Data Driven MOEA/D And Its Applications In The Design Of High Performance EM Structures

Design of electromagnetic structure affects the performance of microwave devices to a great extent.The design schema based on fragment-type structure configuration technique and distribution optimization technique gives us a new opportunity to explore miniaturized and high-performance planar electromagnetic structures.Machine learning(ML)and artificial intelligence(AI)technology can help improve the efficiency of the structure design schema.In this thesis,multi-objective evolutionary algorithm based on decomposition(MOEA/D)is choosen as the representative,by combining with data-driven assisted improvement offspring reproduction(OR)or fitness assignment(FA)operators of heuristic optimization,an efficient intelligent optimization design scheme of high-performance planar electromagnetic structures was proposed.The main research contents of this thesis are as follows:1).Data-driven optimization theory based on MOEA/D.aiming at the MOEA/D heuristic optimization algorithm,a data driven OR strategy was proposed to select appropriate individuals from all the samples obtained in the optimization process as learning objects,and the appropriate aggregation generation method was used to obtain more appropriate offspring individuals,so as to improve the optimization search efficiency.On the other hand,the data-driven FA strategy is adopted to select an appropriate ML/AI model and fit the samples obtained in the optimization process to obtain a high-precision surrogate model,which reduces the cost of simulation solution and improves the optimization efficiency.The numberical test of multiobjective benchmark problems effectively verifies the ability of data-driven theory to improve the efficiency of multi-objective optimization problem.2).Design of chipless radio frequency identification(RFID)tags based on MOEA/D.By proposing MOEA/D combined with loop genetic operator(MOEA/DLGO)and using the prior data-driven based loop genetic operator,the OR strategy of MOEA/D is improved,which can be used to ensure the inheritance(crossover and mutation)of the loop structure in the iterative update.Moreover,the sample data obtained from the FA is used to enrich the prior data set.A high-capacity chipless RFID tag embedded in QR code is designed and tested for verifying the proposed MOEA/D-LGO.3).Design of multi-input multi-output(MIMO)antenna based on MOEA/D.By proposing MOEA/D combined with ensemble offspring reproduction(MOEA/DEOR),data-driven OR strategy is used to improve algorithm convergence efficiency,while the global OR is adopted appropriately to ensure the global performance of the MOEA/D algorithm.Combining with the theory of ensemble learning,the most suitable offspring is selected from several possible effective OR strategies for different design problems and different sample data set.Two compact high-isolation MIMO antenna designs are designed and tested for verifying the proposed MOEA/DEOR.4).Design of reflectiveless filter based on MOEA/D.A MOEA/D optimization technique combined with multi-surrogate model(MOEA/D-MSM)was proposed,for different problems and different sample data set,MOEA/D-MSM selects the most suitable data-driven ML/AI model instead of full-wave solver to evaluate the electromagnetic performance of the structure.At the same time,a normalized Gaussian network(NGnet)configuration technique is introduced to further improve the precision of the data-driven surrogate model,so that lower dimensional decision matrix can be used to describe the necessary fine structure.The proposed MOEA/DMSM is verified by a high frequency wideband reflectiveless filter from theoretical design to practical measurement.To sum up,by combining the data-driven theory to improve the multi-objective heuristic optimization algorithm,different data-driven optimization technologies were proposed,which can provide a more efficient and reliable data-driven MOEA/D optimization framework for high-performance electromagnetic structure design.