Pseudolabel-Guided Disentangled Representation Learning for Cross-Domain Anomalous Sound Detection

Anomalous sound detection (ASD) is widely used for industrial condition monitoring because of its noninvasive sensing capability. However, real-world deployment under extreme conditions remains highly challenging. Distribution shifts, annotation scarcity, and limited anomaly instances significantly hinder generalization. This study presents a cross-domain ASD framework designed for unsupervised and low-resource settings. The approach integrates semantic pseudolabel discovery using a pretrained audio model and graph-based clustering, multiresolution representation learning via a dual-branch encoder, and robustness-enhancing strategies, including adversarial disentanglement and domain-aware scoring. Experimental results demonstrate that the proposed method achieves superior cross-domain robustness on both real-world Antarctic diesel generator data and several public datasets. These findings validate the practical potential of deploying ASD systems in extreme environments.