Research On The Theory And Technology Of Tool Wear Monitoring Based On Multi-kernel Function

With the continuous development of intelligent manufacturing technology,machine learning plays an increasingly crucial role in promoting intelligence and efficiency in the industrial field.In machine learning techniques,the feature representation of data and the selection of models are crucial,especially in monitoring cutting machining quality.The data in this domain often exhibits highly complex nonlinear structures,posing challenges to traditional linear classifiers and regression models,which are difficult to meet the needs of dealing with this complex problem.The kernel function is mainly used to solve the high-dimensional mapping of nonlinear data to realize the linear differentiability of complex nonlinear problems,which plays a vital role in solving the high-dimensional eigenspace operations of complex nonlinear functions and provides a research direction and an application method to monitor the quality of cutting and machining.This paper proposes a series of highly accurate monitoring models and optimization methods to address the problem of cutting quality monitoring in the cutting process.The basic theory of multi-kernel function construction is explored,including the definition of kernel function,the principle of high-dimensional mapping,Mercer theory,common types of kernel functions,and Gram matrix et al.Based on the analysis of critical factors in constructing multikernel functions,two multi-kernel function construction methods,namely the weighted multicore function algorithm and unified weighted multicore function algorithm,are proposed,optimized,and comparatively analyzed;Data collection of turning and milling machining experiments are carried out to develop a series of tool wear monitoring systems with high accuracy and stability,and the main research contributions are as follows:(1)A tool wear monitoring model is proposed based on the weighted multi-kernel function algorithm,combining the probabilistic kernel principal component analysis method based on the integral radial basis function and the weighted multi-kernel relevance vector machine.The tool wear monitoring model based on the relevance vector machine is established using the sparrow search algorithm to optimize kernel weights and hyperparameters.Experimental results show that compared to standard single kernels(Gaussian,polynomial,quadratic,exponential,and Laplacian kernels),the tool wear monitoring model based on the weighted multi-kernel function exhibits higher convergence accuracy and prediction precision.A probabilistic kernel principal component analysis method based on the integral radial basis kernel function is established to enrich the noise information of the cutting signal features,improve the 95%confidence interval validity of the weighted multi-kernel relevance vector machine’s prediction results,and enhance the reliability of the weighted multi-kernel relevance vector machine model for the simulation of noise information.(2)Based on the unified weighted multi-kernel function algorithm,a tool wear monitoring model combining multilayer stacked denoising autoencoder and unified weighted multi-kernel Gaussian process regression is proposed,and an adaptive moment estimation algorithm is introduced to optimize the multicore hyper-parameters of the unified weighted multicore function.The effect of the denoising autoencoder activation function and the number of layers on the performance of the tool wear monitoring model is analyzed.Experiment results show that,compared to standard single kernels(e.g.,Gaussian kernel,Matern52 kernel,periodic kernel,and power exponential kernel),the tool wear prediction results of the unified weighted multi-kernel Gaussian process regression model show significant advantages under four evaluation indicators,with the minimum root mean square error,mean absolute error,mean square error,and maximum Pearson’s correlation coefficient,with the values of 9.8629×103,7.5217×10-3,9.7300×10-5 and 0.9960.Furthermore,after optimizing with the multilayer stacked denoising autoencoder,the unified weighted multi-kernel Gaussian process regression model’s tool wear prediction values show maximum increases of 28.31%in mean absolute error,25.75%in root mean square error,44.90%in mean square error,and 0.180%in Pearson correlation coefficient.(3)Based on the tool wear monitoring model built by two multi-kernel functions,a multitarget monitoring model based on unified weighted multi-kernel Gaussian process autoregression is built for real-time milling tool wear and machined surface roughness monitoring.The milling experiment results demonstrate that,compared with the traditional multi-target prediction models(e.g.,random forest,gradient boosting,support vector regression,and k-nearest neighbors algorithm),the prediction results of the multi-target monitoring model are more accurate,and its comprehensive indicators,including root mean square error,mean absolute error,root mean square error,mean absolute percentage error,pearson correlation coefficient,and spearman correlation coefficient,are 1.5840×10-4,0.8767×10-2,1.2442×10-2,0.9883 and 0.9818,respectively.Aiming at the poor consistency in the confidence intervals of the milling multi-target monitoring model’s prediction results,a two-stage feature fusion technology(principal component analysis+orthogonal neighborhood preserving projections)is proposed to fuse the milling signal features,which effectively improves the accuracy and confidence interval of the prediction results.Under the Jaccard matrix fusion feature,the prediction results’ comprehensive mean square error,root mean square error,mean absolute error,Pearson correlation coefficient,and Spearman correlation coefficient increased by 15.75%、8.21%,4.82%,0.07%and 0.24%respectively,and their confidence intervals’CI＿width＿D and CI＿var＿D are decreased 34% and 41.75% respectively.