Abnormal sound detection for gearboxes based on multiscale spectrum and temporal fusion network

Purpose The purpose of this paper is to solve the challenges arising from abnormal samples and inadequate feature extraction of sound signals in abnormal sound detection. Design/methodology/approach Addressing the issues of insufficient abnormal samples and inadequate feature extraction in gearbox abnormal sound detection, this paper proposes a gearbox abnormal sound detection method based on multiscale spectral and temporal fusion networks. The innovations lie in integrating a multiscale spectral extraction network and a temporal feature extraction network into the baseline network to extract 128-dimensional Mel-spectrogram features and capture reliable interdependencies among temporal sequence features of sound signals, respectively. The Backbone module is used to extract Mel-spectrogram features, which are then enhanced through multiscale context modeling using an Atrous Spatial Pyramid Pooling to augment the representational capacity of limited samples. In addition, the autocorrelation attention mechanism is used to quantify temporal features, thereby enhancing the intrinsic characteristics of sound signals. Finally, the Sub-Cluster AdaCos Loss (SCAdaCos Loss) jointly optimizes the multiscale feature extraction network and the temporal extraction network, yielding optimal features that enhance the model’s sensitivity to abnormal sound samples and improve detection accuracy. Findings Experimental validation was conducted on the DCASE2021 Task2 gearbox data set. Ablation experimental results demonstrate that the proposed abnormal sound detection method for gearboxes based on multiscale spectrum and temporal fusion network can effectively improve the generalization ability of the model. Performance comparison experimental results indicate that, under identical conditions, area under curve (AUC) and partial area under curve (pAUC) of the proposed method reach 80.66% and 62.19%, respectively. When these conditions vary, AUC and pAUC of the proposed method increase to 82.04% and 64.62%, respectively, which are notably higher than the corresponding detection results of the baseline systems SCADACOS and SIAMESE+AE. Originality/value The proposed method is effective for gearbox abnormal sound detection and exhibits superior detection performance. Peer review The peer review history for this article is available at: https://publons.com/publon/10.1108/ILT-11-2024-0431/