CNN-BiLSTM-based IGBT residual life prediction method
By using a hybrid network model based on CNN-BiLSTM, degradation feature sample parameters of IGBT accelerated aging test data are extracted, a feature time series sample set is established and trained, which solves the problem of low accuracy in predicting the remaining lifetime of IGBTs, achieves high-precision lifetime prediction, and improves the predictive maintenance capability of IGBTs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID HEBEI ELECTRIC POWER RES INST
- Filing Date
- 2025-12-04
- Publication Date
- 2026-05-12
AI Technical Summary
Existing methods for predicting the remaining life of IGBTs suffer from low prediction accuracy and poor reliability. Traditional physical model methods rely on idealized assumptions and are sensitive to parameters, while data-driven methods struggle to account for subtle degradation signals and long-term aging trends.
A hybrid network model based on CNN-BiLSTM is adopted. By acquiring IGBT accelerated aging test data, degradation feature sample parameters are extracted, a feature time series sample set is established, and training and testing are performed to build a lifetime prediction model. CNN is used to extract local spatial features, and BiLSTM is used to capture long-term temporal dependencies to achieve accurate prediction.
It significantly improves the accuracy and reliability of IGBT remaining life prediction, can accurately respond to changes in aging characteristics at different stages, provides a reliable basis for predictive maintenance, and ensures the stable operation of power electronic systems.
Smart Images
Figure CN122020502A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power electronic device reliability prediction technology, and in particular to an IGBT remaining lifetime prediction method based on CNN-BiLSTM. Background Technology
[0002] Insulated Gate Bipolar Transistors (IGBTs) are core power switching devices in power electronic systems, widely used in critical fields such as new energy vehicles, smart grids, and rail transportation. Their operational reliability directly determines the safety and efficiency of the entire system. During long-term service, IGBTs are affected by factors such as junction temperature fluctuations and electrical load cycles, leading to progressive aging failures such as bond wire fatigue detachment, solder layer void expansion, and metallization layer reconstruction. These failures can ultimately result in sudden device failures, incurring not only high maintenance costs but also potential safety accidents. Therefore, accurately predicting their remaining service life is a core requirement for predictive maintenance.
[0003] Currently, IGBT remaining lifetime prediction methods are mainly divided into two categories: traditional physical model methods and data-driven methods. Traditional physical model methods, represented by Wiener processes and rainflow counting methods, achieve prediction by establishing mathematical analytical models of the degradation process. On the one hand, these methods rely excessively on idealized assumptions such as monotonic degradation and normal distribution, while the actual IGBT aging process is affected by multiple coupled factors and often exhibits nonlinear and non-monotonic fluctuation characteristics, resulting in poor model adaptability. On the other hand, model parameters are sensitive to data quality, and sensor noise or outliers can easily cause parameter estimation bias, making it difficult to meet the high-precision requirements of engineering scenarios.
[0004] With the development of sensing technology and machine learning, data-driven methods have become a research hotspot. However, prediction schemes relying on a single model architecture are difficult to comprehensively predict subtle degradation signals and long-term aging trends, and the prediction accuracy is difficult to meet practical needs.
[0005] This shows that traditional IGBT remaining life prediction schemes suffer from low prediction accuracy and poor reliability. Summary of the Invention
[0006] This invention provides a CNN-BiLSTM-based method for predicting the remaining lifetime of IGBTs, which addresses the shortcomings of traditional IGBT remaining lifetime prediction schemes, such as low prediction accuracy and poor reliability.
[0007] This invention provides a method for predicting the remaining lifetime of IGBTs based on CNN-BiLSTM, comprising: Acquire IGBT accelerated aging test data and extract degradation characteristic sample parameters from the IGBT accelerated aging test data; Based on the degradation feature sample parameters, a feature time series sample set is established, and the feature time series sample set is divided into a training set and a test set; A hybrid network model based on CNN and BiLSTM is established. The hybrid network model is trained using the training set and tested using the test set to obtain a lifespan prediction model. Obtain the measured parameters of the degradation characteristics of the IGBT under test, input the measured parameters of the degradation characteristics into the lifetime prediction model, and obtain the remaining lifetime prediction result.
