GLOBAL CONVERGENCE RESULTS OF A THREE TERM MEMORY GRADIENT METHOD WITH A NON-MONOTONE LINE SEARCH TECHNIQUE

In this paper, a new class of three term memory gradient method with non-monotone line search technique for unconstrained optimization is presented. Global convergence properties of the new methods are discussed. Combining the quasi-Newton method with the new method, the former is modified to have global convergence property. Numerical results show that the new algorithm is efficient.

A LIMITED MEMORY QUASI-NEWTON METHOD FOR LARGE SCALE PROBLEM

We study how to use the SR1 update to realize minimization methods for problems where the storage is critical. We give an update formula which generates matrices using information from the last m iterations. The numerical tests show that the method is efficent.

Simulation of isothermal precision extrusion of NiTi shape memory alloy pipe coupling by combining finite element method with cellular automaton

In order to present the microstructures of dynamic recrystallization(DRX) in different deformation zones of hot extruded NiTi shape memory alloy(SMA) pipe coupling,a simulation approach combining finite element method(FEM) with cellular automaton(CA) was developed and the relationship between the macroscopic field variables and the microscopic internal variables was established.The results show that there exists a great distinction among the microstructures in different zones of pipe coupling because deformation histories of these regions are diverse.Large plastic deformation may result in fine recrystallized grains,whereas the recrystallized grains may grow very substantially if there is a rigid translation during the deformation,even if the final plastic strain is very large.As a consequence,the deformation history has a significant influence on the evolution path of the DRX as well as the final microstructures of the DRX,including the morphology,the mean grain size and the recrystallization fraction.

Feasibility Research on Fatigue Damage Evaluation Based on Magnetic Memory Method

To confirm whether Magnetic Memory Method can be used to evaluate the fatigue damage of ferromagnetic part or not,this paper stated the principle of magnetic memory method.Also, the relationship between testing parameter of magnetic memory method and fatigue damage was analyzed based on the experiment. The results show that magnetic memory method can be used to evaluate the fatigue damage represented by stress concentration. However, the relationship between the parameter of magnetic memory method and fatigue damage still needs to be studied.

The Model of Asynchronous Parallel Nonlinear Multisplitting Method on Shared Memory System

Nonlinear multisplitting method is known as parallel iterative methods for solving a large-scale system of nonlinear equations F(x) = 0. We extend the idea of nonlinear multisplitting and consider a new model ill which the iteration is executed asynchronously: Each processor calculate the solution of an individual nonlinear system belong to its nonlinear multisplitting and can update the global approximation residing in the shared memory at any time. A local convergence analysis of this model is presented. Finally, we give a uumerical example which shows a 'strange' property that speedup Sp > p and efficiency Ep > 1.

Quantitative Detection of Corrosion State of Concrete Internal Reinforcement Based on Metal Magnetic Memory

Corrosion can be very harmful to the service life and several properties of reinforced concrete structures.The metal magnetic memory(MMM)method,as a newly developed spontaneous magnetic flux leakage(SMFL)non-destructive testing(NDT)technique,is considered a potentially viable method for detecting corrosion damage in reinforced concrete members.To this end,in this paper,the indoor electrochemical method was employed to accelerate the corrosion of outsourced concrete specimens with different steel bar diameters,and the normal components Bz and its gradient of the SMFL fields on the specimen surfaces were investigated based on the metal magnetic memory(MMM)method.The experimental results showed that the SMFL experimental Bz curves are consistent with the analytical results of the theoretical model.Furthermore,the crest-to-trough behavior on the Bz signal curve and its zero-point gradient spacing can more accurately indicate the corroded area’s extent.Then,a magnetic characteristic parameter W based on a large amount of experimental data was established to characterize the degree of corrosion of the steel bars.The magnetic characteristic parameter W is linearly related to the maximum cross-sectional area loss rate
of the corroded reinforcement.This paper will lay the foundation for future research on corrosion detection of reinforced concrete structures based on the MMM method and provide a feasible way for non-destructive detection of corrosion independent of the influence of reinforcement diameter and magnetization history.

Set Programming Method and Performance Improvement of Phase Change Random Access Memory Arrays

A novel slow-down set waveform is proposed to improve the set performance and a 1 kb phase change random access memory chip fabricated with a 13nm CMOS technology is implemented to investigate the set performance by different set programming strategies based on this new set pulse. The amplitude difference （I1 - I2） of the set pulse is proved to be a crucial parameter for set programming. We observe and analyze the cell characteristics with different I1 - I2 by means of thermal simulations and high-resolution transmission electron microscopy, which reveal that an incomplete set programming will occur when the proposed slow-down pulse is set with an improperly high I1 - I2. This will lead to an amorphous residue in the active region. We also discuss the programming method to avoid the set performance degradations.

