Research On Helical Milling Tool Wear Condition Monitoring Based On One-dimensional Convolutional Neural Network

CFRP/Ti stacks can reduce the weight and combine the excellent properties of the two materials,has been used in aviation manufacturing industry widely.The connection of CFRP/Ti stacks is still dominated by bolt connection,in order to ensure the relative position accuracy and processing efficiency of connection hole,integrated hole is often used.Helical milling is a new type of hole-making technology with unique advantages,and has been applied in the integrated hole making of CFRP/Ti stacks preliminarily.However,the tool wear is serious during process,it is necessary to monitor its wear condition in order to ensure the hole surface quality and dimensional accuracy.Therefore,this paper takes the robotic helical milling system as platform to study the helical milling tool wear monitoring method.The main contents and innovations of this dissertation are as follows:The processing principle,advantages and applications of helical milling are introduced,the research status of tool wear condition monitoring is described.The shortcomings of traditional machine learning and two-dimensional convolutional neural network(2D CNN)applied in tool wear condition monitoring and the advantages,basic principle and application status of one-dimensional convolutional neural network(1D CNN)are compared and analysied.On the basis of introducing the composition of robotic helical milling system and analyzing the wear form of helical milling tool,the tool wear stages are divided.The advantages of the current signal as monitoring signal were clarified,current signal acquisition experiment was performed in the robotic helical milling system,and current signal dataset was established.A helical milling tool wear condition monitoring method based on 1D CNN is proposed.A lightweight 1D CNN is designed,the time-domain current signals of spindle,revolution shaft and feed shaft of robotic helical milling end-effector were input into network directly.The network can extract tool wear features and classify different wear stages automatically,and unify feature extraction and classification into the same framework for global optimization.The experiment results carried on the current signal dataset demonstrate that the proposed method can achieve 99.18% monitoring accuracy,which is obviously better than the 96.24% monitoring accuracy of the support vector machine(SVM)model.Aiming at the problem of class imbalance in current signal dataset and low recall of the severe wear stage,a helical milling tool wear monitoring method based on cost sensitive one-dimensional convolutional neural network(1D CS-CNN)is further proposed.1D CNN is improved from the algorithm perspective,by increasing the misclassification cost of the severe wear stage,and integrating the cost matrix into the network loss function to make 1D CNN cost sensitive,and 1D CS-CNN is obtained ultimately.The experiment results carried on the current signal dataset demonstrate that the proposed method can improve the recall of severe wear stage from 98.38% to 99.60%,and the monitoring accuracy reach 99.29%.Finally,the whole research contents are summarized,and the work to be further studied is prospected.