[0008] According to the IGBT remaining lifetime prediction method based on CNN-BiLSTM provided by the present invention, degradation feature sample parameters are extracted from the IGBT accelerated aging test data, including: The peak value of the turn-off voltage between the collector and emitter in each test cycle is obtained from the IGBT accelerated aging test data. Based on the peak value of the turn-off voltage between the collector and emitter in each test cycle, the degradation characteristic sample parameters are determined.
[0009] According to the IGBT remaining lifetime prediction method based on CNN-BiLSTM provided by the present invention, the peak value of the turn-off voltage between the collector and emitter in each test cycle is obtained from the IGBT accelerated aging test data, including: The test data of the IGBT accelerated aging test data in the early stage of degradation is sampled by a time window of the first set length, and the test data of the IGBT accelerated aging test data in the later stage of degradation is sampled by a time window of the second set length, so as to obtain the peak value of the turn-off voltage between the collector and the emitter in each test cycle. Wherein, the first set length is greater than the second set length.
[0010] According to the IGBT remaining lifetime prediction method based on CNN-BiLSTM provided by the present invention, degradation characteristic sample parameters are determined based on the peak value of the turn-off voltage between the collector and emitter in each test cycle, including: The peak value of the turn-off voltage between the collector and emitter in each test cycle is correlated with the gate current and package temperature collected in advance to obtain the correlation characteristic parameters. The peak value of the turn-off voltage between the collector and emitter in each test cycle and the associated characteristic parameters are used as degradation characteristic sample parameters.
[0011] According to the IGBT remaining lifetime prediction method based on CNN-BiLSTM provided by the present invention, a feature time series sample set is established based on the degradation feature sample parameters, including: The degraded feature sample parameters are subjected to outlier removal and data normalization to obtain standard feature sample parameters; The standard feature sample parameters are reconstructed using time-series samples to obtain multiple time-series sample pairs containing input feature parameters and output lifetime labels; Multiple time series sample pairs are constructed into a feature time series sample set.
[0012] According to the IGBT remaining lifetime prediction method based on CNN-BiLSTM provided by the present invention, the hybrid network model includes: The input layer is used to receive input data and transform it into an input feature vector of a specified dimension. The CNN feature extraction layer is used to extract local spatial features from the input feature vector and output a local feature sequence. The BiLSTM temporal modeling layer is used to capture long-term temporal dependencies in the local feature sequences from both positive and negative directions, and output a comprehensive feature sequence. The output layer is used to perform a nonlinear transformation on the comprehensive feature sequence and output the remaining lifetime prediction result.
[0013] According to the CNN-BiLSTM-based IGBT remaining lifetime prediction method provided by the present invention, the hybrid network model is trained using the training set, including: The training set is preprocessed and then input into the CNN feature extraction layer. The network parameters of the CNN feature extraction layer are optimized using the local feature reconstruction error as the loss function until the local feature reconstruction error meets the iteration termination condition, thus obtaining the pre-trained CNN feature extraction layer. The local feature sequence output by the pre-trained CNN feature extraction layer is input into the BiLSTM temporal modeling layer. The feature weights are assigned using a temporal attention mechanism, and the network parameters of the BiLSTM temporal modeling layer are optimized using the temporal prediction loss as the loss function until the temporal prediction loss meets the iteration termination condition, thus obtaining the BiLSTM temporal modeling layer after targeted training. Based on the pre-trained CNN feature extraction layer and the directed-trained BiLSTM temporal modeling layer, the hybrid network model is iteratively trained to obtain the trained hybrid network model.
