Journal of Process Control 82 (2019) 44–57

Contents lists available at ScienceDirect

Journal of Process Control

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

Accelerated multiple alarm flood sequence alignment for abnormality

pattern mining
Shiqi Lai a , Fan Yang b,∗ , Tongwen Chen a , Liang Cao b
a
Department of Electrical and Computer Engineering, University of Alberta, Edmonton T6G 2V4, Canada
b
Beijing National Research Center for Information Science and Technology (BNRist), Department of Automation, 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: Alarm floods can interfere with operators and may therefore cause or aggravate industrial accidents. A
Received 27 March 2017 novel algorithm is proposed for pattern mining in multiple alarm floods. Unlike traditional methods which
Received in revised form 7 May 2019 either cannot deal with multiple sequences with time stamps or suffer from high computational cost,
Accepted 5 June 2019
the computational complexity of this proposed algorithm is reduced significantly by introducing a gen-
Available online 23 August 2019
eralized pairwise sequence alignment method and a progressive multiple sequence alignment approach.
Two types of alignment refinement methods are developed to improve the alignment accuracy. The
Keywords:
effectiveness of the proposed algorithm is tested using a dataset from a real chemical plant.
Alarm flood analysis
Industrial alarm monitoring © 2019 Elsevier Ltd. All rights reserved.
Time-stamped sequences
Multiple sequence alignment
Smith–Waterman algorithm

1. Introduction can only handle the abnormal event based on his/her experience
or even take no actions, likely leading to improper executions for
1.1. Background and motivation important abnormalities. As a result, based on operators’ normal
response time [2], both EEMUA, ISA and IEC standards [3,1,4] rec-
As the scale of industrial plants grows, process monitoring ommend to set the upper limit of alarms announced to an operator
becomes indispensable for ensuring the safety and efficiency of to be 6 alarms/h, and the alarm rate threshold for alarm floods to
operation. An alarm is defined as an audible and/or visible means of be 10 alarms/10 min per operator.
indication to the operator an equipment malfunction, process devi- One way to reduce alarm floods is to study univariate alarms, as
ation, or abnormal condition requiring a responses [1]. In a typical pointed out in [5]: “an alarm flood reduction will almost certainly
modern control system, a Message Processor program is employed require a rationalization exercise to challenge each and reduce
to handle the alarms which are detected by the data acquisition the number of configured alarms.” Usually a majority of univari-
subsystem. An alarm message is a text message or digital message ate alarms are chattering alarms, which can be removed efficiently
generated by the alarm system and stored in the alarm log. To be by applying delay timers or dead-bands. Techniques for designing
more specific, when a process value exceeds one of its predeter- delay timers and dead-bands have been proposed in [6–8].
mined thresholds, an alarm message is generated. Alarms can be Another way to reduce alarm floods is to rationalize conse-
easily configured on a Distributed Control Systems (DCS), which quential alarms. A alarm sequence pattern is usually generated
reduces the cost of alarm design and configuration, but on the other when there is an abnormality propagating through physical con-
hand prolongs the life-cycle of alarm rationalization, targeted to nections or information paths in a process. This type of pattern
avoid poor alarm designs. One consequence of poor alarm designs is can be obtained by applying pattern mining algorithms. The dis-
an increase in both the number and intensity of alarm floods. Alarm covered pattern can be potentially used in the advanced methods
floods can interfere with operators and may therefore cause or mentioned in the EEMUA standard [3], such as predictive alarming,
aggravate industrial accidents because even an experienced oper- online alarm attribute modification, and online alarm suppression.
ator can be overwhelmed by tens or hundreds of alarms raised in a Using pattern alarm sequences, causality analysis, either based on
short period of time. Without enough time for analysis, an operator process data [9–11] or alarm data (event logs) [12], can also be
conducted to investigate the root cause.
The authors in [13] summarized and categorized the existing
∗ Corresponding author. studies on the alarm overloading problem and formulated nine
E-mail addresses: [email protected] (S. Lai), [email protected] (F. Yang), fundamental research problems to be solved. In this paper, we
[email protected] (T. Chen), [email protected] (L. Cao). propose an accelerated multiple sequence alignment algorithm

https://doi.org/10.1016/j.jprocont.2019.06.004
0959-1524/© 2019 Elsevier Ltd. All rights reserved.
 S. Lai et al. / Journal of Process Control 82 (2019) 44–57 45

for pattern mining in multiple alarm floods, which potentially appears in a data set with frequency no less than a user-specified
can be used in solving the 5th and 6th listed problems, namely, threshold. For example, a subsequence, buying first a PC, then a
“whether there exists any nuisance alarms in the historical data, digital camera, and then a memory card, if it occurs frequently
worthy of redesigning alarm generation mechanisms” and “how in a shopping history database, is a sequential pattern. In [16],
to design mechanisms to generate predictive alarms in order to an A priori-like algorithm was proposed to mine frequent pat-
predict upcoming critical abnormal events.” terns by combining short sequences into longer ones. The authors
in [17] developed an algorithm to achieve patterns based on tree
1.2. Alarm flood analysis structures. In [18], a vertical data format was utilized to generate
patterns, which did not require multiple scans over the database.
To give further explanation of our contribution in solving the 5th
Algorithms based on projections were proposed in [19,20], which
and 6th listed problems, we need to have a detailed survey on alarm
had improvements on efficiency.
flood analysis, which contains univariate alarm analysis, offline pat-
In the biology area, a very large number of pattern mining algo-
tern mining of alarm floods, online alarm flood pattern matching,
rithms have been proposed, most of which were modifications of
and causality analysis. Usually, a large proportion of an alarm flood
the pairwise sequence alignment methods [21–24]. Mainly two
are nuisance alarms, of which chattering alarms form an impor-
types of modifications exist: the exhaustive search approach, as
tant part. Some methods, such as delay-timers [14], have been used
in [25], which can guarantee global optimality, and the progressive
to reduce such chattering alarms. However, these methods cannot
pairwise approach, such as [26] and [27], which can approximate
totally suppress nuisance alarms during alarm floods, the major-
global optimal solutions. There is another type of algorithm based
ity of remaining alarms are consequential alarms, which can be
on profile Hidden Markov Models (HMM), such as [28]. However,
caused by three reasons: (1) process state changes such as start-
in [29], the authors pointed out that: “a suitable HMM architec-
up and shutdown; (2) bad alarm configurations such as redundant
ture (the number of states, and how they are connected by state
measurements on a single process; and (3) causal relationships
transactions) must usually be designed manually.”
among measured variables. In this case, alarm flood analysis is use-
While a large amount of pattern mining algorithms exists in
ful to reveal connections of alarm messages and discover possible
the literature, most of them cannot be directly applied to find
patterns in alarm flood sequences.
patterns in alarm and event logs, due to the special format of
Fig. 1 shows the procedures of alarm flood analysis. The purpose
alarm data (every message comes with a time stamp). A variety
of the offline part is to set up a pattern database for offline analysis
of pattern mining methods have been developed to analyze alarm
and the use of online pattern matching. During the offline part,
floods. The authors in [30] manually selected some alarm tags to
data preprocessing is carried out first to remove chattering alarms
be the target tags at first, then applied the extracted patterns as
before the extraction of alarm floods. Usually an off-delay timer is
the basis for discovery of local relationships by identifying pri-
used since it will not introduce detection delays when alarms are
mary and consequential alarms. It is an event-based segmentation
raised. After preprocessing, an alarm rate threshold recommended
strategy which data segmentation takes place in an event-based
by ISA standard is used to extract alarm floods. The periods during
context. In [31], the authors proposed a way to capture the rela-
which the alarm rate is higher than the threshold is extracted. The
tions of alarm messages in an alarm flood using first-order Markov
extracted alarm flood sequences are then numbered and saved for
chains; then, the Euclidean distance between the transition prob-
further pattern analysis.
ability matrices of two alarm floods was used to cluster the alarm
In the next step, pattern mining algorithms can be applied to
floods into groups. The authors in [32] developed a pattern growth
automatically find a common pattern sequence for each of the clus-
method to obtain patterns from alarm floods. However, the authors
ters and save the patterns into a database for offline and online
pointed out “the proposed method was sensitive to disturbances
analysis. The pattern database can provide early prediction for an
so that the pattern in sequences had to be exactly the same in
incoming alarm flood by matching the online alarm sequence with
order to be recognized.” A dissimilarity-based method was pro-
the patterns in the database and obtain their similarity scores; then,
posed in [33] to extract alarm sequence templates of given faults
based on these scores, identify whether the online sequence is sim-
and a Needleman–Wansch based algorithm was developed to iso-
ilar to any of the patterns in the database and thereafter predict the
late alarm sequences caused by certain known faults in [34]. In [35],
incoming alarm flood if a matching can be found [15]. In Fig. 1,
the authors proposed a method to re-order alarms during alarm
potential dynamic alarm management could be predictive alarm-
floods by assigning their priorities values and designed an interface
ing, online alarm attribute modification, and online alarm suppres-
for a better operator support during flood scenarios.
sion. Causality analysis can also be applied once the patterns are
Time information is also very important. On one hand, two
obtained to help recover the connections between the correspond-
sequences with the same alarms and their ordering but different
ing tags in the pattern sequences and target the root cause.
time intervals in between the alarms can be caused by totally dif-
1.3. Literature survey ferent faults. On the other hand, even the orders of some alarms in
two alarm sequences are different, they still can be alarm sequence
This subsection contains a detailed literature survey on recent patterns. Because of this, methods have been proposed to take time
work for sequence pattern mining and alarm flood pattern analysis. intervals between alarms into consideration when studying the
Expert consultation and operator experience, by far, are still the sequences.
two approaches that have been used by most industrial compa- For example, for two alarm sequences, take into account the
nies when dealing with alarm floods. Expert consultation provides time information as shown below:
good and accurate results; however, without doubts, its efficiency
is extremely low because of the involvement of a relatively large S1 = < (4, 00 : 01), (5, 00 : 02),
amount of process knowledge. The approach based on the opera- (3,00:05),(2,00:06),(1,00:08),(3,00:19),(4, 00 : 20) >
tor experience is usually faster, but its accuracy is not guaranteed
S2 = < (3, 00 : 00), (2, 00 : 01),
since it depends heavily on human judgments. Many pattern min-
ing algorithms have been developed to facilitate the study of alarm (3,00:03),(2,00:04),(1,00:05),(3,00:09),(5, 00 : 12) >
floods.
Sequence pattern mining finds relevant parts in data examples where the first one of each pair is the type of alarm and the second
that appear repeatedly. A sequential pattern is a subsequence that one is the corresponding time stamp. By analyzing the time infor-
46 S. Lai et al. / Journal of Process Control 82 (2019) 44–57

