Trend-aware attention LSTM model for remaining useful life prediction of IGBT

Abstract

Insulated Gate Bipolar Transistors (IGBTs) are critical components in power conversion systems, and their aging condition has a direct influence on the operational safety and reliability of electric vehicles, railway traction equipment, renewable energy converters, and industrial drives. To enhance remaining useful life (RUL) prediction under nonlinear degradation, this paper proposes a Trend-Aware Attention Long Short-Term Memory (TAA-LSTM) method. In this framework, Singular Spectrum Analysis (SSA) is first used to extract the main degradation trend from multi-source monitoring data while weakening the effects of random noise and thermal disturbance. The reconstructed trend features are subsequently processed by an LSTM network to learn temporal aging patterns. An attention module is then incorporated to adaptively highlight degradation-sensitive information and key aging intervals. Experimental results based on the NASA IGBT accelerated aging dataset demonstrate that the proposed method achieves more stable and accurate RUL prediction than the compared models, particularly when early-stage degradation information is insufficient.