Physics-informed Multi-task Tabular Transformer for Joint Fault Prediction and Remaining Useful Life Estimation of Industrial Equipment

Abstract

Accurate fault prediction and remaining useful life (RUL) estimation are essential for predictive maintenance of industrial equipment. However, existing methods mainly focus on time-series signals and often ignore the heterogeneous characteristics and degradation knowledge in tabular equipment data. This study proposes a Physics-informed Multi-task Tabular Transformer (PIMTT) for joint fault prediction and RUL estimation. The proposed model integrates numerical, categorical, and physics-informed tokens, where a type-aware physics gate is introduced to adaptively enhance degradation-related representations. A multi-task learning strategy is further developed to jointly optimize fault classification and RUL regression. Experiments on a synthetic industrial-equipment condition dataset with approximately 260,000 samples demonstrate that PIMTT achieves an F1-score of 0.8461 for fault prediction and an R2 of 0.9899 for RUL estimation, outperforming several machine learning, Transformer-based, and multi-task baselines. The results verify the effectiveness of combining physical knowledge with tabular Transformer learning for intelligent equipment health assessment.