IGBT junction temperature online monitoring method based on GAF-CNN

Abstract

The invention discloses an IGBT junction temperature online monitoring method based on a GAF-CNN, and the method is characterized in that the method comprises the following steps: collecting the grid undershoot data of a to-be-detected IGBT; inputting grid undershoot data of an IGBT to be detected into the trained IGBT junction temperature on-line monitoring model for processing to obtain real-time junction temperature; the structure of the IGBT junction temperature on-line monitoring model is a GAF-CNN neural network model. According to the GAF-CNN-based IGBT junction temperature on-line monitoring method, two-dimensional image conversion is carried out on a one-dimensional grid undershoot waveform time sequence by using a GAF, and one-dimensional data and a two-dimensional image are simultaneously used as input to construct a hybrid CNN neural network model to carry out junction temperature prediction on an IGBT. The problems that an existing temperature-sensitive electrical parameter method is affected by aging and dependence of load current is large are solved, and more accurate junction temperature prediction is carried out.

Description

Technical Field

[0001] The present invention belongs to the technical field of semiconductor device junction temperature detection, and relates to an IGBT junction temperature online monitoring method based on GAF-CNN. Background Art

[0002] IGBTs (Insulated Gate Bipolar Transistors), core components of modern high-power power electronics systems, are widely used in renewable energy generation, electric vehicles, industrial inverters, and smart grids. However, under high-speed switching and high-power density conditions, IGBT chips are subjected to long-term dynamic load fluctuations, resulting in drastic fluctuations in junction temperature. The thermal stress caused by junction temperature fluctuations accelerates power device degradation, seriously threatening the lifespan and reliability of IGBT chips. Therefore, accurate online junction temperature prediction is crucial for reliability assessment and health management of IGBT devices.

[0003] The temperature-sensitive electrical parameter method extracts the device's own electrical parameters to invert junction temperature, enabling online monitoring of junction temperature and offering significant advantages over other junction temperature detection methods. However, traditional temperature-sensitive electrical parameters such as saturation voltage drop and threshold voltage drift with IGBT module aging, requiring additional aging compensation. Furthermore, most junction temperature monitoring methods based on the temperature-sensitive electrical parameter method rely heavily on accurate load current measurement, significantly impacting junction temperature monitoring accuracy.

[0004] In summary, the existing technology has the problem that the IGBT temperature-sensitive electrical parameters shift as the module ages and the junction temperature prediction accuracy is not high. Summary of the Invention

[0005] The purpose of the present invention is to provide an IGBT junction temperature online monitoring method based on GAF (Gramian Angular Field)-CNN (Convolutional Neural Network), which solves the problems existing in the prior art that the IGBT temperature-sensitive electrical parameters shift with module aging and the junction temperature prediction accuracy is not high.

[0006] The technical solution adopted by the present invention is an IGBT junction temperature online monitoring method based on GAF-CNN, comprising the following steps: Step 1: Collect gate undershoot data of the IGBT to be tested; Step 2: Input the gate undershoot data of the IGBT to be tested into the trained IGBT junction temperature online monitoring model to obtain the real-time junction temperature; The structure of the IGBT junction temperature online monitoring model is a GAF-CNN neural network model.

[0007] The present invention is also characterized in that: The trained IGBT junction temperature online monitoring model is obtained through the following steps: Step A1: collecting gate undershoot data at different junction temperatures to establish a sample data set; Step A2: Use GAF to convert the sample dataset into a three-dimensional dataset; Step A3: Use the sample data set and the three-dimensional data set to train and test the IGBT junction temperature online monitoring model. If the test results meet the set conditions, a trained IGBT junction temperature online monitoring model is obtained. If the test results do not meet the set conditions, repeat step A3.

[0008] Step A1 includes: The IGBT is heated by a heating module. When the upper tube of the IGBT is turned off, the gate undershoot waveform of the lower tube is collected at a sampling frequency of 100 MHz through the acquisition circuit and the analog-to-digital converter. The junction temperature of the IGBT module is recorded and summarized using an infrared temperature gun to obtain a one-dimensional sample data set. The sample data set is divided into a one-dimensional data training set and a one-dimensional data test set according to a fixed ratio.

