Method for diagnosing open-circuit fault of IGBT (Insulated Gate Bipolar Translator) of T-type three-level inverter and fault of current sensor

Abstract

The invention discloses a method for diagnosing an IGBT open-circuit fault and a current sensor fault of a T-type three-level inverter, belongs to the technical field of inverter fault diagnosis, and is used for solving the problems of diagnosis and positioning when a three-level T-type inverter IGBT switch tube and a sensor have faults. Comprising the following steps: classifying and marking and numbering fault types of an open-circuit fault of a T-type three-level inverter IGBT and a fault of a current sensor; fault samples are collected and divided into a training set and a test set according to the proportion of 7: 3; a TCN-LSTM-SelfAction fault diagnosis model is constructed, and model training is carried out by using the training set; the trained model is used for evaluating the accuracy of fault type diagnosis of the training set and the test set. According to the diagnosis method, the fault types and the fault positions of the IGBT open-circuit fault and the current sensor fault can be diagnosed at the same time, the TCN and the LSTM are selected to be combined, local information and global relation of fault features can be captured at the same time, and meanwhile, the performance of a model is improved by combining a self-attention mechanism in the LSTM, and the accuracy is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of inverter fault diagnosis, and particularly relates to a diagnosis method for IGBT open-circuit fault and current sensor fault of a T-type three-level inverter. BACKGROUND

[0002] With the rapid development of the power industry and the continuous progress of power electronics technology, T-type three-level inverters are widely used in more and more medium and high voltage and high power industry fields. Compared with neutral-point-clamped three-level inverters, T-type inverters have fewer devices, smaller losses and smaller output voltage harmonics, and therefore have great development potential in the field of medium and high voltage and high power inverters. However, complex working environments cause inverters to fail during operation, and open-circuit fault and sensor fault are the most common faults in inverters. Therefore, rapid and accurate diagnosis and positioning of inverter faults play an important role in improving system stability and reducing economic losses.

[0003] At present, there are many methods for inverter fault diagnosis, mainly including model-based methods, knowledge-based methods and signal-based methods. For example, a fault diagnosis method based on voltage intrinsic mode function energy feature extraction is proposed in "Open-Circuit Diagnosis Method for T-Type Three-Level Inverter Based on Extracted IMF Energy Feature", which realizes the diagnosis of single-switch and double-switch open-circuit faults of switching tubes. In "Current Sensor Fault Detection and Identification for PMSM Drives Using Multichannel Global Maximum Pooling CNN", an efficient current sensor fault detection and identification method based on an improved one-dimensional convolutional neural network is proposed, which realizes the diagnosis of sensor faults in permanent magnet synchronous motors. As can be seen from the above, good progress has been made in the diagnosis of open-circuit faults of switching devices and current sensor faults. However, these fault diagnosis methods only consider one kind of fault. This is because switching open-circuit faults and current sensor faults have similar fault characteristics, and both will cause loss or distortion of the system output current. Therefore, when the current sensor fails, the diagnosis method considering only the switching open-circuit fault is prone to false alarm. Similarly, the current sensor fault diagnosis method is prone to interference from open-circuit switching faults.

[0004] To solve the above problems, many scholars have made research in recent years. In "A Simultaneous Diagnosis Method for Power Switch and Current Sensor Faults in Grid-Connected Three-Level NPC Inverters", a fault diagnosis method based on interval SMO is proposed to realize the simultaneous diagnosis of open-circuit faults of grid-connected inverter power switches and current sensor faults. This method designs a new type of interval sliding mode observer, which realizes fast and accurate tracking and estimation of three-phase current. Secondly, a fault phase detection scheme based on adaptive threshold is proposed to obtain sensitive and robust detection results. In addition, whether the sum of three-phase currents is zero is verified to distinguish between power switch open-circuit fault and sensor fault. However, the method can only diagnose single switch tube fault and does not consider the case of double tube fault. In "A Reduced-Order Observer-Based Method for Simultaneous Diagnosis of Open-Switch and Current Sensor Faults of a Grid-Tied NPC Inverter", a reduced-order observer-based simultaneous diagnosis strategy for open-switch and current sensor faults of grid-connected neutral point clamped inverter is proposed. This strategy realizes accurate estimation of phase current and current sensor fault by designing a reduced-order observer. Finally, an adaptive threshold fault diagnosis algorithm is proposed using the estimated results. Although the method can accurately locate and diagnose single IGBT fault and current sensor fault, it still does not consider the case of IGBT double tube fault.

