Physics-Guided Tiny-Mamba Transformer for Machine Fault Detection

A new AI model, the Physics-Guided Tiny-Mamba Transformer (PG-TMT), aims to improve early fault warning in rotating machinery. The compact tri-branch encoder is designed for online condition monitoring under nonstationary conditions, domain shifts, and class imbalance. It uses a depthwise-separable convolutional stem for micro-transients, a Tiny-Mamba state-space branch for long-term degradation, and a lightweight local Transformer for cross-channel resonances. An analytic temporal-to-spectral mapping ties attention to bearing fault-order bands, providing physical plausibility scores. The research was published on arXiv (2601.21293v2) and targets reliability-centered prognostics with low false-alarm rates.