[Objective] To confront the enduring challenges in governing multi-source heterogeneous data from power generation equipment-such as the frequent occurrence of missing fields, semantic ambiguity, and the persistent problem where completion results violate complex business rules-this paper introduces an innovative data quality enhancement method that seamlessly integrates semantic disambiguation with knowledge graph constraints, aiming to achieve highly accurate missing-field completion and standardized normalization of fault-related texts. 

[Methods] Using equipment unique identifiers and timestamps as primary keys, a multi-source corpus is constructed by associating equipment ledgers, SCADA alarm records, and operation-maintenance logs. DeepSeek-V3 is then employed to generate multiple types of candidate completions based on existing fields and similar historical samples, while a domain-adapted BERT model-fine-tuned on power generation datasets with a partial freezing and layer-wise unfreezing strategy-evaluates the semantic consistency between candidates and contexts. The fine-tuning process uses a batch size of 16, the AdamW optimizer, and a maximum sequence length of 512. A local module further calculates candidate support degrees. Concurrently, a specialized knowledge graph for power generation equipment is built to explicitly define mandatory rules-including attribute ranges, equipment hierarchies, and temporal logic-alongside soft constraints. A dual-factor dynamic gating mechanism is designed to dynamically fuse semantic scores, candidate support degrees, and knowledge graph soft compliance scores using a semantic complexity factor S and a domain constraint strength factor C, where hard constraints strictly filter out invalid candidates and comprehensive reranking optimizes the final output. 

[Results] Extensive evaluations are conducted on seventy-six thousand three hundred and sixty actual operational records from thermal power, wind power, and photovoltaic systems of a regional subsidiary of a major power generation group spanning 2020 to 2024. The dataset is divided into training, validation, and test sets in a ratio of 7:1:2. The proposed method achieves macro-average completion accuracy, semantic consistency, and business compliance rates of ninety-five point zero percent, ninety-three point nine percent, and ninety-five point eight percent respectively, significantly outperforming standalone BERT, DeepSeek zero-shot, and fixed-weight fusion methods. Ablation studies further indicate that removing the knowledge graph constraints causes the business compliance rate to plummet to eighty-four point three percent, and removing the dual-factor gating degrades all three metrics, fully confirming the effectiveness of each core module. 

[Conclusion] By deeply synergizing semantic understanding with explicit business rules, the proposed method effectively overcomes semantic ambiguity and rule conflicts in data completion for power generation equipment, significantly improving data quality and text normalization levels, and providing a reliable technical reference for data governance in the energy and power sector.