Research On The Discharge Gap State Of High-speed Wire EDM Machine Tools Based On Deep Learnin

High-speed Wire-cut Electrical Discharge Machining(HS-WEDM)is a specialized processing technology with promising development prospects pioneered in our country.It utilizes high-frequency pulses in the minimal gap between the electrode and the workpiece to generate high temperatures for cutting the workpiece.HS-WEDM is commonly used for highprecision machining of large-sized,highly hard,and conductive materials with complex shapes,holding significant strategic importance for advanced equipment and civil instruments in our country.The introduction of the high-speed wire mechanical structure in HS-WEDM machines enables wire recycling,resulting in substantial cost savings and improved processing efficiency compared to other EDM machines.However,current research primarily focuses on enhancing wire processing speed,with limited attention given to processing quality.In HS-WEDM,detecting the discharge gap state is crucial for servo control and machining quality.While existing studies mainly identify the gap state using voltage and current measurements,few explore the analysis of spark images and acoustic emission(AE)signals generated during the machining process.The invention of EDM was initially inspired by the electrical spark erosion effect on switch contacts.Building upon this inspiration and incorporating modern deep learning technology,this paper conducts a comprehensive analysis of a series of sparks generated during the HS-WEDM process.In addition to spark optical signals,the characteristics of AE signals,including frequency,amplitude,and period,significantly differ in various discharge processes.To accurately predict the discharge state using spark images and AE signals,this paper mines the mapping relationship between AE signals,spark images,and discharge states through signal analysis and the design of a corresponding deep learning model.The optimal model is obtained and verified through experimental predictions,providing a foundation for subsequent process control optimization.The core contributions of this paper can be summarized into the following three aspects:(1)Designing a synchronized data acquisition system for capturing various signals in different dimensions in HS-WEDM,including pulse signals,temperature fields,spectra,natural spark images,and AE signals.Preprocessing and analysis of the data are performed,and a pulse state quantification method based on sliding median filtering and energy is proposed to facilitate the construction of the machining state dataset.Preliminary analysis reveals a maximum temperature field of 229.7℃ under the cooling fluid during spark discharge,a predominance of infrared light in the spectral distribution,and the generation of high-frequency AE signals during the discharge process,with corresponding temporal and spectral characteristics to the pulse signals.(2)Proposing a machining state detection model based on optical flow estimation and residual(2+1)D convolutional neural networks to explore the correspondence between sparks and machining states from static,dynamic,spatial,and temporal perspectives.Through a series of experiments,it is validated that the introduced optical flow estimation module utilizing the2D-GRU iterative method extracts both static and dynamic information from multiple spark images,providing accurate prior knowledge to enhance stability,accelerate convergence,and improve accuracy.Moreover,the residual(2+1)D convolution exhibits superior spatial and temporal feature extraction capabilities compared to regular 3D convolutional networks,enhancing the model’s learning capacity.Transfer learning is also verified to be applicable for the prediction task on the spark dataset.(3)Addressing long-term dependency issues in two types of long sequential data,AE signals,and pulse signals,by proposing a machining state detection model based on batch iteration and temporal convolutional networks.This approach alleviates the difficulties of learning long-term dependent features and achieves accurate prediction of the machining state.Batch iteration incorporates the concept of "divide and conquer," where AE data from different batches are separately processed by temporal convolutional networks to capture their respective patterns.These patterns are then organized into the final solution using a batch iterator,avoiding the problem of gradient diffusion when training a single long sequence model.Experimental results demonstrate that a dual-path AE model with a gate recurrent unit as the batch iterator and a temporal convolutional neural network as the encoder achieves the optimal performance,with a comprehensive mean square error test loss of 0.009.In conclusion,this paper analyzes the signals in HS-WEDM,designs corresponding deep learning models to explore the mapping relationship between acoustic and optical signals and discharge states,and validates the optimal models through experimental verification.The research results demonstrate that the proposed spark model and AE model effectively predict the discharge gap state in HS-WEDM,highlighting the crucial information embedded in these two different-dimensional signals.The obtained optimal models provide valuable information feedback for practical machining control systems.