Studies of electrical capacitance tomography image reconstruction based on improved CycleGAN

Abstract

Electrical Capacitance Tomography (ECT) technology offers significant advantages for the real-time, non-invasive, and visual monitoring of fluids within oil pipelines. Nonetheless, the unique characteristics of the soft field environment present challenges in the ECT image reconstruction process, including issues of non-linearity, under-determinedness, and ill-posedness. These challenges result in notable limitations in conventional reconstruction techniques, such as LBP, Landweber, and Tikhonov regularization, particularly concerning the identification of flow boundaries and the accuracy of reconstructions. To mitigate these challenges, this study introduces an ECT-GAN algorithm, which is based on an enhanced CycleGAN framework, aimed at improving both the quality and robustness of ECT image reconstructions. Building upon the R6K architecture(the generator architecture of resnet_6 blocks and the initial network of Kaiming), ECT-GAN incorporates four significant advancements that further optimize its performance: The integration of Huber loss for the evaluation of model prediction accuracy contributes to an improvement in the robustness of the model while simultaneously diminishing its sensitivity to outliers during the reconstruction process; The Vision Permutator (VIP) attention mechanism enhances the representation of target regions by capturing long-range dependencies along a single spatial dimension while preserving positional information along the other, thereby ensuring better retention of structural details in ECT images, improving flow boundary clarity, and reducing artifacts; The FusedConv convolution module, which optimizes computational efficiency by integrating batch normalization within convolution layers, reducing parameters and inference costs, and the adoption of a Receptive Field Block (RFB) with spatial pyramid pooling to expand the receptive field, thereby improving the capture of multi-scale contextual information. To validate the effectiveness of ECT-GAN, a simulation model was developed using COMSOL Multiphysics, followed by extensive experiments conducted under various flow conditions in both the simulation model and a static experimental system. These experiments aimed to evaluate the algorithm's ability to reconstruct ECT images accurately across different flow regimes, ensuring its reliability in real-world applications. The results confirm that ECT-GAN outperforms conventional methods in ECT image reconstruction, effectively addressing the challenges of underdetermination and nonlinearity in the reconstruction process, demonstrating enhanced imaging accuracy, robustness, and generalization capability in practical ECT systems.