In industrial production,utilizing sound to detect machine fault information is a highly effective and practical means.However,in actual applications,abnormal sound occurrences are rare,and only normal sound data is available.Therefore,it is necessary to extract features that can represent the key physical characteristics of the machine.This paper proposes an unsupervised abnormal sound detection method based on multi-feature representation.Spectral coherence features are extracted to measure the correlation between different frequency-shifted versions,and log-Mel spectral features are extracted to assess the overall energy level of the signal on a Mel-frequency scale,and these features are fused to represent the critical attributes of the machine.Subsequently,training is conducted through auxiliary classification tasks,and the anomaly scores are calculated based on the cosine similarity of the embedding vectors using the K-Nearest Neighbor algorithm.Experimental results on the DCASE 2022 Challenge Task 2 dataset demonstrate a significant improvement in the performance of the proposed multi-feature representation-based detection system compared to the baseline system.
维普
万方数据