en
×

分享给微信好友或者朋友圈

使用微信“扫一扫”功能。
作者简介:

郭昕刚(1979—),男,吉林长春人,副教授,硕士,硕士生导师,E-mail:6889068@qq.com;

程超,男,吉林长春人,副教授,博士,博士生导师,E-mail:125725673@qq.com

通讯作者:

程超,男,吉林长春人,副教授,博士,博士生导师,E-mail:125725673@qq.com

中图分类号:TP277

文献标识码:A

文章编号:1001-2486(2023)06-165-09

DOI:10.11887/j.cn.202306019

参考文献 1
LI Z C,YAN X F.Complex dynamic process monitoring method based on slow feature analysis model of multi-subspace partitioning[J].ISA Transactions,2019,95:68-81.
参考文献 2
HUANG J,ERSOY O K,YAN X F.Fault detection in dynamic plant-wide process by multi-block slow feature analysis and support vector data description[J].ISA Transactions,2019,85:119-128.
参考文献 3
JIANG Q C,GAO F R,YI H,et al.Multivariate statistical monitoring of key operation units of batch processes based on time-slice CCA[J].IEEE Transactions on Control Systems Technology,2019,27(3):1368-1375.
参考文献 4
LI S,ZHOU X F,PAN F C,et al.Correlated and weakly correlated fault detection based on variable division and ICA[J].Computers & Industrial Engineering,2017,112:320-335.
参考文献 5
LI R Y,RONG G.Fault isolation by partial dynamic principal component analysis in dynamic process[J].Chinese Journal of Chemical Engineering,2006,14(4):486-493.
参考文献 6
LEE J M,YOO C Y,LEE I B.Statistical monitoring of dynamic processes based on dynamic independent component analysis[J].Chemical Engineering Science,2004,59(14):2995-3006.
参考文献 7
WISKOTT L,SEJNOWSKI T J.Slow feature analysis:unsupervised learning of invariances[J].Neural Computation,2002,14(4):715-770.
参考文献 8
LI R F,WANG X Z.Dimension reduction of process dynamic trends using independent component analysis[J].Computers & Chemical Engineering,2002,26(3):467-473.
参考文献 9
SHANG C,HUANG B,YANG F,et al.Slow feature analysis for monitoring and diagnosis of control performance[J].Journal of Process Control,2016,39:21-34.
参考文献 10
MACGREGOR J F,JAECKLE C,KIPARISSIDES C,et al.Process monitoring and diagnosis by multiblock PLS methods[J].AIChE Journal,1994,40(5):826-838.
参考文献 11
GE Z Q,SONG Z H.Distributed PCA model for plant-wide process monitoring[J].Industrial & Engineering Chemistry Research,2013,52(5):1947-1957.
参考文献 12
TONG C D,SONG Y,YAN X F.Distributed statistical process monitoring based on four-subspace construction and Bayesian inference[J].Industrial & Engineering Chemistry Research,2013,52(29):9897-9907.
参考文献 13
CONTRERAS-REYES J E,ARELLANO-VALLE R B.Kullback-Leibler divergence measure for multivariate skew-normal distributions[J].Entropy,2012,14(9):1606-1626.
参考文献 14
YOUSSEF A,DELPHA C,DIALLO D.Enhancement of incipient fault detection and estimation using the multivariate Kullback-Leibler Divergence[C]//Proceedings of 24th European Signal Processing Conference(EUSIPCO),2016.
参考文献 15
WANG B,JIANG Q C,YAN X F.Fault detection and identification using a Kullback-Leibler divergence based multi-block principal component analysis and Bayesian inference[J].Korean Journal of Chemical Engineering,2014,31(6):930-943.
参考文献 16
周伟,潘海鹏,吴平,等.基于DPCA和KL散度的微小故障检测方法[J].传感器与微系统,2020,39(3):135-138.ZHOU W,PAN H P,WU P,et al.Tiny fault detection method based on DPCA and KL divergence[J].Transducer and Microsystem Technologies,2020,39(3):135-138.(in Chinese)
参考文献 17
柯亮,熊伟丽,徐保国.基于滑动窗PCA的微小故障检测[J].小型微型计算机系统,2016,37(6):1360-1364.KE L,XIONG W L,XU B G.Small fault detection based on moving windowPCA[J].Journal of Chinese Computer Systems,2016,37(6):1360-1364.(in Chinese)
参考文献 18
SEGHOUANE A K,AMARI S I.The AIC criterion and symmetrizing the Kullback-Leibler divergence[J].IEEE Transactions on Neural Networks,2007,18(1):97-106.
参考文献 19
DOWNS J J,VOGEL E F.A plant-wide industrial process control problem[J].Computers & Chemical Engineering,1993,17(3):245-255.
参考文献 20
YIN J,YAN X F.Mutual information-dynamic stacked sparse autoencoders for fault detection[J].Industrial & Engineering Chemistry Research,2019,58(47):21614-21624.
参考文献 21
LI Z C,YAN X F.Complex dynamic process monitoring method based on slow feature analysis model of multi-subspace partitioning[J].ISA Transactions,2019,95:68-81.
参考文献 22
LI Z C,YAN X F.Fault-relevant optimal ensemble ICA model for non-Gaussian process monitoring[J].IEEE Transactions on Control Systems Technology,2020,28(6):2581-2590.
目录contents

