CN115935244B - Single-phase rectifier fault diagnosis method based on data driving - Google Patents

Abstract

The invention discloses a fault diagnosis method of a single-phase rectifier based on data driving, which comprises the following steps: s1, constructing a model of a single-phase PWM rectifier, and obtaining network current measurement and direct-current side voltage fault data of a normal state of the single-phase PWM rectifier, an IGBT open circuit, an anti-parallel diode open circuit, a series resonance circuit inductance open circuit and a capacitance open circuit; s2, performing VMD decomposition on the obtained normal state data and fault data to obtain an intrinsic mode component IMF serving as an input feature vector of a subsequent fault diagnosis network; s3, constructing a single-phase rectifier fault diagnosis sub-model based on CRNN; and S4, integrating the features extracted by the two sub-models, and finally converting the output data into probability values of corresponding categories through softmax. The invention solves the problems that the fault signals of the IGBT and the anti-parallel diode are insensitive to direct-current side voltage and the fault signals of the elements of the series resonance circuit are insensitive to network side current, and the fault elements are accurately positioned.

Description

A data-driven single-phase rectifier fault diagnosis method

technical field

The invention relates to the technical field of single-phase rectifier fault detection, in particular to a data-driven single-phase rectifier fault diagnosis method.

Background technique

AC-DC-AC traction drive system is widely used in the field of high-speed railway traction system. Power electronic transformers play the role of power conversion in traction drive systems, mainly including high-frequency transformers and power electronic converters. Power electronic transformers have complex structures, a large number of components, and frequent hidden faults. The failure of the main components of power electronic transformers often leads to abnormal fluctuations in voltage and current, which endangers the operation safety of high-speed trains. Most of the existing fault diagnosis methods for power electronic transformers are model-based. The basic idea is to use mathematical models to calculate fault residuals, and then perform fault diagnosis through residual analysis. However, in practical engineering applications, the nonlinearity and discreteness of switching devices limit the accuracy of analytical models. Therefore, in the fault diagnosis of complex nonlinear systems, it is extremely urgent to study fault diagnosis methods that are less dependent on the model. In this process, data-based fault diagnosis methods tend to make full use of the rich equipment status information contained in system detection data, and explore the intrinsic relationship between these data and fault modes.

Single-phase pulse width modulation (PWM) rectifier is an important part of the traction drive system, its performance will directly affect the performance of the drive system, and power semiconductor devices are the most vulnerable. Therefore, open-circuit and short-circuit faults of insulated-gate bipolar transistors (IGBTs) have become common faults in grid-side rectifiers. Generally, when a short-circuit fault occurs, the voltage and current will rise sharply in a short period of time, and then the short-circuit fault will turn into an open-circuit fault. Therefore, the open-circuit fault analysis and diagnosis of IGBT is the core of rectifier fault diagnosis. In addition, the abnormal voltage stress caused by the anti-parallel diode open-circuit fault is much higher than the blocking voltage of the IGBT, which may cause the IGBT to overvoltage fault in a short time. For the intermediate DC circuit, the research on the diagnosis method of the component failure of the series resonant circuit is less.

The model-based method is the most widely used method in the diagnosis of IGBT open-circuit faults. Its diagnostic performance is highly dependent on the model accuracy, but the model accuracy is difficult to obtain, especially considering the complexity of the traction system and the change of current direction on the operation of anti-parallel diodes. Impact. Signal-based and machine learning-based fault diagnosis methods differ from model-based fault diagnosis methods in that precise modeling is not required, but faults are diagnosed by analyzing the amplitude-frequency characteristics of voltage and current. Therefore, the effectiveness of data-driven methods is largely affected by feature extraction performance. The method based on signal processing basically only uses the extraction of shallow features, which include amplitude information, frequency information, energy information, etc. Through the analysis of the above characteristic information, the corresponding relationship between the fault signal and the fault mode can be established. However, it is often difficult to establish such relationships directly through the characteristics of the data itself. At present, there are three key issues that must be solved by using data-driven methods for fault diagnosis of single-phase rectifiers: (1) The method must have good generalization and robustness. For different IGBTs on single-phase rectifiers , Diodes and series resonant circuit components can all be used for fault diagnosis, replacing manual diagnosis to achieve intelligent diagnosis. (2) The method must have the characteristics of short time, because the abnormality of key components will cause large changes in parameters such as module current, which will cause the entire module to work abnormally in a short time. (3) The algorithm model must have high precision and high stability. The single-phase rectifier has a complex system structure and many components. It needs to be able to accurately diagnose faulty devices to ensure the stable operation of the train.

