TL;DR
This paper introduces a Bayesian framework using Gaussian Processes to model and infer temperature and pressure profiles within aeroengines, even with sparse sensor data. It leverages transfer learning to share information across different measurement planes and proposes a novel method for decomposing uncertainty into sampling and sensor error components, leading to more accurate performance assessments.
What Problem Does This Paper Address?
Aeroengines generate complex internal flow fields, and understanding their performance relies on accurate measurements of temperature and pressure at various axial stations. However, practical constraints limit the number and placement of sensors, leading to sparse data. Traditional methods for inferring these profiles and calculating performance metrics like area averages are often simplistic and can be biased or inaccurate due to this sparsity. Furthermore, distinguishing between uncertainty arising from insufficient sampling versus sensor measurement error is crucial for making informed decisions about instrumentation upgrades, but current methods lack a principled way to do this. This paper aims to bridge this gap by providing a statistically rigorous Bayesian approach to inferring these spatial fields and their associated uncertainties.
Methods
The core of the proposed method is a Bayesian framework using Gaussian Processes (GPs).
- Spatial Modeling: For each axial station, the temperature or pressure profile is modeled as a spatial Gaussian random field. This field is characterized by:
- A squared exponential kernel for radial variations.
- A Fourier basis for circumferential variations, with a sparsity-promoting prior (regularized horseshoe prior) to identify dominant harmonics when their exact values are unknown.
- Transfer Learning: To leverage data from multiple axial measurement planes, a novel planar covariance kernel is introduced. This kernel allows information to be transferred across planes, effectively learning from more instrumented planes to improve inference on sparsely instrumented ones. This is framed as a multi-task GP problem.
- Uncertainty Decomposition: The framework naturally decomposes the posterior uncertainty into two key components:
- Spatial sampling uncertainty: Arising from limited sensor coverage.
- Measurement imprecision: Arising from sensor noise and calibration errors.
- Bayesian Area Average: A closed-form solution for the area-weighted average is derived, which is shown to be more representative than traditional sector area averages.
Datasets
The paper uses experimental data from a research turbine rig. Specific details about the dataset include:
- Temperature measurements from an exit station in a high-pressure turbine.
- Data collected at different axial planes and with varying numbers and arrangements of rakes (sensor arrays).
- The paper also discusses simulated data for illustrating specific concepts, such as the ground truth spatial distribution of temperature.
Key Insights and Ideas
- Transfer Learning is Key for Sparse Data: The ability to transfer information across measurement planes significantly improves the accuracy of spatial field estimation, especially when dealing with very limited sensor data. This is crucial for real-world aeroengine applications where instrumentation is expensive and constrained.
- Sparsity Prior Uncovers Dominant Features: The sparsity-promoting prior on Fourier coefficients is effective in identifying the most important circumferential variations, even when the exact frequencies are not known beforehand. This leads to more parsimonious and robust models.
- Principled Uncertainty Decomposition: The ability to disentangle spatial sampling uncertainty from measurement imprecision is a major advancement. This provides actionable insights for engineers, helping them decide whether to add more sensors or improve the precision of existing ones, and where to place them.
- Bayesian Area Average Outperforms Traditional Methods: The proposed Bayesian area average is more robust and less sensitive to sensor placement than traditional sector area averages, offering a more reliable performance metric.
Key Results
- The proposed Bayesian framework, incorporating transfer learning and sparsity priors, successfully interpolates and estimates spatial temperature fields with reduced uncertainty compared to single-plane models.
- In case studies, the multi-plane transfer learning model demonstrated improved accuracy and reduced uncertainty in predicting temperature distributions, especially for sparsely instrumented planes.
- The decomposition of uncertainty showed that spatial sampling uncertainty often dominates measurement imprecision, and that adding more rakes can significantly reduce sampling uncertainty, particularly when strategically placed.
- The Bayesian area average was shown to be more stable and less sensitive to rake arrangement than sector area averages.
Limitations and Future Work
- Manual Similarity Specification: The current transfer learning approach relies on a user-defined similarity vector
sto indicate which planes are similar. The authors suggest that future work could leverage Dirichlet prior models to learn this similarity directly from the data as part of the inference process. - Generalization to Other Engines/Conditions: While the model was tested on a research turbine rig, further validation across a wider range of engine types, operating conditions, and measurement configurations would be beneficial.
- Extending to Pressure: The paper focuses on temperature, but the methodology is applicable to pressure as well. Further exploration of its performance with pressure data and combined temperature-pressure models could be valuable.
- Real-time Applications: Investigating the computational efficiency and scalability of the model for real-time engine health monitoring or control applications would be a significant next step.
Summary updated