# Model Predictive Controller Monitoring Based on Pattern Classiﬁcation and PCA Fred Loquasto III∗ Dale E.

Seborg†

Department of Chemical Engineering, University of California, Santa Barbara, CA 93106

Abstract

A pattern classiﬁcation-based methodology is presented as a practical tool for monitoring Model Predictive Control (MPC) systems. Principal component analysis (PCA) is used, especially PCA and distance similarity factors, to classify a window of current, MPC operating data into one of several classes. Pattern classiﬁers are developed using a comprehensive, simulated database of closed-loop MPC system behavior that includes a wide variety of disturbances and/or plant changes. The pattern classiﬁers can then be employed to classify current MPC performance by determining if the behavior is normal or abnormal, if an unusual plant disturbance is present, or if a signiﬁcant plant change has occurred. The methodology is successfully applied in an extensive case study for the Wood-Berry distillation column model.

2

PCA and Dynamic PCA Methodology

1

Introduction

Principal component analysis is a multivariate statistical technique that has been widely used for both academic research and industrial applications of process monitoring. Its ability to create low-order, data-driven models by accounting for the collinearity of the process data, its modest computational requirements, and its sound theoretical basis make PCA a highly desirable technique upon which to base tools for monitoring processes. Traditional PCA monitoring techniques use the Q and T 2 statistics [11,12] to determine how well a single sample agrees with the PCA model. The monitoring strategy proposed in this paper is based on a different approach; it uses several PCA-based similarity factors [13,14] to compare current operating data with a simulated, closed-loop database in order to classify the current operating data. The dataset is a matrix X with dimensions (n × m), where m is the number of measured variables and n is the number of samples for each variable.

Model predictive control is widely used in the petrochemical industries to control complex processes that have operating constraints on the input and output variables. The MPC controller uses a process model and a constrained, on-line optimization to determine the optimal future control move sequence. The ﬁrst control move is implemented and the calculations are then repeated at the next control calculation interval, the so-called receding horizon approach. Excellent overviews of MPC and comparisons of commercial MPC controllers are available [1,2]. Although MPC control has been widely applied for over 25 years, the problem of monitoring MPC system performance has received relatively little attention until very recently [3–9]. The objective of this research is to develop a MPC monitoring technique that will help plant personnel to answer the following questions: (1) Is the MPC system operating normally? (2) If not, is its poor performance due to an abnormal disturbance or an inaccurate process model (for the current conditions)? The proposed MPC monitoring technique is based on a pattern classiﬁcation approach. This approach was selected because it is desired to be able to identify plant changes, in addition to disturbances, without performing a full model re-identiﬁcation that would require signiﬁcant process excitation. Thus, identifying plant changes in this context is an extremely difﬁcult task. In a previous paper [10], a MPC monitoring strategy was developed using multi-layer perceptron neural networks as the pattern classiﬁers. In this paper, the classiﬁcation is instead based on a novel application of principal component analysis, especially PCA similarity factors and distance similarity factors. The proposed MPC monitoring technique is evaluated in a simulation case study for the Wood-Berry distillation column model.

∗ E-mail: † E-mail:

2.1

PCA models

In PCA a set of uncorrelated variables, the principal components, are calculated from linear combinations of the original (correlated) variables. The principal components are the eigenvectors of the covariance matrix of the data. They correspond to the directions of the data that possess the highest degree of variability [11,15]. For example, PCA models can be calculated using the pcacov.m function in the Statistics Toolbox in MATLAB [16]. For dynamic PCA (DPCA), the X data matrix is augmented with previous “lagged” data [17,12]. Thus, the process dynamics are accounted for by effectively constructing an auto-regressive with exogenous inputs (ARX) model of the process from this data. ˜ If l lags are considered, the DPCA matrix X(l) has dimension (n − l × ml + m): ˜ X(l) = [X(k), X(k − 1), · · · , X(k − l))] (1)

