The application of multiple model fusion based on correctable weight in tool wear pattern recognizing

This paper treats an important problem in how to increase the accuracy of the tool wear condition evaluation. The problem of tool wear condition monitoring is always hard to be solved perfectly, especially the online monitoring. In order to arrive at the destination, the authors make use of artificial neural networks (BP) and support vector machines (RBF-SVM) which are two ways to build predictive models independently and then fuse each model predictions by D-S (Dempster/Shafer) in decision-making level. And a new parameter which is called model correcting weight ρ is introduced in to adjust the final contribution rate of each recognizing model. The weight is got by a black box process. In fact two series independent experiment data for each one tool wear pattern are prepared: Cdata1 and Cdata2. Use the Cdata1 (the first repeat experiment data) to train a model and predict the tool wear pattern. Some recognizing accuracy is got in this process. then the recognizing accuracy is regarded as the weight ρ which is used to adjust the tool wear condition monitoring models which are trained by Cdata2 (the second repeat experiment data that correspond to Cdata1 but have deferent use). And the final experiment results show that the monitoring system clarifies the four states of the tool.