SQL语言
汽车工业
计算机科学
领域(数学)
跟踪(心理语言学)
加入
可再生能源
水准点(测量)
语义学(计算机科学)
排
领域(数学分析)
数据挖掘
地球观测
知识库
数据库
碳足迹
还原(数学)
SPARQL公司
相似性(几何)
精确性和召回率
足迹
电
组分(热力学)
数据建模
能量(信号处理)
作者
Ninghui Sun,Qi Chun Dong,Xiaokang Liu,Yi Wang
出处
期刊:
[SPIE]
日期:2025-12-02
卷期号:: 54-54
摘要
In the dual-carbon era, automotive manufacturing plants deploy thousands of energy meters, PLC loggers and MES tables to trace electricity, thermal and material flows across stamping, welding, painting and final-assembly lines. These systems generate hundreds of inter-linked tables—covering kWh-level renewable share, kg-level CO₂e per vehicle, and t-level carbon-quota balances—that domain engineers query with extremely long and intricate SQL to monitor real-time carbon footprints, optimize green-electricity dispatch and forecast compliance gaps. Yet the corresponding natural language questions remain strikingly short (8–12 tokens): “2024 BEV life-cycle CO₂?” or “top three plants with renewable ratio >60 %?” This profound length/complexity mismatch cripples conventional Text-to-SQL retrieval: dense or sparse encodings collapse the fine-grained semantics of nested, multi-join carbon-accounting queries into a single vector, yielding poor recall and degraded end-to-end accuracy. We present EM-SQL-Carbon, a retrieval method that re-casts the similarity between a terse carbon question and a verbose SQL query as an optimal-transport problem via Earth Mover’s Distance between token-level embeddings. By preserving the compositional structure of carbon-indicators, joins and filters, EM-SQL-Carbon avoids semantic collapse and delivers markedly higher recall. Evaluated on a real-world dual-carbon automotive dataset spanning 2.3 M rows of plant-level emissions and renewable-energy logs, EM-SQL-Carbon improves Hit@10 by 27 % over state-of-the-art baselines, setting a new benchmark for Text-to-SQL in sustainable automotive manufacturing.
科研通智能强力驱动
Strongly Powered by AbleSci AI