A dynamic grid KAN network method for IGBT module degradation trend prediction
By using mutual information and XGBoost to filter features, combined with KPCA dimensionality reduction and dynamic grid KAN network, the problems of feature redundancy and training instability in IGBT module degradation trend prediction are solved, and high-precision degradation trend prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- UNIV OF SCI & TECH BEIJING
- Filing Date
- 2026-04-28
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies struggle to effectively address the feature redundancy problem in predicting the degradation trend of IGBT modules. Fixed architecture models are ill-suited to adapting to non-stationary signal changes, and KAN networks exhibit poor training stability during dynamic grid adjustments.
Mutual information and XGBoost are used to filter degradation-sensitive features, combined with KPCA dimensionality reduction, to construct a dynamic grid KAN network. The grid structure is adaptively adjusted through B-spline function, and a structure-aware training strategy is introduced to improve training stability.
It achieves high-precision prediction of IGBT module degradation trends, reduces feature redundancy, and enhances the model's adaptability and training stability, especially its ability to capture local mutations.
Smart Images
Figure CN122432656A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power electronic device health management and prediction technology, and in particular relates to a dynamic grid KAN network method for predicting the degradation trend of IGBT modules. Background Technology
[0002] As a core power semiconductor device, IGBT modules are widely used in new energy vehicles, industrial motors, and other fields. With the rapid development of related industries, the application demand for IGBTs continues to grow, and their operational reliability issues are becoming increasingly prominent. Because IGBT modules operate under high-voltage, high-current, and thermal cycling conditions for extended periods, internal materials are prone to fatigue damage and performance degradation, which is further reflected in monitoring signals such as collector-emitter voltage, current, and package temperature. Especially in the later stages of their lifespan, these signals typically exhibit more pronounced fluctuations, localized abrupt changes, and distribution drift, displaying significant non-stationary characteristics.
[0003] For predicting IGBT degradation trends, existing research typically employs one approach: extracting time-domain, frequency-domain, or time-frequency-domain features to characterize the device's degradation state. While this approach can reflect degradation information to some extent, it often generates a large number of candidate features. Without an effective filtering mechanism, this can easily introduce feature redundancy and dimensionality inflation, increasing modeling complexity and hindering the effective representation of key degradation information. Another approach uses data-driven models such as convolutional neural networks to directly learn the degradation trend. However, these models mostly employ fixed network architectures, making it difficult to adaptively adjust predictive capabilities based on the stage-specific differences in the IGBT degradation process. This can easily lead to wasted resources in the early stages and insufficient fitting in the later stages.
[0004] KAN networks possess strong nonlinear fitting capabilities and show promise for applications in degradation modeling. However, existing KAN networks typically employ a fixed spline mesh structure, with its mesh resolution remaining constant after training begins. This makes it difficult to dynamically adjust the expressive power based on changes in signal nonstationarity. Furthermore, dynamically changing the mesh structure during training can easily lead to parameter mapping perturbations and training instability.
[0005] Therefore, how to construct an IGBT degradation trend prediction method that can adaptively adjust the grid structure according to the non-stationarity of the signal and maintain training stability during dynamic grid changes has become a technical problem that urgently needs to be solved in this field. Summary of the Invention
[0006] To address the technical problems of redundant degradation features, the inability of fixed architecture models to adapt to the phased non-stationary changes in IGBT degradation signals, and the poor training stability of KAN networks during dynamic grid adjustment, this invention provides a method for predicting the non-stationary degradation trend of IGBT modules, which has good training stability and strong adaptability.
[0007] To achieve the above objectives, the technical solution adopted by this invention is: a dynamic mesh KAN network method for predicting the degradation trend of IGBT modules, the specific steps of which are as follows:
[0008] Step S1: Collect multi-source monitoring signals during the operation of the IGBT module (Insulated Gate Bipolar Transistor). The multi-source monitoring signals include collector-emitter current, collector-emitter voltage, and package temperature.
