基于主动学习的离群点集成挖掘方法研究

离群点检测任务通常缺少可用的标注数据，且离群数据只占整个数据集的很小一部分，相较于其他的数据挖掘任务，离群点检测的难度较大，尚没有单一的算法适合于所有的场景。因此，结合多样性模型集成和主动学习思想，提出了一种基于主动学习的离群点集成检测方法OMAL（Outlier Mining based on Active Learning）。在主动学习框架指导下，根据各种基学习器的对比分析，选择了基于统计的、基于相似性的、基于子空间划分的三个无监督模型作为基学习器。将各基学习器评判的处于离群和正常边界的数据整合后呈现给人类专家进行标注，以最大化人类专家反馈的信息量；从标注的数据集和各基学习器投票产生的数据集中抽样，基于GBM（Gradient BoostingMachine）训练一个有监督二元分类模型，并将该模型应用于全数据集，得出最终的挖掘结果。实验表明，提出方法的AUC有了较为明显的提升，且具有良好的运行效率，具备较好的实用价值。     

面向连续查询的敏感语义位置隐私保护方案

针对连续查询位置服务中构造匿名区域未考虑语义位置信息导致敏感隐私泄露问题，通过设计[(K,θ)]-隐私模型，提出一种路网环境下面向连续查询的敏感语义位置隐私保护方案。该方案利用Voronoi图将城市路网预先划分为独立的Voronoi单元，依据用户的移动路径和移动速度，选择具有相似特性的其他[K-1]个用户，构建匿名用户集；利用匿名用户集用户设定的敏感语义位置类型和语义安全阈值，以及用户所处语义位置的Voronoi单元，构建满足[(K,θ)]-隐私模型的语义安全匿名区域，可以同时防止连续查询追踪攻击和语义推断攻击。实验结果表明，与SCPA算法相比，该方案在隐私保护程度上提升约15%，系统开销上降低约20%。     

基于条件信息卷积生成对抗网络的图像识别

传统的图像识别方法需要大量有标签样本进行训练，且模型训练难以达到稳定。针对这些问题，结合条件生成网络和信息最大化生成网络的结构优势建立了条件信息卷积生成网络（C-Info-DCGAN）。模型增加图像的类别信息和潜在信息作为输入数据，然后利用Q网络去更好地发挥类别信息和潜在信息对训练的引导作用，并且利用深度卷积网络来加强对图像特征的提取能力。实验结果表明，该方法能够加快模型训练收敛速度，并有效提高图像识别的准确率。     

结合预训练模型和语言知识库的文本匹配方法

针对文本匹配任务,该文提出一种大规模预训练模型融合外部语言知识库的方法。该方法在大规模预训练模型的基础上,通过生成基于WordNet的同义—反义词汇知识学习任务和词组—搭配知识学习任务引入外部语言学知识。进而,与MT-DNN多任务学习模型进行联合训练,以进一步提高模型性能。最后利用文本匹配标注数据进行微调。在MRPC和QQP两个公开数据集的实验结果显示,该方法可以在大规模预训练模型和微调的框架基础上,通过引入外部语言知识进行联合训练有效提升文本匹配性能。     

基于语言特征自动获取的反问句识别方法

反问句是以疑问的形式表达强烈情感的修辞方式,对其有效识别可为自然语言处理中的情感分析任务提供技术支持。该文提出了一种基于语言特征自动获取的反问句识别方法。首先,利用标签注意机制,建立了一个数据驱动的特征抽取模型,用于获取与任务相关的词汇、句法结构、符号标记和话题等语言特征。其次,利用Bi-LSTM模型分别对句子和语言特征进行表示,两者的交互注意被用于获取句子的各个词和符号的注意力权重向量。该权重向量作用于句子的表示,用于构建一个强化语言特征的反问句识别模型。在中文微博数据集上的实验结果表明,提出的方法与之前的工作相比,反问句识别性能有显著提升。     

Zone model predictive control for pressure management of water distribution network

