CN-122020502-A - IGBT residual life prediction method based on CNN-BiLSTM

CN122020502ACN 122020502 ACN122020502 ACN 122020502ACN-122020502-A

Abstract

The invention belongs to the technical field of reliability prediction of power electronic devices, and particularly relates to a CNN-BiLSTM-based IGBT residual life prediction method, which comprises the steps of obtaining IGBT accelerated aging test data and extracting degradation characteristic sample parameters in the IGBT accelerated aging test data; establishing a hybrid network model based on CNN and BiLSTM, training the hybrid network model through the training set, testing the trained hybrid network model through the testing set to obtain a life prediction model, acquiring actual measurement parameters of degradation characteristics of the IGBT to be tested, and inputting the actual measurement parameters of the degradation characteristics into the life prediction model to obtain a residual life prediction result. The scheme improves the accuracy and reliability of the IGBT residual life prediction link.

Inventors

GAO SHUGUO
GENG JIANGHAI
LIU HAOYU
XING CHAO
GUO MENG
Fan Ruidie
ZHANG ZHIGANG
Dai Lujian
ZHANG TIANYUE
HAN CHENG

Assignees

国网河北省电力有限公司电力科学研究院
国家电网有限公司

Dates

Publication Date: 20260512
Application Date: 20251204

Claims (10)

1. The IGBT residual life prediction method based on CNN-BiLSTM is characterized by comprising the following steps of: acquiring IGBT accelerated aging test data, and extracting degradation characteristic sample parameters in the IGBT accelerated aging test data; according to the degradation characteristic sample parameters, a characteristic time sequence sample set is established, and the characteristic time sequence sample set is divided into a training set and a testing set; establishing a hybrid network model based on CNN and BiLSTM, training the hybrid network model through the training set, and testing the trained hybrid network model through the testing set to obtain a life prediction model; and acquiring actual measurement parameters of degradation characteristics of the IGBT to be detected, and inputting the actual measurement parameters of the degradation characteristics into the life prediction model to obtain a residual life prediction result.
2. The CNN-BiLSTM-based IGBT residual life prediction method according to claim 1, wherein extracting degradation characteristic sample parameters in the IGBT accelerated aging test data includes: Acquiring a peak value of turn-off voltage between a collector and an emitter in each test period from the IGBT accelerated aging test data; And determining degradation characteristic sample parameters according to the peak value of the turn-off voltage between the collector and the emitter in each test period.
3. The CNN-BiLSTM-based IGBT remaining life prediction method according to claim 2, wherein obtaining off-voltage spikes between collector and emitter in each test period from the IGBT accelerated aging test data includes: the method comprises the steps of adopting a time window with a first set length to conduct sliding window sampling on test data in the early degradation stage in IGBT accelerated aging test data, and adopting a time window with a second set length to conduct sliding window sampling on test data in the later degradation stage in IGBT accelerated aging test data, so as to obtain a peak value of turn-off voltage between a collector and an emitter in each test period; wherein the first set length is greater than the second set length.
4. The method for predicting the residual life of an IGBT based on CNN-BiLSTM according to claim 2, wherein determining the degradation characteristic sample parameter according to the off-voltage spike between the collector and the emitter in each test period includes: Performing correlation calculation on the off voltage peak value between the collector and the emitter in each test period, and the gate current and the packaging temperature which are synchronously collected in advance to obtain correlation characteristic parameters; And taking the off-voltage peak value between the collector and the emitter and the related characteristic parameters in each test period as degradation characteristic sample parameters.
5. The CNN-BiLSTM-based IGBT remaining life prediction method according to claim 1, wherein establishing a feature timing sample set from the degraded feature sample parameters includes: performing outlier rejection and data normalization processing on the degradation characteristic sample parameters to obtain standard characteristic sample parameters; reconstructing time sequence samples of the standard characteristic sample parameters to obtain a plurality of time sequence sample pairs comprising input characteristic parameters and output life labels; A plurality of time series sample pairs are constructed as a feature time series sample set.
6. The CNN-BiLSTM-based IGBT remaining life prediction method according to claim 1, wherein the hybrid network model includes: The input layer is used for receiving input data and converting the input data into input feature vectors with set dimensions; The CNN feature extraction layer is used for extracting local spatial features in the input feature vector and outputting a local feature sequence; BiLSTM a time sequence modeling layer, which is used for capturing the long-term time sequence dependency relationship in the local feature sequence from the positive direction and the negative direction and outputting a comprehensive feature sequence; And the output layer is used for carrying out nonlinear transformation on the comprehensive characteristic sequence and outputting a residual life prediction result.
7. The CNN-BiLSTM-based IGBT remaining life prediction method according to claim 6, wherein training the hybrid network model by the training set includes: Inputting the training set into the CNN feature extraction layer after preprocessing, and optimizing network parameters of the CNN feature extraction layer by taking a local feature reconstruction error as a loss function until the local feature reconstruction error meets an iteration termination condition to obtain a pre-trained CNN feature extraction layer; Inputting the local feature sequence output by the pre-trained CNN feature extraction layer into the BiLSTM time sequence modeling layer, adopting a time sequence attention mechanism to distribute feature weights, and optimizing network parameters of the BiLSTM time sequence modeling layer by taking time sequence prediction loss as a loss function until the time sequence prediction loss meets iteration termination conditions to obtain a BiLSTM time sequence modeling layer after directional training; and performing iterative training on the hybrid network model based on the pre-trained CNN feature extraction layer and the directional trained BiLSTM time sequence modeling layer to obtain a trained hybrid network model.
8. The CNN-BiLSTM-based IGBT remaining life prediction method according to claim 7, wherein iteratively training a hybrid network model based on the pre-trained CNN feature extraction layer and the directionally trained BiLSTM timing modeling layer to obtain a trained hybrid network model, comprising: The pre-trained CNN feature extraction layer and the directional trained BiLSTM time sequence modeling layer are connected in series, and a pre-trained mixed network model is built; And adopting a back propagation algorithm and a self-adaptive moment estimation optimizer to iteratively optimize model parameters of the pre-trained hybrid network model with the aim of minimizing the mean square error between the predicted value and the true value, thereby obtaining the trained hybrid network model.
9. The CNN-BiLSTM-based IGBT remaining life prediction method according to claim 1, wherein testing the trained hybrid network model by the test set includes: Inputting the test set into a trained hybrid network model to obtain a residual life prediction value of each test sample in the test set; according to the predicted value of the residual life and the true value of the residual life of each test sample, respectively calculating to obtain a root mean square error, an average absolute error and an average absolute percentage error; And comparing the root mean square error, the average absolute error and the average absolute percentage error with the respective corresponding error thresholds, and if the root mean square error, the average absolute error and the average absolute percentage error are lower than the respective corresponding error thresholds, judging that the model test is passed.
10. The method for predicting the residual life of an IGBT based on CNN-BiLSTM according to claim 1, wherein obtaining IGBT accelerated aging test data includes: in an aging test scene, square wave signals with the frequency of 1kHz, the duty ratio of 40% and the amplitude of 0-8V are applied to the grid electrode of the IGBT, the packaging temperature is controlled at 260-270 ℃, and after the aging test is finished, the IGBT accelerated aging test data are obtained.