[0009] Step S2: Normalize the multi-source monitoring signals;
[0010] Step S3: Perform sliding window processing on the normalized multi-source monitoring signals, construct a sliding window dataset and extract dimensional and dimensionless statistical features; on this basis, use mutual information and XGBoost (an optimized distributed gradient boosting decision tree machine learning library) to help filter the top ten degradation-sensitive features with the highest score weights.
[0011] Step S4: Perform KPCA (Kernel Principal Component Analysis) to reduce the dimensionality of degradation-sensitive features and construct a health index to characterize the degradation trend of IGBTs;
[0012] Step S5: Input the sliding window dataset into the temporal feature extraction network RNN (Recurrent Neural Network) to extract the latent temporal features in the degradation process;
[0013] Step S6: Input the latent temporal features into the dynamic grid KAN network (neural network architecture), and adaptively adjust the spline grid structure according to the distribution and drift of the degradation signal to obtain the degradation trend prediction result;
[0014] Step S7: Introduce a structure-aware training strategy during the training of the dynamic grid KAN network to suppress parameter perturbations and gradient oscillations caused by grid switching, thereby improving training stability.
[0015] Furthermore, in step S2, the normalization process specifically includes: standardizing the multi-source monitoring signals separately so that different physical quantities can be represented on a unified scale, thereby reducing the impact of dimensional and amplitude differences on subsequent feature extraction and model training.
[0016] Further, step S3 specifically includes: step S31, sliding window processing is to segment the normalized multi-source monitoring signal using a time window of preset length to construct a time series sample for degradation modeling; dimensional statistical features and dimensionless statistical features include at least statistical quantities that can reflect the signal amplitude, fluctuation degree, distribution pattern and impact characteristics.
[0017] Step S32: Use mutual information to evaluate the relevance of the extracted candidate features to measure the degree of association between each candidate feature and the degradation state; use XGBoost to assist in the evaluation of the importance of the candidate features, and select a subset of degradation-sensitive features based on the comprehensive evaluation results of mutual information and XGBoost to reduce feature redundancy and retain key degradation information.
[0018] Further, step S4 specifically includes: using kernel principal component analysis (KPCA) to perform nonlinear dimensionality reduction on the screened degradation-sensitive features, extracting the feature components that dominate degradation changes, and constructing a health index to characterize the degradation trend of IGBT modules; the health index is used to enhance the monotonicity, trend, and predictability of degradation characterization.
[0019] Further, step S5 specifically includes: the temporal feature extraction network is a recurrent neural network; the temporal feature extraction network extracts the potential temporal dependencies in the degradation process to obtain potential temporal features for subsequent dynamic grid KAN network input.
[0020] Furthermore, step S6 specifically includes:
[0021] Step S61: The dynamic mesh KAN network includes a spline mesh structure for performing nonlinear mapping. The spline mesh structure is adaptively adjusted according to the nonstationarity changes or distribution drift of the input features to improve the model's ability to fit signal patterns at different degradation stages.
[0022] Step S62, adaptively adjusting the spline mesh structure includes: when the input signal is in a relatively stable stage in the early stage of degradation, a sparse mesh structure is used for modeling; when the input signal is in a high-fluctuation stage in the later stage of degradation or when there is obvious distribution drift, a dense mesh structure is used for modeling, or the mesh is refined in local areas to improve the ability to express local abrupt changes and high-frequency fluctuations.
[0023] Step S63: The spline function in the dynamic mesh KAN network is constructed using the B-spline function, and the basic activation function is the SiLU function; the mesh range is dynamically determined according to the distribution of input features, and the continuity of the mapping function is maintained during the mesh adjustment process.
[0024] Furthermore, step S7 specifically includes:
[0025] Step S71: The structure-aware training strategy is achieved by constructing a joint optimization objective, which includes a prediction loss term and a regularization term used to constrain the mesh evolution process, so as to reduce parameter perturbations and gradient oscillations caused by dynamic mesh changes.
[0026] Step S72: The structure-aware training strategy includes at least one of the following: L1 regularization constraint, entropy regularization constraint, state constraint term, gradient pruning strategy, or parameter mapping preservation strategy; wherein, the state constraint term is used to enhance the model's ability to identify structural features at different degradation stages.