    摘要

    针对传统分块方法根据经验划分子块导致变量特征信息无法充分利用,其单一的建模方式忽略局部信息以及离线模型无法适应时变特性的问题,提出了一种KL (Kullback-Leibler) 散度多模块滑动窗口慢特征分析方法。在正常工况数据集中,利用KL散度来度量变量间的距离,同时引入最小误差平方和准则进行聚类,分成两个距离最小的子模块;在此基础上利用慢特征分析方法对每个子模块进行建模,结合滑动窗口对每次采样的数据进行更新,得到最优模型,分别计算监测统计信息,利用支持向量数据描述对故障监测结果进行融合,实现故障诊断。并将该方法应用于田纳西伊斯曼过程的监控中,得到了较高的故障检测率和较低的虚警率,验证了该方法的可行性和有效性。

    Abstract

    A KL(Kullback-Leibler) divergence multi-block moving window slow feature analysis method was proposed to solve the problems that the variable feature information cannot be fully utilized by the traditional block segmentation method based on experience, the local information is ignored by a single modeling method, and the off-line model cannot adapt to the time-varying characteristics. KL divergence was used to measure the distance between variables in the normal working condition data set, and the minimum error sum criterion was introduced to cluster, which was divided into two sub-blocks with the minimum distance. On this basis, the slow feature analysis method was utilized to model each sub-block, and the optimal model was obtained by updating the sampled data with moving window. Monitoring statistics were calculated respectively, and the fault monitoring results were fused with support vector data description to achieve fault diagnosis. The proposed method was applied to the monitoring of Tennessee Eastman process, and higher fault detection rate and lower false alarm rate are obtained, verify the feasibility and effectiveness of this method.

  • 随着现代工业的发展,数据驱动方法在过程监控中发挥着重要作用[1-2]。在主流方法中,多元统计过程监控(multivariate statistical process monitoring,MSMP)被用于监控多元复杂工业过程中的故障[3]。传统的多元统计监控主要包括主成分分析法(principal component analysis,PCA)、偏最小二乘法(partial least square,PLS)和独立成分分析[4](independent component analysis,ICA)。

  • 工业过程的复杂性和未知的动态特性使得静态过程的监测效果较差。传统的MSMP方法是静态过程监控方法,即当前时间的样本与过去的样本没有关联,样本数据相互独立,忽略了动态特性。Li等提出了一种部分动态主成分分析(dynamic principal component analysis,DPCA)来提高动态过程监测能力[5]。针对过程动力学和数据非高斯统计的特点,动态独立成分分析(dynamic independent component analysis,DICA)也相继被提出[6-7]。上述方法以消除动力学为目标,往往假设所有变量具有相同的动态特性,然后对原始数据进行扩展,与传统方法相比提高了性能,但产生了大量的冗余信息,处理动态特性较强的变量时性能较差。慢特征分析[8](slow feature analysis,SFA)可以从时间序列中提取缓慢变化的特征,表征变量变化的快慢程度,是一种有效的无监督算法。Shang等提出了基于SFA的动态监控,实现了运行和控制性能的监控[9]。上述方法虽然克服了动态缺陷,但只建立单一模型却忽略了局部信息,导致大规模工业过程的诊断性能较差。

