一种多层前馈网参数可分离学习算法

目前大部分神经网络学习算法都是对网络所有的参数同时进行学习．当网络规模较大时，这种做法常常很耗时．由于许多网络，例如感知器、径向基函数网络、概率广义回归网络以及模糊神经网络，都是一种多层前馈型网络，它们的输入输出映射都可以表示为一组可变基的线性组合．网络的参数也表现为二类：可变基中的参数是非线性的，组合系数是线性的．为此，提出了一个将这二类参数进行分离学习的算法．仿真结果表明，这个学习算法加快了学习过程。

基于 PID 调节的多层前馈网及其应用研究

介绍了多层前馈网的基本算法及其在镦粗压力计算中的应用．针对其基本学习算法（反向误差传播算法，ＢＰ算法）存在的收敛速度慢和易陷入局部最小值的问题，提出了基于ＰＩＤ调节的ＢＰ算法，由此建立起镦粗压力计算的神经网络知识库，经过测试集考核，证明该模型算法正确，能有效地提高模型的判别精度，是一种有效的和有前途的算法．

神经网络的函数逼近理论

分析了将函数逼近理论与方法引入神经网络研究的必要性；从经典函数逼近与统计分析两方面详细地讨论了多层前馈网（ＭＬＰ）逼近能力分析的基本方法及结论；分析了正则理论观点下的径向基函数网络（ＲＢＦ）的逼近能力；讨论了ＲＢＦ网与多层前馈网在最佳逼近特性上的差异。文末指出了神经网络函数逼近的发展方向。

用神经网辨识非线性系统中的模型误差分析(Ⅱ)——随机系统中的噪声影响

针对含简单系统噪声和输出噪声的系统,详细分析了在通常的建模方法下噪声对模型权值及模型预报的影响。

Automated Building Block Extraction and Building Density Classification Using Aerial Imagery and LiDAR Data

This paper examines the utility of high-resolution airborne RGB orthophotos and LiDAR data for mapping residential land uses within the spatial limits of suburb of Athens, Greece. Modem remote sensors deliver ample information from the AOI （area of interest） for the estimation of 2D indicators or with the inclusion of elevation data 3D indicators for the classification of urban land. In this research, two of these indicators, BCR （building coverage ratio） and FAR （floor area ratio） are automatically evaluated. In the pre-processing step, the low resolution elevation data are fused with the high resolution optical data through a mean-shift based discontinuity preserving smoothing algorithm. The outcome is an nDSM （normalized digital surface model） comprised of upsampled elevation data with considerable improvement regarding region filling and ＂straightness＂ of elevation discontinuities. Following this step, a MFNN （multilayer feedforward neural network） is used to classify all pixels of the AOI into building or non-building categories. The information derived from the BCR and FAR building indicators, adapted to landscape characteristics of the test area is used to propose two new indices and an automatic post-classification based on the density of buildings.

Wavelet neural network based watermarking technology of 2D vector maps

A novel lossless information hiding algorithm based on wavelet neural network for digital vector maps is introduced. Wavelet coefficients being manipulated are embedded into a vector map, which could be restored by adjusting the weights of neurons in the designed neural network. When extracting the watermark extraction, those coefficients would be extracted by wavelet decomposition. With the trained multilayer feed forward neural network, the watermark would be obtained finally by measuring the weights of neurons. Experimental results show that the average error coding rate is only 6% for the proposed scheme and compared with other classical algorithms on the same tests, it is indicated that the proposed algorithm hashigher robustness, better invisibility and less loss on precision.

Power Big Data Fusion Prediction

This paper is a research on the characteristics of power big data. According to the characteristics of ＂large volume＂, ＂species diversity＂, ＂sparse value density＂, ＂fast speed＂ of the power big data, a prediction model of multi-source information fusion for large data is established, the fusion prediction of various parameters of the same object is realized. A combined algorithm of Map Reduce and neural network is used in this paper. Using clustering and nonlinear mapping ability of neural network, it can effectively solve the problem of nonlinear objective function approximation, and neural network is applied to the prediction of fusion. In this paper, neural network model using multi layer feed forward network--BP neural network. Simultaneously, to achieve large-scale data sets in parallel computing, the parallelism and real-time property of the algorithm should be considered, further combined with Reduce Map model, to realize the parallel processing of the algorithm, making it more suitable for the study of the fusion of large data. And finally, through simulation, it verifies the feasibility of the proposed model and algorithm.

Temperature and Humidity Control System Identification Based on Neural Network in Heating and Drying System

Artificial neural network has unique advantages for massively parallel processing, distributed storage capacity and self-learning ability. The paper mainly constructs neural network identifier and neural network controller for system identification and control on temperature and hmnidity of heating and drying system of materials. And the paper introduces the structure and principles of neural network, and focuses on analyzing learning algorithm, training algorithm and limitation of the most widely applied multi-layer feed-forward neural network （ BP network） , based on which the paper proposes introducing momentum to improve BP network.

Seismic Health Monitoring of Foundations Using Artificial Neural Networks

Damage identification plays an important role in structural health monitoring systems. Despite variety in damage identification methods, little attention has been paid to the seismic damage identification of foundations. When shear walls serve as the lateral load resistance system of structures, foundations may subject to the high level of concentrated moment and shear forces. Consequently, they can experience severe damage. Since such damage is often internal and not visible, visual inspections cannot identify the location and the severity of damage. Therefore, a robust method is required for damage localization and quantification of foundations. According to the concept of performance-based seismic design of structures, the seismic behavior of foundations is considered as Force-Controlled. Therefore, for damage identification of foundation, internal forces should be estimated during ground motions. In this study, for real-time seismic damage detection of foundations, a method based on artificial neural networks was proposed. A feed-forward multilayer neural network with one hidden layer was selected to map input samples to output parameters. The lateral displacements of stories were considered as the input parameters of the neural network while moment and shear force demands at critical points of foundations were taken into account as the output parameters. In order to prepare well-distributed data sets for training the neural network, several nonlinear time history analyses were carried out. The proposed method was tested on the foundation of a five-story concrete shear wall building. The obtained results revealed that the proposed method was successfully estimated moment and shear force demands at the critical points of the foundation.

A Credit Risk Evaluation Approach to Neural Network Training by Means of Financial Ratios

In recent years artificial neural networks are used to recognize the risk category of investigated companies. The research is based on data from 81 listed enterprises that applied for credit in domestic regional banks operating in China. The backpropagation algorithm-the multilayer feedforward network structure is described. Each firm is described by 9 diagnostic variables and potential borrowers are classified into four classes. The efficiency of classification is evaluated in terms of classification errors calculated from the actual classification made by the credit officers. The results of the experiments show that LevenbergMarque training error is smallest among 4 learning algorithms and its performance is better, and application of artificial neural networks and classification functions can support the creditworthiness evaluation of borrowers.