Accepted Manuscript

Title: Framework for a Smart Data Analytics Platform towards

Process Monitoring and Alarm Management

Authors: Wenkai Hu, Sirish L. Shah, Tongwen Chen

PII: S0098-1354(17)30355-1
DOI: https://doi.org/10.1016/j.compchemeng.2017.10.010
Reference: CACE 5916

To appear in: Computers and Chemical Engineering

Received date: 29-5-2017

Revised date: 10-10-2017
Accepted date: 11-10-2017

Please cite this article as: Hu, Wenkai., Shah, Sirish L., & Chen,
Tongwen., Framework for a Smart Data Analytics Platform towards Process
Monitoring and Alarm Management.Computers and Chemical Engineering
https://doi.org/10.1016/j.compchemeng.2017.10.010

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
4 CONTROLLERS

Framework for a Smart Data Analytics Platform towards Process

Monitoring and Alarm Management

Wenkai Hu1, Sirish L. Shah2, Tongwen Chen1

1Department of Electrical & Computer Engineering, University of Alberta
2Department of Chemical & Material Engineering, University of Alberta, Edmonton, AB, Canada T6G
1H9

Highlights:
 A framework for a smart data analytics platform is presented.
 The platform integrates information from process and alarm databases complimented with
process connectivity.
 Main features include the alarm data analysis, alarm system design, process data analysis, and
causality inference.
 Case studies involving real industrial data demonstrate the practical utility of the tools.

Abstract
The fusion of information from disparate sources of data is the key step in devising strategies for a smart analytics
platform. In the context of the application of analytics in the process industry, this paper provides a framework for
seamless integration of information from process and alarm databases complimented with process connectivity
information. The discovery of information from such diverse data sources can be subsequently used for process
and performance monitoring including alarm rationalization, root cause diagnosis of process faults, hazard and
operability analysis, safe and optimal process operation. The utility of the proposed framework is illustrated by
several successful industrial case studies.

Keywords
Analytics; Big data; Performance monitoring; Process monitoring; Alarm systems; Process data analytics; Fault
detection and diagnosis.

1 Introduction1
The operation of modern industrial facilities has become highly automated with the deployment of computerized
control systems, such as the Distributed Control System (DCS) and the Supervisory Control and Data Acquisition
(SCADA) system. However, due to large scale and complex system dynamics, abnormal situations do occur from time to
time and such events easily propagate along the interconnected pathways to cause significant and catastrophic disruptions
(Wang et al., 2016). This has been the main motivation factors for studies to ensure process safety, reliable operation, and
compliance with environmental requirements, via tools for better process monitoring and alarm management. Massive
amount of data in the DCS and SCADA systems contain rich information about process operations, making it
indispensable assets for decision making processes. However, without effective analytical tools, such data would only be
compressed and archived for record keeping rather than being turned into valuable resources to facilitate decision-making
(Qin, 2014).

This work was supported by the Natural Sciences and Engineering Research Council of Canada.
[Type text]
 Process data analytic methods rely on the notion of sensor fusion whereby data from many sensors and alarm tags are
combined with process information, such as physical connectivity of process units, to give a holistic picture of health of an
integrated plant. In this study, the process data has a broader meaning compared to many existing publications, which
mainly refer it to as the continuous-valued sensor data. In this paper, process data analytics includes the analytics of sensor
data, alarm and event data, as well as the process connectivity information.
The discovery and learning from process and alarm data refer to a set of tools and techniques for modeling and
understanding of complex data sets. Such data sets generally include normal numerical (or non-categorical) data but should
also take into account categorical (or non-numerical or qualitative) data from Alarm and Event (A&E) logs combined with
process connectivity or topology information. The later refers to the capture of material flow streams in process units as
well as information flow-paths in the process due to control loops. This is particularly useful when one is analyzing data
from highly integrated processes to understand propagation of process faults as would be required in HAZard and
OPerability (HAZOP) analysis for safe process operation. Highly interconnected process plants are now a norm and the
analysis of root causes of process abnormality including predictive risk analysis is non-trivial. It is the extraction of
information from the seamless fusion of process data, alarm and event data and process connectivity that should form the
backbone of a viable process data analytics platform. This paper focuses on an attempt to create such a platform. This idea
of information fusion in the context of process data analytics is depicted in Fig. 1.
For efficient and informative analytics, data analysis is ideally carried out in the temporal as well as spectral domains,
on a multitude and NOT singular sensor signals to detect process abnormality, ideally in a predictive mode. With the
explosion of applications of analytics in diverse areas (such as aircraft engine prognosis, medicine, sports, finance,
insurance, social sciences and the advertising industry) statistical learning skills are in high demand. The emphasis in this
study is on tools and techniques that help in the process of understanding data and discovering information that would lead
to predictive monitoring and diagnosis of process faults, alarm rationalization and safe and optimal process operation.
Typical process data analytic methods require the execution of following steps:
1) Data quality assessment, such as outlier detection, data normalization, and noise filtering;
2) Data visualization and segmentation;
3) Process and performance monitoring including root cause detection of faults;
4) Alarm data analysis;
5) Data-based process topology discovery and validation.

The focus of this paper is to introduce a framework for a smart analytics platform supported by industrial case studies
to demonstrate the practical utility of such a tool. This smart analytics platform has the following unique features: 1) rich
functionalities of alarm data analysis, including alarm data visualization, alarm similarity analysis, quantification of
chattering alarms, design of delay timers, oscillating alarm analysis, similarity analysis of alarm floods, and causality
inference for alarms; 2) powerful alarm configuration analysis for univariate alarm systems, including the design of alarm
limits, filters, delay timers, and deadbands based on continuous-valued sensor data; 3) new methods for process data
analysis, such as the spectral envelope and spectral principle components analysis; as well as 4) useful techniques for
connectivity and causality analysis, such as the spectral correlation and transfer entropy approaches.
The rest of this paper is organized as follows: First, main features of this data analytics platform are introduced in
detail, including several functional modules, such as the alarm data analysis, alarm system design, process data analysis,
and causality inference. To demonstrate the utility of these functional modules, several case studies involving real
industrial data are presented. The concluding remarks are given in the last section.
2 Framework for a Data Analytics Platform
A comprehensive platform should integrate a variety of basic statistical functions as well as advanced analytical
features, which are powerful and insightful in analyzing either continuous-valued process data or binary-valued alarm data
as well as additional categorical data. The proposed analytics platform consists of a data loading section and four
functional modules as shown in Fig. 2. “Data Loading” imports, reorganizes, merges, or exports alarm data and/or process
data. The functional module “Alarm Data Analysis” provides analytic and reporting functions to analyze and visualize
alarm data. The remaining three functional modules are based on process data. The “Alarm Configuration Analysis”
module designs univariate alarm systems for specific process variables. The “Process Data Analysis” module visualizes
and analyzes process data from either time or frequency perspective. The “Connectivity & Causality Analysis” module
uncovers correlations and causal relations between process variables. In addition, a data summary section displays the
basic information of loaded alarm data and/or process data. Details and features in each functional module are presented in
the following subsections. The framework allows merging of process and alarm data, and gives exploratory as well as
analytical insights in the extraction of information from such data.
2.1 Data loading
“Data Loading” is the first highlighted part in Fig. 2. It has six functions that can fulfill different functions related to
data loading, including alarm data loading, process data loading, data exporting, preprocessed data loading, data matching,
and data clearing. The descriptions of these functions are listed in Table 1.
To import alarm historian data to the platform, an Alarm & Event (A&E) log file in Microsoft Excel format is needed.
Databases from various vendor systems can be exported into this toolbox by first converting them into Excel files. Several
requirements on the log file should be satisfied: (1) Each row should reflect one event message; (2) each column should be
a certain field of event messages; (3) the same column in all the sheets of the log file should represent the same field. A
typical A&E log usually consists of configuration attributes, e.g., the tag name, alarm identifier, priority and location, and
realtime messages such as alarm occurrences (ALM), return-to-normal instants (RTN), and their time stamps (Izadi et al.,
2010; Kondaveeti et al., 2012). Among these attributes and messages, the following pieces are key and necessary: time
stamp, tag name, tag identifier, and message type. They may have different field headers in the data archived from different
vendor systems, but usually all of these four pieces of information are provided. In addition to this, the priority and unit
information are optional for the data loading depending on whether the two pieces of information are archived or not. An
example of A&E log in Excel format is shown in Fig. 3. The four mandatory attributes and two optional attributes are
highlighted by square-dotted red and fine dashed blue rectangles, respectively.
To import historical process data to the platform, files that store historical measurements of process variables in
Microsoft Excel format are needed. Process data has totally different format compared to alarm data. Several requirements
on the Excel files should be satisfied: (1) The first column should be time stamps of the sampling instants; (2) the
following columns store the historical values of process variables at these sampling instants. If the time stamp information
is not available, an Excel file without the time stamp column is also acceptable. The log file of process data has a much
simpler structure compared with the A&E log, and usually includes three parts, namely, tag names, measurements, and
time stamps. An example of a process data stream is presented in Fig. 4. The tag names, measurements, and time stamps
are highlighted by round-dotted green, fine dashed blue, and square-dotted red rectangles, respectively.
Once the alarm and/or process data are loaded, the platform will reorganize them in a format that can be easily
processed by the analytical functions. Exporting data before closing the platform is recommended, since it is much less
time-consuming to import the preprocessed data in .mat format than to import the raw data stored in Excel, especially for
subsequent analysis. If both the alarm data and process data are loaded, then all the process variables that have
observations during the time-window of the alarm dataset are listed. Macros can be created to associate alarm tags with a
listed process variable. Some basic information about the data set is shown in the information section at the bottom left
corner of Fig. 2. It shows the directories of the alarm data and process data files, respectively. Alarm historian duration,
average alarm rate, number of alarm tags, number of process tags, and number of process tags that have been matched to
alarm tags are also provided if available.

