System and method for extracting principal time series data

Abstract

A method for extracting a set of principal time series data of dynamic latent variables. The method includes detecting, by a plurality of sensors, dynamic samples of data each corresponding to one of a plurality of original variables. The method also includes analyzing, using a controller, the dynamic samples of data to determine a plurality of latent variables that represent variation in the dynamic samples of data. The method also includes selecting, by the controller, at least one inner latent variable that corresponds to at least one of the plurality of original variables. The method also includes estimating an estimated current value of the at least one inner latent variable based on previous values of the at least one inner latent variable.

Claims

1. A method for extracting a set of principal time series data of dynamic latent variables comprising: detecting, by a plurality of sensors, dynamic samples of data each corresponding to one of a plurality of original variables; analyzing, using a controller, the dynamic samples of data to determine a plurality of latent variables that represent variation in the dynamic samples of data; selecting, by the controller, at least one inner latent variable that corresponds to at least one of the plurality of original variables; and estimating an estimated current value of the at least one inner latent variable based on previous values of the at least one inner latent variable, wherein selecting the at least one inner latent variable includes selecting the at least one inner latent variable based on a desire for a relatively large correlation between the estimated current value of the at least one inner latent variable at a selected time and an actual current value of the at least one inner latent variable at the selected time.

2. The method of claim 1 wherein analyzing the dynamic samples of data is performed using at least one of dynamic-inner principal component analysis (DiPCA) or dynamic-inner canonical-correlation analysis (DiCCA).

3. The method of claim 1 wherein: selecting the at least one inner latent variable of a first set of original variables includes selecting the at least one inner latent variable from the first set of original variables based on a desire for a relatively large correlation between the estimated current value of the at least one inner latent variable at a selected time and an actual current value of the at least one inner latent variable at the selected time; and the estimated current value of the at least one inner latent variable at the selected time is estimated from at least one of past values of the at least one inner latent variable or past values of at least another inner latent variable of another set of original variables.

4. The method of claim 1 wherein the at least one inner latent variable corresponds to a dynamic relationship between at least some of the dynamic samples of data.

5. The method of claim 1 wherein the dynamic samples of data correspond to at least one of process data or equipment data used in a data-driven process.

6. The method of claim 1 wherein the dynamic samples of data contain oscillating dynamics.

7. The method of claim 1 wherein the method is utilized to perform Granger causality analysis.

8. The method of claim 1 wherein the dynamic samples of data are inner latent dynamic samples.

9. The method of claim 1 further comprising outputting, by an output device, output data corresponding to the estimated current value of the at least one inner latent variable.

10. A method for extracting a set of principal time series data of dynamic latent variables comprising: receiving, by a controller, dynamic samples of data each corresponding to one of a plurality of original variables; analyzing, using the controller, the dynamic samples of data to determine a plurality of latent variables that represent variation in the dynamic samples of data; selecting, by the controller, at least one inner latent variable that corresponds to at least one of the plurality of original variables based on a desire for: a relatively large correlation between the estimated current value of the at least one inner latent variable at the selected time and the actual current value of the at least one inner latent variable at the selected time; and estimating, by the controller, the estimated current value of the at least one inner latent variable based on previous values of the at least one inner latent variable.

11. The method of claim 10 wherein analyzing the dynamic samples of data is performed using at least one of dynamic-inner principal component analysis (DiPCA) or dynamic-inner canonical-correlation analysis (DiCCA).

12. The method of claim 10 wherein: selecting the at least one inner latent variable of a first set of original variables includes selecting the at least one inner latent variable from the first set of original variables based on a desire for a relatively large correlation between the estimated current value of the at least one inner latent variable at a selected time and an actual current value of the at least one inner latent variable at the selected time; and the estimated current value of the at least one inner latent variable at the selected time is estimated from at least one of past values of the at least one inner latent variable or past values of at least another inner latent variable of another set of original variables.

13. The method of claim 10 wherein the estimated current value of the at least one inner latent variable at the selected time is estimated from at least one its past values of the at least one inner latent variable or past values of at least another inner latent variable of another set of original variables.

14. The method of claim 10 wherein the at least one inner latent variable corresponds to a dynamic relationship between at least some of the dynamic samples of data.

15. A system for extracting a set of principal time series data of dynamic latent variables comprising: a plurality of sensors configured to detect dynamic samples of data each corresponding to one of a plurality of original variables; an output device configured to output data; and a controller coupled to the plurality of sensors and configured to: analyze the dynamic samples of data to determine a plurality of latent variables that represent variation in the dynamic samples of data, select at least one inner latent variable that corresponds to at least one of the plurality of original variables, estimate an estimated current value of the at least one inner latent variable based on previous values of the at least one inner latent variable, and control the output device to output data based on the estimated current value of the at least one inner latent variable, wherein the controller is configured to select the at least one inner latent variable based on a desire for a relatively large correlation between the estimated current value of the at least one inner latent variable at a selected time and an actual current value of the at least one inner latent variable at the selected time.

16. The system of claim 15 wherein the controller is configured to analyze the dynamic samples of data using at least one of dynamic-inner principal component analysis (DiPCA) or dynamic-inner canonical-correlation analysis (DiCCA).

