Unsupervised Segmentation of CNC Milling Sensor Data into Comparable Cutting Conditions

This study aims to identify subsegments of identical cutting conditions in milling manufacturing processes using unsupervised clustering methods. Such a segmentation has a wide range of potential applications, from data compression and data obfuscation to domain specific process optimization tasks like tool wear detection or energy consumption forecasting. While prior research focuses on supervised approaches, in which the program code that instructs the milling machine is enriched by a human expert with special commands that mark the beginning and ending of segments of interest, our focus lies on unsupervised techniques, eliminating the need for such labeling efforts. Our method relies on a time-discretized simulation of the milling process. The main assumption of our method is that the so-called removed volume, i.e. the volume that is removed by the milling tool between two successive time steps, gives valuable and sufficient insight into the cutting condition. We demonstrate that, by analyzing characteristic properties of the removed volume geometries, comparable cutting conditions can be identified, which can be used e.g. to segment collected sensor data into homogeneous pieces. The results showcase the effectiveness of unsupervised methods in the analysis of manufacturing processes.