Professional Documents
Culture Documents
Computers and Chemical Engineering: Ankur Kumar, Apratim Bhattacharya, Jesus Flores-Cerrillo
Computers and Chemical Engineering: Ankur Kumar, Apratim Bhattacharya, Jesus Flores-Cerrillo
Computers and Chemical Engineering: Ankur Kumar, Apratim Bhattacharya, Jesus Flores-Cerrillo
a r t i c l e i n f o a b s t r a c t
Article history: Reformer boxes are complex, integrated, and high-temperature units, subject to various failures during
Received 9 September 2019 continuous operations for extended time periods. Challenges in the development of high-fidelity first
Revised 30 December 2019
principle models, despite easy availability of process measurements motivated the development of data-
Accepted 23 January 2020
driven, automated fault detection (FD) systems. Paucity of plant-wide implementation of FD technologies
Available online 28 January 2020
in the chemical industry, accentuates the absence of relevant practical guidelines and best practices. In
Keywords: this paper, a trivially replicable FD system has been developed for large-scale industrial reformer boxes
Steam-methane reformers of hydrogen manufacturing units. Actual process data from plant historian has been used for training and
Process monitoring validation of a novel model, developed using a combination of partial least squares regression and prin-
Abnormality detection cipal components analysis. Abnormalities based on several important measurements around the reformer
Data-driven modeling were identified. Explicit algorithmic details and insights obtained during development of the expert sys-
tem have been provided for ease of replication and adaptability.
© 2020 Elsevier Ltd. All rights reserved.
1. Introduction detection and diagnosis (FDD), has been an active field of re-
search over the past few decades (MacGregor and Kourti 1995;
Large scale industrial processes warrant increasing demands on Venkatasubramanian et al., 2003c; Venkatasubramanian et al.,
process safety, uniformity in production quality, and overall plant 2003a,b, Severson et al., 2016).
reliability. To achieve these objectives, real-time process monitor- FDD can be broadly classified into three categories: analyt-
ing is employed for early detection of abnormalities in processes ical (model-based), knowledge-based and data-driven methods
and avoidance of severe equipment damages. Early abnormality (Alzghoul et al., 2014). The analytical method relies on first-
detection significantly reduces maintenance and lost-production principles-based mathematical models of the process and incor-
costs (Dash and Venkatasubramanian, 20 0 0). For example, large poration of physical understanding of the system into the fault
scale syngas (gas mixture consisting primarily of hydrogen and car- detection process. While the analytical models are expected to
bon monoxide) production employs reformer furnaces with several provide superior accuracy, development of a high-fidelity math-
hundred catalyst-filled tubes (see the details in Section 2). Since ematical model of complex industrial processes can be diffi-
this process is highly integrated, a failure of a single tube can ini- cult. For example, modeling radiative heat transfer inside large-
tiate a cascade of failures of several other tubes inside the furnace scale furnace boxes in syngas plants can be quite involved
box. Since the catalyst tubes in a reformer box contribute to ap- (Kumar et al., 2015). The computationally-intensive calculations
proximately 15–20% of the total capital cost, the economic impli- make model-based real-time monitoring of complex industrial pro-
cations of early fault detection are significant. Industrial statistics cesses infeasible.
estimate the economic impact due to unplanned outages at petro- Knowledge-based methods are rule-based expert systems
chemical plants alone to be around $20 billion per year (Nimmo where the rules are derived from process engineers’ experience
1995). However, efficient process monitoring and control has re- and plant operators’ intuitive knowledge of the underlying pro-
mained a challenge due to the increasing complexity of process cess. For example, plant operators may notice that openings of
systems and practical limitations on continuous manual monitor- flow-control valves exhibit specific characteristic patterns before
ing. Hence, process monitoring, also commonly termed as fault failure; rules can be framed based on these patterns to de-
tect valve issues to prevent severe process disturbances. How-
∗ ever, it is difficult and time-consuming to create an exhaustive
Corresponding author.
E-mail addresses: ankur_kumar@praxair.com, Ankur.Kumar@linde.com (A. Ku-
collection of rules that covers a wide range of potential process
mar). faults. Additionally, for novel processes, the knowledge base can
https://doi.org/10.1016/j.compchemeng.2020.106756
0098-1354/© 2020 Elsevier Ltd. All rights reserved.
2 A. Kumar, A. Bhattacharya and J. Flores-Cerrillo / Computers and Chemical Engineering 136 (2020) 106756
Fig. 1. Representative flowsheet of a steam methane reforming-based Hydrogen plant. Sensor measurements from the shaded region of the plant were used in process FDD
analysis. Representative process-stream temperatures adopted from (Kumar et al., 2017b). (For interpretation of the references to color in the text, the reader is referred to
the web version of this article.)
be extremely sparse. A good review of knowledge-based meth- available to build accurate classification models. Alternatively, data
ods for fault analysis and diagnostics can be found in the work of from a wide range of normal plant operations can be used to build
(Venkatasubramanian et al., 2003a). statistical models to determine whether the process is operating
Data-driven approach for FDD utilizes historical process data. normally. These multivariate statistical process monitoring (MSPM)
The data contain process information and capture intrinsic process methods. (MacGregor and Kourti, 1995) have recently become
complexities. Hence it can be used for modeling, monitoring, and more popular due to the rapid development of process instrumen-
control (Kano and Nakagawa 2008); (MacGregor et al., 2005). Pro- tation and data acquisition technology, and wide utilization of dis-
cess data from faulty and normal plant operations can be used tributed control system (DCS) in modern industrial processes. Due
to develop classification models that classify process conditions to the ease of accessibility to process data, and the abundance of
into faulty and normal classes; (Yin et al., 2014b) however, large inexpensive data management systems, the volume of data gen-
amount of data from faulty plant operations are generally not erated from large-scale industrial processes have been on the rise
A. Kumar, A. Bhattacharya and J. Flores-Cerrillo / Computers and Chemical Engineering 136 (2020) 106756 3
Fig. 3. Historical data for key process variables. Note that the displayed values have been scaled by different scaling-factors for confidentiality purposes.
