IGBT service life prediction method based on Bayesian enhanced multi-dimensional adaptive BiLSTM

Abstract

The invention discloses an IGBT life prediction method based on Bayesian enhanced multi-dimensional adaptive BiLSTM, and the method comprises the steps: S1, collecting a collector-emitter peak voltage signal of an IGBT, carrying out the adaptive signal enhancement and MDDE, and generating a dynamically adjusted standardized feature sequence; s2, constructing a BEMDA-BiLSTM model, and jointly optimizing feature parameters and model hyper-parameters by using an enhanced Gaussian process and a user-defined expected promotion acquisition function; s3, performing joint training and dynamic prediction based on the optimization parameters, and outputting a high-precision RUL result by adopting an extended mean square error loss function, an Adam optimizer and an early stop mechanism; and S4, calculating RMSE, MAE and a decision coefficient based on the test set, and verifying the performance. According to the method, MDDE, BiLSTM modeling and Bayesian enhancement optimization are fused into an organic whole, and high-precision RUL prediction is realized through dynamic feature-model coupling and joint parameter optimization.

Description

Technical Field

[0001] This invention relates to an IGBT life prediction method, specifically an IGBT remaining useful life (RUL) prediction method. Background Technology

[0002] As a core semiconductor device in power electronic systems, IGBTs are widely used in high-end fields such as renewable energy power generation, electric vehicle drive systems, high-speed railway traction, and industrial-grade frequency converters. With the rapid development of the new energy industry and smart grids, the reliability and lifetime prediction of IGBTs have become key bottlenecks restricting the efficient operation of systems. Traditional lifetime prediction methods mainly rely on accelerated aging tests and physics-based degradation models (such as thermo-electro-mechanical coupling analysis). However, these methods have limitations when facing non-stationary degradation signals and dynamic changes under complex operating conditions, such as insufficient feature extraction, weak model generalization ability, and high computational complexity. In addition, existing research mostly uses single feature or linear regression techniques, which are difficult to capture the nonlinear and multimodal characteristics of the IGBT degradation process, limiting the prediction accuracy and the universality of engineering applications.

[0003] In recent years, the rapid development of artificial intelligence and big data technologies has provided a new paradigm for IGBT lifetime prediction. Especially in the field of deep learning, Bidirectional Long Short-Term Memory (BiLSTM) networks have gradually gained attention due to their superior time-series modeling capabilities. However, the high-dimensional hyperparameter space of standard BiLSTM models significantly impacts performance, making traditional manual tuning or grid search methods inefficient and unsuitable for the real-time requirements of industrial scenarios. Bayesian optimization, as a global optimization technique based on probabilistic models, has shown potential in hyperparameter tuning; however, its application in IGBT lifetime prediction is still limited by the computational resource requirements of Gaussian processes and the ability to effectively integrate multidimensional features. Existing data-driven methods often treat feature extraction, model modeling, and parameter optimization as independent modules, resulting in a loose framework with poor robustness, failing to form an organic closed loop to handle signal mode changes and dependencies. Therefore, there is an urgent need to develop a unified framework integrating multidimensional adaptive embedding, intelligent joint optimization, and deep learning to overcome the limitations of existing technologies and provide cutting-edge support for predictive maintenance of power electronic systems. Summary of the Invention

[0004] Objective of the invention: To address the aforementioned existing technologies, this invention proposes an IGBT lifetime prediction method based on Bayesian enhanced multidimensional adaptive BiLSTM (BEMDA-BiLSTM). This method integrates multidimensional degradation embedding (MDDE), BiLSTM modeling, and Bayesian enhancement optimization into a closed-loop adaptive framework, BEMDA-BiLSTM. Through dynamic feature-model coupling and joint parameter optimization, it captures the non-stationary and multimodal characteristics of IGBT degradation signals, achieving high-precision RUL prediction.

