Abstract: The inconsistency between the distributions of training and testing samples is a major factor contributing to the low accuracy of on-site partial discharge (PD) recognition in deep learning methods. In order to achieve continuous adaptation of PD pattern recognition to changes in data distribution and reduce the workload associated with sample labeling, an online learning method for PD pattern recognition is proposed. First, leveraging features from different layers of the PD pattern recognition model, an inference model is employed to distinguish newly added on-site PD samples within and outside the training set distribution. Soft labels and manual annotation are utilized for online labeling of these two types of samples. Then, to balance the quantity of samples within and outside the training set distribution and enhance the recognition accuracy of newly added samples, the paper employs a conditional wasserstein generative adversarial network with gradient penalty (CWGAN-GP) to augment both types of PD samples. This augmentation is integrated into the joint training of the PD pattern recognition model. Experimental validation using PD samples obtained from both laboratory experiments and on-site measurements demonstrates that the proposed method reduces the labeling workload by 66.68%. After online learning, the recognition accuracy for newly added samples within the training set distribution improves by 4.61%, and for samples outside the distribution, the minimum improvement in accuracy is 22.27%.