Bayesian assessments of aeroengine performance with transfer learning

Pranay Seshadri, A. Duncan, George Thorne, Geoffrey T. Parks, Raul Vázquez Díaz, Mark Girolami

TL;DR

This paper introduces a Bayesian framework using Gaussian Processes (GPs) to reconstruct the complex temperature and pressure profiles inside jet engines from sparse sensor data. By leveraging transfer learning across measurement planes and sparsity-promoting priors, the authors provide a more accurate "Bayesian area average" and a principled method to determine whether an engine needs more sensors or simply better ones.

What Problem Does This Paper Address?

To assess the health and efficiency of a jet engine, engineers need to know the average temperature and pressure at various stages (axial stations) of the engine. However, the interior of a turbine is a hostile, cramped environment. Sensors are expensive and few in number, often arranged in "rakes" that provide only a handful of data points across a 2D cross-section (an annulus).

The industry standard is "sector area-averaging"—essentially a weighted sum of sensor readings. This method is notoriously sensitive to "rake clocking" (the circumferential position of the sensors). If a rake happens to sit in a "hot spot" or a "trough" of the flow, the resulting 1D average will be biased. Furthermore, traditional methods struggle to quantify why an estimate is uncertain: is it because we don't have enough sensors (sampling error), or because the sensors themselves are imprecise (measurement noise)?

Methods

The authors model the spatial temperature/pressure profile as a Gaussian Random Field on an annulus. Their approach combines several sophisticated techniques:

  • Hybrid Kernels: They use a Fourier basis kernel to capture periodic circumferential variations (engine modes) and a squared exponential kernel for radial variations.
  • Transfer Learning: They introduce a novel planar covariance kernel. This allows the model to "transfer" information between different axial planes or even different engines in a fleet, assuming they share similar underlying physical characteristics.
  • Sparsity-Promoting Priors: Since the number of dominant "modes" in an engine is usually small but unknown, they use a regularized horseshoe prior. This encourages the model to find a sparse representation, preventing over-fitting to noise.
  • Bayesian Area Average (BAA): Unlike the discrete sector average, the BAA is derived by integrating the posterior GP over the entire annulus in closed form.
  • Uncertainty Decomposition: Using the law of total covariance, they split the total uncertainty into two distinct metrics: spatial sampling uncertainty and impact of measurement imprecision.

Datasets

The methodology was validated using three distinct types of data:

  1. Synthetic Data: A ground-truth temperature field with five circumferential harmonics used to test convergence and the uncertainty decomposition framework.
  2. Research Turbine Rig Data: High-resolution traverse temperature measurements from a real turbine rig, used to demonstrate the model's ability to identify dominant wave numbers (modes 1, 12, and 24).
  3. Multi-Engine Fleet Data: Temperature measurements from eight different research engines. This demonstrated how the planar kernel can identify similarities between engine "families" to improve predictions across a fleet.

Key Insights and Ideas

  • The "Aha" Moment on Placement: The study shows that adding more sensors doesn't always help if they are placed poorly. The Bayesian framework can actually guide experimental design by identifying regions where "spatial sampling uncertainty" is highest.
  • Better vs. More: The decomposition of uncertainty is a game-changer for instrumentation engineers. It provides a rigorous answer to the question: "Should we spend money on more rakes, or on higher-precision probes?"
  • Information Transfer: The model can take a heavily instrumented "test-bed" engine and use its learned spatial correlations to improve the performance assessment of a sparsely instrumented "in-flight" engine.

Key Results

  • Superior Convergence: The Bayesian area average (BAA) converged much faster than the sector average. In one test case, with 12 rakes, the BAA reduced uncertainty to $\pm 0.04$ K, compared to the sector average's $\pm 1.63$ K.
  • Robustness to Clocking: The BAA proved significantly less sensitive to the physical rotation (clocking) of the sensor rakes, providing a more stable and representative metric for engine efficiency.
  • Fleet Similarity: In the 8-engine study, the model successfully clustered engines from the same family based on their posterior correlation, even when the similarity wasn't explicitly encoded in the priors.

Limitations and Future Work

  • Steady-State Assumption: The current model is designed for steady-state performance. The authors note that extending this to "unsteady" or transient problems would require the development of a temporal kernel.
  • Similarity Inputs: The transfer learning currently relies on a user-defined similarity vector ($s$). The authors suggest that future work could use Dirichlet process priors to "learn" these similarities directly from the data.
  • Mass Averaging: While this paper focuses on area averages, the authors acknowledge that "mass-averaging" (which accounts for velocity/density) is also critical in turbomachinery and is a logical next step for the Bayesian framework.

Summary updated

Abstract

Abstract Aeroengine performance is determined by temperature and pressure profiles along various axial stations within an engine. Given limited sensor measurements, we require a statistically principled approach for inferring these profiles. In this paper we detail a Bayesian methodology for interpolating the spatial temperature or pressure profile at axial stations within an aeroengine. The profile at any given axial station is represented as a spatial Gaussian random field on an annulus, with circumferential variations modelled using a Fourier basis and radial variations modelled with a squared exponential kernel. This Gaussian random field is extended to ingest data from multiple axial measurement planes, with the aim of transferring information across the planes. To facilitate this type of transfer learning , a novel planar covariance kernel is proposed. In the scenario where frequencies comprising the temperature field are unknown, we utilise a sparsity-promoting prior on the frequencies to encourage sparse representations. This easily extends to cases with multiple engine planes whilst accommodating frequency variations between the planes. The main quantity of interest, the spatial area average is readily obtained in closed form. We term this the Bayesian area average and demonstrate how this metric offers far more representative averages than a sector area average---a widely used area averaging approach. Furthermore, the Bayesian area average naturally decomposes the posterior uncertainty into terms characterising insufficient sampling and sensor measurement error respectively. This too provides a significant improvement over prior standard deviation based uncertainty breakdowns.
BibTeX
@inproceedings{seshadri2022bayesian,
  author = {Pranay Seshadri and A. Duncan and George Thorne and Geoffrey T. Parks and Raul Vázquez Díaz and Mark Girolami},
  title = {{Bayesian assessments of aeroengine performance with transfer learning}},
  booktitle = {Data-Centric},
  year = {2022},
  doi = {10.1017/dce.2022.29},
}
Figure