Construction method of IGBT module life prediction model based on GRU neural network

Abstract

The present invention relates to the technical field of reliability analysis of IGBT modules in power semiconductor devices, and discloses a method for constructing an IGBT module life prediction model based on a GRU neural network, an IGBT module life prediction model with optimal parameters constructed using the construction method, and a method for predicting the life of an IGBT module using the prediction model. The present invention proposes a method for constructing a life prediction model based on a GRU neural network based on the aging characteristic quantity of the IGBT module. The present invention is rationally designed. When constructing the model, the grid search method is used to tune the parameters of the GRU neural network life prediction model, which improves the problem of overly random selection of recurrent neural network hyperparameters or only fine-tuning of existing reference neural network parameters. The proposed GRU neural network life prediction model with optimal parameters has higher prediction accuracy and better adaptability to the problem of life prediction of IGBT modules.

Description

Technical Field

[0001] The present invention relates to the technical field of reliability analysis of IGBT modules in power semiconductor devices, and more specifically, to a method for constructing an IGBT module life prediction model based on a GRU neural network, an IGBT module life prediction model with optimal parameters constructed using the construction method, and a method for predicting the life of an IGBT module using the prediction model. Background Art

[0002] With the continuous advancement of power electronics and semiconductor manufacturing technologies, the application of power semiconductor devices in fields such as grid-connected wind and solar power generation, electric vehicles, power systems, and aerospace propulsion systems is expanding. This booming development of new energy vehicles, rail transit, and smart grids is generating massive demand, while also facing a severe shortage of power semiconductor production capacity. This presents new opportunities for the recycling and reuse of used power semiconductor devices.

[0003] The insulated-gate bipolar transistor (IGBT) is a composite power semiconductor device that combines the structures of a power metal-oxide-semiconductor field-effect transistor (MOSFET) and a bipolar junction transistor (BJT). It offers advantages such as fast switching speed, low drive power, simple control circuitry, and rapid on-off times. It is widely used in scenarios where electronic components, such as electric vehicles and electric vehicle AC / DC charging modules, are in use for extended periods or experience frequent power cycling. For example, in electric vehicles and their charging equipment, IGBTs carry large currents, accompanied by significant temperature fluctuations and subject to cyclical thermal stress. This can easily lead to long-term failure accumulation, reducing the reliability of the device and, ultimately, the entire system. Due to their high failure rate, IGBTs are considered the weakest link in electromechanical systems. Therefore, predicting the remaining useful life of IGBTs is crucial for the stable operation of power systems.

[0004] Currently, IGBT lifespan prediction is primarily based on mathematical and physical models. However, these methods are complex, lack generalizability, and have numerous model parameters, making accurate prediction difficult. Therefore, other methods are needed to improve the accuracy and speed of module lifespan prediction. Summary of the Invention

[0005] Based on this, it is necessary to address the problem of low prediction accuracy of existing IGBT life prediction models by providing a method for constructing an IGBT module life prediction model based on a recurrent neural network with optimal parameters, an IGBT module life prediction model constructed using the construction method, and a method for predicting the IGBT module life using the prediction model.

[0006] The present invention is achieved by adopting the following technical solutions:

[0007] In a first aspect, the present invention discloses a method for constructing an IGBT module life prediction model based on a GRU neural network, which is used to construct a GRU neural network IGBT module life prediction model with optimal parameters for the IGBT module.

[0008] The method for constructing an IGBT module life prediction model based on a GRU neural network includes the following steps:

[0009] Step 1: Perform a power cycle aging test on the IGBT module to obtain aging characteristic data of the IGBT module, and obtain an experimental group data set and a validation group data set;

[0010] Step 2: Preprocess the IGBT module aging characteristic data to obtain a processed experimental group data set and a processed test group data set; wherein the processed experimental group data set is divided into a training set and a validation set in a specific ratio;

[0011] Step 3: Determine the GRU neural network framework and preliminarily build the GRU neural network model;

[0012] Step 4: Use the training set and validation set to train the GRU neural network model, using the training loss function and validation loss function of the neural network as validation indicators, first determine the optimal number of hidden layers of the GRU network, then determine the optimal number of hidden layer nodes and the optimal number of training cycles, and form a GRU neural network IGBT module life prediction model based on the optimal parameters;

[0013] Among them, the method for determining the optimal number of hidden layers is:

[0014] The root mean square error function is used as a prediction indicator for the IGBT life prediction problem. Different numbers of hidden layers and different numbers of hidden layer nodes are iterated the same number of times, and the mean root mean square error of different numbers of layers is compared to determine the optimal number of hidden layers.