[0009] Step A2 includes: Step A2.1, representing the collected gate undershoot data as a one-dimensional array X; Step A2.2, normalize the one-dimensional array X; Step A2.3, performing polar coordinate conversion on the normalized one-dimensional array X; Step A2.4: Convert the cosine values of the angle sums in the one-dimensional array X after polar coordinate conversion into a first characteristic matrix, namely, a GASF (Gramian Angular Summation Field) matrix, and convert the sine values of the angle differences into a second characteristic matrix, namely, a GADF (Gramian Angular Difference Field) matrix; Step A2.5: Convert the first feature matrix and the second feature matrix into grayscale images respectively and map them to the range of [0, 255]. Superimpose the layers by adding one layer of the first feature matrix and one layer of the second feature matrix to obtain a three-dimensional image. Summarize the three-dimensional images to obtain a three-dimensional data set and divide it into a three-dimensional data training set and a three-dimensional data test set.

[0010] One-dimensional array X={ x 1, x 2, x 3,…, x i},in xi Indicates that the gate undershoot is 10* after the IGBT module is turned off. i Voltage value in nanoseconds; After polar coordinate conversion, the variables angle and radius are used to represent the value of the time series and its corresponding timestamp. The calculation formula for polar coordinate conversion is as follows: , in, Represents the first i Gate undershoot data, Indicates the n variable angles, Represents the total number of gate undershoot signals in the one-dimensional array X.

[0011] Step A2.2 includes: Normalizing a one-dimensional array X can be achieved using the following formula: , in, represents the gate undershoot data after normalization, Indicates that the gate undershoot is 10* after the IGBT module is turned off. i The voltage value at nanoseconds, Represented as the maximum value in the one-dimensional array X, Represents the minimum value in the one-dimensional array X.

[0012] Step A2.2 includes: Normalizing a one-dimensional array X can be achieved using the following formula: , in, represents the gate undershoot data after normalization, Indicates that the gate undershoot is 10* after the IGBT module is turned off. i The voltage value at nanoseconds, Represented as the maximum value in the one-dimensional array X, Represents the minimum value in the one-dimensional array X.

[0013] The expression of GASF matrix is:

[0014] in, represents the GASF matrix, Indicates the n variable angles; The expression of the GADF matrix is:

[0015] in, represents the GADF matrix, Indicates the n variable angles.

[0016] Step A3 includes: Step A3.1: Use the one-dimensional data training set as input to train the one-dimensional convolutional neural network model CNN1: perform calculations on the convolution layer, pooling layer, activation function, and fully connected layer to obtain the first eigenvector of the hybrid CNN neural network model; Step A3.2: Using the 3D data training set as input, train the 2D convolutional neural network model CNN2 by performing calculations on the convolutional layer, pooling layer, activation function, and fully connected layer to obtain the second eigenvector of the hybrid CNN neural network model. Step A3.3: Use the feature fusion layer to fuse the first and second feature vectors, input the fused feature vectors into the hybrid input model CNN3 for training, and output the junction temperature prediction value; Step A3.4: Input the junction temperature prediction value into the initial GAF-CNN neural network model for training. After calculating the model prediction value through forward propagation, adjust the model hyperparameters or network structure. Step A3.5: Input the test set into the GAF-CNN neural network model with adjusted model hyperparameters or network structure to calculate the model error. If the model error meets the set training accuracy, a trained IGBT junction temperature online monitoring model is obtained; if the model error does not meet the set training accuracy, repeat steps A3.1-A3.5.

[0017] Model hyperparameters include learning rate, batch size, number of iterations, optimizer type, convolution kernel size, and number of channels; The model error is judged using MSE, and the calculation formula of MSE is as follows: , in, represents the model error judgment parameter, Indicates the j The actual junction temperature corresponding to the gate undershoot data is Indicates the j The junction temperature prediction value is obtained by processing the gate undershoot data through the GAF-CNN neural network model. Indicates the total number of gate undershoot signals in this group.

