传感器动态建模FLANN方法改进研究

Research on improved FLANN for sensor dynamic modeling

提出一种改进的函数连接型神经网络(FLANN),并将其应用于传感器动态建模。首先,将单输入单输出(SISO)的传感器系统表达为动态差分方程模型;再充分考虑动态模型输出的历史值与参数之间的关系,对模型输出与参数的偏导数进行重新推导,得到了对权值参数偏导数的更高精度估计;最后,利用该模型梯度进行迭代训练,加快了网络收敛速度并提高了收敛的稳定性。实验结果表明,改进FLANN具有更快的收敛速度和更强的鲁棒性,十分适合传感器动态系统的建模。

An improved functional link artificial neural networks （FLANN） is presented and applied to dynamic modeling for sensor. Firstly, the single-input single-output （SISO） sensor is expressed as a dynamic difference equation model. Secondly, the partial derivatives of the dynamic model output w. r. t its parameter are re-derived and the dependences of the past dynamic model output on the parameters are also considered. Therefore more accurate evaluations of partial derivative of the weight parameters are obtained. Lastly, through iterative training using the novel model gradient, the improved FLANN effectively accelerates the convergence rate and enhances the stability of the network. Experimental results show that the improved FLANN has higher convergence rate and stronger robustness, which is more suitable for sensor dynamic modeling.