loquasto@engineering.ucsb.edu seborg@engineering.ucsb.edu, Corresponding au-

thor

˜ Note that the number of rows of X(l) is reduced by the number of lags, l, compared to the original matrix X. There are various methods available in the literature for choosing k, the number of principal components (PCs) for the PCA model, as well as the number of lags l for the DPCA model. In this paper, k is chosen by two methods: (i) specifying a threshold for the cumulative explained variance or (ii) using parallel analysis [15]. For the ﬁrst approach, k is selected to be the minimum number of principal components whose cumulative explained variance exceeds a speciﬁed threshold. The parallel analysis (PA) technique compares the eigenvalues λ of Σ, the covariance matrix of X, with the eigenvalues µ from the covariance matrix of a similarly-sized data matrix with independent, normally distributed random elements [15]. Usually, the λ values begin with larger magnitudes than the µ values, and the number of PCs is chosen to be the smallest value of k for which λ(k) ≤ µ(k). However, in some cases, λ(1) ≤ µ(1), and k is selected as k = 1.

Proceedings of the American Control Conference Denver, Colorado June 4-6, 2003

0-7803-7896-2/03/$17.00 ©2003 IEEE

1968

the two classes for the AB classiﬁer are “Normal” and “Abnormal”. Colorado June 4-6. Denote the vectors of sample ¯ ¯ means as xC and xT . A candidate pool of size Np is constructed by selecting the Np training datasets that are most similar to the current dataset. An abnormal disturbance is present (D) 3. The training datasets that are most similar to the current dataset are collected into a candidate pool. based on the SF metric. The X data is comprised of the inputs. and C to be the (n × k) PCA model for the current dataset of interest. These same values are then used to scale dataset C. Deﬁne θij as the angle between the ith principal component of T and the j th principal component of C. It is proposed to classify each dataset as being a member of one of four mutually exclusive operating classes: 1. SP CA the PCA similarity factor value for the CORR data X 3. An important aspect of the classiﬁcation is how the different operating classes are deﬁned. Binary classiﬁers Three binary classiﬁers are used. Both a plant change and a disturbance are present (P+D) We use the term “fault” to refer to a member of the D.
Proceedings of the American Control Conference Denver. P. SP CA the PCA similarity factor value for the X data CORR 2.18] SPCA = 1 k
k k
3
Classiﬁcation Strategy
cos2 θij =
i=1 j=1
CT T T T C k
(2)
Although the PCA similarity factor compares the similarity of two datasets. and based on an analysis of the training datasets in the pool. 2003
For the purposes of classifying MPC operation.12]. control error. In general. SPCA . The performance of the individual binary classiﬁers can be improved by using a combination of them in an exclusion strategy. A PCA model is deﬁned to be the matrix that has the ﬁrst k principal component vectors as its columns.2
PCA and Distance Similarity Factors