Contents of the invention

Aiming at the above-mentioned problems existing in the existing data-driven method for single-phase rectifier fault diagnosis, the present invention provides a data-driven single-phase rectifier fault diagnosis method. Comprehensive fault diagnosis of single-phase rectifiers is achieved by combining variational mode decomposition (VMD) method and dual-model convolutional recurrent neural network (CRNN).

The data-driven single-phase rectifier fault diagnosis method provided by the present invention has the following steps:

S1. Build a single-phase PWM rectifier model through the dSPACE hardware circuit test platform, and use this platform to obtain the normal state of the single-phase PWM rectifier, IGBT open circuit, anti-parallel diode open circuit, series resonant circuit inductor open short circuit, and capacitor open short circuit network measurement current and DC link voltage fault data.

S2. Perform VMD decomposition on the obtained normal state data and fault data to obtain the intrinsic mode component IMF, which is used as an input feature vector of the subsequent fault diagnosis network. Specifically include the following sub-steps:

S21. Find K IMF components with specific sparsity by constraining the variational model, so that the estimated bandwidth sum of each component is the smallest, and the constrained condition is that the sum of each component is equal to the original signal. where specific sparsity refers to the percentage of zero elements in a matrix or dataset relative to the total number of elements in the dataset.

相乘，将模态的频谱调制到相应基频带，再计算解析信号梯度平方范数L²，构造变分模态。S22, in order to obtain K IMFs with limited bandwidth, first obtain the single-marginal spectrum of each IMF component u _k (t) by Hilbert transform, then estimate the center frequency ω _k of each IMF and its exponential signal

Multiply, modulate the frequency spectrum of the mode to the corresponding base frequency band, and then calculate the analytical signal gradient square norm L ² to construct the variational mode.

S23. Introduce the penalty factor α and the Lagrange multiplier λ to transform the constrained variational problem into an unconstrained variational problem, so as to solve the above-mentioned variational problem and obtain an augmented Lagrange expression.

S24. Using the alternate direction multiplier algorithm to update and iteratively solve the saddle point, and obtain an optimal solution, so as to decompose the original signal into K IMF components.

S3. Building a CRNN-based single-phase rectifier fault diagnosis sub-model, including a CRNN current sub-model and a CRNN voltage sub-model.

Step S3 includes the following sub-steps:

S31. Build a feature extraction module based on 1D-CNN. The convolution layer of CNN is composed of two parts: the first part is convolution layer C1 and C2, which perform convolution operation to extract the structural features of adjacent domains; the second part is The pooling layers S1 and S2 perform downsampling operations to remove redundant information of feature maps.

In this step, each signal under normal state and fault is passed as input to the CNN layer through a one-dimensional convolution filter; the feature map is obtained through the convolution operation of the l-th convolution layer

和

分别代表第j层卷积滤波器的权重和偏置值，M_j是输入特征图的数量；in,

and

Represent the weight and bias value of the j-th layer convolution filter respectively, M _j is the number of input feature maps;

The pooling process is after the convolution process, and the maximum pooling operation formula is as follows:

和

代表最大池化层的权重和偏置值，dowm()代表最大池化函数。in,

and

Represents the weight and bias values of the maximum pooling layer, and dowm() represents the maximum pooling function.