MEMORY-EFFICIENT PATH METRIC UPDATE METHOD IN MAP DECODER IMPLEMENTATION

A novel memory efficient path metric update is proposed for Maximum A Posteriori(MAP) decoder of turbo codes to reduce the memory requirement of state metric information calcu-lation. For MAP decoder,the same memory can be shared by the forward and backward metrics with this metric update scheme. The forward and backward metrics update can be performed at the same time. And all of the extrinsic information can be calculated at the end of metric update. Therefore,the latency and area in the implementation will be reduced with the proposed metric update method.

基于数据驱动的动车组镍镉电池记忆效应消除策略研究

在碱性镍镉蓄电池长期未彻底充放电的情况下,蓄电池会产生次级放电平台,称为“记忆效应”。该效应会导致电池可释放容量降低,需要定期消除。在动车组蓄电池检修中,传统的“记忆效应”消除方法耗时过长,会对检修工作带来很大压力。为了降低储能检修成本,对动车组镍镉电池“记忆效应”消除策略优化方案开展研究。以实际动车组三级修镍镉蓄电池电池为研究对象,通过分析电池的充放电循环中的参数特征,结合大量的电池充放电实验,提取出与电池容量密切相关的特征数据。首先,通过特征筛选和降维组合,搭建线性模型和非线性模型的交叉加权耦合方法,创建近似模型以反映电池的真实可释放容量,消除镍镉电池不一致性对优化方案实验设计带来的系统误差。其次,在此基础上,通过调整影响电池状态的高影响因子变量,构建正交实验方案。研究动车组镍镉蓄电池“记忆效应”快速、安全的消除策略。最后,通过重复验证实验和安全性检测实验对该策略进行实证,研究结果表明,提出的镍镉蓄电池“记忆效应”消除策略能够有效恢复镍镉电池真实容量,并且能够极大地缩减时间成本(约44.15%),并且有利于缓解蓄电池因过充导致电池劣化现象。该消除方案对于提升动车组镍镉电池的维护效率及确保其安全性具有重要意义。

基于循环神经网络的工程专业语义智能分析方法研究

针对传统翻译方法学习能力差、翻译质量较低的问题,提出了一种基于循环神经网络的专业英语机器翻译方法。该方法以编码器-解码器为模型框架,利用改进的循环卷积神经网络对输入数据加以训练。同时编码器使用多头注意力机制对输入数据进行共同训练,进而使算法兼具局部与全局特性。解码器单层则采用三子层结构,分别为多头注意力子层、上下文信息子层及全连接子层,可保证句子翻译的流畅性。在实验测试中,所提算法的BLEU值与其他算法相比提升了2.7;而在专业语料翻译测试中,相较于网络翻译,该算法的准确性和流畅度均更优,由此表明其性能较好,具有一定的工程应用价值。

选区激光熔化成形NiTi合金工艺参数对表面粗糙度的影响规律

目的针对选区激光熔化(SLM)制备NiTi形状记忆合金表面粗糙度难以满足实际应用要求,通过优化工艺参数(激光功率、扫描速度、扫描间距)以有效地降低表面粗糙度以及研究各工艺参数对表面粗糙度的影响规律。方法采用L16正交阵列的田口模型设计选区激光熔化制备NiTi样品的工艺参数,通过对表面粗糙度信噪比值进行统计方法分析以及样品表面形貌的表征,研究不同工艺参数对表面粗糙度的影响程度以及影响机理,最终优化出制备低表面粗糙度的工艺参数组合。结果在激光功率为20 W和30 W时,NiTi粉末不能够充分熔化造成熔道不连续,使得样品表面起伏增大,粗糙度值最大到7.8μm;增大激光功率到40W和50W时,粉末充分熔化,样品表面形貌明显改善;在相同功率下,扫描速度从200mm/s增加到500mm/s时,样品的粗糙度值也随之增大。结论工艺参数对表面粗糙度影响的重要性顺序依次为激光功率、扫描速度、扫描间距;最终优化出的工艺参数组合为激光功率50W、扫描速度200mm/s、扫描间距0.07mm,并在该工艺参数下制备的样品表面粗糙度值为1.38μm,与模型预测的值1.43μm接近,相差仅为9.97%。