[0005] Although there are many methods for simultaneous diagnosis of NPC three-level inverter IGBT open-circuit fault and current sensor fault, the methods for T-type three-level inverter IGBT and current sensor fault diagnosis are still focused on separate research. Therefore, in view of the above problems, it is urgent to propose a method for T-type three-level inverter that can simultaneously diagnose and locate IGBT open-circuit fault and current sensor fault. SUMMARY

[0006] The purpose of the present application is to solve the problems existing in the prior art, provide a T-type three-level inverter IGBT open-circuit fault and current sensor fault diagnosis method, which can quickly and accurately diagnose the fault type and fault position of IGBT open-circuit fault and current sensor fault at the same time, and has high diagnosis accuracy.

[0007] To achieve the above purpose, the present application adopts the following technical scheme: a T-type three-level inverter IGBT open-circuit fault and current sensor fault diagnosis method, comprising the following steps:

[0008] S1, classify the fault types of T-type three-level inverter IGBT open-circuit fault and current sensor fault respectively, and mark them with different digital tags, wherein IGBT open-circuit fault is divided into 78 kinds, and sensor fault is divided into 9 kinds;

[0009] S2, collect three-phase current data sets under different faults as fault samples, process the fault samples for abnormal values and repeated values, then select the same number of partial samples from each fault type to add Gaussian white noise pollution for pretreatment simulation of real working environment, and shuffle the data of all fault samples and divide them into training set and test set according to the ratio of 7:3;

[0010] S3, construct a TCN-LSTM-SelfAttention fault diagnosis model, including a TCN part for extracting local features and an LSTM-SelfAattention for extracting global connections and performing fault diagnosis, wherein the TCN part is connected by three layers of residual blocks, and each residual block includes an inflation causal convolution layer of the main path, a normalization layer, an activation function, a regularization layer for preventing overfitting, and a convolution layer for cross-layer connection;

[0011] The output of the residual block is represented as: In the formula, wherein, y represents the output of the residual module, Activation represents the ReLU nonlinear activation function, F(x) represents the output through the main path, and x represents the input data;

[0012] The Relu activation function is represented as:

[0013] The inflation causal convolution operation is represented as: In the formula, * represents convolution operation, f:{0,...,k-1} represents convolution kernel, k represents convolution kernel length, d represents hole parameter, i is network layer number, s is time step number, x={x0,x1,...,x n} represents input sequence;

[0014] The normalization layer is represented as: In the formula, is denoted as a batch normalization process, γ, β are training parameters of the batch normalization, a l is denoted as the output of the l-th layer after the batch normalization, and f is an activation function.

[0015] The LSTM layer is unfolded along the time dimension:

[0016] h (t) =f(h (t-1) ,X (t) ;w)=…=g (t) (X (t) ,X (t-1) ,X (t-2) ,...,X (2) X (1) ),

[0017] wherein X (t) denotes the signal data collected at time t of the LSTM network input, h (t) denotes the fault feature extracted at time t of the network output, w denotes the network parameter, g (t) denotes the output of the entire network at time t.

[0018] The gating mechanism of each LSTM unit body structure is composed of an input gate, a forget gate and an output gate, and the calculation formula of the LSTM model is as follows:

[0019] f t =σ(W f ·[x t ,h t-1 ]+b f ),

[0020] i t =σ(W i ·[x t ,h t-1 ]+b i ),

[0021] O ι =σ(W0·[x ι ,h ι-1 ]+b0),

[0022] C t =f t ⊙C t-1 +i t ⊙tanh(W C ·[x t ,h t-1 ]+b c ),

[0023] h t =O ttanh(C t ),

[0024] where f t denotes the forget gate, i t denotes the input gate, O t denotes the output gate; W f , W i , W o , b f , b i , b o represent the weights and biases of the forget gate, input gate and output gate respectively, W c , b c denote the weights and biases of the memory cell; C (t) , h (t) are updated from the cell state C (t-1) and hidden state h (t-1) at time t-1; ⊙ denotes element-wise multiplication, and σ and tanh represent sigmoid and tanh activation functions respectively;

