Real-time Object Detection for Substation Security Early-warning with Deep Neural Network based on YOLO-V5

The substations have rigid security management regulations, including timely detection of fire hazards, detection of unauthorized persons and engineering vehicles in equipment areas. Conventional substation early-warning requires installing additional hardware such as infrared equipment, which is uneconomic and cannot simultaneously detect multiple types of security threats. In this paper, with the video surveillance information, a substation early-warning system is developed based on a real-time object detection algorithm and YOLO-v5 deep neural network. The proposed early-warning system establishes a regression model with a deep convolutional neural network combining the Backbone structure and PANet structure. The substation image dataset is augmented under multiple weather conditions and labeled to indicate multiple hazards. By minimizing the YOLO-v5 integrated loss functions with stochastic gradient descent, the deep neural network is trained to identify early-stage fires/smokes, unauthorized objects, and abnormal positions of vehicles in the substations. Experiments show that the proposed method can automatically identify multiple safety hazards in substations in real-time and enhance the security level of substations.