基于迁移学习和Bi-LSTM神经网络的桥梁温度-应变映射建模方法

为快速构建并准确预测温度作用引起的斜拉桥主梁应变用于结构状态评估，基于某大跨度斜拉桥主梁超过1年的温度和应变监测数据，提出了一种基于迁移学习和双向长短时记忆(bi-directional long short-term memory, Bi-LSTM)神经网络的斜拉桥温度-应变映射模型建立方法。首先，利用解析模态分解(analytical mode decomposition, AMD)去噪应变数据，得到仅由温度引起的应变响应；其次，选择温度和某一测点应变数据构成数据集，采用Bi-LSTM神经网络训练该数据集，并通过网络结构和超参数优化建立温度-应变Bi-LSTM基准模型；最后，利用迁移学习方法，将已训练好的基准模型中部分参数迁移到其他温度-应变数据集，建立相应的温度-应变映射被迁移模型，并与未采用迁移学习的神经网络训练方法进行对比。研究结果表明，相比直接建立的温度-应变Bi-LSTM神经网络映射模型，采用迁移学习方法建立的被迁移模型，其拟合精度均高于所用的基准模型，且训练时间短，预测误差小。     

基于RBF和BP神经网络的低真空管道高速列车气动阻力预测对比研究北大核心CSCD

为了实现对低真空管道中运行列车的最大阻力预测研究,本文采用数值仿真和神经网络结合的方法。选取不同阻塞比、运行速度和管道压力,利用流体仿真软件计算100种运行工况下列车的最大阻力;以96组仿真数据作为网络模型训练样本,选取RBF和BP两种三层神经网络,经多次调试确定最佳隐层神经元数目,利用训练函数训练两种预测模型;利用随机选取的4组验证样本验证两种网络模型。研究表明:RBF和BP神经网络模型能较好的预测列车在真空管道中运行的最大阻力,其中RBF神经网络预测值的最大误差不高于5%,相比BP神经网络,RBF预测精度更优。     

基于RBF和BP神经网络的低真空管道高速列车气动阻力预测对比研究

为了实现对低真空管道中运行列车的最大阻力预测研究,本文采用数值仿真和神经网络结合的方法。选取不同阻塞比、运行速度和管道压力,利用流体仿真软件计算100种运行工况下列车的最大阻力;以96组仿真数据作为网络模型训练样本,选取RBF和BP两种三层神经网络,经多次调试确定最佳隐层神经元数目,利用训练函数训练两种预测模型;利用随机选取的4组验证样本验证两种网络模型。研究表明:RBF和BP神经网络模型能较好的预测列车在真空管道中运行的最大阻力,其中RBF神经网络预测值的最大误差不高于5%,相比BP神经网络,RBF预测精度更优。     

基于神经网络的输电塔钢管构件涡激振动幅值预测方法

输电塔中长细比较大的钢管构件容易发生低风速下的涡激振动，鉴于传统风洞试验和数值模拟研究方法存在的成本高、周期长的局限，该文提出了一种基于神经网络的输电塔钢管构件涡激振动幅值高效预测方法。为获取训练模型所需的数据集，发展了适用于任意插板形式、几何尺寸的钢管构件涡激振动响应分析方法；结合多种神经网络模型(BPNN、PSO-BPNN、RBFNN、GRNN)以及性能评价指标，建立了基于神经网络的输电塔钢管构件涡激振动幅值预测方法；通过算例对某C型插板和十字型插板钢管构件涡激振动幅值进行了预测。研究表明：通过与试验结果的对比，验证了该文输电塔钢管构件涡激振动响应分析方法的准确性，对于C型和十字型插板钢管构件VIV幅值的相对误差分别为3.84%和5.87%，利用该方法可为神经网络模型提供可靠样本；通过7折10次交叉验证优化超参数后的4种神经网络模型，均表现出较好的预测精度；相比之下，GRNN在C型插板和十字型插板钢管构件算例中均呈现出最佳的泛化能力，其R2值分别为0.989和0.992；采用GRNN方法可以较好地预测C型和十字型插板钢管构件在不同质量阻尼比参数下的VIV幅值，且在计算效率上相比于C...     