Fig. 1. Flowchart of alarm flood analysis.

mation, we may find that the bold part may not be alarm sequence ple sequence alignment approach. Based on this work, a pattern
pattern due to the consideration of time stamps. database can be set up, which can provide early prediction for an
For two alarm sequences, incoming alarm flood by matching the online alarm sequence with
the patterns in the database and predict an incoming alarm flood
S3 = < (4, 00 : 01), (5, 00 : 02), if a matching can be found. The discovered patterns can help train
(3,00:05),(1,00:06),(2,00:08),(3,00:13),(4, 00 : 20) > operators to handle corresponding series of alarm messages more
efficiently in order to prevent the overwhelming situation during
S4 = < (3, 00 : 00), (2, 00 : 01),
alarm floods. Some badly configured parts in alarm systems can be
(3,00:05),(2,00:06),(1,00:08),(3,00:12),(5, 00 : 17) > revealed by the alarm flood patterns as well.
The rest of the paper is organized as follows. The problem
even the orders of some alarms in two alarm sequences are differ- description and algorithm principle are given in section 2. In sec-
ent, they may be an alarm sequence pattern due to the fact that the tion 3, an industrial case study is provided to test both efficiency and
time interval of alarms occurrences is relatively close, the similarity accuracy of the proposed algorithm and comparisons are made with
of these two alarm sequences may be high. the exhaustive search approach in [39]. Detailed discussions on the
In [36], Generalized Sequential Patterns (GSP) was applied to accuracy, computational complexity, algorithm convergence, and
search for pattern alarm sequences in alarm floods; the algorithm some potential problems are carried out in section 4, followed by
was able to blur the orders of alarms if they are raised in quick the conclusions in section 5.
succession and thus their order becomes less important. Algo-
rithms for pattern matching of two sequences were developed in 2. Algorithm principle
[37] and [38], based on the Smith-Waterman [22] and BLAST [23]
algorithms, respectively. In both papers, a weighted time distance Our intended problem is to find the optimal alignment for a
and vector were used to take the relative time between alarms cluster of alarm floods, so that based on this alignment an alarm
into account during the alignment. In [39], the authors proposed a sequence pattern can be easily found.
method for pattern mining in multiple alarm flood sequences using
a sequence alignment approach. However, the computational cost 2.1. Problem formulation
of the algorithm does not scale well with the number and length of
sequences to be aligned. 2.1.1. Alarm system and alarm message analysis
To offer a comprehensive explanation of alarms and alarm mes-
1.4. Contribution and organization of the paper sages, we give the figure of the alarm system dataflow and the figure
of alarm message log from industry. Fig. 2 shows a typical schematic
In this work, we propose an accelerated multiple pattern min- of alarm system dataflow. The basic Process Control System (BPCS)
ing algorithm to find a suitable alignment of multiple alarm flood and Safety Instrumented System (SIS) are the two important com-
sequences and obtain the pattern for a selected alarm flood clus- ponents that control the process and generate alarms based on
ter. Unlike traditional methods which either cannot deal with sensor measurements and predefined logics. In our work, we focus
multiple sequences with time stamps or suffer from high com- on the alarms from safety instrumented systems based on the
putational cost, the computational complexity of this proposed standard IEC 62682 [4], the design and management of safety
algorithm is reduced significantly by introducing a generalized instrumented systems are excluded from this standard. Please refer
pairwise sequence alignment method and a progressive multi- to the standard IEC 61511 [40]. The panel and the Human Machine
 S. Lai et al. / Journal of Process Control 82 (2019) 44–57 47

Fig. 2. Alarm system dataflow [1].

