An IGBT lifetime prediction method based on a GA-Elman-LSTM combined model

Abstract

This invention discloses an IGBT lifetime prediction method based on a GA-Elman-LSTM combined model. Specifically, the method involves: acquiring IGBT aging data; preprocessing the data; using a GA-optimized Elman neural network to import the training dataset into the GA-Elman model for training; constructing error correction training and validation sets using the trained GA-Elman preliminary prediction model; training the LSTM error correction prediction model; and building a GA-Elman-LSTM error correction combined model based on the GA-Elman preliminary prediction model and the LSTM error correction prediction model for final combined prediction. This invention conveniently and efficiently enables the model to converge to the global optimum, effectively improving the model's prediction accuracy; it also improves the model's prediction stability and reduces prediction risk.

Description

Technical Field

[0001] This invention pertains to IGBT maintenance technology, and particularly relates to an IGBT lifetime prediction method based on a GA-Elman-LSTM combined model. Background Technology

[0002] Traction converters are core equipment in urban rail vehicles, and their service condition directly affects the safe and reliable operation of urban rail transit. According to industry surveys, IGBT failures account for 31% of converter system failures, making them the most vulnerable components in the converter system. Furthermore, the complex and variable operating conditions of urban rail vehicles lead to an even higher IGBT failure rate. Accurately predicting IGBT lifespan can effectively avoid unplanned maintenance, reduce operation and maintenance costs, and improve the intelligence level of urban rail transit system operation and maintenance. Currently, IGBT lifespan prediction technologies mainly include two categories: model-based lifespan prediction and data-driven lifespan prediction.

[0003] Model-based IGBT lifetime prediction is a method that uses multiphysics simulation to obtain thermal stress data of the device, processes it, and then inputs it into a lifetime model to predict the IGBT lifetime. There are two main types of IGBT lifetime models: analytical models and physical models. Analytical models are empirical models that establish the relationship between the number of failure cycles and the physical quantities of failure based on accelerated testing data, while physical models utilize microscopic changes such as strain during aging failure of the module's internal structure. Both analytical and physical models require key coefficients obtained through extensive experiments or simulations, making modeling complex and time-consuming. Furthermore, due to the differences in the physical properties of IGBT modules, their universality is relatively poor.

[0004] Data-driven IGBT lifetime prediction extracts feature information from IGBT aging data using various data processing methods, and then combines this information with a prediction model to predict the IGBT's lifetime. This approach has become the current development direction for IGBT lifetime prediction technology due to its advantages such as easy data acquisition, simple modeling, and the ability to perform online prediction. However, because the values ​​of IGBT aging parameters are relatively small, and the degradation amount during the entire aging process is also small, traditional single-item prediction models struggle to comprehensively and accurately extract information from IGBT aging data, resulting in low prediction accuracy and poor versatility. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an IGBT lifetime prediction method based on a GA-Elman-LSTM combined model. This method solves the problems of inaccurate and incomplete extraction of IGBT aging data by single-item prediction models, effectively improving model prediction accuracy and reducing prediction risks.

[0006] The present invention provides a method for predicting IGBT lifetime based on a GA-Elman-LSTM combined model, comprising the following steps:

[0007] Step 1: Acquisition of IGBT aging data.

[0008] The condition monitoring circuit installed in the traction converter of the rail vehicle is used to collect IGBT aging failure data and transmit the data to the host computer for storage.

[0009] Step 2: Data preprocessing.

[0010] The IGBT aging data was smoothed and denoised. Then, the autocorrelation time series data sample set was constructed using the sliding time window method. The time window was set to 10 and the step size was set to 1. Finally, the data sample set was divided into 70% training set, 15% validation set, and 15% prediction set.

[0011] Step 3: Initial training of the GA-Elman prediction model. Import the training dataset from Step 2 into the GA-Elman model for training. The specific training steps are as follows:

[0012] S31. Elman Model Initialization: Determine the Elman model topology and basic network parameters. Based on the training dataset, the number of neurons in the input layer of the GA-Elman model should be set to 10, the number of neurons in the output layer to 1, the number of neurons in the hidden layer to 15, the iteration period to 10,000, and the training precision to 0.01. Finally, randomly obtain a set of initial weights and thresholds. The mathematical model of the Elman neural network at time k is:

[0013] y(k)=g(ω3x(k)+b2) (1)

[0014] x(k)=f(ω1(u(k-1))+ω2z(k)+b1) (2)

[0015] z(k)=x(k-1) (3)

[0016] In the formula: u is an r-dimensional input vector; y is a p-dimensional output vector; x is a q-dimensional hidden layer node vector; z is a q-dimensional feedback state vector; g(·) and f(·) are the transfer functions of the output layer and the hidden layer, respectively; ω1, ω2 and ω3 are the connection weights from the input layer to the hidden layer, from the receiving layer to the hidden layer and from the hidden layer to the output layer, respectively; b1 and b2 represent the thresholds of the hidden layer and the output layer, respectively.

