A method and system for monitoring the state of an insulated gate bipolar transistor

Abstract

This application relates to a state monitoring method and system for insulated gate bipolar transistors (IGBTs). The method includes: standardizing peak voltage time-series data and dividing the dataset; performing adaptive variational mode decomposition (VMD) on the training set sequence to generate intrinsic mode function (IMF) components and residual components; performing wavelet packet denoising on high-frequency noise components based on zero-crossing rate and center frequency; weighting all processed components with attention weights based on energy proportion; constructing a global training set and validation set based on the weighted component sequence and a preset time window; optimizing an LSTM network through PSO coarse search and Bayesian fine-tuning of hyperparameters to obtain an LSTM feature extraction layer, and training the model with an SVR regression prediction layer to obtain a prediction model; inputting the test set into the prediction model to output a future standardized peak voltage prediction sequence; and determining the first sampling period below a preset failure threshold based on a preset failure threshold to generate the predicted failure period. This method can effectively handle the non-stationarity and noise interference of peak voltage sequences, improving prediction accuracy and the reliability of lifetime assessment.

Description

Technical Field

[0001] This invention belongs to the field of power electronics, and in particular relates to a state monitoring method and system for an insulated gate bipolar transistor. Background Technology

[0002] Insulated Gate Bipolar Transistor (IGBT) As a core switching device in power electronic systems, it is widely used in new energy vehicles, rail transit, smart grids and other fields. During long-term operation, factors such as temperature cycling and electrical stress can cause performance degradation, eventually leading to failure and system malfunction. Peak voltage ( As a reflection Key indicators of degradation status, and their changing trends are directly related to device lifetime. Therefore, achieving... Accurate prediction and lifetime assessment of peak voltage are of great significance for ensuring the reliability of power electronic systems and reducing maintenance costs.

[0003] existing Lifetime prediction methods mainly include physical model-based methods and data-driven methods. Physical model-based methods require in-depth analysis. While complex mathematical models can be established to understand the degradation mechanism, they struggle to fully account for the multi-factor interference in actual operation, limiting their applicability. Data-driven methods, on the other hand, do not require explicit understanding of the degradation mechanism; they achieve prediction by mining patterns in historical data. However, traditional data-driven methods (such as support vector machines, ...) Neural networks struggle to effectively handle the nonlinear and non-stationary characteristics of peak voltage sequences, resulting in low prediction accuracy.

[0004] Variational Mode Decomposition As a signal processing technique, it can decompose non-stationary signals into multiple stationary intrinsic mode functions. Effectively extracts different frequency features from signals; Long Short-Term Memory Network It has the ability to process long-series data and capture long-term dependencies in time series; Support Vector Regression Machine It exhibits good generalization ability in small-sample, nonlinear regression problems. Combining them can fully leverage the advantages of all three and enhance their effectiveness. Voltage prediction accuracy is important, but a complete system based on [the following] is currently lacking. Peak voltage prediction and lifetime assessment schemes are difficult to meet the needs of actual engineering. Summary of the Invention

[0005] Therefore, it is necessary to provide a state monitoring method and system for insulated gate bipolar transistors (IGBTs) that can predict peak voltage and assess lifetime based on SVMD-LSTM-SVR, addressing the aforementioned technical problems.

[0006] In a first aspect, this application provides a state monitoring method for an insulated-gate bipolar transistor, comprising:

[0007] S1. Standardize the peak voltage timing data of the insulated gate bipolar transistor to generate a standardized voltage sequence; the standardized voltage sequence is divided into training set, validation set and test set according to time order;

[0008] S2. Perform data-adaptive variational mode decomposition on the standardized voltage sequences in the training set to generate intrinsic mode function components and a residual component;

[0009] S3. Based on the zero-crossing rate and center frequency, wavelet packet denoising is performed on the high-frequency noise of the intrinsic mode function components to obtain all processed components; the processed components include the denoised high-frequency components, the mid-to-low frequency intrinsic mode function components, and a residual component.

[0010] S4. Perform attention weighting calculation based on energy proportion on all processed components to generate a weighted feature component sequence;

[0011] S5. Construct time series samples based on weighted feature component sequences and preset time windows to generate a global training set and a global validation set;

[0012] S6. Train the LSTM network feature extraction layer and SVR regression prediction layer based on the global training set and global validation set to obtain the prediction model; optimize the LSTM network through a two-stage hyperparameter optimization using PSO and Bayesian methods.

[0013] S7. Input the test set into the prediction model to make predictions and output the future standardized peak voltage prediction sequence.

[0014] S8. Based on the preset failure threshold, determine the sampling period where the first voltage value in the actual peak voltage prediction sequence is lower than the preset failure threshold, and generate the predicted failure period of the insulated gate bipolar transistor.

[0015] In one embodiment, S2 includes:

[0016] S21. Set the standardized voltage sequence in the training set as the current residual signal;

[0017] S22. Perform power spectrum analysis on the current residual signal to generate the corresponding power spectral density distribution;

[0018] S23. Identify the frequency corresponding to the maximum power in the power spectral density distribution and generate the initial center frequency of the current mode to be extracted;

[0019] S24. Based on the triple constraints of modal concentration, spectral overlap and modal independence, variational mode extraction iterative optimization is performed on the current residual signal containing the initial center frequency to generate an intrinsic mode function component.

[0020] S25. Subtract the generated intrinsic mode function components from the current residual signal to generate the updated current residual signal;

[0021] S26. Using the updated current residual signal as input, repeat S22-S25 until the updated current residual signal has no effective oscillation modes that can be extracted or the number of extracted modes reaches the preset maximum mode number threshold, and generate intrinsic mode function components and a residual component sorted by frequency from high to low.

[0022] In one embodiment, S3 includes:

[0023] S31. Calculate the zero-crossing rate of the intrinsic mode function components, generating the zero-crossing rate values ​​for each component. The zero-crossing rate is calculated according to the formula:

[0024]

[0025] in, Let be the zero-crossing rate of each component, and IMF be the intrinsic mode function component. for The length of the sequence, This represents the difference in sign changes across IMF sequences.

[0026] S32. Based on the center frequencies determined by each component during the adaptive variational mode decomposition of data, the condition that the zero-crossing rate is greater than the first threshold and the center frequency is greater than the second threshold is determined, and a high-frequency noise component identifier is generated.

[0027] S33. For the component specified by the high-frequency noise component identifier, perform multi-level wavelet packet decomposition through wavelet packet transform to generate high-frequency data packet coefficients for each sub-band.

[0028] S34. Perform soft threshold quantization on the coefficients of high-frequency data packets based on dynamic thresholds to generate coefficients after threshold processing;

[0029] S35. Perform inverse wavelet packet reconstruction on the coefficients after thresholding to generate denoised high-frequency components, and merge the denoised high-frequency components, the mid-to-low frequency intrinsic mode function components not identified as high-frequency noise, and a residual component into all processed components.

[0030] In one embodiment, S4 includes:

[0031] S41. Calculate the energy percentage of all processed components and generate the energy value of each component.

[0032] S42. Based on the energy values ​​of each component, attention weights are assigned to generate attention weights for each component; the formula for calculating the attention weights is:

[0033]

[0034] in, Let i be the attention weight of the i-th component. , For the first j processed IMF sequences;

[0035] S43. Based on the attention weights, perform weighted scaling on the sequence of each component to generate a weighted feature component sequence.

[0036] In one embodiment, S6 includes:

[0037] S51. Using the root mean square error of prediction on the global validation set as the fitness function, perform PSO search on the number of hidden layer units, learning rate, batch size and dropout rate of the LSTM network in the preset hyperparameter space to generate a coarse-optimized hyperparameter combination.

[0038] S52. Centered on the coarse optimization hyperparameter combination, the preset hyperparameter space is compressed to generate a local fine search space.

[0039] S53. Within the local fine search space, the hyperparameters of the LSTM network are tuned using a Bayesian optimization method with random forest as the surrogate model and expected improvement as the acquisition function, to generate the final hyperparameter combination.

[0040] S54. Configure the LSTM network with the final hyperparameter combination to obtain the optimized LSTM feature extraction layer;

[0041] S55. The input data of the global training set is extracted through the LSTM feature extraction layer to generate a high-dimensional temporal feature vector;

[0042] S56. Based on high-dimensional time-series feature vectors, Bayesian optimization is performed on the penalty parameter, error tolerance parameter, and kernel function scaling parameter of the SVR regression prediction layer to generate a prediction model.

[0043] In one embodiment, S54 includes:

[0044] S541. Based on the number of hidden layer units and the dropout rate in the final hyperparameter combination, construct the network structure of the LSTM feature extraction layer; the network structure includes a sequence input layer, an LSTM layer, a dropout layer, and a fully connected layer.

[0045] S542. Initialize the weight parameters of the network structure to obtain the optimized LSTM feature extraction layer.

[0046] In one embodiment, S7 includes:

[0047] S61. Based on the test set, construct the initial model input vector according to the same time window rule used in the global training set;

[0048] S62. Input the initial model input vector into the prediction model for single-step forward calculation to generate the standardized peak voltage prediction value at the next sampling time.

[0049] S63. The newly generated standardized peak voltage prediction value is used as a new data point and fed back to the end of the historical data sequence. At the same time, the oldest data point is removed to generate an updated historical sequence.

[0050] S64. Based on the updated historical sequence, construct a new model input vector according to the same time window rule;

[0051] S65. Using the new model input vector as input, iteratively execute S62-S64 to generate a future standardized peak voltage prediction sequence of a specified length.

[0052] In one embodiment, S8 includes:

[0053] S71. Obtain the mean and standard deviation calculated when standardizing the original peak voltage time series data;

[0054] S72. Based on the mean and standard deviation, perform an inverse standardization transformation on the future standardized peak voltage prediction sequence to generate the actual peak voltage prediction sequence; the inverse standardization transformation calculation formula is:

[0055]

[0056] in, This is the actual voltage value. For standardized voltage values, , These are the mean and standard deviation of the standardized voltage values, respectively.