Interface (HMI) allow the operator to intervene the control of the where e1m , e2n ,..., eJo ∈  (the set of alarm types) and t1m , t2n ,...,
process and view alarm logs. tJo are the corresponding time stamps, respectively. S1 , S2 ,..., SJ are
In a DCS engaged plant, all monitored process variables and similar sequences detected by running pairwise sequence match-
alarms are collected onto the DCS server, where alarms are for- ing algorithms such as [37]. The objective is to obtain the optimal
matted and stored in alarm logs. Generally, in alarm logs, an alarm alignment of these sequences by adding gaps (represented by “[]”)
message contains important information such as time stamp, tag and deleting unrelated alarm messages, for example:
name, alarm type (identifier), priority, and acknowledgment state.
An alarm message may also include descriptive information such as < (e16 , t16 ), (e17 , t17 ), (e18 , t18 ), [ ] , (e19 , t19 ) >,
trip value, event, and description. Time stamp indicates the occur- < (e22 , t22 ), [ ] , (e23 , t23 ), (e24 , t24 ) , (e25 , t25 ) >,
rence time of an alarm. Tag name is a unique code of a system
variable that is being monitored; it indicates the process vari- ...
able that has caused the alarm. Alarm type, or identifier, is the <[ ] , (eJ1 , tJ1 ), (eJ2 , tJ2 ), [ ] , (eJ3 , tJ3 ) > .
alarm type related to the monitored variable (e.g., PVLO when
the variable is under its low limit). Priority shows the importance
of alarms. Acknowledgment state shows the status of an alarm, 2.2. Generalized pairwise sequence alignment
namely, Return to Normal (RTN) or Alarm (ALM).
Fig. 3 gives an example of a segment of an alarm log. It is nec- In this subsection, to explain the principles of generalized pair-
essary to point out that there are different kinds of alarm logs in wise sequence alignment, three elements, including time distance
different process or situations. Therefore, we need to give spe- and weight vectors, generalized similarity score and generalized
cific explanations according to different situations. The definitions dynamic programming procedure, will be introduced.
in this example cannot be applied directly to other examples. In
the example, a PVLO alarm was raised for tag A at 2013-09-01
2.2.1. Time distance and weight vectors
0:13 because its corresponding process variable went below the
Time stamps are important components of alarm messages; two
trip value, which was set as 131. Then, this PVLO alarm for tag
alarm messages with different time intervals could indicate totally
A returned to normal at 2013-09-01 0:13 since the correspond-
different problems in the process. In [37], it introduces the time
ing process variable went back to its normal range. The priority
information into the Smith-Waterman algorithm to evaluate alarm
of this alarm is JOURNAL, which means the lowest priority. Nor-
sequence alignment, which is called the modified Smith-Waterman
mally, we connect the tag name and the tag identifier of an alarm
algorithm. The Smith-Waterman algorithm is proposed for the local
together with a dot in between while processing alarm messages,
alignment of sequences in the field of biological informatics. Its
e.g., A. PVLO, and replace ALM and RTN messages with 1 and 0 for
objective is to find a pair of segments, one from each of two long
computing purposes.
sequences, such that there is no other pair of segments with greater
similarity (homology) [22].
2.1.2. Mathematical representation The Smith-Waterman algorithm is a dynamic programming
Consider the following alarm message sequences: algorithm, which contains two main steps. The first step is calculat-
ing the similarity score of two sequences and forming a similarity
S1 = < (e11 , t11 ), (e12 , t12 ), ..., (e1m , t1m ), ..., (e1M , t1M ) >, matrix or table, the similarity score of the mismatch and gap-match
is negative as a penalty. The similarity score function can be defined
S2 = < (e21 , t21 ), (e22 , t22 ), ..., (e2n , t2n ), ..., (e2N , t2N ) >, as follows:
... 
1, ifexi = eyj
SJ = < (eJ1 , tJ1 ), (eJ2 , tJ2 ), ..., (eJo , tJo ), ..., (eJO , tJO ) >, (xi , yi ) = (1)
, ifexi =
/ eyj
48 S. Lai et al. / Journal of Process Control 82 (2019) 44–57

Fig. 3. An example of a segment of an Alarm log.

where w is the time weight vector, w is a time weighting function

with respect to the time distance. The time distance vector for an
alarm message (em , tm ) is defined as:

1 , d2 , ..., dk , ..., dK ]
dm = [dm m m m
(4)
k = |t − t |,
dm m k

Fig. 4. Smith–Waterman algorithm. where K is the total number of alarm messages in the sequence.
Thus, dm k gives the absolute time distance between (e , t ) and (e ,
m m k
tk ) in the sequence. The time weight vector for (em , tm ) is defined
as:

wm = 1 , w 2 , ..., w k , ..., w K ]
[wm m m m
(5)
1 ), f (d2 ), ..., f (dk ), ..., f (dK )],
= [f (dm m m m

where f ( · ) : R → R is the weighting function. A scaled Gaussian

function is selected in this paper:
Fig. 5. Modified Smith–Waterman algorithm.
2 /2 2
f (x) = e−x , (6)

The element of similarity matrix H can be recursively calculated by where  is the standard deviation, controlling how much weight to
the following equation: be put on the close-by alarm messages to blur their orders in the
    alignment. The function also normalizes the real time distance to [0,
Hi+1,j+1 = max Hi,j + s xi , yj , Hi,j+1 + ı, Hi+1,j + ı, 0 (2)
1]; a large output indicates a short time distance. Correspondingly,
where ı is uniform gap penalty. The second step is backward path for modified Smith-Waterman algorithm, the element of similarity
search based on similarity table. Go backward from this highest matrix H can be recursively calculated by the following equation
value until meet an entry with value 0 (Figs. 4 and 5). with a uniform gap penalty ı:
For the modified Smith-Waterman algorithm in [37], it intro-     
duces the time information in sequences, which can be reflected Hi+1,j+1 = max Hi,j + s (ei , ti ) , ej , tj , Hi,j+1 + ı, Hi+1,j + ı, 0
by the change of similarity score function:
(7)
 
s ((ea , ta ) , (eb , tb )) = max wak × wbk (1 − ) +  (3)
1≤k≤K
 S. Lai et al. / Journal of Process Control 82 (2019) 44–57 49

Fig. 6. An example of tuples in two alarm sequence alignments.

2.2.2. Worked example of time distance and weight vectors they are from the same alignment, while P and Q may not be the
Here is a demonstration of how the time distance and weight same value.
vectors are calculated for an alarm sequence. Consider the following The formula for calculating the similarity score S(TAp , TBq )
alarm sequence between the tuples TAp and TBq is as follows:

1 −  
M N
(2, 2), (1, 3), (3, 3.5), (1, 5), (4, 5.2),
p p q q
S(TAp , TBq ) = + × S((eAi , tAi ), (eBj , tBj )), (8)
where there are 5 alarm messages and 4 alarm types. The time MN
i=1 j=1
distance vector d1 for message (2, 2) is calculated as: d11 = 2 − 2 =
p p q q
0, d12 = 3 − 2 = 1, d13 = 3.5 − 2 = 1.5, d14 = 5 − 2 = 3, and d15 = 5.2 − where S((eAi , tAi ), (eBj , tBj )) is the similarity score between two
2 = 3.2. The time distance matrix formed by vectors d1 , d2 , d3 , d4 , alarm messages and is obtained by
and d5 is p p q q
S((eAi , tAi ), (eBj , tBj ))
⎡ ⎤
0.0 1.0 1.5 3.0 3.2 p p q q
= max{ s((eAi , tAi ), (eBj , tBj )), (9)
⎢ 1.0 0.0 0.5 2.0 2.2 ⎥
⎢ ⎥ q q p p
[d1 , d2 , d3 , d4 , d5 ] = ⎢ ⎥
T T T T T s((eBj , tBj ), (eAi , tAi ))},
⎢ 1.5 0.5 0.0 1.5 1.7 ⎥ .
⎣ 3.0 2.0 1.5 0.0 0.2 ⎦ where
3.2 2.2 1.7 0.2 0.0 p p q q
s((eAi , tAi ), (eBj , tBj ))
⎧ q
By applying the weighting function (with  = 1) on the time distance ⎪ k
max {wAi,p k
: eAi = eBj }, if the set isn’t empty
⎪
⎪
matrix, the time weight matrix is obtained: ⎪
⎨
1≤k≤P
 
