Comprehensive Evaluation of Techniques for Intelligent Chatter Detection in Micro-Milling Processes

Frictional and regenerative chatter are the most relevant unstable dynamic conditions in micro-milling. Both phenomena promote chatter, affecting the surface finishing and reducing the tool life. Therefore, in-process chatter detection strategies are of utmost importance to avoid adverse effects on productivity. These strategies can use machine learning (ML) and deep learning (DL) classifiers, but they should be able to lead with tight computational requirements. A smaller yet relevant set of features could aid these classifiers in coping with these requirements. This work proposed using feature selection to evaluate the impact of several statistical features on the performance of ML classifiers for chatter detection during micro-milling operations, compare them to the performance of the Convolutional Neural Network algorithm, and discuss the employability of the techniques on the STM32F446RE microcontroller. This study exploits experimental data from chatter and chatter-free cuts during machining operations with commercial COSAR-60 low-carbon steel under two different grain sizes (as received and ultra-fined). Random Forest, Pearson’s Correlation, and collinearity analysis are the strategies used to identify the most relevant features in these operations. The performance of several ML classifiers is compared in each feature reduction stage with the Deep Learning algorithm. The results discuss the need for applying complex algorithms since the accuracy and F1-score indices presented similar values regardless of the number of features. Moreover, reducing the dataset dimension significantly decreased training/testing time, and mainly in execution latency (FLOPs). This reduction is essential for implementing strategies for in-process chatter detection.