基于双子空间并行回归的化工过程质量相关故障检测方法  

A chemical process quality-related fault detection method based on twin-space parallel regression

作　　者：宋冰[1] 郭涛[1] 侍洪波[1] 谭帅[1] 陶阳[1] 马浴阳 SONG Bing;GUO Tao;SHI Hongbo;TAN Shuai;TAO Yang;MA Yuyang(Key Laboratory of Smart Manufacturing in Energy Chemical Process,Ministry of Education,Shanghai 200237,China;China Railway 14th Bureau Group Corporation Limited,Jinan 250000,Shandong,China)

机构地区：[1]华东理工大学能源化工过程智能制造教育部重点实验室,上海200237 [2]中铁十四局集团有限公司,山东济南250000

出　　处：《化工学报》2023年第11期4600-4610,共11页CIESC Journal

基　　金：国家自然科学基金项目(62073140,62073141,62103149,62273147);上海市青年科技启明星计划项目(21QA1401800);上海市晨光计划项目(21CGA37);国家重点研发计划项目(2020YFC1522502,2020YFC1522505);中国铁建股份有限公司科研计划课题(2018-B06)。

摘　　要：邻域保持嵌入(neighborhood preserving embedding,NPE)是一种常用的无监督学习方法,在故障检测领域得到了广泛应用。由于NPE提取的数据特征无法解释输入数据和输出数据之间的关系,因此在化工过程质量相关故障检测方面存在局限性。另外,NPE在提取数据流形结构时忽略了动态信息的表征。为了解决上述问题,基于NPE和慢特征分析(slow feature analysis,SFA)算法提出了一种名为双子空间并行回归(twin-space parallel regression,TSPR)的质量相关故障检测方法,该方法能够同时提取数据的流形特征和变化速度信息。首先,通过基于互信息的策略将原始过程空间分为序列相关子空间和序列无关子空间,以应对变量在时间序列相关性的差异。其次,在两个子空间中分别应用提出的邻域保持-慢特征嵌入算法(neighborhood preserving-slow feature embedding regression,NP-SFE)和NPE算法提取数据的有效结构特征,并同时用最小二乘回归在两个特征子空间中构建过程变量与质量变量的回归关系。随后,通过对回归系数的协方差矩阵分解,得到质量相关子空间和质量无关子空间,进而在相应子空间建立统计量并估计其控制限。最后,将所提方法在典型案例上进行测试验证,以说明所提方法的有效性和合理性。Neighborhood preserving embedding(NPE)is a commonly used unsupervised learning method and has been widely applied in fault detection.However,the features extracted by NPE cannot explain the relationship between input and output data,which limits its application in quality-related fault detection in chemical processes.Moreover,NPE ignores the representation of dynamic information while extracting data manifold structure.To address these issues,this paper proposes a quality-related fault detection method called twin-space parallel regression(TSPR)based on NPE and slow feature analysis(SFA)algorithm,which can simultaneously extract manifold features and velocity information.First,the original process space is divided into serial correlated subspaces and serial correlated subspaces based on mutual information strategy to deal with the differences in time series correlation caused by sensors.Secondly,the proposed neighborhood preserving-slow feature embedding(NPSFE)algorithm and NPE algorithm are used to extract the effective structural features in two subspaces,and the regression relationship between process variables and quality variables is constructed by using least square regression in both feature subspaces to characterize the change trend of process variables and quality variables.Then,the covariance matrix of the regression coefficients is decomposed to obtain the quality-related subspace and quality-unrelated subspace,and monitoring statistics and control limits are established and estimated respectively.Finally,the proposed method is tested and verified on typical cases to illustrate the effectiveness and rationality of the proposed method.

关 键 词：邻域保持嵌入 慢特征分析 质量相关 最小二乘回归 故障检测 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]