[0057] S73. Compare each actual voltage value in the actual peak voltage prediction sequence with the preset failure threshold in sequence, and identify the index position of the first voltage value in the actual peak voltage prediction sequence that is lower than the preset failure threshold.

[0058] S74. Determine the corresponding future sampling period based on the index position to generate the predicted failure period of the insulated gate bipolar transistor.

[0059] In one embodiment, after S6, the following is also included:

[0060] S81. The prediction model performs forward prediction on the input data of the global training set and the global validation set respectively to generate the corresponding standardized voltage prediction value.

[0061] S82. Perform denormalization on the generated standardized voltage prediction value to generate the corresponding actual voltage prediction value.

[0062] S83. Based on the actual voltage prediction value and the corresponding real voltage value, calculate the prediction performance index for the training and validation phases respectively, and generate performance evaluation results; the performance index includes root mean square error and coefficient of determination, and the formula for calculating root mean square error is:

[0063]

[0064] Where RMSE is the root mean square error, and n is the sample size. Let i be the true value of the i-th sample. Let be the predicted value for the i-th sample;

[0065] The formula for calculating the coefficient of determination is:

[0066]

[0067] in, As the coefficient of determination, The mean of the true values;

[0068] S84. Compare and analyze the performance evaluation results with the pre-stored benchmark model performance results to generate a model performance comparison report; the benchmark models include a single long short-term memory network model and a combination model of variational mode decomposition and long short-term memory network.

[0069] Secondly, this application also provides a state monitoring system for an insulated gate bipolar transistor, comprising:

[0070] The voltage sequence normalization module is used to normalize the peak voltage timing data of insulated gate bipolar transistors to generate a normalized voltage sequence. The normalized voltage sequence is divided into training set, validation set and test set according to time sequence.

[0071] The variational mode decomposition module is used to perform data-adaptive variational mode decomposition on the standardized voltage sequences in the training set, generating intrinsic mode function components and a residual component.

[0072] The high-frequency noise removal module is used to perform wavelet packet denoising on the high-frequency noise of the intrinsic mode function components based on the zero-crossing rate and center frequency, to obtain all processed components. The processed components include the denoised high-frequency components, mid- and low-frequency intrinsic mode function components, and a residual component.

[0073] The weighted feature calculation module is used to perform attention weight calculation based on energy proportion on all processed components to generate a weighted feature component sequence;

[0074] The time series sample construction module is used to construct time series samples based on weighted feature component sequences and preset time windows, and generate a global training set and a global validation set.

[0075] The prediction model training module is used to train the LSTM network feature extraction layer and SVR regression prediction layer based on the global training set and global validation set to obtain the prediction model; the LSTM network is optimized through a two-stage hyperparameter optimization using PSO and Bayesian methods.

[0076] The voltage prediction module is used to input the test set into the prediction model for prediction and output a future standardized peak voltage prediction sequence.

[0077] The failure cycle prediction module is used to determine the sampling period in the actual peak voltage prediction sequence where the first voltage value is lower than the preset failure threshold based on a preset failure threshold, and to generate the predicted failure cycle of the insulated gate bipolar transistor.

[0078] The aforementioned state monitoring method and system for insulated-gate bipolar transistors (IGBTs) involves sequentially standardizing and partitioning the peak voltage time-series data of the IGBTs. Then, adaptive variational mode decomposition (VMD) is applied to the training set data to adaptively decompose the non-stationary voltage sequence into automatically generated intrinsic mode function (IMF) components of different frequencies and a residual component. The high-frequency noise components obtained from the decomposition are denoised, and all processed components are weighted by attention weights based on their energy proportions to enhance key trend characteristics representing device degradation. Finally, a time series is constructed based on the weighted feature component sequence. The method generates a global dataset for model training by taking samples and inputting it into an LSTM feature extraction layer optimized by PSO and Bayesian two-stage hyperparameters. The model is then trained by combining the LSTM feature extraction layer with an SVR regression prediction layer to obtain a prediction model. The prediction model is used to predict the test set and outputs a future standardized peak voltage prediction sequence. The lifetime is then assessed by comparing the prediction with a preset failure threshold. This method effectively overcomes the problems of noise and trend mixing and strong non-stationarity in the original signal, and significantly improves the prediction accuracy and stability of the peak voltage change trend, thereby achieving a more reliable and objective assessment of the remaining lifetime of insulated gate bipolar transistors. Attached Figure Description

[0079] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0080] Figure 1 A schematic flowchart illustrating a state monitoring method for an insulated gate bipolar transistor provided by the present invention;

[0081] Figure 2 A schematic diagram of the original peak voltage degradation of a state monitoring method for an insulated gate bipolar transistor provided by the present invention;

[0082] Figure 3 The original peak voltage of the state monitoring method for an insulated gate bipolar transistor provided by this invention is... Schematic diagram of decomposition results;

[0083] Figure 4 A schematic diagram of the noise reduction result of the original IGBT peak voltage after high-frequency component processing for a state monitoring method of an insulated gate bipolar transistor provided by the present invention;

[0084] Figure 5 IGBT data iteration diagram of the state monitoring method for an insulated gate bipolar transistor provided by the present invention. Figure 1 ;

[0085] Figure 6 IGBT data iteration diagram of the state monitoring method for an insulated gate bipolar transistor provided by the present invention. Figure 2 ;

[0086] Figure 7 A schematic diagram of IGBT peak voltage drop prediction trained by a hybrid model for a state monitoring method of an insulated gate bipolar transistor provided by the present invention;

[0087] Figure 8 A schematic diagram illustrating the goodness of fit between predicted and actual values ​​in a state monitoring method for an insulated gate bipolar transistor provided by the present invention.

[0088] Figure 9 The state monitoring method for an insulated gate bipolar transistor provided by this invention is based on model prediction. Schematic diagram of peak voltage drop residual distribution;

[0089] Figure 10 The model-trained method for state monitoring of an insulated gate bipolar transistor provided by this invention Schematic diagram of peak pressure drop trend prediction;

[0090] Figure 11 This is a schematic diagram of the state monitoring system for an insulated gate bipolar transistor provided by the present invention. Detailed Implementation

[0091] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0092] In one embodiment, such as Figure 1 As shown, a state monitoring method for an insulated gate bipolar transistor (IGBT) is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps S1 to S8:

[0093] S1. Standardize the peak voltage timing data of the insulated gate bipolar transistor to generate a standardized voltage sequence; the standardized voltage sequence is divided into training set, validation set and test set according to time order.

[0094] Optionally, standardization is used to eliminate dimensional interference, and the dataset is divided into reasonable proportions to provide high-quality, unbiased input data for model training and validation. First, the IGBT peak voltage degradation data stored in the specified path is read. The data format is a CSV file, containing two columns of core information: the first column is the sampling period, representing the time dimension and reflecting the order of data acquisition; the second column is the peak voltage. As a key electrical indicator reflecting the degradation state of IGBTs, outliers are removed using the industry-standard 3σ criterion to ensure data validity. The specific steps are as follows: First, calculate the mean of the peak voltage sequence. and standard deviation If a certain peak voltage data point satisfies | - | > 3× If a data point is found to be outlier (caused by measurement error or sudden change in electrical stress), it is removed. Z-score standardization is then performed, using the following formula: =( - ) / ,in This is the original peak voltage sequence. The mean of the peak voltage sequence. The standard deviation of the peak voltage sequence is used to eliminate the influence of dimensions on voltage values ​​of different orders of magnitude, allowing model training to focus on the degradation trend of voltage rather than absolute numerical magnitude, thus improving model convergence speed and prediction stability. Finally, the standardized voltage sequence is divided into training, validation, and test sets in chronological order, strictly adhering to a 7:2:1 ratio. The training set accounts for 70% and is used for the model to learn the degradation pattern of IGBT peak voltage; the validation set accounts for 20% and is used for hyperparameter optimization and overfitting monitoring during training; the test set accounts for 10% and is used for final evaluation of the model's generalization ability. The partitioning process strictly follows chronological order, extracting data from the beginning to the end, ensuring that the training, validation, and test sets correspond to degradation data at different stages of IGBT operation, preventing data from future periods from being leaked into historical training data, ensuring the logical rationality of model training and evaluation, and adapting to the data volume requirements of models in small sample scenarios.

[0095] S2. Perform data-adaptive variational mode decomposition on the standardized voltage sequences in the training set to generate intrinsic mode function components and a residual component.

[0096] Optionally, SVMD decomposition is used to perform hierarchical processing of non-stationary voltage sequences, decomposing complex degenerate signals into modal components of different frequency dimensions. First, the current residual is initialized as a standardized voltage sequence from the training set, and a maximum mode number threshold is set to avoid over-decomposition leading to modal redundancy and feature confusion. The iterative decomposition process is executed according to the following logic: First, the pwelch function is used to perform power spectrum analysis on the current residual. This function calculates the distribution of signal power with frequency by segmented windowing (using the Hanning window function), locates the frequency corresponding to the maximum power, and converts it to an angular frequency, which is used as the center frequency of the mode to be extracted, ensuring that the extracted mode focuses on the main energy distribution range of the signal; Second, iterative optimization is performed based on the triple constraints of "modal concentration, spectral overlap, and modal independence," where modal concentration requires that the extracted modes fluctuate around the center frequency without significant deviation; frequency... Spectral overlap requires that the spectrum of the current mode and the remaining residual have no overlapping regions to avoid feature crossing; modal independence requires that the newly extracted mode does not interfere with the extracted modes to ensure the uniqueness of each modal feature. The maximum number of iterations during the iteration process is set to 500 times, and the convergence threshold is set to 1e-6. When the iteration error is less than the convergence threshold, the extraction of the mode is stopped; the third step is to update the current residual to "original residual - extracted IMF component" and repeat the above two steps until there are no effective oscillating modes to be extracted in the current residual (i.e. the residual signal tends to be stable) or the number of extracted modes reaches the preset maximum number of modes threshold. The final decomposition yields N Intrinsic Mode Functions (IMFs) and 1 residual component. All IMFs are sorted from high to low frequency. The high-frequency IMFs mainly reflect the high-frequency electromagnetic noise and short-term random fluctuations during IGBT operation, the mid-frequency IMFs reflect the mid-term voltage fluctuations caused by the periodic changes in operating conditions, and the low-frequency IMFs and residual component reflect the degradation trend of IGBTs caused by long-term temperature cycling and electrical stress accumulation. This achieves effective layering and feature stripping of non-stationary degradation signals.

