An uncertainty-aware deep neural network framework integrated with adapted engine modeling for robust sensor fault detection in aero gas turbine engines: validation using flight test data

초록

This study proposes an uncertainty-aware deep neural network (DNN) framework integrated with adapted engine modeling for robust sensor fault detection in aero gas turbine engines. In practical applications, sensor fault detection is challenged by the absence of true reference values and the presence of engine-to-engine performance variation caused by manufacturing tolerances, degradation, and maintenance conditions. To address this limitation, a physics-based engine model is employed to generate large-scale datasets incorporating component performance uncertainty. These analytically generated datasets are used to train DNN models that learn the nonlinear relationship between operating conditions and expected sensor behavior under healthy conditions. A key feature of the proposed approach is that no flight test data are used during model training. The model is trained exclusively on simulation-based datasets, eliminating the need for large volumes of experimentally accumulated data. The trained models are first evaluated using simulation datasets not used during training, demonstrating strong regression performance with normalized root mean squared errors below approximately 1.5% under ±1% component performance variation. To validate real-world applicability, flight test data are used solely for independent validation and are not involved in the training process. Sensor fault scenarios are constructed by introducing controlled bias errors into measured signals. When evaluated on these flight test–based datasets, the model achieves classification accuracies of 93.4%, 93.5%, and 87.5% for DNN models trained under ±1%, ±3%, and ±5% component performance variation, respectively. Furthermore, because the proposed method relies only on a pre-trained DNN model during operation, no physics-based engine model is required in the on-board environment. This enables fast and stable inference suitable for real-time applications. The results demonstrate that the proposed framework successfully generalizes from simulation-based training to real-world conditions while maintaining reliable sensor fault detection performance. These characteristics make the approach highly suitable for practical digital twin and real-time health monitoring applications in aero gas turbine engines. © 2026 Elsevier Masson SAS.

키워드