Engineformer: A digital twin model for predicting aero-engine performance and degradation

Health management plays a critical role in ensuring the safety and reliability of engine operation. This study proposed Engineformer, a novel digital twin modelling architecture for aero-engines that integrates the transformer framework with prior physical knowledge. The compression components were modelled using the transformer encoder, whereas the expansion components were modelled by the decoder. A cross-attention mechanism was employed to extract the interactive features between compression and expansion components. To better adapt to engine data, the feedforward layers in both encoder and decoder used one-dimensional convolutional layers. The proposed Engineformer model was evaluated through two case studies: exhaust gas temperature (EGT) and remaining useful life (RUL) prediction. In EGT prediction, the model was tested on flight datasets from four civil high-bypass-ratio engines and achieved average mean absolute relative errors of 1.88%, 1.78%, 0.48%, and 0.49% across 10 training rounds. In RUL prediction, based on the DS02 subset of the N-CMAPSS dataset, Engineformer achieved a minimum root mean square error of 4.142 over 10 training rounds. This reduced further to 3.708 after optimization. The results demonstrated that Engineformer achieved state-of-the-art performance in both tasks. This verifies its reliability for predicting aero-engine performance and degradation.