Professional Documents
Culture Documents
Dynamic Modeling of Complex Industrial Processes Data Driven Methods and Application Research 1st Edition Chao Shang (Auth.)
Dynamic Modeling of Complex Industrial Processes Data Driven Methods and Application Research 1st Edition Chao Shang (Auth.)
https://textbookfull.com/product/dynamic-mode-decomposition-data-
driven-modeling-of-complex-systems-1st-edition-j-nathan-kutz/
https://textbookfull.com/product/data-driven-fault-detection-for-
industrial-processes-canonical-correlation-analysis-and-
projection-based-methods-1st-edition-zhiwen-chen-auth/
https://textbookfull.com/product/group-processes-data-driven-
computational-approaches-1st-edition-andrew-pilny/
https://textbookfull.com/product/analysis-and-data-based-
reconstruction-of-complex-nonlinear-dynamical-systems-using-the-
methods-of-stochastic-processes-m-reza-rahimi-tabar/
Methods and Applications for Modeling and Simulation of
Complex Systems 18th Asia Simulation Conference AsiaSim
2018 Kyoto Japan October 27 29 2018 Proceedings Liang
Li
https://textbookfull.com/product/methods-and-applications-for-
modeling-and-simulation-of-complex-systems-18th-asia-simulation-
conference-asiasim-2018-kyoto-japan-
october-27-29-2018-proceedings-liang-li/
https://textbookfull.com/product/modeling-in-membranes-and-
membrane-based-processes-industrial-scale-separations-1st-
edition-anirban-roy-editor/
https://textbookfull.com/product/industrial-mathematics-and-
complex-systems-emerging-mathematical-models-methods-and-
algorithms-1st-edition-pammy-manchanda/
https://textbookfull.com/product/exploring-the-health-state-of-a-
population-by-dynamic-modeling-methods-1st-edition-christos-h-
skiadas/
https://textbookfull.com/product/monitoring-multimode-continuous-
processes-a-data-driven-approach-marcos-quinones-grueiro/
Springer Theses
Recognizing Outstanding Ph.D. Research
Chao Shang
Dynamic Modeling
of Complex Industrial
Processes: Data-
driven Methods and
Application Research
Springer Theses
The series “Springer Theses” brings together a selection of the very best Ph.D.
theses from around the world and across the physical sciences. Nominated and
endorsed by two recognized specialists, each published volume has been selected
for its scientific excellence and the high impact of its contents for the pertinent field
of research. For greater accessibility to non-specialists, the published versions
include an extended introduction, as well as a foreword by the student’s supervisor
explaining the special relevance of the work for the field. As a whole, the series will
provide a valuable resource both for newcomers to the research fields described,
and for other scientists seeking detailed background information on special
questions. Finally, it provides an accredited documentation of the valuable
contributions made by today’s younger generation of scientists.
Dynamic Modeling
of Complex Industrial
Processes: Data-driven
Methods and Application
Research
Doctoral Thesis accepted by
Tsinghua University, Beijing, China
123
Author Supervisor
Dr. Chao Shang Prof. Dexian Huang
Department of Automation Department of Automation
Tsinghua University Tsinghua University
Beijing Beijing
China China
This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
part of Springer Nature
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Supervisor’s Foreword
v
vi Supervisor’s Foreword
incorporated a priori into states and parameters of soft sensing models, respectively,
yielding two new dynamic soft sensing models: Probabilistic slow feature regres-
sion is proposed, which extracts useful dynamic features from time series data and
synthesizes information of multi-rate data; the classical dynamic partial least
squares model is enhanced with temporal smoothness imposed, and the over-fitting
phenomenon gets alleviated. Finally, a new dynamic nonlinear soft sensing
approach is developed, which describes process dynamics and nonlinearity by
first-order systems with pure delays and support vector machines, respectively.
Extensive simulations and real industrial case studies are adopted to testify the
effectiveness of the proposed methods. This thesis not only contributes to a new
theoretical framework of process data analytics but also provides effective strategies
for process monitoring, fault diagnosis, control performance assessment, and soft
sensing in industrial practice.