[0018] The beneficial effects of the present invention are as follows: the present invention adopts a new type of temperature-sensitive electrical parameter, namely the gate undershoot signal, which is not affected by the aging of the IGBT module bonding wire. The parameter properties will not change due to the aging of the module, and can effectively maintain the accuracy of junction temperature monitoring throughout the service life of the module; compared with traditional physical measurement methods and thermal model methods, the present method only needs to collect the gate undershoot parameter to perform junction temperature monitoring, which effectively reduces the complexity of hardware and experimental equipment; the present invention uses GAF to convert the one-dimensional gate undershoot waveform into a two-dimensional image, and retains the global dependency and timing structure of the time series through polar coordinate encoding, while CNN can extract local features from the generated image and capture subtle changes in the time series. The combination of the two simultaneously utilizes global timing information and local dynamic features, can effectively extract the features of the gate undershoot waveform locally and globally, and effectively solve the load current dependence problem of the temperature-sensitive electrical parameter; at the same time, CNN can effectively model the complex nonlinear dynamics in the time series through multi-layer convolution and nonlinear activation functions; through the above two points, the GAF-CNN neural network can effectively improve the prediction accuracy of the junction temperature model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the IGBT junction temperature online monitoring method based on GAF-CNN of the present invention; Figure 2 It is a structural schematic diagram of the IGBT junction temperature online monitoring model in the present invention. DETAILED DESCRIPTION

[0020] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] IGBT junction temperature online monitoring method based on GAF-CNN, such as Figure 1 As shown, the following steps are included: Step 1: Collect gate undershoot data of the IGBT to be tested; Step 2: Input the gate undershoot data of the IGBT to be tested into the trained IGBT junction temperature online monitoring model for processing to obtain the real-time junction temperature.

[0022] The structure of the IGBT junction temperature online monitoring model is a GAF-CNN neural network model; Figure 2 As shown, the GAF-CNN neural network model includes a hybrid CNN neural network model, and the hybrid CNN neural network model includes a feature fusion layer, the output end of the feature fusion layer is connected to the input end of the CNN3 module, and the input end of the feature fusion layer is connected to the output end of the CNN1 module and the output end of the CNN2 module respectively; The CNN1 module includes an input layer, a convolutional layer, an activation layer, a pooling layer, a convolutional layer, an activation layer, a pooling layer, and a fully connected layer connected in sequence. The input of the CNN1 module is one-dimensional gate undershoot data, the input layer input is one-dimensional single-channel data, the convolutional layer uses a 1*3 convolution kernel for 1-D convolution, the activation layer uses the ReLU activation function, the pooling layer uses a 1*2 maximum pooling, and the output of the fully connected layer is the first feature vector. The CNN2 module includes an input layer, a convolutional layer, an activation layer, a pooling layer, a convolutional layer, an activation layer, a pooling layer, and a fully connected layer connected in sequence; the input of the CNN2 module is the GASF image and the GADF image converted from the gate undershoot data, the input of the input layer is two-dimensional dual-channel data, the convolution layer uses a 3*3 convolution kernel for 2-D convolution, the activation layer uses the Relu activation function, the pooling layer uses a 2*2 maximum pooling, and the output of the fully connected layer is the second feature vector.

[0023] The CNN3 module includes an input layer, a convolutional layer, an activation layer, a pooling layer, a convolutional layer, an activation layer, a pooling layer, and a fully connected layer connected in sequence. The input of the CNN3 module is a multimodal fusion vector obtained by concatenating the output vectors of the CNN1 module and the fully connected layer of the CNN2 module after processing by the feature fusion layer. The size of the input vector is the sum of the output vectors of the fully connected layer of the CNN1 module and the CNN2 module. The convolution layer uses a 1*3 convolution kernel for 1-D convolution, the activation layer uses ReLU, and the pooling layer uses 1*2 maximum pooling. Finally, the output of the fully connected layer is the predicted value of the junction temperature.

[0024] The trained IGBT junction temperature online monitoring model is obtained through the following steps: Step A1: collecting gate undershoot data at different junction temperatures to establish a sample data set; The IGBT is heated by a heating module. When the upper tube of the IGBT is turned off, the gate undershoot waveform of the lower tube is collected at a sampling frequency of 100 MHz through the acquisition circuit and the analog-to-digital converter. The junction temperature of the IGBT module is recorded and summarized using an infrared temperature gun to obtain a one-dimensional sample data set. The sample data set is divided into a one-dimensional data training set and a one-dimensional data test set according to a fixed ratio.

