神经网络在参数估计中的应用

将神经网络用于线性连续不变系统（包括一种带时延的线性系统）参数估计中。线性系统分别用状态方程及传递函数来表示,给出了相应的神经网络结构及学习算法。     

Analysing non-linear systems in the frequency domain--I. The transfer function

Both continuous and discrete time transfer functions of non-linear systems are analysed and interpreted in the frequency domain by investigating the properties and graphical representation of these functions. The contributions that some typical terms from non-linear time domain models make to the transfer functions is illustrated to provide a better understanding of the frequency response behaviour of complex non-linear dynamic systems.     

NON-LINEAR SYSTEM IDENTIFICATION USING LUMPED PARAMETER MODELS WITH EMBEDDED FEEDFORWARD NEURAL NETWORKS

This paper describes a new methodology to identify multi-degree-of-freedom non-linear systems from the system's operating data. The methodology includes a new non-linear model architecture which embeds feedforward neural networks to represent unknown non-linearities in a lumped parameter model, and a learning algorithm to train the embedded neural networks as well as the other model parameters to obtain model fidelity. Three simulated and experimental examples are used to validate the proposed methodology.     

压电振动系统的神经网络建模和预测研究

本文应用MATALB／XPC实时仿真工具测量了贴有压电元件的复合材料薄壁结构的振动响应。并对其进行神经网络的离线建模和预测。比较了几种网络的优缺点。选择了引进外部反馈的前向BP网络作为非线性系统建模的方法，有望推广用于智能结构的健康监测和振动主动控制。     

基于遗传算法和神经网络的塔机结构动态优化设计

利用遗传算法和BP神经网络建立复杂结构系统动态优化的计算模型,该模型可代替系统原来的有限元模型,用于振动系统的快速重分析。首先对塔式起重机结构系统进行模态分析及谐响应动力学分析,找出对结构动态特性影响最大的模态频率,再利用灵敏度分析,确定对动态特性较敏感的设计变量作为神经网络的输入变量,并利用正交试验法确定神经网络训练样本,用有限元模型计算出样本点数据,建立反映结构振动特性的人工神经网络模型,最后利用遗传算法对所建立的神经网络模型寻优,得到使结构动态性能最优的设计参数。     

基于改进变异蚁群算法神经网络的空间大型可展开天线动态响应辨识

针对空间大型可展开天线柔性大、展开过程中弹性变形与刚体运动相互耦合、机构运动参数时变的特点,提出了基于改进变异蚁群算法神经网络的辨识模型用于可展开天线动态响应辨识的方法。该方法采用改进变异蚁群算法优化神经网络权值,将变异机制引入蚁群算法,解决了蚁群算法收敛慢的缺点,对变异蚁群算法进行改进,提高了算法跳出局部最优的能力,进一步加快了收敛速度。仿真结果表明,该辨识模型兼具神经网络和蚁群算法的优点,不仅具有优异的非线性逼近能力,还具有高的运算效率。该辨识模型能够准确地辨识天线的动态响应,辨识的收敛速度快且精度高。     

Neuro-fuzzy control strategy for an offshore steel jacket platform subjected to wave-induced forces using magnetorheological dampers

Magnetorheological (MR) damper is a prominent semi-active control device to vibrate mitigation of structures. Due to the inherent non-linear nature of MR damper, an intelligent non-linear neuro-fuzzy control strategy is designed to control wave-induced vibration of an offshore steel jacket platform equipped with MR dampers. In the proposed control system, a dynamic-feedback neural network is adapted to model non-linear dynamic system, and the fuzzy logic controller is used to determine the control forces of MR dampers. By use of two feedforward neural networks required voltages and actual MR damper forces are obtained, in which the first neural network and the second one acts as the inverse dynamics model, and the forward dynamics model of the MR dampers, respectively. The most important characteristic of the proposed intelligent control strategy is its inherent robustness and its ability to handle the non-linear behavior of the system. Besides, no mathematical model needed to calculate forces produced by MR dampers. According to linearized Morison equation, wave-induced forces are determined. The performance of the proposed neuro-fuzzy control system is compared with that of a traditional semi-active control strategy, i.e., clipped optimal control system with LQG-target controller, through computer simulations, while the uncontrolled system response is used as the baseline. It is demonstrated that the design of proposed control system framework is more effective than that of the clipped optimal control scheme with LQG-target controller to reduce the vibration of offshore structure. Furthermore, the control strategy is very important for semi-active control.     

System identification using associated linear equations

In this paper, analytical models of non-linear systems are obtained by identifying the frequency response functions (FRFs) of their associated linear equations (ALEs). This allows the use of several methods of identification in the frequency domain usually applicable to linear systems. Among other advantages, the cumbersome multidimensional Fourier Transformation required in higher-order frequency response functions (HFRFs) analysis is eliminated. Two theoretical systems are used here as examples, an electrostrictive actuator and a Duffing oscillator. The concept of the non-linear gain constant arises as a simple means of identification.     