Fig. 4. Filtered historical data for the variables of interest for fault analysis. Note that the displayed values have been scaled by different scaling-factors for confidentiality
purposes.
4 A. Kumar, A. Bhattacharya and J. Flores-Cerrillo / Computers and Chemical Engineering 136 (2020) 106756
Fig. 5. Extracted steady-state and historical data for key process variables. Note that the displayed values have been scaled by different scaling-factors for confidentiality
purposes.
(Alzghoul et al., 2012). Principal component analysis (PCA) and par- on real large-scale industrial systems and strategies to overcome
tial least squares (PLS) are among the most popular MSPM tech- the practical challenges (Kano and Nakagawa, 2008) (related to
niques and rely on the projection of the high-dimensional pro- model adaptation, model maintainability, data pre-processing, etc.)
cess data onto a lower dimensional space through latent vari- faced during industrial applications are some of the reasons for
ables for extraction of key process information (AlGhazzawi and low industrial adoption of these techniques. Additionally, current
Lennox 20 08; Qin 20 03; Flores-Cerrillo and MacGregor 2004). Sev- research hints towards context specificity, suggesting that model-
eral other data-based methods, such as independent component ers need to create appropriate combination of tools designed for
analysis (ICA), artificial neural networks (ANNs), support vector each application. Albeit, in industry, it is common to find a conser-
machines (SVM), kernel PCA/PLS, recursive PCA/PLS, gaussian mix- vative process monitoring approach where static upper and lower
ture models (GMMs), etc., have been explored to deal with issues alarm thresholds are used for a few key process variables, however,
such as non-gaussianity, non-linearity, non-stationarity, and multi- as shown later in the text, this conservative approach leads to rel-
ple operating modes (Lee et al., 2004; Cho et al., 2005; Qin 1998). atively delayed detection of faults. Relatively faster detection is de-
Excellent reviews on the state-of-the-art data-driven FDD meth- pendent on plant operators serendipitously identifying these faults
ods for industrial processes can be found in literature (Qin, 2012; from process graphs on HMI (human machine interface) screens in
Ge, et al., 2013; Yin et al., 2014a; Ge, 2017; Reis and Gins 2017). the control room. Of course, this is not an ideal solution to the
Empirical/data-driven models are especially useful in fault detection problem.
industrial settings where low development time and payback In this paper, results from development and application of
period, speed of implementation, and robustness to practical an expert process monitoring system, based on data-driven
issues such as missing data are important. As discussed above, a FDD methodologies, for monitoring reformer boxes in hydro-
plethora of data-driven techniques is available for the develop- gen plants are presented. Reformer box is a physically large
ment of a process monitoring tool. However, each method has (16 m × 16 m × 12 m) scale unit operating at ~ 1800◦ F to
its advantages and shortcomings; a method that works well for convert natural gas (methane) into syngas (Kumar et al., 2015).
one system might not exhibit satisfactory performance for another Several methods available in literature are compared on their fault
(Dash and Venkatasubramanian, 20 0 0; Ge et al., 2013 Ng and detection capabilities. Methods for both steady-state and dynamic
Srinivasan, 2010; Perk et al., 2010). have proposed combining process are applied and detailed step-by-step procedures have
multiple FDD methods in a multi-agent system for process been provided. Further, the developed expert system is successfully
monitoring; these multi-agent systems, however, do not make tested on two separate hydrogen plants to demonstrate the sys-
selection of specific FDD techniques any easier. Most of the FDD tem’s replicability. The paper is organized as follows. Section 2 pro-
techniques in literature are benchmarked against its performance vides a brief overview of the hydrogen production process. Re-
on the Tennessee Eastman (TE) process (Downs and Vogel 1993; former monitoring workflow is then described followed by a brief
Howell et al., (1997). However, as noted by Chiang et al. (2017), discussion on historical process data and data pre-processing.
this problem is somewhat antiquated. Sections 5 and 6 provide the details of algorithm and subsequent
Very few studies are available on application of FDD methods results from the application of steady-state and dynamic monitor-
on real large-scale industrial systems. Paucity of publicly available ing algorithms respectively.
literature on demonstrated success of data-driven FDD techniques
A. Kumar, A. Bhattacharya and J. Flores-Cerrillo / Computers and Chemical Engineering 136 (2020) 106756 5
CO + H2 O CO2 + H2 (2)
Reactions (1) and (2) are the steam methane reforming and
water gas shift reactions respectively. Steam reforming is a highly
endothermic reaction, favored by high temperature and low pres-
sure. Water gas shift reaction is mildly exothermic and favored
by low temperature. Both reactions are equilibrium limited with
overall conversion depending on the reformer outlet temperature
and pressure. The overall reaction is highly endothermic. The re-
quired heat is supplied by the burners through the combustion of
tail/recycle gas from the PSA unit and supplemental natural gas
fuel. Depending on the design, there are different burner arrange-
ments for a furnace. In Fig. 1, a top-fired reformer design is shown,
in which the combustion products are introduced co-current with
the process gas stream at the top of the furnace box.
