Non-Intrusive Efficiency Determination of

In-Service Induction Motors using Genetic Algorithm

and Air-Gap Torque Methods

Bin Lu Wenping Cao

Innovation Center School of Science and Technology
Eaton Corporation University of Teesside
Milwaukee, WI 53216, U.S.A. Tees Valley, TS1 3BA, England

Ian French Keith J. Bradley Thomas G. Habetler

School of Science and Technology School of Electrical and Electronic Eng. School of Electrical and Computer Eng.
University of Teesside University of Nottingham Georgia Institute of Technology
Tees Valley, TS1 3BA, England Nottingham, NG7 2RD, England Atlanta, GA 30332, U.S.A.

Abstract — In-service testing poses particular difficulties for the testing is, the more accurately the machine’s electrical
experimentally determining induction machine efficiency. This parameters can be measured. However, to reduce disruption to
paper focuses on non-intrusive methods for testing in-service normal machine operation to an acceptable level, several
machines and proposes a hybrid method based on the air-gap empirical factors have been introduced. For instance, stray-load
torque method and genetic algorithms. The proposed method loss (SSL) is predicted using a fixed ratio of the rated output
has been verified from the experimental results from three power depending on the machine’s ratings [2]. Also, the rotor
induction motors rated at 7.5 hp, 100 hp and 225 kW. The leakage reactance can be calculated by a fixed ratio of stator
overall efficiency estimation accuracy is approximately within 4- leakage reactance depending on the NEMA code and the type
5% errors. of rotor, according to the NEMA MG-1 and IEEE 112
standards [2, 5].
Keywords — efficiency estimation; equivalent circuit; induction
motors; in-service testing; genetic algorithms; air-gap torque This paper presents some preliminary findings in the
method; parameter estimation development of one such non-intrusive method for testing in-
service induction machines. The method is based on a non-
I. INTRODUCTION intrusive air-gap torque method and uses genetic algorithm to
estimate machine parameters. Experimental results are
The increase in energy awareness is accelerating the provided from three induction motors rated at 7.5 hp, 100 hp
requirement for improvements in electrical machine efficiency. and 225 kW, labeled as machines A, B and C, respectively.
Among the common machine types, induction motors are the The key nameplate information of these motors is listed in the
most important consumers of electrical energy. Thus, if Appendix.
accurate and non-intrusive methods of estimating the
efficiencies of these machines can be established, the door to a
relatively straightforward means of cost-effective energy
savings will be opened. II. EXISTING METHODS FOR IN-SERVICE INDUCTION
In the literature, numerous methods have been proposed to MOTORS
evaluate the efficiencies of induction motors [1]. However, the In [4], 30 different existing methods for determining motor
need to determine efficiency under normal operating conditions efficiency are compared. A more comprehensive and recent
poses particular difficulties when experimentally determining literature survey on motor efficiency estimation is presented in
efficiency by conventional methods. For example, in-service [1], with special focus on in-service induction motors,
machines cannot be readily decoupled to perform a no-load considering the recent advances in online rotor speed and stator
test, as is defined in the standards [2, 3]. As a result, the resistance estimations. Some existing methods can be
existing efficiency evaluation methods vary significantly in considered to be low intrusion methods. However, few are
complexity, intrusion level, and accuracy, when used in non-intrusive. Generally, these methods can be categorized
“normal” plant conditions [1, 4]. In general, the more intrusive into three groups.

0197-2618/07/$25.00 © 2007 IEEE 1186

A. Torque methods invert its binary code from 0 to 1, or vice versa. By
These methods directly measure the input line voltages and manipulating the degree of each process involving the three
currents, input power, and stator winding resistance. The shaft operators, some global optimal solutions from the population
torque is either directly measured using a torque transducer or can be achieved.
estimated from the average air-gap torque, which is calculated In this paper, the Matlab environment is used to implement
using the measured variables. Stray-load loss, and windage a bespoke GA that searches for the machine’s equivalent circuit
and friction losses (WFL) are subtracted from the air-gap parameters, based on measurements made via the machine
power to find the output power. The major drawback is that terminals. In the results presented in section IV, the GA
the speed and stator resistance have to be accurately measured predictions are made off-line in an attempt to reduce any errors
by using additional sensors and additional tests such as no-load that may arise from experimental measurements.
and un-power tests are required, which prevent its wide
application in in-service motors.

