基于扩展输入空间的迟滞神经网络模型

Neural network based hysteresis model on expanded input space

针对迟滞的非光滑、多值映射特性进行分析,提出了反映迟滞主要特征的迟滞算子的构造方法,并将迟滞算子引入扩展输入空间,实现了将迟滞的多值映射转换为一一映射,建立了关于迟滞的神经网络模型.避免了对非光滑的迟滞求梯度,获得了简单的模型结构和较高的建模精度.

An approach to construct the hysteretic operator is proposed based on the analysis on the characteristic of hysteresis with non-smooth and multi-valued mapping. By introducing the proposed hysteretie operator into the expanded input space, the multi-valued mapping of hysteresis can be transferred into a one-to-one mapping. Thus, the neural network based model can be derived based on the expanded input space. This method avoids the calculation of gradients of the non-smooth hysteresis with respect to its input. Then, an accurate model with simpler architecture has been obtained.