The hot reformed gas exits the reformer from the lower end
of the tubes at a controlled outlet temperature around 1500 ◦ F
(Kumar et al., 2015) and is cooled against boiling water in the
process gas boiler (PGB). The reformed gas enters the high tem-
perature shift-reactor, where carbon monoxide reacts with excess
steam present in the reformed gas to produce additional hydro-
gen via the water gas shift reaction (reaction (2)). Since the wa-
ter gas shift reaction is mildly exothermic, the HTS reactor effluent
passes through a series of heat exchangers for heat recovery. The
process gas is first cooled against the natural gas feedstock within
the Natural Gas Preheater. The process gas is then used to preheat
boiler feed water in the Boiler Feedwater (BFW) Heater, and to pre-
heat deaerator makeup water in the Demineralized Water Heater.
The Steam Separator, separates process condensate from the pro-
Fig. 6. Training PCA model on steady-state data: (a) Cumulative variance explained cess gas stream which is then sent to the Pressure Swing Adsor-
by principal components (b) SPE statistic and 99.5% control limit (c) T2 statistic and bers (PSA), where final product purification takes place. The pro-
99.5% control limit. duction of a high purity hydrogen product is accomplished by the
removal of impurities (unconverted CH4 and CO) from the process
2. Reformer box and hydrogen plant gas stream within the PSA unit.
In the convection section of the reformer, ambient air is drawn
In this section, we describe the reforming process in a large- from the atmosphere using a Forced Draft Fan (FD Fan), and heated
scale hydrogen plant typically producing more than a hundred mil- in the Combustion Air Preheater. The heated air and fuel mixture
lion square cubic-feet per day of high-purity hydrogen. The pur- are ignited in the burners located within the radiant section of the
pose of this section is to highlight the heat-intensive, integrated reformer. The resultant heat of combustion is partly transferred to
nature of the HyCO process, which justifies the utility of intelli- the reformer tubes. A portion of the remaining heat is recovered
gent data-driven FD technologies, for ease of replication. We begin from the flue gas stream within the tube banks of the Mixed Feed
by describing a representative flowsheet, followed by description Preheater, the Steam Superheater, the Flue Gas Steam Generator,
of a repertoire of faults that warrant replicable, process-dependent and the Combustion Air Preheater. The Induced Draft Fan (ID-Fan),
alarm systems, as opposed to simplistic, rule-based alarms. assists the flow of the flue gas from the convective section to the
Fig. 1 shows a representative flowsheet of a steam methane Flue Gas Stack, prior to discharge into the atmosphere.
reforming (SMR) based Hydrogen plant. Steam methane reform- The reforming process is highly integrated and fault-prone.
ing processes convert natural gas feedstock and steam into hydro- Large-scale hydrogen plants typically operate continuously for pro-
gen. The reformer or the furnace box, has two major sections - longed periods, usually dictated by the service contracts which re-
the radiant section (colored red in Fig. 1) and the convection sec- quire continued supply of hydrogen. The tubes experience a num-
tion (colored blue). The radiant section contains several burners ber of temperature and pressure cycles, the cumulative effect of
and catalyst-filled reformer tubes suspended vertically in a furnace which can be extremely damaging to the tubes. The failure of one
6 A. Kumar, A. Bhattacharya and J. Flores-Cerrillo / Computers and Chemical Engineering 136 (2020) 106756
Fig. 7. Testing PCA model. Note that the displayed process values have been scaled by different scaling-factors for confidentiality purposes. (a) Fault imposed only in Tfg,reformer
(b) Fault imposed only in Tfg, conv (c) Fault imposed only in Tshift,out .
tube usually has a domino effect on all the tubes in the reformer, In this work, a total of 36 variables/sensor-measurements are
resulting in lower hydrogen yield. In extreme cases, intense heat- used to build the monitoring expert system. Out of the 36 vari-
ing from abnormally high temperatures within the furnace can also ables, 27 are temperature or pressure measurements of several
cause failure of the refractory walls of the furnace. These failures high-temperature process streams around and within the reformer
must be detected as early as possible, such that cascading failures box. Some of these include, amongst others, temperatures of flue
can be avoided by timely planned shutdowns for repairs. Flue gas gas at several locations within the radiant section, temperatures
in top-fired reformers is collected by tunnels (Kumar et al., 2015). at several locations in the convection section, and plant produc-
Maldistribution of flue gas in the tunnels can arise due to improper tion rate. Note that monitoring the high-temperature regions of the
designs and mechanical collapse due to fatigue which can give rise plant was of prime importance and therefore temperatures of rel-
to concentration of heat within a small region of the furnace lead- atively colder process streams are not included. For confidentiality
ing to structural failures. Other common source of faults include reasons, the locations of all the output sensors cannot be disclosed.
leakage of process gas through leaks in pipes. Leakage of high- The rest of the variables include fuel flow, PSA-recycle flow, and
temperature process streams pose fire and plant operator safety combustion air flow into the burners, process feed and steam flow
hazards. Albeit, the conventional alarm systems in the plant can into the reformer tubes, heat content of natural gas, and tempera-
detect some of these inconsistencies, they are mostly reactive in tures of a few reformer input streams.
nature and do not attribute any deeper process knowledge.
