Evaluation Of Tool Wear Of Milling Machine Based On Deep Learning

With the progress of technology,intelligent manufacturing has developed rapidly as the core content of the manufacturing industry.Numerical control machine tool as the manufacturing industry "industrial mother machine",the tool as the key executive parts,in the process of industrial processing is very important,the wear state and remaining life of the tool directly affect the processing quality and processing efficiency.Therefore,the realization of tool wear condition monitoring and life prediction is of great significance for improving machining efficiency and quality.This paper takes milling machine tool as the research object,analyzes the mechanism of milling cutter wear degradation,combined with advanced deep learning theory,provides a new scheme for milling cutter wear condition monitoring and life prediction.This paper is based on the data set of PHM2010 milling cutter.Firstly,the tool original signal data is preprocessed.The high frequency of original data acquisition leads to long signal data length,large scale and redundant information,so dimension reduction and resampling pre-processing are required.The advantages and disadvantages of existing super long time series data preprocessing technology were analyzed,and the original time series data was resampling equidistant.Through expert experience method,different resampling frequencies were determined under the specific application scenarios of tool wear condition monitoring and life prediction,and the dimensionality reduction preprocessing of the original data was completed.Secondly,the wear condition monitoring and life prediction algorithm of milling cutter are discussed.On the basis of the existing deep learning technology,CaAt-Resnet-ld algorithm was proposed to monitor the tool wear condition.The channel attention mechanism was introduced into the residual network,and the one-dimensional convolutional neural network was redesigned.In order to predict tool life,the MAG-AE algorithm was proposed,and the cod-decoding model architecture was adopted.The encoder included multiple convolutional neural networks,the decoder included multiple attention mechanisms and GRU cyclic neural networks,and a monomonotonicity loss function was newly defined.Finally,the experimental evaluation and analysis were completed.The preprocessed signal data is reconstructed into multiple data sets according to different tasks.In the monitoring and evaluation experiment of milling cutter wear state,compared with LSTM and Gated-Transformer models,the Accuracy evaluation index reaches 89.27%,about 3%higher than the second place.The results prove that the proposed model achieves improved performance.In the milling cutter life prediction evaluation experiment,compared with GRU and DCNN models,the MSE and MAE reached 65.25 and 9.55,respectively,about 18 and 1 better than the second place,and the prediction curves were analyzed to prove that the performance of the proposed model and the newly defined monomonotonity loss function was improved.