【作者】 陈斌；

【导师】 万江文；

【作者基本信息】 北京邮电大学 ， 机械电子工程， 2009， 博士

【摘要】 天然气管网是现代城市的“生命线”之一。由于管道的劣化、老化、自然灾害和建筑施工破坏等原因,管道泄漏以及由此引发的爆炸事故时有发生,严重威胁着城市天然气供给体系的安全。利用无线传感器网络技术,可对城市天然气管网进行在线、实时安全监测,解决人工巡检效率低和现有泄漏诊断方法难以准确地识别小泄漏量、多泄漏源等问题。受测量噪声、传感器类型、网络节点数量和监测位置等因素的影响,传感网络中的检测信息表现出形式上的不确定性、多样性、数量的巨大性和关系的复杂性。为了能够及时、准确地识别出管道泄漏,需要解决以下问题:(a)传感器采集的原始泄漏信号的噪声剔除;(b)单节点上各类传感器检测信息的互补性处理;(c)处于不同监测位置的多个传感器节点诊断结果间联合决策。为此,从信息融合的角度,系统研究了管道泄漏监测网络内多源检测数据的处理方法。主要工作如下:(1)通过对管道泄漏监测网络的结构特点和检测数据特征分析,引入小波神经网络和D-S证据理论,建立了一种从数据级、特征级到决策级的层级式多源检测数据融合模型。在簇内成员节点处,采用小波神经网络方法对泄漏信号进行数据预处理和特征参数融合;在簇头处,利用改进的证据理论对多个节点的初始识别结果进行联合决策。(2)针对城市环境下传感器采集泄漏信号时受强噪声干扰的问题,优化选取Symlets小波基对声发射泄漏信号进行多层小波分解,采用启发式的小波阈值法剔除信号中的干扰噪声,然后从降噪信号中提取时频域内对泄漏敏感的特征参数。为了提高对泄漏源的定位精度,提出一种多传感器节点重复定位算法,该方法依据小波分解得到的单模态声发射信号的平均幅值对所有信号进行分组和配对,通过波形互相关分析得到每对信号定位出的泄漏点位置,最后进行加权平均。实验结果表明,多节点重复定位方法提高了对泄漏源的定位精度。(3)针对BP神经网络存在的收敛速度慢、识别率低且易收敛于局部最优解的不足,引入了蚁群算法全局优化网络的权值;为了保证网络同时具有较高的训练速度和识别准确率,采用“试优法”确定网络所需的隐含层神经元数,在此基础上建立了管道泄漏特征参数的蚁群神经网络融合结构,完成传感器节点对泄漏的初始识别。实验结果表明,相比于BP神经网络,蚁群神经网络能够极大地提高网络的训练速度,避免收敛于局部最小值,同时提高了泄漏识别的准确率。(4)考虑到处于不同监测位置的传感器节点关于泄漏事件的诊断结果间可能存在冲突,致使直接利用D-S及其修证据正组合规则可能得出与事实相悖的结论,为此,提出了一种基于可靠度和一致强度的冲突证据组合算法CECARCI。该方法依据源节点的可靠度对证据集进行预处理,引入证据的一致强度和基元支持度,合理地分配冲突和优化证据的组合次序。实验结果表明,CECARCI算法减弱了不可靠证据对组合结果造成的影响,提高了证据集对正确命题的聚焦度。(5)为了降低簇头决策管道是否发生泄漏时的风险,提出一种基于集合属性和优先度的D-S证据决策方法。该方法将证据决策问题分解成精细信度区间的构造和优先度比较两个层面。在构造层面上,引入集合的不确定性测度和焦元间的属性支持度,获取命题集合的精细信度区间值;在比较层面上,引入优先度评价不同命题的精细信度区间值,在优先度排序的基础上构建了证据决策模型。实验结果表明,该方法可充分利用信度区间所蕴含的信息,克服其它单点值证据决策方法所存在的误决策或不做决策的问题。(6)针对D-S证据理论无法处理管道泄漏监测网络内的模糊信息问题,提出一种基于距离测度的模糊证据理论扩展方法。该方法从模糊集合间距离的角度,确定模糊焦元对其它焦元的信任度、似真度函数的贡献程度,并建立了有效的模糊证据组合规则。实验结果表明,相比于其它的扩展方法,该方法能够从模糊焦元变化中获取更多的信息,避免模糊信任度函数对某些焦点元素变化不敏感的问题。通过对数据级预处理、异类泄漏特征参数融合和多节点联合决策等关键技术的研究,系统解决了管道泄漏监测网络的多源检测数据处理问题。基于小波神经网络和证据理论的层级式数据融合方法,可降低单传感器和单节点泄漏识别的不确定性,提高识别的准确率。更多还原

