基于神经网络的建筑工程造价预测研究

采用误差反向传播人工神经网络模型（BP网络模型），以建筑特征参数为输入变量，通过实际资料对网络进行训练和模拟，并用贡献分析法筛选输入变量，对网络结构进行优化，结果显示了该模型在建筑工程造价预测中的有效性     

径向基函数预测网络拓扑确定新方法

预测能力相对薄弱,已经成为制约PHM(Prognostics and Health Management)技术发展和应用的瓶颈.随着传感器和BIT(Built-in Test)设计技术的日益进步,采用序列分析的方法对复杂系统装备进行故障预测已经成为可能.在基于序列分析的预测方法研究中,径向基函数预测网络具有结构简单、学习速度快、具备非线性建模能力等诸多优点.为了改进其预测性能,在深入分析网络拓扑对模型性能及建模时间影响的基础上,综合考察了序列最佳线性自相关长度、建模精度和模型复杂度等多种因素,提出了基于偏自相关函数统计检测的输入层节点数目确定算法和基于BIC(Bayesian Information Criteria)准则的隐层节点数目确定算法,用以构建径向基函数预测网络;并对算法的有效性进行了分析.仿真结果表明,同传统建模算法相比较,由新算法构建的径向基函数预测网络具有最佳的预测性能,且建模时间不足传统算法的3%.     

基于PF-LSTM的锂电池剩余使用寿命预测

针对锂电池剩余使用寿命(RUL)难以准确预测的问题，提出一种考虑多种寿命衰退特征与数据时序性的基于粒子滤波改进长短期记忆网络(PF-LSTM)的预测模型，并应用于锂电池的RUL预测。从电池历史充放电老化数据中提取与容量衰退密切相关的健康因子作为LSTM网络的输入，利用PF算法全局优化的能力寻优超参数，包括神经元个数、学习率、节点丢弃率、批尺寸大小、训练步数等6个参数，提高网络的预测能力；引入Dropout层，避免网络过拟合，提高模型的泛化能力。基于NASA PCoE电池数据集进行实验验证，对4块电池在不同预测起始点下的容量估计和寿命情况进行预测，并与经网格搜索的LSTM,SVR等算法进行比较。实验结果表明，PF-LSTM容量估计的RMSE与MAE均在2%以内，且寿命预测误差在3个循环以内，相比于其他算法精度最高。     

反应式容侵系统入侵预测的混合式贝叶斯网络方法

为解决反应式容忍入侵系统中的入侵预测问题,提出了新的混合式贝叶斯网络方法。该方法中,提出了一种基于系统安全状态的入侵模型,以攻击者能力上升的过程来描述入侵,关注入侵对系统的影响,适合于反应式容侵系统根据当前状态选择合适的响应机制。提出了基于入侵模型的混合式贝叶斯网络（HyBN,hybrid bayesian network）模型,将入侵模型中攻击行为和系统安全状态节点分离为攻击层和状态层两个网络层次,两层间使用收敛连接,而两层内部的节点间使用连续连接。在特定的信度更新算法的支持下,实验说明该贝叶斯网络方法用于入侵预测的有效性,比较说明HyBN方法的优点。     

改进生成对抗网络及其在结构非线性模型修正中的应用

提出改进生成对抗网络（Generative Adversarial Network，GAN）并在结构非线性模型修正中成功应用。在改进的GAN中，通过引入代理模型的方式，增强网络判别器对非线性结构各节点响应关系特征的学习能力；为避免传统GAN存在的梯度消失问题，使用跳跃连接和密集连接等方式加强网络层之间的信息交流，并且通过引入组合目标函数，构建模型输入响应与输出参数之间的映射关系实现网络训练。在进行结构非线性模型修正时，结构的动力响应作为网络模型的输入，训练好的GAN模型能够根据输入数据的特征，输出非线性模型参数的最优值，从而实现结构非线性模型修正。通过对地震荷载作用下的12层钢筋混凝土框架结构进行数值模拟，验证了方法的可行性，并通过对比基于卷积神经网络的非线性模型修正结果，验证所提方法的优越性；最后进一步结合地震荷载作用下的悬臂铝梁振动台实验，验证了该非线性模型修正方法的可靠性。     