[0015] The method for determining the optimal number of hidden layer nodes and the optimal number of training cycles is:

[0016] First, use the empirical formula to determine the number of all candidate hidden layer nodes in the model, and then use the grid search method to tune the parameters; the grid search method is to calculate the prediction results for each case by looping through all candidate parameter selections;

[0017] The mean square error is selected as the loss function. Based on the simple cross-validation method, the optimal number of hidden layer nodes and the optimal number of training cycles are determined by the trend of the graph of training loss and validation loss.

[0018] Step 5: Substitute the processed test group data set into the GRU neural network IGBT module life prediction model based on the optimal parameters to verify the accuracy and effectiveness of its network model.

[0019] The method for constructing an IGBT module life prediction model based on a GRU neural network implements a method or process according to an embodiment of the present disclosure.

[0020] In a second aspect, the present invention provides a GRU neural network IGBT module life prediction model with optimal parameters, which is constructed using the method for constructing an IGBT module life prediction model based on a GRU neural network in the first aspect.

[0021] The present invention also provides an IGBT module life prediction method based on a GRU neural network, which uses the above-mentioned GRU neural network IGBT module life prediction model with optimal parameters;

[0022] The IGBT module life prediction method based on the GRU neural network includes:

[0023] The on-state voltage drop of the IGBT module is obtained, and the GRU neural network IGBT module life prediction model with the optimal parameters is imported to obtain the aging degree of the IGBT module.

[0024] In a third aspect, the present invention discloses a readable storage medium, which stores computer program instructions. When the computer program instructions are read and run by a processor, the method for constructing the IGBT module life prediction model based on the GRU neural network is executed.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. Based on the aging characteristics of IGBT modules, the present invention proposes a method for constructing a life prediction model based on a GRU (gated recurrent neuron) neural network. The present invention is rationally designed. During model construction, the grid search method is used to optimize the parameters of the GRU neural network life prediction model. This improves the problem of overly random selection of recurrent neural network hyperparameters or only fine-tuning existing reference neural network parameters. The proposed GRU neural network life prediction model with optimal parameters has higher prediction accuracy and better adaptability to the problem of IGBT module life prediction.

[0027] 2. The construction method proposed in the present invention is based on a data-driven neural network, which is more data-based and intelligent. It improves the problems of traditional mathematical and physical model prediction methods, such as the overly complex modeling process, insufficient generalization ability of the prediction model, and many model parameters that are difficult to obtain accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a modeling and testing flow chart of the IGBT module life prediction model provided in Example 1 of the present invention;

[0029] Figure 2 yes Figure 1 Schematic diagram of the module structure of the IGBT module, which is the life prediction object of the IGBT life prediction model;

[0030] Figure 3 yes Figure 2 Power cycle test curve of IGBT module;

[0031] Figure 4 yes Figure 2 The original data set and evaluation criteria diagram of the saturated on-state voltage drop of the IGBT module;

[0032] Figure 5 yes Figure 2 Comparison of the IGBT module saturation on-state voltage drop dataset before and after local filtering;

[0033] Figure 6 yes Figure 1 GRU recurrent neural network structure diagram of the IGBT module life prediction model;

[0034] Figure 7 yes Figure 6 Root mean square error plot for different numbers of hidden layers of the GRU neural network in

[15] ;

[0035] Figure 8 yes Figure 6 Loss trend chart of single-layer GRU neural network with different number of nodes in GRU neural network;

[0036] Figure 9 yes Figure 8 A magnified schematic diagram of area a in the middle;

[0037] Figure 10 yes Figure 8 A magnified schematic diagram of area b;

[0038] Figure 11 yes Figure 8 A magnified schematic diagram of area c in the middle;

[0039] Figure 12 yes Figure 8 Enlarged schematic diagram of the middle d area;

[0040] Figure 13 yes Figure 8 Enlarged schematic diagram of the middle e area;

[0041] Figure 14 yes Figure 8 Enlarged schematic diagram of the middle f area;

[0042] Figure 15 yes Figure 8 Enlarged schematic diagram of the middle g area;

[0043] Figure 16 yes Figure 8 Enlarged schematic diagram of the h region in the middle;

[0044] Figure 17 yes Figure 8 A magnified schematic diagram of the middle i region;

[0045] Figure 18 yes Figure 1 The prediction results of the IGBT module life prediction model using the GRU neural network with optimal parameters. DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] It should be noted that when a component is referred to as being "mounted on" another component, it may be directly on the other component or there may be a central component. When a component is considered to be "set on" another component, it may be directly set on the other component or there may be a central component. When a component is considered to be "fixed to" another component, it may be directly fixed to the other component or there may be a central component.