A. Kumar, A. Bhattacharya and J. Flores-Cerrillo / Computers and Chemical Engineering 136 (2020) 106756 7
Fig. 10. Testing external-analysis model. Note that the displayed process values have been scaled by different scaling-factors for confidentiality purposes. (a) Fault imposed
only in Tfg,reformer (b) Fault imposed only in Tfg, conv (c) Fault imposed only in Tshift,out .
work process for replication of the tool to other sites within the that can change due to fluctuations in demand. In such cases, ei-
organization simple and convenient. ther the normal operating range has to be wide spanning multi-
ple regimes, or the operating regimes need to be reset every time
4.1. Variables of interest for fault analysis a change in operating conditions is encountered. In the following
sections FD methodologies are investigated that overcome the lim-
In this work, 3 of the output variables - temperature of flue gas itations of this simple approach for reformer monitoring.
from reformer (Tfg,reformer ), temperature of flue gas near the end of
convection section (Tfg, conv ), and shift-reactor outlet temperature 5. Steady-state monitoring algorithm
(Tshift,out ) - are chosen for fault analysis. The choice is arbitrary and
any other monitored variable could also have been chosen to illus- Steady-state monitoring algorithms train fault-models using
trate the technology discussed in this article. To mimic the fault steady-state training data to identify normal operating states. Dur-
scenarios generally observed during plant operations, a simulated ing model-training, control limits on one or more fault detection
fault manifested by increase in temperature values have been in- indices are determined. During testing, faulty/abnormal process
troduced in these variables. Fig. 4 shows the historical data for data that deviates from the normal process behavior and conse-
these variables; all the variables display considerable variations in quently, violate the control limits are flagged as faults.
operating values. A common approach, typically followed in pro-
cess industry for abnormality detection in these variables, is to put 5.1. Steady-state extraction
upper and lower alarm limits spanning the observed range of val-
ues. This approach, however, is inadequate since several potential For monitoring based on steady-state data, steady-state periods
faults where the faulty variable remains within the normal oper- need to be extracted out of the historical time-series data which
ating range can remain undetected. This is primarily due to the is then used as the training data-set. The approach by Simon and
multivariate nature of the process. Moreover, the attribution of the Litt (2011) for identification of steady-state operating points has
normal operating range is based upon specific operating conditions been adopted in this work. An overview of the methodology is
A. Kumar, A. Bhattacharya and J. Flores-Cerrillo / Computers and Chemical Engineering 136 (2020) 106756 9
Fig. 11. Normal dynamic changes in key process variables. Note that the displayed values have been scaled by different scaling-factors for confidentiality purposes.
Fig. 12. Normal dynamic changes in Tfg,reformer , Tfg, conv , and Tshift,out . Note that the displayed values have been scaled by different scaling-factors for confidentiality purposes.
Fig. 14. Augmenting a matrix with lagged measurements (yij refers to the value of jth variable from the ith measurement sample). Note that if any value in a row in the
augmented input matrix or output matrix is invalid (due to bad measurements or sensor issues), then the entire row from both the augmented input matrix and output
matrix is removed.
outlined in Algorithm 1. A 5h sample (filtered) data-set is used to subspace. It is a metric that measures the variance within the
calculate nominal standard deviations of a key reformer variables, model.
such as fuel flow to burners, reformed-gas temperature, etc. Sam-
ples belonging to steady-state operation periods are identified by 5.2.1. Training model & application
comparing the respective standard deviations calculated in a slid- Fig. 6a shows the cumulative variance in training data captured
ing test-window of size 5h ( ~ 300 samples), with nominal values. by the principal components (PCs). 12 PCs capturing 99% variance
Fig. 5 shows the extracted steady-state operation periods ob- were retained. Fig. 6b and c show the resulting Q and T2 statistics
tained from application of Algorithm 1; about 0.7 million sam- for the training data; the control limits of 2.8 and 84 are deter-
ples are retained. It can be observed that operation periods mined such that the number of samples outside the control limits
with significant increase or decrease in process values have been is only 0.5% of the entire training samples (Yamamoto et al., 2004;
excluded. Kano et al., 2004). For gauging the fault-detection capability of the
developed model, the 5-hour sample steady state operation data
that was used in Fig. 5 is considered and 5% faults are imposed
5.2. PCA-based monitoring in Tfg,reformer , Tfg, conv , and Tshift,out , separately in 3 different simu-
lations, as shown in Fig. 7. The fault in each of the 3 variables is
PCA models are primarily used to extract correlations among introduced at the start of the 5th h and is linearly increased till the
process variables from data. Abnormal/faulty process data do not end of the hour. It can be seen that while the fault in Tshift,out leads
follow normal process correlations and violate one or more fault to breach of SPE control limit, faults in Tfg,reformer and Tfg, conv go
detection indices. While the reader is referred to the work by undetected. This method has been discussed in this paper despite
Qin (2012) for the underlying mathematical details, step-by-step inadequate monitoring performance so that it serves as a reference
procedure for practical application of the method is provided in while assessing the performance of the methodologies discussed in
Algorithm 3 in Appendix section (A.1). The fault detection indices the following sections.
used in PCA are T2 and Q statistics. The Q statistic, defined as the
sum of squared residuals or the sum of prediction errors (SPE), 5.3. External analysis
measures abnormality in data that breaks the normal process cor-
relations. The T2 statistic measures the distance of the projec- Kano et al. (2004) proposed external analysis for multivariate
tion of the sample from the origin, in the principal component statistical process monitoring to explicitly take into account the
A. Kumar, A. Bhattacharya and J. Flores-Cerrillo / Computers and Chemical Engineering 136 (2020) 106756 11
1: Training model:
2: Generate output matrix (Y ∈ N×q ) and input matrix
(X ∈ N×p ) where N, q, p are numbers of measurement
samples, output variables, and input variables, respec-
tively
3: Calculate mean (y ∈ 1×q , x ∈ 1×p ) and standard de-
viation (σy ∈ 1×q , σx ∈ 1×p ) vectors: ith element of y
is mean of ith column of Y . Calculate similarly for other
vectors.