[0025] Step A2: Use GAF to convert the sample dataset into a three-dimensional dataset; Step A2.1, representing the collected gate undershoot data as a one-dimensional array X; One-dimensional array X={ x 1, x 2, x 3,…, x i},in x iIndicates that the gate undershoot is 10* after the IGBT module is turned off. i Voltage value in nanoseconds; Step A2.2, normalize the one-dimensional array X; Normalizing the one-dimensional array X can be achieved using the following formula: , in, represents the gate undershoot data after normalization, Indicates that the gate undershoot is 10* after the IGBT module is turned off. i The voltage value at nanoseconds, Represented as the maximum value in the one-dimensional array X, Represented as the minimum value in the one-dimensional array X; Step A2.3, performing polar coordinate conversion on the normalized one-dimensional array X; After polar coordinate conversion, the variables angle and radius are used to represent the value of the time series and its corresponding timestamp. The calculation formula for polar coordinate conversion is as follows: , in, Represents the first i Gate undershoot data, Indicates the n variable angles, Represents the total number of gate undershoot signals in the one-dimensional array X; Step A2.4, convert the cosine value of the sum of the angles in the one-dimensional array X after polar coordinate conversion into a first characteristic matrix, i.e., a GASF matrix, and convert the sine value of the difference of the angles into a second characteristic matrix, i.e., a GADF matrix; The expression of GASF matrix is:

[0026] in, represents the GASF matrix, Indicates the n variable angles; The expression of the GADF matrix is:

[0027] in, represents the GADF matrix, Indicates the n variable angles.

[0028] Step A2.5: Convert the first and second feature matrices into grayscale images and map them to the range [0, 255]. Overlay the images by adding one layer of the first feature matrix to another layer of the second feature matrix to produce a three-dimensional image. The three-dimensional images are aggregated to produce a three-dimensional dataset, which is then divided into a three-dimensional data training set and a three-dimensional data test set. Step A3: Use the sample data set and the three-dimensional data set to train and test the IGBT junction temperature online monitoring model. If the test result meets the set conditions, a trained IGBT junction temperature online monitoring model is obtained. If the test result does not meet the set conditions, repeat step A3. Step A3.1: Use the one-dimensional data training set as input to train the one-dimensional convolutional neural network model CNN1: perform calculations on the convolution layer, pooling layer, activation function, and fully connected layer to obtain the first eigenvector of the hybrid CNN neural network model; Step A3.2: Using the 3D data training set as input, train the 2D convolutional neural network model CNN2 by performing calculations on the convolutional layer, pooling layer, activation function, and fully connected layer to obtain the second eigenvector of the hybrid CNN neural network model. Step A3.3: Use the feature fusion layer to fuse the first and second feature vectors, input the fused feature vectors into the hybrid input model CNN3 for training, and output the junction temperature prediction value; Step A3.4: Input the junction temperature prediction value into the initial GAF-CNN neural network model for training. After calculating the model prediction value through forward propagation, adjust the model hyperparameters or network structure. Step A3.5: Input the test set into the GAF-CNN neural network model after adjusting the model hyperparameters or network structure to calculate the model error. If the model error meets the set training accuracy, a trained IGBT junction temperature online monitoring model is obtained. If the model error does not meet the set training accuracy, repeat steps A3.1-A3.5. Model hyperparameters include learning rate, batch size, number of epochs, optimizer type, convolution kernel size, and number of channels; The model error is judged using MSE (Mean Squared Error). The calculation formula of MSE is as follows: , in, represents the model error judgment parameter, Indicates the j The actual junction temperature corresponding to the gate undershoot data is Indicates the jThe junction temperature prediction value is obtained by processing the gate undershoot data through the GAF-CNN neural network model. Indicates the total number of gate undershoot signals in this group.

[0029] The present invention uses GAF to convert the one-dimensional gate undershoot waveform time series into a two-dimensional image, and uses the one-dimensional data and the two-dimensional image as input to construct a hybrid CNN neural network model to predict the junction temperature of the IGBT, so as to solve the problems of the existing temperature-sensitive electrical parameter method being affected by aging and having high dependence on load current, and to perform more accurate junction temperature prediction.

