动态递归神经网络及其在机敏机构辨识中的应用

System Identification of Smart Mechanisms Based on Dynamic Recurrent Neural Networks

介绍了动态递归神经网络的数学模型 ,提出了一种学习因子自适应调节的改进算法。利用修剪法确定拓扑结构 ,提高了神经网络的泛化能力。以实验数据为样本 ,采用复合辨识方法离线设计了动态递归神经网络辨识器 ,获得了机敏机构的非线性动力学模型 ,其精度明显高于传统的

In this paper, the dynamic recurrent neural network (DRNN) was applied to the estimation of a smart mechanism featuring piezoceramic actuators and strain gage sensors. The mathematical model of a 4-layered DRNN was presented firstly. To guarantee convergence, a fast learning algorithm VLR was proposed for the DRNN whose learning rate could be regulated adaptively in accordance with Equation (28). Moreover, the suitable topology of the DRNN was determined utilizing the pruning algorithm so as to increase its generalization capability. On the basis of these improvements, a DRNN identifier was designed off-line by means of the compound identification method shown in Figure 3. As can be seen from Figures 4 and 5, the identifier obtained is proved to be more accurate than the KED theoretical model and may be used for suppressing the elastodynamic responses of flexible linkage mechanisms with resort to the neural networks based model reference adaptive control (MRAC) strategy.