17. The system of claim 15 wherein the estimated current value of the at least one inner latent variable at the selected time is estimated from at least one of past values of the at least one inner latent variable or past values of at least another inner latent variable of another set of original variables.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 illustrates a system for extracting principal time series data according to an embodiment of the present invention;

(2) FIG. 2 is a flowchart illustrating a method for extracting principal time series data according to an embodiment of the present invention;

(3) FIG. 3 illustrates an exemplary dynamic whitening filter structure using DiCCA according to an embodiment of the present invention;

(4) FIG. 4 illustrates a system to show how dynamic latent variable analytics distill dynamic latent components according to an embodiment of the present invention;

(5) FIG. 5 illustrates a chemical plant from the Eastman Chemical Company for chemical manufacturing that was used for DiPCA and DiCCA experimentation according to an embodiment of the present invention;

(6) FIG. 6 is a plot illustrating the five dynamic PCs that were identified using DiPCA according to an embodiment of the present invention;

(7) FIG. 7 is a plot illustrating auto-correlation and cross-autocorrelation for the five dynamic PCs of FIG. 6 according to an embodiment of the present invention;

(8) FIG. 8 is a plot illustrating the first two DiPCA PCs of FIG. 6 and predictions from their past scores according to an embodiment of the present invention;

(9) FIG. 9 is a plot illustrating five dynamic PCs that were identified using DiCCA according to an embodiment of the present invention;

(10) FIG. 10 is a plot illustrating autocorrelation and cross-autocorrelation for the five DiCCA PCs of FIG. 9 according to an embodiment of the present invention; and

(11) FIG. 11 is a plot illustrating the first two DiCCA PCs of FIG. 9 and predictions from their past scores according to an embodiment of the present invention.

DETAILED DESCRIPTION

(12) Industrial process data may be massive and may be relatively high dimensional due to the complexity of process operations and control. Although these measurements are high dimensional, the measured variables usually do not act independently due to process operation requirements and physical constraints. This data is often highly collinear and correlated, making traditional regression methods, such as least squares, unreliable due to ill-conditioning. Regularized least squares methods such as ridge regression may be tuned to achieve reliable prediction with bias and variance trade-off. However, these models are not easily interpretable.

(13) Latent variables methods (LVM), including principal component analysis (PCA), projection to latent structures (PLS), and canonical correlation analysis (CCA), may be preferable for analyzing this high-dimensional and correlated data efficiently. Efficiency of the analysis may be improved by analyzing the dynamics of the latent variables.

(14) Because a relatively large amount of process data is collected in the form of time series, with sampling intervals from seconds to milliseconds, dynamics or time correlations are often strong among the data. These dynamics make static data analytics inadequate, however, they may be modeled appropriately so that they can be useful for prediction and monitoring. Considering the large dimensional time series data that are both cross-correlated and auto-correlated over time, it is desirable to develop dynamic extensions of the latent variable methods such that their current values are best predicted by their past data, using a reduced number of dynamic latent variables. The extracted data of these dynamic latent variables may be referred to as principal time series with reduced dimensions.

(15) Referring to FIG. 1, a system 100 for analyzing high dimensional complex data is shown. The system 100 includes a machine learning server 102 along with a data source 104. The machine learning server 102 may include a machine learning controller 106, a machine learning memory 108, and an output device 110. Where used herein, controller may refer to any logic device capable of receiving one or more input and outputting one or more output. For example, a controller may include a PI controller, a processor, an application specific integrated circuit (ASIC), a digital signal processor (DSP), a field-programmable gate array (FPGA), or the like. In some embodiments, the controller 106 may be specifically designed to process relatively large amounts of data.

(16) The memory 108 may include any non-transitory memory known in the art. In that regard, the memory 108 may store data usable by the controller 106. For example, the memory 108 may store an algorithm usable by the controller 106 to model the data source 104 using a set of principal time series data of dynamic latent variables.

(17) The output device 110 may include any output device such as a display, a touchscreen, a speaker, or the like.

(18) The data source 104 may include, for example, an industrial machine or system such as a chemical manufacturing plant, a warehouse, or the like. The data source 104 may include a plurality of sensors 112. Time series data from the plurality of sensors 112 may be transmitted to the machine learning server 102. The machine learning server 102 may make predictions regarding the data source 104 based on the data detected from the plurality of sensors 112 and based on an algorithm that uses dynamic latent variables.

(19) Turning now to FIG. 2, a method 200 for extracting a set of principal time series data of dynamic latent variables is shown. The method 200 may be performed by components of a system similar to the system 100 of FIG. 1.

(20) In block 202, the controller may receive dynamic samples of data. The data may be detected by one or more sensor of a system or process. For example, the data may include one or more of a volume, a flow rate, a temperature, a pressure, or the like. The received data may be dynamic as the values may change over time. Each item of the received data may correspond to one of a plurality of original variables.

(21) In block 204, the dynamic samples of data may be analyzed to determine or identify latent variables. For example, the dynamic samples of data may be analyzed using an algorithm such as dynamic-inner principal component analysis (DiPCA) or dynamic-inner canonical-correlation analysis (DiCCA), as will be described below.

(22) In block 206, at least one inner latent variable may be selected by the controller. The at least one inner latent variable may correspond to at least one of the plurality of original variables. The at least one inner latent variable may be selected based on a desire for a relatively large covariance over time or a relatively large correlation over time. For example, the at least one inner latent variable may be selected based on a desire for a relatively large covariance between an estimated current value of the latent variable at a selected time and an actual value of the latent variable at the selected time. As another example, the at least one inner latent variable may be selected based on a desire for a relatively large correlation between an estimated current value of the latent variable at a selected time and an actual value of the latent variable at the selected time. The at least one inner latent variable may be selected as having a maximum covariance or a maximum correlation over time.