[0048] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used herein in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0049] Example 1

[0050] See Figure 1 , Figure 1 The figure is a flow chart of modeling and testing of the IGBT module life prediction model of the present invention. The IGBT module is the life prediction object of the IGBT module life prediction model.

[0051] Please combine Figure 2 , Figure 2The diagram below is a schematic diagram of the IGBT module structure. The IGBT module is a complex multi-layer structure consisting of bonding wires, an IGBT chip, a freewheeling diode (FWD) chip, upper and lower solder layers, upper and lower copper layers of the DBC layer, a ceramic layer between the DBC layer, a copper substrate connected to the DBC layer, and a heat sink at the outermost layer. During IGBT operation, the heat transfer mechanism is relatively simple. The bulk of the heat is generated primarily by the IGBT chip and the freewheeling diode (FWD). This heat first transfers downward to the solder layer in direct contact with the two chips, then to the various layers of the DBC layer, and then further downward to the copper substrate. Finally, some heat is transferred through the outer casing to the air, or most of it is transferred to the heat sink below. Heat is then dissipated by convection between the heat sink and the air. Analysis shows that heat transfer occurs from top to bottom through thermal conduction. Heat radiation also occurs between the IGBT chip and the freewheeling diode (FWD) chip.

[0052] The primary failure mode in IGBT modules is solder layer aging. IGBT modules have two solder layers: the first between the upper copper layer of the DBC and the bottom of the IGBT chip and freewheeling diode (FWD) chip; the second between the lower copper layer of the DBC and the underlying copper substrate. Because each layer in the module is made of different materials and has a different coefficient of thermal expansion (CTE), the thermal cycling stress generated during operation can cause physical offsets between the layers, leading to physical damage.

[0053] Based on the above analysis of IGBT module failure modes, the present invention proposes an IGBT module life prediction model and a model construction method. Specifically, it provides a method for constructing an IGBT module life prediction model based on a GRU neural network, and a life preset model constructed using this construction method.

[0054] See Figure 1 Specifically, the method for constructing the IGBT module life prediction model based on the GRU neural network includes the following steps:

[0055] Step 1: Perform a power cycle aging test on the IGBT module to obtain the aging characteristic data of the IGBT module, and obtain the experimental group data set and the verification group data set.

[0056] Step 1 is to collect the aging characteristic quantity of the IGBT module. The main method is to measure the K factor of the IGBT module through a power semiconductor K factor tester, then select the control method of the power cycle aging test, determine the parameters for the power aging test (that is, select a power loading method with constant on and off time for the power cycle aging test), and obtain the aging characteristic quantity change data and the number of power cycles (that is, the total number of cycle aging tests). Among them, a set of IGBT module aging characteristic quantity data is collected after each cycle is completed. The aging characteristic quantity data includes module on-state voltage drop data, junction temperature data, case temperature data, thermal resistance data, etc.

[0057] In this embodiment, a power cycling aging test was conducted using two IGBT modules of the same model. This yielded two sets of IGBT module aging characteristic data, which were set as the experimental and test datasets, respectively. Alternatively, N IGBT modules of the same model (N > 2) could be selected for testing to obtain more sets of data. N-1 sets of data would be used as the experimental dataset for training the neural network model, and the remaining set of data would be used as the test dataset for testing model accuracy.

[0058] Power cycling aging tests were conducted on IGBT modules using a power semiconductor K-factor tester and a power semiconductor aging test platform, including a test platform, a power semiconductor clamping platform, a wiring console, a PC host, a driver power supply, and a circulating liquid cooling system. The test subjects selected for this study were IGBT modules disassembled from electric vehicles. These modules were model IKW30N65H5, designed and manufactured by Infineon Technologies. Some performance parameters of these modules are shown in Table 1.

[0059] Table 1 Some performance parameters of IKW30N65H5 module

[0060]

[0061] The test design uses room temperature (25°C) as the starting junction temperature of the test, takes 80% of the rated maximum junction temperature as the highest junction temperature of the test, and takes the rated maximum current (T c =100℃) as the test current, leaving a test margin to prevent module breakdown, and determine the specific time of opening and closing according to the working and cooling time of the test equipment. Figure 3 ,The power cycle test design is shown in Table 2.

