【作者】 朱培逸；

【导师】 徐保国；

【作者基本信息】 江南大学 ， 控制理论与控制工程， 2013， 博士

【摘要】 随着各种物理、生化类先进传感器的发展，特别是某些传感器技术还不是很成熟，使得测量的信息呈现量大、多维、动态、不确定、不完整和时空变异等特点，使得系统对后续信息处理提出了更高的要求。由于传统的数据分析方法已经很难适应要求，因此如何将这些传感器测量数据进行分析，最大限度的降低单个传感器所提供信息的误差以及多传感器所提供信息的不确定性，是亟待解决的问题。本文主要对不确定信息的融合方法进行了研究，获得了一些有意义的成果，其主要研究内容如下：1．针对土壤参数野外测量时存在多种因素干扰问题，采用粗糙集理论对多传感器采集的原始数据进行属性约简和目标约简，减少采样时的噪声和冗余，再根据预处理的信息结构构建支持向量机的信息预测系统，解决小样本和不确定条件下的多传感器数据融合问题，为了获得最优的融合精度，采用粒子群算法来优化融合参数。2．针对DS证据理论处理不确定信息存在无法解决冲突信息的缺陷，提出一种基于加权证据距离的数据融合来提高融合效率。首先采用加权的DS理论对证据进行处理，然后以各加权证据与期望证据构成修正的证据距离。其次采用最优化的思想，将修正的证据距离作为改进粒子群优化(Particle Swarm Optimization, PSO)算法优化的目标函数，在此基础上获得证据距离最小时的各个证据权值，最后采用DS组合方法进行加权合成，得出融合结果。3．针对递推加权最小二乘融合存在算法鲁棒性较差等缺陷，在传统最小二乘算法研究的基础上，提出一种基于自适应权值的鲁棒加权最小二乘算法。按照不同的权值设计准则，得出所需的测量估计值，并且采用鲁棒权因子和学习因子的加权原则，充分利用测量数据中冗余信息，进一步提高冗余系统测量数据的估计精度。4．针对传感器管理在数据融合系统中地位与作用，提出一种智能化的传感器管理算法，通过进化算法自下而上地评价每一个控制决策使得系统的不确定性最小。首先为了取得最优的传感器测量参数将采用粒子群算法在参数空间中进行自下而上地搜索，再采用贝叶斯网络直接设计系统目标与全局性能指标之间的关系，从而使得算法更适合评估当前环境形势以及正确选择基于系统条件和传感器输入的合适性能指标，最后建立全局性能指标，并采用粒子群算法搜索最佳选择的传感器参数，最终实现对传感器资源的优化配置。更多还原

【Abstract】 With the development of various kinds of advanced physical, chemical and biologicalsensors, and some sensing technologies are not mature, so that measurement information wasbeing quantities, multidimensional, dynamic, uncertain, incomplete and spatiotemporalvariation, and thus these technologies put forward the higher request for the cominginformation processing. It is difficult to meet the requirement with the traditional data analysis.Therefore, it is an urgent problem for processing these sensor measurement data that reducethe error from single sensor or the uncertainty from multiple sensors. In this dissertation, anumber of key issues of uncertain information fusion technology were studied, and got somesignificant results. The main content and contributions of this dissertation are summarized asfollows:1. In order to solve the problem of the soil properties measurement error because ofmultiple factors interference, the dissertation adopts Rough Sets for reduction of the attributeset and target set for the raw data from various sensors, thus the noise and redundancy will bereduced in sampling. Then constructs information prediction system of Support VectorMachine according to the preprocessing information structure, and solves the problem ofmultisensor data fusion in the situation of small sample and uncertainty. In order to get theoptimal fusion accuracy, it uses Particle Swarm Optimization (PSO) for fusion parameters. Tomake operate faster, and also to increase the accuracy of the fusion, a feature selection processwith PSO is used in this dissertation to optimize the fusion accuracy by its the superiority ofoptimal search ability.2. Although DS evidence theory has many advantages, but the uncertainty formaldescription method based on DS theory has some defects. To reduce the defects and enhancethe fusion efficiency, a new data fusion algorithm based on weighed evidence synthetictechnique is presented. Firstly, the concept of weight of sensor evidence itself and evidencedistance based on a quantification of the similarity between sets to acquire the reliabilityweight of the relationship between evidences is set up. Then considering the disadvantages ofthe improved DS theory, a best method of obtaining evidence weight value is presented by animproved PSO. Finally, we use the improved PSO to acquire the reliability weight of therelationship between evidences to modify DS theory.3. Considering uncertainty measurement and the principle of Least Squares, recursiveweighted least squares fusion algorithm is deduced, and sensor weight coefficient is closerelationship with measuring variance. However, weight coefficient often is deduced frommeasuring variance, due to measurement attribute weight is different, and the dimensionalinconsistencies, then it is difficult to realize the effective measure for the sensor precision. Meanwhile, the sensor itself variance parameter is assigned or given in advance according tothe experience. Thus the anti-interference ability of fusion results is weak, and the robustnessof the algorithms is low. The robust least squares estimation approach proposed in thisdissertation employs an adaptive weighted viewpoint. For the different weight design criteria,the measurement estimation value is got. Due to introduce robust weight factor and learningfactor, we take full advantages of redundancy information of measurement data, and improvethe estimated precision of redundant system measurement data.4. Sensor management is very essential to the multisensor system in the targetassignment, and become the core element of the information fusion system to improve itsperformance. This dissertation describes an intelligent sensor management paradigm thatminimizes system uncertainties with an evolutionary algorithm that evaluates each controldecision from the bottom up. PSO algorithm works from the bottom up searching theparameter space for the best sensor measurement setting. The Bayesian network is moresuited towards assessing the situation and selecting proper performance goals based onconditions and operator inputs. Bayesian networks allow the designer to direct therelationship between the system goals and global performance values providing the flexibilityto define these performance values to fit any situation. Once a relationship between the globalperformance values is well-defined, PSO algorithm searches to optimally select the sensorparameters that achieve the desired global performance, and then the problem ofautomatically allocating the resources of multi-sensor is solved.更多还原