CSTNet: A Dual-Branch Convolutional Network for Imaging of Reactive Flows using Chemical Species Tomography

Yunfan Jiang, Jingjing Si, Rui Zhang, Godwin Enemali, Bin Zhou, H. McCann, Chang LiuarXiv 2020 · DOI · arXiv

Abstract

Chemical Species Tomography (CST) has been widely used for in situ imaging of critical parameters, e.g. species concentration and temperature, in reactive flows. However, even with state-of-the-art computational algorithms the method is limited due to the inherently ill-posed and rank-deficient tomographic data inversion, and by high computational cost. These issues hinder its application for real-time flow diagnosis. To address them, we present here a novel CST-based convolutional neural Network (CSTNet) for high-fidelity, rapid, and simultaneous imaging of species concentration and temperature. CSTNet introduces a shared feature extractor that incorporates the CST measurement and sensor layout into the learning network. In addition, a dual-branch architecture is proposed for image reconstruction with crosstalk decoders that automatically learn the naturally correlated distributions of species concentration and temperature. The proposed CSTNet is validated both with simulated datasets, and with measured data from real flames in experiments using an industry-oriented sensor. Superior performance is found relative to previous approaches, in terms of robustness to measurement noise and millisecond-level computing time. This is the first time, to the best of our knowledge, that a deep learning-based algorithm for CST has been experimentally validated for simultaneous imaging of multiple critical parameters in reactive flows using a low-complexity optical sensor with severely limited number of laser beams.

Introduction

In combustion research and industrial monitoring, Chemical Species Tomography (CST) is a vital non-intrusive tool for measuring water vapor concentration and temperature. However, CST faces a significant mathematical challenge: the "inverse problem." Reconstructing high-resolution 2D images from a limited number of laser beams—often necessitated by restricted optical access—results in an ill-posed and rank-deficient system. Traditional iterative algorithms like Algebraic Reconstruction Technique (ART) or Tikhonov regularization are computationally intensive and prone to artifacts, making real-time monitoring of fast-moving flames difficult.

Methods: The CSTNet Architecture

To overcome these limitations, we developed CSTNet, a Convolutional Neural Network (CNN) specifically tailored for the physics of tomography. Unlike standard CNNs, CSTNet incorporates domain-specific knowledge through several key innovations:

  • Lambda Layer: This layer pre-processes raw projection data into heatmaps representing physical properties like centrosymmetry and smoothness, effectively embedding the sensor layout into the network's input space.
  • Shared Feature Extractor: A common set of convolutional layers learns fundamental patterns related to the sensor geometry and measurement noise.
  • Dual-Branch Architecture with Crosstalk Decoders: Recognizing that gas concentration and temperature are physically coupled in flames, the network employs two parallel branches. "Crosstalk" connections allow these branches to exchange information, leveraging the natural correlation between parameters to improve reconstruction fidelity.
The CSTNet architecture, illustrating the flow from CST projection data through the Lambda layer and shared feature extractor to the dual-branch crosstalk decoders.

Datasets and Validation

CSTNet was validated using both simulated and experimental data:

  1. Simulated Dataset: A library of 19,227 examples featuring one, two, or three "hot spots" (inhomogeneities) modeled with 2D Gaussian profiles to simulate diverse flame morphologies.
  2. Experimental Data: Real-world measurements from a laboratory setup using an industry-oriented 32-beam CST sensor. Tests included single flames in various positions and dual-flame configurations.

Key Results

CSTNet demonstrates significant improvements over traditional methods in speed, accuracy, and robustness:

  • Real-Time Performance: The network achieves an average reconstruction speed of 3,134 frames per second (fps) on a GPU, enabling true real-time monitoring.
  • High Fidelity and Noise Robustness: Even at a practical signal-to-noise ratio (SNR) of 35 dB, the Image Error (IE) remained low (approximately 0.12 for concentration and 0.07 for temperature). The model maintains stable performance across varying noise levels.
  • Experimental Validation: In laboratory tests, CSTNet precisely localized multiple flames and accurately retrieved peak values without the "ghosting" artifacts typical of older algorithms.
Quantitative performance of CSTNet across different SNRs and qualitative examples of the complex H2O concentration and temperature distributions used for training and testing.

Conclusion and Future Work

CSTNet represents a significant step forward in real-time flame diagnosis. By encoding physical constraints and correlations directly into the deep learning architecture, it achieves high-fidelity imaging from sparse sensor data. Future work will focus on deploying CSTNet to FPGAs for low-power embedded applications and testing its robustness in harsher, more turbulent industrial environments where physical correlations may be more complex.

Summary updated

Abstract

Chemical Species Tomography (CST) has been widely used for in situ imaging of critical parameters, e.g. species concentration and temperature, in reactive flows. However, even with state-of-the-art computational algorithms the method is limited due to the inherently ill-posed and rank-deficient tomographic data inversion, and by high computational cost. These issues hinder its application for real-time flow diagnosis. To address them, we present here a novel CST-based convolutional neural Network (CSTNet) for high-fidelity, rapid, and simultaneous imaging of species concentration and temperature. CSTNet introduces a shared feature extractor that incorporates the CST measurement and sensor layout into the learning network. In addition, a dual-branch architecture is proposed for image reconstruction with crosstalk decoders that automatically learn the naturally correlated distributions of species concentration and temperature. The proposed CSTNet is validated both with simulated datasets, and with measured data from real flames in experiments using an industry-oriented sensor. Superior performance is found relative to previous approaches, in terms of robustness to measurement noise and millisecond-level computing time. This is the first time, to the best of our knowledge, that a deep learning-based algorithm for CST has been experimentally validated for simultaneous imaging of multiple critical parameters in reactive flows using a low-complexity optical sensor with severely limited number of laser beams.
BibTeX
@misc{jiang2020cstnet,
  author = {Yunfan Jiang and Jingjing Si and Rui Zhang and Godwin Enemali and Bin Zhou and H. McCann and Chang Liu},
  title = {{CSTNet: A Dual-Branch Convolutional Network for Imaging of Reactive Flows using Chemical Species Tomography}},
  year = {2020},
  eprint = {2010.03868},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2010.03868},
}
Figure