[0097] S3. Based on the zero-crossing rate and center frequency, perform wavelet packet denoising on the high-frequency noise of the intrinsic mode function components to obtain all processed components; the processed components include the denoised high-frequency components, the mid-to-low frequency intrinsic mode function components, and a residual component.

[0098] Optionally, noise suppression is performed on the high-frequency IMF components after SVMD decomposition to prevent high-frequency interference from masking the true degradation trend of the IGBT. First, the zero-crossing rate of each IMF component is calculated. The formula for calculating the zero-crossing rate is: ,in for The sign function of a sequence returns 1 for a positive value, -1 for a negative value, and 0 for zero. The diff function is used to calculate the difference between two adjacent signs. Used to count the number of sign changes in a sequence, i.e., the number of times it crosses zero. for The longer the sequence, the higher the zero-crossing rate, indicating more frequent signal fluctuations and a greater likelihood of containing invalid high-frequency noise. Combined with various... The center frequency of the component determined during the SVMD decomposition process is determined using the dual conditions of "zero crossing rate > 0.05 and center frequency > 0.1Hz" to identify high-frequency noise components. This criterion is based on a large amount of IGBT accelerated aging test data and can accurately distinguish between high-frequency noise and effective signal components. For IMF components identified as high-frequency noise, wavelet packet transform is used for denoising. The specific process is as follows: First, the db4 wavelet basis is used to perform two-level decomposition of the high-frequency modes using the wpdec function. The db4 wavelet basis has good time-frequency localization characteristics and can focus signal features in both the time and frequency domains. The two-level decomposition can divide the high-frequency components into four data packets with different frequency bandwidths, achieving frequency separation between noise and effective signals. Second, the coefficients of the decomposed high-frequency data packets are subjected to soft thresholding using the wthresh function. The threshold is set to 0.05 × the maximum value of the data packet coefficients. This dynamic thresholding strategy can adaptively match the noise intensity of different high-frequency components, avoiding the loss of effective signals or noise residue caused by a fixed threshold. Third, the coefficients after thresholding are reconstructed using inverse wavelet packet transform (implemented by the wprec function) to generate denoised high-frequency IMF components, significantly reducing the interference of high-frequency noise on subsequent model training. For IMF components with a zero-crossing rate ≤ 0.05 and a center frequency ≤ 0.1Hz, they are determined to be stationary components (including mid-frequency periodic fluctuations and low-frequency degradation trends). No wavelet packet denoising is required, and they are directly retained for subsequent feature construction. The denoised high-frequency components, mid- and low-frequency intrinsic mode function components, and a residual component constitute all the processed components.

[0099] S4. Perform attention weighting calculation based on energy proportion on all processed components to generate a weighted feature component sequence.

[0100] Optionally, attention weights can be allocated based on energy proportions to strengthen the low-frequency degradation characteristics crucial for IGBT lifetime prediction, weaken the influence of low-energy noise components, and improve the model's ability to capture core trends. First, the energy of each processed IMF component (including denoised high-frequency and mid-to-low-frequency IMFs) and residual component is calculated. The energy calculation uses the sum of squares of each element in the sequence, i.e. ,in For the i-th processed IMF sequence, the core logic of this calculation method is: signal energy is positively correlated with the strength of the effective information contained in that component; the higher the energy, the stronger the correlation between that component and the IGBT degradation state. Attention weights are assigned based on the energy values ​​of each component, and the weight calculation formula is: Where j ranges from 1 to M, and M is the total number of components, including all IMF components and one residual component. That is, the attention weight of a single component is equal to the proportion of its own energy to the total energy of all components. This allocation rule allows the model to automatically focus on high-energy effective feature components, such as low-frequency IMFs and residual components reflecting long-term degradation trends, while weakening the contribution of low-energy noise components (such as denoised high-frequency IMFs). Finally, all data points in each component sequence are multiplied by their corresponding attention weights to complete the weighted scaling process, generating a weighted feature component sequence.

[0101] S5. Construct time series samples based on weighted feature component sequences and preset time windows to generate a global training set and a global validation set.

[0102] Optionally, by constructing time-series samples adapted to the LSTM network, the long- and short-term dependencies of the peak voltage sequence can be captured, providing a data format that meets the network input requirements for the model to learn time-series features. First, the time series length is set ( The length is set to 10, determined based on the degradation characteristics of IGBT peak voltage. This length covers the correlation information in short-term fluctuations while avoiding the surge in model training complexity and gradient vanishing problems caused by excessively long sequences. For each weighted IMF component and residual component, time-series samples are constructed using a sliding window method, starting from the first data point at the beginning of the sequence: 10 consecutive data points are used as input samples, i.e. Where i is the sample index, and the 11th data point is used as the output sample, i.e. The entire component sequence is iterated sequentially until no complete sample can be constructed at the end of the sequence, ensuring that each sample reflects the temporal correlation between the voltage values ​​at the first 10 time points and the voltage value at the 11th time point. To ensure the effectiveness of model validation, a validation set is divided from the temporal dataset of each component at a ratio of 20%, truncated from the end of the dataset, consistent with the data distribution logic of "predicting future states based on historical data" in actual engineering. The remaining 80% is used as the training set. Finally, the training set inputs, training set outputs, validation set inputs, and validation set outputs of all components are merged to obtain the global training set and global validation set. To avoid model overfitting and without disrupting the time step order within a single sample, the randperm function is used to only shuffle the order between samples in the global training set, while keeping the 10 time step data within a single sample continuous, ensuring that the model can learn the true temporal dependencies.

[0103] S6. Based on the global training set and global validation set, train the LSTM network feature extraction layer and SVR regression prediction layer to obtain the prediction model; the LSTM network is optimized through a two-stage hyperparameter optimization using PSO and Bayesian methods.

[0104] Optionally, Particle Swarm Optimization (PSO) is used for the first stage of coarse hyperparameter search. This involves simulating the collaborative behavior of the particle swarm and the individual learning behavior to search for the optimal solution. Each particle corresponds to a set of Long Short-Term Memory (LSTM) hyperparameters, including the number of hidden layer units, learning rate, batch size, and dropout rate. Using the prediction error of the global validation set as the fitness function, after initializing the particle swarm, the position and velocity of each particle are iteratively updated according to preset rules. During iteration, each particle retains its own optimal position information while also moving towards the optimal position of the swarm. After iteration, the globally optimal hyperparameter combination is selected as the coarse search result. Further, based on the coarse search result, the hyperparameter search space is compressed, and Bayesian optimization is used for the second stage of fine tuning. A surrogate model is used to predict the model performance corresponding to the hyperparameter combination, and a sampling function guides the hyperparameter sampling direction, gradually approaching the optimal hyperparameters and finally determining the hyperparameter configuration of the LSTM network.

[0105] An LSTM network feature extraction layer is constructed based on the final hyperparameters. The input layer dimension is explicitly matched to the global training set sample dimension. Temporal dependencies are captured through internal forget gates, input gates, and output gates. An appropriate optimizer is selected to iteratively update the network parameters. During training, model performance is monitored using a global validation set to avoid overfitting. After training, the global training set is input into the LSTM network feature extraction layer, which outputs a high-dimensional temporal feature vector. This vector condenses the key temporal information of the weighted feature component sequences. The high-dimensional temporal feature vector is used as input to train a Support Vector Regression (SVR) model. For example, a radial basis function is chosen as the kernel function to map the feature vector to a high-dimensional feature space. The SVR model parameters are then optimized based on the global training set, and the regression accuracy is verified using a global validation set. Finally, the trained LSTM network feature extraction layer and the SVR regression prediction layer are connected in the input-output logical order to form a complete prediction model.

[0106] S7. Input the test set into the prediction model for prediction, and output the future standardized peak voltage prediction sequence.

[0107] Optionally, based on the trained prediction model, a rolling window prediction strategy is adopted to generate standardized peak voltage sequences for multiple future sampling periods, providing data support for lifetime assessment. First, based on the test set data, an initial model input vector is constructed according to the time window rules of the global training set (sequence length 10), ensuring that the dimension and temporal structure of the input data are completely consistent with the input format during model training, avoiding prediction errors caused by input mismatch. The initial input vector is input into a Long Short-Term Memory (LSTM) feature extractor for single-step forward computation. An LSTM layer captures the temporal dependencies of the 10 time steps in the input vector, extracting high-dimensional temporal features. The prediction model then performs nonlinear regression on the features, outputting the standardized peak voltage prediction value for the next sampling moment. A rolling update mechanism is used to update the input vector: the newly generated standardized prediction value is used as a new data point and fed back to the end of the historical data sequence. Simultaneously, the oldest data point in the sequence is removed, generating an updated historical sequence. This mechanism simulates the application scenario of "predicting the next state based on the latest collected data" in actual engineering, improving the timeliness and accuracy of the prediction results. Based on the updated historical sequence, a new model input vector is reconstructed according to the same time window rules. The single-step prediction and sequence update operations are iteratively executed until a future standardized peak voltage prediction sequence of a specified length is generated. The number of future prediction steps is preset to 20~30, which can cover the typical remaining life range of IGBTs and balance prediction accuracy and computational efficiency.