S32. Build a secondary feature extraction module based on SRU. The SRU network uses matrix multiplication to perform operations. Each gating structure needs to process the input x _t through an activation function, and finally reset the gate and the internal state of the cell and input x _t to obtain the output h _t ; the calculation formula is as follows:

f _t =α _f (W _f *x _t )+b _f

r _t =α _r (W _r *x _t )+b _r

h _t ＝r _t ⊙tanh(c _t )+(1-r _t )⊙x _t

线性表示、f_t为遗忘门、r_t为重置门，c_t代表的是内在状态；α_f和α_r分别是遗忘门gate_f和重置门gate_r的S激活函数；W、W_f和W_r分别是线性表示

遗忘门gate_f和重置门gate_r的权重；b_f和b_r分别表示遗忘门gate_f和重置门gate_r的偏差；⊙表示对应元素相乘；tanh()表示为隐状态的双曲正切激活函数。In the formula, for

Linear representation, f _t is the forget gate, r _t is the reset gate, c _t represents the internal state; α _f and α _r are the S activation functions of the forget gate gate _f and the reset gate gate _r respectively; W, W _f and W _r are linear representations respectively

The weight of the forgetting gate _f and the reset gate gate _r ; b _f and b _r respectively represent the deviation of the forgetting gate _f and the reset gate gate _r ; ⊙ represents the multiplication of corresponding elements; tanh() represents the hyperbolic of the hidden state Tangent activation function.

S33. Introduce an attention mechanism before the fully connected layer. The present invention uses SEnet (Squeeze-and-Excitation Network), which considers the relationship between feature channels, and adds an attention mechanism to feature channels. First, through the squeeze operation, each feature map is globally pooled and averaged into a real value; then the excitaton operation is performed, in which the dimensionality of the C channels is first reduced and then expanded back to the C channel; finally, the output of exciation is regarded as The importance of each channel after feature selection is multiplied to the previous features by multiplicative weighting to realize the function of promoting important features and suppressing unimportant features.

S34. Use the IMF obtained after VMD decomposition of the grid-side current and the DC-side voltage as the input of the built network model to form a dual-model architecture, that is, the CRNN current sub-model and the CRNN voltage sub-model. In the current and voltage sub-models, model for training and computation.

S4. The features extracted by the two sub-models are integrated through the Flatten layer, and the extracted feature information is converted to the label space through the fully connected layer to complete the data classification. The output is expressed as:

分别为输入数据和输出数据；In the formula, D _i and b _i are the learning parameters of the fully connected layer,

are input data and output data respectively;

The output data is converted into the probability value of the corresponding category by softmax, and its expression is:

In the formula, y _i is the ith parameter value of vector y (i=1,2,...,j).

Confusion matrix and T-SNE visualization are used as visualization tools to analyze the recognition results of test samples. The former visualizes the algorithm performance in matrix form, and the latter presents the visualization effect of sample distribution through feature maps.

Compared with the prior art, the benefits of the present invention are:

1. The present invention obtains data by building a model of a single-phase PWM rectifier on a dSPACE hardware circuit test platform, which solves the problem of difficulty in obtaining fault data and a small amount of data, and improves the comprehensiveness and reliability of fault diagnosis on the premise of ensuring diagnostic safety. reliability.

2. Through the signal processing method of VMD, the present invention increases the difference between the original similar fault signals, obtains stable fault characteristics from massive data, and solves the problem of the same bridge arm when an open circuit fault occurs in the IGBT or diode. IGBTs and diodes on the device are difficult to identify due to similar fault signatures. This method can specify the number of modes that need to be decomposed, avoid the situation of useless components, reduce the steps of subsequent feature processing, and solve the endpoint effect and mode mixing that are easy to occur in conventional signal processing methods through the method of mirror extension. overlapping problem.

3. The present invention designs a CRNN network, which solves the problem that when the 1D-CNN network extracts the structural features of the adjacent domain of the signal, the time series features contained in it are not sufficiently mined and the recursive structure of the RNN network is conducive to the model extraction of time series characteristics. However, at the expense of model calculation speed, it is easy to produce the two major pain points of gradient disappearance and over-fitting phenomenon. The CRNN network enables the fault recognition model to have the two advantages of fast training speed of CNN and high recognition accuracy of RNN.