[0025] S4, set the initial model parameters according to the TCN-LSTM-SelfAttention fault diagnosis model constructed in step S3, the parameters including optimization algorithm, maximum iteration number MaxEpochs, minimum sample size MiniBatchSize and initial learning rate InitialLearnRate, then input the training set fault data and corresponding fault labels into the TCN-LSTM-SelfAttention fault diagnosis model for training, and adjust the model parameters according to the training result, obtain the best network parameters of the single training sample number, maximum iteration number and initial learning rate of the model, and save the best training parameter model for subsequent test set verification model accuracy;

[0026] S5, input the fault data of all training sets and test sets into the model trained in step S4, and output the fault type through the fault diagnosis of the model on the training set and test set samples, and compare the model diagnosis category with the actual fault category to judge the accuracy of the model.

[0027] In step S1, the method for marking the fault type and label of the open circuit fault of the T-type three-level inverter IGBT is as follows: the switch tube fault is divided into single tube and double tube faults, and the fault types include:

[0028]

[0029] The fault type of the current sensor fault is divided into sticking fault, gain fault and open fault, and the definitions of the corresponding current sensor faults are as follows:

[0030]

[0031] In the formula, i(t) is the output current, i n (t) is the output current in normal operation state, C1 is a constant, C2 is a constant less than 0.8 or greater than 1.2, and t1 is the time of fault occurrence;

[0032] The 78 kinds of IGBT open circuit faults and 9 kinds of sensor faults are numbered as follows:

[0033]

[0034] In the step S2, the expression of the Gaussian white noise is: In the formula, z represents the gray value, represents the average value of z, and σ represents the standard deviation of z, σ 2 represents the variance of z.

[0035] In the step S3, the calculation formula of the self-attention mechanism SelfAattention is as follows:

[0036]

[0037] In the formula, Q, K, and V represent the query vector (Query), the key vector (Key), and the value vector (Value), respectively, d k represents the dimension of the key vector, and softmax represents the normalization of the relevant features to obtain the weight coefficient;

[0038] Q = W Q X, K = W K X, V = W V X,

[0039] In the formula, X represents the input data, W Q represents the learnable matrix for the query vector Q, W K represents the learnable matrix for the key vector K, W V represents the learnable matrix for the value vector V, and W V represents the parameter matrix for output learning.

[0040] In the step S3, the method for classifying faults in the model LSTM-SelfAattention is to use the Softmax layer and the classification layer classificationLayer to classify faults;

[0041] In the Softmax layer, the Softmax activation function is used to convert the fault categories into a probability distribution between 0 and 1, and the expression is as follows:

[0042]

[0043] Where x=(x1,x2,...,x n ) represents the input vector, x i represents the original score of the i-th category, Represents input x i exponential function of ; Represents the sum of the exponentials of all categories, used to normalize the output;

[0044] The classification layer selects the final classification result by the size of the probability, and helps the model to backpropagate and optimize the weights by calculating the cross entropy loss function. The cross entropy loss function is expressed as follows:

[0045]

[0046] In the formula, N represents the number of samples, K is the number of categories, and w i is the weight of category i, t ni Indicates that if the true category of the nth sample is equal to i, it takes 1, otherwise it takes 0; y ni represents the predicted probability that the nth sample belongs to category i.

[0047] In step S4, the accuracy evaluation formula is:

[0048]

[0049] Where, X i represents the true category of the fault, Y i Represents the category of model diagnosis, M represents the total number of samples; where (X i = = Y i ) means that if the true category is equal to the diagnosis category, the output is 1, otherwise the output is 0; the Sum function is used to calculate the total number of correct predictions.

[0050] The beneficial effects of the present invention are:

[0051] 1) The diagnostic method of the present invention can simultaneously diagnose the fault type and fault location of IGBT open circuit fault and current sensor fault, and by combining TCN with LSTM, it can simultaneously capture the local information and global connection of fault characteristics. At the same time, the self-attention mechanism is combined with LSTM to improve the performance of the model and improve the accuracy.

