Answers are grounded in this paper and cite their sources.
FFT amplitude thresholds raise false alarms when machine load changes, and storing a full spectral history exceeds microcontroller memory. This paper presents edge-AI firmware that runs the whole decision, from spectrum to alarm, on a microcontroller-class node without a network. Three modules process each vibration frame. An exponential weighted moving spectrum stores a normal fingerprint in one vector. Spectral energy distribution shift scores the frame by the Kullback–Leibler divergence of the normalized spectrum against that fingerprint, so a global gain change cancels out. The alarm threshold is updated only from frames that did not alarm. Tests used 400 synthetic frames of normal running, bearing fault, imbalance, and load drift, against an FFT-peak rule and a total-energy z-score. The 16-bit fingerprint used 512 bytes, 400 times smaller than a 200-frame floating-point buffer. Bearing-fault scores averaged 1,651 times the normal score, while a 30 percent load rise stayed at or below normal. On a 350-frame timeline the proposed threshold kept a 100 percent true-positive rate through the fault, against 4 percent for a conventional adaptive threshold. The mixed-set area under the ROC curve was 1.000 and the F1 score was 0.976, with no missed faults. Each element has precedents; the contribution is their combination, which keeps a sustained fault alarmed with a 512-byte fingerprint. All results are from simulation; the firmware has not yet run on a physical microcontroller.
No takes yet. Share an insight, caveat, or question.
Faisal et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: