|
Laser metal deposition (LMD) is an additive manufacturing technique that utilizes powder as its material. The powder is transported through the nozzle coaxially, where it converges with the laser beam onto the surface of the substrate. With the assistance of auxiliary gas, the powder melts upon laser irradiation and is deposited layer by layer onto the substrate, forming the desired component. Real-time monitoring of the deposition height plays a crucial role in enhancing the precision of LMD, reducing defects such as edge collapse and surface unevenness. It represents one of the fundamental aspects in achieving high-quality metal additive manufacturing. In this study, a laser metal deposition height prediction method based on a multi-modal neural network was proposed. The network architecture consisted of a convolutional neural network (CNN) and a fully connected network (FCN). The CNN extracted and analyzed the characteristics of the molten pool, generating a feature vector. This feature vector, along with the molten pool temperature, was fused as input into the FCN, ultimately predicting the deposited height. Compared with the predicted results of support vector regression (SVR), multi-modal neural networks can quickly predict the deposited height and track their changing trends. The model achieves a remarkable prediction accuracy of 95% and exhibits robustness in handling outlier values. The proposed network framework holds considerable potential in facilitating real-time control and fine-tuning of the laser metal deposition process.
|