(23) In block 208, the dynamic samples of data may be dynamically distilled by the controller. For example, the dynamic samples of data may be dynamically distilled to maximize the covariance or the correlation, based on whether the relatively large covariance or the relatively large correlation is desired in block 206.

(24) In block 210, the controller may estimate an estimated current value of the at least one inner latent variable based on previous values of the at least one inner latent variable.

(25) In block 212, the controller may control an output device to output data corresponding to the estimated current value of the at least one latent variable. For example, the output device may output the estimated current value. As another example, the output device may output a prediction that is made based on the estimated current value of the at least one latent variable, such as a remaining life of a corresponding system, or a malfunctioning portion of the corresponding system.

(26) In various embodiments, the controller may control an actuator to take an action based on the estimated current value of the at least one latent variable. For example, the controller may identify a fault within a manufacturing or other process based on the estimated current value of the at least one latent variable, may identify a change to an actuator state that will correct the fault based on the estimated current value of the at least one latent variable, and may control the actuator to change to the actuator state to correct the fault.

(27) As described above, the dynamic samples of data may be analyzed to determine a plurality of latent variables that represent variation in the dynamic samples of data using DiPCA.

(28) DiPCA builds dynamic relations of inner latent variables that have maximized auto-covariance. DiPCA may extract one latent variable after another. The extracted variables may be a linear combination of the original variables. In a sense, the current values of the original variables are best predicted from their past values. As a consequence, the residuals after extracting the most predictable latent variables from the data will be least predictable data. In the limiting case, the remaining data tends to be white noise, and, thus, DiPCA may be used as a whitening filter. The DiPCA method overcomes drawbacks of existing dynamic PCA methods that simply perform static PCA on augmented data with time lags.

(29) Various advantages of the DiPCA algorithm that extracts principal time series include that: (i) the dynamic components can be predicted from their past data as known information, so any uncertainty corresponds to the prediction errors only; (ii) the extracted dynamic components may highlight useful dynamic features for data interpretation and diagnosis, which are otherwise difficult to observe from the original data; and (iii) the prediction errors after all dynamics are effectively extracted are essentially not time-correlated and can be further modeled as static data with a traditional PCA method.

(30) In general, it is desirable to extract dynamics in a latent variable t.sub.k so that the current value can be predicted from the past using the following auto-regressive (AR) model shown in Equation 1 below:
t.sub.k=.sub.1t.sub.k1+ . . . +.sub.st.sub.ks+r.sub.kEquation 1
with the latent variable as a linear combination of the original variables t.sub.k=x.sub.k.sup.Tw, where =1, w=1, x.sub.k is the sample vector at time k, and w is the weight vector. If the quantity of time lags, s, is chosen to be sufficiently long that the residual r.sub.k is essentially white noise for each latent variable. The prediction from the dynamic inner model may be shown by Equation 2 below:

(31) $\begin{array}{cc}\begin{array}{c}{\hat{t}}_k=x_{k-1}^{T}w{}_1+\mathrm{.Math.}+x_{k-s}^{T}w{}_s\\ =\left[x_{k-1}^T\mathrm{.Math.}{x}_{k-s}^T\right]\left(.Math.w\right)\end{array}& \mathrm{Equation}2\end{array}$

(32) In Equation 2, =[.sub.1 .sub.2 . . . .sub.s].sup.T and w are constrained to be the unit norm without loss of generality, and .Math.w is the Kronecker product. The objective of the dynamic inner PCA algorithm is to maximize the covariance between the extracted data t.sub.k and the prediction {circumflex over (t)}.sub.k, as shown in Equation 3 below:

(33) $\begin{array}{cc}\underset{w,}{\max }\frac{1}{N}\underset{k=s+1}{\overset{s+N}{.Math.}}{w}^T{x}_k\left[x_{k-1}^T\mathrm{.Math.}{x}_{k-s}^T\right]\left(.Math.w\right)& \mathrm{Equation}3\end{array}$
for (N+s) observations. The data matrix (corresponding to the dynamic samples of data) may be denoted as X=[x.sub.1 x.sub.2 . . . x.sub.N+s], thus forming the data matrices shown in Equation 4 below:
X.sub.i=[x.sub.i x.sub.i+1 . . . x.sub.N+i1].sup.T for i=1,2, . . . ,s+1
Z.sub.s=[X.sub.s X.sub.s1 . . . X.sub.1]Equation 4

(34) The objective of DiPCA may also be formulated as shown below in Equation 5:

(35) $\begin{array}{cc}\underset{w,}{\max }{w}^T{X}_{s+1}^T{Z}_s\left(.Math.w\right)s.t..Math.w.Math.=1,.Math..Math.=1& \mathrm{Equation}5\end{array}$

(36) In Equation 5, S is the dynamic order of the model. The dimension of w is the same as the quantity of variables, which does not increase with the dynamic order of the model. After the weighting vector w is extracted, the latent score t.sub.k is calculated as tk=x.sub.k.sup.Tw. It is clear that the most co-varying dynamic relationship is extracted between t.sub.k and t.sub.k1, . . . , t.sub.ks by the objective function. Therefore, an explicit dynamic model is built between the latent variables. Compared to other dynamic PCA algorithms, the objective function of DiPCA leads to the extraction of only dynamic latent relations. The residuals after all dynamic components are extracted contain little dynamic information and, therefore, can be analyzed using static PCA if so desired.

