Detection of weld groove edge based on multilayer convolution neural network

Abstract

To detect the weld groove edge interfered by noise, a lightweight multilayer convolution neural network is designed. The network can extract multi-layer features, which improve the resolution of weld groove edge position detection and have strong anti-interference ability. The position of weld groove edge is recognized and located by generating prior box region. The global detection and location of weld groove edge is transformed into confidence prediction and location coordinate value regression in local prior box region, which ensures the accuracy and robustness. The training data set covers different interference noise and groove deformation to improve the generalization ability of the network under different interference conditions. The experiments show that the designed network has good detection accuracy under the condition of strong interference, with an average error of 0.238 mm and a maximum error of 0.702 mm. The processing time is 0.09 s, which shows a good application prospect in field welding.