B. Current methods IV. THE PROPOSED METHOD AND TESTING PROCEDURES

These methods largely rely on the machine nameplate or Based on the GA as an optimized search tool, this paper
manufacturer’s data to obtain information on machine currents proposes a hybrid efficiency estimation method incorporating
and speed. They often assume that load current varies linearly features that can be found in the air-gap torque methods, loss
with percentage load [1, 4]. This, however, produces large segregation methods and equivalent circuit methods.
errors due to the actual nonlinear relationship, which excludes a
large range of load conditions. Air-gap torque methods have proven themselves to be
highly precise methods under field conditions with reported
accuracy of 0.5-1% errors [10, 11]. An improved non-intrusive
C. Empirical methods version of these methods has been recently developed in [12].
These methods are derived from testing a large number of The test results have shown that it is capable of estimating
machines. Typical examples are the Oak Ridge Motor efficiency at an accuracy of 2-3% errors. To obtain the
Efficiency and Load (ORMEL) 96 method [6], and the equivalent circuit parameters and in turn allow the prediction of
Stanford Empirical Method [1, 4]. These methods vary machine efficiency at any load condition, input voltages, input
significantly in determining machine parameters and losses. currents, input power, stator resistance, and rotor speed need to
Nevertheless, the testers’ experience can play a significant role be determined. Since non-intrusiveness is desired, the only
and the resultant errors can be as much as 10% or even more in information available is the motor terminal quantities.
efficiency estimation [7]. Measurements of input line voltages, line currents, and input
power can be easily obtained. However, stator resistance and
rotor speed are rather difficult, as the resistance is traditionally
measured after shutting down the machine and rotor speed is
III. GENETIC ALGORITHMS obtained by using additional sensors. Consequently, methods
The genetic algorithm (GA) is a stochastic global for the online measurements of stator resistance and rotor speed
optimization method that is inspired by the principles of are needed. This paper adopts the online methods
evolution found in nature [8]. It is gaining in popularity in recommended in [1]. The specific methods are briefly
optimizing system design, parameter identification, and in described in the following sections.
solving nonlinear equations. The GA works with a population
of individuals represented by bit strings and modifies the A. Online stator resistance measurement
population with random search and selection operations. The dc signal injection method proposed in [13] has been
Compared with other numeric methods such as Newton- employed in this study. As shown in Fig. 1, a simple
Raphson, the GA is capable of finding the global minimum, MOSFET-controlled circuit is connected in one of the three
and requires neither initial estimates nor the use of any phases between the power source and the machine under test.
derivative of a function [9]. A controllable dc bias is intermittently injected into the
The significance of this approach lies in number machine, causing dc components in the motor input voltages
manipulation and natural selection. The key operators in the and currents. Since the injected dc current signal can produce a
computation consist of reproduction, crossover, and mutation. small level of unbalances in the stator voltages and currents,
Generally, the GA is implemented by firstly assigning each leading to torque distortion and additional power dissipation,
unknown parameter a random binary code (string). A the dc bias is controlled to be as low as possible to minimize
population of strings can be generated through repeating this any adverse effects. Overall, this method is capable of
procedure. Reproduction involves a process that each string is estimating stator resistance to an accuracy of less than 0.5%
assessed against predetermined criteria of fitness (fitness errors, despite any load variation, as reported in [13].
function). The fitters are more likely to survive and to produce Moreover, for inverter-fed and softstarter-connected
offspring in the next generation. Crossover represents mating machines, this additional dc signal injection circuit is not even
and swapping of some information of the two strings that are required, since signal injection can be readily implemented by
also randomly selected from the survivors of the selection adjusting the existing switching strategy of the semiconductor
process. Mutation occurs to one randomly selected string to devices.

1187
B. Online rotor speed estimation Stray-load loss under rated load is assumed in accordance with
Commonly, direct measurement of the rotor speed requires the IEEE standard [2] (i.e., k = 1.8% of the rated output power
a shaft-mounted speed encoder or an optical tachometer to be for machines A and B, and 1.5% for machine C), depending on
installed. However, in many circumstances, installing their power ratings. For any loads other than the rated, SSL is
additional sensors is not possible under field conditions. In this calculated by
study, an online sensorless rotor speed estimation method,
developed in [14], is used for determining the rotor speed. This Prated − ssl
Pssl = 2
×I2. (3)
method is based on the motor magnetic saliency harmonics, I rated
which arise from rotor slotting and eccentricity. The major
procedure of this method is given in Fig. 2. As a result, the SSL resistance is given by