[0005] Technical Solution: An IGBT lifetime prediction method based on Bayesian enhanced multidimensional adaptive BiLSTM, which integrates multidimensional degradation embedding, bidirectional long short-term memory modeling, and Bayesian enhancement optimization into a closed-loop adaptive framework BEMDA-BiLSTM. Through dynamic feature-model coupling and joint parameter optimization, it captures the non-stationary and multimodal characteristics of IGBT degradation signals, including the following steps:

[0006] S1: Extract the collector-emitter peak voltage V of the IGBT. CE-p The signal is subjected to adaptive signal enhancement and multidimensional degradation embedding to generate a dynamically adjusted standardized feature sequence;

[0007] S2: Construct a Bayesian enhanced adaptive bidirectional long short-term memory model, and achieve joint optimization of feature parameters and model hyperparameters by enhancing the Gaussian process and customizing the acquisition function;

[0008] S3: Based on joint optimization parameters, joint training and dynamic prediction generation are performed. The extended mean square error loss function, Adam optimizer and early stopping mechanism are used to output high-precision remaining lifetime prediction results.

[0009] S4: Calculate the root mean square error, mean absolute error, and coefficient of determination based on the test set to verify the model performance.

[0010] Furthermore, in the closed-loop adaptive framework BEMDA-BiLSTM, the multidimensional degenerate embedding relies on the hidden state feedback of the bidirectional long short-term memory for dynamic embedding, the Bayesian enhancement optimization relies on the joint loss for parameter adjustment, and the training process relies on optimization feedback to form an adaptive loop, thereby achieving overall robust prediction of IGBT non-stationary signals.

[0011] Furthermore, in step S1, the adaptive signal enhancement and multidimensional degradation embedding includes:

[0012] S101: Apply a convolution kernel W of size k. conv For V CE-p The original signal X raw (t) is used for denoising to obtain a preliminary high-quality sequence X. smooth(t), where k is the joint parameter of the Bayesian enhancement optimization;

[0013] S102: In the sliding window size W f Multidimensional features were extracted, including mean μ(t), standard deviation σ(t), and peak-to-peak value v. pp (t) and the time-dependent weighted linear regression slope a are used to construct a dynamic feature vector sequence;

[0014] S103: The constructed dynamic feature vector sequence is independently normalized by dimension, and the hidden state feedback of the bidirectional long short-term memory is embedded through a multidimensional degenerate embedding mechanism to realize the dynamic projection of features.

[0015] Furthermore, in step S101, the convolution denoising formula is:

[0016] ;

[0017] The boundary adopts a reflection-filled mode; the kernel size k is dynamically adjusted by Bayesian enhancement optimization.

[0018] Furthermore, in step S102, the formulas for multidimensional feature extraction are as follows:

[0019] ;

[0020] Among them, w attn The attention weight scalar is derived from the BiLSTM backward hidden state. Obtained through mapping via a fully connected layer; W a and b a These are the weight matrix and bias term of the attention mechanism, respectively; sigmoid is the activation function that maps the weights to the (0, 1) interval; The average of the time indices; This represents the average signal value within the window.

[0021] Furthermore, in step S103, the normalization formula is:

[0022] ;

[0023] The multidimensional degenerate embedding formula is:

[0024] ;

[0025] in, Let be a certain eigenvalue in the eigenvector at time t. Normalized eigenvalues , These are the global maximum and minimum values ​​of this feature dimension on the training set, respectively; z t The original multidimensional feature vector, This represents the enhanced feature vector after embedding temporal dependency feedback; α is the adaptive weight, learned through Bayesian enhancement optimization; ⊕ indicates concatenation, W proj The dimension projection matrix is ​​used to map the concatenated high-dimensional hidden state to the feature vector z. t Same dimensional space.

[0026] Furthermore, in step S2, the Bayesian augmented adaptive bidirectional long short-term memory model simultaneously utilizes forward and backward LSTM units to process time-series data in order to capture the sequential dependencies in the IGBT degradation signal; specifically, the forward LSTM unit processes the input sequence sequentially from time step t=1 to t=n, generating the forward hidden state. The backward LSTM unit processes data from t=n to t=1 in reverse order to generate the backward hidden state. The final hidden state h t The calculation based on splicing is as follows:

[0027] ;

[0028] In the formula, β is the gradient feedback term.

