结构紧固连接件损伤的神经网络诊断研究

在结构紧固连接件损伤诊断中采用了函数连接型神经网络。与常用的ＢＰ网络不同，函数连接网络通过对输入模式的增强来实现复杂模式的分类。它借助某种函数变换，将原始的模式特征向量变换成另一个更高维的形式，使原始空间里不能区分的类别，在这个高维空间中得以区分，减少了神经网络的层数，提高了网络的学习速度。诊断时选用正弦激振下结构中多个位置上传感器稳态响应的幅值构成表征损伤的特征向量，简化了特征提取的过程。数值模拟和实验结果表明，这一方法能对多个位置的损伤进行准确的识别     

分区式最小二乘法模式识别研究

提供了一种改进的关于智能板结构受载位置的模式识别方法．在设置有限个应变传感器和进行有限点加载的实验基础上，通过分区式最小二乘法，可对板上任意点受载的位置进行识别     

面向星敏感器的星模式识别算法

介绍了到目前为止出现的所有面向星敏感器的星模式识别算法,它们是概率统计算法、三角形算法、匹配组算法、网格算法、奇异值分解算法、神经网络算法和遗传算法,并分两组比较了它们的性能,给出了具有指导意义的结论。     

神经网络时滞系统非共振双Hopf分岔及其广义同步

本文建立了具有自连接和抑制-兴奋型他连接的两个同性神经元模型。其中自连接是由于兴奋型的突触产生,而他连接则分别对应于两神经元兴奋、抑制型的突触。发现如果有兴奋型自连接就会有双Hopf分岔,而没有时滞自连接时双Hopf分岔就会消失,因此自连接引起了双Hopf分岔。作为一个例子,通过变动连接中的时滞和他连接中的比重,1／√2双Hopf分岔得到了详细研究。通过中心流形约化,分岔点邻域内各种不同的动力学行为得到了分类,并以解析形式表出。神经元活动的分岔路径得以表明。从得到的解析近似解可以发现,本文所研究的具有兴奋一抑制型他连接的两相同神经元的节律不能完全同步而只能广义同步。时滞也可以使其节律消失,两神经元变为非活动的。这些结果在控制神经网络关联记忆和设计人工神经网络方面有着潜在的应用。     

框架结构简图的模式识别

提出了基于工程结构简图语义知识的计算机结构模式识别方法。根据框架结构特征,融快速模板匹配、语义识别和树文法遍历方法为一体,对结构计算简图几何拓扑形式和荷载环境进行识别,进一步生成应用软件所需的数据文件,从而屏蔽了应用软件原始数据文件的具体格式。     

Engine Knock Detection from Vibration Signals Using Pattern Recognition

The paper deals with a diagnostic method that allows to detect engineknock. The developed algorithm differentiates three kinds of engine cycles:absence of knock, increasing knock and heavy knock. The decision is takenfrom a block vibration signal. The diagnostic method is based on patternrecognition. Three models of different data shapes provided from theaccelerometer are elaborated. This is done using a time-scale analysis toolcalled a wavelet network. It allows to extract relevant features from thesignal. The aim of the method is then to partition the feature space intoclasses representing the knock states. Experimental results are reported.     

基于粗集—模糊理论的柴油机磨损模式识别

在总结柴油机磨损模式的磨粒分析基础上,首次提出在柴油机磨损模式识别中采用基于粗集和模糊理论相结合的模式识别方法。在建立磨损形式标准样本库的基础上,利用粗集理论对标准样本知识库进行简化和权重集计算,然后利用模糊理论确定隶属函数,最后给出结论。     

Load Extraction from Photoelastic Images Using Neural Networks

Photoelastic materials develop colored fringes under white light when subjected to mechanical stresses, which can be viewed through a polariscope. This technique has traditionally been used for stress analysis of loaded components, however, this can also be potentially used in sensing applications where the requirement may be measurement of the stimulating forces causing the generation of fringes. This leads to inverse photoelastic problem where the developed image can be analyzed for the input forces. However, there could be infinite number of possible solutions which cannot be determined by conventional techniques. This paper presents neural networks based approach to solve this problem. Experiments conducted to prove the principle have been verified with theoretical results and finite element analysis of loaded specimens. The developed technique, if generalized, can be implemented for whole-field analysis of the stress patterns involving complex fringes under different loading conditions. This can also provide direct visualization of the stress field, which may find application in a variety of specialized areas including biomedical engineering and robotics.

