In order to effectively identify the multi-source partial discharge phenomenon and solve the problems of high complexity of the model network, false detection and missed detection of small targets, a multi-source partial discharge type identification method based on improved DQ-BETR was proposed. Firstly, a multi physical signal detection platform for partial discharge was built, and the data were measured using ultrasonic, transient voltage and UHF detection methods. The typical single source and multi-source data of partial discharge were collected, and the partial discharge spectrum data set was constructed. Secondly, a new visual backbone network(VManba) was introduced to replace the original ResNet50 backbone network of DQ-DETR to extract more abundant image features. Based on the above improvements, a hybrid matching strategy was introduced to improve the efficiency of the training phase of DQ-DETR model and solve the problem of scarcity of positive samples. Finally, single source and multi-source partial discharge were identified by using different sensor data, and the identification results of the three sensors were fused by using D-S evidence theory. The experimental results show that good recognition results are achieved on the three different sensor data. Compared with the single sensor recognition results, the performance indexes of P, R, F₁ and mAP (IOU = 0.5) are improved by 2.03%, 2.01%, 1.95% and 1.76% on average. At the same time, compared with SSD, Yolov5, fast RCNN, DETR, anchor DETR and conditional DETR, the performance indexes of P, R, F₁ and mAP (IOU = 0.5) of the proposed method are improved by 1.52%, 8.24%, 5.07% and 2.99% on average. Therefore, the proposed method can not only effectively identify the categories of single source and multi-source partial discharge, but also significantly improve the convergence speed and accuracy of partial discharge category identification, and has good computational efficiency and versatility.