Description

IGBT residual life prediction method based on CNN-BiLSTM Technical Field The invention relates to the technical field of reliability prediction of power electronic devices, in particular to a CNN-BiLSTM-based IGBT residual life prediction method. Background The IGBT (Insulated Gate Bipolar Transistor ) is used as a core power switch device of a power electronic system, is widely applied to key fields of new energy automobiles, smart grids, rail transit and the like, and the running reliability of the IGBT directly determines the safety and the efficiency of the whole system. In the long-term service process of the IGBT, the IGBT is influenced by factors such as junction temperature fluctuation, electrical load circulation and the like, progressive aging failure such as fatigue falling of bonding wires, hole expansion of a welding layer, reconstruction of a metallization layer and the like can occur in the IGBT, and finally, sudden failure of the device is caused, high maintenance cost is brought, safety accidents are more likely to be caused, and therefore, the residual service life of the IGBT is accurately predicted to become a core requirement for realizing predictive maintenance. At present, the IGBT residual life prediction method is mainly divided into a traditional physical model method and a data driving method. The traditional physical model method is represented by a wiener process, a rain flow counting method and the like, and prediction is realized by establishing a mathematical analysis model of a degradation process, on one hand, the method excessively depends on idealized assumptions such as monotone degradation, normal distribution and the like, and the actual IGBT aging process is influenced by multi-factor coupling, and often presents nonlinear and non-monotone fluctuation characteristics, so that the model suitability is poor, on the other hand, model parameters are sensitive to data quality, sensor noise or abnormal values easily cause parameter estimation deviation, and the high-precision requirement of engineering scenes is difficult to meet. Along with the development of sensing technology and machine learning, a data driving method becomes a research hot spot, but a prediction scheme relying on a single model architecture is difficult to comprehensively predict a fine degradation signal and a long-term aging trend, and the prediction precision is difficult to meet the actual demands. Therefore, the traditional IGBT residual life prediction scheme has the technical problems of low prediction precision and poor reliability. Disclosure of Invention The invention provides a CNN-BiLSTM-based IGBT residual life prediction method, which is used for solving the defects of low prediction precision and poor reliability of a traditional IGBT residual life prediction scheme. The invention provides a CNN-BiLSTM-based IGBT residual life prediction method, which comprises the following steps: acquiring IGBT accelerated aging test data, and extracting degradation characteristic sample parameters in the IGBT accelerated aging test data; according to the degradation characteristic sample parameters, a characteristic time sequence sample set is established, and the characteristic time sequence sample set is divided into a training set and a testing set; establishing a hybrid network model based on CNN and BiLSTM, training the hybrid network model through the training set, and testing the trained hybrid network model through the testing set to obtain a life prediction model; and acquiring actual measurement parameters of degradation characteristics of the IGBT to be detected, and inputting the actual measurement parameters of the degradation characteristics into the life prediction model to obtain a residual life prediction result. According to the IGBT residual life prediction method based on CNN-BiLSTM provided by the invention, the degradation characteristic sample parameters in the IGBT accelerated aging test data are extracted, and the method comprises the following steps: Acquiring a peak value of turn-off voltage between a collector and an emitter in each test period from the IGBT accelerated aging test data; And determining degradation characteristic sample parameters according to the peak value of the turn-off voltage between the collector and the emitter in each test period. According to the method for predicting the residual life of the IGBT based on the CNN-BiLSTM, which is provided by the invention, the off-voltage peak value between the collector and the emitter in each test period is obtained from the accelerated aging test data of the IGBT, and the method comprises the following steps: the method comprises the steps of adopting a time window with a first set length to conduct sliding window sampling on test data in the early degradation stage in IGBT accelerated aging test data, and adopting a time window with a second set length to conduct sliding window sampling on test data in the lat