Target detection of substation electrical equipment from infrared images using an improved faster regions with convolutional neural network features algorithm

Abstract

The failure of substation equipment can cause incalculable losses to the economy and power consumption of the whole country. The use of infrared images is a powerful tool to obtain equipment temperature, which can then be used directly to diagnose substation equipment without stopping the operation of the equipment. In this paper, the authors focus on the correct identification of different types of electrical equipment from the infrared images. An improved faster regions with convolutional neural network features (faster R-CNN) algorithm is proposed, which shows very high detection accuracy for substation equipment. Firstly, the backbone of the faster R-CNN is optimised. A new network, the ResNet-30 network, is designed to reduce the redundancy of the ResNet-34 network and increases the proportion of residual blocks in the network in the previous stages. Next, the deep receptive field is combined with the shallow receptive field of the network and a double-shortcut structure with a large convolutional kernel is proposed. This enhances the ability of network feature extraction. A cross-channel shortcut is proposed at the channel transition of the network based on the channel number relationship between the dual-shortcut structures. Finally, the proposed method is compared with faster R-CNNs whose backbones are ResNet-50 plus a feature pyramid network (ResNet-50+FPN), you only look once v3 plus spatial pyramid pooling (YOLOv3+SPP) and a single-shot multibox detector (SSD). The results show that the improved model not only has a smaller number of parameters and low requirements for graphics processing unit (GPU) equipment, but also has the highest mean average precision (mAP) for mostly substation equipment in the test-set. This lays a foundation for fault diagnosis of substation equipment in the future.