In the era of Industry 4.0 and smart manufacturing, Wire Arc Additive Manufacturing (WAAM) stands at the forefront, driving a paradigm shift towards automated, digitalized production. However, online simulation remains a technical barrier toward building a Digital Twin (DT) for metallic AM due to the prolonged computing time of numerical simulations and limitations in accuracy of current data-driven models. This study addresses these issues by introducing an adaptive online simulation model for predicting distortion fields, utilizing a diffusion model architecture for distortion process modelling with a Vector Quantized Variational AutoEncoder coupled with Generative Adversarial Network (VQVAE-GAN) backbone for spatial feature extraction, complemented by a Recurrent Neural Network (RNN) for time-scale result fusion. Pretrained offline with Finite Element Method (FEM) simulated distortion fields, the model successfully predicts distortion fields online using laser-scanned point clouds during the deposition process. Experimental validation on seven thin-wall structures demonstrated its superior performance, achieving a Root Mean Square Error (RMSE) below 0.9 m, outperforming FEM by 143 % and Artificial Neural Networks (ANN) based methods by 151 %, marking a significant stride towards realizing an AM-DT.

在工业4.0和智能制造时代，电弧增材制造（WAAM）站在最前沿，推动着自动化、数字化生产的范式转变。然而，由于数值模拟的计算时间较长以及当前数据驱动模型的准确性有限，在线模拟仍然是构建金属增材制造数字孪生（DT）的技术障碍。本研究通过引入用于预测失真场的自适应在线仿真模型来解决这些问题，利用扩散模型架构进行失真过程建模，并使用矢量量化变分自动编码器与生成对抗网络（VQVAE-GAN）主干进行空间特征提取，并辅以用于时间尺度结果融合的循环神经网络（RNN）。该模型通过有限元法 (FEM) 模拟畸变场进行离线预训练，并在沉积过程中使用激光扫描点云成功在线预测畸变场。对七个薄壁结构的实验验证证明了其优越的性能，实现了低于 0.9 m 的均方根误差 (RMSE)，比 FEM 好 143 %，比基于人工神经网络 (ANN) 的方法好 151 %，标志着朝着实现这一目标迈出了一大步AM-DT。