State Identification And Time Series Prediction For Milling Chatter Based On Deep Learning

In the development process of agricultural machine,the application of intelligent production technology has become the main trend.In order to improve the processing efficiency and precision of agricultural machinery parts,various intelligent equipments and advanced manufacturing methods have been applied to the design and processing of agricultural machinery workpiece.In agricultural machinery,such as crankshaft,camshaft and cylinder liner of tractor engine,gear and bearing of transmission system parts of agricultural tractor,blade and knife arm of rice harvester,pump body and impeller of agricultural sprayer,etc.,the working efficiency and service life of agricultural machinery equipment are directly determined by the shape,size and quality of these workpieces.Milling chatter leads to the increase of roughness,dimensional accuracy and roundness errors of workpiece surface,and may even lead to machine damage and operator injury.In order to realize efficient machining,it is necessary to avoid chatter during milling.Detection,prediction and avoidance of milling chatter are still the key challeges to be solved by modern intelligent manufacturing technology.The specific research work of this thesis is summarized as follows:(1)The analytical model of milling dynamics has been established to obtain the relationship between spindle speed and flutter frequency.The modal parameters of milling system are obtained by hammer experiments,and the milling stability lobe diagram is drawn.The milling test is designed according to the milling stability lobe diagram,the milling vibration signal data is collected,and the milling vibration signal is analyzed from the time domain,frequency domain and time frequency domain.Finally,milling stability analysis is carried out,and the influence of different parameters on milling flutter stability is analyzed by changing the parameters.The milling stability lobe diagram provides guidance for the parameter selection of the subsequent cutting test.(2)The required vibration signals are obtained through milling experiments,and the data are preprocessed to obtain the two-dimensional time-frequency graph required by the deep learning model.Two traditional deep learning methods,CNN and Alexnet,are used to identify the milling chatter state,and the test set accuracy and confusion matrix of the traditional deep learning methods are obtained,which provides a comparison for the model proposed in the following paper.(3)A complex online identification method for milling chatter based on convolutional neural network(CNN),attention mechanical-WFKNN(CNN-Attention MechanicalWFKNN)is proposed.CNN,attention mechanism and weighted fusion K-nearest neighbor(WFKNN)are used to monitor and recognize chatter in time.Six milling lifetime data sets of PHM2010 are used to verify the feasibility of the proposed method,the average accuracy of test sets,confusion matrix and cluster graph of several methods are obtained,and the results are analyzed.The experimental results show that,compared with the existing methods,the identification results of the proposed method are not affected by the quality of the datasets,and the identification accuracy of different datasets is higher than that of other models,which can obtain higher identification accuracy.and the average accuracy is increased by 0.24% compared with traditional KNN classifier,1.5% Alexnet model;Compared with the CNN model and the CNN-Attention mechanism model,the average accuracy is increased by 1.15% and 0.74%,respectively.(4)A composite prediction method for milling chatter time series based on principal component analysis,convolutional neural network and bidirectional Long and Short term memory network(PCA-CNN-Bi LSTM)is proposed.The input of the model is multidimensional time series data extracted in the time domain and frequency domain.The PCA is used for dimensionality reduction fusion of multidimensional data.The fused time series is used as the prediction object,and CNN adaptive feature is used to extract the feature information of the time series.Six sets of open datasets of PHM2010 are used to verify the feasibility of the proposed method.The proposed method is compared with existing methods to obtain the prediction time of four methods.The model is evaluated by four evaluation indexes,such as RMSE、MAE、R2 and MSE.The results are analyzed.The average values of the four evaluation indexes of the proposed PCA-CNN-Bi LSTM model in the six datasets are RMSE=0.2187 、 MAE=0.1650 、 R2=0.9144 and MSE=0.0533,which are better than other models.The experimental results show that the prediction data of the proposed model is closer to the actual data,it can better predict the milling chatter time series,and can be used as an effective method to predict the milling chatter time series.