Professional Documents
Culture Documents
4531
Advances in instrumentation and data storage technologies have allowed the process industries to collect
extensive operating data which can be used to extract information about the underlying process and provide
online decision support. One of the fundamental problems in data-based decision support is comparison of
time-series data. Many signal comparison methods require signals that are of the same length and synchronized.
Synchronization of varying length signals is usually achieved using dynamic time warping (DTW). Major
limitations of DTW include computational cost and the tendency to link operationally different points.
Previously, we proposed singular points augmented time warping to overcome these shortcomings during
offline signal comparison (Srinivasan and Qian Ind. Eng. Chem. Res. 2005, 44, 4697). Landmarks in process
data such as extreme values and sharp changes, called singular points, are used to segment a signal into
regions with homogeneous properties, called episodes. Singular points of two signals are linked by dynamic
programming; time warping is used to synchronize the episodes. A locally optimal equivalent of DTW called
extrapolative time warping (XTW) with better computational performance was also proposed. In this paper,
we present the extension of this approach to online signal comparison. The online signal comparison problem
is a generalization of the offline problem and has two additional challenges: (1) the reference signal for
comparison is not known a priori and has to be selected from a library, and (2) the starting point of the
reference and real-time signal would not coincide in general, and the corresponding points have to be identified.
The approach proposed here addresses these by extending dynamic locus analysis (Srinivasan and Qian Chem.
Eng. Sci. 2006, 61, 6109) and singular point augmented XTW using anchoring and flanking strategies. The
application of the proposed approach is illustrated using two different case studies: online operating mode
identification in the Tennessee Eastman process and online fault identification in a lab-scale distillation column.
The results show that the proposed approach is robust, efficient, and suitable for online signal comparison.
1. Introduction
As a result of significant advances in data collection and
storage, vast amounts of historical operating data are becoming
commonly available in the chemical process industry. This
data is a rich source of information about the process that can
be used to improve the plant operation. Given the parallel
developments in pattern classification3 and statistical, information, and systems theories,4 data-based approaches have been
gaining in popularity. Potential areas of application of datadriven methods include regulatory and sequence control,
visualization, operation improvement, state identification, and
fault diagnosis. Despite these developments in extracting
information and knowledge from data, many important and
challenging problems persist in data analysis and knowledge
extraction. In this paper, we address one such problem: online
temporal signal comparison.
Temporal signals with time-varying properties commonly
arise in chemical plants during transition states such as startup,
grade change, shutdown, maintenance, and other abnormal
operations. The precept of signal comparison-based approaches
is that similar states result in similar temporal signals. So, if a
historical database of representative signals is available, the root
cause of a process change occurring in real-time can be
* Corresponding author. E-mail: chergs@nus.edu.sg. Tel.: +65
65168041. Fax: +65 67791936.
4532
(1)
(2)
(3)
(4)
c(1) ) (1, 1)
(5)
c(p) ) (r, t)
(6)
Ind. Eng. Chem. Res., Vol. 46, No. 13, 2007 4533
DA(i, j) )
DA(i - 1, j) + d(i, j) or [ if condition (A*)]
min DA(i - 1, j - 1) + d(i, j)
DA(i - 1, j - 2) + d(i, j)
(7)
4534
SP
but m > l (that is, RSP
m does not temporally precede Rl ). As
described in detail in Srinivasan and Qian,1 the singular points
of two signals are said to correspond and can be linked if they
are of the same type and if the linkage does not result in any
sequence violations. For a given linkage of singular points, the
distance between the two signals is calculated as the sum of
episode-wise distances. Time warping is used to calculate the
distance between the corresponding episodes so as to account
for magnitude and duration differences between the episodes.
Among the various linkages possible between two singular point
sequences, the linkage that results in minimum signal distance
is considered optimal.
Srinivasan and Qian1 also proposed a new time warping
strategy called extrapolative time warping (XTW), which is a
greedy search modification of classical DTW with Itakura local
constraint. The XTW method obviates dynamic programming
for each local point by optimizing each point locally. In contrast
to DTW, search in XTW proceeds in the forward direction
starting from the first point of the signal to the last. Given the
warping assignment (i, j), the optimal warping for the subsequent
step, that is, the location of j* that corresponds to (i + 1), is
based only on the previous decision and the current distance.
