基于Kubernetes的AI调度引擎平台

文中介绍了基于Kubernetes的AI调度引擎平台的设计与实现, 针对当前人工智能调度系统中存在的服务配置复杂, 集群中各节点计算资源利用率不均衡以及系统运维成本高等问题, 本文提出了基于Kubernetes实现容器调度和服务管理的解决方案. 结合AI调度引擎平台的需求, 从功能实现和平台架构等方面设计该平台的各个模块. 同时, 针对Kubernetes无法感知GPU资源的问题, 引入device plugin收集集群中每个节点上的GPU信息并上报给调度器. 此外, 针对Kubernetes调度策略中优选算法只考虑节点本身的资源使用率和均衡度, 未考虑不同类型的应用对节点资源的需求差异, 提出了基于皮尔逊相关系数 (Pearson correlation coefficient, PCC)的优选算法, 通过计算容器资源需求量与节点资源使用率的互补度来决定Pod的调度, 从而保证调度完成后各节点的资源均衡性.     

智能路灯控制系统缺失数据流关联修复仿真

为了避免智能路灯控制系统受缺失数据影响,研究了智能路灯控制系统缺失数据流关联修复仿真方法。选取嵌套窗口的流数据处理模型检测智能路灯控制系统的缺失数据流,嵌套窗口的流数据处理模型将滑动窗口分割为多个嵌套滑动窗口,滑动窗口过程中,利用Pearson相关系数确定智能路灯控制系统中相邻数据的相关性,依据获取的数据相关性重构相关图,利用相关图检测智能路灯控制系统缺失数据流。选取GMM算法划分存在缺失数据的智能路灯控制系统的数据集,选取EM插补算法关联修复完成初始数据集划分后的数据流。仿真测试结果表明,该方法可以有效修复智能路灯控制系统缺失数据流,不同数据缺失率情况下,均具有较高的缺失数据流关联修复精度。     

从测角后交最佳点位规律看规范规定之不足

本文分析了文献[1]所总结的测角后交两类公式在取得最佳点位时的规律,指出当前规范对此问题规定之不足.     

一种计算机专业课程考核方法

科学合理的考核方法对提高高校课程教学质量、检验教学效果起着重要的作用,课程的教学效果主要是通过考核来反映的。文章以"计算机组成原理与系统结构"课程为例,对高校计算机专业课程的考核模式进行探索,提出通过建立考核评价指标体系、调整指标权重系数,得到一种科学、公正的课程考核方法。     

An approach to quantify depth-resolved marine photochemical fluxes using remote sensing: Application to carbon monoxide (CO) photoproduction

This paper presents a model to calculate depth-resolved marine photochemical fluxes from remotely sensed ocean color and modeled solar irradiance. The basic approach uses three components: 1) below-sea-surface spectral downward scalar irradiance calculated from a radiative transfer model (STAR) and corrected for clouds using TOMS UV reflectivities; 2) surface-ocean spectral diffuse attenuation coefficients and absorption coefficients for chromophoric dissolved organic matter retrieved from SeaWiFS ocean color using the SeaUV/SeaUVc algorithms; and (3) spectral apparent quantum yield for the photochemical reaction considered. The output of the model is a photochemical rate profile, Ψ_PR(z), where z represents depth.We implemented the model for carbon monoxide (CO) photochemistry using an average apparent quantum yield spectrum and generated a monthly climatology of depth-resolved CO photoproduction rates in the global ocean. The climatology was used to compute global budgets and investigate the spatial and seasonal variabilities of CO photoproduction in the ocean. The model predicts a global CO photoproduction rate of about 41 TgC yr^− 1, in good agreement with other recent published estimates ranging from 30 to 84 TgC yr^− 1. The fate of photochemically derived CO and its role in global biogeochemical cycles remains uncertain however, with biological consumption and sea-air exchange competing for its removal in the surface ocean. Knowledge of the vertical distribution of CO photoproduction is critical in the quantification of the relative magnitudes of these sink mechanisms. The depth-resolution capabilities of this model, together with US Naval Research Laboratory climatology for mixed layer depths allowed further estimation that > 95% of the total water-column CO photoproduction occurs within the mixed layer on a global, yearly basis. Despite this compelling figure, the model also suggests significant spatio-temporal variability in the vertical distribution of CO photoproduction in the subtropical gyres, where up to 40% of water-column CO can be produced below the mixed layer during summertime.While the approach can be applied to other photochemical fluxes (e.g. DIC formation or DMS removal), accurate quantification of such processes with remote sensing will be limited until the mechanisms regulating observed oceanic variability in the apparent quantum yields are better understood. Minor modification to this model can also make it applicable for the determination of the effects of UV and visible solar radiation on sensitive biological systems.     