4. The present invention designs a dual-model architecture. By establishing voltage and current sub-models separately, the grid-side current and DC-side voltage IMF components decomposed by VMD are respectively used as the input of the sub-models. It solves the problem that the IGBT and anti-parallel diode fault signals are not sensitive to the DC side voltage and the series resonant circuit component fault signal is not sensitive to the grid side current, making full use of the grid side current and DC side voltage to achieve the purpose of accurately locating faulty components.

Other advantages, objectives and features of the present invention will partly be embodied through the following descriptions, and partly will be understood by those skilled in the art through the research and practice of the present invention.

Description of drawings

FIG. 1 is a flow chart of the data-driven single-phase rectifier fault diagnosis method of the present invention.

Figure 2 is a topology diagram of a single-phase PWM rectifier.

Figure 3 is a flow chart of VMD decomposition.

Figure 4 is a schematic diagram of the dual-model CRNN framework.

Figure 5 is a diagram of the confusion matrix results.

Detailed ways

The preferred embodiments of the present invention will be described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described here are only used to illustrate and explain the present invention, and are not intended to limit the present invention.

As shown in Figures 1-4, the data-driven single-phase rectifier fault diagnosis method of the present invention comprises the following steps:

Step S1: Build a single-phase PWM rectifier model through the dSPACE hardware circuit test platform, and obtain the normal state of the single-phase PWM rectifier, IGBT open circuit, anti-parallel diode open circuit, series resonant circuit inductor open-short circuit and capacitor open-short circuit network measurement current through this platform and DC link voltage fault data.

dSPACE can control virtual objects with real controllers to realize extreme simulation tests of real environments. Therefore, using the dSPACE-based data as the measured data has a good verification effect on the effectiveness of the proposed method. It consists of a digital signal processor (DSP) as the executor of the control algorithm, in which the control chip is TMS320F28335, a dSPACE simulator and a host computer as the control interface. The PWM pulse signal is input to dSPACE by the DS2103 board. The grid side voltage UN, the input current iN of the rectifier circuit and the main circuit DC side voltage Udc are output to the AD sampling module of the DSP by the DS2002 board and displayed by the oscilloscope. The topological circuit of the single-phase PWM rectifier is shown in Figure 2. The control strategy of DQ decoupling is adopted, so that the phase of the grid-side voltage and grid-side current of the rectifier is consistent, and both can be kept in the state of unity power factor, and the grid-side current harmonics Less content, superior steady-state performance.

相乘将模态的频谱调制到相应基频带，再计算解析信号梯度平方范数L²，构造变分模态。引入惩罚因子α及Lagrange乘子λ将约束变分问题转变为非约束变分问题,以求解上述变分问题,得到增广Lagrange表达式。采用交替方向乘子算法更新迭代求解鞍，获得最优解,以将原始信号分解为K个IMF分量。分别对网侧电流和直流侧电压进行先进行频谱分析得到K的最佳取值是3，然后再进行VMD分解得到对应的IMF。Step S2: Perform VMD decomposition on the obtained normal data and fault data to obtain the intrinsic mode component (IMF), which is used as the input feature vector of the subsequent fault diagnosis network. The set sampling frequency is 100000HZ, that is, 100000 data can be collected in 1 second, and there will be many useless information and features, so data preprocessing is required. An obvious feature of VMD is that the number K of the modal components needs to be set before the signal is decomposed by VMD. For some scenarios where the hidden mode number of the signal is not predicted, the determination of the K value is very important for VMD. Therefore, considering that the decomposed IMFs have independent center frequencies, the original data can be transformed by fast Fourier transform first, and the spectrogram of the data can be obtained. It is generally believed that the components with larger amplitudes have larger energy, which means Contains more features, so it is reasonable to choose such components as features. After determining the value of K according to the spectrogram, VMD decomposition can be performed. The flowchart of VMD decomposition is shown in Figure 3. In the figure, {u _k }={u ₁ ,u ₂ ,u ₃ ,…,u _K } represent K IMFs, and {ω _k }={ω ₁ ,ω ₂ ,ω ₃ ,…,ω _K } represent each The center frequency of the component; λ is the Lagrangian multiplier, α is the penalty factor, τ is the fidelity coefficient, ∧ is the Fourier transform, and n is the number of iterations. In the final satisfaction condition, ε is the discriminant accuracy, which is 10e-6 here. By constraining the variational model, seek K IMF components with specific sparsity, so that the estimated bandwidth sum of each component is the smallest, and the constraint condition is that the sum of each component is equal to the original signal. In order to obtain K IMFs with limited bandwidth, the single marginal spectrum of each IMF component u _k (t) is obtained through Hilbert transform first, and then the center frequency ω _k and its exponential signal of each IMF are estimated

