| Modern processing and manufacturing lines are gradually becoming highly automated,and the production mode of "unmanned workshop" or "less workshop" is becoming more and more common.In this mode,if a tool of a certain process in the production process breaks,it is likely that other tools and subsequent parts will be scrapped due to abnormal processing,which will bring certain economic losses to the enterprise.Therefore,the CNC system needs to be able to monitor the tool breakage fault online and make a broken tool alarm in time to reduce the loss of the enterprise,improve the reliability of the production line,and further realize the automation of production processing.Current tool monitoring methods based on acoustic emission,machine vision,etc.,usually require additional data acquisition equipment.This article chooses to directly use the internal spindle power data of the CNC system as the basis for monitoring whether the tool is broken,which reduces the cost of increasing external equipment.In order to process the power data as the input samples of the neural network model,it is necessary to filter out a piece of power data with a small amount of data that may contain broken tool features from the original huge time-domain data.In this paper,the tool power data in machining is preliminarily segmented based on the command field,and the spindle power data containing the command line corresponding to the actual cutting command is used as candidate data.Based on the last sampling point of the candidate data,the appropriate data points are shifted forward to intercept the appropriate data points as the characteristic data for monitoring tool breakage.Finally,the feature data is normalized from 0 to 255 numerically,and the dimension is reorganized from the original one-dimensional data to three-dimensional data,and converted into image format samples,making it more suitable as an input sample of the neural network model.A broken tool monitoring algorithm based on neural network model requires a large number of tool samples for training.However,there are too few negative samples of broken tool data collected by processing,which is seriously unbalanced with the number of positive samples of normal processing data.Using generative adversarial networks(GAN)can generate a large number of broken knife samples with broken knife features,but the traditional GAN structure is simple and unstable during training.Based on the traditional GAN,this paper proposes a generative adversarial network(JGAN)based on the joint distribution of random variables and broken knife samples,and uses certain training techniques in training to improve the stability during training and the quality of the generated samples,And use JGAN to expand the broken knife sample,to solve the problem of sample imbalance.Monitoring whether the spindle power during machining is broken data is a two-class problem.Using a tool positive and negative sample set to train a deep convolutional neural network(CNN)can automatically extract the features of the tool data and realize the binary classification of the input tool samples.However,the internal parameters of the deep CNN are too many,the time cost of training the model from scratch is too high,and the tool samples are not sufficient,the model is prone to overfitting.In this paper,the migration training for tool samples is used to fix the weights of a large number of convolutional layers in the pre-trained CNN model with feature extraction,retrain the remaining fully connected layer parameters,and successfully migrate the original CNN model to the broken tool monitoring task.In the end,this paper uses the CNN model after migration training to develop a complete broken tool monitoring function module,and conducts a broken tool monitoring and processing experiment for the drill bit.The accuracy of the experimental results verifies the feasibility of a broken tool monitoring algorithm that uses JGAN to generate broken tool samples and depth CNN to monitor the broken tool. |
PDF Full Text Request