2.2 Alarm data analysis

This functional module provides a variety of functions for the analysis of alarm data, including basic statistical
features, such as the reporting of top bad actors, calculation of average/peak alarm rates, and OPerator Acknowledgement
Analysis (OPAA), advanced data analytics, such as the Run Length Distribution & Delay Timer Analysis (RLD&DTA),
Chattering Index (CI), Oscillating Alarm Analysis (OAA), Alarm Flood Analysis (AFA), Causality Inference for Alarms
(CIA), and Mode-Dependent Alarm Analysis (MDAA), plus powerful visualization plots, such as the High Density Alarm
Plot (HDAP), Alarm Burst Plot (ABP) and Alarm Similarity Color Map (ASCM). This functional part is helpful for
issuing weekly or monthly assessments/recommendations of alarm management. But it is not merely a simple Key
Performance Indicator (KPI) calculator as it has many advanced features that can help in alarm rationalization. The
descriptions of these functions are listed in Table 2.
To observe top bad actors and visualize changes of KPIs, HDAP and ABP can be used. A High Density Alarm Plot
(HDAP) is useful for visualizing large amounts of alarm data over a selected period (Kondaveeti et al., 2012). It displays
alarm counts for top bad actors using a color map and provides an overall picture of alarm data without getting into details
of each alarm variable. Using a sliding time window of 10 minutes, the peak alarm rate is calculated along with time and
can be visualized using a line graph, namely, the Alarm Burst Plot (ABP) (Hollifield & Habibi, 2011).
Correlated alarms are referred to as alarms occurring within a short time period of each other. They could be either
redundant or overlapping in indicating the same abnormality. Therefore, the detection and quantification of correlated
alarms are important. Based on the alarm correlations, redundant alarms can be removed and related alarms can be
grouped. A variety of methods have been developed and demonstrated to be effective using industrial case studies (Noda et
al., 2011; Kondaveeti et al., 2012; Yang et al., 2013; Hu et al., 2015). In this functional part, the correlation metrics are
visualized by an Alarm Similarity Color Map (ASCM), where correlated alarms are clustered, and thus can be easily
identified (Kondaveeti et al., 2012; Yang et al., 2012).
 Chattering alarms are major contributors of alarm overloading. According to ANSI/ISA-18.2 (2009), any alarm
occurring more than 3 times over a 60 seconds period is likely to be chattering. To identify chattering alarms and quantify
their severities, a Chattering Index (CI) was developed based on Run Length Distribution (RLD) (Naghoosi et al., 2011;
Kondaveeti et al., 2013). CI takes values between 0 and 1, with a value closer to 1 corresponding to a more serious
chattering problem. To reduce chattering alarms, delay timers are effective tools (Kondaveeti et al., 2013; Wang & Chen,
2013 & 2014). Kondaveeti et al. (2013) and Adnan et al. (2013) provided effective ways for the design of delay-timers,
which is available in RL&DTA.
An oscillating alarm is a special case of chattering or repeating alarms, which is caused by oscillatory processes. An
oscillating alarm, as identified by periodicity in the alarm tag, is an indication of some underlying process oscillation. The
oscillating alarms can be easily detected offline using the method in (Cheng, 2013) and online using the method in (Wang
& Chen, 2013). Specifically, the offline method, namely OAA, is incorporated in this platform, to detect oscillating alarms
based on the periodicity in alarm states.
Alarm floods typically arise during a situation when a process abnormality propagates leading to triggering of a large
number of annunciated alarms over a short period that often exceeds the operator’s ability to respond in a timely manner to
mitigate the fault(s). An alarm flood is said to be raised when the number of annunciated alarms reaches 10 alarms over a
10 minutes period per operator, and be cleared when the number drops below 5 alarms over a 10 minutes period (ISA-18.2,
2009; EEMUA-191, 2013; IEC, 2014). Alarm floods are common in alarm systems and have various negative effects
(Nimmo, 2005; Timms, 2009; Beebe et al., 2013). To analyze alarm floods, AFA functions enable analysis in two steps.
First, alarm floods are identified from historical alarm data and highlighted in an alarm burst plot. Second, alarm floods are
compared in pairwise using sequence alignment algorithms, such as Dynamic Time Warping (DTW) (Müller, 2007;
Ahmed et al., 2013), modified Smith Waterman (SW) algorithm (Smith & Waterman, 1981; Cheng et al., 2013a), and
BLAST-like algorithm (Altschul et al., 1990; Hu et al., 2016a). Furthermore, sequence patterns of alarm floods can be
found using a multiple sequence alignment algorithm in (Lai & Chen, 2015).
To identify abnormality propagation paths and assist in the root cause detection, the CIA function provides a practical
way to detect causal relations between alarms. Based on the detection results, users are able to make a more reliable
judgement on how an abnormality propagates from one alarm to another and relate it to experiences from similar abnormal
events in the past (Hu et al., 2017). In this way the alarm flood event clustering analysis allows one to develop a canonical
fingerprint of common faults that trigger an alarm flood and also identify appropriate corrective actions that need to be
taken to mitigate the abnormality rapidly.
To analyze the interactions between alarms and operator responses, the OPAA and MDAA functions have been
developed. The OPAA includes three plots to visualize the alarm states and operator acknowledgements (Ack’s) in a high
density color map, compare the alarm count and Ack’s count for each alarm tag using a bar chart, and show the response
time for acknowledging an alarm using a boxplot. The MDAA function discovers the association rules of mode-dependent
alarms from A&E logs, where both the alarm data and operator actions should be available. The results can be used to
assist in configuring state-based alarming strategies. Moreover, another advanced technique to analyze the interaction
between alarms and operator responses is process discovery of operational procedures (Hu et al., 2016b). The results can
be used to provide decision supports by analyzing operator actions from historical data.
In addition to the above analytical functions, reporting functions are also useful. Table 3 lists three reporting functions,
including the Bad Actor List, Performance Calculator, and Bad Actor Comparison. A bad actor list contains statistical
results of alarm data. It does not only indicate top bad actors, but also tells distributions of alarm priorities, alarm
identifiers, and unit areas. Performance Calculator gives results of how many alarms can be reduced by applying delay
timers to bad actors and suppressing specified nuisance alarms. Bad Actor Comparison compares the bad actor list of the
loaded alarm historian with another one for the same alarm system but at a different time period, thus allowing audit
checks to determine if implemented changes have helped.

2.3 Alarm configuration analysis

This functional part is used to simulate the configuration of alarm systems for univariate process signals, as is the
current practice in industry. Techniques, such as filters, delay timers, and deadbands, are provided. Moreover, there are
several ways to determine alarm limits. The user can manually define them, set a certain alarm rate, select the maximum or
minimum historical value, or make the platform find an optimal limit automatically. The platform then can calculate,
display, and compare the design results based on different techniques. The descriptions of these functions for alarm
configuration are listed in Table 4.
Filters are used to reduce noises from process signals. As a result, the process data during normal and abnormal
situations will be easily separated, which is an effective way to minimize false and missed alarms. Six commonly used
industrial filters are provided, including the moving average filter, moving variance filter, moving norm filter, rank order
filter, low pass filter, and Exponentially Weighted Moving Average (EWMA) filter. The mathematical principles of these
filters can be found in (Izadi et al., 2009; Cheng et al., 2013b). In contrast to filters, delay timers work on alarm signals
rather than process signals. Two types of delay timers are provided, namely, the off-delay timer and on-delay timer. The
off-delay timer reduces chattering alarms by delaying return-to-normal instants while the on-delay timer reduces chattering
alarms or removes fleeting alarms by delaying alarm occurrences. The principle and design of delay timers can be found in
(Adnan et al., 2011; Adnan et al., 2013). Adding a deadband is another widely used technique in industry. This requires
two different alarm limits for the raising and clearing of alarms (Adnan et al., 2011). Chattering alarms caused by noise can
be effectively reduced by deadbands.
To design the above techniques, three performance specifications are used commonly, including the False Alarm Rate
(FAR), Missed Alarm Rate (MAR) and Averaged Alarm Delay (AAD). FAR and MAR measure the accuracy in detecting
abnormal situations, and AAD denotes the alarm latency (Adnan et al., 2011; Xu et al., 2012). However, tradeoff exists
between FAR and MAR. It is impossible to reduce FAR and MAR simultaneously, by just adjusting the alarm limit. Thus,
filters, delay timers, and deadbands can be applied to improve the performance of alarm systems. A Receiver Operating
Characteristic (ROC) curve is provided to visualize the tradeoff between FAR and MAR, for the design of alarm systems.