  • 多模块算法最早由Macgregor等提出,有效利用局部信息改进过程分析将整个模型分为多个子模型[10]。Ge等根据主成分分析的不同方向对原始数据进行划分,将线性变量分配到同一个块中,分块过程中存在缺失和重叠的问题[11]。Tong等分别分析了变量与主元子空间和残差空间的相关性,再将变量分配到相应的子空间中实现分块[12]。这些方法都以先验知识作为前提,极大程度限制了该方法的应用。KL(Kullback-Leibler)散度作为一种新的概率测度,可以测量两个统计变量之间的差值,用来解决故障诊断问题[13-14]。Wang等利用KL散度将具有类似统计特征的变量划分为一块,在每个低维子空间中建立PCA模型,使用贝叶斯策略进行融合[15]。周伟等在DPCA建模的基础上,利用 KL 散度量化模型得到分向量概率分布之间的相似度,从而建立多块模型实现对微小故障的诊断[16]

  • 在实际的工业生产过程中,大多数缓慢变化的过程数据也具有非线性时变的特性,其使正常数据出现偏移,导致出现误警现象。柯亮等提出了基于滑动窗PCA的微小故障检测方法,利用滑动窗口的策略对数据实时更新,使得故障数据与非故障数据分化明显,实现了对微小故障的放大,提高了模型的自适应能力[17]

  • 综合以上问题,提出一种基于KL散度的多模块滑动窗口慢特征分析故障诊断方法,利用KL散度的统计特性,建立多块模型,在每个块中应用SFA来提取过程数据的不同动态,利用滑动窗口得到最优模型,克服依据先验知识分块的策略、单一模型不稳定等问题。在田纳西伊斯曼(Tennessee Eastman,TE)过程监控中取得了较好的诊断效果。

  • 1 理论基础

  • 1.1 KL散度

  • KL散度通常用来测量两个概率密度分布的差异,广泛应用于模型的选择[18]。两个连续变量v1x)和v2x)的KL散度DKLv1v2表示为

  • DKLv1v2=v1(x)lnv1(x)v2(x)dx
    (1)
  • 由于KL散度的定义具有不对称性,不能作用于距离的度量。在实际的应用中,人们将其改进为对称形式

  • DKLv1v2=v1(x)lnv1(x)v2(x)dx+v2(x)lnv2(x)v1(x)dx
    (2)
  • 式(2)可得到对称的距离度量结果,具有非负性,v1x)和v2x)相似性越大,DKLv1v2值越接近于0,反之,DKLv1v2值越大,相似性越小。

  • 1.2 慢特征分析

  • 假设存在m维时间序列输入信号x(t)=[x1t),x2t),···,xmt)]T,SFA的目标是找到一组特征函数gt)=[g1t),g2t),···,gmt)]T,使得特征st)=gxt))缓慢变化[2]st)=[s1t),s2t),···,smt)]T用来表示缓慢变化的特征,SFA的优化问题表示如下

  • Δsi=mingi() s˙i2t
    (3)
  • 约束条件为

  • sit=0
    (4)
  • si2t=0
    (5)
  • ij,sisjt=0
    (6)
  • 其中,〈·〉ts˙分别表示均值和s对时间的导数。上述约束条件既排除了常数解,又保证所有解独立。

