This presentation was held at the 2020 NAFEMS UK Conference "Inspiring Innovation through Engineering Simulation". The conference covered topics ranging from traditional FEA and CFD, to new and emerging areas including artificial intelligence, machine learning and EDA.

Resource Abstract

In the context of global warming, highly competitive markets, and usage of multiple sources of energy, it becomes more and more important to optimize the operation of gas turbine power plants. This means that starts and stops of these engines become more and more frequent. On the other hand, engine manufacturers provide the operators conservative guidelines on how long to let the engine cool down before starting a new cycle, to ensure temperatures do not exceed certain thresholds.

While stationary components of the engine can be monitored easily, it is very difficult and expensive, if not impossible, to monitor the temperature of some rotating parts which determine the engine conditions (cold / warm / hot). These temperatures are critical to predict clearances, and have a potentially significant impact on the life span of the engine. The goal of this paper is to present an approach for real-time monitoring of these internal temperatures using AI, based on accurate and validated simulation data.

More specifically:

a) Siemens Simcenter 3D Thermal is used to build a complete representation of the whole engine through the use of thousands of separate boundary conditions that adequately represent the transient behavior of the whole engine. These boundary conditions represent the flow conditions, in various regimes, for a range of rotational speeds and in different sections of the engine. They make use of built-in thermal correlations, of "expressions" that can automatically access specific model data (e.g., fluid temperature, material properties, solve-time computed results, rotational speed or radius), or of proprietary correlations obtained from in-house measurements, experience or CFD.

b) This "digital twin" is validated against experimental data where scaling factors on the boundary conditions are used to obtain very precise thermal predictions.

c) "Virtual sensors" are created at specific locations that need to be monitored in real-time during operations and that cannot be instrumented easily.

d) Machine learning (ML) algorithms and reduced order techniques are used to correlate the temperature time-response of these virtual sensors against real physical sensors where temperature can be easily monitored. Since the response is encapsulated in a much reduced set, it runs very fast compared to the digital twin.

e) The time response of the virtual sensors could then be implemented "on the edge" in a real-time data acquisition system that instruments the engine.

The advantage of the method is that it is general enough to quickly create engine-specific virtual sensors while taking into account all of the specific physical characteristics of the engine (fluid temperatures, mass flow rates, engine operating conditions, thermal inertia, etc.). This method will be illustrated on a real engine for demonstration purposes.