A Lightweight Transformer-Based Early-Stage Fault Identification Method for Tunnel Boring Machine Main Bearings With Limited Data

Tunnel Boring Machine (TBM) main bearings are critical components that require continuous condition monitoring but are physically inaccessible during tunneling. They feature ultra-large structural dimensions, ultra-low speeds, and alternating loads, creating extreme operational conditions. These factors lead to exceptionally weak, rapidly attenuated vibration signatures and severe difficulty in acquiring labeled measurement data. Conventional vibration-based diagnostic methods struggle to achieve reliable early-stage fault identification under these constraints. To address this, we propose ASAFormer, a lightweight Transformer-based measurement framework designed for weak-signal enhancement and data-efficient fault identification. ASAFormer integrates two core modules: the Weak-Signal Aware Sparse Attention (WSASA) leverages token-level variance modulation with single-head sparse attention to selectively amplify low-energy transient features while suppressing noise. The Selective Convolutional Knowledge Distillation (SCKD) module extracts multi-scale local patterns through parallel convolutional filters and selectively distills the most discriminative measurement features. Stacking WSASA and SCKD builds a hierarchical representation that fuses global weak-signal awareness with local sensitivity under severe noise and data scarcity. Experimental validation on a custom-designed TBM main bearing test rig demonstrates the feasibility of this measurement-driven approach for the early-stage fault identification. ASAFormer achieves over 90.5% accuracy with only 100 samples per class and maintains performance stability when training data is halved.