[0014] According to the CNN-BiLSTM-based IGBT remaining lifetime prediction method provided by the present invention, a hybrid network model is iteratively trained based on the pre-trained CNN feature extraction layer and the directed-trained BiLSTM temporal modeling layer to obtain a trained hybrid network model, including: The pre-trained CNN feature extraction layer and the targeted trained BiLSTM temporal modeling layer are concatenated to establish a pre-trained hybrid network model. The backpropagation algorithm and adaptive moment estimation optimizer are used to iteratively optimize the model parameters of the pre-trained hybrid network model with the goal of minimizing the mean square error between the predicted and the true values, so as to obtain the trained hybrid network model.
[0015] According to the CNN-BiLSTM-based IGBT remaining lifetime prediction method provided by the present invention, the trained hybrid network model is tested using the test set, including: The test set is input into the trained hybrid network model to obtain the remaining lifetime prediction value for each test sample in the test set; Based on the predicted remaining lifetime and the actual remaining lifetime for each test sample, the root mean square error, mean absolute error, and mean absolute percentage error are calculated respectively. The root mean square error, mean absolute error, and mean absolute percentage error are compared with their respective error thresholds. If the root mean square error, mean absolute error, and mean absolute percentage error are all lower than their respective error thresholds, the model test is deemed to have passed.
[0016] According to the IGBT remaining lifetime prediction method based on CNN-BiLSTM provided by the present invention, IGBT accelerated aging test data are obtained, including: In the aging test scenario, a square wave signal with a frequency of 1kHz, a duty cycle of 40%, and an amplitude of 0-8V is applied to the gate of the IGBT, and the package temperature is controlled at 260-270℃. After the aging test is completed, the accelerated aging test data of the IGBT is obtained.
[0017] The CNN-BiLSTM-based IGBT remaining lifetime prediction method provided by this invention can accurately identify key information reflecting the aging state of IGBTs by acquiring accelerated aging test data of IGBTs and specifically extracting degradation feature sample parameters, thus laying a high-quality data foundation for subsequent prediction. By establishing a feature time-series sample set based on the degradation feature sample parameters and dividing it into training and test sets, the method ensures that the sample data can fully support model training and performance verification. The hybrid network model based on CNN and BiLSTM fully leverages the advantages of CNN in extracting local features and BiLSTM in capturing temporal dependencies, achieving comprehensive learning of complex features during the IGBT aging process. Compared with a single network model, this significantly improves the model's perception and prediction accuracy of aging trends, especially accurately addressing the changes in aging characteristics at different stages throughout the IGBT's lifespan. By acquiring the measured degradation feature parameters of the IGBT under test and inputting them into the trained lifetime prediction model, the remaining lifetime prediction results can be obtained quickly and accurately, providing a reliable basis for predictive maintenance of IGBTs and improving the accuracy and reliability of the IGBT remaining lifetime prediction process, thereby ensuring the stable and efficient operation of power electronic systems. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0019] Figure 1 This is a flowchart illustrating the IGBT remaining lifetime prediction method based on CNN-BiLSTM provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the numerical curve of the peak value of the turn-off voltage in the degraded feature sample parameters before data normalization processing in this embodiment of the invention. Figure 3 This is a schematic diagram of the numerical curve of the peak value of the turn-off voltage in the degradation feature sample parameters after data normalization processing in an embodiment of the present invention. Figure 4 This is a schematic diagram of the data processing principle of the BiLSTM timing modeling layer in an embodiment of the present invention. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0021] The following is combined with Figures 1 to 4 This paper describes the specific implementation process of the IGBT remaining lifetime prediction method based on CNN-BiLSTM provided in the embodiments of the present invention.
[0022] like Figure 1 As shown, the IGBT remaining lifetime prediction method based on CNN-BiLSTM provided in this embodiment of the invention mainly includes the following steps: Step 110: Obtain IGBT accelerated aging test data and extract degradation characteristic sample parameters from the IGBT accelerated aging test data.
[0023] It is understandable that degradation characteristic sample parameters refer to key degradation characteristic information that reflects the aging process of IGBTs. In this embodiment, the accelerated aging test data can be obtained from publicly available accelerated aging test datasets, such as the overcurrent high-temperature aging test dataset published by NASAPCoE Research Center, or obtained through an online monitoring system under actual operating conditions.
