An Edge AI Test Bench for Unsupervised Anomaly Detection

Abstract

Condition monitoring of rotating machinery requires time-series representations that are compact enough to run on edge hardware, stable enough to generalize across production runs, and informative enough to support downstream anomaly detection without labeled fault examples. This thesis addresses all three requirements on a purpose-built, low-cost CNC-inspired test bench equipped with a brushless DC spindle and three stepper feed axes, instrumented for synchronized current, vibration, and speed sensing at approximately 91 Hz. A three-tier edge platform— ESP32-P4 acquisition node, Raspberry Pi 3 gateway, and Raspberry Pi 5 inference node—acquires a dataset of 94 636 samples across ten labeled operating cycles. Four encoders spanning a wide capacity range—per-channel summary statistics, FFT amplitude bins, the self-supervised TS2Vec encoder, and three sizes of the pre-trained MOMENT transformer—are evaluated on 20 public UCR and UEA datasets as a cross-benchmark reference and on the CNC bench as the target domain, using four axes: a supervised linear probe, unsupervised clustering, six mode-aware geometry metrics, and a CPU edge-deployment benchmark. The main findings are as follows. First, encoder rankings are dataset-dependent: TS2Vec leads on cross-benchmark accuracy but is outperformed on CNC by both MOMENT-large (0.850 accuracy) and the parameter-free Summary baseline (0.844), a reversal explained by the small CNC training set and the high mode-discriminability of the raw sensor channels. Second, geometric stability and label-aware accuracy rank encoders differently: MOMENT’s embedding space is roughly an order of magnitude more stable across production runs than Summary’s, making it the better foundation for run-disjoint anomaly detection despite similar classification scores. Third, post-training INT8 quantization collapses MOMENT’s accuracy to chance when applied naively; restricting INT8 to the FFN linears preserves FP32 accuracy at all three model sizes with a 1.33–1.82× disk reduction and 1.25–1.55× latency reduction, and the small and base variants run comfortably within the per-window budget on the reference CPU.