【作者】 李广军；

【导师】 张翠芳；

【作者基本信息】 西南交通大学 ， 计算机应用技术， 2005， 硕士

【摘要】 PID调节器的特点是结构简单,适应性强,应用性广。但是单的PID控制往往不能达到令人满意的程度,对于时变对象和非线性系统,传统的PID控制更是显得无能为力。对于非线性系统来说,神经网络PID具有良好的控制效果,该控制器是将神经网络和PID控制技术融为一体,即具有常规PID控制器结构简单,物理意义明确之优点,同时又具有神经网络自学习、自适应的功能。 在神经网络PID控制器中,神经网络辨识器作为PID控制器参数整定和优化的基础,必须选择辨别精度高的神经网络结构,才能为参数整定和优化提供可靠精确的对象模型。径向基函数神经网络(RBFNN)是一种具有单隐层的三层前馈网络,其网络结构和学习算法与BP网络有着很大的差别,在一定程度上克服了BP网络的缺点,因此本文就在RBF神经网络辨识的基础上实现神经网络PID参数的自整定。论文的主要内容如下: (1) 介绍了PID参数整定的基本方法,并引入基于神经网络的PID参数自整定,提出了一种改进的单神经元PID整定方法,细致地研究了BP神经网络PID参数整定方法。 (2) 提出了一种改进的梯度下降法,并用该方法优化RBF神经网络。然后在RBF神经网络辨识的基础上,实现了基于梯度下降算法、单神经元和BP神经网络的PID参数自整定和优化。 (3) 提出了一种混合递阶遗传算法,用该算法优化RBF神经网络的隐层结构、隐层节点的中心值、核宽度和输出的线性权值,并通过仿真,证明了该方法优于基于梯度下降法的辨识结果。更多还原

【Abstract】 The character of PID controller is simple structure , good adaptability and great robustness. But the simple PID controllers can’t get the satisfied degree , especially for the time-varying objects and non-linear systems,the traditional PID controllers can do nothing for them. To non-linear systems, the NN PID controller has a good control effect in the on-line parameter turning and optimizing. The NN PID controller can make both neural network and PID control into an organic whole , which has the merit of any PID controller for its simple construction and definite physical meaning of parameters ,and also has the self- learning and adaptive functions of a neural network .The NN system structure must be high precision of identification to provide a more accurate object mode. Radial basis function neural network (RBFNN) is a kind of three-layer feedforward neural network with single hidden layer, there is great difference between it’s tructure and learning algorithms with BP neural network’s. So, in the paper, the NN PID is used to achieve PID parameters self-adjustment on RBFNN identification .The works are listed as follows:(1) PID control and its basic parameters auto-tuning methods and the NN PID controller are introduced, then an improved single neural adptive PID controller is presented and PID control based on BPNN is studied in detail.(2) The improved adaptive Gradient-descent algoritms is proposed to RBFNN identification. At the same time , three kinds of controllers is studied. There the adaptive Gradient-descent algoritms, a single neural element and a BP neural network is utilized to achive PID parameters self-adjustment, a RBF neural network is used to identify the controlled plant on-line.(3) Hybrid hierarchy genetic algorithms is introduced to configure the structure and parameters of of RBFNN, and the RBFNN identification results are compared with which produced by Gradient-descent algorithms. The simulation results show the identification effects of RBFNN which is optimized by hybrid hierarchy genetic algorithms is better than those of which is optimized更多还原