【Abstract】 Natural gas pipeline is one of the important lifelines in the city. Considering the problems of pipeline’s deterioration, aging, natural disasters, construction damage and so on, the leakage and the resulted explosion accidents occur frequently, which seriously threaten the security of gas supply system. However, it is difficult for the existed leakage detection methods to identify small and multi-sources leakage accurately. By using wireless sensor networks, it can achieve on-line and real-time security monitoring of gas supply system, and solve the problem of low efficiency of manual inspection.For the effects of measurement noise, sensor types, node numbers and monitoring location, the information shows the form of uncertainties, diversities, enormous quantity and complex relationship. In order to identify the pipeline leakage timely and accurately, such problems should be resolved: (a) removing noise from leakage signal, (b) dealing with the relevance of information from heterogeneous sensors, (c) making united decisions with multi-nodes’ diagnosis results.Therefore, as viewed from information fusion, this thesis systematically studies the method of multi-source data processing in pipeline monitoring network. The main work is provided as follows:Firstly, by analysis of the network structure and data characteristics, a hierarchical multi-source leakage monitoring data fusion model based on wavelet neural network and D-S evidence theory is established. At sensor node, wavelet neural network is used for data-level preprocessing and feature-level fusion. Then, at the cluster head, the united decision was made to the orignial multi-nodes’ diagnosis results using improved D-S evidence theory. Secondly, considering that the noise interference of leakage detecting signal is very strong under urban environment, the symlets wavelet and heuristic wavelet threshold method are selected for wavelet decomposition and signal-to-noise separation respectively. And the leakage sensitive characteristics of time-frequency domain can be extracted from wavelet decomposition signal. In order to improve the precision of leakage location, a multi-nodes location algorithm is proposed. The algorithm divides all signals into two groups according to the average amplitude of single-state acoustic emission signal. The leak position can be obtained from each pair of signal using waveform cross-correlation method, and then processes it with weighted average. The experiment result shows that the proposed method improves the precision of pipeline leakage location.Thirdly, BP neural network has the shortage of slow convergence, low recognition rate and easy to converge to local minimum value. So the ant colony algorithm is introduced to optimize network’s weights. To insure that the network has high training speed and recognition accuracy, the necessary number of hidden neurons is chosen by the excellent test method. Then the fusion structure of leakage characteristic parameters based on ant neural network is established for initial leak identification. Compared with BP neural network, the experiment result shows that the ant colony neural network model can not only improve the training speed, avoid network convergence to local optimal solution effectively, but also improve the accuracy rate of leakage recognition.Fourthly, considering that the diagnoisis results from different nodes may seriously conflict, it is difficult for the cluster head to make correct decisions using D-S or its modified combination rules. So a novel conflicting evidence combination algorithm based on reliability and coherence intensity (CECARCI) is proposed. The algorithm preprocesses the evidence set according to nodes’ reliability. By introducing the coherence intensity of evidence and support degree of base element, it obtains reasonable combination sequences and manages the conflict. The numerical example shows that proposed method not only decreases the effect of unreliable evidences on the fusion result, but also can obtain more reasonable results with good convergence.Fifthly, a novel evidence decision rule based on set attribute and preference degree is proposed for reducing the risk of cluster head’s decision-making. The method divides the decision problem into construction and evaluation level of belief interval. At the construction level, refined belief interval of base elements in power set is calculated based on the set’s uncertainty measure and support degrees between two focal elements. At the evaluation level, preference degree is used to estimate and rank the defined belief interval. So the decision-making model is constructed. Comparing with other endpoint value decision rules, the numerical example shows that the proposed method can make full use of interval information and avoid the wrong decision-making.Sixthly, considering the D-S evidence theory can’t deal with the fuzzy information in pipeline leakage monitoring network, novel fuzzy evidence reasoning method based on distance measure is proposed. In the method, as viewed from the point of distance between two fuzzy sets, the contribution of one focal element to the other elements’ belief or plausibility function is defined, and the fuzzy evidence combination rule is established. Compared with other methods, experimental results show that proposed generalization method can catch more information from focal elements’ change, and avoid the insensitivity problem of fuzzy belief function to focal elements’ change.In summary, by researching on the key technologies of data pre-processing, heterogeneous characteristic parameters fusion and multi-nodes united decision-making, this paper presents the systemic solution to the multi-source data processing in pipeline leakage monitoring network. The hierarchical data fusion method based on wavelet neural network and D-S evidence theory can reduce the identification uncertainty of single sensor and node, and improve the recognition accuracy greatly.更多还原