Effect of cooling rates on dendrite spacings of directionally solidified DZ125 alloy under high thermal gradient

The dendrite morphologies and spacings of directionally solidified DZ125 superalloy were investigated under high thermal gradient about 500 K/cm. The results reveal that, with increasing cooling rate, both the spacings of primary and secondary dendrite arms decrease, and the dendrite morphologies transit from coarse to superfme dendrite. The secondary dendrite arms trend to be refined and be well developed, and the tertiary dendrite will occur. The predictions of the Kurz/Fisher model and the Hunt/Lu model accord basically with the experimental data for primary dendrite arm spacing. The regression equation of the primary dendrite arm spacings 21 and the cooling rate Vc is λ1 = 0.013 Vc^-0.32. The regression equation of the secondary dendrite arm spacing λ2 and the cooling rate Vc is λ2 = 0.00258 Vc^-0.31, which gives good agreement with the Feurer/Wunderlin model.

Effects of Temperature Gradient and Cooling Rate on the Formation of Methane Hydrates in Coarse Sand

Temperature gradient and cooling rate have an obvious effect on formation of methane hydrate. The process for formation of methane hydrate in coarse sand is monitored to understand the relationship between temperature gradient and cooling rate and nucleation, growth and distribution of methane hydrate by using the electrical resistivity method. The results show that the change of resistivity can better reflect the nucleation and growth and distribution of methane hydrate. Temperature gradient promotes the nucleation, formation, and formation rate of methane hydrate. At a temperature gradient of 0.11 ℃/cm, the rate of methane hydrate formation and saturation reaches a maximum. Cooling rate has little effect on the methane hydrate formation process. Judging from the outcome of final spatial distribution of methane hydrate, the cooling rate has an obvious but irregular effect in coarse sand. The effect of temperature gradient on distribution of methane hydrate in coarse sand is less than that of cooling rate. At a temperature gradient of 0.07 ℃/cm, methane hydrate is distributed uniformly in the sample. If the temperature gradient is higher or lower than this value, the hydrate is enriched in the upper layer of sample.

EFFICIENT GRADIENT DESCENT METHOD OFRBF NEURAL ENTWORKS WITHADAPTIVE LEARNING RATE

A new algorithm to exploit the learning rates of gradient descent method is presented, based on the second-order Taylor expansion of the error energy function with respect to learning rate, at some values decided by "award-punish" strategy. Detailed deduction of the algorithm applied to RBF networks is given. Simulation studies show that this algorithm can increase the rate of convergence and improve the performance of the gradient descent method.

分布式训练系统及其优化算法综述

人工智能利用各种优化技术从海量训练样本中学习关键特征或知识以提高解的质量,这对训练方法提出了更高要求.然而,传统单机训练无法满足存储与计算性能等方面的需求.因此,利用多个计算节点协同的分布式训练系统成为热点研究方向之一.本文首先阐述了单机训练面临的主要挑战.其次,分析了分布式训练系统亟需解决的三个关键问题.基于上述问题归纳了分布式训练系统的通用框架与四个核心组件.围绕各个组件涉及的技术,梳理了代表性研究成果.在此基础之上,总结了基于并行随机梯度下降算法的中心化与去中心化架构研究分支,并对各研究分支优化算法与应用进行综述.最后,提出了未来可能的研究方向.

流体加热融雪系统运行对道路结构温度分布特性影响分析

为了明确流体加热道路融雪系统运行对道路结构短期/长期温度分布特性的影响,利用室外足尺试验系统对比分析了普通水泥混凝土路面结构与流体加热道路融雪系统道路结构在温度、温变速率及温度梯度分布特性等方面的差异,阐明了系统运行对道路结构短期/长期温度分布特性的影响规律。结果表明:融雪系统运行初期,快速上升的流体温度对管壁-路面材料界面存在强烈的温度冲击作用;随着系统持续运行,管壁-路面材料界面的温变速率逐渐减小并趋于稳定,且小于普通路面最大温变速率;同时,融雪系统的运行增大了路面结构内部的温度梯度,改变了路面结构一定深度范围内的热量传递方向。融雪系统道路结构年周期温度分布的观测结果表明:融雪系统的运行可在较短的时间内释放大量热能,降低了道路结构内部热量的散失,有效延缓了季节性冰冻地区冬季道路结构温度的降低及冻结深度的发展。