(10)
⎡ ⎤ =
q q
and eBj , tBj is not a gap
1.00 0.61 0.32 0.01 0.01
⎪
⎪
⎢ 0.61 1.00 0.88 0.14 0.09 ⎥ ⎪
⎪
⎢ ⎥ ⎩
[wT1 , wT2 , wT3 , wT4 , wT5 ] = ⎢ ⎥
⎢ 0.32 0.88 1.00 0.32 0.24 ⎥ .
0, otherwise,
⎣ 0.01 0.14 0.32 1.00 0.98 ⎦ q q p p
s((eBj , tBj ), (eAi , tAi ))
0.01 0.09 0.24 0.98 1.00 ⎧ k k p
⎪ max {wBj,q : eBj = eAi }, if the set isn’t empty
⎪
⎪ 1≤k≤Q
 p p
⎨ (11)
and eAi , tAi is not a gap
2.2.3. Generalized similarity score =
⎪
⎪
In [37], the authors proposed a way to calculate the similarity ⎪
⎩
score between two alarm messages in two alarm sequences. Here 0, otherwise.
we generalize it to calculate the similarity score between two tuples
in two alarm sequence alignments. The definition of a tuple is illus- The negative parameter  in Eq. (8) is the mismatch penalty. The
trated by the example in Fig. 6. In the figure, there are two alarm similarity score S(TAp , TBq ) between the two tuples TAp and TBq
sequence alignments (A1 , A2 , . . ., AM and B1 , B2 , . . ., BN ); the tuples is obtained by averaging all the pairwise similarity scores of the
(TA1 , TA2 ,. . ., TAP and TB1 , TB2 ,. . ., TBQ ) are formed by alarm mes- alarms in the two tuples. The value of S(TAp , TBq ) is always within
sages or gaps located at the same position of the corresponding [, 1]; a larger value means a stronger similarity. The similarity
p p q q p p
alignment. The lengths of the alignments are P and Q, respectively. score S((eAi , tAi ), (eBj , tBj )) between two alarm messages (eAi , tAi )
q q
Notice the lengths of A1 , A2 , . . ., AM are always the same because and (eBj , tBj ) is achieved by selecting the larger value between
50 S. Lai et al. / Journal of Process Control 82 (2019) 44–57

Fig. 7. A numerical example for the calculation of the generalized similarity scores.

p p q q q q p p
s((eAi , tAi ), (eBj , tBj )) and s((eBj , tBj ), (eAi , tAi )), as the commutative
p p q q
law does not hold. When both (eAi , tAi )
and (eBj , tBj ) are gaps, their
score will be assigned as 0. Note that the pairwise similarity score
formula in [37] is a special case of the generalized similarity score
calculation proposed in this paper; when the two alignments are
reduced to two alarm sequences (e.g., M = N = 1), the two formulae
will give exactly the same output.
Fig. 8. An illustrative example on how to obtain the generalized dynamic program-
ming matrix.
2.2.4. Worked example of generalized similarity score
Consider the example of two alarm sequence alignments in
Fig. 7, where M = N = 2 and the lengths of the two sequence align- WA2 = [wTA , wTA , wTA , wTA , wTA ]
2 ,1 2 ,2 2 ,3 2 ,4 2 ,5
ments are both 5. The time weight matrices of A1 , A2 , B1 , and B2 are ⎡ ⎤
obtained as WA1 , WA2 , WB1 , and WB2 , given  = 1. The fourth row 1.00 0.88 0.14 0.01 0
and column of WB2 are empty since the fourth element in B2 is a ⎢ 0.88 1.00 0.32 0.04 0 ⎥
⎢ ⎥
gap. Let  = −1, the similarity score for the pair of tuples TA4 and =⎢ ⎥
⎢ 0.14 0.32 1.00 0.61 0.14 ⎥
TB4 is calculated as ⎣ 0.01 0.04 0.61 1.00 0.61 ⎦
0 0 0.14 0.61 1.00
1 + 1 
2 2
4 4 4 4
S(TA4 , TB4 ) = −1+ × S((eAi , tAi ), (eBj , tBj )) = −0.12,
2×2
i=1 j=1
WB1 = [wTB , wTB , wTB , wTB , wTB ]
1 ,1 1 ,2 1 ,3 1 ,4 1 ,5
⎡ ⎤
where 1.00 0.61 0.14 0.04 0.01
⎢ 0.61 1.00 0.61 0.32 0.14 ⎥
4
S((eA1 4 4 4
)) = max{wA3 , wB3 } = 0.88, ⎢ ⎥
=⎢ ⎥
, tA1 ), (eB1 , tB1
1 ,4 1 ,4
⎢ 0.14 0.61 1.00 0.88 0.61 ⎥
⎣ 0.04 0.32 0.88 1.00 0.88 ⎦
4 4 4 4 0.01 0.14 0.61 0.88 1.00
S((eA2 , tA2 ), (eB1 , tB1 )) = max{wA3 , wB3 } = 0.88,
2 ,4 1 ,4

WB2 = [wTB , wTB , wTB , wTB , wTB ]

2 ,1 2 ,2 2 ,3 2 ,4 2 ,5
4
S((eA1 4
, tA1 4
), (eB2 4
, tB2 )) = max{0, 0} = 0, ⎡ ⎤
1.00 0.88 0.14 [] 0.01
and ⎢ 0.88 1.00 0.32 [] 0.04 ⎥
⎢ ⎥
=⎢ ⎥
⎢ 0.14 0.32 1.00 [] 0.61 ⎥
4 4 4 4
S((eA2 , tA2 ), (eB2 , tB2 )) = max{0, 0} = 0. ⎣ [] [] [] [] [] ⎦
0.01 0.04 0.61 [] 1.00
Intuitively, this negative similarity score also makes sense since
there is a gap in tuple TB4 and the type of the two alarm messages
in tuple TA4 is “1” while the type of the alarm message in tuple TB4
2.2.5. Generalized dynamic programming procedure
is “3”. The reason why this similarity score is −0.12 instead of the
Similarly, the dynamic programming procedure proposed in
minimum value −1 is because there exists an alarm message (1, 14)
[37] for aligning two alarm sequences is generalized to aligning
in tuple TB3 that was raised slightly ahead of (3, 14.5) and its alarm
two alarm sequence alignments.
type matches the alarm messages in tuple TA4 .
Fig. 8 illustrates the way to achieve the generalized dynamic
WA1 = [wTA , wTA , wTA , wTA , wTA ] programming matrix from the tuples’ point of view. To begin with,
1 ,1 1 ,2 1 ,3 1 ,4 1 ,5 the first row and column of the matrix are filled in with zeros. Then,
⎡ ⎤
1.00 0.61 0.32 0.01 0.01 iteratively fill the rest of cells of the matrix with Hx,y obtained from
⎢ 0.61 1.00 0.88 0.14 0.09 ⎥
⎢ ⎥ Hx+1,y+1 = max (I(TA,i:x , TB,j:y ), 0)
=⎢ ⎥
⎢ 0.32 0.88 1.00 0.32 0.24 ⎥ 1≤i≤x,1≤j≤y
⎣ 0.01 0.14 0.32 1.00 0.98 ⎦ = max{ Hx,y + S(TA,x+1 , TB,y+1 ), (12)
0.01 0.09 0.24 0.98 1.00 Hx,y+1 + ı, Hx+1,y + ı, 0},
 S. Lai et al. / Journal of Process Control 82 (2019) 44–57 51

the proposed method approximates the optimal alignment by pro-

gressively aligning all the sequences. This new approach greatly
improves the efficiency of the algorithm; as the cost, however, the
global optimum is no longer guaranteed.

2.4. Iterative alignment refinement methods

Two iterative alignment refinement methods are developed to

improve the accuracy of the primitive alignment generated by the
progressive sequence alignment approach.

