Gas-Insulated Switchgear (GIS) partial discharge (PD) condition assessment is a critical task in ensuring the operational reliability of power equipment. However, existing models face two primary challenges. First, they fail to effectively leverage the shared and differential learning between the interconnected tasks of diagnosis, localization, and severity assessment, which limits the overall performance of the model. Second, these models typically rely on a large volume of labeled data for training, making them unsuitable for real-world scenarios where only a limited number of labeled samples are available, such as in-field operations. To overcome these limitations, this paper proposes a novel self-supervised representation learning approach for GIS PD condition assessment. We introduce a multi-task network that jointly addresses the diagnosis, localization, and severity assessment tasks by exploiting the commonalities and distinctions between these tasks, thus improving the model’s overall performance. Furthermore, we incorporate a self-supervised representation learning strategy to enable effective model training with minimal labeled data. This approach not only enhances the accuracy of GIS PD severity assessment with limited labeled samples but also significantly reduces the dependency on large annotated datasets. Experimental results demonstrate that the proposed self-supervised representation learning method achieves performance comparable to supervised learning, even in scenarios with scarce labeled data. This work offers a promising solution for GIS PD condition assessment in practical field applications, especially where labeled data is limited, thereby contributing to the efficient monitoring and maintenance of GIS.