  • 线性SFA中,每个缓慢变化的特征(slow features,SFs)是所有输入数据的线性组合

  • s=Wx
    (7)
  • 式中,W=[w1w2,···,wm]T是权重矩阵。

  • R=x(t)x(t)Tt=UΛUT
    (8)
  • 式中,UR的特征值,Λ是由特征值组成的对角矩阵。白化矩阵可表示为Q=Λ-1/2UT,白化过程为

  • z=Λ-1/2UTx=Qx
    (9)
  • 结合式(7)和式(9),SFs进一步表示为

  • s=Wx=WQ-1z=Pz
    (10)
  • 式中,P=WQ-1zzTt=QxxTQT=Izt=0。优化问题(3)和约束条件进一步表示为

  • Δsi=mingi() s˙i2=minpiTz˙z˙Tpi
    (11)
  • sit=piTzt=0
    (12)
  • si2t=piTzzTtpi=piTpi=0
    (13)
  • ij,sisjt=piTzzTtpj=piTpj=0
    (14)
  • 为了满足式(11)~(14),优化问题再次转化为求解正交矩阵P

  • z˙z˙T=PTΩP
    (15)
  • 式中,P=[p1 p2,···,pm]是正交特征向量矩阵,对应的特征值为Ω=diag(λ1λ2,···,λm),且λ1λ2≤···≤λm

  • 权重W表示为

  • W=PQ=PΛ-1/2UT
    (16)
  • 最后,得到m个缓慢程度由大到小排列的SFs。

  • s=Wx=PQx=PΛ-1/2UTx
    (17)
  • Δsi=mingi() s˙i2=minpiTz˙z˙Tpi=λi
    (18)
  • 2 基于KL散度多模块滑动窗口慢特征分析方法及过程监测

  • 2.1 KL散度分块策略

  • 使用KL散度度量任意两个变量间的相关性,并进一步构造KL散度分量,作为分块的基础。

  • 对于训练数据X=[x1x2,···,xm]TRn×mmn表示变量数和样本个数。取任意两个变量xixj,且xiXxjX,在正态分布的条件下,KL散度分量表示为

  • di,jKL(x)=12σiσj+σjσi+μi-μj21σi+1σj-2
    (19)
  • 式中,μi=1mm=1i xiσi=1mm=1i xi-μi2。得到度量变量的对称矩阵dKLRm×m,该对称矩阵为变量的相关性矩阵。随机选取两类数值c1c2dKL作为初始中心,分别计算对称矩阵到每个初始中心的欧氏距离,将离初始中心最近的类确定为新的初始中心,其他类也被重新分配,依次迭代进行,直到满足式(20)时,分块结果收敛。

  • J(c,i)=i=12 cidKL dKL-ci2
    (20)
  • 最后将训练数据分成两个子模块X=[x1x2,···,xm]T = [X1X2]T

  • 2.2 多模块滑动窗口慢特征分析方法

  • 根据工业过程的动态特性,将子模块X=[X1X2]T扩展d个时延,使当前样本与过去样本相关联,得到增强的过程矩阵,在SFA建模过程中扩展了动态特性,矩阵增强过程如式(21)所示。

  • Xi=x(t)Tx(t-1)Tx(t-d)Tx(t+1)Tx(t)Tx(t+1-d)Tx(t+n-1)Tx(t+n-2)Tx(t+n-d-1)T
    (21)
  • 式中,i是子模块数量,n是过程变量的样本量。

  • 随后进行SFA离线建模,局部建模过程中使用滑动窗口算法训练最优的离线模型,如算法1所示。

  • 算法1 滑动窗口训练最优模型

  • Alg.1 Moving window trains the optimal model

  • 以上模型通过将样本数据根据欧氏距离升序排列,使用滑动窗口算法对样本进行筛选,得到最优的局部模型及相应的权重矩阵W1W2,提高了模型的自适应能力。

  • 为了实现对故障的检测,分别构造T2S2统计量对故障进行在线监测,T2表示统计量在慢特征空间中的静态变化,S2表示过程统计量的动态变化分布。T2统计量定义为

  • T12=s1Ts1
    (22)
  • T22=s2Ts2
    (23)
  • 其中,子块的SFs为si=Wixii=1,2。S2统计量定义为

  • S12=s˙1Ts˙1s˙1Tt-1s˙1
    (24)
  • S22=s˙2Ts˙2s˙2Tt-1s˙2
    (25)
  • 根据定义式分别得到子块对应的阈值和测试样本,由于多个统计量很难同时被监测,利用支持向量数据描述(support vector data description,SVDD)集成策略将子块结果转换为更加直观的监测指标。将阈值Y=[T21T22S21S22]作为输入,建立SVDD模型,如算法2所示,进行融合。

