工业场景多源异构数据异常自适应检测研究

工业数据异常检测是保障工业安全的关键基础工作，具有重要的研究意义。工业场景数据来源多样且数据结构动态多变，黑白名单等传统异常检测方法不能针对多源异构数据进行分析，也无法针对工业数据及异常攻击变化进行自动化适应检测。本项目面向工业场景数据的异常检测需求，研究多源异构数据异常特征检测机理，多角度全面描述系统状态，构造高维复杂工业数据特征空间自适应模型，并在此基础上建立自适应检测导向的异常检测分析模型，从特征空间优化和异常分析检测两个方面实现自适应调节。不同于固定规则的检测方式，本项目创新性地提出一种应对工业数据动态变化的自适应异常检测算法，支持多源异构工业数据处理，能够自动适应工业数据的变化进行异常分析，并选用能源电力、航天航空等工业场景进行试验验证。本项目研究有望为工业场景异常检测提供新的技术路线。

Anomaly detection of multi-source heterogeneous industrial data is very important to ensure the security of industry, and it attracts more and more attention from academic and industry. In industrial scenes, the sources of industry data are diverse and the structures of data space are dynamic. Traditional anomaly detection methods such as the black-and-white list cannot be used for the analysis of multi-source heterogeneous industrial data. Moreover, these methods cannot detect the attacks from dynamic industrial data and industrial anomalies automatically and adaptively. In this project, we focus on the need for anomaly detection among industrial data. We investigate the feature mechanism of the multi-source heterogeneous data with a comprehensive description of the system state from multi-angles. With the extracted data, we build the feature space of the high-dimensional and complex industrial data adaptively. Then we develop the corresponding programming models and algorithms for adaptive detection of abnormal industrial data, which have the ability to optimize the feature space and the anomaly detection analysis model according to the dynamic industrial data and industrial anomalies automatically. Instead of setting fixed rules to detect anomalies, this project innovatively constructs a new adaptive anomaly detection scheme for detecting abnormal industrial data, which has the ability to process the multi-source heterogeneous industrial data. The algorithms and structures proposed in this work will be validated by typical industrial scenarios such as energy, electricity and aeronautics. The results of this work provide a new solution for industrial anomaly detection.