[0027] Step S73: The structure-aware training strategy jointly optimizes the model through the total loss function, which includes at least the mean squared error loss term, the regularization loss term, and the state constraint term.
[0028] Step S74, the state constraint term is constructed as follows: based on the health index sequence obtained in step S4, the corresponding degradation stage state label is generated by the clustering algorithm, and training constraints are constructed according to the state label to enhance the model's ability to perceive changes in the degradation stage during training; the clustering algorithm is preferably the K-means++ algorithm.
[0029] Step S75: After the spline mesh structure is switched or refined, the parameters before and after the adjustment are mapped and initialized or smoothed to reduce the disturbance caused by the mesh change to the model training process and improve the stability and convergence performance of the training process.
[0030] Furthermore, the degradation trend prediction results obtained in steps S6 and S7 specifically include at least one of the current degradation state value, the degradation trend value at future times, and the predicted value of health indicators.
[0031] Compared with the prior art, the specific beneficial effects of this invention are reflected in:
[0032] I. This invention effectively reduces feature redundancy through a feature selection mechanism that combines mutual information and XGBoost, and enhances the physical consistency of degradation representation by constructing a health index with monotonicity and trend through KPCA, thus providing high-quality input for subsequent prediction models.
[0033] Second, this invention embeds a dynamic mesh adaptive mechanism into the B-spline function approximation layer of the KAN network, realizing dynamic switching between sparse and dense meshes according to the non-stationary changes of the signal. This solves the problem that the existing fixed mesh structure of the KAN network is difficult to adapt to the modeling needs of the stable stage in the early stage of degradation and the high-fluctuation stage in the later stage of degradation at the same time.
[0034] Third, this invention introduces a structure-aware joint regularization strategy, which suppresses redundant spline coefficients through L1 sparse regularization, alleviates instability during grid evolution through entropy regularization, and stabilizes the optimization process by combining gradient pruning technology. This effectively reduces parameter disturbances and gradient oscillations caused by dynamic grid switching, ensuring the stability and convergence performance of model training.
[0035] Fourth, this invention extracts potential temporal dependencies through a temporal feature extraction network and combines the adaptive modeling capability of the dynamic grid KAN network to improve the prediction accuracy in the IGBT non-stationary degradation prediction task, especially enhancing the ability to capture local mutation features in the near-failure stage. Attached Figure Description
[0036] Figure 1 This is a flowchart of a dynamic mesh KAN network method for predicting the degradation trend of IGBT modules.
[0037] Figure 2 This is a flowchart illustrating the feature selection and health indicator construction process of a dynamic mesh KAN network method for predicting the degradation trend of IGBT modules.
[0038] Figure 3 This is a flowchart of the dynamic mesh KAN network prediction and training process for a dynamic mesh KAN network method for predicting the degradation trend of IGBT modules.
[0039] Figure 4 This is a flowchart of the mesh switching process for a dynamic mesh KAN network method for predicting the degradation trend of IGBT modules. Detailed Implementation
[0040] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.
[0041] A dynamic mesh KAN network method for predicting IGBT module degradation trends, according to Figure 1 The main flowchart shown first executes step S100, which collects multi-source monitoring signals during the operation of the IGBT module. In this embodiment, the multi-source monitoring signals include collector-emitter current, collector-emitter voltage, and package temperature. A high-precision data acquisition card is used for data acquisition, with a sampling frequency set to 10kHz. The acquisition duration covers the entire life cycle of the IGBT module from its healthy state to failure.
[0042] Step S110 is executed to normalize the acquired multi-source monitoring signals. Specifically, the Z-score normalization method is used to normalize each signal separately, so that different physical quantities can be characterized on a uniform scale. The normalization formula is:
[0043]
[0044] in, For the k-th standardized monitoring signal value, For the kth original detection signal value, The sample mean. The standard deviation is the sample value. After standardization, the mean of each signal is 0 and the standard deviation is 1, eliminating the influence of differences in dimensions and amplitude on subsequent feature extraction and model training.
