面向海上风机登乘机构相对位置估计方法研究  

Research on relative position estimation method for boarding mechanism of offshore wind turbines

作　　者：刘娇 刘积昊[1] 林志良[1] LIU Jiao;LIU Jihao;LIN Zhiliang(State Key Laboratory of Ocean Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学海洋工程全国重点实验室,上海200240

出　　处：《海洋工程》2024年第5期68-77,共10页The Ocean Engineering

基　　金：国家自然科学基金资助项目(51979162,52371287);上海交通大学“深蓝计划”基金资助项目(SL2021MS019)。

摘　　要：面向海上风机运维人员登乘场景,为获取准确的船舶与风机平台相对位置姿态关系,提出采用端-端的深度学习网络方法来训练基于风机平台RGB图像的位置估计器。首先,设计搭建能够精准控制位置姿态的图像数据采集装置,用以采集带有圆柱坐标标签的训练、测试和验证图像数据集,其中,训练集图像12 586张,测试集图像3 198张。训练集样本标签的空间为:径向2.1~18.9 cm,高度1.2~9.0 cm,角度66°~114°;采样间隔密度为:径向0.6 cm,高度0.6 cm,角度1.6°。然后,通过学习训练获得基于ResNet50网络的船舶-风机平台相对位置估计器,利用多组数据测试集评估了该位置估计器包括误差分布情况和运动模糊条件下的鲁棒性等性能。根据数据分析,相对位置估计器在径向、高度和角度3个自由度的平均绝对误差分别为0.10、0.07 cm和0.28°,且对轻度运动模糊具备鲁棒性。In the context of offshore wind turbine maintenance personnel boarding scenarios,an end to end deep learning network approach is proposed to train a position estimator based on RGB images of wind turbine platforms.First,an image data acquisition device capable of precise control over position and attitude is designed and constructed.This device is used to collect training,testing,and validation image datasets with cylindrical coordinate labels.The training dataset consists of 12586 images,while the test dataset contains 3198 images.The spatial labels for the training dataset include radial distances ranging from 2.1 to 18.9 cm,heights from 1.2 to 9.0 cm,and angles from 66°to 114°.The sampling intervals are 0.6 cm radially,0.6 cm in height,and 1.6°in angle.Subsequently,a relative position estimator of the ship-wind turbine platform based on the ResNet50 network is learned through training.The performance of this estimator is evaluated using multiple sets of test data,considering error distribution and robustness under mild motion blur conditions.Data analysis reveals that the relative position estimator achieves an average absolute error of 0.10 cm in radial distance,0.07 cm in height,and 0.28°in angle across the three degrees of freedom.It also demonstrates robustness against slight motion blur.

关 键 词：风机登乘机构 位置估计 深度学习 ResNet50 单目视觉 

分 类 号：P752[天文地球—海洋科学]