[0024] In the aging test scenario, a square wave signal with a frequency of 1kHz, a duty cycle of 40%, and an amplitude of 0-8V can be applied to the gate of the IGBT, and the package temperature can be controlled at 260-270℃. After the aging test is completed, the accelerated aging test data of the IGBT can be obtained.
[0025] Step 120: Based on the parameters of the degraded feature samples, establish a feature time series sample set, and divide the feature time series sample set into a training set and a test set.
[0026] In practical applications, 80% of the feature time series sample set can be divided into the training set and 20% into the test set.
[0027] Step 130: Establish a hybrid network model based on CNN and BiLSTM. Train the hybrid network model using the training set and test the trained hybrid network model using the test set to obtain the lifespan prediction model.
[0028] In this embodiment, a hybrid network model with two grid architectures, CNN (Convolutional Neural Network) and BiLSTM (Bidirectional Long Short Term Memory network), is adopted. CNN can deeply mine local spatial features and short-term patterns in degradation data, and BiLSTM can learn the long-term dependencies of feature sequences from both positive and negative directions, so as to achieve a more comprehensive and accurate modeling of the degradation process, which significantly improves the generalization and robustness of the remaining lifetime prediction process.
[0029] Step 140: Obtain the measured parameters of the degradation characteristics of the IGBT under test, input the measured parameters of the degradation characteristics into the lifetime prediction model, and obtain the remaining lifetime prediction results.
[0030] In practical prediction scenarios, by inputting the measured parameters of the degradation characteristics of the IGBT under test into the lifetime prediction model, the remaining lifetime prediction results can be obtained directly.
[0031] In one embodiment, the degradation characteristic sample parameters are extracted from the IGBT accelerated aging test data, specifically including: First, the peak value of the turn-off voltage between the collector and emitter in each test cycle is obtained from the IGBT accelerated aging test data.
[0032] Understandably, the peak value of the turn-off voltage can effectively reflect the degradation pattern of the parasitic parameters inside the IGBT as it ages, showing a clear monotonic change trend and possessing good lifespan characterization capabilities.
[0033] In one specific implementation, the peak value of the turn-off voltage between the collector and emitter within each test cycle is obtained from the IGBT accelerated aging test data, specifically including: The test data of the IGBT accelerated aging test in the early stage of degradation is sampled by a time window of the first set length, and the test data of the IGBT accelerated aging test in the later stage of degradation is sampled by a time window of the second set length, so as to obtain the peak value of the turn-off voltage between the collector and the emitter in each test cycle.
[0034] The first set length is greater than the second set length.
[0035] In this embodiment, to address the non-uniform characteristics of gradual degradation in the early stages and abrupt changes in the later stages of IGBT acceleration, an adaptive time window sampling method can be used to acquire the peak value of the turn-off voltage. Specifically, for the test data in the early stages of degradation, sliding window sampling is performed using a first predetermined time window, which can be 20 consecutive sampling points, to capture subtle changes in the early stages of degradation. For the test data in the later stages of degradation, sliding window sampling is performed using a second predetermined time window, which can be 10 consecutive sampling points, to focus on the abrupt changes in the later stages of degradation. The above sampling method can ensure both sampling efficiency and the accuracy of the sampled data, effectively solving the problem of poor adaptability of traditional fixed time window sampling methods to non-uniform degradation characteristics.
[0036] Then, based on the peak value of the turn-off voltage between the collector and emitter in each test cycle, the degradation characteristic sample parameters are determined.
[0037] In one specific implementation, degradation characteristic sample parameters are determined based on the peak value of the turn-off voltage between the collector and emitter during each test cycle, specifically including: First, the peak value of the turn-off voltage between the collector and emitter in each test cycle is correlated with the gate current and package temperature collected in advance to obtain the correlation characteristic parameters.
