AESim: A Data-Driven Aircraft Engine Simulator

Abdellah Madane, Florent Forest, Hanane Azzag, Mustapha Lebbah, Jérôme LacailleIJCAI 2024

Introduction

In the aviation industry, Prognostics and Health Management (PHM) is vital for ensuring flight safety and optimizing maintenance schedules. PHM relies heavily on Continuous Engine Operational Data (CEOD)—multivariate time series from sensors recording parameters like rotor speeds and exhaust temperatures. However, researchers face two significant barriers: the proprietary nature of CEOD, which limits data sharing, and the prevalence of incomplete datasets due to sensor failures or transmission issues.

To address these challenges, the authors introduce AESim, a data-driven simulator capable of generating high-fidelity synthetic engine data. By conditioning generation on specific flight mission profiles, AESim creates "digital twins" that maintain the statistical and physical properties of real-world engines without compromising sensitive intellectual property.

The AESim Pipeline

The AESim framework employs a structured pipeline to transform a mission profile—consisting of variables like altitude, Mach number, and throttle lever angle—into realistic sensor outputs.

Overview of the AESim pipeline, illustrating the transition from mission profile input to phase-partitioned segment generation and final sensor output.

As shown in Figure 1, the process begins with Temporal Phase Partitioning. Engine physics vary significantly across flight stages; therefore, the system divides the flight into "Before Cruise," "During Cruise," and "After Cruise" phases. To handle long-duration sequences, the data is further divided into 300-second windowed segments with a 20-second overlap. This segmentation allows the model to manage the computational complexity of Transformer architectures while maintaining temporal context.

MTS-CGAN Architecture

At the heart of AESim is the Multivariate Time Series Conditional Generative Adversarial Net (MTS-CGAN). This architecture is specifically designed to handle the conditional nature of engine behavior and ensure continuity between generated segments.

Detailed architecture of the MTS-CGAN Generator (featuring Context and Adjustment Encoders) and the Patch-based Transformer Discriminator.

The Generator

The generator (Figure 2a) comprises two primary components:

  1. Context Encoder: This module processes the environmental "context" (the mission profile). Using self-attention mechanisms, it learns how the engine should respond to specific pilot inputs and atmospheric conditions.
  2. Adjustment Encoder: To solve the common "stitching" problem where transitions between generated segments appear disjointed, the Adjustment Encoder incorporates information from the previously generated segment. This ensures temporal smoothness across the entire flight duration.

The Discriminator

The discriminator (Figure 2b) is a patch-based Transformer. It evaluates both real and synthetic CEOD, learning to identify subtle discrepancies in dynamics. This adversarial training forces the generator to produce outputs that are indistinguishable from real sensor readings.

Training and Results

The model was trained on proprietary CEOD from Safran Aircraft Engines, focusing on generating critical sensors: Low-pressure rotor speed (N1), combustion chamber entry temperature (T), and Exhaust Gas Temperature (EGT). To stabilize training, the team utilized the Least Squares GAN (LSGAN) loss and a dedicated continuity loss to ensure overlapping segments align perfectly.

AESim demonstrates that engine behavior is fundamentally a response to its environment. By making the GAN strictly conditional, the simulator mimics physical reality: for a given mission profile, the generated N1 and EGT curves follow the expected physical trajectories. A major achievement of this work is the ability to perform "what-if" simulations, allowing researchers to compare how different engine models would react to the exact same mission profile.

Conclusion and Future Directions

AESim provides a robust solution for democratizing aviation research by providing high-quality synthetic datasets. While currently focused on specific engine types at 1 Hz resolution, future work aims to expand the model's reach to diverse engine architectures and higher-frequency transient behaviors. Currently being piloted at Safran’s DataLab, AESim represents a significant step toward fully realized digital twins in the aerospace sector.

Summary updated

BibTeX
@inproceedings{madane2024aesim,
  author = {Abdellah Madane and Florent Forest and Hanane Azzag and Mustapha Lebbah and Jérôme Lacaille},
  title = {{AESim: A Data-Driven Aircraft Engine Simulator}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure