Multi-information Fusion Based On Gaussian Process Regression And Its Application In Tool Condition Monitoring

In the high-speed cutting process,the machining process is complex and changeable.As the direct execution part of metal cutting,the accurate monitoring of the tool state plays an increasingly important role in ensuring product processing quality,improving economic efficiency,and ensuring system safety.To establish an accurate and reliable tool condition monitoring system can create huge value for various related industries and advanced areas.For this purpose,this dissertation combines the advantages of Gaussian process in machining process modeling and tool wear prediction performance.According to the improved feature selection method,intelligent online tool monitoring models and implementation method based on Gaussian process regression are proposed.(1)In order to overcome the shortcomings of conventional filtering feature selection methods in terms of feature quantity,filtering threshold,information redundancy,etc.,a two-step feature selection algorithm based on improved Fisher Score combined with genetic algorithm is proposed.Based on the multi-source heterogeneous sensor signals collected in high-speed machining experiments,a lot of representative features are extracted through signal analysis methods.The improved discriminative score is used to initialize the population in the form of probability and the fitness function can be set according to the actual problems.Through the evolution of the population,a subset of features with better comprehensive performance is automatically obtained.(2)To solve the problem that the current common offline data modeling methods are based on offline data and are interpolation prediction by randomly dividing the data set in proportion,and these pure data-driven methods require lots of labeled data to train,then a Gaussian process regression model method based on the physical process of tool wear is proposed.(3)To further overcome the shortcomings of commonly used pure data methods such as interpolation prediction and no consideration of the tool wear process,a recursive Gaussian process regression model method with feedback structure is proposed,which can achieve more accurate and smooth prediction results,according to the fact that tool wear is a gradual process,the current wear state is always based on the historical wear state.To deal with the problem of cumulative error of the feedback structure model in the forward prediction,solutions such as model update,"forgotten" factor,combined kernel function,and deviation correction are proposed,which significantly reduce the prediction error.