[0038] When calculating associated characteristic parameters in each test cycle, the synchronous alignment of multi-source data can be completed first. Taking the acquisition time of the peak value of the turn-off voltage between the collector and emitter as the reference timestamp, the gate current and package temperature acquired synchronously within ±5ms before and after this timestamp are extracted. The average value of the gate current and the average value of the package temperature are obtained by averaging, ensuring that the peak value of the turn-off voltage, the average value of the gate current, and the average value of the package temperature correspond one-to-one in the time dimension.
[0039] Based on this, three types of correlation calculations can be used to generate correlation feature parameters. First, a linear correlation feature is constructed by calculating the product of the peak value of the turn-off voltage with the average gate current and the average package temperature, quantifying the synergistic effect of electrical and thermal parameters on the peak characteristics. Second, a normalized correlation feature is designed by dividing the peak value of the turn-off voltage by the average gate current and the average package temperature, respectively, to obtain the peak value of the turn-off voltage per unit gate current and the peak value of the turn-off voltage per unit temperature, eliminating the interference of dimensional differences on the effectiveness of the feature. Third, a statistical correlation feature is introduced by calculating the difference between the peak value of the turn-off voltage in the current test cycle and the average peak value of the turn-off voltage in the previous 5 cycles, and then multiplying it by the rate of change of the average gate current and the average package temperature, respectively, to obtain a dynamic correlation feature, capturing the coupling trend of multiple parameters changing over time.
[0040] Finally, the aforementioned linear correlation features, normalized correlation features, and dynamic correlation features are used as correlation feature parameters for each test cycle to improve the ability to characterize the aging state of IGBTs.
[0041] Then, the peak value of the turn-off voltage between the collector and emitter and the associated characteristic parameters in each test cycle are used as degradation characteristic sample parameters.
[0042] In this embodiment, the linear correlation features, normalized correlation features, and dynamic correlation features can be constructed with the peak value of the turn-off voltage into a one-dimensional parameter vector, namely the degradation feature sample parameters.
[0043] In one embodiment, a feature time series sample set is established based on the degradation feature sample parameters, including: First, outlier removal and data normalization are performed on the parameters of the degraded feature samples to obtain the parameters of the standard feature samples.
[0044] Understandably, in the outlier removal stage, abnormal data points caused by measurement noise or interference can be identified and removed; in the data normalization stage, methods such as min-max normalization can be used to linearly transform the parameters of degenerate feature samples to a specific interval, so as to accelerate model training convergence and improve numerical stability.
[0045] Figure 2 The numerical curves of the turn-off voltage spikes in the degraded feature sample parameters before data normalization are shown. Figure 3 The numerical curves of the peak values of the turn-off voltage in the degraded characteristic sample parameters are shown after data normalization.
[0046] Then, time-series sample reconstruction is performed on the standard feature sample parameters to obtain multiple time-series sample pairs containing input feature parameters and output lifetime labels.
[0047] In the temporal sample reconstruction stage, standard feature sample parameters can be reconstructed into supervised learning samples with a fixed time step, that is, temporal sample pairs containing input feature parameters and output lifetime labels, for subsequent model training and testing.
[0048] Finally, multiple time series sample pairs are constructed into a feature time series sample set.
[0049] In one embodiment, the hybrid network model specifically includes: The input layer is used to receive input data and transform it into an input feature vector of a specified dimension.
[0050] The CNN feature extraction layer is used to extract local spatial features from the input feature vector and output a local feature sequence.
[0051] In this embodiment, the CNN feature extraction layer consists of one or more one-dimensional convolutional sub-layers followed by pooling sub-layers stacked alternately. The one-dimensional convolutional sub-layers are responsible for extracting local features from the input feature vector by sliding the convolutional kernel; the pooling sub-layers are responsible for downsampling the local features to retain the main information and reduce the data dimensionality.
[0052] The BiLSTM temporal modeling layer is used to capture long-term temporal dependencies in local feature sequences from both positive and negative directions, and output a comprehensive feature sequence.