4: Normalize Y , X to Y ∈ N×q , X ∈ N×p : ith element of
each row of Y is subtracted by ith element of y and then
divided by ith element of σy . Calculate similarly for X Fig. 16. Training dynamic external-analysis model: (a) SPE statistic and 99.5% con-
5:
Perform PLS regression with X and Y . Choose number
trol limit (b) SPE statistic and 99.5% control limit.
Fig. 17. Testing dynamic external-analysis model. Note that the displayed process values have been scaled by different scaling-factors for confidentiality purposes. (a) Fault
imposed only in Tfg,reformer (b) Fault imposed only in Tfg, conv (c) Fault imposed only in Tshift,out (d) Fault imposed in Tfg,reformer and Tfg, conv .
A. Kumar, A. Bhattacharya and J. Flores-Cerrillo / Computers and Chemical Engineering 136 (2020) 106756 13
Fig. 18. Testing dynamic external-analysis model for data during normal process 6.1. Dynamic external analysis
changes.
Fig. 19. Contribution plots from testing dynamic external-analysis model. (a) Fault imposed only in Tfg,reformer (b) Fault imposed only in Tfg, conv (c) Fault imposed only in
Tshift,out (d) Fault imposed in Tfg,reformer and Tfg, conv .
Fig. 20. Monitoring workflow based on dynamic and steady-state external analysis.
that, unlike steady-state external analysis, dynamic external analy- from each output variable. Qin (2003) For example, step 13 in
sis results in Q (and T2 ) values that are well below the threshold Algorithm 4 can be decomposed as follows:
value. Note that this period of process data was removed from the
q
training data-set. SP E = y∗r 2 =
yr − SP Ei (3)
i=1
Fig. 21. Historical data for process variables form another the second hydrogen manufacturing site. Note that the displayed values have been scaled by different scaling-
factors for confidentiality purposes.
tributors to the fault. A known shortcoming of this approach is Post-deployment of the process monitoring tool, false positives
’smearing’ effect, where contribution from one variable is propa- (fault alarm without actual fault) and false negative (no fault alarm
gated to other variables during calculation of the PCA projections when there are actual faults) are both undesirable; while the for-
onto latent directions; this, correspondingly, can result in mis- mer leads to a loss of user’s confidence in the tool, the latter
diagnosis Qin (2003). An alternate approach, reconstruction-based leads to delayed fault detection by the plant operators. In pre-
contribution, has been proposed by Alcala and Qin (2009) where vious sections, it was observed that while steady-state external
16 A. Kumar, A. Bhattacharya and J. Flores-Cerrillo / Computers and Chemical Engineering 136 (2020) 106756
Fig. 23. Testing dynamic external-analysis model at the second hydrogen manufacturing site. Note that the displayed process values have been scaled by different scaling-
factors for confidentiality purposes. (a) Fault imposed only in Tfg,reformer (b) Fault imposed only in Tfg, conv (c) Fault imposed only in Tshift,out .
analysis can lead to false positives during normal process tran- 8. FD application at identical plant
sients, dynamic external analysis can lead to false negatives due
to lower fault sensitivity. To combine the benefits of both these An advantage offered by the data-based monitoring method-
approaches, the workflow depicted in Fig. 20 serves as the mon- ologies is that once a method has been found that gives satis-
itoring algorithm for the reformer unit. Depending upon whether factory performance for a particular system, the method can po-
the inputs are at steady state or not, fault-detection is performed tentially be applied without many modifications for monitoring
using steady-state or dynamic external analysis. For steady-state other similar systems. This keeps the investment of labor, time,
determination, a methodology similar to that in Algorithm 1 is and money for model development during tool replication low. For
utilized. Moreover, to avoid false positives from point fluctu- reformer monitoring, the external analysis-based method was re-
ations in sensor measurements, Q and T2 are required to be peated for monitoring another identical reformer-based hydrogen
above QCL and TCL2 , respectively, continuously for some period of manufacturing plant. Satisfactory fault-detection performance was
time. again obtained. Fig. 21 shows the filtered historical data (sampled
every minute over 3 years) from the other site for the key re-
A. Kumar, A. Bhattacharya and J. Flores-Cerrillo / Computers and Chemical Engineering 136 (2020) 106756 17
former variables along with the fault-relevant variables. Consider- ogy. A combination of steady-state external analysis and dynamic
able variations can be noticed in the values of all the process vari- external analysis is finally proposed to overcome the trade-off
ables. Fig. 22 shows the Q plot for the training data. Fault detection between false-positives (from steady-state external analysis) and
tests (Fig. 23) confirms good fault sensitivity where 1.5%, 3%, and false-negatives (from dynamic external analysis). Several aspects
4% faults in Tfg,reformer , Tfg, conv , and Tshift,out , respectively, cause QCL which affect the adoption and success of any analytic product
breach. within an organization have been highlighted: these aspects in-
clude ease of replication, time to replicate, frequency of false pos-
9. Discussion and recommendations itives and negatives, interpretability of results from the tool. The
detailed description of application of data-driven process moni-
In the previous sections, it was shown how process relation- toring methodologies for large-scale systems is expected to guide
ships can be extracted from process data to build process and other industrial practitioners and facilitate wider industrial adop-
fault-models for reformer boxes quickly. While this is very con- tion.
