EfficientNetB7: Video image model for identifying potential risk in power transmission line corridor

Abstract

More frequently occurring severe weather events are a result of climate change, and power line operations are vulnerable to a variety of natural catastrophes. Unpredictable line failures due to natural disasters require additional maintenance efforts. To account for the cumulative impact of multiple natural disasters, this article proposes a method for assessing the reliability of transmission lines. This approach develops a theory of catastrophe risk quantification that considers the likelihood of the risk occurring, the extent to which the risk impacts the line, and the severity of the potential disaster. By integrating hierarchical analysis with entropy weighting techniques, we can determine the risk weights associated with various natural catastrophes. The usefulness of the proposed method is demonstrated by an example of a transmission line risk assessment for the Sichuan area, which is subjected to the combined impacts of multiple natural disasters. The technique successfully evaluates operational dangers to transmission lines from the combined impacts of natural catastrophes, according to the findings. Disaster recovery and the avoidance of line risks may both benefit from the evaluation outcomes. This paper presents a dataset that includes external damage hazards in power line passageways and proposes an efficient NetB7-based detection technique to mitigate the number of external destruction incidents caused by building equipment, such as excavators and lifting equipment. First, the SqueezeNet module enhances the exchange of data between feature levels in the spatial pyramid pooling layer, thereby minimizing background noise on exterior damage hazard targets. Ghost-shuffle convolution (GSConv) replaces regular convolution in the neck network, resulting in a highly lightweight and slim-neck feature merger structure. This enhances the extraction of tiny object features by combining deep semantic information with shallow detail characteristics, thereby decreasing the model's computational burden and parameter count. The dynamic head replaces efficientNetB7 and improves model identification by combining size, geographical, and demand attention methods. Finally, the smart intersection over union (WIoU) loss function optimizes model convergence and detection. The new method significantly improves upon the self-constructed dataset of transmission line corridors affected by external threats. The model meets real-time detection needs with a high recognition speed of 96.2 images per second, while also reducing the computing burden and parameter count.