Beyond this thesis, the work done by Chao Shang during his Ph.D. program is
also excellent in other aspects. For example, the new modeling methods introduced
in this thesis have already been implemented in commercial softwares for advanced
process control. The deep learning technique for process data modeling developed
by Chao Shang, albeit not included in this thesis, has been widely applied to quality
prediction, crude classification, and carbon capture process modeling, and the
associated journal article has been among the most downloaded and cited articles in
Journal of Process Control. Up to now, his publications have been cited about 200
times in total according to the statistics of Google Scholar, which well reflect the
academic influence of Chao Shang’s doctoral research.
vii
Part of this thesis has been published in the following journal articles:
1. Chao Shang, Fan Yang, Xinqing Gao, Xiaolin Huang, Johan A.K. Suykens,
Dex- ian Huang, Concurrent monitoring of operating condition deviations and
process dynamics anomalies with slow feature analysis, AIChE Journal,
2015,61(11):3666–3682. (Reproduced with permission).
2. Chao Shang, Biao Huang, Fan Yang, Dexian Huang, Slow feature analysis for
monitoring and diagnosis of control performance, Journal of Process Control,
2016, 39:21–34. (Reproduced with permission).
3. Chao Shang, Biao Huang, Fan Yang, Dexian Huang, Probabilistic slow feature
analysis-based representation learning from massive process data for soft sensor
modeling, 2015, AIChE Journal, 61(12):4126–4139. (Reproduced with
permission).
4. Chao Shang, Xiaolin Huang, Johan A.K. Suykens, Dexian Huang, Enhancing
dynamic soft sensors based on DPLS: A temporal smoothness regularization
approach, 2015, Journal of Process Control, 28:17–26. (Reproduced with
permission).
5. Chao Shang, Xinqing Gao, Fan Yang, Dexian Huang, Novel Bayesian frame-
work for dynamic soft sensor based on support vector machine with finite
impulse response, 2014, IEEE Transactions on Control Systems Technology, 22
(4):1550–1557. (Reproduced with permission).
ix
Acknowledgements
xi
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ....... 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ....... 1
1.2 Literature Review on Multivariate Statistical
Process Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Principal Component Analysis . . . . . . . . . . . . . . . . . . . . . 3
1.2.2 Current Research Status of Process Monitoring . . . . . . . . . 5
1.3 Literature Review on Data-driven Soft Sensor Modeling . . . . . . . . 7
1.3.1 Partial Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.2 Current Research Status of Soft Sensing . . . . . . . . . . . . . . 10
1.4 Challenges and Opportunities . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.5 Contents and Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2 Monitoring of Operating Condition and Process
Dynamics with Slow Feature Analysis . . . . . . . . . . . . . . . . . . . . . . . 21
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 Slow Feature Analysis—Revisit . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.1 Definition of Slowness and Some Notations . . . . . . . . . . . 23
2.2.2 Optimization Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2.3 Solution Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2.4 Geometric and Statistical Properties of SFA . . . . . . . . . . . 27
2.2.5 Comparison with Classic Latent Variable Models . . . . . . . 28
2.3 Process Monitoring with SFA . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.3.1 Dimension Reduction of SFA . . . . . . . . . . . . . . . . . . . . . . 29
2.3.2 Monitoring Statistics Design with Slow Features . . . . . . . . 31
2.4 Case Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.1 Simulated CSTR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.2 TE Benchmark Process . . . . . . . . . . . . . . . . . . . . . . . . . . 38
xiii
xiv Contents
2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3 Control Performance Monitoring and Diagnosis
Based on SFA and Contribution Plot . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2 Control Performance Monitoring with SFA . . . . . . . . . . . . . . . . . 50
3.3 Control Performance Diagnosis with Contribution Analysis . . . . . . 51
3.3.1 Revisiting Contribution Analysis . . . . . . . . . . . . . . . . . . . 51
3.3.2 Control Performance Diagnosis . . . . . . . . . . . . . . . . . . . . 52
3.3.3 Comparison with the Covariance-Based Approach . . . . . . . 53
3.4 Case Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.4.1 A Simple Simulated Multivariate Process . . . . . . . . . . . . . 56
3.4.2 TE Benchmark Process . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4 Recursive SFA Algorithm and Adaptive Monitoring System
Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2 Updating Law for Covariance Matrices . . . . . . . . . . . . . . . . . . . . 66
4.3 The Recursive SFA Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.3.1 Improved Monitoring Statistics . . . . . . . . . . . . . . . . . . . . . 69
4.3.2 Rank-One Modification for the First Decomposition . . . . . 69
4.3.3 OIP Algorithm for the Second Decomposition . . . . . . . . . . 70
4.3.4 Complexities of the RSFA Algorithm . . . . . . . . . . . . . . . . 71
4.4 Adaptive Monitoring System Design . . . . . . . . . . . . . . . . . . . . . . 72
4.4.1 Control Limits Updating . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.4.2 An Improved Stopping Criterion for Model Updating . . . . 73
4.5 Case Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.5.1 Simulated CSTR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.5.2 An Industrial Heating Furnace System . . . . . . . . . . . . . . . 76
4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5 Probabilistic Slow Feature Regression for Dynamic
Soft Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.2 Probabilistic SFA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.2.1 Mathematical Definition . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.2.2 Parameter Estimation Using the EM Algorithm . . . . . . . . . 88
Contents xv
xvii
xviii Acronyms
Abstract This chapter provides an overview of the entire thesis. First, the research
background of process data analytics for modeling, monitoring and fault diagnosis in
industrial processes is introduced. Then an extensive review of existing research on
multivariate statistical process monitoring and soft sensor modeling is given. Third,
the opportunities and challenges in advancing industrial applications of process data
analytics are highlighted. Finally, the main contents and the layout of this thesis are
provided.