2.4.1. Leave-one-out alignment refinement

The first one is Leave-one-out alignment refinement. Similar to
Fig. 9. An example of dendrogram and the progressive multiple sequence alignment
the leave-one-out cross validation, which leaves one dataset out
procedure based on it. for testing and use the rest for training, during each iteration of the
leave-one-out alignment refinement, one sequence in the align-
ment is replaced by its original sequence and then be re-aligned
where the negative parameter ı is the gap penalty, and the simi- with the rest of sequences in the alignment (if any, common gaps
larity index I(TA,i:x , TB,j:y ) is the alignment score of the segment pair in the rest of the alignment sequences should be removed) using
(TA,i:x , TB,j:y ). the generalized pairwise sequence alignment method. After re-
Once the whole dynamic programming matrix is achieved, the aligned, the new alignment score is compared with the old one
optimal alignment of the two alignments can be generated based on before the current refinement iteration, and the alignment with the
back-tracking. During back-tracking, the position of the maximum higher score will be kept for the next iteration. Iteratively repeat
value in the dynamic programming matrix is located at first. Then, this procedure until the alignment score converges.
the algorithm tracks backwards to obtain the path along which the Fig. 10a shows an illustration of one iteration of the leave-
maximum score is generated. At last, one forward pass through the one-out refinement; respectively, the solid lines and “[]” represent
back tracking path gives the optimal alignment. alarm sequences and gaps in the alignments. In the example, align-
The difference between the generalized and the standard ment 1 is left out and replaced by its original alarm sequence, as
dynamic programming procedures is that the former one treats indicated by the red bold line. Then the common gaps in the rest
the “sequence” from a “tuple” point of view, which allows the of the alignments (2, 3, 4, and 5) are removed before they are re-
“sequence” to be either an alarm sequence or an alignment of mul- aligned with the original alarm sequence 1.
tiple alarm sequences. Thus, the standard dynamic programming Note the convergence is guaranteed by the leave-one-out refine-
procedure can be seen as a special case of the generalized one. ment method. However, it is not guaranteed to converge to the
global optimum. Details of convergence will be provided in the
2.3. Progressive multiple sequence alignment method discussion section.

The computational cost of the exhaustive search approach in

[39] grows exponentially with the number of sequences and their 2.4.2. Random-division alignment refinement
lengths, making the algorithm easily overwhelmed by real data. To The second is Random-division alignment refinement. Com-
improve the efficiency, we propose a progressive multiple sequence pared to the leave-one-out approach, it has a relatively lower
alignment method based on the generalized dynamic programming convergence speed, but in some cases, it could still improve the
to find the optimal sequence alignment. alignment result even when the leave-one-out alignment con-
In the first step of our algorithm, the modified Smith-Waterman verges. The detailed reason will be provided in the discussion
is applied to calculate all the pairwise similarity scores between section.
the alarm sequences. Then, a dendrogram is built based on these As illustrated in Fig. 10b, the primitive alignment is randomly
pairwise similarity scores using UPGMA (Unweighed Pair Group divided into two groups (1, 2, 4, and 3, 5). Then, if any, the common
Method with Arithmetic Mean). Fig. 9 shows an example of a den- gaps in each of the two groups of sequences should be removed,
drogram, in which there are 6 alarm sequences and the height of a as shown in the example. Next, the two groups of sequences
link indicates the dissimilarity between two connected sequences. are re-aligned using the proposed generalized pairwise sequence
Next, guided by the dendrogram, the primitive alignment of the alignment method. After re-aligned, the new alignment score is
sequences is obtained by progressively aligning all the sequences, compared with the old one before the current refinement itera-
from the most similar pair to the most dissimilar ones. In the exam- tion, and the alignment with the higher score will be kept for the
ple shown in Fig. 9, the link connecting sequences 1 and 2 is the next iteration. Iteratively repeat this procedure until the upper limit
lowest one, indicating them to be the most similar sequence pair. number of iterations is reached.
Thus, these two sequences are aligned first and thus their align- Similar to the leave-one-out refinement, the convergence is
ment a is obtained. Similarly, sequences 4 and 5 are aligned and guaranteed for the random-division approach when no upper limit
thus alignment b is obtained. Then, since there is a connection iteration number is set. However, the convergence is hard to be
between sequence 3 and alignment b, the generalized pairwise identified without knowing the optimal alignment score before-
sequence alignment method is conducted to get alignment c by hand, also the convergence value is not guaranteed to be global
aligning sequence 3 and alignment b. By iteratively aligning the optimum. Moreover, since no sequence in the alignment will be
sequences based on the dendrogram, the primitive alignment of all replaced by its original sequence during the refinement, the primi-
the 6 alarm sequences can be achieved. tive alignment becomes much more critical; and the deleted alarm
Compared to the exhaustive search approach in [39], which messages in the sequence of an alignment will never be brought
conducts exhaustive search to find the exact optimal alignment, back.
52 S. Lai et al. / Journal of Process Control 82 (2019) 44–57

Table 1
Statistics of the dataset

Description Number

Total time period 336 days

Total number of tags 1502
Total number of alarms 109,393
Average alarm rate 14/h
Highest peak alarm rate 334/10 min
Number of alarm floods 359
Average length of alarm floods 39

Table 2
Information of the alarm sequences in each selected cluster

Cluster A B C D

Number of alarm floods 19 12 9 17

Average sequence length 26.6 14.6 31.6 20.1
Longest sequence length 46 23 64 36
Shortest sequence length 15 10 14 10
Pattern length 14 4 11 2

have been applied as a preprocessing step at the beginning of the

analysis to remove chattering alarms. The reason for using off-
delay rather than on-delay is because it does not introduce any
delay when raising an alarm. Alarm floods were extracted based on
the ISA standard: 10 alarms per 10 minutes per operator. General
descriptions of the dataset are shown in Table 1.
The flowchart of the algorithm is shown in Fig. 11. The param-
eters were chosen as:  = 0.2,  = −1, ı = −1, and the upper limit
iteration number for random-division refinement is set as 50. All
the tests were carried out on a 64 bit Windows PC with Intel i7-4770
3.40GHz CPU and 24.0 GB memory.
Pairwise similarity scores were calculated using the method
in [37] and the UPGMA clustering was conducted based on the
obtained similarity scores thereafter. Similar to the proposed
approach, the method in [37] finds the similarity score between a
pair of sequences by searching for their best alignment. It is able to
find the exact optimal alignment; however, it can only be applied
on a pair of sequences rather than multiple ones. The clustering
result is shown in Fig. 12, where each pixel represents the sim-
ilarity score between two corresponding alarm flood sequences;
darker color means a higher similarity. Four groups of alarm floods
were selected to be the test dataset, denoted as clusters A, B, C, and
D. The reason for choosing them as the test datasets was because
they each contains sufficient number of alarm floods for testing, and
a pattern of alarm sequence exists in each of the clusters. Detailed
information about the four selected clusters can be found in Table 2.
Both accuracy and efficiency tests have been conducted to com-
pare the proposed algorithm with the exhaustive search approach
in [39]. Same parameters were set for both algorithms.

3.1. Accuracy tests

For each of the four clusters, 3 sequences were randomly

selected, and the two algorithms were applied respectively to find
the best alignment. Since the exhaustive search approach guaran-
tees the global optimum, the accuracy of the proposed algorithm
can be evaluated by the alignment score difference between the two
algorithms. Repeat the whole procedure for 50 times and obtain
Fig. 10. Illustrations of the two types of alignment refinement methods.
the average alignment score difference for each cluster; the result
is shown by the green line in Fig. 13.
3. Industrial case study From the left to the right, the four points on the green line
indicates the average alignment score difference between the two
The proposed algorithm has been tested on an industrial dataset algorithms in clusters B, D, A, and C respectively (in the ascendant
from Suncor Energy Inc. The dataset comes from a plant that is order of the average sequence length in each cluster). Similarly, the
already in production but need rationalization to deal with chat- average alignment score difference between the two algorithms
tering alarms and alarm floods. Off-delay timers of 300 seconds was obtained when 4 and 5 sequences were randomly selected
 S. Lai et al. / Journal of Process Control 82 (2019) 44–57 53

Fig. 12. Clustering result of the extracted alarm floods and the four selected clusters.

