[Objective] The initial feature of single-phase grounding faults in distribution networks is weak, with a low signal-to-noise ratio. Traditional fault detection methods suffer from low detection accuracy and insufficient generalization capability when labeled data samples are limited. To address this issue, this paper proposes a TransTCN semi-supervised collaborative learning framework that integrates Transformer and temporal convolutional network (TCN). [Methods] Firstly, the improved complementary ensemble empirical mode decomposition (ICEEMD) method was employed to perform adaptive mode decomposition on fault zero-sequence current signal, thereby selecting the optimal feature components. Secondly, model training was initialized with a small number of labelled data sample, and the unlabeled dataset expanded through a high-confidence pseudo-label generation mechanism, and combined with a loss function featuring weight-adaptive allocation to achieve iterative optimization of model parameters. Finally, a single-phase grounding fault model for a 10 kV distribution network was constructed using PSCAD. The detection performance of the proposed TransTCN semi-supervised model was validated under varying grounding resistances, initial fault angles, and operational conditions. [Results] Under conditions where labeled data constituted merely 15% of the dataset, the proposed TransTCN semi-supervised model achieved an identification accuracy of 95.31% for weak feature single-phase grounding fault. [Conclusion] TransTCN semi-supervised model has significant advantages in weak feature extraction and few-sample learning scenarios. It performs well in terms of fault identification accuracy, convergence stability, and cross-condition generalization ability, and has certain engineering application value.