Physics-Guided Tiny-Mamba Transformer for Reliability-Aware Early Fault Warning

TL;DR

This work tackles the problem of reliable, early fault warning for rotating machinery under nonstationary operation, drift, and severe class imbalance. It introduces PG-TMT, a compact tri-branch encoder that fuses a convolutional stem, Tiny-Mamba state-space dynamics, and a local Transformer, guided by physics priors that align attention with fault-order bands. Decision-making combines EVT-based thresholds with hysteresis to achieve a target false-alarm intensity while stabilizing alarms and enabling interpretable, physics-aligned evidence. Across leakage-free streaming tests on CWRU, Paderborn, XJTU-SY, and an industrial pilot, PG-TMT delivers improved PR--AUC and timely detection at matched false-alarm levels, with strong cross-domain transfer and on-device deployability, underlining its practical impact for reliability-centered PHM schedules and ROI.

Abstract

Reliability-centered prognostics for rotating machinery requires early warning signals that remain accurate under nonstationary operating conditions, domain shifts across speed/load/sensors, and severe class imbalance, while keeping the false-alarm rate small and predictable. We propose the Physics-Guided Tiny-Mamba Transformer (PG-TMT), a compact tri-branch encoder tailored for online condition monitoring. A depthwise-separable convolutional stem captures micro-transients, a Tiny-Mamba state-space branch models near-linear long-range dynamics, and a lightweight local Transformer encodes cross-channel resonances. We derive an analytic temporal-to-spectral mapping that ties the model's attention spectrum to classical bearing fault-order bands, yielding a band-alignment score that quantifies physical plausibility and provides physics-grounded explanations. To ensure decision reliability, healthy-score exceedances are modeled with extreme-value theory (EVT), which yields an on-threshold achieving a target false-alarm intensity (events/hour); a dual-threshold hysteresis with a minimum hold time further suppresses chatter. Under a leakage-free streaming protocol with right-censoring of missed detections on CWRU, Paderborn, XJTU-SY, and an industrial pilot, PG-TMT attains higher precision-recall AUC (primary under imbalance), competitive or better ROC AUC, and shorter mean time-to-detect at matched false-alarm intensity, together with strong cross-domain transfer. By coupling physics-aligned representations with EVT-calibrated decision rules, PG-TMT delivers calibrated, interpretable, and deployment-ready early warnings for reliability-centric prognostics and health management.

Figures (9)

• Figure 1: PG--TMT overview. A Tiny--Mamba state-space branch, a compact Transformer (cross-channel couplings), and a convolutional stem are fused. The fusion output $\mathbf{r}_t$ and association $p_t$ produce a calibrated score $s_t$, which is converted to online alarms via the EVT-based decision layer with hysteresis.
• Figure 2: Physics--learning alignment. Heat shows agreement between $M_t(f)$ and the learned $A_t(f)$; bright bands near BPFI/BPFO/BSF/FTF (and sidebands) indicate physically meaningful focus.
• Figure 3: Streaming timeline. $t_{\mathrm{phys}}$: first physically detectable deviation (label or expert annotation); $t_0$: first issued decision under hysteresis (Sec. \ref{['sec:evt']}). Windows contributing to MTTD/FAR/PR--AUC are highlighted. A refractory interval $\Delta T_{\mathrm{merge}}$ merges nearby onsets; runs with no alarm by horizon end are right-censored.
• Figure 4: Deployment metrics at batch$=1$. (a) p50/p90/p99 latency on CPU and Jetson. (b) Sustainable frame rate over time. Narrow tails indicate predictable real-time behavior suitable for on-device EVT and logging.
• Figure 5: Noise robustness. (a) Latency CDF (CPU vs. Jetson) at batch$=1$. (b) PR--AUC vs. SNR. (c) MTTD vs. SNR. (d) FAR vs. SNR (threshold calibrated at a high-SNR reference). Curves are aggregated across datasets under the protocol in Section \ref{['sec:protocols']}.