An Improved Artificial Neural Network Model for Predicting Silicon Content of Blast Furnace Hot Metal

Based on the skills of initializing weight distribution, adding an impulse in a neural network and expanding the ideal of plural weights, an artificial neural network model with three connection weights between one and another neural unit was established to predict silicon content of blast furnace hot metal. After the neural network was trained in the off-line state on the basis of a large number of practical data of a commercial blast furnace and making many learning patterns, satisfactory testing and simulating results of the model were obtained.     

An Interval-valued Fuzzy Competitive Neural Network

Because interval value is quite natural in clustering, an interval-valued fuzzy competitive neural network is proposed. Firstly, this paper proposes several definitions of distance relating to interval number. And then, it indicates the method of preprocessing input data, the structure of the network and the learning algorithm of the interval-valued fuzzy competitive neural network. This paper also analyses the principle of the learning algorithm. At last, an experiment is used to test the validity of the network.     

An ERP study on shift of spatial attention resulting from number processing

Using the “stimulus-delay-target” experimental model, the brain mechanism of visual-spatial attention shift after number processing was conducted on 13 participants by using event-related potentials (ERP) technology. Results show that influence of the number processing on the spatial attention is indexed by the P3 component; parietal lobes at both sides may be the main neural substrate for number processing and spatial attention, and the numerical-spatial interaction mainly occurs in the response selection stage.     

An ERP study on shift of spatial attention resulting from number processing

Using the “stimulus-delay-target” experimental model, the brain mechanism of visual-spatial attention shift after number processing was conducted on 13 participants by using event-related potentials (ERP) technology. Results show that influence of the number processing on the spatial attention is indexed by the P3 component; parietal lobes at both sides may be the main neural substrate for number processing and spatial attention, and the numerical-spatial interaction mainly occurs in the response selection stage.     

Colorant Formulation Using 3 - layer Artificial Neural Networks

The colorant formulation using artificial neural networks (ANN) was investigated in this study. A simple 3 -layer, input - hidden - output system was constructed for the recipe formulation of one - , two - , and three -dye mixtures. Comprehensive tests were carried out to explore the properties of a 3 - layer simple ANN systematically . These properties include number of neurons in the hidden layer, learning rate of the network, momentum factor of the network, as well as the number of epochs for the learning process. The tests show accurate results for one - and two - dye mixtures while less accurate but comparable results to conventional colorant formulation systems for three - dye mixtures. It is also found that the optimum values of the neural network parameters are important towards the accuracy of the colorant formulation.     

Identification model of multi-layered neural network parameters and its applications in the petroleum production

This paper creates a LM （Levenberg-Marquardt） algorithm model which is appropriate to solve the problem about weights value of feedforward neural network. On the base of this model, we provide two applications in the oilfield production. Firstly, we simulated the functional relationships between the petrophysical and electrical properties of the rock by neural networks model, and studied oil saturation. Under the precision of data is confirmed, this method can reduce the number of experiments. Secondly, we simulated the relationships between investment and income by the neural networks model, and studied invest saturation point and income growth rate. It is very significant to guide the investment decision. The research result shows that the model is suitable for the modeling and identification of nonlinear systems due to the great fit characteristic of neural network and very fast convergence speed of LM algorithm.     

Techniques of Image Processing Based on Artificial Neural Networks

This paper presented an online quality inspection system based on artificial neural networks. Chromatism classification and edge detection are two difficult problems in glass steel surface quality inspection. Two artificial neural networks were made and the two problems were solved. The one solved chromatism classification. Hue, saturation and their probability of three colors, whose appearing probabilities were maximum in color histogram, were selected as input parameters, and the number of output node could be adjusted with the change of requirement. The other solved edge detection. In this neutral network, edge detection of gray scale image was able to be tested with trained neural networks for a binary image. It prevent the difficulty that the number of needed training samples was too large if gray scale images were directly regarded as training samples. This system is able to be applied to not only glass steel fault inspection but also other product online quality inspection and classification.     