Multiplication modulates the frequency spectrum of the mode to the corresponding baseband, and then calculates the gradient square norm L ² of the analytical signal to construct the variational mode. A penalty factor α and a Lagrange multiplier λ are introduced to transform the constrained variational problem into an unconstrained variational problem, so as to solve the above variational problem and get the augmented Lagrange expression. The alternate direction multiplier algorithm is used to update the iterative solution saddle to obtain the optimal solution to decompose the original signal into K IMF components. The optimal value of K is 3 by performing spectrum analysis on the grid side current and the DC side voltage respectively, and then performing VMD decomposition to obtain the corresponding IMF.

Step S3: Build a CRNN-based single-phase rectifier fault diagnosis sub-model, including a CRNN current sub-model and a CRNN voltage sub-model.

First, a downsampling layer is introduced. Due to the large sample length, it is easy to generate too many network parameters and lead to overfitting, which greatly reduces the scale of network parameters, enhances the robustness of the algorithm, and greatly shortens the network operation time. CNN has the property of sparse weights, which can detect small but meaningful features by using convolutional filters much smaller than the input. This means that CNN reduces the number of parameters that need to be stored and significantly improves the efficiency of feature extraction. The convolutional layer of CNN generally consists of two parts: (1) the first part performs convolution operation to extract features; (2) the second part performs pooling operation to adjust the output of the convolutional layer. In CRNN, convolutional layer functions are considered as feature extractors. The individual signals in normal state and fault are passed as input to the CNN layer through a 1D convolutional filter. The feature map is obtained by the convolution operation of the l-th convolutional layer (l∈l _c )

和

分别代表第j层卷积滤波器的权重和偏置值，M_j是输入特征图的数量。in

and

Represent the weight and bias value of the j-th convolutional filter, respectively, and _Mj is the number of input feature maps.

The pooling process plays the role of secondary extraction after the convolution process. Max pooling is used to reduce the dimensionality of the data and preserve useful information:

和

代表最大池化层的权重和偏置值，dowm()代表最大池化函数。in

and

Represents the weight and bias values of the maximum pooling layer, and dowm() represents the maximum pooling function.

不再依赖历史迭代状态h_t-1，引入重置门状态r_t改进时序特征提取方式。因此，SRU从结构、运算优化两方面实现并行化加速。SRU的优势具体表现为：与LSTM相比，具有同样层数的SRU提取特征更快，信息量损耗更少。RNNs are suitable for tasks such as natural semantic analysis and time series modeling. The RNN module of CRNN uses SRU unit, which has faster computing power than LSTM. Build a secondary feature extraction module based on SRU. SRU realizes parallel acceleration in terms of structure and operation optimization. SRU preprocesses the input data relatively independently for each recursive extraction of time series features, and then performs recursive feature extraction through relatively lightweight parallel operations. SRU simplifies the feature extraction process of historical information h _t , expands the parallelism of operations similar to 1D-CNN, adopts the update mechanism of the retained information storage unit cell, and constructs a reset gating unit gate _r to dynamically adjust the recursive step size and eliminate the output gate Gate _o produces gradient disappearance phenomenon at long recursive steps. In the SRU, the gated unit state

Instead of relying on the historical iteration state h _t-1 , the reset gate state r _t is introduced to improve the timing feature extraction method. Therefore, SRU achieves parallel acceleration from two aspects of structure and operation optimization. The advantages of SRU are as follows: Compared with LSTM, SRU with the same number of layers extracts features faster and has less information loss.