  • 算法2 SVDD建模及监测

  • Alg.2 SVDD modeling and monitoring

  • 监测结果融合后,得到阈值控制限Dt=1和半径中心DR。当DDt时,检测出故障,反之DDt,正常工作。该方法通过增强过程数据,实现对故障的动态监测,局部建模的方式避免了全局建模中忽略局部信息的问题,通过KL散度度量变量相似性,迭代实现无监督分块,得到两个相关性更高的样本数据,监测过程中有效解决了单一模型不稳定的问题。

  • 2.3 KL-MWSFA过程监测

  • 基于KL散度的多模块滑动窗口慢特征分析方法(multi-block moving window slow feature analysis method based on KL divergence,KL-MWSFA)的过程监测详细步骤如下。

  • 离线训练:

  • 步骤1:对训练数据X标准化,计算变量间的KL散度分量dKL

  • 步骤2:随机取dKL中的两个值作为中心,根据损失函数依次迭代更新中心数值,直到收敛,划分为两个子块X=[X1X2]T

  • 步骤3:将子块进行d个时延扩展为动态矩阵,计算与测试数据的欧氏距离,将其升序排列。

  • 步骤4:利用滑动窗口对每个子块进行SFA离线建模,分别计算T2S2的阈值。

  • 在线监测:

  • 步骤5:对故障数据标准化,扩展d个时延,然后分块。

  • 步骤6:使用SFA进行训练,得到T2S2的阈值,作为SVDD的输入Y

  • 步骤7:建立SVDD模型,计算测试向量Yt到超球半径的距离,得到DR

  • 步骤8:控制限Dt=1,当DDt时,发生故障,反之正常。

  • 3 TE过程仿真实验

  • 3.1 TE过程简要介绍

  • TE过程是美国化工公司在大量实际工程经验的基础上,由Downs和Vogel[19]提出的一个化工过程模型,其产生的数据具有时变性、强耦合性以及非线性,广泛应用于过程控制、监测和故障诊断研究,工艺流程如图1所示。主要由反应器、冷凝器、压缩机、气液分离器以及汽提塔5个操作控制部分构成。其共有33个变量,正常情况下的基准数据集包含500个样本。故障数据共有960个样本,前160个是正常数据,第161到960个样本为故障数据。本实验选择22个过程变量和11个操作变量作为输入,用500个正常样本建立模型。仿真共有21个故障类型,其中16种是已知故障,剩余5种是未知故障,实际的建模和监测过程中,一般不包括由搅拌速度引起的故障21,故采用15种已知故障进行测试,具体描述如表1所示。

  • 图1 TE过程流程图

  • Fig.1 TE process flow chart

  • 表1 TE过程故障类型表

  • Tab.1 TE process fault type table

  • 3.2 分块

  • 随机选取两个变量作为初始中心,计算KL散度分量构成的对称矩阵到初始中心的欧氏距离,离初始中心最近的分量重新确定为新的中心,依次迭代,最后收敛时将变量自动分成两个子块。具体分块策略见第2.1节,变量间的分布如图2所示,分块结果如表2所示。

  • 3.3 TE仿真实验结果分析

  • 将KL-MWSFA和核主成分分析(kernel principal component analysis,KPCA),ICA,SVDD进行对比分析,故障检测率如表3所示,可以看出,对于故障1,2,4,5,6,7,8,9,10,11,12,13,KL-MWSFA方法都能有效检测到这些故障。由于故障3,9,15的震级非常小,几乎所有的方法都无法检测到故障。但对于故障5,10,12,该方法具有较好的检测性能。

