You are on page 1of 7

Monitoring and Diagnosis of External Faults in Three

Phase Induction Motors Using Artificial Neural


Network
ElSayed M. Tag Eldin, Hassan R. Emara, Essam M. Aboul-Zahab, Shady S. Refaat

Abstract—This paper addresses the possibility of integration of ann have numerous advantages over conventional diagnosis
an external motor faults (e.g., phase failure, unbalanced voltage, techniques. The main advantage of ANN method over the
locked rotor, undervoltage, overvoltage, phase sequence reversal conventional method is the non-algorithmic parallel
of supply voltage, mechanical overload) monitoring and distributed architecture for information processing. In general,
diagnostic technique into batch simulation with a digital they could improve the diagnosis performance. They are easy
protection set by using an artificial neural network (ANN) for a
three-phase induction motor. The proposed set-up has been
to extend and modify, and they could be easily adapted by the
simulated using”Matlab/Simulink” Software and tested for incorporation of new data as they became available.
external motor faults. The simulated results clearly show that Moreover, their design does not require a complete
well-trained neural networks can precisely of early fault mathematical model of the induction motor [7]–[12].
detection, diagnosis of external faults induction motor, also In the ANN-based systems, several quantities are utilized as
validating the proposed setup as a simple, reliable and effective input signals such as: stator currents and voltages, magnetic
protection for the three-phase induction motor fault fields and frame vibrations, etc. In general, stator currents and
identification scheme using an artificial neural network (ANN). voltages are preferred because they allow for the realization of
Index Terms— Artificial neural networks, Fault diagnostics, noninvasive diagnostic systems and the sensors required are
induction motor, rotor-cage monitoring and external motor fault. usually present in the drive considered.
This paper presents a further step toward the protection of
I. INTRODUCTION three phase induction motors with reliable system for

Large three phase induction motors are very commonly


employed in power system play a very important role in
detection and diagnosis of external motor faults. This
technique is utilizes results of the RMS measurement of stator
electricity production because of their good self-starting voltages, currents and motor speed to employ to train a neural
capability, offers users simple, rugged construction, easy network for monitor and diagnosis of external motor faults.
maintenance, low cost and reliability. These factors have led This methodology is applied to a three phase induction motor,
to use of induction motor in industry very extensively. As the using the well-know Matlab/Simulink software. In each
result, the motors can be exposed to different hostile simulation of the system, the induction motor faults are varied,
environments, misoperation, manufacturing defects, etc. including the fault types, fault positions.
Different internal motor faults (e.g., interturn short circuits, The simulated results presented clearly show that the proposed
ground faults, mechanical failure) along with external motor technique can accurately early detection, diagnosis of external
faults. These faults include: a) mechanical overloads, which faults of induction motor. The ability of the system to provide
define prolonged starting or locked rotor and stalling; b) a complete protection set for the three-phase induction motor
abnormal supply conditions such as loss of supply voltage, under different hazard conditions. Also external motor faults
unbalanced supply voltage, phase sequence reversal of supply diagnosis can be achieved more efficiently and at a reduced
voltage, over voltage, under voltage, and under frequency; c) cost.
faults in starting supply / Circuit, which include Interruptions
in phases or Blowing of fuse / single phasing, and short circuit II. FAULT DIAGNOSIS METHODOLOGY
in supply cable. These faults are expected to happen sooner or Since the analysis and design of induction motor is extremely
later [1]–[3]. critical in terms of the cost of both production and
Correct diagnosis and early detection of faults result in fast maintenance, it is not surprising that the fault diagnosis of
unscheduled maintenance and short down time for the induction motor is a crucial aspect of the subject, receiving
machine under consideration. It also avoids harms and helps ever more attention. Detection of faults addresses the binary
to reduce financial loss. Many of the motor fault detection and decision of whether or not a condition exists that is outside the
diagnosis schemes can be applied non-invasively on-line scope of normal system operation, elimination of possible
without the need of expensive monitoring equipment by using sources of false alarms. Fault diagnosis addresses the further
a microprocessor, microprocessor-based monitoring system is refinement of information regarding a condition that is known
of particular interest because it can be used for regular to be a fault, such as its nature and root cause [13]. In order to
analysis of machine variables and to predict possible fault use ANN for identifying induction motor fault and no fault
conditions[4]–[6]. conditions, it is necessary to select proper inputs and outputs
Recent developments in diagnosis systems have been led to of the network, structure of the network, and train it with
the consideration of radically different diagnosis strategies by appropriate data.
making extensive use of artificial neural networks (ANN’s),

