Video-based mechanical state monitoring of isolating switches using conductive-tube included-angle trajectories

Abstract Conventional vision-based recognition of isolating switches mainly outputs discrete opening or closing states and therefore provides limited information about continuous mechanical motion during switching operations. To address this limitation, this paper proposes a video-based mechanical state monitoring method for isolating switches using conductive-tube included-angle trajectories. The time-varying included angle between the centerlines of the upper and lower conductive tubes is used as the primary visual mechanical feature. For a fixed-view video, the region of interest, joint center, and reference separating the two conductive tubes are initialized in the first frame and reused in all subsequent frames. Planar perspective rectification is applied to establish a common motion plane. The conductive-tube centerlines are then extracted using Canny-Devernay sub-pixel edge detection, Hough-based coarse localization, and common-normal total-least-squares fitting of paired outer boundaries. During motion, two-frame differencing generates a motion mask that constrains candidate edges extracted from the current frame. At stationary endpoints, such as the fully open and fully closed states, the procedure automatically switches to a single-frame geometric-edge extraction path. Field surveillance videos acquired in an operating high-voltage substation are used to demonstrate the processing procedure and trajectory extraction, whereas quantitative validation is conducted on a 384-frame closing video of a GW23A-252 isolating switch acquired using a fixed camera under indoor lighting. A nonlinear piecewise physical model incorporating the four-bar linkage, rack-and-pinion mechanism, and final-stage clamping mechanism is developed to map the measured spindle angle to a reference conductive-tube included angle. Frame-by-frame evaluation shows that the proposed method produces valid results for all 384 frames and achieves a mean absolute error of 2.06°, a root-mean-square error of 3.24°, and a coefficient of determination of 0.9967, compared with 2.82°, 4.08°, and 0.9948, respectively, for the YOLOv5s-based comparison method. Under strong controlled nonuniform illumination disturbance, the illumination-induced mean absolute error of two-frame differencing is 0.26°, lower than the 0.49° obtained using MOG2. These results demonstrate that the conductive-tube included-angle trajectory provides a geometrically interpretable visual feature for quantitative characterization of continuous isolating-switch motion.