Integrating Mixup and ArcFace for enhanced anomalous sound detection

In machine condition monitoring, accurate detection of anomalous sounds is essential for timely maintenance interventions. Traditional deep neural network models employ classification-based self-supervised learning to discriminate normal and abnormal operational sounds using classification confidence. However, these models often suffer from overfitting or excessive confidence, limiting their effectiveness. To overcome this, we introduce a hybrid technique that combines the strengths of Mixup and ArcFace techniques. Mixup enhances model generalization by training on mixed sound data, while ArcFace promotes greater inter-class distances with an angular margin loss. However, there is a challenge in combining these methods because the angular margin for ArcFace is defined for a single target class, and hence, can be hardly defined for the mixed data utilized in Mixup. The proposed technique employs the class-dependent angular margins based on the mixing ratio of data. SoftMax losses for individual classes are constructed using the class-dependent margins and then integrated into the Mixup loss. When applied to the DCASE2020 Task2 dataset, the proposed method significantly outperforms the state-of-the-art NoisyArcMix model in anomaly detection. This approach demonstrates the potential of leveraging class-dependent angular margins in mixed data for improved performance in machine condition monitoring.