Study On End Mill Wear State Identification Method Under Sample Imbalance And Different Process Conditions

Due to the wide range of processing,the end mill is a kind of milling cutter used more often on CNC milling machines.In actual production,measures are usually taken to change tools in advance within the effective working cycle of the end mill to avoid excessive wear of the end mill affecting the quality of the product machining and leading to an increase in the frequency of tool changes and tool costs.Therefore,accurate identification of the wear status of end mills is of great significance to improve production efficiency and reduce production costs under the premise of ensuring the quality of product production.The deep learning method represented by convolutional neural network provides an “endto-end” solution for end mill wear states recognition because of its powerful nonlinear feature extraction capability.However,there is non-uniform degradation of end mill wear during machining.The manual marking of end mill wear status is costly and affects productivity,resulting in the deep learning-based model to identify the accurate identification of end mill wear status still suffering from the following challenges: 1)the imbalanced number of samples with different wear status during end mill milling,which makes it difficult for deep learning models to identify the wear status of minority class samples;2)end mills are often used in under different process conditions,the lack of generalization performance of current deep learning models leads to their low accuracy in identifying the wear status of end mills under different process conditions.For this reason,this thesis carries out the research on the end mill wear state identification method under sample imbalance and different process conditions,and the main research works are as follows:(1)The end mill wear test was conducted,and the end mill wear state characterization method was studied.Based on the production site conditions,a milling test bench was built,milling tests were carried out,spindle vibration data were collected,and the end mill wear state was classified into four categories based on the width of the wear zone on the rear tool face to establish the end mill wear vibration data set.Then,a convolutional neural network was constructed to extract the features of end mill wear vibration samples.Based on the model,end mill wear state representation experiments were conducted under constant process conditions,sample imbalance conditions,and different process conditions,and multi-level feature visualization analysis was performed.It can be found from these experiments that the convolutional neural network can achieve a good characterization effect on the wear state of the balanced data set samples under stable process conditions.However,the model is less effective in clustering the wear features under imbalanced samples and different process conditions,and it is difficult to accurately characterize the wear state of end mills because the sample features of different wear states are seriously mixed,and the recognition ability of the convolutional neural network needs to be further improved.(2)Aiming at the problem that convolutional neural networks and classical cost-sensitive recognition models have low accuracy in recognizing the wear status of minority class samples under the condition of sample imbalance,a method is proposed for the recognition of the wear status of end mills based on a deep feature weighted convolutional neural network.The method improves the original cross-entropy loss function in the convolutional neural network by using the balance impact factor and variance impact factor and re-weighting the feature information.So as to reduce the interference of the majority class samples to the classification decision boundary and improve the aggregation of the features of the various wear status samples.The experimental results show that the proposed method can effectively improve the recognition accuracy of the end mill wear state under the condition of sample imbalance,and achieve better recognition results than other classical methods.(3)Aiming at the problem of low accuracy of convolutional neural network and the classical transfer recognition model in recognizing the wear status of end mills under different process conditions,a wear state identification method of end mills based on multi class domain adaptive convolution neural network is proposed.The method first constructs a convolutional neural network to extract the migratable features of wear vibration samples,and uses the maximum mean discrepancy to reduce the overall distribution discrepancy of samples with different process conditions.Then uses the inter-class–intra-class distance constraint to enhance the feature distribution alignment ability of the model.Finally,the output probability matrix of the target domain data is subjected to a strategy of maximizing the kernel parametrization in order to extract highly discriminative target domain sample wear features.The experimental results show that the multi class domain adaptive convolutional neural network eliminates reliance on manually labeled data and achieves higher recognition accuracy than other typical deep learning models.(4)Based on the above research results,a set of end mill wear recognition monitoring system with offline training and online testing functions is developed based on the demand for end mill wear recognition under imbalanced samples and different process conditions,which can recognize the end mill wear status online based on the collected data and provide a reference for staff to make tool change decisions.