[0017] S32. GA optimizes Elman weights and thresholds: First, GA obtains the initial weight thresholds of the network and encodes them. Then, it constructs the fitness function and performs genetic operations. Finally, it calculates the fitness value and determines whether the termination condition is met. If it is met, the optimal weight threshold is assigned to the Elman model; otherwise, the genetic operations continue.

[0018] S33, GA-Elman model training and prediction: Elman uses the optimal initial weights and thresholds obtained by GA search for training. After training, 15% of the validation set is imported into the model for validation. Once the prediction accuracy of the validation set meets the set threshold, the GA-Elman preliminary prediction model is output.

[0019] Step 4: Construction of error correction training and validation sets.

[0020] The trained GA-Elman preliminary prediction model is used to fit the training dataset and the validation dataset respectively, and the prediction error is calculated. The original IGBT aging failure data is used as the input of the LSTM model, and the corresponding error data is used as the output to construct the error correction training dataset and validation set.

[0021] Step 5: Training the LSTM error correction prediction model.

[0022] Determine the topology and basic network parameters of the LSTM model. Similarly, set the number of LSTM input neurons and output neurons to 10 and 1 respectively, the number of hidden neurons to 200, the number of iterations to 400, and the training accuracy to 0.01. Input the error correction training set constructed in step 4 into the LSTM model for training. After training, import the error correction validation set into the model for validation. After the prediction accuracy of the validation set meets the set threshold, output the LSTM error correction prediction model.

[0023] The calculation process for the LSTM model is shown below:

[0024] f t =sigmoid(W f ·[h t-1 ,x t ]+b f (4)

[0025] i t =sigmoid(W i ·[h t-1 ,x t ]+b i (5)

[0026]

[0027]

[0028] o t =sigmoid(W o ·[h t-1 ,x t ]+b o (8)

[0029] ht =o t ⊙tanh(C t (9)

[0030] In the formula: f t i t o t , These are the forget gate, input gate, output gate, and intermediate unit state values, respectively; W f W i W o W c These are the weight matrices for the corresponding gate structures; b f b i b o b c These are the bias matrices for the corresponding gate structures; x t Given the input sequence, h t-1 For the final output of the cell unit at the previous time step, [h t-1 ,x t [] indicates that two vectors are concatenated to form a longer vector; C t h represents the current cell state value. t This represents the final output at the current time step; ⊙ indicates element-wise multiplication of the vector elements; sigmoid and tanh are the activation functions of the LSTM model.

[0031] Step 6: Output the final combined prediction model.

[0032] Based on the GA-Elman preliminary prediction model and the LSTM error correction prediction model trained in steps 3 and 5, a GA-Elman-LSTM error correction combined model is constructed. 15% of the prediction set is imported into the combined model for lifetime prediction. The output of the GA-Elman model is used as the preliminary prediction value, and the output of the LSTM model is used as the error correction value. The final model prediction value is obtained by subtracting the preliminary prediction value from the error correction value.

[0033] The beneficial technical effects of this invention are as follows:

[0034] 1. This invention optimizes the training process of the Elman neural network model by utilizing three operators—selection, crossover, and mutation—in a genetic algorithm. This effectively solves the problem of traditional Elman neural network models easily getting trapped in local minima, enabling the model to converge to the global optimum efficiently and effectively improving the model's prediction accuracy.

[0035] 2. This invention combines a GA-optimized Elman neural network with an LSTM network to establish a GA-Elman-LSTM error correction combined prediction model, which is then applied to IGBT lifetime prediction. By leveraging the advantages of the combined prediction model, the accuracy and precision of the model prediction are further improved, while also enhancing the model prediction stability and reducing prediction risks. Attached Figure Description

[0036] Figure 1 This is a flowchart of the IGBT lifetime prediction method based on the GA-Elman-LSTM combined model of the present invention.

[0037] Figure 2 The flowchart for GA-Elman training.

[0038] Figure 3 This is a diagram of the Elman neural network topology.

[0039] Figure 4 This is a diagram of the internal structure of an LSTM cell. Detailed Implementation

[0040] The present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0041] The flowchart of the IGBT lifetime prediction method based on the GA-Elman-LSTM combined model of the present invention is as follows: Figure 1 As shown, the specific steps include:

[0042] Step 1: Acquisition of IGBT aging data.

[0043] The condition monitoring circuit installed in the traction converter of the rail vehicle is used to collect IGBT aging failure data and transmit the data to the host computer for storage.

[0044] Failure parameter selection criteria: 1. The selected failure parameters should be able to effectively reflect the degree of IGBT aging failure; 2. The difficulty and cost of collecting the selected failure parameters should be low.

