SIA OpenIR  > 光电信息技术研究室
基于Agent的海上搜救任务规划方法研究
其他题名Maritime search and rescue planning based on Agent
吕进锋1,2
导师赵怀慈
关键词海上搜救 粒子群 析因 启发式 流形学习
学位专业模式识别与智能系统
学位名称博士
2018-05-18
学位授予单位中国科学院沈阳自动化研究所
学位授予地点沈阳
作者部门光电信息技术研究室
摘要本文首先对现有海上搜救方法进行研究总结,在分析现有搜救技术难点的基础上,提出一种基于Agent的海上搜救任务规划方法框架;其次针对不同类型的搜救任务,提出相应的启发式优化方法进行任务规划;最后提出一种基于流形学习的海上搜救目标感知方法,旨在对搜救任务中获取的航拍海面图像进行快速有效的检测。本文主要贡献如下:(1)提出一种基于Agent的海上搜救任务规划方法框架,该框架采用混合式MAS结构,包含指挥Agent、船舶Agent、航空器Agent、执行Agent四种类型的Agent,对不同Agent的属性、功能及Agent之间的关系、通信方式进行定义。指挥Agent利用蒙特卡罗方法生成目标位置信息;指挥Agent、船舶Agent、航空器Agent采用合同网招标法确定参与搜救任务的Agent;指挥Agent采用启发式优化算法为船舶Agent、航空器Agent等制定搜救方案;船舶Agent、航空器Agent等采用基于流形学习的搜救目标感知方法对航拍海面图像进行搜救目标感知。该搜救任务规划方法框架可以有效减少数据冗余,提高搜救任务中信息传递效率。(2)针对静态概率分布特征的搜救任务,提出一种记忆库粒子群算法。记忆库粒子群算法首先利用组合优化算法的思想,在构建记忆库的基础上采用概率组合的方式生成新的备选搜救方案;其次采用自适应网格法更新记忆库,基于此对解空间进行有效的全局搜索;最后当记忆库中的备选搜救方案所在网格稳定,利用种群围绕记忆库中的备选搜救方案进行初始化,采用改进的粒子群迭代策略,根据适应度值将粒子分组,使粒子在组内围绕较好质量的解进行有效的局部搜索,完成一定次数的组内更新后将种群重新分组,利用此策略完成对解空间的搜索。对协作任务规划,记忆库粒子群算法组合多个差异较大的备选搜救方案生成初始协作搜救方案,在此基础上利用种群搜索最优协作方案。实验结果表明相较其余几种优化算法,记忆库粒子群算法可生成具有更高任务成功率的搜救计划。(3)针对动态概率分布特征的搜救任务,提出一种析因启发式算法。析因启发式算法引入析因实验设计思想,通过利用随机化及区组化策略,对任一参数设定多个水平,随机选择其余参数在不同水平的组合,得到相应的适应度值,获取各个参数的适应度曲线并形成区组。通过分析区组内参数变化对适应度值的影响以及参数间的交互作用,设定种群初始化策略及每个参数的更新策略。种群中个体按距离进行分组,形成多个多样性较好的族群。族群内部更新过程中,针对交互作用明显的参数侧重于全局搜索,针对交互作用不明显的参数侧重于局部搜索。族群内部完成一定程度的更新后重新分组。对协作搜救任务,析因启发算法组合多个差异性较大的备选搜救方案生成初始协作搜救方案,种群中个体按小聚类结构进行通信并更新。实验结果表明该策略可稳定为动态概率分布特征的搜救任务生成高质量的搜救计划。(4)针对海上搜救中利用航拍海面图像感知检测目标问题,提出一种基于流形学习的搜救目标感知方法。在获取包含目标及不包含目标两类带标签的样本图像数据基础上,利用线性判别式分析方法将高维图像数据映射到低维空间,使类间距离最大,类内距离最小,获得映射矩阵和分类器;对船舶Agent/航空器Agent采集的航拍图像首先利用Canny算子和Hough变换提取海天线,其次利用搜救目标(落水人员、救生艇)与海面背景像素差提取目标可能存在区域,最后利用映射矩阵对提取的图像块进行映射分类,判断图像是否包含搜救目标。根据感知结果利用贝叶斯后验概率更新搜救目标位置信息,预测目标运动轨迹。实验结果表明,基于流形学习的搜救目标感知方法可在降低虚警率要求的前提下,有效检测海上搜救航拍图像,提高搜救任务效率。
其他摘要This dissertation makes a review on the existing maritime search and rescue methods. After analyzing the technical difficulties in maritime search and rescue, this dissertation proposes a maritime search and rescue planning method framework at first. Based on the analysis of different kinds of SAR problems, then corresponding heuristic optimization algorithms are proposed to generate SAR plans. In order to make rapid and effective sea image detection, a SAR target sensing method based on manifold learning is proposed at last. The major contributions of the dissertation include: (1) A SAR planning method framework is proposed. This framework uses a hybrid multi Agents structure, which includes four types of Agents:Command Agent, Ship Agent, Plane Agent and Unmanned Agent. This dissertation defines the properties of each type of Agent, the relationships between different Agents, and the communication modes in the framework. Command Agent generates the location information of each SAR target based on Monte Carlo method. Command Agent, Ship Agent and Plane Agent use contract bidding method to determine the Agents participating in each task. By using heuristic algorithms, Command Agent generates SAR plans for Ship Agent, Plane Agent and Unmanned Agent. Ship Agent, Plane Agent and Unmanned Agent employ the sensing method based on manifold learning to detect sea images. The SAR planning framework can reduce the data redundancy and increase the efficiency of information transmission in SAR. (2) Aiming at maritime static probability distribution SAR planning, a memory bank particle swarm optimization (MBPSO) is proposed. MBPSO uses the idea of combination optimization at first. It constructs a memory bank and generates new candidate solutions based on memory consideration and random selection. MBPSO updates its memory bank based on a self-adaptive grid method, based on which the population can make effective global search. When the grids of the solutions in the memory