[0030] Example 1 This embodiment proposes an IGBT junction temperature online monitoring method based on GAF-CNN, such as Figure 1 As shown, the following steps are included: Step 1: Collect gate undershoot data of the IGBT to be tested; Step 2: Input the gate undershoot data of the IGBT to be tested into the trained IGBT junction temperature online monitoring model to obtain the real-time junction temperature; The structure of the IGBT junction temperature online monitoring model is a GAF-CNN neural network model.

[0031] Example 2 This embodiment proposes an IGBT junction temperature online monitoring method based on GAF-CNN, such as Figure 1 As shown, the following steps are included: Step 1: Collect gate undershoot data of the IGBT to be tested; Step 2: Input the gate undershoot data of the IGBT to be tested into the trained IGBT junction temperature online monitoring model to obtain the real-time junction temperature; The structure of the IGBT junction temperature online monitoring model is a GAF-CNN neural network model.

[0032] The trained IGBT junction temperature online monitoring model is obtained through the following steps: Step A1: collecting gate undershoot data at different junction temperatures to establish a sample data set; Step A2: Use GAF to convert the sample dataset into a three-dimensional dataset; Step A3: Use the sample data set and the three-dimensional data set to train and test the IGBT junction temperature online monitoring model. If the test results meet the set conditions, a trained IGBT junction temperature online monitoring model is obtained. If the test results do not meet the set conditions, repeat step A3.

[0033] Example 3 This embodiment proposes an IGBT junction temperature online monitoring method based on GAF-CNN, such as Figure 1 As shown, the following steps are included: Step 1: Collect gate undershoot data of the IGBT to be tested; Step 2: Input the gate undershoot data of the IGBT to be tested into the trained IGBT junction temperature online monitoring model to obtain the real-time junction temperature; The structure of the IGBT junction temperature online monitoring model is a GAF-CNN neural network model.

[0034] The trained IGBT junction temperature online monitoring model is obtained through the following steps: Step A1: collecting gate undershoot data at different junction temperatures to establish a sample data set; The IGBT is heated by a heating module. When the upper tube of the IGBT is turned off, the gate undershoot waveform of the lower tube is collected at a sampling frequency of 100 MHz through the acquisition circuit and the analog-to-digital converter. The junction temperature of the IGBT module is recorded and summarized using an infrared temperature gun to obtain a one-dimensional sample data set. The sample data set is divided into a one-dimensional data training set and a one-dimensional data test set according to a fixed ratio.

[0035] Step A2: Use GAF to convert the sample dataset into a three-dimensional dataset; Step A3: Use the sample data set and the three-dimensional data set to train and test the IGBT junction temperature online monitoring model. If the test results meet the set conditions, a trained IGBT junction temperature online monitoring model is obtained. If the test results do not meet the set conditions, repeat step A3.

[0036] Example 4 This embodiment proposes an IGBT junction temperature online monitoring method based on GAF-CNN, such as Figure 1 As shown, the following steps are included: Step 1: Collect gate undershoot data of the IGBT to be tested; Step 2: Input the gate undershoot data of the IGBT to be tested into the trained IGBT junction temperature online monitoring model to obtain the real-time junction temperature; The structure of the IGBT junction temperature online monitoring model is a GAF-CNN neural network model.

[0037] The trained IGBT junction temperature online monitoring model is obtained through the following steps: Step A1: collecting gate undershoot data at different junction temperatures to establish a sample data set; Step A2: Use GAF to convert the sample dataset into a three-dimensional dataset; Step A2.1, representing the collected gate undershoot data as a one-dimensional array X; Step A2.2, normalize the one-dimensional array X; Step A2.3, performing polar coordinate conversion on the normalized one-dimensional array X; Step A2.4, convert the cosine value of the sum of the angles in the one-dimensional array X after polar coordinate conversion into a first characteristic matrix, i.e., a GASF matrix, and convert the sine value of the difference of the angles into a second characteristic matrix, i.e., a GADF matrix; Step A2.5: Convert the first feature matrix and the second feature matrix into grayscale images respectively and map them to the range of [0, 255]. Superimpose the layers by adding one layer of the first feature matrix and one layer of the second feature matrix to obtain a three-dimensional image. Summarize the three-dimensional images to obtain a three-dimensional data set and divide it into a three-dimensional data training set and a three-dimensional data test set.