聚丙烯酰胺对高黏油-水环状流流动特性的影响

针对稠油水环输送的水环流动改进问题,提出稠油流动边界层在聚丙烯酰胺(PAM)水环作用下的模拟实验方法,基于自主研发设计的室内油-水两相流可视化环道实验装置,采用500#白油模拟旅大稠油,实验研究了稠油-水环状流外环水相中添加PAM前后的流型特征与阻力特性,评价了PAM对水环润滑的协同减阻效果,分析了PAM对水环流动稳定性的作用机理。实验结果表明:PAM对稠油-水环状流的流型特征与阻力特性均有显著影响,使环状流的分布区域缩小,其缩减区域转变为分层流和塞状流,当其为分层流时减阻作用丧失;当油相表观流速为0.23~0.90m/s时,PAM减阻率随含水率增加先迅速增大后缓慢减小,在测试范围内最高达35%。研究结果可为稠油水环输送工艺优化提供理论指导和技术支持,也可为稠油管输流动改进提供新理念与新方法。

降雨、汇流和坡度对黑土浅沟坡面侵蚀影响的试验研究

[目的]为研究降雨、汇流和坡度对浅沟侵蚀的影响,实施精确的水土保持措施。[方法]通过室内模拟降雨和汇流试验,设计2个降雨强度(50,100 mm/h)和2个坡度(3°,7°),并在4种降雨强度和坡度组合下设置5个依次增加的汇流强度(15,30,45,60,75 L/min),分析降雨、汇流和坡度及其交互作用对浅沟坡面侵蚀的影响,明确浅沟侵蚀对坡面侵蚀的贡献。[结果](1)坡度和汇流强度对浅沟坡面侵蚀的影响均大于降雨强度。当降雨强度由50 mm/h增加到100 mm/h时,3°和7°对应的坡面侵蚀量分别增加52.3%~81.8%和29.4%~88.4%;而当坡度从3°增大到7°,50,100 mm/h降雨强度对应的坡面侵蚀量分别增加114.3%~395.5%和130.0%~320.9%;当汇流强度由15 L/min增加至75 L/min,坡面侵蚀量增加4.6~13.5倍。同时,汇流强度的增加加剧坡度对坡面侵蚀的作用,而减弱降雨强度对坡面侵蚀的影响。(2)降雨、汇流、坡度双因子和三因子交互作用对浅沟坡面侵蚀的影响以汇流强度-坡度交互作用和降雨强度-汇流强度-坡度三者的交互作用的影响最大,其次为降雨强度-汇流强度交互作用,而以降雨强度-坡度的交互作用的影响最小。(3)不同试验条件下,浅沟侵蚀量占坡面侵蚀量的比例平均在85%以上,且其比例随汇流强度和坡度的增加呈增大趋势。(4)坡面侵蚀速率与水流流速、径流剪切力和水流功率之间均表现为极显著正相关,与阻力系数之间呈显著负相关;且坡面侵蚀速率与水流功率相关关系最优。[结论]研究结果为基于侵蚀动力因子和地形因子及其交互作用的浅沟侵蚀预测模型开发、浅沟侵蚀贡献分离和浅沟侵蚀机理研究提供了科学依据。

电渗技术中水头压力与电压相互关系试验研究

电渗透防水技术是一种新型的防水技术,具有维修成本低、应用范围广等特点。利用电渗法,通过室内试验测定了不同电压、电势梯度下4种不同土样的排水速率、电极电势降、有效电势及其利用率,分析了水头压力与电压对电渗的影响。研究结果表明:电压越大,电极电势降越大,有效电势、有效电势利用率减小,有效电势利用率衰减越快;在同一水力梯度下,随着水力渗透系数增大,电渗排水速率也在逐渐增大;且当电势梯度与水力梯度的比值接近k值(水力渗透系数与电渗透系数的比值)时,电渗防水成功。研究可为电渗法的工程应用提供指导。

模拟降雨条件下坡度和降雨强度对桂南崩积体侵蚀产沙的影响

[目的]探究坡度和降雨强度对崩岗崩积体坡面侵蚀产沙特征的影响,为桂南地区崩积体水土流失预测及其防治提供科学依据。[方法]采用室内模拟降雨试验,研究不同坡度(15°,20°,25°)和降雨强度(60,90,120 mm/h)条件下崩积体坡面侵蚀产沙过程。[结果]各坡度条件下,60 mm/h降雨强度时,崩积体坡面侵蚀速率变化过程较为稳定,且侵蚀速率多低于5.0 g/(m^(2)·s),并呈现出一定的下降趋势;90~120 mm/h降雨强度下,侵蚀速率呈先波动增大后波动减小的变化趋势。降雨过程中侵蚀速率最大值随降雨强度和坡度的增大而增大,侵蚀速率最大值出现的时间集中在0~25 min内,且随坡度的增大有提前的趋势。次降雨产沙量(3.64~48.07 kg/m^(2))随降雨强度和坡度的增大而增大,与降雨强度和坡度之间的非线性回归结果(幂函数)优于线性回归结果。次降雨产沙量与坡度和降雨强度的交互项相关性最为显著,二者呈极显著线性函数关系。[结论]次降雨产沙量对降雨强度的敏感系数要高于对坡度的敏感系数,降雨强度对次降雨产沙量的影响强于坡度,崩积体侵蚀防治中建议做好坡面排水措施。

