Computers and Chemical Engineering 106 (2017) 1–16

Contents lists available at ScienceDirect

Computers and Chemical Engineering

journal homepage: www.elsevier.com/locate/compchemeng

Refined convergent cross-mapping for disturbance propagation

analysis of chemical processes
Lei Luo a,b , Feifan Cheng a,b , Tong Qiu a,b,∗ , Jinsong Zhao a,b
a
Department of Chemical Engineering, Tsinghua University, Beijing 100084, China
b
Beijing Key Laboratory of Industrial Big Data System and Application, Tsinghua University, Beijing 100084, China

a r t i c l e i n f o a b s t r a c t

Article history: In chemical processes, determining the root causes of faults and disturbances is important for improving
Received 11 July 2016 process safety and economic profit. Recently proposed convergent cross-mapping (CCM) is suitable for
Received in revised form 15 March 2017 both linear and nonlinear systems. However, it cannot be directly applied to chemical processes. This is
Accepted 19 March 2017
because chemical processes possess characteristics, such as nonlinearity, high dimensions, continuity,
Available online 29 April 2017
and time delay, that are significantly different from ecosystems, where CCM is applicable. In this paper,
we propose refined CCM (RCCM), which integrates fast embedding dimension determinations with a
Keywords:
modified procedure to improve the speed of the calculation and accuracy of the analysis. Examples show
Convergent cross-mapping
Root cause
that RCCM fits chemical processes with the above characteristics. In addition, it shows better performance
Nonlinearity than other methods in finding the propagation paths of disturbances.
Time delay © 2017 Elsevier Ltd. All rights reserved.
Chemical process

1. Introduction
In chemical processes, disturbance propagation paths describe
ways in which disturbance signals consisting of stream flows and
control paths travel through the process. Root-cause identifica-
tion of plant-wide faults and disturbances is important for safe
and efficient operation of chemical processes. However, because
of the process complexity and existence of control loops and
time delays, it is still difficult to determine the correct propa-
gation paths. Several methods have been proposed to solve this
problem. Most of these methods originate from cause-and-effect
analysis approaches and are then utilized to identify the time-
delay relationships between the variables. Among the approaches
to determine the correct propagation paths, the simplest approach
is called cross-correlation analysis (Bauer and Thornhill, 2008). In
this method, different values of the cross-correlation function (CCF)
are obtained at different time delays. The best estimated time delay
and its corresponding cross-correlation are determined by the max-
imum absolute value of the CCF. Although this method is practical
and computationally efficient, it is too simple to identify nonlinear
causal relationships existing in the system. For example, Yang et al.
(2010) found that if y = x2 with a time delay of one sample, where x
∗ Corresponding author at: Department of Chemical Engineering, Tsinghua Uni-
versity, Beijing 100084, China.
E-mail address: qiutong@tsinghua.edu.cn (T. Qiu).
is a superposition of a sine signal and white noise, then the CCF can-
not determine this causality because all of the values are too small
compared with the detecting threshold. In fact, the true correlation
is zero, which means that the CCF cannot find the result. In another
method called the Granger causality (GC) test, variable x “Granger
causes” y if the prediction ability for y decreases when x is removed
from all possible causative variables of y. However, the GC test is
not suitable for coupled systems because it requires that the two
variables for the test should be mutually independent, so the influ-
ence of one variable, such as x, on the other variable, such as y, could
be eliminated by simply removing x from the original system. This
is impossible in a coupled system and the result is trustless when
the separability cannot be satisfied (Sugihara et al., 2012). Schreiber
(2000) proposed the transfer entropy based on the concept of infor-
mation entropy, aiming to pairwise measure the interactions of
variables in a system. The transfer entropy is model-free and it
is widely used in cause-and-effect analysis of chemical processes,
atmospheric science, civil engineering, neuroscience, and other
fields (Bauer et al., 2007; Vicente et al., 2011; Xie et al., 2012; Zhang
et al., 2013). However, because this method is a probability-based
method, the result accuracy is based on estimation of probability
density functions and the whole computational burden is very high.
Furthermore, the transfer entropy cannot always ensure the cor-
rect time-delay relationships between variables, so the arc sign in
signed directed graphs cannot be obtained (Yang et al., 2010). The k-
nearest neighbor (KNN) method proposed by Stockmann and Haber
(2010) uses information about the nearest points that are close to

