Enhancing Feature Discriminability for Anomalous Sound Detection Based on Improved One-Class Softmax

In the fields of equipment maintenance, abnormal behavior detection, industrial manufacturing and safety monitoring, Anomalous sound detection (ASD) is a very important technology. Due to limited anomalous sound data, the existing methods often rely on one-class classification and leverage selfsupervised approaches with multi-class auxiliary tasks to enhance feature learning. However, existing methods face challenges in feature discriminability. To solve this problem, this paper extends the One-Class Softmax (OC-Softmax) to multi-class scenarios and proposes a novel loss function MultiOC-Softmax(MOCSoftmax), increasing compactness within classes and enlarging margins between classes. Combined with ArcFace, MOC-Softmax is applied to train the STgram-MFN model, which integrates spectral-temporal features for robust ASD. Experiments on the DCASE2020 Task 2 dataset demonstrate significant improvements, Achieving state-of-the-art performance accompanied by a 0.73 % improvement in AUC and a 0.46 % gain in pAUC compared to baseline methods. The results highlight the effectiveness of MOC-Softmax in enhancing feature discriminability and detection accuracy.