Abstract: Objective In response to the challenges posed by the diverse states and complex imaging conditions for nuts in real inspection environments, it is aimed to achieve automated detection of loosened nuts in rigid catenary systems, thereby improving inspection efficiency and reducing manual labor costs. Method A nut loosening detection method based on image segmentation and a three-stage progressive serial classification algorithm is proposed. First, the YOLOv8 backbone network is employed to segment catenary images, identify and locate bolt-nut regions. Then a parameter-sharing serial classification network is constructed to discriminate nut cleanliness, judge anti-loosening wire conditions, and the alignment status of these wires. The classification tasks share the same backbone network for feature extraction through a parameter-sharing mechanism, significantly reducing computational load and increasing detection speed. Result ＆ Conclusion Multi-angle catenary images captured by the roof camera of the comprehensive inspection vehicles from China Academy of Railway Sciences are used as the test dataset. The proposed method is tested and validated through actual catenary image datasets. The results show that the proposed method achieves a precision rate of 86.73% and a recall rate of 97.70% for detecting loosened nuts, meeting engineering application requirements. Compared with conventional methods, the parameter-sharing serial classification framework established by the proposed method can efficiently and accurately detect loosened nuts in rigid catenary systems, providing technical support for catenary condition monitoring and maintenance.