1.1 Background
Process industry has been playing an indispensable role in the national economy of
China. Typical examples include petroleum industry, power systems, metallurgical
industry, and paper-making industry, to name just a few. With the development of
technology, society and economy, an increasing number of industrial enterprises
have been emerging in China. What’s more, production plants erected become more
complicated and more densely distributed, which poses significant challenges for
operations and management. Therefore, any accidental situation in industrial practice
may lead to disastrous consequences, such as serious casualties, economic losses
and environmental pollution. Hence the safety and efficiency of operations in the
process industry are of great importance in safeguarding economic interests and
social benefits.
Control technologies have been extensively adopted in the process industry.
Although various disturbances universally exist, processes can still operate around
ideal working points that are specified in advance, since influence of disturbances
can be well compensated by the feedback behavior of control systems. It is hence
crucial for process operators to monitor the operating states of processes in an online
manner. For the sake of maintenance, any deviation from the nominal operating
condition shall be detected timely. Moreover, some key variables in industrial pro-
ductions, such as boiling point of petroleum products and pollutants content in emis-
sions, are directly related to final product qualities or environmental indices, which
shall receive more attentions than other variables and be carefully controlled. To
achieve this goal, real-time measurements are necessary. However, due to techno-
logical limitations or economic concerns, such key variables are often difficult to
measure directly. The only feasible source of information is the laboratory analy-
sis, which is typically time-consuming and involves significant delays. Therefore,
it cannot satisfy the requirement of real-time monitoring and control in practice. In
this sense, online monitoring of operating status of processes, and real-time predic-
tion of key variables, are two quintessential issues in productions of modern process
industries [1].
Model-based methods can provide accurate descriptions to process dynamics
and input-output relationships. Based on this, techniques for parameter estimations
and state estimations have been systematically developed [2–5], and successfully
applied to power systems and mechanical engineering, in which research objects are
amenable to clear physical mechanisms. Nevertheless, model industrial processes
typically involve a great number of measurements, and the inherent mechanism is
extremely complicated. It erects obvious obstacles for establishing accurate math-
ematical models, thereby heavily compromising the applicability of model-based
methods. Since 1990, with the ongoing advancement of sensor technology and
distributed control system (DCS), massive data in the process industry have been
collected. The explosive availability of process data provides an asset for decision-
making in the process industry. Under such circumstance, statistical process control
approaches emerged and became more and more prevalent. The most representative
techniques are multivariate statistical process monitoring (MSPM) and soft sensing,
which have received tremendous attentions in both academia and industries [6–9].
The standpoint of MSPM is to describe the probability distribution of process data
via multivariate statistical analysis. Based on statistical properties of data, one can
achieve online monitoring and evaluation of operating condition of industrial pro-
cesses, including abnormal event detection, fault diagnosis and fault reconstructions.
As for soft sensing technology, the core idea is to establish the mathematical mapping
between fast-sampled process variables and key variables that are hard to measure,
which can be used for online estimations as an alternative to the hardware sensor. To
build a soft sensor model, one typically first down-samples the process variables to
synchronize with the key variables, and regards it as a trivial regression task, which
can be tackled using pattern recognition and machine learning techniques. Next, we
proceed with a comprehensive overview for MSPM and soft sensing techniques.
