Insulated gate bipolar transistors (IGBTs) are the core components of power electronic systems for converting and controlling electrical energy. However, the reliability of IGBT is lower than expected due to the complex environment and operating conditions, and the sudden failure of IGBT will lead to unplanned downtime of the entire system. Therefore, assessing the remaining useful lifetime (RUL) of IGBT will help guide regular maintenance and reduce economic losses. To prevent the sudden failure of IGBT, it is urgent to accurately predict the RUL of IGBT, but most existing methods have low prediction accuracy and high uncertainty. Therefore, this paper proposes an IGBT life prediction model based on optimized long short-term memory (LSTM). Starting from the two cores of the data-driven model, “data” and “model” are optimized and upgraded, which can effectively improve the accuracy and reduce the uncertainty of the model prediction. Firstly, the original condition monitoring (CM) data often contain many contaminated data that appear abnormal due to environmental interference and limitations of measurement technology. Meanwhile, CM data may also appear abnormal when IGBT devices degrade or fail, containing important information to characterize the degradation and failure of IGBT. It cannot be processed simultaneously with contaminated data. The proposed model extracts and enhances degraded features by decomposing the IGBT degraded data into multiple modes using the successive variational mode decomposition (SVMD) technique and then reconstructing the useful modes. Secondly, selecting the model’s hyperparameters will greatly affect the model’s learning ability and training effect. Traditionally, the selection of hyperparameters by the empirical trial-and-error method has contingency and randomness, seriously affecting the performance of the model. The proposed model uses the Bayesian optimization (BO) method to realize the global optimization of multiple hyperparameters in the model through the Gaussian process (GP) proxy model and expectation improvement (EI) acquisition function. Finally, the effectiveness and superiority of the LSTM prediction model based on SVMD and BO are verified with real data. The results show that the predicted RUL is not close to the real RUL by the BO+LSTM method and cannot even meet the 30% error requirement at CM is 160 cycles. In contrast, the errors of the conventional LSTM and RNN methods are large, while the predicted RUL errors using the proposed model meet the requirements for all CM cycles. In addition, the evaluation of the overall performance of the model shows that as an improvement on the RNN, the average relative accuracy (YARA) of the LSTM method improves from 34.65% of RNN to 50.53%, and the average width of prediction interval (WAPI) reduces from 365.3 cycles to 272 cycles. In comparison, the BO+LSTM method has a better prediction performance. Furthermore, the YARA of the proposed model improves to 90.91%, and the WAPI decreases to 169.3 cycles, which is the best performance among several models. Quantitative analysis shows that the proposed model improves the lifetime prediction accuracy by 13% and reduces the prediction uncertainty by 34% compared to the BO+LSTM model. The conclusions can be drawn: (1) The BO algorithm is used to optimize the hyperparameters of the LSTM, which improves the prediction accuracy of the model. (2) The SVMD is used to extract the degraded features of the IGBT, which reduces the uncertainty and improves the accuracy of the model prediction. (3) Compared with other models, the proposed model can maintain a high prediction accuracy with less CM data, and its long-term prediction performance is better. © 2024 China Machine Press. All rights reserved.