[0052] 2) In the diagnostic method of the application, the collected three-phase current data samples are preprocessed by adding noise and the like and input into the established TCN-LSTM-SelfAttention fault diagnosis model. The model combines TCN and LSTM to extract the local features and global relationship of the fault signal, and has strong anti-noise performance. Among them, the TCN time convolutional neural network can perform efficient parallel operation, ensure that the input and output lengths are the same, and will not leak future data in advance for local feature extraction of fault data; the LSTM model can effectively capture long-time dependence for global relationship extraction of fault data; finally, the self-attention mechanism can calculate the relationship between each part of the input sequence, assign appropriate weights to different features or time steps, dynamically focus on the most important information, and make the model have stronger expression ability; the diagnostic method has good diagnostic results for the open circuit fault of the T-type three-level inverter IGBT, and the sticking fault, gain fault and open fault of the current sensor, which is better than the existing diagnostic method.

[0053] 3) In the diagnostic method of the application, two faults are diagnosed by one method, which can avoid misdiagnosis. In addition, since the analysis, calculation and code of the open circuit and current sensor fault diagnosis method can be partially shared, it is much simpler to handle the two faults in one algorithm than in two different algorithms. On the other hand, one method solves two faults, which helps to locate the specific fault equipment in time and facilitates subsequent maintenance, thereby improving the maintenance efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The flowchart of the diagnostic method of the application;

[0055] Figure 2 The structure topology diagram of the T-type three-level inverter used in the diagnostic method of the application;

[0056] Figure 3 The structure diagram of the TCN residual block constructed in the diagnostic method of the application;

[0057] Figure 4 The fault diagnosis model TCN-LSTM-SelfAttention constructed in the diagnostic method of the application is shown in the schematic diagram;

[0058] Figure 5 The three-phase current diagram when the T1 switch tube open circuit fault and the sensor A gain fault occur in the diagnostic method of the application;

[0059] Figure 6 The result comparison diagram of the training set and the test set for experimental verification in the diagnostic method of the application. DETAILED DESCRIPTION

[0060] The application will be further explained in connection with the accompanying drawings and specific embodiments.

[0061] Embodiment: As shown in the flow chart of the diagnostic method, the application provides a diagnostic method for T-type three-level inverter IGBT open circuit fault and current sensor fault, comprising the following steps: Figure 1

[0062] S1, classify the fault types of T-type three-level inverter IGBT open circuit fault and current sensor fault respectively, and mark them with different digital labels, wherein the IGBT open circuit fault is divided into 78 types, and the sensor fault is divided into 9 types.

[0063] The structure topology of the T-type three-level inverter is shown in Figure 2 Compared with the two-level topology, the T-type three-level topology only adds one clamping bridge arm, realizing the conversion from two-level to three-level. Each phase of the three-level inverter has four power switching tubes with anti-parallel diodes and one current sensor. The IGBT switching tube fault can be divided into the following Table 1 according to different bridge arms and different phase faults.

[0064] Table 1 is the fault type of IGBT switching tube

[0065]

[0066] In addition to the switching tube fault of the inverter, the sensor is also prone to failure. The current sensor fault may cause the closed-loop system to lose control. Among them, the fault types of three common current sensor faults are divided into stuck fault, gain fault and open fault, and the definitions of the corresponding current sensor faults are as follows:

[0067]

[0068] In the formula, i(t) is the output current, i n (t) is the output current in normal operation state, C1 represents a constant, C2 is a constant less than 0.8 or greater than 1.2, t1 is the time of fault occurrence;

[0069] In order to accurately diagnose and locate different types of faults, different digital labels are used to number 78 types of IGBT open circuit faults and 9 types of sensor faults, as shown in Table 2 below:

[0070] Table 2 is the fault label of the inverter and the sensor

[0071]

[0072] ​S2, collect three-phase current data sets under different faults as fault samples, process the abnormal values and repeated values of the fault samples, and then select ten samples from each fault type to add Gaussian white noise pollution for pretreatment simulation of real working environment, and shuffle the data of all fault samples and divide them into training set and test set according to the ratio of 7:3. Because A, B and C three-phase are symmetrical, the fault samples are collected and analyzed by taking A phase as an example. The collection frequency is set to 50 kHz, and there are 1000 collection points every 0.02 seconds.

[0073] The expression of Gaussian white noise is: In the formula, z represents the gray value, represents the average value of z, and σ 2 represents the variance of z.

