Asymptotic domain adaptive detection for abnormal targets in transmission lines under complex weather conditions

The edge intelligence models commonly used to identify abnormal targets in transmission lines are easily affected by complex weather changes, triggering significant declines in detection accuracy and even detection failure for abnormal targets, thus compromising the safe and stable operation of transmission lines. An asymptotic domain adaptive detection method for abnormal targets in transmission lines under complex weather conditions is thus proposed in this paper. An approximate domain closer to the target domain is first constructed, utilising an image style transfer method; this is followed by a proposed multi-layer domain transfer method that integrates a spatial attention mechanism that causes the model to pay more attention to the feature information of abnormal targets, further shortening the distance between the approximate domain and the target domain. Finally, a SAM strategy is introduced into the domain adaptive task, creating a detection model for abnormal targets of transmission lines that is driven to access additional effective domain invariant information to make the landscape of the task loss function smoother. In this study, detailed experiments were thus carried out based on an abnormal target dataset from various transmission lines. The experimental results showed that the method improves the accuracy of abnormal target recognition under complex weather conditions to 91.97%, offering recognition accuracy obviously superior to that of a single-step domain adaptive SOTA method or a variety of SOTA-based edge detection models for abnormal targets in transmission lines, and thus displaying good application prospects.