<![CDATA[Acoustic monitoring offered a low-cost, non-invasive approach to industrial predictive maintenance, but anomalous sound detection (ASD) models trained under one operating condition were shown in prior work to lose substantial accuracy when operating conditions shifted, and lightweight models suitable for edge deployment were rarely evaluated jointly for cross-domain generalization and computational efficiency. This study proposed LDIAAD (Lightweight Domain Invariant Acoustic Anomaly Detector), a MobileNetV3-Small encoder trained on a multi-resolution log-Mel representation with a self-supervised auxiliary classification objective, an explicit domain-alignment loss, knowledge distillation from a larger EfficientNet-B0 teacher, and a normal-prototype compactness regularizer. Experiments followed the official DCASE 2022 Task 2 protocol on the MIMII DG domain-generalization benchmark. A three-way pilot selected MMD as the best-performing alignment loss (overall AUC 0.608, versus 0.576 for CORAL and 0.529 for DANN). Across ten independent random seeds, LDIAAD reached a mean overall AUC of 0.6001 (± 0.0234), significantly higher than the identical backbone trained without any domain-generalization component (0.5536 ± 0.0332; paired t-test p = 0.0029, Cohen's d = 1.28, surviving Holm-Bonferroni correction). LDIAAD was also more robust than the plain backbone under all nine tested audio degradations. A much smaller custom lightweight CNN baseline, however, significantly outperformed LDIAAD on raw AUC (0.6871 ± 0.0149; p 0.001), a finding reported plainly rather than downplayed. LDIAAD is proposed as the preferable choice where consistent cross-domain behavior matters more than peak same-condition accuracy.]]>
Copyrights © 2026