The attention mechanism is introduced before the fully connected layer, which can focus on important information with high weight and ignore irrelevant information with low weight, and can also continuously adjust the weight so that important information can be selected in different situations, so It has better scalability and robustness. The present invention uses SEnet (Squeeze-and-ExcitationNetwork), which considers the relationship between feature channels, and adds an attention mechanism to feature channels. First, through the squeeze operation, global pooling is performed on each feature map, and the average is a real value. This real number somehow has a global receptive field. This operation enables features close to the data input to also have a global receptive field. Next is the excitaton operation. After the squeeze operation, the network outputs a feature map of 1xC size, and the weight w is used to learn the direct correlation of C channels. In this process, the dimensions of C channels are first reduced and then expanded back to C channels. The advantage is that on the one hand, it reduces the amount of network calculation, and on the other hand, it increases the nonlinear capability of the network. Finally, the output of exciation is regarded as the importance of each channel after feature selection, which is multiplied to the previous features by multiplicative weighting, so as to realize the function of improving important features and suppressing unimportant features.

The primary feature extraction module of the CRNN framework consists of two alternately stacked 1D-CNN convolution and pooling layers, and the secondary feature extraction module consists of four SRU cyclic layers. The feature extraction module based on 1D-CNN includes a convolutional layer and a pooling layer, and extracts a structural feature map from a single-phase rectifier fault signal rectifier _t (t=1, 2, ..., T):

F _i ＝(F _i.1 ,F _i,2 ,...F _i,j ),j=1,2,...,m _k

The SRU-based feature primary extraction module receives the structural feature map F _i and extracts time series features h _t (t=1, 2,...T).

The IMFs of the grid-side current and DC-side voltage decomposed by VMD are used as input, respectively, and are trained and calculated in the current and voltage sub-models.

Step S4: integrate and output the features extracted by the two sub-models.

The structural block diagram of DCRNN is shown in Figure 4. The features extracted by the two sub-models are integrated through the Flatten layer, and the extracted feature information is converted into the label space through the fully connected layer to complete the data classification. The output can be expressed as:

In the formula, D _i and b _i are the learning parameters of the fully connected layer. Then, the output data is transformed into the probability value of the corresponding category by softmax, and its expression is:

In the formula, y _i is the ith parameter value of vector y (i=1,2,...,j).

Confusion matrix and T-SNE visualization are used as visualization tools to analyze the recognition results of test samples. The former visualizes the algorithm performance in matrix form, and the latter presents the visualization effect of sample distribution through feature maps.

The visualization results of the confusion matrix show that DCRNN is very sensitive to the fault characteristics of the single-phase rectifier, and can effectively distinguish whether the operating state of the single-phase rectifier is faulty, and which specific component of the single-phase rectifier is faulty.

T-SNE is a stochastic optimization dimensionality reduction algorithm, which is mainly used for dimensionality reduction visualization of nonlinear high-dimensional data, and visualizes feature distribution in low-dimensional spaces such as two-dimensional or three-dimensional. The core of the T-SNE algorithm is to minimize the original data distribution and KL divergence, and to find a suitable low-dimensional mapping. As a key adjustable parameter of T-SNE, the complexity (preplexity, prep) characterizes the estimated range of potential adjacent points of the sample point. The setting interval of perp value is [5-10]. For single-phase rectifier fault identification, let the high-dimensional feature extracted by DCRNN be X _bogie ={x ₁ ,x ₂ ,…,x _n }, and obtain the low-dimensional feature distribution Y _bogie ={y ₁ after dimensionality reduction based on T-SNE ,y ₂ ,...,y _n }. Visualization of single-phase rectifier fault identification based on T-SNE, in the low-dimensional feature space, the sample boundaries between different fault categories are obvious. This shows that DCRNN has excellent feature extraction ability from the aspect of feature distribution, which makes it achieve better identification accuracy and speed in single-phase rectifier fault identification.

The overall scheme of the present invention has a complete diagnosis process. In the model-based method, its diagnostic performance is highly dependent on the accuracy of the model, and accurate modeling is very difficult. Therefore, a data-driven single-phase rectifier fault diagnosis method is adopted.