[0074] S3, construct a TCN-LSTM-SelfAttention fault diagnosis model, including a TCN part for extracting local features and an LSTM-SelfAattention for extracting global connections and performing fault diagnosis, wherein the TCN part is connected by three residual blocks, each residual block includes an expansion causal convolution layer of the main path, a normalization layer, an activation function, a regularization layer to prevent overfitting, and a convolution layer for cross-layer connection.

[0075] The collected current data is time series fault data, and the combination of TCN and LSTM can extract local features and global connections of the data, and the combination of the two can improve the anti-interference ability of the model to noise. When facing longer time series data, LSTM still has some problems, so adding a self-attention mechanism layer to LSTM makes the model output can dynamically focus on important features to capture and enhance the expression ability of the model.

[0076] The addition of causal convolution in TCN can ensure that fault information does not leak from the future to the past, and the use of dilated convolution can make TCN have a larger receptive field with fewer layers, that is, receive longer time series. Specifically, for the input x of a certain layer, the residual connection adds it to the output F(x) of the layer, and the sum of the two is taken as the input of the next layer. The introduction of residual connection can make the network stable even in deep depth, making the network easier to train and faster feedback and convergence.

[0077] The output of the residual block is represented as: In the formula, represents the output of the residual module, Activation represents the ReLU nonlinear activation function, F(x) represents the output through the main path, and x represents the input data.

[0078] The Relu activation function is represented as:

[0079] The dilated causal convolution operation is expressed as: where * represents convolution operation, f:{0,...,k-1} represents the convolution kernel, k represents the length of the convolution kernel, d represents the dilation parameter, i represents the network layer, s represents the time step, x={x0,x1,...,x n} represents the input sequence.

[0080] The normalization layer is expressed as: where is expressed as a batch normalization process, γ, β are the training parameters of the batch normalization, a l is expressed as the output of the l-th layer after batch normalization after the output of the activation function f.

[0081] The LSTM layer is unfolded according to the time dimension:

[0082] h (t) =f(h (t-1) ,X (t) ;w)=…=g (t) (X (t) ,X (t-1) ,X (t-2) ,...,X (2) X (1) ),

[0083] where X (t) represents the signal data collected at time t of the LSTM network input, h (t) represents the fault feature extracted at time t of the network output, w represents the network parameter, g (t) represents the output of the entire network at time t.

[0084] The gating mechanism of each LSTM unit body structure is composed of an input gate, a forget gate and an output gate, and the calculation formula of the LSTM model is as follows:

[0085] f t =σ(W f ·[x t ,h t-1 ]+b f ),

[0086] i t =σ(W i ·[x t ,h t-1 ]+b i ),

[0087] O ι =σ(W0·[x ι ,h ι-1 ]+b0),

[0088] C t = f t ⊙C t-1 +i t ⊙tanh(W C ·[x t ,h t-1 ]+b c ),

[0089] h t = O t ⊙tanh(C t ),

[0090] In the formula, f t represents the forget gate, i t represents the input gate, and O t represents the output gate; W f , W i , W o , b f , b i , b o represent the weights and biases of the forget gate, the input gate, and the output gate, respectively; W c , b c represent the weights and biases of the storage unit; C (t) , h (t) are updated from the cell state C (t-1) and the hidden state h (t-1) at time t-1; ⊙ represents element-wise multiplication, and σ and tanh represent sigmoid and tanh activation functions, respectively.

[0091] An input sequence X1, X2, X3, …, X t is defined, and linear layers are used to obtain Q, K, and V of the sequence data, where Q is a query vector, K is a key value of information, and V is a matching information amount, then the similarity between Q and K is calculated and normalized to obtain an importance weight matrix between information, and finally the weight matrix is multiplied by V to obtain feature information of different importance.

[0092] The calculation formula of the self-attention mechanism SelfAattention is as follows:

[0093]

[0094] In the formula, Q, K, and V represent the query vector (Query), the key vector (Key), and the value vector (Value), respectively, d k represents the dimension of the key vector, and softmax represents the normalization of the relevant features to obtain the weight coefficient;

[0095] Q = W Q X, K = WK X, V = W V X,

[0096] In the formula, X represents input data, W Q represents a learnable matrix for query vector Q, W K represents a learnable matrix for key vector K, W V represents a learnable matrix for value vector V, W V represents a learnable parameter matrix for output.

