Research On Tool Wear Monitoring Method Based On Hidden Markov Model

Tool wear will directly influence the processing quality of workpiece,and reduce the manufacturing efficiency and economic effectiveness of manufacturing system.In order to reduce these negative effects,it is crucial to establish a dependable and precise tool wear monitoring system.The existing problems in the current research of tool wear monitoring include:1)The phase information of the monitoring signal can reflect the tool wear degree to some extent,while the time-frequency phase information of the signal is often ignored in the previous research.2)For tool wear monitoring under a single working condition,the hidden state transition probability learned from the traditional hidden Markov model cannot accurately describe the tool wear change,so the classification and recognition results are not ideal.3)For tool wear monitoring in various working conditions,the pure data-driven hidden Markov model cannot accurately describe the tool wear law in each working condition simultaneously.In view of the above three problems,the research of this paper is as follows:1)A feature extraction method based on s-transform time-frequency amplitude and phase information is proposed.Firstly,the time-frequency matrix of the monitoring signal are obtained by using S transform,and then the feature is extracted from the whole time-frequency amplitude/phase matrix,time-frequency amplitude contour,time-frequency maximum value/phase curve,frequency maximum value/phase curve and other aspects by using statistical method.Using mutual information feature selection algorithm,the optimal subset that can represent tool wear is selected from the high dimensional feature set,which provides reliable input for subsequent decision model.2)According to the characteristics of unbalanced distribution of tool wear data,this paper proposes an enhanced integrated hidden Markov model with multi-classification focused positive samples combined with ensemble learning algorithm.Based on the traditional AdaBoost algorithm,this paper constructs a classification evaluation index suitable for unbalanced data,and optimizes the loss function and the weight update function of the base learner.The average recognition rate of the proposed integration model is 3.6%higher than that of the traditional hidden Markov model.3)By introducing the physical model of tool wear to constrain the classification of hidden states,and adding the influence of real-time features into the output function of wear value,this paper proposes a hidden Markov model of physical information.Experience labels of process parameters and tool wear physical model production were added to the training set,and the training strategy of physical information fusion was used to complete the training of the model.The average recognition rate of this method in classification problems reached 99.5%,and the average MSE of regression problems reached 7.052.This paper presents an effective and reliable tool wear monitoring method.The feature extraction method based on S-transform provides reliable feature input for the decision model,and two methods of integrated hidden Markov model and physical information hidden Markov model are proposed to deal with single working condition and multiple working condition,respectively,to achieve accurate and reliable tool wear monitoring.