http://dx.doi.org/10.1016/j.compchemeng.2017.03.026
0098-1354/© 2017 Elsevier Ltd. All rights reserved.
2 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16
input sample points to predict the actual outputs of those input
samples, and it then finds the time-delay peaks by comparing the
predicted and actual values along the time axis. This method is
thought to be faster than the transfer entropy and applicable for
nonlinear systems. However, in the KNN method, the important
imputation number kimp is chosen to be 2, 4, or 6. This subjective
selection cannot always ensure accurate results and it is difficult
to apply in automatic detection. In addition, the KNN method can-
not avoid the false-nearest-neighbor problem in high-dimensional
situations because of the absence of phase-space reconstruction.
For the above methods, the result accuracy and calculation effi-
ciency are two crucial aspects for evaluating their performance.
Most of these methods have shortcomings such as high computa-
tional burden or limited performance for linear or weak nonlinear
systems. These methods cannot ensure correct results for causal-
ity identification of coupled variables in strong nonlinear systems,
while recently proposed convergent cross-mapping (CCM) can
overcome this problem (Sugihara et al., 2012). CCM uses the dif-
feomorphic features of coupled systems and converts the original
causality detecting problem into comparisons between the mutual
prediction effects of embedding manifolds. Recently, Ye et al. (2015)
found that CCM is also capable of distinguishing time-delayed inter-
actions between different process variables. In their method, if
a time delay exists in the effect of an upstream variable y on a
downstream variable x then xt = f(yt−t ). CCM should have the best
detecting performance when two variable series are matched in
accordance with t. However, in our previous study (Luo et al.,
2016), we found that because of significant differences in the
embedding parameters between variables of continuous chemical
processes, directly applying CCM usually cannot obtain as accurate
results as when CCM is applied to ecosystems, especially when the
process is high-dimensional with complicated dynamic behaviors.
In this paper, we refine the original CCM procedure and apply it
to a chemical process to determine the propagation paths and root
causes of process faults and disturbances. In Section 2, we intro-
duce the original CCM and its refinement for chemical processes
(RCCM). Compared with CCM, RCCM includes two modifications.
First, it refines the manifold reconstructing process by using the
time lag of the predicted variable to build both of the reconstructed
manifolds, which enables it to be more suitable for variables with
significant differences in time lags, like the situation in chemical
processes where the data is usually from the distributed control
system and each variable fluctuates at different frequencies. In
addition, it integrates fast and accurate parametric determinations
for  and E in the flowchart. This step is important because there are
tens to hundreds of variables in a chemical process, so efficient and
automatic parameter determination procedures will greatly facili-
tate the analysis. At the end of Section 2, we compare RCCM with
original CCM and provide the program flowchart of RCCM. In Sec-
tion 3, we test RCCM on three examples to show its applicability
for chemical processes and its advantages over other well-known
methods. In example 1, we apply RCCM to a continuous stirred tank
reactor to show its applicability for chemical processes. In examples
2 and 3, we test RCCM for four different linear and nonlinear single-
input-single-output (SISO) systems and a benchmark model with
nonlinear subprocesses. In Section 4, we test RCCM on a benchmark
process: the Tennessee Eastman process.
2. CCM and RCCM
CCM was proposed by Sugihara et al. (2012) to determine the
interaction strengths between different variables in a nonlinear
ecosystem. This method is based on diffeomorphism theory and
uses the characteristics of the interactions between coupled vari-
ables of nonlinear systems to provide reliable detection of causal
relationships.
According to the Takens embedding theorem, if two variables
are causally linked in a dynamic system, their corresponding
shadow manifolds Mx and My would be diffeomorphic to the man-
ifold of the original system M. Here, Mx and My are reconstructed
using lagged-coordinate embedding of time series x and y with
embedding dimension E and time lags  x and  y . The elements of
Mx and My are denoted as xt and yt . Therefore, there is a consecutive
one-to-one mapping relationship between the points in Mx and My .
For some specific points xt,1 , xt,2 ,. . .,xt,i in Mx , there must be corre-
sponding yt,1 , yt,2 ,. . .,yt,i points in My . If xt,1 converges to a specific
point xt0 , yt,i would also converge to the corresponding point yt0 in
My . Thus, once the process data is available, CCM attempts to find
the nearest points for each xt,i in Mx and their mapping points in My
to predict the corresponding value of yt,i , which is denoted as ŷt,i .
The cross-mapping ability of x for y canbe  evaluated by the correla-
tion coefficient x→y between {yt } and ŷt . As the sample number
L increases, ŷt,i converges, and so does y . Thus, x can be used to
estimate y, and vice versa. However, if x and y are not coupled, the
accuracy of the prediction decreases owing to asymmetric infor-
mation exchange between the two observed variables. Note that
the unilateral causality in CCM is counterintuitive compared with
other methods, such as the transfer entropy. A higher value of x→y
means that the variable x is more affected by y, so y tends to be a
cause of x.
Here, we describe the mathematical mechanism of CCM. The
variables are produced by an n-dimensional dynamic system:
dX
= f (X, U) (1)
dt
Here, X and U are vectors of the variables and parameters: X = [x1 ,
x2 , . . ., xn ]T and U = [u1 , u2 , . . ., ul ]T . For example, taking the two
variables x and y of X, each element of the corresponding shadow
manifolds of x and y are constructed according to
 
M x,t = xt , xt−x , xt−2x , . . ., xt−(E−1)x (2)
 
M y,t = yt , yt−y , yt−2y , . . ., yt−(E−1)y (3)
First, Mx is used to predict My and evaluate the influence of y
on x. If parameter u is stable according to the Takens embedding
theorem, Mx and My would be stable and they are diffeomorphic to
the manifold of the original system M. This will lead to a consecutive
one-to-one mapping relationship between the points on in Mx and
My :
lim M y,i → M y,k0 = 0 (4)
M x,i →M x,k0
Thus, if we determine E + 1 points in Mx then
{M x,ki } = {M x,k1 , M x,k2 , . . ., M x,k(E+1) } (5)
which are close to Mx,k0 . The value of My,k0 can then be estimated
by {My,ki }:
 
