Desarrollo de técnicas de Inteligencia Artificial (IA) para su implementación en los procesos de aporte por deposición de energía focalizada (DED) para fabricación aditiva y soldadura

The metallic products manufacturing sector has consolidated itself as a strategic area within the European industry, undergoing a constant transformation aimed at improving the competitiveness, reliability, and sustainability of processes and products. Within this context, the present Doctoral Thesis seeks to implement novel Artificial Intelligence (AI) systems to enable the optimization of Advanced Manufacturing (AM) processes. The main objective of this research is to expand knowledge—particularly in the prediction of part geometry, the monitoring of the process, and the real-time control of additive manufacturing production. This approach aims to overcome the current limitations in dimensional accuracy, prevent the appearance of defects, and provide reliable and cost-effective industrial monitoring systems for the market. The Thesis is structured as a compendium of three scientific publications, each focused on a specific objective but all contributing to the common goal of the research. In the initial phase, a predictive model was developed that achieved highly reliable results in estimating the bead geometry of Invar deposits fabricated on the Addilan V0.1 system, representing the first step toward the digitalization of WAAM. In the second phase, the influence and significance of process parameters and substrate thickness on bead geometry and the Heat-Affected Zone (HAZ) were analyzed through the implementation of a symmetric neural network. This architecture was designed to enhance prediction robustness under varying deposition conditions by correlating input parameters with bead dimensions and thermal phenomena linked to material integrity. The results demonstrate that the network accurately predicts the main geometric parameters (height, width, and penetration) while also capturing critical thermal effects that influence structural stability. In the final phase, a non-intrusive, low-cost in-line quality control system was developed, based on the transformation of electrical signals into spatial representations. The signals recorded during the process were converted into heatmaps and Markov Transition Images (MTI), which were processed using a pre-trained ResNet-18 convolutional neural network. This method enabled accurate real-time defect detection, achieving 94% classification accuracy with inference times below 20 ms, fully compatible with layer-by-layer manufacturing. The integration of the three studies provides a comprehensive view of the digitalization of WAAM through Artificial Intelligence, structured around a natural and continuous prediction → modelling → control cycle. In the first stage, the virtual sensor predicts bead geometry; in the second, the symmetric neural network contributes to process modelling by linking deposition parameters with thermal behaviour and material response; and finally, the convolutional network closes the loop by performing real-time quality control and defect classification. Experimental validation was carried out on the Addilan V0.1 machine using two key materials—Invar and stainless steel 316L—demonstrating the applicability and robustness of the proposed methods and contributing to the technological advancement of metal additive manufacturing. Overall, the results of this Thesis pave the way for the development of industrializable, precise, and cost-effective monitoring systems, capable of ensuring real-time quality, reducing rework, and improving the overall efficiency of WAAM processes. In this way, the research contributes to both the theoretical and practical advancement of the technology, reinforcing its significance within the framework of Industry 4.0 and its role in the transition toward intelligent manufacturing systems.