卷积神经网络在装备磨损颗粒识别中的研究综述

铁谱法是用于装备故障诊断的1种重要方法,其中铁谱法的重点是铁谱图像的分析,即磨损磨粒分析. 卷积神经网络是当下最流行的深度学习算法之一,其广泛应用于图像识别领域,使得图像识别领域得到突破. 随着卷积神经网络的快速发展,磨损颗粒在智能识别方面的技术取得了重大的突破. 本文中首先简述了卷积神经网络与磨粒智能识别的发展历史,针对基于卷积神经网络的磨粒识别方法进行了从图像数据集处理到模型优化技术方面的介绍,并详细说明了这些技术在磨粒识别中的具体应用实例. 然后从现有网络和自设计网络两方面分类,整理了近年来卷积神经网络应用于磨粒智能识别的代表性文献,综述了这些工作所提出的模型结构和特点,分析并阐述了各个模型主要的识别原理,各个网络结构存在的优缺点,以及它们的数据采用情况等,并对未来磨粒智能识别的主要研究方向进行了展望. 最后肯定了卷积神经网络方法在磨粒智能识别方面的重要性,同时指出了基于此方法的磨粒识别模型的缺点,并提出了应紧跟图像识别领域的最新技术以促进磨粒智能识别水平提高等建议,对磨粒智能识别的发展具有一定的意义.      

Neural networks for computing in the elastoplastic analysis of structures

A neural network model is proposed and studied for the treatment of elastoplastic analysis problems. These problems are formulated as Q.P.P.s with inequality subsidiary conditions. In order to treat these conditions the Hopfield model is appropriately generalized and a neural model is proposed covering the case of inequalities. Finally, the parameter identification problem is formulated as a supervised learning problem. Numerical applications close the presentation of the theory and the advantages of the neural network approach are illustrated.

---

Sommario Si propone un modello di rete neurale con l'obiettivo di usarlo per la trattazione di problemi di analisi elastoplastica, formulati come problemi di programmazione quadratica con disequazioni sussidiarie. Allo scopo di trattare queste condizioni si generalizza il modello di Hopfield e si propone un modello neurale che copre il caso di disequazioni. Inoltre il problema di identificazione parametrica viene formulato come un problema di apprendimento guidato. La presentazione della teoria è seguita da esempi di applicazioni numeriche e dalla illustrazione dei vantaggi dell'uso delle reti neurali.

     

Single-Pulse Solutions for Oscillatory Coupling Functions in Neural Networks

We study the existence and linear stability of stationary pulse solutions of an integro-differential equation modeling the coarse-grained averaged activity of a single layer of interconnected neurons. The neuronal connections considered feature lateral oscillations with an exponential rate of decay and variable period. We identify regions in the parameter space where solutions exhibit areas of excitation with single- and dimpled-pulses. When the gain function reduces to the Heaviside function, we establish existence of single-pulse solutions analytically. For a more general gain function, we include numerical support of the existence of pulse-like solutions. We then study the linear stability behavior of these solutions.     

基于模式运算的工程设计方法

研究了复杂的工程设计决策过程,讨论了设计问题的五个性质,提出了两类不同速度的设计过程:以模式识别为基础的快速设计过程和以模式加法、模式联想为基础的慢速设计过程。此外,还讨论了模蝴集合论和人工神经网络在实现模式运算中的应用。     

基于径向基函数神经网络的磨粒识别系统

应用磨粒形状特征参数、颜色特征参数和表面纹理特征参数对磨粒形态进行量化表征，并以此为输入矢量，引入径向基函数神经网络对磨损微粒进行自动分类识别，建立了适用于磨粒识别的径向基函数神经网络模型，并给出了具体算法．应用实例表明，径向基函数神经网络的收敛速度和识别率优于传统的BP神经网络．     

神经网络在工程爆破应力波规律探讨中的应用

采用BP（Back Propagation)前馈人工神经网络模型，对工程爆破中柱状震源的自由场应力波传播规律进行了探讨。结果表明：利用人神经网络模型的非线性映射功能，可以较好地给出工程爆破引起的近区自由场力学规律，对于同类型问题的研究，也有着很重要的意义。     

Engine Control using Neural Networks: A New Method in Engine Management Systems

To be able to meet the demands of low emissions and fuel consumption ofmodern combustion engines, new ways have to be found to control thecombustion. We use new sensors to measure the pressure in the combustionchamber and analyze this signal with a neural network in order to receiveseveral form factors which can be used to control the ignition timing. Theneural network is trained off line with measured data and used on line toderive the form factors. The proposed algorithm can be computed in real timeon conventional digital signal processors and adapted to new engines withvery little effort.     

