A Neural Network based Approach for the Detection of Faults in the Brushless Excitation of a Synchronous Motor
Donald Gray, Member, IEEE, Ziang Zhang, Constantin Apostoaia, Member, IEEE, and Chang Xu
Department of Electrical and Computer Engineering, Purdue University Calumet, Hammond, IN, 46323, USA gray@calumet.purdue.edu, zjohn99@hotmail.com, apostoai@calumet.purdue.edu
, xcviva@gmail.com
Abstract

This paper presents an neural network based approach to identify in real time faulty components found on industrial brushless exciters. A brushless exciter or “rotating rectifier” is a key component of a synchronous motor. Improper operation of this component can prove costly for the motor’s owner. A method is based on Fourier analysis combined with the use of neural networks is presented to detect some common failures involving a three phase rotating rectifier. A laboratory setup was constructed to create fault condition data sets. These data sets were used to determine a preprocessing technique in conjunction with an appropriate neural net structure and training algorithm. Robustness of the system was tested using various levels of measurement noise to good result.
Index Terms
 brushless exciter, rotating rectifier, faulty diodes, Fourier analysis, harmonic spectrum, neural network, pattern recognition, pattern classification
I. I
NTRODUCTION
Large synchronous motors commonly use brushless exciters to supply the machine's field current primarily due to the fact that much less maintenance is required compared to other methods. This efficient system, on the other hand, unfortunately can create significant difficulties in the sensing and monitoring the components of the rotating parts. The circuit diagram of a brushless exciter is shown in Figure 1. This exciter system consists of a small ac generator mounted on the synchronous motor's rotor shaft, and a threephase rectifier which is also mounted on the same shaft. The three phase output voltage of the exciter can be converted to a dc voltage to supply the field circuit for synchronous operation. In a typical system, a Field Application Circuit (FAC) performs three control functions: (a) during starting of the synchronous motor, provides a discharge path for the current caused by the voltage induced in the field winding of the motor, through the field discharge resistor (FDR) using the silicon controlled rectifier SCR2; (b) when the motor reaches an adequate pull in speed, applies the field with dc excitation through SCR1 at the with such polarity that maximum torque will be obtained; (c) if the motor pullsout of step, the FAC removes the excitation and immediately reapplies the FDR. A rotating rectifier wheel is the main component located in the motor which carries the exciter electronics. Some existing industrial exciters have a system consisting of LEDs installed on the control unit. This provides a limited diagnosis of the rotating rectifier wheel. Transient faults occurring during motor start are almost impossible to detect with this system. At best in dark conditions with the access panels removed the lights can be observed while the motor is in operation. This paper examines a system ultimately meant to provide real time information about faults occurring in the rotating brushless exciter. While this monitoring system does not provide information regarding a specific component’s failure it does detect faults requiring shutdown and maintenance of the brushless exciter system.
Figure 1. A brushless exciter circuit diagram of the synchronous motor.
II. EXPERIMENTAL DATA COLLECTION The normal bridge operation with all intact diodes, and also two failing diode modes were simulated with one diode opencircuit and one diode shortcircuit. The data collection for the experimental threephase diode bridge operation was performed on the lab setup depicted in Figure 2. To facilitate of the exciter system as compared to the normal rotating diode assembly, a diode bridge was built external to the synchronous motor on the test bed, and connected between the threephase exciter armature and the main field of the synchronous motor. The alternator depicted on the right side of Figure 2 emulates the exciter of the real system. A wound rotor induction machine was chosen to serve as the synchronous motor which is shown on the left side of the figure. In order to
3phaselow powerinputExciterfield
Rf50%
If
Exciter armatureFIELD APPLICATION CIRCUITSCR1SCR2
FDR5%
MotorfieldMotor armatureEXCITER SIXPULSERECTIFIERSYNCHRONOUSMOTOR3phase ACmain powerinputROTATING COMPONENTSControlUnit
4239781424433551/09/$25.00©2009 IEEE
test the startup process of the synchronous motor as well as normal and abnormal rectifier operation, switches were added to the rectifier circuit as shown below.
Figure 2. Circuit diagram of the lab setup.