The frequencies of these rotor and dynamic eccentricity 1 − srated ,

harmonics are related to the rotor speed by Rssl = kRr (4)
srated
 1− s  where srated is the slip under rated load.
f seh = f1 (kR + nd ) + nw  (1)
 p/2  The stator current and input power can be calculated using
where fseh is the frequency of rotor slot harmonic components; the equations
f1 is the supply frequency; k = 0, 1, 2, …; R is the number of 1 1 1 1
rotor slots; p is the number of poles; s is the motor slip; nd = 0, ×( + + )
R + jX s Rr / s + Rssl + jX r Rm jX m ,
±1, ±2, …, is the order of rotor eccentricity; and nw = ±1, I s = Vs × s (5)
1 1 1 1
±3, …, is the air-gap MMF harmonics order. + + +
Rs + jX s Rr / s + Rssl + jX r Rm jX m
The speed estimation process consists of two algorithms: an and
initialization algorithm and an online speed detection algorithm.
The initialization algorithm employs the eccentricity harmonics
Rr (1 − S ) kRr (1 − Srated ) ,
to determine the value of R and an optimal set of numbers for k, Pin = 3I s2 Rs + 3I m2 Rm + 3I r2 [ Rr + + ] + k wf n (6)
nd, and nw , which are usually unknown quantities. After that, S S rated
the subsequent online speed detection algorithm estimates the
rotor speed by
where k wf n denotes the windage and friction losses.
f seh
p − nw
f1 The magnetizing and rotor currents are calculated as
s = 1− ⋅ (2)
2 kR + nd 1
The slip estimate is independent of motor parameters, only V R s + jX s ,
requiring the number of poles. The determination of slot Im = s × (7)
Rm 1 1 1 1
harmonic frequency requires the motor to be in steady state so + + +
that f1 and s are nearly constant; thus, this method is ideal for R s + jX s Rr / s + Rssl + jX r R m jX m
online efficiency estimation, where only steady states are
concerned. This method provides robust speed estimate down 1 1
to 1 Hz operation with a high accuracy of within 5 rpm (for a 4- ×
Rs + jX s Rr / s + Rssl + jX r .
pole machine) at high speeds and 0.005 p.u. slip at low speeds. I r = Vs × (8)
1 1 1 1
+ + +
C. Implementation of genetic algorithms Rs + jX s Rr / s + Rssl + jX r R m jX m
The method of determining machine efficiency or loss The GA used to solve the four unknown variables is
components for an induction motor relies largely on the implemented within the Matlab environment. Each unknown
machine model used in the analysis. Figure 3 shows a parameters is represented by a 25-bit binary string (with 15 bits
modified equivalent circuit employed in this study. Of the after the decimal point). Together, they are then combined to
parameters shown in the model, the stator resistance Rs can be form a 100-bit individual. The GA mechanisms are fairly
measured online with a satisfactory accuracy, and the rotor standard. The population size is selected to be 500 and a
leakage reactance Xr can be derived from stator leakage maximum of 10,000 generations are permitted. Uniform
reactance Xs . This empirical factor is given in NEMA MG1 selection with an elite count of 2 and stochastic multipoint
[5] and IEEE 112-F1 [2]. Namely, Xr = Xs / 0.67 for the NEMA crossover form the basis of the algorithm. Mutation is also
design B motors used in this case. As a consequence, there are uniform, with relatively high rate to keep algorithm “awake”.
only four parameters that remain unknown and need to be Fitness scaling is proportional. After the initial runs, parameter
estimated: (i) stator leakage reactance Xs, (ii) magnetizing normalization is applied and simple penalty functions are used
resistance Rm , (iii) magnetizing reactance Xm, and (iv) rotor to maintain sensible operation. Moreover, it is observed that,
resistance Rr . in addition to the four unknown parameters, the WFL
In Fig. 3, the output power is calculated by subtracting coefficient kwf could also be predicted as a by-product of the
windage and friction losses from the estimated output power. GA calculations if included in equation (6).

1188
 _V
sw,dc +
a' a

Ias,dc
Rsource Rs

n Rsource s Rs
Rsource Rs
c' b' b c

Figure 1. Online stator resistance measurement circuit

Figure 2. Online rotor speed detection algorithm

Rs Xs Xr Rr

Im
Vs (1 − s )Rr
s
Is Rm Xm Ir

RSSL

Figure 3. Modified equivalent circuit of induction motors

1189
 V. RESULTS AND DISCUSSIONS 140

130
A series of tests have been conducted to assess the
capabilities of the GA approach in determining the required 120

machine parameters. In the beginning of these tests, several 110

Real component
fitness functions are constructed based of the sum of the 100

squares (or in some cases sum of the absolute) differences 90

between the input power measurement and input power 80

calculated by equation (6).
70

The application of this fitness function to the test data for 60

machine B proves somewhat disappointing. In total, several 50

tens of runs are performed using slightly different variations of 50 60 70 80 90
% Load
100 110 120 130

the GA. Table I presents a typical series of the results.

