Research On Inner Surface Defect Detection Method Of Pressure Pipe Based On Video Ball

The traditional pressure pipe inner surface defect detection device has a higher demand for the shape and structure of the detection pipeline,and the detection method is time-consuming and laborious.This thesis designs a kind of spherical video internal detection device,and a method based on depth convolution neural network and region convolution neural network is carried out to detect the inner surface defects of pipelines,the specific work includes the following aspects:(1)The video ball detection device is designed and produced,and the inside of the pressure pipe is taken.The device comprises a camera,a light source,a power supply battery,a spherical shell and a bracket,which can be used for 720 times to record the defects of the inner surface of the pipe and the weld seam,etc..The video ball suspends in the water by buoyancy,the requirements of the shape and structure of the pipe are not strict,so it can take photo in any pipeline as long as the water flow speed is appropriate.(2)Research determined the appropriate range of travel speed of the video ball in pressure pipe.In the shooting process,two consecutive frames of video balls are required to not be omitted as far as possible,and the repeated parts of the two consecutive frames are required to be no more than 50%.Based on the camera angle and frame rate data,according to the speed formula,the minimum velocity of 0.038m/s and the maximum 0.042m/s of the video ball running in the pipeline is calculated in this thesis,which provides guidance for the subsequent flow velocity setting.(3)The improved model of convolution neural network based on deep learning is used to classify the pipeline images and find out the pipeline images with defects and weld seams.On the one hand,the pipeline image database is established,and the convolution neural network model based on deep learning is trained,and the test set is verified.On the other hand,the HOG+SVM,LBP+SVM,SIFT+SVM algorithms are adopted,detection effects are compared,results show that the convolution neural network classification method based on deep learning has the highest accuracy rate of 78.3% for pipeline image classification.(4)Two kinds of models of Fast-RCNN(fast regional convolution neural network)and Faster-RCNN(faster regional convolution neural network)are used to mark pipeline defects,which can be used to determine the defect grade by pipeline inspection personnel.Manually annotating 6000 pipeline defect images for training and modifying the full connection layer parameters of these two models make the two models more suitable for pipeline image processing.The experimental results show that the accuracy of the two models has little difference,which is both about 70%,but the detection speed of the latter is faster than that of the former,which is more suitable for real-time detection.