Physics-informed multi-fidelity learning-driven imaging method for electrical capacitance tomography.

The electrical capacitance tomography is considered to be a promising non-invasive visualization method for the measurement of multiphase flow parameters, but low-quality tomograms limit the reliability and accuracy of measurements. In order to overcome this limitation and bottleneck, the physics-informed data-dependent prior learned from the given samples is introduced and coupled with the measurement mechanism and the domain knowledge modeled by a new L1/L2 norm-based regularizer into a new optimization imaging model in this study. To improve the convergence and reduce the computational load, the built imaging model is solved by a new numerical scheme driven by the split Bregman method and the forward-backward splitting method. A new physics-informed robust sparse extreme learning machine method that not only obeys the measurement mechanism in training but also promotes the robustness and performance of the model is proposed and used to build a novel multi-fidelity learning method to infer the physics-informed data-dependent prior. The new imaging method achieves multi-source data fusion, leads to new changes in reconstruction algorithms, increases the complementarity and diversity of image priors, unifies the image reconstruction and multi-fidelity learning method, and improves robustness, numerical stability and flexibility of the model. The reconstruction results show that the new imaging method achieves more accurate reconstruction with reduced sensitivity to noise compared to important and popular imaging methods. This study leads to new changes in imaging algorithms and mining and use of image priors, and serves as a catalyst for the paradigm shift in image reconstruction.