Evaluation of multi-location accelerometer sensor configurations for vibration-based classification of CNC milling conditions

Abstract In CNC milling processes, vibration signals are a crucial data source that directly reflects machining dynamics and changes in cutting conditions. This study aims to classify material type, feed rate, and depth-of-cut using data-driven vibration signals obtained during CNC milling of Al6061 and Al7075 aluminum alloys. Vibration data were collected from the bed and tool holder areas; a sensor-fusion approach, in which these two data sources are evaluated together, was also investigated. The dataset comprised 16 independent runs and 288 derived five-channel windows; evaluation used experiment-ID-based 4-fold grouped cross-validation. Classical machine learning models, Random Forest, Extra Trees, and Support Vector Machine, were compared, while 1D-CNN, GRU, and CNN-GRU were compared as deep learning models. The results showed that vibration signals carry strong discriminative information, especially for the classification of cutting parameters. The five-channel fused SVM achieved 0.9062 accuracy for depth-of-cut classification, while a two-channel cross-location ablation further improved accuracy to 0.9306, supporting the complementary value of spatially separated sensor information. In feed rate classification, the best performance was again obtained with the SVM model under sensor fusion, achieving an accuracy of 0.8403 and an F1 score of 0.8401. In contrast, material classification remained a more challenging task; the highest accuracy of 0.6181 was obtained with the Random Forest model using tool holder data. The findings show that sensor position and sensor fusion are decisive factors in model success. Furthermore, it was observed that classical machine learning models produced more stable results compared to deep learning models in the current data structure.