Research On Tool Wear Monitoring Technology For Robot Milling

Since the release of the national development strategic plan of "Made in China 2025",China has entered a period of information and intelligent development in the manufacturing industry.As the foundation and core of the manufacturing industry,production and manufacturing have received more attention.Machine tools are an extremely important part of production and manufacturing,shouldering the important mission of high-precision,highefficiency,and high-quality processing,and are mature representatives of processing equipment.However,traditional machine tool processing has shortcomings such as high cost and insufficient flexibility,especially due to the limitations of processing space,which leads to the inability to process large and thin-walled aviation parts.Therefore,some studies have proposed alternative solutions for industrial robots to process large parts.Industrial robots have many advantages,such as low cost,flexible angles,and greater flexibility.However,their process systems exhibit multiple flexible attributes such as open chain series multi rigid elastomer.Under the impact of high-frequency cutting loads,the machining process vibrates strongly and has poor stability,resulting in severe wear and tear of cutting tools,affecting the surface accuracy and processing quality of machined parts,and restricting the application and promotion of industrial robot machining.This study takes milling tools equipped with industrial robots as the research object,proposes a deep learning state monitoring model based on cutting force signals and vibration signals,and verifies through experiments that the proposed model method has high detection accuracy.The main contents of this study are as follows:(1)A robot milling experimental platform has been built.Using a 6-axis industrial robot to grip a high-performance motorized spindle,the Al7050-T7451 workpiece was milled in a stable attitude.A tool wear experiment and signal acquisition experiment were designed to collect the full life milling signals(cutting force signals and vibration signals)of two cutters under nine different cutting parameters.A modulus maximum noise reduction method and a data normalization method were introduced to reduce the noise of the original data.(2)The time domain information of tool life milling signals in robotic machining and traditional machine tool machining is analyzed,and the data differences between cutting force signals and vibration signals in the same milling state are compared and analyzed.The results show that both the cutting force signal and the vibration signal have a significant increasing trend as the degree of tool wear increases,and the sensitivity of the cutting force signal to tool wear is much greater than that of the vibration signal.(3)A tool breakage monitoring model based on deep learning was constructed.Firstly,local features(time domain,frequency domain,and time frequency domain)of input data are extracted through a multi domain feature extraction layer.A new deep learning model with a two branch structure,parallel Bi LSTM(vibration branch and cutting force branch),is introduced to fuse local features in multiple domains and learn the time-dependent change patterns of signals.The attention mechanism is used to improve the self decision-making ability of neural networks,Finally,the regression layer analyzes the mapping relationship between local time characteristics and wear amount to achieve damage status recognition.(4)The recognition accuracy of traditional machine tool data sets and robotic milling data sets based on the same network hyperparameters is summarized.The results show that the average RMSE and MAE for traditional machine tool recognition under a single layer network are 9.0 and 11.4,respectively,and the average RMSE and MAE for robotic milling data are17.98 and 14.76,respectively.However,the average RMSE and MAE for traditional machine tool data under a parallel network are 6.6 and 4.8,respectively,and the average RMSE and MAE for robotic milling data are 10.61 and 9.104,respectively.Obviously,parallel networks have better feature extraction capabilities.