venient, model-developers (and the end-users) should be aware
of the limitations of data-driven approaches. If the historical pro- Declaration of Competing Interest
cess data, for some reason, do not accurately represent the be-
havior of the current system, frequency of false alarms will in- The authors declare that they have no known competing finan-
crease significantly. For example, in the hydrogen plant, cata- cial interests or personal relationships that could have appeared to
lyst activity decreases over time which affects the temperature influence the work reported in this paper.
of several process streams. Equipment, such as heat exchangers,
may be replaced causing step-changes in variable relationships Appendix A
around the equipment which impact the overall process correla-
tions. Methodologies should be in place to deal with such sce- A1. PCA-based monitoring
narios. Since the catalyst deactivation is a slow change, it can be
handled by re-training the model at regular intervals and provid-
ing higher weightage to latest data. Step changes in process cor- Algorithm 3 PCA model.
relations may necessitate taking the monitoring tool offline un-
til enough data have been gathered to re-develop the model. The 1: Training model:
monitoring system presented in this work identifies the need for 2: Generate data matrix (Z ∈ N×(q+ p) ) where N, q, p
model update by tracking the user feedback on false positives. If are numbers of measurement samples, output variables,
frequency of false positives is high, the model is re-trained us- and input variables, respectively
ing more recent data. However, if false positives are isolated oc- 3: Calculate mean (z ∈ 1×(q+ p) ) and standard deviation
currences, then a bank of false positive scenarios may be main-
(σz ∈ 1×(q+ p) ) vectors: ith element of z is mean of ith
tained which can be used to identify and avoid repeating false
column of Z. Calculate similarly for standard deviation.
alarm.
4: Normalize Z to Z ∈ N×(q+ p) : ith element of each row
Another important consideration during model development
of Z is subtracted by ith element of z and then divided
is choosing values for several model parameters. These param-
eters include number of lags used for data augmentation (l),
by ith element of σz .
number of latent components to retain in PCA and PLS, control 5: Perform PCA on Z . PCA generates:
limit for SPE (QCL ). Several studies have been published that pro- – Score matrix T ∈ N×(q+ p)
vide guidance on proper selection of values for these parameters. – Coefficient matrix C ∈ (q+ p)×(q+ p) : ith column of C
Kano et al. (2004) used process time-constant to determine the lag contains coefficients for the ith principal component
period (l) for a CSTR process. The number of components in PCA (PC)
and PLS can be determined using a scree plot, cross validation, or 6: Select number of PCs (ncomp ) such that 99% variance
the cumulative percentage of variance De Ketelaere et al. (2015), in
Z is captured.
Li et al. (20 0 0). An alternate approach is to collectively determine 7: Reconstruct: Z ∗ = Tncomp ∗ (Cncomp )transpose : Tncomp , Cncomp
these parameters via automated trial and error; a minimum fault- are the first ncomp columns of T and C, respectively.
sensitivity criteria can be set and the parameter space can be ex- 8: Calculate error matrix: E = Z −Y Z ∗
haustively searched for the values that meet the specified criteria.
9: Compute Q ∈ N×1 vector: ith element of Q is dot-
product of ith row of E matrix
10. Conclusion
10: Plot Q and find control limit QCL such that only 0.5%
of Qs are above QCL
In this work, development and application of process data-
based process monitoring has been reported for a large-scale 11: Testing model:
reformer-box unit of a hydrogen manufacturing plant. A complete
expert system workflow, from retrieving data from data historian 12: Let z ∈ 1×(q+ p) be the test measurement vector.
to displaying fault details on plant operator’s screen, has been 13: Compute z = (z − z )/σy : operations are performed
provided. External analysis was found to provide the best fault- element-wise.
detection performance. Similar fault-detection performance was 14: Calculate
z∗ = z ∗ Cncomp ∗ (Cncomp )transpose
obtained during direct application of the method at another hy- 15: Compute Q = z −z ∗ 2
drogen plant with little modifications in the overall methodol-
18 A. Kumar, A. Bhattacharya and J. Flores-Cerrillo / Computers and Chemical Engineering 136 (2020) 106756
Alcala, C.F., Qin, S.J., 2009. Reconstruction-based contribution for process monitor-
ing. Automatica 45 (7), 1593–1600.
Algorithm 4 Dynamic external analysis. Alcala, C.F., Qin, S.J., 2011. Analysis and generalization of fault diagnosis methods for
process monitoring. J. Process Control 21 (3), 322–330.
1: Training model: AlGhazzawi, A., Lennox, B., 2008. Monitoring a complex refining process using mul-
tivariate statistics. Control Eng. Pract. 16 (3), 294–307.
2: Generate output matrix (Y ∈ (N−l+1)×q ) and aug- Alzghoul, A., Backe, B., Löfstrand, M., Byström, A., Liljedahl, B., 2014. Comparing a
knowledge-based and a data-driven method in querying data streams for sys-
mented input matrix (X ∈ (N−l+1)×pl ) where N, q, p are tem fault detection: a hydraulic drive system application. Comput. Ind. 65 (8),
numbers of augmented measurement samples, output 1126–1135. doi:10.1016/j.compind.2014.06.003.
