MAFCD: Multi-level and adaptive conditional diffusion model for anomaly detection

作     者：Wu, Zhichao Zhu, Li Yin, Zitao Xu, Xirong Zhu, Jianmin Wei, Xiaopeng Yang, Xin 

作者机构：Dalian Univ Technol Sch Comp Sci & Technol Dalian 116024 Peoples R China Dalian Univ Technol Key Lab Social Comp & Cognit Intelligence Minist Educ Dalian 116024 Peoples R China Dalian Univ Technol Sch Control Sci & Engn Dalian 116024 Peoples R China Dalian Univ Technol Key Lab Intelligent Control & Optimizat Ind Equipm Minist Educ Dalian 116024 Peoples R China Liao Ning Oxiranchem Inc Liaoyang 111003 Peoples R China 

出 版 物：《INFORMATION FUSION》 (Inf. Fusion)

年 卷 期：2025年第118卷

核心收录：

学科分类：08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：Natural Science Founda-tion of China [U21A20491] Liaoning Provincial Natural Science Foundation of China [2023-MSBA-001] 

主　　题：Multi-level feature Adaptive feature fusion Conditional diffusion model Multi-step sampling ensemble 

摘      要：In the real-world Internet of Things (IoT) systems, a variety of Internet-connected sensory devices, spanning from chemical processing equipment to material handling machinery and server machines are typically monitored with multivariate time series. Anomaly detection in these systems is pivotal for identifying potentially dangerous or unsafe conditions and implementing timely preventive measures. However, the complex contextual dependencies and diversified patterns inherent multivariate time series, such as seasonal fluctuations and trends in industrial processes, present significant challenges for existing anomaly detection methods, which strike a balance between fidelity and diversity in multivariate time series analysis. To address these issues, a novel Multi-level and Adaptive Conditional Diffusion model, called MAFCD, is proposed for anomaly detection across various industrial devices. The architecture of MAFCD is built upon a conditional diffusion model framework, guaranteeing both high-fidelity and diversity in generated multiple time series samples through adaptive fusion strategy and multi-level feature information. In particular, the model offers real-time anomaly occurrences by dynamically adjusting fusion weights across multiple features. Moreover, to enhance model stability, anomaly recognition results undergo weighted aggregation using exponential and symbolic ensemble function through multi-step sampling. Empirical evaluation across four public datasets and its application in an ethylene oxide production process demonstrates the superior performance and practical utility of the proposed MAFCD, underscoring its robust generalization ability.