With the rotor switch K in the upper position (1), the wound rotor coils are shortcircuited and the machine starts as an induction motor and is allowed to accelerates up to near the rated synchronous speed. Then, switch K is thrown to the lower position (2) causing the rectified dc voltage of the diode bridge to be applied to the main field. The excitation field is now formed by the equivalent of one rotor winding in series with the other two parallel connected rotor windings. In this configuration the machine’s rotor is able to pullin and operates as a synchronous fashion. The characteristics of both machines that were used in this lab setup, as well as nominal measurements are given in the Appendix. The specific tests run included normal bridge operation and diode failures by opening as well as shorting. While several voltage and current waveforms were captured special attention was focused on the field voltage waveform in an attempt to keep the system simple. Various views of the data captured are shown in Figures.3, 4, and 5. In particular, Figure 3 depicts a normal bridge with all six diodes intact forming six pulses for each cycle of the exciter output voltage. The waveform of the rectifier's output dc voltage shown in Fig.4, was captured with one diode opencircuited and shows a significant shape modification compared to the normal operating case. Despite these qualitative waveform changes, the synchronous motor was still able to operate at the synchronous speed while being loaded with the rotor of the alternator. This scenario is one reason that in practice some component failure can go undetected for a period of time.
Figure 3. DC output voltage of the rectifier measured under normal diode bridge operation Figure 4. DC output voltage of the rectifier measured with one opencircuited diode
In Figure 5 the voltage waveform depicted indicates of one diode shortcircuited, and possesses a very different shape from those above. The large ac components that may be observed are induced by the torque pulsations which forced the synchronous motor at the limit to pullout of synchronism.
Figure 5. DC output voltage of the rectifier with one shortcircuited diode
The recorded data was exported to MATLAB to begin a more detailed analysis and characterization.
424
III. D
IODE
B
RIDGE
F
AULT
D
ETECTION
B
ASED
O
N
F
OURIER
A
NALYSIS
The output voltage amplitude spectrums of each of the various rectifier conditions were plotted as a percentage relative to the fundamental harmonic, and are shown in Figures 6 through 8.
Figure 6. Harmonic spectrum of the captured dc output voltage of the rectifier under normal diode bridge operation
In Figure 6, the relative harmonic content of the dc output voltage of the rectifier showed that the amplitude relative ratios for harmonics numbered 2 and 9 are always between 20% and 30%, when the diode bridge is operating under normal conditions. In Figure 7, for the case with one diode opencircuited, the harmonic Fourier analysis revealed a different pattern: the amplitude relative ratios for harmonics numbered 2 and 9 are always between 40% and 50%.
Figure 7. Harmonic spectrum of the captured dc output voltage of the rectifier with one opencircuit failing diode
In Fig.8, for the case with one diode shortcircuited, the harmonic content shows a new pattern: that the amplitude relative ratios for harmonics numbered 2 and 9 are always between 30% and 40%.
Figure 8. Harmonic spectrum of the captured dc output voltage of the rectifier with one shortcircuit failing diode
Thus a possible separation mechanism has been identified for the various failure modes using the spectrum of the voltage waveform. Next a pattern recognition routine is required to detect each specific failure condition. We next examine the application of neural networks to this problem. Once the data is appropriately preprocessed we need to survey a variety of neural network structures and complexities as well as corresponding training algorithms for their applicability to this problem. IV.
PATTERN
CLASSIFICATION
APPROACH
BASED
ON
NEURAL
NETWORK
S
A variety of fault detection systems have been developed, and several of them are being extensively used in industry. These techniques have achieved a certain degree of success, but are either cost inefficient, unreliable or too difficult to use. One of the major goals of fault detection is to develop the ability to collect relevant information and use the appropriate technology to effectively analyze the information to provide accurate and reliable results. There are two well known methods in the motor fault detection problem. One of them has always been called as Modelbased methods. In this particular method, an accurate mathematical model based on a set of system parameters is required. The disadvantage of the Modelbased method is obviously; not only the high requirement of elaborate understanding of the system dynamics, but also the different operating environment makes every model different. The other method is parameterbased method. An experienced engineer can diagnose a motor’s conditions based on its operation environments and measurements without knowing the exact mathematical model of the motor. But experienced engineers are expensive and difficult to train. With artificial neural networks, human expertise can be partially mimicked and automated, which is a major advantage of using neural networks in motor fault detection problem.
425
A.
Data Preprocessing
From pervious section we have narrowed down the manifestation of the fault to the voltage output of the rectifier. All data used herein was collected at a 500 millisecond period using a sample rate is 4 kHz producing 2000 data point snapshots. In order to avoid the weights of network having a large dynamic range; the data has been scaled to be in the interval to 0 to 1 by using equation (1):
(1)
Where X is srcinal data sequence, Y is normalized data. The srcinal data was is time domain, which needs to be transformed to the frequency domain. The Fast Fourier Transformation (FFT), was used to produce the spectrums for the three different conditions of interest. Sample spectrums are shown in Figure 9. The advantage of using spectrums as neural networks input data not only makes the signal easy to classify, but also could minimize the effect of white noise.