In the water distribution network (WDN), although the water demand of the node is changing constantly, the water quantity and water pressure of the node need to be met at each moment. To realize energy saving and consumption reduction, it is proposed to control the nodal water head of WDN in an appropriate narrow range. The frequent large fluctuation of the water demand, which may lead the water pressure exceed the expected range, increases the difficulty of the zone control. To realize optimized WND control, a novel zone predictive control is proposed, where two switching cases are considered. The switching condition is whether there are feasible solutions to keep the pressure within the expected region over the prediction horizon. When the condition is satisfied, the controller minimizes the variation of inputs with constraints of pressure range for ensuring the tank level staying within the expected zone and obtaining optimal economic cost. When the current pressure is out of the expected region or the condition is not satisfied due to the large variation of water demand, a reference trajectory of outputs is introduced, which is combined with the inputs as an optimization variable, and the constraints of expected zone are moved from the output to the introduced reference trajectory. Through minimizing the distance between reference trajectory and output, the controller will keep the tank level from deviating too far from the expected zone and will drive the tank level rapidly into the expected zone once the tank level exceeds the expected range. An application of the proposed zone MPC to WDN in Shinan District of Shanghai is given to illustrate its effectiveness.     

2种灰色模型在矿区地表沉陷预计中的适用性

目前矿区地表单点沉陷动态预计方法主要基于传统的水准测量数据,监测方法单一,成本高,观测点易破坏,不能保证地表形变信息的实时性,且采用灰色模型进行地表沉陷预计时只针对单一模型的应用,没有结合模型自身特点分析其适用性。以袁店二矿7221工作面为试验区域,采用合成孔径雷达差分干涉测量技术监测矿区地表沉陷量,分别建立了描述沉陷量与时间关系的GM(1,1)与灰色Verhulst模型进行地表沉陷量预计,实现了矿区地表沉陷监测与动态预计一体化。通过比较、分析GM(1,1)与灰色Verhulst模型对地表沉陷量的拟合及预计结果,得出了2种灰色模型在矿区地表沉陷预计中的适用性:在矿区开采沉陷开始至活跃前期,若地表单点沉陷量曲线呈近似单峰型,则宜采用GM(1,1)进行短期预计;当矿区地表沉陷进入衰退阶段,单点沉陷量曲线呈平底饱和状态,则宜采用灰色Verhulst模型进行中长期预计。     

海上稠油油田蒸汽吞吐产量确定新方法

针对目前蒸汽吞吐产量预测模型假设条件简单、普适性差等问题,一般采用测试法和类比法综合确定海上稠油油田蒸汽吞吐初期产量。由于目前海上油田通常只开展常规测试,无法直接获得热采开发初期产量。笔者提出海上稠油油田蒸汽吞吐初期产量确定新方法,建立蒸汽吞吐相对于常规开发的初期产量倍数预测模型,通过蒸汽吞吐产量倍数,将常规测试确定的产量转化为蒸汽吞吐产量。研究表明,蒸汽吞吐初期产量倍数主要受储集层渗透率、原油黏度、注入强度、蒸汽干度等因素影响,利用正交试验设计和多元回归等方法,建立海上稠油油田蒸汽吞吐初期产量倍数与油藏地质参数及注入参数之间的非线性预测模型,该模型经实际生产数据验证,预测误差小于5%,可靠性高,能够为海上稠油油田蒸汽吞吐初期产量的确定提供依据。     

Performance-based modelling of long-term deterioration to support rehabilitation and investment decisions in drinking water distribution systems

Managing the urban drinking water system in the long term in order to maintain system performance can be challenging due to the difficulty of modelling future deterioration of the networks. This paper establishes a methodology for cohort survival models where historical (empirical) data on decommissioning ages of pipes are used to calibrate survival functions of pipe cohorts according to service level targets. The benefit of the approach is that remaining useful life of pipes, future renewal rates and investment needs can be governed by a required level of service in the network. A case study shows how the methodology can be applied to a cohort of drinking water pipes to create a ‘calibration curve’, which is a survival function calibrated with empirical data.     

A comparative study for consistency-based decision making with interval multiplicative preference relations

ABSTRACT

To ensure the reasonable application and perfect the theory of decision making with interval multiplicative preference relations (IMPRs), this paper continues to discuss decision making with IMPRs. After reviewing previous consistency concepts for IMPRs, we find that Krej?í’s consistency concept is more flexible and natural than others. However, it is insufficient to address IMPRs only using this concept. Considering this fact, this paper researches inconsistent and incomplete IMPRs that are usually encountered. First, programming models for addressing inconsistent and incomplete IMPRs are constructed. Then, this paper studies the consensus of individual IMPRs and defines a consensus index using the defined correlation coefficient. When the consensus requirement does not satisfy requirement, a programming model for improving consensus level is built, which can ensure the consistency. Subsequently, a procedure for group decision making with IMPRs is offered, and associated examples are provided to specifically show the application of main theoretical results.