| 引用本文: | 薛佳帆,张航维,何广华,姜泽成.基于深度学习的海工平台运动预测[J].哈尔滨工业大学学报,2024,56(8):163.DOI:10.11918/202307085 |
| XUE Jiafan,ZHANG Hangwei,HE Guanghua,JIANG Zecheng.Motion prediction of offshore platforms based on deep learning[J].Journal of Harbin Institute of Technology,2024,56(8):163.DOI:10.11918/202307085 |
|
| |
|
|
| 本文已被：浏览 1046次 下载 1277次 |
码上扫一扫！ |
|
|
| 基于深度学习的海工平台运动预测 |
|
薛佳帆1,张航维1,何广华1,2,3,姜泽成2
|
|
(1.哈尔滨工业大学(威海) 海洋工程学院,山东 威海 264209; 2.哈尔滨工业大学 机电工程学院, 哈尔滨 150001; 3.山东船舶技术研究院, 山东 威海 264209)
|
|
| 摘要: |
| 为提高海上作业装备的安全性能,实现海上结构物运动的实时预报,本研究利用了一种卷积神经网络（convolutional neural network, CNN）方法与长短期记忆（long short-term memory, LSTM）方法混合的深度学习模型用于运动预测,该混合模型运用CNN提取运动数据的特征,采用LSTM学习提取出的特征间的时间关系,并结合贝叶斯(Bayesian)优化算法优化混合模型的超参数。首先,对海工平台进行数值仿真,将得到的纵荡运动数据作为实验数据。其次,将数据集划分为训练集、验证集与测试集,利用训练集与验证集进行训练与验证,获得6、12、18 s运动最优预测模型,并利用测试集与LSTM模型进行比较。结果表明:混合模型在6、12、18 s预测上相比于LSTM等模型,预测精度可以提高15%~30%。除此之外,本研究还分别探究了预测精度与输入时长、预测时长之间的关系,结果显示预测精度受输入时长的影响较小,但随着预测时长的增加近似呈线性下降趋势。最后结合训练耗时证明了混合模型相比于LSTM等模型更具有优势。 |
| 关键词: 贝叶斯优化 自动机器学习 运动预测 CNN LSTM |
| DOI：10.11918/202307085 |
| 分类号:TP391.9 |
| 文献标识码:A |
| 基金项目:山东省泰山学者工程专项经费 (tsqn201909172)；山东省高等学校青年创新团队科技计划(2019KJN003) |
|
| Motion prediction of offshore platforms based on deep learning |
|
XUE Jiafan1,ZHANG Hangwei1,HE Guanghua1,2,3,JIANG Zecheng2
|
|
(1.School of Ocean Engineering, Harbin Institute of Technology (Weihai), Weihai 264209, Shandong, China; 2.School of Mechatronics Engineering, Harbin Institute of Technology, Harbin 150001, China; 3.Shandong Institute of Shipbuilding Technology, Weihai 264209, Shandong, China)
|
| Abstract: |
| To improve the safety performance of offshore operation equipment and realize the real-time prediction of motion of offshore structures, a hybrid deep learning model combining convolutional neural network (CNN) and long short-term memory (LSTM) methods is used in this study. The hybrid model extracts the features from motion data by CNN, and utilizes LSTM to learn the temporal relationship among the extracted features. Additionally, Bayesian optimization algorithm is introduced to optimize the hyperparameters of the hybrid model. Firstly, the numerical simulation of the offshore platform is carried out, and the obtained surge motion data is used as experimental data. Secondly, the experimental dataset is divided into training set, verification set and test set. The training and verification set are used for model training and validation to obtain the optimal prediction models for 6 s, 12 s and 18 s of motion. The performance of the developed models is compared with that of the LSTM model using the testing set. The results show that the hybrid model, compared with the LSTM model, can improve the prediction accuracy by 15% to 30% for 6 s, 12 s and 18 s predictions. Furthermore, this study also investigates the relationship between prediction accuracy and input duration as well as prediction duration. The results suggest that the input duration has a minimal impact on the prediction accuracy, while the prediction accuracy shows a linear downward trend with the increase of the prediction duration. Finally, combined with the training time, the hybrid model in this paper demonstrates advantages over LSTM and other models. |
| Key words: Bayesian optimization automatic machine learning motion prediction CNN LSTM |
|
|
|
|