[0062] Table 2 Power cycle test design

[0063]

[0064] First, the IGBT modules were connected to the power semiconductor K-factor tester's test board and a pulsed test current was applied. The IGBT module conduction voltage-to-voltage (VT) images were plotted based on the test results, and the specific K-factors of the two IGBT modules were calculated. After determining their specific K-factors, the two IGBT modules were clamped on a power aging test platform and then subjected to power cycling aging tests.

[0065] During the power cycle aging test, a constant on-time and off-time is selected as the control mode of the aging test. Each time the process from on-time to off-time is completed is regarded as a cycle. After each cycle, a set of IGBT module aging characteristic data is collected, including the on-state voltage drop V when the device is at the highest junction temperature. ce_hot ; The device is at the lowest junction temperature when the state voltage drop V ce_cold ; Maximum junction temperature T j，max ; Minimum junction temperature T j，min ; Junction temperature swing ΔT j ; Maximum shell temperature T c，max ; Minimum shell temperature T c，min ;Case temperature swing ΔT c ; Thermal resistance between junction and shell, that is, the thermal resistance between the internal silicon chip and the package shell R jc The test results are shown in Table 3.

[0066] Table 3 Some power cycle test results

[0067]

[0068] It should be noted that this embodiment uses the saturation on-state voltage drop V when the module is at the highest junction temperature. ce_hot As a characteristic measure of device failure, record the saturation on-state voltage drop of a new, uncycled IGBT module and calibrate its remaining useful life to 100%. Repeat the above cycles until the saturation on-state voltage drop reaches the aging failure threshold, at which point the IGBT module is deemed failed. Record the total number of power cycles and the saturation on-state voltage drop of the IGBT module after the last cycle, and calibrate the corresponding remaining useful life to 0%. This data set is calibrated one-to-one with the number of cycles, saturation on-state voltage drop, and remaining useful life.

[0069] Step 2: preprocess the IGBT module aging characteristic data to obtain a processed experimental group data set and a processed test group data set; wherein, the processed experimental group data set is divided into a training set and a validation set in a specific ratio.

[0070] The second step mainly involves data preprocessing, including filtering and normalizing the experimental and test group data sets, determining the network input and output sequences, and dividing the experimental group data set into training and validation sets in a specific ratio.

[0071] Specifically, the preprocessing method in step 2 includes:

[0072] Step 2.1: Use the SG filtering method to filter the IGBT module aging characteristic data to reduce data volatility, enhance data periodicity, and improve the training efficiency of the neural network prediction model.

[0073] The saturation on-state voltage drop V at the highest junction temperature in the IGBT power cycle aging test is selected. ce_hot As the device aging characteristic. The commonly used evaluation standard is: a 3% increase in saturation on-state voltage drop is assessed as the beginning of aging; a 5% increase in saturation on-state voltage drop is assessed as the early stage of aging; a 10% increase in saturation on-state voltage drop is assessed as the late stage of aging; and a 20% increase in saturation on-state voltage drop is assessed as critical failure. Based on this, the aging data obtained in step 1 is plotted and the failure location is marked as follows: Figure 4 , where the vertical axis is the saturation on-state voltage drop value and the horizontal axis is the number of cycles.

[0074] To mitigate the negative impact of noise on neural network training and improve lifespan prediction accuracy, this paper uses the SG filtering method to reduce data volatility and enhance data periodicity. The SG filtering algorithm performs weighted smoothing on the data within a sliding window of a specified width (2m+1), with the weights applied using the least squares method to fit a given high-order polynomial.

[0075] First, the observation value x at the current moment t Fit a k-1 order polynomial:

[0076] x t =a0+a1·t+a2·t 2 +...+a k-1 ·t k-1 (1)

[0077] Where: t = (-m, -m+1, ..., 0, 1, ..., m-1, m), represents different moments; k is the order of the linear equations; a is the fitting parameter.

[0078] Similarly, polynomial fitting is performed on the previous and next observation points, and then they are combined into a matrix:

[0079]

[0080] Where: x represents the observation value; t represents the time; n is the width of the sliding window, that is, 2m+1; ε is a constant.

[0081] This reduces the matrix to an overdetermined system of equations:

[0082] X (2n+1)×1 =T (2n+1)×k·A K+1 +E (2n+1)×1 (2)

[0083] Where: X, T, A, and E all represent matrices.

[0084] Then the filtering result formula can be obtained:

[0085] P=T·(T t ·T) -1 ·T t ·X=B·X (3)

[0086] Where: P is the filter result matrix; B is the filter coefficient matrix, which is determined by and only by the T matrix, and the B matrix is ​​a (2m+1)×(2m+1) order matrix.