[0029] Furthermore, in step S2, Bayesian augmentation optimization enhances the acquisition function by augmenting the Gaussian process and customizing the expected value, jointly optimizing the window size W. f The formulas for kernel k, number of hidden units, learning rate, and batch size are as follows:

[0030] ;

[0031] Wherein, GP(θ) represents a Gaussian process, used to fit the relationship between hyperparameters and loss function; Let μ(θ) represent a Gaussian distribution, where μ(θ) is the mean function; k(θ,θ') is the covariance function; γ and λ are adjustment parameters, respectively; Δ mode (θ) represents the variance of mode mutation; EI(θ) is the expected boost; E[ ] represents the expectation operation; L(θ) is the hyperparameter. The loss function is L(θ). best Var(ΔV) is the current optimal loss; CE-p ) represents V CE-p The first-order difference variance of the signal.

[0032] Furthermore, in step S3, the joint training employs an extended mean squared error loss function:

[0033] ;

[0034] Where N is the size of the training batch, y i It is the actual value. The first term is the predicted value; the second term is the feature smoothness constraint, where M is the time series length. Let be the embedded feature vector at the j-th time step, δ be the weight coefficient of the feature embedding smoothness loss, and the Adam optimizer is used for optimization training.

[0035] Beneficial effects: (1) By using the MDDE mechanism, the V of the IGBT is reduced. CE-p The original signal is dynamically projected into a high-dimensional feature space and incorporated into the time-dependent weighted slope calculation of the BiLSTM hidden state feedback, which effectively captures the nonlinear trend and sequential dependence in the IGBT degradation process and improves the sensitivity to latch-up effect.

[0036] (2) Bayesian enhancement optimization is used to jointly adjust parameters, and a custom acquisition function is integrated into V. CE-p Signal variance ensures that the optimization process addresses mode mutations in IGBT accelerated aging tests, enabling automatic parameter configuration and reducing errors in traditional parameter tuning across 418 sets of cyclic data.

[0037] (3) By adopting an adaptive BiLSTM core engine with extended loss function and gradient feedback, the capture of front-to-back dependencies is enhanced when processing IGBT degradation signals. Experiments show that the stability of RUL prediction is improved and the error caused by non-stationary characteristics is reduced.

[0038] (4) The closed-loop design of the framework deeply integrates the nonlinear modeling capability of deep learning with the empirical data of the IGBT accelerated aging test platform, overcoming the inherent defects of traditional prediction models based on fixed physical equations or empirical formulas, which have poor generalization ability and cannot adapt to complex working conditions. It forms an overall system and provides an efficient and robust solution for predictive maintenance of new energy, electric vehicles and industrial power electronic systems. Attached Figure Description

[0039] Figure 1 A flowchart of the IGBT lifetime prediction method provided by the present invention;

[0040] Figure 2 A schematic diagram of the voltage change at the moment of IGBT turn-off provided by the present invention;

[0041] Figure 3 The convolution denoising comparison diagram provided by this invention;

[0042] Figure 4 This is a schematic diagram of the optimized IGBT lifetime prediction model provided by the present invention.

[0043] Figure 5 The schematic diagram of the BiLSTM network provided for this invention;

[0044] Figure 6The LSTM memory cell structure diagram provided by this invention;

[0045] Figure 7 A comparison chart of IGBT time series curves predicted by the method provided in this invention and other models. Detailed Implementation

[0046] The invention will now be further explained with reference to the accompanying drawings.

[0047] Figure 1 This is a flowchart of an IGBT lifetime prediction method based on Bayesian enhanced multidimensional adaptive BiLSTM provided by this invention. This method integrates MDDE, BiLSTM modeling, and Bayesian enhancement optimization into a closed-loop adaptive framework BEMDA-BiLSTM, ensuring dynamic coupling between components and achieving high-precision RUL prediction of IGBT non-stationary degradation signals. The method steps are as follows:

[0048] S1: Extract the collector-emitter peak voltage V of the IGBT. CE-p The signal undergoes adaptive signal enhancement and MDDE to generate a dynamically adjusted standardized feature sequence.

