Low-Power Detection and Classification for In-Sensor Predictive Maintenance Based on Vibration Monitoring

In this work, a new custom design of an anomaly detection and classification system is proposed. It composed convolutional Auto-Encoder (AE) hardware to perform which cooperates with mixed HW/SW Convolutional Neural Network (CNN) the detected anomalies. The AE features partial binarization, so that weights are binarized while activations, associated some selected layers, non-binarized. This has been necessary meet severe area energy constraints allow it be integrated on same die as MEMS sensors for serves neural accelerator. CNN shares feature extraction module AE, whereas SW classifier triggered by when fault detected, working asynchronously it. mapped Xilinx Artix-7 FPGA, featuring Output Data Rate (ODR) 365 kHz achieving power dissipation $333 \boldsymbol {\mu } $ W/MHz. Logic synthesis targeted TSMC CMOS 65 nm, 90 130 nm standard cells. Best results achieved highlight consumption notation="LaTeX">$138 W/MHz occupation 0.49 mm 2 real-time operations set. These enable integration complete accelerator in circuitry typically sits inertial silicon die. Comparisons related works suggest proposed capable state-of-the-art performances accuracy.