When an industrial process has a relatively simple structure and a small number
of process variables, it is a practical approach to individually set the control limit
for each process variable to detect the abnormal behavior of the process. For exam-
ple, for a certain process variable x, an upper limit xmax and a lower limit xmin can
1.2 Literature Review on Multivariate Statistical Process Monitoring 3
be set, and when x > xmax or x < xmin an abnormal condition is considered to
occur. However, in modern industrial processes, there are usually hundreds of pro-
cess variables measured simultaneously. It hence becomes unrealistic to set control
limits for process variables individually. Moreover, under closed-loop control, pro-
cess variables may show significant correlations. On some occasions, it is likely that
correlations between variables are violated but control limits for individual variables
are still met. To address this challenge, the principal component analysis (PCA)-
based MSPM technique could help extracting low-dimensional information within
high-dimensional data, based on which simple monitoring statistics can be defined
and applied. In the past twenty years, it has been successfully applied in chemical
engineering, pharmaceutical industry, and iron-making industry, and it has laid a nice
foundation for implementation of advanced process control (APC) technology.
PCA was first proposed by Pearson in the early 20th century, which is the most
widely adopted method in multivariate statistical analysis. Next, we introduce its
principle and the resulted monitoring scheme, and then revisit some recent advances
in MSPM.
Assumed that process data constitute a matrix X ∈ R N ×m , where N and m stand for
the number of data sample and data dimension, respectively. Meanwhile, each dimen-
sion of data has already been normalized.1 In PCA model, the matrix X decomposes
as [10]:
X= X + E = TPT + E (1.1)
That is, matrix P is formed by singular vector associated with the first A singular
value. According to the property of SVD, we know that the projection directions of
different PCs are orthogonal, and different PCs are linear independent. There are a
variety of approaches for selection of the number A of PCs, and a comprehensive
introduction in this vein is given by [11].
1 Correlation-based
PCA requires that each dimension has zero mean and unit variance, while
covariance-based PCA only asks for zero-mean normalization.
4 1 Introduction
x1
Direction of 2nd PC p2
A(N 2 − 1)
Tα2 = FA,N −A,α (1.5)
N (N − A)
where FA,N −A,α denotes the α-confidence level of the F-distribution with A and
N − A degrees of freedom. In addition, the variations in the residual subspace can
be characterized by the squared prediction error (SPE) statistic:
⎡ ⎤ h10
cα 2θ2 h 20 θ2 h 0 (h 0 − 1) ⎦
Q A,α = θ1 ⎣ +1+ (1.7)
θ1 θ12
m
2θ1 θ3 j
h0 = 1 − , θj = λi , j = 1, 2, 3 (1.8)
3θ22 i=A+1
Q A,α = gχh,α
2
, (1.9)
where g = θ2 /θ1 , h = θ12 /θ2 . Different from the T 2 statistic, the SPE statistic
describes the violation of correlations between process variables. The T 2 and SPE
statistics are complementary to each other, in that they provide a complete allowance
for variations in the data space. For convenience in practical use, Yue and Qin [14]
put forward a combined index, which synthesize information within both T 2 and
SPE statistics:
SPE(x) T 2 (x)
ϕ(x) = + xT Φx. (1.10)
δα2 Tα2
Figure 1.2 summarizes the framework of existing studies on MSPM, which primarily
includes two research streamlines. The first streamline is to design statistical mod-
els based on different properties of process data. Although linear correlation has
already been well addressed by PCA-based monitoring approaches, its underlying
assumption is still prone to some obvious limitations.
In order to address the non-Gaussianity of process data, many researchers have
proposed to use independent component analysis (ICA) and its extensions for pro-
cess data modeling and monitoring [15–17]. High-order moment information is
adopted in ICA to measure the independence between latent variables. Due to the
non-Gaussianity, the control limits in ICA-based monitoring approaches are rather
difficult to establish. One needs to resort non-parametric methods such as kernel
density estimation [18, 19] to get reliable estimations of control limits.
Focusing on the nonlinear correlations in process variables, Kramer [20] first pro-
posed the concept of autoassociative neural networks to achieve nonlinear dimension
reduction of high-dimensional data. Later, Dong and McAvoy [21] developed the
principal curve model, in which nonlinear curves and neural networks are used to
characterize the most information of process data. With the prosperity of machine
6 1 Introduction
Non-Gaussianity
Nonlinearity
Dynamics
Time-Varying
Behaviors
learning techniques, kernel learning methods, notably kernel PCA (KPCA) [22], has
been widely adopted for process monitoring [23–25]. The basic idea of kernel learn-
ing methods is to map the original input to a high-dimensional reproducing kernel
Hilbert space (RKHS), and translate the problem of nonlinear dimension reduction
into a linear PCA modeling task, which can be tackled conveniently by means of the
kernel trick [22]. A potential limitation of kernel models is that kernel parameters
are commonly difficult to choose. Meanwhile, the model complexity soars linearly
with the number of available data samples, thereby inevitably incurring high costs
in practical implementations.
