To address the common issues of false and missed detections in the current process of foreign object detection on power transmission lines, as well as the limited detection accuracy due to the variability in the size of foreign objects, this paper proposes an AAGV-YOLOX model for the detection of foreign objects on power transmission lines. The model first designs an Adaptive Dilated Convolution (Adaptive Dilated Convolution, ADConv) and constructs a feature extraction module (Adaptive Dilated Convolution Module, ADCM) to effectively distinguish the widely distributed foreign objects from background information, thereby enhancing the model's feature extraction capabilities. Subsequently, an Adaptive Receptive Field Feature Fusion (Adaptive Receptive Field Feature Fusion, ARFFF) module is introduced into the neck network to fully integrate features of different scales, further improving detection accuracy. Finally, the GVFL loss function is proposed, which not only increases the convergence speed of the proposed network but also enhances the localization accuracy. Experimental results show that the average precision mean of this model on the self-built dataset of foreign objects on power transmission lines reaches 90.34%, with a 5.56 percentage points improvement over the YOLOXs, demonstrating the effectiveness of the proposed method in improving the detection of foreign objects on power transmission lines.

In the context of digital operation and maintenance of power grids, the surge in unstructured transformer data has made defect information extraction and fault tracing challenging, hindering the transition to intelligent maintenance. Knowledge graph technology offers a potential solution by leveraging its structured nature to integrate operational data and improve efficiency. Inspired by this, this paper proposes a unified short-text format for electrical equipment defects and constructs a high-quality transformer defect relationship dataset. The ET-FSUIE (Electrical Transformer-Fuzzy Span Universal Information Extraction) model is introduced, which integrates a 20% pruned Roformer v2 pre-trained language model. By utilizing its rotary position encoding, the model effectively handles variations in defect description text lengths, enhancing text comprehension. Additionally, a W-FSL (Wasserstein-Fuzzy Span Loss) loss function based on Wasserstein distance is proposed to overcome the limitations of traditional loss functions and improve extraction accuracy. Experimental results on both public and self-built datasets demonstrate the superior performance of the ET-FSUIE model, achieving F₁ scores of 81.84% and 88.67%. Finally, a knowledge graph for power transformer defect relationships is constructed using the extracted triplets, providing robust support for the intelligent transformation of power equipment operation and maintenance.