GA-BASED PID NEURAL NETWORK CONTROL FOR MAGNETIC BEARING SYSTEMS

In order to overcome the system non-linearity and uncertainty inherent in magnetic bearing systems, a GA(genetic algorithm)-based PID neural network controller is designed and trained to emulate the operation of a complete system (magnetic beating, controller, and power amplifiers).The feasibility of using a neural network to control nonlinear magnetic beating systems with unknown dynamics is demonstrated. The key concept of the control scheme is to use GA to evaluate the candidate solutions (chromosomes), increase the generalization ability of PID neural network and avoid suffering from the local minima problem in network learning due to the use of gradient descent learning method. The simulation results show that the proposed architecture provides well robust performance and better reinforcement learning capability in controlling magnetic bearing systems.     

ADAPTIVE LEARNING CONTROL OF CUTTING PARAMETERS FOR SCULPTURED SURFACE CUTTING BASED ON GENETIC ALGORITHMS AND NEURAL NETWORK

An adaptive learning control scheme intended to the on-line optimization of sculptured surface cutting process is presented. The scheme uses a back-propagation neural network to learn the relationships between process inputs and process states. The cutting parameters of the process model are optimized through a genetic algorithms(GA). The capacity of the proposed scheme for determining optimum process inputs under a variety of process conditions and optimization strategies is evaluated on the basis of milling of a sculptured surface using a ball-end mill. The experimental results show that the neural network could model the cutting process efficiently, and the cutting conditions such as spindle speed could be regulated for achieving high efficiency and high quality. Therefore the proposed approach can be well applied to the manufacturing of dies and molds.     

FUZZY NEURAL NETWORK FOR MACHINE PARTS RECOGNITION SYSTEM

The primary purpose is to develop a robust adaptive machine parts recognition system. A fuzzy neural network classifier is proposed for machine parts classifier. It is an efficient modeling method. Through learning, it can approach a random nonlinear function. A fuzzy neural network classifier is presented based on fuzzy mapping model. It is used for machine parts classification. The experimental system of machine parts classification is introduced. A robust least square back-propagation (RLSBP) training algorithm which combines robust least square (RLS) with back-propagation (BP) algorithm is put forward. Simulation and experimental results show that the learning property of RLSBP is superior to BP.     

基于模糊神经网络和遗传算法的仿人智能PID控制器设计

阐述一种新型的模糊神经网络加遗传算法的智能PID控制器     

ARTIFICIAL NEURAL NETWORK AND FUZZY LOGIC CONTROLLER FOR GTAW MODELING AND CONTROL

0 INTRODUCTION(The satisfied control of the overall weld process is not easily accomplished, largely due to the inadequacies of the available process models. Without exceptions, most welding control methods are based upon the analytical welding models. Although these models are derived directly from the physical laws that govern the main features of the weld pool, a number of assumptions are made to obtain the mathematical solutions and some variables are ignored due to the complexity of t…     

复合材料加筋结构的神经网络响应面优化设计

针对复合材料加筋结构优化设计的复杂性,提出利用人工神经网络结构近似分析响应面来反映结构设计输入与结构响应输出的全局映射关系的优化方法。通过正交试验设计选取合适的结构有限元分析样本点,进行神经网络响应面的构建和训练;将神经网络响应面作为目标函数或者约束条件,汇同其他常规约束条件完成优化模型的建立,并应用遗传算法(GA)进行优化,从而形成一套适应性强的的高效优化方法。以复合材料翼身融合体帽型加筋板的质量优化为实例,建立加筋板模型的重量响应面目标函数、强度和翘曲稳定性响应面约束条件;通过PATRAN/NASTRAN有限元软件进行有限元计算,获取用于响应面训练的样本点数值。算例结果表明,该方法能以很少的有限元分析次数取得高精度的响应面近似模型,并且使优化计算耗时大为减少,优化效率大大提高。     

神经网络与遗传算法在拉延筋参数反求中的应用

以某车型前地板角支撑板的拉延工序为例,讨论BP神经网络技术与遗传算法在拉延筋几何参数反求中的综合应用问题,建立能描述反映成形效果的三个参数与半圆形拉延筋几何参数之间非线性映射关系的神经网络模型,并运用遗传算法对神经网络结构进行了优化。提出逐次局部密化样本点的样本点设计方法。该方法有助于加快神经网络的设计进程,提高神经网络的模拟精度。当训练样本数据可通过有限元法自动获得时,使用该方案则更为便利。     

进化小波网络及其在设备状态预测中的应用

结合小波网络和进化计算提出了进化小波网络策略,该策略采用控制级基因和参数级基因分别对网络结构和网络参数进行编码,并将遗传算法与进化规划结合进行进化操作,实现同时对网络结构与网络参数进行进化设计和学习训练。该策略不仅克服了网络训练中的局部极小和不收敛问题,也使网络结构更优,从而提高了网络训练和工作性能。最后分别就函数逼近问题、太阳黑子数预测问题及水轮机组的状态预测问题进行了事例研究,验证了所提出的进化小波网络策略的优越性能和可行性。     

基于遗传神经网络的机器人腕力传感器动态建模与补偿方法

介绍用于MotomamV3X机器人上的新型多维腕力传感器,比较遗传算法与人工神经网络的特点,将遗传算法的交叉和变异操作进行改进,提出一种融合改进遗传算法(Genetic algorithm, GA)的函数连接型人工神经网络(Functional link artificial neural network, FLANN),并将其用于所介绍的新型机器人腕力传感器动态建模与动态性能补偿中。介绍动态建模与动态补偿原理及改进遗传神经网络算法,给出该传感器的动态模型和动态补偿模型。该方法利用腕力传感器的动态标定数据,采用改进遗传神经网络搜索和优化模型参数,保留了遗传算法的全局搜索能力和FLANN结构简单,鲁棒性好,且具备自学习能力的特点,克服了FLANN容易陷入局部极小的缺陷,具有快的网络训练速度及高的动态建模精度。理论分析和试验结果都证实了所提出的动态建模与动态补偿方法的有效性。