孪生神经网络在抽油系统故障诊断中的应用  

Application of Siamese Neural Network in Fault Diagnosis of Pumping System

作　　者：温后珍[1] 王浩宇 栾仪广 于双锴 陈德斌 Wen Houzhen;Wang Haoyu;Luan Yiguang;Yu Shuangkai;Chen Debin(School of Mechanical Science and Engineering,Northeast Petroleum University;Shandong Yulong Petrochemical Industrial Park Development Co.,Ltd.;Automobile and High-end Equipment Industry Promotion Center of Daqing High-tech Industrial Development Zone)

机构地区：[1]东北石油大学机械科学与工程学院 [2]山东裕龙石化产业园发展有限公司 [3]大庆高新技术产业开发区汽车和高端装备产业促进中心

出　　处：《石油机械》2023年第11期20-26,60,共8页China Petroleum Machinery

基　　金：黑龙江省自然科学基金项目“圆管螺旋流近壁相干结构和减阻性能研究”(QC2016003);东北石油大学青年科学基金项目“管柱力学软件开发”(HBHZX202004)。

摘　　要：人工智能诊断技术可在不同的数据和环境下进行自适应,从而大大提高故障诊断效率。由此提出了一种智能故障诊断方法,用于油田抽油系统的故障检测。现有的方法多采用神经网络技术,通过分析油井示功图来实现诊断。然而,实际采集到的油井示功图数据非常有限且类别不平衡,导致深度卷积神经网络容易出现过拟合。为了解决这个问题,提出采用预训练孪生神经网络方法。在一个较大的数据集上训练一个比较模型,用于判断图像之间的相似度。这个模型能够输出不同图片之间的相似度。利用预训练好的模型,在功图识别任务上进行微调,通过提取和融合2张图片的特征向量,输出它们之间的相似度。研究结果表明,预训练孪生网络模型能够很好地解决小样本问题,特别适用于功图识别这类任务。试验结果显示,该方法在小样本量功图识别任务上表现出色,具有高精度的故障诊断能力,满足抽油系统智能故障诊断要求。预训练孪生网络模型在小样本量功图识别任务上表现良好,为油田抽油系统的智能故障诊断提供了有效的解决方案。Artificial intelligence diagnostic technology can conduct self-adaptation under different data and environments,greatly improving the diagnostic efficiency.However,the existing methods mostly use neural network technology to achieve diagnosis by analyzing the oil well indicator diagram,but the actually collected oil well indicator diagram data are very limited and their categories are imbalanced,easily leading to overfitting of deep convolutional neural network.To solve this problem,a pre-trained siamese neural network method was proposed to be used for fault diagnosis of pumping system.First,a comparison model was trained on a larger dataset to judge the similarity between images.Then,the pre-trained model was used to perform fine-tuning on the recognition task of indicator diagram,and finally output the similarity of 2 pictures by extracting and fusing the feature vectors of them.The research results show that the pre-trained siamese network model can effectively solve the small sample problem,and is especially suitable for tasks such as indicator diagram recognition.The test results show that the method performs well in the small sample size indicator diagram recognition task,has high-precision fault diagnosis ability,and meets the requirements of intelligent fault diagnosis of pumping system.The pre-trained siamese network model provides an effective solution for the intelligent fault diagnosis of the pumping system.

关 键 词：抽油系统 孪生神经网络 示功图 故障诊断 预训练 小样本 

分 类 号：TE933[石油与天然气工程—石油机械设备]