Hierarchical Neural Networks Method for Fault Diagnosis of Large-Scale Analog Circuits

A novel hierarchical neural networks (HNNs) method for fault diagnosis of large-scale circuits is proposed. The presented techniques using neural networks(NNs) approaches require a large amount of computation for simulating various faulty component possibilities. For large scale circuits, the number of possible faults, and hence the simulations, grow rapidly and become tedious and sometimes even impractical. Some NNs are distributed to the torn sub-blocks according to the proposed torn principles of large scale circuits. And the NNs are trained in batches by different patterns in the light of the presented rules of various patterns when the DC, AC and transient responses of the circuit are available. The method is characterized by decreasing the over-lapped feasible domains of responses of circuits with tolerance and leads to better performance and higher correct classification. The methodology is illustrated by means of diagnosis examples.     

Gender Recognition with Face Images Based on Partially Connected Neural Evolutionary

In this paper,a new type of neural network model - Partially Connected Neural Evolutionary (PARCONE) was introduced to recognize a face gender. The neural network has a mesh structure in which each neuron didn't connect to all other neurons but maintain a fixed number of connections with other neurons. In training,the evolutionary computation method was used to improve the neural network performance by change the connection neurons and its connection weights. With this new model,no feature extraction is needed and all of the pixels of a sample image can be used as the inputs of the neural network. The gender recognition experiment was made on 490 face images (245 females and 245 males from Color FERET database),which include not only frontal faces but also the faces rotated from-40°-40° in the direction of horizontal. After 300-600 generations' evolution,the gender recognition rate,rejection rate and error rate of the positive examples respectively are 96.2%,1.1%,and 2.7%. Furthermore,a large-scale GPU parallel computing method was used to accelerate neural network training. The experimental results show that the new neural model has a better pattern recognition ability and may be applied to many other pattern recognitions which need a large amount of input information.     

Neural network approach for modification and fitting of digitized data in reverse engineering

Reverse engineering in the manufacturing field is a process in which the digitized data are obtained from an existing object model or a part of it, and then the CAD model is reconstructed. This paper presents an RBF neural network approach to modify and fit the digitized data. The centers for the RBF are selected by using the orthogonal least squares learning algorithm. A mathematically known surface is used for generating a number of samples for training the networks. The trained networks then generated a number of new points which were compared with the calculating points from the equations. Moreover, a series of practice digitizing curves are used to test the approach. The results showed that this approach is effective in modifying and fitting digitized data and generating data points to reconstruct the surface model.     

人工神经网络混合剪枝算法

目前人工神经网络(ANN)应用中所遇到的挑战之一就是如何针对特定问题确定相应网络。基于进化算法和局部搜索算法两类策略的特点和不足,文中提出了混合剪枝算法HAP(HybridAlgorithmofPruning)。算法首先联合进化算法代表之一遗传算法(GA)和反向传播算法BP的不同优势完成ANN网络结构和权重进化的初步阶段;然后应用多权重剪枝策略(MW-OBS)进一步简化、确定网络结构。结合案例与以往的混合策略算法进行对比研究,结果表明HAP在寻优能力、简化网络结构、保证稳定性等方面均有明显优势,更加适合大规模ANN的优化问题。     

基于多尺度跃层卷积神经网络的精细车型识别

为解决精细车型识别中特征不具有代表性,且识别准确率低的问题,提出了基于多尺度跃层卷积神经网络(CNN)的车型识别方法。通过多个不同尺度的跃层卷积神经网络,提取适用于精细车型识别的低层局部特征和高层全局特征,并分别训练Softmax分类器。利用自适应方式融合方法,将多个单一尺度跃层卷积神经网络的识别结果进行融合,调整不同网络对识别结果的贡献。实验中车型识别准确率达到97.59%。实验结果表明多尺度跃层卷积神经网络适用于精细的车型识别,并能提高识别的准确率。     

基于混沌神经网络的语音识别方法

基于语音信号的时变特性,研究了神经网络语音识别的方法．把混沌特性引入到神经元,构造了一种新的多层混沌神经网络结构,同时推导了相应的学习算法．把这种混沌神经网络用于语音识别,并与常用的神经网络语音识别方法作了比较．实验结果表明,混沌神经网络方法的平均识别率要高于同等条件下常用神经网络方法的识别率．     

利用串/并联神经网络建立化学反应器模型

本文研究使用线性动态神经网络与非线性的静态网络相结合的混合建模方式解决复杂非线性系统的建模问题。使用混合神经网络建模，可以降低单个网络的训练难度，基于此，也可将非线性系统控制策略的求解分解，转换为线性系统的求解。从而改善使用单一神经网络建模存在的精度不高以及训练时间长等不足，也为非线性系统控制策略的求解提供方便。本文以一个典型多变量系统——连续搅拌釜式反应器（CSTR）作为仿真对象，详细研究和实现了两类神经网络串联和并联的混合建模方法，并对结果进行了比较。     