[0053] In this embodiment, the BiLSTM temporal modeling layer consists of a forward LSTM sublayer and a backward LSTM sublayer in parallel. It can process local feature sequences from both forward and reverse directions to learn their long-term temporal dependencies. Finally, the outputs from the two directions are integrated at each time step to output a comprehensive feature sequence.
[0054] Figure 4 The specific structure of the BiLSTM timing modeling layer is shown, as follows: Figure 4 As shown, X t-1 X t X t+1 Representing the input features at different time steps in the local feature sequence, Forward LSTM represents the forward LSTM sublayer, which can process the input features in chronological order from t-1 to t+1, capturing information from the past to the present.
[0055] Backward LST refers to the backward LSTM sublayer, which processes input features in reverse time order from t+1 to t-1, capturing information from the future to the present. The two LSTMs in each direction pass their hidden states to the Calculate module at the corresponding time step. The Calculate module fuses the forward and reverse LSTM hidden states at the same time step, usually by concatenation, ultimately obtaining the output features Y at different time steps. t-1 Y t Y t+1 .
[0056] The output layer is used to perform nonlinear transformations on the comprehensive feature sequence and output the remaining lifetime prediction results.
[0057] In this embodiment, the output layer mainly performs further nonlinear combination and dimensional transformation on the comprehensive feature sequence, and finally outputs the remaining lifetime prediction result.
[0058] In one embodiment, training the hybrid network model using a training set specifically includes: First, the training set is preprocessed and then input into the CNN feature extraction layer. The network parameters of the CNN feature extraction layer are optimized using the local feature reconstruction error as the loss function until the local feature reconstruction error meets the iteration termination condition, thus obtaining the pre-trained CNN feature extraction layer.
[0059] Understandably, this embodiment first achieves step-by-step optimization of the CNN feature extraction layer and the BiLSTM temporal modeling layer through layered pre-training before formal training, which can avoid gradient conflict problems caused by direct joint training.
[0060] This step allows the CNN feature extraction layer to accurately capture local degradation patterns in the voltage spike sequence, such as short-term fluctuations and local peak changes. By pre-training the CNN feature extraction layer separately, the interference of the temporal gradient of the BiLSTM temporal modeling layer on the local feature learning of the CNN feature extraction layer during direct joint training can be avoided, ensuring that the features extracted by the CNN feature extraction layer are more in line with the local degradation patterns of IGBTs.
[0061] Then, the local feature sequence output by the pre-trained CNN feature extraction layer is input into the BiLSTM temporal modeling layer. The feature weights are assigned using a temporal attention mechanism, and the network parameters of the BiLSTM temporal modeling layer are optimized using the temporal prediction loss as the loss function until the temporal prediction loss meets the iteration termination condition, thus obtaining the BiLSTM temporal modeling layer after targeted training.
[0062] This step allows the BiLSTM temporal modeling layer to focus on learning the long-term temporal correlations of local features output by the CNN feature extraction layer, especially the rapid degradation trend in the later stages of degradation. In the step of assigning weights using a temporal attention mechanism, feature sequences in the later stages of degradation with less than 20% remaining lifetime are given 1.5 times the attention weight, and feature sequences in the early stages of degradation are given 0.8 times the attention weight. The network parameters of the BiLSTM temporal modeling layer are optimized using temporal prediction loss, and iterated until the loss converges. At this point, the BiLSTM temporal modeling layer is able to capture the temporal dependencies of key degradation stages.
[0063] Finally, based on the pre-trained CNN feature extraction layer and the directed-trained BiLSTM temporal modeling layer, the hybrid network model is iteratively trained to obtain the trained hybrid network model.
[0064] In one specific implementation, the hybrid network model is iteratively trained based on the pre-trained CNN feature extraction layer and the directionally trained BiLSTM temporal modeling layer to obtain the trained hybrid network model, specifically including: First, the pre-trained CNN feature extraction layer and the targeted-trained BiLSTM temporal modeling layer are concatenated to establish a pre-trained hybrid network model.