[0087] The length of the sliding window is set to 15, and due to the large amount of original data, the order of the polynomial fitting is set to 1. The local 500 groups of data after 20,000 cycles are intercepted as follows: Figure 5 , Figure 5 This is a comparison chart of the IGBT module saturated on-state voltage drop data set before and after local filtering.

[0088] Step 2.2: In order to further improve the training efficiency of the neural network prediction model and accelerate the network convergence speed, it is necessary to normalize the filtered IGBT module aging characteristic data.

[0089] The above aging characteristic data are normalized using the following formula:

[0090]

[0091] Among them, N i represents the normalized aging characteristic, D i represents the aging characteristic quantity before normalization, μ represents the mean value of the aging characteristic quantity data, and σ represents the standard deviation of the aging characteristic quantity data.

[0092] Step 2.3: Use the sliding window method to take values ​​from left to right on the preprocessed IGBT module aging feature data according to a certain window size. The values ​​inside the window are used as the input sequence X, and the first value on the right side of the value window is used as the output sequence Y. Establish a mapping relationship between the aging feature data from input to output.

[0093] This embodiment uses a sliding window with a window size of 60, takes the first 60 aging feature data as the input sequence X, and takes the 61st aging feature data as the output sequence Y to establish an input and output sequence.

[0094] The above aging characteristic data is used to establish an input and output sequence using the following formula:

[0095]

[0096] The experimental group data set and the validation group data set are processed in the same way using steps 2.1, 2.2, and 2.3 to obtain the processed experimental group data set and the processed validation group data set.

[0097] Step 2.4: Split the input and output sequences of the experimental group data set into training sets and validation sets according to the corresponding proportions.

[0098] Specifically, in this embodiment, the first 70% of the input and output sequences obtained from the original aging feature dataset of the experimental group in step 2.3 are set as a training set, and the last 30% are set as a validation set.

[0099] Step 3: Determine the GRU neural network framework and preliminarily build the GRU neural network model.

[0100] The complete GRU recurrent neural network structure is as follows Figure 6 , Figure 6 yes Figure 1 The GRU recurrent neural network structure diagram of the IGBT module life prediction model in

[15] consists of a sequence input layer, a hidden layer, a fully connected layer, and a regression output layer.

[0101] The sequence input layer imports the input sequence of the IGBT module aging feature data into the network. The hidden layer learns the high-level features of the IGBT module cyclic aging test timing data through input weights, recursive weights and bias weights. The hidden layer includes several groups of sub-hidden layers, and any group of sub-hidden layers includes several GRU layers with a configurable number of hidden layer nodes and a Dropout layer with a set initial value of the drop probability. Then there is a fully connected layer and a regression output layer, which outputs the predicted value of the IGBT module aging feature predicted by the GRU network. This embodiment uses TensorFlow.Keras as the framework and Python as the programming language to write the algorithm to create an initial GRU neural network with the above structure. Among them, the specific number of sub-hidden layers and the specific number of nodes in each GRU layer are what are to be obtained in the subsequent model construction.

[0102] Step 4: Use the training set and validation set to train the GRU neural network model. The training set is used for the initial fitting of the model and gradient descent of the training error to obtain the weight parameters; the validation set is used to adjust the model's hyperparameters and determine its generalization ability.

[0103] Taking the training loss function and verification loss function of the neural network as verification indicators, the optimal number of hidden layers of the GRU network is first determined, and then the optimal number of hidden layer nodes and the optimal number of training cycles are determined to form a GRU neural network IGBT module life prediction model based on the optimal parameters.

[0104] This step is the core of the present invention, and its purpose is to obtain the optimal parameters of the GRU neural network.

[0105] Among them, the method for determining the optimal number of hidden layers is: using the root mean square error function (RMSE) as the prediction indicator of the IGBT life prediction problem, iterating the same number of times with different numbers of hidden layers and different numbers of hidden layer nodes, comparing the mean root mean square errors of different numbers of layers, and determining the optimal number of hidden layers.

[0106] Choosing different numbers of hidden layers will greatly affect the performance of the neural network, and the choice of the number of layers is generally determined by experiment. The experimental setting of this embodiment is as follows: the number of hidden layers is 1, 2, and 3 layers, totaling 3 parameters, the number of nodes in each hidden layer is 50, 55, 60, 65, 70, 75, 80, 85, 90, 95, and 100, totaling 11 parameters, the number of single training iterations epoch is 5 times, and a loop program is written in Python to test the GRU network model with the above different parameters respectively. The root mean square error is used as an indicator to determine the choice of the number of hidden layers by comparison. Specifically, the optimal number of hidden layers is the one with the smallest mean root mean square error.

