【作者】 章文俊；

【导师】 刘正江；

【作者基本信息】 大连海事大学 ， 交通信息工程及控制， 2014， 博士

【摘要】 小波神经网络结合了神经网络的自学习、自适应、鲁棒性、容错性和泛化推广能力以及小波变换的时频局部和变焦等特性,具有全局最优逼近和运算速度快等优点,避免了BP网络等传统的神经网络类型在这些方面的不足,已成功应用于系统辨识、模式识别和控制等领域。在实际应用中,小波神经网络尚存在对系统动态反映能力不足以及泛化能力难以保证等问题,制约了其在实际工程中的应用。为了提高小波神经网络的泛化能力,提出基于Akaike信息准则改进的余值选择算法。通过设定最优学习停止标准,在保证辨识精度的同时精简网络的规模,避免了神经网络在学习过程中出现的过拟和与欠拟合现象,提高了神经网络的泛化能力。作为小波神经网络的构造算法,余值选择算法通过正交选择方法高效地衡量了隐层节点对输出的贡献,有利于网络规模的自适应调整。仿真实验表明该算法提高了网络的泛化能力。为了更好地反映系统动态的变化,将系统历史信息引入网络输入层构造时滞小波神经网络模型,为弥补由此带来的输入变量膨胀的缺陷,利用基于相对贡献率的灵敏度分析方法确定与系统输出相关性强的变量作为输入,优化了输入层结构,解决了网络模型失配的问题,提高了网络对系统动态变化的反映能力。针对船舶海上运动非线性、大惯性和动态时变等特点,构建基于改进时滞小波神经网络的预测PID控制器。其中,利用小波神经网络进行船舶运动动态的在线辨识和预测,利用预测控制策略克服船舶运动的大惯性对控制效果的不利影响。基于该控制器进行了船舶航向跟踪控制的仿真实验,并与传统的PID控制器进行了对比研究,结果表明该控制系统具有较高的控制精度和较强的抗干扰能力。以上研究结果显示,改进的小波神经网络有针对性地提高了系统的动态反映能力和泛化能力,其运算快速性和非线性拟合能力适应船舶海上运动特点,在船舶运动控制领域有广阔的应用前景。更多还原

【Abstract】 Wavelet neural network combines the virtues of neural network such as the self-learning ability, adaptivity, robustness, fault-tolerance ability as well as the time-frequency local characteristics and zooming characteristics of wavelet transform, possesses virturs of global optimum approximating ability and fast processing speed. It avoids drawbacks of conventional BP neural network such as local minima and slow convergence speed, has been successfully applied in areas of system identification, pattern recognition and control. However, it is found in practical applications that there are several problems such as deficiencies in dynamic representing ability and generalization ability, which frustrates its practical applications.To improve the generalization ability of wavelet neural network, an improved residual selection learning algorithm is proposed based on Akaike information criterion. By setting optimal stop criterion for learning process, the achieved wavelet neural network enables satisfying identification accuracy as well as compact network structure, which avoids unfavorable phenomenon of over-fitting and under-fitting and guarantees the generalization ability of wavelet neural network. As a learning algorithm for wavelet neural network, residual selection algorithm evaluates the contribution of hidden neurons to the output respectively, which facilitate the adaptive adjustment of network dimension. Simulation results shows that the proposed algorithm improves the generalization capability of the wavelet neural network.To better represent the changes of system dynamics, the history information of system is introduced in the network input layer leading to the time-delay wavelet neural network. To settle the subsequent problem of too much variables in input layer, sensitivity analysis method based on relative contribution ratio is applied to decide the optimal inputs by selecting variables which have higher correlation with output, which resolves the problem of model mismatch and improve the network’s representing ability for changes of dynamical system. Aiming at the complex characteristics of ship motion such as nonlinearity, large inertia and time-varying dynamics, predictive PID controller is constructed based on the improved time-delay wavelet neural network. The wavelet neural network is performed for the online ship dynamic identification and prediction. The predictive control strategy is utilized to overcome the unfavorable effects of large inerita. Simulations of ship heading course following control were conducted and coparison study was processed with the conventional PID controller, the results shows that the proposed controller possesses higher control accuracy as well as stronger anti-interference ability.The aforesaid study results demonstrate that the improved wavelet neural network enhances the dynamic representing ability and generalization ability of wavelet neural network respectively. Its fast processing speed and nonlinear approximation ability enable that it can represent the characteristics of ship motion at sea, and can be implemented widely in field of ship motion control.更多还原