[0065] Then, the backpropagation algorithm and adaptive moment estimation optimizer are used to iteratively optimize the model parameters of the pre-trained hybrid network model with the goal of minimizing the mean square error between the predicted and the true values, so as to obtain the trained hybrid network model.
[0066] In this embodiment, joint training is performed after the important layer structures have been pre-trained, which can improve the prediction accuracy of the hybrid network model obtained by training.
[0067] In one embodiment, the trained hybrid network model is tested using a test set, specifically including: First, the test set is input into the trained hybrid network model to obtain the remaining lifetime prediction value for each test sample in the test set.
[0068] Then, based on the predicted remaining lifetime and the actual remaining lifetime for each test sample, the root mean square error, mean absolute error, and mean absolute percentage error are calculated respectively.
[0069] Finally, the root mean square error, mean absolute error, and mean absolute percentage error are compared with their respective error thresholds. If the root mean square error, mean absolute error, and mean absolute percentage error are all lower than their respective error thresholds, the model test is considered to have passed.
[0070] For example, the error threshold corresponding to the root mean square error can be set to 0.04, the error threshold for the mean absolute error can be set to 0.035, and the error threshold for the mean absolute percentage error can be set to 0.45, so as to ensure that the prediction accuracy of the hybrid network model based on CNN and BiLSTM provided in this embodiment is better than that of a single LSTM model or a hybrid model of CNN and LSTM.
[0071] In practical applications, the hybrid network model based on CNN and BiLSTM proposed in this invention can be compared and tested with a single LSTM model or a hybrid model of CNN and LSTM on the same dataset. Specifically, the root mean square error, mean absolute error, and mean absolute percentage error are used as evaluation criteria to verify the superiority of each model in terms of prediction accuracy and robustness.
[0072] In summary, the CNN-BiLSTM-based IGBT remaining lifetime prediction method provided in this embodiment of the invention fully utilizes the strong feature extraction capability of CNN and the contextual temporal modeling advantage of BiLSTM. Compared with single models or other hybrid models, it significantly improves the prediction accuracy and robustness of IGBT remaining lifetime.
[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting the remaining lifetime of IGBTs based on CNN-BiLSTM, characterized in that, include: Acquire IGBT accelerated aging test data and extract degradation characteristic sample parameters from the IGBT accelerated aging test data; Based on the degradation feature sample parameters, a feature time series sample set is established, and the feature time series sample set is divided into a training set and a test set; A hybrid network model based on CNN and BiLSTM is established. The hybrid network model is trained using the training set and tested using the test set to obtain a lifespan prediction model. Obtain the measured parameters of the degradation characteristics of the IGBT under test, input the measured parameters of the degradation characteristics into the lifetime prediction model, and obtain the remaining lifetime prediction result.
2. The IGBT remaining lifetime prediction method based on CNN-BiLSTM according to claim 1, characterized in that, Extracting degradation characteristic sample parameters from the IGBT accelerated aging test data includes: The peak value of the turn-off voltage between the collector and emitter in each test cycle is obtained from the IGBT accelerated aging test data. Based on the peak value of the turn-off voltage between the collector and emitter in each test cycle, the degradation characteristic sample parameters are determined.
3. The IGBT remaining lifetime prediction method based on CNN-BiLSTM according to claim 2, characterized in that, The peak value of the turn-off voltage between the collector and emitter within each test cycle is obtained from the IGBT accelerated aging test data, including: The test data of the IGBT accelerated aging test data in the early stage of degradation is sampled by a time window of the first set length, and the test data of the IGBT accelerated aging test data in the later stage of degradation is sampled by a time window of the second set length, so as to obtain the peak value of the turn-off voltage between the collector and the emitter in each test cycle. Wherein, the first set length is greater than the second set length.
4. The IGBT remaining lifetime prediction method based on CNN-BiLSTM according to claim 2, characterized in that, Based on the peak value of the turn-off voltage between the collector and emitter during each test cycle, degradation characteristic sample parameters are determined, including: The peak value of the turn-off voltage between the collector and emitter in each test cycle is correlated with the gate current and package temperature collected in advance to obtain the correlation characteristic parameters. The peak value of the turn-off voltage between the collector and emitter in each test cycle and the associated characteristic parameters are used as degradation characteristic sample parameters.
