Internal interactions within the human circadian system: the masking effect

Summary In the realm of human circadian rhythms, the masking effect is defined as the change in the course of deep body temperature induced by changes in the degree of physical activity, or by the alteration between sleep and wake. This effect is particularly obvious during internal desynchronization where the rhythms of deep body temperature, and the sleep-wake sleep-wake sleep cycle — i.e. one of the masking factors — run with different periods.Every sleep onset is accompanied by a rapid drop, and wake onset by a rapid rise in deep body temperature, each one with an overshoot of about 50% of the steady state variations. When rhythms are calculated, with the dominant temperature period as the screening period, exclusively from data obtained during sleep episodes, on the one hand, and from those obtained exclusively during wake, on the other, two average cycles emerge: the sleep temperature curve and the wake temperature curve. Both run in parallel but are separated by the masking effcct. As derived from many experiments, the mean masking effect amounts to 0.28±0.06°C. The masking effect also depends to some extent on the phase of the temperature rhtthm; it is larger than average around the temperature maximum and during the descending phase of the temperature cycle, where the alertness commonly is highest and the probability to sleep, in general, and the REM sleep propensity, in particular, are smaller than average. This also can be interpreted to indicate that the sleep temperature curve is phase advanced relative to the wake temperature curve; this, on the average, by 0.9±0.3 h.If the individually determined amount of masking is added to the temperature data obtained during sleep, or substracted from the temperature data obtained during wake, a temperature curve emerges that can be though of as being purified of the masking effect. Analyses of this artificial curve allow estimation of that part of the internal interactions uninfluenced by the masking effect. On the average, about half of the amount of interaction between the rhythm of sleep-wake and that of deep body temperature is explained by the masking effect, whereas the other half is oscillatory interaction. Both types of interaction are inherent and inseparable parts of the circadian clock mechanism, as can be deduced from model considerations.     

Learning attention-based strategies to cooperate for multi-agent path finding

Multi-agent path finding (MAPF) is a challenging multi-agent systems problem where all agents are required to effectively reach their goals concurrently with not colliding with each other and avoiding obstacles. In MAPF, it is a challenge to effectively express the observation of agents, utilize historical information, and effectively communicate with neighbor agents. To tackle these issues, in this work, we proposed a well-designed model that utilizes the local states of nearby agents and outputs an optimal action for each agent to execute. We build the local observation encoder by using residual attention CNN to extract local observations and use the Transformer architecture to build an interaction layer to combine local observations of agents. With the purpose of overcoming the deficiency of success rate, we also designed a new evaluation index, namely extra time rate (ETR). The experimental results show that our model is superior to most previous models in terms of success rate and ETR. In addition, we also completed the ablation study on the model, and the effectiveness of each component of the model was proved.     

基于强化学习的电磁悬浮型磁浮列车悬浮控制

为了保证磁浮列车的安全、可靠运行,研究了悬浮系统在参数摄动条件下的悬浮控制问题。首先,对电磁悬浮（EMS）型磁浮列车的基本悬浮单元建模,给出了电流控制模型;然后,建立了悬浮系统的强化学习环境以及软演员-评论家（SAC）智能体,并设计了加速训练的奖励函数与“吸死”处理方案;最后,提出了基于强化学习的悬浮控制方法。与传统比例-积分-微分（PID）控制方法的对比结果表明,本方法具有更快的动态响应,在损失50%线圈匝数或磁极面积变化时具有更好的跟踪精度。     

基于深度学习的红外与可见光图像融合综述

图像融合技术是指从不同的源图像中提取并融合互补的信息,生成一幅信息量更丰富、对后续高级视觉任务提供足够支持的图像.红外与可见光图像融合（Infrared and Visible Image Fusion,IVIF）是图像融合领域的一个重要分支.近年来,深度学习技术在视觉计算领域表现出了良好的性能,尤其是基于自编码器、卷积神经网络、生成对抗网络等几种基于深度学习的IVIF技术得到了蓬勃发展.为此,对基于深度学习的IVIF算法的方法、数据集和评估指标等进行了总结和阐述;通过大量的实验,进行定性和定量的结果分析,对比了各类基于深度学习IVIF算法的性能;最后,讨论了该领域未来发展的一些前景和研究方向.     