[0108] S8. Based on the preset failure threshold, determine the sampling period where the first voltage value in the actual peak voltage prediction sequence is lower than the preset failure threshold, and generate the predicted failure period of the insulated gate bipolar transistor.

[0109] Optionally, by comparing the inverse normalization transformation with the failure threshold, a quantitative assessment of IGBT life can be achieved, providing a clear basis for the formulation of power electronic system maintenance plans. First, the mean of the peak voltage sequence calculated during the normalization process is obtained. with standard deviation Based on this mean and standard deviation, the future standardized peak voltage prediction sequence is inversely standardized. The inverse standardization formula is as follows: ,in This is the actual peak voltage value. To standardize the voltage values, the core purpose of this conversion is to restore the dimensionless standardized values ​​output by the model to actual voltage values ​​that can be directly determined in engineering. A preset peak voltage failure threshold for the IGBT is established. This threshold needs to be determined based on the rated parameters in the IGBT device datasheet or the failure judgment criteria in engineering applications, ensuring that the threshold is consistent with the actual failure judgment logic of the device under operating conditions. The `find` function is used to traverse the predicted actual peak voltage sequence after destandardization, sequentially comparing the actual voltage value of each sampling period with the preset failure threshold. The process stops when the first index position where the voltage value is lower than the failure threshold is identified; the future sampling period corresponding to this index position is the predicted failure period of the IGBT. If no voltage value lower than the failure threshold is found within the preset number of future prediction steps, the number of future prediction steps needs to be extended until the threshold crossover point is found, ensuring the completeness and practicality of the lifetime assessment.

[0110] The aforementioned state monitoring method for insulated gate bipolar transistors (IGBTs) utilizes IGBT peak voltage data preprocessing, SVMD adaptive decomposition, high-frequency component denoising and attention weighting, time-series sample construction, LSTM-SVR training (two-stage hyperparameter optimization), future voltage prediction, and failure cycle determination. This enables layered signal extraction, precise noise suppression, key feature enhancement, and LSTM-SVR collaborative modeling, effectively addressing the nonlinear and non-stationary characteristics of voltage sequences. It significantly improves the accuracy of peak voltage prediction and the reliability of life assessment, providing a scientific basis for power system maintenance.

[0111] In one embodiment, S2 includes:

[0112] S21. Set the standardized voltage sequence in the training set as the current residual signal.

[0113] Optionally, the initial conditions and constraints of SVMD decomposition are clearly defined, and the standardized voltage sequence of the training set is directly used as the current residual signal without additional smoothing or filtering. This ensures that the decomposition is based on the original preprocessed data and guarantees the authenticity of the signal features. Simultaneously, a maximum mode number threshold is set. The core purpose is to avoid excessive decomposition leading to mode redundancy and to prevent information overlap or invalid calculations in subsequent feature processing. This threshold is determined based on the frequency distribution characteristics of the IGBT peak voltage sequence and engineering experience, adapting to the hierarchical requirements of non-stationary signals.

[0114] S22. Perform power spectrum analysis on the current residual signal to generate the corresponding power spectral density distribution.

[0115] Optionally, the pwelch function is used to perform power spectrum analysis on the current residual signal. Through the inherent operational logic of this function, the power distribution of the residual signal with frequency is analyzed, generating a power spectral density distribution. The core objective is to accurately locate the frequency corresponding to the maximum power in the distribution, and then convert this frequency into an angular frequency. This provides data support for determining the center frequency of the mode to be extracted, ensuring that mode extraction focuses on the main energy range of the residual signal and avoiding feature bias caused by blind extraction.

[0116] S23. Identify the frequency corresponding to the maximum power in the power spectral density distribution and generate the initial center frequency of the current mode to be extracted.

[0117] Optionally, the frequency value corresponding to the maximum power in the power spectral density distribution is identified and converted into an angular frequency using the angular frequency conversion formula: ω=2πf, where f is the frequency of maximum power. This angular frequency is then used as the initial center frequency of the mode to be extracted. The core of this conversion logic is to adapt to the mathematical model requirements of variational mode decomposition, ensuring that the mode extraction process revolves around the main energy range of the signal, thereby improving the physical correlation and effectiveness of the modes.

[0118] S24. Based on the triple constraints of modal concentration, spectral overlap and modal independence, variational mode extraction iterative optimization is performed on the current residual signal containing the initial center frequency to generate an intrinsic mode function component.

[0119] Optionally, iterative optimization is carried out based on the triple constraints of "modal concentration, spectral overlap, and modal independence": modal concentration requires that the extracted modes fluctuate stably around the initial center frequency to ensure the focus of the modes; spectral overlap requires that the spectrum of the currently extracted mode has no obvious intersection with the spectrum of the remaining residual to avoid frequency feature confusion; modal independence requires that the newly extracted modes have no redundant information with the extracted modes to ensure the uniqueness of each mode. The maximum number of iterations is fixed at 500, and the convergence threshold is set to 1e-6. When the iteration error meets the convergence threshold or all three constraints are met, the iteration is stopped, and one intrinsic mode function (IMF) component is extracted to ensure the effectiveness and stability of mode extraction.

[0120] S25. Subtract the generated intrinsic mode function components from the current residual signal to generate the updated current residual signal.

[0121] Optionally, the residual signal is updated according to the logic of "updated residual = original residual - extracted IMF component". By stripping the extracted effective modes, the remaining residual is focused on the frequency features that have not been captured, providing clean input data for the next round of mode extraction and ensuring the effectiveness of the layering of each mode component.

[0122] S26. Using the updated current residual signal as input, repeat S22-S25 until the updated current residual signal has no effective oscillation modes that can be extracted or the number of extracted modes reaches the preset maximum mode number threshold, and generate intrinsic mode function components and a residual component sorted by frequency from high to low.

[0123] Optionally, S22-S25 are repeated until no effective oscillation modes can be extracted from the updated current residual (the residual signal tends to be stable and the power spectrum has no obvious peaks) or the number of extracted modes reaches the preset maximum mode number threshold. The final output consists of N IMF components and 1 residual component sorted by frequency from high to low, where the high-frequency IMF corresponds to high-frequency noise and short-term fluctuations, the mid-frequency IMF corresponds to medium-term periodic fluctuations, and the low-frequency IMF and residual correspond to long-term degradation trends.

[0124] In the above embodiments, the SVMD decomposition process is refined, including residual initialization, power spectrum analysis to determine the center frequency, iterative extraction of IMF under triple constraints, residual update, and termination condition determination. The quantization parameters and execution standards for each step are clearly defined. The operational logic and constraint boundaries of the decomposition process are standardized to avoid over-decomposition or modal redundancy, ensuring that the layered signals can accurately distinguish high-frequency noise, mid-frequency fluctuations, and low-frequency degradation trends, providing high-quality input for subsequent feature processing.

[0125] In one embodiment, S3 includes:

[0126] S31. Calculate the zero-crossing rate of the intrinsic mode function components, generating the zero-crossing rate values ​​for each component. The zero-crossing rate is calculated according to the formula:

[0127]

[0128] in, Let be the zero-crossing rate of each component, and IMF be the intrinsic mode function component. for The length of the sequence, This represents the difference in sign changes of the IMF sequence.

[0129] Alternatively, the zero-crossing rate is determined by the formula: Calculation, where This is the sign function for the IMF sequence, meaning it returns 1 for a positive value, -1 for a negative value, and 0 for a zero value. The function is used to calculate the difference between adjacent signs; a difference of ±2 indicates crossing the zero value. The total number of changes in the statistical sign. The length of the IMF sequence is given. The core purpose of this calculation is to quantify how frequently the signal crosses zero, providing a quantitative indicator for determining high-frequency noise components and avoiding subjective judgment errors.

[0130] S32. Based on the center frequencies determined by each component during the adaptive variational mode decomposition of the data, the condition that the zero-crossing rate is greater than the first threshold and the center frequency is greater than the second threshold is determined, and a high-frequency noise component identifier is generated.

[0131] Optionally, based on the center frequency determined in the SVMD decomposition of each IMF component, a dual criterion of "zero-crossing rate > 0.05 and center frequency > 0.1Hz" is used to screen all IMF components. This criterion is determined based on the distribution characteristics of IGBT peak voltage noise. A zero-crossing rate > 0.05 indicates frequent signal fluctuations, and a center frequency > 0.1Hz indicates that the signal is in the high-frequency range. Combining the two can accurately locate high-frequency noise components and generate corresponding high-frequency noise component identifiers, typically IMF1.

[0132] S33. For the component specified by the high-frequency noise component identifier, perform multi-level wavelet packet decomposition through wavelet packet transform to generate the high-frequency data packet coefficients of each sub-band.

[0133] Optionally, the high-frequency components of the identifier are subjected to two-level wavelet packet decomposition using the db4 wavelet basis and the wpdec function. After decomposition, four data packets with different frequency bandwidths are obtained, among which the high-frequency data packets correspond to the 0.25~0.5 times sampling frequency range, which is the main distribution range of high-frequency noise of IGBT peak voltage. The frequency separation of noise and effective signal is achieved through multi-level decomposition.

[0134] S34. Perform soft threshold quantization on the coefficients of high-frequency data packets based on dynamic thresholds to generate coefficients after threshold processing.

[0135] Optionally, a soft thresholding process is applied to the coefficients of high-frequency data packets. The threshold is set to 0.05 × the maximum value of the data packet coefficient. Quantization is performed using the wthresh function: coefficients with absolute values ​​less than the threshold are set to 0; those with values ​​greater than or equal to the threshold are retained and the threshold is subtracted. This dynamic thresholding strategy can adaptively match the noise intensity of high-frequency components, eliminating small noise coefficients while preserving large coefficients of the effective signal to the greatest extent possible, thus avoiding signal distortion.

