Research On Tool Wear Prediction Methods In Batch Machining Based On Stacked Sparse Autoencoders

With the rapid development of Chinese equipment manufacturing,the batch processing industry of metal mobile phone shells and molds has put forward higher and higher requirements on product quality and manufacturing cost control.The cutting tool is the essential component of the cutting system of the machine tools,its wear has significant influence on the surface quality and dimensional accuracy of the products,and even may influence the machining efficiency and stability of machine tool,as well as reduce the qualification rate of the products and increase the cost of production.It is of great importance in increasing the tool usage and the qualification rate of the products,reducing the cost of production and increasing productivity by monitoring the tool wear in real time.Many researchers in academia and industry focus on predicting the tool wear accuracy and efficiency.Thus,the tool wear condition and value prediction in the batch processing of metal mobile phone shells and molds is researched in this thesis,the main contents are as follows:The relationship between the tool wear and machining quality in the metal mobile phone shells and molds processing is studied.The relationship between the tool wear and machining quality is demonstrated by the statistical analysis of tool wear and machining quality data.Taking the measurement data of machining quality as the evaluation index of the tool wear,and taking the vibration signal as the monitoring signal,a tool wear prediction scheme for machining quality control is established.A method used for tool wear sensitivity feature extraction based on stacked sparse autoencoders is established.Considering the influence of information loss in the process of reducing the dimension of signals to extract features,the automatic features extraction of vibration signals in time domain,frequency domain and wavelet domain is realized based on stacked sparse autoencoders.The Pearson correlation coefficient between the proposed features and the tool wear values is calculated,and the sensitivity of the proposed features is analyzed to verify the effectiveness of the method.A supervised tool wear state prediction model based on multidimensional stack sparse autoencoders is proposed.Considering the fact that different tool wear information reflected by vibration signals in time domain,frequency domain and wavelet domain,the parallel learning and feature fusion networks are constructed to realize the parallel extraction and fusion of multi domain signal features.And the iterative optimization of the features is realized based on the back propagation algorithm,which is helpful to predict the tool wear state more accurately.Furthermore,based on the framework of multidimensional stack sparse autoencoders model,a semi-supervised tool wear state prediction model considering a small number of labeled samples is presented.Considering the fact that a large number of samples are not labeled that make them impossible to back propagate,the feature extraction of labeled samples and unlabeled samples is realized based on the parameters sharing method.A hyper parameter is introduced into the loss function,the dependence of the model on labeled samples and unlabeled samples is balanced by optimizing the value of .The accuracy and generalization of these two models are verified by the example of metal mobile phone shells processing.The machining surface quality defects can be accurately predicted based on the tool wear state prediction result of the models.A tool wear quantitative prediction model based on regression multidimensional stack sparse autoencoders is proposed.Considering that the multidimensional stack sparse auto encoders model cannot achieve the numerical prediction,the Softmax function of the model output layer is abandoned,and the nonlinear regression function based on the increase and decrease factor and change rate is introduced.The method of the increase and decrease factor is presented to update the model parameters,and the quantitative prediction of tool wear is realized based on the nonlinear regression function.The prediction accuracy of the model can be kept within 1 μm through the example of molds processing,and the effective prediction of machining dimension accuracy can be realized.