期刊文献+

基于贝叶斯单源域领域泛化算法的天然气管道故障智能诊断

Natual gas pipeline fault intelligent diagnosis based on the Bayesian single-source domain generalization algorithm
下载PDF
导出
摘要 基于深度学习算法的故障智能诊断模型已被广泛应用于天然气管道运输安全领域,然而管道通常处于准稳态,使得训练集中的故障样本量受限。为此,针对天然气管道故障诊断中因训练集故障样本量有限,导致难以准确诊断的问题,提出了一种基于贝叶斯单源域领域泛化(BSDG)算法,部署了一种攻击防御策略,通过在攻击阶段明确伪目标域增强路径,并在防御阶段引导模型参数的后验分布向伪域样本得分更高的方向调整,增强模型在面对不同域扰动时的适应性和鲁棒性。研究结果表明:(1)基于贝叶斯网络建立的非定向攻击模型确保伪域样本既保留了与源域的相关性,又引入了足够的域差异来模拟潜在的目标域,由此提升了多源域和单源域设置下的领域泛化诊断准确率;(2)测试结果显示,BSDG算法在多源域泛化任务及两项单源域泛化任务中,相较于性能最优的对比算法,其准确率分别提高了9.79%、5.09%和27.98%;(3)裕度差异损失通过在学习决策边界的过程中引入不确定性,令分类器可以灵活且有效应对频繁的分布变化,显著性测试结果表明BSDG算法在多数场景下显著优于先进对比算法;(4)贝叶斯神经网络通过在权重上引入不确定性,有效提升了BSDG算法的泛化稳定性。结论认为,BSDG算法通过使用基于贝叶斯推理的攻击防御策略,有效扩展了源域模型的决策边界,解决了实际场景数据匮乏导致的深度神经网络泛化能力差的问题,为样本受限情形下的天然气管道故障诊断模型设计提供了理论支撑。 Deep learning-based fault intelligent diagnosis models have been widely used in the research on gas pipeline transportation safety.However,gas pipelines usually operate in a quasi-steady state,resulting in a limited number of fault samples in the training set,which may impede the accuracy of fault diagnosis.In this regard,this paper presents a new fault diagnosis method based on Bayesian single-domain generalization(BSDG).The core of the BSDG algorithm lies in deploying an attack-defense strategy that enhances the model's adaptability and robustness under different domain perturbation settings by specifying the pseudo-target domain augmentation paths in the attack phase and adjusting the posterior distributions of the model's parameters in the direction of the higher pseudo-domain samples scores in the defense phase.The results show that the Bayesian network-based untargeted attack model ensures that the pseudo-domain samples retain correlation with the source domains while introducing enough domain discrepancies to simulate the potential target domains,which aims to improve the diagnostic accuracy in both the multi-source and single-source domain settings.The testing results indicate that the BSDG algorithm improves the performance over the best contrast algorithm(SSAA algorithm)by 9.79%,5.09%,and 27.98%in the multi-source domain generalization task and the two single-source domain generalization tasks,respectively.The margin discrepancy loss allows the classifier to be flexible and effective in dealing with frequent distribution variations by introducing uncertainty in the process of learning the decision boundaries.The significance tests show that the BSDG algorithm significantly outperforms the state-of-the-art contrast algorithms in most scenarios.The Bayesian neural network effectively improves the generalization stability of the BSDG algorithm by introducing uncertainty on the weight.It is concluded that the BSDG algorithm effectively extends the decision boundary of the source-domain model and improves the diagnostic generalization performance by using the attack-defense strategy based on the Bayesian inference algorithm,which provides theoretical support for the design of fault diagnosis models of gas pipelines with limited samples.
作者 董宏丽 商柔 汪涵博 王闯 陈双庆 管闯 DONG Hongli;SHANG Rou;WANG Hanbo;WANG Chuang;CHEN Shuangqing;GUAN Chuang(Sanya Offshore Oil&Gas Research Institute,Northeast Petroleum University,Sanya,Hainan 572025,China;Artificial Intelligence Energy Research Institute,Northeast Petroleum University,Daqing,Heilongjiang 163318,China;Heilongjiang Provincial Key Laboratory of Networking and Intelligent Control,Daqing,Heilongjiang 163318,China;School of Petroleum Engineering,Northeast Petroleum University,Daqing,Heilongjiang 163318,China)
出处 《天然气工业》 EI CAS CSCD 北大核心 2024年第9期27-37,共11页 Natural Gas Industry
基金 国家自然科学基金区域创新发展联合基金项目“基于分布式算法及大数据驱动的微地震信号去噪与反演研究”(编号:U21A2019) 国家资助博士后研究人员计划B档资助项目“面向深海油气管道典型小样本场景的知识迁移方法研究”(编号:GZB20240136) 中国博士后科学基金第75批面上资助“地区专项支持计划”项目“非完备数据约束下的油气管网智能运维关键技术研究”(编号:2024MD753911)。
关键词 天然气管道 故障智能诊断 迁移学习 贝叶斯神经网络 小样本问题 泛化能力 Natural gas pipelines Fault intelligent diagnosis Transfer learning Bayesian neural network Small sample issue Generalization ability
  • 相关文献

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部