Streamlining multi-hole probe calibration using artificial neural networks

Abstract

Accurately measuring the three-dimensional flow field characteristics in complex flow fields, particularly in turbomachines, is of utmost importance and is commonly achieved through experimental methods. Multi-hole pressure probes are a proven measuring technique to provide precise readings of flow characteristics in both two-dimensional (2D) and three-dimensional (3D) flow fields. However, manufacturing imperfections have an impact on their measuring accuracy so each probe has a slightly different measuring behavior. In order to take these specific characteristics into account a thorough calibration process prior to utilization has to be applied. This calibration process involves positioning the probe in various flow conditions in a wind tunnel and collecting all relevant data. A model is then established to correlate the measured pressure distribution across the probe holes with known flow conditions. Unfortunately, this calibration process is time-consuming and requires costly equipment.To address these challenges, this study introduces a novel calibration technique that leverages advanced machine learning methods, specifically artificial neural networks (ANNs), to significantly streamline the process. ANNs are renowned for their ability to model diverse systems and adapt and generalize well, making them a promising solution for simplifying the calibration of multi-hole probes. In this research, an ANN calibration model was developed specifically for a 5-hole probe, and its configuration was optimized through an automated hyperparameter optimization process. The performance of the ANN model was then compared to that of a conventional polynomial model. After confirming its accuracy, the ANN model was successfully re-adapted and applied to new probes, requiring minimal calibration data collection. The results highlight the potential of ANNs in probe calibration, as they enable a simplified process that demands fewer calibration data points.