A neural network model based predictive control approach: application to a semi-batch reactor

Neural networks can be considered to be new modelling tools in process control and especially in non-linear dynamical systems cases. Their ability to approximate non-linear functions has been very often demonstrated and tested by simulation and experimental studies. In this paper, a predictive control strategy of a semi-batch reactor based on neural network models is proposed. Results of a non-linear control of the reactant temperature of a semi-batch reactor are presented. The process identification is composed of an off-line phase that consists in training the network, and of an on-line phase that corresponds to the neural model adaptation so that it fits any modification of the process dynamics. Experimental results when using this method to control a semi-batch reactor are reported and show the great potential of this strategy in controlling non-linear processes.     

基于神经网络的壳体结构损伤诊断研究

神经网络输入参数的选择将直接影响工程结构损伤识别的精度和准确性。本文提出以反映结构损伤位置和程度的固有频率与频率下降率的组合作为神经网络输入的特征参数，以增加对损务程度敏感的参数项，克服单独使用某种参数的缺陷。针对使用BP算法的多层感知器中存在的网络收敛速度慢，容易陷入局部极小点等问题，采用一个改进算法。并以门座起重机筒形支柱－－圆柱壳结构损伤为例，进行计算分析，从中可以看出，采用此组合特征参数和改进算法提高了诊断的精度，加快了网络收敛的程度。     

Sensitivity analysis of side slip angle for a front wheel steering vehicle: a frequency domain approach

In this paper, sensitivity analysis of side slip angle for a front wheel steering vehicle is performed in the frequency domain. For the derivation of the transfer function, a simple vehicle model with two degrees of freedom is used in the initial modeling stage. This model exhibits the simplest lateral dynamic effect, and is useful for understanding the dynamic characteristics and control aspects of the target system. Vehicle mass, inertia, cornering stiffness, and wheel base are taken to be the design variables. Sensitivity functions of the transfer function with respect to the design variables are derived. From this study, we see that a transition speed exists in the frequency response of side slip angle. This implies that the characteristics are changed from minimum phase to non-minimum phase as the vehicle speed increases. The objective of this study is to propose a basis for design and re-design of the vehicle by checking the side slip angle variations with respect to design variable changes in the frequency domain. Finally, dominant design variables are suggested based on the sensitivity analysis.     

Fast reduced model of non-linear dynamic Euler-Bernoulli beam behaviour

This paper introduces a new way to obtain the dynamical response of Euler-Bernoulli beams under large deflections. They are simulated with parsimonious models based on a two integrated neural networks. Our article proves that neural network based-reduced modelling allows mechanical simulations to benefit shorter time processing without a great loss of accuracy. Our model reduction technique ensures a good error-speed compromise. It rapidly leads to algebraic generic model matching interactive simulation or high-realistic multi-sensorial virtual protoyping requirements. The originality of our solution consists in the linking of two neural networks. They are integrated in a conditional loop. First, the article presents a systematic numerical technique that enables users to efficiently build an optimal neural network. This approach is based on the handling of discrete optimization evolutionary technique. Our optimization process is applied to the construction of the two neural networks. From an exhaustive learning base created from non-linear finite element analysis series, we secondly clearly describe our two neural network-based reduced models. The first neural network that enables nodal displacements to be determined from boundary conditions is linked to the second one that aims to set boundary conditions from a deformation state. The new relationship is finally analysed for justifying the successes of our proposal. The article completely details a new numerical process supervising the making of an integrated neural network loop being a new from of model for beam non-linear behaviour representation.     

一种基于脑网络的分层模块化回声状态网络

针对回声状态网络(Echo state network, ESN)结构设计问题,提出一种基于脑网络的分层模块化回声状态网络(Hierarchical modular echo state network, HMESN)。脑网络的拓扑结构使功能网络具有丰富的动力学特性,因此,从生物仿生学角度出发,对HMESN的储备池进行分层设计,各层级上的神经元采用小世界网络构建算法生成模块化结构,并引入层级连接。基于脑网络分层模块化的拓扑特征弱化了神经元间的耦合程度,从而使神经元的动力学特性更为丰富,在功能与结构上更接近于真实生物神经网络,有效地提高了网络处理问题的能力。采用Mackey-Glass时间序列预测和非线性系统辨识对网络进行验证,证明该网络的有效性和可行性。     