[0136] S35. Perform inverse wavelet packet reconstruction on the coefficients after thresholding to generate denoised high-frequency components, and merge the denoised high-frequency components, the mid-to-low frequency intrinsic mode function components not identified as high-frequency noise, and a residual component into all processed components.

[0137] Optionally, the coefficients after thresholding are reconstructed using the wprec function via inverse wavelet packet decomposition to restore the time-domain signal, resulting in the denoised high-frequency IMF component sequence. The reconstruction process strictly follows the inverse operation logic of wavelet packet decomposition, ensuring that the denoised signal retains the basic characteristics of the original high-frequency modes, removing only meaningless noise interference, and combining it with the mid-to-low frequency intrinsic mode function components not identified as high-frequency noise and a residual component to form the processed complete components.

[0138] In the above embodiments, denoising is achieved by quantizing the zero-crossing rate calculation, determining high-frequency components using dual standards, employing db4 wavelet basis level 2 decomposition, dynamic soft thresholding, and inverse reconstruction. This accurately identifies and removes high-frequency noise, preserving effective signal features to the greatest extent possible, avoiding noise interference with model training, and standardizing the denoising process to improve the consistency and effectiveness of feature processing.

[0139] In one embodiment, S4 includes:

[0140] S41. Calculate the energy percentage of all processed components to generate the energy value of each component.

[0141] Optionally, the energy calculation covers the denoised high-frequency IMF components, mid-to-low-frequency IMF components, and one residual component, with the energy of each component being the sum of the squares of the elements in its sequence (…). This calculation method can objectively reflect the signal strength of each component. The higher the energy, the richer the effective information related to IGBT degradation contained in the component, providing a quantitative basis for attention weight allocation.

[0142] S42. Based on the energy values ​​of each component, attention weights are assigned to generate attention weights for each component; the formula for calculating the attention weights is:

[0143]

[0144] in, Let i be the attention weight of the i-th component. , For the first j processed IMF sequences.

[0145] Optionally, based on the energy values ​​of each component, according to the formula... Calculate attention weights, where Let j be the i-th processed IMF sequence, with a value ranging from 1 to 3, corresponding to 3 IMF components. If the number of IMFs obtained from the decomposition is different, the value range of j is adjusted accordingly. The core logic of this formula is "energy proportion determines weight," which gives higher weight to high-energy effective components, such as low-frequency IMFs reflecting long-term degradation trends, and lower weight to low-energy noise components, thereby achieving automatic enhancement of key features.

[0146] S43. Based on the attention weights, perform weighted scaling on the sequence of each component to generate a weighted feature component sequence.

[0147] Optionally, all data points in each component sequence are multiplied by their corresponding attention weights to complete the weighted scaling process. For example, when the attention weight of the low-frequency IMF component is 0.674, all data points in its sequence are multiplied by 0.674. This operation amplifies the contribution of key features to model training while weakening the interference of noise components, generating a more targeted weighted feature component sequence.

[0148] In the above embodiments, the energy of each component is calculated based on the sum of squares of the sequence, attention weights are allocated according to the energy proportion, key features are enhanced through weighted scaling, and the quantitative calculation logic of energy and weights is clarified. This automatically focuses on high-energy components with long-term degradation trends, weakens low-energy noise interference, makes model training more targeted, effectively improves the ability to capture IGBT degradation patterns, and helps improve prediction accuracy.

[0149] In one embodiment, S6 includes:

[0150] S51. Using the root mean square error of prediction on the global validation set as the fitness function, perform PSO search on the number of hidden layer units, learning rate, batch size and dropout rate of the LSTM network in the preset hyperparameter space to generate a coarse-optimized hyperparameter combination.

[0151] Optionally, the fitness function is set using the predicted RMSE of the global validation set. Twenty particles are initialized, each corresponding to a set of LSTM hyperparameters, including 10-100 hidden layer units, a learning rate of 0.001-0.1, a batch size of 8-32, and a dropout ratio of 0.2-0.5. The PSO iteration parameters are configured as follows: inertia weight 0.7, cognitive factor 1.5, and social factor 1.5. After 50 iterations, the globally optimal hyperparameters are selected. The exploration and development capabilities of the particle swarm optimization (PSO) rapidly compress the hyperparameter search space, avoiding getting bogged down in a large range of ineffective searches during fine-tuning and improving optimization efficiency.

[0152] S52. Centering on the coarse-optimized hyperparameter combination, the preset hyperparameter space is compressed to generate a local fine-search space.

[0153] Optionally, using the optimal hyperparameters obtained through coarse PSO as the center, the search range of each hyperparameter is reduced by 50%. For example, when the coarse learning rate is 0.012, the fine-tuning range is set to 0.006~0.018, generating a local fine-tuning search space. This mainly focuses on the optimal region of the hyperparameters, reducing the computational cost of Bayesian fine-tuning and improving the accuracy of hyperparameter optimization.

[0154] S53. Within the local fine search space, the hyperparameters of the LSTM network are tuned using a Bayesian optimization method with random forest as the surrogate model and expected improvement as the acquisition function, to generate the final hyperparameter combination.

[0155] Optionally, within the locally refined search space, a Random Forest (RF) is used as a surrogate model, with 100 decision trees, a maximum tree depth of 10, and a minimum number of samples per leaf node of 5. This replaces the traditional Gaussian process to improve the fit of the nonlinear hyperparameter space. Expected Improvement (EI) is used as the acquisition function to complete 30 hyperparameter evaluations. By predicting hyperparameter performance through the surrogate model, the number of actual model training iterations is reduced, ultimately obtaining a suitable combination of LSTM hyperparameters to ensure optimal LSTM network performance.

[0156] S54. Configure the LSTM network with the final hyperparameter combination to obtain the optimized LSTM feature extraction layer.

[0157] Optionally, the LSTM network is configured according to the final hyperparameter combination: the input dimension of the sequence input layer matches "time series length 10 + total number of components (IMF number + 1 residual)"; the number of hidden units in the LSTM layer is an optimized value (e.g., 96), and the output mode is set to "last" (only outputting the hidden state of the last time step); the dropout rate of the Dropout layer is 0.2; and the number of neurons in the fully connected layer is adapted to the feature dimension. This structure ensures that the network can effectively capture temporal dependencies while suppressing overfitting, resulting in an optimized LSTM feature extraction layer.

[0158] S55. The input data of the global training set is extracted through the LSTM feature extraction layer to generate a high-dimensional temporal feature vector.

[0159] Optionally, the input data of the global training set is input into the optimized LSTM feature extraction layer. Through the forward propagation process, the LSTM layer captures the long-short-term time series dependencies, and the fully connected layer performs feature dimension adaptation, finally outputting a high-dimensional time series feature vector. This vector condenses the key time series information of the voltage sequence, providing high-quality input for the fine regression of the SVR regression prediction layer.

[0160] S56. Based on high-dimensional time-series feature vectors, Bayesian optimization is performed on the penalty parameter, error tolerance parameter, and kernel function scaling parameter of the SVR regression prediction layer to generate a prediction model.

[0161] Optionally, the RBF kernel function can be used as input, taking the high-dimensional temporal feature vector as input. The `fitrsvm` function calls Bayesian optimization to find the optimal parameter within the specified range (penalty parameter C: 1~100, error tolerance parameter ε: 0.01~1, kernel scaling parameter γ: 0.01~1), completing 15 parameter evaluations with the goal of minimizing the 5-fold cross-validation error. The training process reuses the learning rate, training epochs, and batch size of the LSTM to ensure LSTM-SVR co-adaptation and generate the prediction model.

[0162] In the above embodiments, the hyperparameter space is compressed by PSO coarse search, and the hyperparameters of the LSTM network are optimized by Bayesian fine-tuning of the random forest surrogate model. The model training of the LSTM feature extraction layer and the SVR regression prediction layer reuses the training parameters and triggers an early stopping mechanism. This achieves efficient and accurate hyperparameter optimization, avoids performance loss caused by improper parameters, and the LSTM-SVR collaborative architecture fully leverages the advantages of temporal feature capture and nonlinear regression while suppressing overfitting and improving the model's generalization ability.

[0163] In one embodiment, S54 includes:

[0164] S541. Based on the number of hidden layer units and the dropout rate in the final hyperparameter combination, construct the network structure of the LSTM feature extraction layer; the network structure includes a sequence input layer, an LSTM layer, a dropout layer, and a fully connected layer.

[0165] Optionally, the network structure is constructed based on the final hyperparameter combination: the output dimension of the sequence input layer is consistent with "time series length × total number of components"; the LSTM layer uses the tanh activation function, and the forget gate, input gate, and output gate use the sigmoid activation function; the dropout rate of the Dropout layer is 0.2, which is a fixed value and has been verified in engineering to meet the overfitting suppression requirements; the fully connected layer uses the linear activation function, and the number of neurons is determined by hyperparameter optimization. This structural design balances the ability to capture temporal features with model stability and is adapted to the temporal data characteristics of IGBT peak voltage.

[0166] S542. Initialize the weight parameters of the network structure to obtain the optimized LSTM feature extraction layer.

[0167] Optionally, after configuring the LSTM network structure based on the final hyperparameter combination, the weight parameters of the network structure are initialized. Specifically, an appropriate parameter initialization strategy is selected to ensure the rationality of the initial values ​​of the weight parameters and the adaptability of their distribution. After the weight parameter initialization operation is completed, the optimized LSTM feature extraction layer is obtained.

[0168] In the above embodiments, the LSTM network structure is configured according to optimized hyperparameters, the network layer configuration and initial weight parameters are clearly defined, ensuring that the LSTM network structure adapts to the data characteristics, the parameter updates are efficient and stable, and the accuracy of temporal feature extraction and the efficiency of model training are improved.

