Research On Oil Field Instruments Reading Recogni-Tion Algorithm Based On Machine Vision

As an important raw material for modern industrial production,petroleum is an important strategic material of the country.It is difficult to estimate the consequences of accidents due to the existence of a large number of pressure vessels,valves,pipelines and other equipment in the oil site,which is flammable and explosive under the condition of high temperature,high pressure.Therefore,to ensure safety in production is the top priority of every oil production unit.As the most important security measures,patrol inspection is mainly manual inspection,which has high labor intensity,low work efficiency,missed inspection,false inspection and other problems.In order to ensure efficient and reliable inspection,domestic and foreign major oilfield production units vigorously develop intelligent inspection mode with robot as the main body.Automatic recognition of oilfield instrument readings is one of the core functions of the inspection robot.Due to the complexity of the oilfield environment,it is difficult for conventional instrument reading recognition methods to be applied to oilfield inspection robots.Therefore,it is of great practical significance to study the algorithms of oilfield instrument recognition for the realization of the oilfield inspection robot.As the core of intelligent inspection,the automatic recognition of instrument reading develops rapidly,but it still faces some difficulties and shortcomings as follows.(1)the current algorithms of instrument reading recognition require high quality of visual environment,the recognition results are easy to make mistakes when the visual environment is poor.(2)the reading recognition algorithms for pointer meter has many steps,low fault tolerance and poor robustness.In response to the above problems,this paper uses image processing technology and deep neural network to conduct a detailed study on the recognition of the readings of commonly used meters in oilfields.The main research contents are listed as follows.(1)Research on recognition of oilfield instrument reading based on image processing technology.Aiming at the problems of fuzzy,deformation,high light and low light caused by the complex environment in the oilfield instrument image,the image denoising,image enhancement and other technologies are used to process the instrument image,which effectively improves the quality of the instrument image.According to the characteristics of digital instruments,liquid level instruments and pointer instruments,K-Nearest-Neighbor classification(KNN)method,proportion method and coordinate mapping method are used to identify the readings,and the recognition accuracy meets the requirements of oil field patrol meter reading.(2)Research on reading recognition of pointer meter based on deep neural network.Aiming at the problems of many steps,many parameters,low fault tolerance rate and poor robustness in the pointer instrument reading recognition methods based on image processing technology,and the problems of only using the network for instrument detection,reducing environmental interference and failing to fully exert the powerful fitting ability of the network,A pointer meter reading recognition method based on the combination of generative adversarial network(GAN)and fine-grained classification network is proposed.Firstly,the deep convolutional generative adversarial network(DCGAN)is used to generate large-scale sample sets for fine-grained classification network training on the basis of limited pointer meter images.Secondly,the generated large-scale sample set is used to train the fine-grained classification network.Finally,the trained fine-grained classification network automatically recognizes the readings according to the input pointer meter image.This method transforms the reading recognition task into fine-grained classification task,completes the end-to-end recognition process of "image-reading",simplifies the recognition steps,and has high robustness and recognition accuracy.In order to verify the effectiveness of the instrument reading recognition algorithms studied in this paper,the oilfield instrument reading recognition method based on image processing technology and the pointer instrument reading recognition method based on deep neural network are verified by applying the real instruments collected in the field.The experimental results show that the 94.93%recognition accuracy can be achieved by KNN,the relative error of proportion method and coordinate mapping method are only 3.68%and 0.43%respectively compared with manual meter reading,which meets the inspection requirements of oilfield development site.The accuracy of the recognition method based on deep neural network reaches 98.23%,which simplifies the process and improves the robustness.