M x,ki − M x,k0
d(M x,ki , M x,k0 ) = exp − (6)
M x,k1 − M x,k0

E+1
d(M x,ki , M x,k0 )
M̂ y,k0 |M x = E+1
M y,ki (7)
i=1 j=1
d(M x,kj , M x,k0 )
If L is sufficiently large to ensure that {Mx,ki } is sufficiently close
to My,k0 , the predicting ability of Mx for My will converge to a
positive constant between 0 and 1:
lim cov M y , M̂ y |M x = x→y (8)
L→+∞
 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16
This procedure is the same for My predicting Mx , and the differ-
ence between the predicting abilities x→y and y→x indicates an
asymmetric interaction between x and y.
Fig. 1 shows a schematic illustration of CCM. Fig. 1A shows the
bidirectional causality between x and y, in which Mx is diffeomor-
phic to My . When making predictions from x to y, the nearest points
in Mx (the yellow triangles in Fig. 1) are mapped to My , and the
uncertainty of ŷt is low, which is also the same for the prediction
from y to x. Fig. 1B shows the asymmetric causality between two
variables. Here, y is the result of x, so y does not have any influence
on x and My is not diffeomorphic to Mx . The prediction from x to
y fails with a high uncertainty of E owing to incomplete informa-
tion about y in x, whereas the prediction from y to x still has good
accuracy because y contains information about x.
Compared with the KNN method, which is also based on nearest
points information, CCM introduces a phase-space reconstruction
step that enables it to avoid the false-nearest-neighbor problem, as
shown in Fig. 2. The left graph in Fig. 2 shows the manifold of a chaos
system, where the best embedding dimension is E = 3. However, if
the embedding dimension decreases to E = 2, as in middle graph, the
nearest points are incorrect. The overlapping red and green points
in the middle graph are not the nearest points compared with the
left original manifold and the right graph (E = 3).
In Sugihara’s paper (Sugihara et al., 2012), the dynamic model is
assumed to be an immediate action model, that is, there is no time
delay for information transferring from one variable to another,
which is impractical, especially for other large-scale systems. For
example, in chemical processes, all of the units are connected by
pipelines through which raw materials and products flow from one
unit to another downstream destination. The result measured at
time t may be a consequence of an upstream variable at t − t, that
is, variabledownstream,t = f (variableupstream,t−t ). These time delays
between different variables are the result of a comprehensive func-
tion of multiple factors and they reflect the information of faults or
disturbance propagation. Ye et al. (2015) proposed an approach to
determine this time-delay relationship with CCM, and the main
idea is shown in Fig. 3. When using a downstream variable x to
predict an upstream variable y, if there is no time delay between
these variables, xt+1 would be a consequence of yt and the cross-
mapping ability of x for y will reach the maximum when the time
lag is 1. However, if there is a significant time delay t between the
two variables, xt+1 would be a consequence of yt−t and the peak of
the cross-mapping ability curve will move toward the right by t,
which is exactly the time delay from y to x.
2.1. Refinement 1: refined manifold reconstruction
In the method of Ye et al. (2015), all of the reconstructed man-
ifolds share the same time lag of  = 1. This does not affect the
accuracy of the results because the data is from an ecosystem or
other discrete models with relatively close time lags of variables.
However, for chemical processes, the time lags become much larger
with significant differences. Their method is not able to obtain
a precise estimation of the time delays and it cannot be directly
applied in this situation. In their method, manifolds Mx and My are
reconstructed according to Eqs. (2) and (3), respectively, and then
included in the calculation. This treatment can obtain good results
when  x and  y are close. However, if there are significant differ-
ences between the values of these time lags, the results will be
inaccurate (see the example shown on the left of Fig. 5). Here, we
introduce a refinement on the original CCM procedure when using
x to predict y or to find if y causes x. The manifolds Mx and My are
both reconstructed with the time lag of the predicted variable  y :
 
M x,t = xt , xt−y , xt−2y , . . ., xt−(E−1)y
3
 
M y,t = yt , yt−y , yt−2y , . . ., yt−(E−1)y (10)
Compared with CCM, this refinement is more reasonable because
when time series x is used to predict another series y, that is, cal-
culating the influence of y on x, only the effects of y on x in the
full frequency range of y should be considered. Thus, if there is a
significant difference between the two frequencies, which means a
significant difference between  x and  y , the corresponding man-
ifolds should be reconstructed with the time lag of the predicted
variable  y .
This assertion can be confirmed by the following example:
xt+1 = 0.3xt + (1 + 0.3ut − yt · xt /5) (11)
yt+1 = 0.9yt + 0.02(xt−20 − 5.5) (12)
where ut is a random external disturbance in accordance with the
0–1 uniform distribution. Fig. 4 shows all of the outputs. It clearly
shows that x has a much smoother variation than x and u owing to
the different strengths of the interactions between these variables,
which indicates that these variables should share different time lag
parameters. Fig. 5 compares the original reconstruction with the
refined reconstruction, and it shows that the refined reconstruction
produces better results. It shows that the time delay from x to y
tx→y = 0 while that from y to x ty→x =+20 points (the positive
time delay means y is behind x in time or y is caused by x after a
time length of 20 points), which is in accordance with Eqs. (11) and
(12).
2.2. Refinement 2: efficient and automatic parameter
determination
There are two key parameters in the method: embedding
dimension E and time lag . For each chemical process, tens to
hundreds of variables need to be analyzed, which is a much larger
number of variables than in ecosystems. This requires efficient and
accurate parameter determination approaches. Here, we introduce
Cao’s improved FNN method (Cao, 1997) and linear autocorrela-
tion function to determine E and  because of their relatively low
computation cost and good accuracy.
2.3. Determination of the embedding dimension
For a specific point M x,k0 , if a point in Mx can be found, like
M x,n(k0 ,E) , where n(k0 , E) is the integer that makes M x,n(k0 ,E) the
nearest point to M x,k0 when the embedding dimension is E, the
nearest distance will be
(E)
distE,k0 = M x,k0 − M x,n(k (13)
0 ,E)
When the embedding dimension increases to E + 1, the distance
between these two points will increase to
(E+1)
distE+1,k0 = M x,k0 − M x,n(k (14)
0 ,E)
For R between 10 and 50, if
(distE+1,k0 − distE,k0 )
>R (15)
distE,k0
then M x,n(k0 ,E) is thought to be a false nearest neighbor of M x,k0 . For
an actual observed sequence, E starts from 2. For all points in Mx ,
if the ratio of the false nearest neighbor no longer increases, the
corresponding E should be chosen as the appropriate embedding
dimension.
Some papers have pointed out that the original FNN method is
subjective because the threshold of R is difficult to determine and it
has a significant influence on the final result. Furthermore, the FNN
method may be not applicable for some systems under random
(9)
disturbances. Several modifications of the original FNN procedure
4 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16

Fig. 1. Schematic illustration of CCM (Sugihara et al., 2012).

