Chatter detection for milling using novel p-leader multifractal features

Abstract

Chatter in machining results in poor workpiece surface quality and short tool life. A reliable chatter detection method is needed to monitor this self-excited vibration before its complete development. This paper applies multifractal features extracted from a novel p-leader multifractal formalism for chatter detection in milling processes. This novel formalism is able to discover a complex singular behavior rising on unstable chatter signals, and improves chatter detection performance. The multifractal features are selected using a multivariate filter, and assessed in terms of their dynamic monitoring abilities and classification accuracies under wide cutting conditions. The results show that the multifractal features cannot only detect chatter with high accuracy and short computation time, but are also efficient to distinguish stable tests with large amplitudes from unstable tests. A small p is recommended to highlight dynamic variations and obtain high classification accuracies. For further verification, the

A Novel Unsupervised Machine Learning-Based Method for Chatter Detection in the Milling of Thin-Walled Parts

“…It can be seen through the time-domain data that the time-domain characteristics of these three signals are obviously different due to the different acquisition methods. Through Figure 3 and Figure 4 , the acceleration, sound, and bending moment signals in both stable and chatter conditions have obvious scale-free intervals, indicating that each sensing signal of the milling process has fractal characteristics, and similar conclusions were obtained in references [ 27 , 29 ], which is the basis for applying fractal geometry to extract fractal features.…”

“…If the correlation between the extraction and the identification results is strong, even if the machine learning algorithm with low theoretical classification accuracy, such as unsupervised machine learning, can obtain better identification performance. The physical nature of the fractal feature, which describes the self-similar characteristics of signals, determines that the extracted feature is not easily influenced by the cutting process parameters, which provides a unique accuracy advantage for monitoring the nonlinear and non-stationary properties of chatter during the milling process [ 27 ]. At present, the method of extracting the fractal dimension of the signal is mostly based on box counting, such as Diykh et al [ 28 ], who extracted the feature of fractal dimension of electronephhallgraphy (EEG) signals by box counting method, and classified the extracted datasets by combining support vector machine (SVM).…”

Chatter detection in milling processes—a review on signal processing and condition classification

“…Fractal dimension (FD) is also an employed feature to measure the complexity of a pattern and the intrinsic properties of a signal, although fractal properties can also be detected by the CWT [38]. Zhuo et al [354] employed the FD in the time and frequency domains, while Chen et al [180] and Liu et al [355] evaluated multifractal-based features and Feng et al [356] utilized a dichotomy-binary strategy to reduce the time consumption required by fractal methods. Jing et al [357] designed two indicators based on the p-leader multifractal spectrum to identify the stable, weak-chatter and chatter occurrence for a micro-milling scenario, where the high spindle speed over 20,000 rpm, the reduced-sized of the cutter and the miniature dimension of the workpiece affect the process dynamics.…”

“…In the cases of using multiple features, or assessment of the same feature from multiple signals, the techniques reported for feature selection include ReliefF [180], t-SNE [153,195,236] and the recursive feature elimination (RFE) [ 145,151,200,372,373], among others. In a study by Wang et al [151], different features of the amplitude domain, frequency domain and nonlinear domain were extracted from acceleration sensors.…”

Chatter Monitoring of Machining Center Using Head Stock Structural Vibration Analyzed with a 1D Convolutional Neural Network

“…Machine learning classifiers of k-nearest neighbor (KNN), artificial neural networks (ANNs), and support vector machine (SVM) have been used for chatter detection [ 12 , 13 ]. The classifiers enable efficient chatter detection with small amounts of data and low computing power.…”