Abstract
Knowledge graph construction typically involves data preprocessing, entity recognition, relation extraction, knowledge fusion, conflict resolution, and quality assessment. However, heterogeneous documents in substation operation and maintenance (O&M), including technical standards, defect records, work tickets, operational logs, and scanned or mixed-format documents, exhibit dense domain terminology, inconsistent expressions, cross-source conflicts, and incomplete provenance, which make accurate and traceable knowledge construction challenging. Existing large-language-model-based and pipeline-based methods have improved automatic knowledge extraction, but their processing stages are usually weakly coupled, making it difficult to reuse cross-stage evidence, trace downstream errors to responsible upstream tasks, and perform targeted rework. To address these limitations, this paper proposes a shared-task-state-driven multi-agent collaborative knowledge graph construction framework (MAC-CKG). The framework introduces a task orchestrator and a shared task-state pool to coordinate five agents for data processing, entity recognition, relation extraction, knowledge fusion, and quality assessment; provenance, confidence, conflict, and review information are preserved across stages, while ontology constraints, multifactor confidence, and the two-threshold quality-routing mechanism jointly support knowledge insertion, human review, and targeted rework. For evaluation, we constructed a heterogeneous substation O&M dataset comprising 96 heterogeneous source items (including documents and structured knowledge resources) and 17,220 standardized input units, yielding a provenance-verified candidate subset of 9519 entities and 9258 relations. An independently annotated 20-document end-to-end benchmark was separately used for raw-document extraction evaluation. Across three independent runs on this benchmark, MAC-CKG achieves mean entity, conditional relation, and strict triple F1 scores of 85.96%, 62.50%, and 62.31%, respectively, outperforming SAC-KG by 2.83, 5.77, and 6.34 percentage points and PiVe by 4.48, 8.32, and 8.94 percentage points. Independent-benchmark ablation further reduces triple F1 from 62.31% to 59.68% without task orchestration and to 57.41% without the shared task-state pool, showing sensitivity to these components under the evaluated configurations, with inference cost considered separately.