The auroral arc is the typical track of the interaction between the solar wind and the Earth's magnetosphere. A sketch of skeletons for arc‐like aurora is usually used to describe auroral structures, such as vortex, fold and curl structures, etc. With artificial intelligence technologies, sketching auroral skeleton structure (AuroSS) in all‐sky images enables automatic detection and measurement of aurora arcs in very large amounts of ground‐based auroral observation data. The skeleton is a highly characterizing topological structure that has been extensively studied in the field of computer vision. However, AuroSS is not the medial axis of auroral shapes and a large number of accurate AuroSS annotations are not available. It is difficult to detect AuroSS by using an unsupervised or fully‐supervised method. In this paper, we formulate the automatic AuroSS extraction to learn a mapping from an all‐sky auroral image to a ridge style AuroSS. Without accurate AuroSS annotations, emission ridge and coarse localization of aurora are incorporated to generate pseudo‐labels of AuroSS. A series of functional weakly supervised models are trained and cascaded to achieve AuroSS detection. Experimental results on auroral images obtained from all‐sky imagers at Yellow River Station (YRS) show that the detected AuroSS is consistent with that of human visual perception. Based on the obtained AuroSS, the orientations and lengths of auroral arcs can be estimated automatically. By browsing the temporal variation in arc orientation from dusk to dawn, we can acquire synoptic observations of auroral activities at YRS.