Tunable diode laser absorption spectroscopy (TDLAS) tomography is an important optical non-invasive combustion-detection technique that enables the imaging of critical flow-field parameters in the combustion field. The existing network-based TDLAS temperature tomographic algorithms are typically constructed with a fixed spatial resolution. If the target resolution is changed, then the network structure should be adjusted and the network retrained accordingly. Images reconstructed by these separate networks do not present a clear spatial correspondence, which renders it inconvenient to combine features in images with different spatial resolutions for combustion diagnosis. Hence, multiresolution spatial discretization modeling and multiresolution temperature-distribution reconstruction were introduced into TDLAS tomography. Based on reconstruction at two spatial resolutions as an example, a dual-domain multiscale feature-based multiresolution temperature tomographic network (DMFMMnet) was constructed.

The proposed DMFMMnet extracts and adaptively merges multiscale spatial features in the TDLAS measurement and image domains to reconstruct low-resolution and high-resolution temperature images with spatial correspondence. First, it extracts multiscale spatial correlation features from TDLAS measurements and reconstructs the overall profile of the temperature distribution rapidly at a low resolution using a low-resolution reconstruction block (LBlock). Second, it performs multiscale feature extraction and adaptive feature merging on the reconstructed low-resolution temperature image using an image-domain multiscale feature extraction and merging block (IMBlock). Third, it combines multiscale spatial features extracted in the TDLAS measurement and image domains to reconstruct a high-resolution temperature image, which presents detailed features of the temperature distribution via feature enhancement and a high-resolution reconstruction block (HBlock).

To examine the performance of the proposed DMFMMnet, it was compared with two existing network-based temperature tomographic algorithms. One is based on a convolutional neural network (H-CNN) and the other is based on a hierarchical vision Transformer and multiscale feature merging (HVTMFnet). These two networks were adjusted and trained for low-resolution reconstruction (with suffix -LR) and high-resolution reconstruction (with suffix -HR), separately. Furthermore, to examine the super-resolution reconstruction performance of IMBlock and HBlock in DMFMMnet, they were compared with the classical bicubic linear interpolation (bicubic) and two super-resolution networks, i.e., super-resolution using deep convolutional networks (SRCNNs) and Transformer for single-image super-resolution (ESRT). The resulting high-resolution reconstruction networks combined with LBlock are referred to as LBlock+Bicubic, LBlock+SRCNN, and LBlock+ESRT. In the simulations, the dataset was generated using a fire dynamics simulator (FDS). Tests were performed in the signal-to-noise ratio (SNR) range of 25 dB‒45 dB. The peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) were used to measure the reconstruction quality. The simulation results show that, for low-resolution reconstruction, the average PSNR values obtained by DMFMMnet are higher than those obtained by H-CNN-LR and HVTMFnet-LR at any SNR (Table 2). For high-resolution reconstruction, the average PSNR values obtained by DMFMMnet are higher than those obtained by H-CNN-HR, HVTMFnet-HR, LBlock+Bicubic, LBlock+SRCNN, and LBlock+ESRT in the SNR range of 30 dB‒45 dB (Table 3). In terms of subjective quality, the low-resolution temperature image reconstructed by DMFMMnet reflects the overall contour of the temperature distribution more clearly than those reconstructed by H-CNN-LR and HVTMFnet-LR (Fig. 7), whereas the high-resolution temperature image reconstructed by DMFMMnet shows more accurate and detailed information than those reconstructed by H-CNN-HR, HVTMFnet-HR, LBlock+Bicubic, LBlock+SRCNN, and LBlock+ESRT (Fig. 8). In the multiresolution temperature reconstruction experiments based on actual TDLAS measurements, the flame contour in the low-resolution temperature image reconstructed by DMFMMnet is more consistent with that of an annular burner, as compared with those reconstructed by H-CNN-LR and HVTMFnet-LR. The high-resolution temperature image reconstructed by DMFMMnet shows more explicit thermal-diffusion characteristics in the combustion field than those reconstructed by H-CNN-HR, HVTMFnet-HR, LBlock+Bicubic, LBlock+SRCNN, and LBlock+ESRT (Fig. 9). Additionally, compared with the peak temperature values retrieved by the other algorithms, the peak temperature values retrieved by DMFMMnet are more similar to the highest temperature values measured by the thermocouples.

Multiresolution spatial discretization modeling and multiresolution temperature reconstruction were introduced into TDLAS tomography. A multiresolution temperature tomography network (DMFMMnet) based on dual-domain multiscale feature merging was constructed. This network reconstructs temperature images at two spatial resolutions with different computing costs, which can balance between imaging time and resolution. The simulation results show that the average PSNR values of the low-resolution temperature images reconstructed by DMFMMnet are 24.95%‒32.79% and 0.66%‒3.28% higher than those reconstructed by H-CNN-LR and HVTMFnet-LR, respectively, in the SNR range of 25 dB‒40 dB. The average PSNR values of the high-resolution temperature images reconstructed by DMFMMnet are 32.63%‒38.67%, 2.34%‒6.18%, 3.18%‒9.24%, 3.22%‒6.48%, and 0.29%‒1.18% higher than those reconstructed by H-CNN-HR, HVTMFnet-HR, LBlock+Bicubic, LBlock+SRCNN, and LBlock+ESRT, respectively, in the SNR range of 30 dB‒45 dB. The flame contours in the low-resolution temperature images and the detailed features in the high-resolution temperature images reconstructed by DMFMMnet are more similar to the ground-truth phantoms, as compared with the algorithms investigated. Experiments based on actual measurements obtained from the TDLAS sensor show that the temperature images reconstructed by DMFMMnet can more accurately reflect the actual heat conduction in the combustion field, as compared with the algorithms investigated.