船舶主动干扰下的水面无人艇脱逃机理研究

Obstacle avoidance ability is the basic safeguard for an unmanned surface vessel(USV) to complete its tasks. At present, the obstacles are mostly being considered as COLREGS ruled boats. However, the researches have shown that the USV’s can not getting away from the initiative threating boats has come to be the prior problem restricting its autonomous naviagation capacity. The project is aimed at solving this problem based on the present foundation of research. Refer to the driving intention recognition method in autonomous driving area, we propose a theory which build multivariate normal distribution models of various interfere based on test samples, then study the mathematics calculation model of matching between boat motion and interference characteristics with supervised learning theory. About the question that the synergistic action of multi-ships is difficult to distinguish, we proposed a Logical Hierarchical Hidden Semi-Markov Model to deal with its complexity, partial observability, and variability. Besides, For the restricted condition of obstacle avoidance planning can not be expressed and conducted effectively, a Augmented Lagrangian recurrent neural network is proposed to proceed multi-constraint fusion and solution. Finally, the simulation and experiment are developed to explore the escaping planning mechanism. The research results will help to improve the obstacle avoidance and escaping ability for USV in full complex initiative threating environment, and to promoto the development of the USV practical application. It can also provide reference for other robot control system for avoidance planning.

避碰能力是无人艇执行任务的基础，现有研究多考虑船舶按规则航行，而在船舶主动干扰时，无人艇难以开展有效的避碰与脱逃，这是限制无人艇全自主航行的重要因素。课题拟在已有的研究基础上，研究无人艇在船舶主动干扰下的脱逃机理。船舶的主动干扰方式难以有效辨识，导致无人艇无法制定针对性的避碰策略，为此基于试验样本建立不同干扰的多元高斯分布模型，利用监督学习理论进行干扰类型的匹配；针对多船协同干扰意图辨识复杂、观测缺失及变化性强的特点，提出了一种逻辑层次化隐半马尔可夫模型，实现协同意图的辨识和各自独立动作的预测；针对多约束条件与规划目标存在隔阂导致脱逃规划求解效果不佳的问题，采用增广拉格朗日协同递归神经网络算法进行多约束与脱逃规划的融合与求解；另外，通过仿真与试验进行理论验证。研究成果将有助于提升无人艇在船舶复杂干扰下的避碰与脱逃能力，推动无人艇全自主航行的发展，也为其他智能平台的自主安全性探索了新方法。

避碰能力是无人艇执行任务的基础，现有研究多考虑船舶按规则航行，而在船舶主动干扰时，无人艇难以开展有效的避碰与脱逃，这是限制无人艇全自主航行的重要因素。本课题在已有的研究基础上，研究无人艇在船舶主动干扰下的脱逃机理。船舶的主动干扰方式难以有效辨识，导致无人艇无法制定针对性的避碰策略，为此基于试验样本建立不同干扰的多元高斯分布模型，利用监督学习理论进行干扰类型的匹配，提出了一种结合主成分分析（Principal Component Analysis，PCA）和支持向量机（Support Vector Machine，SVM）的分类算法.首先利用PCA对船舶模型运动得到的高维运动特征数据集进行降维，获取特征集的主成分，再通过SVM对经过处理得到的低维主成分进行分类处理，以识别障碍对无人艇的干扰意图类型为主动干扰还是非主动干扰.最后根据障碍的不同干扰类型，采用不同的算法进行避碰、脱逃，通过这种差异化的处理，提升无人艇运动的安全性。同时提出了基于高斯混合模型（Gaussian mixture model，GMM）和连续隐马尔科夫模型（Continuous Hidden Markov Model，CHMM）的船舶干扰意图识别模型。首先，考虑到意图识别的复杂性，基于USV与船舶的速度障碍模型，提出了多周期干扰系数等运动特征参数，用作模型的输入；考虑到船舶运动的连续性，使用GMM作为观测-状态转移概率分布。然后，通过Baum-Welch算法对训练数据进行训练，得到意图识别模型。最后，获取预测样本对模型的准确性进行评估。针对多船协同干扰意图辨识复杂、观测缺失及变化性强的特点，提出了基于DQN的无人艇在协同围捕下的脱逃算法并进行训练和仿真、试验验证。研究成果有助于提升无人艇在船舶复杂干扰下的避碰与脱逃能力，推动无人艇全自主航行的发展，也为其他智能平台的自主安全性探索了新方法。