基于MREMD-NGRU深度学习的非线性非平稳船舶运动实时预报方法研究

Ship motions due to ocean environmental disturbances bring negative effects to maritime operational safety. Especially in sever sea states, the ship motions are nonlinear and non-stationary. The nonlinearity is coupled with the non-stationarity. Hence, the non-stationary effects should be decoupled when developing a fully nonlinear prediction method. Which makes the real-time prediction of ship motions at real seas becomes a challenge research subject. This project aims to develop a hybrid deep learning network for real-time forecasting of nonlinear and non-stationary ship motions. Nonlinearity and non-stationarity processing methods are explored separately for developing a hybrid prediction model. The nested gated recurrent unit (NGRU) deep learning network architecture design and training algorithms are investigated for nonlinear time series modeling. Breakthrough in nonlinearity processing by using deep learning method is expected. Non-stationary effects on the prediction performance and their online processing technique are studied to overcome the non-stationarity challenge. A midpoint and regression-based empirical mode decomposition (MREMD) method is to be developed for online processing of non-stationarity. The negative effects on the deep learning network due to the non-stationarity can be eliminated. On the basis of the research works on nonlinearity and non-stationarity processing, a novel MREMD-NGRU deep learning network for real-time prediction of nonlinear and non-stationary ship motions is investigated. Numerical algorithm is to be developed and validations will be made based on collected ship motion data by simulation, towing test and sea trail. Research works of this project can provide theoretical and technical foundation support for both real-time forecasting of ship motions and operation decision making at sever seas.

船舶运动实时预报技术通过预测未来几秒到十几秒内确定性船舶运动时历，为船舶海上航行和任务作业提供决策信息，提升作业安全与能力，是海上作业决策关键技术。高海况下实船运动具有非线性和非平稳特征，二者同时存在。非线性时间序列建模中需同步考虑非平稳性在线处理，这使得实船运动实时预报成为了极具挑战的研究领域。本项目聚焦非线性与非平稳特征所引发的关键科学问题，提出一种复合深度学习网络方法。其中，针对非线性时序建模问题，研究NGRU网络架构设计及其训练算法，利用深度学习技术在非线性处理上获得突破；针对非平稳在线处理问题，研究MREMD在线处理方法，以克服非平稳性对预报精度的不利影响。在此基础上，针对非线性非平稳船舶摇荡运动实时预报问题，研究复合MREMD-NGRU深度学习网络，基于数值、水池和实船数据完成验证。研究成果将为高海况下船舶摇荡运动实时预报及海上作业决策技术研究提供理论参考和基础性支撑。

船舶在海上的大幅运动会导致船舶的航行安全问题，船舶运动预报能够为船舶提供安全预警，辅助船舶开展各类作业。因此船舶运动预报对于保障船舶海上作业有着重要意义。实际船舶运动带由明显的非线性与非平稳性，深度学习模型能够有效处理船舶运动的非线性.但是对于非平稳性的处理缺缺少能力，且现有时序深度学习模型还存在无法兼顾过拟合与记忆时间长度不足的问题。针对以上问题，本项目提出在GRU模型的基础上加入编码-解码模型，建立NGRU模型，解决模型的过拟合与记忆时间长度不足问题，并且也实现了多自由度的输入与多自由度的输出。在非平稳性的处理上，基于EMD建立优化端点处理的MREMD算法。在此即俗称上结合MREMD与NGRU模型，建立MREMD-NGRU模型，基于数值仿真数据与水池试验数据对模型进行验证分析。同时也对深度学习模型泛化性进行分析提出泛化性优化方法。本项目的主要创新点包括（1）引入编码解码结构，建NGRU模型，解决模型的过拟合与记忆时间长度不足问题，实现了多自由度的输入与多自由度的输出。（2）结合MREMD与NGRU模型，建立MREMD-NGRU模型，改善非线性、非平稳运动时历的预报。（3）提出频带分离的数据处理方法，改善深度学习模型提取不规则复杂运动特征不佳的问题。（4）对船舶运动预报深度学习模型的泛化性进行分析，并提出基于数据混合的泛化性优化方法。