The rapid adoption of smart factories has increased the availability of CNC machine data, while the practical use of manufacturing AI remains limited. Tool wear prediction is a critical task because improper tool replacement degrades productivity and ...
The rapid adoption of smart factories has increased the availability of CNC machine data, while the practical use of manufacturing AI remains limited. Tool wear prediction is a critical task because improper tool replacement degrades productivity and product quality. This thesis proposes a tool-wear cycle prediction model that detects tool-change (TC) events using only time-series CNC operational logs, without additional sensors.
Tool-change detection is formulated as a multivariate time-series binary classification problem (TC vs. non-tool-change, NTC). CNC logs are segmented into machining cycles and converted into fixed-length subsequences, and TC labels are defined using the Unclamp signal with consideration of the pre-change wear phase. GRU, LSTM, and Transformer models are evaluated under identical conditions. Experimental results on real production data show that LSTM achieves the highest overall accuracy, while GRU provides balanced performance across classes. The results demonstrate the feasibility of reliable and low-cost tool-change prediction using built-in CNC logs.