[0045] Step S120 is executed to perform sliding window processing on the normalized monitoring signal to construct a sliding window dataset. In this embodiment, a time window of preset length is used to segment the normalized signal. The window length is set to 32 time points, and the sliding step size is set to 1, generating approximately 300,000 sliding window samples for subsequent degradation modeling.
[0046] like Figure 2 As shown, this embodiment first executes step S210 to extract time-domain statistical features from the sliding window dataset. These time-domain statistical features include dimensional features and dimensionless features. Dimensional features include mean, variance, root mean square, and peak value, which reflect the energy fluctuations and amplitude growth of the signal. Dimensionless features include skewness, kurtosis, peak factor, and impulse factor, which characterize the distribution pattern and impulse characteristics of the signal. They are insensitive to changes in signal amplitude and can effectively detect early anomalies.
[0047] In step S220, mutual information is used to evaluate the relevance of the extracted candidate features to measure the degree of association between each candidate feature and the degradation state. The formula for calculating mutual information is:
[0048]
[0049] in, For mutual information between random variables x and y, Indicates double integral, Let be the differential element of the integration variables x and y, representing a double integration over the domains of x and y. Let be the joint probability density function. and These are the edge probability density functions. Features are ranked according to their mutual information scores, and the top 10 features with the highest scores are selected as a candidate feature subset. XGBoost is then used to further evaluate the importance of the candidate features. Combining the evaluation results of mutual information and XGBoost, a degradation-sensitive feature subset is finally selected, thereby reducing feature redundancy and retaining key degradation information.
[0050] Step S230 involves using kernel principal component analysis (KPCA) to perform nonlinear dimensionality reduction on the screened degradation-sensitive features. In this embodiment, a Gaussian radial basis kernel function is used, and the kernel parameters are determined through cross-validation. The first principal component is extracted as the fusion health index (HI), calculated using the following formula:
[0051]
[0052] in, Let be the health indicator value at time t. It is a nonlinear mapping function. Let be the degradation-sensitive feature vector at time t. The first principal component is the eigenvector; the constructed health index has good monotonicity, trend and predictability, and can effectively characterize the degradation trend of IGBT modules.
[0053] Step S240 involves generating degradation stage state labels based on the health indicator sequence using the K-means++ clustering algorithm. In this embodiment, the health indicator sequence is clustered into four stages: initial stage, middle stage, early-to-late stage, and near-failure stage. Training constraints are constructed based on the generated state labels for subsequent model training to enhance the model's ability to perceive changes in degradation stages.
[0054] like Figure 3 As shown, in this embodiment, step S310 is first executed, whereby the sliding window dataset is input into the temporal feature extraction network RNN to extract latent temporal features during the degradation process. In this embodiment, the RNN adopts a two-layer structure with 64 and 128 hidden units respectively, and the activation function is Tanh. The update formula of the RNN is:
[0055]
[0056] in, Let be the hidden state vector of the recurrent neural network at time t. Let be the input vector at time t. The hyperbolic tangent activation function is used. The input is the hidden weight matrix. The weight matrix is used to hide the elements. and Both are bias vectors. Let be the hidden state vector at time t-1. The latent temporal dependencies in the degradation process are extracted using an RNN to obtain the latent temporal feature vector for subsequent input to the dynamic mesh KAN network.
[0057] Execute step S320 to initialize the spline mesh. In this embodiment, the initial mesh is set to sparse mode, the number of mesh intervals G=5, and the mesh range is initialized to... The KAN network uses third-order B-spline basis functions, with SiLU as the fundamental activation function. The parameterization of the spline function is as follows:
[0058]
[0059] in, Let spline activation functions be used on the edges of the KAN network. For third-order B-spline basis functions, For learnable spline coefficients, These are the weighting coefficients of the basis functions. These are the weighting coefficients of the spline basis functions. This represents the total number of intervals in the spline grid. is the learnable coefficient of the k-th mesh interval.
[0060] Execute step S330 to enter the training loop. In each training round, first calculate the predicted value using the current grid, and then calculate the validation loss, which uses mean squared error (MSE) as the evaluation metric.
