Research On Key Technologies Of Weld Bead Diagnosis For Supercritical Unit

Weld beads are usually formed inside water-cooling pipes when the pipes are connected out of supercritical boilers,reducing the operating efficiency of supercritical units.Radiographic imaging technique is commonly employed to observe the shape of weld beads and manual observation method is used to evaluate their effects.Aiming at avoiding the low efficiency and randomness results of manual observation,this paper researches on algorithms of welding area detection,weld seam segmentation,weld surface reconstruction and weld bead diagnosis.An intelligent diagnosis system for weld beads on radiographic images of water-cooled pipe is constructed based on computer vision and deep learning technology to independently judge the influence of weld beads.The main contents of the paper are as follows:Firstly,the composition of radiographic system and the principle of double-wall doubleshadow radiography are studied based on Beer’s law.The morphological features of the projected weld beads in the welding area of the radiographic image are analyzed.To deal with the small scale and complex posture of welding area in the radiographic image,preprocessing technologies like image inversion,filtering,and enhancement are employed to improve the image quality.Secondly,a welding area detection model is established by combining rotated rectangular anchor and circular step label,and the model is simplified by regression of the shitf value of angle.YOLO5 network is improved to realize the automatic prediction of position and angle of welding areas,and perspective correction technology is used to extract welding areas from entire radiographic image.Then,to solve the problem of fuzzy and discontinuous boundary of weld seam,an boundaryaware semantic segmentation model is established by introducing signed boundary distance field,which improves model’s distinguish ability between object and edge.Semantic segmentation network is improved to realize the prediction of signed boundary distance fields and weld seam masks.At the same time,results of welding area detection and weld seam segmentation are combined to construct the feature space,and a support vector classifier is designed to realize the preliminary diagnosis of weld beads,eliminating the influence of extreme welding beads on intelligent system.Finally,to figure out the problem of shape distortion in radiographic imaging process,level set theory is introduced to build a three-dimensional weld surface model,and homography matrix is employed to simulate the radiographic imaging process.By combining two-dimensional image features and prior knowledge,several energy functions are designed to realize the autonomous optimization of level set function and reconstruct the three-dimensional weld surface.Furthermore,height error and area error indicators based on concept of flow surface are proposed to analyze the characteristics of three-dimensional weld beads,realizing the quantitative diagnosis of weld beads.Experiments show that the automatic diagnosis system based on computer vision and deep learning can overcome the randomness and subjectivity of manual evaluation,and make the diagnosis results of weld bead more objective and stable.At the same time,the speed and efficiency of the system are much higher than manual evaluation system,which has great significance for industrial practice.