The PCA similarity factor. Four-class classiﬁers This classiﬁer classiﬁes a dataset as being one of the four exclusive operating classes (#1–4). The distance similarity factor. SP CA . This measure can be obtained by using the distance similarity factor. the variables in the T dataset are scaled to zero mean and unit variance. P. where the classes are deﬁned based on whether the binary classiﬁers are or the four-class classiﬁer is used.14. it is also beneﬁcial to have a similarity measure based on the “distance” between two datasets in the m-dimensional data space. Sdist .21] (the “CORR” data). none (N or D) II. The CORR data is deﬁned to be the ﬁrst n/4 correlation coefﬁcients (as per the recommendations in [20]) of the residuals. proposed in [14]. provides a useful characterization of the degree of similarity for two datasets. Abnormal (D. It is based on the similarity of the directions of the principal component vectors for the two corresponding PCA models. Each one classiﬁes a dataset as being in one of two categories: AB: Normal (N) vs. For each classiﬁer. The inputs were differenced to help increase the similarity measure’s sensitivity to higher-frequency dynamic behavior. the classiﬁcation label for the dataset is determined by using one of two alternative techniques.In order to specify the number of lags l for DPCA. different methods have been presented in the literature [17. above. our experience has been that the AB and DIST classiﬁers perform very well while the MPM classiﬁer is less successful. For k principal components. The second type of data consists of sample autocorrelation (ACF) and partial autocorrelation (PACF) function coefﬁcients [20. Sdist the distance similarity factor for the X data CORR 4. or MPM (P) 4. The ﬁrst type is the standard. while reducing the sensitivity to low-frequency disturbances. thus there is signiﬁcant modelplant mismatch.
X CORR X CORR α1 SP CA + α2 SP CA + α3 Sdist + α4 Sdist
(5)
2. the current dataset is classiﬁed. a combination of four individual similarity factors:
X 1. time-series measurement data (the “X” data). is then deﬁned as [13. With these deﬁnitions. in order to take advantage of the model used in the controller. First. the PCA similarity factor. Then. none (N or P) MPM: Plant change (P or P+D) vs. The class with the highest frequency is selected as the classiﬁcation label for the current dataset. SF . A plant change has occurred. The overall classiﬁcation of MPC performance is based on a composite similarity factor. the ﬁrst option involves calculating class-average similarity values by ﬁnding the average of the SF values for each class present in the pool. The method used here is based on the method presented in [17]. or P+D)
where Σ† is the pseudo-inverse of the covariance matrix Σ of the T dataset. the classiﬁcation can be performed using two types of classiﬁers: I. The second option is based on the frequency (or number of times) that each class is present in the candidate pool. similarity factors were calculated for two types of data. and differenced inputs. Deﬁne T to be the (n × k) PCA model for a training dataset. Sdist is deﬁned to be [14]: Sdist = 2 π
∞ Φ
e
−z 2 /2
dz
(4)
DIST: Abnormal disturbance (D or P+D) vs. or P+D classes. the basis for the proposed PCA approach is to use the composite similarity factor SF to determine the similarity between a current dataset and a group of training datasets that contain a wide variety of closed-loop process responses. The class in the candidate pool with the highest class-average SF value is selected as the classiﬁcation for the current dataset. Normal operation (N) 2. Sdist the distance similarity factor for the CORR data
1969
. For example. and one-step ahead residuals. outputs. The Mahalanobis distance is [19] Φ= ¯ ¯ (¯ T − xC )T Σ† (¯ T − xC ) x T x (3)
In summary.
The composite similarity factor is deﬁned by the following equation SF where α1 + α2 + α3 + α4 = 1 (6) The αi values are tuning parameters for the classiﬁer that can be used to weight the individual similarity factors differently.

