基于径向基神经网络的水下自主航行器寻源算法研究  

Autonomous underwater vehicle source-seeking algorithm based on radial basis function neural network

作　　者：吴宗秀 吴超[1,2] WU Zongxiu;WU Chao(School of Naval Architecture,Ocean and Civil Engineering,Shanghai Jiao Tong University,Shanghai 200240,China;State Key Laboratory of Ocean Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学船舶海洋与建筑工程学院,上海200240 [2]上海交通大学海洋工程国家重点实验室,上海200240

出　　处：《海洋工程》2021年第6期99-110,共12页The Ocean Engineering

基　　金：国家自然科学基金重大科研仪器研制项目(41427806);国家重点研发计划项目(2016YFC0300700)。

摘　　要：针对水下航行器在二维信号场中的场源搜索问题,提出了一种基于径向基函数神经网络(radial basis function neural network,简称RBFNN)的在线自主寻源算法。在神经网络中引入全局正则化参数以保证泛化性和稳定性,通过最小化广义交叉验证误差(generalized cross-validation,简称GCV)进行正则化参数的迭代优化,并利用增量式奇异值分解(incremental SVD)对迭代过程进行加速,此外通过基于样本新颖性的资源分配网络算法(resource-allocating network,简称RAN)进行径向基函数的分配,在此基础上使用动量梯度算法进行航行器运动方向的规划。最后,以热泉区硫化氢浓度分布场中的搜索作业为背景,使用该算法与其他研究中的算法进行单峰值信号场的场源搜索模拟计算对比,结果显示该算法对于信号场梯度的估计更加准确,且搜索过程的路径更短。此外在多峰值信号场的寻源模拟中该算法能够以较高的成功率通过局部最大值区域。证明该算法具有良好的拟合、预测性能以及稳定性,并且能在一定程度上避免陷入局部最优解。Aiming at the problem of field source search for underwater vehicles in a two-dimensional signal field,an online autonomous source search algorithm based on radial basis function neural network(RBFNN)is proposed.Introduce global regularization parameters into the neural network to ensure generalization and stability.Iteratively optimize the regularization parameters by minimizing generalized cross-validation(GCV),and use incremental singular value decomposition to accelerate the iterative process.In addition,the radial basis function is allocated through the resource-allocating network(RAN)algorithm based on sample novelty.On this basis,the momentum gradient algorithm is used to determine the direction of the vehicle's movement.planning.Finally,based on the search operation in the distribution field of hydrogen sulfide concentration in the hot spring area,the algorithm is compared with other algorithms in the field source search simulation of single peak signal field.The results show that the algorithm estimates the signal field gradient more accurate,and the path of the search process is shorter.In addition,the algorithm can pass through the local maximum area with a higher success rate in the source-seeking simulation of multi-peak signal field.It is proved that the algorithm has good fitting,prediction performance and stability,and can avoid falling into the local optimal solution to a certain extent.

关 键 词：径向基函数神经网络 水下自主航行器 自动寻源 奇异值分解 资源分配网络 

分 类 号：P754[天文地球—海洋科学] TP249[自动化与计算机技术—检测技术与自动化装置]