Research On Tool Wear State Prediction Method Based On Deep Learning

In the cutting process,the cutting force and friction heat generated by the interaction between the tool and the workpiece and the chip cause tool wear,which has a decisive influence on the surface quality and dimensional accuracy of the parts.If the wear state cannot be accurately grasped,the corresponding process flow cannot be reasonably adjusted.In particular,the excessive wear of the tool will lead to scrapping and unplanned downtime.At present,the replacement of the tool is still mainly based on the operator’s experience or the set use time,which is easy to cause unnecessary tool waste and decline in processing quality.During the machining process,the real-time identification of tool wear state by using the physical signal reflecting the cutting state can timely repair the tool to ensure the machining quality or avoid the occurrence of broken edge events.At present,the tool wear condition monitoring technology is mostly suitable for a single working condition,but there is little research on the tool condition monitoring under the multi-variety and small-batch production mode to meet the diversified needs of users.This thesis takes milling tool wear as the research object,and combines multisensor fusion techology to establish a tool wear prediction model based on deep learning,in order to achieve accurate prediction of tool wear under multiple working conditions.The main research contents of this thesis are as follows:(1)The power and vibration sensors are selected to collect the orthogonal experimental cutting signal of tool wear.The mapping model of cutting power and cutting parameters is established by calculating the instantaneous differential cutting power caused by tool rotation and feed at the micro cutting edge.The collected signal is preliminarily preprocessed,and the Chebyshev Ⅱ high-pass filter is used to digitally filter the low-frequency interference signal in the vibration signal.Based on the wavelet threshold denoising method,the filtered signal is subjected to two-layer wavelet decomposition of the heuristic threshold to eliminate the highfrequency noise component in the signal.Through the preprocessing of the original data collected directly from the experiment,a large amount of noise information unrelated to tool wear is removed.(2)The time-frequency domain eigenvalues are extracted from the signal time-frequency domain by calculating statistical features and autocorrelation features.The wavelet packet transform is used to decompose the signal into multiple frequency bands to better capture the multi-scale energy time-frequency characteristics of the signal.The endpoint effect and mode mixing are avoided by adaptive noise complete ensemble empirical mode decomposition.The signal is decomposed into the intrinsic mode function from the time-frequency domain,and the energy entropy and sample entropy eigenvalues are extracted.The above eigenvalues and the average value of power are jointly evaluated by two feature evaluation methods based on crosscorrelation method and Pearson correlation coefficient method,and finally the eigenvalues most related to tool wear are selected.The original data is enhanced by translation and adding Gaussian white noise,and the multi-domain feature joint matrix of tool wear is constructed.(3)The multi-condition tool wear prediction is studied by using convolutional neural network,long-term and short-term memory network and attention mechanism,and a multicondition tool wear prediction model based on 1DCNN-BiLSTM-Attention is designed.By tuning the hyperparameters of the model and the number of neural network layers,the learning efficiency,expression ability,fitting ability and prediction performance of the model are improved on the basis of avoiding overfitting.The multi-domain feature joint matrix of nine sets of working condition data is used to form a training set and a test set to verify the multi-condition tool wear prediction ability of the model,so as to achieve accurate prediction of tool wear.By comparing three traditional machine learning algorithms and four deep learning models,this model has the best accuracy and robustness in multi-condition tool wear prediction.