Context

Bearings fail gradually and failure labels are rare. Vibration windows are long (4,096 points), which makes attention-based models expensive, and many published pipelines leak future information when they pick an alarm threshold.

My internship task was a detector that needs no fault labels, runs cheaply enough for edge hardware, and calibrates its threshold without leakage.

What I did

• Built PhyDecoMamba, a series-decomposed hybrid Mamba-CNN forecaster trained only on healthy data.
• Added a physics-informed head that fuses eight time-domain indicators (RMS, kurtosis and others) into the forecast.
• Calibrated alarms with Peak-Over-Threshold on early healthy segments only, so no future labels reach the threshold.
• Benchmarked against LSTM, Simple-Mamba and PatchTST at matched parameter budgets (~338k) and ran a four-variant ablation; wrote it up as a paper draft.

How it works

• Split each 4,096-point window into a slow trend and a fast seasonal part.
• Forecast the trend with a linear layer.
• Patch the seasonal part, suppress noise with a 1D CNN, then model it with a Mamba selective-scan backbone in linear time.
• Fuse the eight physical statistics, mix both branches with a learned sigmoid gate and forecast the next 512 points.
• Score each window by forecast error; raise an alarm when it crosses the POT threshold.

Results

• Lowest test forecasting error of the four models (MSE 4.2488) and F1 75.65% under POT, second to PatchTST (76.49%).
• At batch 1,024, 1.01 ms per sample: 14.1× faster than LSTM, 5.6× than Simple-Mamba, 2.0× than PatchTST.
• At batch 64, 215.9 MB of VRAM, 83.8% less than PatchTST.
• In the ablation, the physical-statistics head cut validation MSE by 5.56%, and POT reached a 0.18% false-alarm rate with 12.12 ms of calibration.

Lessons

Detection accuracy was close across models; the clear win was cost — memory and latency — which is what decides whether a model can run next to the machine. Next I would validate on more datasets and model interactions between sensor axes.