2.4 Process data analysis

This functional module provides a variety of techniques to analyze and visualize process signals from different
perspectives of view. Prior to analysis of process data, data preprocessing is usually required. Four commonly used data
preprocessing techniques are incorporated in the platform, including data detrending, data smoothing, data normalization,
and outlier removal. The descriptions of these data preprocessing functions are listed in Table 5.
The data analytical functions are classified into two groups as listed in Table 6. The analytical functions and
visualization plots for univariate process signals include time trend, frequency spectrum, power spectrum density, and
spectrogram, and those for multivariate process signals include high density plots, parallel coordinate plots, scatter
matrices, spectral envelopes, and temporal and spectral Principle Components Analysis (PCA and SPCA).
Among the univariate plots, the frequency spectrum and power spectrum density are in frequency domain, and the
spectrogram is in the time-frequency or wavelet domain. A spectrogram plot is a 2 dimensional colormap that can reflect
both the temporal and spectral information simultaneously for a single process variable. High density plots visualize the
plots of a multitude of process variables in a compact form, namely, in a single plot. Parallel coordinate plots and scatter
matrices are commonly used to visualize multivariate process signals, so that the relations between process variables can
be easily observed. PCA is a commonly used technique to analyze large multivariate datasets (Jolliffe, 2002). It reduces
data dimensionality and distills important information based on the dependencies between process variables. Spectral
envelopes (Stoffer, 1999; Stoffer et al., 2000) and spectral PCA (Thornhill et al., 2002) are effective in analyzing the
spectral behavior of process signals, the results of which could help to diagnose plant oscillations (Jiang et al., 2007,
Tangirala et al., 2007). In addition, the platform also provides functions to visualize process signals and their associated
alarm signals in a single plot, and to calculate statistical values, such as the mean, median, variance, compression factor,
and quantization factor. The latter functions are particularly useful for assessing data quality.

2.5 Connectivity and causality analysis

An often overlooked part of analytics is causation analysis. While this information is available in a Piping and
Instrumentation Diagrams (P&ID), the connectivity or causality information is not always up to-date and not easily
available in mathematical forms. This data-based functional part is used to detect correlations and causal relations between
process variables to capture material and information flow paths in the process. Four functions are provided, including the
correlation coefficient, spectral correlation, Granger causality, and Transfer Entropy (TE). The calculated correlations and
causal relations are visualized using correlation color maps, and Signed Directed Graphs (Yang et al., 2010). Table 7
describes the functionalities and capacities of these functions.
Correlation coefficient is calculated by assuming a certain time lag between two time series. Accordingly, the real
correlation between two process variables is achieved as the absolute maximum value of correlation coefficients (Welch,
1974; Yang et al., 2012). Spectral correlation is a specific application of correlation coefficient to the power spectrums of
process signals (Tangirala et al., 2005). It indicates the similarities in the spectral ‘shapes’ of all signals; for example, this
can reveal all process variables that are oscillating at the same frequency.
However, correlation does not indicate a causal relation. The directions of interactions between process variables are
usually unknown from correlations. To detect how process variables influence each other and how abnormalities propagate
through processes, causality inference is effective. There are a variety of causality inference techniques based on different
resources (Chiang & Braatz, 2003; Thambirajah et al., 2009; Jiang et al., 2009; Schleburg et al., 2013; Yang et al., 2014).
The techniques, namely, Granger causality and transfer entropy provided in this platform are based on process history.
Granger causality was first proposed by Granger in 1969, based on two assumptions: (1) the cause occurs before the effect;
(2) the cause contains information about the effect that is unique (Granger, 1969). However, this method only works for
linear processes. By contrast, the transfer entropy method is effective for both linear and nonlinear processes (Schreiber,
2000; Kaiser & Schreiber, 2002; Bauer et al., 2007). Furthermore, Direct Transfer Entropy (DTE) can be used to find
direct dependencies between process variables (Duan et al., 2013). To circumvent the assumptions of stationary processes
and Gaussian distributions, a Trasfer 0-Entropy (T0E) was proposed (Nair, 2013; Duan et al., 2015).

3 Case Studies for Alarm Data Analytics

This section illustrates the application of alarm data analytics of the smart platform based on case studies involving
real industrial A&E data sets.

3.1 Analysis of bad actors and removal of nuisance alarms

The alarm data used in this case study was collected from an oil plant over a time period of 10 days. Totally, 173
alarm tags were found to occur with an average of 8.5 alarms over a 10 minutes period, which was much higher than the
benchmark threshold of an efficient alarm system, namely, 1 alarm over 10 minutes. To explore why this alarm system had
such a high alarm rate over the selected time period, the alarm data analytical functions of this smart platform are applied.
Fig. 5 shows a high density alarm plot of the top 20 bad actors over the selected time period of 10 days. This Fig.
contains 864,000 bits of alarm data for each of the 20 tags. The color bar denotes the number of alarms in each 10 min
time bin. The red and orange colors indicate high alarm rates, implying chattering or repeating alarms. It is obvious that
“Tag102.CFN” and “Tag101.CFN” were likely chattering over some short time periods, and “Tag192.FAILED” kept
repeating for the whole time period. Moreover, it can be seen that two alarms, “Tag64.COMM” and “Tag60.IOF”, were
annunciated almost simultaneously in the first two days of the selected time period. Around 11pm on Aug. 23 rd, there was a
high chance that most top bad actors occurred, implying a plant upset around this time instant.

Fig. 6 shows an alarm similarity color map of the top 20 bad actors. Alarm tags are clustered based on their
correlations. The darker color of a block indicates a higher correlation between alarms. The diagonal of the map consists of
1’s (black squares), indicating the highest correlations of alarm tags with themselves. In this case study, only one pair of
alarms was found to be correlated, namely, “Tag64.COMM” and “Tag60.IOF”. The correlation value was 0.9, indicating a
strong relation.
Fig. 7 displays the chattering indices of the top 20 bad actors using red bars. The green line denotes the threshold of
chattering alarms based on ANSI/ISA-18.2 (2009) standard (no more than 3 alarms over a 60 second period). Any
chattering index that exceeds this threshold indicates a chattering alarm. Among these bad actors, 7 alarms were
determined to have chattering problems. To design delay timers for these chattering alarms, the run length distribution is
used. Fig. 8 shows an example of designing off-delay timer for the topmost bad actor “Tag102.CFN”. The red curve
indicates the alarm count that can be reduced by implementing an off-delay timer with the value of the delay timer on the
horizontal axis. For instance, an off-delay timer of 10 sec reduces 82% of the alarm occurrences. The off-delay timer turns
the chattering alarms into standing alarms. Accordingly, the alarm count was reduced from 4856 to 875 over 10 days.
In the same manner, the types and values of delay timers were recommended for all the seven chattering alarm tags as
shown in Table 8. It is noteworthy that small off-delay timers were very effective in reducing most chattering instants for
these alarm tags. By applying the recommended off-delay timers, the average alarm rate can be reduced to 4.7 alarms over
10 minutes, which is 45% lower than the original alarm rate, namely, 8.5 alarms per 10 minutes.

