Reading Recognition On Pointer Instrument In Power Plant Based On Automatic Inspection Vehicle

Pointer instrument is widely used in electric power,chemical industry and other fields because of its high reliability,convenient and fast use and maintenance,low price and suitable for all kinds of harsh use environment.At present,there are still many pointer instruments in the power plant,such as barometer,thermometer,oil temperature gauge,oil pressure gauge,etc.Now,the data recording of these pointer instruments mainly rely on the close observation manually.It has many problems,such as high labor cost,scattered data and strong subjectivity,and hard to risk information in time.According to the requirements of information management of power enterprises,it is necessary to automatically read the readings of these instruments.Moreover,these pointer instruments cannot realize the automatic collection and transmission of measurement data because they do not have communication interface.Therefore,it is urgent to carry out the application research of automatic patrol data and collection technology of pointer instruments.This paper combines image processing technology with cloud technology,and takes the raspberry cart as the automatic inspection hardware platform,studies and designs the instrument reading recognition system and cloud platform communication system,in order to provide technical support for the automatic inspection and data reading of pointer instruments in power plant.The specific research contents are as follows:(1)A lightweight YOLOv4 instrument panel detection model is designed.Aiming at the automatic patrol inspection platform with limited computing power,the lightweight improvement is made on the YOLOv4 target detection model,the original backbone feature extraction network cspdarknet53 of YOLOv4 is eliminated,the lightweight MobileNetv3 network is integrated into the YOLOv4 network structure,and the standard convolution in the PANet network is replaced by the deep separable convolution.The lightweight model has a resolution of 320 on the NVIDIA Geforce GTX 1050ti 4G graphics card×320 images can reach the detection rate of 22Frames Per second/s(FPS),and the mean average accuracy(mAP)can reach 99.92%;(2)The algorithm of dividing instrument pointer and scale line and the algorithm of instrument reading are designed.For the reason that the image quality is not high due to the influence of external factors,the segmentation pointer and scale line algorithm not only ensure the accuracy,but also take into account the computing power of the automatic inspection platform,replace the Xception part of the backbone network of DeepLabv3+with the lightweight network MobileNetv2 integrated with hole convolution.The improved model has a resolution of 320 on NVIDIA Geforce GTX 1050Ti 4G graphics card×The image detection rate of 320 is 21 FPS,and the average intersection union ratio(mIoU)is 89.77%.Based on the segmentation information,the tilt correction is carried out by perspective transformation,and the corrected image is spatially transformed and combined with the distance method,which can realize the reading of the instrument and the reference error is less than 4%;(3)Build a cloud platform test system.Take the cloud server as the core,realize the communication between the server and the inspection car platform,and store the data obtained from the inspection car image processing to the cloud server database or local;Develop WeChat public platform to read instrument inspection results according to instructions issued by users;(4)An instrument reading system based on raspberry pie car is developed.In the laboratory,imitate the position of the power plant instruments,and realize the functions of the automatic inspection platform in the simulated power plant environment,such as automatic reading of instrument reading data,data transmission to the cloud server and viewing inspection information on the mobile WeChat client through the industrial camera,camera pan tilt and other hardware equipment carried by the automatic inspection platform,and the reference error of the final instrument reading is less than 4%.