STWWgram-ODCBAM: Multimodal feature fusion and dynamic attention mechanism for anomalous sound detection

Abstract

Anomalous sound detection (ASD) aims to identify abnormal acoustic patterns emitted by machines or devices, enabling the timely detection of potential malfunctions. In recent years, various approaches have been proposed to extract both temporal and spectral features from audio data to improve detection performance. However, simply concatenating these features often leads to high-dimensional representations containing redundant information, which increases the risk of overfitting and hinders model performance. To address this issue, we propose a novel model based on a dynamic attention mechanism that adaptively selects and emphasizes informative temporal and spectral features while suppressing irrelevant noise. This enhances the quality of feature representation and improves the accuracy of anomaly detection. Moreover, we design a joint learning architecture that simultaneously captures multimodal features from both time and frequency domains, enabling the model to better capture the complex nature of audio signals and enrich the expressiveness of acoustic features. Experimental results demonstrate that the proposed method significantly outperforms state-of-the-art approaches on the DCASE 2020 Challenge Task 2 dataset, achieving AUC and mAUC improvements of 0.40% and 0.88%, respectively. Notably, for the challenging ToyConveyor machine type, our method achieves a remarkable 5.2% improvement in AUC, demonstrating strong robustness and generalization capability.