To enhance the "anti-forgetting" ability and incremental sample recognition capability of the transformer partial discharge (PD) recognition model during learning new samples

a PD incremental recognition method based on lossless estimation and balanced training is proposed. Firstly

during the model-update training

a lossless estimation of historical task gradients is introduced. High-precision estimation of the training gradient for the complete historical PD sample set is achieved by weighting a few critical sample gradients

thereby enhancing the model's "anti-forgetting" ability during updates. Secondly

a PD incremental recognition method based on balanced training is proposed

which adjusts the aggregation weights of new and historical tasks to keep the updates beneficial and maintain forward progress for both tasks. It can enable accurate recognition of samples from both tasks. The proposed method is validated using PD samples obtained from experiments and field sampling. The results show that the proposed method has low computational requirements

exhibiting strong "anti-forgetting" ability and incremental sample learning ability. The recognition accuracy of the updated model for test samples is no less than 98.18%

improving by at least 9.51%.