Aero-PHM-KGQA: A Multi-Granularity QA Dataset with Multimodal Knowledge Graph for Full-Cycle Aerospace PHM

Published: 18 August 2026| Version 1 | DOI: 10.17632/ycd38r2j6v.1
Contributors:
,
,
,

Description

KG extracted from 247 Chinese aerospace standards (GB, GJB, JB, QJ, DB, HB, JJG, YDT, YHB series) using ontology-supervised LLM-based extraction with posterior completeness reinforcement. QA dataset of 1,000,000 pairs generated via causal traversal over the knowledge graph and manually reviewed. Images extracted and curated from source standards, filtered for PHM relevance, and linked to knowledge graph entities.

Files

Steps to reproduce

To ensure full transparency and reproducibility, this section describes how to access and use the Aero‑PHM‑KGQA dataset. All data are provided in standard JSON format and can be read with any programming language that supports JSON parsing. No specialized KG software is required. + Dataset Components: The dataset consists of three main parts: 1. Knowledge Graph: `KG_data_en_zh.json` (10,066 entities, 12,559 relations) and `KG_data_with_images_base64.json` (with Base64‑encoded images). 2. QA Dataset: 500 volume files (`qa_dataset_7themes_expanded/vol_0001.json` … `vol_0500.json`), each containing 2,000 QA pairs. 3. Documentation: `kg_manager.py` (Python module for KG operations) and `to_doc.py` (document export script). + Reading the Knowledge Graph: The KG JSON file contains two top‑level fields: `"entities"` and `"relations"`. Each entity includes `entity_id`, `name_zh`/`name_en`, `class_zh`/`class_en`, `description`, `context`, `judgement`, `treatment`, `source_ref`, and (optionally) `imageUuids`. Each relation includes `source_id`, `target_id`, `class_zh`/`class_en`, `description`, `strength`, and `source_ref`. ```python import json with open('KG_data_en_zh.json', 'r', encoding='utf-8') as f: data = json.load(f) entities = data['entities'] # list of entity objects relations = data['relations'] # list of relation objects ``` The `kg_manager.py` module provides a higher‑level interface for querying the KG, including BFS‑based causal reasoning, product tree extraction, and FMECA generation. + Reading the QA Dataset: Each volume file is a JSON array of QA pairs. Each QA pair contains `question_zh`/`question_en`, `answers` (with C, B, and A levels), and `metadata` (theme, entity ID, target subset). Each answer level includes `answer_zh`/`answer_en`, `knowledge_elements_zh`/`knowledge_elements_en`, and `sourceRefs`. ```python import glob qa_pairs = [] for fname in sorted(glob.glob('qa_dataset_7themes_expanded/vol_*.json')): with open(fname, 'r', encoding='utf-8') as f: qa_pairs.extend(json.load(f)) ``` All code and instructions are provided to enable independent use and verification of the dataset.

Institutions

Categories

Aerospace Engineering, System Management, Agent-Based Technology, Knowledge Graph

Licence