Transmission lines often pass through mountainous and hilly areas with complex topography, where extreme climate and complex landscape lead to ice-covered lines, foreign objects and other faults, which threaten the safe operation of power grids. In this paper, for the problems of small target detection difficulty and insufficient multidimensional data fusion in transmission line hidden danger monitoring, a transmission line hidden danger monitoring method based on improved deep neural network is proposed. In the method, three key improvements are made on the basis of the YOLOX model: firstly, the loss function is improved, and the combination of IOU logarithmic operation and power function operation is used to enhance the punishment of difficult samples; secondly, the CBAM attention mechanism is introduced in the P2 layer of the backbone network to enhance the model’s ability to perceive small targets; finally, the upsampling structure is added to the neck network to promote the deep semantic information and shallow representation information fusion. The experimental results show that the improved model achieves an average accuracy mean value of 98.61% on the transmission line hazards dataset, which is 12.07% higher than that of the original YOLOX algorithm; the detection speed reaches 46.47 images per second, which is an improvement of 8.06 images per second; the size of the model is 148.81 MB, which is a reduction of 3.33 MB; and the training time increases by only 0.38 hours. The method effectively improves the accuracy and real-time performance of transmission line hazard monitoring, and provides a new technical support for power grid safety monitoring in multi-dimensional data environment.