Also. the classiﬁcation strategy presented in this paper is identical for both constrained and unconstrained MPC. The manipulated variables (MV’s) were R and S. N.7s + 1 21.) Input and output disturbances were chosen randomly and independently from four types of disturbances: step.9e−3s −19.0s + 1 + 14.4s + 1 13. (It was used to generate the test database.However. and thus was not used in the generation of either the training or validation databases. The optimal values of these design parameters were determined by classifying a second independent database containing validation data. The column model is shown in Eq. The unmeasured feed ﬂow rate. sinusoidal. upon comparing the case study results. or P+D). The validation and test databases represent the “current data”. An important aspect of the proposed approach is that it does not require that a specially-designed perturbation signal be applied to the plant.4e−3s S xB 10. (7).25 − 3. Each dataset contained 100 one-minute samples of setpoint response data.8e−s −18. For example. the ﬁnal magnitude of the ramp disturbance. The prediction horizon was P = 30 and the control horizon was M = 5. Results are then presented for both the PCA and DPCAbased classiﬁers. better classiﬁcation occurs if each dataset contains the same known excitation (such as a setpoint change or measured disturbance) as in the current dataset. the number of lags for DPCA (l). constant step-disturbance approach was employed [23. in the case study. as described below. validation. It should be noted that the step and ramp output disturbances could also be used to represent sensor biases and drifts. F . The feed ﬂow rate disturbance model in Eq. (7) was assumed to be unknown.9s + 1 F 6.2
Database Generation for the Case Study
Three types of databases are generated for different purposes: training.
4. “false alarms”). and the SF weights (αi ’s). In order to reduce the simulation time needed to explore a variety of design options and design parameters.2s + 1 (7) MPC tuning parameters. The magnitude of the step disturbance.1
Wood-Berry distillation column model
The Wood-Berry model is a 2 × 2 transfer function model of a pilotplant distillation column that separates methanol and water [22].8e−8s xD R = 16. both the training data and the “current” data are assumed to be setpoint response data for a single setpoint change. the pool size (Np ). The disadvantage of this technique is that a P+D condition will be classiﬁed as a ‘Disturbance’ if perfect classiﬁers are used. The training and validation data were created using input disturbances (added to u) and output disturbances (added to y) in order to mimic a wide range of actual process operation when unknown disturbances are present.9e−3s 3.516lb/min. P.
publications and in the MATLAB MPC Toolbox [23]. it must be scaled appropriately. which are controlled by the reﬂux and steam ﬂow rates. The MPC controller was tuned by trial-and-error to give a reasonably fast setpoint response. The error weighting and move suppression matrices. and stochastic. acts as a process disturbance. 1.833 lb/min and −0. High and low saturation limits were imposed on the inputs. the current dataset is classiﬁed as a ‘Plant Change’ condition.e. D. unconstrained MPC was employed. The ramp disturbances began between t = 25 and t = 50 min and had a randomly chosen duration between 25 and 50 min. if the AB classiﬁer indicates an ‘Abnormal’ condition. The magnitudes of the input disturbances
Proceedings of the American Control Conference Denver. xD and xB [wt %]. each dataset in the training database was scaled to zero mean and unit variance based on the overall mean and standard deviation of the class to which that dataset belonged (i. Colorado June 4-6. It is suggested that the length of the dataset is approximately the longest open-loop settling time of the process so that all important dynamics can be represented.6e−7s 4. and the controlled variables (CV’s) were xD and xB . the common “DMC”. For this paper. The choice of the variable whose setpoint was stepped was based on the fact that xD is the faster responding of the two controlled variables.516 ≤ S ≤ +0.1].
3. After the simulated data has been generated. it is apparent that the beneﬁt of this approach is that it signiﬁcantly reduces the amount of false ‘P’ classiﬁcations (i. the AB and DIST classiﬁers can also be used to detect plant changes with a low false alarm rate when used as part of the exclusion strategy. The databases are made up of individual datasets. 2003
4
MPC Case study
This section presents speciﬁc details about the simulated case study. In this paper. The control execution period was ∆t = 1 min. one could argue that this result is acceptable because it is partially correct. A unit xD setpoint change was made at t = 0. Q and R. The test database takes the place of actual process data for the purpose of evaluating the proposed monitoring approach. the Wood-Berry distillation column model and its MPC system.
4.1
Generation of the Training Database
The ﬁrst step in developing these pattern classiﬁers is to create the simulated databases. However. However. and the DIST classiﬁer indicates a ‘No Disturbance’ condition. The closed-loop MPC simulation was performed using MATLAB and the Model Predictive Control Toolbox [23].9 + 1 14.25} wt%. The generation of the simulated databases and the classiﬁers is also described. were chosen to be the identity matrix. The system outputs are the distillate and bottoms compositions.833 ≤ R ≤ +0. 12. The purpose of the training database was discussed earlier. Training and validation databases. The Wood-Berry model is a classical example used in many previous
1970
. where each dataset consists of closed-loop response data. The saturation limits on the inputs were −0. No special disturbance modeling was used. Step disturbances had random starting times between t = 25 and t = 75 min. This strategy is shown in Fig. ramp. R and S [lb/min]. and the amplitude of the sinusoidal oscillations were chosen randomly from the range ±{0. These databases consist of many individual datasets. rather than as ‘P+D’. and test.e. However. Then several design and tuning parameters must be speciﬁed to design the classiﬁers: the number of principal components in the PCA models (k). “class scaling” was performed. The validation data is used to obtain the optimal design and tuning parameters for the classiﬁer.

