Milling Cutter Wear State Prediction Based On Integration Fusion Method Study

In the process of high-precision milling,considering that the milling cutter rotation speed is very fast and the joint with the surface of the processed workpiece will not continue,it is easy to cause milling cutter jitter and wear failure.Therefore,it is of great significance to realize the intelligent detection of the wear state of the milling cutter and control the dynamic change of milling cutter wear in the milling process in a timely manner,so as to provide a basis for tool change decision-making.This thesis aims to explore the key methods of milling cutter wear state prediction,take the milling force signal and vibration signal as health state monitoring signals,establish a variety of different milling cutter wear status prediction models,and combine with integrated fusion algorithms to aggregate and improve a single learner with different energy efficiency,in order to verify milling cutter grinding externally in the prediction.Excellent results were achieved in the loss data experiment.The research content of this article includes:(1)The wear mechanism of the milling cutter is studied,and the different wear forms of the milling cutter and the signal change characteristics of each wear stage are summarized.By using a variety of signal processing technologies,the characteristic matrix of milling force signal and vibration signal in the time domain,frequency domain and time frequency domain are extracted.Aiming at the problem that the dimension of the milling cutter wear state feature matrix is too large,a multi-level feature joint dimension reduction scheme is designed to combine the order results of the feature importance of multiple base learners,and select the characteristic factors of the milling cutter wear state prediction model in combination with correlation analysis.In view of the uneven sample distribution in different periods of milling cutter wear,this thesis adopts the adaptive comprehensive sampling method to supplement the sample size of the milling cutter.(2)By building a variety of single milling cutter wear state prediction learner models,including Extreme Trees(ET),Support Vector Machine(SVM),Logical Regression(LR),Adaptive Boosting(Ada Boost)and K Nearest Neighbors(KNN),it is found that the ET model can effectively predict milling cutter wear.Status,and the effectiveness of the ET model is proved by verifying evaluation indicators such as accuracy rate,recall rate,F1-Score and AUC.Aiming at the problem that ET is vulnerable to noise interference,resulting in reduced robustness and overfitting,a Blending-ET integrated fusion milling cutter wear state prediction model is proposed.Four different energy-efficient base learners are used to train different weights to predict new samples,so as to optimize the learning performance of ET.Through comparison and verification with a single learner model,the results show that the Blending-ET integrated fusion milling cutter wear state prediction model is on 6 classifiers and 4 evaluation indicators.Blending-ET is significantly better than the other five mainstream algorithms.(3)In view of the insufficient hyperparameter optimization of the two lightweight gradient hoist models of LightGBM and CatBoost,the RS-LightGBM and RS-CatBoost milling cutter wear prediction models are built in combination with random search algorithms(RS).Using the evaluation indicators of mean root error(RMSE)and average absolute error percentage(MAPE),the external test set of milling cutter wear is used as a comparative verification.The experimental results show that both the LightGBM model and the CatBoost model can predict the amount of milling cutter wear well.Considering the differences in the operating mechanisms of each model,in order to make full use of their advantages,this thesis proposes a Bagging-LightGBM and Bagging-CatBoost milling cutter wear prediction model based on the integration of Bagging.By comparing with the prediction of a single model and combined with the external verification of the c1 milling cutter,the experiment shows that the Bagging-LightGBM and BaggingCatBoost models for milling cutter wear prediction have better generalization performance on the basis of improved accuracy.