Efficient GAN-Based Unsupervised Anomaly Sound Detection for Refrigeration Units

A smart factory or Industry 4.0 is creating an epoch for manufacturing and its production lines. It reduces the total cost by monitoring and predicting the expected faults of factory lines and products. One of the essential challenges is to develop a technology to detect and predict abnormalities at an early stage without human resources. For this reason, the automation of anomaly detection is now attracting attention. Many statistical and machine-learning methods have been studied for anomaly detection. In this study, we focus on a refrigeration system for large storage, where the failures of the system will cause enormous losses. Moreover, this type of system was independently designed according to the environment, location, and storage items. Under this condition, it is difficult to train discriminative models for anomaly detection using training data that include failure data. In addition, it is indispensable to provide a basis for determining whether the system is abnormal to achieve future treatments. Therefore, deep generative models are used to achieve unsupervised abnormality detection. Because the sensing system's cost for detecting system failures should be reduced, the proposed system uses low-cost microphone arrays to monitor sounds and source locations. The system also provides a rationale by visualizing and mentioning irregular sounds. Furthermore, this study compared various deep generative models in terms of accuracy and showed that the Efficient GAN-based method achieved the highest accuracy.