Quantum Parallel Transformer Model for Detecting Multiclass Insulator Faults

The mainstream detection model suffers from low insulator fault detection accuracy caused by the limitations of the complex background of unmanned aerial vehicle images and the large variation of fault scales. Therefore, this study presents a multi-class insulator faults DETR (MCIF-DETR) based on real-time detection Transformer (RT-DETR) for locating multi-class insulator fault regions in complex backgrounds. Firstly, this study proposes a multi-directional feature enhancement module to reduce information loss during network delivery. Secondly, to address the problem that the original backbone network cannot extract target features effectively, an efficient partially reparameterized convolutional network (EPRC-Net) is proposed to focus on target features. Then, a pyramidal feature aggregation distribution network is proposed to enhance the multi-scale target detection capability of the model by passing the aggregated global information to each feature level. Finally, this study introduces a quantum convolutional neural network in parallel with EPRC-Net for feature extraction to further augment the learning capability of MCIF-DETR. The results indicate that the average precision of the MCIF-DETR proposed in this study for the three classes of faults on the insulator dataset containing breakage, flashover, and drop faults are 96.7%, 94.1%, and 95.1%, respectively. Compared to existing methods, MCIF-DETR achieves a more balanced detection performance for different fault types while maintaining high detection accuracy.