The method of data acquisition uses the dSPACE hardware circuit test platform to build a model of a single-phase PWM rectifier. It is safer to set faults and avoid the collapse of the entire circuit due to a component failure. And using this collection method, you can obtain detailed data of grid-side current and DC-side voltage without installing additional sensors, which solves the problem that fault data is difficult to obtain and the amount of data is small, and improves the accuracy of fault diagnosis on the premise of ensuring diagnostic safety. comprehensiveness and reliability.

Due to the topology of the single-phase rectifier, the fault conditions are similar on the IGBTs and anti-parallel diodes on the same leg, with four IGBTs and four anti-parallel diodes on the two legs. Through the VMD decomposition method based on the signal processing method, the difference between the original similar fault signals is increased, and the number of modes that need to be decomposed is determined by comparing with the spectrum diagram, and the most significant feature of each fault signal is accurately extracted to overcome the existence of the EMD method. Endpoint effects and modal component aliasing issues.

The method of the invention combines the advantages of the CNN network and the RNN network, taking into account stability, accuracy, generalization and rapidity. Since the fault characteristics of IGBT and anti-parallel diodes are mainly reflected in the grid side current and the fault characteristics of series resonant circuit components are mainly reflected in the DC side voltage, a dual-model architecture is designed. By establishing voltage and current sub-models respectively, the The grid-side current and DC-side voltage IMF components decomposed by VMD are respectively used as the input of the sub-model. Make full use of the grid side current and DC side voltage to effectively distinguish the faults of IGBTs, anti-parallel diodes and series resonant circuit components, and achieve the purpose of accurately locating faulty components. The visualization results of the confusion matrix (Figure 5) show that DCRNN is very sensitive to the fault characteristics of single-phase rectifiers, and the accuracy rate of distinguishing whether the operating status of single-phase rectifiers is faulty is 100%, and the accuracy of identifying each fault of single-phase rectifiers is 100%. 95.83%, meet the diagnostic requirements.

The above description is only a preferred embodiment of the present invention, and does not limit the present invention in any form. Although the present invention has been disclosed as above with preferred embodiments, it is not intended to limit the present invention. Anyone familiar with this field Those skilled in the art, without departing from the scope of the technical solution of the present invention, can use the technical content disclosed above to make some changes or modify equivalent embodiments with equivalent changes. Any simple modifications, equivalent changes and modifications made to the above embodiments by the technical essence still belong to the scope of the technical solutions of the present invention.

Claims (7)