[0049] This embodiment uses accelerated aging data provided by NASA as an example. The IGBT model is IRG4BC30K, and the experimental environment is set as shown in Table 1. Parameters such as collector-emitter voltage were measured, and a total of 418 sets of cyclic experiments were conducted. Each set of data contained 100,000 sampling points until the device experienced latch-up, at which point the device was considered to have failed.

[0050] Table 1 Experimental Environment Setup

[0051]

[0052] Figure 2 To visualize the voltage change at the moment of IGBT turn-off, a convolution kernel W of size k is applied. conv For V CE-p The original signal X raw (t) is used for denoising to obtain a preliminary high-quality sequence X. smooth (t):

[0053] ;

[0054] Filtering effect as follows Figure 3 As shown, the boundary adopts a reflection-filled mode, where k is the convolution kernel size, dynamically adjusted by Bayesian enhancement optimization; W conv The weight distribution is Gaussian smoothed, and this step ensures that the denoising process adapts to the abrupt changes in IGBT degradation modes.

[0055] In the sliding window size W fMultidimensional features were extracted, including mean μ(t), standard deviation σ(t), and peak-to-peak value v. pp Given the time-dependent weighted linear regression slope a(t) and the time-dependent weighted linear regression slope a(t), a dynamic feature vector sequence is constructed, and the extraction formulas are as follows:

[0056] ;

[0057] Among them, w attn (t) represents the attention weight scalar, derived from the BiLSTM backward hidden state. Obtained through mapping via a fully connected layer; W a and b a These are the weight matrix and bias term of the attention mechanism, respectively; sigmoid is the activation function that maps the weights to the (0, 1) interval; The average of the time indices. The weighting is the average signal value within the window; this weighting mechanism ensures that feature extraction depends on the temporal dependence of the model, forming an inherent coupling within the framework.

[0058] The constructed dynamic feature vector sequence is normalized dimension-independently using MinMaxScaler. The normalization formula is as follows:

[0059] ;

[0060] in, Let be a certain eigenvalue in the eigenvector at time t. These are the normalized eigenvalues. , These represent the global maximum and minimum values ​​of this feature dimension on the training set, respectively. Then, the hidden state feedback of bidirectional long short-term memory is embedded through the MDDE mechanism to achieve dynamic projection of the features. The MDDE embedding formula is:

[0061] ;

[0062] Among them, z t Represents the original multidimensional feature vector; The enhanced feature vector is embedded with temporal dependency feedback; α is the adaptive weight, which is learned by the Bayesian enhancement optimization. This is the forward-hidden state; ⊕ indicates splicing, W proj The dimension projection matrix is ​​used to map the concatenated high-dimensional hidden state to the feature vector z. t Same dimensional space.

[0063] This embedding mechanism projects features into the BiLSTM input space, ensuring dynamic capture of dependencies.

[0064] S2: Construct a Bayesian augmented adaptive BiLSTM model, and achieve joint optimization of feature parameters and model hyperparameters through augmented Gaussian process (EGP) and custom expected boost (EI) acquisition function.

[0065] Figure 4 The diagram shows the optimization principle of the IGBT lifetime prediction model provided by this invention. By integrating MDDE, BiLSTM modeling and Bayesian enhancement optimization into a closed-loop adaptive framework BEMDA-BiLSTM, dynamic coupling between components is ensured, and high-precision RUL prediction of IGBT non-stationary degradation signals is achieved.

[0066] Figure 5 This is a schematic diagram of a BiLSTM network. Its core lies in simultaneously utilizing forward and backward LSTM units to process time-series data, capturing the dependencies between different time steps in IGBT degradation signals. The forward LSTM unit processes the input sequence sequentially from time step t=1 to t=n, generating the forward hidden state. Then, the LSTM units are processed in reverse from t=n to t=1 to generate the backward hidden state. The final hidden state h t The calculation based on splicing is as follows:

[0067] ;

[0068] Where β is the gradient feedback term.

