科研管理

科研管理 ›› 2026, Vol. 47 ›› Issue (1): 181-192.DOI: 10.19571/j.cnki.1000-2995.2026.01.018

• 论文 • 上一篇    

多特征融合的人工智能沉睡专利识别与测度

贾怡炜1，胡剑1，戚湧1,2   

  1. 1.南京理工大学知识产权学院，江苏 南京210094；
    2.南京理工大学经济管理学院，江苏 南京210094
  • 收稿日期:2024-06-05 修回日期:2024-12-07 接受日期:2024-12-09 出版日期:2026-01-20 发布日期:2026-01-12
  • 通讯作者: 戚湧
  • 基金资助:
    国家社科基金重点项目：“新型举国体制下打赢关键核心技术攻坚战的目标、主攻方向与对策研究”（23AZD038,2023.01-2025.12)。

Identification and measurement of sleeping patents of artificial intelligence with multi-feature fusion

Jia Yiwei1, Hu Jian1, Qi Yong1,2

  1. 1. School of Intellectual Property, Nanjing University of Science and Technology, Nanjing 210094, Jiangsu, China;
    2. School of Economics and Management, Nanjing University of Science and Technology, Nanjing 210094, Jiangsu, China

  • Received:2024-06-05 Revised:2024-12-07 Accepted:2024-12-09 Online:2026-01-20 Published:2026-01-12
  • Contact: qi, yong

摘要：    探索构建人工智能沉睡专利识别模型，挖掘潜在具有市场价值的专利，对提升专利运营效率和促进产业数智化转型具有重要意义。基于中国人工智能发明专利数据，对专利运营形势和睡眠特征进行全貌分析，利用AdaBoost算法和SHAP解释方法构建多特征融合的沉睡专利识别模型并进行实证分析。研究发现：中国人工智能运营专利数量呈稳步增长趋势，总体呈现出“睡眠时长短、唤醒机制灵活、唤醒强度低”的睡眠特征；基于AdaBoost算法的沉睡专利识别模型性能最优，将算法特征纳入指标体系能够明显提高模型的识别准确度；不同专利特征对沉睡专利唤醒产生交互效应，交通运输和电子设备制造领域的人工智能专利市场运营潜力较大；应用场景中预测结果的召回率为0.982，具备市场价值的专利比重为21.23%，验证了模型的有效性。本研究不仅丰富了沉睡专利识别与评估的研究体系，还为创新主体优化专利运营决策提供了实践方案。

关键词: 人工智能, 沉睡专利, AdaBoost算法, SHAP解释方法, 专利识别, 专利运营

Abstract: Exploring the construction of artificial intelligence sleeping patent identification model and mining potential patents with market value is of great significance to improve the patent operation efficiency and promote the transformation of industrial digital intelligence. Based on China′s AI invention patent data, this paper conducted a holistic analysis of the patent operation situation and sleeping characteristics, and used the AdaBoost algorithm and SHAP interpretation method to construct a multi-featured fusion sleeping patent identification model for an empirical analysis. The results showed that (1) the number of Chinese AI operation patents shows a steady growth trend, and the overall sleep characteristics of "short sleep duration, flexible wake-up mechanism, and low wake-up intensity"; (2) the sleeping patent identification model based on AdaBoost algorithm has the best performance, and the inclusion of algorithmic features in the metrics system can significantly improve the identification accuracy of the model; (3) different patent features have interactive effects on the awakening of sleeping patents, and AI patents in the field of transportation and electronic equipment manufacturing have high potential for market operations; and (4) the recall rate of prediction results in the application scenarios is 0.982, and the proportion of patents with market value is 21.23%, which verifies the validity of the model. The above conclusions will enrich the research system of identification and evaluation of sleeping patents, and provide practical solutions for optimizing the patent operation decision-making of innovation subjects.

Key words: artificial intelligence, sleeping patent, AdaBoost algorithm, SHAP interpretation method, patent identification, patent operation