Random Forest on an Embedded Device for Real-time Machine State Classification

Abstract

Heavy machine tools are used in numerous industrial manufacturing processes. Avoiding unplanned maintenance time is crucial and can be achieved by continuous condition monitoring. A more targeted condition monitoring is possible if the operating state of a machine tool is known (e.g. standstill, neutral or cut). However, many manufacturers of control units do not release any information about the machine's operating state. For this reason, we investigated a system-independent approach to determine the operating state in real-time in less than 1 ms. This low runtime is necessary for machine state classification. Since most machine tools vary in individual components, the proposed state detection uses learning algorithms based on the machine's vibration characteristics. In this work we propose the Random Forest algorithm because of its reliable classification performance while keeping explainability of each prediction. Facing the real-time requirements, the Random Forest model has been adapted to an embedded device with very limited resources. Thus, the model prediction has to work with low resource consumption in a very low runtime and without loss in accuracy. We show that this approach is suitable for determining machine operating states in real-time in less than 700 μs with an average accuracy of 96 %.