A feature extraction method for intelligent chatter detection in the milling process

In machining, especially in milling, chatter refers to self-induced vibrations that arise and grow due to the dynamic interaction of the cutting process. It can result in poor surface finish, inaccurate dimensions, and increased tool wear. Although the development of strategies for extracting chatter features in milling processes has improved over time, the milling environment’s complexity continues to pose challenges. Variational mode decomposition (VMD) is a commonly employed method for extracting features from chatter signals and has gained widespread use. However, the optimal determination of critical VMD parameters, such as the parameter of penalty (α) and the modes (K), is challenging because they significantly affect the decomposition result. To address this limitation, the basic VMD algorithm is embedded into the Bayesian Optimization Algorithm (BO), called the VMD-BO algorithm, to automatically select the best parameter combination of (α) and (K). To enhance the capabilities of extracting chatter features, the VMD-BO algorithm is combined with slope entropy (SlopEn), and the decision tree (DT) algorithm is used to intelligently detect the severity level of chatter. Therefore, this study introduces a novel hybrid method for extracting chatter features, combining VMD-BO, SlopEn, and DT algorithms to enhance intelligent chatter detection. Optimization results of the simulation chatter indicate that the VMD-BO algorithm achieves the highest kurtosis value, demonstrating superior decomposition performance compared to existing optimization methods. Additionally, results of the experiment with the measured vibration signals reveal that SlopEn outperforms three other entropy measures—fuzzy entropy (FuzzEn), sample entropy (SampEn), and permutation entropy (PermEn)—in terms of detection accuracy. However, its detection rate can be improved with an increase in the number of features. Finally, the proposed method can outperform other methods with two features, where the achievement of classification accuracy on the validation set is 96.67%. In addition, a milling test is done on a stepped work-piece with varying machining conditions to confirm the efficacy of chatter detection for online monitoring. In model application, the built model accurately identifies all machining states, including those in transition.