A data-driven approach for chatter detection in machining process through feature optimization

Chatter is a self-induced vibration originating at the tool-workpiece interface that negatively impacts machining performance, making in-process detection essential. Raw time-domain features eliminate complex transformations, enabling faster and more efficient chatter detection than frequency and time-frequency domain techniques. Therefore, this study proposes a systematic approach to optimize chatter detection by selecting the most relevant features from raw triaxial cutting force signals. Statistical features are first extracted and then refined by eliminating redundant features using correlation heatmap analysis. Then, the optimal feature combinations are obtained using various feature selection techniques, including the logistic regression model, recursive feature elimination, random forest, and lasso (L1 regularization). The subsequent performance analysis is carried out by training and testing different machine learning models, such as logistic regression, support vector classifier, k-nearest neighbours, decision tree, and random forest. The proposed method efficiently reduces the input features from 51 to 3 features for chatter classification, without compromising the performance of machine learning models. The highest accuracy is obtained using three key features which are kurtosis index from the feed direction, peak-to-peak and pulse index from the velocity direction of cutting force. The proposed novel data-driven approach using feature optimization enhances the chatter detection with less computational time and cost.