Indrnn-Based Data-Driven Modeling Integrated With Physical Knowledge for Engine Performance Monitoring

Abstract Monitoring the whole performance status of aircraft engines is of paramount importance for ensuring flight safety, control system, and prognostic health management. This work introduced an aircraft engine deep learning (DL) model that integrated with engine physical knowledge. First, component networks were established for each engine component (e.g., fan, turbine, nozzle) using the independently recurrent neural network (IndRNN), self-attention mechanism, and residual network. Subsequently, based on the physical spatial alignment of engine components, the data transfer between component networks was determined to establish the whole engine model. Case studies were conducted on exhaust gas temperature (EGT) prediction for two civil aircraft engines and thrust prediction for another two turbofan engines. When processing the actual engine running data, the data augmentation method was invested to address the issue of nonuniform distribution of engine working states in the training data. Compared with three pure data-driven models based on IndRNN, recurrent neural network, and long short-term memory (LSTM), the model introduced in this work demonstrated superior precision in both steady states and transient states. Specifically, the achieved mean absolute relative error (MARE) was 0.54% for EGT prediction and 0.41% for thrust prediction. When adjusting the time-steps, the introduced model showed steadier predictions with minimal MARE fluctuation compared to the three pure data-driven models, enhancing overall predictive stability.