| The laser additive manufacturing technology of metals has been listed as one of the most cutting-edge and promising technologies by its characteristics of rapid,customized,digital,and networking.Compared with traditional manufacturing of reduced material,it can process complex structural parts with high density and precision without tooling and cutting tools,and greatly reduce the production cost.Selective laser melting technology is the most widely used and useful rapid prototyping technology in laser additive manufacturing,which is widely studied at home and abroad.In the process of selective laser melting forming,there are many interference factors in the molten pool images,such as small diameter of molten pool spot,easy occurrence of mirror-like refraction,liquid splash and plume around,and insufficient hardware equipment conditions.There is a large amount of noise and low resolution in the molten pool images collected by the online monitoring system.If the molten pool features are extracted directly,there will be a large error,which will affect the real-time control of processing parameters.In order to solve the above problems,this dissertation puts forward the research of super-resolution reconstruction of molten pool images by selective laser melting,and the main research contents are as follows:(1)In Chapter 3,we propose a kind of residual channel attention network which can reconstruct high resolution molten pool images.This network solves the difficult problem of deep network training by introducing the residual module,and uses the residual nesting to construct deeper networks to provide a larger receptive field.In order to reduce the number of parameters in the network,the BN layer in the traditional residual module is removed and the ECA-Net attention mechanism is added.Compared with the images in natural scenes,it is difficult to collect paired data sets of high and low resolution molten pool images at the same moment and from the same Angle.Therefore,in this dissertation,high-speed near-infrared camera is used to collect HR molten pool images and act as the label of super resolution reconstruction network.The corresponding LR molten pool images are obtained by degradation.(2)In Chapter 4,we propose an attention and feature fusion network which can reconstruct high resolution molten pool images.Since the single chain structure in the super-resolution reconstruction network of molten pool images proposed in Chapter 3 ignores the fusion between shallow features and deep features,in Chapter 4,a superresolution reconstruction network of molten pool images based on attention and feature fusion is designed based on U-Net network structure.In order to maximize the utilization rate of output features of each network layer,the shallow network on the left and deep network on the right of U-shaped network structure are connected by skip connection mechanism,and the shallow and deep features are integrated.Since it is easy to extract a large amount of noise information during network training,in order to reduce the influence of noise on the quality of molten pool images reconstruction,the attention mechanism of channel and space transformation is added to each layer and skip connection part of the network.(3)In order to verify that the molten pool images reconstructed by super-resolution not only improves the resolution,but also greatly improves the extraction accuracy of molten pool features,this dissertation innovatively proposes an evaluation index of molten pool feature error rate.The HR,LR and SR molten pool images are preprocessed respectively and the geometrical and shape features of molten pools are extracted.The molten pool features of HR molten pool images are taken as reference standards,and the error rates between the molten pool features of LR and SR molten pool images are calculated in turn.The results show that the reconstruction of molten pool images using super-resolution network can greatly improve the accuracy of molten pool features extraction. |
PDF Full Text Request