[0038] Step A3: Use the sample data set and the three-dimensional data set to train and test the IGBT junction temperature online monitoring model. If the test results meet the set conditions, a trained IGBT junction temperature online monitoring model is obtained. If the test results do not meet the set conditions, repeat step A3.

[0039] Example 5 This embodiment proposes an IGBT junction temperature online monitoring method based on GAF-CNN, such as Figure 1 As shown, the following steps are included: Step 1: Collect gate undershoot data of the IGBT to be tested; Step 2: Input the gate undershoot data of the IGBT to be tested into the trained IGBT junction temperature online monitoring model to obtain the real-time junction temperature; The structure of the IGBT junction temperature online monitoring model is a GAF-CNN neural network model.

[0040] The trained IGBT junction temperature online monitoring model is obtained through the following steps: Step A1: collecting gate undershoot data at different junction temperatures to establish a sample data set; Step A2: Use GAF to convert the sample dataset into a three-dimensional dataset; Step A2.1, representing the collected gate undershoot data as a one-dimensional array X; One-dimensional array X={ x 1, x 2, x 3,…, x i},in x i Indicates that the gate undershoot is 10* after the IGBT module is turned off. i Voltage value in nanoseconds; Step A2.2, normalize the one-dimensional array X; Normalizing a one-dimensional array X can be achieved using the following formula: , in, represents the gate undershoot data after normalization, Indicates that the gate undershoot is 10* after the IGBT module is turned off. i The voltage value at nanoseconds, Represented as the maximum value in the one-dimensional array X, Represents the minimum value in the one-dimensional array X.

[0041] Normalizing the one-dimensional array X can also be achieved using the following formula: , in, represents the gate undershoot data after normalization, Indicates that the gate undershoot is 10* after the IGBT module is turned off. i The voltage value at nanoseconds, Represented as the maximum value in the one-dimensional array X, Represents the minimum value in the one-dimensional array X.

[0042] Step A2.3, performing polar coordinate conversion on the normalized one-dimensional array X; After polar coordinate conversion, the variables angle and radius are used to represent the value of the time series and its corresponding timestamp. The calculation formula for polar coordinate conversion is as follows: , in, Represents the first i Gate undershoot data, Indicates the n variable angles, Represents the total number of gate undershoot signals in the one-dimensional array X.

[0043] Step A2.4, convert the cosine value of the sum of the angles in the one-dimensional array X after polar coordinate conversion into a first characteristic matrix, i.e., a GASF matrix, and convert the sine value of the difference of the angles into a second characteristic matrix, i.e., a GADF matrix; Step A2.5: Convert the first feature matrix and the second feature matrix into grayscale images respectively and map them to the range of [0, 255]. Superimpose the layers by adding one layer of the first feature matrix and one layer of the second feature matrix to obtain a three-dimensional image. Summarize the three-dimensional images to obtain a three-dimensional data set and divide it into a three-dimensional data training set and a three-dimensional data test set.

[0044] Step A3: Use the sample data set and the three-dimensional data set to train and test the IGBT junction temperature online monitoring model. If the test results meet the set conditions, a trained IGBT junction temperature online monitoring model is obtained. If the test results do not meet the set conditions, repeat step A3.

[0045] Example 6 This embodiment proposes an IGBT junction temperature online monitoring method based on GAF-CNN, such as Figure 1 As shown, the following steps are included: Step 1: Collect gate undershoot data of the IGBT to be tested; Step 2: Input the gate undershoot data of the IGBT to be tested into the trained IGBT junction temperature online monitoring model to obtain the real-time junction temperature; The structure of the IGBT junction temperature online monitoring model is a GAF-CNN neural network model.

[0046] The trained IGBT junction temperature online monitoring model is obtained through the following steps: Step A1: collecting gate undershoot data at different junction temperatures to establish a sample data set; Step A2: Use GAF to convert the sample dataset into a three-dimensional dataset; Step A3: Use the sample data set and the three-dimensional data set to train and test the IGBT junction temperature online monitoring model. If the test results meet the set conditions, a trained IGBT junction temperature online monitoring model is obtained. If the test results do not meet the set conditions, repeat step A3.

