基于深度学习的工业水系统异常检测算法研究  

Research on deep learning-based anomaly detection algorithm for industrial water systems

作　　者：刘永利[1] 郭明媛 晁浩[1] LIU Yongli;GUO Mingyuan;CHAO Hao(College of Computer Science and Technology,Henan Polytechnic University,Jiaozuo 454003,China)

机构地区：[1]河南理工大学计算机科学与技术学院,焦作河南454003

出　　处：《水利信息化》2025年第2期31-38,44,共9页Water Resources Informatization

基　　金：国家自然科学基金项目(62273290);河南理工大学基本科研业务费专项项目(NSFRF240310)。

摘　　要：随着工业水系统智能化的发展,对运行安全与稳定的需求也日益增加,异常检测成为关键的研究方向,深度学习凭借在动态环境中的建模优势,逐渐成为异常检测的重要技术。综述深度学习在工业水系统异常检测中的应用,分别从有监督、半监督、无监督和自监督4个角度分析深度学习在工业水系统领域的研究进展,重点介绍AnomalyTrans、DCdetector和THOC等模型的特点及适用场景,并通过实验对比9种异常检测算法或模型,评估各算法或模型在工业水系统数据集中的表现。结果表明:深度学习算法在工业水系统中检测准确率整体优于传统的统计与机器学习算法,表现出更强的鲁棒性和更高的检测精度,其中AnomalyTrans模型适用于高稳定性场景,DCdetector模型在资源受限环境中表现突出,THOC模型在实时性和低计算开销的应用中具有显著优势,证明深度学习在不同规模工业水系统中的实用性,为工业水系统安全防护工程的落地提供创新路径。With the development of intelligent industrial water systems,the demand for operational safety and stability has increased,making anomaly detection a key research direction.Deep learning,with its advantages in modeling dynamic environments,has gradually become an important technology for anomaly detection.This study reviewed the application of deep learning in anomaly detection for industrial water systems,analyzing the research progress from four perspectives:supervised,semi-supervised,unsupervised,and self-supervised learning.The study focused on the characteristics and applicable scenarios of models such as AnomalyTrans,DCdetector,and THOC,and compared nine anomaly detection algorithms or models through experiments to evaluate their performance on industrial water system datasets.The results showed that deep learning algorithms have overall better detection accuracy in industrial water systems than traditional statistical and machine learning algorithms,demonstrating stronger robustness and higher detection accuracy.Specifi cally,AnomalyTrans model was suitable for high-stability scenarios,DCdetector model excelled in resource-constrained environments,and THOC model had signifi cant advantages in real-time applications with low computational overhead.These fi ndings confi rm the practicality of deep learning in industrial water systems of different scales and provide innovative pathways for the engineering implementation of safety protection in industrial water systems.

关 键 词：工业水系统 异常检测 深度学习 机器学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TV213.4[自动化与计算机技术—控制科学与工程]