Abstract
Power equipment maintenance work orders, a critical type of power system texts, are rich in operational details such as faulty components and maintenance procedures. However, automated information extraction from these orders is impeded by complex domain-specific terminology and intricate semantic structures. This paper proposes a novel multi-level feature enhancement framework to overcome these challenges. The framework's contributions are threefold: Firstly, a Hierarchical Knowledge-Driven Data Completion method is proposed to construct the Power Equipment Maintenance Named Entity Recognition (PEM-NER) dataset, leveraging raw data from the State Grid Corporation. Secondly, a Position-Aware Global Attention mechanism is developed and integrated within the transformer architecture. This mechanism effectively captures relative positional information and dataset-scale features, significantly enhancing contextual understanding for NER tasks. Thirdly, a Fine-Grained Information Enhancement Module is designed to refine character-level dependency analysis, thereby improving the precision of entity boundary detection. Extensive evaluations on the PEM-NER dataset and three public benchmarks demonstrate the proposed model's superior performance, especially in recognizing entities within power system texts. The framework exhibits promising applications in knowledge graph construction and question-answering systems within the field of power equipment maintenance.