bank are stable, MBPSO generates initiate SAR plans based on the solutions in the memory bank. MBPSO uses an improved particle swarm optimization strategy. Particles are divided into multi groups. In each group, particles exchange information to make effective local search around the solutions with high qualities. Population will be regrouped after each group finishes updating. MBPSO searches the solution space based on this strategy. For cooperative SAR planning, MBPSO combines multi candidate plans with great differences to generate initial cooperative plans. Based on that, particles exchange information to search the optimal plan. The experimental results show that compared with other heuristic algorithms, MBPSO can generate SAR plans with higher success probability. (3) Aiming at maritime moving probability distribution SAR planning, a factorial heuristic optimization algorithm (FHOA) is proposed. FHOA introduces the idea of factorial experimental design, employs the idea of randomization and blocking. FHOA sets the values of each parameter, generates multi combinations of the other parameters with different values randomly. The fitness value curves of each parameter can be acquired by computing the success probability of each candidate plan. By analyzing the impact of each parameter on the fitness and the interaction between factors, FHOA generates initialization and update strategies. Individuals are divided into multi groups based on their distances, based on which each group can have good diversity. During the updating process, population will focus on global search for the parameters with strong interaction and local search for the parameters with weak interaction. Population will be regrouped after each group finishes updating. For cooperative SAR planning, FHOA combines multi candidate plans with great differences to generate initial cooperative plans. Individuals communicate with each other based on cluster structure to update its position. The experimental results show that FHOA can generate SAR plans with high quality stably for moving probability distribution SAR planning. (4) Aiming at detecting SAR target in sea images, a SAR target sensing method based on manifold learning is proposed. The proposed method acquires two classes of samples with labels, one class with samples containing the SAR target, the other class with samples not containing the SAR target. The proposed method uses linear discriminant analysis to map the samples into low dimensional data. The distance within the class is minimized, and the distance between the classes is maximized. The mapping matrix and classifier can be acquired. The proposed method uses the Canny operator and Hough transform to extract sea-sky lines in the sea images and extracts the regions which may contain SAR target based on the gray difference between SAR target and background. SAR target can be sensed and detected by mapping the extracted regions. The information of the target location is updated and the drift trajectory is predicted based on the sensing results and Bayesian posterior probability. The experimental results show that, on the basis of reducing the requirements of false alarm, LDATS can detect sea images effectively, which can increase SAR efficiency.
语种中文
产权排序1
文献类型学位论文
条目标识符http://ir.sia.cn/handle/173321/21822
专题光电信息技术研究室
作者单位1.中国科学院沈阳自动化研究所
2.中国科学院大学
推荐引用方式
GB/T 7714
吕进锋. 基于Agent的海上搜救任务规划方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2018.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
基于Agent的海上搜救任务规划方法研究(215KB)学位论文 开放获取CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[吕进锋]的文章
百度学术
百度学术中相似的文章
[吕进锋]的文章
必应学术
必应学术中相似的文章
[吕进锋]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。