Abstract: In order to realize the accurate identification and location of insulator anomaly, an improved lightweight network model YOLOv5 insulator state anomaly detection method TR-YOLOv5 is proposed. The Transformer-Encoder module based on the self-attention mechanism was used to improve the feature extraction network to improve the detection accuracy of the model. A prediction layer for small targets was added, and K-means clustering algorithm was used to design target anchor frame parameters to provide more low-level feature information for subsequent feature fusion. The EIoU was used as the loss function to optimize the loss value curve, and the multi-scale data enhancement strategy was combined to achieve high-precision positioning of the target. The experimental results show that the accuracy of the TR- YOLOv5 model can reach 94. 2%, which can effectively identify abnormal insulator targets.