PWM Rectifier Fault Diagnosis Based on Multi-Feature Information Fusion

1. Introduction

Since the 21st century, new urban rail transit has developed rapidly in my country and has become an important symbol of my country's national economic development and people's living standards. It has a series of advantages such as low pollution, high efficiency and simple structure. PWM rectifier [1] is a key component of the new energy-fed traction power supply system. At present, domestic and foreign scholars have conducted little research on the fault diagnosis of new PWM rectifiers. Traditional fault diagnosis algorithms cannot accurately and quickly diagnose faults. Therefore, this paper proposes a fusion fault diagnosis method.

[2], which can quickly, accurately and in real time diagnose the fault of PWM rectifier switch tube online, thus facilitating fault-tolerant control and ensuring the smooth and safe operation of the train. The accurate extraction of fault characteristics is the key to the success of fault diagnosis. Since power electronic circuits are complex systems with multiple variables, nonlinearity and strong coupling, it is difficult to establish accurate and effective mathematical models. Traditional fault diagnosis methods cannot meet the requirements of current technical indicators at all, and the fault diagnosis of a single intelligent diagnosis method is not very effective. Therefore, based on theoretical analysis and MATLAB simulation, this paper proposes to use wavelet decomposition to extract wavelet energy spectrum as fault feature quantity, and input the normalized feature quantity into BP neural network to complete fault identification and diagnosis.

2.1 Extracting fault features using wavelet analysis

When a PWM rectifier switch tube fails, the current or voltage characteristic quantity changes suddenly, and the signal contains non-stationary time-varying information. The traditional Fourier transform can only perform local analysis on the frequency domain of the signal. It is an integral of the entire time domain and is suitable for analyzing steady-state signals, but not for non-steady-state signals. The wavelet transform has localization capabilities in both the time domain and the frequency domain. Its window size can be automatically adjusted according to the frequency of the signal, and it is a time-frequency analysis method based on "frequency bands", so it is very suitable for the analysis of transient or non-steady-state signals [3]. Binary wavelet transform is implemented through a multi-resolution analysis algorithm, which decomposes the signal ()ft into approximations and details at different scales, that is, the corresponding low-frequency and high-frequency parts. The decomposition formula [4] can be expressed as:

3 Experiments

3.1 MATLAB fault simulation and analysis The PWM rectifier circuit is selected as a diagnostic example. The schematic diagram is shown in Figure 3. MATLAB is used for modeling and simulation. The circuit parameters are set as follows: the input three-phase AC voltage is 380V, the operating frequency is 50Hz, the resistance is 0.1Ω, the inductance is 1mH, the carrier frequency is 10000Hz, and the modulation coefficient is 0.4. MATLAB is used to simulate the normal operation and switch tube failure of the PWM rectifier. The simulation time is set to 0.2s. At 0.1s, the switch tube open circuit fault occurs. Next, we extract the wavelet energy spectrum as a neural network training sample, and then extract the fault features again at 0.12s, 0.08s, etc., to test the neural network, thereby completing the diagnosis of the switch tube fault and the verification of the diagnostic algorithm. In order to simplify and explain the problem, we only take the single switch tube fault as an example, and other cases are similar. Basic working principle of PWM rectifier [1]:

3.2 Fault feature extraction

By comparing Figure 4 and Figure 5, it is not difficult to find that when the switch tube fails, the output current is greatly distorted. The output current voltage is decomposed into 5 layers by using the db3 wavelet to extract 1 low-frequency coefficient and 5 high-frequency coefficients. Then, the energy spectrum of each frequency band is obtained according to the wavelet decomposition coefficients, and a column vector is arranged in order. This vector is the feature vector corresponding to a certain fault. Next, the voltage signal is decomposed into 5 layers to obtain the wavelet coefficients of 6 frequency bands. The wavelet decomposition coefficients of each node are reconstructed, and the total signal can be expressed as [7]:

4. Conclusion

This paper decomposes the output current of the PWM rectifier with wavelet decomposition. Through comparative analysis, it is found that the wavelet decomposition coefficients of the PWM rectifier are significantly different when it is normal and when it is faulty. Therefore, the current under normal conditions and various fault conditions is decomposed and its wavelet energy spectrum is calculated. It is found that the energy spectrum of each frequency band of different faults is significantly different. In order to facilitate the subsequent analysis and comparison, it is normalized and then input into the trained neural network for fault identification and diagnosis [9]. The simulation results show that the diagnostic accuracy of the algorithm is 100%. It is an accurate and efficient diagnostic algorithm, which provides certain guidance for the rapid and accurate diagnosis and fault-tolerant control of PWM rectifier faults in engineering [10].

References

[1] Liu Zhigang, Ye Bin, Liang Hui. Power Electronics[M]. Beijing: Beijing Jiaotong University Press, 2004.

[2] Luo Hui, Wang Youren, et al. Multi-source feature layer fusion fault diagnosis method for power electronic circuits[J]. Journal of Electric Machines and Control. 2010, 4, Vol. 14(Issue 4): 92-95.

[3] Xu Xin, Fu Xuan. Research on analog circuit fault diagnosis based on wavelet decomposition and BP neural network [J]. Modern Electronic Technology. 2011, 10, Vol. 34 (Issue 19): 171-175.

[4] Wang Yunliang, Meng Qingxue, et al. Fault diagnosis of power electronic devices based on wavelet energy method and neural network [J]. Intelligent Control Technology, 2009, Vol. 31 (Issue 2): 25-27.

[5] Meng Linghui, Wang Lei, et al. Traction converter fault diagnosis based on improved BP neural network [J]. Electronic Design Engineering, 2012, 3, Vol. 20 (Issue 6): 61-63.

[6] Wang Yi. Locomotive traction converter fault diagnosis based on data mining[D]. Chengdu: Southwest Jiaotong University, 2005, 4.10

[7] Ming Tingfeng, Yao Xiaoshan, et al. Centrifugal pump fault diagnosis method based on wavelet-principal component analysis [J]. Journal of Wuhan University of Technology, 010, 12

[8] BingLi and Peilin Zhang.FeatureExtraction and Selection for Diagnosis Gear Using Wavelet Entropy and Mutual Information.[A].2008

[9] Zhimin Dong,Xinqiao Jin,YunyuYang.Fault diagnose for temperature,flow rate and pressure sensors in VAV systems using wavelet neural network[J].Applied Energy,2009,86:1624-1631.

[10] Yaguo Lei, Zhengjia He, Yanyang Zi. Expert Systems with Applications, 2009.

Author profile: Meng Linghui (1988-), male, from Shulan, Jilin Province, doctoral student.

Research direction: power electronics and electrical transmission .