Fig. 2. Illustration of the false-nearest-neighbor problem.

Fig. 3. Schematic of the time-delay detection method of Ye et al. (2015).

 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16 5

Fig. 4. Outputs of Eqs. (11) and (12).

Fig. 5. Comparison of Ye’s method and RCCM.

have been performed, and the modified method is used in this paper By investigating the variation of the variable
(Cao, 1997; Shun et al., 2015).
ā(E + 1)
To overcome the shortcomings of the FNN method, we define A(E) = (18)
ā(E)
the best embedding dimension where A(E) stops changing and con-
distE+1,k0
a(k0 , E) = , k0 = 1, 2, 3, . . ., L − E (16) verges to 1 can be determined.
distE,k0
2.4. Determination of the time lag of each variable
where L is the element number of manifold Mx . The mean value of
a is As previously discussed, chemical process data from the
distributed control system is continuous and very large with signif-
icantly different time lags between different variables. The choice
1 
L−E
of the time lags significantly affects the results. If the time lags are
ā(E) = a(k0 , E) (17)
L − E too small, the points in each reconstructed manifold will be highly
k0 =1 correlated and the manifold cannot sufficiently unfold. If the time
6 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16

lags are too large, there will be no relationships between the points Table 1
Parameters of the CSTR model.
and the reconstructed manifold will lose its continuity. There are
mainly two ways to determine the time lags of time series: the Parameter Value Unit
mutual information method and the linear autocorrelation func- F 100 L/min
tion. In this paper, the linear autocorrelation function is chosen V 100 L
because of its lower calculation burden. k1 7.2 × 1010 1/min
The time lags of x are determined by the following function: k2 5.2 × 1010 1/min
E1 /R 8750 K
E2 /R 8750 K
1 N−
N n=1 (xn+
− x̄) (xn − x̄)
C() = N−
(19)
1
N n=1 (xn
− x̄)2 Table 2
Parameters of RCCM for example 1.

Tfl cin cA cB cc
where
Time lag  1 1 25 46 43
Embedding dimension E 7 7 7 7 7
1
N

x̄ = xn (20)
N
n=1 3. Application of RCCM to chemical processes

 0 , whose corresponding C() reaches zero first, is chosen to be
the time lag of the embedding manifold. This ensures sufficient
extension of the reconstructed manifold Mx and simultaneously
maintains the relationships of the variables.
The above two refinements enable RCCM to process the detect-
ing ability of the chemical process. Its accuracy is determined by
the sample length L. As shown in the left graph of Fig. 6, a suffi-
ciently large L value will ensure that the results of RCCM converge
to an accurate value. However, a large L value will also result in
a heavy computational burden. RCCM also adopts other methods,
like the k–d tree algorithm, to accelerate the process of finding the
nearest points, which helps to reduce the overall computation time,
as shown in the right graph of Fig. 6. Although the time cost also
increases, it growth is linear with increasing L. Compared with the
transfer entropy, RCCM is much faster. In our program, the transfer
entropy takes tens of seconds to obtain the result at each time delay
point (L = 500, embedding dimensions k = 1 and l = 1, and prediction
horizon h = 20).
2.5. Computational procedure of RCCM
The complete flowchart of RCCM is shown in Fig. 7. It adopts
a pairwise detecting procedure, which is also used by the transfer
entropy and other methods.
Before performing the calculation to identify the time delay rela-
tionship between two variables such as x0 and y0 , parameters, such
as the sample length L, time interval dt, and maximum iteration
number N, must be set. The time lags x0 and y0 and the embedding
dimension E should also be calculated.
In the first step, variable x0 will be used to predict y0 . Time
delay detection starts at t1 and finishes at t2 = t1 + Ndt. After each
step of the calculation, the prediction variable x0 will be translated
toward the positive direction of time t for a time interval dt, and
then Mx0 will be reconstructed again and input into the calculation.
When the iteration number reaches N, the results are recorded as
x0 →y0 ,t1 +ndt , where n = 1, 2, . . ., N, and the two variables x0 and y0
are exchanged and the time-delay detecting calculation is repeated
starting at −t2 and finishing at −t1 . In this step, after each calcu-
lation, the prediction variable y0 will move toward the negative
direction of time by −dt. The results of the second step are recorded
as y0 →x0 ,−t1 −ndt , where n = 1, 2, . . ., N. Finally, the cross-mapping
values at each time delays are
x0 →y0 ,t1 +ndt + y0 →x0 ,−t1 −ndt
¯ x0 →y0 ,t1 +ndt =
2 Tfl and cA , cB , or cc , peaks appear when the time delay equals −1.
3.1. Example 1: a CSTR model
Before further application, we tested RCCM on some simple
examples to verify its applicability to chemical processes and to
show its advantages over other methods, such as CCF and the trans-
fer entropy.
Example 1 is from the paper of Kuehnert and Beyerer (2014).
This example is for a continuously stirred tank reactor, in which
reactant A reacts to give immediate product B and then product
C. The concentrations of the reactants are cA , cB , and cC , the feed
concentration is cin , and the fluid temperature is Tfl . The differential
equations of the CSTR reaction model are
F
ċA (t) = − k1 cA e−E1 /RTfl (22)
V (cin (t) − cA (t))
F
ċB (t) = k1 cA e−E1 /(RTfl ) − k2 cB e−E2 /(RTfl ) − (23)
VcB (t)
F
ċC (t) = k2 cB e−E2 /(RTfl ) − (24)
VcC (t)
The relevant parameters are given in Table 1.
The mean values of cin and Tfl are 1 mol/L and 350 K, respectively.
cin is superposed with white noise having N(0,3 K2 ) and Tfl is super-
imposed with white noise having N(0, 0.1(mol/L)2 ). All of the inputs
and outputs are shown in Fig. 8. As the disturbance propagates
toward the downstream variables, the outputs become smoother
and cB and cC are strongly related, while cA is negatively correlated
with these two variables. There is no other information provided
in this figure and a calculation based on RCCM was performed to
reveal the potential relations between the variables.
According to the flowchart in Fig. 7, parameters  and E should be
appropriately determined before the main calculation. Fig. 9 shows
that the downstream variables in example 1 have greater time lags
than those of upper streams, which is in agreement with the out-
puts in Fig. 8, and all of the curves of A(E) are almost the same
and converge to 1 after E = 7. Therefore, in this model, the best E
value for each variable is chosen as E = 7 and relevant computational
parameters are summarized in Table 2.
Table 3 shows the results of RCCM for this example. In each dia-
gram, the x axis is the integration time with unit dt = 0.1, while the y
axis is the cross-mapping ability ¯ x→y . In this table, the variables in
the columns denote the predicted y variables, while the variables in
the rows are the predicting x variables. There is no significant value
in the whole delay time interval when including Tfl and cin in the
calculation, which indicates that there is no relationship between
these two variables. When calculating the relationship between
(21)
 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16 7