1-4244-1298-6/07/$25.00 ©2007 IEEE.


in the hidden and output layers to adjust the weighted sum
input of each neuron. The output of the jth neuron is obtain as
⎡ ⎤
o j = f j (net j ) = f j ⎢ ∑ ( w ji xi ) + b j ⎥
⎣ i ⎦
Where x1 , x2 ,….. xi are the input signals, w j1 , w j 2 ,…
w ji are the synaptic weights of neuron i ; f j is the activation
function; and o j is the output signal of the neuron. Generally,
the sigmoid function is used as the activation function. The
output of each FFNN is connected to a simple perceptron
neural network having threshold effectively adds an offset to
Fig. 1. Illustrates the inputs and outputs of the ANN the weighted sum.
Condition monitoring techniques for the diagnosis of faults in ⎡ n ⎤
induction motor need to be improved in order to be able to yo = f h ⎢ ∑ wi o j − θ ⎥
identify, as quickly as possible, the different kinds of faults ⎣ i =1 ⎦
which may occur in the motor. The RMS value of the line Where yo is final output simple perceptron neural network, f h
stator voltages, currents and motor speed signals are used as
the input to neural network, hence the RMS value is the only is a heaviside function. The total input of neurons has to be
ac reading that does not depend on the shape of the signal. compared to a certain value in the neuron, the threshold value.
The RMS is often the most useful measurement for real-world This thresholding process is accomplished by comparison; if
waveforms and motor speed signal can logically be utilize to the sum is greater than the threshold value, it threshold at θ.
discriminate between motor starts and short circuits, and more The value of θ is therefore known as the neuron’s bias, the
accurate fault identification. Therefore, there are seven input resulting value compared to -0.2; if the result is positive, then
neurons. There are eight outputs corresponding to eight output a 1, else output a 0.
external motor faults mentioned before. Figure 1 illustrates the The perceptron neural network has a hard limit function a 0
inputs and outputs of the ANN. The output stage of the or a 1, the target output of each ANN was built in such way
classifier selects the maximum discriminant fault and responds that the value 1 represents an external motor fault (trip signal
with the number of the discriminant fault having the largest to Circuit Breaker); the value 0 represents a normal operation
value. condition for induction motor. The ANN structure for the
The output "Zero" will correspond to "No Fault" according to motor fault detection and diagnosis system is shown in Fig.3.
normal data achieved through monitored input measurements. Each ANN is trained with the data sets obtained from
The output “one” will correspond to condition "Faulty motor" simulated faults on induction motor.
if that particular condition in the monitored currents, voltages
and motor speed are exists. Consequently, that will be
detection and diagnosis. The diagnosis process (fault
identification) is illustrated by Fig. 2. In fact, this diagnosis
process has several advantages including accuracy, reliability,
modularity, and flexibility. This architecture is achieved
satisfactory distributed implementation of new types of faults
to the initial NN system [9], [12].

