Aeroengine

Auditing of Spatial Model for Pressure and Temperature distribution within an Aeroengine, based on test-rig data.

 

Context:  A model had been developed to study unexpected spatial distribution of temperature across temperature sensors at HP compressor exit, positioned to miss OGV-to-OGV variation.   A spatial distribution model had been developed to attempt to capture the unexplained cyclic temperature variation.

Our Assessment:   Following a static-code analysis of the existing model, we proceeded to analyse the statistical model.   Based on the historical data provided, and the predictions of the model based on this data, we identified subspaces of the input data in which the model would consistently underperform.  These blind-spots in the model were mostly arising from the sparsity of the sensor distribution along with the Nyquist-Shannon constraint (a condition for a sample rate that permits a discrete sequence of samples to capture all the information from a continuous-time signal of finite bandwidth).   Based on this we constructed a collection of adversarial test cases which would demonstrate high-impact failures of the model.   We made several recommendations in a report, including:

  1. A recommendation to attempt Bayesian modelling approaches to be able to deal with the sparsity of the data and overcome the NS constraint.
  2. We highlighted that any model would be limited by the lack of sensor data, and therefore one should not expect a high degree of certainty in any prediction. Therefore, any prediction should be provided with associated degree of uncertainty.
  3. We recommended undergoing a sensor position optimisation exercise to maximise the information gained through the sensors, as this was currently not being selected optimally.
  4. We identified issues in how the sub-component efficiency based on this model was assessed, noting that this was currently at risk of being significantly different from what was predicted, and that this would have a negative influence on subsequent decision making.

Follow-On Based on Recommendations:  We developed the core for a new digital twin for the gas path of an engine, based on the recommendations we had made, offering full and detailed uncertainty quantification of any predictions made.  We provided a framework in which optimal design can be performed, resulting in us recommending new sensor configurations for the suite.  This model was subsequently used to provide assessments of sub-component efficiency within the engine.

Timescales and Effort:  This project predates AQ, and thus the model analysis was performed entirely manually.  The assessment of error process to identify underperforming features etc involved 6 man-months.   With the new AQ tools, this could be reduced drastically, down to days.  The subsequent follow-on work is still ongoing (for the last 2 years).

 

https://aq.live/wp-content/uploads/2021/12/cropped-AQ-01cropped.png
Get in touch

    Copyright 2021. AQ Live. All rights reserved.