[0169] In one embodiment, S7 includes:

[0170] S61. Based on the test set, construct the initial model input vector according to the same time window rule used in the global training set.

[0171] Optionally, based on the test set data, and following the time window rule of the global training set with a sequence length of 10, the last 10 data points of the test set are selected to construct the initial model input vector. This vector reflects the latest degradation state of the IGBT, ensuring the reliability of the first prediction. At the same time, the input dimension remains consistent with that during model training, avoiding prediction errors caused by format mismatch.

[0172] S62. Input the initial model input vector into the prediction model for single-step forward calculation to generate the standardized peak voltage prediction value at the next sampling time.

[0173] Optionally, the initial input vector is fed into the prediction model. The LSTM feature extraction layer first captures the temporal dependencies over 10 time steps, extracting a high-dimensional temporal feature vector. Then, the SVR regression prediction layer performs nonlinear regression on the high-dimensional temporal feature vector, outputting the standardized peak voltage prediction value for the next sampling time. Single-step prediction avoids the error accumulation caused by direct multi-step prediction, improving the accuracy of each prediction step.

[0174] S63. The newly generated standardized peak voltage prediction value is used as a new data point and fed back to the end of the historical data sequence. At the same time, the oldest data point is removed to generate an updated historical sequence.

[0175] Optionally, the newly generated standardized predicted value is used as a new data point and fed back to the end of the historical data sequence, while the oldest data point in the sequence is removed to generate an updated historical sequence. This rolling update mechanism simulates the application scenario of "predicting the state of the next moment based on the latest collected data" in actual engineering, ensuring the timeliness and accuracy of the prediction results.

[0176] S64. Based on the updated historical sequence, construct a new model input vector according to the same time window rule.

[0177] Optionally, based on the updated historical sequence, a new model input vector is reconstructed according to the rule of "time series length 10" to ensure that the temporal structure of the new model input vector is consistent with that during model training, so as to provide input data that meets the requirements for the next round of single-step prediction.

[0178] S65. Using the new model input vector as input, iteratively execute S62-S64 to generate a future standardized peak voltage prediction sequence of a specified length.

[0179] Optionally, using the new model input vector as input, S62-S64 are iteratively executed until a future standardized peak voltage prediction sequence of a specified length (20-30 sampling periods) is generated. This length covers the typical remaining lifespan of IGBTs while balancing prediction accuracy and computational efficiency, meeting the needs of engineering applications.

[0180] In the above embodiments, an initial input vector is constructed based on the test set. Through single-step prediction, rolling updates of historical sequences, and iterative calculations, a future standardized voltage prediction sequence of a specified length is generated. A rolling window prediction strategy is adopted to avoid the accumulation of errors from multi-step direct prediction, ensuring the timeliness and accuracy of future voltage predictions and adapting to the actual needs of "predicting based on the latest data" in engineering projects.

[0181] In one embodiment, S8 includes:

[0182] S71. Obtain the mean and standard deviation calculated when standardizing the original peak voltage time series data.

[0183] Optionally, the mean of the peak voltage sequence calculated during S1 normalization can be directly obtained. with standard deviation This ensures the consistency and accuracy of the denormalization conversion and avoids voltage value restoration errors caused by inconsistent parameters.

[0184] S72. Based on the mean and standard deviation, perform an inverse standardization transformation on the future standardized peak voltage prediction sequence to generate the actual peak voltage prediction sequence; the inverse standardization transformation calculation formula is:

[0185]

[0186] in, This is the actual voltage value. For standardized voltage values, , These are the mean and standard deviation of the standardized voltage values, respectively.

[0187] Optionally, according to the formula The future standardized peak voltage prediction sequence is destandardized, whereby... This is the actual peak voltage value. This is a standardized voltage value. This conversion can restore the dimensionless standardized value output by the model to the actual voltage value that can be directly determined in engineering, providing an intuitive basis for determining the failure cycle.

[0188] S73. Compare each actual voltage value in the actual peak voltage prediction sequence with the preset failure threshold in sequence, and identify the index position of the first voltage value in the actual peak voltage prediction sequence that is lower than the preset failure threshold.

[0189] Optionally, a preset peak voltage failure threshold for the IGBT, such as 11.0V, is determined according to the rated parameters in the device datasheet or engineering application standards. It needs to match the actual voltage degradation direction. The actual peak voltage prediction sequence is traversed by the find function, and the voltage value of each sampling period is compared with the failure threshold in sequence. The process stops when the index position of the first voltage value lower than the preset failure threshold is identified. This index position directly corresponds to the sequence position of the failure period.

[0190] S74. Determine the corresponding future sampling period based on the index position to generate the predicted failure period of the insulated gate bipolar transistor.

[0191] Optionally, based on the identified index position, the corresponding future sampling period is determined, which is the predicted failure period of the IGBT. If no voltage value falls below the failure threshold within the preset number of future prediction steps (20~30), the number of future prediction steps needs to be extended until the threshold crossover point is found, ensuring the completeness and practicality of the life assessment and providing a clear time basis for the formulation of power electronic system maintenance plans.

[0192] In the above embodiments, the mean and standard deviation from the preprocessing stage are reused to complete the destandardization, restoring the predicted value to the actual voltage. By comparing it with a preset failure threshold, the first sampling period that meets the standard is located as the failure period. This realizes the transformation of the predicted value into an actual, verifiable voltage in engineering, standardizes the failure period determination logic, avoids errors from human experience, and makes life assessment quantitative and traceable, providing a clear time basis for maintenance plan formulation.

[0193] In one embodiment, after S6, the following is also included:

[0194] S81. The prediction model performs forward prediction on the input data of the global training set and the global validation set respectively to generate the corresponding standardized voltage prediction value.

[0195] Optionally, the prediction model performs forward prediction on the input data of the global training set and the global validation set respectively to generate the corresponding standardized voltage prediction values, and then uses the inverse standardization formula... This is converted into actual voltage prediction values ​​to ensure that performance evaluation is based on actual, referable voltage values ​​in engineering, avoiding evaluation bias caused by standardized values.

[0196] S82. Perform denormalization on the generated standardized voltage prediction value to generate the corresponding actual voltage prediction value.

[0197] Optionally, calculate the core performance metrics for the training and validation phases: RMSE and R². RMSE is calculated using the formula... The calculation reflects the average deviation between the predicted and actual values; R² is calculated using the formula... The calculation reflects the model's ability to interpret data. The output performance evaluation results provide a quantitative basis for the model's accuracy; the smaller the RMSE and the closer R² is to 1, the better the model's performance.

[0198] S83. Based on the actual voltage prediction value and the corresponding real voltage value, calculate the prediction performance index for the training and validation phases respectively, and generate performance evaluation results; the performance index includes root mean square error and coefficient of determination, and the formula for calculating root mean square error is:

[0199]

[0200] Where RMSE is the root mean square error, and n is the sample size. Let i be the true value of the i-th sample. Let be the predicted value for the i-th sample.

[0201] The formula for calculating the coefficient of determination is:

[0202]

[0203] in, As the coefficient of determination, The mean of the true values;

[0204] Optionally, the performance metrics can be compared with the performance results of pre-stored benchmark models, including a single LSTM network model (RMSE=0.1209, R²=0.5416), an SVMD-LSTM combined model (RMSE=0.0957, R²=0.7131), and a hybrid (VMD-LSTM-SVR) model (RMSE=0.0411, R²=0.7511). The comparison dimensions cover RMSE, R², training time, and generalization ability, clearly highlighting the technical advantages of this solution in prediction accuracy and generalization ability.

[0205] S84. Compare and analyze the performance evaluation results with the pre-stored benchmark model performance results to generate model performance comparison results; the benchmark models include a single long short-term memory network model and a combination model of variational mode decomposition and long short-term memory network.

[0206] Optionally, based on prior performance evaluation data, the RMSE (0.02988) and R² (0.9638) of the hybrid model are specifically compared with the performance metrics of three pre-existing benchmark models. Specifically, the RMSE of the single Long Short-Term Memory (LSTM) network model is 0.1209 and the R² is 0.5416; the RMSE of the variational mode decomposition and LSTM network combination model is 0.0957 and the R² is 0.7131; and the RMSE of the VMD-LSTM-SVR model is 0.0411 and the R² is 0.7511. The comparison results clearly demonstrate that the RMSE of the prediction model of this invention is significantly lower than all benchmark models, and the R² is closer to 1. It outperforms existing methods in prediction accuracy, data interpretation ability, and generalization ability, clearly highlighting the performance improvement and providing direct quantitative evidence for the engineering application of the solution.

[0207] In the above embodiments, the actual voltage prediction value is obtained through model prediction and inverse standardization. RMSE and R² quantization performance are calculated to generate performance evaluation results. This objectively verifies the model's prediction accuracy and superiority, clarifies performance improvements, provides quantitative support for engineering applications, and enhances the credibility and promotional value of the solution. It effectively solves the technical problems of limited applicability of existing physical models and the difficulty of traditional data-driven methods in dealing with nonlinear, non-stationary characteristics and high-frequency noise interference. This achieves high-precision prediction of IGBT peak voltage and reliable lifetime assessment, providing effective technical support for ensuring the reliability of power electronic systems and reducing maintenance costs.

[0208] The aforementioned state monitoring method and system for insulated gate bipolar transistors (IGBTs) utilizes data preprocessing, feature optimization, model training, performance evaluation, and lifetime determination. It addresses the non-stationary signal layering problem through SVMD decomposition, strengthens key features with attention weighting, and overcomes the prediction accuracy bottleneck of traditional methods through two-stage hyperparameter optimization and prediction model modeling. Multi-stage quantization standards enable more comprehensive feature extraction, superior prediction accuracy, and more reliable lifetime assessment in this technical solution.

