Digitization

Complex test benches are required for the experimental investigation of modern engine and gas turbine compressors. Realistic operating conditions of a transonic axial compressor must be enabled, monitored and regulated. A wide variety of measurement variables must be recorded in numerous measurement levels. Special time-resolving measurement techniques are necessary for the simultaneous investigation of aerodynamic and aeroelastic phenomena in modern compressor designs. Transonic compressor rotors reach blade tip speeds of over 400 m/s and therefore a supersonic relative flow. In addition, the rotors are subjected to higher aerodynamic loads, which makes behavior at the stability limit extremely relevant. There, unsteady flow phenomena interact with the structure, which induces high-frequency vibrations.

In order to be able to take up research-relevant topics in compressor research, a broad spectrum of measurement data is recorded with specific, high-precision measurement systems.

Both for online monitoring and for postprocessing, over 300 measuring points for stationary and transient instrumentation are recorded synchronously with more than 45,000,000 measured values per second and stored in a structured manner in databases. In order to be able to resolve the interaction mechanisms of the compressor blades in time, sampling rates of up to 500,000 Hz are necessary. Automated online monitoring ensures data quality during the measurement campaign. For targeted evaluation, in-house algorithms and multidisciplinary evaluation methods are implemented using different programming languages.

Various optical measurement methods are used on the turbine test bench 'Large Scale Turbine Rig' for the experimental investigation of turbine stages. As a first example, the Particle Image Velocimetry method, or PIV for short, for the contactless determination of speed fields should be mentioned. For this purpose, particles are added to the flow, which are then excited by means of a pulsating laser sheet. During a pulse, two consecutive images are taken with the help of a digital camera, and with this the light reflected by the particles. By correlating the two images, 2D flow field information can be inferred. The evaluation algorithms contain a large number of digital signal processing steps that are very computationally intensive. Approaches from various areas of image processing are used, which quickly push a conventional workstation to its limits when evaluating entire flow configurations in a turbomachine. Large memory requirements and high processor performance must be given.

Similar to the investigation using the PIV method, two-dimensional thermograms are obtained during heat transfer studies at the LSTR, which are returned to the thermal quantities sought in specially developed evaluation routines. The thermograms first undergo a local calibration in order to assign the examined surface in the turbine to the individual pixels of the camera. Here, projection methods are used to determine the transformation matrices. The method also includes a linear fit method for every single pixel of the camera, in which a whole series of thermogram sequences are processed. This processing is computationally intensive and can be significantly accelerated by appropriately efficient programming (stack and matrix processing). The use of parallel programming approaches could lead to a significant reduction in evaluation times in the future.

The approaches which are used in the processing of the optical measurement methods mentioned are related to those which are used, for example, in the processing of satellite images or in the field of autonomous driving. All of the measurement systems mentioned have immense amounts of data that are archived on central hard disk systems.

The local and temporal development of flow phenomena, which can either not be recorded at all or only locally, can be determined very well using a numerical model. It is only through the additional application of a numerical model that a coherent interpretation of the stall-influencing flow events and the loss-causing mechanisms is possible.

As part of the interdisciplinary research at the Institute of Gasturbines and Aerospace Propulsion, numerical simulations of the measured compressor and turbine configurations are carried out. In order to be able to understand the measurement results as precisely as possible, numerical models are required that have a high level of detail and take all geometry features into account. For example, a digital twin of the Large Scale Turbine Rig (LSTR) has been built in recent years, which now enables the turbine flow, including the combustion chamber swirl and cooling air injection, to be numerically predicted.

In order to further reduce the NOx emissions of modern aircraft engines, new types of combustion chamber concepts are used. These are characterized by a strongly swirled flow, temperature inhomogeneities and high turbulence at the combustion chamber outlet. All of these phenomena have a negative impact on the efficiency and service life of the downstream high-pressure turbine. The interactions between the two components of the combustion chamber and turbine and, above all, the effects of variability in the flow fields at the combustion chamber-turbine interface on the flow in the first stage of the high-pressure turbine have not yet been fully understood. Due to the very high fluid temperatures at the combustion chamber outlet, detailed measurements can only be carried out to a very limited extent under real conditions, since these are significantly above the melting temperature of most metals. Therefore, numerical simulations play a very important role in investigations at the combustion chamber-turbine interface.

In this context, a program was developed at the Institute of Gasturbines and Aerospace Propulsion in cooperation with Rolls-Royce Germany, with which closed entry conditions of high-pressure turbines for RANS CFD can be created. The two-dimensional distributions of speed, pressure, temperature and turbulent quantities are generated using parameters. In this way it is possible to make changes to the boundary condition flexibly and sensitivity studies of the turbine with respect to variations in the entry boundary condition can be carried out. These studies should be based on entry traverses that are as realistic as possible. It is therefore of great relevance to adjust the parameterized fields to reference traverses. This is done with the help of artificial intelligence. An optimization algorithm tests thousands of parameter sets and finds the one that best fits the reference traverse.

This combination of fundamental methods of aerodynamics and state-of-the-art technology helps to increase the efficiency of the next generation of aircraft engines and thus contributes to the development of environmentally friendly aircraft engines of the future.

At the institute Gas Turbines and Aerospace Propulsion aerodynamic optimizations of compressors are carried out while at the same time considering the structural mechanics. An attempt is made to change the geometry of the compressor in such a way that aerodynamic variables such as efficiency or map width become maximum. At the same time, stresses and natural frequencies of the compressor wheel are checked in order to ensure the structural integrity of the optimized geometry.

Modern, evolutionary algorithms are used for the optimizations. Such algorithms have the advantage of being more likely to find the global maximum within the parameter space. In comparison, gradient-based methods tend to find only a local maximum of the parameter space, depending on the starting solution. The major disadvantage of the evolutionary algorithms is the large number of function calls required until the optimization has converged. In the case of compressor optimization, a function call means the evaluation of various aerodynamic variables, such as efficiency. Such evaluations are only possible using CFD simulations, which, depending on the geometry and objectives under consideration, can take several hours. Since evolutionary algorithms require several thousand, if not several tens of thousands of function calls, and even today the available computer capacities are limited, the use of evolutionary algorithms without additional aids for compressor optimization is difficult to use.

The tools are modern algorithms from the field of machine learning. Such algorithms are used to generate replacement models (Response Surface Models, RSM). Based on a large database, which contains several hundred compressor geometries with their corresponding aerodynamic sizes, the replacement models are trained. It is then possible to predict the corresponding aerodynamic quantities for new, unknown compressor geometries within seconds using the replacement model without having to carry out CFD simulations. Modern algorithms of machine learning thus enable optimizations to be carried out which would have far too long computing times without the use of such replacement models.

In addition, the trained replacement models make it possible, for example, to make statements about the influence of the underlying geometry parameters on aerodynamic variables (analysis of variance). Or new models can be generated which, based on the flow field, allow statements to be made about global variables in operating behavior. The possible uses are wide and it takes the imagination of the engineer to make full use of them.