Research on the milling tool wear and life prediction by establishing an integrated predictive model

• Published in: Measurement : journal of the International Measurement Confederation 2019-10, Vol.145, p.178-189
• Main Authors: Yang, Yinfei, Guo, Yuelong, Huang, Zhiping, Chen, Ni, Li, Liang, Jiang, Yifan, He, Ning
• Format: Article
• Language: English
• Subjects: Algorithms; Eigenvectors; Integrated prediction model; Life prediction; Machine tools; Machinery condition monitoring; Milling (machining); Predictions; Regression analysis; Remaining useful life; Support vector machines; Support vector regression; Surface properties; Time domain analysis; Tool life; Tool wear; Trajectory similarity; Wavelet analysis; Wavelet transforms; Wear resistance; Workpieces
• ISSN: 0263-2241; 1873-412X
• DOI: 10.1016/j.measurement.2019.05.009

•The prediction model is established to predict tool wear and remaining life.•The model uses trajectory similarity and support vector regression algorithms.•Time domain and wavelet analysis of the cutting force signal are implemented.•Five eigenvectors are selected as the input vectors of the prediction model.•The integrated prediction model is better than other four models in general. As the tool wear increases, the surface quality of the workpiece will decrease, and even the workpiece will be scrapped. Therefore, in order to obtain a better machined workpiece quality, monitoring the tool wear is necessary. By monitoring the machining condition, the degree of the tool wear and the remaining useful life (RUL) can be obtained in time. This paper establishes an integrated prediction model based on trajectory similarity and support vector regression, which can predict the tool wear and life. The time domain and wavelet analysis are carried out. The relationship between the signal characteristic quantity and the tool wear is studied. Five eigenvectors are selected as the input vectors of the prediction model by studying the correlation between 45 characteristic quantities and the tool wear. The model training is carried out by using the PHM public data set. The relative errors of VB value prediction accuracy in the stable stage of the sample tool is above 88% and the prediction accuracy of the stable stage of Tool 1, 2, and 3 is 88.5%, 87.5%, and 90.5% respectively, by using this integrated prediction model, which is better than other four single algorithms.