B.
Noise Generation
Since the robustness of the pattern classification system needs to be subsequently tested, a controllable noise level generation function becomes necessary. The noise generation function is defined by equation (2):
R
(2)
Where X is srcinal data matrix, E is noise level control parameter, R is random number form 0 to 1 and Y is the data matrix with noise. The above data combined with a 100% noise level (E=1.0) is shown in Figure 10.
C.
Training Network Design
The 2000 data points data sets were first transferred into a set of matrices. The first column of matrix contain 256 srcinal data points which from 1 to 256; in the second column we have data points from 2 to 257 and so on to create shifted windows of data. Totally we could have a 256 by 1000 matrix as neural network input for each data capture, where 256 is the number of inputs, 1000 is number of possible data sets. The basic structure of neural network is shown in Figure 11. Where W is the weights between neurons, b is bias point and N is the number of hidden layer neurons. Because we have three type of wave form need to be classifying, two neurons are set for the output layer. We set normal condition as (0, 0), open condition as (0, 1) and short condition as (1, 1).
1)
Training parameters need to be determined:
Accurately choosing these parameters could optimize the training reuslt
a)
Training algrothim:
Scaled conjugate gradient backpropagation and LevenbergMarquardt backpropagation has been tested in this paper.
b)
Two sets of transfer function
: Including hidden layer and output layer. Logsigmoid is chose for both layers.
c)
Performace funtion:
This also has been considered as training error. In this paper we tested Mean Square Error (MSE), and Mean Absolute Error (MAE)
d)
N: Number of neurons in hidden layer 2)
Network training a)
Training algrothim:
The problem we have is a function approximation problem. The three FFT spectrum could be considered as three target function, if the data we aquried from motor could fit one of the target function with an acceptable error, then the condition of the motor could be defined. Thus we narrowed down the training algrothims to the LevenbergMarquardt (LM) training algrothim and Scaled Conjugate Gradient (SCG) training algrothim. Both of the algorithms are well known for the function approximation
Figure 10. Comparison of original data and data with noise Figure 11. Neural Network Structure Figure 9. Voltage output after FFT
426
problem. Because of the usage of a second derivative to update the weights, the LM algorithm is able to obtain lower mean square errors than any of the other algorithms tested. But when the input data set is large, the calculation not only determinesan accurate result but becomes a time consuming process. The SCG algorithm is almost as fast as the LM algorithm on function approximation problems. The SCG algorithm takes less calculation than the LM method, but it can not reach the same precision level as LMwith the tradeoff that it can deal with large sets of data with only modest memory requirements. The data has been trained by both LM and SCG algorithm based on different number of hidden layer neurons. In this test the noise level has been set as 35%. For each condition we trained the network 50 times and taken the one which has the minimum error. The MAE error for each condition is shown in Figure12. As the plot shown above, all MAE error is less than
. And the MAE error from LM algorithm is even less. The last plot in Figure 12 is the single training execution time for each condition. Both of them are relatively close, which means the size of the data set in this problem is still acceptable for the LM algorithm. Thus, consider the error and execution time comprehensively we chose LevenbergMarquardt as the training algorithm.
b)
Performace function:
Based on the same data set, using LM algorithm to train the network with MAE and MSE, we can obtain the following results shown in Figure 13. As the plots shown, the MAE are always little bigger than the MSE in most case. Thus, in order to get a more accurate result, we chose MAE as our performance function.
c)
Number of neurons in hidden layer
The final step is to determine the number of neurons in hidden layer. There is no doubt that more neurons we have in the hidden layer more accurate result we could obtain. But we need find a balance between acceptable error and execution time. We train the network with those parameters has been decided from the previous section. The result is shown in Figure 14. From the output error and single training execution time we acquired, six hidden neurons have a relatively accurate result and don’t cost a lot of training time. V.
SIMULATION
RESULTS The final training process is run with the LevenbergMarquardt training algorithm, Mean Absolute Error as a performance function, the Logsigmoid as neurons’ transfer functions and there are six neurons in the hidden layer. The network was tested with 400 sets of data in each condition. Because there are three conditions, we have 1200
Figure 12. Error and execution time comparison of LM and SCG
Figure 13. Comparison of MAE and MSE
Figure 14. . Error and execution time comparison of hidden neurons
427