Explainable Artificial Intelligence for Exhaust Gas Temperature of Turbofan Engines

Marios Kefalas, Juan de Santiago Rojo Jr., Asteris Apostolidis, Dirk van den Herik, Bas van Stein, Thomas Bäck

TL;DR

This paper applies Symbolic Regression (SR) to real-world turbofan engine data to uncover interpretable algebraic relationships between Exhaust Gas Temperature (EGT) and other engine parameters. The method successfully generates accurate, human-understandable models, offering a valuable alternative to "black box" AI for critical engineering applications.

What Problem Does This Paper Address?

In industries like aviation, understanding the underlying processes driving system behavior is crucial, especially for safety-critical components like turbofan engines. Traditional physics-based models can be time-consuming and resource-intensive to develop. While data-driven models, particularly deep neural networks (DNNs), can achieve high accuracy, they often act as "black boxes," lacking interpretability. This lack of transparency makes it difficult to trust their predictions, diagnose issues, or gain new engineering insights. The paper addresses this by seeking an interpretable data-driven approach to model a key engine health indicator, Exhaust Gas Temperature (EGT), using real-world flight data.

Methods

The core methodology employed is Symbolic Regression (SR), specifically using Genetic Programming (GP). Unlike conventional regression, SR doesn't assume a predefined model structure. Instead, it searches for the optimal algebraic expression that best fits the data. The authors utilize a variant called Multi-Gene Genetic Programming (MGGP), which creates models as linear combinations of simpler symbolic expressions (genes). This approach aims to balance model performance with complexity, making the resulting formulas more interpretable. The process involves:

  • Data Pre-processing: Cleaning and selecting relevant parameters from the continuous engine operational data (CEOD).
  • Dimensionality Reduction: Using correlation analysis to remove highly correlated input variables.
  • SR with MGGP: Training the GP algorithm to evolve symbolic expressions that predict EGT.
  • Model Evaluation: Assessing model performance using standard regression metrics like RMSE, MAE, and R².

Datasets

The study uses real-world, continuous engine operational data (CEOD) collected from four flights of a Boeing 787-10 aircraft equipped with a GEnx turbofan engine during July 2019. This data consists of several hundred parameters recorded at a 1Hz frequency throughout the entire flight duration. For experiments, three flights were concatenated for training and testing, and a fourth flight was used for validation.

Key Insights and Ideas

  • Interpretability is Achievable with Complex Data: The paper demonstrates that even with high-dimensional, real-world operational data, SR can uncover meaningful and interpretable algebraic relationships, challenging the notion that complex systems necessitate black-box models.
  • EGT is Predictable from Operational Parameters: The results show that EGT can be accurately predicted using a combination of other engine operational parameters, highlighting the interconnectedness of engine health indicators.
  • Symbolic Regression Offers Engineering Value: The generated algebraic expressions are not just accurate but also physically plausible and validated by field experts, suggesting SR can provide actionable engineering insights beyond simple prediction.
  • CEOD Data is Crucial for Dynamics: The use of continuous, high-frequency CEOD data, as opposed to traditional snapshot data, allows for a more complete representation of engine behavior and the capture of dynamic relationships.

Key Results

The SR models achieved excellent performance:

  • High Accuracy: R² values close to 1 were observed for training and test sets, indicating near-perfect fits.
  • Low Error: The average deviation from the ground truth EGT was less than 1°C for training and test sets (MAE ≈ 0.77°C).
  • Acceptable Validation Error: The validation set showed a slightly higher average error of around 3°C, which is considered negligible from an engineering perspective, especially given that factors like flight level and aircraft weight were not explicitly included.
  • Meaningful Variables Identified: The analysis identified key engine parameters (e.g., High Pressure Turbine clearance, Gas Temperature at Station 25, corrected Fan Speed) that significantly influence EGT, aligning with engineering knowledge.

Limitations and Future Work

The authors acknowledge several limitations:

  • Cruising Phase Only: The study focused solely on the cruising phase of the flight, potentially missing dynamics from other phases (take-off, climb, descent) where engine performance is more transient.
  • Data Pre-processing Impact: The removal of parameters with NaN values, while necessary, led to some data loss.
  • Single Output Variable: Only EGT was modeled; other parameters were not investigated.

Future work suggested includes:

  • Modeling Transient Phases: Extending the analysis to include take-off and other dynamic flight phases.
  • Improved Data Pre-processing: Investigating techniques like NaN imputation to retain more data.
  • Incorporating Additional Information: Augmenting data with factors like weather conditions, flight level, and aircraft weight.
  • Hyperparameter Optimization: Performing more extensive optimization of GP hyperparameters.
  • Ensemble Modeling: Creating meta-models or ensemble models from multiple derived formulas.
  • Predictive Maintenance: Leveraging interpretable models for fault detection and predictive maintenance.
  • Forecasting: Developing models for parameter forecasting to enable prognostics.

Summary updated

Abstract

Data-driven modeling is an imperative tool in various industrial applications, including many applications in the sectors of aeronautics and commercial aviation. These models are in charge of providing key insights, such as which parameters are important on a specific measured outcome or which parameter values we should expect to observe given a set of input parameters. At the same time, however, these models rely heavily on assumptions (e.g., stationarity) or are "black box" (e.g., deep neural networks), meaning that they lack interpretability of their internal working and can be viewed only in terms of their inputs and outputs. An interpretable alternative to the "black box" models and with considerably less assumptions is symbolic regression (SR). SR searches for the optimal model structure while simultaneously optimizing the model's parameters without relying on an a-priori model structure. In this work, we apply SR on real-life exhaust gas temperature (EGT) data, collected at high frequencies through the entire flight, in order to uncover meaningful algebraic relationships between the EGT and other measurable engine parameters. The experimental results exhibit promising model accuracy, as well as explainability returning an absolute difference of 3{\deg}C compared to the ground truth and demonstrating consistency from an engineering perspective.
BibTeX
@article{kefalas2022explainable,
  author = {Marios Kefalas and Juan de Santiago Rojo Jr. and Asteris Apostolidis and Dirk van den Herik and Bas van Stein and Thomas Bäck},
  title = {{Explainable Artificial Intelligence for Exhaust Gas Temperature of Turbofan Engines}},
  journal = {J. Aerosp. Inf. Syst.},
  year = {2022},
  url = {https://openreview.net/forum?id=z42cqnuvYa},
}
Figure