III. NEURAL NETWORK ARCHITECTURE AND TRAINING


The architecture of a neural network specifies the arrangement
of the neural connection as well as the type of units
characterized by an activation function. The processing
algorithm specifies how the neuron calculates the output
vector for any input vector and for a given set of weights. The
training algorithm specifies how the NN adapts its weights w
for all given input vectors, called training vectors. Thus, the
neural network can acquire knowledge through the training
algorithm and store the knowledge in synaptic weights.
A. Neural Network Structure Fig. 2. The general structure of a Diagnosis system of a three-phase induction
motor.
In this study, a multilayer feedforword neural network
(FFNN) that has two hidden layers is used. The bias unit,
whose activity level is fixed at one, is connected to all neurons
a three-phase squirrel-cage induction motor of a 2250 HP,
2400 V, 60 Hz, and 1725 rpm through a three-pole circuit
breaker. A RMS measurements device is connected in series
to an measure motor currents and connected in parallel to
measured voltages applied to motor, motor speed signal from
a speed sensor device is feed to an ANN module containing
all the necessary knowledge to perform the fault detection and
trip the three-phase switch in the main module according to
the fault conditions as shown in Figure 4. The protection
scheme is implemented through the ANN to provide a
complete protection set to the induction motor against any
external faults. It should be noted that for each protection
Fig.3. Basic structure of each ANN Structure. scheme, the time delay and setting values could be
individually adjusted according to the rating and characteristic
B. Neural Network Training of the protected motor, taking into account the coordination
The objective of training the network is to adjust the weights between different protection schemes
so that application of a set of inputs produces the desired set
of outputs. At the beginning of neural network training, the V. SIMULATION RESPONSE EVALUATION
neural network fault detector's diagnosis of the motor's In the design and study of the external motor faults diagnosis
conditions will not be accurate an error quantity is measured schemes, it is important to evaluate if the designed fault
and used to adjust the neural network's internal parameter in diagnosis algorithm is able to correctly classify different
order to produce a more accurate output decision. This process external faults conditions, so In order to confirm the effective
is repeated until a suitable error is achieved. Once the network performance of the proposed technique a high number of
is sufficiently trained and the network parameters have been computer simulations were performed using Matlab/Simulink
saved, the neural network contains all the necessary (considering different kinds of conditions).
knowledge to perform the fault detection. The performance
function for NN’s was mean square error (mse) between the A. Line to Ground Fault (ANSI code 50/51)
network outputs and the target outputs. If a short circuit occurs at the connection point between the
The data set for training neural networks have been feeder and motor (down stream from the monitoring device),
constructed by computer simulation of Matlab models the monitored current may be rise to between 10 and 15 rated
measurements. This data is first used to train a back- current and then drop to zero upon the tripping of the circuit
propagation feed-forward network using Matlab Neural protection.
Network Toolbox [14]. Levenberg-marquardt algorithm in The typical responses of the protection scheme for line to
MATLAB Toolbox was used for training because it is the ground fault for phase is 4.16msec as shown in Fig. 5. Clearly
fastest training algorithm for networks of moderate size and from Fig. 5, the neural network is detected the line to ground
has a memory reduction feature. The trained network is placed fault from the first fault sample input to the neural network.
in a Matlab model that monitors the voltages, currents and
motor speed and displays the fault conditions.
The step length of the moving data window is 1/4 cycle B. Phase Reverse Fault (ANSI code 47)
(4.16ms at 60 Hz frequency). When the connection of any two phases in a three phase
power supply is reversed, motor will run backward. This
condition is destructive to rotation sensitive load; a reversal in
the phase sequence of the supply voltage causes the motor
reverse direction of rotation. For most motors, this is not
damage to the motor, but it may be a cause for serious concern
for the driven load, and this may conceivably be dangerous to
nearby personnel.
The typical response of the protection scheme for phase
reverse fault of two phases (b and c) is 4.16msec as shown in
fig. 6.
C. Locked Rotor Fault (ANSI code 51LR)
Fig.4. Overall block diagram of external motor faults detection system
If for some mechanical reason, the motor rotor becomes
obstructed and prevented from rotation, this is a locked-rotor
IV. SIMULATION OF FAULT DETECTION SYSTEM condition. The motor current changes from its normal run
An extensive series of simulation studies have been carried value to the stall or Startup value, which will usually be from
out to obtain various motor faults for analysis. five to eight times greater than normal full load run current.
The simulation of external motor faults detection system The motor is not design to carry this much current
can be conceptually consist of a three-phase AC supply feeds
continuously and protection is required to ensure that the sustain this kind of overload for a period of time before
motor is disconnected before permanent damage occurs. tripping.
The response of the protection scheme for over load is
depicted in Fig. 9, and we can see that the response of relay is
41ms, which indicates that this protection scheme can protect
the motor insulation from excessive thermal stresses.
F. Over Voltage Fault (ANSI code 59)
When the supply voltage is more than 10% greater than the
motor rated voltage it often results in motor overheating [15]
from increased core losses associated with saturation.
For (10%) over voltage the motor will be overloaded by
approximately (10%). The core loss is (20 to 30%) greater
than normal and could cause the machine to overheat. The
response of the protection scheme for (12%) over voltage fault
is 8.3msec as shown by Fig. 10.

