Application of Time Series Model of TCN-BiLSTM Based on Multi-head Attention Mechanism for IGBT Remaining Life Prediction

To address the needs of precise maintenance of power electronic devices and situational awareness of semiconductor power devices, A method for predicting the remaining life of IGBTs based on the fusion of multi-head attention mechanismand TCN-BiLSTM is proposed. Firstly, based on an in-depth analysis of the aging mechanism of IGBT packaging modules, an accelerated aging experimental platform is designed and built. By controlling the junction temperature fluctuations during the power cycling process and applying current to accelerate the aging process of the IGBT module, high-precision data acquisition systems are used to obtain the feature parameter set — the aging data of the collector-emitter saturation voltage V_ce(sat). Secondly, based on the TCN model, the multi-head attention mechanism and BiLSTM neural network are introduced to build the prediction model for predicting and verifying the degradation feature sequence of IGBTs. The results show that under the same conditions, compared with traditional time-series prediction models, the proposed model has higher accuracy without significantly increasing the model complexity and computational burden, which fully verifies the feasibility and efficiency of applying the model to online prediction of IGBT remaining life in engineering practice.