联合实体边界检测的命名实体识别方法

针对传统命名实体识别方法无法有效利用实体边界信息的问题,提出一种联合实体边界检测的命名实体识别方法,即将实体边界检测作为辅助任务,增强模型对实体边界的判断能力,进而提升模型对实体的识别效果。首先,利用Bert预训练语言模型对原始文本进行特征嵌入获取词向量,并引入自注意力机制增强词对上下文信息的利用;其次,在命名实体识别任务的基础上,添加实体边界检测辅助任务,增强模型对实体边界的识别能力;再次,对比联合实体边界检测的命名实体识别方法与基线方法的有效性,并对测试结果进行消融实验;最后,进行样例分析,分析损失权重β对实体边界检测的影响。实验结果表明,在英文社交媒体数据集Twitter-2015上,联合实体边界检测的命名实体识别方法相较于基线模型取得了更高的精准率、召回率和F1值,其中F1值达到了73.57%;并且,边界检测辅助任务提升了基线方法的检测效果。所提方法能有效利用实体边界信息,从而获得更好的实体识别效果,促进了人机交互系统的发展,对自然语言处理下游任务有重要意义。     

数据驱动下交通异常行为的双层识别模型

为实现对车辆异常行为的准确识别,提高车辆行驶的安全性,提出了一种基于支持向量机(support vector machine, SVM)、监督学习与长短期记忆(long short term memory, LSTM)深度学习的交通异常驾驶行为双层识别模型。首先,对车辆轨迹数据筛除和滤波,构建异常行为数据集;其次,从异常行为轨迹特征中提取出特定异常行为的特征标签,并人为标定在训练集中;再次,构建SVM模型对训练集进行粗识别,基于SVM的二分法原理,从测试集中筛选出异常行为;最后,通过LSTM时间序列模型构建具体种类的异常行为模型,并通过深度学习的方法,从异常行为数据中细分为蛇形驾驶、急速变向、侧滑、大半径转弯、快速U型转弯、急刹车等具体的异常驾驶行为。本次实验选用下一代仿真(next generation simulation, NGSIM)数据中US-101高速公路和peachtree城市道路的数据集的轨迹数据验证SVM和LSTM双层识别模型的性能,包括均方根误差、识别准确率等。结果表明,构建的双层识别模型在第一层有98%的识别准确率,第二层有超过80%的识别准确率,可以较为准确地识...     

考虑应力集中的吊车梁多轴疲劳寿命预测

钢结构吊车梁在工业建筑中被广泛应用,应力集中的存在导致几何不连续位置在复杂应力下服役.借助有限元法分析吊车梁应力-应变状态,确定危险点位置,并提取应力、应变分量;以应变能密度作为损伤参量并以最大平面为临界面,基于能量准则建立临界面位置数值计算方法,结合有限元结果给出吊车梁临界面位置;考虑吊车梁在非对称载荷下服役,借助Goodman方程进行平均应力修正,并结合Q355D钢近似S-N曲线计算疲劳寿命.该方法考虑了应力集中处多轴应力对疲劳损伤的影响,可以为复杂应力下几何不连续钢构件的疲劳寿命评估提供新方法.     

扩展裂纹缺陷评定方法的优化研究

裂纹扩展的缺陷评定方法,越益受到了人们的重视,美国ＥＰＲＩ提出的“裂纹 推力图”方法,虽然准确合理,但方法偏于繁复,尚难于向工程界推广。清华大学力学 研究所曾先后提出几种评定方法,以期使评定过程得以简化。在此基础上,文章提出 “相容应力增量”法作为优化方法,它将使全部评定过程大为简化,从而为扩展裂纹的 缺陷评定方法向工程界推广创造便利条件。     

光伏模组积灰与阴影特性分析及识别方法

针对光伏模组积灰与阴影特性识别问题,详细分析了积灰和阴影的光伏特性曲线差异,揭示了阴影光伏曲线的拐点时变特性。提出由特性曲线的拐点数量及电流电压特性条件共同构成训练模型的输入特征量,基于CatBoost算法训练积灰和阴影识别模型。最后,利用光伏模组实测数据对CatBoost算法、ID3和GA-BP算法训练出的识别模型进行性能分析和对比测试,结果表明基于CatBoost训练出的识别模型输入量区分性强、诊断精度高,极具工程应用价值。     

基于压缩感知的在线多示例学习目标追踪

近年来提出的多示例学习算法在一定程度上能够克服模板漂移问题。然而,在线学习需要获取足够多的有用数据才能达到稳定的追踪效果,但是这却增加了算法的复杂度。为了解决这一问题,在压缩感知理论的基础上,运用随机观测的方法对多尺度图像特征进行降维,提取的这些低维特征中包含大量的有用信息。因此,我们提出的算法是先利用压缩感知理论提取目标特征之后,再使用在线多示例学习算法分类器对这些特征进行分类从而实现目标的稳定跟踪。通过对不同的图像序列进行实验,结果表明基于压缩感知的在线多示例学习算法对实时的目标追踪有很好的适应性。