Study On Neural Network Optimisation Of Substrate Integrated Waveguide Components

In recent years, a new technique, substrate integrated waveguide (SIW) has been constructed by two parallel rows of via holes in a metalized planar substrate. It can be fabricated as a printed circuit board assembly or fabircated using low-temperature-cofires-ceramic(LTCC) process, so it is easy to be integrated in modern planar microwave and millimetre-wave circuits. The SIW has merits of low cost, low profile, and easy integration with planar circuits. It’s convenient to use SIW to form microwave and millimetre-wave components such as filters, power dividers, power combiners, directional couplers, resonators, antennas etc. A simulation of a SIW structure often needs a relative long time, for there are many cylinders (the vias) in the structure which need to be meshed, this increases the complexity of calculation. Unfortunately, generally speaking we need optimize two or even more parameters of the component, that will cost plenty of time, and it’s not easy to find the optimized value. So it’s necessary to introduce neural networks to help optimizing the parameters.Based on Maxwell’s equations and Floquet mode functions the electromagnetic interaction between TE incident wave and periodic conductive wall is analyzed and calculated. The electromagnetic leakage represented by the transmission coefficients of different order of Floquet mode is obtained. The results obtained are fundamental for analyzing the structures and devices with periodic conductive walls.A Vibration Gravity Field method is proposed. Periodic normally distributed random ajustment of the weight and gradient descent training help the neural network jumping out of the local minimal value area and falling into the lower region of generalization error. The training time control method and parameters criterion method are discussed. The effect of vibration parameter is analyzed and the appropriate interval of vibration parameter is recommended. Calculation show that Vibration Gravity Field method has better performance than standard BP algrithm, early stop method and Bayesian Regularization algorithm in finding better solutions which have lower generalization error, and it can enhance the neural network generalization capacity.We proposed a new method that uses π to generate random number. An improved bailey-borwein-plouffe algorithm is proposed to make the continous expanding of π and generating of random numbers more efficiently. Calculation shows that the improved bailey-borwein-plouffe algorithm calculates faster than standard bailey-borwein-plouffe algorithm in continuous expanding,and can improve the output efficiency of the random number generator. The expansion of π is naturally unpredictable, which can be a good random source. Combined with a linear congruential generator, we creat a new random number generator. Tests show it has excellent statistic property and better uniformity than a linear congruential generator, and it has no max period limit. It may be applied in applications where high randomness is desired.A training method based on Ebbinghaus Memory Curve is proposed, which can learn new samples in less training than batch samples training and has almost the same performance as it. Training result shows that it’s a high efficiency incremental training method, it can protect the previous training result and the generalization capability. It is appropriate for the online training of neural network.We use neural networks to optimize some key parameters in a SIW power divider, a post rectangular waveguide band-pass filter and a rectangular waveguide slot antenna components design. The results show that the neural network works well, helps to get the right parameters and make the optimization process more convenient.