用粗集-模糊神经网络评定空袭目标威胁程度

针对模糊神经网络运算过程中,当模糊规则较多时,网络学习速度慢,方法实时性差的缺点,本文提出采用粗糙集理论对该模型进行优化,该方法利用粗集数据分析方法,通过知识约简从数据中推理逻辑规则,并用约简后规则集作为模糊神经网络的规则将输入映射到输出的子空间上：在这个子空间上用改进的BP算法训练进行逼近。实验结果表明：通过粗集数据挖掘后提取的规则,不仅规则数目减少,且规则是不完全规则,减少了网络输入维数和各层神经元的个数,提高了网络运算速度,满足了系统实时性要求。     

用粗集-模糊神经网络评定空袭目标威胁程度

针对模糊神经网络运算过程中,当模糊规则较多时,网络学习速度慢,方法实时性差的缺点,本文提出采用粗糙集理论对该模型进行优化,该方法利用粗集数据分析方法,通过知识约简从数据中推理逻辑规则,并用约简后规则集作为模糊神经网络的规则将输入映射到输出的子空间上:在这个子空间上用改进的BP算法训练进行逼近.实验结果表明:通过粗集数据挖掘后提取的规则,不仅规则数目减少,且规则是不完全规则,减少了网络输入维数和各层神经元的个数,提高了网络运算速度,满足了系统实时性要求.     

基于BP神经网络算法的316L不锈钢极薄带热处理力学性能预测

目的 为了预测不锈钢极薄带热处理后的力学性能、优化热处理工艺以及实现热处理工艺的智能控制,构建基于BP算法的神经网络模型。方法 以316L不锈钢极薄带为研究对象,进行热处理试验和拉伸试验,通过以热处理的退火温度、保温时间和取样方向作为输入层参数,以屈服强度、抗拉强度、断后伸长率作为输出层参数,采用BP算法构建了316L不锈钢极薄带力学性能预测的思维进化算法优化BP神经网络模型,并进行模型的预测和应用验证,考虑不同隐含层节点数及不同BP神经网络模型对性能的影响。结果 思维进化算法优化的BP神经网络模型测试集的屈服强度、抗拉强度和断后伸长率的平均相对误差分别为8.92%、5.21%和9.28%,训练集相关系数为0.980 94。思维进化算法优化BP网络单、双隐含层误差总和最低分别为0.578 6和0.546 9,BP网络与思维进化算法优化的BP网络误差总和最低分别为0.579 9和0.546 9。结论 思维进化算法优化BP神经网络模型具有较好的预测能力和泛化能力,以及较高的预测精度。与企业现用生产工艺相比,采用模型优化后热处理工艺的综合力学性能有显著提高。     

工位空调热舒适指标PMV人工神经网络模型研究

建立了预测工位空调微环境热舒适指标PMV的人工神经网络模型。模型的7938组输入向量数据选自ISO7730中推荐的PMV及其参数范围,并考虑工作位形成的微环境的参数区间,以及ASHRAE标准舒适区域。编制程序计算出输入向量对应的PMV值作为模型的输出量。对网络进行训练和测试的结果表明,用人工神经网络建立的模型能够迅速准确地预测工位空调微环境的热舒适指标PMV。     

偏最小二乘回归神经网络模型在爆破振动峰值速度预测中的应用

神经网络方法是处理非线性问题的有力工具,但当输入变量较多,输入变量间存在的多重共线性性会使得网络的建模效率下降。偏最小二乘回归方法通过提取对因变量解释性较强的成分,能较好地克服变量间的多重共线性。将两种方法相结合,建立了爆破振动峰值速度的偏最小二乘回归BP神经网络预测模型。利用偏最小二乘法对影响爆破振动的因素进行分析,提取出3个新综合变量,使BP网络的输入层节点数目由9个减少到3个,简化了网络结构,提高了计算速度,增强了网络稳定性。分析结果表明,耦合模型的平均预测误差为7.62%,相较于传统的萨氏公式及标准的BP神经网络模型其预测精度有了明显提高。     

Rule extraction from an optimized neural network for traffic crash frequency modeling

This study develops a neural network (NN) model to explore the nonlinear relationship between crash frequency and risk factors. To eliminate the possibility of over-fitting and to deal with the black-box characteristic, a network structure optimization algorithm and a rule extraction method are proposed. A case study compares the performance of the trained and modified NN models with that of the traditional negative binomial (NB) model for analyzing crash frequency on road segments in Hong Kong. The results indicate that the optimized NNs have somewhat better fitting and predictive performance than the NB models. Moreover, the smaller training/testing errors in the optimized NNs with pruned input and hidden nodes demonstrate the ability of the structure optimization algorithm to identify the insignificant factors and to improve the model generalization capacity. Furthermore, the rule-set extracted from the optimized NN model can reveal the effect of each explanatory variable on the crash frequency under different conditions, and implies the existence of nonlinear relationship between factors and crash frequency. With the structure optimization algorithm and rule extraction method, the modified NN model has great potential for modeling crash frequency, and may be considered as a good alternative for road safety analysis.     

