Simultaneous 5 kHz OH-PLIF/PIV for the study of turbulent combustion at engine conditions

Abstract

Simultaneous measurements of velocity and scalar fields were performed in turbulent nonpremixed flames at gas turbine engine-operating conditions using 5 kHz particle image velocimetry (PIV) and OH planar laser-induced fluorescence (OH-PLIF). The experimental systems and the challenges associated with acquiring useful data at high pressures and high thermal powers are discussed. In this work, a wide range of operating conditions were studied, with a maximum pressure and thermal power of 1.8 MPa and 950 kW, respectively. In the PIV measurements, the high thermal power conditions were shown to cause significant defocusing of the particle images. This was the result of variations in the optical refractive index of the gas which were caused by strong temperature gradients within the inner structure of the flame. High flame luminosity also led to decreased SNR with increasing flame power. The OH-PLIF measurements did not show indication of strong laser sheet absorption at any condition tested. However, a decrease in the peak SNR was observed with increasing chamber pressure. An analysis of the true measurement resolution with respect to the scales of the flow is also given. Based on the resolved scales, the present dataset was used to study the time-averaged flow structure and its effect on flame behavior. Heat release conditioned flow statistics were studied to elucidate the flow–flame interactions in high-power flames.

Appendices

Appendix 1: Spectral filtering

In this work, a two-dimensional (2D) spectral filtering technique was developed to detect and suppress the variation in signal intensity resulting from nonuniformity in the beam energy profile. Similar methods have been shown to successfully mask known spatial frequencies in two-dimensional data, such as in the Structured Laser Illumination Planar Imaging (SLIPI) where an intensity-modulated laser sheet is used to delineate signal photons from noise in Rayleigh scattering measurements [46]. In this application, calibration images taken in the homogeneous fluorescing medium are used as calibration data to detect the characteristic spatial frequencies of the PLIF laser-excitation sheet. The process is illustrated in Fig. 30 as the operations are demonstrated on a single-shot acetone PLIF calibration image. The raw image (a) is transformed into the frequency domain with a 2D fast Fourier transform (FFT). In the centered 2D Fourier spectrum (b), the coefficients with the low wave numbers are shifted to the middle of the image while the highest frequencies occur at the periphery. The 1D spectral density extracted along $n=0$ (c) indicates peaks in the power of multiple effective frequencies spatially oriented along the x direction. Suppression of these peaks, followed by computation of the inverse 2D FFT, yields a corrected image (d) with a significant reduction in variation of intensity along the x direction. The computed difference between the corrected data (d) and the raw image (a) is given in Fig. 30e, showing the level of intensity correction performed.

Fig. 30

PLIF sheet intensity correction procedure: raw calibration image (a), centered 2D Fourier spectrum (b), extracted spectrum for horizontal direction (c), corrected calibration image (d), difference between corrected and original (e)

Through development of the 2D spectral filter using the homogeneous-field acetone PLIF images, it was determined that the low-wave-number spatial frequencies measured across the width of the sheet were sufficiently resolved and consistent to be robustly masked on a single-shot basis. Hence, these dominant frequencies were identified using the calibration image set, then applied as a spectral mask on the PLIF data images. The results are shown in Fig. 20 for a sequence of images to show the effective removal of the strong intensity variations previously observed in Fig. 19. It is noted that some high wave number vertical structures have remained in the corrected images (also seen in Fig. 30d). It was found that robust suppression of these features was not possible without causing undesirable effects on the corrected images given the spatial resolution of the detection system used in this work.

Appendix 2: Flame surface topography extraction

Quantitative detection of the reaction zones using OH-PLIF measurements is a challenging task. While it has been shown that the gradient in the OH-PLIF signal can be utilized as a marker to delineate regions of burned and unburned gas near the reaction zone, these computations are highly sensitive to the image SNR. When utilizing high-repetition-rate, DPSS laser systems, the difficulties stemming from low SNR are exacerbated by the characteristically low laser pulse energies and, consequently, low fluorescence signal levels. High concentrations of OH are also known to remain far downstream in the postflame gases while three-body recombination reactions consume OH on a much slower time scale [18, 44, 47]. In highly turbulent, high-power swirl flames, multiple vortex breakdown recirculation zones are generated (by design) to transport heat and radical back to the flame root for stabilization. In these cases, it can become very challenging to distinguish the OH created at the reaction from the old OH that has been recirculated.

In this study, extensive algorithm development was required to accurately and robustly detect the reaction fronts based on the gradient in the OH-PLIF signal. Building on the foundation outlined in Boxx et al. [22], the sheet intensity-corrected OH-PLIF images were first binned with a 2 pixel by 2 pixel square kernel to improve the SNR (Fig. 31a). An edge-preserving nonlinear diffusion filter was then applied to remove uncorrelated detection system noise and low magnitude signal gradients in the burnt gases [48, 49]. A Gaussian-smoothed spatial gradient was computed from the nonlinear diffusion-filtered image, given in Fig. 31b. It can be seen that the strong gradients correspond well with the boundaries of the regions with high OH signal. However, it can also be seen that there are some regions detected where the signal gradient is quite low, which likely do not represent a flame edge; instead, these edges are likely a pocket of burnt gas or even just signal noise (common near the outer boundaries of the image where convolution filtering operations can cause problems). Hence, the gradient image is binarized with a user-defined threshold on the signal-gradient magnitude, resulting in the logical image given in Fig. 31c. At this stage, a morphological thinning operation was utilized to reduce the broad regions of high OH signal gradient to mathematically treatable contours (Fig. 31d). The utilization of morphological operations in lieu of more physics-based methods (e.g., signal-gradient profile curvature) was necessitated by the low PLIF SNR. The result of the procedure is a map of the flame surface topography, defined by the transition for reactants to high OH-PLIF signal. A final filtering operation removes the reaction fronts below a minimum length and multiply-connected bridge or chad artifacts from the morphological thinning operation. A parametric contour function, $f_i = x_{fi}\left( \xi _i\right) \hat{x} + y_{fi}\left( \xi _i\right) \hat{y}$, is then generated to describe the spatial coordinates of the flame front within the measurement plane. It should be noted that $f_i$ contains only spatial information about the location of the reaction zone; it is not a direct measure of the reaction rate. As it is seen in Fig. 31, the routine accurately identified the flame edges in the low SNR OH-PLIF images

An uncertainty analysis was performed to capture the effects of varying the user-defined thresholds on the defined flame surface topography. Reporting on condition B, which had a peak SNR of approximately 12.7, variation of the signal-gradient minimum threshold by $\pm$10 % resulted in a $\pm$3.485 % maximum response in the total flame contour length and a $\pm$3.92 % change in the peak flame surface density. Sweeping the gradient threshold $\pm$25 % resulted in a difference as high as 13 % in the summed total flame-front length was as much at 28 % change in the computed RMS. With large datasets, convergence of the mean standard error was $<$2 % for statistical quantities computed from $f_{i}$, including the mean flame surface density (sampled at multiple distinct points) as well as the summed total flame contour length.

Fig. 31

Flame surface topography extraction