(37) An algorithm for DiPCA can be performed as follows. Step 1: first, X may be scaled to have a zero mean and unit variance, and w may be initialized to a random unit vector. Latent variables w and may then be extracted in step 2. For example, the following relations shown as Equation 6 may be iterated until convergence:

(38) $\begin{array}{cc}t=\mathrm{Xw}\mathrm{and}\mathrm{form}{t}_i\mathrm{from}t|\mathrm{similar}\mathrm{to}\mathrm{the}\mathrm{formation}\mathrm{of}{X}_i\mathrm{for}i=1,2,\mathrm{.Math.},s+1={\left[\begin{array}{cccc}{t}_1& t_2& \mathrm{.Math.}& t_s\end{array}\right]}^T{t}_{s+1}w=\underset{i=1}{\overset{s}{.Math.}}{}_i\left(X_{s+1}^T{t}_i+X_j^T{t}_{s+1}\right)w:=w/.Math.w.Math.:=/.Math..Math.& \mathrm{Equation}6\end{array}$

(39) Next and after the weight vector w and latent scores t are obtained using the above, in step 3, X may be deflated using Equation 7 below, where the loading vector is p:
X:=Xtp.sup.T; p=X.sup.Tt/t.sup.TtEquation 7

(40) After deflation and in step 4, the algorithm returns to step 2 to extract the next latent variable, until l latent variables are extracted.

(41) In step 5, dynamic inner modeling may occur. For example, a VAR model may be built for latent scores, and T.sub.s+1 may be predicted using Equation 8 below:
{circumflex over (T)}.sub.s+1=T.sub.s{circumflex over ()}Equation 8

(42) Furthermore, X.sub.S+1 may be predicted as X.sub.s+1={circumflex over (T)}.sub.s+1P.sup.T where P=[P.sub.1P.sub.2 . . . P.sub.l] is the loading matrix with each P.sub.i defined in Equation 7.

(43) In step 6, static modeling of prediction errors may be performed. For example, traditional PCA may be performed on the prediction error matrix E.sub.s+1 as shown in Equation 9 below:
E.sub.s+1=X.sub.s+1{circumflex over (T)}.sub.s+1P.sup.T=T.sub.rP.sub.r.sup.T+E.sub.rEquation 9

(44) Therefore, X.sub.S+1 may be decomposed as X.sub.s+1={circumflex over (T)}.sub.s+1P.sup.T+T.sub.rP.sub.r.sup.T+E.sub.r where the first term on the right hand side of the equal symbol is a prediction using past data, while other terms are projections of data involving the current data.

(45) In DiPCA modeling, three parameters need to be determined: dynamic order s, the quantity of dynamic latent variables l, and the quantity of static latent variables. First, assuming s is determined, then l can be chosen such that 95 percent of auto-covariance is captured by the first l dynamic latent variables. Therefore, l is viewed as a function of s, which can be written as l=l(s). To determine the optimal s, a DiPCA model is built based on training the data first. Then, applying the model to the validation dataset facilitates obtaining the prediction error matrix E.sub.S+1.sup.V of the validation data matrix. According to the previous analysis, little dynamic relationships are left in E.sub.S+1.sup.V. Therefore, the sample cross-correlation of any two variables in E.sub.S+1.sup.V should be close to 0, except when lag=0. The calculation of the corresponding confidence bounds may then be found. When all the pairs of variables are considered, the total violations of the confidence bounds can be obtained for any (s, l(s)). The parameter (s, l(s)) corresponding to the minimum violations is determined to be optimal. To determine the quantity of static components, cumulative percentage of variance (CPV) may be applied.

(46) The DiPCA algorithm has a different structure from static PCA and other dynamic PCA algorithms. Therefore, it is important to understand the geometric properties of the DiPCA projections and how the data space is partitioned. To explore the DiPCA geometric properties with j latent variables being extracted, subscript may be used to denote the succession from one latent variable to the next as follows: X.sub.j+1=X.sub.jt.sub.jp.sub.j.sup.T with p.sub.j=X.sub.j.sup.Tt.sub.j/t.sub.j.sup.Tt.sub.j.

(47) The DiPCA algorithm may also be viewed as a whitening filter applied to the data. After all DiPCA components are extracted, the prediction errors are essentially white, as virtually all of the dynamic relationships in the data have been extracted. An important notion of this whitening filter is that it has a reduced quantity of latent variables relative to the quantity of variables in the data, and may be appropriate for modeling the common case of highly collinear data from real world problems. This solution is different from a full dimensional vector autoregressive (VAR) model that requires inversion of a covariance matrix, which can be ill-conditioned with highly correlated data. Furthermore, the DiPCA latent variables have a clear objective and can provide useful features for data based interpretation, visualization, and diagnosis.

