| Deep learning has been proven to hold superior learning ability and representation ability,and then is widely used in computer vision,natural language processing and many other fields.However,data drivendeep learning has two drawbacks,including:1)high-quality data samples with rich types,abundant quantities and good annotations are required for model training.2)poor interpretability of the model with "black box"characteristic prevents the good understanding of the working mechanism of the model,the effect on the result by the extracted features,and the relationship between the features.In face of these problems,in this thesis an intelligent computing method is studied by fusing mechanism with data driven deep learning model.Through mechanism analysis and modeling,we can:1)understand the physical mechanism of the task,which guides not only the extraction of the semantic features but also the design of the network architecture;2)with the understanding of the task mechanism and semantic features,construct the inverse mapping model of physical feature for the difficult inverse problem.In this thesis,the investigation of mechanism and data fusion are focused on the field of electromagnetic calculation and stress-strain calculation,and the forward and inverse prediction models are developed for the electromagnetic(EM)detection of flight targets and the stress-strain prediction pile foundation under static load.The main contributions include:(1)For the task of EM detection of flight target,the physical mechanism of radar cross section(RCS)signal of the flight target with regular geometry is analyzed firstly,and the analytical model via finite element analysis(FEA)is established under the condition of far-field radar plane wave and single excitation source.Then,a mesh data preprocessing method is designed,which can reduce the mesh parameters and input parameters.Based on grid structure of the analytical RCS generation model and the understanding of EM mechanism,a generative adversarial network(GAN)model is designed to generate the RCS signals according to the key factor of scattering coefficients.The proposed GAN model is evaluated,and the experimental results verify that it can simulate the FEA model with great generalization performance and greatly reduced computational complexity.Moreover,a deep learning network is proposed to recover the EM scattering coefficients on the base of grid structure of FEA model,and the prediction performance is verified by experiments.(2)Aiming at the prediction task of stress-strain distribution of pile foundation under static load,the physical mechanism is analyzed,and then the suitable constitutive model is determined and the FEA model is constructed.In order to facilitate the design of deep learning model,the grid structure of the FEA model is translated from three-dimensional solid structure into a two-dimensional matrix while the connection relationship among grid nodes is maintained as far as possible.Then,a Resnet type deep learning model is developed for stress-strain prediction of foundation pile,and its excellent performance is sufficiently verified.At last,a deep learning model combining U-Net and diffusion model is designed to recover the accurate elastic modulus and friction coefficient of soil layers from coarse values,and the experimental results verify the effectiveness of the proposed model.Through the study,the intelligent computing method via mechanism and data fusion can effectively distill and combine the advantage of both deep learning and FEA.The accurate understanding of mechanism from FEA can guide the deep learning model design,which not only promote the efficiency of both architecture design and dataset construction,but also greatly improve the generalization performance of deep learning models. |
