Physics-Guided Multi-Source Representation Learning for Aero-Engine Bearing Fault Diagnosis under Data Scarcity

Abstract This paper proposes the physics-informed dual-stream contrastive diagnostic network (PIDC-Net), a unified framework resolving both difficulties through three integrated mechanisms. The fault-physics encoding stream (FPES) constructs multi-scale depth-wise separable convolutional encoders whose receptive field widths are derived from bearing fault characteristic frequencies and mechanical impulse decay envelopes, embedding kinematic inductive bias without requiring additional labeled data. The cross-sensor alignment stream (CSAS) applies a learnable affine domain bridge before each convolutional block, learning a parametric cross-sensor normalization that eliminates inter-channel distributional discrepancies. An asymmetric InfoNCE-inspired contrastive objective aligns CSAS embeddings with the physically anchored FPES embeddings, thereby multiplying the supervisory signal available from each labeled window. A sparrow-search-algorithm-optimized bidirectional gated recurrent unit classifies the fused representation. On the XJTU-SY and CWRU datasets, PIDC-Net attains 99.89% and 99.92% accuracy under pooled five-fold cross-validation, retains 94.18% accuracy with 10% training data, and maintains 99.76% accuracy at a 2 dB signal-to-noise ratio. Under a leakage-controlled cross-recording split, it achieves 93.67% and 94.52%, leading the strongest baseline by 3.74 and 3.34 points, and maintains accuracy above 97% under ±20% physical-parameter mis-specification.