Physics-guided generative data augmentation for vision-based signal processing under class-imbalanced conditions in directed energy deposition monitoring system

• Addresses signal processing limits from class-imbalanced DED molten pool signals. • Physics-guided CVAE-GAN synthesizes high-fidelity visual signals. • Thermodynamic losses enforce spatial and spectral signal consistency. • Augmented data improves decision boundaries for rare defect detection. • Validated on a real-world aerospace component with complex geometry. Vision-based signal processing is critical in metal additive manufacturing (AM), particularly within Directed Energy Deposition (DED), for capturing the rapid thermo-fluid dynamics essential for detecting process instabilities and ensuring structural integrity. While DED operates as an open, inherently non-stationary system and molten pool visual signals carry rich spatiotemporal information, effective processing is fundamentally constrained by the scarcity of anomalous events relative to stable deposition. This data sparsity leads to imbalanced training sets that fail to represent the full spectrum of transient process dynamics, causing classifiers to misclassify rare but critical failure modes and degrading overall monitoring performance. Crucially, conventional data augmentation techniques, being agnostic to the underlying physics, often introduce signal artefacts that violate governing thermodynamic constraints, thereby degrading the physical validity of the training data. To bridge the gap between data-driven feature learning and physical principles, this study proposes a physics-guided generative data augmentation framework based on a Conditional Variational Autoencoder–Generative Adversarial Network (CVAE-GAN). The framework explicitly learns the intrinsic physical manifold of the process, incorporating thermodynamically motivated loss functions to ensure the synthesis of high-fidelity, physically coherent molten pool visual signals. Validation through both spectral analysis and downstream classification confirms that the augmented signals preserve essential morphological and textural properties while significantly improving the decision boundaries of sequence classifiers. A case study on a complex aerospace component demonstrates the framework’s practical utility in enhancing robust in-situ monitoring under varying geometric conditions. By addressing the sparse representation of critical dynamics through a physics-aware methodology, this work offers a reliable pathway for advanced signal processing in data-scarce engineering systems.