基于策略梯度的WRSN公平能量补充方案

无线可充电传感网络(WRSN)中节点能耗是动态变化的,在调度移动充电器(MC)为传感器充电时,估算节点的实时能耗来计算节点的剩余生存时间十分重要。针对WRSN的问题,提出一种基于策略梯度的公平能量补充策略(PFSS),利用过往的节点能量记录对节点实时能耗进行估算,采用seq2seq结合Attention的网络结构求解待充电序列,为节点充电获得一个充电权重来评估该次充电的质量,其由MC到节点间的距离、节点等待时间和节点剩余生存时间组成,利用策略梯度对网络进行训练,从而求得具有最优充电质量的充电序列近似解,仿真结果表明,与现有的算法相比,PFSS能够有效降低网络中传感器节点失效数量、MC的移动成本和充电延迟。

带有校正项的自适应梯度下降优化算法

基于批处理的随机梯度下降(SGD)优化算法通常用于训练卷积神经网络(CNNs),其性能的优劣直接影响神经网络收敛的速度.近年来,一些自适应梯度下降优化算法被提出,如Adam、Radam算法等.然而,这些优化算法既没有利用历史迭代的梯度范数,也没有利用随机子样本中梯度的二阶矩,这些导致自适应梯度下降优化算法收敛速度较慢,性能也不稳定.结合历史梯度范数和梯度的二阶矩,提出了一种新的自适应梯度下降优化算法normEve.通过模拟仿真实验,实验结果表明,提出的新算法在结合历史梯度范数和梯度二阶矩的情形下能有效地提高算法的收敛速度.通过实例验证新算法与Adam优化算法比较,新算法的测试准确率大于Adam优化算法,验证了新算法的优越性.

一种基于条件梯度的加速分布式在线学习算法

由于容易实施,基于投影梯度的分布式在线优化模型逐渐成为一种主流的在线学习方法.然而,在处理大数据应用时,投影步骤成为该方法的计算瓶颈.近年来,研究者提出了面向凸代价函数的分布式在线条件梯度算法,其悔界为O(T^(3/4)),其中T是一个时间范围.该算法存在两方面的问题,一是其悔界劣于公认的悔界O(/T);二是没有分析非凸代价函数的收敛性能,而实际应用中代价函数大部分是非凸函数.因此,提出一种基于条件梯度的加速分布式在线学习算法,使用Frank-Wolfe步骤替代投影步骤,避免昂贵的投影计算.文中证明当局部代价函数为凸函数时,所提算法达到公认的悔界O(/T);当局部代价函数为潜在非凸函数时,所提算法以速率O(/T)收敛到平稳点.最后,仿真实验验证了所提算法的性能与理论证明的结论.

基于斜坡渐变扰动的新型电力系统等效惯量评估

基于阶跃扰动的惯量评估方式依赖难以准确测量的初始频率变化率(RoCoF)与稀少的频率事件,而基于类噪声扰动的惯量评估方法无法评估电流源型虚拟惯量且对数据要求高。对此,提出基于可再生能源机组主动输出斜坡渐变扰动功率的惯量评估方法。建立含RoCoF与斜坡渐变扰动功率的等效惯量评估基础模型。考虑RoCoF噪声阶跃易导致等效惯量评估产生较大误差,推导双二阶广义积分锁相环中的q轴电压分量变化量与RoCoF的线性关系,并将其代入基础评估模型替代RoCoF。采用改进的非线性最小二乘拟合斜坡渐变扰动下q轴电压分量变化量与时间的非线性表达式,从拟合的表达式中提取系统等效惯量。在改进的EPRI-36中验证了所提评估方法相较于常规方法的优越性。