用人工神经网络预测冰蓄冷系统蓄冰时间

蓄冰时间的预测对于冰蓄冷空调系统的设计和运行控制十分重要。在本文中,作者以理论计算数据作训练集、验证集、测试集,采用BP型人工神经网络预测了板单元冰蓄冷系统的蓄冰时间．取得了令人满意的结果。与采用差分数值计算相比,用神经网络可大大缩短计算时间。     

低温余热回收ORC系统工质的筛选

利用有机朗肯循环(ORC)技术高效回收低温余热的关键之一是选用合适的工质。本文针对热源温度介于120~220℃区间内的ORC系统,选用R123,R245fa,R600和R1233zd(E)四种工质为研究对象,通过对50 kW的ORC系统的运行分析并结合模拟计算,详细讨论不同工质的热力特性以及蒸发温度、蒸发压力、蒸发器出口过热度对ORC系统热效率的影响。结果表明,对于透平膨胀机入口工质温度在100~150℃区间、热源温度在120~220℃区间的低温余热回收ORC系统,工质R600性能表现最佳,但易燃;从不可燃性、热力特性、环境友好性及设备成本方面考虑,R1233zd(E)具有优势,但工质价格较高。     

基于WOA–BP神经网络的液滴铺展预测

目的 提高BP神经网络对电喷印过程中液滴铺展行为的预测能力。方法 提出一种鲸鱼优化算法(WOA)优化BP神经网络的液滴铺展预测模型。首先，采用相场方法建立电场作用下液滴铺展的数值模型，并通过实验验证仿真结果的准确性。然后，选取初始直径、撞击速度、接触角和电场强度作为神经网络的输入参数，将最大铺展直径作为神经网络的输出参数，利用鲸鱼优化算法优化神经网络中的初始权值和阈值，构建液滴铺展预测模型。最后，基于仿真结果对预测模型进行训练与测试，并将其与传统的BP神经网络模型进行对比分析。结果 相较于传统BP神经网络预测模型，WOA–BP神经网络预测模型的平均绝对误差、均方根误差分别降低了72.60%、77.60%，而平均绝对百分比误差则从15.029 3%减小为4.585 3%。结论 WOA–BP神经网络预测模型可以更好地预测液滴铺展，可为液滴铺展的预测提供新的方法。     

大型结冰风洞制冷系统蒸发压力预测与降温过程优化研究

构建了液氨蒸发器热力学仿真模型,进行了蒸发压力以及风洞降温的预测与验证。采用BP神经网络算法对试验数据和仿真数据进行训练,建立了氨制冷系统蒸发压力快速预测模型,分析了试验高度和试验风速对风洞降温速度的影响,提出了变风速降温控制方式,对比分析了不同降温控制过程的降温时间及系统能耗。结果表明:对于高空环境试验,先进行风洞减压再降温,可以有效缩短降温时间、减小系统能耗;试验风速75 m/s时风洞系统降温速度最快,高于75 m/s时可以通过降温过程变风速控制方法降低系统能耗。     

类金刚石结构的分子动力学与人工神经网络耦合研究

掺氮类金刚石薄膜（N-DLC）可以改善零件表面的摩擦学性能,近些年来对N-DLC薄膜摩擦学特性研究的热度居高不下。由于计算资源与计算机运行时间有限,难以获得大量数据对N-DLC薄膜摩擦实验中界面结构演化规律进行微观模拟。为了探究分子动力学和人工神经网络交叉使用的可行性,全面了解N-DLC的摩擦学性质及规律,将BP神经网络、KELM神经网络引用到N-DLC的研究中。通过LAMMPS软件对N-DLC进行建模,将分子动力学模拟的数据作为人工神经网络的数据来源,对两种神经网络进行训练。利用验证样本对训练好的两种模型进行验证,将两种神经网络的预测结果进行对比,选出性能最佳的网络模型。结果表明,采用神经网络可以预测N-DLC内部杂化键的变化趋势,且效率更高,所需计算资源更少,在一定程度上可以代替分子动力学模拟结果,为人们提供进一步的分析判断。研究为促进分子动力学与人工神经网络两种方法的共同发展提供了有益探索。     

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。     

Application of artificial neural network for prediction of the oxidation behavior of aluminized nano-crystalline nickel