Ind. Eng. Chem. Res., Vol. 46, No. 13, 2007 4535
Figure 5. Temporal development and translation of flanking segments during optimal reference signal identification.
variable
number
base case
value
units
XMEAS (1)
XMEAS (2)
XMEAS (3)
XMEAS (4)
XMEAS (5)
XMEAS (6)
XMEAS (7)
XMEAS (8)
XMEAS (9)
XMEAS (10)
XMEAS (11)
0.25052
3664.0
4509.3
9.3477
26.902
42.339
2705.0
75.0
120.40
0.33712
80.109
kscmh
kg h-1
kg h-1
kscmh
kscmh
kscmh
kPa gauge
%
C
kscmh
C
XMEAS (12)
XMEAS (13)
XMEAS (14)
50.000
2633.7
25.160
%
kPa gauge
m3 h-1
XMEAS (15)
XMEAS (16)
XMEAS (17)
50.000
3102.2
22.949
%
kPa gauge
m3 h-1
XMEAS (18)
XMEAS (19)
XMEAS (20)
XMEAS (21)
65.731
230.31
341.43
94.599
XMEAS (22)
77.297
C
kg h-1
kW
C
C
4536
Figure 6. Test signal T and reference signals (R1 and R2) for the illustrative example.
(9)
Also, yl* is called the corresponding point of x1, and yj* is the
corresponding point of xm. The brute force approach of
considering each possible l in Y and performing an independent
comparison will result in an unacceptable computational load.
DLA overcomes this by extending Smith and Watermans18
dynamic programming approach for comparing protein sequences. In DLA, the locus of X is identified by using a
dissimilarity matrix, DS. Let i and j be the time indices of X
and Y, respectively. The (i, j) element of DS measures the
minimal difference between the sub-segment {x1, x2, x3, ..., xi}
in X and the sub-segment {yl, yl+1, yl+2, ..., yj} in Y. In the general
case, l is unknown and is determined using dynamic programming.
i
DS(i, j) ) min{
F
(xd, yj(d))}
d)1
(10)
where yj(d) is the time warped point that matches with xd and
(xd, yj(d)) ) |yj(d) - xd| is the difference between xd and yj(d).
Because the optimal search should allow for compression and
elongation in Y relative to X, time warping is used to synchronize
X and Y. Following DTW with Itakura local constraint, eq 10
reduces to
DS(i, j) ) min{DS(i - 1, j - 1) +
(xi, yj), Fi,j, Gi,j}i [2 m] j [2 n] (11)
Fi,j ) DS(i - 1, j - 2) + (xi, yj)
Gi,j ) DS(i - 1, j) + (xi, yj) or if G*
where G* indicates that the predecessor of point (i - 1, j) is
the point (i - 2, j). Note that DS(i, j) is not the total minimum
distance between {x1, x2, x3, ..., xi} and{y1, y2, ..., yj}, rather it
is the total minimum distance between X and its locus in Y.
Segments of Y that are similar to X would lead to small values
of DS. The optimal match between X and the locus in Y is given
by DS(m, j*) where j* ) argminj{DS(m, j)} j [1 n]. More
details of DLA can be found in Srinivasan and Qian.2 In this
paper, we extend DLA and singular point augmented time
warping for online signal comparison.
3. Online Signal Comparison Using Singular Points
Augmented Time Warping
The online signal comparison problem can be stated as
follows: Given a set of reference signals K and a real-time signal
T emanating from the process operating at an unknown state,
(1) identify the reference signal that best matches the current
state of the process and (2) identify the progress of the process
Ind. Eng. Chem. Res., Vol. 46, No. 13, 2007 4537
Figure 7. Comparison of real-time signal T at T ) 8 with R1 (shown in part b) reveals a minimum at R1 ) 199 (shown in part c). Similar comparison with
R2 depicted in part d shows a minimum at R2 ) 1083 as shown in part e.
as its first points, that is, XA ) {x1, x2, ..., x} where is the
flank length. Similarly, the posterior flanking segment of X is
defined as the last points, that is, XP ) {xm-+1, xm-+2, ...,
xm}. The inner m - 2 points of X comprise the core segment
as illustrated in Figure 4 for a sample signal.
