Toward efficient digital twin simulation: A causal representation learning approach

In recent years, digital twin (DT) technology has emerged as a focal point in the field of shaft system prognostics and health management. To reduce simulation time cost and computational overhead, data-driven intelligent data generation algorithms have been employed as surrogates for traditional finite element simulations. However, such algorithms are typically constrained to generating in-distribution data within known operational domains and fail to generalize to out-of-distribution data under unseen conditions, which significantly hindering the development of DT model under variable operating scenarios. To address this limitation, this paper proposes a novel causal factorization–recombination network (CFRN) for generating shaft vibration responses under previously unseen operating conditions. Firstly, the structural causal model (SCM) for shaft vibration response is constructed to encode the causal mechanisms linking two critical operational parameters with vibration responses. Based on the SCM, a dual-encoder architecture is developed. By optimizing causal consistency loss, causal independence loss, and reconstruction loss, the model identifies latent mediators associated with the two causal factors. Additionally, a novel bidirectional cross-attention mechanism is introduced to equitably integrate mediators corresponding to different combinations of causal factors, enabling robust feature representation under unseen operational conditions. Finally, the recombined features are utilized to synthesize vibration response data. The proposed CFRN is validated using a shaft system simulation dataset. Extensive comparative experiments demonstrate that the generated data under unseen conditions by CFRN achieves 98.06% accuracy on crucial frequency. The proposed approach offers a novel paradigm for accelerating simulation response in DT frameworks.