基于神经网络PID的冗余伺服系统自适应控制

建立冗余直接驱动式电液伺服系统的数学模型。针对电液伺服系统时变、强非线性的特点以及冗余伺服系统在余度降级过程中的故障瞬态现象和余度降级后的性能降级现象,考虑传统PID控制器自适应能力不强、鲁棒性差等缺陷,提出神经网络自适应控制方案。根据冗余电液伺服系统的特点和目前神经网络控制的发展水平,采用基于径向基函数神经网络的智能PID控制器实现冗余伺服系统的自适应控制。研究结果表明：该控制器能够根据控制指令、被控对象结构参数等因素的变化实时调整控制器参数,和传统PID控制器相比具有控制精度高、鲁棒性强的特点,可以有效地克服冗余伺服系统余度切换时的故障瞬态现象和余度降级后的性能降低现象。     

一种残差注意力迁移学习方法及其在滚动轴承故障诊断中的应用

提出了一种基于残差注意力卷积神经网络(CSRA-CNN)的迁移学习算法,用于提高滚动轴承的故障诊断精度。在卷积神经网络模型中加入残差注意力机制,使模型在训练过程中更加注重故障特征的提取,从而有效提高迁移准确率。为了测评基于残差注意力卷积神经网络的性能,将其与传统卷积神经网络在不同迁移学习策略下的结果进行对比。用动力传动故障诊断综合实验台和高速列车综合实验台对所提算法进行了验证,该方法可以完成变转速以及变转速变载荷下轴承不同健康状态的迁移学习,且迁移效果均优于传统的卷积神经网络。     

RBF神经网络用于家用空调系统匹配特性的研究

RBF神经网络是目前应用较多的一种神经网络。它能以任意精度逼近任意非线性函数,具有良好的逼近性能,并且结构简单,是一种性能优良的神经网络。因此,将RBF神经网络应用于家用空调匹配仿真研究时具有独特的优势。提出采用RBF神经网络估算制冷量和压力来优化研发过程,仿真结果表明,RBF神经网络运用于家用空调匹配仿真,能够精确仿真空调制冷量和低压力等参数,并预测制冷量和压力,能有效地减少家用空调匹配时间,提高研究效率。     

基于神经网络的机翼振动主动控制

根据DARMA模型提出了简单易用的神经网络控制方案,该方法采用线性人工神经网络对系统动态特性进行在线辨识,并利用辨识得到的信息,采用BP神经网络对系统进行控制,将该算法应用于飞机机翼振动主动控制数值仿真。仿真结果表明,该方法能减少算法的计算量,压缩计算时间,便于提高系统采样频率,使得自由振动和调频振动的抑制成为可能。     

Practical applications of artificial neural networks in chemical process development

Artificial neural networks (ANN) have the ability to map non-linear relationships without a-priori information about process or system models. This significant feature allows the network to “learn” the behavior of a system by example when it may be difficult or impractical to complete a rigorous mathematical solution. Recently ANN technology has been leaving the academic arena and placed in user-friendly software packages. This paper will offer an introduction to artificial neural networks and present a case history of two problems in chemical process development that were approached with ANN. Both optimal PID control tuning parameters and product particle size predictions were constructed from process information using neural networks. The ANN provides a rapid solution to many applications with little physical insight into the underlying system function. The amount of data preparation and performance limitations using a neural network will be discussed. However, the properly applied ANN will generally provide insight to which variables are most influential to the model and evolve dynamically to the minimum performance surface squared error. Neural networks have been used successfully with non-linear dynamic systems and can be applied to chemical process development for system identification and multivariate optimization problems.     

应用迭代补偿方法的台车碰撞仿真研究

采用频域迭代的方法逐步修正驱动信号,实现对输出信号在幅值和相位上的补偿,使系统的输出逼近期望响应信号。驱动信号迭代算法是迭代补偿算法的核心内容,为了消除非线性的影响,给出了详细的驱动信号频域迭代算法,并运用到模型中,验证了此算法的可行性和有效性。     

On-line fan blade damage detection using neural networks

This paper presents a methodology for monitoring the on-line condition of axial-flow fan blades with the use of neural networks. In developing this methodology, the first stage was to utilise neural networks trained on features extracted from on-line blade vibration signals measured on an experimental test structure. Results from a stationary experimental modal analysis of the structure were used for identifying global blade mode shapes and their corresponding frequencies. These in turn were used to assist in identifying vibration-related features suitable for neural network training. The features were extracted from on-line blade vibration and strain signals which were measured using a number of sensors.The second stage in the development of the methodology entails utilising neural networks trained on numerical Frequency Response Function (FRF) features obtained from a Finite Element Model (FEM) of the test structure. Frequency domain features obtained from on-line experimental measurements were used to normalise the numerical FRF features prior to neural network training. Following training, the networks were tested using experimental frequency domain features. This approach makes it unnecessary to damage the structure in order to train the neural networks.The paper shows that it is possible to classify damage for several fan blades by using neural networks with on-line vibration measurements from sensors not necessarily installed on the damaged blades themselves. The significance of this is that it proves the possibility to perform on-line fan blade damage classification using less than one sensor per blade. Even more significant is the demonstration that an on-line damage detection system for a fan can be developed without having to damage the actual structure.