基于神经网络的船体运动位姿预测方法

Deep learning for predicting ship motion from images,pitches and rolls

海上运输在全球贸易中发挥着重要作用,依靠船舶自动驾驶系统的高速环境感知和自主决策能力,可有效减少由于人为操控失误造成事故的发生率。针对远洋航行中由于船体型号多变、海域情况复杂等造成的数据集缺失问题,本文提出一种通过物理引擎人工生成波浪图像与船体运动位姿数据集的方法。同时,针对波浪与船体运动姿态的时序性特点,以及经典循环神经网络面临梯度爆炸、输入输出等长等问题,提出一种基于CNN卷积神经网络和GRU门控循环神经网络的船体运动姿态预测模型,通过卷积神经网络获取图片特征,并借助Encoder-De-coder编码解码器结构,成功实现了以较短时间的数据对未来船体运动姿态(纵摇和横摇)的长时间和高精度预测。

Maritime transportation plays an important role in global trade.Relying on the high-speed environment perception capability of the ship’s automatic driving system,it can effectively reduce the incidence of accidents caused by human control errors.The changes of waves are the key factors affecting the movement of the vague.Aiming at the problem of missing data sets due to variable vague models and complex sea conditions during ocean voyages,this paper proposes a method of artificially generating wave images and vague motion pose data sets through a physics engine.At the same time,in view of the temporal characteristics of wave and ship motion posture,as well as the difficulties such as gradient explosion and equality of length of input and output,a vague motion posture prediction model based on CNN and GRU is proposed,with the help of Encoder-Decoder structure,which successfully realized the long-term and high-precision prediction of the future ship motion(pitch and roll)with a short time of data.