基于神经网络补偿的转台伺服系统控制研究

Research on servo control system of flight simulator based on neural network compensation

针对飞行仿真转台伺服系统中存在的非线性摩擦干扰进行了研究,采用一种基于RBF神经网络进行误差补偿的在线自适应控制策略。在基于逆动力学的计算力矩控制方法的基础上,利用RBF神经网络的万能逼近特性在线辨识模型误差,从而对系统进行补偿,其权值自适应律根据Lyapunov稳定性理论推导,保证了系统跟踪误差的收敛及稳定。仿真结果表明该控制策略可使位置MAE指标从0.0087 m提高到0.0016 m,使位置MSE指标从1.0128E-4 m提高到3.3002E-6 m,具有较高的鲁棒性和稳态控制精度。最后分别从隐层节点数及节点中心学习算法的变化两方面提出两种改进方案,仿真结果表明隐层节点数的增加可以进一步减小位置误差,而采用K-means聚类算法解决了神经网络节点中心按经验选取或试凑的困难。

Considering nonlinear friction force in servo system of flight simulator,this paper adopted a new self-adaptive control strategy based on model error compensatation by RBF neural networks. By means of computed torque method based on inverse dynamics,due to the universal approximation propety,the RBF neural network could identify the model error on-line so as to compensate for the system. The adaptive learning law of network weights was developed based on Lyapunov stability theory,therefore it guaranteed the convergence and the stability of tracking error. Simulation experiments show that the control strategy improves the MAE performance indicator of position tracking error from 0. 0087 m to 0. 0016 m,the MSE performance indicator from 1. 0128E- 4 m to 3. 3002E- 6 m. The compensation control strategy is proven to have robustness and high steady state control accuracy. Finally,it proposed two improved scheme respectively from the number of hidden layer nodes and the learning algorithm of node centre. Simulation experiments show that the increase of hidden layer nodes has smaller error range,while K-means clustering algorithm is adopted,which can solve the difficulty to choose the neural network parameters by experience.