[0107] The test results are shown in Figure 7 , Figure 7 yes Figure 6 The root mean square error diagram of different hidden layers of the GRU neural network in . Figure 7 The GRU results are shown in the horizontal axis, where the horizontal axis represents the selection of the 11 different numbers of nodes, and 1, 2, and 3 represent the number of hidden layers. With a single hidden layer, the average RMSE is 7.09E-3, with two hidden layers, the average RMSE is 9.11E-3, and with three hidden layers, the average RMSE is 1.26E-2. This value increases with the number of layers, and the fluctuation of the root mean square error increases with the number of hidden layers. Therefore, a single hidden layer is used for the GRU neural network in this embodiment.

[0108] The optimal number of hidden layer nodes and the optimal number of training cycles are determined by first using an empirical formula to determine all candidate hidden layer node numbers in the model, and then using a grid search method to perform parameter tuning. It should be noted that the grid search method is to calculate the prediction results for each case by traversing all candidate parameter selections.

[0109] The mean square error (MSE) is selected as the loss function. According to the simple cross-validation method, the optimal number of hidden layer nodes and the optimal number of training cycles are determined based on the trend of the graph of training loss and validation loss.

[0110] In addition to the number of hidden layers, the parameters that have a decisive impact on the accuracy of the GRU neural network IGBT module life prediction model include the number of nodes in each layer and the number of single training iterations.

[0111] In application, there is no specific theoretical basis for determining the number of hidden layer neural nodes. Generally, a large number of experiments are conducted according to specific prediction goals and empirical formulas to determine the parameters suitable for this model. The empirical formula is as follows:

[0112]

[0113] Among them, N h is the number of hidden layer neural nodes; N s is the number of samples in the training set; N i is the number of neurons in the input layer; N0 is the number of neurons in the output layer; a is a coefficient, which can be selected according to the model, usually in the range of 2 to 10.

[0114] To accurately determine the parameters and model variations under different parameters, this paper uses a grid search method for parameter optimization. This method iterates through all candidate parameter choices, trying every possible option. The best performing parameter is the final result. The test settings for the optimized parameter experiment are shown in Table 4.

[0115] Table 4 Optimization parameter test settings

[0116]

[0117] Set the number of epochs for cyclic training to 30, substitute the number of hidden layer neural nodes mentioned above, select mean square error (MSE) as the loss function, and use simple cross-validation to observe the changing trends of training loss and validation loss. If the training loss continues to decrease, the validation loss also continues to decrease, indicating that the network is underfitting; if the training loss continues to decrease and the validation loss tends to be stable, it means that the network is overfitting; if the training loss and validation loss tend to coincide and change stably, it means that this parameter is suitable, and the coincidence point is the optimal number of cyclic training. In order to better observe the trend of the line change, the experiment starts recording from epoch = 3, and the results are shown in Figure 8 , Figure 8 yes Figure 6 The loss trend chart of the single-layer GRU neural network with different number of nodes in the GRU neural network is shown in Figure 2. Figure 8 In this embodiment, the details of each area in Figure 8 The 9 areas in the image are locally enlarged, such as Figures 9 to 17 As shown: Figure 9 yes Figure 8 A magnified schematic diagram of area a in the middle; Figure 10 yes Figure 8 A magnified schematic diagram of area b; Figure 11yes Figure 8 Enlarged schematic diagram of area c in the middle; Figure 12 yes Figure 8 Enlarged schematic diagram of the middle d area; Figure 13 yes Figure 8 Enlarged schematic diagram of the middle e area; Figure 14 yes Figure 8 Enlarged schematic diagram of the middle f area; Figure 15 yes Figure 8 Enlarged schematic diagram of the middle g area; Figure 16 yes Figure 8 Enlarged schematic diagram of the h region in the middle; Figure 17 yes Figure 8 Enlarged schematic diagram of the middle i region.

[0118] in accordance with Figure 8 As can be seen, the model is overfitting when the number of nodes is 265, 180, 130, and 105. Models with other node numbers converge and are usable. However, as the number of hidden layer neural nodes increases, the number of optimal cycle training increases, and the time cost also increases accordingly. The time taken is recorded in Table 5 below.

[0119] Table 5 Single-layer GRU network running time and number of records

[0120]

[0121] The optimal model parameters were obtained by selecting the model that converged and used the lowest time cost as the metric. Specifically, Table 5 shows that the single-layer GRU neural network with 60 hidden layer nodes and 23 cycles had the shortest total completion time (138 seconds), representing the optimal model parameters obtained in this example.

