摘要
针对物联网异常检测中特征重叠和类不平衡导致的检测模型性能退化问题,提出了一种基于特征选择和梯度提升的检测方法。首先,从特征相关性、重要性和独特性等不同维度对特征进行了分组和排序,选择相关性高、重要性低的特征进行删除,提高了训练精度;然后,通过单边梯度采样和互斥特征捆绑,进一步减少无效或低效训练样本,提高了训练效率;最后,通过对比试验验证了该方法的有效性。试验结果表明,该方法在异常检测性能、稳定性和计算开销等方面具有显著优势,在大规模数据集的小样本异常检测中,具有良好的可用性。
Aiming at performance degradation in internet of things(IoT)anomaly detection models caused by feature overlap and class imbalance,an anomaly detection method based on feature selection and gradient boosting(AD-FS-GB)is proposed.Firstly,features are grouped and ranked across dimensions including correlation,importance,and uniqueness.Highly correlated yet low-importance features are selected and eliminated.Thus,training accuracy is enhanced.Then,through one-sided gradient sampling and exclusive feature bundling,invalid or inefficient training samples are further reduced,and training efficiency is improved.Finally,the method is validated with contrasting experiment.Experimental results demonstrate significant advantages on anomaly detection performance,stability,and computational overhead.The approach exhibits strong applicability for small-sample anomaly detection in large-scale dataset.
作者
胡杰
曹佃光
龙岭春
高冰
隋翯
HU Jie;CAO Dianguang;LONG Lingchun;GAO Bing;SUI He(The 28th Research Institute of China Electronics Technology Group Corporation,Nanjing 210023,China;CCCC First Harbor Engineering Company Ltd.,Tianjin 300461,China;College of Aeronautical Engineering,Civil Aviation University of China,Tianjin 300300,China;School of College Science and Artificial Intelligence,Civil Aviation University of China,Tianjin 300300,China)
出处
《指挥信息系统与技术》
2025年第5期40-51,共12页
Command Information System and Technology
基金
国家自然科学基金(U2333201)
空中交通管理系统全国重点实验室开放基金(SKLATM202306)资助项目。
关键词
物联网
异常检测
特征选择
梯度提升
internet of things(IoT)
anomaly detection
feature selection
gradient boosting
