[2Win5] Poster session 2

[2Win5-65] Classification for anomaly sound of machine in noisy environments using deep learning models

〇Yutaro Tsuchisaka¹, Ito Toru¹, Sakai Wataru¹, Morita Katsuaki¹ (1.Mitsubishi Heavy Industries, Ltd.)

Keywords:Deep learning, Anomaly Detection, Sound Classification

Various abnormal sounds generated by machinery are caused by defects and anomaly condition in the machines and their components. Early detection of these sounds can help prevent serious failures and improve maintenance efficiency. However, these abnormal sounds vary depending on the specific machine products and their component types, and they usually occur in noisy environments. In this study, we investigated deep learning models capable of robustly detecting abnormal sounds even in noisy environments. Using a dataset focused on the operational sounds of industrial machinery (normal and abnormal data for four types of machine sounds) , we constructed deep learning models, including ACRNN (Attention-based Convolutional Recurrent Neural Network) and a Liquid Neural Network. We compared the effectiveness of these models with a baseline model (ResNet18) . The results confirmed that the ACRNN is effective for intermittent machine sounds (valves and slide rails) and that the LNN is effective for continuous machine sounds (pumps) .