【作者】 王晶；

【导师】 王宏健；

【作者基本信息】 哈尔滨工程大学 ， 模式识别与智能系统， 2013， 博士

【摘要】 导航是自主水下航行器（Autonomous Underwater Vehicle，AUV）安全、有效执行任务的前提和基础。惯性导航和航位推算等方法的导航误差随时间增长而累积，通常需要定期上浮至水面通过GPS实现导航位置校正，不适合于AUV长航时隐蔽作业。本论文致力于解决结构化环境中，部分已知或无先验信息情况下AUV自主导航问题，基于同步定位与构图（Simultaneous Localization and Mapping，SLAM），依靠AUV携带的环境感知传感器和位姿测量传感器实现位置估计与环境地图构建，对于AUV长航时安全、可靠作业具有重要的理论意义和实际应用价值。首先，设计SLAM研究的基本框架，建立环境地图模型、坐标系统、特征模型、传感器测量模型及AUV运动学模型，为后续SLAM研究奠定基础；然后，深入研究SLAM过程中特征提取问题，针对Hough变换特征提取方法中所存在的投票量大、提取效率低的问题，提出基于模糊自适应Hough变换的海洋环境特征提取方法，根据梯度方向信息，模糊化处理声呐数据点，采用极小极大模糊推理方法评判数据点隶属于直线特征的概率，自适应地选择参与投票的数据点并提取港口环境的线特征。与传统Hough变换方法相比，降低了投票次数，具有存储空间小、计算效率高、实用性强等优点；其次，深入研究SLAM的数据关联问题，针对数据关联过程中所存在的关联精度与计算效率之间的矛盾，提出灰色预测ICNN-JCBB快速切换数据关联方法，利用灰色理论对环境特征密度进行预测，通过设定密度阈值实现数据关联算法的快速切换选择，仿真实验结果表明，所提出算法提高了关联效率，保证了关联精度；再次，深入研究SLAM中AUV位置估计问题，针对AUV运动学模型与实际模型无法完全匹配且噪声统计特性不准确所导致的EKF-SLAM导航精度降低的问题，提出Sage-Husa自适应EKF-SLAM方法，将模型及噪声统计特性的不确定性虚拟化为系统的过程噪声，利用噪声统计特性估值器实时有效预测噪声统计特性，并对其进行校正；基于AUV海试数据的试验结果表明，选择不同的噪声初值对Sage-Husa自适应EKF-SLAM位置估计准确性影响较大；为了避免上述初值选取问题，基于Sage-Husa自适应EKF-SLAM和强跟踪EKF-SLAM提出组合自适应EKF-SLAM方法，设计残差收敛判据判断滤波估计发散，从而实施强跟踪EKF-SLAM估计AUV位置参数。基于海试数据的试验结果表明，组合自适应EKF-SLAM不受噪声初值选取的影响，可一定程度上保证AUV及地图中特征的位置估计精度；最后，深入研究基于FastSLAM的AUV位置估计问题，以解决EKF-SLAM中运动模型非线性、噪声非高斯的影响，针对FastSLAM中存在的粒子退化及粒子贫化现象，提出基于线性优化重采样的FastSLAM方法，在重采样过程中将复制的粒子与符合一定条件的被抛弃粒子进行线性组合，从产生的新粒子集合中选取权值增大者，减轻粒子的简单复制压力，一定程度上保留更多粒子携带的信息。基于海试数据的试验结果表明线性优化重采样FastSLAM可有效地降低粒子贫化现象，相对于标准FastSLAM方法，可在一定程度上提高AUV及特征的位置估计精度，但其估计结果仍受少量小权值粒子丢失的影响；针对粒子丢失问题，提出基于粒子权值方差缩减的FastSLAM方法，通过模拟退火算法的退温函数产生自适应指数渐消因子，通过小权值粒子权值的升高、大权值粒子权值的降低，实现粒子权值方差的缩减，提高有效粒子数。基于海试数据的试验结果表明，所提出的模拟退火方差缩减FastSLAM方法避免了粒子的退化，提高了AUV位置估计及地图构建精度。更多还原

