Unsupervised adversarial domain adaptation abnormal sound detection for machine condition monitoring under domain shift conditions

Relying on mechanical sound signals to carry out anomaly detection is a challenging task. Due to the stability of the production process of complex industrial mechanical systems, there are very few or no abnormalities, and the types of mechanical failures are also difficult to describe in detail. In addition, the sound characteristics of the machine itself will change with the change of production operating conditions, and traditional anomaly detection models are prone to misjudge normal sounds as abnormal. We recommend that the change of mechanical conditions in similar situations be regarded as a domain shift between the source domain and the target domain. For unsupervised anomaly detection under the premise of domain shift, we propose an unsupervised adversarial domain adaptation method (UADA-OCSVM) based on Adversarial Domain Adaptation and One-Class SVM. Through adversarial learning strategy, the source domain and target domain data are aligned in an unsupervised method. Meanwhile, a special loss is introduced for the feature extraction layer. Finally, the anomaly detection based only on normal data is regarded as the one class classification problem, and the anomaly detection task after feature extraction is performed by OCSVM.We applied the proposed method to the MIMII DUE dataset for verification, and compared it with the autoencoder-based anomaly detection method. Experiments show that the AUC of our method is better than the method based on the autoencoder in different mechanical types, especially on the Value data set, the average AUC is increased by 15.31%, indicating that the method we proposed is better than the method based on AE Significant improvement.