Sensorless classification of severe tool wear in CNC milling using ResNet

Tool Condition Monitoring (TCM) is crucial in CNC milling to ensure machining quality and prevent costly downtime caused by tool failure. While traditional TCM approaches often rely on expensive external sensors, such as dynamometers for measuring cutting forces, this thesis investigates the feasibility of using internal machine signals, specifically spindle torque, as a cost-effective, sensorless alternative for detecting the severe tool wear phase. The experiments involve two datasets comprising a total of 13 tools across four different CNC milling machines. A novel method for defining binary ground-truth labels is introduced that mathematically determines the onset of the severe tool wear phase by fitting a cubic polynomial to flank wear measurements, thereby avoiding the limitations of fixed-thresholding. Two deep learning architectures, Long Short-Term Memory (LSTM) networks and ResNet-50, are evaluated. The LSTM processes raw time-series sequences, while the ResNet-50 is trained on Short-Time Fourier Transform (STFT) spectrograms. The performance of spindle torque-based models is compared against that of those trained on high-frequency cutting force data. The results demonstrate that spindle torque is a viable and effective signal for classifying the severe tool wear phase, achieving classification accuracy comparable to force-based models when using the ResNet-50 architecture. This finding suggests that modern CNC machines can leverage existing internal data for effective tool monitoring without requiring additional hardware, thereby offering a scalable and accessible solution for the industry.