Proper orthogonal decomposition and smooth orthogonal decomposition approaches for pattern recognition: Application to a gas turbine rub-impact fault

2021 ◽  
pp. 116344
Author(s):  
L.V. Pereira ◽  
T.G. Ritto
1997 ◽  
Vol 119 (2) ◽  
pp. 289-296 ◽  
Author(s):  
G. A. Kopp ◽  
J. A. Ferre´ ◽  
Francesc Giralt

The eigenvectors obtained from proper orthogonal decomposition (POD) are used as the selection criteria to classify the individual events contained in date files of two-component velocity signals recorded in the non-homogeneous (vertical) and homogeneous (horizontal) planes of a fully-developed turbulent cylinder wake. This procedure uses the dominant eigenvectors from POD as initial templates to perform a pattern recognition (PR) analysis of the signals so that the individual coherent events appearing randomly in the signals can be educed more objectively. It is found that the prototype or ensemble average of the group of events classified as the large-scale structure in the vertical plane has a circulatory motion with strong negative streamwise and outward lateral velocity fluctuations. In the horizontal plane, the average structure is a double roller with negative streamwise velocity fluctuations in its centerplane. The class of instantaneous events selected contribute significantly to the variance in the outer intermittent region, but much less in the fully turbulent core. While the first eigenvector from POD captures some of the features shown by this prototype, it is not enough to depict by itself all the characteristics shown by the ensemble average of the repetitive three-dimensional instantaneous events that occur in the fully-developed turbulent wake.


Meccanica ◽  
2021 ◽  
Author(s):  
Matteo Dellacasagrande ◽  
Dario Barsi ◽  
Patrizia Bagnerini ◽  
Davide Lengani ◽  
Daniele Simoni

AbstractA different version of the classic proper orthogonal decomposition (POD) procedure introducing spatial and temporal weighting matrices is proposed. Furthermore, a newly defined non-Euclidean (NE) inner product that retain similarities with the POD is introduced in the paper. The aim is to emphasize fluctuation events localized in spatio-temporal regions with low kinetic energy magnitude, which are not highlighted by the classic POD. The different variants proposed in this work are applied to numerical and experimental data, highlighting analogies and differences with respect to the classic and other normalized variants of POD available in the literature. The numerical test case provides a noise-free environment of the strongly organized vortex shedding behind a cylinder. Conversely, experimental data describing transitional boundary layers are used to test the capability of the procedures in strongly not uniform flows. By-pass and separated flow transition processes developing with high free-stream disturbances have been considered. In both cases streaky structures are expected to interact with other vortical structures (i.e. free-stream vortices in the by-pass case and Kelvin–Helmholtz rolls in the separated type) that carry a significant different amount of energy. Modes obtained by the non-Euclidean POD (NE-POD) procedure (where weighted projections are considered) are shown to better extract low energy events sparse in time and space with respect to modes extracted by other variants. Moreover, NE-POD modes are further decomposed as a combination of Fourier transforms of the related temporal coefficients and the normalized data ensemble to isolate the frequency content of each mode.


Sign in / Sign up

Export Citation Format

Share Document