The actual operation of power electronic circuit shows that most of the faults are damage of power switching devices, among which open circuit and straight circuit are common. The fault diagnosis of power electronic circuit is quite different from the fault diagnosis of general analog circuit and digital circuit. Due to the small overload capacity and fast damage speed of power electronic devices, their fault information only exists within tens of milliseconds before the failure to power outage. Therefore, real-time monitoring and online diagnosis are required. In addition, the power of the power electronic circuit has reached thousands of kilowatts, and the method of changing the input to see the output used in the diagnosis of analog circuit and digital circuit is no longer applicable. Only the output waveform can be used to diagnose whether and what kind of fault exists in the power electronic circuit.

The actual operation of power electronic circuit shows that most of the faults are damage of power switching devices, among which open circuit and straight circuit are common. The fault diagnosis of power electronic circuit is quite different from the fault diagnosis of general analog circuit and digital circuit. Due to the small overload capacity and fast damage speed of power electronic devices, their fault information only exists within tens of milliseconds before the failure to power outage. Therefore, real-time monitoring and online diagnosis are required. In addition, the power of the power electronic circuit has reached thousands of kilowatts, and the method of changing the input to see the output used in the diagnosis of analog circuit and digital circuit is no longer applicable. Only the output waveform can be used to diagnose whether and what kind of fault exists in the power electronic circuit.

The key of fault diagnosis is to extract fault features. Fault feature refers to the comprehensive quantity that reflects the type, location and degree of fault of equipment and system after the signal reflecting fault symptom is processed. According to the difference of feature extraction methods, fault diagnosis methods can be divided into spectral analysis method, mathematical model based dynamic system method, pattern recognition method, neural network based method, expert system method, wavelet transform method and genetic algorithm. These methods are described in detail below.

Spectral analysis method in fault diagnosis

Spectrum analysis is a common signal processing method in fault diagnosis. Commonly used Fourier spectrum, Walsh spectrum, filtering, correlation analysis and so on. The purpose of spectrum analysis is to extract features from signals containing noise. The time domain waveform of the fault signal cannot clearly reflect the characteristics of the fault. However, the key point signal containing fault information in power electronic circuit is usually periodic, so the fault waveform can be transformed from time domain to frequency domain by Fourier transform to highlight the fault characteristics and realize fault diagnosis.

The Fourier transform is the decomposition of a periodic function into sinusoidal components of various frequencies, and similarly, the Walsh transform is the decomposition of a function into a set of Components of the Walsh function. Adaptive filter is a kind of digital signal processing statistical method, it does not need to know the first and second order prior statistical knowledge of the signal, directly use the observation data, through operation to change some parameters of the filter, and the output of the adaptive filter can automatically track the change of signal characteristics. In fault diagnosis of power electronic system, adaptive processing can be used to realize noise cancellation, spectral line enhancement and other functions, and fault features can be extracted from noise background, so as to achieve accurate diagnosis.

Parameter model and fault diagnosis

If the mathematical model of the system is known, the state and parameters of the system can be estimated by measurement, and whether the state variables and system parameters change can be determined. The method of fault diagnosis based on system mathematical model can estimate multiple states or system parameters from fewer measurement points, so as to realize fault diagnosis.

Furthermore, it can be divided into three methods: detection filter method, state estimation method and parameter identification method.

1. Detection filter method

It fixes the output direction of the fault of the component, actuator and sensor in a specific direction or plane.

2. State estimation method

By monitoring the state change of the system, it can also reflect the fault caused by the change of system parameters and diagnose the fault. Different from general state estimation, in fault diagnosis, the unknown state information is not estimated, but the system output is reconstructed by means of observer or Kalman filter, so as to obtain the estimated value of system output. The difference between this estimate and the actual output is called the measurement residual. Residual error contains a lot of information about the internal changes of the system, so it can be used as a basis for fault diagnosis. The advantages of the state estimation method are that the on-line calculation is small and the diagnosis is fast.

3. Parameter identification method

Identify the parameters of the system model in real time, and compare with the parameters of the normal model to determine the fault. The common one is the square method.

