To address the issues of high computational complexity and poor real-time performance of traditional inversion methods in electrical resistance tomography (ERT), which make it difficult to meet the high-precision detection requirements of industrial two-phase flow, this paper proposes a machine learning-based approach for reconstructing conductivity distribution in ERT. A dataset containing 10,000 single-target and dual-target samples is constructed via simulations on the EIDORS platform. Three models with different methodological principles — extreme learning machine (ELM), ridge regression, and random forest — are selected for comparative experiments. The imaging quality is evaluated using root mean square error (RMSE) and image correlation coefficient (ICC). Experimental results show that ELM achieves the best performance on both single-target and dual-target test sets, with RMSE values of 0.0353 and 0.0374, respectively, significantly outperforming the other two models. This method effectively alleviates the ill-posedness of the ERT inverse problem and provides a feasible solution for high-precision real-time imaging in industrial two-phase flow.