Conceptual
Login

Hierarchical Feature Disentanglement for Anomalous Sound Detection

Gradient Reversal-based Hierarchical feature Disentanglement (GRHD) for anomalous sound detection under domain shift: a gradient reversal classifier strips domain-unrelated information from learned features, while a hierarchical metadata structure (section IDs organized over machine and recording attribute groups) guides fine-grained, domain-specific feature learning. Evaluated on the DCASE 2022 Challenge Task 2 machine-sound dataset with AUC/pAUC gains over parallel section-ID/attribute training (ICASSP 2025; arXiv:2501.01604).