Fig.5. Responses of relay to line to ground fault

However, when Investigating on the 3-ph stator currents,


voltages and speed. It will be found that the stator current will
increase and the speed will rapidly decrease. The typical
responses of the protection scheme for Locked rotor fault is
8msec as illustrated by fig.7
D. Phase losses fault (ANSI code 46)
Single phasing or phase loss is the operation of a three
phase motor on only two phases due to the loss of voltage on
one phase. Phase loss is the maximum condition of voltage
Fig. 6. Responses of relay to phase reverse fault
unbalance, since the voltage of one of the three phases is not
just reduced, but becomes zero.
G. Unbalanced Voltage Fault (ANSI code 46)
The three-phase motor now operates on a two line supply.
This condition is phases reflected by an increase in current on When line voltages applied to the motor are not exactly the
the two healthy phases. Fig. 8 presents a three phase stator same, unbalanced currents will flow in the stator winding, the
currents, voltages waveform, speed of the motor and the magnitude depending upon the amount of unbalance .A small
typical response of the relay for phase C losses faults is amount of voltage unbalance may increase the current an
12msec excessive amount. The effect on the motor can be severe and
the motor may overheat to the point of burn out [16].
E. Over Load Fault (ANSI code 49) The response of the protection scheme for unbalanced
The motor is overloaded by up to 25% of its rated (full-load) voltage fault is 8.3msec. when unbalanced voltages exceed
current. For most motors, the design will allow the motor to
3% to prevent overheating of the windings as defined by Fig according to the importance of the fault and effect on the
11 motor, the protection technique develop correctly.

Fig.7. Responses of relay to locked rotor fault Fig.8. Responses of relay to phase losses fault

H. Under Voltage Fault (ANSI code 27) An extensive training of neural network for short circuit
and phase reversal, instantaneous tripping is utilized to
Under voltage is an average line-to-line voltage that is
provide the fastest protection scheme, while other protection
lower than the minimum acceptable operating voltage. The
schemes are utilized for better performance. It is apparent that
motor speed also increases proportionally to the square of the
an ANN with 7 neuron in the first hidden layer and 1 in the
voltage drop. As a result, the motor will be running slower
second, gives the best performance.
with a lower output, a large drop in voltage may cause the
In all cases, the ANN will be successfully detect the fault
motor to stall. This suggests that low voltage is serious
from the measured current, voltage and motor speed signals
enough to warrant protection in certain cases.
then tripe the circuit breaker, accurately identifying the faults
So, if the supply voltage reduces below 80% of rated
on a three-phase induction motor.
voltage, a drastic increase in the motor current will occur. The
effect of under voltage is graphically illustrated in shown Fig.
12, where the responses of the protection scheme for under VI. NEURAL NETWORK TESTING
voltage fault is 4.16msec. The complete system is successfully tested by creating
It should be noted that for each protection scheme, the different faults on the simulated motor, testing measurement
setting values could be individually adjusted according to the data sets to test the accuracy of the trained neural network to
rating and characteristic of the protected motor, taking into diagnose different external motor faults. The result
account the coordination between different protection demonstrates that with proper processing of the measured data
schemes. and proper training procedure, the neural network motor fault
The typical responses of the protection scheme for diagnosis schema can diagnose external motor faults with
randomly created external motor fault to evaluate the desired accuracy.
performance of the created neural network. From these results, Testing of the system involves measuring the RMS values
it can be seen that 4.16-41 ms after the occurrence of a fault of stator voltage, current and motor speed and passes the
values to the neural network. The motor is turned-off using
the circuit breaker, if a fault is encountered, that enable to test preventive maintenance diagnostics. The simulation results
the trained network show that the trained neural networks can early detection,
diagnosis of external motor faults of induction motor.