【Abstract】 Navigation is the premise and basement of performing tasks for autonomous underwatervehicle (AUV). Error of inertial navigation and dead reckoning accumulates over time, soAUV must float to the suface of water periodically to correct the position by GPS which isnot fit for hidden mission. This paper concentrates on navigation of AUV with partial or nonepriori information in structured environment. AUV could achieve autonomous navigation andbuilds the environment map by environmental perception sensor, position and attitude sensors,which has a great theory significance and practice value to longtime and safe work of AUV.Firstly, the basic framework of SLAM was designed. Environment map model, featuremodel, coordinate system, kinematic model of AUV and measurement model of sensors wereestablished. The work above is the basis of the research for the following SLAM.Secondly, feature extraction of SLAM was reasearched deeply. Large memory capacityand low efficiency exist in traditional Hough transform. To solve the problem, featureextraction of marine environment was proposed based on fuzzy adaptive Hough transform.Sonar data was processed fuzzily with the information from gradient direction. Minimaxfuzzy reasoning was used to judge the probability that one data belongs to one line. Data thatparticipat in voting were selected adaptively. The line features of the port were extracted.Compared with traditional Hough transform, the method proposed has the advantage of smallmemory capacity and high efficiency and strong practicality.Thirdly, data association of SLAM was studied. To solve the contradiction betweenaccuracy and computational efficiency, ICNN-JCBB rapid swich data associatin based ongray prediction was designed. Gray theory was used to predict the feature density ofenvironment. A threshold was set to switch association method quickly. Simulations showthat the method proposed can improve data association efficiency with high accuracy.Fourthly, AUV position estimation method for SLAM was researched. In EKF SLAM,kinematic model of AUV can not match the actual model perfectly, and noise statisticalproperties are not accurate which makes the navigation accuracy of EKF-SLAM low.Sage-Husa adaptive EKF-SLAM was proposed to solve the problem above. The uncertaintyof model and statistics of noise were considered as process noise of system. Recursivefiltering was carried on based on observation data. With the estimatior of time-varying noise,the noise statistical properties were estimated and revised, which ruduces the impact of modelerror, and accuray of filter is improved. Experiment with trial data shows that different initial values had big influence on Sage-Husa adaptive EKF-SLAM. To sovle the prolem of initialnoise value, combined adaptive EKF-SLAM was designed based on Sage-Husa adaptiveEKF-SLAM and strong tracking EKF-SLAM. The convergence criterion of residual was usedto judge the estimation divergence. Simulation result with trial data shows that combinedadaptive EKF-SLAM isn’t affected by the initial noise value which can ensure the estimationaccuracy of the position of AUV and features to some extent.Finally, AUV position estimation method based on FastSLAM was researched to solvethe influence of nonlinear of model and non-Gaussian noise in EKF-SLAM. Particledegeneracy and impoverishment exist in FastSLAM. Linear optimization resamplingFastSLAM was designed. In the process of resamplng, the combinations of coied and lostparticles were done. Particles with bigger weight were selected from particles produced whichcan reduce the pressure of simple copy. The information carried by particles with smallweight can be reserved. Simulations with trial data show that linear optimization resamplingFastSLAM can reduce the particle impoverishment. Compared with standard FastSLAM, theposition estimated accuracy of AUV and features are enhanced. But the accuracy is stillinfluenced by the losing of little particles. Variance reduction of particle weights FastSLAMwas designed to avoid losing particle. An adaptive exponential fading factor was produced bycooling function of simulated annealing. With the weight rising of small weight particle andreducing of big weight particle, the variance of particles is reduced, and the effective particlenumber is improved. Simulation based on trial data shows that the method proposed can avoidparticle degeneracy, the accuracy of AUV navigation and map building were improved.更多还原