Abstract：
With the development of intelligent oil fields, the technology for intelligent surveillance of equipment in oil fields
is required to reach a higher level. At present, there is a large number of surveillance videos stored in the databases of oil fields.
There has been research on video interpretation for face recognition, leak detection and other fields, but the video for rotating
equipment has not been fully exploited. In order to solve the problem of serious noise and various kinds of interference targets
in the surveillance video images of water injection pumps in oil fields, an intelligent machine vision method based on the Faster
Region Convolution Neural Network (Faster R-CNN) algorithm is introduced. Through the Feature Extraction Network (FEN),
the image features of the plunger region of the input image are effectively extracted. Through the Region Proposal Network
(RPN), a series of candidate regions are generated based on the extracted feature maps. Through the Object Detection Network
(ODN), the feature maps extracted by FEN and the candidate regions generated by RPN are integrated to identify the plunger
region and determine the coordinates of the plunger region. Therefore, automatic detection of the precise position of the plunger
region of the pump in the changing background is realized. By using a binarization algorithm and a Gaussian filtering algorithm,
each image of the plunger region is preprocessed to reduce image noise, which can help the frame difference during plunger
movement become significantly larger. The current motion state of the plunger region in each frame is determined by the
inter-frame difference method, and the overall motion state of the plunger region is determined based on the determination of the
standard motion state through the multiple frame differences, so as to realize the intelligent surveillance of the movement state of
the pump plunger. Compared with the traditional surveillance method based on numerical parameters from Supervisory Control
and Data Acquisition (SCADA), this method is more accurate and intuitive. Based on the machine vision method, the character
istics of serious noise and various kinds of interfering targets in the video image of the water injection pump can be responded
to effectively. The surveillance video in actual oil field production sites is utilized to verify the outstanding performance of this
intelligent surveillance technology for water injection pump. The total accuracy reached 96.75
%
, which is significantly higher
than the traditional inter-frame difference method and the optical flow method. The proposed method can provide technical
support for the intelligent management of oil field equipment.
