Toward Individualized Digital Twins for Aero-Engines: A Physics-Informed Framework With Fleet-Individual Cross-Attention and Feature Transfer

Abstract

Abstract Digital Twin (DT) technology has become a cornerstone for cyber-physical aero-engine systems, enabling full lifecycle digital replication and intelligent management. To address challenges of model adaptability and generalization under restricted access to fleet data conditions, this paper proposes a physics-informed DT modeling framework integrating fleet feature transfer and a fleet-individual cross-attention fusion mechanism. The framework is composed of three main components: (1) a physics-embedded network, where a long short-term memory (LSTM) network captures temporal component dynamics, coupled with a multi-component attention module to model inter-component interactions; (2) a fleet-individual cross-attention module designed to extract interactive features between individual engines and the entire fleet; and (3) a pretraining and finetuning strategy that transfers generalized knowledge from fleet-level models to individual engines, enabling fast deployment without relying on large-scale historical data. Validation on thrust and exhaust gas temperature (EGT) prediction demonstrates superior accuracy and robustness, with best-case mean absolute percentage errors (MAPE) below 0.5% for thrust and 2.6% for EGT across three datasets. Compared with baselines (a general fleet model and individual models without fleet features), the proposed approach reduces average MAPE by 47.80-75.49% (thrust) and 7.47-41.33% (EGT), improving both precision and stability. This framework enhances the scalability and digital management efficiency of aero-engine DT systems.