数据驱动与机理耦合的冷轧电机功率智能预测方法

电机功率计算在保证生产安全、制定轧制规程以及发挥设备能力等方面起到了重要作用。为提高计算精度,本文提出一种机理模型与数据驱动方法相结合的电机功率预测模型。首先,机理模型部分采用上界法得到解析解,然后将部分功率解析解(塑性变形功率、张力功率和剪切功率)与相关变量输入长短时记忆(long short-term memory,LSTM)网络中训练,以提取变量数据之间的深度特征和时序相关性。结果表明:本文提出的时序耦合预测模型相对误差不超过±3.9%,且在相邻卷带钢轧制规程变化较大时,预测效果明显优于上界法和人工神经网络(artificial neural networks,ANN);从残差分布直方图中可以看到,上界法由于遗漏了摩擦系数和损失功率与其他参数间的强耦合关系,导致残差集中在±50 kW附近且不符合正态分布,而ANN和本模型均基本遵循正态分布,但本模型拟合精度更高;对比更广泛的模型评估指标,本模型具有更好的综合预测性能。此外,通过模型输出结果分析了电机功率与轧制长度、轧制速度之间的关系,表明本模型除具有高预测精度外,还与已知的参数间物理规律有较强一致性。

基于Distance-2算法的并行Jacobian矩阵计算及其在耦合问题中的应用

并行Newton-Krylov方法是求解大规模多物理耦合问题的有效方法,如何高效自动计算Jacobian矩阵是一大难点。利用有限差分方法,可避免推导Jacobian矩阵的表达式,实现矩阵的自动计算。现有工作表明,在串行环境下利用矩阵的稀疏性和图着色算法,Jacobian矩阵的计算效率可提高至少1个量级。但在并行环境下,串行着色算法失效,需采用相应的并行着色算法。本研究将图论领域的Distance-2算法应用于Jacobian矩阵的并行着色。通过求解一个简化多物理耦合问题检验了该并行算法的正确性和计算效率。测试结果表明,该并行算法得到的Jacobian矩阵完全正确;着色数随着并行核数的增加略微有所增加,100个进程下并行效率为56%;基于该算法求解多物理耦合问题,其计算时间和Krylov迭代次数较JFNK减少了约1/2。

Schwarzschild时空中带记忆项的波动方程耦合方程组解的奇性

本文研究Schwarzschild时空中非线性波动方程耦合方程组的Cauchy问题解的破裂性态.问题的非线性项包含混合型记忆项、组合和幂次型记忆项、组合和导数型记忆项以及组合型记忆项.当非线性项的指数满足一定假设时,利用迭代方法建立解的生命跨度的上界估计.创新之处是在Schwarzschild度量下分析非线性记忆项对解的生命跨度估计的影响.据已有文献所知,定理1.1-1.4中的结果是新的.

基于多源数据的山区小流域降水融合模型

为准确获取山区小流域的降水空间分布及其资源量,采用Kriging插值法对低分辨率卫星数据进行空间降尺度处理,通过长短期记忆网络(Long Short-Term Memory,LSTM)将局部卫星与观测数据进行降水融合,引入前期降水信息加强卫星与观测降水之间的时间相关性,并利用该模型进行流域降水空间分布估计。结果表明:从空间分布来看,融合模型对暴雨中心位置的捕捉更加精确;从降水量来看,所提模型在短时强降水下的探测率和临界成功指数分别为0.60和0.50,能够改善原始低分辨率卫星降水数据,使其更接近实际情况;从雨量站数量来看,融合降水的精度随着站点数量的增加而提高,当站点数量达到某个临界值时,融合降水的精度趋于稳定。Kriging-LSTM模型为准确获取山区小流域的降水资源提供了新思路。

基于IECSDE算法的PEMFC改进分数阶子空间辨识模型

为准确描述质子交换膜燃料电池(PEMFC)在其发电过程中的特性及变量影响关系,提出一种基于信息交流布谷鸟搜索差分进化(IECSDE)算法的改进分数阶子空间辨识方法来建立PEMFC分数阶模型。首先基于状态空间方程建立PEMFC模型,为了描述PEMFC的分数阶特性,将分数阶微分理论融入到模型中,引入Poisson滤波函数预处理实验数据,解决数据多阶不可导的问题,同时引入变步长记忆法处理分数阶微分时的权系数,提高子空间辨识精度。其次在辨识过程中的参数对于建模效果具有重大影响,因此基于IECSDE算法并对其进行优化,对布谷鸟搜索(CS)算法中的控制参数进行自适应处理,受到粒子群优化(PSO)算法的启发,改进随机游走方式提高收敛精度和速度,并引入差分进化(DE)算法与改进CS算法分别对种群进行优化,同时在寻优过程中进行信息交流提高种群的多样性和算法的鲁棒性。仿真结果表明,IECSDE算法的寻优能力在8种测试函数下比其他5种优化算法至少提升了10倍;通过对PEMFC测控平台收集到的实验数据进行模型辨识,所建立的模型将误差缩小到基于短记忆法的分数阶子空间辨识方法误差的20%,输出功率误差控制在0~0.1之间,输出电压误差控制在0~0.2之间,能够精准地模拟PEMFC发电过程。

