摘要
示功图识别是有杆泵故障诊断的常用方法,随着神经网络技术的发展,虽已能准确识别典型示功图,但存在以下问题:代表不同故障的相似示功图识别精度有待提高;卷积神经网络(CNN)经典AlexNet模型识别示功图运算效率偏低。基于示功图,设计叠加示功图作为诊断对象,增加分类特征。改良的AlexNet模型:学习LeNet-5模型结构简单,运算快速,删减2层卷积、2层池化和1层全连接,提高模型运算速度;选用3×3的卷积核和2×2的池化核,增强特征学习能力,保证模型识别精度。经叠加示功图样本集测试:新模型相比LeNet-5模型、AlexNet模型,在进一步提高识别精度的同时显著提高运算效率。经现场应用:新模型可以准确高效识别代表不同故障的相似示功图,提液单耗减少16%,设备工作效率提高33%等。
The indicator diagram recognition is a common method for fault diagnosis of rod pump.With the development of neural network technology,it can accurately identify typical indicator diagrams.However,the following problems exist:the accuracy of identifying approximate indicator diagrams caused by different faults needs to be improved;and the operation efficiency of identifying indicator diagrams based on convolutional neural network(CNN)classical AlexNet model is low.On the basis of the indicator diagram,the superimposed indicator diagram was used as diagnostic object to add classification features.Modified the AlexNet model:Absorbing the characteristics of LeNet-5 model which is simple in structure and fast in operation,deleting two convolution layers,two pooling layers and one full connection layer to improve the operation efficiency;using 3×3 convolution cores and 2×2 pooling cores to enhance the feature learning ability and ensure the recognition accuracy.After testing the data sets of superimposed indicator diagrams,the improved AlexNet model is compared with LeNet-5 model and AlexNet model,which could further improve the recognition accuracy and significantly improve the operational efficiency.Field application showed that approximate indicator diagrams generated by different faults could be identified accurately and efficiently,the average unit consumption of liquid extraction could be reduced by 16%,and the equipment working efficiency could be increased by 33%.
作者
何岩峰
刘成
王相
HE Yanfeng;LIU Cheng;WANG Xiang(School of Petroleum Engineering,Changzhou University Changzhou,Jiangsu 213164)
出处
《工业安全与环保》
2020年第7期22-26,共5页
Industrial Safety and Environmental Protection
基金
江苏省高等学校自然科学研究项目(17KJB440001)。