Anomalous Sound Detection System in Manufacturing Industry Using Unsupervised Learning

In the manufacturing industry, skilled employees have conventionally judged abnormalities in products and machines by listening to the operating sounds of the machines. However, there are problems with the human ear, including human error such as missing sounds and the need for skill in distinguishing between sounds. By automating these human tasks with anomalous sound detection technology, there is a need for stable inspections and inspections that do not rely on employees. In the manufacturing industry, the occurrence of abnormal samples is limited to a small number. Therefore, it is difficult to construct an AI that learns from normal and abnormal samples. Therefore, unsupervised learning for anomalous sound detection is being considered. In this study, an abnormal sound detection system that detects abnormalities by using Multi-Layer Feature Sparse Cording (MLF-SC), which is an abnormality detection method, is investigated. The discrimination accuracy exceeds 80%. We investigated the change in accuracy depending on the number of training data, and found that good results were obtained with a small amount of training data. In addition, the discrimination accuracy of VGGNet used in MLF-SC was investigated, and it was confirmed that VGG19BN produced the highest accuracy.