采油系统故障诊断的增广数据深度特征域适应方法研究

Affected by factors such as geology, equipment and environment, the oil production conditions are complicated and the faults occur frequently, resulting in a decline in output efficiency. Timely and accurate fault diagnosis is of great significance for reducing the cost of the oil field. Fault diagnosis of the oil production system is a typical pattern classification problem. Since the normal samples are far more than the fault samples, the unlabeled samples are far more than the labeled data, and the differences in the samples under different regions, working conditions and collection conditions are large, the existing fault diagnosis methods have poor adaptability. To address this issue, this project tries to study the fault diagnosis method of oil production system based on deep convolutional neural network. Considering the hierarchical characteristics of oil production faults, multi-scale fusion of deep features is studied to improve the accuracy of deep convolutional neural network based fault diagnosis methods; Considering the large intra-domain differences caused by the biased characteristics of training samples, the augmentation techniques of source data are studied to establish a unified multi-classification data augmentation framework; Considering the large difference between the source domain and the target domain data, the generation adversarial network and domain adaptation methods are innovatively integrated, and the semi-supervised domain adaptive learning method is established. The development of the project can expand the theory of deep transfer learning and promote its application in the field of industrial big data analysis, especially industrial fault diagnosis. The project has significant theoretical and practical value.

受地质、设备和环境等因素影响，采油系统工况复杂、故障频发，导致产量效益下滑。及时、准确判断采油系统故障类型对油田降本增效意义重大。采油系统故障诊断是典型的模式分类问题，由于系统数据中正常样本远大于故障样本，未标注样本远大于标注样本，不同地域、工况和采集条件下样本差异性大等问题，现有故障诊断方法的适应性差。为此，本项目以深度卷积神经网络为基础，开展采油过程故障诊断的增广数据深度特征域适应方法研究。针对采油系统故障层次化特性，开展深度特征多尺度融合研究，提升深度卷积神经网络故障诊断方法的精度；针对因训练样本有偏特性导致的域内差异大问题，开展源域数据增广研究，建立统一的多分类数据增广框架；针对源域与目标域数据差异大的问题，创新地融合生成对抗和域适应方法，建立半监督域适应学习方法。项目的开展可促进深度迁移学习理论的拓展，推动其在工业大数据分析尤其是工业故障诊断领域的应用，具有理论价值和实用意义。

采油系统故障诊断对油田降本增效意义重大。采用深度神经网络可以挖掘生产过程参数与故障工况之间的潜在规律，从而为采油系统安全、稳定、高效生产提供辅助决策。本项目利用深度学习和迁移学习理论建立构建具有高精度和高泛化能力的新型采油系统故障诊断模型。具体开展了如下工作：.(1)采油系统故障诊断的多尺度特征融合深度卷积神经网络策略研究。提出一种双源信息融合和多层次特征融合网络进行采油系统故障诊断的新型算法MFF_CNN。为解决模型中传统卷积的一些缺陷问题和数据不平衡问题，提出一种基于空洞卷积和Focal loss的故障诊断算法DIMFF_FL_CNN，提升卷积感受野和对难分类样本的诊断准确率。.(2)深度卷积神经网络采油系统故障诊断的数据增广研究。本研究在Mixup方法的基础上引入标签平滑策略，实现训练数据真实标签的正则化，增加数据多样性，改进Mixup以适用于多分类数据。提出了一种基于标签平滑和Plane-Mixup结合的双源多层次信息融合深度卷积神经网络，并应用于故障诊断。.(3)基于对抗学习的深度卷积神经网络采油系统故障诊断域适应方法研究。利用域适应理论克服故障诊断模型泛化性差的问题，同时引入基于聚类的伪标签生成方法抑制伪标签噪声影响，提高故障诊断模型在不同工况或场景下的适应性，实现模型的高精度和高泛化性。设计了实现源域和目标域类别层次匹配的结构信息保持域适应网络算法SIP_DAN；并考虑到更加细粒度的域对齐，减小类内距离并增大类间距离，引入加权对比域差异，系统提出深度子结构域适应网络算法DSDAN。此外，本研究提出对抗部分域适应网络算法APDAN以解决部分域适应问题，引入对抗式策略学习域不变特征，通过非对抗式域鉴别器加权衡量源域样本重要性，降低源域私有类样本的影响，抑制负迁移。.本项目利用通过智能科学方法建立采油系统故障诊断模型，推动深度迁移学习理论在工业故障领域诊断领域的应用，为提高采油系统智能故障诊断水平奠定了精准模型基础。