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; and 3) there is inconsistency in the normal region feature information between training and testing images. To address these issues, this article proposed a novel adaptive anomaly detection framework for catenary support component an anomaly detection model based on invariant normal region prototype extraction (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 (INRP) features 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. In addition, a soft mining loss is introduced to further optimize the training process of the multiclass anomaly detection model. Finally, we established a real-world catenary support component dataset, catenary system component UAV 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%–16.6% across various metrics.

The core task of intelligent detection in electrified railway catenaries is to detect their supporting components. However, due to the large number and variety of components in catenary inspection images, labeled catenary data is often limited, and few studies have focused on leveraging large amounts of unlabeled datasets in this field. This article proposes a novel self-supervised pretraining model catenary support rod masking-based masked image modeling (CSRM-MIM) for detecting catenary support components (CSCs), which effectively utilizes the valuable information in unlabeled catenary data. Specifically, a new semantic catenary support rod area masking (CSRA masking) strategy based on the CSRA features is proposed to guide the model in learning meaningful catenary semantic information during the pretraining process. Additionally, a new Siamese pretrained network framework is designed, incorporating an information interaction enhancement module (IIEM) and a dual reconstruction network to perform dual reconstruction on the semantic mask, thereby extracting global features in the catenary domain. Finally, a multiscale knowledge distillation (MSKD) strategy optimized by the cross-layer fusion module cross-layer fusion decoding (CLFD) is introduced to assist the self-supervised model in acquiring specific-general representations for the catenary field and transferring multiscale information, which benefits subsequent detection tasks. Experimental results demonstrate that this method significantly enhances the detector’s performance in identifying CSCs, owing to the retained self-supervised learning (SSL) strategy.

The Catenary Support Component (CSC) is a critical element in ensuring the safe operation of electrified railway systems. In recent years, the use of UAVs for railway maintenance and inspection has steadily increased, making them an increasingly important technological tool. However, UAVs face several challenges during railway inspections. These challenges include energy consumption constraints, which limit their prolonged use, as well as complex backgrounds, structural occlusions, and varying component scales that hinder detection accuracy. To address these challenges, a low-energy consumption, graph-guided spike-based brain-inspired neural network algorithm (GSINet) was proposed. First, a novel backbone called the spatiotemporal multiscale spike fusion network was designed. This network leverages spatiotemporal spike encoding for feature modeling and enables cross-scale feature fusion. Temporal and channel attention mechanisms, along with SpikeASPP technology, were used to enhance the network’s ability to handle diverse scales and complex environmental conditions. Next, a Spike-inspired detection head was designed with a position guidance module (PGM) based on graph convolutional networks. The PGM utilizes prior category relationships, integrating both semantic and geometric information to mitigate challenges such as occlusion and multi-component measurement. Finally, a spiking network architecture was implemented using the improved Leaky Integrate-and-Fire (IB-LIF) Neuron. This reduces energy consumption by exploiting sparse computations and an event-driven mechanism, while ensuring high detection accuracy. Extensive experiments on a CSC dataset captured by UAVs demonstrate that the model achieves superior measurement performance while consuming significantly lower energy.

As an essential component of the train system, the wheelset is prone to defects on its tread surface due to the influence of high-frequency impacts during operation. However, due to the complex and changeable wheel-rail contact conditions, tread defects face problems such as complex backgrounds, diverse scales, and small defects. An interactive fusion distillation network embedded with Poisson-coded images (RP-IFDN) is proposed to address the above issues for tread defect detection. Specifically, a dual backbone network (DStarNet) embedded with Poisson coded images is designed to extract and fuse the texture features of RGB images and the edge features of Poisson (POS) images, thereby alleviating the problem of complex background; An adaptive interactive squeeze-excitation module (AISE) is designed to extract the interactive information between RGB features and POS features and enhance the focus on essential areas, thereby improving the efficiency of feature fusion and suppressing background information; A global-context feature pyramid (GC-FPN) is proposed to capture global-context information when fusing cross-layer features, thereby alleviating the multi-scale problem and enhancing detection capability for small target defect; A straightforward knowledge distillation strategy is proposed to transfer the knowledge from a dual-backbone detector to a single-backbone detector. This approach addresses the challenges of excessive parameters and computational complexity associated with dual-backbone architectures. The effectiveness of the proposed method has been demonstrated through a series of comparative experiments, ablation experiments, visualization analysis, and generalization performance experiments. Index Terms-Defect detection of train wheelsets, image fusion, multi-scale features module, interactive attention mechanism, knowledge distillation.

Catenary support components (CSCs) are essential for the safe and efficient operation of electrified railway systems. However, detecting CSCs in images presents significant challenges due to the scarcity of labeled data, the presence of complex and diverse backgrounds, and the difficulties associated with multi-scale variations. To tackle these issues, this paper introduces a novel detection framework designed explicitly for CSCs. First, a Siamese-based self-supervised learning framework is designed as a pre-training strategy to reduce the reliance on labeled data, effectively leveraging unlabeled images and significantly lowering annotation costs. This pre-training approach enables the model to focus on identifying and extracting relevant features from prior knowledge, honing its ability to discern key patterns and structures within the data. Second, the Vision Attention-based Intrascale Feature Interaction (Vision-AIFI) and Relation Vision Module (RVM) are proposed to enhance the model, which can strengthen its ability to extract multi-scale features and effectively address challenges posed by complex backgrounds and scale variations. Third, a Dempster–Shafer (DS) evidence theory-based detection head is inserted to improve classification confidence and localization precision, ensuring accurate detection results in complex inspection scenarios. Finally, a UAV-based dataset for CSCs is constructed and validation experiments are performed. To evaluate the model, we used several standard COCO metrics, including mAP (77.84), APs (67.84), APm (70.31), and APl (90.04). In addition, the framework is further evaluated for Domain Generalization, which can demonstrate its strong adaptability and high detection accuracy for real-world CSC detection tasks.