基于遗传算法和模拟退火算法优化神经网络的铁路营业里程预测

提出应用遗传算法(GA)和模拟退火(SA)优化神经网络预测铁路营业里程.采用3层前馈神经网络实现铁路营业里程的时间序列预测,输入节点数为5,隐层节点数为8,输出节点数为1.对神经网络的连接权重和节点阈值的确定,采用GA和SA算法相结合的混合优化学习策略.两种算法结合时,SA算法处于外层,GA处于内层.GA采用实数编码,把要确定的神经网络节点连接权重和节点阈值作为基因串.数值计算结果表明混合优化的神经网络的学习速度和精度都比单纯BP算法得出的结果好.因此,用GA-SA混合优化的神经网络预测铁路营业里程是可行的.     

神经网络控制系统仿真

利用基于神经网络修正误差ＢＰ学习算法的多层网络和间接学习或专门学习的动态逆特性控制方法＾［１］编制的神经网络控制系统的仿真软件（ＳＣＳＢＮＮ），给出了调节时间和最大超调量与神经网络中间层节点数的关系曲线，同时给出了各种学习率和神经元作用函数增益的响应曲线。ＳＣＳＢＮＮ也可用于神经网络非线性控制系统。仿真结果说明神经网络非线性控制系统具有良好的控制性能。     

基于遗传算法的进化神经网络

提出了一种基于遗传算法的前馈神经网络的自动化设计方法 (genetic m ultilayer neural network,GMNN ) ,用以同时完成对网络结构空间和权值空间的搜索。该算法利用模拟退火算法、 BP算法和小生境技术来加快算法的收敛速度 ,改善解的性能。初步实验结果表明 ,该方法的收敛速度较快 ,由此得到的神经网络的泛化能力也较好 ,能够达到根据训练样本自动优化设计多层前馈式神经网络的目的。     

一种基于蚁群聚类的径向基神经网络

提出了一种基于蚁群聚类算法的径向基神经网络．利用蚁群算法的并行寻优特征和挥发系数方法的自适应更改信息量的能力，并以球面聚类的方式确定了径向基神经网络中基函数的位置，同时通过比较隐层神经元的相似性、合并相似性较为接近的2个神经元来约简隐含层的神经元，以达到简化径向基神经网络结构的目的．实验比较了几种不同聚类算法的径向基神经网络，结果表明，所提神经网络的整体训练时间至少可缩短40％，学习的准确率可提高1%以上，而且网络结构更加精简．     

基于WNN的双目摄像机标定方法

针对传统摄像机标定方法需要建立复杂的数学模型,且计算量大、实时性不好的问题,引入了人工神经网络来有效处理非线性映射问题,准确地建立起立体视觉中三维空间特征点与它在图像平面上像点之间的非线性关系;但现有的神经网络标定法仍存在实时性差、标定精度不够、泛化能力差的缺点,于是该文提出了一种基于小波神经网络（waveletneuralnetwork,WNN）的方法,同时用粒子群优化算法对学习算法进行改进,并对小波网络与BP网络的标定结果进行比较．实验结果表明,基于小波神经网络的双目视觉标定方法能够达到较高的实时性、标定精度和泛化能力的要求．     

增量构造负相关异构神经网络集成的方法

基于负相关异构网络,提出了一种增量构造异构神经网络集成(NNE)的方法.该方法在训练成员网络时,不仅调整网络的连接权值,而且动态调整网络的结构,从而在提高单个网络精度的同时增加各成员网络之间的差异度,减小网络集成的泛化误差.该方法包括构造最佳异构网络(BHNN)和构造异构网络集成(HNNE)两个部分,BHNN基于负相关学习动态构造多个最佳网络,HNNE利用训练好的最佳网络增量地构造异构NNE.使用网络泛化误差和集成泛化误差,整个集成过程可自动完成,无需预先确定成员网络的结构.分别对回归和分类问题进行了实验,相对于单个网络,该方法在测试数据集上的错误率降低了17%～85%,与已有的Boosting、Bagging等网络集成方法相比,错误率也有不同程度的改善.