Fig. 6. Variation of y→x and the time cost with the sample length.

Table 3
RCCM results of the CSTR model (x is the predicting variable and y is the predicted variable).

x y

cin cA cB cC

Tfl

cin −

cA − −

cB − − −

According to Fig. 3, negative time delays indicate that disturbance
passes from the prediction variable Tfl to cA , cB and cc after a time
unit dt. In other words, these three variables are all influenced by Tfl
and the corresponding time delay is dt. However, when including
cin in the calculation with cA , cB and cc , only cA is influenced by the
disturbance of cin and the time delay is also −dt, while cB and cc do
not affect cin . This table also shows that cA , cB and cc all simultane-
ously vary and there is no time delay relationship between them.
The results are summarized in Fig. 10, which are in agreement with
the structure of Eqs. (22)–(24), because the feed cin only appears in
the first equation but the temperature Tfl has a direct influence on
cA , cB and cc .
Fig. 10 shows the detected disturbance propagation paths in
the CSTR model. The first number in the parentheses is the cross
mapping ability when using a downstream variable to predict an
upstream variable, which also indicates how much the downstream
variable is affected by the predicted upstream variable. The sec-
ond number in the parentheses is the corresponding time delay
between these two variables.
Example 1 shows that for a simple nonlinear CSTR model, RCCM
can easily determine accurate time lags and embedding dimensions
for each variable with the aid of the linear autocorrelation function
and Cao’s FNN method. It then determines the correct root cause of
the disturbance in a dynamic system as well as the strengths of the
upstream variables influencing the downstream variables. Because
8 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16

Fig. 7. Flowchart of RCCM.

the other examples given in previous CCM papers mainly focus on
ecosystems or other discrete systems, this example shows that the
principle of CCM is also suitable for continuous processes, such as
chemical processes, after some necessary modifications.
3.2. Example 2: time delay detection for four SISO systems
Stockmann and Haber (2010) evaluated the performance of dif-
ferent time-delay detecting methods for four linear and nonlinear
SISO systems. In this part, RCCM will be added to the calculation
and its performance for time delay detection will be evaluated and
compared with other methods. Schematic diagrams of each sys-
tem and their corresponding outputs are shown in Figs. 11 and 12,
respectively.
In all of the simulations, the input u(k) randomly varies, between
−1 and 1. White noise ε(k) with a mean value of 0 and a standard
deviation 0.05 is superposed on the output. Table 4 lists the final
results of the four systems detected by RCCM. It shows that this
method is suitable for both linear and nonlinear systems and all of
the time-delay detecting results are equal to the true values.
Table 5 also compares the results of different methods: the ARX
method, the CCF, the transfer entropy, the KNN method, and RCCM,
where some results are from Stockmann et al. (2012).
From Table 5, among these five time-delay detection methods,
the ARX method and CCF are only suitable methods for linear sys-
tems. The transfer entropy can determine the correct time delays
for both types of systems with the aid of manual parameter adjust-
ment, so it is difficult to apply this method to fully automated
 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16 9

Fig. 8. Inputs and outputs of the CSTR model.

Fig. 9. RCCM calculation parameters of the CSTR model.

