基于改进Adam-DBN的油井工况诊断方法  

Diagnosis method based on improved Adam-DBNfor working conditions of oil well

作　　者：王通[1] 熊涛理 WANG Tong;XIONG Tao-li(School of Electrical Engineering,Shenyang University of Technology,Shenyang 110870,China)

机构地区：[1]沈阳工业大学电气工程学院,沈阳110870

出　　处：《沈阳工业大学学报》2023年第3期330-335,共6页Journal of Shenyang University of Technology

基　　金：国家自然科学基金项目(62173073)。

摘　　要：针对深度信念网络在油井工况诊断中由于梯度扩散导致训练效果差,模型诊断准确率不高的问题,提出一种基于改进Adam优化算法的深度信念网络油井工况诊断方法.以二值化处理后的示功图作为深度信念网络输入,利用对比散度算法对网络进行无监督预训练,以获取较优的初始权值;在反向传播微调网络过程中,运用动量法预测梯度下降位置,更新梯度下降方向,并通过学习率自适应选择下降步长,避免梯度扩散降低模型训练效果.某采油平台油井上的仿真实验结果表明,基于改进Adam优化算法的深度信念网络对油井工况的识别准确率较高,能更好地满足油田生产实际需求.Aiming at the problem of poor training effect and low accuracy of model diagnosis caused by gradient diffusion of deep belief network in diagnosis of working conditions of oil well,a deep belief network diagnosis method based on improved Adam optimization algorithm for working conditions of oil well was proposed.The dynamometer card after binarization processing was taken as the input of deep belief network,and the network was subjected to unsupervised pre-training by using a contrast divergence algorithm to obtain a better initial weight value.In the process of back propagation fine-tuning network,a momentum method was used to predict the gradient descent position and update the gradient descent direction,and adaptively select the descent step size by learning rate to avoid the reduction of model training effect caused by gradient diffusion.The results of simulation experiments on the oil well of oil production platform in a certain oilfield show that the depth belief network based on the improved Adam optimization algorithm has a higher recognition accuracy for working conditions of oil well,and can better meet the actual needs of oil field production.

关 键 词：工况诊断 特征提取 深度信念网络 受限玻尔兹曼机 示功图 梯度下降 优化算法 油井 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]