(48) In process monitoring based on PCA, squared prediction error (SPE) and Hotelling's T.sup.2 are typically used for detecting abnormal situations. SPE monitors the variations in the residual subspace, and Hotelling's T.sup.2 monitors variations in the principal component sub-space. Process monitoring based on DiPCA can be divided into two parts: the monitoring of dynamic relationships and the monitoring of static relationships. Dynamic latent scores, dynamic residuals, static latent scores, and static residuals can be calculated respectively as follows:

(49) $t_k=R^T{x}_k$$v_k=t_k-{\hat{t}}_k;{\hat{t}}_k=\underset{i=1}{\overset{s}{.Math.}}{}_i^T{t}_{k-i}$$e_k=x_k-P{\hat{t}}_k$$t_{r,k}=P_r^T{e}_k$$e_{r,k}=\left(I-P_r{P}_r^T\right)e_k$

(50) It is clear that process monitoring should be performed on {circumflex over (t)}.sub.k, t.sub.r,k and e.sub.rk, respectively. However, since {circumflex over (t)}.sub.k is dynamic and could even be non-stationary, monitoring t can result in a high false alarm rate. Therefore, the monitoring of {circumflex over (t)}.sub.k can be performed through v.sub.k. Since v.sub.k can be cross-correlated, it is appropriate to build another PCA model on v.sub.k and construct a combined index to monitor v.sub.k. The combined index for v.sub.k is defined as follows:

(51) ${}_v=v_k^T{}_v{v}_k=\frac{T_v^2}{{}_v^2}+\frac{Q_v}{{}_v^2}$${}_v=\frac{P_v{}_v^{-1}{P}_v^T}{{}_v^2}+\frac{I-P_v{P}_v^T}{{}_v^2}\mathrm{where}{}_v=\frac{1}{N-1}{V}^{T}V,T_v^2$
and Q.sub.v are the fault detection indices for PCA based fault detection on v.sub.k, and X.sub.v.sup.2 and .sub.v.sup.2 are the corresponding control limits. In this way, the control region of the prediction {circumflex over (t)}.sub.k can also be indicated by the control region of the combined index coy.

(52) The DiPCA algorithm builds inherent dynamics in the latent variables with explicit projections from the data space to the latent space. However, its objective function, to maximize covariance, may not be as efficient as the canonical-correlation analysis (CCA) objective in maximizing prediction with the least latent dimensions. To obtain a principal time series that can be best predicted from its past values, a dynamic-inner CCA (DiCCA) algorithm is proposed. The objective function of DiCCA is to maximize correlation between latent variables.

(53) Mathematically, it may be desirable to ensure that the dynamic latent variable t.sub.k is best predicted by {circumflex over (t)}.sub.k. This is done by maximizing the correlation between t.sub.k and {circumflex over (t)}.sub.k, which is represented by Equation 10 below:

(54) $\begin{array}{cc}\max \frac{\underset{k=s+1}{\overset{s+N}{.Math.}}{t}_k{\hat{t}}_k}{\sqrt{\underset{k=s+1}{\overset{s+N}{.Math.}}{t}_k^2}\sqrt{\underset{k=s+1}{\overset{s+N}{.Math.}}{\hat{t}}_k^2}}& \mathrm{Equation}10\end{array}$

(55) It can be shown that when restricting .sub.k=s+1.sup.s+Nt.sub.k.sup.2=1 and .sub.k=s+1.sup.s+N{circumflex over (t)}.sub.k.sup.2=1, maximizing Equation 10 is equivalent to minimizing .sub.k=s+1.sup.s+N(t.sub.k{circumflex over (t)}.sub.k).sup.2, the residual sum of squares of the prediction model under these constraints. Therefore, with the same prediction model and matrix notation as Equations 2 and 4, the objective of DiCCA may be rewritten as Equation 11 below:

(56) $\begin{array}{cc}\max \frac{w^T{X}_{s+1}^T\left(X_{s}w{}_1+\mathrm{.Math.}+X_{1}w{}_s\right)}{.Math.X_{s+1}w.Math..Math.X_{s}w{}_1+\mathrm{.Math.}+X_{1}w{}_s.Math.}& \mathrm{Equation}11\end{array}$

(57) Equation 11 can be reformulated into the DiCCA objective function as shown below in Equation 12:

(58) $\begin{array}{cc}\underset{w,}{\max }J=w^T{X}_{s+1}^T{Z}_s\left(.Math.w\right)s.t..Math.X_{s+1}w.Math.=1,.Math.Z_s\left(.Math.w\right).Math.=1& \mathrm{Equation}12\end{array}$

(59) In Equation 12, Xi and Zs are defined by Equation 4.

(60) One dynamic correlation component may be extracted as follows. To solve the optimization problem in Equation 12, LaGrange multipliers are applied as shown in Equation 13 below:
L=w.sup.TX.sub.s+1.sup.TZ.sub.s(.Math.w)+.sub.1(1w.sup.TX.sub.s+1.sup.TX.sub.s+1w)|+.sub.2(1(.Math.w).sup.TZ.sub.s.sup.TZ.sub.s(.Math.w))Equation 13

(61) Making use of the identities (.Math.w)=(.Math.I)w=(I.Math.w) and taking derivatives of L with respect to w and results in Equations 14 and 15 below:

(62) 0$\begin{array}{cc}\frac{L}{w}=X_{s+1}^T{Z}_s\left(.Math.w\right)+{\left(.Math.I\right)}^T{Z}_s^T{X}_{s+1}w-{}_1{X}_{s+1}^T{X}_{s+1}w-{{}_2\left(.Math.I\right)}^T{Z}_s^T{Z}_s\left(.Math.I\right)w=0& \mathrm{Equation}14\\ \frac{L}{}={\left(I.Math.w\right)}^T{Z}_s^T{X}_{s+1}w-{{}_2\left(I.Math.w\right)}^T{Z}_s^T{Z}_s\left(I.Math.w\right)=0& \mathrm{Equation}15\end{array}$

(63) Pre-multiplying Equation 15 by .sup.T and using the constraint in Equation 12 results in J=.sub.2. Pre-multiplying Equation 15 by w.sup.T and referring to the constraint in Equation 12 results in 2J.sub.1.sub.2=0, leading to .sub.1=J Therefore, let =.sub.1.sub.2. Additionally, T.sub.s may be defined as follows in Equation 16:

