Interpretable generative machine learning model based in-situ process monitoring in robotic wire arc based directed energy deposition of aluminum alloys

WA-DED using CMT is emerging as a high-throughput metal AM strategy, yet it remains susceptible to a variety of thermomechanical instabilities and metallurgical discontinuities. In this study, we present an advanced AE based in-situ monitoring utilizing the generative ML framework to robustly detect and characterize anomalous conditions that compromise part integrity. Specifically, we examine five critical fault scenarios which are overcurrent, high travel speed, insufficient shielding gas flow, combination of overcurrent and low shielding gas flow rate and combination of high travel speed and low shielding gas flow rate elucidate their distinct signatures in the acoustic domain. A rigorous selection of time and frequency domain descriptors is leveraged to train the variational autoencoder, enabling accurate reconstruction of normal process states and efficient outlier detection. Microstructural analyses, encompassing FESEM, Micro-CT, and XRD, validate the detrimental influence of these faults on porosity evolution, grain morphology, and mechanical properties such as UTS. The proposed VAE model demonstrated robust performance across multiple defect types, achieving peak detection accuracies of 87% for overcurrent-induced faults, 85% for high travel speed anomalies, 81% for defects caused by insufficient shielding gas flow, 87% for combined effect of overcurrent and low gas flow rate, and 84% for combined effect of high travel speed and low gas flow rate. Overcurrent anomalies induce coarse columnar grains and high porosity content, while high travel speed amplifies geometric irregularities. Low gas flow conditions foster oxidation induced porosity. The proposed approach achieves high fidelity in detection of these defects, underscoring the synergy between data driven reconstruction errors and material characterization. By integrating unsupervised generative deep learning with domain specific interpretability through feature sensitivity analysis, this acoustic monitoring paradigm provides a scalable and cost effective pathway to detect defects and ensure structural reliability in WA-DED manufactured components. The comprehensive experimental validations and multi-physics correlational insights position this framework as a robust framework for in-situ process monitoring in WA-DED.