Generative model-based transmission health monitoring of construction machinery in real factory

Abstract

In real factory production processes, transmission systems occasionally produce abnormal noises that deviate from their normal sound patterns. Detecting these anomalies is crucial for identifying the underlying causes and ensuring the quality of products. The traditional health monitoring of the transmission system in construction machinery relies on the expertise of skilled workers. In order to enhance detection capabilities during instances of abnormal noise occurrence, conserve human resources, and provide a technological foundation for future automation in production, we propose a transmission system sound detection method based on a generative model. In this study, normal transmission system sounds were collected from Komatsu Corporation using professional equipment, and log-Mel features were extracted from the raw audio data. The generative model was employed as a classifier and trained on normal transmission system sounds, with reconstruction error serving as the threshold to distinguish between normal and abnormal sounds. The experimental comparison with traditional deep learning methods verified the effectiveness of our approach. The results demonstrate the feasibility of applying the generative model as a classifier for health monitoring of the transmission system in Komatsu's construction machinery.