[0122] Step 5: Substitute the processed test group data set into the GRU neural network IGBT module life prediction model based on the optimal parameters to test the accuracy and effectiveness of its network model.

[0123] The purpose of step five is to test the progress and effectiveness of the GRU neural network IGBT module life prediction model based on the optimal parameters.

[0124] The test group data set is also preprocessed in step 2 to obtain the processed test group data set. It is imported into the GRU neural network IGBT module life prediction model with optimal parameters to test the model prediction accuracy and adaptability. In this embodiment, in order to test the ability of the GRU neural network IGBT module life prediction model with optimal parameters to predict the remaining life of the IGBT module, 120 continuous saturated on-state voltage drop data are randomly extracted from the test group data set, and then after the preprocessing in step 2, they are substituted into the GRU neural network IGBT module life prediction model with optimal parameters to obtain 60 predicted data. The results are as follows. Figure 18 As shown, Figure 18 yes Figure 1 The prediction results of the IGBT module life prediction model using the GRU neural network with optimal parameters are shown in the figure. The last 60 data points reflect the prediction trend.

[0125] In addition, in order to quantitatively analyze the prediction ability of the GRU neural network IGBT module life prediction model based on the optimal parameters, its parameters such as root mean square error, mean absolute error, and determination coefficient are calculated.

[0126] The specific performance of the model can be analyzed by using the root mean square error (RMSE), mean absolute error (MAE) and determination coefficient R 2 The prediction ability of the parameter quantitative analysis model is as follows:

[0127]

[0128]

[0129]

[0130] Where N is the number of samples, y i is the true value, is the predicted value, is the average value of the true value. RMSE and MAE indicate the degree of deviation between the predicted value and the true value. The closer the result is to 0, the better the model fit is. 2 Evaluate the goodness of fit of the regression model coefficients. The closer the result is to 1, the higher the degree of explanation is and the better the effect of the regression model is.

[0131] In order to eliminate the randomness of the experimental results, multiple experiments were conducted to obtain the average value. The quantitative parameter results are shown in Table 6.

[0132] Table 6 GRU network model performance indicators

[0133]

[0134] According to Table 6, the GRU neural network IGBT module life prediction model has a high accuracy when the optimal parameters are used, and the average RMSE and MAE are less than 0.5%, and the determination coefficient R 2 The goodness of fit is over 99.9%, so the model in this embodiment has an extremely high goodness of fit, which is better than traditional mathematical fitting models and physical simulation models.

[0135] This embodiment also provides an IGBT module life prediction model based on a GRU neural network, which is constructed using the above-mentioned method for constructing an IGBT module life prediction model based on a GRU neural network. This embodiment also provides an IGBT module life prediction method based on a GRU neural network, which uses the above-mentioned IGBT module life prediction model based on a GRU neural network. Specifically, the IGBT module life prediction method based on a GRU neural network includes: obtaining the on-state voltage drop of the IGBT module, importing the IGBT module life prediction model based on the GRU neural network, and obtaining the aging degree of the IGBT module, which also characterizes the remaining life of the IGBT module. Based on the above verification, it can be seen that the aging degree of the IGBT module can be accurately obtained using this life prediction model.

[0136] Of course, it should be noted that the IGBT module used in this embodiment is the IKW30N65H5 module, so the life prediction model established is also for the IKW30N65H5 module. If other signals are to be predicted, it is necessary to rely on the above model establishment method to establish a corresponding life prediction model.

[0137] Example 2

[0138] This embodiment discloses a readable storage medium, in which computer program instructions are stored. When the computer program instructions are read and executed by a processor, the method for constructing the IGBT module life prediction model based on the GRU neural network is executed.