In this paper, the applicability of artificial neural network (ANN) for the prediction of the oxidation kinetics of aluminized coating is presented. For developing the model, a consistent set of experimental data i.e. nanocrystalline Ni samples were aluminized by two steps aluminizing process and oxidized at 800, 900 and 1000 °C for various times are used. The exposure time and temperature of oxidation were used as the inputs of the model and the resulting mass gain of oxidized samples as the output of the model. Multi-layer perceptron neural network structure and back-propagation algorithm are used for the training of the model. After testing many different ANN architectures an optimal structure of the model i.e. 2-5-6-1 is obtained. Comparison of experimental and predicted values using the proposed ANN model shows that there is a good agreement between them with mean relative error less than 1.2%. This shows that the ANN model is an accurate and reliable approach to predict the oxidation behavior of aluminized nanocrystalline coatings.     

Application of artificial neural networks for prediction of the retention indices of alkylbenzenes

Artificial neural networks (ANN) with extended delta–bar–delta (EDBD) learning algorithms were used to predict the retention indices of alkylbenzenes. The data used in this paper include 96 retention indices of 32 alkylbenzenes on three different stationary phases. Four parameters: temperature, boiling point, molar volume and the kind of stationary phase, were used as input parameters. These three stationary phases are: PEG, SE-30, SQ. The 96 group data were randomly divided into two sets: a training set (including 64 group data) and a testing set (including 32 group data). The structures of networks and the learning times were optimized. The best network structure is 4–7–1. The optimum number of learning time is about 20 000. It is shown that the maximum relative error is no more than 3%. The result illustrated that the prediction performance of ANN in the field of investigating the retention behaviors of alkylbenzenes is very satisfactory.     

基于改进神经网络的干点软测量

针对传统的神经网络收敛判断以模型的拟合精度为指标造成训练时间过长和过拟合等缺点,提出了一种改进神经网络(M-ANN).M-ANN将样本分成训练样本和校验样本,并提出了过拟合判据参数.通过训练样本采用误差反传算法对网络进行训练,训练过程中以模型对校验样本的预测性能为指标,通过过拟合判据参数的计算自适应地在获得具有最佳预测性能模型时终止网络训练.同时,针对影响初馏塔塔顶石脑油干点的因素众多且呈高度非线性的特征,应用M-ANN建立初顶石脑油干点软测量模型,获得模型的预测相对误差平方和均值比传统神经网络模型降低了27.5%.     

基于预测的复合地源热泵系统控制方法研究

在复合式地源热泵系统中控制策略存在着极大的优化空间,提出一种新的更为有效的控制方法,即在并联系统中直接比较冷却塔和地埋管出口温度的方法,然而在实际运行中只能实时测得一个出口水温,因此需要建立一个可靠的土壤换热器模型预测其出口水温。运用人工神经网络（ANN）实现该做法,利用FLUENT软件模拟动态复合式地源热泵系统为ANN模型提供训练、测试样本。为获得最优模型,土壤换热器的人工神经网络进行优化。结果表明,在LM算法下,隐层神经元数目为14的网络结构最为理想,预测结果绝对误差不超过0.15℃。     

Prediction of flow stress in dynamic strain aging regime of austenitic stainless steel 316 using artificial neural network

Flow stress during hot deformation depends mainly on the strain, strain rate and temperature, and shows a complex nonlinear relationship with them. A number of semi empirical models were reported by others to predict the flow stress during deformation. In this work, an artificial neural network is used for the estimation of flow stress of austenitic stainless steel 316 particularly in dynamic strain aging regime that occurs at certain strain rates and certain temperatures and varies flow stress behavior of metal being deformed. Based on the input variables strain, strain rate and temperature, this work attempts to develop a back propagation neural network model to predict the flow stress as output. In the first stage, the appearance and terminal of dynamic strain aging are determined with the aid of tensile testing at various temperatures and strain rates and subsequently for the serrated flow domain an artificial neural network is constructed. The whole experimental data is randomly divided in two parts: 90% data as training data and 10% data as testing data. The artificial neural network is successfully trained based on the training data and employed to predict the flow stress values for the testing data, which were compared with the experimental values. It was found that the maximum percentage error between predicted and experimental data is less than 8.67% and the correlation coefficient between them is 0.9955, which shows that predicted flow stress by artificial neural network is in good agreement with experimental results. The comparison between the two sets of results indicates the reliability of the predictions.     

