MaxEnt feature-based reliability model method for real-time detection of early chatter in high-speed milling

Abstract

Real-time detection of early chatter is a vital strategy to improve machining quality and material removal rate in the high-speed milling processes. This paper proposes a maximum entropy (MaxEnt) feature-based reliability model method for real-time detection of early chatter based on multiple sampling per revolution (MSPR) technique and second-order reliability method (SORM). To enhance the detection reliability, the MSPR is used to acquire multiple sets of once-per-revolution sampled data (i.e., MSPR data) and to overcome the shortcoming of the once-per-revolution sampling. The proposed MaxEnt feature-based reliability model method solves the issue of the real-time detection of early chatter while ensuring its reliability. The failure hazard function (FHF) is estimated as a chatter indicator by using the SORM with the MaxEnt feature. The proposed method consists of five steps. First, set the prior parameters. Then collect data by using the MSPR technique. Next, calculate a set of the standard deviation of the data collected as a chatter feature and estimate the chatter indicator FHF by applying the SORM with the MaxEnt feature. Finally, implement the real-time detection of early chatter based on the estimated chatter indicator FHF and the threshold FHF0. The proposed method is applied to the high-speed milling process. Two examples prove that the proposed method can detect two kinds of early chatter: the early-stage of a severe chatter and the slightly intolerable chatter.

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

“…Simulated signals have been widely used as a convenient input to assess signal processing algorithms for chatter detection or to validate some proposed chatter features, for instance, as shown in [120,121,130,141,164,194,234,235,[261][262][263]. Using model-based simulation for AI training and experimental signals for testing has also been reported, such as by Ozgur and Sener [89] and Vashisht and Peng [199].…”

“…Schmitz et al [322] evaluated a chatter detection technique based on the statistical variance in the once per revolution sampling (OPRS) audio signal during milling, which uses the synchronous and asynchronous nature of stable and unstable cuts, respectively, to identify chatter. Zhao et al [141,323] used multiple sampling per revolution (MSPR) to improve the reliability in representing the stability characteristics and to achieve real-time detection of early chatter in high-speed milling. A study on assessing some key statistical features in the time domain from multiple sensors in monitoring titanium milling was presented by Navarro-Devia et al [324], in which signal segmentation methods were considered, as the window size affects computational efficiency and accuracy of TCM [141,324].…”

“…Zhao et al [141,323] used multiple sampling per revolution (MSPR) to improve the reliability in representing the stability characteristics and to achieve real-time detection of early chatter in high-speed milling. A study on assessing some key statistical features in the time domain from multiple sensors in monitoring titanium milling was presented by Navarro-Devia et al [324], in which signal segmentation methods were considered, as the window size affects computational efficiency and accuracy of TCM [141,324].…”

“…Liu et al [177] defined a feature where the threshold value varies with the tool geometry. Dong and Zhang [120] proposed the complexity index, while Zhao et al [141] used a maximum entropy-based feature.…”

“…Approximate entropy [88,133,169,170,183] Sample entropy and others [133,141,142,162,171,186,198,223] Power spectral entropy [90,111,121,134,138,157,175,176] Permutation entropy [134,157] Complexity and hand-crafted [76, 78, 80, 83, 84, 90, 115, 118, 120, 121, 123, 126, 128, 129, 132, 144, 146, 147, 149, 150, 163-165, 169, 174, 177, 179, 180, 184, 193, 197, 202, 204, 217, 220, 227, 239, 244, 255, 259, 326, 353] showed that the accuracy of each model varies according to the utilized feature, and SVM had the highest susceptibility compared to the other methods, while the probabilistic neural network (PNN) had a slightly lower accuracy. The performance of NNs is similar to SVM, and it has some advantages such as lower training time [145,180].…”