment by maintenance events
Fig. 13: Recall
of BM and PM which also can be seen in the figure as their
trend in feature space also abruptly changes accordingly. The
1.0 Pump 4, 5 and 8 are replaced according to replacement plan
which leaves 390, 310, and 280 days of wasted RUL, which
are translated into the total cost in US$ 98,000 (US$ 39,000,
0.8 US$ 31,000 and US$ 28,000 on each pump, respectively) given
US$ 100 of daily value depreciation of pump by RUL model
0.6 I and the usual price of a pump at US$ 55,000. Note that
accuracy

Pump 7 has been misclassified for a long-term operation in

the unrecognised alarming health condition which eventually
0.4 lead causes breakdown of a pump. As a short summary, our
prediction technique saves 22% of the operation cost on pumps
0.2 Peak harmonic dist. in Model I (with average lifetime of 18 months), and 7.4% of
Euclidian dist. the operation cost on pumps in Model II (with average lifetime
Mahal dist. of 6 months). It is a significant improvement over conventional
0.0 Temp. replacement strategies.
5 10 15 20 25 30 35 40 45 50
Number of training samples VI. R ELATED W ORK
Fig. 14: Accuracy Predictive maintenance is a hot topic in mechanical engi-
neering community, which provides new opportunities to re-
unacceptable to most of the Fab managers. duce the operational cost on equipment replacements. In [18],
You et al. propose updated sequential predictive maintenance
2) Prediction: As discussed in previous sections, same (USPM) model, used optimize the scheduling on equipment
model of equipments may have different completely different servicing. In [14], Sun et al. design proportional covariate
ageing patterns when deployed in practice. In Fig. 15, we model (PCM) to predict the risk of hazards, based on extremely
presents the lifetime models (equivalently RUL model), which sparse data on historical records of hazards. In [15], Tan et
we call model I and model II in the rest of the section, found by al. target at the predictive maintenance problem over multi-
our analysis system. A pump’s maximum operation lifetimes state systems. In [16], Tinga et al. present a physical failure
are approximately 6 months for the RUL model I, and 1.5 model. In [12], Simeu et al. tackles the predictive maintenance
year for RUL model II. In the figure, we map 155520 vibration problem by estimating the fault latency. In [17], Wu et al.
measurements from 12 pumps into feature space of a harmonic propose the very first neural network approach for predictive
peak distance from zone A (or Da ) given their total service maintenance. In [7], Kaiser et al. also apply artificial neural

1280
1268
1282

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on August 29,2023 at 19:25:27 UTC from IEEE Xplore. Restrictions apply.
 Fig. 15: Two lifetime models (Da = bj1 Dx + bj0 for RUL Model I and Model II) found by our recursive RANSAC algorithm

Fig. 16: RUL model prediction on Pump 0 -11

TABLE IV: Savings of our RUL prediction over diagnostic-based approaches.eps

Pump ID 0 1 2 3 4 5 6 7 8 9 10 11
Estimated RUL Model 1 1 2 1 1 1 2 2 1 1 1 2
Replacement Event - - - - PM PM - BM PM - - -
Wasted RUL - - - - 390 310 - -80 280 - - -
Predicted RUL (days) 458 458 -87 384 530 574 51 118 532 401 495 -3
Diagnosed RUL > 1 yr. > 1 yr. < 1 wk. < 1yr. > 1 yr. > 1 yr. < 3 mth. < 6 mth. > 1 yr. < 1 yr. > 1yr. 1 < wk.
BM: Breakdown Maintenance, PM: Planned Maintenance, RUL Model I / Model II: lifetime model for long-term (1 >yr.) / short-term (< 6 mth.)operation.