[0069] Figure 6 The BiLSTM model is an LSTM network structure consisting of input gates, output gates, and forget gates. It contains multiple hidden layers, where neurons are replaced with gating memory cells. These memory cells are the core components of the LSTM network, containing the time-t sequence input x. t The state of the hidden layer cells at time (t-1) is h. t-1 and memory unit c t-1 The output includes the memory cell state c. t With hidden layer state h t , where c t with h t Each contains the model's long-term and short-term memory information. The reading and modification of memory cells are achieved by controlling the input gate, forget gate, and output gate, facilitating information flow between networks. The input gate uses the sigmoid activation function to input parameters, controlling the variable between [0, 1] to achieve x. t For c t The forgetting gate is the selective forgetting of the neuronal state at the previous moment, specifically manifested by utilizing memory units c. t-1 For c tThe control; the output gate is used to output and control parameter variables, that is, using c t For h t The degree of influence of the memory module at time t in the LSTM model is determined by both the output gate and the cell state, as shown in the following formula:

[0070] ;

[0071] Among them, i t f t o t These represent the state calculation results of the input gate, forget gate, and output gate, respectively; W i W f W o With b i b f b o These represent the weight matrix and bias term of the corresponding gate, respectively; h t It represents the input hidden state; σ represents the sigmoid activation function; The cell state input at time t represents the state input; tanh is the hyperbolic tangent activation function; W c and b c These represent the state weight matrix and bias term of the input layer, respectively; ⊙ indicates that the elements are multiplied according to their positions.

[0072] BiLSTM predictions for:

[0073] ;

[0074] Among them, W fc and b fc It is the weight matrix and bias term of the fully connected layer.

[0075] Bayesian enhancement optimization uses EGP and EI to improve the acquisition function and jointly optimizes the window size W. f The formulas for kernel size k, number of hidden units, learning rate, and batch size are as follows:

[0076] ;

[0077] Wherein, GP(θ) represents a Gaussian process, used to fit the relationship between hyperparameters and loss function; Let μ(θ) represent a Gaussian distribution, where μ(θ) is the mean function, k(θ,θ') is the covariance function, γ and λ are the adjustment parameters, respectively; Δ mode (θ) represents the variance of mode mutation; EI(θ) represents the expected improvement; E[ ] represents the expected operation; L(θ) is the hyperparameter. The loss function is L(θ). best Var(ΔV) is the current optimal loss;CE-p ) represents V CE-p The first-order difference variance of the signal; this user-defined function ensures that the optimization process is coupled with the IGBT signal characteristics. The optimal hyperparameters searched in this embodiment are shown in Table 2.

[0078] Table 2 Bayesian Hyperparameter Optimization

[0079] hyperparameters Search scope optimal value convolution kernel [1,10] 5 Window size [24,100] 24 Hidden unit number [32,128] 105 Learning rate [0.0001,0.01] 0.005213 Batch size [16,64] 16

[0080] S3: Based on joint optimization parameters, joint training and dynamic prediction generation are performed. The extended mean squared error loss function (MSE), Adam optimizer and early stopping mechanism are used to output high-precision remaining lifetime (RUL) prediction results.

[0081] Joint training employs an extended mean squared error loss function:

[0082] ;

[0083] Where N is the size of the training batch, y i It is the actual value. The first term is the predicted value; the second term is the feature smoothness constraint, where M is the time series length. Let be the embedded feature vector at the j-th time step, and δ be the weight coefficient of the feature embedding smoothness loss. This is learned through Bayesian reinforcement learning. The Adam optimizer is used for training, and its update rule is as follows:

[0084] ;

[0085] Where, m t β1 is the first-order momentum; β1 is the momentum decay rate. It is the gradient of the loss with respect to the parameters; v t It is the second momentum; β2 is the second momentum decay rate; θ t η is the current parameter; η is the learning rate; ε is the adaptive step size; ε is the smoothing term. In this embodiment, the number of training set samples N trian =0.7N, where N is the total number of samples; the number of samples in the validation set N val =0.2N trian Test set sample size N test =0.3N; The early stopping mechanism monitors and verifies the loss to ensure the dynamic convergence of the framework.

[0086] S4: Calculate the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) based on the test set to verify the model performance.