5. The IGBT remaining lifetime prediction method based on CNN-BiLSTM according to claim 1, characterized in that, Based on the degradation feature sample parameters, a feature time series sample set is established, including: The degraded feature sample parameters are subjected to outlier removal and data normalization to obtain standard feature sample parameters; The standard feature sample parameters are reconstructed using time-series samples to obtain multiple time-series sample pairs containing input feature parameters and output lifetime labels; Multiple time series sample pairs are constructed into a feature time series sample set.
6. The IGBT remaining lifetime prediction method based on CNN-BiLSTM according to claim 1, characterized in that, The hybrid network model includes: The input layer is used to receive input data and transform it into an input feature vector of a specified dimension. The CNN feature extraction layer is used to extract local spatial features from the input feature vector and output a local feature sequence. The BiLSTM temporal modeling layer is used to capture long-term temporal dependencies in the local feature sequences from both positive and negative directions, and output a comprehensive feature sequence. The output layer is used to perform a nonlinear transformation on the comprehensive feature sequence and output the remaining lifetime prediction result.
7. The IGBT remaining lifetime prediction method based on CNN-BiLSTM according to claim 6, characterized in that, Training the hybrid network model using the training set includes: The training set is preprocessed and then input into the CNN feature extraction layer. The network parameters of the CNN feature extraction layer are optimized using the local feature reconstruction error as the loss function until the local feature reconstruction error meets the iteration termination condition, thus obtaining the pre-trained CNN feature extraction layer. The local feature sequence output by the pre-trained CNN feature extraction layer is input into the BiLSTM temporal modeling layer. The feature weights are assigned using a temporal attention mechanism, and the network parameters of the BiLSTM temporal modeling layer are optimized using the temporal prediction loss as the loss function until the temporal prediction loss meets the iteration termination condition, thus obtaining the BiLSTM temporal modeling layer after targeted training. Based on the pre-trained CNN feature extraction layer and the directed-trained BiLSTM temporal modeling layer, the hybrid network model is iteratively trained to obtain the trained hybrid network model.
8. The IGBT remaining lifetime prediction method based on CNN-BiLSTM according to claim 7, characterized in that, Based on the pre-trained CNN feature extraction layer and the targeted-trained BiLSTM temporal modeling layer, the hybrid network model is iteratively trained to obtain the trained hybrid network model, including: The pre-trained CNN feature extraction layer and the targeted trained BiLSTM temporal modeling layer are concatenated to establish a pre-trained hybrid network model. The backpropagation algorithm and adaptive moment estimation optimizer are used to iteratively optimize the model parameters of the pre-trained hybrid network model with the goal of minimizing the mean square error between the predicted and the true values, so as to obtain the trained hybrid network model.
9. The IGBT remaining lifetime prediction method based on CNN-BiLSTM according to claim 1, characterized in that, The trained hybrid network model is tested using the test set, including: The test set is input into the trained hybrid network model to obtain the remaining lifetime prediction value for each test sample in the test set; Based on the predicted remaining lifetime and the actual remaining lifetime for each test sample, the root mean square error, mean absolute error, and mean absolute percentage error are calculated respectively. The root mean square error, mean absolute error, and mean absolute percentage error are compared with their respective error thresholds. If the root mean square error, mean absolute error, and mean absolute percentage error are all lower than their respective error thresholds, the model test is deemed to have passed.
10. The IGBT remaining lifetime prediction method based on CNN-BiLSTM according to claim 1, characterized in that, Obtain IGBT accelerated aging test data, including: In the aging test scenario, a square wave signal with a frequency of 1kHz, a duty cycle of 40%, and an amplitude of 0-8V is applied to the gate of the IGBT, and the package temperature is controlled at 260-270℃. After the aging test is completed, the accelerated aging test data of the IGBT is obtained.