Against the background of the deep coupling between the global energy structure transition and the national energy security strategy, offshore oil and gas development is regarded as a core field for ensuring China’s independent energy supply. However, the long-term development of this field has been hindered by three major technical bottlenecks: low efficiency in multimodal data recognition; insufficient accuracy in extracting geological parameters from small-sample scenarios; and weak capabilities for structurally integrating knowledge across different sources.A full-process integrated scheme of data acquisition-knowledge transformation-application implementation was innovatively constructed. Multimodal recognition technology, the fine-tuning strategy of the Qwen 3.0 large model, and the knowledge graph construction method were combined. Chinese and English geological reservoir reports were used as the core data sources. Digital characterization and structured integration of four typical oil and gas reservoirs were completed. These reservoirs include fluvial facies reservoirs, delta facies reservoirs, marine facies oil reservoirs, and marine facies gas reservoirs. Results show that the conversion rate from unstructured data to structured knowledge is over 90%. The extraction accuracy of core geological parameters is proven to reach 87.07%. Data processing efficiency is increased by 85% compared with the traditional manual method.Reliable knowledge support is provided for the optimization of offshore oil and gas development plans and scientific decision-making. Important theoretical and practical value is offered for the promotion of digital transformation in the energy sector.
DING Zu-peng, DONG Yin-tao, ZHONG Qin-wan, CHEN Long, WANG Jun-jie, ZHONG Zhi, WANG Fen.
Construction Method of Offshore Geological Reservoir Knowledge Base Based on Multimodal Recognition and Large Model Fine-tuning[J].
Science Technology and Engineering, 2026, 26(27): 11740-11751 DOI:10.12404/j.issn.1671-1815.2600031
传统全参数微调方法在海上地质油藏小样本场景下存在明显局限性:一方面需要更新模型的所有参数,计算资源消耗巨大;另一方面小样本数据难以支撑模型充分训练,易出现过拟合现象。为此,采用“预训练+SFT(监督微调)+LoRA(低秩适应)”的组合微调方案,具体流程如下:预训练阶段,以海上油气领域公开文献与未标注地质油藏报告为语料,通过无监督学习对Qwen 3.0模型进行领域适配,使模型初步掌握海上地质油藏领域的专业词汇、语法结构与语义逻辑,该阶段采用掩码语言模型MLM(masked language modeling)任务,其损失函数公式为
ChenBin, ZouLiang, XuGang, et al. Current status and prospect of China’s marine natural resources survey[J]. Geological Survey of China, 2024, 11(4): 1-16.
FanBaitao, WuYi, YuanJunliang, et al. Progress and development suggestions of petroleum engineering technologies of CNOOC during the 14th Five-Year Plan[J]. Petroleum Drilling Techniques, 2025, 53(6): 28-36.
LiuChenglin, JinLianwen, BaiXiang, et al. Frontiers of intelligent document analysis and recognition: review and prospect[J]. Journal of Image and Graphics, 2023, 28(8): 2223-2252.
ZhangWenqi, LiuYuangang, LiShaohua, et al. Research on geological entity recognition based on BERTwwm and data augmentation[J]. Geological Review, 2024, 70(6): 2303-2314.
SunYuqin, XiaoJingting, WangHaichao. Named entity recognition of power operation and inspection based on multi-model fusion[J]. Science Technology and Engineering, 2023, 23(36): 15545-15552.
[15]
HuE J, ShenY, WallisP, et al. LoRA: low-rank adaptation of large language models[EB/OL]. arXiv Preprint, arXiv: 2106. 09685, 2022.
XingXintao, ChenPeng. Key element extraction method for 110 police cases based on open-source large model[J]. Science Technology and Engineering, 2025, 25(29): 12547-12559.
HeTingting, ZhangQiang. Application status and development trend of knowledge graph in oil and gas exploration and development[J]. Natural Gas Industry, 2024, 44(9): 55-67.
PengYurong, QiuQinjun, MaKai, et al. Research on geological information extraction combining domain knowledge guidance and prompt engineering[J]. Chinese Journal of Geology, 2026, 61(2): 654-663.
[22]
ZhangZ, LiuS, HuP, et al. UniTabNet: bridging vision and language models for enhanced table structure recognition[EB/OL]. arXiv Preprint, arXiv: 2409. 13148, 2024.
LinZening, WangJiapeng, JinLianwen. Survey on deep learning methods for visual information extraction[J]. Journal of Image and Graphics, 2023, 28(8): 2276-2297.
LiuZihao, ZhuZhaopeng, ZhuLin, et al. Intelligent monitoring and analysis method of drilling anomalies based on knowledge graph and large model[J]. Petroleum Drilling Techniques, 2026, 54(2): 31-44.
WangWenjun, ChenYouwang, ZhuYingru, et al. Intelligent auxiliary decision-making method for energy consumption anomaly of oilfield gathering and water injection system based on knowledge graph[J]. Petroleum Science Bulletin, 2025, 10(3): 620-632.
ZhangZhanyong, WeiHaotian, DongShaohua. Risk early-warning analysis of accident evolution in complex gas pipeline network based on knowledge graph[J]. Science Technology and Engineering, 2025, 25(31): 13651-13659.
CaoQiaozhuoran, WangSisi, ChenZugang, et al. Extraction method of geoscience scientific data names based on regular expressions[J]. Journal of Geo-Information Science, 2023, 25(8): 1601-1610.
[35]
YangL, ChenH, WangX, et al. Two heads are better than one: integrating knowledge from knowledge graphs and large language models for entity alignment[EB/OL]. arXiv Preprint, arXiv: 2401. 16960, 2024.