[0209] To further illustrate the solutions of the embodiments of this application, a specific example is provided below.

[0210] This embodiment takes the IGBT (model FF450R12ME4, rated voltage 1200V, rated current 450A) used in new energy vehicle inverters as the research object. Peak voltage degradation data are obtained through laboratory accelerated aging tests to fully verify the effectiveness and engineering applicability of the method of the present invention.

[0211] 1. Experimental parameters and data preparation

[0212] like Figure 2As shown, the horizontal axis represents the sampling period (0~450), and the vertical axis represents the peak voltage (V), visually presenting the overall degradation trend of the original peak voltage over the sampling period. The accelerated aging test aims to simulate the degradation process of IGBTs under extreme conditions in new energy vehicles. The test parameters are set as follows: the temperature cycling range is -40℃ to 125℃ (covering the low temperatures of winter and the high temperatures of summer in northern China), the temperature cycling cycle is 1 hour / cycle (ensuring data validity while accelerating aging), and the applied electrical stress is 1.2 times the rated voltage (1440V) to accelerate the voltage degradation process of the IGBT and shorten the test cycle. In the data acquisition phase, the sampling interval is set to 10 hours / cycle. This interval ensures sufficient data density to capture the degradation trend while controlling the total data volume to avoid redundancy. A total of 200 sets of peak voltage degradation data were collected in CSV format, containing two columns: "Sampling Period" and "Peak Voltage".

[0213] The data preprocessing process is as follows: outliers are removed using the 3σ criterion, and the mean of the peak voltage sequence is calculated. =9.2V, standard deviation =0.85V, discard 2 groups that satisfy | -9.2| > 3×0.85 outlier data (1 group 12.1V, 1 group 5.8V), leaving 198 valid data groups; Z-score standardization was performed on the valid data, and the standardized sequence range was concentrated in [-1.41, 1.53]; the dataset was divided into a 7:2:1 time sequence, with the training set being period 1~139 (139 data groups), the validation set being period 140~178 (39 data groups), and the test set being period 179~198 (20 data groups), strictly avoiding future data leakage.

[0214] 2. SVMD Decomposition and Feature Optimization

[0215] Perform SVMD decomposition on the standardized training data, such as Figure 3 As shown, from top to bottom are the original standardized electrical values. The residual components, with the horizontal axis representing the sampling period. The vertical axis represents the standardized voltage value, which shows... Frequent fluctuations The residuals show a slow upward trend and are stable, achieving effective stratification of non-stationary signals: the current residuals are initialized as a standardized voltage sequence of the training set, the maximum number of modes threshold is set to 5, power spectrum analysis is performed using the pwelch function, and IMF components are iteratively extracted by combining the triple constraints of "modal concentration, spectral overlap, and modal independence". The number of iterations is set to 500, and the convergence threshold is 1e-6. Finally, 3 IMF components and 1 residual component are obtained. The high-frequency component determination results are as follows: IMF1 has a zero-crossing rate of 0.08 and a center frequency of 0.15Hz, which satisfies "zero-crossing rate > 0.05 and center frequency > 0.1Hz", and is determined to be a high-frequency noise component; IMF2 has a zero-crossing rate of 0.03 and a center frequency of 0.08Hz; IMF3 has a zero-crossing rate of 0.01 and a center frequency of 0.02Hz, all of which are stationary components.

[0216] Wavelet packet denoising is performed on IMF1, such as... Figure 4 As shown, the left figure shows the IMF1 before denoising (fluctuation amplitude). The image on the right shows the result after noise reduction. The horizontal axis represents the sampling period, and the vertical axis represents the standardized voltage value. It is clearly observed that high-frequency noise is significantly suppressed after denoising, while preserving... Basic fluctuation characteristics: Two-level decomposition using the db4 wavelet basis was adopted, with the soft threshold set to 0.05 × the maximum value of the data packet coefficients. The denoised IMF1 was obtained by inverse wavelet packet reconstruction. The energy of each component was calculated: IMF1 (0.32), IMF2 (1.56), IMF3 (3.88), and residual (0.24), with a total energy of 6.0. The attention weights were assigned as IMF1 (0.055), IMF2 (0.271), IMF3 (0.674), and residual (0.04). Weighted scaling was performed on each component to enhance the low-frequency degradation characteristics.

[0217] 3. Model Training and Performance Evaluation

[0218] Time series samples were constructed: the time window length was set to 10, and a global training and validation set were generated. The order of the training set samples was shuffled using the randperm function. The hyperparameters of the LSTM network were determined through a combination of PSO coarse search and Bayesian fine-tuning: after compressing the space using PSO coarse search (20 particles, 50 iterations), a random forest was used as the Bayesian surrogate model, and the expected improvement was used as the acquisition function for fine-tuning. The final optimal hyperparameters were 96 hidden units, a Dropout ratio of 0.3, a learning rate of 0.012, and a batch size of 16. The SVR regression prediction layer was optimized using Bayesian optimization via the fitrsvm function, with BoxConstraint... Epsilon Setting KernelScale=0.1 results in fast convergence and avoids local optima. Radial basis functions (RBFs) are used as kernel functions because they are suitable for nonlinear regression and have strong generalization ability. They can be obtained through grid search. , The optimal parameters are determined within the specified range, where the optimal value for the penalty parameter C is 50, at which point the cross-validation error is minimized. The kernel function parameters... The optimal value is 0.1, at which point the prediction accuracy is highest. For example... Figure 5 and Figure 6 As shown, the horizontal axis represents the number of iterations (0~2500), and the vertical axis represents RMSE, demonstrating the trend of RMSE gradually converging during training and verifying the stability of model training. After the prediction model is trained, the training set... , The value is 0.971, for the validation set. The prediction model has excellent fitting performance with no overfitting, strong generalization ability, and meets the accuracy requirements of engineering projects.

[0219] 4. Future Voltage Prediction and Lifetime Assessment

[0220] During lifetime assessment, the last 20 optimized features in the training set were selected. Inputting the period into the prediction model directly generates the next 30 sampling periods. The peak voltage prediction value (period) is set to 30 sampling periods; the future prediction step is set to 30 sampling periods, covering... The model's remaining lifespan is considered while also taking practicality into account; the failure threshold is based on... The device datasheet is confirmed as follows Exceeding this value will trigger overvoltage protection. The predicted failure period is 168 cycles, meaning the future... China reached its first time The period is, and the current period (end of the validation set) is. Due to the premature degradation of accelerated test data, the actual remaining lifetime is negative.

[0221] like Figure 7 As shown in the graph (comparison of actual and predicted values ​​in the training set), the solid line represents the actual voltage values ​​in the training set, the dashed line represents the model's predicted values, and the horizontal axis represents the sampling period. The vertical axis represents the peak voltage. As can be seen, the predicted values ​​almost perfectly match the actual values, validating the model training effect. As shown in Figure 8 (scatter plot of actual and predicted values ​​in the validation set), the horizontal axis represents the actual peak voltage. The vertical axis represents the predicted peak voltage. Label the goodness of fit The scattered points are concentrated near the diagonal with no significant deviation, demonstrating the model's generalization ability. For example... Figure 9 As shown, the horizontal axis represents the residuals, and the vertical axis represents the frequency. The residuals are concentrated around 0, approximating a normal distribution, proving that the model has no systematic bias. Figure 10In the lifetime prediction analysis diagram shown, the dots represent the last training set data. The solid line represents the next 30 periods. Voltage prediction trend, horizontal line represents failure threshold The horizontal axis represents the sampling period, and the vertical axis represents the peak voltage. The predicted voltage can be clearly observed at Achieve failure cycle location.

[0222] The parameter verification and result analysis in this embodiment show that: in the data preprocessing stage, Guidelines and Time-order partitioning effectively ensures data quality; in the feature optimization stage, filter( )and It can achieve effective hierarchical classification and noise suppression of non-stationary voltage sequences, and attention weighting further enhances key features; in the model and evaluation stages, 5 The architecture, combined Polynomial fitting and Failure threshold enables high-precision prediction of peak voltage. With quantitative assessment of lifespan, it fully meets the requirements of power electronic systems. Engineering requirements for dynamic monitoring and life management.

[0223] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0224] Based on the same inventive concept, this application also provides a system for implementing the state monitoring method of an insulated gate bipolar transistor (IGBT) as described above. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more IGBT state monitoring system embodiments provided below can be found in the above-described limitations of the IGBT state monitoring method, and will not be repeated here.

[0225] In one exemplary embodiment, such as Figure 11As shown, a state monitoring system 10 for an insulated gate bipolar transistor is provided, comprising:

[0226] The voltage sequence standardization module 11 is used to standardize the peak voltage timing data of the insulated gate bipolar transistor to generate a standardized voltage sequence; the standardized voltage sequence is divided into a training set, a validation set, and a test set according to the time sequence.

[0227] The variational mode decomposition module 12 is used to perform data-adaptive variational mode decomposition on the standardized voltage sequence in the training set to generate intrinsic mode function components and a residual component.

[0228] The high-frequency noise removal module 13 is used to perform wavelet packet denoising on the high-frequency noise of the intrinsic mode function components based on the zero-crossing rate and center frequency to obtain all processed components; the processed components include the denoised high-frequency components, the mid- and low-frequency intrinsic mode function components, and a residual component.

[0229] The weighted feature calculation module 14 is used to perform attention weight calculation based on energy proportion on all processed components to generate a weighted feature component sequence;

[0230] The time series sample construction module 15 is used to construct time series samples based on weighted feature component sequences and preset time windows, and generate a global training set and a global validation set.

[0231] The prediction model training module 16 is used to train the LSTM network feature extraction layer and SVR regression prediction layer based on the global training set and global validation set to obtain the prediction model; the LSTM network is optimized through a two-stage hyperparameter optimization using PSO and Bayesian methods.