基于动量梯度下降的回声消除算法

针对极端环境话音系统下声学回波影响工作人员正常施工,且常规声学回声消除算法收敛速度慢的问题,提出一种基于动量梯度下降的基于l0范数的改进系数成比例归一化最小均方误差算法(L0⁃IPNLMS)。该算法将动量因子引入L0⁃IPNLMS算法中,解决在算法运行过程中梯度下降时梯度摆动幅度可能过大的问题,也提高了自适应滤波器的收敛速度,且残余回声下降明显,声学回波抑制效果更好。仿真实验表明,与L0⁃IPNLMS算法相比,新算法在模拟随机多音信号与真实语音信号输入时,均方误差(MSE)可以降低3.47 dB和3.69 dB,回波抑制比(ERLE)提高了3.46 dB和3.68 dB,在低信噪比情况下,使用新算法对真实语音信号进行回声消除,收敛速度高于L0⁃IPNLMS等算法,且收敛效果有明显改进。

基于动态学习率边界的隐私保护算法

深度学习优化算法在对数据进行训练时容易导致隐私泄露,卷积神经网络在进行隐私计算时会因为计算每个样本的梯度而带来巨大的内存开销,针对以上问题,提出一种结合混合重影剪裁的差分隐私动态学习率边界算法.将AdaBound优化算法与差分隐私相结合,缓解了算法在训练时的极端学习率和不稳定现象,减少了在反向传播过程中因为加入噪声而对模型收敛速度产生的影响.在卷积层上使用混合重影剪裁,简化了更新中对于梯度的直接计算所带来的开销成本,可以有效地训练差分隐私模型.最后,通过仿真实验,与其他经典的差分隐私算法进行对比,实验表明,算法实现了在相同隐私预算下更高的准确率,具有更优的性能,对模型的隐私保护效果更好.

Observation of single neutral atoms in a large-magnetic-gradient vapour-cell magneto-optical trap

Single caesium atoms in a large-magnetic-gradient vapour-cell magneto-optical trap have been identified. The trapping of individual atoms is marked by the steps in fluorescence signal corresponding to the capture or loss of single atoms. The typical magnetic gradient is about 29 mT/cm, which evidently reduces the capture rate of magneto-optical trap.

Vibration and wave propagation analysis of twisted micro-beam using strain gradient theory

In this research, vibration and wave propagation analysis of a twisted micro- beam on Pasternak foundation is investigated. The strain-displacement relations （kine-matic equations） are calculated by the displacement fields of the twisted micro-beam. The strain gradient theory （SGT） is used to implement the size dependent effect at micro-scale. Finally, using an energy method and Hamilton＇s principle, the governing equations of motion for the twisted micro-beam are derived. Natural frequencies and the wave prop- agation speed of the twisted micro-beam are calculated with an analytical method. Also, the natural frequency, the phase speed, the cut-off frequency, and the wave number of the twisted micro-beam are obtained by considering three material length scale parameters, the rate of twist angle, the thickness, the length of twisted micro-beam, and the elastic medium. The results of this work indicate that the phase speed in a twisted micro-beam increases with an increase in the rate of twist angle. Moreover, the wave number is in- versely related with the thickness of micro-beam. Meanwhile, it is directly related to the wave propagation frequency. Increasing the rate of twist angle causes the increase in the natural frequency especially with higher thickness. The effect of the twist angle rate on the group velocity is observed at a lower wave propagation frequency.

Dynamic analysis of fault rockburst based on gradient-dependent plasticity and energy criterion

Fault rockburst is treated as a strain localization problem under dynamicloading condition considering strain gradient and strain rate. As a kind of dynamic fracturephenomena, rockburst has characteristics of strain localization, which is considered as aone-dimensional shear problem subjected to normal compressive stress and tangential shear stress.The constitutive relation of rock material is bilinear (elastic and strain softening) and sensitiveto shear strain rate. The solutions proposed based on gradient-dependent plasticity show thatintense plastic strain is concentrated in fault band and the thickness of the band depends on thecharacteristic length of rock material. The post-peak stiffness of the fault band was determinedaccording to the constitutive parameters of rock material and shear strain rate. Fault bandundergoing strain softening and elastic rock mass outside the band constitute a system and theinstability criterion of the system was proposed based on energy theory. The criterion depends onthe constitutive relation of rock material, the structural size and the strain rate. The staticresult regardless of the strain rate is the special case of the present analytical solution. Highstrain rate can lead to instability of the system.

Simple Predicting Method for Fatigue Crack Growth Rate Based on Tensile Strength of Carbon Steel

Three types of fatigue tests for an annealed carbon steel containing carbon of 0.42%were carried out on smooth specimens and specimens with a small blind hole in order to investigate the fatigue crack growth law.A simple predicting method for crack growth rates has been proposed involving strengthσband the relation between cyclic stress and strain.The validity of proposed method has been confirmed by experiments on several carbon steels with different loadings.