[Paper Review] Measurement of Hybrid Rocket Solid Fuel Regression Rate for a Slab Burner using Deep Learning

Gabriel Surina|arXiv (Cornell University)|Aug 25, 2021

Rocket and propulsion systems research57 references16 citations

TL;DR

This study introduces a U-net convolutional neural network to automate fuel regression rate measurement in hybrid rocket slab burners using high-noise experimental images. By replacing manual segmentation with deep learning, the method achieves <10% error compared to literature values and quantifies uncertainty via Monte-Carlo Dropout, demonstrating robust performance across varying oxidizer fluxes and superior noise filtering over traditional image processing techniques.

ABSTRACT

This study presents an imaging-based deep learning tool to measure the fuel regression rate in a 2D slab burner experiment for hybrid rocket fuels. The slab burner experiment is designed to verify mechanistic models of reacting boundary layer combustion in hybrid rockets by the measurement of fuel regression rates. A DSLR camera with a high intensity flash is used to capture images throughout the burn and the images are then used to find the fuel boundary to calculate the regression rate. A U-net convolutional neural network architecture is explored to segment the fuel from the experimental images. A Monte-Carlo Dropout process is used to quantify the regression rate uncertainty produced from the network. The U-net computed regression rates are compared with values from other techniques from literature and show error less than 10%. An oxidizer flux dependency study is performed and shows the U-net predictions of regression rates are accurate and independent of the oxidizer flux, when the images in the training set are not over-saturated. Training with monochrome images is explored and is not successful at predicting the fuel regression rate from images with high noise. The network is superior at filtering out noise introduced by soot, pitting, and wax deposition on the chamber glass as well as the flame when compared to traditional image processing techniques, such as threshold binary conversion and spatial filtering. U-net consistently provides low error image segmentations to allow accurate computation of the regression rate of the fuel.

Motivation & Objective

• Automate the manual segmentation of fuel boundaries in 2D slab burner images to improve efficiency and repeatability in regression rate measurement.
• Address challenges in image segmentation caused by soot, pitting, wax deposition, and flame interference in high-noise experimental images.
• Quantify model uncertainty in regression rate predictions using Monte-Carlo Dropout to enhance reliability of deep learning-based measurements.
• Evaluate the performance of U-net under limited data and monochrome imaging conditions, relevant to cost-constrained experimental setups.
• Validate the deep learning approach against manually traced ground truth and published literature values to ensure accuracy and consistency.

Proposed method

• Employ a U-net convolutional neural network architecture for semantic segmentation of fuel regions in RGB and grayscale images from a 2D slab burner experiment.
• Train the U-net model on 150 annotated image-mask pairs collected from four oxidizer flux conditions (5.91–22.19 kg/m²·s) using stochastic gradient descent with early stopping.
• Apply Monte-Carlo Dropout during inference to estimate predictive uncertainty in regression rate predictions, providing confidence intervals.
• Compare U-net performance against traditional image processing techniques such as threshold-based binary segmentation and spatial filtering.
• Train separate models on individual oxidizer fluxes and a combined dataset to assess generalization and flux independence.
• Convert RGB images to grayscale to evaluate feasibility of using low-cost monochrome cameras in regression rate measurement.

Experimental results

Research questions

• RQ1 Can a U-net deep learning model accurately segment fuel boundaries in high-noise hybrid rocket combustion images, outperforming traditional image processing?
• RQ2 How does the U-net model’s regression rate prediction accuracy compare to manually traced ground truth and published literature values?
• RQ3 To what extent does the model’s performance depend on oxidizer flux, and can it generalize across different flux conditions?
• RQ4 How effective is the U-net in filtering out noise from soot, pitting, wax, and flame compared to conventional segmentation methods?
• RQ5 Can the model achieve reliable performance when trained on monochrome images, and what are the limitations of such low-depth data?

Key findings

• The U-net model achieved regression rate predictions with less than 10% error when compared to manually traced ground truth and literature values.
• The U-net demonstrated robustness to image noise, effectively filtering soot, pitting, wax deposition, and flame interference better than thresholding and spatial filtering methods.
• Model uncertainty was successfully quantified using Monte-Carlo Dropout, providing reliable confidence intervals for regression rate predictions.
• The U-net model trained on data from all oxidizer fluxes showed the lowest error (<10−2) across all flux conditions and was independent of oxidizer flux, indicating generalization capability.
• Training on monochrome images led to higher errors, especially for the 22.19 kg/m²·s flux case, due to over-saturation and insufficient feature detail, though the combined model still outperformed individual flux models.
• The network trained exclusively on the 22.19 kg/m²·s flux data performed poorly on its own data due to image over-saturation, highlighting the importance of image quality in training data.