Mechanical Abnormal Sound Detection Based on Self-Supervised Feature Extraction

Abnormal state detection of mechanical equipment based on acoustic diagnosis is of great significance in the field of industrial automation. At present, unsupervised abnormal sound detection of mechanical equipment is mainly based on artificial construction algorithms to extract sound signal features, and then use these features for further anomaly detection, which is greatly influenced by the human factors and the lack of universality of the artificial extraction method. To solve these problems, a new feature extraction method based on self-supervised learning is proposed, and the feature is input into the autoencoder (AE) for abnormal sound detection of mechanical equipment. In this method, the sound sample is firstly converted into a time-frequency spectrum, and the time-frequency spectrum of the normal equipment is used as the training sample, then the self-supervised feature extractor (SSFE) is constructed by using the normal time-frequency spectrum and the artificially constructed abnormal time-frequency spectrum. AE is trained by the features of normal samples extracted by SSFE to realize abnormal sound recognition of the unsupervised mechanical equipment. Experiments are carried out with MIMII open data set, and the results show that the proposed method can adaptively extract the sound features of four kinds of mechanical equipment, including fans, pumps, sliders and valves. The average area under curve (AUC) result obtained by the proposed method is 88.5%, which is significantly improved compared with those of the artificial feature extraction methods such as linear sonogram, logarithmic Mel spectrum, and Mel- frequency cepstral coefficients.