Autoignition Temperature Prediction Using an Artificial Neural Network with Particle Swarm Optimization

The autoignition temperatures of organic compounds were estimated using a hybrid method that includes a simple group contribution method (GCM) implemented in an artificial neural network (ANN) replacing a standard back-propagation algorithm with particle swarm optimization (PSO). A data set of 250 compounds was used for training the network. The optimal condition of the network was obtained by adjusting various parameters by trial-and-error. The capabilities of the designed network were tested in the prediction of the autoignition temperature of 93 compounds not considered during the training step. The proposed model is shown to be more accurate than those of other published works. The results show that the proposed GCM + ANN + PSO method represent an excellent alternative for the estimation of this property with acceptable accuracy (AARD = 1.7%; AAE = 10K).     

Development a multi-layer perceptron artificial neural network model to estimate the Vickers hardness of Mn–Ni–Cu–Mo austempered ductile iron

The hardness of austempered ductile irons is relative to its microstructure, strength, ductility, machinability and wear resistance properties. Therefore, hardness measurement can be used as a simple tool to control the heat treatment, chemical composition and mechanical properties of ADI parts during the production process. The aim of this study is to develop an Artificial Neural Network (ANN) model for estimating the Vickers hardness of ADIs after austempering treatment. A Multi-Layer Perceptron model (MLP–ANN) was used with Mo%, Cu%, austempering time and temperature as inputs and the Vickers hardness of samples after austempering as the output of the model. A variety of samples were prepared in different conditions of chemical composition and heat treatment cycle. The obtained experimental results were used for training the neural network. Efficiency test of the model showed reasonably good agreement between experimental and numerical results, so the synthesized ANN model can estimate the hardness of the castings with a small error in the range of the experimental results standard deviation.     

Developing a model for hardness prediction in water-quenched and tempered AISI 1045 steel through an artificial neural network

The aim of the current study was to develop an artificial neural network (ANN) model to predict the hardness drop of the water-quenched and tempered AISI 1045 steel specimens, as a function of tempering temperature and time parameters. In the first stage, the effects of selected tempering parameters on the hardness drop value were investigated. In the second stage, a group of data, which have been obtained from experiments, was used for training of the ANN model. Likewise, another group of experimental data was utilized for the ANN model validation. Ultimately, maximum error of the ANN prediction was determined. The agreement between the predicted values of the ANN model with the experimental data was found to be reasonably good.     

基于粒子群支持向量回归的原位自生工艺参数优化

综合应用激光熔覆和原位反应增强金属基复合材料,是当前金属基复合材料研究领域的一个热点,本文采用该工艺制备铁基表面复合材料,重点考虑该工艺参数的确定问题.根据在不同工艺参数下合成的铁基表面的WC体积分数实测数据集,提出建立不同工艺参数下WC体积分数的支持向量回归预测模型,并与基于人工神经网络模型(ANN)的预测结果进行比较.结果显示:对于相同的训练样本和检验样本,SVR预测模型比ANN预测模型具有更强的泛化能力.最后根据建立的预测模型,应用粒子群算法寻优得到最优工艺参数,该工艺参数在实际实验过程中的应用,验证了该方法的有效性.     

Modelling of a thermal insulation system based on the coldest temperature conditions by using artificial neural networks to determine performance of building for wall types in Turkey

In formation of building external envelope, as two important criteria, climatic data and wall types must be taken into consideration. In the selection of wall type, the thickness of thermal insulation layer (d_i) must be calculated. As a new approach, this study proposes determining the thermal insulation layer by using artificial neural network (ANN) technique. In this technique five different wall types in four different climatic regions in Turkey have been selected. The ANN was trained and tested by using MATLAB toolbox on a personal computer. As ANN input parameters, U_w, T_e,Met, T_e,TSE, R_wt, and q_TSE were used, while d_i was the output parameter. It was found that the maximum mean absolute percentage error (MRE, %) is less than 7.658%. R² (%) for the training data were found ranging about from 99.68 to 99.98 and R² for the testing data varied between 97.55 and 99.96. These results show that ANN model can be used as a reliable modeling method of d_i studies.