Abstract

To address the challenges of scarce anomaly samples, inadequate modeling of temporal dynamic features, and limited feature selection capability of lightweight models in industrial anomalous sound detection, this paper proposes a method under an unsupervised framework. In the time-domain feature extraction branch, a Lightweight Temporal Pyramid Module (LTPM) is introduced to enhance the multi-scale temporal modeling capability of TgramNet, capturing both short-term and long-term temporal dependencies. In the classification network, the Efficient Channel Attention (ECA) mechanism is embedded into the bottleneck structure of MobileFaceNet to enable adaptive channel recalibration. Furthermore, three waveform-level data augmentation strategies—noise perturbation, time shifting, and amplitude scaling—are adopted. Experimental results on the DCASE 2020 Task 2 dataset demonstrate that the proposed method achieves competitive performance compared with existing approaches, attaining optimal or highly competitive results across multiple machine types. The minimum Area Under the Curve (mAUC) across different machine IDs, along with ROC curve analysis, verifies the stability and generalization capability of the model. This method offers a promising lightweight approach for industrial anomalous sound detection and condition monitoring applications.

IPC Classification

Keywords