[0061] Execute step S340 to determine whether mesh adjustment is triggered. This embodiment employs a two-stage triggering mechanism:
[0062] Early stage: In the early training stage (first 10 epochs), a periodic triggering strategy is adopted, triggering grid adjustment once every 5 epochs to quickly adapt to the initial data distribution.
[0063] Mid-to-late stage: In the mid-to-late stage of training (after 10 epochs), a performance-driven triggering strategy is adopted. When the validation loss does not decrease for 5 consecutive epochs, it is determined that distribution drift has occurred and grid adjustment is triggered.
[0064] If grid adjustment is triggered, proceed to step S350 and enter... Figure 4 The mesh refinement and mode switching process is shown; if mesh adjustment is not triggered, step S360 is executed to construct the total loss function.
[0065] like Figure 4 As shown, the mesh refinement and mode switching process includes the following steps:
[0066] Step S410: Obtain the current spline mesh parameters, including the number of mesh intervals G=5 and the mesh range. spline coefficients Simultaneously, obtain the hidden layer representation vector output by the RNN. .
[0067] Step S420: Calculate the 5th and 95th percentiles of the hidden layer representation to determine the dense distribution area of the data.
[0068] Step S430: Insert new grid nodes in densely distributed areas to divide the original grid intervals in two and increase the number of grid intervals.
[0069] Step S440: Expand the grid range outward by 10% to ensure coverage of possible future input values.
[0070] Step S450: Update the grid coefficients using the least squares method to maintain the continuity of the function shape.
[0071] Step S460: Switch the spline mesh from sparse mode to dense mode, with the number of mesh intervals G=15.
[0072] Step S470: Output the updated mesh parameters, including the number of mesh intervals, the expanded mesh range, and the updated spline coefficients, and return. Figure 3 Step S330 continues the forward propagation.
[0073] Execute step S360 to construct the total loss function. In this embodiment, the total loss function includes three parts: prediction error loss, regularization loss, and state constraint loss, calculated as follows:
[0074]
[0075] in: For the total loss function, The loss is the mean square error prediction loss; This is an L1 sparse regularization term used to suppress redundant spline coefficients; This is the control point entropy regularization term, used to prevent instability during grid evolution; For state constraint loss, based on Figure 2 Construction of state labels generated by K-means++ clustering; This is the hyperparameter for regularization balancing.
[0076] Step S370 is executed: backpropagation and parameter update. The Adam optimizer is used to update the network parameters, with an initial learning rate of 0.0003 and a batch size of 64. Gradient clipping is used to stabilize the optimization process, and the maximum gradient norm is set to 1. After the spline mesh structure is switched or refined, the least squares method is used to map and initialize the parameters before and after the adjustment, maintaining the continuity of the function shape and reducing the disturbance of mesh changes to the training process.
[0077] Execute step S380 to determine whether the iteration termination condition is met. In this embodiment, the iteration termination condition is reaching the maximum number of training epochs (50 epochs) or the validation loss not decreasing for 10 consecutive epochs. If the termination condition is not met, return to step S330 to continue the next round of training; if the termination condition is met, exit the training loop, execute step S390, and output the degradation trend prediction result.
[0078] The degradation trend prediction results include at least one of the following: the degradation state value at the current moment, the degradation trend value at future moments, and the predicted value of health indicators.
[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the scope of the present invention.
Claims
1. A dynamic mesh KAN network method for predicting the degradation trend of IGBT modules, characterized in that, The specific steps are as follows: Step S1: Collect multi-source monitoring signals during the operation of the IGBT module. The multi-source monitoring signals include collector-emitter current, collector-emitter voltage, and package temperature. Step S2: Normalize the multi-source monitoring signals; Step S3: Perform sliding window processing on the normalized multi-source monitoring signals, construct a sliding window dataset, and extract dimensional and dimensionless statistical features; based on this, use mutual information and XGBoost to assist in screening to obtain degradation-sensitive features; Step S4: Perform KPCA dimensionality reduction on the degradation-sensitive features to construct a health index for characterizing the degradation trend of IGBTs; Step S5: Input the sliding window dataset into the temporal feature extraction network RNN to extract the latent temporal features in the degradation process; Step S6: Input the potential temporal features into the dynamic mesh KAN network, and adaptively adjust the spline mesh structure according to the distribution and drift of the degradation signal to obtain the degradation trend prediction result; Step S7: Introduce a structure-aware training strategy during the training of the dynamic grid KAN network to suppress parameter perturbations and gradient oscillations caused by grid switching.