Fig. 10. Detected disturbance propagation paths in the CSTR model.
detection. The KNN method and RCMM are the only two methods
that are adequate for both linear and nonlinear systems.
The KNN method and RCCM both have similar principles. They
attempt to find the nearest approximation to a specific input and
then compare the predicted output with the actual output. If the
embedding dimension of RCCM decreases to 1, it will degenerate
to the KNN method to some extent. However, some papers have
shown that the estimation of the nearest approximation in the
KNN method may be incorrect for some high dimensional systems
with strong nonlinear characteristics, which is called the false near-
est neighbor problem (Krakovská and Mezeiová, 2015; Rhodes and
Morari, 1997). This problem can be solved by an appropriate esti-
mation of E in the RCCM procedure, while the KNN method cannot
avoid this problem and may be inaccurate for some high dimen-
sional nonlinear cases. In the next example, a significant difference
between these two methods will be shown.
10 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16

Fig. 11. Process types. The sample time is t = 1 in all of the simulations and x(k)q−d = x(k − d).

Fig. 12. Outputs of the four SISO systems.

Table 4
Time-delay detection results of four different SISO systems calculated by RCCM (L = 800).

Linear static SISO Linear dynamic SISO Nonlinear static SISO Nonlinear dynamic SISO
 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16 11

In general, example 2 shows the advantages of RCCM over other Table 5

Results of time delay analysis.
methods, such as the ARX method, CCF, and transfer entropy, in
time delay detection for both linear and nonlinear systems, and it Process Method
can also be applied to automated detection.
ARX CCF Transfer KNNb RCCMc True
entropy
3.3. Example 3: a benchmark process with unmeasured entry and
Linear static SISO 5 5 5a 5 5 5
nonlinear static subprocesses Linear dynamic SISO 5 5 6a 5 6 6
Nonlinear static SISO 3 3 5a 5 5 5
Fig. 13 shows a benchmark process taken from Stockmann et al. Nonlinear dynamic SISO 1 −4 5a 5 6 6
(2012), which is a generalized plant consisting of the unmeasured a
Manual adjustment must be performed for the results.
entry S0 and five sub-processes S1 − S5 . Many industrial processes b
kimp = 4.
c
are believed to be able to be expressed by this generalized scheme, All of the time lags are set to 1 and E = 8.
although the number of subprocesses can be different. Because of
the existence of a loop in the flowchart from S5 to S1 , a distur- Table 6
bance arising in one subprocess will eventually lead to plant-wide Parameters of each subprocess.
fluctuation. Here, S0 is uniformly distributed between −1 and 1 Parameter Number
and all of the outputs of subprocesses S1 − S5 are superposed with
white noise with a mean value of 0 and a standard deviation of 1 2 3 4 5

0.1. Subprocesses S1 − S5 are all in accordance with nonlinear static Ai 0.2 0.2 0.2 2 6
functions: Bi 0.8 0.8 0.8 0.8 0.8

Ai q2
Si = , i = 1, 2, . . ., 4 (25)
1 − Bi q2
where parameters Ai and Bi of each subprocess are listed in Table 6.
Table 7 lists the final results of example 3, where L = 1000.
The diagrams on the main diagonal show the significant effects of
the upstream processes on the downstream processes, and these
effects rapidly decrease in the following processes because of the
disturbances introduced by ε in each output of Si . Diagrams S1 − S5
and S2 − S5 reveal the existence of a loop in the flowchart, because
in each diagram, different peaks are simultaneously present at posi-
tions with positive and negative time delay. Table 8 gives all of the
detected time delays and the corresponding mapping abilities. The
results indicate a circle path of disturbance propagation, which is
in accordance with the parameters in Fig. 13.
The CCF, KNN method, and transfer entropy were also consid-
ered, and Table 9 lists their final results when including the output
of S1 in the calculation with S2 − S5 . It shows that the former two
methods cannot obtain the correct time delays. Although there
is a significant peak when using CCF to calculate the relationship

Table 7
Results of example 3 calculated by RCCM.

