An Electrified Railway Catenary Component Anomaly Detection Frame Based on Invariant Normal Region Prototype With Segment Anything Model
Haonan Yang, Zhigang Liu, Haorui Cui, Ning Ma, Hui Wang, Chen Zhang, Yang Song
Abstract
The existing drone-based detection systems for electrified railway catenary support components face the following challenges: (1) The background of aerial images of the catenary support component is complex; (2) A unified model struggles to handle anomaly detection for various component types; (3) There is inconsistency in the normal region feature information between training and testing images. To address these issues, this paper proposed a novel adaptive anomaly detection framework for catenary support components (INRP-ADer). First, we proposed a new segmentation model (CSC-SAM) that embeds key catenary component location information to extract foreground images. Next, we design an anomaly detection model that directly extracts invariant normal region prototype features (INRP) from the test images. This model includes an INRP extractor constrained by INRP smoothness loss and an INRP-guided decoder, aimed at solving the problem of inconsistency in normal region features between training and testing images and the challenge of adapting a unified model to multiple component types. Additionally, a soft mining loss is introduced to further optimize the training process of the multi-class anomaly detection model. Finally, we established a real-world catenary support component dataset (CSCUD) collected by drones, and achieved detection performance of 99.1/98.2/98.2 in image-level metrics (I-AUROC/I-AP/I-F1) and 96.4/37.1/55.7 in pixel-level metrics (P-AUROC/P-AP/P-F1). The proposed method outperforms traditional methods by approximately 0.4% to 16.6% across various metrics.