[0232] Voltage prediction module 17 is used to input the test set into the prediction model for prediction and output the future standardized peak voltage prediction sequence.

[0233] The failure cycle prediction module 18 is used to determine the sampling period in the actual peak voltage prediction sequence where the first voltage value is lower than the preset failure threshold based on a preset failure threshold, and to generate the predicted failure cycle of the insulated gate bipolar transistor.

[0234] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a state monitoring method for an insulated gate bipolar transistor as described above.

[0235] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of a state monitoring method for an insulated gate bipolar transistor as described above.

[0236] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0237] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method of monitoring the state of an insulated gate bipolar transistor, characterized by, The method includes: S1. Standardize the peak voltage timing data of the insulated gate bipolar transistor to generate a standardized voltage sequence; the standardized voltage sequence is divided into a training set, a validation set, and a test set according to time sequence. S2. Perform data-adaptive variational mode decomposition on the standardized voltage sequences in the training set to generate intrinsic mode function components and a residual component; S3. Based on the zero-crossing rate and center frequency, perform wavelet packet denoising on the high-frequency noise of the intrinsic mode function components to obtain all processed components; the processed components include the denoised high-frequency components, mid- and low-frequency intrinsic mode function components, and the residual component. S4. Perform attention weighting calculation based on energy proportion on all processed components to generate a weighted feature component sequence; S5. Construct time series samples based on the weighted feature component sequence and the preset time window to generate a global training set and a global validation set; S6. Based on the global training set and global validation set, train the LSTM network feature extraction layer and SVR regression prediction layer to obtain the prediction model; the LSTM network is optimized through a two-stage hyperparameter optimization using PSO and Bayesian methods. S7. Input the test set into the prediction model for prediction, and output the future standardized peak voltage prediction sequence. S8. Based on a preset failure threshold, determine the sampling period in the actual peak voltage prediction sequence where the first voltage value is lower than the preset failure threshold, and generate the predicted failure period of the insulated gate bipolar transistor.

2. The method of claim 1, wherein, S2 includes: S21. Set the standardized voltage sequence in the training set as the current residual signal; S22. Perform power spectrum analysis on the current residual signal to generate the corresponding power spectral density distribution; S23. Identify the frequency corresponding to the maximum power in the power spectral density distribution and generate the initial center frequency of the current mode to be extracted; S24. Based on the triple constraints of modal concentration, spectral overlap and modal independence, the variational mode extraction iterative optimization is performed on the current residual signal containing the initial center frequency to generate an intrinsic mode function component. S25. Subtract the generated intrinsic mode function component from the current residual signal to generate an updated current residual signal; S26. Using the updated current residual signal as input, repeat S22-S25 until the updated current residual signal has no effective oscillation modes that can be extracted or the number of extracted modes reaches the preset maximum mode number threshold, and generate the intrinsic mode function components and a residual component sorted by frequency from high to low.

3. The method of claim 1, wherein, S3 includes: S31. Calculate the zero-crossing rate of the intrinsic mode function components to generate the zero-crossing rate values ​​for each component. The zero-crossing rate is calculated according to the formula: wherein, is the zero-crossing rate for each component, IMF is the intrinsic mode function component, is the length of the sequence, is the difference of IMF sequence sign changes; S32. Based on the center frequencies determined by each component during the adaptive variational mode decomposition of the data, the condition that the zero-crossing rate is greater than the first threshold and the center frequency is greater than the second threshold is determined, and a high-frequency noise component identifier is generated. S33. For the component specified by the high-frequency noise component identifier, perform multi-level wavelet packet decomposition through wavelet packet transform to generate high-frequency data packet coefficients for each sub-band. S34. Perform soft threshold quantization based on dynamic threshold on the coefficients of the high-frequency data packets to generate coefficients after threshold processing; S35. Perform inverse wavelet packet reconstruction on the coefficients after threshold processing to generate the denoised high-frequency components, and merge the denoised high-frequency components, the mid-to-low frequency intrinsic mode function components not identified as high-frequency noise, and the residual component into all the processed components.

4. The method of claim 1, wherein, S4 includes: S41. Calculate the energy percentage of all processed components to generate the energy value of each component. S42. Based on the energy values ​​of each component, attention weights are assigned to generate attention weights for each component; the formula for calculating the attention weights is: wherein, is the attention weight for the i-th component, , is the i-th , j-th processed IMF sequence; S43. Based on the attention weights, perform weighted scaling on the sequence of each component to generate the weighted feature component sequence.

5. The method of claim 1, wherein, S6 includes: S51. Using the root mean square error of the prediction of the global validation set as the fitness function, perform PSO search on the number of hidden layer units, learning rate, batch size and dropout rate of the LSTM network in the preset hyperparameter space to generate a coarse-optimized hyperparameter combination. S52. Using the coarse optimization hyperparameter combination as the center, the preset hyperparameter space is compressed to generate a local fine search space. S53. Within the local fine search space, the hyperparameters of the LSTM network are tuned using a Bayesian optimization method with random forest as the surrogate model and desired improvement as the acquisition function, to generate the final hyperparameter combination. S54. Configure the LSTM network with the final hyperparameter combination to obtain an optimized LSTM feature extraction layer; S55. The input data of the global training set is extracted through the LSTM feature extraction layer to generate a high-dimensional temporal feature vector; S56. Based on the high-dimensional time-series feature vector, Bayesian optimization is performed on the penalty parameter, error tolerance parameter, and kernel function scaling parameter of the SVR regression prediction layer to generate the prediction model.

6. The method of claim 5, wherein, S54 includes: S541. Based on the number of hidden layer units and the dropout rate in the final hyperparameter combination, construct the network structure of the LSTM feature extraction layer; the network structure includes a sequence input layer, an LSTM layer, a dropout layer, and a fully connected layer; S542. Initialize the weight parameters of the network structure to obtain the optimized LSTM feature extraction layer.

7. The method of claim 1, wherein, S7 includes: S61. Based on the test set, construct the initial model input vector according to the same time window rule used in the global training set; S62. Input the initial model input vector into the prediction model for single-step forward calculation to generate the standardized peak voltage prediction value at the next sampling time. S63. The newly generated standardized peak voltage prediction value is used as a new data point and fed back to the end of the historical data sequence. At the same time, the oldest data point is removed to generate an updated historical sequence. S64. Based on the updated historical sequence, construct a new model input vector according to the same time window rule; S65. Using the new model input vector as input, iteratively execute S62-S64 to generate the future standardized peak voltage prediction sequence of a specified length.

8. The method of claim 1, wherein, S8 includes: S71. Obtain the mean and standard deviation calculated when standardizing the original peak voltage time series data; S72. Based on the mean and standard deviation, perform denormalization transformation on the future standardized peak voltage prediction sequence to generate the actual peak voltage prediction sequence; the denormalization transformation calculation formula is: wherein, is the actual voltage value, is the normalized voltage value, , are the mean and standard deviation of the normalized voltage value, respectively; S73. Compare each actual voltage value in the actual peak voltage prediction sequence with the preset failure threshold in sequence, and identify the index position of the first voltage value in the actual peak voltage prediction sequence that is lower than the preset failure threshold. S74. Determine the corresponding future sampling period based on the index position, and generate the predicted failure period of the insulated gate bipolar transistor.

9. The method of claim 1, wherein, Following S6, the following is also included: S81. The prediction model is used to perform forward prediction on the input data of the global training set and the global validation set respectively to generate the corresponding standardized voltage prediction value. S82. Perform denormalization on the generated standardized voltage prediction value to generate the corresponding actual voltage prediction value; S83. Based on the actual voltage prediction value and the corresponding real voltage value, calculate the prediction performance indicators for the training and validation phases respectively, and generate performance evaluation results; the performance evaluation results are used to evaluate the model's prediction accuracy; the performance indicators include root mean square error and coefficient of determination, and the formula for calculating the root mean square error is: wherein RMSE is the root mean square error, and n is the number of samples, is the true value of the i-th sample, is the predicted value of the i-th sample; The formula for calculating the coefficient of determination is: wherein, R2is the correlation coefficient, R is the mean of the true values; S84. Compare and analyze the performance evaluation results with the pre-stored benchmark model performance results to generate a model performance comparison report; the benchmark models include a single long short-term memory network model and a combination model of variational mode decomposition and long short-term memory network.

10. A state monitoring system for an insulated gate bipolar transistor, characterized in that, The system includes: The voltage sequence standardization module is used to standardize the peak voltage timing data of insulated gate bipolar transistors to generate a standardized voltage sequence; the standardized voltage sequence is divided into a training set, a validation set, and a test set in chronological order. The variational mode decomposition module is used to perform data-adaptive variational mode decomposition on the standardized voltage sequences in the training set to generate intrinsic mode function components and a residual component. A high-frequency noise removal module is used to perform wavelet packet denoising on the high-frequency noise of the intrinsic mode function components based on the zero-crossing rate and center frequency to obtain all processed components; the processed components include denoised high-frequency components, mid- and low-frequency intrinsic mode function components, and a residual component; a weighted feature calculation module is used to perform attention weighting calculation on the processed components based on energy proportion to generate a weighted feature component sequence; The time series sample construction module is used to construct time series samples based on the weighted feature component sequence and a preset time window, and generate a global training set and a global validation set. The prediction model training module is used to train the LSTM network feature extraction layer and SVR regression prediction layer based on the global training set and global validation set to obtain the prediction model; the LSTM network is optimized through a two-stage hyperparameter optimization using PSO and Bayesian methods. The voltage prediction module is used to input the test set into the prediction model for prediction and output a future standardized peak voltage prediction sequence. The failure cycle prediction module is used to determine the sampling period in the actual peak voltage prediction sequence where the first voltage value is lower than the preset failure threshold based on a preset failure threshold, and to generate the predicted failure cycle of the insulated gate bipolar transistor.

Images