Sensor Data Analytics for Tool Condition Anomaly Detection with Machine Learning Techniques

Tool condition anomaly detection is a critical aspect of machining processes, ensuring product quality, cost effectiveness, and operational safety. This paper presents a study of tool condition anomaly detection in computer numerical control (CNC) machining using multiple sensor signals and machine learning models. The study employed four sensors, namely spindle vibration, tool vibration, tool force, and acoustic emission, to collect data for ten different signals. Eight process parameters, involving different cutting speeds, feed rates, and depths of cut, were tested one at a time until the tool wear size reached 0.15mm. The tool wear size was periodically measured using a digital microscope. The sensor data was collected during each machine run at a sampling rate of 25600Hz. The experiments were conducted in sets of eight, with each set repeated four times. Tool wear exceeding 0.1mm was considered a worn condition, while anything below was considered healthy. The data was processed by computing the Welch Power Spectral Density (PSD) for each signal and summing the PSD within predefined frequency ranges. Feature selection was performed using Recursive Feature Elimination (RFE) combined with Random Forest Classifier (RFC). The selected features were then used to train and test Support Vector Machine (SVM), k-Nearest Neighbors (kNN), Multilayer Perceptron (MLP), Naive Bayes (NB), and Decision Tree (DT) models. The prediction performance was evaluated using classification reports and confusion matrices. The study explored the impact of changing the bin size of the sum PSD and repeating the feature selection and model training and testing. The results demonstrated that with force x-axis, a bin size of 50Hz for summing PSD yielded the best performance, with the MLP model achieving high precision, recall, and F1-score values.