新颖的神经网络逆建模方法及其应用

NOVEL NEURAL NETWORK INVERSE MODELLING METHOD AND ITS APPLICATION

使用神经网络设计微波器件时,经常用到神经网络逆向模型。对于复杂的器件输入输出关系,直接逆向建模方法无法满足精度的要求,而其他精度可以满足要求的逆建模方法又具有结构复杂、计算量大等缺点,提出一种新颖的设计微波器件的逆建模方法。该方法只需建立神经网络正向模型,并在保持其权值不变的基础上,通过自适应调节最速下降法学习速率更新正模型的输入参数,使模型输出与理想输出误差达到最小,实现逆模型的功能。此方法没有单独建立逆模型,却能实现逆模型的功能,因此比其他方法简单很多。自适应学习速率的引入进一步改善了模型的速度和精度,将该方法应用到阻抗变换器的设计中,结果表明,相对于直接逆模型,此模型的运行时间减少了7.49%,所求变换器长度和频率的均方误差分别改善了99.95%和98.81%,可以解决多解的问题并用于实际微波器件的设计。

When using neural network （NN） to design microwave devices, the inverse model of NN is used frequently. For the complex input-output relationship of devices, direct inverse modelling approach cannot meet the accuracy requirements, and other inverse modelling approaches that can meet the accuracy requirements have the shortcomings of complex structure and heavy computation load, so in this paper we proposed a novel inverse modelling method for designing the microwave devices. The method only needs to establish NN forward model, then based on keeping its weight unchanged, it updates the input parameters of the model by adaptively adjusting the learning rate of fastest descent method, makes the error between model＇s output and ideal output to a minimum and thus realises the function of inverse model. Although the method doesn＇t build a separate inverse model, it can realise its function, so it is much simpler than other methods. The introduction of adaptive learning rate further improves the speed and accuracy of the model. Applying the method to designing impedance converter, results showed that relative to direct inverse model, the running time of this model reduced by 7.49%, the MSE of the calculating length and frequency of the converter improved by 99.95% and 98.81% respectively, it could solve the problem of multi-solution and be applied to the design of practical microwave devices.