3
Case Study Results
This section presents classiﬁcation results for both the PCA and DPCA techniques.5}% of the nominal feed ﬂow rate (2. only the frequency-based classiﬁer results are presented. P. To evaluate the effect of using less training data for the PCA technique. choose how many model parameters (i) will be perturbed. steps. for each plant change dataset. 2003
1971
. 4. P. For the feed disturbances. and sinusoids were chosen randomly in the range ±{12. For the binary Type I and Type II errors. the percentage of all test cases in which an error was caused by falsely identifying a non-fault condition as a fault. P+D). an individual dataset could have contained as many as 12 parameter changes that were chosen randomly. we assume no a priori knowledge of the likelihood
of any of the fault conditions (D. For example. an average of the multiple results is presented. Note that for the binary classiﬁers. P+D) present in the training database. or P+D) and have the same number of datasets for each of the four operating conditions for all three databases.003 lb/min.03 − 1. or (ii) either the P or P+D class for the p(P. T2 = Type II errors. For the case study. is the correct classiﬁcation. τ . respectively.05 wt% was added to each output. validation. To analyze the classiﬁer performance. we consider the distribution of the four independent operating classes (N. Colorado June 4-6. To create the plant change datasets. 3. sinusoids. smaller training databases with N =500 or 200 datasets were created by randomly choosing a subset of the original 2000 datasets. the percentage of all test datasets in which an error was caused by missing a fault condition. the percentage of all cases in which the indicated class is not correct. In these cases. where K0 is the nominal value. T2 = 100% − T1 . (7) contain a total of 12 model parameters (K.5% − 37. . Determine the i speciﬁc model parameters (K. Test database. and a third transfer function was used for the validation data. and the subsequent output magnitude were scaled to give the same approximate effect on the output as the deterministic disturbances. The number of validation and training datasets. 1000 and 800.006 lb/min and 0. But in this case study. Then. ramps. The four transfer functions in the Wood-Berry model of Eq. or P+D}
DIST µ0 : True Class ∈ {D or P+D} MPM µ0 : True Class ∈ {P or P+D} Four-class classiﬁers: T1 = 1 4 T1j . PCA Results. These changes occur at the beginning of the simulated setpoint change and remain in effect for the entire window. Note that in some cases. Gaussian measurement noise with an approximate standard deviation of 0. For each perturbed model. The magnitudes for the steps. P+D) metric. P. . a “false positive”. Thus. θ) to perturb for each model. for j = 1. P+D). the null hypotheses are deﬁned as follows: AB µ0 : True Class ∈ {D.
4. Random plant changes were included in the simulated database in the same manner for the training.45 lb/min).were adjusted so that the open-loop changes in the outputs were approximately the same as for the output disturbance cases. the number of datasets in which j process models are to be perturbed. i. The classiﬁcation results on the validation database (not shown) indicated that using PCA models explaining 85% of the variance leads to the best performance for
Proceedings of the American Control Conference Denver. and stochastic sequences were used as inputs to the disturbance model. 100% − p(P. accuracy of the data sets classiﬁed as a plant change. ramps. Next. a dataset classiﬁed to be a plant change must belong to either (i) the P class for the p(P only) metric. P. the unmeasured feed ﬂow rate disturbance model (that was ignored in creating the training and validation data) was utilized to generate disturbances.5 − 37. where T1j represents the percentage of j=1 4 all cases in which a class j. In order to make the test case more realistic. For an accurate classiﬁcation. and Gaussian process noise was added to the two inputs with approximate standard deviations of 0. Plant changes. Specify which j speciﬁc models will be perturbed. the magnitude of the additive parameter perturbation.
False Alarm Rate Binary classiﬁers:
T1 = Type I errors. η + T1 + T2 = 100%. D. . and θ for each transfer function). were chosen arbitrarily to be less than the number of training datasets. The total amount of training data. was based on a neural network design criteria for a previous study [10]. the following parameters were chosen in a uniformly random manner: 1.03} rad/min. as indicated by the classiﬁer. the following performance measures (in %) are deﬁned for the test datasets: η ηi p efﬁciency = percentage of all datasets correctly classiﬁed. τ . Operating class distribution.5%}. this was based on the number of possible combinations in which j models could be chosen out of the four models. For the sinusoidal oscillations.e. . where ∆ = ±{12. Two ARMA transfer functions were used in creating the stochastic sequences in the training data. These values were speciﬁed so that the process noise had approximately the same open-loop magnitude effect on the outputs as the measurement noise. must be speciﬁed. D. Because the classiﬁers based on the frequencypool method almost always performed better than the pool-average method. the magnitude of the input sequence to the transfer functions. For “stochastic” input and output disturbances. 2. Specify ∆. Increasing the relative amount of a particular operating class can sometimes result in better performance for one classiﬁer at the potential expense of another. multiple pool sizes or weight values may give the same classiﬁcation accuracy. N = 2000 datasets. 4. the frequency had a uniformly random distribution in the range of {0. percent of all class i datasets classiﬁed correctly (i = N. K = K0 (1 + ∆). A different ARMA transfer function was used to generate the stochastic sequences for the test database. and test databases.

