Research Of On-line Tool Wear Monitoring Method In Turning Based On Cutting Force

There inevitably exists tool wear during cutting process, and the flank wear affects the machined surface quality and dimensional accuracy, what's worse, severe flank wear can scrap the workpiece. In the past, tool wear state is mainly determined by the skilled workers, and in recent years with automated and unmanned machinng process development, it seems particularly urgent to realize on-line tool wear monitoring to get rid of dependence on skilled workers. This paper mainly studies on the methods and techniques of on-line tool wear monitoring through static and dynamic cutting force.Firstly, we conducted tool wear tests by cutting 50 steel under various cutting parameters, and focused on acquiring static and dynamic cutting force in different wear states, but also measured the corresponding surface roughness. Then we have studied the actual tool wear and its effect on the force, chips and quality of the machined surface.Secondly, in order to obtain the sensitive features with tool wear, we have analyzed cutting force of different wear states in time and frequency domain. So we have studied the mean, variance, RMS, skewness, kurtosis and other time-domain features change with the gradual tool wear based on time-domain statistical analysis, and have studied frequency gravity, frequency variance change with the gradual tool wear based on the requency domain analysis, have studied the average energy band, band energy ratio of total energy change with the gradual tool wear based on wavelet multi-resolution analysis, and also have studied the variation of the model residual variance, the model coefficients change with the gradual tool wear based on AR analysis. Then sensitive features with tool wear are preliminary selected by correlation coefficient method, a total of 27 features without cutting parameters.Afterwards, the extracted sample features are randomly divided into two groups, one for training and the other for testing. The final training sample features and four projection direction vectors are getted by the Kernel Principal Component Analysis(KPCA) from training sample features. Then a centered kernel matrix is constructed with testing sample features, and projects to the four projection direction vectors to get final testing sample features. The result shows that 27-dimensional feature vectors become into 4-dimensional feature vectors after KPCA, so it can greatly improve the follow-solving speed.Finally, kernel ?-support vector regression(?-SVR) is trained with the final training sample features using grid search method and cross-validation method, and the best model coefficients are obtained eventually. After that, the obtained model is tested by the final testing sample features, the result shows the root mean square error(RMSE) between predicted and real value is only 0.0229 mm, and the square of the correlation coefficient is 0.9628. Then we study that under declining of training numbers, the calculated model can still maintain a high recognition accuracy. At last, a simulation monitoring system is designed based on Matlab GUI, and the back-end monitoring system designing is described in detail.