x y

S2 S3 S4 S5

S1

S2 –

S3 – –

S4 – – −
12 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16
Fig. 13. Benchmark process with unmeasured entry S0 and five nonlinear static subprocesses.
Table 8
Time-delay detection results of example 3 (x→y , tdelay ).
x y
S2 S3 S4 S5
S1 (0.18,−2) (0.072,−6) (0.063,−8) (0.10,−12),(0.22,6)
S2 – (0.51,−4) (0.15,−6) (0.092,−10),(0.074,8)
S3 – – (0.60,-2) (0.16,−6)
S4 – – – (0.19,−4)
between S1 and S4 , its position is not in agreement with the model
in Fig. 13. The KNN method loses its sensitivity in this case and all of
the results are almost the same. Only the transfer entropy can pro-
duce as accurate time delay results as RCCM (Table 9). However, it
takes nearly 10 min to obtain the results of one figure, while RCCM
takes only 5 s. For the transfer entropy, to find the time-delay peak,
L = 500, the embedding dimensions of x and y are k = 1 and l = 1, and
h varies from −14 to 14. This example shows that both RCCM and
the transfer entropy are reliable approaches for time-delay analy-
sis of nonlinear systems, but RCCM is much faster than the transfer
entropy.
In conclusion, we have used examples 1–3 to show that RCCM
is suitable for detection of disturbance-propagation paths in both
linear and nonlinear systems. With the aid of the linear autocorre-
lation function and FNN method, the calculation parameters  and
E can be easily determined and give accurate time delay results,
which are in agreement with the parameters given in the model. In
the next section, to testify its applicability for chemical processes,
RCCM will be applied to a benchmark example in chemical engi-
neering with a much more complex structure than the above cases:
the Tennessee Eastman process.
3.4. Case study of the Tennessee Eastman process
The Tennessee Eastman process is a model proposed by Downs
and Vogel (1993) based on an actual chemical process. It is a bench-
mark process for fault detection and diagnosis. Its flowchart is
shown in Fig. 14. There are five major unit operations in this pro-
cess: a reactor, a product condenser, a vapor–liquid separator, a
recycle compressor, and a product stripper. First, reactants A, D,
and E are mixed and fed into the reactor, generating liquid prod-
ucts G and E. The products are then cooled through the condenser
and flow into the vapor–liquid separator. Some of the vapor prod-
uct is compressed and recycled to the reactor and the rest is purged
to avoid accumulation of inert product B and byproducts. The liq-
uid product of the separator is pumped into a striper to strip the
reactants and deliver them back to the reactor. The products of the
stripper (G and H) are then sent to downstream units.
The data used in this case was generated by the Matlab program
“Mode 1 Init”.1 This program was realized in Matlab Simulink and
the data was generated with normal operating conditions (sample
time 1.8 s and time length 720 h). The model has 41 measured vari-
ables and 12 manipulated variables. However, in this case, only the
measurements with “XMEAS” are considered and variables 23 to
41 are excluded because of their long sampling intervals. There-
fore, only measurements 1 to 22, which are process variables such
as flowrates, the temperature, the pressure, and the liquid level,
are considered (Table 10). The threshold for the significant RCCM
mapping ability is set to ¯ x→y = 0.5 and E = 6. The time lag of each
variable is shown in Fig. 16.
The actual outputs of XMEAS (1)–(22) are listed in Fig. 15. In
each graph, the x axis is the time in hours. Although some of the
variables are easily determined to have causal relations with other
variables, it is still necessary to use RCCM to determine the hidden
plant-wide propagation paths of the disturbance.
After eliminating redundant relations, most of the variables
have significant time-delay relationships with each other and some
underlying loops also appear (Fig. 16). For example, the reactor feed
rate, which is influenced by the reactor cooling water outlet tem-
perature and the feed rate of reactant A, has a significant effect on
the reactor level, compressor work, and recycle flow. The compres-
sor recycle valve can also be manipulated to control the reactor feed
rate to maintain a steady reactor level.
Compared with the transfer entropy (Shu and Zhao, 2013),
RCCM can find more information hidden in the Tennessee East-
man process, especially some potential loops in the system. The
final result shows that when operating at the normal steady state,
the disturbance of the whole plant is probably caused by some inlet
streams, such as the feed rates of A (stream 1 in Fig. 14) and A and
C (stream 4 in Fig. 14). This is because when these feed rates are
used to perform a RCCM calculation with other variables, the max-
imum value is always obtained at the negative time-delay interval,
which means that they tend to be the upstream variables of other
variables and there are always pathways from these two sources to
other variables. However, some downstream variables also affect
the upstream variables because of the feed-back control loop.
It should be point out that some causal relations that should
exist according to the inference based on the flowchart are not
captured by RCCM. For example, the stripper level, which should
have a significant effect on the product flow rate, actually has lit-
tle relation with stream 11 in the results of RCCM. The value of
the cross-mapping ability calculated from the stripper level to the
product stream (stream 11) is only about 0.4, which is below the
threshold of 0.5 and is thus rejected by the program. The reason for
1
http://depts.washington.edu/control/LARRY/TE/download.html.
 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16 13

Table 9
Comparison of the different methods (the x axis is the time delay and the y axis is the detected result).

Methods Variables

x : S1 , y : S2 x : S1 , y : S3 x : S1 , y : S4 x : S1 , y : S5

CCF

KNN

Transfer entropy

Fig. 14. Flowchart of the Tennessee Eastman process.

this detection failure is mainly the noise existing in the output of
stream 11.
This benchmark example shows the effectiveness of RCCM for
complex nonlinear chemical processes. However, there are still
some limitations for its application. As pointed out by Ye et al.
(2015), this method is more suitable for systems with weak to mod-
erate coupling strengths. For systems under strong unidirectional
force, it can lead to the “generalized synchrony” phenomenon,
which will make RCCM unable to distinguish between bidirectional
and unidirectional causalities. Furthermore, our calculations also
find that a large value of E will decrease the significance of the
results and cause redundant peaks during detection, which can
increase the difficulty of causal-relationship extraction. We will pay
more attention to these problems in future work.
14 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16

Table 10
Variables used in the Tennessee Eastman case study.

Variable name Variable number Variable name Variable number

A feed (stream 1) XMEAS(1) Product separator level XMEAS(12)

D feed (stream 2) XMEAS(2) Product separator pressure XMEAS(13)
E feed (stream 3) XMEAS(3) Product separator under flow (stream 10) XMEAS(14)
A and C feed (stream 4) XMEAS(4) Stripper level XMEAS(15)
Recycle flow (stream 8) XMEAS(5) Striper pressure XMEAS(16)
Reactor feed rate (steam 6) XMEAS(6) Stripper underflow (stream 11) XMEAS(17)
Reactor pressure XMEAS(7) Stripper temperature XMEAS(18)
Reactor level XMEAS(8) Stripper steam flow XMEAS(19)
Reactor temperature XMEAS(9) Compressor work XMEAS(20)
Purge rate (stream 9) XMEAS(10) Reactor cooling water outlet temperature XMEAS(21)
Product separator temperature XMEAS(11) Separator cooling water outlet temperature XMEAS(22)