3.2 Identification, comparison and clustering of alarm floods

This subsection illustrates the functions of alarm flood analysis, including the identification, comparison, and
clustering of alarm floods. The same alarm data set in the previous subsection is used here. Based on benchmark thresholds
of occurrence and clearing of an alarm flood (ISA-18.2 2009; EEMUA-191 2013; IEC 2014), alarm floods were easily
identified using an alarm burst plot shown in Fig. 9. The black line indicates the threshold of identifying the occurrence of
an alarm flood, namely 10 alarms over 10 minutes. The alarm floods are highlighted by blue blocks. It can be seen that this
plant was in the alarm flood situation for a considerable time period.
Fig. 10 shows the alarm burst plot for the alarm data with chattering alarms reduced using a uniform off-delay-timer
(40 sec) applied to all alarm tags. It can be seen that significant reduction was achieved by applying the off-delay-timer.
Originally, there were 66 alarm floods in Fig. 9 and they occurred for almost 25.4% of the entire time period of 10 days.
After applying the off-delay timer, there were only 38 alarm floods left and they occurred during 6.4% of this time period.
Having removed the chattering alarms, these alarm floods can be considered as true alarm floods, requiring further analysis,
so as to prevent the occurrence of the same root cause.
To detect frequent alarm sequences of the 38 alarm floods, sequence alignment algorithms are used. The similarity
indices for all pairs of alarm floods are calculated and shown as a similarity color map in Fig. 11, where similar alarm
floods are clustered. The vertical and horizontal axes display the event number of the 38 alarm floods. A smaller event
number index refers to an earlier occurrence in time. The color bar at the right side of the cluster map indicates the strength
of similarity indices. The diagonal of the color map represents the similarity between each alarm flood event and itself, and
is naturally 1. This functionality exploits sequence alignment algorithms for similarity analysis between alarm floods. More
specifically, the results in this case study are obtained based on the modified Smith-Waterman algorithm (Cheng et al.,
2013a), which has high computational complexity as discussed in (Cheng et al., 2013a; Hu et al., 2016a). In this case study,
the computation for alarm flood similarity analysis took only 15 sec (on a computer with 4 GB RAM, 2.3 GHz CPU, and
64-bit Windows System), which was quick because there were only 38 short alarm floods. However, if there are many
alarm floods or alarm flood sequences are long, the computation might not be able to finish within a tolerable period. Then,
we can exploits the BLAST-like algorithm in (Hu et al., 2016a) as an alternative.
To show how similar alarm floods resemble each other, an example is given in Fig. 12. The No. 12 and No. 14 alarm
floods (as highlighted in the red rectangle in Fig. 11) were found to be very similar. Fig. 12 presents the sequence
alignment between the two alarm floods. It can be seen that they share a long sequence of common alarms. According to
their time stamps, this may indicate that the abnormality causing No. 12 alarm flood had not been well solved, and thus
appeared again in the afternoon of the same day, causing No. 14 alarm flood.

3.3 Causality inference using alarm data

This subsection illustrates the alarm data based causality inference technique. The alarm data was collected from a
hydrogen plant over 3.5 days. Totally, 288 alarm tags were found from the A&E log. The causal relations between each
pair of alarms are detected by applying the TE based causality inference. Six alarm signals are selected to demonstrate the
causality inference technique. The historical data samples of the six alarm tags are shown in Fig. 13. Over the selected time
period, the six alarm signals were annunciated 653, 242, 242, 242, 219, and 106 times, respectively. More detailed results
of this case study appear in (Hu et al., 2017).
By setting a maximum value of the time lag (100 sec in this case study), the Normalized Transfer Entropies (NTEs)
under different time delays between each pair of alarm signals were calculated and shown in Fig. 14. The maximum values
of NTEs are regarded as the causal strengths.
Meanwhile, significance thresholds are calculated from Monte Carlo tests using the method in (Hu et al., 2017). The
maximum NTEs and their corresponding significance thresholds (in brackets) are given in Table 9.
Based on Table 9, the causal relations are found between alarms with NTEs larger than the significance thresholds. As
a result, a causal map describing the information flow paths is drawn in Fig. 15.
Causal relations between two variables can be direct or indirect if mediated by an intermediate variable. It is therefore
important to be able to differentiate between such relations as has been done in (Hu et al., 2017). The Normalized Direct
Transfer Entropies (NDTEs) are calculated for alarm pairs with causal relations in Fig. 15. By comparing the NDTEs with
the corresponding significance thresholds, indirect causalities are found and excluded. Accordingly, a causal map
describing all direct information flow paths is shown in Fig. 16. This conclusion is consistent with the process knowledge
presented in (Hu et al., 2017).

4 Case Studies for Alarm Configuration Analysis

This section illustrates the techniques for the design of univariate alarm systems in the functional part “Alarm
Configuration Analysis” of the smart platform. An example is given based on the process signal shown in Fig. 17. In the
foremost step, the normal and abnormal parts of the process signal are specified as the blue and red sections in Fig. 17.
Assuming the original high alarm limit to be 1 (HAL =1), alarms are generated as the vertical red lines in the left plot
of Fig. 18. The distributions of normal data and abnormal data are shown in the right plot of Fig. 18. As a result, there are
793 alarm occurrences. The FAR and MAR rates are 1.22% and 14.36%, respectively.
In the first design scenario, the high alarm limit is redesigned using an optimization function. The optimal alarm limit
for this process signal is 0.72. The alarm signal is generated and shown in the left plot of Fig. 19. As a result, the number
of alarm occurrences is reduced to 579. The FAR and MAR rates are 3.00% and 1.80%. Compared with the original
setting, the FAR rate is increased slightly while the MAR is reduced significantly.
In the second design scenario, an off delay timer of 7 samples is used based on the original setting (HAL=1). This off-
delay timer is very effective in reducing chattering alarms. The alarm signal with chattering alarms reduced is shown in the
left plot of Fig. 20. It can be seen that there is almost no alarm occurrence in the abnormal part of the process signal. The
total number of alarm occurrences is reduced to 183. The MAR decreases to 0.03%, which is a significant reduction
compared to the original setting. However, the FAR grows to 6.69% at the same time.
In the third design scenario, a moving average filter of 10 samples is used. Compared with the original setting and the
previous design scenarios, the normal and abnormal parts of the process signal are more separated, as shown in the right
plot of Fig. 21. With the original alarm limit (HAL=1), there are only 35 alarm occurrences in Fig. 21. The FAR and MAR
are 0 and 1.77%, respectively. The two metrics are much lower than those in the original setting.
To visualize the tradeoff between FAR and MAR in the design of alarm system, a ROC curve is drawn in Fig. 22. The
blue and dashed blue curves are the ROC curves for the design scenario with a filter and the original setting. The high
alarm limits for the two cases are set the same. It is obvious that the ROC based on the application of a filter is much closer
to the origin, indicating better alarm system design.

5 Case Studies for Process Data Analytics

This section illustrates the application of process data analytics of this smart platform using case studies involving real
industrial process data.

5.1 Detection of root cause of plant-wide oscillations

This subsection considers the application of fusing process data and process connectivity information to detect and
locate the root-cause of plant-wide oscillations using process data analytical methods.
Detection and diagnosis of plant-wide disturbances is an important issue in many process industries (Qin, 1998,
Desborough & Miller, 2001). Thornhill and HÄagglund (1997) used zero-crossings of the control error signal to calculate
integral absolute error (IAE) in order to detect oscillations in a control loop. This method has poor performance in the
cases of noisy error signals. Miao and Seborg (1999) suggested a method based on the auto-correlation function to detect
excessively oscillatory feedback loops. The auto-covariance function (ACF) of a signal was utilized in Thornhill et al.
(2003a) to detect oscillation(s) present in a signal. This method needs a minimum of five cycles in the auto-covariance
function to detect oscillation, which is often hard to obtain, particularly in the case of long oscillations (e.g., an oscillation
with a period of 400 samples). Although the data set can be down sampled, down sampling may introduce aliasing in the
data. Thornhill et al. (2002) have also proposed Spectral Principal Component Analysis (SPCA) to detect oscillations and
categorize the variables having similar oscillations. This method does not provide any diagnosis of the root cause of the
oscillations which is generally the main objective of the exercise.
A more efficient procedure based on the spectral envelope method for detection and diagnosis of plant-wide
oscillations was proposed by Jiang et al. (2007). The spectral envelope method is a frequency domain technique that was
originally introduced by Stoffer et al. (1993) to explore the periodic nature in time series. The idea is to assign numerical
values to each of the categories followed by a spectral analysis of the resulting discrete-valued time series. Later
McDougall et al. (1997) extended the concept of spectral envelope to real-valued series. In exploring the periodic nature of
a real-valued series, one can do spectral analysis of not only the original series, but also transformed series. The key idea in
McDougall et al. (1997) was to select optimal transformations of a real-valued series that emphasize any periodic nature in
the frequency domain.
In this case study, the process data was collected from a distillation plant of Eastman Chemical Company, USA
(Thornhill et al., 2003b). The Piping and Instrumentation Diagram (P&ID) of this plant is shown in Fig. 23. The collected
data has 48 process variables and 5040 observations sampled at the 20 sec interval. The high density plots of time trends
and power spectra of these process variables are shown in Figs. 24 and 25, respectively. It can be seen from Fig. 25 that
many peaks appear in the power spectra, which indicates the presence and propagation of plant-wide oscillations in many
variables.