[0097] The method for classifying faults in the LSTM-SelfAttention model is: using the Softmax layer and the classification layer classificationLayer to classify faults;

[0098] In the Softmax layer, the Softmax activation function is used to convert the fault category into a probability distribution between 0 and 1, and the expression is as follows:

[0099]

[0100] In the formula, x = (x1, x2,..., x n ) represents an input vector, x i represents the original score of the i-th category, represents the exponential function of input x i ; represents the sum of the exponentials of all categories, used for normalizing the output;

[0101] The classification layer classificationLayer selects the final result of classification by the size of the probability, and helps the model to optimize the weight through the calculation of the cross-entropy loss function. The cross-entropy loss function is expressed as follows:

[0102]

[0103] In the formula, N represents the number of samples, K is the number of categories, w i is the weight of category i, t ni represents 1 if the true category of the n-th sample is equal to i, otherwise 0; y ni represents the predicted probability that the n-th sample belongs to category i.

[0104] S4, set initial model parameters according to the TCN-LSTM-SelfAttention fault diagnosis model constructed in step S3, the parameters include optimization algorithm, maximum iteration number MaxEpochs, minimum sample size MiniBatchSize and initial learning rate InitialLearnRate, then input the training set fault data and corresponding fault labels into the TCN-LSTM-SelfAttention fault diagnosis model for training, and adjust the model parameters according to the training result, obtain the best network parameters of the single training sample number, the maximum iteration number and the initial learning rate of the model, and save the best training parameter model for subsequent test set verification model accuracy;

[0105] Among them, there are 30 kinds of fault types related to A, including IGBT single tube open circuit fault, IGBT double tube open circuit fault and current sensor fault, each fault type has 60 samples, including 50 original fault samples and 10 fault samples with added noise interference, a total of 1800 samples.

[0106] S5, input the fault data of all training sets and test sets into the model trained in step S4, and output the fault type through the fault diagnosis of the model to the training set and test set samples, and compare the model diagnosis category with the actual fault category to judge the accuracy of the model.

[0107] The accuracy evaluation formula is:

[0108]

[0109] In the formula, X i represents the true category of the fault, Y i represents the category diagnosed by the model, and M represents the total number of samples; wherein (X i ==Y i ) represents that if the true category is equal to the diagnosed category, 1 is output, otherwise 0 is output; the total number of correct predictions is calculated by using the Sum function;

[0110] The root mean square error formula is: In the formula, N represents the number of fault samples, y i and represent the i-th true value and the i-th predicted value.

[0111] The present application uses matlab software to realize the programming of the fault diagnosis model, and the program design flow chart of the model is as shown in Figure 1 The TCN-LSTM-SelfAttention fault diagnosis model is as shown in Figure 4The diagnostic model includes a stacked 3-layer residual module, an LSTM layer and a self-attention layer; the connection of the 3-layer residual block adopts a Relu activation function which can reduce the amount of calculation, and a Softmax layer and a classification layer classificationLayer are used at the end of the diagnostic model to realize classification of the fault.

[0112] As shown in the three-level inverter structure, the circuit model is built by using Simulink in Matlab, and the fault samples are collected, the method is verified, 1000 sampling points per 0.02 seconds are collected as a sample, 60 samples of each fault state are collected, a total of 1800 samples, and 10 samples of each fault type are selected to add Gaussian noise to simulate the noise pollution in the actual working environment, then all the samples are shuffled, and the training set and the test set are divided according to the ratio of 7:3. Figure 2 As shown in the three-level inverter structure, the circuit model is built by using Simulink in Matlab, and the fault samples are collected, the method is verified, 1000 sampling points per 0.02 seconds are collected as a sample, 60 samples of each fault state are collected, a total of 1800 samples, and 10 samples of each fault type are selected to add Gaussian noise to simulate the noise pollution in the actual working environment, then all the samples are shuffled, and the training set and the test set are divided according to the ratio of 7:3.

[0113] Figure 3 As shown in the three-level inverter structure, the circuit model is built by using Simulink in Matlab, and the fault samples are collected, the method is verified, 1000 sampling points per 0.02 seconds are collected as a sample, 60 samples of each fault state are collected, a total of 1800 samples, and 10 samples of each fault type are selected to add Gaussian noise to simulate the noise pollution in the actual working environment, then all the samples are shuffled, and the training set and the test set are divided according to the ratio of 7:3.

