【作者】 王雷；

【导师】 张化光；

【作者基本信息】 东北大学 ， 控制科学与工程， 2021， 硕士

【摘要】 随着国民对石油、天然气等能源需求的增加,管道运输方式因其成本低、效率高、安全性好而被广泛采用。然而部分管道由于铺设时间长且受运输介质、自然环境等影响产生管壁破损等现象,具有很大的潜在危险。科学合理的管道故障诊断是能源安全运输的重要保证。缺陷尺寸是管道安全状态评估的重要参考指标,因此实现对管道缺陷信息的定量化评估具有重要意义。当前以深度学习和机器学习为主流的先进方法发展日趋成熟,给管道故障诊断技术的发展提供了新方向。本文结合人工智能技术实现了对管道缺陷深层特征的自适应提取和缺陷尺寸的智能化反演,主要的创新型工作如下:第一,针对传统数据预处理方案使得缺陷特征相对关系发生畸变,给缺陷反演引入人工误差的问题,提出了一种特定于缺陷深层特征提取的数据预处理方法。首先分析了基于漏磁信号的管道缺陷数据采集原理和经典数据预处理方案存在的问题,然后设计中值法则对数据进行异常检测和基值校正,再设计自适应数据对齐方法对缺陷数据进行规范化处理,为缺陷深层特征提取提供数据支持,最后通过消融实验证明所提预处理方法的有效性。第二,针对缺陷显性特征提取方法受主观因素影响大、反演精度低、稳定性差的问题,研究了一种缺陷深层特征自适应提取方法。通过建立一种新型多路径堆栈自编码模型以无监督方式实现了缺陷特征的跨层传输和多层信息共享,扩展了各个自编码单元的信息来源,从而获取到优质的深层特征。最后通过对比实验验证了所提出的方法较传统自编码模型相比具有更小的标准偏差和更好的实用性。第三,针对多路径堆栈自编码模型导致网络复杂性增加的问题,提出了深层特征融合和稀疏表达策略。首先以自编码模型中编码部分的结构作为预训练模型设计多分类神经网络以有监督的方式对缺陷特征进行整体微调,然后设计合理的融合权重实现子模型特征融合确保输出特征维度的一致性,再采用神经元随机失活和激活函数改进等方法来降低网络的复杂性,减少过拟合。最后,将所提方法与现有方法进行对比分析,实验结果表明所提方法具有更高的反演精度,验证了深层特征的优越性。第四,针对缺陷特征和缺陷尺寸之间非线性关系复杂,难以被单一尺寸估计模型所刻画的问题,提出了一种基于深度级联Stacking的管道缺陷尺寸估计方法。首先分析了集成学习方法较单一回归模型相比的优势,其次以集成学习思想为基础设计了级联Stacking缺陷尺寸估计网络,其中通过设计多尺寸的滑动窗口对输入特征进行多粒度扫描以增强网络的鲁棒性,通过级联结构将多个单一的尺寸估计模型层层堆栈,来拟合缺陷特征和尺寸间复杂的关系。最后,将该方法与其他方法进行尺寸估计性能对比,以MAE和RMSE等为指标从多个角度验证所提方法的先进性。本文研究的管道缺陷智能化反演方法有效地降低了对人工经验的依赖,提升了缺陷反演的精度和稳定性,为管道缺陷的定量化评估提供了有力的技术支撑,对我国长输管道故障诊断和精准维护工作有重大意义。更多还原

【Abstract】 With the increasing demand for energy such as oil and natural gas,pipeline transportation is widely used because of its low cost,high efficiency and good safety.However,some pipelines are likely to be damaged due to the long buried time and the influence of natural environment.Scientific and reasonable pipeline fault diagnosis is an important guarantee for safe energy transportation.Defect size is an important reference index for pipeline safety condition assessment,so it is of great significance to realize quantitative assessment of pipeline defect information.The development of some advanced methods based on deep learning and machine learning is becoming more and more mature,which provides a new direction for the development of pipeline fault diagnosis technology.In this thesis,combined with artificial intelligence technology,the adaptive extraction of deep features of pipeline defects and the intelligent inversion of defect size are realized.The main work is summarized as follows:Firstly,a data preprocessing method specific to deep defect feature extraction is proposed aiming at the problem that the traditional data preprocessing scheme distorts the relative relationship of defect features and.introduces artificial error to defect inversion.At first,the principle of pipeline defect data acquisition based on magnetic flux leakage(MFL)signal and the problems of classical data preprocessing scheme are analyzed.Then,the median rule is designed for anomaly detection and base value correction,and the adaptive data alignment method is designed for standardized processing of defect data,which provides data support for deep feature extraction of defects.Finally,the effectiveness of the proposed pre-processing method is proved by ablation experiments.Secondly,considering that the dominant feature extraction method of defects is greatly affected by subjective factors with low inversion accuracy and poor stability,an adaptive method for deep feature extraction of defects is conducted.A novel multi-path stack auto-encoder is proposed to realize cross layer transmission of defect features and information sharing among multi-layers in an unsupervised way.In addition,the information source of each auto-encoder unit is broadened to obtain high quality deep features.Finally,compared with the traditional auto-encoder,the proposed method has smaller standard deviation and better practicability.Thirdly,in order to solve the problem of increasing network complexity caused by multi-path stack auto-encoder,deep feature fusion and sparse representation strategies are proposed.At first,the coding structure is adopted as a pre-training model to design a multi classification neural network achieving the fine-tuning of defect features in a supervised way.Then,a reasonable fusion weight is designed to ensure the uniformity of the data structure of output feature.Morever,in order to reduce the complexity and over-fitting of the network,neuron random deactivation and activation function improvement are designed.Finally,the proposed method is compared with other state-of-art methods,and the experimental results show that the proposed method has higher inversion accuracy verifying advantages of deep features.Fourthly,due to the complex nonlinear relationship between defect feature and defect size,it is difficult to be described by a single size estimation model.A depth cascading Stacking based pipeline defect size estimation method is proposed.At first,some points of ensemble learning method over single regression model are analyzed.Then based on the idea of ensemble learning,a cascaded Stacking defect size estimation network is designed,in which a multi-dimensional sliding window is designed to scan the input features with multi-grained to enhance the robustness of the network,and a cascade structure is used to stack multiple single size estimation models layer by layer to fit the complex relationship between defect features and size.Finally,the performance of this method is compared with other methods for size estimation.The advanced properties of the proposed method are verified from multiple angles such as MAE and RMSE.The intelligent inversion method of pipeline defects studied in this thesis effectively reduces the dependence on artificial experience and improves the accuracy of defect inversion,which provides strong technical support for quantitative evaluation of pipeline defects.It is of great significance for fault diagnosis and accurate maintenance of long-distance pipelines.更多还原