Research And Application Of Milling Cutter State Detection Method Based On Deep Learning

Milling,a typical processing method,is widely used in industrial processing.During the processing,severe friction between the milling cutter and the workpiece can lead to wears and breakage of the milling cutter.Once the wear and the breakage degree exceed the critical value,it decreases the machining accuracy of the workpiece,and in severe cases,it leads to the breakage of the milling cutter and the shutdown of the machine tool,which costs huge economic losses.Therefore,milling cutter state detection has significant research values.Currently,the commonly used methods include signal-based indirect detection and image-based direct detection,which are poor in expression and difficult to handle the chips in the images.This paper combines the deep learning and machine vision methods to conduct research on milling cutter state detection.The major contributions and innovative works are as below:(1)In response to lacking a dataset that makes deep learning methods difficult to effectively apply to tool state detection,this paper selects several commonly used milling cutters in actual milling processing and constructs a dataset comprised of 78 shank milling cutter end face images.In order to collect milling cutter images accurately,this paper establishes an image acquisition device,which to the greatest extent simulates the image acquisition situation in actual processing.To ensure that the dataset can be effectively applied to image semantic segmentation,the paper annotates the cutting-edge area of each image in the dataset.The annotation results are divided into two categories: cutting-edge and background.The binary classification can reduce time on annotation and model training,and increase segmentation accuracy.The final results of the dataset is available on Github(https://github.com/bitouwu/end-mills-image.git)。(2)In response to the problem that the end mill end face image dataset is rich in sample types and poor in its sizes,this paper proposes a W-ACGAN-GP method,which combines gradient and classification penalty,to generate end mill end face image data.This method combines Wasserstein distance and gradient penalty term to solve gradient vanishing problems appeared during the model training.Meanwhile,classification penalty enables the network to generate different types of milling cutter images so that the stability and accuracy of the results is secured.It is a method that can enrich the sample types and expand the dataset.In the experimental analysis,multiple evaluation indicators are used to analyze the richness and accuracy of the results of the two-edge,three-edge,four-edge,and overall end milling cutter image datasets.The visual analysis is also conducted.(3)In response to the difficulty in accurately extracting the Region of Interest(ROI)of milling cutter images under chip interference,the paper adopts Deeplabv3+semantic segmentation method based on different backbone networks(Mobile Net V2,Xception,Res Net)to extract the ROI region of milling cutter images.This method accurately extracts the cutting-edge part from the end face image of the end mill under chip interference by labeling the milling cutter cutting-edge part in the dataset and inputting it into the Deeplabv3+model for training.The results show that the mean intersection over union of the ROI extraction results and the real annotation results reaches 96.23%,and the F-score reaches 97.9%.(4)In response to the current research only focusing on the tool wear state detection and faling to fully measure the tool state,the paper proposes a Wear-Breakage State Detection(WBSD)method for end mills.This method detects the wear state of the end mill while incorporates breakage state detections.The tool state is sufficiently measured.This method firstly uses the Welzl’s algorithm and Hough line detection to fit the minimum circumscribed circle and detect the longest line in the milling cutter ROI image to achieve cutting-edge position;after image segmentation,uses the Intuitionistic Fuzzy C-means(IFCM)clustering algorithm based on intuitionistic fuzzy entropy to cluster each cutting-edge and obtains its wear area;thirdly,obtains the damaged areas by proposing a method determining cutting-edge boundaries based on Edge Point Rebuild(EPR);finally,effective evaluation is achieved through the wear and damage state of each cutting-edge.Compared with the expert annotation results,by using this method,the mean pixel accuracy of the wear detection reached 93.51%,and the mean intersection over union reached 91.30%;the mean pixel accuracy of the damage detection reaches 94.63%,and the mean intersection over union reaches 90.06%.The paper’s method effectiveness in detecting the wear and damage state of end mills is verified.