Research On X-ray Weld Defect Detection System Based On Parallel Computing

Pipeline safety is the basis of safe and reliable operation of energy industry.With the increase of environmental protection in our country,more and more urban thermal power plants are replacing coal with gas,and pipeline construction is ushered in a great development period.Welding quality is an important factor affecting pipeline safety.Among all kinds of welding quality testing methods,X-ray testing is the most widely used technology.However,the traditional manual film evaluation mode not only requires a lot of work,but also the detection results are affected by subjective factors of the detection personnel.In order to improve the detection accuracy,automatic detection using image processing technology has become a hot topic in current research.In this paper,the X-ray girth weld image of pipeline was taken as the research object,and sparse description technique and deep convolutional network were proposed to detect different types of defects.In order to improve the computing speed and meet the actual needs of the field,parallel computing is introduced to improve the real-time and accuracy of detection.This paper mainly studies from the following five aspects:(1)In view of the problems of excessive noise,low contrast and difficult direct processing of X-ray weld images,this paper firstly adopts the median filtering method for noise reduction,combined with gamma transform enhancement,to improve the image clarity.Then,the maximum inter-class variance method is combined with the Sobel operator to extract the region of interest accurately.Finally,the density clustering algorithm is used to segment the suspected defect region accurately.(2)Aiming at the problems of small relative area such as cracks and circular defects represented by pins and needles,which are difficult to be identified by conventional methods,the sparse description technique is combined with dictionary learning technique,and the dictionary model construction algorithm based on expectation projection is used to construct the dictionary matrix of small defects.In recognition,sparse solving technology is used to fit the image to be detected,and the sparse coefficient is used to determine whether the suspected image is a defect.(3)Aiming at the identification problem of defects with large area such as concave,underwelded and unfused,the Inception module was combined with Res Net network to design a deep convolutional neural network model suitable for the identification of defects with large area.The model can extract the deep features of the image well,and the network parameters are optimized by using the cross entropy loss function and the adaptive moment estimation algorithm.Experimental results show that the model has good robustness and high defect recognition rate.(4)In order to improve the real-time performance of defect detection,this paper adopts GPU parallel acceleration technology based on CUDA platform to optimize the image preprocessing module and improve the image processing speed.In addition,the parallel running of the two recognition algorithms is realized by using the function of CPU multi-core and multi-threading,which promotes the high efficiency of the system.(5)According to the actual industrial testing requirements,this paper developed an intelligent identification system for weld defects and passed the eyewitness test of SGS-CSTC Standard Technical Services Co.,LTD.The system can accurately detect defects without manual intervention,with comprehensive coverage of detection types,fast detection rate and high recognition accuracy,which effectively relieves the working pressure of detection personnel.