Study On The Calculation Problem Of Microwave Field Based On Artificial Neural Network

Nowadays, artificial neural network (ANN) recently has received extensive attention as a fast and accurate modeling tool in electromagnetic engineering. Generally speaking, it contained forward problems and reverse problems. Due to the non-linear nature of the artificial and parallel treatment on data, the resolution of the microwave electromagnetic field inverse problem can be made without calculation models but only with the simple couple data about their microwave field system, the model can be simulated by the self-study of the network. Moreover, thanks to the robust character and resistant bothered capacity of the network, the result will not be bothered even if there are some mistakes in data. The above feathers of the artificial neural network will offer a broad future in the application of the microwave electromagnetic field inverse problem. This paper includes two aspects, the research of forward problems and inverse problem.For the research of the forward problems, training dada sets and prior knowledge have to be prepared in the first step of the development of neural models with traditional modeling methods. In the paper, we apply the analytical solution approach basing on the operation's eigenfunction to solve the computation of the microwave electromagnetic. Electromagnetic-ANN models not only preserve the accuracy of the electromagnetic simulations, but also simplify their CPU and memory requirements, and at the same time keep good extrapolation capability.For the research of the inverse problem, conventional method to solve this problem has many disadvantages, such as large computation, great complexity and so on. In the paper we use ANN to solve this problem. For the accuracy and the efficiency of neural models, some techniques are performed, including finer re-sampling method, reordering training patterns in every iteration cycle, logarithmical normalization of training sets, selection of training methodology and so on. Besides, a comparison between the performance of BPNN and RBFNN are presented.