The best validation classiﬁers for the full data (85% variance explained) are underlined. It appears that classiﬁer performance tends to deteriorate when less training data is used.
Proceedings of the American Control Conference Denver. Parallel analysis design achieves very high p values of p(P only) = 94. L. at the expense of a lower η value.” in Proc. (in press). the performance of all classiﬁers was still reasonable for reduced amounts of training data. Choosing the number of PC’s to retain by parallel analysis gave mixed results but is still recommended because it is an automatic technique. because the goal of this strategy is to obtain high p values. Vol. an abnormal operating condition was detected with over 98% accuracy.” Control Eng. However. the best overall test database classiﬁers are shown in boldface. Now. while the disturbance classiﬁer achieved over 92% correct. Shah. l. by far. Control-VI. Germany). MPM) performed worse for DPCA. The best overall results (based on the test database) are shown in boldface. if one “cheats” and looks at the test database results. Use of this classiﬁer thus results in a very low false alarm rate for identifying plant changes of less than 3%. 47. 2001. For example. [4] B. the ﬁrst author would like to thank the Control Engineering Laboratory at the Helsinki University of Technology for providing him with a visiting researcher position. based on the validation database results (not shown). The exclusion strategy of Section 3 can be applied in order to detect plant changes more accurately (with less false positives). the AB classiﬁer. 15th IFAC Triennial World Congress.6%. it seems that one should choose the PCA method over DPCA for best performance. both 85% and 95% thresholds were acceptable. 2002.
Acknowledgements
Funding for the research presented in this paper. M. AIChE Symposium Series No.” in Proc. 2003
1972
. Test database results using the best validation classiﬁers (i. a smaller amount of training data had a detrimental effect on the classiﬁcation results. which was one of the indicated best validation cases. for the full training database.. the αi . P+D) = 97. When the best validation classiﬁers were used to evaluate the test database. The authors would like to thank Ron Sorensen (CTRTC) and Jim Gunderman (ChevronTexaco) for discussions about industrial MPC systems. while the 4-class classiﬁer was slightly better. Comparing the test database performance of the optimal validation classiﬁers that used DPCA (Table 5) with the PCA-based classiﬁers (Table 1). 190 – 207. and challenges.the full training data (N =2000) case. 2002. used to detect abnormal operating conditions. Table 1 presents the classiﬁcation performance for the test database. Test database results for the exclusion strategy are presented in Table 7.” AIChE J. Although the 75% explained variance cases achieved the best p values for the validation database. Proc. [5] D.” Chem. Again. [2] S. and k values giving the best performance) are shown in Table 5. 98. vol. indicates that the binary classiﬁers (AB. 2002. 2001. applications. where many of the ideas in this paper were initially developed. A comparison of the PCA and DPCA validation database classiﬁcation performance indicates that the DIST and 4-class classiﬁers performed slightly better using DPCA than the PCA classiﬁers discussed earlier. Proc. the p values are signiﬁcantly lower for DPCA. the MPC-controlled Wood-Berry distillation column model. 