It is often the case that industrial processes have obvious dynamic characteris-
tics. To address this issue, Ku et al. [26] proposed dynamic PCA (DPCA) model
for process monitoring by including some lagged measurements into the classic
PCA model. State-space models with Markov properties pave an efficient way for
describing process dynamics. To build state-space models from time series data,
researchers have proposed to employ canonical variate analysis (CVA) [27, 28] and
PLS [29]. Dunia and Qin discussed the problem of fault diagnosis based on state-
space models [30].
Industrial processes are time-varying themselves. To accommodate such time-
varying behaviors, adaptive monitoring approaches have been proposed and well
studied. Wold [31] first proposed to utilize exponentially weighted moving average
(EWMA) approach to update the model in an online manner. An alternative method
called moving window (MW) was proposed by Wang et al. [32]. The common stand-
point of these adaptive methods is to gradually alleviate the effect of historical data
onto the model, thereby enabling the model to adapt to recent data and track inherent
changes within the process. In order to reduce the computational burden in online
implementation, Li et al. [33] proposed two different algorithms for recursive PCA.
A comprehensive summary of existing adaptive monitoring methods is given by
Jeng [34].
1.2 Literature Review on Multivariate Statistical Process Monitoring 7
The other streamline in Fig. 1.2 proceeds with the functionality of monitoring
systems, from fault detection, fault diagnosis to fault reconstruction and fault prog-
nosis. After detecting anomalies with MSPM, we need to locate the possible faulty
variables as the “root cause”, or identify the fault type with some prior faulty infor-
mation at hand. Therefore, useful information within abnormal data can be further
excavated and leveraged to assist the decision-making of process operators. The
contribution plot is the most-used method for fault diagnosis, which quantitatively
evaluates the contribution of each process variable to the fault by decomposing mon-
itoring statistics. Fault reconstruction aims to determine the magnitude of the fault
after the fault has been correctly classified. Fault prognosis focuses on modeling
the evolving process of incipient faults and further predicting the remaining useful
life, which is helpful for taking maintenance actions and avoiding severe dangerous
events. Notice that methodologies of fault diagnosis, fault reconstruction and fault
prognosis heavily depend on the statistical model that is used for modeling process
data. Existing studies are primarily based on the basic PCA model and its extensions.
For other models, it is not a trivial task to repeat the methodologies applicable to the
PCA case, and one needs to design tailored techniques to achieve fault diagnosis,
reconstruction, and prognosis.
The idea of soft sensing was first proposed by Brosilow in 1978 when he and his
colleagues studied the inferential control problem of multi-rate systems [35, 36].
Figure 1.3 illustrates the typical multi-rate sampling phenomenon in industrial pro-
cesses. Process variables, such as temperature, pressure, and flow-rate, are usually
sampled based on a short sampling interval, shown as circles in Fig. 1.3. Different
from process variables, measurements of quality variables are unavailable most of the
time, as indicated by red crosses in Fig. 1.3. Typically they can be obtained only by
manual laboratory analysis, which leads to large, sometimes even irregular, sampling
intervals. In order to obtain real-time predictions for quality variables at all snapshots,
we can establish function mappings from process variables to quality variables by
analyzing limited historical data. Such mathematical models can be used in place of
hardware sensors, and this is why they are commonly referred to as “soft sensors”.
In this way, real-time monitoring and control of key quality indices become feasible.
Åström and McAvoy regarded soft sensing as the most challenging and fundamental
problem in chemical process control [37]. After decades of development, significant
progress has been made in both theory and industrial applications of soft sensing
technology. International enterprises such as Aspen and Honeywell have launched
commercial software products that have been widely adopted in a variety of industrial
production processes.
8 1 Introduction
Quality
Variables x x x x
As a basic approach in soft sensor modeling, PLS first emerged in the field of
multivariate statistical analysis. In 1975, Wold, a famous Swedish statistician, first
introduced PLS to path analysis problems in economics and social science [38].
Next few decades witnessed remarkable flourish in practical applications of PLS,
especially in economics and chemometrics. In process industries, PLS-based soft
sensor models have been applied to both continuous production processes [39] and
batch processes [40]. In the subsection to follow, we first briefly introduce the basic
principle behind PLS model, and then review current research status of soft sensing
methodologies.
