现有特征分析方法往往仅关注时间或空间单一特征,忽视特征间关联和分布差异,导致难以分析高维数据复杂结构的问题。针对上述问题本文提出一种面向高维复杂数据的时空特征分析方法(SCASRCAE-Unsupervised Domain-Adversarial Neural Network,SS-UDANN)。该方法使用稀疏卷积自编码器(sparse regularization convolutional auto-encoder,SRCAE)作为无监督特征提取器,应用于改进的领域对抗网络(Domain-Adversarial Neural Network,DANN),并引入轻量级跨纬度空间通道注意力(Spatial Channel Attention,SCA)机制,在低计算成本下提取时空相关特征；同时,为解决特征分布差异问题,采用最大均值差异(Maximum Mean Discrepancy,MMD)进行正则化,训练特征提取器与域鉴别器获取域不变特征,实现数据增强。以工业领域网络异常数据检测为验证背景,应用SS-UDANN分析CICIDS2018数据集的时空特征,精确率为98.66%。并对实验室油气集输全流程攻防靶场数据进行特征分析,精确率为95.39%,验证了SS-UDANN的有效性,表明本数据处理方法具有应用推广价值。

To address the limitations of existing methods that only focus on temporal or spatial features, this study proposes a novel approach for analyzing high-dimensional complex data using spatiotemporal features. The method, named Spatiotemporal Sparse Regularization Convolutional Autoencoder Unsupervised Domain-Adversarial Neural Network (SS-UDANN), employs a Sparse Regularization Convolutional Autoencoder (SRCAE) as an unsupervised feature extractor and incorporates an enhanced Domain-Adversarial Neural Network (DANN). A lightweight cross-dimensional Spatial Channel Attention (SCA) mechanism is integrated to extract spatiotemporal features efficiently while keeping computational costs low. Maximum Mean Discrepancy (MMD) is applied to regularize the feature extractor and domain discriminator, facilitating the extraction of domain-invariant features for effective data augmentation. Validation of SS-UDANN on the CICIDS2018 dataset for industrial network anomaly detection shows an accuracy of 98.66%. Additionally, the model achieves 95.39% accuracy when applied to data from a laboratory oil and gas full-process attack-defense testbed, further confirming the method"s effectiveness and demonstrating its potential for broader application in data processing.

TP393???????

辽宁省自然科学基金项目（2023-MSLH-273）；辽宁省科学技术计划项目(2023JH1/10400082)；辽宁省人工智能创新发展计划项目（2023JH26/1030008）；辽宁省科技创新平台建设计划项目([2022]36号)