Investigation of the mechanisms of structure-induced acoustic modulation in WA-DED metal additive manufacturing with an adaptive modelling approach

As metal additive manufacturing (AM) advances toward intelligent and data-driven paradigms, reliable sensing has become essential for enabling real-time monitoring, process control and decision making. Among various sensing modalities, acoustic sensing offers unique advantages including high temporal resolution, sensitivity to arc dynamics, and low deployment cost, which have resulted in its widespread adoption in industrial applications. However, most existing acoustic-based methods are developed under static or simplified conditions, without accounting for the dynamic structural evolution inherent in metal AM. This often leads to significant performance degradation, especially when applied to large-scale, high-layer components. This study first identifies and investigates a critical but underexplored phenomenon— structure-induced acoustic modulation —in which the progressive accumulation of volume and variation of geometry throughout the build alters the vibration characteristics and acoustic transmission paths, thereby modulating the emitted acoustic signals. Subsequently, the existence of this modulation and its adverse impact on acoustic-based applications are confirmed through controlled experimental validation. To address this challenge, a dedicated framework based on a ConvNeXt-Adapter-Transformer architecture is then proposed to model the temporal evolution of acoustic signals and enable adaptive generalization across varying builds. The framework is evaluated through the fabrication of a 203-layer circular hollow section component as a case study, demonstrating its ability to effectively adapt to long-duration, structure-evolving metal AM processes. Ultimately, building on this result, a practical acoustic-based application—fabrication process tracing—is deployed and integrated with the proposed framework. The integrated system is further validated through the fabrication of a gear shaft component with complex structures, where results show that the integrated system maintains reliable performance under conditions where baseline systems fail. Specifically, the integrated system achieves a 39.3% improvement in process tracing accuracy compared to the baseline. Overall, this pioneer study presents the first systematic analysis of structure-induced acoustic modulation in metal AM and introduces an adaptive modelling framework to effectively mitigate the adverse effects of this modulation. Beyond the specific case studies, the identification of structure-induced acoustic modulation itself constitutes a generic scientific finding, revealing that acoustic signals in layer-by-layer AM are inherently modulated by structural evolution. The proposed ConvNeXt-Adapter-Transformer framework further provides a transferable modelling strategy with strong potential for broad adoption in acoustic signal-based applications across the metal AM domain and beyond, enabling reliable performance throughout the manufacturing process and enhancing system robustness under industrial conditions. • Structure-induced acoustic modulation identified as a generic metal additive manufacturing phenomenon. • Adaptive ConvNeXt-Adapter-Transformer enables robust and transferable acoustic signal modelling. • Macro photographs confirm defects and align with waveform and Log-Mel spectrogram features. • Conditional variational autoencoder generates high-layer anomalous spectrogram data to address scarcity. • Two large wire arc directed energy deposition parts validate robustness and reliability of the framework.