Alzghoul, A., Löfstrand, M., Backe, B., 2012. Data stream forecasting for system fault
variables, and input variables, respectively, and l is the
prediction. Comput. Ind. Eng. 62 (4), 972–978. doi:10.1016/j.cie.2011.12.023.
lag-period. Chen, G., McAvoy, T.J., 1998. Predictive on-line monitoring of continuous processes.
3: Calculate mean (y ∈ 1×ql , x ∈ 1×pl ) and standard de- J. Process Control 8 (5–6), 409–420.
viation (σy ∈ 1×ql , σx ∈ 1×pl ) vectors: ith element of y Chiang, L., Lu, B., Castillo, I., 2017. Big data analytics in chemical en-
gineering. Annu. Rev. Chem. Biomol. Eng. 8 (1), 63–85. doi:10.1146/
is mean of ith column of Y . Calculate similarly for other annurev- chembioeng- 060816- 101555.
vectors. Cho, J.-H., Lee, J.-M., Choi, S.W., Lee, D., Lee, I.-B., 2005. Fault identification for pro-
4: Normalize Y , X to Y ∈ N×ql , X ∈ N×pl : ith element of cess monitoring using kernel principal component analysis. Chem. Eng. Sci. 60
(1), 279–288.
each row of Y is subtracted by ith element of y and then Dash, S., Venkatasubramanian, V., 20 0 0. Challenges in the industrial applications of
divided by ith element of σy . Calculate similarly for X fault diagnostic systems. Comput. Chem. Eng. 24 (2–7), 785–791.
De Ketelaere, B., Hubert, M., Schmitt, E., 2015. Overview of PCA-based statistical pro-
5:
Perform PLS regression with X and Y . Choose number cess-monitoring methods for time-dependent, high-dimensional data. J. Qual.
of latent variables such that 99% variance in input vari- Technol. 47 (4), 318–335.
Dong, Y., Qin, S.J., 2018. Dynamic latent variable analytics for process operations and
ables is captured. PLS regression gives coefficient matrix control. Comput. Chem. Eng. 114, 69–80.
β ∈ pl×ql Dong, Y., Qin, S.J., 2018. A novel dynamic PCA algorithm for dynamic data modeling
6: Reconstruct: Y∗ = X × β . De-normalize Y ∗ to Y ∗ . ith and process monitoring. J. Process Control 67, 1–11.
Dong, Y., Qin, S.J., 2018. Regression on dynamic PLS structures for supervised learn-
∗
element of each row of Y is added by ith element of ing of dynamic data. J. Process Control 68, 64–72.
y and then multiplied by ith element of σy . Downs, J.J., Vogel, E., 1993. A plant-wide industrial process control problem. Com-
put. Chem. Eng. 17 (3), 245–255. doi:10.1016/0 098-1354(93)80 018-I.
7: Compute output residuals: Yr = Y − Y ∗ Flores-Cerrillo, J., MacGregor, J.F., 2004. Multivariate monitoring of batch processes
8: Normalize Yr to Y r . Let y , σy be the mean and stan- using batch-to-batch information. AlChE J. 50 (6), 1219–1228.
r r
dard deviation vectors. Ge, Z., 2017. Review on data-driven modeling and monitoring for plant-wide indus-
9: Perform PCA on Y r . PCA generates: trial processes. Chemometr. Intell. Lab. Syst. 171, 16–25.
Ge, Z., Song, Z., Gao, F., 2013. Review of recent research on data-based process mon-
– Score matrix T ∈ N×ql itoring. Ind. Eng. Chem. Res. 52 (10), 3543–3562. doi:10.1021/ie302069q.
Ge, Z., Yang, C., Song, Z., Wang, H., 2008. Robust online monitoring for multimode
– Coefficient matrix C ∈ ql×ql : ith column of C con- processes based on nonlinear external analysis. Ind. Eng. Chem. Res. 47 (14),
tains coefficients for the ith principal component (PC) 4775–4783.
10: Select number of PCs (ncomp ) such that 99% variance Howell, J., Chen, J., Zhang, J., 1997. The tennessee eastman problem as a pro-
r is captured.
in Y
cess monitoring benchmark. In: Proceedings of IFAC Symposium Safeprocess’97,
pp. 222–227. doi:10.1016/S1474- 6670(17)42405- 0.
∗ = Tn
comp ∗ (Cncomp )
transpose
11: Reconstruct: Y r : Tncomp , Cncomp Kano, M., Hasebe, S., Hashimoto, I., Ohno, H., 2004. Evolution of multivariate statis-
tical process control: application of independent component analysis and exter-
are the first ncomp columns of T and C, respectively. nal analysis. Comput. Chem. Eng. 28 (6–7), 1157–1166.
12: Calculate error matrix: E = Y r − Y
∗ Kano, M., Nakagawa, Y., 2008. Data-based process monitoring, process control, and
r
13: Compute Q ∈ N×1
vector: ith element of Q is dot- quality improvement: recent developments and applications in steel industry.
Comput. Chem. Eng. 32 (1–2), 12–24.
product of ith row of E matrix Ku, W., Storer, R.H., Georgakis, C., 1995. Disturbance detection and isolation by
14: Plot Q and find control limit QCL such that only 0.5% dynamic principal component analysis. Chemometr. Intell. Lab. Syst. 30 (1),
179–196.
of Qs are above QCL Kumar, A., Baldea, M., Edgar, T.F., 2017. A physics-based model for industrial
steam-methane reformer optimization with non-uniform temperature field.
15: Testing model: Comput. Chem. Eng. 105, 224–236.
