Anomalous Sound Detection for Industrial Machines Using Acoustical Features Related to Timbral Metrics

Abstract

This paper proposes an anomalous sound detection (ASD) method that uses a combination of timbral metrics and short-term features tailored to industrial machine faults to identify whether the sound emitted from a target machine is anomalous. The timbral-feature-based ASD (TF-ASD) method involves using five timbral metrics and two developed features as auditory features and a support vector machine (SVM) for classification. We develop two types of short-term features to estimate the change in the fluctuation of sound waves and pitch in terms of harmonics to improve the time resolution of the timbral analysis. This combination of timbral metrics and our two short-term features is based on an investigation of timbral association with industrial machine malfunction from the viewpoint of 'noticeable difference in hearing' that is the human ability to discriminate differences in sounds. We evaluated the ASD performance of our method in terms of SVM classification using the MIMII (Malfunctioning Industrial Machine Investigation and Inspection) dataset. The results indicate that the proposed method has excellent classification performance with an accuracy of 0.984 on average for emitted sounds of 16 machine types and models. This demonstrates that the combination of timbral metrics and our short-term features in accordance with the 'noticeable difference in hearing' is effective for ASD.