Fig. 15. Outputs XMEAS (1)–(22) of the Tennessee Eastman process (the x-axis is the sample number).
 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16
Fig. 16. Time delay results of the Tennessee Eastman process.
4. Conclusion
We have proposed RCCM. RCCM includes some refinements
of the original CCM procedure and integrates embedding param-
eter determinations, such as linear autocorrelation and the FNN
method. These modifications make RCCM more automatic and
faster in determining the disturbance propagation paths in chemi-
cal processes. We also tested RCCM for some examples to show its
applicability for both linear and nonlinear systems. RCCM can more
accurately and rapidly obtain the time delays compared with other
existing methods, such as the CCF and transfer entropy. We also
applied RCCM to the Tennessee Eastman process to verify its appli-
cability for complex nonlinear processes. When the whole system
is operated normally without faults, the plant-wide disturbances
are the result of two inlet feed streams of the reactants (feed A
and feed A and C). Some problems still need to be overcome, such
as the detection sensitivity and time-delay relationship extraction,
which will be the focus of future research. We look forward to fur-
ther application of RCCM to larger and more complex chemical
processes and other fields.
Acknowledgments
The authors gratefully acknowledge the National Natural Sci-
ence Foundation of China for its financial support (Grant No.
U1462206).
References
Bauer, M., Cox, J.W., Caveness, M.H., Downs, J.J., Thornhill, N.F., 2007. Finding the
direction of disturbance propagation in a chemical process using transfer
15
entropy. IEEE Trans. Control Syst. Technol. 15, 12–21, http://dx.doi.org/10.
1109/TCST. 2006.883234.
Bauer, M., Thornhill, N.F., 2008. A practical method for identifying the propagation
path of plant-wide disturbances. J. Process Control 18, 707–719, http://dx.doi.
org/10.1016/j.jprocont.2007.11.007.
Cao, L., 1997. Practical method for determining the minimum embedding
dimension of a scalar time series. Physica D 110, 43–50, http://dx.doi.org/10.
1016/S0167-2789(97)00118-8.
Downs, J.J., Vogel, E.F., 1993. A plant-wide industrial process control problem.
Comput. Chem. Eng. 17, 245–255.
Krakovská, A., Mezeiová, K.I.H.B., 2015. Use of false nearest neighbours for selecting
variables and embedding parameters for state space reconstruction, 2015.
Kuehnert, C., Beyerer, J., 2014. Data-driven methods for the detection of causal
structures in process technology. Machines 2, 255–274, http://dx.doi.org/10.
3390/machines2040255.
Luo, L., Cheng, F., Qiu, T., Zhao, J., 2016. An improved convergent cross mapping
algorithm for causality identification and time delay analysis between
systemic variables under external disturbance. CIESC J 67, 5122–5130, http://
dx.doi.org/10.11949/j.issn.0438-1157.20161300.
Rhodes, C., Morari, M., 1997. False-nearest-neighbors algorithm and
noise-corrupted time series. Phys. Rev. E 55, 6162–6170, http://dx.doi.org/10.
1103/PhysRevE.55.6162.
Schreiber, T., 2000. Measuring information transfer. Phys. Rev. Lett. 85, 461–464,
http://dx.doi.org/10.1103/PhysRevLett.85.461.
Shu, Y., Zhao, J., 2013. Data-driven causal inference based on a modified transfer
entropy. Comput. Chem. Eng. 57, 173–180, http://dx.doi.org/10.1016/j.
compchemeng.2013.05.011.
Shun, Y., Xiaoqi, L., Changzhi, Z., 2015. Determining the sigular spectrum
embedding dimension based on an improved Cao algorithm. Eng. Surv. Mapp.
24, 64–68.
Stockmann, M., Haber, R., 2010. Determination of fault propagation by time delay
estimation using k nearest neighbour imputation, 2010 Conf. Control
Fault-Tolerant Syst., 185–190, doi:10.1109/SYSTOL. 2010.5676081.
Stockmann, M., Haber, R., Schmitz, U., 2012. Source identification of plant-wide
faults based on k nearest neighbor time delay estimation. J. Process Control 22,
583–598.
Sugihara, G., May, R., Ye, H., Hsieh, C.-h., Deyle, E., Fogarty, M., Munch, S., 2012.
Detecting causality in complex ecosystems. Science 338, 496–500, http://dx.
doi.org/10.1126/science.1227079.
16 L. Luo et al. / Computers and Chemical Engineering 106 (2017) 1–16
Vicente, R., Wibral, M., Lindner, M., Pipa, G., 2011. Transfer entropy-a model-free
measure of effective connectivity for the neurosciences. J. Comput. Neurosci.
30, 45–67.
Xie, Z., LIU, G., WU, Z., 2012. Dynamic damage identification for beam structures
based on transfer entropy. J. Zhejiang Univ. (Engineering Sci. 46).
Yang, F., Shah, S.L., Xiao, D., 2010. Signed directed graph modeling of industrial
processes and their validation by data-based methods. Conf. Control
Fault-Tolerant Syst. SysTol’10 – Final Progr. B. Abstr., 387–392, http://dx.doi.
org/10.1109/SYSTOL. 2010.5676059.
Ye, H., Deyle, E.R., Gilarranz, L.J., Sugihara, G., 2015. Distinguishing time-delayed
causal interactions using convergent cross mapping. Sci. Rep. 5, 14750, http://
dx.doi.org/10.1038/srep14750.
Zhang, Z.S., Gong, Z., Qiang, Zhi, R., 2013. Analysis of the direction of information
transfer of Lorenz system and Walker circulation with transfer entropy. Acta
Phys. Sin., 62.