Research On Weld Feature Extraction Methods Based On Deep Learning

With the proposal of the goal of "manufacturing power","quality power" and the development of machine vision technology,automation and intelligence become the inevitable direction of the welding development.The realization of intelligent welding operation depends on a powerful weld tracking system,and weld feature extraction,as a key link in weld tracking technology,directly affects the quality and efficiency of welding.Therefore,it is very important to ensure the accuracy of weld feature extraction.At present,most studies start from the gray distribution feature of weld structural light image,and obtain the structural light fringe with traditional image processing algorithm,and then extract the weld feature points.This method has the advantages of low cost,fast processing speed and good real-time performance,but insufficient anti-interference ability and poor robustness.For weld images with strong arc,splash,smoke and other noise interference or reflective materials,the accuracy of traditional weld feature extraction algorithm is low.Therefore,on the basis of analyzing and studying various traditional weld feature extraction algorithms,this paper focuses on exploring weld feature extraction methods based on deep learning by focusing on two perspectives of weld feature segmentation and weld feature point positioning.(1)Based on traditional weld feature extraction steps,several typical image preprocessing,optical strip centerline extraction and weld feature point extraction methods are analyzed and studied.Then,according to the characteristics of splicing,lap and V-shaped welds,the algorithm with better processing effect is selected to carry out the weld feature extraction experiment respectively,and the shortcomings and limitations of the traditional algorithm are analyzed and summarized.(2)Aiming at the problem that the traditional threshold segmentation method has poor anti-interference ability,the deep learn-based weld feature segmentation method is studied.Combining with the characteristics of weld image,two improvements were made to FCN algorithm.On the one hand,a dual attention mechanism is designed to improve the sensitivity of the algorithm to the weld feature area.On the other hand,the pooling operation is reduced to preserve more details,and the hollow space convolutional pooling pyramid network is introduced to further improve the feature extraction ability of the algorithm.Compared with the classical semantic segmentation algorithms,the experimental results show that the improved FCN algorithm has significant advantages in each evaluation index,stronger robustness and higher segmentation accuracy.(3)In view of the low accuracy and complex process of the traditional extraction method of weld feature points,the extraction method of weld feature points based on deep learning was studied.Taking the location information of weld feature points as the output of the deep learning model,an algorithm based on the RetinaNet is proposed to extract weld feature points.Firstly,a large number of priori boxes are preset for rough positioning of weld feature points,and then the exact location information is obtained through model training.At the same time,take MobileNetV3 as the main feature extraction network of the RetinaNet to reduce the number of model parameters and running time.Experimental results show that compared with traditional algorithms,the proposed algorithm has simple processing flow,higher positioning accuracy,and meets the real-time requirements.(4)In order to realize the independent implementation and application of the proposed algorithm,a software system for weld feature extraction based on deep learning was built,and an online weld feature extraction experiment was carried out combined with related hardware platforms.The results show that the proposed algorithm has strong noise suppression ability and good segmentation effect.In terms of feature point extraction,the average error of the proposed algorithm is 0.133mm,the maximum error is 0.6mm,and the average processing time of each frame is 45ms,which meets the actual demand.