[0114] In order to evaluate the performance and generalization ability of the model, the training set and the test set used for model training are respectively input into the established fault diagnosis model, and the accuracy of the model for known faults and unknown faults is verified. As shown in Figure 6 As shown in the three-level inverter structure, the circuit model is built by using Simulink in Matlab, and the fault samples are collected, the method is verified, 1000 sampling points per 0.02 seconds are collected as a sample, 60 samples of each fault state are collected, a total of 1800 samples, and 10 samples of each fault type are selected to add Gaussian noise to simulate the noise pollution in the actual working environment, then all the samples are shuffled, and the training set and the test set are divided according to the ratio of 7:3.

[0115] The diagnostic method of the application can diagnose the fault type and fault position of IGBT open circuit fault and current sensor fault at the same time, and by selecting the combination of TCN and LSTM, the local information and global connection of the fault features can be captured at the same time, and the self-attention mechanism is combined in the LSTM to improve the performance of the model and improve the accuracy.

[0116] The above description is only used to illustrate the technical solutions of the application, not to limit, and other modifications or equivalent replacements of the technical solutions of the application made by those skilled in the art should be covered in the scope of the claims of the application.​

Claims

1. A method for diagnosing IGBT open-circuit faults and current sensor faults in a T-type three-level inverter, characterized by: The following steps are involved: S1. Classify the fault types of IGBT open-circuit faults and current sensor faults of T-type three-level inverters respectively and mark them with different digital labels. IGBT open-circuit faults are divided into 78 types and sensor faults are divided into 9 types. S2. Collect three-phase current data sets under different fault conditions as fault samples. After processing the fault samples for outliers and duplicates, select the same number of samples from each fault type and add Gaussian white noise pollution for preprocessing to simulate a real working environment. Then, shuffle the data of all fault samples and divide them into training and test sets in a ratio of 7:

3. S3. Build a TCN-LSTM-SelfAttention fault diagnosis model, which includes a TCN part for extracting local features and an LSTM-SelfAttention part for extracting global connections and performing fault diagnosis. The TCN part is connected by three layers of residual blocks. Each residual block includes a dilated causal convolution layer in the main path, a normalization layer, an activation function, a regularization layer to prevent overfitting, and a convolution layer for cross-layer connections. The output of the residual block is expressed as: Where, Represents the output of the residual module, Activation represents the ReLU nonlinear activation function, F(x) represents the output of the main path, and x represents the input data; The Relu activation function is expressed as: The dilated causal convolution operation is expressed as: Where * represents the convolution operation, f:{0,...,k-1} represents the convolution kernel, k represents the convolution kernel length, d represents the hole parameter, i represents the number of network layers, s represents the number of time steps, and x = {x0,x1,...,x n } represents the input sequence; The normalization layer is expressed as: Where, It is expressed as a batch normalization process, γ and β are the training parameters of batch normalization, and a l It is represented as the output of the batch normalized output of the lth layer after the activation function f; The LSTM layer expands it along the time dimension: h (t) =f(h (t-1) ,X (t) ;w)=…=g (t) (X (t) ,X (t-1) ,X (t-2) ,...,X (2) X (1) ), Where, X (t) represents the signal data collected at time t input to the LSTM network, h (t) represents the fault feature extracted at time t of the network output, w represents the network parameters, g (t) represents the output of the entire network at time t; The gating mechanism of each LSTM unit structure consists of an input gate, a forget gate, and an output gate. The calculation formula of the LSTM model is as follows: f t =σ(W f ·[x t ,h t-1 ]+b f ), i t =σ(W i ·[x t ,h t-1 ]+b i ), ABOUT ι =σ(W0 [x ι ,h ι-1 ]+b0), C t =f t ⊙C t-1 +i t ⊙tanh(W C ·[x t ,h t-1 ]+b c ), h t =O t ⊙tanh(C t ), Where, f t represents the forget gate, i t represents the input gate, O t Represents the output gate; W f 、W i 、W o , b f 、b i 、b o Represent the weights and biases of the forget gate, input gate, and output gate, respectively. c 、b c Represents the weight and bias of the storage unit; C (t) 、h (t) The unit state C at time t-1 (t-1) and hidden state h (t-1) Updated; ⊙ represents element-by-element multiplication, σ and tanh represent sigmoid and tanh activation functions respectively; S4. Set the initial model parameters according to the TCN-LSTM-SelfAttention fault diagnosis model constructed in step S3. The parameters include the optimization algorithm, the maximum number of iterations MaxEpochs, the minimum number of samples MiniBatchSize, and the initial learning rate InitialLearnRate. Then, use the training set fault data and the corresponding fault labels to input the TCN-LSTM-SelfAttention fault diagnosis model for training, and adjust the model parameters according to the training results to obtain the optimal network parameters of the model for the number of single training samples, the maximum number of iterations, and the initial learning rate. Save the optimal training parameter model for subsequent test sets to verify the accuracy of the model. S5. Input the fault data of all training sets and test sets into the model trained in step S4 respectively, diagnose the faults of the training set and test set samples through the model, output the fault type, and judge the accuracy of the model by comparing the category diagnosed by the model with the actual fault category.

