Detecting Nonsynchronous Vibration in Transonic Fans Using Machine Learning Techniques

Due to manufacturing tolerance and deterioration during operation, different blades in a fan assembly exhibit geometric variability. This leads to asymmetry which will be amplified in the running geometry by centrifugal and aerodynamic loads. This study investigates a phenomenon known as Alternative Passage Divergence (APD), where the blade untwist creates an alternating pattern in passage geometry and stagger angle around the circumference. After the formation of alternating tip stagger pattern, APD’s unsteady effect, Non-Synchronous Vibration (NSV), can cause the blades from one group to switch to the other creating a travelling wave pattern around the circumference. Thus, it can potentially lead to high cycle fatigue issues. More importantly, this phenomenon occurs close to, or at, peak efficiency conditions and can significantly reduce overall efficiency. Therefore, it is vital to attenuate the NSV behaviour.

In a previous study (Lu et al, 2019), an idealised mis-stagger setup was used to gain an initial understanding of the NSV phenomenon. However, random mis-staggering patterns due to manufacturing variability complicate the evolution of NSV significantly, making it difficult to draw general conclusions from parametric studies. Thus, machine learning techniques are used to analyse mis-stagger patterns to identify patterns that can lead to NSV and thus help avoid it. Numerical results from 1.6 million CPU hours of computation are used to train and test the classifiers. From the results, two parameters contributing to NSV behaviour have been identified with one of them enhancing the understanding found in Lu et al, 2019.

The NSV phenomenon is examined numerically by a partially coupled aeroelastic solver. The solver, AU3D, is developed by Imperial College London and has been proven to be proficient in conducting aerodynamic and aeroelastic prediction at off-design conditions for numerous fans and compressors.

The machine learning component of this study is conducted through the open-source library scikit-learn. Two machine learning algorithms are used: logistic regression classifier and the Classification And Regression Tree (CART). The CART algorithm is specifically chosen for its interpretability.

[1] M. J. Wilson, M. Imregun, and A. I. Sayma, 2007. “The effect of stagger variability in gas turbine fan assemblies,” J. Turbomach., 129(2), 404-411.

[2] Choi, M., Smith, N.H.S and Vahdati. M, 2013. “Validation of numerical simulation for rotating stall in a transonic fan”, J. Turbomach., 135 (2), 021004-021004-8.

[3] Stapelfeldt, S.C. and Vahdati, M. 2018, “On the Importance of Engine-Representative Models for Fan Flutter Predictions,” J. Turbomach, 140(8), p.081005.

[4] Lu, Y., Lad, B., Vahdati, M., and Stapelfeldt, S. C., 2019. “Nonsynchronous Vibration Associated with Transonic Fan Blade Untwist". In ASME Turbo Expo 2019: Turbomachinery Conference and Exposition, Phoenix AZ, USA, American Society of Mechanical Engineers.

[5] Pedregosa et al.,2011. “Scikit-learn: Machine Learning in Python”. JMLR 12, pp. 2825-2830.