Assume that process data and quality data are included in matrices X ∈ R N ×m and
Y ∈ R N × p , respectively. Here p denotes the number of quality variables. In industrial
practice, the sampling frequency of process data is much higher than that of quality
data. Therefore, the actual number of available process data samples is much greater
than N . It indicates that matrix X should be obtained by down-sampling the fast-
rate process data. Assuming that each dimension of both X and Y has already been
scaled to zero mean and unit variance in advance, the PLS model makes use of an
A-dimensional subspace that explains information in both input X ∈ R N ×m and
output Y ∈ R N × p . Mathematically, the PLS model is described as follows:
X = TPT + E
(1.11)
Y = TQT + F
where T ∈ R N ×A is the score matrix. P ∈ Rm×A and Q ∈ R p×A stand for the loading
matrices of X ∈ R N ×m and Y ∈ R N × p , respectively. The optimal number of PCs A
can determined by means of cross-validation. To derive parameters the PLS model,
the nonlinear iterative partial least squares (NIPALS) is a basic and widely adopted
approach, whose implementation details are given in Table 1.1.
After obtaining model parameters W = [w1 , . . . , w A ], P = [p1 , . . . , p A ] and
Q = [q1 , . . . , q A ], given a new input data sample xnew , the corresponding prediction
is given by:
1.3 Literature Review on Data-driven Soft Sensor Modeling 9
which is a linear function of input data and can be directed adopted for online
prediction of key variables.
It is worth mentioning that, we can also achieve process monitoring and fault
diagnosis based on the subspace decomposition made by PLS. In comparison with
the PCA-based monitoring scheme, here we can incorporate additional information
about quality variables of special interests. First, given an input data x, we need to
calculate the score and residual in the low-dimensional subspace:
t = RT x (1.13)
x̃ = (Im − PR )x T
(1.14)
where R = W(PT W)−1 . Then the Hotelling’s T 2 statistic and the Q-statistic can be
defined as follows [42]:
A(N 2 − 1)
T 2 = tT Λ−1 t ≤ FA,N −A,α (1.15)
N (N − A)
Q = ||x̃||2 ≤ gχh,α
2
(1.16)
It seems that the T 2 and Q-statistics in PLS are similar to the T 2 and SPE statistics
in PCA; however, their physical interpretations are fundamentally different. In the
context of PLS, the T 2 statistics is commonly conceived as abstracting variations that
are highly correlated with quality variables, while the Q-statistic is defined based
on variations that are irrelevant with quality variables [43]. Since the main focus of
this thesis is on the application of unsupervised learning models in MSPM, we are
not meant to cover further details about the PLS-based monitoring and diagnosis
approaches, and interested readers are referred to [44, 45].
10 1 Introduction
In this thesis, we pay particular attentions to the application of PLS in soft sensor
modeling. Under closed-loop control, both process variables and quality variables
are affected by the intrinsic disturbance underlying the process, thereby exhibiting
obvious correlations and cross-correlations. In the PLS model, “common cause”
disturbances are interpreted as variations in the low-dimensional subspace by maxi-
mizing the covariance between input scores and output scores. Therefore, similar to
the application of PCA in MSPM, PLS-based soft sensors embody profound physical
meanings as well.
Studies on soft sensing can also be categorized into two different streams, as shown
in Fig. 1.4. The first streamline concentrates on making various assumptions on sta-
tistical properties of models according to different process characteristics. Let us take
PLS as an example. Strictly speaking, PLS essentially delineates the linear cross-
correlation between process variables and quality variables under a single operating
condition. However, it is possible that nonlinear correlation relationship exists, and
hence many nonlinear generalizations of PLS have been proposed, for instance, the
kernel PLS (KPLS) [46] and the neural network PLS (NNPLS) [47]. Qin [48] first
pointed out the great potential of ANN in solving soft sensing problems. According
to the review article by Kadlec et al. [6], about one third of soft sensors implemented
in practice are developed based on ANN. Despite such prevalence, the training pro-
cess of classic ANN is often subject to local optimum, and it is usually difficult to
determine an optimal network structure. Support vector machine (SVM) based on
statistical learning theory is advantageous to ANN due to its ease in global optimiza-
tion and parameter selection. Because of these merits, SVM has been a hot topic
in machine learning community [49]. Besides, SVM has an excellent generalization
ability in dealing with small sample problems, thereby being particularly suitable for
soft sensor modeling [50, 51].
