Research On End-milling Tool Wear State Recognition Method Based On Deep Learning

Impeller is the core part of turbine compressor,pump of nuclear power plant and other large rotating machinery equipment.The overall processing quality of impeller channel structure will directly determine the operation performance of the equipment.As the main part of impeller runner milling,when the tool is worn to a certain extent in the process,the overall processing quality of the workpiece will be adversely affected,and the production efficiency of the workshop will be reduced.Therefore,the research on tool wear state identification in milling process has great practical engineering application value.In this paper,by monitoring the motor current signal of CNC milling machine spindle,the deep learning algorithm is used to realize adaptive feature extraction and wear state recognition.The main research contents of this paper are as follows:(1)The research status of tool wear condition monitoring at home and abroad is fully discussed.The wear mechanism of milling cutter is studied.Through theoretical analysis,the rationality of using the motor current signal of machine tool spindle as the detection signal of tool wear state is verified.Different collection schemes were designed,and one group of data at different wear stages and the other one group of production site data were collected respectively.A total of two data sets were obtained as data support for the deep learning algorithm.(2)Two different preprocessing methods were used to split and screen the samples of the two data sets.For different wear state experiments,the wear grade was divided according to the wear quantity of the rear cutter face,and then the samples were separated.For the production field experiment,the original long sequence signal is automatically split by the sliding window,and then the mean value of each sample after splitting is calculated.Finally,the effective sample screening is realized by searching the local peak value and setting the threshold.(3)A one-dimensional deep convolution autoencoder is used to adaptively extract the characteristic information of tool wear state based on the original time domain signal itself.Firstly,the three-phase current signals were fused into current effective values and normalized,which were used as the input of the model.Then,the feature information based on the signal itself is extracted by unsupervised pre-training.Finally,the second supervised fine-tuning is carried out to realize the recognition of different tool wear states.(4)In order to solve the problem that the one-dimensional deep convolution autoencoder ignores the characteristics of time series and causes the loss of characteristic information,a gated loop unit and an attention mechanism are introduced based on the original model.At the same time of extracting the multi-dimensional feature information from the convolutional layer,the time characteristics of the time series are extracted synchronously,so as to comprehensively and deeply excavate the feature information of different tool wear states.(5)Based on Lab VIEW platform,the milling cutter condition monitoring system was built,and the current signal of milling cutter in the milling process was collected in real time,and the waveform of current effective value in time domain and frequency domain was visualized.At the same time,the deep neural network which has been trained is used to monitor the tool condition,and the judgment result of tool wear condition is obtained directly.