Quantification of ternary mixtures of heavy metal cations from metallochromic absorbance spectra using neural network inversion

A new method based on artificial neural networks (ANN) for the processing of spectrophotometric data is proposed and illustrated on the example of the simultaneous quantification of ternary mixtures of zinc, cadmium, and mercury cations in aqueous solutions. Three types of commercially available metallochromic indicators were used as a simple model setup to create spectral data analogous to those normally received from an optical sensor array. In conventional ANN training methods for chemical sensors based on spectrophotometric data, a calibration is established by mathematically correlating the measured optical signal as network input with the concentration of the calibration sample as network output. In several situations, however, especially when dealing with mixed sample solutions, the relationship between a measured absorption spectrum and the corresponding ion concentrations is ambiguous, resulting in an "ill-posed problem". On the other hand, if the training direction is reversed by correlating known sample concentrations with measured optical signals, the relationship becomes reasonable for the ANN to obtain its structure. The proposed model illustrated in this paper is based on a more reasonable direct mapping and estimation by artificial neural network inversion (ANNI). In the training step, sample mixtures of known concentrations are optically measured to construct networks correlating the input data (ion concentrations) and the output data (absorption spectra). In the estimation step, the ion concentrations of unknown samples are estimated using the constructed ANN. The measured spectra of the unknown samples are fed to the output layer, and the appropriate input concentrations are determined by ANNI. When training the ANN system with 143 ternary mixtures of Zn2+, Cd2+, and Hg2+ in a concentration range from 1 to 100 microM, root-mean-square errors of prediction (RMSEP) of 0.45 (Zn2+), 0.96 (Cd2+), and 0.32 microM (Hg2+) were observed for the estimation of concentrations in 30 test samples, using the ANNI procedure. This newly proposed model, which involves the construction of an ANN based on direct mapping and estimation by ANNI, opens up one way to overcome the limitations of nonselective sensors, allowing the use of more easily accessible semiselective receptors to realize smart chemical sensing systems.     

Reliability analysis of structures using neural network method

In order to predict the failure probability of a complicated structure, the structural responses usually need to be estimated by a numerical procedure, such as finite element method. To reduce the computational effort required for reliability analysis, response surface method could be used. However the conventional response surface method is still time consuming especially when the number of random variables is large. In this paper, an artificial neural network (ANN)-based response surface method is proposed. In this method, the relation between the random variables (input) and structural responses is established using ANN models. ANN model is then connected to a reliability method, such as first order and second moment (FORM), or Monte Carlo simulation method (MCS), to predict the failure probability. The proposed method is applied to four examples to validate its accuracy and efficiency. The obtained results show that the ANN-based response surface method is more efficient and accurate than the conventional response surface method.     

Fault detection based on Gaussian process latent variable models

Gaussian processes, GPs, can be used to approximate complex non-linear functions with relative simplicity. Their regression performance is, at least, comparable to that achieved via artificial neural networks (ANN) and, in fact, both methods are intrinsically related. They are both non-parametric and, as Neal (1994) [1] has shown, when the number of nodes in the hidden layer of a neural network tends to infinity the ANN converge to a Gaussian process.In most of the cases, the GP will map a multivariate input into a univariate response. In this paper, however, we present an approach to process monitoring that combines several GPs so that multivariate responses can be appropriately modeled. We review a similar approach recently proposed in the literature and highlight some concerns related to it that needs to be taken into consideration. Additionally, we propose an alternative procedure to the way in which new observations are mapped into the non-linear model. A simulation study is provided that will help understand the method flexibility. Furthermore, results from a real example are also discussed.     

