Servo Control System Of WEDM Based On Acoustic-optical Signal And Deep Learning

The block cutting of large annular parts in tire mold and aero-engine adopts special wire cutting single slot sequential cutting,which has uneven stress distribution,easy deformation and low processing efficiency.The new method of single wire multi-station synchronous Wire electrical discharge machining(WEDM)is a very ideal solution.However,in the single-wire multi-station synchronous machining,there is a highly nonlinear and strong coupling relationship between the electrical signals reflecting the gap size of each station,which brings great difficulties to the independent servo control of each station.The solution of its theoretical and technical problems has become the key to the popularization and application of new technologies.Therefore,this thesis proposes to use acousto-optic signal instead of electrical signal for wire cutting servo control.In this thesis,the basic theoretical issues are studied,such as spark characteristics and machining gap state modeling method,acoustic emission signal and machining gap state modeling method,and machining gap state modeling method,and servo feed control scheme of WEDM.A new mechanism of WEDM control mode based on acoustic-optical signal and deep learning is established.Finally,the scientificity and reliability of the proposed control theory and method are verified by the actual machining data,and the machining efficiency and quality are improved.The main research contents and innovations of this thesis are as follows:(1)Research on processing state prediction algorithm model of CNN-GRU based on spark visual recognition.Firstly,the traditional algorithm is used for image feature extraction,and the relationship between spark image and discharge state is established.Secondly,the discharge state is predicted based on the shape and kinematics characteristics of spark images.And the proposed ' sequence to sequence ' model explores the relationship between spark characteristics and discharge state.On this basis,an innovative ' image to sequence ' model is proposed for training.CNN is used to extract the features of spark images and GRU is used to identify the discharge state.The processing state prediction algorithm model of CNN-GRU is established.(2)Research on the processing state prediction algorithm model of BRTCN based on acoustic emission.Based on the acoustic emission phenomenon in WEDM processing,a dualchannel acoustic emission physical propagation model is established.According to the propagation model and orthogonal correlation characteristics of the main signal and noise signal,based on the structure of batch correlation(BR)and time convolution network(TCN),a batch correlation time convolution neural network(BRTCN)is proposed for the first time.In order to solve the problems of large amount of data,high dimension and label imbalance,a new data set labeling method based on pulse statistical distribution and logarithmic smoothing is developed.Thus the BRTCN processing state prediction algorithm model based on acoustic emission is established.(3)Process experiment verification of machining state prediction algorithm model.The proposed ‘ sequence-to-sequence ' and ‘ image-to-sequence ' models were trained,respectively.CNN was used to extract the characteristics of spark images,and GRU was used to identify the discharge state.The experimental results show that spark images can accurately predict and track the processing state.The accuracy of ‘ sequence-to-sequence ' model is 90 %,and that of ‘ image-to-sequence ' model is 95 %.Through comparative experiments,the effectiveness of BRTCN backbone network for WEDM training task is verified.TCN encoder in BRTCN,through causal convolution,expanded convolution and residual connection,has the advantages of extracting AE sequence characteristics,grasping long dependence and reducing calculation amount.The batch correlation(BR)module in BRTCN mines the batch features obtained by TCN parallel computing.It is found that the BR module mainly affects the final stability loss value,while the encoder mainly affects the stability of the training process.The model with BR set as GRU and encoder set as TCN is the best model with the final MSE loss of 0.009 in the test set.