To analyze plant oscillations, the power spectral correlations are first calculated and shown as a power spectral
correlation color map in Fig. 26, where process variables with similar power spectra are clustered. The color bar on the
right side of the color map indicates the strength of correlation. The red and orange colors indicate strong correlations and
the green color indicates a weak correlation. The diagonal of the color map represents the correlation between one variable
and itself. Based on Fig. 26, it is easy to identify all process variables that share the same oscillating feature. These would
include Process Variables (PVs) as well as the corresponding Manipulative Variables (MVs). If necessary, the MVs can be
omitted to obtain a short-list of all oscillating PVs.
In this case study the spectral envelope method is used to diagnose the plant oscillation. Fig. 27 shows the calculated
spectral envelope, where a clear peak is observed at the frequency of 0.003175 cycles per sample, which was the frequency
of concern for plant engineers.
The Chi-squared test statistics of the 48 process variables at the oscillation frequency of 0.003175 cycles per sample
are calculated and shown using a bar chart in Fig. 28. The dashed red line denotes the significance threshold of 13.82 at the
significance level of 0.001. As a result, the process variables with Chi-squared values larger than this threshold are
identified to be oscillating at the frequency of 0.003175 cycles per sample.
Fig. 29 shows the Oscillation Contribution Indices (OCIs) of the 48 process variables at the oscillation frequency of
0.003175 cycles per sample. The dashed red line denotes the OCI threshold of 1. Variables that have OCIs larger than 1
are regarded as root cause candidates. Among these variables, “LC2.PV” and “LC2.OP” have the largest OCIs, indicating
the loop associated to tag “LC2” contributes most to the spectral envelope at the frequency of 0.003175 cycles per sample.
Thus, this particular loop should be examined as the first root cause candidate.
Fig. 30 shows the scatter plot between “LC2.PV” and “LC2.OP”. The elliptical pattern indicates a valve stiction, that
caused limit cycles in the loop and then propagated to many other variables. Based on a plant test, there indeed existed a
4% stiction in the valve. Thus, this was exactly the root-cause of the plant-wide oscillations. This root cause was also
validated by the connectivity analysis (Jiang et al., 2009).

5.2 Process connectivity analysis

This subsection illustrates root cause diagnosis of plant-wide oscillations based on process connectivity information.
The process data and connectivity diagram were collected from a process plant operated by Mitsubishi Chemical
Corporation (MCC), Japan (Jiang et al., 2009). The process flow diagraph of this plant is shown in Fig. 31. Taking the
controllers as nodes and connecting them based on their direct interactions, a control loop diagraph is identified as the lines
with red arrows in Fig. 31. Using the spectral envelope method in (Jiang et al., 2007), process oscillations were detected in
19 of the 28 loops. Furthermore to diagnose the cause of these plant-wide oscillations, the analytical method based
connectivity information is used.
Based on direct interactions in Fig. 31, an adjacency matrix is built and shown in Fig. 32. Furthermore, a control
reachability matrix from the adjacency matrix is built and shown in Fig. 33. The 19 oscillatory process variables are
highlighted in blue in Fig. 33.
According to this reachability matrix in Fig. 33, loops 13 (PC1) and 24 (LC2) are the only loops that can reach all the
oscillatory variables. Moreover, these two loops are located in the same area and physically close to each other. Thus,
loops 13 (PC1) and 24 (LC2) can be isolated as root cause candidates. Eventually, a sticky valve was found in loop 13
(PC1). This was identified as the root cause of the plant-wide oscillations as presented in (Jiang et al., 2009).