2002. 6. [6] T.5% and p(P. AIChE Symposium Series No. Huang. is greatly appreciated. over 77% of the plant changes in the test database were identiﬁed as a plant change with a false alarm rate of less than 3%. In general. but the MPM classiﬁer has a quite large Type II error value. the ones in boldface). Control-VI. Patwardhan. Edgar. Prac. The classiﬁcation accuracy for each type of operating condition and the Type I and II errors are shown. provided by the ChevronTexaco Research and Technology Company (CTRTC). The classiﬁers performed slightly better when they were based on PCA as opposed to DPCA. it is apparent that. Table 3 summarizes these results for the test database. 326. “Controller performance monitoring and diagnosis. The performance is reasonable. Seppala. Colorado June 4-6. Harlow. pp. (Karlsruhe. could accurately classify over 98% of the independent test datasets. 1999. “A survey of industrial model predictive control technology.
5
Summary
Several PCA-based pattern classiﬁers have been developed for monitoring model predictive control systems. “Recent developments in controller performance monitoring and assessment techniques. pp. “Performance assessment of constrained model predictive control systems. A. Analysis results for the DPCA classiﬁers are now presented. DIST.e. J.. (Barcelona. F. Harris and C. Badgwell. The highest test database p values for the DPCA case occur for 95% variance explained. European Control Conf. Predictive Control with Constraints. Once again. and B. the best p values occurred for parallel analysis.” Chem. J. the choice of which design to select for this task is not as clear cut as for the PCA case. no. 208 – 222. [3] Y. the best performing classiﬁers for the validation database (not shown) were obtained based on 85% variance explained.
References
[1] J. the AB and MPM results were worse. with both at a threshold of 85%. an important consideration in industrial applications. plant change classiﬁer. “A performance measure for constrained model predictive controllers. These PCA-based classiﬁers gave accurate classiﬁcations for a 2 × 2 simulated case study. Lastly. 98. thus. Henson. while the four-class classiﬁer achieved an overall correct classiﬁcation rate of 64%. and it produced reasonably accurate AB and DIST classiﬁers along with a very accurate exclusion-based. A. Kozub.-S. where the best validation classiﬁers are underlined. Table 6 shows detailed results for the best overall test database classiﬁers (those in boldface in Table 5). DPCA Results. 326. Zhang and M. pp. Using the best exclusion classiﬁers. an industrial perspective. T. Qin and T. However. Although the η values for DPCA in Table 7 are similar to the PCA results in Table 3.e. R. Ko and T. Vol. Maciejowski. Table 2 presents more detailed test database results for the best test database classiﬁers from Table 1 (i. England: Prentice Hall. [7] S. “Multivariate controller performance analysis: Methods. J. 1363 – 1371. The plant change classiﬁer detected plant changes correctly 73% of the time. Spain). Table 4 presents a class distribution table that shows how the true classes of the test database are classiﬁed when using the parallel analysis exclusion classiﬁers.

