A Physics‐Aware Lightweight Transformer Network for Intelligent Bearing Fault Diagnosis Under Variable Operating Conditions

ABSTRACT Existing bearing fault diagnosis methods fail to generalise across variable operating conditions while remaining lightweight for industrial edge deployment. We propose PLT‐Bearing, a physics‐aware transformer that adapts vision transformer tokenisation to vibration signals through overlapping convolutional patch embedding, sampling‐frequency‐guided patch sizing and analytically proved amplitude‐invariant self‐attention. With only 0.52 M parameters and 8.2 ms GPU inference latency, PLT‐bearing achieves 99.2% and 95.8% accuracy on the CWRU and Paderborn University benchmarks, respectively, and 90.3% zero‐shot cross‐load transfer accuracy, outperforming recent lightweight CNNs and Mamba‐based state‐space models.