Droplet transfer behavior is a key factor affecting the forming quality and final performance of wire arc additive manufacturing (WAAM). Therefore

online monitoring based on arc voltage signals is of considerable significance. However

under complex process conditions

the distribution of arc voltage signals tends to shift significantly

while conventional fixed-threshold denoising methods are unable to adapt to dynamic variations in noise characteristics

thereby limiting the accuracy and robustness of cross-condition recognition. To tackle these challenges

this paper proposes a domain similarity-aware adaptive denoising model (DSAAD). The model incorporates an adaptive wavelet denoising (AWD) module that introduces a learnable soft-thresholding mechanism to achieve data-driven noise suppression. It further integrates feature extraction

domain-aware mechanisms

and a joint optimization strategy combining local maximum mean discrepancy (LMMD) with triplet loss

thereby enhancing cross-domain feature alignment and class discriminability. Four cross-domain recognition tasks are constructed based on multi-source working conditions to systematically evaluate the generalization capability of the proposed method. Experimental results demonstrate that the DSAAD model achieves high droplet transition mode recognition accuracy on both chromium-zirconium-copper and pure copper datasets

with the best task accuracy reaching up to 96.72%. Moreover

t-SNE visualization and confusion matrix analyses further validate the effectiveness of the proposed model in optimizing feature distribution and mitigating domain shifts. This study provides new theoretical support and technical pathways for intelligent monitoring and control of droplet behavior in WAAM processes under complex working conditions.