2. The method for diagnosing IGBT open circuit fault and current sensor fault in a T-type three-level inverter according to claim 1, characterized in that: In step S1, the method for marking the fault type and label of the T-type three-level inverter IGBT open circuit fault is as follows: switch tube faults are divided into single tube faults and double tube faults, and the fault types include: The fault types of current sensor faults are divided into stuck faults, gain faults and disconnection faults. The corresponding current sensor fault definitions are as follows: Where i(t) is the output current, i n (t) is the output current under normal operating conditions, C1 is a constant, C2 is a constant less than 0.8 or greater than 1.2, and t1 is the time when the fault occurs; The 78 IGBT open circuit faults and 9 sensor faults are numbered as follows:

3. The method for diagnosing IGBT open circuit fault and current sensor fault in a T-type three-level inverter according to claim 1, characterized in that: In step S2, the expression of Gaussian white noise is: Where z represents the grayscale value, represents the mean value of z, σ represents the standard deviation of z, σ 2 represents the variance of z.

4. The method for diagnosing IGBT open circuit fault and current sensor fault in a T-type three-level inverter according to claim 1, characterized in that: In step S3, the calculation formula of the self-attention mechanism SelfAattention is as follows: In the formula, Q, K, and V represent the query vector (Query), key vector (Key), and value vector (Value), respectively. k Represents the dimension of the key vector, and softmax represents the weight coefficient obtained by normalizing the relevant features; Q=W Q X,K=W K X,V=W V X, Where X represents the input data, W Q represents the learnable matrix for query vector Q, W K represents the learnable matrix for the key vector K, W V represents the learnable matrix for the value vector V, W V Represents the parameter matrix used to output the learnable 5. The method for diagnosing IGBT open circuit fault and current sensor fault in a T-type three-level inverter according to claim 1, characterized in that: In step S3, the method of LSTM-SelfAattention in the model to classify faults is: using the Softmax layer and the classification layer classificationLayer to classify the faults; The Softmax activation function is used in the Softmax layer to convert the fault type into a probability distribution between 0 and 1. The expression is as follows: Where x=(x1,x2,...,x n ) represents the input vector, xi represents the original score of the i-th category, represents the exponential function of the input xi; Represents the sum of the exponentials of all categories, used to normalize the output; The classification layer selects the final classification result by the size of the probability, and helps the model to backpropagate and optimize the weights by calculating the cross entropy loss function. The cross entropy loss function is expressed as follows: In the formula, N represents the number of samples, K is the number of categories, and w i is the weight of category i, t ni Indicates that if the true category of the nth sample is equal to i, it takes 1, otherwise it takes 0; y ni represents the predicted probability that the nth sample belongs to category i.

6. The method for diagnosing IGBT open circuit fault and current sensor fault in a T-type three-level inverter according to claim 1, characterized in that: In step S4, the accuracy evaluation formula is: Where, X i represents the true category of the fault, Y i Represents the category of model diagnosis, M represents the total number of samples; where (X i = = Y i ) means that if the true category is equal to the diagnosis category, the output is 1, otherwise the output is 0; the Sum function is used to calculate the total number of correct predictions.

Images