Implicitly labeled Forecasting based Tool Condition Monitoring in Machining Processes

Abstract

Self-optimizing machining systems are cyber-physical systems that enable autonomous monitoring of CNC machining processes. The realization of such a system is achieved through a variety of efforts in the areas of CNC control digitization, sensor integration and machine learning. In terms of sensor signal usage in machining, approaches that aim to use existing data sources on the machine are an efficient way to implement this in practice. However, multiple machine learning approaches in process and tool condition monitoring rely heavily on classical supervised approaches, which demand machining-technology-specific feature engineering tasks such as process-dependent statistical values and labeling efforts. Both tasks require expert knowledge as well as a solid data infrastructure for storing labeled data. This paper aims to present an implicitly labeled time series forecasting approach to eliminate the need for labeling and feature engineering, which poses positive effects for real world applications based on CNC-internal control data. By using histogram-based outlier scores on the prediction residual after comparing the prediction with the real signal pattern, the approach offers further possibilities for building a fully functional, self-optimizing system. Different drilling operations are considered in which the cutting speed, feed rate, tool diameter and material are varied. In addition, the transfer of the models to another machine with other parameter variations is discussed and whether the net can be used to detect both anomalies and wear.