Two new control signal approaches for obtaining the MRAS-CDM and a real-time application

The coefficient diagram method (CDM) is one of the most effective control design methods. It creates control systems that are very stable and robust with responses without the overshoot and small settling time. Furthermore, all control parameters of the control systems are changed by varying some adjustment parameters in CDM depending on the demands. The model reference adaptive systems (MRAS) are the systems that follow and change the control parameters according to a given model reference system. There are several methods to combine the CDM with MRAS. One of these is to use the MRAS parameters as a gain of the CDM parameters. Another is to directly use the CDM parameters as the MRAS parameters. In the industrial applications, the system parameters can be changed frequently, but if the controller, by self-tuning, recalculates and develops its own parameters continuously, the system becomes more robust. Also, if the poles of the controlled systems approach the jw axis, the response of the closed-loop MRAS becomes more and more insufficient. In order to obtain better results, CDM is combined with a self-tuning model reference adaptive system. Systems controlled by a model reference adaptive controller give responses with small or without overshoot, have small settling times, and are more robust. Thus, in this paper, a hybrid combination of MRAS and CDM is developed and two different control structures of the control signal are investigated. The two methods are compared with MRAS and applied to real-time process control systems.     

室性早搏的多模板匹配自适应识别算法

为实现室性早搏的自动判别,提出一种基于多模板匹配的自适应识别算法。采用滤波提高数据信噪比,利用样本数据通过特定的自学习机制建立模板库,设计模板的压缩算法实现待测心搏与模板心搏宽度的对齐,采用模板队列匹配和模板库匹配相结合的匹配策略,设计自适应更新规则实时优化队列与库的结构,利用相关系数检测室早。使用MIT-BIH数据库进行测试,达到99.39%的灵敏度和98.79%的特异度。     

一种基于联合轮廓特征矢量的目标识别方法

针对基于形状特征进行目标识别的方法存在的不足,提出一种联合轮廓不变矩特征和轮廓几何特征的识别方法;针对不同的待识别目标,采集足够数量的训练样本,统计每一个轮廓特征的均值、标准差和变异系数;并据此对这些轮廓特征进行动态筛选和加权,建立起待识别目标的联合轮廓特征矢量模型.在线识别时,提取场景目标的联合轮廓特征矢量,对其进行...     

Fast, automatic and fine-grained tampered JPEG image detection via DCT coefficient analysis

The quick advance in image/video editing techniques has enabled people to synthesize realistic images/videos conveniently. Some legal issues may arise when a tampered image cannot be distinguished from a real one by visual examination. In this paper, we focus on JPEG images and propose detecting tampered images by examining the double quantization effect hidden among the discrete cosine transform (DCT) coefficients. To our knowledge, our approach is the only one to date that can automatically locate the tampered region, while it has several additional advantages: fine-grained detection at the scale of 8×8 DCT blocks, insensitivity to different kinds of forgery methods (such as alpha matting and inpainting, in addition to simple image cut/paste), the ability to work without fully decompressing the JPEG images, and the fast speed. Experimental results on JPEG images are promising.