基于CNN-LSTM的工业出水水质预测模型

工业废水含有多种污染物,提前预测工业废水水质从而快速对其进行相应处理具有重要意义。为此,研究提出了一种新的卷积神经网络(convolutional neural network,CNN)和长短期记忆网络(long short-term memory,LSTM)融合的工业废水水质污染物指标预测模型(CNN-LSTM)。为了更好地捕捉工业废水数据的时序性和动态性,模型设置了多个滑动窗口。使用CNN算法将时间序列数据进行高维特征提取,利用LSTM模型学习时间序列数据的时序特征,建立CNN-LSTM工业废水预测模型,并对废水水质中的化学需氧量(COD_(Cr))、氨氮、总磷(TP)3项指标进行预测分析。结果表明,与CNN和LSTM两个基准模型相比,CNN-LSTM预测模型的平均绝对值误差率(MAE)和均方误差率(MSE)均较小,预测效果较优。该模型能较好地实现对工业废水出水水质的准确预测,可为工业废水水质的在线监测和精准控制提供有效的、可行的技术支持和决策依据。

基于SVM-STL-LSTM的区域短期电力负荷预测研究

针对区域电力负荷的时间序列数据随机性强、预测精度低及单一模型的数据特征提取能力差等问题,提出了一种支持向量机(SVM)、STL时序分解法、长短期记忆神经网络(LSTM)组合的电力负荷预测模型。该模型利用SVM对时间序列的电力负荷数据进行初始预测,并通过STL时序分解法对残差序列进行时序分解,从而提高残差序列的稳定性,减小其随机性,最后用LSTM对SVM的预测误差进行修正。试验结果证明,该方法利用误差修正可有效处理随机性强的数据,有利于预测结果的稳定性,提高预测精度。

基于APF-LSTM-DDPG算法的移动机器人局部路径规划

针对深度强化学习算法存在训练时间长、收敛速度慢的问题,将深度确定性策略梯度(deep deterministic policy gradient,DDPG)算法和人工势场(artifical potential field,APF)法相融合,引入长短期记忆(long short-term memory,LSTM)神经网络结构,提出了APF-LSTM-DDPG算法。首先在DDPG算法中添加LSTM,通过记忆单元和遗忘单元将奖励较高的样本优先学习,使模型更快地收敛;其次引入人工势场法,设计相应势场函数,解决环境奖励稀疏的缺点,加速模型收敛进程;然后通过人工势场法修正算法的动作选择,缩短路径长度;最后在机器人仿真平台(robot operating system,ROS)中搭建不同障碍物环境对算法进行了仿真验证。仿真结果表明,APF-LSTM-DDPG算法在搭建的仿真环境中训练时平均奖励能够更快地稳定,提高了算法的成功率并减少了规划路径中的冗余。

基于ESMD-FE-AJSO-LSTM算法的水闸深基坑变形预测

水闸深基坑开挖变形具有明显的非线性和非稳定性特征,基于此,引入极点对称模态分解算法(extremepoint symmetric mode decomposition method,ESMD)对水闸深基坑开挖变形原型监测序列进行多模态分解,并基于模糊熵(fuzzy entropy, FE)理论对各分解分量进行模糊多模态相空间重构,从而有效甄别水闸基坑变形不同时间尺度有效物理特征。构建基于人工水母搜索算法(artificial jellyfish search optimizer,AJSO)优化的长短期记忆(long short-term memory,LSTM)人工神经网络模型,以重构后的各重构子序列为基础进行优化训练,并把训练后的各预测模态分量合并,实现对水闸基坑开挖变形动态预测和分析。以张家港市十一圩江边枢纽改建工程基坑开挖变形监测为例,采用上述方法对该枢纽工程基坑开挖过程变形进行预测和分析。结果表明:基于ESMD-FE-AJSOLSTM算法的水闸深基坑变形预测方法能够有效预测基坑开挖变形非线性特征,相比传统LSTM、循环神经网络(recurrent neural network, RNN)和支持向量机(support vector machine,SVM)等算法具有更高的预测精度和稳定性,为实现对基坑开挖安全性态实时科学诊断和分析提供技术参考。