Clearly, whilst the estimates for Rr and Xs are reasonably Figure 4. Real component of stator current for machine B
consistent; the estimates for Xm and, in particular, Rm are not.
What is more, a similar pattern emerges when the data from -20
machine A and C are used.
-30

The reason for this is that the fitness function, based on -40
equation (6), uses only magnitude information of Is, Im and Ir.

Imaginary component
-50
Consequently, for the magnetizing branch (where no apriori
information is available and where both parameters need to be -60

estimated), the solution is clearly not unique. -70

-80
To overcome this problem, a modified fitness function is
proposed, in which the measured line voltage and input power -90

are used to resolve the stator current into its real and imaginary -100

components. Whilst Fig. 4 and 5 show a comparison of the -110

50 60 70 80 90 100 110 120 130
measured and calculated resolved currents for machine B based % Load

on the estimates made in Run 3 (‘o’ measured; ‘-’ calculated).

Fitness is then calculated based on a weighted sum of the Figure 5. Imaginary component of stator current for machine B
normalized absolute differences between the resolved
components of the measured current and those calculated using
equation (5). Table II presents the results of a typical series of TABLE I. GA RESULTS FOR MACHINE B USING A FITNESS FUNCTION
BASED ON EQUATION (6)
results for machine B.
Rr Rm Xm Xs
Table II shows a considerable consistency in the estimates Run 1 0.0414 35.946 47.6 0.372
for Rr, Xs (with the values for Rr and Xs also agreeing strongly Run 2 0.041 330.737 26.175 0.37
with those shown in Table I) and Xm, with Rm once again Run 3 0.0415 107.146 56.789 0.373
proving poorly consistent. The problem seems to be associated Run 4 0.0408 188.831 23.922 0.369
with the proportion of parameters. If the relative values of Rr/s Run 5 0.0413 959.682 38.777 0.372
and Rm are in proportion to the currents flowing in the rotor and
iron loss branches then less than 0.5% of the current flows TABLE II. GA RESULTS FOR MACHINE B USING A FITNESS FUNCTION
BASED ON EQUATION (5)
through Rm .
Rr Rm Xm Xs
Therefore, to reasonably estimate Rm , far greater Run 1 0.0399 151.06 12.975 0.363
measurement accuracy is necessary. To overcome this Run 2 0.0399 503.85 13.075 0.363
problem, an alternative method based on Gauss-Seidel and Run 3 0.0399 410.49 12.986 0.363
empirical factors is employed. These results are given in Table Run 4 0.0399 162.5 12.941 0.361
III. Run 5 0.0399 542.42 12.936 0.361

Based on the above results, efficiency can be predicted over TABLE III. MEASURED AND PREDICTED ELECTRICAL PARAMETERS
a wide range of load conditions, this is shown graphically in FOR THE THREE MACHINES
Fig. 6 to 8, where the GA estimated efficiencies under various Parameter A B C Method used
load conditions are compared with those obtained by the air- Rr 0.167 0.039 0.431 GA
gap torque method. Xm 15.33 12.93 139.79 GA
Xs 0.79 0.36 4.39 GA
From these figures, it can be seen that the overall error in Rs 0.103 0.047 0.345 Measurement
efficiency determination is approximately 4-5%. This level of Rm 4841 440 5538 Gauss-Seidel
estimation may be effective when the actual machine efficiency Xr 1.18 0.54 6.55 Empirical
is not very high. However, for highly efficient machine, the
GA prediction is rather limited over a wide range of loads.

1190
 100

95

90

efficiency (%)
85

air-gap t orque
80 GA met hod

75

70
0 20 40 60 80 100 120
load (%)

Figure 6. Comparison of efficiency for machine A

100

98

96

94
efficiency (%)

92

90

88 air-gap t orque
GA met hod
86

84

82

80
0 20 40 60 80 100 120 140
load (%)

Figure 7. Comparison of efficiency for machine B

100

98

96

94
efficiency (%)

92

90

88 air-gap t orque
GA met hod
86

84

82

80
0 20 40 60 80 100 120 140
load (%)

