A monitoring framework for predicting laser directed energy deposition property

Abstract

In Laser directed energy deposition (LDED) additive manufacturing, challenges such as porosity, surface defects, cracks, and the complex relationship between melt pool dynamics and mechanical properties still impede consistent quality control. Traditional monitoring and prediction remain fragmented and signal-specific, limiting early defect discovery and degrading reliability in safety-critical parts. To tackle these limitations, this study introduces a novel data-driven framework integrating multi-level feature fusion and dual-task learning, which significantly improves LDED process monitoring and prediction. This work proposes ResCIFNN, a ResNet-based framework that couples unsupervised, defect-aware clustering with supervised regression under a sliding time-window, enabling defect-informed prediction of tensile behavior. Utilizing melt pool infrared images, simulation data, and quantitative features, ResCIFNN achieves a precise mapping of melt pool dynamics to mechanical properties. Five-fold validation shows robust clustering (Silhouette = 0.7588; DBI = 0.3480). For tensile property prediction, ResCIFNN delivers an RMSE of 0.1113 and R² of 0.9875, surpassing ResNet18 (RMSE = 0.2821, R² = 0.9207) by reducing RMSE by 60.5 % and improving R² by 0.0668. Robustness tests under noise/occlusion yield RMSE ≤ 0.2011; Grad-CAM highlights high-temperature cores and edges, reinforcing interpretability. This pioneering approach not only elevates defect classification and mechanical property prediction but also provides a scalable, interpretable solution for quality assurance, with broad potential for additive manufacturing.