[0047] Step A3.1: Use the one-dimensional data training set as input to train the one-dimensional convolutional neural network model CNN1: perform calculations on the convolution layer, pooling layer, activation function, and fully connected layer to obtain the first eigenvector of the hybrid CNN neural network model; Step A3.2: Using the 3D data training set as input, train the 2D convolutional neural network model CNN2 by performing calculations on the convolutional layer, pooling layer, activation function, and fully connected layer to obtain the second eigenvector of the hybrid CNN neural network model. Step A3.3: Use the feature fusion layer to fuse the first and second feature vectors, input the fused feature vectors into the hybrid input model CNN3 for training, and output the junction temperature prediction value; Step A3.4: Input the junction temperature prediction value into the initial GAF-CNN neural network model for training. After calculating the model prediction value through forward propagation, adjust the model hyperparameters or network structure. Step A3.5: Input the test set into the GAF-CNN neural network model with adjusted model hyperparameters or network structure to calculate the model error. If the model error meets the set training accuracy, a trained IGBT junction temperature online monitoring model is obtained; if the model error does not meet the set training accuracy, repeat steps A3.1-A3.5.

[0048] By utilizing gate undershoot and a GAF-CNN neural network, this invention overcomes the shortcomings of conventional methods for monitoring junction temperature of temperature-sensitive electrical parameters, such as parameter drift with aging, which affects junction temperature prediction and the load current dependence of temperature-sensitive electrical parameters. The gate undershoot signal proposed in this invention is unaffected by aging of the IGBT module's bond wires, and its parameter properties remain unchanged due to module aging. Furthermore, the GAF-CNN neural network proposed in this invention effectively addresses the load current dependence of temperature-sensitive electrical parameters, improving the accuracy of junction temperature monitoring.

[0049] In the GAF-CNN neural network model of the present invention, the size of the convolution kernel in the convolution layer is adjusted by the model output error; the activation function of the activation layer is the Relu function; the pooling method in the pooling layer adopts maximum pooling; a Batch Normalization (BN) layer is added after each fully connected layer to perform a BatchNormlization operation; the output of the CNN3 output layer is the predicted junction temperature value.

Claims

1. The IGBT junction temperature online monitoring method based on GAF-CNN is characterized by: The following steps are involved: Step 1: Collect gate undershoot data of the IGBT to be tested; Step 2: Input the gate undershoot data of the IGBT to be tested into the trained IGBT junction temperature online monitoring model to obtain the real-time junction temperature; The structure of the IGBT junction temperature online monitoring model is a GAF-CNN neural network model.

2. The method for online monitoring of IGBT junction temperature based on GAF-CNN according to claim 1, characterized in that: The trained IGBT junction temperature online monitoring model is obtained by the following steps: Step A1: collecting gate undershoot data at different junction temperatures to establish a sample data set; Step A2: Use GAF to convert the sample dataset into a three-dimensional dataset; Step A3: Use the sample data set and the three-dimensional data set to train and test the IGBT junction temperature online monitoring model. If the test results meet the set conditions, a trained IGBT junction temperature online monitoring model is obtained. If the test results do not meet the set conditions, repeat step A3.

3. The IGBT junction temperature online monitoring method based on GAF-CNN according to claim 2, characterized in that: The step A1 comprises: The IGBT is heated by a heating module. When the upper tube of the IGBT is turned off, the gate undershoot waveform of the lower tube is collected at a sampling frequency of 100 MHz through the acquisition circuit and the analog-to-digital converter. The junction temperature of the IGBT module is recorded and summarized using an infrared temperature gun to obtain a one-dimensional sample data set. The sample data set is divided into a one-dimensional data training set and a one-dimensional data test set according to a fixed ratio.