2. The dynamic mesh KAN network method for predicting the degradation trend of IGBT modules according to claim 1, characterized in that, In step S2, the multi-source monitoring signals are standardized so that different physical quantities can be characterized on a unified scale.
3. A dynamic mesh KAN network method for predicting the degradation trend of IGBT modules according to claim 2, characterized in that, Step S3 specifically includes: Step S31: The sliding window processing involves segmenting the normalized multi-source monitoring signal using a time window of a preset length to construct a time-series sample for degradation modeling; the dimensional and dimensionless statistical features include at least statistical quantities reflecting the signal amplitude, fluctuation degree, distribution pattern and impact characteristics. Step S32: Use mutual information to evaluate the relevance of the extracted candidate features to measure the degree of association between each candidate feature and the degradation state; use XGBoost to assist in the evaluation of the importance of the candidate features, and select a subset of degradation-sensitive features based on the comprehensive evaluation results of mutual information and XGBoost.
4. A dynamic mesh KAN network method for predicting the degradation trend of IGBT modules according to claim 3, characterized in that, In step S4, kernel principal component analysis (KPCA) is used to perform nonlinear dimensionality reduction on the screened degradation-sensitive features, extract the feature components that dominate degradation changes, and construct a health index to characterize the degradation trend of IGBT modules.
5. A dynamic mesh KAN network method for predicting the degradation trend of IGBT modules according to claim 4, characterized in that, In step S5, the temporal feature extraction network is a recurrent neural network; the temporal feature extraction network extracts the potential temporal dependencies in the degradation process to obtain potential temporal features for subsequent dynamic grid KAN network input.
6. A dynamic mesh KAN network method for predicting the degradation trend of IGBT modules according to claim 5, characterized in that, Step S6 specifically includes: Step S61: The dynamic mesh KAN network includes a spline mesh structure for performing nonlinear mapping, which adaptively adjusts according to the nonstationarity changes or distribution drift of the input features; Step S62, the adaptive adjustment of the spline mesh structure includes: when the input signal is in a relatively stable stage in the early stage of degradation, a sparse mesh structure is used for modeling; when the input signal is in a high-fluctuation stage in the late stage of degradation or when there is obvious distribution drift, a dense mesh structure is used for modeling, or the mesh is refined in local areas; Step S63: The spline function in the dynamic mesh KAN network is constructed using the B-spline function, and the basic activation function is the SiLU function; the mesh range is dynamically determined according to the input feature distribution, and the continuity of the mapping function is maintained during the mesh adjustment process.
7. A dynamic mesh KAN network method for predicting the degradation trend of IGBT modules according to claim 6, characterized in that, Step S7 specifically includes: Step S71: The structure-aware training strategy is achieved by constructing a joint optimization objective, which includes a prediction loss term and a regularization term used to constrain the mesh evolution process. Step S72: The structure-aware training strategy includes at least one of the following: L1 regularization constraint, entropy regularization constraint, state constraint term, gradient pruning strategy or parameter mapping preservation strategy; Step S73: The structure-aware training strategy jointly optimizes the model through a total loss function, which includes at least a mean squared error loss term, a regularization loss term, and a state constraint term. Step S74: The state constraint term is constructed as follows: Based on the health indicator sequence obtained in step S4, a corresponding degradation stage state label is generated through a clustering algorithm, and training constraints are constructed according to the state label. The clustering algorithm is preferably the K-means++ algorithm. Step S75: After the spline mesh structure is switched or refined, perform mapping initialization or smoothing constraints on the parameters before and after the adjustment.
8. The dynamic mesh KAN network method for predicting the degradation trend of IGBT modules according to claim 7, characterized in that, The degradation trend prediction results obtained in steps S6 and S7 specifically include at least one of the current degradation state value, the degradation trend value at future times, and the predicted value of health indicators.