[0087] The performance metrics used to estimate the model are calculated using the following formulas:

[0088]

[0089] Among them, y i It is the actual value; It is a predicted value; These are the true mean values. These metrics collectively validate the framework's ability to capture IGBT mode mutations.

[0090] The time series data obtained after MDDE preprocessing is input into a Bayesian augmented adaptive BiLSTM model, and the Adam optimizer is used to update the optimal hyperparameters obtained by the Bayesian algorithm. To facilitate comparison of the prediction accuracy of different models, this invention conducts comparative experiments on IGBT time series prediction using the LSTM model, BiLSTNM model, and BEMDA-BiLSTM model respectively. Performance comparisons are shown in Table 3, and the prediction results are visualized as follows: Figure 7 As shown.

[0091] Table 3 Performance Comparison

[0092] Model RMSE(V) MAE(V) <![CDATA[R 2 ]]> LSTM 0.0687 0.0554 0.5861 BiLSTM 0.0456 0.0360 0.8143 BEMDA-BiLSTM 0.0331 0.0275 0.9005

[0093] R of the BEMDA-BiLSTM model 2 The value increased from 0.5861 to 0.9005, an increase of approximately 54%; the RMSE decreased from 0.0687V to 0.0331V, a decrease of approximately 52%; and the MAE decreased from 0.0554V to 0.0275V, a decrease of approximately 50%.

[0094] As can be seen from the above scheme: First, the present invention uses the MDDE mechanism to transfer the V of the IGBT. CE-p The original signal is dynamically projected into a high-dimensional feature space and incorporated into the time-dependent weighted slope calculation of BiLSTM hidden state feedback, effectively capturing the nonlinear trend and sequential dependencies in the IGBT degradation process and improving the sensitivity to latch-up effects. Secondly, parameters are jointly adjusted through Bayesian enhancement optimization, and a custom acquisition function is incorporated into V. CE-pThe system addresses several key aspects: First, it improves signal variance to ensure the optimization process addresses mode mutations in IGBT accelerated aging tests, enabling automatic parameter configuration and reducing errors in traditional parameter tuning across 418 sets of cyclic data. Second, by employing an adaptive BiLSTM core engine with an extended loss function and gradient feedback, it enhances the capture of pre- and post-dependent dependencies when processing IGBT degradation signals. Experiments show improved stability of RUL prediction and reduced errors caused by non-stationary characteristics. Third, the closed-loop design of the framework deeply integrates the nonlinear modeling capabilities of deep learning with empirical data from the IGBT accelerated aging test platform. This overcomes the inherent defects of traditional prediction models based on fixed physical equations or empirical formulas, which have poor generalization ability and cannot adapt to complex operating conditions. This forms a holistic system that provides an efficient and robust solution for predictive maintenance of new energy, electric vehicles, and industrial power electronic systems.

[0095] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting IGBT lifetime based on Bayesian enhanced multidimensional adaptive BiLSTM, characterized in that, By fusing multidimensional degradation embedding, bidirectional long short-term memory modeling, and Bayesian enhancement optimization into a closed-loop adaptive framework BEMDA-BiLSTM, the non-stationary and multimodal characteristics of IGBT degradation signals are captured through dynamic feature-model coupling and joint parameter optimization, including the following steps: S1: Extract the collector-emitter peak voltage V of the IGBT. CE-p The signal is subjected to adaptive signal enhancement and multidimensional degradation embedding to generate a dynamically adjusted standardized feature sequence; S2: Construct a Bayesian enhanced adaptive bidirectional long short-term memory model, and achieve joint optimization of feature parameters and model hyperparameters by enhancing the Gaussian process and customizing the acquisition function; S3: Based on joint optimization parameters, joint training and dynamic prediction generation are performed. The extended mean square error loss function, Adam optimizer and early stopping mechanism are used to output high-precision remaining lifetime prediction results. S4: Calculate the root mean square error, mean absolute error, and coefficient of determination based on the test set to verify the model performance.