The flanking segments of the signal can be used to identify
the locus of X efficiently based on the recognition that any locii
of X should also have segments that have high similarity with
the flanking segments XA and XP. The flanking strategy exploits
this property. Given a long reference signal Y, all the segments
in Y, say YA, that closely match XA can be identified. Similarly,
all adequate matches for XP, say YP, can also be identified. Note
that the lengths of YA and YP may not be equal to the flank
length due to run-to-run variations between the real-time and
the reference signals. Each pair of YA and YP where YA precedes
YP in Y can be used to construct a unique composite segment Z
sandwiched by YA and YP, Z Y. Each composite segment is
a possible locus of X. By eq 9, the locus of X in Y is the
composite segment Z* which has the least difference with T.
The flanking strategy is thus a generalization of the DLA for
identifying the locus of a longer signal. It recognizes that the
computational complexity of any signal comparison method
depends directly on the length of the two signals to be compared.
The flanking strategy is computational efficient because the two
flanking segments are short; therefore, the cost of the first phase
of comparisons with all the reference signals is small and
4538
Ind. Eng. Chem. Res., Vol. 46, No. 13, 2007 4539
Figure 10. Comparison of real-time signal T at T ) 685 with R1 (shown in part b) reveals a minima at R1 ) 185 (shown in part c). Similar comparison
with R2 depicted in part d shows a minimum at R2 ) 1087 as shown in part e.
observations {x1, ..., xtXSP-1} before the anchor point and the
ones after, {xtXSP, ..., xm}. Similarly, Y is split by the anchor
point into two segments: {y1, ..., ytYSP-1} and {ytXSP, ..., yn}.
Denoting the difference between X and Y by (X, Y),
(12)
4540
Figure 11. Schematic of the Tennessee Eastman19 process with control system.
Table 2. Disturbance Profiles for TE Process (a) XD1 (b) XD2 (c) XD3 (d) XD4 (e) XD5.
target
time
(min)
target
XD1-A
XD1-B
XD1-C
1.20base value
1.15base value
1.10base value
180
240
300
1.40base value
1.35base value
1.30base value
XD2-A
XD2-B
XD2-C
1.03base value
1.025base value
1.02base value
180
240
300
XD3-A
XD3-B
XD3-C
1.05base value
1.045base value
1.04base value
XD3-A
XD3-B
XD3-C
XD5-A
XD5-B
XD5-C
time
(min)
target
time
(min)
target
time
(min)
(a) XD1
190
254
318
1.60base value
1.55base value
1.50base value
200
268
336
1.0base value
1.0base value
1.0base value
780
900
1020
1.05base value
1.045base value
1.04base value
(b) XD2
190
254
318
1.07base value
1.065base value
1.06base value
200
268
336
1.0base value
1.0base value
1.0base value
1020
1080
1200
180
240
300
1.10base value
1.09base value
1.08base value
(c) XD3
190
254
318
1.15base value
1.135base value
1.12base value
200
268
336
1.0base value
1.0base value
1.0base value
780
900
1020
1.05base value
1.045base value
1.04base value
180
240
300
1.10base value
1.09base value
1.08base value
(d) XD4
190
254
318
1.15base value
1.135base value
1.12base value
200
268
336
1.0base value
1.0base value
1.0base value
780
900
1020
0.95base value
0.955base value
0.96base value
180
240
300
0.90base value
0.91base value
0.92base value
(e) XD5
190
254
318
0.85base value
0.865base value
0.88base value
200
268
336
1.0base value
1.0base value
1.0base value
780
900
1020
Ind. Eng. Chem. Res., Vol. 46, No. 13, 2007 4541
Figure 12. Three runs of XD1 in the TE case study with different magnitudes and duration.
(Tw, R/w) )
|rj(i) - ti|
m-
SP
T
i)TSP
(14)
((Tw, R/w) < max) and ((tm, rR*) < 2 max) (16)
Here, (Tw, R/w) is the difference between Tw and R/w while (tm, rR*) is the distance between tm and the base point as
calculated using XTWSP. Note that (Tw, R/w) e (Tw, R/w)
because XTWSP relies on singular point linkage. If eq 16 is
satisfied, the process is considered to continue in the same state
(i.e., no change to the reference signal), and the base point is
updated. One byproduct of performing XTWSP is that new
singular points could have been identified in Tw and linked with
R*. The last singular point in T and its corresponding linked
singular point in R* are subsequently used as the new anchor
point which results in the shortening of the evaluation window.
So, future Step A and Step B calculations become more efficient
and accurate. If condition 16 is not satisfied, the process is
considered to have moved to a new state, and another optimal
+1
4542
TD )
if m g POD
+
T
(17)
(18)
R)
(TD, R*)
min
RK,R*R*
((TD, R))
(19)
(20)
Ind. Eng. Chem. Res., Vol. 46, No. 13, 2007 4543
Figure 14. Process signals from Run-03 of the lab-scale distillation column.