1. A data-driven single-phase rectifier fault diagnosis method, characterized in that the steps are as follows: S1. Build a single-phase PWM rectifier model through the dSPACE hardware circuit test platform, and use this platform to obtain the normal state of the single-phase PWM rectifier, IGBT open circuit, anti-parallel diode open circuit, series resonant circuit inductor open short circuit, and capacitor open short circuit network measurement current and DC side voltage fault data; S2. Decompose the obtained normal state data and fault data into VMD to obtain the intrinsic mode component IMF, and use it as the input eigenvector of the subsequent fault diagnosis network; S3. Build a CRNN-based single-phase rectifier fault diagnosis sub-model, including a CRNN current sub-model and a CRNN voltage sub-model; S4. The features extracted by the two sub-models are integrated through the Flatten layer, and the extracted feature information is converted into the label space through the fully connected layer to complete the data classification, and the final output data is converted into the probability value of the corresponding category by softmax. 2. The single-phase rectifier fault diagnosis method based on data drive as claimed in claim 1, is characterized in that, described step S2 comprises the following sub-steps: S21. By constraining the variational model, seek K IMF components with specific sparsity, so that the estimated bandwidth sum of each component is the smallest, and the constraint condition is defined as the sum of each component, and equal to the original signal; wherein, the specific sparsity refers to the matrix or the percentage of zero elements in the dataset relative to the total number of elements in the dataset; S22, in order to obtain K IMFs with limited bandwidth, first obtain the single-marginal spectrum of each IMF component u _k (t) by Hilbert transform, then estimate the center frequency ω _k of each IMF and its exponential signal Multiply, modulate the frequency spectrum of the mode to the corresponding base frequency band, and then calculate the analytical signal gradient square norm L ² to construct the variational mode; S23. Introduce the penalty factor α and the Lagrange multiplier λ to convert the constrained variational problem into an unconstrained variational problem, so as to solve the above-mentioned variational problem and obtain an augmented Lagrange expression; S24. Using the alternate direction multiplier algorithm to update and iteratively solve the saddle point, and obtain an optimal solution, so as to decompose the original signal into K IMF components. 3. The data-driven single-phase rectifier fault diagnosis method as claimed in claim 1, wherein said step S3 comprises the following sub-steps: S31. Building a feature extraction module based on 1D-CNN. The first part is the convolution layer C1 and C2, which perform convolution operations to extract the structural features of adjacent domains; the second part is the pooling layer S1 and S2, which perform downsampling Operate to remove redundant information of the feature map; S32. Build a secondary feature extraction module based on SRU. The SRU network uses matrix multiplication to perform operations. Each gating structure needs to process the input x _t through an activation function, and finally reset the gate and the internal state of the cell and input x _t to obtain the output h _t ; the calculation formula is as follows: f _t =α _f (W _f *x _t )+b _f r _t =α _r (W _r *x _t )+b _r h _t ＝r _t ⊙tanh(c _t )+(1-r _t )⊙x _t In the formula, is a linear representation, f _t is the forget gate, rt _is the reset gate, c _t represents the internal state; α _f and α _r are the S activation functions of the forget gate _f and reset gate _r respectively; W, W _f and W _r are linear representations respectively The weight of the forgetting gate _f and the reset gate gate _r ; b _f and b _r respectively represent the deviation of the forgetting gate _f and the reset gate gate _r ; ⊙ represents the multiplication of corresponding elements; tanh() represents the hyperbolic of the hidden state Tangent activation function; S33. Introduce an attention mechanism before the fully connected layer; S34. Use the IMF obtained after VMD decomposition of the grid-side current and the DC-side voltage as the input of the built network model to form a dual-model architecture, that is, the CRNN current sub-model and the CRNN voltage sub-model. In the current and voltage sub-models, model for training and computation. 4. The data-driven single-phase rectifier fault diagnosis method as claimed in claim 3, wherein in said step S31, each signal under normal state and fault is passed to CNN as input through a one-dimensional convolution filter layer; the feature map is obtained through the convolution operation of the l-th convolutional layer Among them, k _ij ^l and b _j ^l represent the weight and bias value of the j-th layer convolution filter respectively, M _j is the number of input feature maps, is the feature map of the previous layer; The pooling process is after the convolution process, and the maximum pooling operation formula is as follows: Among them, β _j ^l and b _j ^l represent the weight and bias value of the maximum pooling layer, dowm() represents the maximum pooling function, is the feature map after pooling. 5. The data-driven single-phase rectifier fault diagnosis method as claimed in claim 3, characterized in that, the specific method of the step S33 is: firstly perform global pooling on each feature map through the squeeze operation, and average into one Real value; then perform the excitaton operation, in which the dimensionality of C channels is first reduced and then expanded back to C channels; finally, the output of exciation is used as the importance of each channel after feature selection, and multiplied to the previous by multiplicative weighting In terms of features, it is possible to enhance important features and suppress unimportant features. 6. the single-phase rectifier fault diagnosis method based on data drive as claimed in claim 1, is characterized in that, described step S4 is specifically as follows: The features extracted by the two sub-models are integrated through the Flatten layer, and the extracted feature information is converted to the label space through the fully connected layer to complete the data classification. The output is expressed as: In the formula, D _i and b _i are the learning parameters of the fully connected layer, are input data and output data respectively; The output data is converted into the probability value of the corresponding category by softmax, and its expression is: In the formula, y _i is the ith parameter value of vector y, i=1,2,...,j. 7. The data-driven single-phase rectifier fault diagnosis method according to claim 6, characterized in that, in the step S4, the confusion matrix and T-SNE visualization are used as visualization tools for analyzing the identification results of test samples.

Images