Fig. 13. Percentage of alignment score difference between the proposed algorithm
and the exhaustive search approach.

3.2. Efficiency tests

Execution times of both algorithms during the accuracy tests

were recorded to compare their efficiency.

Fig. 11. Flowchart of the proposed algorithm. 3.2.1. Efficiency comparison of two algorithms
Fig. 14 shows the efficiency of the two algorithms as the num-
ber of sequences to be aligned grows. The execution times of both
from each of the four clusters, presented by the red and blue lines algorithms increases when the number of sequences grows; how-
respectively. ever, the growth rate of the proposed algorithm is much smaller
A very noticeable spike can be found in all the three lines than that of the exhaustive search approach. When the number of
in Fig. 13, showing that the average alignment score difference sequences reached 5, the proposed algorithm was almost 10,000
between the two algorithms was large for the tests in cluster D. times faster than the exhaustive search approach.
The reason for this is because the pattern length in cluster D was Fig. 15 indicates there is a positive relation between the com-
too short (only 2). Thus, the alignment can be easily influenced putational cost of the exhaustive search approach and the average
by the similar terms among the sequences that are not related to sequence length, but this trend is not evident for the proposed algo-
the pattern. Also, since the pattern length is too short, the per- rithm. The reason is because the pattern length, as numbered on
centage of alignment score difference caused by a small defect in the figure, influences the execution time of the proposed algorithm
the alignment would be amplified. In addition to the big spike as well. Detailed explanation will be provided in the discussion
caused by cluster D, the accuracy of the proposed algorithm is section.
high, with no more than 10% difference compared to the exhaustive
search approach. Fig. 13 also shows that there is no evident rela- 3.2.2. Execution time analysis of the proposed algorithm
tion between the accuracy of the proposed algorithm and number More efficiency tests were conducted to cover the testing region
of sequences to be aligned or the average sequence length. that had not been touched by the accuracy tests. Since the exhaus-
54 S. Lai et al. / Journal of Process Control 82 (2019) 44–57

Fig. 14. Efficiency comparison of the two algorithms when the number of sequences Fig. 17. Execution time of the proposed algorithm to align 10 sequences as the
increases. average sequence length grows.

Fig. 15. Efficiency comparison of the two algorithms when the average sequence Fig. 18. Execution time of the algorithm with respect to the upper limit iteration
length increases. number for the random-division refinement.

grows. Even when the sequence number exceeds 250, as shown

in Fig. 17, the execution time of the proposed algorithm was still
within 1 minute.
Fig. 17 reveals the relationship between execution time of the
algorithm and average sequence length. The tests were carried out
by applying the proposed algorithm to align 10 alarm sequences
randomly selected from cluster A. In order to increase the sequence
length in the each test, the sequences were extended by duplicat-
ing and connecting themselves with their duplicates. Each test was
repeated 50 times and both the average and the longest execution
time were recorded. The result shows the longest execution time
grows quite fast with the growth of average sequence length; how-
ever, the average execution time grows much slower and is always
less than 50 seconds. The reason is because the sequence length
in cluster A varies from 15 to 46 and thus the sequence growth
rates were quite different when the sequences were extended by
Fig. 16. Execution time of the proposed algorithm when the number of sequences duplicating themselves.
grows. Fig. 18 shows the influence of the upper limit of iteration num-
ber for the random-division refinement on the execution time of
tive search approach could hardly finish in most of the tests, only the algorithm. The results were obtained by aligning 10 sequences
the result of the proposed algorithm is shown. randomly selected from cluster A, with different upper limits of
Fig. 16 shows the execution time of the proposed algorithm iteration numbers for the random-division refinement in the algo-
when the number of sequences to be aligned increases. The tests rithm. Each test was repeated 50 times and both the average and
were conducted by randomly selecting sequences from cluster A the longest execution times were recorded. The result shows a lin-
(average sequence length 26.6); 50 repeated tests were carried out ear increasing trend of the execution time of the algorithm when
for each number of sequences. The result reveals a quadratic trend the upper limit increases, as is expected. The detailed reason will
in the computational cost as the number of sequences to be aligned be explained in the discussion section.
 S. Lai et al. / Journal of Process Control 82 (2019) 44–57 55

Fig. 19. The number of refinement iterations needed before reaching the conver-
gence when the number of sequences to be aligned increases.

Fig. 21. Examples of the optimums achieved by the two types of refinement meth-
ods: (a) when the random-division refinement method was better; (b) when the
leave-one-out refinement method was better.

3.3.2. Comparison of two refinement algorithms

The two refinement methods were also compared by evaluating
Fig. 20. An example of convergence curves when the order of sequences to be left
out for the leave-one-out refinement method varies.
the amount of improvement they can make on a given primitive
alignment. Since for a certain primitive alignment, the leave-one-
out refinement always has the same convergence trail, while the
random-division method doesn’t, the final alignment score of the
3.3. Convergence tests leave-one-out refinement was compared with the results from 30
runs of the random-division refinement.
3.3.1. Convergence analysis of the leave-one-out refinement Two cases where the two methods achieved a higher alignment
Since the convergence speed of the random-division refinement score compared with each other are shown in Fig. 21. The red bold
is much slower than the leave-one-out refinement and can hardly lines represent the converging trails of the leave-one-out method,
be identified, only the convergence speed of the leave-one-out while the blue lines show the trails of the random division method.
refinement was tested. Fig. 19 shows the refinement iteration num- In case (a), the leave-one-out refinement was stuck in a local opti-
ber needed before the convergence was reached by leave-one-out mum and outperformed by most of the runs of the random division
refinement when aligning different numbers of sequences in each refinement. However, in case (b), the leave-one-out refinement
cluster. As the result shows, more iterations were needed when was better than all runs using the random-division refinement. The
the number of sequences to be aligned was large. However, there reason for this was because the leave-one-out refinement method
is no clear relationship between the refinement iteration num- could recover from the incorrectly deleted part for the primitive
ber needed for convergence and the average sequence length, as alignment, but the random-division refinement method could not.
the descending order of average sequence length in each cluster is Thus, during the procedure of forming the primitive alignment, if a
C > A > D > B. part of the sequence that should appear in the optimal alignment
When applying the leave-one-out refinement method, for each was deleted incorrectly because of the approximation, then we
iteration we always select the sequence by their orders in the prim- should expect it to be recovered by the leave-one-out refinement
itive alignment. However, will the order we select the sequences method rather than the random-division refinement.
to be left out make a difference in the final alignment given by the
leave-one-out refinement method? The answer is yes, as the result 4. Discussions
in Fig. 20 shows, where 10 different orders were tried to improve
a certain primitive alignment using the leave-one-out refinement 4.1. Accuracy discussion
method. The result shows the order we use to select the sequences
will influence not only the convergence value but also the speed for Compared to the exhaustive search approach in [39], the pro-
the leave-one-out refinement method. posed algorithm does not guarantee the global optimum. However,
56 S. Lai et al. / Journal of Process Control 82 (2019) 44–57