(TD, R) )
|rj(i) - ti|
i)m-+1
(21)
(TD, R) )
|rj(i) - ti|
i)TPOD
m - POD
+1
T
(22)
4544
Table 3. Offline Signal Difference between Different Disturbance Instances in the TE Process Using Direct Comparison (10-1)
XD1-A
XD1-B
XD1-C
XD2-A
XD2-B
XD2-C
XD3-A
XD3-B
XD3-C
XD4-A
XD4-B
XD4-C
XD5-A
XD5-B
XD5-C
XD1-A
XD1-B
XD1-C
XD2-A
XD2-B
XD2-C
XD3-A
XD3-B
XD3-C
XD4-A
XD4-B
XD4-C
XD5-A
XD5-B
XD5-C
0.0686
0
0.074
0.0629
0
0.2228
0.2235
0.2241
0
0.2052
0.2028
0.2009
0.1446
0
0.1983
0.1953
0.1885
0.166
0.11
0
0.1735
0.1735
0.1847
0.2327
0.2631
0.2465
0
0.1647
0.1564
0.1587
0.2309
0.2342
0.2424
0.1991
0
0.161
0.1476
0.1383
0.1892
0.2413
0.2157
0.1816
0.1736
0
0.3631
0.3739
0.3694
0.4558
0.4555
0.4556
0.4143
0.4482
0.4294
0
0.3589
0.3483
0.357
0.4544
0.4219
0.4351
0.4318
0.3953
0.4175
0.2139
0
0.3443
0.3387
0.3278
0.4107
0.4247
0.4001
0.4113
0.4007
0.3697
0.2909
0.2003
0
0.0977
0.1136
0.1114
0.2004
0.1914
0.1875
0.1591
0.176
0.1681
0.3386
0.3616
0.3589
0
0.1108
0.0903
0.1018
0.2015
0.1791
0.1786
0.1718
0.1415
0.1521
0.3605
0.3262
0.337
0.1005
0
0.1054
0.1023
0.0824
0.1739
0.1801
0.1624
0.1626
0.1565
0.124
0.3812
0.3425
0.3063
0.0868
0.0864
0
R1
R3
R4
R5
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
0.0501
0.0592
0.0659
0.0693
0.0723
0.0755
0.0782
0.0779
0.0773
0.0767
0.0760
0.0752
0.0745
0.0737
0.0729
0.0511
0.0593
0.0656
0.0654
0.0604
0.0556
0.0510
0.0506
0.0508
0.0507
0.0506
0.0504
0.0500
0.0505
0.0504
0.0514
0.0589
0.0654
0.0572
0.0408
0.0240
0.0072
0.0070
0.0068
0.0065
0.0064
0.0061
0.0061
0.0060
0.0060
0.0512
0.0581
0.0647
0.0683
0.0702
0.0723
0.0742
0.0739
0.0732
0.0729
0.0723
0.0716
0.0712
0.0707
0.0703
0.9804
0.9864
0.9970
0.8746
0.6755
0.4317
0.1412
0.1383
0.1339
0.1282
0.1265
0.1210
0.1220
0.1188
0.1190
Ind. Eng. Chem. Res., Vol. 46, No. 13, 2007 4545
Table 5. Online Process Disturbance Detection in the TE Process
Run-1
Run-2
Run-3
Run-4
Run-5
Run-6
Run-7
Run-8
Run-9
Run-10
Run-11
Run-12
Run-13
Run-14
Run-15
average
disturbance
introduction
time
disturbance
identification
time
bestmatching
reference
9
9
9
9
9
9
9
9
9
9
9
9
9
9
9
15
13
13
17
13
16
16
13
13
14
13
17
13
13
13
XD1
XD1
XD1
XD2
XD2
XD2
XD3
XD3
XD3
XD4
XD4
XD4
XD5
XD5
XD5
identification
delay
(sample)
second
disturbance
introduction
time
disturbance
identification
time
bestmatching
reference
6
4
4
8
4
7
7
4
4
5
4
8
4
4
4
5.1333
1020
670
680
1030
1100
420
860
970
1100
820
380
400
500
600
1080
1025
675
684
1034
1107
427
867
974
1105
825
387
407
505
604
1084
XD4
XD4
XD4
XD4
XD5
XD5
XD1
XD1
XD1
XD2
XD2
XD2
XD3
XD3
XD3
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
identification
delay
(sample)
average
time
cost
(CPU s)
average
time cost
of DTW
(CPU s)
5
5
4
4
7
7
7
4
5
5
7
7
5
4
4
5.3333
0.0501
0.0699
0.1233
0.0273
0.0734
0.1289
0.0367
0.0689
0.2509
0.0373
0.0535
0.2024
0.0784
0.0712
0.2115
0.0989
55.3446
55.2385
55.2589
55.2721
55.3044
55.3314
55.3578
55.3446
55.3374
55.2985
55.3051
55.2517
55.2919
55.3051
55.2517
55.2996
case
disturbance
type
DST01
DST02
DST03
DST04
DST05
DST06
DST07
DST08
DST09
DST10
step
step
step
step
random variation
step
step
sticking
slow drift
step
studied here (see Table 2). Different instances (runs) of the same
disturbance class have different start times, duration, and
