Knowledge-based Neural Network And Its Application Study In Fast Modelling Of The Electromagnetic

Within the Computer Aided Design (CAD) related to microwave engineering, theutilization of Artificial Neural Network (ANN), which is a fast and accurate nonlinearmodel, becomes very common. But in order to establish a precise model for ANN, a largenumber of training samples are necessary. Therefore people must pay a great price. At thesame time, the researchers point out to the usage of previously established EM knowledge(as empirical models) in conjunction with neural networks, due to the improvement of theANN generalization capability and reduction in the training data set size, it has greatpotential applications in electromagnetic modeling problem. In this thesis, we directly useKnowledge-based neural network (KBNN) in complex electromagnetic engineering on thebasis of ANN. The main research works are discussed as follows:(1) KBNN theory is described in detail, including source difference method, priorknowledge input (PKI) and knowledge-based neural network. We put forward an improvedneural network. It includes three neural networks: two ordinary neural networks and aParticle Swarm Optimization (PSO) with wavelet mutation Wavelet Neural Network(WNN). The focus of the KBNN is to use previously trained neural network to supply theprior information in this thesis.(2) Aiming at the complexity of the radio signal in the complex electromagneticengineering, high demands are put forward for radio direction technology. In this thesis,Direction of Arrival (DOA) estimation based on KBNN and Locality Preserving Projection(LPP) is studied. The focus is to preprocess the upper triangular matrix of covariance matrixby LPP, the processed data is used as the input of RBF neural network, and the output of thepreviously RBF neural network is used as a part of the PSO-based ANN, thus we completethe DOA estimation based on KBNN and LPP with a smaller error range.(3) Microwave low-pass filter based on KBNN is studied. Firstly microwave circuit isused as the prior knowledge, the training expectation is the difference value between priorcircuit and HFSS electromagnetic simulation. The final experiment simulation proves thefeasibility and advantages of the application of KBNN to electromagnetics including filteretc.(4) An Electromagnetic Band Gap (EBG) structure with bow-tie units proposed by Shasha in our research group based on KBNN and a double layer EBG filter based on KBNN are researched in this thesis. For the EBG structure with bow-tie units, using the ABCDmatrix as a prior knowledge, the source difference method model is a choice. For the doublelayer EBG filter, the last wavelet mutation PSO-based WNN has two additional inputs, inorder to receive prior knowledge information supplied by the outputs of the previouslytrained neural networks. The simulation results show that the model has small differencewith the electromagnetic simulation.