[0045] The collector-emitter saturation voltage drop of IGBTs gradually increases as the device failure process progresses. Moreover, the data acquisition does not require damage to the IGBT module packaging structure, the acquisition technology is mature, and the acquisition cost is low. Therefore, the collector-emitter saturation voltage drop of IGBTs is selected as its aging failure parameter, and an increase of 5% in the collector-emitter saturation voltage drop is taken as the standard for IGBT module failure.

[0046] Step 2: Data preprocessing.

[0047] The IGBT aging data was smoothed and denoised. Then, the autocorrelation time series data sample set was constructed using the sliding time window method. The time window was set to 10 and the step size was set to 1. Finally, the data sample set was divided into 70% training set, 15% validation set, and 15% prediction set.

[0048] Step 3: Initial training of the GA-Elman prediction model. Import the training dataset from Step 2 into the GA-Elman model for training, such as... Figure 2 As shown, the specific training steps are as follows:

[0049] S31. Elman Model Initialization: Determine the Elman model topology and basic network parameters (e.g., ...). Figure 3 Based on the training dataset, the GA-Elman model should have 10 input layer neurons, 1 output layer neuron, and 15 hidden layer neurons. The iteration cycle should be 10,000 times, and the training precision should be 0.01. Finally, a random set of initial weights and thresholds is obtained. The Elman neural network mathematical model at time k is:

[0050] y(k)=g(ω3x(k)+b2) (1)

[0051] x(k)=f(ω1(u(k-1))+ω2z(k)+b1) (2)

[0052] z(k)=x(k-1) (3)

[0053] In the formula: u is an r-dimensional input vector; y is a p-dimensional output vector; x is a q-dimensional hidden layer node vector; z is a q-dimensional feedback state vector; g(·) and f(·) are the transfer functions of the output layer and the hidden layer, respectively; ω1, ω2 and ω3 are the connection weights from the input layer to the hidden layer, from the receiving layer to the hidden layer and from the hidden layer to the output layer, respectively; b1 and b2 represent the thresholds of the hidden layer and the output layer, respectively.

[0054] S32. GA optimizes Elman weights and thresholds: First, GA obtains the initial weight thresholds of the network and encodes them. Then, it constructs the fitness function and performs genetic operations. Finally, it calculates the fitness value and determines whether the termination condition is met. If it is met, the optimal weight threshold is assigned to the Elman model; otherwise, the genetic operations continue.

[0055] S33, GA-Elman model training and prediction: Elman uses the optimal initial weights and thresholds obtained by GA search for training. After training, 15% of the validation set is imported into the model for validation. Once the prediction accuracy of the validation set meets the set threshold, the GA-Elman preliminary prediction model is output.

[0056] Step 4: Construction of error correction training and validation sets.

[0057] The trained GA-Elman preliminary prediction model is used to fit the training dataset and the validation dataset respectively, and the prediction error is calculated. The original IGBT aging failure data is used as the input of the LSTM model, and the corresponding error data is used as the output to construct the error correction training dataset and validation set.

[0058] Step 5: Training the LSTM error correction prediction model.

[0059] Determine the topology and basic network parameters of the LSTM model. Similarly, set the number of LSTM input neurons and output neurons to 10 and 1 respectively, the number of hidden neurons to 200, the number of iterations to 400, and the training accuracy to 0.01. Input the error correction training set constructed in step 4 into the LSTM model for training. After training, import the error correction validation set into the model for validation. After the prediction accuracy of the validation set meets the set threshold, output the LSTM error correction prediction model.

[0060] The internal structure of the LSTM unit is shown in the attached figure. Figure 4 As shown, the corresponding calculation process is as follows:

[0061] f t =sigmoid(W f ·[h t-1 ,x t ]+b f (4)

[0062] i t =sigmoid(W i ·[h t-1 ,x t ]+b i (5)

[0063]

[0064]

[0065] o t =sigmoid(W o ·[h t-1 ,x t ]+b o (8)

[0066] h t =o t ⊙tanh(C t (9)

[0067] In the formula: f t i t ot , These are the forget gate, input gate, output gate, and intermediate unit state values, respectively; W f W i W o W c These are the weight matrices for the corresponding gate structures; b f b i b o b c These are the bias matrices for the corresponding gate structures; x t Given the input sequence, h t-1 For the final output of the cell unit at the previous time step, [h t-1 ,x t [] indicates that two vectors are concatenated to form a longer vector; C t h represents the current cell state value. t This represents the final output at the current time step; ⊙ indicates element-wise multiplication of the vector elements; sigmoid and tanh are the activation functions of the LSTM model.

[0068] Step 6: Output the final combined prediction model.