5.3 Causality inference of complex processes

This subsection demonstrates data-based discovery of causal relations using causality inference methods in the
proposed analytics platform. The sampled data of 5 variables and 3544 observations (sampling time=1 minute) were
collected from a Flue Gas Desulfurization (FGD) process (Duan 2013), the schematic of which is shown in Fig. 34. The
process variables y1, y2, y3, y4, and y5, represent the liquid level of the reactor, the level of Tank 1, the level of Tank 2, the
outlet flow rate of Pump 2, and the outlet flow rate of Pump 3, respectively. The time trends of these variables are shown
in Fig. 35.
The NTEs between all pairs of process variables are calculated and listed in Table 10. By choosing the threshold as
0.02, the causal relations are found as those with NTEs larger than 0.02 (Duan et al., 2013). As a result, the information
flow paths are identified and represented using a signed directed graph in Fig. 36. This Fig. also includes bidirectional
connectivity as detected by the application of transfer entropy method. In many cases, bidirectional connectivity is due to
the presence of material flow paths as well as feedback control loops.
It is also important to know if the detected causal relations are direct or indirect. To achieve this, the NDTEs are
calculated step by step. For example, in the first step, the NDTE from y1 to y3 based on y2 and y4 is calculated to be 0.024
in Table 11. The value is small, indicating no direct causality from y1 to y3. Thus, the path from y1 to y3 is pruned in step 1
in Fig. 37. Analogously, other indirect causalities are confirmed. In this case the causality from y2 to y1 is confirmed to be
direct.
Eventually, the direct information flow paths are detected and shown as a signed directed graph in Fig. 38. The result
is confirmed to be correct according to (Duan et al., 2013). By highlighting these information flow paths in Fig. 39, it is
clear how the process variables affect each other. Based on this, the propagation of abnormalities can be easily identified.
The dotted red arrows in Fig. 38 correspond to feedback loops as validated in Fig. 39. The causality inference in this case
study is based on the transfer entropy approach, which includes two critical steps, including the kernel estimation and the
calculation of transfer entropy metric. Both steps have high computational complexity (Duan et al., 2013). For example,
the computation to determine the causal relations in Fig. 38 took about 10 min on a computer with 4 GB RAM, 2.3 GHz
CPU, and 64-bit Windows Operating System. Towards practical usage, two strategies can be adopted to reduce
computation time: 1) the number of samples and the embedding dimension should not be large; 2) the transfer entropy
approach should be applied to smaller units with a smaller number of variables, or a large-scale complex system should be
broken down into smaller sections prior to applying the transfer entropy method (Duan et al., 2013).
6 Conclusions
To improve process monitoring and alarm management in process industries, a smart platform for the alarm and
process data analytics has been developed. The smart platform consists of a “Data Loading” section and four functional
modules. Specifically, “Alarm Data Analysis” helps to detect and remove nuisance alarms, identify and cluster alarm
floods, discover causal relations, and find mode-dependent alarms. “Alarm Configuration Analysis” provides a variety of
techniques for design of better alarm systems. “Process Data Analysis” analyzes and visualizes the process data from
different perspectives. “Connectivity & Causality Analysis” uncovers causal relations between process variables. This
paper has introduced the major features in each functional module of the platform, and illustrates the effectiveness and
practicability of these features using case studies involving real industrial data. According to the application results, this
framework has illustrated the value in fusing information from disparate data sources and in this respect the proposed
platform is comprehensive in functionalities and powerful in providing insightful conclusions.
The development of new features is ongoing so as to enhance the practical utility of the smart platform. One promising
future work is the process discovery of operator actions in response to alarm notifications. The pattern extracted from the
A&E log is supposed to provide decision support for operators. A preliminary work can be found in (Hu et al., 2016b).
Another promising direction is the root cause analysis of alarm floods. In alarm flood situations, operators may fail to
identify the root causes and miss the critical alarms, which is the main reason of many accidents (Wang et al., 2016). Thus,
if the root causes of alarm floods can be quickly identified in an on-line manner then the operator would be able to
confidently handle critical alarms and make correct responses.
References
Adnan, N.A., Izadi, I., & Chen, T. (2011). On expected detection delays for alarm systems with deadbands and delay-timers. Journal of
Process Control, 21, 1318-1331.
Adnan, N.A., Cheng, Y., Izadi, I., & Chen, T. (2013). Study of generalized delay-timers in alarm configuration. Journal of Process
Control, 23(3), 382-395.
Ahmed, K., Izadi, I., Chen, T., Joe, D., & Burton, T. (2013). Similarity analysis of industrial alarm flood data. IEEE Transactions on
Automation Science and Engineering, 10, 452-457.
Altschul, S. F., Gish, W., Miller, W., Myers, E. W., & Lipman, D. J. (1990). Basic local alignment search tool. Journal of Molecular
Biology, 215, 403–410.
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 entropy. IEEE Trans. Control Systems Technology, 15(1), 12-21.
Beebe, D., Ferrer, S., & Logerot, D. (2013). The connection of peak alarm rates to plant incidents and what you can do to minimize.
Process Safety Progress, 32(1), 72-77.
Cheng, Y., Izadi, I., & Chen, T. (2013a). Pattern matching of alarm flood sequences by a modified Smith-Waterman algorithm.
Chemical Engineering Research and Design, 91, 1085-1094.
Cheng, Y., Izadi, I., & Chen, T. (2013b). Optimal alarm signal processing: filter design and performance analysis. IEEE Transactions
on Automation Science and Engineering, 10, 446-451.
Cheng, Y. (2013). Data-driven Techniques on Alarm System Analysis and Improvement. Doctoral Thesis, University of Alberta.
Chiang, L. H., & Braatz, R. D. (2003). Process monitoring using causal map and multivariate statistics: fault detection and
identification. Chemometrics and Intelligent Laboratory Systems, 65(2), 159-178.
Desborough, L. and R. Miller (2001). Increasing customer value of industrial control performance monitoring - honeywell's experience.
In Proc. of CPC VI. Tuscon, Arizona. 172-192.
Duan, P., Yang, F., Chen, T., & Shah, S. L. (2013). Direct causality detection via the transfer entropy approach. IEEE Trans. Control
Systems Technology, 21(6), 2052-2066.
Duan, P., Yang, F., Shah, S.L., & Chen, T. (2015). Transfer zero-entropy and its application for capturing cause and effect relationship
between variables. IEEE Trans. Control Systems Technology, 23(3), 855-867.
EEMUA (Engineering Equipment and Materials Users' Association) (2013). Alarm Systems: A Guide to Design, Management and
Procurement, Edition 3. London: EEMUA Publication 191.
Granger, C. W. (1969). Investigating causal relations by econometric models and cross-spectral methods. Econometrica: Journal of the
Econometric Society, 424-438.
Hollifield, B., & Habibi, E. (2011). Alarm Management: A Comprehensuve Guide. Research Traingle Park, NC: ISA.
Hu, W., Wang, J., & Chen, T. (2015). A new method to detect and quantify correlated alarms with occurrence delays. Computers &
Chemical Engineering, 80, 189-198.
Hu, W., Wang, J., & Chen, T. (2016a). A local alignment approach to similarity analysis of industrial alarm flood sequences. Control
Engineering Practice, 55, 13-25.
Hu, W., Ahmad W. A., Chen, T., & Shah, S. L. (2016b). Process discovery of operator actions in response to univariate alarms. In
DYCOPS-CAB2016, 1026-1031.
Hu, W., Wang, J., Chen, T., & Shah, S. L. (2017). Cause-effect analysis of industrial alarm variables using transfer entropies. Control
Engineering Practice, 64, 205-214.
IEC (International Electrotechnical Commission) (2014). Management of Alarm Systems for the Process Industries. IEC 62682.
ISA (International Society of Automation) (2009). Management of Alarm Systems for the Process Industries. North Carolina: ISA
18.02.
Izadi, I., Shah, S.L., Shook, D. S., Kondaveeti, S.R., & Chen, T. (2009). A Framework for Optimal Design of Alarm Systems. In
proceedings of the 7th IFAC SAFEPROCESS, Barcelona, Spain.
Izadi, I., Shah, S. L., & Chen, T. (2010). Effective resource utilization for alarm management. The 49th IEEE Conference on Decision
and Control (CDC2010), 6803-6808.
Jiang, H., Patwardhan, R., & Shah, S. L. (2009). Root cause diagnosis of plant-wide oscillations using the concept of adjacency matrix.
J. Process Control, 19(8), 1347-1354.
Jiang, H., Choudhury, M. S., & Shah, S. L. (2007). Detection and diagnosis of plant-wide oscillations from industrial data using the
spectral envelope method. J. Process Control, 17(2), 143-155.
Jolliffe, I. (2002). Principal component analysis. John Wiley & Sons, Ltd.
Kaiser, A., & Schreiber, T. (2002). Information transfer in continuous processes. Physica D: Nonlinear Phenomena, 166(1), 43-62.
Kondaveeti, S.R., Izadi, I., Shah, S.L., Black, T., & Chen, T. (2012). Graphical tools for routine assessment of industrial alarm systems.
Computers and Chemical Engineering, 46, 39-47.
Kondaveeti, S. R., Izadi, I., Shah, S. L., Shook, D. S., Kadali, R., & Chen, T. (2013). Quantification of alarm chatter based on run
length distributions. Chemical Engineering Research and Design, 91, 2550-2558.
Lai, S., & Chen, T. (2015). A method for pattern mining in multiple alarm flood sequences. Chemical Engineering Research and
Design, in press.
McDougall, A.J., D.S. Stoffer & D.E. Tyler (1997). Optimal transformations and the spectral envelop for real-valued time series.
Journal of Statistical Planning and Inference 57, 195-214.
Miao, T., & Seborg, D. E. (1999). Automatic detection of excessively oscillatory feedback control loops. In Proc. of IEEE international
Conference on Control Applications, 359-364.
Müller, M. (2007). Dynamic time warping, in Information Retrieval for Music and Motion. New York: Springer-Verlag, ch. 4, 69–82.
Naghoosi, E., Izadi, I., & Chen, T. (2011). Estimation of alarm chattering. Journal of Process Control, 21, 1243-1249.
Nair, G. N. (2013). A nonstochastic information theory for communication and state estimation. IEEE Trans. Automatic Control, 58(6),
1497-1510.
Nimmo, I. (2005). Rescue your plant from alarm overload. Chemical Processing, 28–33.
Noda, M., Higuchi, F., Takai, T., & Nishitani, H. (2011). Event correlation analysis for alarm system rationalization. Asia-Pacific
Journal of Chemical Engineering, 6(3), 497-502.
Qin, S. J. (1998). Control performance monitoring - a review and assessment. Computer and Chemical Engineering, 23, 173-186.
Qin, S. J. (2014). Process data analytics in the era of big data. AIChE Journal, 60(9), 3092-3100.
Schleburg, M., Christiansen, L., Thornhill, N. F., & Fay, A. (2013). A combined analysis of plant connectivity and alarm logs to reduce
the number of alerts in an automation system. J. Process Control, 23(6), 839-851.
Schreiber, T. (2000). Measuring information transfer. Physical Review Letters, 85(2), 461.
Smith, T. F., & Waterman, M. S. (1981). Identification of common molecular subsequences. Journal of Molecular Biology, 147, 195–
197.
Stoffer, David S. (1999). Detecting common signals in multiple time series using the spectral envelope. Journal of the American
Statistical Association, 94, 1341-1356.
Stoffer, David S., David E. Tyler & Andrew J. McDougall (1993). Spectral analysis for categorical time series: Scaling and spectral
envelope. Biometrika, 80, 611-622.
Stoffer, David S., Tyler, David E. & Wendt, David A. (2000). The spectral envelope and its applications. Statistical Science, 15, 224-
253.
Tangirala, A. K., Kanodia, J., & Shah, S. L. (2007). Non-negative matrix factorization for detection and diagnosis of plantwide
oscillations. Industrial & engineering chemistry research, 46(3), 801-817.
Tangirala, A. K., Shah, S. L., & Thornhill, N. F. (2005). PSCMAP: A new tool for plant-wide oscillation detection. Journal of Process
Control, 15(8), 931-941.
Thambirajah, J., Benabbas, L., Bauer, M., & Thornhill, N. F. (2009). Cause-and-effect analysis in chemical processes utilizing XML,
plant connectivity and quantitative process history. Computers & Chemical Engineering, 33(2), 503-512.
Thornhill, N. F., Huang, B., & Zhang, H. (2003a). Detection of multiple oscillations in control loops. Journal of Process Control,
13(1), 91-100.
Thornhill, N. F., Cox, J. W., & Paulonis, M. A. (2003b). Diagnosis of plant-wide oscillation through data-driven analysis and process
understanding. Control Engineering Practice, 11(12), 1481-1490.
Thornhill, N. F., & Hägglund, T. (1997). Detection and diagnosis of oscillation in control loops. Control Engineering Practice, 5(10),
1343-1354.
Thornhill, N.F., Shah, S.L., Huang, B., & Vishnubhotla, A. (2002). Spectral principal component analysis of dynamic process data.
Control Engineering Practice, 10, 833–846.
Timms, C. (2009). Hazards equal trips or alarms or both. Process Safety and Environmental Protection, 87, 3–13.
Wang, J., & Chen, T. (2013). An online method for detection and reduction of chattering alarms due to oscillation. Computers and
Chemical Engineering, 54, 140-150.
Wang, J., & Chen, T. (2014). An online method to remove chattering and repeating alarms based on alarm durations and intervals.
Computers and Chemical Engineering, 67, 43-52.
Wang, J., Yang, F., Chen, T., & Shah, S.L. (2016). An Overview of Industrial Alarm Systems: Main Causes for Alarm Overloading,
Research Status, and Open Problems. IEEE Transactions on Automation Science and Engineering, 13(2), 1045-1061.
Welch, L. (1974). Lower bounds on the maximum cross correlation of signals (Corresp.). IEEE Transactions on Information theory,
20(3), 397-399.
Xu, J., Wang, J., Izadi, I., & Chen, T. (2012). Performance assessment and design for univariate alarm systems based on FAR, MAR,
and AAD. IEEE Transactions on Automation Science and Engineering, 9, 296-307.
Yang, F., Duan, P., Shah, S. L., & Chen, T. (2014). Capturing Connectivity and Causality in Complex Industrial Processes. Springer
Science & Business Media.
Yang, F., Xiao, D., & Shah, S. L. (2010).Qualitative fault detection and hazard analysis based on signed directed graphs for large-scale
complex systems. Fault Detection. INTECH Open Access Publisher.
Yang, F., Shah, S.L., Xiao, D., & Chen, T. (2012). Improved correlation analysis and visualization of industrial alarm data. ISA
Transactions, 51, 499-506.
Yang, Z., Wang, J., & Chen, T. (2013). Detection of correlated alarms based on similarity coefficients of binary data. IEEE
Transactions on Automation Science and Engineering, 10, 1014-1025.
Fig. 1. Ingredients of a smart process data analytics platform to enable seamless fusion of information from
process and alarm data combined with process connectivity information.

