【摘要】 现有智能故障诊断技术在设备识别和分割方面存在精度差、召回率低、推理速度慢的问题，在故障诊断方面存在诊断效率低，无法对具体设备类别进行诊断分析等不足。针对此问题，本文采用先分割后诊断的自诊断方法对变电设备进行红外故障诊断。在实例分割方面，选用推理速度较快的Efficient ViT网络进行改进：在骨干中改进Transformer级联结构增强网络对弱小目标的敏感性，添加改进的SK(Selective Kernel)注意力增加网络对变电设备的感兴趣能力；在颈部采用GD(Gather and Distribute)机制提升模型对不同尺寸特征的利用能力；在头部优化掩膜损失函数增强模型对变电设备边界的分割能力，为后续故障诊断提供准确的分割区域。在故障诊断方面，对分割后的设备区域采用面域相对温差法进行缺陷自诊断。实验结果表明，改进后的网络性能虽然在推理速度上略有下降，但在测试集上检测m AP值和分割m AP值分别提升了8.4%、6.0%，同时故障诊断准确率达到了91.3%。更多还原

【Abstract】 Existing intelligent fault diagnosis techniques exhibit poor precision, low recall, slow inference speed in equipment identification and segmentation, and low diagnostic efficiency in fault diagnosis as well as cannot diagnose and analyze specific equipment categories. To overcome these drawbacks, in this study, we perform infrared fault diagnosis of substation equipment using the self-diagnosis method of segmentation before diagnosis. An efficient Vi T network with faster reasoning speed is used to improve instance segmentation. The transformer cascade structure in the backbone is improved to increase the sensitivity of the network to weak targets, and the improved SK attention method is added to increase the network’s interest in the substation equipment. The GD mechanism is used in the neck to improve the model’s ability to use features of different sizes, and the optimized mask loss function is used in the head to improve the model’s ability to diagnose infrared faults in substation equipment. The loss function improves the segmentation ability of the model at the boundary of the substation equipment, providing accurate segmentation regions for subsequent fault diagnosis. In terms of fault diagnosis, the relative temperature difference in the area is used for fault selfdiagnosis in the segmented equipment region. The experimental results show that the performance of the improved network improves the detection and segmentation mAP values on the test set by 8.4% and 6.0%, respectively, although a slight decrease in the inference speed is observed. In addition, a fault diagnosis accuracy of up to 91.3% is obtained.更多还原