Free vibration analysis of functionally graded annular plates by state-space based differential quadrature method and comparative modeling by ANN

This paper deals with three-dimensional analysis of functionally graded annular plates through using state-space based differential quadrature method (SSDQM) and comparative behavior modeling by artificial neural network (ANN) for different boundary conditions. The material properties are assumed to have an exponent-law variation along the thickness. A semi-analytical approach which makes use of state-space method in thickness direction and one-dimensional differential quadrature method in radial direction is used to obtain the vibration frequencies. The state variables include a combination of three displacement parameters and three stress parameters. Numerical results are given to demonstrate the convergency and accuracy of the present method. Once the semi-analytical method is validated, an optimal ANN is selected, trained and tested by the obtained numerical results. In addition to the quantitative input parameters, support type is also considered as a qualitative input in NN modeling. Eventually the results of SSDQM and ANN are compared and the influence of thickness of the annular plate, material property graded index and circumferential wave number on the non-dimensional natural frequency of annular functionally graded material (FGM) plates with different boundary conditions are investigated. The results show that ANN can acceptably model the behavior of FG annular plates with different boundary conditions.     

Giant Magnetoimpedance Effect: Concept and Prediction in Amorphous Materials

Giant magneto impedance (GMI) effect was experimentally measured on as-cast, post-production and coated with chemical technique amorphous wire and ribbon materials consisted of varied chemical composition over a frequency range from 0.1 to 8 MHz under a static magnetic field between ?8 and +8 kA/m. The results show that each amorphous sample has a certain operational frequency for which the GMI effect has maximum magnitude and the other parameters such as annealing and coating have a significant influence on the GMI effect. It is believed that the domain structure and wall mechanism in the material are responsible for this behaviour. A 3-node input layer, 1-node output layer artificial neural network (ANN) model with three hidden layers including 30 neurons and full connectivity between the nodes was developed. A total of 1600 input vectors obtained from varied treated samples was available in the training data set. After the network was trained, better results were obtained from the network formed by the hyperbolic tangent transfer function in the hidden layers, there was a sigmoid transfer function in the output layer and we predicted the GMI. Comparing the predicted values obtained from the ANN model with the experimental data indicates that a well-trained neural network model provides very accurate results.     

航天光学遥感器信噪比的人工神经网络评价

提出基于人工神经网络进行航天光学遥感器信噪比评价的方法,首先对航天遥感图像进行分析,从图像中将与景物结构和噪声有关的特征向量分别提取出来,作为ANN的输入。网络通过对大量信噪比已知的图像样本训练后,可完成对航天光学遥感器传输下来的任意一幅地面景物图像进行系统的信噪比测试,从而避免了采用特定景物目标进行测量中的诸多弊端,测量平均误差低于10%。     

基于子结构和遗传神经网络的递推模型修正方法

根据实际动力响应对结构有限元模型进行修正,是实现损伤识别和健康监测的必要前提。针对基于神经网络的模型修正方法的不足,选用均匀设计法构造样本从而有效减少所需样本数量,而且计算效率高。采用遗传算法优化神经网络权值,提高了运算速度。基于上述研究,提出了基于子结构和神经网络的递推模型修正方法。该方法将结构分解成多层次的子结构,选取适当的损伤因素逐步实现逐级的修正。应用该方法对一网壳结构进行了模型修正,修正中首先采用固有频率作为损伤因素,结果表明遗传算法明显地提高了神经网络的计算速度,最后的递推修正效果令人满意;其次提出了采用小波包频带能量作为损伤因素的修正方法,该方法同样准确有效,并且不再依赖传统的模态分析技术,更为实用便捷。     

A stable and optimized neural network model for crash injury severity prediction

The study proposes a convex combination (CC) algorithm to fast and stably train a neural network (NN) model for crash injury severity prediction, and a modified NN pruning for function approximation (N2PFA) algorithm to optimize the network structure. To demonstrate the proposed approaches and to compare them with the NN trained by traditional back-propagation (BP) algorithm and an ordered logit (OL) model, a two-vehicle crash dataset in 2006 provided by the Florida Department of Highway Safety and Motor Vehicles (DHSMV) was employed. According to the results, the CC algorithm outperforms the BP algorithm both in convergence ability and training speed. Compared with a fully connected NN, the optimized NN contains much less network nodes and achieves comparable classification accuracy. Both of them have better fitting and predicting performance than the OL model, which again demonstrates the NN’s superiority over statistical models for predicting crash injury severity. The pruned input nodes also justify the ability of the structure optimization method for identifying the factors irrelevant to crash-injury outcomes. A sensitivity analysis of the optimized NN is further conducted to determine the explanatory variables’ impact on each injury severity outcome. While most of the results conform to the coefficient estimation in the OL model and previous studies, some variables are found to have non-linear relationships with injury severity, which further verifies the strength of the proposed method.