摘要
为解决海上油田注水系统故障诊断问题,提出基于KShape数据增广与混合神经网络的故障诊断模型。首先采用KShape时间序列聚类算法实现样本数据增广,然后通过融合双向长短时记忆网络(bidirectional long short-term memory network, BiLSTM)和全卷积神经网络(fully convolutional neural network, FCN)构建混合神经网络模型,最后应用贝叶斯优化算法对模型参数进行全局寻优。结果表明:提出的数据增广方法在数字油田大数据的基础上,能够有效扩充注水系统工况样本。提出的混合神经网络模型较BiLSTM和FCN单一模型效果更好,该模型综合了BiLSTM网络对时间序列数据依赖关系的良好捕捉能力以及FCN网络对局部特征的有效提取能力,较两者在准确率上分别提升4.9%和1.8%。贝叶斯优化方法在寻找更优超参数组合方面有显著效果,为提高模型的鲁棒性和泛化性能起到了重要作用。该方法较传统调参方法在准确率上提升5%,较网格搜索和随机搜索方法分别提升3.7%和1.9%。同时,该方法产生的不同超参数组合下的模型准确率中位数为84.5%,模型准确率在90%以上的占比达到18%。所提出的故障诊断模型,可有效识别地层堵塞、配水器堵塞和油管漏失等故障,为海上油田注水系统故障诊断提供了新的解决方案和有效的技术支持。
In order to solve the problem of fault diagnosis of water injection system in offshore oilfield,a fault diagnosis model based on KShape data augmentation and hybrid neural network was proposed.Firstly,the KShape time series clustering algorithm was used to augment the sample data,then the hybrid neural network model was constructed by fusing bidirectional long short-term memory network(BiLSTM)and fully convolutional neural network(FCN),and finally the Bayesian optimization algorithm was used to optimize the model parameters globally.The results show the proposed data augmentation method can effectively expand the working condition samples of the water injection system on the basis of digital oilfield big data.The introduced hybrid neural network model outperforms individual BiLSTM and FCN models.This integrated model combines the temporal dependency capture capability of BiLSTM networks with the local feature extraction ability of FCN networks,resulting in a respective improvement of 4.9%and 1.8%in accuracy compared to the two separate models.Bayesian optimization proves significantly effective in discovering optimal hyperparameter combinations,playing a crucial role in enhancing the model’s robustness and generalization performance.The method outperforms traditional tuning approaches with a 5%accuracy improvement.Moreover,it surpasses grid search and random search methods by 3.7%and 1.9%,respectively.The model accuracy median under different hyperparameter combinations reaches 84.5%,with an 18%proportion achieving accuracy above 90%.The proposed fault diagnosis model effectively identifies faults such as zone blockage,chock blockage,and wellbore leakage.This model offers a novel solution and technical support for diagnosing faults in offshore oilfield water injection systems.
作者
于继飞
姬煜晨
常振宁
隋先富
曹砚锋
杨阳
彭建霖
李昂
YU Ji-fei;JI Yu-chen;CHANG Zhen-ning;SUI Xian-fu;CAO Yan-feng;YANG Yang;PENG Jian-lin;LI Ang(National Key Laboratory of Offshore Oil and Gas Exploitation,Beijing 100028,China;CNOOC Research Institute Ltd.,Beijing 100028,China;CNOOC EnerTech-Drilling&Production Co.,Tianjin 300452,China)
出处
《科学技术与工程》
北大核心
2024年第24期10235-10243,共9页
Science Technology and Engineering
基金
中国海洋石油集团公司“十四五”重大科技项目(KJGG2021-0502)。
关键词
KShape
数据增广
混合神经网络
贝叶斯优化
故障诊断
注水系统
KShape
data augmentation
hybrid neural networks
Bayesian optimization
fault diagnosis
water injection system
