Physics‑Powered Fault Detection with a New Transformer Mix
A novel method addresses the challenge of machines producing divergent signals when operating under varied conditions.
Instead of treating each signal as a single block, the approach splits it into two complementary parts:
- Quick bursts that capture sudden changes
- Steady waves that reveal regular patterns
These parts are fed into two distinct processing paths:
- Local‑detail path – focuses on fine‑grained, short‑range features.
- Global‑pattern path – scans the entire signal for broad, long‑range patterns.
The two paths share information through a gated fusion mechanism, allowing the model to learn both fine‑grained and broad features at multiple abstraction levels.
Physics‑Based Alignment
When aligning data from different operating conditions, the model goes beyond statistical similarity.
It incorporates a rule derived from bearing physics: the fault‑indicating frequencies remain in a fixed order relative to the machine’s rotation speed.
By enforcing this rule, the model constructs positive and negative pairs that respect real physical relationships, resulting in more meaningful alignment.
Multi‑Objective Training
Training optimizes three goals simultaneously:
- Fault classification accuracy
- Cross‑condition pair consistency – pairs are kept distinct or similar as physics demands.
- Physical consistency preservation
The relative weights of these objectives shift gradually during training, ensuring smooth learning without any single task dominating.
Results
- Accuracy: ~95 % on one benchmark dataset, 90 % on another.
- Transfer tasks: Outperforms all major competing techniques across 18 different transfer scenarios.
- Robustness: Maintains performance under heavy data shifts and noise.
Ablation studies confirm the importance of each component. Visual inspections of learned features show patterns that align with known fault signatures, validating the effectiveness of the physics guidance.