[0139] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0140] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A method for constructing an IGBT module life prediction model based on a GRU neural network, which is used to construct a GRU neural network IGBT module life prediction model with optimal parameters for the IGBT module, characterized in that: The construction method comprises the following steps: Step 1: Perform a power cycle aging test on the IGBT module to obtain aging characteristic data of the IGBT module, and obtain an experimental group data set and a test group data set; Step 2: Preprocess the IGBT module aging characteristic data to obtain a processed experimental group data set and a processed test group data set; wherein the processed experimental group data set is divided into a training set and a validation set in a specific ratio; The pretreatment method comprises: Step 2.1, use the SG filtering method to filter the IGBT module aging characteristic data; Step 2.2, normalizing the filtered IGBT module aging characteristic data; Step 2.3: Use a sliding window method to sequentially select values ​​from left to right on the preprocessed IGBT module aging characteristic quantity data according to a certain window size. The values ​​within the window are used as the input sequence X, and the first value on the right side of the value window is used as the output sequence Y. Establish a mapping relationship between the aging characteristic quantity data from input to output; Step 3: Determine the GRU neural network framework and preliminarily build the GRU neural network model; Step 4: Use the training set and validation set to train the GRU neural network model; using the training loss function and validation loss function of the neural network as validation indicators, first determine the optimal number of hidden layers of the GRU network, then determine the optimal number of hidden layer nodes and the optimal number of training cycles, and form a GRU neural network IGBT module life prediction model based on the optimal parameters; The method for determining the optimal number of hidden layers is as follows: The root mean square error function is used as a prediction indicator for the IGBT life prediction problem. Different numbers of hidden layers and different numbers of hidden layer nodes are iterated the same number of times, and the mean root mean square error of different numbers of layers is compared to determine the optimal number of hidden layers. The method for determining the optimal number of hidden layer nodes and the optimal number of training cycles is: First, an empirical formula is used to determine the number of all candidate hidden layer nodes in the model, and then a grid search method is used to tune the parameters; the grid search method is to calculate the prediction results for each case by looping through all candidate parameter selections; The empirical formula is: ; N h is the number of hidden layer neural nodes; N s is the number of samples in the training set; N i is the number of neurons in the input layer; N 0 is the number of neurons in the output layer; a is the coefficient; The mean square error is selected as the loss function. Based on the simple cross-validation method, the optimal number of hidden layer nodes and the optimal number of training cycles are determined by the trend of the graph of training loss and validation loss. Step 5: Substitute the processed test group data set into the GRU neural network IGBT module life prediction model based on the optimal parameters to verify the accuracy and effectiveness of its network model.

2. The method for constructing an IGBT module life prediction model based on a GRU neural network according to claim 1, characterized in that: Step one includes: The K factor value is obtained by measuring the characteristic diagram of the IGBT module using a power semiconductor K factor tester; Select the control method of power cycle aging test and determine the parameters for power aging test; After each cycle is completed, a set of IGBT module aging characteristic quantity data is collected to obtain aging characteristic quantity data and power cycle number; The aging characteristic data include module on-state voltage drop data, junction temperature data, case temperature data, and thermal resistance data.

3. The method for constructing an IGBT module life prediction model based on a GRU neural network according to claim 2, characterized in that: Step 2 also includes: Step 2.4: Split the input and output sequences of the experimental group data set into training sets and validation sets according to the corresponding proportions.

4. The method for constructing an IGBT module life prediction model based on a GRU neural network according to claim 3, characterized in that: Step three includes: Use TensorFlow.Keras as the framework and Python as the programming language to write the algorithm and create a GRU neural network; The GRU neural network includes: A sequence input layer, used to import the IGBT module aging characteristic data input sequence into the network; a hidden layer for learning high-level features of IGBT module cycle aging test time series data through input weights, recursive weights, and bias weights; the hidden layer includes several groups of sub-hidden layers, each group of sub-hidden layers includes several GRU layers with a configurable number of hidden layer nodes and a Dropout layer with a set initial dropout probability; and A fully connected layer and a regression output layer are used to output the predicted value of the IGBT module aging feature quantity predicted by the GRU network.

5. The method for constructing an IGBT module life prediction model based on a GRU neural network according to claim 4, characterized in that: Step five includes: The processed validation group data set is imported into the GRU neural network IGBT module life prediction model with optimal parameters to test the model prediction accuracy and adaptability.

6. The method for constructing an IGBT module life prediction model based on a GRU neural network according to claim 5, characterized in that: Step five also includes: The prediction ability of the GRU neural network IGBT module life prediction model based on optimal parameters is quantitatively analyzed using root mean square error, mean absolute error, and determination coefficient.

7. The optimal parameter GRU neural network IGBT module life prediction model is characterized by: It is constructed using the method for constructing an IGBT module life prediction model based on a GRU neural network as described in any one of claims 1-6.

8. The IGBT module life prediction method based on GRU neural network is characterized by: It uses the GRU neural network IGBT module life prediction model with optimal parameters as described in claim 7; The IGBT module life prediction method based on the GRU neural network includes: The on-state voltage drop of the IGBT module is obtained, and the GRU neural network IGBT module life prediction model with the optimal parameters is imported to obtain the aging degree of the IGBT module.

9. A readable storage medium, characterized in that: The readable storage medium stores computer program instructions. When the computer program instructions are read and executed by a processor, the method for constructing an IGBT module life prediction model based on a GRU neural network according to any one of claims 1 to 6 is executed.

Images