  • 图2 变量的分布

  • Fig.2 Distribution of sub-blocks

  • 表2 分块结果

  • Tab.2 Block the results

  • 故障5是冷凝器冷却水入口温度的阶跃变化,导致冷凝器温度发生变化[20]。故障5的监测图如图3所示,图3(a)在第300个采样点以后,无法检测出故障。图3(b)I2统计量在第161个采样点之前发生虚警现象,图3(c)在故障正常检测一段时间后,大部分统计量向控制限下方波动,检测效果一般。图3(d)在第161个采样点检测到故障后,一直持续到最后一个样本,检测效果较好。

  • 表3 KPCA,DICA,SVDD和 KL-MWSFA的故障检测率

  • Tab.3 Fault detection rate of KPCA, DICA, SVDD and KL-MWSFA

  • 图3 故障5监测结果

  • Fig.3 Monitoring results for fault 5

  • 故障10是C进料温度的随机变化,主要影响汽提塔压力[21]。图4为KPCA,DICA,SVDD和KL-MWSFA的监测图,图4(a)的T2统计样本大部分在控制限下面,检测效果不佳,图4(c)的监控数据在控制限附近波动,也无法清楚地检测到故障的发生,由表3可见DICA和KL-MWSFA检测效果较好,但在图4(b)的子图中,我们可以看到个别正常样本出现虚警,只有图4(d)在检测故障的同时没有发生虚警状况。

  • 故障12是冷凝器冷却水入口温度的随机变化[22]。由图5(a)可知,T2监测性能不好且在阈值附近上下波动,图5(b)虽然可以检测到故障,但在正常工况时发生虚警,图5(c)在发生虚警的同时监测效果没有图5(b)的效果好,而图5(d)的监测性能在检测中均得到改善,没有虚警并检测到所有故障。

  • 图4 故障10监测结果

  • Fig.4 Monitoring results for fault 10

  • 图5 故障12监测结果

  • Fig.5 Monitoring results for fault 12

  • 4 结论

  • 本文提出了一种基于KL散度的多模块滑动窗口慢特征分析方法,该方法使用KL散度算法度量每两个变量间的相似度,利用最小误差平方和准则对变量间的距离依次迭代,实现了对样本的无监督分块,并在每个子块中建立SFA监测模型,解决了单一模块的不稳定性问题,同时引入滑动窗口建立最优子模型,利用SVDD将子块监测结果融合。并将所提方法应用于TE过程,验证了该方法的有效性。

  • 参考文献

    • [1] LI Z C,YAN X F.Complex dynamic process monitoring method based on slow feature analysis model of multi-subspace partitioning[J].ISA Transactions,2019,95:68-81.

    • [2] HUANG J,ERSOY O K,YAN X F.Fault detection in dynamic plant-wide process by multi-block slow feature analysis and support vector data description[J].ISA Transactions,2019,85:119-128.

    • [3] JIANG Q C,GAO F R,YI H,et al.Multivariate statistical monitoring of key operation units of batch processes based on time-slice CCA[J].IEEE Transactions on Control Systems Technology,2019,27(3):1368-1375.

    • [4] LI S,ZHOU X F,PAN F C,et al.Correlated and weakly correlated fault detection based on variable division and ICA[J].Computers & Industrial Engineering,2017,112:320-335.

    • [5] LI R Y,RONG G.Fault isolation by partial dynamic principal component analysis in dynamic process[J].Chinese Journal of Chemical Engineering,2006,14(4):486-493.

    • [6] LEE J M,YOO C Y,LEE I B.Statistical monitoring of dynamic processes based on dynamic independent component analysis[J].Chemical Engineering Science,2004,59(14):2995-3006.

    • [7] WISKOTT L,SEJNOWSKI T J.Slow feature analysis:unsupervised learning of invariances[J].Neural Computation,2002,14(4):715-770.

    • [8] LI R F,WANG X Z.Dimension reduction of process dynamic trends using independent component analysis[J].Computers & Chemical Engineering,2002,26(3):467-473.

    • [9] SHANG C,HUANG B,YANG F,et al.Slow feature analysis for monitoring and diagnosis of control performance[J].Journal of Process Control,2016,39:21-34.

    • [10] MACGREGOR J F,JAECKLE C,KIPARISSIDES C,et al.Process monitoring and diagnosis by multiblock PLS methods[J].AIChE Journal,1994,40(5):826-838.