Synchronization in the Lorenz system: Stability and robustness

We present necessary and sufficient conditions for the existence of synchronization in a class of continuous-time nonlinear systems: the so-called -affine systems. We apply the results to the Lorenz attractor. The robustness of the synchronization against parameter value variations is discussed using the Lyapunov stability theory for perturbed systems. We obtain sufficient conditions that guarantee a bounded steady-state error. This technique gives conservative results; however, in some systems like that of Lorenz, it provides definitive results about the existence of the synchronization. Furthermore, we give estimates of the maximal error as a function of the difference between the parameter values of the systems to be synchronized.     

Nondestructive Evaluation of Crack Depth in Concrete Using PCA-compressed Wave Transmission Function and Neural Networks

Cracks in concrete are common defects that may enable rapid deterioration and failure of structures. Determination of a crack’s depth using surface wave transmission measurement and the cut-off frequency in the transmission function (TRF) is difficult, in part due to variability of the measurement data. In this study, use of complete TRF data as features for crack depth assessment is proposed. A principal component analysis (PCA) is employed to generate a basis for the measured TRFs for various simulated crack (notch) cases in concrete. The measured TRFs are represented by their projections onto the most significant PCs. Then neural networks (NN), using the PCA-compressed TRFs, are applied to estimate the crack depth. An experimental study is carried out for five different artificial crack (notch) cases to investigate the effectiveness of the proposed method. Results reveal that the proposed method can effectively estimate the artificial crack depth in concrete structures, even with incomplete NN training.     

A Turbulent-Wake Estimation Using Radial Basis Function Neural Networks

An estimation of the fluctuations in the passive-tracer concentration for the turbulent wake behind an airfoil is presented. The estimation is based on experimental modelling using Radial Basis Function Neural Networks. For the experiment the fluctuations of the concentration in the turbulent wake were recorded with a visualization method. The records of the concentration in the selected regions of the turbulent wake were used as the input and output regions for the training and estimation with neural networks. The uncertainty of the estimation increased with increasing distance between the input and the output regions. The power spectra, the spatial correlation functions and the profiles of the concentration were calculated from the measured and estimated fluctuations of the concentration. The measured and estimated concentration power spectra were in reasonable agreement. The measured and estimated spatial correlation functions and the profiles of the concentration showed a similar agreement.     

基于神经网络的结构变形估计和形状控制

准确的变形估计是智能结构形状控制的前提。本文基于人工神经网络（ＡＮＮ）方法设计了智能桁架结构的变形估计器和形状控制器，根据结构系统有限数目的测量值可以估计结构变形并用于形状控制。该方法同时适于处理结构非线性问题。算例表明方法的可行性与有效性。     

Performance comparison of artificial neural networks and expert systems applied to flow pattern identification in vertical ascendant gas–liquid flows

Instantaneous readouts of an electrical resistivity probe are taken in an upward vertical air–water mixture. The signals are further processed to render the statistical moments and the probability density functions here used as objective flow pattern indicators. A series of 73 experimental runs have its flow pattern identified by visual inspection assisted by the analyses of the void fraction’s trace and associated probability density function. The flow patterns are classified into six groups and labeled as: bubbly, spherical cap, slug, unstable slug, semi-annular and annular. This work compares and analyzes the performance of artificial neural networks, ANN, and expert systems to flow pattern identification. The employed ANNs are Multiple Layer Perceptrons, Radial Basis Functions and Probabilistic Neural Network, with single and multiple outputs. The performance is gauged by the percentage of right identifications based on experimental observation. The analysis is extended to clustering algorithms to assist the formation of knowledge base employed during the learning stages of the ANNs and expert systems. The performance of the following clustering algorithms: self organized maps, K-means and Fuzzy C-means are also tested against experimental data.