Physics‑Powered Fault Detection with a New Transformer Mix

A new method tackles the problem of machines giving different signals when they run under varied conditions. Instead of treating each signal as a single block, the approach splits it into two parts: quick bursts that show sudden changes and steady waves that reveal regular patterns. These parts are fed into two separate processing paths—one that looks closely at local details and another that scans the whole signal for global patterns. The two paths then share information through a gated fusion mechanism, letting the model learn both fine‑grained and broad features at several abstraction levels. When aligning data from different operating conditions, the model does not rely only on statistical similarity. It uses a rule from bearing physics: the frequencies that indicate faults stay in a fixed order relative to the machine’s rotation speed.By enforcing this rule, the model builds positive and negative pairs that respect real physical relationships, making the alignment more meaningful. Training combines three goals: correctly classifying faults, keeping cross‑condition pairs distinct or similar as physics demands, and maintaining physical consistency. Weights for these goals shift gradually during training so that the model learns smoothly without one task overpowering another. Tests on two well‑known bearing datasets show the method reaches about 95 % accuracy on one set and 90 % on the other, beating all major competing techniques across 18 different transfer tasks. It also keeps its performance when the data are heavily shifted or noisy, proving robustness. Further experiments that remove parts of the system confirm each component’s importance. Visual inspections of the learned features reveal patterns that match known fault signatures, giving confidence that the physics guidance works as intended.