In a small-current grounded system, the fault characteristics are very obscure when a single-phase ground fault occurs; therefore, the faulty line must be selected to remove it. This paper proposes a fault-line selection (FLS) method based on multi-classifier, which transforms FLS into a multi-classification problem. It solves the problems in traditional methods, such as low accuracy and high equipment cost. Multi-classifiers based on denoising Autoencoder(DAE) are used to reduce the dimension of historical dispatching data and extract single-phase ground-fault features. Firstly, the dispatching data are preprocessed to eliminate useless data and fill in vacancies. Then, the fault segments are marked and labeled samples containing steady-state and transient information of single-phase ground faults are obtained. Finally, a multi-classifier based on DAE is built, and this model is trained with labeled fault samples to obtain a high-accuracy FLS model. The experiments show that the accuracy of the proposed method exceeds 97%, which is much better than other data-driven models and traditional methods. The proposed method has been operating for over two years in a real power system south of China. The excellent performance of the proposed method for FLS in practice and simulation indicates a vast application potential.