Multi-view time–frequency dynamic modeling and asymmetric convolutional networks for industrial anomalous sound detection

Abstract In recent years, industrial anomalous sound detection (ASD) methods based on self-supervised learning have achieved remarkable progress; however, existing approaches still provide insufficient characterization of the time–frequency dynamics and anisotropic structures of acoustic measurement signals. To address this limitation, this paper proposes an ASD method for non-contact machine-condition monitoring that integrates multi-view time–frequency representations with asymmetric convolutional structures. Specifically, log-Mel spectrograms, raw waveforms, and spectral-centroid-based statistical features are jointly exploited to construct multi-view acoustic representations, thereby simultaneously capturing local time–frequency textures, fine-grained temporal structures, and global frequency evolution. On this basis, a global energy trend encoding module and a temporal adaptive attention module are designed to explicitly model the energy variation sequences in time–frequency representations and their long-range temporal dependencies. In parallel, a multi-scale temporal context modeling module is introduced to capture both short-term transient components and long-term periodic patterns. To account for the anisotropy of acoustic features along the time and frequency axes, multi-branch fusion and residual structures based on asymmetric convolutions are further employed to enhance the modeling capability of high-level time–frequency dynamic features. Experimental results on the DCASE 2020 Task 2 dataset show that the proposed method achieves average area under the receiver operating characteristic curve and partial area under the ROC curve scores of 96.06% and 90.41%, respectively. These results consistently outperform multiple baseline methods, demonstrating the effectiveness of the proposed approach under the Task 2 evaluation protocol and suggesting its potential applicability to industrial acoustic measurements and ASD for non-contact machine-condition monitoring.