Physics-Informed Multi-Scale Network with Loss-Guided Curriculum Learning for Robust Fault Diagnosis

Abstract

Reliable fault diagnosis of rotating machinery is critical, yet early weak fault impulses are frequently buried in severe compound interference from mechanical harmonics and environmental noise. To address this, a novel Physics-Informed Multi-Scale Network (PI-MSN) is proposed. Variational Mode Decomposition (VMD) is first employed to decouple raw signals into distinct physical frequency bands. Subsequently, a Physics-Informed Channel Attention (PICA) module jointly evaluates the Kurtosis and Root Mean Square of each channel to autonomously highlight fault impulses and suppress harmonic interference. A Multi-Scale Feature Extractor then captures comprehensive fault characteristics. Furthermore, a closed-loop Loss-Guided Smooth Interference Scheduler (LGSIS) dynamically regulates injected interference during training based on real-time loss, fundamentally eradicating catastrophic forgetting. Extensive experiments on the CWRU and HUST datasets demonstrate the framework's exceptional robustness. The highly lightweight PI-MSN achieves state-of-the-art diagnostic accuracy, sustaining over 98\% accuracy even under severe -4 dB compound interference, proving that physical interpretability effectively eliminates the reliance on massive parameter stacking.