<p>This paper proposes a noninvasive, image-based approach for detecting maneuvers and anomalous cases of incompletely closed high-voltage switches, a critical component of transmission substations. A challenge often encountered when training deep learning models for applications in substations is the acquisition of representative and balanced datasets. Notably, it is difficult to maneuver switches under a variety of lighting and weather conditions without disturbing services or raising safety concerns. We address this challenge by employing single-class anomaly localization models, which are trained with samples showing the switches only in their usual closed state, and can point out image regions that deviate from the norm in anomaly heatmaps. We compare 13 anomaly localization methods and propose improvements to one of them, leading to a solution that outperformed the others, achieving an F-score of 97.07% for image-level detection, while also producing explicit reconstructions that support explainability. To the best of our knowledge, this is the first work to propose a general approach to detect incompletely closed switches, and the first to address switch state identification as an anomaly localization/unsupervised learning problem. Our dataset will be made public, providing a new benchmark for testing other anomaly detection methods.</p>