Multi-Spectral and Multi-Temporal Features Fusion With SE Network for Anomalous Sound Detection

Unsupervised anomalous sound detection (ASD) identifies anomalies by learning or estimating normal operational patterns and detecting deviations. This capability is crucial for tasks, such as equipment maintenance and quality control. In this study, we propose a novel unsupervised ASD method that integrates multi-spectral and multi-temporal features to enhance the detection performance. Specifically, in the frequency domain, along with Log-Mel features, we introduce a concise convolutional neural network (SpecNet) to capture spectral information. In the time domain, in addition to TgramNet, we employ WaveNet to model temporal dependencies and extract causal relationships from time-series signals. These newly generated features are combined with the original features and processed using the MobileFaceNet (MFN) backbone. To further improve the extraction of latent features, a squeeze-and-excitation (SE) module is integrated into the MFN backbone. Additionally, an auxiliary loss function based on machine type is introduced to preserve commonalities among similar machines, thereby improving the generalization. We validated our approach on the DCASE2020 Task 2 dataset, achieving improvements of 1.32% in AUC and 2.39% in partial AUC (pAUC) compared to state-of-the-art methods. These results demonstrate that our method offers a robust solution for enhancing the anomaly detection performance in industrial applications.