摘要
工控系统的安全防护研究中面临数据特征矩阵分布相关性不确定、特征相关性难以判定等难题,导致深度学习模型异常检测准确性存在波动。通过对仿真油料储运系统数据的研究,提出了CNNreorder特征螺旋式重排序方法。将每个样本螺旋式生成满足空间局部最大相关性的合成矩阵,并降低冗余序列。首先,构建特征列间相关系数矩阵、特征与标签列相关系数向量。其次,改进了mRMR排序约束框架,由矩阵中心开始环状螺旋式生成排序合成矩阵,使得模型在空间相邻区域能够提取最大相关特征。在2个公开和3个自建数据集的实验表明,排序后得到的数据特征热力图更趋合理,数据量减少为同类算法的1/4;对比4种经典分类算法精度提升平均25%。提高了异常检测的准确度和稳定性。
The security protection of industrial control systems is an important key to ensuring the security of the national infrastructure.However,it faces challenges such as uncertainty in the distribution correlation of data feature matrices and difficulty in determining feature correlations,leading to fluctuations in the accuracy of deep learning models.A CNN reorder feature spiral reordering method was proposed throughthe study of simulated oil storage and transportation system data.Each sample was generated in a spiral manner to generate a synthetic matrix satisfying the local maximum correlation,and the redundant sequences are reduced.Firstly,the correlation coefficient matrix between feature columns and the correlation coefficient vector between features and label columns were constructed.An improved mRMR sorting constraint framework was proposed,which generates a composite matrix by performing a circular spiral sorting of the matrix center according to the constraints,enabling the model to extract the maximum relevant features in adjacent spatial regions.Compared with similar algorithms,experiments on 2 publicly available and 3 self-built datasets showed that the sorted data resulted in a more reasonable feature heatmap,with a reduction in data volume to the top 1/4 of the improvement.Compared to four classic classification algorithms,the average accuracy improvement is 25%.
作者
刘学君
王汝墨
江逸楠
刘家
LIU Xue-jun;WANG Ru-mo;JIANG Yi-nan;LIU Jia-yi(School of Information Engineering,Beijing Institute of Petrochemical Technology,Beijing 102617,China)
出处
《计算机仿真》
2025年第11期430-437,共8页
Computer Simulation
基金
北京石油化工学院致远基金项目(2023006)
2023年北京石油化工学院重点教改项目(ZD202303002)。
关键词
工业控制系统
特征螺旋式重排序
深度学习
异常检测
样本升维
Industrial control systems
Feature spiral reordering
Deeplearning
Anomaly detection
Sample dimensionality enhancement