It is often the case that industrial processes are characterized by long settling
times, and the casual relationship between process variables and quality variables
may have evident dynamics. In other words, quality measurements can be affected
by process measurements in a period of time. Kano et al. [52] first extended PLS
to its dynamic counterpart, namely dynamic PLS (DPLS), to enhance the predic-
tion accuracy of product quality in industrial distillation columns. Similar to DPCA,
DPLS incorporates some lagged historical input measurements into the model, which
takes process dynamics into account. Ma et al. [8] first proposed to filter the input by
means of impulse response (IR), and then feed the filtered input to classic static soft
sensing models. This strategy describes dynamics and steady relationship between
process variables and quality variables separately, which has a typical Wiener struc-
ture, and many research efforts have been made in this vein [53, 54]. Some other
dynamic soft sensing methods include recurrent neural networks [55, 56] and wavelet
transformations [57].
To meet varying product demands in the market, there are always multiple oper-
ating modes in industrial processes. Under such circumstance, soft sensor models
based on single operating condition have obvious limitations. Hence, multi-mode
soft sensor models have been proposed. In order to realize partitioning of different
operating modes, one has to first carry out clustering of process data based on certain
similarity criterion, and then establish an individual soft sensor model for each data
cluster. The most popular approach for data clustering is the C-means clustering
[58, 59]. An alternative strategy is to employ probabilistic models such as the Gaus-
sian mixture model [60–62]. Recently, Ge et al. [63] proposed to use external analysis
to extract common information from multi-mode data, which finally yields a unified
prediction model for multi-mode soft sensing.
The other research streamline in soft sensing focuses on the implementation prob-
lems in industrial practice, such as variable selection, data preprocessing and model
maintenance, as indicated by the horizontal axis in Fig. 1.4. If all available pro-
cess variables are used for soft sensor modeling, the resulted model will tend to be
excessively complicated and perform poorly in terms of generalization capabilities.
Meanwhile, a high implementation cost will be incurred. Therefore, it is necessary
to pick out a small percentage of dominant process variables as soft sensor inputs for
the sake of model simplifications. A pragmatic approach is to resort to both expert
knowledge and correlation analysis [39]. In a recent work [64], variable selection
methodologies for PLS-based soft sensor models are comprehensively revisited.
Moreover, before fitting a soft sensor model, one typically needs to carry out data
preprocessing to handle potential outliers and missing data due to the imperfection
of industrial data. The classic 3σ -approach is believed to be practically efficient in
detecting outliers [65]. Later, quantile estimation-based robust methods were fur-
ther proposed, for example, the Hampel test [66]. Besides, a promising approach is
to model outliers statistically in virtue of the tail effect of Student’s t distribution
[67, 68]. Hodge and Austin [69] provided a comprehensive survey of commonly used
techniques for outlier detection and identification. Targeting on the issue of missing
12 1 Introduction
From the literature review in Sects. 1.2 and 1.3, one can see that in the past twenty
years techniques of MSPM and soft sensing have experienced rapid development and
have been pervasively applied in a large variety of industrial scenarios. Nevertheless,
there still exists room to further enhance their applicabilities, with the aim to further
promote the level of automation and intelligent decision-making in process industry.
Next we highlight two primary research challenges.
(1) Research challenges in MSPM
Classic MSPM techniques show significant limitations in dealing with frequent pro-
cess operating condition switches in modern process industry. One one hand, the
process objectives could be oriented actively according to planning and scheduling
with the aim to meet changing demands for its products. For example, the setpoint of
the melt index in a propylene production process may be adjusted to meet the real-
time requirements of the market. On the other hand, some variations or disturbances,
for instance, temperature changes in the ambient environment or the operating con-
dition changes in the upstream unit, might as well exert influence on the operating
condition of the process. In both cases, the steady distribution of process data under
normal operating condition is disrupted, and therefore, alarms would be consistently
triggered by basic MSPM methods; however, if the process remains well controlled
1.4 Challenges and Opportunities 13
variables are related to process data within the process settling time; however,
the latent variables themselves are still supposed to be temporally independent in
the model structure. We know that latent variables essentially stand for the process
states, and thus shall be temporally correlated. In this sense, the generic indepen-
dence assumption in latent variable models is inadequate. Moreover, model parame-
ters should have some temporal smoothness since model dynamics is quite unlikely
to have abrupt changes due to the inertia of industrial processes. Unfortunately, this
issue remains unaddressed by existing dynamic data-driven models.
The contents to be covered in this thesis are summarized in Fig. 1.5. We intend to
impose prior knowledge of temporal smoothness on model states and model parame-
ters, respectively, to devise new data-based methods for dynamic modeling of indus-
trial processes. An in-depth integration of both prior knowledge and information
within massive process data is to be achieved, based on which the applicability of
MSPM and quality prediction techniques can be further enhanced. By incorporating
temporal smoothness information on model states, we first propose a new approach
to process data modeling and the associated MSPM technology based on slow feature
PRINSSI
IMANDRA
PRINSSI
Katsokaas, näin!