(64) $\begin{array}{cc}{T}_s=Z_s.Math.\left(I.Math.w\right)=\left[\begin{array}{cccc}{X}_{s}w& X_{s-1}w& \mathrm{.Math.}& X_{1}w\end{array}\right]=\left[\begin{array}{cccc}{t}_s& t_{s-1}& \mathrm{.Math.}& t_1\end{array}\right]& \mathrm{Equation}16\\ {\hat{X}}_{s+1}=Z_s\left(.Math.I\right)=\underset{i=1}{\overset{s}{.Math.}}{}_i{X}_{s-i+1}& \mathrm{Equation}17\end{array}$

(65) By combining Equations 16 and 17, Equation 18 is derived:

(66) $\begin{array}{cc}{Z}_s\left(.Math.w\right)=\underset{i=1}{\overset{s}{.Math.}}{}_i{X}_{s-i+1}w={\hat{X}}_{s+1}w={\hat{t}}_{s+1}& \mathrm{Equation}18\end{array}$
where:
t=Xw.sup.s+N
t.sub.i=X.sub.iw.sup.N, for i=1,2, . . . ,s+1Equation 19

(67) Equation 15 may be rewritten as Equation 20 below:
T.sub.s.sup.TX.sub.s+1w=T.sub.s.sup.TT.sub.s
or,
(T.sub.s.sup.TT.sub.s).sup.1T.sub.s.sup.TX.sub.s+1 w=Equation 20

(68) Similarly, Equation 14 may be rewritten as Equation 21 below:
X.sub.s+1.sup.T{circumflex over (X)}.sub.s+1w+{circumflex over (X)}.sub.s+1.sup.TX.sub.s+1w|=(X.sub.s+1.sup.TX.sub.s+1+{circumflex over (X)}.sub.s+1.sup.T{circumflex over (X)}.sub.s+1)w
or,
(X.sub.s+1.sup.TX.sub.s+1+{circumflex over (X)}.sub.s+1.sup.T{circumflex over (X)}.sub.s+1).sup.+(X.sub.s+1.sup.T{circumflex over (X)}.sub.s+1+{circumflex over (X)}.sub.s+1.sup.TX.sub.s+1)w=w Equation 21

(69) Where ( ).sup.+ denotes the Moore-Penrose pseudo-inverse, Equations 20 and 21 imply that w is the Eigenvector of (X.sub.s+1.sup.TX.sub.s+1+{circumflex over (X)}.sub.s+1.sup.T{circumflex over (X)}.sub.s+1).sup.+(X.sub.s+1.sup.T{circumflex over (X)}.sub.s+1+{circumflex over (X)}.sub.s+1.sup.TX.sub.s+1) corresponding to the largest Eigenvalue. However, since {circumflex over (X)}.sub.s+1 depends on and therefore and w are coupled together, there is no analytical solution to the optimization problem in Equation 12. Equations 20 and 21 can be reorganized as follows using Equations 18 and 19, as shown in Equations 22 and 23 below:

(70) $\begin{array}{cc}\begin{array}{c}w={\left(X_{s+1}^T{X}_{s+1}+{\hat{X}}_{s+1}^T{\hat{X}}_{s+1}\right)}^{+}\left(X_{s+1}^T{\hat{t}}_{s+1}+{\hat{X}}_{s+1}^T{t}_{s+1}\right)\\ ={{\left({\left[\begin{array}{c}{X}_{s+1}\\ {\hat{X}}_{s+1}\end{array}\right]}^T\left[\begin{array}{c}{X}_{s+1}\\ {\hat{X}}_{s+1}\end{array}\right]\right)}^{+}\left[\begin{array}{c}{X}_{s+1}\\ {\hat{X}}_{s+1}\end{array}\right]}^T\left[\begin{array}{c}{\hat{t}}_{s+1}\\ t_{s+1}\end{array}\right]\\ ={\left[\begin{array}{c}{X}_{s+1}\\ {\hat{X}}_{s+1}\end{array}\right]}^{+}\left[\begin{array}{c}{\hat{t}}_{s+1}\\ t_{s+1}\end{array}\right]\end{array}& \mathrm{Equation}22\\ ={\left(T_s^T{T}_s\right)}^{-1}{T}_s^T{t}_{s+1}& \mathrm{Equation}23\end{array}$

(71) It is clear from Equation 23 that depends on w implicitly through T.sub.i. Also, is proportional to the least squares solution of the AR model parameters of time series {t.sub.k}.sub.k=1, 2, . . . . Since w is an Eigenvector based on Equation 21, the norm of w is scaled to one. Therefore, can be found to be the least squares solution from Equation 23.