magnitude. For example, during XD1-A, the flowrate of A feed
from upstream is increased from the base case value of 0.25052
kscmh to 0.3902 kscmh (a 60% change) in three steps starting
at t ) 180 min as shown in Table 2a. After the process recovers
from these, the inverse change, decreasing the A feed flow, is
introduced at t ) 780 min. The process is then allowed to return
to a steady state. The effect on the A flow rate (XMEAS(1))
and the downstream pressures (XMEAS(13) and XMEAS (16))
is shown in Figure 12. Two other instances XD1-B and XD1-C
with changes of magnitude of 55% and 50% were also
introduced. As described in Srinivasan et al.8 similar changes
were introduced to bring forth the other disturbance classes.
One instance from each of the classes, XD1-B, XD2-B, XD3B, XD4-B, and XD5-B, was used to develop the reference
disturbance database for online disturbance identification.
The difficulty in identifying the disturbance online can be
estimated by a preliminary analysis. Table 3 shows the difference between the 15 disturbances calculated as the average
difference among signals. Comparison is made after signals have
been normalized to [0 1] on the basis of the range of the sensor
(see Srinivasan et al.8). As can be seen from the table, the
minimum inter-class distance is 0.0824 (between XD1-C and
XD5-C) and the maximum intra-class distance is 0.2909 (class
XD4). Therefore, difference between the classes by direct
comparison, even if the complete signal is available, is a
nontrivial exercise. In this work, we consider the even more
difficult task of differentiating between the disturbances as they
evolve.
The proposed online signal comparison method is used for
online disturbance identification as follows. Consider Run-4
where the process is in state XD2 until T ) 1030. An unknown
disturbance occurs starting at T ) 1030 which is initially
4546
Table 8. Faults Diagnosis Results for Lab-scale Distillation Column Case Study
case
time fault
introduced
(sample)
detection
time
(sample)
Run-01
Run-02
Run-03
Run-04
Run-05
Run-06
Run-07
Run-08
Run-09
Run-10
1
1
359
356
425
350
345
470
1
300
6
6
370
357
426
353
346
472
6
302
detection
delay
(sample)
identification
time
(sample)
5
5
11
1
1
3
1
2
5
2
average
y
(sample)
6
6
371
360
430
355
347
473
6
309
6
6
13
5
6
4
3
4
7
5
3.6
5.9
1
1
359
356
425
350
345
470
1
300
6
6
371
360
430
354
347
473
6
308
identification delay
(sample)
2%
3%
4%
5%
6
6
371
360
430
355
347
473
6
308
6
6
371
360
430
355
347
473
6
309
6
6
371
360
430
355
347
473
6
309
6 5
5
5
5
5
6 5
5
5
5
5
371 12
12
12
12
12
360 4
4
4
4
4
438 5
5
5
5
13
355 4
5
5
5
5
348 2
2
2
2
3
473 3
3
3
3
3
6 5
5
5
5
5
311 8
8
9
9
11
5.3 5.4 5.5 5.5 6.6
1%
2%
3%
4%
5%
best
matching
reference
DST01
DST02
DST03
DST04
DST05
DST06
DST07
DST08
DST09
DST10
0.0162
0.0097
0.0101
0.0217
0.0109
0.0268
0.0319
0.0100
0.0103
0.0119
0.2817
0.1091
0.3456
0.6663
0.4113
0.2379
0.6940
0.2330
0.1582
0.6910
0.0156
0.3828
identification
delay (sample)
5
5
12
4
5
5
2
2
5
9
5.4
time cost
(s)
0.1716
0.1028
0.0317
0.0342
0.0256
0.0368
0.0264
0.0256
0.0556
0.0840
0.0594
Ind. Eng. Chem. Res., Vol. 46, No. 13, 2007 4547
Table 10. Effect of Noise on Disturbance Identification in TE Case Study
identification time (sample)
Run-1
Run-2
Run-3
Run-4
Run-5
Run-6
Run-7
