A Deep Attention Fusion Corrected Conformal Prediction Method for IGBT Remaining Useful Life Prediction

Abstract

Insulated gate bipolar transistors (IGBTs) are critical components in power converters, yet their degradation exhibits profound nonlinearity. Conventional prognostic methods consistently struggle with these nonstationary dynamics, environmental noise, data dropouts, and a lack of reliable uncertainty quantification (UQ). To overcome this, this article proposes a remaining useful life (RUL) forecasting framework: a deep attention fusion corrected conformal prediction (CP). After constructing health indicators (HIs), a parallel dual-stream architecture concurrently extracts multiscale temporal fluctuations and latent spectral representations. A bidirectional time–frequency cross-attention mechanism then hierarchically integrates these features to isolate salient degradation patterns. Point forecasts are iteratively refined by a residual correction (RC) module, while a computationally lightweight CP wrapper generates distribution-free prediction intervals (PIs) with strict statistical guarantees. Validation on NASA datasets demonstrates the state-of-the-art predictive accuracy of the proposed framework. Crucially, it maintains remarkable robustness against severe environmental noise and data missingness. Moreover, compared to existing UQ benchmarks, this approach delivers optimally calibrated and tightly concentrated uncertainty bounds, providing a highly scalable solution for risk-aware predictive maintenance.