(72) The core DiCCA algorithm may be described in the following steps. In step 1, w may be initialized with a column of an identity matrix. In step 2, w and may be calculated by iterating the following relations shown in Equation 24 below until convergence:

(73) $\begin{array}{cc}t=\mathrm{Xw}\mathrm{and}\mathrm{form}{t}_i\mathrm{from}t\mathrm{according}\mathrm{to}\mathrm{Eq}.19\mathrm{Form}{T}_s=\left[t_s\mathrm{.Math.}{t}_1\right].={\left(T_s^T{T}_s\right)}^{-1}{T}_s^T{t}_{s+1},{\hat{X}}_{s+1}=\underset{i=1}{\overset{s}{.Math.}}{}_i{X}_{s-i+1},{\hat{t}}_{s+1}=\underset{i=1}{\overset{s}{.Math.}}{}_i{t}_{s-i+1},w={\left[\begin{array}{c}{X}_{s+1}\\ {\hat{X}}_{s+1}\end{array}\right]}^{+}\left[\begin{array}{c}{\hat{t}}_{s+1}\\ t_{s+1}\end{array}\right]\mathrm{and}w:=w/.Math.w.Math.& \mathrm{Equation}24\end{array}$

(74) In step 3, J may be calculated using Equation 25 below:

(75) $\begin{array}{cc}J==\frac{{\hat{t}}_{s+1}^T{t}_{s+1}}{.Math.{\hat{t}}_{s+1}.Math..Math.t_{s+1}.Math.}& \mathrm{Equation}25\end{array}$

(76) To extract the next dynamic latent variable, the same iteration procedure can be applied to the deflated matrices of X.sub.s+1 and Z.sub.s, which will be discussed below.

(77) After the loading vector w and latent scores t are obtained from the iteration procedure, x is deflated using Equation 26 below:
X:=Xtp.sup.TEquation 26

(78) Where the loading vector p is defined in Equation 27 below:
p=X.sup.Tt/t.sup.TtEquation 27

(79) The deflated matrix X is then used to repeat the same iteration procedure to extract the next latent variable. The deflation procedure leads to desirable geometric properties of the DiCCA algorithm. For instance, the score vectors t from different latent variables are orthogonal.

(80) After obtaining the latent variable t.sub.k, an AR model can be built to describe the dynamics in t.sub.k as shown below in Equation 28:
t.sub.k=.sub.1t.sub.k1+ . . . +.sub.st.sub.ks+.sub.kEquation 28

(81) The solution to the estimate is ordinary least squares, which coincidentally is already solved in the iterative algorithm as . Therefore, there is no need to fit another AR model.

(82) With other dynamic data modeling algorithms, a re-estimation of may be performed after an outer model projection. The extraction of the latent variables and dynamic modeling of the latent variables are achieved simultaneously in DiCCA, because DiCCA employs consistent outer modeling and inner modeling objectives. This is a unique property of DiCCA and makes it a more efficient dynamic modeling algorithm than the others.

(83) The DiCCA algorithm extracts latent time series variables one by one with descending predictability, or R.sup.2, values. After l latent time series variables are extracted, the next latent time series variable that is extracted will have a R.sup.2 value close to 0, which implies that there are little or no dynamics left in the residuals. The orthogonality of the latent scores guarantees that the quantity of latent time series variables required to extract all dynamics is fewer than the quantity of variables, which will be shown below. Mathematically, by using t.sub.k.sup.(j) to denote the j.sup.th latent score at time k, and .sub.ji for i=1, 2, . . . s to denote the AR coefficients for the j.sup.th latent score, the prediction model for each score may be shown in Equation 29 below:

(84) $\begin{array}{cc}\begin{array}{c}{\hat{t}}_k^{\left(j\right)}=\left({}_{j1}{q}^{-1}+{}_{j2}{q}^{-2}+\mathrm{.Math.}+{}_{\mathrm{js}}{q}^{-s}\right)t_k^{\left(j\right)}\\ =G_j\left(q^{-1}\right)t_k^{\left(j\right)}\end{array}& \mathrm{Equation}29\end{array}$

(85) In Equation 29, q.sup.1 is a backwards shift operator. By combining l prediction models together, a prediction model for the latent score vector tk=[t.sub.k.sup.(1) t.sub.k.sup.(2) . . . t.sub.k.sup.(l)].sup.T may be shown in Equation 30 below:

(86) $\begin{array}{cc}\begin{array}{c}{\hat{t}}_k=G\left(q^{-1}\right)t_k\\ =\mathrm{diag}\left(G_1\left(q^{-1}\right),G_2\left(q^{-1}\right),\mathrm{.Math.},G_l\left(q^{-1}\right)\right)t_k\end{array}& \mathrm{Equation}30\end{array}$

(87) Equation 30 may be used to derive the one step prediction error as shown in Equation 31 below, which may correspond to a whitening filtered error.
e.sub.k=x.sub.k P{circumflex over (t)}.sub.k=x.sub.k PG(q.sup.1)t.sub.kEquation 31

(88) When the quantity of dynamic latent variables l is selected to extract all dynamics in the data, little or no dynamics will remain in e.sub.k. Furthermore, PCA modeling of e.sub.k may be appropriate if it is desired to extract a covariance structure in the prediction error.

(89) Additionally, since little or no dynamics remain in e.sub.k after removal of the prediction error, DiCCA may be interpreted as a whitening filter which removes all dynamics in the data. The whitening filter may be written as shown in Equation 32 below:
e.sub.k=(IPG(q.sup.1)R.sup.T)x.sub.kEquation 32

(90) FIG. 3 illustrates an exemplary dynamic whitening filter structure 300 using DiCCA. As shown, the filter structure 300 receives the dynamic samples of data 302 x.sub.k. The data is applied to the matrix 304 R.sup.T, the dynamic transfer function 306 G(q.sup.1), and the projection 308 P. The result of the projection 308 is subtracted from the dynamic samples of data 302 x.sub.k to provide the white noise 310 e.sub.k

(91) With the objective of maximizing covariance or correlation between the latent time series and its prediction from the past, DiPCA and DiCCA each perform dynamic data distillation from all measured data such that the extracted dynamic components co-vary or correlate the most with their past. These most predictable dynamic components are referred to as principal time series that are latent in the original data. The prediction errors of the data after the first predicted component is extracted are then used to extract the second most predictable latent component, until all significant dynamic components are extracted. This procedure is analogous to a multi-stage binary distillation process, with each stage separating the most dynamic component from others. The leftovers after all dynamic components are extracted are essentially un-correlated in time, resembling white noise.

