Session:

Manufacturing Modeling and Dynamic Systems
Monday March 01, 2004, 17.15 – 17.35

Session Chair:
Vice Chair:

Fatima Rateb
Stefan Trzcielinski

Paper Title:

Bayesian Support Vector Regression for Tool Condition Monitoring and Feature Selection

Author(s):

J. Dong, National University of Singapore, Singapore
Geok Soon Hong, National University of Singapore, Singapore
Yoke San Wong, National University of Singapore, Singapore

Abstract:

This paper introduces the application of Bayesian support vector regression (SVR) and automatic relevance determination (ARD) methods for the selection of relevant features derived from force signal for tool condition monitoring (TCM) during face milling processes. 7 primary features used by other researchers are considered, including the power spectral density, skewness, kurtosis, average and maximum force, root mean square of force, and the residual error based on the AR1 model. A two-step approach is applied to extract the features. In the first step, the 7 primary features are derived. And then a moving window is used to calculate the mean and variance value of each primary feature. As a result, 14 features are obtained and fed into the ARD model. Different features have been found to be sensitive to two different phenomena, micro-chipping and gradual wear. The successful features of all the experiments are combined together to make them applicable for different cases. An additional set of experimental data is used to test the generalization capability of the features. The comparison between the selected features and the rejected ones prove that the selected features are really more useful. Finally, a moving average approach is proposed to further process the regression results. And fairly good estimation result has been achieved using the selected features.