【作者】 姜延吉；

【导师】 黄凤岗；

【作者基本信息】 哈尔滨工程大学 ， 计算机应用技术， 2010， 博士

【摘要】 从上个世纪中后期开始,随着信息论、控制论、计算机技术、网络技术以及传感器技术等的快速发展,多传感数据融合技术在军事和民事领域都有着极其广泛的应用。如:复杂工业控制、机器人、自动目标识别、交通管制、海洋监测和管理、农业、遥感、医疗诊断、图像处理、模式识别等领域。与单传感器相比,运用多传感器数据融合技术在解决探测、跟踪和目标识别等问题方面,能够提高系统的可靠性和鲁棒性,增强数据的可信度,提高精度,扩展系统的时间、空间覆盖率,增加系统的实时性和信息利用率等。对多个传感器的数据多级别、多方面、多层次的处理所产生出的信息比单个传感器获得的信息更加有意义,为各种应用系统提供准确信息和决策依据。因此,数据融合服务已经成为传感器网络最重要的应用服务之一。本文针对多传感器数据融合技术的若干关键问题进行研究,主要包括:异构信息的统一描述和建模;非完备信息系统的空值属性估算与特征约简技术;分布式数据融合技术等。首先,在介绍了随机集基本理论及其与传统不确定信息融合方法D-S证据理论以及模糊集方法的相互转化关系的基础上,提出了一种利用随机集理论解决状态监测与故障诊断中异类信息表示和融合的方法。第一步,利用多传感器对影响状态的某一属性进行监测的情况下引入全局传感器的概念,并给出了随机集运算拟合全局传感器测量值的方法;第二步,将传感器获得信息与专家经验信息在随机集框架下进行统一描述,采用似然测度和随机集运算的方法分别得到传感器信息和专家主观意见的基本概率分配;第三步,在随机集框架下对传感器数据和专家意见进行统一融合,获得状态监测诊断结果。其次,引入了协同过滤的评分预测的知识,将协同过滤中处理稀疏数据的方法结合到空值预测中,以解决稀疏数据的问题;通过稀疏度控制估值算法的择优性,相似权值保证空值预测的准确性,并在此基础上提出了基于相似关系的改进空值估算方法,在一定程度上解决了稀疏数据条件下的空值插补不准确问题。再次,提出基于存在型空值插补的限制容差关系,该关系模型能够同时处理含有存在型空值和不存在型空值的非完备信息系统,并引入知识粒度的相关概念,给出了针对该关系模型的属性重要度计算方法和特征约简算法;通过实验对该模型的时间复杂性进行了分析,并通过与其他关系模型的对比验证了模型的有效性。第四,提出一种基于预测可信度的分布式D-S证据理论融合方法,首先,在已有的证据源可信度系数算法基础上提出了预测可信度系数的概念并给出了具体计算公式,并且介绍了训练可信度系数平衡因子的方法;其次,将给出的预测可信度系数的计算公式与原D-S证据理论合成规则进行结合,在保证原合成规则的所有性质的基础上加入了预测可信度系数来解决证据冲突问题;并通过仿真实现证明了算法的有效性。多传感器数据融合技术的研究目前还存在着很多关键性问题,本文针对其中的部分问题进行了研究和探讨,提出的基于随机集理论的异构信息统一表示和建模的方法,为实现异构多源信息融合提供了前提;基于存在型空值插补的特征约简技术可以剔除数据中的冗余信息,有效降低融合的时空复杂度;基于预测可信度的分布式D-S证据理论融合方法在保证融合结果的前提下提高融合了的效率。因此本文的研究具有重要理论和应用价值。更多还原

【Abstract】 With the rapid development of Control Theory, Information Theory, Microelectronic Technology, Computer Technology, Network Technology and Sensor Technology since 1950s, multi-sensor data fusion technology has an extremely wide use in the domain of military and civilian, for example: complex industrial control, robotics, automatic target recognition, traffic control, marine monitoring and management, agriculture, remote sensing, medical diagnostics, image processing, pattern recognition, etc.Compared with the single sensor, using multi-sensor data fusion technology to solve the problems of exploration, tracking and target recognition, can improve system reliability and robustness, enhance data reliability, improve accuracy, extend the system time and space coverage, and increase the system’s real-time performance and information utilization. Through the multi-level, multifaceted and multi-level processing of the data from multiple sensors, we can obtain moremeaningful information which can not be provided by a single sensor and accurate information, decision-making basis for a variety of application system can be provided. Therefore, data fusion service has become one of the most important applications in the sensor network.In this dissertation, a number of key issues of multi-sensor data fusion technology have been studied, including: a unified description and modeling of heterogeneous information; null value attributes estimating and feature reduction technology of incomplete information system; distributed data fusion technology, etc.Firstly,the basic theory of random sets and the relationship of mutual transformation between random sets and traditional methods of uncertain information fusion including D-S evidence theory and fuzzy sets are introduced, then a method that using random sets theory to descript heterogeneous information in condition monitoring and fault diagnosis is proposed. The first step, the concept of the global sensor is introduced into the case of using multi-sensor to monitor one of factors that affect the condition, and then obtain the value of theglobal sensor by curve fitting; the second step, using the random sets theory to descript the information provided by the sensors and the experts, and the plausibility measure random sets probability to descript basic probability distribution; The third step, do the fusion of sensors data and the experiences of experts in the framework of random sets, get the final result.Secondly, the knowledge of the rating prediction in the collaborative filtering is introduced, the method that solve the sparse data problem in collaborative filtering is combined with null-value estimation in order to solve the sparse data problem; using sparsity to control the selectivity of the estimating algorithm and the similarity weight ensure accuracy of the null-value estimation, and then an improved method based on similarity relation of null-value estimation is proposed; finally, the capability of the algorithm is validated and analysed through classic data sets and real data sets.Thirdly, a limited tolerance relation based on existence null-values interpolated is proposed. The relationship model can deal with the incomplete information system including both existence null-values and inexistence null-value. The concept of knowledge granularity is introduced, and the calculation method of attribute importance and feature reduction algorithm used in this relationship model is given; Time complexity of the model is analyzed by experiments; the validity of the model is verified by comparing with other relational model.Finally, a distributed D-S evidence theory data fusion method based on the predicted of credibility is proposed. First, the conception of predicted of credibility coefficient and formula are given based on the existing evidence source credibility coefficient algorithm, and the method of training balancing factor of the credibility coefficient is introduced; Second, combining the formula with original D-S evidence theory in order to solve the evidence conflict problem without changing the properties of original combination rules; Finally, the effectiveness of the algorithm is proved by the simulation.At present, there are still a lot of key issues in the research of multi-sensor data fusion technology. In this dissertation, some of these issues are studied and discussed, and the method based on the random sets theory, solving the problem that the unified description and modeling of heterogeneous information is proposed. This method provides a prerequisite for heterogeneous multi-source data fusion; the feature reduction algorithm based on existential null-values interpolated can remove redundant information in the data and effectively reduce the time complexity and space complexity; The distributed D-S evidence theory data fusion method based on the predicted of credibility can improve the efficiency of data fusion under the premise of ensuring the results. Therefore, the study in this dissertation has important theoretical value and application value.更多还原