2. The method according to claim 1, characterized in that, In the closed-loop adaptive framework BEMDA-BiLSTM, the multidimensional degenerate embedding relies on the hidden state feedback of the bidirectional long short-term memory for dynamic embedding, the Bayesian enhancement optimization relies on the joint loss for parameter adjustment, and the training process relies on optimization feedback to form an adaptive loop, thereby achieving overall robust prediction of IGBT non-stationary signals.

3. The method according to claim 1, characterized in that, In step S1, the adaptive signal enhancement and multidimensional degradation embedding includes: S101: Apply a convolution kernel W of size k. conv For V CE-p The original signal X raw (t) is used for denoising to obtain a preliminary high-quality sequence X. smooth (t), where k is the joint parameter of the Bayesian enhancement optimization; S102: In the sliding window size W f Multidimensional features were extracted, including mean μ(t), standard deviation σ(t), and peak-to-peak value v. pp (t) and the time-dependent weighted linear regression slope a are used to construct a dynamic feature vector sequence; S103: The constructed dynamic feature vector sequence is independently normalized by dimension, and the hidden state feedback of the bidirectional long short-term memory is embedded through a multidimensional degenerate embedding mechanism to realize the dynamic projection of features.

4. The method according to claim 3, characterized in that, In step S101, the convolution denoising formula is: ； The boundary adopts a reflection-filled mode; the kernel size k is dynamically adjusted by Bayesian enhancement optimization.

5. The method according to claim 3, characterized in that, In step S102, the formulas for multidimensional feature extraction are as follows: ； Among them, w attn The attention weight scalar is derived from the BiLSTM backward hidden state. Obtained through mapping via a fully connected layer; W a and b a These are the weight matrix and bias term of the attention mechanism, respectively; sigmoid is the activation function that maps the weights to the (0, 1) interval; The average of the time indices; This represents the average signal value within the window.

6. The method according to claim 3, characterized in that, In step S103, the normalization formula is: ； The multidimensional degenerate embedding formula is: ； in, Let be a certain eigenvalue in the eigenvector at time t. Normalized eigenvalues , These are the global maximum and minimum values ​​of this feature dimension on the training set, respectively; z t The original multidimensional feature vector, This represents the enhanced feature vector after embedding temporal dependency feedback; α is the adaptive weight, learned through Bayesian enhancement optimization; ⊕ indicates concatenation, W proj The dimension projection matrix is ​​used to map the concatenated high-dimensional hidden state to the feature vector z. t Same dimensional space.

7. The method according to claim 1, characterized in that, In step S2, the Bayesian augmented adaptive bidirectional long short-term memory model simultaneously utilizes forward and backward LSTM units to process time-series data in order to capture the sequential dependencies in the IGBT degradation signal; specifically, the forward LSTM unit processes the input sequence sequentially from time step t=1 to t=n, generating the forward hidden state. The backward LSTM unit processes data from t=n to t=1 in reverse order to generate the backward hidden state. The final hidden state h t The calculation based on splicing is as follows: ； In the formula, β is the gradient feedback term.

8. The method according to claim 7, characterized in that, In step S2, Bayesian augmentation optimization improves the acquisition function by enhancing the Gaussian process and customizing the expected value, and jointly optimizes the window size W. f The formulas for kernel k, number of hidden units, learning rate, and batch size are as follows: ； Wherein, GP(θ) represents a Gaussian process, used to fit the relationship between hyperparameters and loss function; Let μ(θ) represent a Gaussian distribution, where μ(θ) is the mean function; k(θ,θ') is the covariance function; γ and λ are adjustment parameters, respectively; Δ mode (θ) represents the variance of mode mutation; EI(θ) is the expected boost; E[ ] represents the expectation operation; L(θ) is the hyperparameter. The loss function is L(θ). best Var(ΔV) is the current optimal loss; CE-p ) represents V CE-p The first-order difference variance of the signal.

9. The method according to claim 1, characterized in that, In step S3, the joint training uses the extended mean squared error loss function: ； Where N is the size of the training batch, y i It is the actual value. The first term is the predicted value; the second term is the feature smoothness constraint, where M is the time series length. Let be the embedded feature vector at the j-th time step, δ be the weight coefficient of the feature embedding smoothness loss, and the Adam optimizer is used for optimization training.

Images