A porosity prediction method for laser cladding with imbalanced multi-modal data

Abstract

Porosity is one of the most important defects in laser cladding (LC); therefore, the prediction of porosity during LC is crucial. Recently, the prediction process has been conducted on images or temperatures from only the melt pool, which hinders the prediction performance due to the incomplete characterization of defects from single-modal data. Multi-modal methods can combine images and temperature fields to provide a more complete characterization of porosity defects, but this approach is limited by data imbalance and data interaction. In this study, a novel multi-modal imbalanced data prediction network (MIDPN), which contains a data mapping module, a feature extraction module and a feature interaction module, is proposed to address these issues. For data imbalance, the data mapping module of the MIDPN uses an improved generative adversarial network that maps image data to the temperature field, with the purpose of generating missing temperature field data to resolve data imbalance. In the feature extraction module, convolutional neural networks are utilized to extract the image feature and temperature field feature of the melt pool. For data interaction, the MIDPN establishes a guided attention mechanism for image-temperature field features to facilitate the interaction of features from both modalities and utilizes the fully connected layers to obtain porosity prediction results. The experimental results show that the proposed method achieves an accuracy of 96% in predicting porosity, which is better than those of the other methods. Overall, this study significantly improved the reliability of porosity prediction, enabling more accurate quality assessment for LC monitoring, and promoting proactive defect prevention.