OTRO
Hahhaa!
IMANDRA
Kuka nauroi!
OTRO
Harakka!
IMANDRA
HOVIHERRA
OTRO
OTRO
IMANDRA
HOVIROUVA
Kultakruunuineen ja kultakengät jaloissa täytyy prinsessan heti
palata pyytämään anteeksi suur'armolliselta, hänen korkeudeltaan
Kaukovallan prinssiltä. (Menee hullunkurisesti niiaten.)
OTRO
PRINSSI
OTRO
PRINSSI
HOVIHERRA
OTRO (palaa)
PRINSSI
HOVIHERRA
OTRO
Minä tulin pyytämään… Ei! Oh, minä nauran, vaikka pitäisi itkeä.
Kas, eihän täällä ole prinssiä! Paimenpoika, mitä sinä haet?
PRINSSI
IMANDRA
PRINSSI
IMANDRA
PRINSSI
IMANDRA
Minäkö häijy! Mutta missä minä olen ennen kuullut sinun äänesi?
PRINSSI
Ehkä unissanne.
IMANDRA
PRINSSI
IMANDRA (veitikkamaisesti)
PRINSSI
Niin, arvatkaas!
IMANDRA
PRINSSI (huokaillen)
Nii — niin.
IMANDRA
PRINSSI
IMANDRA (huokaillen)
Hohoo, niin. Mutta merkillistä, kuinka sinä olet Kaukovallan
prinssin näköinen.
PRINSSI
IMANDRA
PRINSSI
IMANDRA
PRINSSI
IMANDRA
PRINSSI
IMANDRA
En ymmärrä.
PRINSSI
IMANDRA
PRINSSI
IMANDRA
PRINSSI
IMANDRA
Entä sitten?
PRINSSI
IMANDRA
Poika parka!
PRINSSI
IMANDRA
Todellakin! Sinä olet kuin prinssin kuva, eikä sinua erottaisi juuri
muusta kuin puvusta ja tukasta.
PRINSSI
IMANDRA
IMANDRA
IMANDRA
Kuules poika! Soitapa huilulla, niin minä tanssin. Tästä tulee
hauskaa!
PRINSSI
IMANDRA
Sano se pian!
PRINSSI
En minä kehtaa.
IMANDRA
PRINSSI
IMANDRA
PRINSSI
IMANDRA
Eikö sen kummempaa. No!
PRINSSI
IMANDRA
PRINSSI
IMANDRA
HOVIROUVA
IMANDRA
HOVIHERRA
Holhoojana otan minä kultakruunun päästänne, te ette ole enään
arvokas sitä kantamaan.
IMANDRA
HOVIROUVA
IMANDRA
HOVIROUVA (suuttuvinaan)
IMANDRA
HOVIHERRA
IMANDRA
HOVIHERRA
HOVIROUVA
Mitä te teitte?
HOVIHERRA
HOVIROUVA
Oo, te…
HOVIHERRA
Teidän hajuvetenne…
HOVIROUVA (aivastaa)
Äst! Terveydeksenne!
NÄYTÖS II
Metsä.
PRINSSI
IMANDRA
PRINSSI
IMANDRA
PRINSSI
Tämä on ihanaa unta valveilla.
IMANDRA
PRINSSI
Olemme eksyksissä.
IMANDRA
PRINSSI
Levätkäämme!
IMANDRA
IMANDRA
IMANDRA
PRINSSI
IMANDRA
PRINSSI
IMANDRA
PRINSSI
IMANDRA
Luulikohan prinssi tällä noitapeilillä lumoovansa? Siinä erehtyi!
PRINSSI
Et uskalla.
IMANDRA
IMANDRA
PRINSSI
IMANDRA
PRINSSI
Se on pieni peikko.
IMANDRA
Peikko, uh!
PRINSSI
Nälkä.
IMANDRA
PRINSSI
Kiitä kyökkimestaria.
IMANDRA
PRINSSI
IMANDRA
PRINSSI
Anna, anna! Minä syön ja syön, mutta se peikko vain parkuu. Oh,
kuinka minun on nälkä! Ja minä luulin, että tämä oli niin ihanaa. Sinä
olet paha poika. Tuo paikalla mesikakkuja!
PRINSSI
IMANDRA
PRINSSI
IMANDRA
PRINSSI
IMANDRA
PRINSSI