Research On Rough Surface Modeling Based On Generative Adverserial Networks

In order to solve the problem which exist in rough surface modeling method,the characterization parameters couldn’t characterize surface processing method or reconstruct rough surface with certain processing method.The fractal parameters and the surface statistical features cannot be satisfied at the same time in modeling method.The above problems have led to the difficult in simulation rough surface with requirements of scale and prameters in certain process.In order to solve the problem,characterization methods and modeling methods which based on the deep convolutional neural networks were introduced to characterize and reconstruct rough surfaces.The main research contents include:optimizing the method of fractal surface modeling and improving the Ssk distribution of fractal surface topography,modeling rough surface topography of turning and milling based on multi-direction wavelet.Compressed sensing simulation is applied to reduce the rough surface sampling rate.Super resolution reconstruction of rough surface scaning process is carried out.A turning and milling surface data set is constructed to train a processing method classification neural network.The generative adversarial network is trained to model rough turning and milling surface.The fractal parameter discriminant network is trained to identify fractal parameters and dimensions of fractal surface topography.The tribological parameters of rough surface topography could be calculated by training tribology parameter discriminating network.The fractal surface modeling network is trained to match the specified fractal parameters and roughness parameters.Research findings:the linear superposition of the fractal modeling process of discrete Fourier transform can improve the Ssk distribution of the modeling,but it will have an effect on the fractal characteristics of the modeling.Through a large number of parameters adjustment,multi direction wavelet modeling can simulate the rough surface topography produced by different processing methods at any position,direction and scale.The improvement of fractal modeling network has effectively expanded the distribution of fractal modeling in Ssk.Based on the depth learning,the neural network can be constructed effectively to discriminate the milling process,and the network can guide the modeling process of the modeling network to the rough surface,and complete the closed loop from the representation to the modeling.By introducing the full link layer in the generator of the generative advertorial networks,we can effectively enhance the modeling effect and modeling level.of the confrontation network.The generative adverserial network can meet the requirement of multi target modeling on the rough surface fractal modeling,and the fractal parameter and the tribological parameter can be completed at the same time in the modeling process.The training process of this method is simple.Compared with manual design,rough surface is constructed with stronger and wider adaptability.Based on the discriminator,the generator can accomplish different tasks and model and analyze different processing methods.