The application of pattern recognition in fault diagnosis

Fault pattern recognition is to extract fault features from system information and classify faults according to different attributes of these features. Fault diagnosis by pattern recognition method is based on the mathematical characteristics of samples, so it does not need accurate mathematical model. The pattern recognition method is also suitable for some problems where the mathematical model of the diagnosed object is too complex to be solved. In addition, in the fault diagnosis of industrial systems, we should try to make use of non-mathematical (including physical and structural) features, and design a variety of feature extractors, which will be beneficial to use the knowledge of existing systems and reduce the workload of calculation. Because feature selection and extraction are closely related to the pattern to be recognized, it is difficult to follow a general rule. At present, the commonly used methods include distance classification, Bayes classification, Fisher discriminant, feature extraction from parameter model, k-L transform, etc.

4. Fault diagnosis method based on neural network

Based on the self-learning and self-induction ability of neural network, the image relation between fault signal and fault classification is established after certain training. The learning neural network is used to realize fault diagnosis. Neural network is a network of extensive interconnection of a large number of neurons. Here the BP network is introduced as an example. BP network is a one-way propagation multi-layer forward network, which is composed of input layer, intermediate layer and output layer. The intermediate layer can have several layers, and the neurons of each layer only accept the output of the neurons of the previous layer. There is no feedback in BP network, there is no coupling between nodes of the same layer, and nodes of each layer only affect the input of nodes of the next layer.

The general learning algorithm adopted by BP network is to compare the output of the network with the desired output, and then adjust the weight of the network according to the difference between the two, so as to reduce the error. When power electronic circuit failure occurs, if you can use the learning ability of neural network, the relationship between fault waveform and the cause of the problem through the study of neural network after stored in its structure and power, and will learn better neural network used in fault diagnosis, neural network can be through the analysis of the current voltage or current waveform, it is concluded that the cause of the problem. In this way, the fault can be automatically diagnosed online.

Expert system

Because the fault diagnosis is from the monitoring and diagnosis of the object characterization to find the cause of the fault, the location, and determine the severity of the fault, therefore, if the known fault to analyze the system or equipment operation characteristics and characterization is called the positive problem, then the fault diagnosis is the inverse problem. The solution of this inverse problem is obviously different from the solution of the positive problem, and the Expert System ES in Artificial Intelligence technology is a favorable tool to solve this inverse problem. Expert system is a branch of artificial intelligence. It realizes fault diagnosis by simulating expert's experience. The structure of the expert system is shown in the following table: A typical diagnostic expert system collects and stores data through online monitoring, and then transmits it to the diagnostic operation center, where the expert system processes, analyzes and diagnoses, and automatically feeds back the diagnosis results and processing suggestions to the operation site. Therefore, expert system is the core part of diagnostic system. In the following part of this paper, the author introduces the application of expert system method to fault diagnosis.

Six, wavelet transform method

In fault diagnosis, sudden signal usually corresponds to some kind of fault of equipment. It is one of the effective methods to analyze and identify various waveform signals generated in the system and judge their status. When the equipment runs normally, the signal is relatively stable. Once the equipment fails, the dynamic non-stationary signal with singularity will be sent out. In order to realize the fast and accurate detection of equipment fault, it is necessary to identify the non-stationary signal of the moment when the fault occurs effectively. Signal processing and analysis is the basis of fault prediction and diagnosis. Improving the accuracy of diagnosis requires signal processing and analysis methods. Wavelet transform provides a new and powerful analysis means for fault diagnosis with its prominent advantages of local analysis of non-stationary signals and good time-frequency positioning function. It makes up for the "bottleneck" problem of knowledge acquisition in traditional fault diagnosis because it is difficult to accurately describe the experience and knowledge of experts.

Genetic algorithm

Genetic Algorithm (GA) is a newly developed optimization Algorithm. At present, GA has become a new idea and method to solve highly complex problems. Based on the evolutionary rules of survival of the fittest and survival of the fittest, it carries out genetics-based operations on the populations containing possible solutions, constantly producing new populations and constantly evolving the populations, and at the same time optimizing the best individuals in the populations by global parallel search to obtain the optimal solutions that meet the requirements. GA is also used in power electronic fault diagnosis system due to its advantages of obtaining global optimal solution with high probability, less calculation time and strong robustness. The collected information is rationally used (that is, the collected information is divided into three layers), and the hierarchical information fault diagnosis is carried out by genetic algorithm. The application of genetic algorithm in reasoning and self-learning of fault diagnosis expert system can overcome the obstacles of slow reasoning speed and difficult knowledge acquisition in the case of little prior knowledge, and improve the adaptability of expert system.