Figure 8. Comparison of efficiency for machine C

1191
 VI. CONCLUSIONS AND FUTURE WORK REFERENCES
This paper has outlined a preliminary investigation on a [1] B. Lu, T. G. Habetler, and R. G. Harley, “A survey of
hybrid non-intrusive method for in-service induction motor efficiency-estimation methods of in-service induction motors,”
IEEE Transactions on Industry Applications, vol. 42, no. 4, pp.
efficiency estimation based on genetic algorithms. The non- 924-933, July/Aug. 2006.
intrusive characteristic of the proposed method requires that it [2] IEEE Standard Test Procedure for Polyphase Induction Motors
can only use information that can be obtained from the and Generators, IEEE Standard 112-2004, Nov. 2004.
machine terminals and the nameplate to calculate the [3] Method for determining losses and efficiency of three-phase
parameters of the equivalent circuit. The proposed method has cage induction motors, IEC Standard 61972 (2G/125/FDIS),
been verified by experimental results from three induction Sept. 2002.
motors with various physical configurations, rated at 7.5 hp, [4] J. S. Hsu, J. D. Kueck, M. Olszewski, D. A. Casada, P. J.
100 hp, and 225 kW. The overall efficiency estimation Otaduy, and L. M. Tolbert, “Comparison of induction motor
accuracy is approximately within 4-5% errors. field efficiency evaluation methods,” IEEE Transactions on
Industry Applications, vol 34, no. 1, pp. 117-125. Jan./Feb.
However, whilst results so far are promising, the estimation 1998.
of the magnetizing resistance, Rm, has proved to be a challenge. [5] NEMA – MG 1 Standard, 2003.
The difficulty lies in the relatively small impact of the [6] P. Otaduy, “ORMEL 96 (Oak Ridge Motor Efficiency and
magnetizing current (in the order of 0.5% in this case) on the Load), A computer program for in-service estimation of motor
efficiency and load with minimum intrusion: user’s guide,”
terminal measurements. Moreover, whilst it is possible that ORNL/MC-ORMEL1, March 15, 1996.
modifications to the relatively simple GA used in this study [7] P. Pillay, V. Levin, P. Otaduy, and J. Kueck, “In-situ induction
may overcome this problem, it is believed that problem lies in motor efficiency determination using the genetic algorithm,”
the measurements themselves. Therefore, an alternative IEEE Transactions on Energy Conversion, vol. 13, no. 4, pp.
method of estimating Rm is in need. One possibility is to take a 326-333, Dec. 1998.
lead from measurement systems theory and to use some form [8] N. Okaeme, P. Zanchetta, and M. Sumner, “Automated online
of differential measurement to identify such a small quantity design of robust speed digital controllers for variable speed
drives,” in Proc. 41st IEEE Industry Applications Society Annual
hidden by a large offset. Meeting, vol. 2, Oct. 2006, pp. 658-663.
[9] P. Pillay, R. Nolan, R, and T. Haque, “Application of genetic
algorithms to motor parameter determination for transient torque
calculations”, IEEE Transactions on Industry Applications, vol.
ACKNOWLEDGMENT 33, no. 5, pp. 1273-1282, Sept./Oct. 1997.
This research was supported, in part, by a U.S. Department [10] J. S. Hsu and P. L. Sorenson, “Field assessment of induction
motor efficiency through air-gap torque,” IEEE Transactions on
of Energy award (DE-FC36-04GO14000). The authors thank Energy Conversion, vol. 11, no. 3, pp. 489-494, Sept. 1996.
Mr. Pinjia Zhang for his assistance in experimental setup and [11] J. S. Hsu and B. P. Scoggins, “Field test of motor efficiency and
data collection. load changes through air-gap torque,” IEEE Transactions on
Energy Conversion, vol. 10, no. 3, pp. 477-483, Sept. 1995.
[12] B. Lu, T. G. Habetler, and R. G. Harley, “A nonintrusive and in-
service motor efficiency estimation method using air-gap torque
APPENDIX with considerations of condition monitoring,” in Proc. 41st IEEE
Industry Applications Society Annual Meeting, vol. 3, Oct.
2006, pp. 1533-1540.
THE LIST OF INDUCTION MOTORS UNDER TEST [13] S. B. Lee and T. G. Habetler, “An online stator winding
resistance estimation technique for temperature monitoring of
Machine Manufacturer Power Hz V A RPM line-connected induction machines,” IEEE Trans. on Industrial
Applications, vol.39, no. 3, pp. 685-694, May/June 2003.
A Emerson 7.5 hp 60 460 10 1765
[14] K. D. Hurst and T. G. Habetler, “Sensorless speed measurement
B US Electric 100 hp 60 460 114 1760 using current harmonic spectral estimation in induction machine
C GEC Alsthom 225 kW 50 3300 47 1483 drives,” IEEE Transactions on Power Electronics, vol. 11, no. 1,
pp. 66-73, Jan. 1996.

1192