Abnormality detection of sliding surface and exploration suitable sensor data for condition monitoring by calculating contribution using machine learning

As machines become more sophisticated and manufacturing processes more complex in the manufacturing industry, there is a need to build simple systems that enable condition monitoring without specialized knowledge. Therefore, AI-based condition monitoring is attracting attention. To improve the accuracy and interpretation of sliding surface condition monitoring, this study aims to elucidate the relationship between wear mechanisms and data obtained from each sensor such as friction coefficient, acoustic emission, vibration acceleration, and contact electrical resistance. In this paper, we investigated a method to reduce the opacity of anomaly detection using machine learning and explored sensor data useful for condition monitoring of the sliding surface. The usefulness of SHAP (SHapley Additive exPlanations), one of the explanatory methods of machine learning, was verified by using it on the analysis results of anomaly detection. As a result, it was found that SHAP is useful as a method to give interpretation to the analysis results of machine learning for monitoring the condition of sliding surfaces, because the condition of sliding surfaces over time, which cannot be understood by unsupervised learning, can be understood in more detail than before based on the contribution of each sensor data. In addition, a comparison of the contribution of each sensor data to the wear behavior for each sliding distance and the predictions by SHAP revealed the possibility that the data from the vibration acceleration sensor mounted perpendicularly to the sliding direction might effectively detect signs of seizure.