8 4.6 94. vol.” Chem.4 100.0
AB DIST MPM 4-class
Table 3. 2001. New York: John Wiley. Sys.5 97. III and D. Assoc. % Var.0 92.5 88. Res.” J. London: Springer. G. Chiang.0 100.8 73.0 77.9 58. Ku..5 81. 41.5 84. vol.” Chemom. Control.6 74. E. pp.8 95. Morari and N. 2003
1973
. Model Predictive Control Toolbox User’s Guide.3 97. Singhal and D. E. 2002.8 78.6 DIST 91. [10] F. Sci.” Chem.0 73.0 62. C. E.1 0.0 71.
DPCA exclusion strategy plant change classiﬁcation results. η p (P only) p (P. Stork. vol.
Proceedings of the American Control Conference Denver. A. and D.. 1979..0 0. Statistics Toolbox Ver. 1973.0 87.3 79.5 76. D. [9] B.3 2. 1991. [17] W.6 75.1 93.0 74. 3822 – 3838. 1995.5 37. Tamayo. Braatz. P+D) 75 82. A User’s Guide to Principal Components. Englewood Cliffs.0 95. E. and G.0 2. 55.8 MPM 68. pp.7 98.5 97. Eng. 2022 – 2038..0 0.5 74. and D. Jenkins.0 5. 2000.1 63. E.0 67.5 94. H. pp. Chem. L.6 61. Gallagher.4 59. [19] R.0 26.5 93.5
Disturbance Classifier (DIST)
No Disturbance Disturbance
NOT CLASSIFIED CLASSIFY AS: Normal
Table 1.0 95. 3.8 92.. E.5 T1 T2 2. [11] B.6
200 97. P. Wood and M.2 64. (Reno. [14] A.0 94. NY: John Wiley.0 68.8 77. Natick. “Multivariable model validation in the presence of time-variant disturbance dynamics. M. 30. “Model valiation for industrial model predictive control systems.5 94. 2000. [16] The MathWorks. L.” Chem. [23] M. and R.0 97.8 63.0 0.5 P 3. Variance Explained 75% 85% 95% PA N= 2000 500 200 2000 500 200 2000 500 200 2000 500 AB 98.1 97. M. Seborg.0 87. 1995. Huang and E.5 81. Duda. 74.5 0. 3rd ed.” J. H.5 0. Stat. Ricker.6 73.6 Table 4. O. 77.6 72.8 4. B.9 59. 4583 – 4595. Inc. [12] L.0 90.0.3 68. Wise and N.. “Monitoring model predictive control systems using a novel neural network approach.9 DIST 90.8 98.. 1997. 6.5 98. 2001.0 90. J.3 20.3 55. P+D) 75% 82. Sci.9 70.3 97. Luce˜ o.3 84.0 76. 48. vol. Box and A.8 100. 2nd ed.0 Table 5. C. η p (P only) p (P. “Terminal composition control of a binary distillation column.8 MPM 74. R.8 85% 83.1 72. PCA exclusion strategy plant change classiﬁcation results.1 96. Anal.3 PA 76.0 68. Seborg. MA: The MathWorks.8 96.4 89. “Between-groups comparison of principal components.6 65. Int. New York: John Wiley. vol. vol. W.3 4-class 68.8 62. pp. Pattern Classiﬁcation.8 96.4 88. [21] G. “Pattern matching in historical data. Lab. Seborg. pp. P. Jackson. 179 – 196. [18] M.6 85 80. 329 – 348.4 89. 703 – 707.5 19. Berry. 2315 – 2327. Storer. 1994. Hart.0 84.1 98.5 72. pp.[8] B. vol. % Var. True and Indicated classiﬁcations using the PCA exclusion classiﬁers (parallel analysis case).1 85.1
Normal Normal / Abnormal Classifier (AB) Abnormal
CLASSIFY AS: Plant Change
CLASSIFY AS: Disturbance
Figure 1.5 0.6 58.0 22. Proc. vol. Singhal. Box. 28. Huang.0 93. Expl. NV). Loquasto.4 95 82.0 97. 2001.5 2.6 31. Expl.3 76.0 29.4
Table 6. Am. Krzanowski. 55. & Eng. 1996.8 7. DPCA test results: η values for different variance thresholds.3 19. Sci.5 92.4 84.3 93.0 99.5 64.0 72.0 82.5 73. and C.5 68.
Table 2.0 T1 T2 1.0 88.1 92. PCA results: η values for different training sizes and variance thresholds.1 65.4 5.0 98.0 60.0 99. Colorado June 4-6.4 74.” AIChE J. Eng. E.0 19.5 73. Fault Detection and Diagnosis in Industrial Systems. Variance Explained 75% 85% 95% PA AB 96.0 P+D 0.6 88.0 86..0 97. [22] R. Detailed results for the best test database PCA classiﬁers. C. NJ: Prentice Hall. Johannesmeyer.” Ind.4 95% 84.5 67. pp. η ηN ηD ηP ηP +D 98. P.7 99. η ηN ηD ηP ηP +D 96. Classiﬁed as (%) Actual Dataset N D P Not Classiﬁed N 98. Russell.5 Par.” in AIChE Annual Meeting. “Disturbance detection and isolation by dynamic principal component analysis. 2002. K.0 65. E.8 71.3 94. [13] W. Reinsel. pp.0 80.0 100. 2000.4
AB DIST MPM 4-class Table 7. Eng.0 D 0.5 100.3 0. Detailed results for the best test database DPCA classiﬁers.7 86. 1707–1717.5 100.0 4-class 61.5 98. [20] G.9 90. Illustration of the exclusion strategy. “The process chemometrics approach to process monitoring and fault detection. [15] J. Time Series Analysis Forecasting and Control.. G. Georgakis. “Pattern matching in multivariate time series databases using a moving-window approach. Statistical Control by Monitoring n and Feedback Adjustment. Inc.