Edge-AI Label-Free Anomaly Detection for Home-Appliance Predictive Maintenance : Representation, Quantization, and the Limits of Cross-Appliance Transfer

Predictive maintenance for home appliances is increasingly expected to run on the appliance itself, which forces two simultaneous constraints: the detector must be label-free at commissioning, when only normal operating data exist, and it must fit on resource- constrained edge hardware. We study both constraints on the public SMART-PDM dataset (washing machines and refrigerators instrumented for current and vibration at 2048 Hz), evaluating seven classical detectors and recurrent autoencoders under two signal representations and a leakage-free protocol with validation-selected thresholds, threshold-free ranking metrics, and bootstrap confidence intervals over ten seeds. The decisive factor for detection is the representation, not the detector. On refrigerators, windowed-spectral features raise every detector from near-random to useful and lift the autoencoders from 0.30 to 0.66 ROC-AUC; on washing machines they yield the single best detector, a principal-component reconstruction at 0.741, though their benefit there is selective rather than uniform. Replacing that lightweight baseline with deep autoencoders yields no gain, and the best detectors (0.741 on washing machines, 0.718 on refrigerators) clear a trivial current-RMS control at 0.52. Cross-type transfer is fragile and flips sign with representation (0.29–0.60 ROC-AUC), so a model of “normal” for one appliance does not define it for another. For edge deployment, dynamic int8 quantization of the recurrent autoencoder shrinks the model by 3.4× with no loss in detection quality but no CPU speedup on our host, isolating quantization’s edge benefit to memory rather than latency. The result is a deployable, honestly bounded edgeAI baseline for label-free appliance monitoring.