as shown in the industrial case study section, the accuracy of the a higher score will be fed to the next iteration. Thus, the align-
proposed algorithm is acceptable (within 10% range) when the pat- ment score is monotonically increasing, while there must be an
tern length is not too small (larger than 3). For the cases where the upper bound for the alignment score of certain sequences. There-
pattern sequence is too short, the algorithm may not perform well fore, the convergence is guaranteed for both of the refinement
since the alignment can be easily influenced by the similar terms methods. The leave-one-out refinement method converges when
among the sequences that are not related to the pattern, and the no more improvement can be made to the alignment for a full round
improvement on the primitive alignment that could be made by of iterations (every sequence has been left out once). However,
the two refinement methods will be very limited as well. the convergence of the random-division refinement can hardly be
We need to point out that there is a trade-off between algo- identified, as there does not exist a full round of iterations. Thus,
rithm’s efficiency and accuracy. Our algorithm does not guarantee an upper limit number of iterations should be configured for the
to find the optimal solution like the method in [39] does. However, random-division refinement method.
based on our tests, the accuracy of the proposed algorithm is still Convergence speeds of the two refinement methods are worthy
suitable for applications, averaging 5% difference when compared studying too. The leave-one-out refinement method usually, but
to the results given by the method in [39]. The efficiency of the pro- not always, converges faster than the random-division method. The
posed algorithm is higher than the method in [39] by several orders reason is because for both of the refinement methods, there are two
of magnitude. Thus, given a large scale dataset, the proposed algo- main steps: separate the alignment into two bunches of sequences,
rithm is able to complete fast with usable results while the method and re-align them. The leave-one-out refinement method has a cer-
in [39] may stuck for hours or even days. tain order to separate the alignment, and only one sequence is
separated from the rest during one iteration. While the random-
4.2. Computational complexity discussion division refinement method does not; thus there is a much bigger
pool of ways of separating the sequences in the alignment that the
The computational complexity of the proposed algorithm is random-division refinement method can select from.
composed of five parts: (1) calculation of time weight matrices Regarding the convergence value, both refinement methods
N cannot guarantee the global optimum, and neither of them can
O( i=1 Li2 ), where N is the sequence number and Li is the sequence
always achieve a better alignment score than the other. How-
length; (2) formation of the dendrogram O(N 3 ); (3) generation of
ever, by combining them together we could have a better chance
the primitive alignment using progressive multiple sequence align-
N−1 to reach the global optimum, because each method improves the
ment method O( j=1 Aj1 Aj2 ), where Aj1 and Aj2 are the lengths of
alignment from a different perspective. The leave-one-out refine-
the two sequences (can be either a sequence or an alignment) for ment method can recover the part of the sequence that has been
the generalizedpairwise sequence alignment; (4) leave-one-out incorrectly deleted during the generation of the primitive align-
I1
refinement O( k=1 Ak1 Ak2 ), where I1 is the number of iterations ment because in each iteration, one sequence in the alignment is
before convergence is reached, and Ak1 and Ak2 are the lengths of the replaced by its original sequence and re-aligned with the rest of the
two sequences (can be either a sequence or an alignment) for the sequences in the alignment. However, the random-division refine-
generalized pairwise
I2 sequence alignment; and (5) random-division ment method cannot do so because no separated sequences will be
refinement O( t=1 At1 At2 ), where I2 is the upper limit iteration replaced by their original ones. Even though it cannot recover the
number, and At1 and At2 are the lengths of the two sequences (can incorrectly deleted sequence part, it still has an advantage over the
be either a sequence or an alignment) for the generalized pairwise leave-one-out refinement method — a much bigger pool of ways of
sequence alignment. Thus the whole computational complexity is separating the sequences in the alignment to choose from, which
means more chances to improve the alignment by re-arranging the
⎛ ⎞ gaps and matchings. Thus, by applying the leave-one-out refine-

N

N−1

I1

I2
ment followed by the random-division refinement method, we will
O⎝ Li2 + N 3 + Aj1 Aj2 + Ak1 Ak2 + At1 At2 ⎠ .
be able to first recover the incorrectly deleted sequence parts, and
i=1 j=1 k=1 t=1
then modify the ways of placing gaps and matchings in the existing
alignment to achieve a better alignment score. This is why in the
Based on the computational complexity analysis, the efficiency
algorithm flowchart shown in Fig. 11, the leave-one-out refinement
of the proposed algorithm is determined by many factors, among
is applied before the random-division refinement.
which the number of sequences, sequence length, and alignment
length are the three main elements. The highest order of sequence
4.4. Parameter tuning discussion
or pattern length in the computational complexity is only quadratic
for the proposed algorithm. Notice the computational complexity
Parameter tuning can be an obstacle for applying the proposed
of the exhaustive search approach in [39] is
algorithm. As mentioned in [37], the variance of the Gaussian func-

N 
N tion , the gap and mismatch penalties ı and  need to be tuned
O(( Lk ) × ( Lk )), to meet users’ requirements. The variance of the Gaussian function
k=1 k=1  controls the size of the time span within which the algorithm
blurs the occurrence orders between the alarm messages. When the
which increases exponentially with the sequence number and value of  goes to infinity, the algorithm totally ignores the orders of
length. Thus, the proposed algorithm achieves a significant alarms in the alignment and simply counts the alarm occurrences.
improvement in efficiency, which has also been confirmed in the On the contrary, if  = 0, the orders of alarms have to be exactly
industrial case study section. the same in the two sequences in order to get a match. The gap
and mis-match penalties,  and , determine the algorithm’s tol-
4.3. Convergence discussion erance to including gaps and mis-match terms in alignments. When
these two parameters get larger, the algorithm places more toler-
The convergence can be guaranteed by both refinement meth- ance on the irrelevant alarms raised within a pattern sequence.
ods. The proof is simple and does not require any mathematical Besides the ones in [37], a new parameter is introduced in the
derivations. Since in every refinement iteration, the new alignment proposed algorithm: the upper limit number of iterations I2 for
score will be compared with the old one, and the alignment with the random-division refinement. With a higher I2 , there will be a
 S. Lai et al. / Journal of Process Control 82 (2019) 44–57 57