    • [11] GE Z Q,SONG Z H.Distributed PCA model for plant-wide process monitoring[J].Industrial & Engineering Chemistry Research,2013,52(5):1947-1957.

    • [12] TONG C D,SONG Y,YAN X F.Distributed statistical process monitoring based on four-subspace construction and Bayesian inference[J].Industrial & Engineering Chemistry Research,2013,52(29):9897-9907.

    • [13] CONTRERAS-REYES J E,ARELLANO-VALLE R B.Kullback-Leibler divergence measure for multivariate skew-normal distributions[J].Entropy,2012,14(9):1606-1626.

    • [14] YOUSSEF A,DELPHA C,DIALLO D.Enhancement of incipient fault detection and estimation using the multivariate Kullback-Leibler Divergence[C]//Proceedings of 24th European Signal Processing Conference(EUSIPCO),2016.

    • [15] WANG B,JIANG Q C,YAN X F.Fault detection and identification using a Kullback-Leibler divergence based multi-block principal component analysis and Bayesian inference[J].Korean Journal of Chemical Engineering,2014,31(6):930-943.

    • [16] 周伟,潘海鹏,吴平,等.基于DPCA和KL散度的微小故障检测方法[J].传感器与微系统,2020,39(3):135-138.ZHOU W,PAN H P,WU P,et al.Tiny fault detection method based on DPCA and KL divergence[J].Transducer and Microsystem Technologies,2020,39(3):135-138.(in Chinese)

    • [17] 柯亮,熊伟丽,徐保国.基于滑动窗PCA的微小故障检测[J].小型微型计算机系统,2016,37(6):1360-1364.KE L,XIONG W L,XU B G.Small fault detection based on moving windowPCA[J].Journal of Chinese Computer Systems,2016,37(6):1360-1364.(in Chinese)

    • [18] SEGHOUANE A K,AMARI S I.The AIC criterion and symmetrizing the Kullback-Leibler divergence[J].IEEE Transactions on Neural Networks,2007,18(1):97-106.

    • [19] DOWNS J J,VOGEL E F.A plant-wide industrial process control problem[J].Computers & Chemical Engineering,1993,17(3):245-255.

    • [20] YIN J,YAN X F.Mutual information-dynamic stacked sparse autoencoders for fault detection[J].Industrial & Engineering Chemistry Research,2019,58(47):21614-21624.

    • [21] LI Z C,YAN X F.Complex dynamic process monitoring method based on slow feature analysis model of multi-subspace partitioning[J].ISA Transactions,2019,95:68-81.

    • [22] LI Z C,YAN X F.Fault-relevant optimal ensemble ICA model for non-Gaussian process monitoring[J].IEEE Transactions on Control Systems Technology,2020,28(6):2581-2590.

图1 TE过程流程图

Fig.1 TE process flow chart

表1 TE过程故障类型表

Tab.1 TE process fault type table

图2 变量的分布

Fig.2 Distribution of sub-blocks

表2 分块结果

Tab.2 Block the results

表3 KPCA,DICA,SVDD和 KL-MWSFA的故障检测率

Tab.3 Fault detection rate of KPCA, DICA, SVDD and KL-MWSFA

图3 故障5监测结果

Fig.3 Monitoring results for fault 5

图4 故障10监测结果

Fig.4 Monitoring results for fault 10

图5 故障12监测结果

Fig.5 Monitoring results for fault 12

图表 1/1

  • 参考文献

    • [1] LI Z C,YAN X F.Complex dynamic process monitoring method based on slow feature analysis model of multi-subspace partitioning[J].ISA Transactions,2019,95:68-81.

    • [2] HUANG J,ERSOY O K,YAN X F.Fault detection in dynamic plant-wide process by multi-block slow feature analysis and support vector data description[J].ISA Transactions,2019,85:119-128.

    • [3] JIANG Q C,GAO F R,YI H,et al.Multivariate statistical monitoring of key operation units of batch processes based on time-slice CCA[J].IEEE Transactions on Control Systems Technology,2019,27(3):1368-1375.