1281
1269
1283

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on August 29,2023 at 19:25:27 UTC from IEEE Xplore. Restrictions apply.
network on the problem of degradation modeling. While all Technology (KIAT) grant funded by the Korea government
these approaches above work well in their problem settings, Ministry of Trade, Industry and Energy. (No. N050600080
they share one common drawback. They are all based on exper- ). Zhang and Winslett are supported by the research grant
iments over well-controlled equipment samples, and therefore for the HCCS Programme at the Advanced Digital Sciences
impractical when used in real world applications. Our proposal Center from Singapore’s Agency for Science, Technology and
tackles this problem with a purely data-driven solution, without Research (A*STAR).
any assumption on the equipments under monitoring.
R EFERENCES
In data mining community, predictive maintenance is re-
cently capturing more and more attentions and efforts. [13], [1] ABB discrete manufacturing operations management software.
Sipos et al. propose to analyse logs to understand the equip- http://new.abb.com/cpm/industry-software/industry-specific-solutions/
discrete-manufacturing. Accessed: 2017-01-20.
ment status. Basically, they model log history as a series of
[2] Discrete Manufacturing Software, Powered by the IIoT , GE digital.
discrete events. In vibration sensor, however, all raw data are http://www.ge.com/digital/products/discrete-manufacturing. Accessed:
acceleration readings, which cannot be easily translated into 2017-01-20.
discrete events. [3] SAP Manufacturing Execution System. http://www.sap.com/product/
enterprise-management/execution-mes.html. Accessed: 2017-01-20.
Remaining Usefulness Lifetime prediction is a well known
[4] Siemens SIMATIC IT for Discrete Manufacturing. http://w3.siemens.
problem in industrial engineering area. [10] reviews a wide com/mcms/mes/en/industry/discretemanufacturing/Pages/Default.aspx.
spectrum of approaches used to model RUL on different types Accessed: 2017-01-20.
of devices. However, all these approaches are not directly [5] D. Comaniciu and P. Meer. Mean shift: A robust approach toward
applicable to our problem setting, because the characteristics of feature space analysis. IEEE Trans. Pattern Anal. Mach. Intell., 24:603–
new generation vibration sensor are different from any existing 619, 2002.
sensor. In practice, most of the engineers simply follow simple [6] M. A. Fischler and R. C. Bolles. Random sample consensus: A
prognostic procedure to estimate the RUL of the equipment. paradigm for model fitting with applications to image analysis and
automated cartography. Commun. ACM, 24(6):381–395, June 1981.
In[11], Sikorska et al. present a number of such procedures.
[7] K. A. Kaiser and N. Z. Gebraeel. Predictive maintenance management
using sensor-based degradation models. IEEE Transactions on Systems,
VII. C ONCLUSION AND F UTURE W ORK Man, and Cybernetics-Part A: Systems and Humans, 39(4):840–849,
In this paper, we present a holistic analysis framework and 2009.
a new data analytic engine to support predictive maintenance [8] S. Kim, R. Fonseca, P. Dutta, A. Tavakoli, D. Culler, P. Levis,
S. Shenker, and I. Stoica. Flush: A reliable bulk transport protocol
over replaceable equipments in manufacturing management for multihop wireless networks. In SenSys, pages 351–365, 2007.
systems. It is designed based on the characteristics of new [9] J. Polastre, R. Szewczyk, and D. E. Culler. Telos: enabling ultra-low
generation vibration sensor, which is used to periodically power wireless research. In IPSN, pages 364–369, 2005.
sample the acceleration of the target equipments in different [10] X. Si, W. Wang, C. Hu, and D. Zhou. Remaining useful life estimation
directions. Our analysis algorithm exploits the new feature - A review on the statistical data driven approaches. European Journal
generation scheme to build reliable models over meaningful of Operational Research, 213(1):1–14, 2011.
feature domain. It can be used to identify different equip- [11] J. Sikorska, M. Hodkiewicz, and L. Ma. Prognostic modelling options
ment classes with completely different ageing model. Our for remaining useful life estimation by industry. Mechanical Systems
and Signal Processing, 25(5):1803–1836, 2011.
preliminary experimental evaluation shows accurate prediction
outcomes and high usefulness on saving unnecessary opera- [12] Z. Simeu-Abazi and Z. Bouredji. Monitoring and predictive mainte-
nance: Modeling and analyse of fault latency. Computers in industry,
tional cost in existing equipment servicing and replacement 57(6):504–515, 2006.
strategies. [13] R. Sipos, D. Fradkin, F. Mörchen, and Z. Wang. Log-based predictive
In the future, we will continue our exploration on the fol- maintenance. In KDD, pages 1867–1876, 2014.
lowing two directions. Firstly, we target to include sequential [14] Y. Sun, L. Ma, J. Mathew, W. Wang, and S. Zhang. Mechanical systems
hazard estimation using condition monitoring. Mechanical systems and
model into our analytical engine. A sequential model, such as signal processing, 20(5):1189–1201, 2006.
RNN and LSTM, could help identify more accurate model on [15] C. M. Tan and N. Raghavan. A framework to practical predictive
equipment ageing and status evolution. Secondly, we attempt maintenance modeling for multi-state systems. Reliability Engineering
to use more flexible sampling scheme over the vibration & System Safety, 93(8):1138–1150, 2008.
sensors. Currently, the sampling frequency is fixed when the [16] T. Tinga. Application of physical failure models to enable usage and
sensor is physically attached to the target equipment. We load based maintenance. Reliability Engineering & System Safety,
believe dynamic sampling with varying frequency could further 95(10):1061–1075, 2010.
save the energy consumption, when the system can capture [17] S.-j. Wu, N. Gebraeel, M. A. Lawley, and Y. Yih. A neural network
integrated decision support system for condition-based optimal predic-
sufficient information over an adaptive frequency, based on tive maintenance policy. IEEE Transactions on Systems, Man, and
the analytical results. Cybernetics-Part A: Systems and Humans, 37(2):226–236, 2007.
[18] M.-Y. You, L. Li, G. Meng, and J. Ni. Cost-effective updated sequential
ACKNOWLEDGMENTS predictive maintenance policy for continuously monitored degrading
systems. IEEE Transactions on Automation Science and Engineering,
The deployment of the experimental testbed of this work 7(2):257–265, 2010.
was supported by the Korea Institute for Advancement of

1282
1270
1284

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on August 29,2023 at 19:25:27 UTC from IEEE Xplore. Restrictions apply.