Worst-case-guided acoustic feature selection for chatter detection in thin-walled milling

Abstract Feature selection methods for data-driven chatter detection are commonly evaluated within fixed data distributions. Their performance under changes in operating conditions and workpiece geometry remains insufficiently examined. This study addresses this limitation using a Genetic Algorithm (GA)-guided framework that selects acoustic features based on worst-case performance. A total of 160 acoustic recordings were acquired from 40 milling conditions on two thin-walled aluminium workpiece geometries, covering different spindle speeds, feed rates, and depths of cut. Thirty-seven acoustic descriptors were evaluated under three group-disjoint scenarios that prevent window-level data leakage: Stability Boundary, Unseen Speed, and Cross-Geometry Generalisation. Given the class imbalance, the Matthews Correlation Coefficient (MCC) was used as the primary metric. A unified GA fitness function was developed to maximise the worst-case MCC across the three scenarios. The selected feature triplet-Impulse Factor, Spectral Energy, and Energy Ratio-achieved MCC values of 0.875, 0.906, and 0.860, respectively. This reduced the descriptor set by 92% while maintaining performance with support vector machine (SVM) and Random Forest classifiers. The proposed approach outperformed principal component analysis (PCA), minimum redundancy–maximum relevance (mRMR), and recursive feature elimination (RFE), indicating that synergistic feature subsets can support diagnostic performance across the evaluated operating conditions and workpiece geometries.