文献类型：会议论文

英文题名：AUTOMATIC RECOGNITION OF OIL INDUSTRY FACILITIES BASED ON DEEP LEARNING

作者：Zhang, Nannan[1];Liu, Yang[1];Zou, Liqun[1];Zhao, Hang[1];Dong, Wentong[1];Zhou, Hongying[1];Guo, Hongyan[1];Huang, Miaofen[2]

机构：[1]PetroChina, Res Inst Petr Explorat & Dev, Beijing 100080, Peoples R China;[2]Guangdong Ocean Univ, Fac Math & Comp Sci, Zhanjiang 524088, Guangdong, Peoples R China

会议论文集：38th IEEE International Geoscience and Remote Sensing Symposium (IGARSS)

会议日期：JUL 22-27, 2018

会议地点：Valencia, SPAIN

语种：英文

外文关键词：petroleum industry facility; deep learning; machine learning

外文摘要：Effectively monitoring the real-time position and status of oil facilities (mainly well-site) in oil field is very important for the safety production. Considering the low efficiency of traditional visual interpretation method and the high demands of preset feature for machine learning method, one of the object detection methods in Deep learning (YOLOv2) was introduced to recognize oil industry facilities automatically. After establishing the dataset of oil facility samples, 90 percent of samples are used for model training while 10 percent are for validating. Comparing with the results extracted by machine learning (Adaboost model based on Haar-like), YOLOv2 recognition results of oil facilities indicated that: Deep learning improve the recognition efficiency and accuracy of oil facilities. The accuracy can be as high as 92% while the error rate and omission rate can be maintained in a low level. At the same time, the constructed model was applied in an oilfield in eastern part of China, and the result shows that the model can identify most of the oilfield facilities correctly with only 4% omission rate, which is much lower comparing with manual interpretation. However, the 11% error rate, caused by insufficient sample types and sample quantities, is relatively high especially in city area.