greater chance that the global optimum is reached, but at the cost [11] P. Duan, F. Yang, S.L. Shah, T. Chen, Transfer zero-entropy and its application
of higher computational cost. for capturing cause and effect relationship between variables, IEEE Trans.
Control Syst. Technol. 23 (3) (2015) 855–867.
During the generation of the primitive alignment, the sequences [12] F. Yang, S.L. Shah, D. Xiao, T. Chen, Improved correlation analysis and
are progressively aligned by the descending order of their pairwise visualization of industrial alarm data, ISA Trans. 51 (4) (2012) 499–506.
similarity. The reason for choosing this order is because once an [13] J. Wang, F. Yang, T. Chen, S.L. Shah, An overview of industrial alarm systems:
main causes for alarm overloading, research status, and open problems, IEEE
alignment is obtained, the gaps and the matchings between the Trans. Autom. Sci. Eng. 13 (2) (2016) 1045–1061.
two sequences are fixed and will be propagated into the primitive [14] N. Ahmed Adnan, I. Izadi, T. Chen, On expected detectiondelays for alarm
alignment. Thus, by aligning the most similar sequence pair first systems with deadbands and delay-timers, J. Process Control 21 (9) (2011)
1318–1331.
and leaving the most divergent pair to the last would improve the
[15] S. Lai, F. Yang, T. Chen, Online pattern matching and prediction of incoming
correctness of placing gaps and matchings in the alignment. alarm floods, J. Process Control 56 (2017) 69–78.
[16] R. Agrawal, R. Srikant, Mining sequential patterns, in: Proceedings of the 11th
International Conference on Data Engineering, IEEE, 1995, pp. 3–14.
5. Conclusions
[17] J. Han, J. Pei, Y. Yin, Mining frequent patterns without candidate generation,
ACM SIGMOD Record, no. 2 (2000) 1–12.
An accelerated multiple sequence alignment algorithm for pat- [18] M.J. Zaki, Efficient enumeration of frequent sequences, in: Proceedings of the
seventh International Conference on Information and Knowledge
tern mining in multiple alarm sequences has been proposed. A
Management, ACM, 1998, pp. 68–75.
progressive multiple sequence alignment mechanism has been [19] J. Han, J. Pei, B. Mortazavi-Asl, Q. Chen, U. Dayal, M.-C. Hsu, Freespan: frequent
introduced to accelerate the generation of the primitive alignment. pattern-projected sequential pattern mining, in: Proceedings of the Sixth
Two types of refinement methods have been developed to improve International Conference on Knowledge Discovery and Data Mining, ACM,
2000, pp. 355–359.
the primitive alignment. A dataset from a real chemical plant has [20] J. Pei, J. Han, B. Mortazavi-Asl, H. Pinto, Q. Chen, U. Dayal, M.-C. Hsu,
been used to test the effectiveness of the proposed algorithm and PrefixSpan: Mining sequential patterns efficiently by prefix-projected pattern
compare it with the exhaustive search approach in [39]. The results growth, in: Proceedings of 29th International Conference on Data
Engineering, IEEE, 2001, pp. 215–224.
show the proposed algorithm has a significant improvement in [21] S.B. Needleman, C.D. Wunsch, A general method applicable to the search for
efficiency with a small accuracy cost. Based on this work, we can similarities in the amino acid sequence of two proteins, J. Mol. Biol. 48 (3)
provide early prediction for an incoming alarm flood by matching (1970) 443–453.
[22] T.F. Smith, M.S. Waterman, Identification of common molecular
an online alarm sequence with a pattern database and conduct- subsequences, J. Mol. Biol. 147 (1) (1981) 195–197.
ing similarity calculation. The discovered patterns are also helpful [23] S.F. Altschul, W. Gish, W. Miller, E.W. Myers, D.J. Lipman, Basic local
in alarm management and causality analysis. In the future, tech- alignment search tool, J. Mol. Biol. 215 (3) (1990) 403–410.
[24] W.R. Pearson, D.J. Lipman, Improved tools for biological sequence
niques to improve the accuracy and methods for parameter tuning
comparison, Proc. Natl. Acad. Sci. 85 (8) (1988) 2444–2448.
may be studied. [25] M.S. Johnson, R.F. Doolittle, A method for the simultaneous alignment of three
or more amino acid sequences, J. Mol. Evol. 23 (3) (1986) 267–278.
[26] D.-F. Feng, R.F. Doolittle, Progressive sequence alignment as a prerequisitetto
Acknowledgments
correct phylogenetic trees, J. Mol. Evol. 25 (4) (1987) 351–360.
[27] J.D. Thompson, D.G. Higgins, T.J. Gibson, CLUSTAL W: improving the
This work was supported by NSERC and the National Natural Sci- sensitivity of progressive multiple sequence alignment through sequence
weighting, position-specific gap penalties and weight matrix choice, Nucleic
ence Foundation of China (Grant No. 61433001 and 61873142). We
Acids Res. 22 (22) (1994) 4673–4680.
would like to thank Suncor for providing the data in the industrial [28] S.R. Eddy, Multiple alignment using hidden Markov models, Proceedings of
case study. International Conference on Intelligent Systems for Molecular Biology (1995)
114–120.
[29] S.R. Eddy, Profile hidden Markov models, Bioinformatics 14 (9) (1998)
References 755–763.
[30] S. Kordic, P. Lam, J. Xiao, H. Li, Analysis of alarm sequences in a chemical plant,
[1] ISA 18.2, Management of Alarm Systems for the Process Industries, The in: Lecture Notes in Computer Science, Springer, 2008, pp. 135–146.
International Society of Automation, Research Triangle Park, 2016. [31] K. Ahmed, I. Izadi, T. Chen, D. Joe, T. Burton, Similarity analysis of industrial
[2] C.R. Dal Vernon, J.L. Downs, D. Bayn, Human performance models for response alarm flood data, IEEE Trans. Automat. Sci. Eng. 10 (2) (2013) 452–457.
to alarm notifications in the process industries: an industrial case study, in: [32] J. Folmer, B. Vogel-Heuser, Computing dependent industrial alarms for alarm
Proceedings of the Human Factors and Ergonomics Society Annual Meeting, flood reduction, in: Proceedings of 9th International Multi-Conference on
SAGE Publications, 2004, pp. 1189–1193. Systems, Signals and Devices, IEEE, 2012, pp. 1–6.
[3] EEMUA, Alarm Systems — A Guide to Design, Management and Procurement, [33] S. Charbonnier, N. Bouchair, P. Gayet, Fault template extraction to assist
Engineering Equipment and Materials Users Association, London, 2013. operators during industrial alarm floods, Eng. Appl. Artif. Intel. 50 (2016)
[4] IEC 62682, Management of Alarm Systems for the Process Industries, 32–44.
International Electrotechnical Commission, 2014. [34] S. Charbonnier, N. Bouchair, P. Gayet, A weighted dissimilarity index to isolate
[5] C. Timms, Hazards equal trips or alarms or both, Process Saf. Environ. Protect. faults during alarm floods, Control Eng. Pract. 45 (2015) 110–122.
87 (1) (2009) 3–13. [35] E.N. Satuf, E. Kaszkurewicz, R. Schirru, M.C.M.M. de Campos, Situation
[6] S.R. Kondaveeti, I. Izadi, S.L. Shah, T. Chen, On the use of delay timers and awareness measurement of an ecological interface designed to operator
latches for efficient alarm design, in: 19th Mediterranean Conference on support during alarm floods, Int. J. Ind. Ergon. 53 (2016) 179–192.
Control & Automation, IEEE, 2011, pp. 970–975. [36] P. Cisar, E. Hostalkova, P. Stluka, Alarm rationalization support via correlation
[7] S.R. Kondaveeti, I. Izadi, S.L. Shah, D.S. Shook, R. Kadali, T. Chen, Quantification analysis of alarm history, in: Proceedings of 19th International Congress of
of alarm chatter based on run length distributions, Chem. Eng. Res. Design 91 Chemical and Process Engineering, Prague, Czech Republic, 2010, pp. 1–6.
(12) (2013) 2550–2558. [37] Y. Cheng, I. Izadi, T. Chen, Pattern matching of alarm flood sequences by a
[8] E. Naghoosi, I. Izadi, T. Chen, A study on the relation between alarm modified Smith-Waterman algorithm, Chem. Eng. Res. Des. 91 (6) (2013)
deadbands and optimal alarm limits, in: Proceedings of 2011 American 1085–1094.
Control Conference, IEEE, 2011, pp. 3627–3632. [38] W. Hu, J. Wang, T. Chen, A local alignment approach to similarity analysis of
[9] F. Yang, S. Shah, D. Xiao, Signed directed graph based modeling and its industrial alarm flood sequences, Control Eng. Pract. 55 (2016) 13–25.
validation from process knowledge and process data, Int. J. Appl. Math. [39] S. Lai, T. Chen, A method for pattern mining in multiple alarm flood
Comput. Sci. 22 (1) (2012) 41–53. sequences, Chem. Eng. Res. Des. 117 (2017) 831–839.
[10] P. Duan, F. Yang, S.L. Shah, T. Chen, Direct causality detection via the transfer [40] IEC 61511, Functional Safety: Safety Instrumented Systems for the Process
entropy approach, IEEE Trans. Control Syst. Technol. 21 (6) (2013) 2052–2066. Industry Sector, International Electrotechnical Commission, 2015.