Fig. 2. Functional modules for a process data analytics toolbox.

Fig. 3. An example of A&E log.

 Fig. 4. An example of process data.

High Density Alarm Plot

Tag102.CFN 75
Tag192.FAILED
Tag101.CFN
Tag108.LOW
60
Tag119.IOF
Tag48.LOW
Tag197.HIGH
Tag96.IOF
45
Tag50.LOW
Tag106.IOF
Tag103.CFN
Tag120.IOF
31
Tag57.IOF
Tag162.CFN
Tag64.COMM
Tag60.IOF
16
Tag53.LOW
Tag46.LOW
Tag98.IOF
Tag198.IOF
1
15-Aug 8PM 17-Aug 8PM 19-Aug 9PM 21-Aug 10PM 23-Aug 11PM 25-Aug 11PM
Time (Bins, each bin = 10 mins)

Fig. 5. High density alarm plot.

 Alarm Similarity Color Map
1
Tag103.CFN
Tag57.IOF
Tag96.IOF
Tag106.IOF
0.8
Tag46.LOW
Tag198.IOF
Tag162.CFN
Tag120.IOF
0.6
Tag50.LOW
Tag53.LOW
Tag48.LOW
Tag98.IOF
0.4
Tag119.IOF
Tag64.COMM
Tag60.IOF
Tag108.LOW
0.2
Tag197.HIGH
Tag101.CFN
Tag102.CFN
Tag192.FAILED
0

Fig. 6. Alarm similarity color map.

Chattering Index using Inverse Weighting

Tag102.CFN
Tag192.FAILED
Tag101.CFN
Tag108.LOW
Tag119.IOF
Tag48.LOW
Tag197.HIGH
Tag96.IOF
Tag50.LOW
Tag106.IOF
Tag103.CFN
Tag120.IOF
Tag57.IOF
Tag162.CFN
Tag64.COMM
Tag60.IOF
Tag53.LOW
Tag46.LOW
Tag98.IOF
Tag198.IOF

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Chattering index

Fig. 7. Chattering indices of top bad actors.

Tag102.CFN
2000 1
Alarm count reduction by off delay timer
RTN to ALM run Length distribution

X: 10
Y: 0.82

1000 0.5

0 0
0 10 20 30 40 50 60
Run length (s)

Fig. 8. Design of off-delay timer based on the run length distribution.

 80

70

60
alarms/10min

50

40

30

20

10

0
15-Aug 8 PM 18-Aug 8 AM 20-Aug 9 PM 23-Aug 10 AM 25-Aug 11 PM

Fig. 9. Alarm burst plot for raw alarm data.

45

40

35

30
alarms/10min

25

20

15

10

0
15-Aug 8 PM 18-Aug 8 AM 20-Aug 9 PM 23-Aug 10 AM 25-Aug 11 PM

Fig. 10. Alarm burst plot for alarm data with chattering alarms reduced.

Cluster using Agglomerative Hierarchical Cluster Tree

1
20
38
19
10 0.9
9
36
28
27 0.8
11
17
16
8 0.7
18
29
7
25 0.6
24
23
35
30 0.5
31
22
15 0.4
13
14
12
37 0.3
33
34
32
26 0.2
21
1
3
2 0.1
5
6
4
0
4 6 5 2 3 1 2126323433371214131522313035232425 7 2918 8 161711272836 9 10193820

Fig. 11. Similarity color map of clustered alarm floods.

Fig. 12. Sequence alignment between two similar alarm flood sequences.
 Alarm
Tag1

Normal
Alarm
Tag2

Normal
Alarm
Tag3

Normal
Alarm
Tag4

Normal
Alarm
Tag5

Normal
Alarm
Tag6

Normal
0.5 1 1.5 2 2.5 3
Time(s) 5
x 10

Fig. 13. Alarm signals over 3.5 days.

Fig. 14. Trends of transfer entropies versus time lags.

Fig. 15. Causal map of information flow paths.

Fig. 16. Causal map of direct information flow paths.

Fig. 17. Normal (blue) and abnormal (red) parts of a process signal.

Fig. 18. Original setting with HAL=1.

Fig. 19. Design scenario with an optimized alarm limit.

Fig. 20. Design scenario with an off-delay timer.

Fig. 21. Design scenario with a moving average filter.

Fig. 22. ROC curves.

Fig. 23. P&ID of the distillation plant of Eastman Chemical Company.

 High Density Time Trend High Density Time Trend
FC5.OP FC8.SP

LC3.OP FC5.SP

PC2.OP FC6.SP

FC6.OP FC1.SP

TC1.OP FC3.SP

FC4.OP LI1.PV

FC1.OP FI3.PV

LC1.OP PI2.PV

FC3.OP TI7.PV

PC1.OP TI8.PV

FC7.PV FI4.PV

TC2.PV Tags TI3.PV

Tags

FC8.PV FI5.PV

LC2.PV TI6.PV

FC5.PV TI4.PV

LC3.PV TI5.PV

PC2.PV PI1.PV

FC6.PV FI1.PV

TC1.PV TI2.PV

FC4.PV TI1.PV

FC1.PV FC7.OP

LC1.PV TC2.OP

FC3.PV FC8.OP

PC1.PV LC2.OP
0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000
Samples Samples

Fig. 24. High density time trends.

High Density Power Spectra High Density Power Spectra

FC5.OP FC8.SP

LC3.OP FC5.SP

PC2.OP FC6.SP

FC6.OP FC1.SP

TC1.OP FC3.SP

FC4.OP LI1.PV

FC1.OP FI3.PV

LC1.OP PI2.PV

FC3.OP TI7.PV

PC1.OP TI8.PV

FC7.PV FI4.PV

TC2.PV TI3.PV
Tags

Tags

FC8.PV FI5.PV

LC2.PV TI6.PV

FC5.PV TI4.PV

LC3.PV TI5.PV

PC2.PV PI1.PV

FC6.PV FI1.PV

TC1.PV TI2.PV

FC4.PV TI1.PV

FC1.PV FC7.OP

LC1.PV TC2.OP

FC3.PV FC8.OP

PC1.PV LC2.OP
-3 -2 -1 -3 -2 -1
10 10 10 10 10 10
Frequency (Cycles/sample) Frequency (Cycles/sample)

Fig. 25. High density power spectra.

 Power Spectral Correlation Map
1
'TI6.PV'
'PC2.PV'
'LC3.PV'
'FC6.SP'
'LC3.OP' 0.8
'FC6.PV'
'PC1.PV'
'FI1.PV'
'FC4.PV'
'TI1.PV' 0.6
'FC4.OP'
'TI3.PV'
'FC6.OP'
'LI1.PV'
'TI2.PV' 0.4
'PI1.PV'
'FC3.OP'
'FC3.SP'
'PC1.OP'
'FC3.PV' 0.2
'FC7.OP'
'FC7.PV'
'FC1.OP'
'PC2.OP'
'PI2.PV' 0
'FC5.OP'
'TI8.PV'
'FI4.PV'
'TC2.PV' -0.2
'FC8.OP'
'FC8.SP'
'TC2.OP'
'FC8.PV'
'FI3.PV' -0.4
'LC2.OP'
'LC2.PV'
'TI5.PV'
'TI4.PV'
'TC1.PV' -0.6
'FC1.SP'
'LC1.OP'
'FC1.PV'
'LC1.PV'
'TI7.PV' -0.8
'FI5.PV'
'FC5.PV'
'FC5.SP'
'TC1.OP'
-1

Fig. 26. Power spectral correlation map.