4. The method for online monitoring of IGBT junction temperature based on GAF-CNN according to claim 2, characterized in that: The step A2 comprises: Step A2.1, representing the collected gate undershoot data as a one-dimensional array X; Step A2.2, normalize the one-dimensional array X; Step A2.3, performing polar coordinate conversion on the normalized one-dimensional array X; Step A2.4, convert the cosine value of the sum of the angles in the one-dimensional array X after polar coordinate conversion into a first characteristic matrix, i.e., a GASF matrix, and convert the sine value of the difference of the angles into a second characteristic matrix, i.e., a GADF matrix; Step A2.5: Convert the first feature matrix and the second feature matrix into grayscale images respectively and map them to the range of [0, 255]. Superimpose the layers by adding one layer of the first feature matrix and one layer of the second feature matrix to obtain a three-dimensional image. Summarize the three-dimensional images to obtain a three-dimensional data set and divide it into a three-dimensional data training set and a three-dimensional data test set.

5. The method for online monitoring of IGBT junction temperature based on GAF-CNN according to claim 4, characterized in that: The one-dimensional array X={ x 1, x 2, x 3,…, x i },in x i Indicates that the gate undershoot is 10* after the IGBT module is turned off. i Voltage value in nanoseconds; After the polar coordinate conversion, the values of the time series and their corresponding timestamps are represented by the variables angle and radius respectively; the calculation formula of the polar coordinate conversion is as follows: ， in, Represents the first i Gate undershoot data, Indicates the n variable angles, Represents the total number of gate undershoot signals in the one-dimensional array X.

6. The method for online monitoring of IGBT junction temperature based on GAF-CNN according to claim 4, characterized in that: The step A2.2 includes: Normalizing a one-dimensional array X can be achieved using the following formula: ， in, represents the gate undershoot data after normalization, Indicates that the gate undershoot is 10* after the IGBT module is turned off. i The voltage value at nanoseconds, Represented as the maximum value in the one-dimensional array X, Represents the minimum value in the one-dimensional array X.

7. The method for online monitoring of IGBT junction temperature based on GAF-CNN according to claim 4, characterized in that: The step A2.2 includes: Normalizing a one-dimensional array X can be achieved using the following formula: ， in, represents the gate undershoot data after normalization, Indicates that the gate undershoot is 10* after the IGBT module is turned off. i The voltage value at nanoseconds, Represented as the maximum value in the one-dimensional array X, Represents the minimum value in the one-dimensional array X.

8. The method for online monitoring of IGBT junction temperature based on GAF-CNN according to claim 4, characterized in that: The expression of the GASF matrix is: in, represents the GASF matrix, Indicates the n variable angles; The expression of the GADF matrix is: in, represents the GADF matrix, Indicates the n variable angles.

9. The method for online monitoring of IGBT junction temperature based on GAF-CNN according to claim 2, characterized in that: Step A3 includes: Step A3.1: Use the one-dimensional data training set as input to train the one-dimensional convolutional neural network model CNN1: perform calculations on the convolution layer, pooling layer, activation function, and fully connected layer to obtain the first eigenvector of the hybrid CNN neural network model; Step A3.2: Use the 3D data training set as input to train the 2D convolutional neural network model CNN2: perform calculations on the convolution layer, pooling layer, activation function, and fully connected layer to obtain the second eigenvector of the hybrid CNN neural network model; Step A3.3: Use the feature fusion layer to fuse the first and second feature vectors, input the fused feature vectors into the hybrid input model CNN3 for training, and output the junction temperature prediction value; Step A3.4: Input the junction temperature prediction value into the initial GAF-CNN neural network model for training. After calculating the model prediction value through forward propagation, adjust the model hyperparameters or network structure. Step A3.5: Input the test set into the GAF-CNN neural network model with adjusted model hyperparameters or network structure to calculate the model error. If the model error meets the set training accuracy, a trained IGBT junction temperature online monitoring model is obtained; if the model error does not meet the set training accuracy, repeat steps A3.1-A3.

5.

10. The method for online monitoring of IGBT junction temperature based on GAF-CNN according to claim 9, characterized in that: The model hyperparameters include learning rate, batch size, number of iterations, optimizer type, convolution kernel size, and number of channels; The model error is judged using MSE, and the calculation formula of MSE is as follows: ， in, represents the model error judgment parameter, Indicates the j The actual junction temperature corresponding to the gate undershoot data is Indicates the j The junction temperature prediction value is obtained by processing the gate undershoot data through the GAF-CNN neural network model. Indicates the total number of gate undershoot signals in this group.

Images