Extraction Of Molten Pool Morphology During Laser Addictive Manufacturing Processing Using Fully Convolutional Neural Network

Laser additive manufacturing is an advanced manufacturing technology.It is formed by layer-by-layer cladding,which provides an effective means to produce complex shapes,functional gradients or customized parts.For laser additive manufacturing of metal materials,high temperature molten pool melt the metal powder transported in situ or the metal powder spread on the surface of the substrate to complete the multilayer cladding.For the laser additive manufacturing of metal materials,the high-temperature molten pool melts the metal powder conveyed in situ or the metal powder spread on the surface of the substrate to complete the multilayer cladding.The shape information of the molten pool such as width and length is related to the process parameters such as laser power,scanning speed,etc.The fluctuation of the molten pool will directly reflect the instability of the processing process;in addition,the performance of the forming process can also be predicted through the information of the molten pool shape Indicators,such as the width of the molten pool can reflect the degree of dilution and the quality of metallurgical bonding,the shape of the tail of the molten pool can predict the direction of grain growth during solidification,and the length of the molten pool can be used as an evaluation index for heat accumulation at the same scanning speed.Therefore,accurate acquisition of molten pool morphology information is of vital importance to laser additive manufacturing.Laser additive manufacturing conditions are complex,and the monitoring of molten pool images will be interfered by powder,mushy zone(oxide or incompletely melted powder),reflections,plasma metal fumes,etc.,and there are huge challenges in monitoring the morphology of the molten pool.Compared with the traditional passive molten pool monitoring method without auxiliary lighting,the image obtained through active lighting and band-pass filtering can effectively improve the image quality.In this paper,by selecting the appropriate auxiliary light source,filter,adjusting the incident angle of the auxiliary light source,etc.,a synchronization monitoring system for the laser processing molten pool was built to achieve the acquisition of clear molten pool images.Traditional algorithms such as threshold segmentation and phase consistency suitable for passive monitoring images are difficult to extract accurate molten pool shape information due to the reflection of the auxiliary light source.The image segmentation algorithm based on the fully convolutional neural network can learn the local and global features of the image,and realize the robust feature extraction of the molten pool.In this paper,PSPNet is selected as the baseline network after considering the morphological characteristics of the brightness,size,and shape of the molten pool,and the original baseline network is improved by adding sc SE Block,adding a selfdesigned decoding module,and improving the loss function.In the case of multi-layer cladding with mushy zone and powder occlusion,and high power with reflection and plasma interference,the algorithm in this paper and the original PSPNet and U-net are accurately extracted for the shape,length,width and other information of the molten pool.The results show that the improved fully convolutional neural network algorithm effectively improves the prediction of the width,length,shape of the molten pool,especially the shape of the tail.In addition,the algorithm in this paper is used to monitor the morphology of single-layer cladding and multi-layer cladding under different power conditions.The results show that this algorithm can accurately extract the morphology of single-layer and multi-layer cladding.The length and width of the molten pool.