基于CNN-WSVM的阀门粘滞检测方法  

A Valve Stiction Detection Method Based on CNN-WSVM

作　　者：田静 王志国 刘飞 TIAN Jing;WANG Zhi-guo;LIU Fei(MOE Key Laboratory of Advanced Process Control for Light Industry,Jiangnan University)

机构地区：[1]江南大学轻工过程先进控制教育部重点实验室

出　　处：《化工自动化及仪表》2024年第5期805-810,829,共7页Control and Instruments in Chemical Industry

基　　金：国家自然科学基金重点项目(批准号:61833007)资助的课题。

摘　　要：针对实际中粘滞与非粘滞数据不均衡问题,提出基于卷积神经网络和加权支持向量机融合的阀门粘滞检测方法(CNN-WSVM)。首先搭建仿真控制回路生成不同工况的回路运行数据,然后利用一维卷积神经网络模型提取特征,再使用惩罚项系数可调的加权支持向量机实现对阀门状态的分类。在仿真数据集和开源工业数据集中的验证结果表明:该方法相较于仅使用卷积神经网络,在数据不均衡情况下,能够显著提高粘滞数据检测准确率,同时在真实工业场景中具有较好的泛化能力。Considering disequilibrium between stiction and non-stiction data in practice,a valve stiction detection method based on convolutional neural networks and weighted support vector machines(CNN-WSVM)was proposed.Firstly,a simulation control loop was made to generate operational data under various conditions;and then,a one-dimensional CNN model for feature extraction was adopted,including making use of WSVM with adjustable penalty coefficients classify the stiction states.The verification results on simulation datasets and open source industrial datasets show that,compared with using CNN only,the method proposed can significantly improve detection accuracy of sticky data in the case of unbalanced data,and has good generalization ability in real industrial scenes.

关 键 词：阀门 粘滞检测 不均衡数据 卷积神经网络 加权支持向量机 阀门状态分类 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]