Spectral Envelope
16000

14000

12000
Spectral Envelope, ()

X: 0.003175
Y: 9706
10000

8000

6000

4000

2000

-3 -2 -1
10 10 10
Frequency, 

Fig. 27. Spectral envelope.

 0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
'LC2.OP'

0
500
1000
1500
2000
2500
3000
'PC 1.
P 'PC 1.
'FC 3. V' P

6
7
8
9
10
11
12
13
'FC 3. V'
'LC1.PV'
P 'LC1.PV'
'FC 1. V' P

41.5
'FC 1. V'
'FC 4.PV' 'FC 4.PV'
'TC 1.PV' 'TC 1.PV'
'FC 6. PV' P
'FC 6. V'
'PC 2.PV'
P 'PC 2.PV'
P
'LC3. V' 'LC3. V'
'FC 5. PV' 'FC 5. PV'
'LC2.PV' P

42
P 'LC2. V'
'FC 8. V' 'FC 8. PV'
'TC 2.PV' 'TC 2.PV'
'FC 7. PV' P
'PC 1. V'P 'FC 7. V'
O 'PC 1. PV'
'FC 3. P'
O 'FC 3.O P'
O
'LC1. P' 'LC1. P'
'FC 1. OP' O
O 'FC 1. P'
'FC 4. P' 'FC 4.OP'
O

42.5
'TC 1. P' O
'TC 1. P'
'FC 6. OP' 'FC 6. OP'
'PC 2.OP' 'PC 2.OP'
O O P'
'LC3. P' 'LC3.
'FC 5. OP' 'FC 5. OP'
O O
'LC2. P' 'LC2. P'

43
'FC 8. OP' 'FC 8. OP'

Tag Index
Tag Index
O O
'TC 2. P' 'TC 2. P'

'LC2.PV'
'FC 7. OP' 'FC 7. OP'
O O
'TI1.PP' 'TI1.PP'
'TI2.PV' 'TI2.PV'
'FI1.P V' 'FI1.P V'
Chi-Square Test Statistics

'PI1.PV' 'PI1.PV'

Oscillation Contribution Index

'TI5.PV' 'TI5.PV'
'TI4.PV'

43.5
'TI4.PV'
'TI6.PV' 'TI6.PV'
'FI5.P V' 'FI5.P V'
'TI3.PV' 'TI3.PV'
'FI4.P V' 'FI4.P V'
'TI8.PV' 'TI8.PV'

Fig. 28. Chi-squared test statistics.

'TI7.PV' 'TI7.PV'

44
'PI2.P V' 'PI2.P V'
'FI3.PV' 'FI3.PV'
'LI1 .PV' 'LI1 .PV'
'FC 3. V' 'FC 3. V'
'FC 1.SP'

Fig. 29. Oscillation contribution indices.

'FC 1.SP'
S S
'FC 6. P' 'FC 6. P'
'FC 5.SP' 'FC 5.SP'
'FC 8.SP' 'FC 8.SP'
SP' SP'

44.5

Fig. 30. Scatter plot between “LC2.PV” and “LC2.OP”.

Fig. 31. Process flow diagram of the MCC plant.

Fig. 32. Adjacency matrix.

Fig. 33. Reachability matrix.

Fig. 34. Schematic of the FGD process.

Fig. 35. Time trends of the process variables.

Fig. 36. Information flow paths based on NTEs.

Fig. 37. Exclusion of the indirect causalities based on the calculation of NDTEs.

Fig. 38. Direct information flow paths based on NDTEs in Table 11.

Fig. 39. Information flow paths in the FGD process.

 Table 1. Tasks in “Data Loading”.

Function Task
Load Alarm Data Load alarm data from an Excel file with structured data format.
Load Process Data Load process data from an Excel file with structured data format.
Load alarm and/or process data from a MATLAB data file with a reorganized data
Load Preprocessed Data
format.
Merge Data Associate the tags of process variables with their corresponding alarm tags.
Export the alarm data and/or process data as a MATLAB data file with a reorganized
Export
data format.
Clear Clear alarm and/or process data.

Table 2. Functions in “Alarm Data Analysis”.

Function Task
HDAP Visualize alarm data over a selected time period using a high density color map.
RL&DTA Design on/off delay timers based on run length distributions.
CI Detect chattering alarms and calculate chattering indices.
ABP Calculate and visualize the peak alarm rate.
OAA Discover alarms caused by process oscillations.
ASCM Detect correlated alarms and visualize their similarity indices using a color map.
Analyze alarm floods, including identification, comparison, and clustering of alarm
AFA
floods.
Analyze the operator acknowledgement, including the acknowledgement rate and
OPAA
response time.
CIA Detect causal relations between alarm variables.
MDAA Detect mode-dependent nuisance alarms.

Table 3. Report Functions in “Alarm Data Analysis”.

Report Function Task

Generate Excel reports of basic statistical information of alarms by top bad
Bad Actor List
actors, alarm identifiers, alarm priorities, and locations.
Generate Excel reports of alarm reduction by simulating the implementation of
Performance Calculator
delay-timers and alarm suppression techniques.
Generate Excel reports of changes of alarm count for each alarm variable during
Bad Actor Comparison
two different time periods.
 Table 4. Functions in “Alarm Configuration Analysis”.

Function Task
Moving Average
Moving Variance
Moving Norm Reduce noises, remove bad data, or modify statistical
Filter
Rank Order distributions of process signals.
Low Pass
EWMA
Off-Delay Timer Reduce chattering or fleeting alarms by delaying alarm
Delay Timer
On-Delay Timer raising or clearing instants.
Reduce false or missed alarms by applying different
Deadband
thresholds for alarm raising and clearing.
Alarm Limit Optimization Optimize high or low alarm limit automatically.

Table 5. Preprocessing Functions for Process Signals.

Function Task
Detrend Data Make the mean of the process signals be zero.
Data Smoothing Remove the noise from process signals using a band filter.
Data Normalization Normalize process signals.
Remove Outliers Detect and remove outliers.

Table 6. Functions to Analyze and Visualize Process Signals.

Type Function
Time trend
Frequency spectrum
Univariate
Power spectrum density
Spectrogram
High density plot
Parallel coordinate
Multivariate Scatter matrix
Spectral envelope
PCA and Spectral PCA
 Table 7. Functions in “Connectivity and Causality Analysis”.
Function Task
Detect correlations between process signals and visualize the
Correlation Coefficient
results using a correlation color map.
Detect the correlations between power spectrums of process
Spectral Correlation
signals and visualize the results using a correlation color map.
Detect causal relations between process signals and visualize the
Granger Causality results using a signed directed graph. This method is fast, but
only effective for linear processes.
Detect causal relations between process signals and visualize the
results using a signed directed graph. This method is effective for
Transfer Entropy
either linear or nonlinear processes, but quite computational
burdensome.

Table 8. Recommendations for setting of delay timers.

Length of delay- Percentage of alarms
Alarm Tag Type of delay-timer No. of alarms reduced
timer(sec) reduced
Tag102.CFN Off-delay 10 3981 82%
Tag101.CFN Off-delay 10 856 79%
Tag108.LOW Off-delay 10 408 77%
Tag96.IOF Off-delay 8 156 91%
Tag103.CFN Off-delay 8 141 96%
Tag98.IOF Off-delay 10 65 73%
Tag198.IOF Off-delay 2 76 87%
 Table 9. Maximum NTEs and significance thresholds.

Tag1 Tag2 Tag3 Tag4 Tag5 Tag6

0.258 0.258 0.258 0.129 0.012
Tag1
(0.016) (0.015) (0.016) (0.015) (0.022)
0.0003 1 1 0.434 0.008
Tag2
(0.013) (0.001) (0.001) (0.014) (0.019)
0.001 0.0001 1 0.434 0.008
Tag3
(0.013) (0.014) (0.001) (0.014) (0.018)
0.0011 0.0001 0.0001 0.434 0.008
Tag4
(0.013) (0.014) (0.014) (0.014) (0.020)
0.0004 0.0004 0.0005 0.0005 0.009
Tag5
(0.012) (0.013) (0.013) (0.013) (0.017)
0.007 0.006 0.006 0.006 0.005
Tag6
(0.009) (0.009) (0.009) (0.009) (0.010)

Table 10. NTEs between process variables.

y1 y2 y3 y4 y5
y1 0.001 0.089 0.177 0.014
y2 0.131 0.117 0.154 0.010
y3 0.078 0.005 0.008 0.105
y4 0.128 0.005 0.095 0.019
y5 0.016 0.001 0.130 0.012

Table 11. NDTE between each pair of process variables with causal relations.

Step Relation Intermediate variables NDTE

1 y1 → y3 y2, y4 0.024
2 y3 → y1 y2, y4 0.023
3 y2 → y1 y4 0.425
4 y2→ y4 y1 0.025
5 y2 → y3 y1, y4 0.021