This study proposes a novel autonomous inspection framework for insulator strings in transmission facilities with an unmanned aerial vehicle (UAV). The proposed framework aims to overcome practical limitations of current inspection using the UAV for transmission facilities. Current inspection relies on manual piloting, which makes it difficult in consistent acquisition of inspection data under structural and environmental variability in transmission facilities. Reliable anomaly detection also remains challenging in real deployments because of scarce abnormal samples and diverse site conditions. To address these challenges, the proposed framework hybridizes an adaptive flight strategy with a novel deep neural network to autonomously inspect insulator strings without manual intervention. The proposed framework features three key characteristics. First, the adaptive flight strategy determines the position and orientation of the UAV to acquire high-quality images of insulator strings by analyzing the structural features of transmission facilities with multimodal information. Second, a novel deep neural network, titled hybrid anomaly detection via multi-scale variational autoencoder and classifier (HAD-VAEC), is proposed to detect abnormal insulator strings by effectively learning the distinct features between normal and abnormal patterns in a latent hyperspace. Third, synthetic insulator images are co-trained with real images using a novel co-training strategy, which are generated through three-dimensional computer-aided design modeling and generative adversarial networks. This feature aims to address data imbalance and improve the robustness of the HAD-VAEC. Extensive experiments on both virtual and real-world environments clearly demonstrate that the proposed framework enables efficient, safe, and autonomous inspection of insulator strings by addressing core technologies of the 4th industrial resolution.
Jeon et al. (Sun,) studied this question.