摘要
异常检测作为信息安全的重要环节,在早期发现问题、降低安全风险中发挥关键作用。针对石油炼化分馏单元的时序信号异常检测问题,设计了一种基于改进掩膜自编码器的时间序列异常检测方法。提出的模型采用U-Transformer结构的编解码器,增强了时间序列特征的多层次解析能力。为更有效利用时间序列的上下文信息,引入因果掩码策略和时滞可变长输出窗口策略,显著提升了异常检测精度。最后,基于Aspen Dynamics仿真平台生成的时序信号数据集,通过对比实验和消融实验验证了模型的有效性和可靠性。
As an essential aspect of information security,anomaly detection plays a vital role in early problem identification and risk mitigation.To address the challenges of time-series signal anomaly detection in petroleum refining and fractionation units,this paper proposes a time-series anomaly detection method based on an improved masked autoencoder.The proposed model incorporates a U-Transformer encoder-decoder structure to enhance the multi-level feature extraction capabilities of time series.To effectively utilize the contextual information of time series,this paper introduces a causal masking strategy and a variable-length latency output window strategy,significantly improving anomaly detection accuracy.Finally,using a time-series signal dataset generated from a petroleum fractionation simulation platform built with Aspen Dynamics software.
出处
《工业控制计算机》
2025年第12期90-92,共3页
Industrial Control Computer
基金
国家自然科学基金项目:面向石油炼化装置的攻击检测及动态安全失效分析仪器(62127808)资助。
关键词
异常检测
掩膜自编码器
因果掩码策略
时滞可变长输出窗口
anomaly detection
masked autoencoder
causal masking strategy
variable-length latency output window
