Broken Power Strand Detection with Aerial Images: A Machine Learning based Approach

Abstract

Power lines are essential for electricity transmission between power plant and consumption point. Periodical inspection and assessment of the power line damages are critical to ensure the uninterrupted power delivery and grid stability. With the recent development of the unmanned aerial vehicles technology, the aerial images of power lines are adopted for broken strand detection. A huge challenge is the lack of the fatal but rare broken strand images. Thus, an oversampling strategy is proposed to increase the data diversity and reduce the data imbalance between the normal and broken lines. In addition, image background noises are filtered through image transformation to facilitate the anomaly detection. After that, five popular machine learning models are trained on four representative views of the aerial images. The experiments results show that the models can achieve a remarkable performance when they are trained and specialized for images from the same view. Second, the model can be generalized from one view to the other views sharing similar features, where the neural network solutions show remarkable knowledge transfer capability. Third, the impact of data size is discussed. More data does help promote the detection accuracy, but the performance gain diminishes in data size.