[0069] Based on the GA-Elman preliminary prediction model and the LSTM error correction prediction model trained in steps 3 and 5, a GA-Elman-LSTM error correction combined model is constructed. 15% of the prediction set is imported into the combined model for lifetime prediction. The output of the GA-Elman model is used as the preliminary prediction value, and the output of the LSTM model is used as the error correction value. The final model prediction value is obtained by subtracting the preliminary prediction value from the error correction value.

[0070] At the system level, this invention provides a method for predicting the lifespan of IGBT modules that can be deployed on a large scale. It effectively accelerates the detection speed of sleeper defects and enables deployment and online detection in embedded devices. At the model level, this invention effectively corrects the prediction error of the GA-Elman model by training it with an LSTM model, offering advantages of simple principle and easy operation.

Claims

1. A method for predicting IGBT lifetime based on a GA-Elman-LSTM combined model, characterized in that, Includes the following steps: Step 1: Acquiring IGBT aging data; The condition monitoring circuit installed in the traction converter of the rail vehicle is used to collect IGBT aging failure data and transmit the data to the host computer for storage; Step 2: Data preprocessing; The IGBT aging data was smoothed and denoised. Then, the autocorrelation time series data sample set was constructed using the sliding time window method. The time window was set to 10 and the step size was set to 1. Finally, the data sample set was divided into 70% training set, 15% validation set, and 15% prediction set. Step 3: Initial training of the GA-Elman prediction model. Import the training dataset from Step 2 into the GA-Elman model for training. The specific training steps are as follows: S31. Elman Model Initialization: Determine the Elman model topology and basic network parameters. Based on the training dataset, the number of neurons in the input layer of the GA-Elman model should be set to 10, the number of neurons in the output layer to 1, the number of neurons in the hidden layer to 15, the iteration period to 10,000, and the training precision to 0.

01. Finally, randomly obtain a set of initial weights and thresholds. The mathematical model of the Elman neural network at time k is: ； ； ； In the formula: u is an r-dimensional input vector; y is a p-dimensional output vector; x is a q-dimensional hidden layer node vector; z is a q-dimensional feedback state vector; g(·) and f(·) are the transfer functions of the output layer and the hidden layer, respectively; b1 and b2 represent the connection weights from the input layer to the hidden layer, from the continuation layer to the hidden layer, and from the hidden layer to the output layer, respectively; b1 and b2 represent the thresholds of the hidden layer and the output layer, respectively. S32. GA optimizes Elman weights and thresholds: First, GA obtains the initial weight thresholds of the network and encodes them. Then, it constructs the fitness function and performs genetic operations. Finally, it calculates the fitness value and determines whether the termination condition is met. If it is met, the optimal weight threshold is assigned to the Elman model. Otherwise, the genetic operations continue. S33, GA-Elman model training and prediction: Elman uses the optimal initial weights and thresholds obtained by GA search for training. After training, 15% of the validation set is imported into the model for validation. After the prediction accuracy of the validation set meets the set threshold, the GA-Elman preliminary prediction model is output. Step 4: Construction of training and validation sets for error correction; The trained GA-Elman preliminary prediction model is fitted to the training dataset and the validation dataset respectively, and the prediction error is calculated. The original IGBT aging failure data is used as the input of the LSTM model, and the corresponding error data is used as the output to construct the error correction training dataset and validation set. Step 5: Training the LSTM error correction prediction model; Determine the topology and basic network parameters of the LSTM model. Similarly, set the number of LSTM input neurons and output neurons to 10 and 1 respectively, the number of hidden neurons to 200, the number of iterations to 400, and the training accuracy to 0.

01. Input the error correction training set constructed in step 4 into the LSTM model for training. After training, import the error correction validation set into the model for validation. After the prediction accuracy of the validation set meets the set threshold, output the LSTM error correction prediction model. The calculation process for the LSTM model is shown below: ； ； ； ； ； ； In the formula: , , , These are the forget gate, input gate, output gate, and intermediate unit state values, respectively. , , , These are the weight matrices for the corresponding gate structures; , , , These are the bias matrices for the corresponding gate structures; Given the input sequence, This is the final output of the cell unit from the previous moment. This indicates that two vectors are concatenated to form a longer vector. This represents the current cell state value. This is the final output at the current moment; This indicates that the elements in the vector are multiplied element by element; sigmoid and tanh are the activation functions of the LSTM model; Step 6: Output the final combined prediction model; Based on the GA-Elman preliminary prediction model and the LSTM error correction prediction model trained in steps 3 and 5, a GA-Elman-LSTM error correction combined model is constructed. 15% of the prediction set is imported into the combined model for lifetime prediction. The output of the GA-Elman model is used as the preliminary prediction value, and the output of the LSTM model is used as the error correction value. The final model prediction value is obtained by subtracting the preliminary prediction value from the error correction value.

Images