Fig.10. Responses of relay to Over Voltage Fault


Fig.9. Responses of relay to over load fault

VIII. REFRENCES
VII. CONCLUSION
[1] M. Y. Chow, R. N. Sharpe, and J. C. Hung, “On the application and
This paper has described and illustrated a methodology for design of artificial neural networks for motor fault detection—Part 1,”
monitoring and diagnosing of three-phase induction motors IEEE Trans. Ind. Electron., vol. 40, pp. 181–188, Apr. 1993.
[2] M. Y. Chow, Methodologies of Using Artificial Neural Network and
external faults. Motor fault detection and diagnosis involves
Fuzzy Logic Technologies for Motor Incipient Fault Detection.
processing a large amount of information of the motor system. Singapore: World Scientific, 1997.
The proposed methodology is based on artificial neural [3] K. F. Martin, “Review by discussion of condition monitoring and fault
network ANN. diagnosis in machine tools,” Int. J. Mach. Tools Manuf., vol. 34, no. 4,
The monitoring and operational diagnosis of a three-phase pp. 527–551, 1994.
[4] A. K. Sood, A. A. Fahs, and N. A. Henein, “Engine fault analysis, part I:
induction motor was achieved by using eight outputs, each for Statistical methods,” IEEE Trans. Ind. Electron., vol. 32, pp. 294–300,
one fault type. The inputs to the specified Artificial Neural Nov. 1985.
Networks were the RMS three phase currents, voltages and [5] “Engine fault analysis part II: Parameter estimation approach,” IEEE
motor speed. The outputs were whether a fault occurs or not. Trans. Ind. Electron., vol. 32, pp. 301–307, Nov. 1985.
[6] R. Isermann and B. Freyermuth, “Process fault diagnosis based on
The neural network model is training and its performance was process model knowledge—Part 1,” J. Dynam. Syst., Meas., Contr., vol.
tested using back-propagation feed-forward network function 113, pp. 620–626, 1991.
network. [7] F. Filippetti et al., “Current pattern analysis to detect induction machine
In this technique we use the ability of ANN in monitoring non rotational anomalies,” in Proc. 1998 Int. Conf. Electrical Machines,
vol. 1, Istanbul, Turkey, pp. 448–453.
and diagnosis different types of external motor faults that
cause unbalance in motor performance and failure to motor
parts. Extensive simulations were carried out using MATLAB
to full assess usefulness of the proposed technique for the
Fig.11 Responses of relay to unbalanced voltage fault Fig.12. Responses of relay to under voltage fault
[8] M. E. H. Benbouzid et al., “A decentralized neural network-based
architecture for induction motor faults detection,” in Proc. 1998 Int. [12] M. Y. Chow, “Motor fault detection and diagnosis,” IEEE Ind.
Conf. Electrical Machines, vol. 3, Istanbul, Turkey, pp. 1855–1860. Electron.Soc. Newslett., vol. 42, pp. 4–7, Dec. 1997.
[9] F. Filippetti et al., “AI techniques in induction machines diagnosis [13] R. J. Patton, P. M. Frank and R. N. Clark 2000 Issue of fault Diagnosis
including the speed ripple effect,” IEEE Trans. Ind. Applicat., vol. 34, for Dynamic Systems. London: Springer-Verlag.
pp.98–108, Jan./Feb. 1998. [14] MATLAB Version 6.5, Mathwork Corporation 2002.
[10] M. Y. Chow, “Intelligent motor fault detection,” in Intelligent [15] Martiny, W. J., R. M. MeCoy, and H. B. Margolis, "Thermal
Techniques in Industry, L. C. Jain, Ed. Boca Raton, FL: CRC Press, Relationships in an Induction Motor Under Normal and Abnormal
1998. Operation," AIEE Trans., 79, 1960.
[11] M. E. H. Benbouzid et al., “Induction motor interturn short-circuit [16] Gafford. B. N., W.C. Duesterhoeft, jr., and c. c. Mosher,III, "Heating of
andbearing wear detection using artificial neural networks,” Induction Motors on Unbalanced Voltages," AIEE Trans., 78, 1959,
Electromotion,vol. 5, no. 1, pp. 15–20, Jan./June 1998. pp.282-287.

You might also like