    • [4] LI S,ZHOU X F,PAN F C,et al.Correlated and weakly correlated fault detection based on variable division and ICA[J].Computers & Industrial Engineering,2017,112:320-335.

    • [5] LI R Y,RONG G.Fault isolation by partial dynamic principal component analysis in dynamic process[J].Chinese Journal of Chemical Engineering,2006,14(4):486-493.

    • [6] LEE J M,YOO C Y,LEE I B.Statistical monitoring of dynamic processes based on dynamic independent component analysis[J].Chemical Engineering Science,2004,59(14):2995-3006.

    • [7] WISKOTT L,SEJNOWSKI T J.Slow feature analysis:unsupervised learning of invariances[J].Neural Computation,2002,14(4):715-770.

    • [8] LI R F,WANG X Z.Dimension reduction of process dynamic trends using independent component analysis[J].Computers & Chemical Engineering,2002,26(3):467-473.

    • [9] SHANG C,HUANG B,YANG F,et al.Slow feature analysis for monitoring and diagnosis of control performance[J].Journal of Process Control,2016,39:21-34.

    • [10] MACGREGOR J F,JAECKLE C,KIPARISSIDES C,et al.Process monitoring and diagnosis by multiblock PLS methods[J].AIChE Journal,1994,40(5):826-838.

    • [11] GE Z Q,SONG Z H.Distributed PCA model for plant-wide process monitoring[J].Industrial & Engineering Chemistry Research,2013,52(5):1947-1957.

    • [12] TONG C D,SONG Y,YAN X F.Distributed statistical process monitoring based on four-subspace construction and Bayesian inference[J].Industrial & Engineering Chemistry Research,2013,52(29):9897-9907.

    • [13] CONTRERAS-REYES J E,ARELLANO-VALLE R B.Kullback-Leibler divergence measure for multivariate skew-normal distributions[J].Entropy,2012,14(9):1606-1626.

    • [14] YOUSSEF A,DELPHA C,DIALLO D.Enhancement of incipient fault detection and estimation using the multivariate Kullback-Leibler Divergence[C]//Proceedings of 24th European Signal Processing Conference(EUSIPCO),2016.

    • [15] WANG B,JIANG Q C,YAN X F.Fault detection and identification using a Kullback-Leibler divergence based multi-block principal component analysis and Bayesian inference[J].Korean Journal of Chemical Engineering,2014,31(6):930-943.

    • [16] 周伟,潘海鹏,吴平,等.基于DPCA和KL散度的微小故障检测方法[J].传感器与微系统,2020,39(3):135-138.ZHOU W,PAN H P,WU P,et al.Tiny fault detection method based on DPCA and KL divergence[J].Transducer and Microsystem Technologies,2020,39(3):135-138.(in Chinese)

    • [17] 柯亮,熊伟丽,徐保国.基于滑动窗PCA的微小故障检测[J].小型微型计算机系统,2016,37(6):1360-1364.KE L,XIONG W L,XU B G.Small fault detection based on moving windowPCA[J].Journal of Chinese Computer Systems,2016,37(6):1360-1364.(in Chinese)

    • [18] SEGHOUANE A K,AMARI S I.The AIC criterion and symmetrizing the Kullback-Leibler divergence[J].IEEE Transactions on Neural Networks,2007,18(1):97-106.

    • [19] DOWNS J J,VOGEL E F.A plant-wide industrial process control problem[J].Computers & Chemical Engineering,1993,17(3):245-255.

    • [20] YIN J,YAN X F.Mutual information-dynamic stacked sparse autoencoders for fault detection[J].Industrial & Engineering Chemistry Research,2019,58(47):21614-21624.

    • [21] LI Z C,YAN X F.Complex dynamic process monitoring method based on slow feature analysis model of multi-subspace partitioning[J].ISA Transactions,2019,95:68-81.

    • [22] LI Z C,YAN X F.Fault-relevant optimal ensemble ICA model for non-Gaussian process monitoring[J].IEEE Transactions on Control Systems Technology,2020,28(6):2581-2590.