Abstract: Rock bolt support applies pretension through the interaction between the tray and the surrounding rock, and the development characteristics of the tray and surrounding rock fracture structures are the key to predicting anchorage void failure. Based on this, a method integrating Res-UNet and YOLOv5n is proposed for key information recognition of bolt trays and surrounding rock fractures, as well as for predicting anchorage void failure. Field images of roadway bolt support in underground coal mines were collected, involving a total of 1 579 bolts, among which 79 were manually identified as anchorage void failures, 147 as potential failures, and the rest as normal support. A joint labeling system of bolt trays and surrounding rock fractures was constructed, and the Albumentations algorithm was employed for image augmentation to build a training set suitable for both detection and segmentation tasks. The YOLOv5n model was then used for automatic localization of bolt trays, while a residual U-Net was introduced to achieve automatic recognition and segmentation of surrounding rock fracture regions. Finally, the spatial geometric relationship between fracture regions and the bolt trays (including their effective range) was analyzed to classify anchorage void failure risk levels. Experimental results show that the proposed Res-UNet–YOLO model achieved a precision of 0.993 and recall of 0.932 for bolt tray detection; and for fracture segmentation, a precision of 0.895, recall of 0.893, F1-score of 0.915, accuracy of 0.988, and mean IoU of 0.918. Taking the minimum Euclidean distance D_min from the tray center to the fracture boundary, the tray radius R, and the effective range of a single bolt (6r, with r being the bolt radius) as criteria: when D_min<R, the bolt is classified as anchorage void failure (Level I, severe risk); when R≤D_min<6r, it is classified as potential failure (Level II, high risk); and when D_min≥6r, it is classified as no failure (Level III, stable). The model predicted 72 bolts as Level I, 135 bolts as Level II, and 1 372 bolts as Level III, with an overall prediction accuracy of 97.72%. By level, the prediction precision for failure bolts was 0.847, for potential failure bolts 0.867, and for stable bolts 0.995. This method demonstrates significant application value in bolt support quality inspection and surrounding rock stability assessment, providing a new approach and technical support for intelligent monitoring and risk-level decision-making in coal mine support systems.