Kumar, A., Baldea, M., Edgar, T.F., Ezekoye, O.A., 2015. Smart manufacturing approach
16: Let y ∈ 1×ql and x ∈ 1×pl be the augmented output for efficient operation of industrial steam-methane reformers. Ind. Eng. Chem.
and input test vectors Res. 54 (16), 4360–4370.
17: Compute y = (y − y )/σy and x = (x − x )/σx : opera- Kumar, A., Edgar, T.F., Baldea, M., 2017. Multi-resolution model of an indus-
trial hydrogen plant for plantwide operational optimization with non-uniform
tions are performed element-wise. steam-methane reformer temperature field. Comput. Chem. Eng. 107, 271–283.
18: Calculate test-output residual (yr ): x × β ; y∗ =
y∗ = Lee, J.-M., Yoo, C., Lee, I.-B., 2004. Statistical process monitoring with independent
y ∗ σy + y; yr = y − y∗
∗ component analysis. J. Process Control 14 (5), 467–485.
Li, W., Yue, H.H., Valle-Cervantes, S., Qin, S.J., 20 0 0. Recursive PCA for adaptive pro-
19: Compute yr = (yr − yr )/σyr cess monitoring. J. Process Control 10 (5), 471–486.
20: Calculate yr ∗ Cncomp ∗ (Cncomp )transpose
y∗r = MacGregor, J.F., Kourti, T., 1995. Statistical process control of multivariate processes.
Control Eng. Pract. 3 (3), 403–414.
21: Compute Q = y∗r 2
yr − MacGregor, J.F., Yu, H., Muñoz, S.G., Flores-Cerrillo, J., 2005. Data-based latent vari-
able methods for process analysis, monitoring and control. Comput. Chem. Eng.
29 (6), 1217–1223.
MATLAB and Statistics Toolbox, 2016b. The MathWorks, Inc., Natick, Massachusetts,
United States.
Ng, Y.S., Srinivasan, R., 2010. Multi-agent based collaborative fault detection and
CRediT authorship contribution statement identification in chemical processes. Eng. Appl. Artif. Intell. 23 (6), 934–949.
Nimmo, I., 1995. Adequately address abnormal operations. Chem. Eng. Prog. 91 (9).
Perk, S., Teymour, F., Cinar, A., 2010. Statistical monitoring of complex chemical pro-
Ankur Kumar: Conceptualization, Methodology, Software, For- cesses using agent-based systems. Ind. Eng. Chem. Res. 49 (11), 5080–5093.
mal analysis, Data curation, Writing - original draft, Visualiza- Qin, S.J., 1998. Recursive PLS algorithms for adaptive data modeling. Comput. Chem.
Eng. 22 (4–5), 503–514.
tion, Project administration. Apratim Bhattacharya: Conceptualiza-
Qin, S.J., 2003. Statistical process monitoring: basics and beyond. J. Chemom. 17
tion, Methodology, Formal analysis, Writing - original draft. Jesus (8–9), 480–502.
Flores-Cerrillo: Writing - review & editing, Supervision.
A. Kumar, A. Bhattacharya and J. Flores-Cerrillo / Computers and Chemical Engineering 136 (2020) 106756 19
Qin, S.J., 2012. Survey on data-driven industrial process monitoring and diagnosis. Venkatasubramanian, V., Rengaswamy, R., Kavuri, S.N., Yin, K., 2003. A review of
Annu. Rev. Control 36 (2), 220–234. process fault detection and diagnosis: part iii: process history based methods.
Reis, M., Gins, G., 2017. Industrial process monitoring in the big data/industry 4.0 Comput. Chem. Eng. 27 (3), 327–346.
era: from detection, to diagnosis, to prognosis. Processes 5 (3), 35. doi:10.3390/ Venkatasubramanian, V., Rengaswamy, R., Yin, K., Kavuri, S.N., 2003. A review of
pr5030035. process fault detection and diagnosis: part i: quantitative model-based methods.
Ricker, N.L., 1988. The use of biased least-squares estimators for parameters in dis- Comput. Chem. Eng. 27 (3), 293–311.
crete-time pulse-response models. Ind. Eng. Chem. Res. 27 (2), 343–350. Yamamoto, T., Shimameguri, A., Ogawa, M., Hashimoto, I., Kano, M., 2004. Appli-
Severson, K., Chaiwatanodom, P., Braatz, R.D., 2016. Perspectives on process moni- cation of statistical process monitoring with external analysis to an industrial
toring of industrial systems. Annu. Rev. Control 42, 190–200. monomer plant. IFAC Proc. Vol. 37 (1), 379–384.
Simon, D.L., Litt, J.S., 2011. A data filter for identifying steady-state operating points Yin, S., Ding, S.X., Xie, X., Luo, H., 2014. A review on basic data-driven approaches
in engine flight data for condition monitoring applications. J. Eng. Gas Turbine for industrial process monitoring. IEEE Trans. Ind. Electron. 61 (11), 1–10.
Power 133 (7), 71603. Yin, S., Gao, X., Karimi, H.R., Zhu, X., 2014. Study on support vector machine-based
Tsung, F., 20 0 0. Statistical monitoring and diagnosis of automatic controlled pro- fault detection in tennessee eastman process. Abstract and Applied Analysis, vol.
cesses using dynamic PCA. Int. J. Prod. Res. 38 (3), 625–637. 2014. Hindawi.
Venkatasubramanian, V., Rengaswamy, R., Kavuri, S.N., 2003. A review of process
fault detection and diagnosis: part ii: qualitative models and search strategies.
Comput. Chem. Eng. 27 (3), 313–326.