(92) FIG. 4 shows a system 400 to illustrate how these dynamic latent variable analytics distill dynamic latent components one after another, with the objective to maximize the covariance or correlation between the component and their prediction from the past data. High dimensional time series data are considered as a mixture of a number of dynamic latent components, which are not measured directly, and static errors. DiPCA and DiCCA distill the multi-dimensional data into dynamic components in descending order of covariance or correlation, respectively.

(93) FIG. 5 illustrates a chemical plant 500 from the Eastman Chemical Company for chemical manufacturing that was used for DiPCA and DiCCA experimentation. The chemical plant 500 includes controllers 502 (that include sensors or otherwise measure and output data) and sensors, or indicators, 504 that output measured data. For example, the controllers 502 and indicators 504 may output temperatures, flow rates, fluid levels, or the like.

(94) The Eastman Chemical Company had identified a need to diagnose a common oscillation of about two hours (with about 340 samples per cycle). Early work focused on the diagnosis of this oscillation. Using Granger causality analysis, five process variables were identified to have strong oscillations. These five variables were selected for this disclosure to illustrate how the dynamic data and features can be modeled using DiPCA and DiCCA.

(95) Using DiPCA on the five process variables leads to five dynamic principal components (PCs). FIG. 6 is a plot 600 illustrating the five dynamic PCs 602 with the most dynamic latent variable (with maximum covariance) at the top and the least dynamic latent variable at the bottom. The auto-regression order of the dynamics is chosen as 21, which makes the prediction errors of the dynamic principal components essentially white.

(96) FIG. 7 is a plot 700 illustrating the auto-correlation and cross-autocorrelation for the five dynamic PCs. Referring to FIGS. 6 and 7, it is clear that the first two PCs 604, 606 are very oscillatory, while the third PC 608 is still somewhat oscillatory and co-varies with the first two PCs 604, 606. To visualize how the DiPCA model predicts the PCs, the first two DiPCA PCs and the predictions from their past scores are shown in a plot 800 in FIG. 8. The circular shape shows the co-varying oscillations at the same frequency.

(97) DiCCA was then used to model the five process variables, which again leads to five dynamic PCs. FIG. 9 is a plot 900 illustrating the five PCs 902 which were distilled from the five process variables. The order of the dynamics was chosen as 22, which was selected such that the errors predicted with the dynamic PCs were essentially white. FIG. 10 is a plot 1000 illustrating the autocorrelation and cross-autocorrelation for the five DiCCA PCs. Referring to FIGS. 9 and 10, is clear that the first two PCs 904, 906 are very oscillatory, while the third PC 908 has little correlation to the first two PCs 904, 906.

(98) To visualize how the DiCCA model predicts the PCs, the first two DiCCA PCs and the predictions from their past scores are shown in a plot 1100 of FIG. 11. While the large circular shape illustrates co-varying oscillations at the same frequency, there is a clear smaller oscillation with a higher frequency that is best captured by the second PC. This feature is not observed at all using DiPCA analysis. The DiCCA scatterplot has clear but faster oscillations on top of the large circular shape, indicating that there is another high frequency oscillation component. This frequency is likely caused by another control loop with valve stiction. The fact that DiCCA detects a new feature may indicate that DiCCA is better suited for extracting dynamic features than DiPCA. However, DiPCA may be better suited for extracting dynamic features in some situations.

(99) Process data analytics have been applied in chemical process operations for decades. Although latent variable methods have been successfully applied to static data analysis, the methods presented in this disclosure (DiPCA and DiCCA) appear to be the first class of methods that maximize the dynamic content of the projected latent scores. The principal dynamic latent variables are most predictable components in the whole data space, and therefore have the least prediction errors. In the industrial data example from the Eastman Chemical Company disclosed herein, the sustained dynamic oscillatory content is undesirable in process operations. Further diagnostic analysis reveals what they are and where they come from. By fixing the causes, the excessive oscillations may be eliminated.

(100) With the advancement of analytics in other sectors of industries and business operations, there appear to be many more research opportunities in the future. While physical and chemical sciences develop principles which enable mechanistic models to be established for process understanding, data analytics provide real and up-to-date information that reflects changes and uncertainty in the operation, and provide a reliable source of information to detect emerging situations.

(101) Prediction, visualization, and interpretation of massive data with latent variables are powerful tools for dealing with high dimensional and highly correlated data. The benefit of data analytics is to turn data into knowledge and support effective operations and decision-making, which help push beyond the traditional boundaries.

(102) Exemplary embodiments of the methods/systems have been disclosed in an illustrative style. Accordingly, the terminology employed throughout should be read in a non-limiting manner. Although minor modifications to the teachings herein will occur to those well versed in the art, it shall be understood that what is intended to be circumscribed within the scope of the patent warranted hereon are all such embodiments that reasonably fall within the scope of the advancement to the art hereby contributed, and that that scope shall not be restricted, except in light of the appended claims and their equivalents. Where used throughout the disclosure and claims, at least one of A or B includes A only, B only, or A and B.