Run-8
Run-9
Run-10
Run-11
Run-12
Run-13
Run-14
Run-15
average
disturbance
introduction time
1%
2%
3%
4%
5%
1%
2%
3%
4%
5%
181
241
301
181
241
301
181
241
301
181
241
301
181
241
301
187
245
305
189
245
308
188
245
305
186
245
309
185
245
305
192
246
306
196
244
316
199
246
317
191
253
314
188
245
313
194
245
313
198
254
317
201
249
318
205
253
314
196
247
315
194
245
313
201
256
323
195
256
326
220
253
315
193
247
314
199
246
313
205
265
316
203
255
329
220
253
314
198
249
316
6
4
4
8
4
7
7
4
4
5
4
8
4
4
4
5.1333
11
5
5
15
3
15
18
5
16
10
12
13
7
4
12
10.0667
13
4
12
17
13
16
20
8
17
24
12
13
15
6
14
13.6000
13
4
12
20
15
22
14
15
25
39
12
14
12
6
13
15.7333
18
5
12
24
24
15
22
14
28
39
12
13
17
8
15
17.7333
Run-1
Run-2
Run-3
Run-4
Run-5
Run-6
Run-7
Run-8
Run-9
Run-10
Run-11
Run-12
Run-13
Run-14
Run-15
average
disturbance
introduction time
Rmax )
0.80
Rmax )
0.75
Rmax )
0.70
Rmax )
0.65
Rmax )
0.60
Rmax )
0.80
Rmax )
0.75
Rmax )
0.70
Rmax )
0.65
Rmax )
0.60
181
241
301
181
241
301
181
241
301
181
241
301
181
241
301
187
245
305
189
245
304
186
245
305
185
245
305
184
245
305
187
245
305
189
245
305
187
245
305
185
245
308
184
245
305
187
245
305
189
245
308
188
245
305
186
245
309
185
245
305
187
245
311
190
245
313
189
249
315
186
247
313
185
245
313
187
245
313
191
267
313
189
249
316
186
247
327
185
282
355
6
4
4
8
4
3
5
4
4
4
4
4
3
4
4
4.3333
6
4
4
8
4
4
6
4
4
4
4
7
3
4
4
4.6667
6
4
4
8
4
7
7
4
4
5
4
8
4
4
4
5.1333
6
4
10
9
4
12
8
8
14
5
6
12
4
4
12
7.8667
6
4
12
10
26
12
8
8
15
5
6
26
4
41
54
15.8000
Table 12. Effect of rmax on Identification Delay in Lab-scale Distillation Column Case Study
identification time (sample)
Run-1
Run-2
Run-3
Run-4
Run-5
Run-6
Run-7
Run-8
Run-9
Run-10
average
Rmax )
0.80
Rmax )
0.75
Rmax )
0.70
Rmax )
0.65
Rmax )
0.60
Rmax )
0.80
Rmax )
0.75
Rmax )
0.70
Rmax )
0.65
Rmax )
0.60
1
1
359
356
425
350
345
470
1
300
6
6
371
359
430
355
347
473
6
308
6
6
371
359
430
355
347
473
6
308
6
6
371
360
430
355
347
473
6
309
6
6
371
360
432
355
347
473
6
310
6
6
371
361
432
355
348
473
6
311
5
5
12
3
5
5
2
3
5
8
5.3
5
5
12
3
5
5
2
3
5
8
5.3
5
5
12
4
5
5
2
3
5
9
5.5
5
5
12
4
7
5
2
3
5
10
5.8
5
5
12
5
7
5
3
3
5
11
6.1
5. Summary
Online signal comparison is important for process monitoring,
fault diagnosis, and process state identification. In this paper,
we have proposed a signal comparison-based strategy for online
disturbance or fault identification. Given a suitably annotated
historical database of process states, normal and abnormal, the
proposed method finds the best matching state at any given time
by comparing the real-time sensor measurements with the signals
in the database. In contrast to signal comparison strategies
reported in literature, which are designed for offline signal
comparison, the proposed method does not require any a priori
knowledge about the online signal; specifically, the beginning
and end of the real-time signal do not need to coincide with
4548
R ) time index in R
T ) time index in T
) time index of point of divergence in T
POD
T
SP
T ) time index of last singular point in T
Literature Cited
(1) Srinivasan, R.; Qian, M. S. Offline temporal signal comparison using
singular points augmented time warping. Ind. Eng. Chem. Res. 2005, 44
(13), 4697-4716.
