Energy-efficient deep learning-based vibration monitoring on parallel ultra-low-power microcontrollers

Structural health monitoring (SHM) through vibration analysis is essential for ensuring safety, reliability, and minimized maintenance costs in sensor-connected environments. Realizing this at the extreme edge, directly on sensor nodes, requires advanced data analytics, ultra-energy-efficient hardware, and high-fidelity datasets. This paper addresses these challenges by developing an Artificial Intelligence (AI)-based deep learning framework for vibration-based structural health monitoring tailored to resource-constrained devices. We present a benchmarking dataset for vibration SHM, encompassing nine abnormal conditions that capture narrowband structural anomalies. Leveraging this dataset, we design supervised and unsupervised deep-learning models, integrating a system identification (SysId) preprocessing stage to derive structurally informed features that enhance model input quality. This improves algorithmic results while boosting hardware performances. These models are deployed on a parallel ultra-low-power microcontroller with dedicated hardware acceleration. Our unsupervised anomaly detectors deliver 89.90 percent (%) accuracy, 4.1 percentage points higher than the state-of-the-art attention-based baseline, favoring recall and missed-detection reduction at the cost of a higher false-alarm rate while consuming 13 microjoules (µJ) in low-power mode and requiring 370 microseconds (µs) in high-performance mode. This marks one of the first experimental validations of unsupervised on-device vibration diagnostics. Supervised counterparts achieve 96.06 % binary and 91.5 % multiclass accuracy, rivaling Residual Network (ResNet) with 15.8 × fewer parameters, at less than 235 µJ and not more than 6.3 ms (ms) inference. Relative to existing edge implementations, our approach is one to three orders of magnitude more energy-efficient and two orders faster, unlocking AI-enabled autonomous, low-latency predictive maintenance for large-scale distributed infrastructure monitoring.