An Anomalous Sound Detection Network Based on Time-Frequency Attention and Improved One-Class Softmax

Anomalous sound detection (ASD) is crucial for the condition monitoring of industrial communication equipment. However, due to the scarcity of anomalous data, existing unsupervised or self-supervised methods often struggle with feature extraction and generalization. We propose STAgram-MDOCS, a novel self-supervised framework that integrates a time-frequency attention mechanism and an improved Multi-class Dispersion One-Class Softmax (MDOC-Softmax) loss, while remaining lightweight and suitable for deployment in industrial scenarios. The attention module enhances feature representation by focusing on key temporal and frequency regions, while MDOC-Softmax not only incorporates dispersion loss to improve robustness, but also extends DOC-Softmax from binary to multi-class classification, making it suitable for self-supervised learning tasks. Experimental results on the DCASE 2020 Challenge Task 2 dataset show that our method achieves AUC, pAUC, and mAUC scores of 96.11%, 90.59%, and 92.30%, respectively, demonstrating its strong feature extraction and generalization capabilities.