(2) Srinivasan, R.; Qian, M. S. Online fault diagnosis and state
identification during process transition using dynamic locus analysis. Chem.
Eng. Sci. 2006, 61, 6109-6132.
(3) Webb, A R. Statistical pattern recognition; Wiley: West Sussex,
U.K., 2002.
(4) Chiang, L. H.; Russell, E. L.; Braatz, R. D. Fault detection and
diagnosis in industrial systems; Springer: London, New York, 2001.
(5) Krzanowski, W. J. Between-group comparison of principal components. J. Am. Stat. Assoc. 1979, 74 (367), 703-707.
(6) Raich, A.; Cinar, A. Diagnosis of process disturbances by statistical
distance and angle measures. Comput. Chem. Eng. 1997, 21 (6), 661-673.
(7) Singhal, A.; Seborg, D. E. Pattern matching in historical batch data
using PCA. IEEE Control Syst. Mag. 2002, 22 (5), 53-63.
(8) Srinivasan, R.; Wang, C.; Ho, W. K.; Lim, K. W. Dynamic PCA
based methodology for clustering process states in agile chemical plants.
Ind. Eng. Chem. Res. 2004, 43 (9), 2123-2139.
(9) Cheung, J. T-Y.; Stephanopoulos, G. Representation of process trends
- Part I. A formal representation framework. Comput. Chem. Eng. 1990,
14 (4/5), 495-510.
(10) Bakshi, B. R.; Stephanopoulos, G. Representation of process trends
- IV. Induction of real-time patterns from operating data for diagnosis
and supervisory control. Comput. Chem. Eng. 1994, 18 (4), 303-332.
(11) Rengaswamy, R.; Venkatasubramanian, V. A syntactic pattern
recognition approach for process monitoring and fault diagnosis. Eng. Appl.
Artif. Intell. 1995, 8 (1), 35-51.
(12) Venkatasubramanian, V.; Rengaswamy, R.; Kavuri, S. N.; Yin, K.
A review of process fault detection and diagnosis Part III: Process history
based methods. Comput. Chem. Eng. 2003, 27, 327-346.
(13) Maurya M. R.; Rengaswamy, R.; Venkatasubramanian, V. Fault
diagnosis using dynamic trend analysis: A review and recent developments.
Eng. Appl. Artif. Intell. 2007, 20, 133-146.
(14) Sundarraman, A.; Srinivasan, R. Monitoring transitions in chemical
plants using enhanced trend analysis. Comput. Chem. Eng. 2003, 27 (10),
1455-1472.
(15) Sankoff, D.; Kruskal, J. B. Time Warps, String Edits, and
Macromolecules: The Theory and Practice of Sequence Comparison;
Addison-Wesley: Reading, MA, 1983.
(16) Kassidas, A.; MacGregor, J. F.; Taylor, P. A. Synchronization of
batch trajectories using dynamic time warping. AIChE J. 1998, 44 (4), 864875.
(17) Qian, M. S. Ph.D. thesis, National University of Singapore,
Singapore, 2004.
(18) Waterman, M. S.; Eggert, M. A New Algorithm for Best Subsequence Alignments with Application to tRNA-rRNA Comparisons. J. Mol.
Biol. 1987, 197, 723-728.
(19) Downs, J. J.; Vogel, E. F. A plant-wide industrial process control
problem. Comput. Chem. Eng. 1993, 17 (3), 245-255.
(20) Singhal, A. Tennessee Eastman simulation model. http://www.
chemengr.ucsb.edu/ceweb/computing/TE/tesimulation.htm
(accessed
2003).
(21) McAvoy, T. J.; Ye, N. Base control for the Tennessee Eastman
problem. Comput. Chem. Eng. 1994, 18 (5), 383-413.