12868 IEEE SENSORS JOURNAL, VOL. 21, NO.

11, JUNE 1, 2021

A Review on Soft Sensors for Monitoring,

Control, and Optimization of Industrial Processes
Yuchen Jiang , Student Member, IEEE, Shen Yin , Senior Member, IEEE, Jingwei Dong,
and Okyay Kaynak , Life Fellow, IEEE

Abstract —Over the past twenty years, numerous research

outcomes have been published, related to the design
and implementation of soft sensors. In modern industrial
processes, various types of soft sensors are used, which play
essential roles in process monitoring, control and optimiza-
tion. Emerging new theories, advanced techniques and the
information infrastructure have enabled the elevation of the
performance of soft sensing. However, novel opportunities
are accompanied by novel challenges. This work is motivated
by these observations and aims to present a comprehensive
review of the developments since the start of the millennium.
While a few books and review articles are published on
the related topics, more focus on the most up-to-the-date
advancement is put in this work, from the perspective of systems and control.
Index Terms — Soft sensor, soft sensing, process monitoring, optimization and control, industrial process.

I. I NTRODUCTION the interaction with the users and to provide intuitive

context-related data, smart (intelligent) sensors with wire-
I N INDUSTRIAL processes, sensing enables the Silicon-
based monitoring and control systems to gain access to
the internal information about the operating statuses, the state
less communication capabilities and standalone processing
functionalities are proposed. As observed from the industrial
trajectories, and the external variations reflected by the envi- practice, numerous sensors and sensor networks are installed
ronmental factors. The traditional sensing devices are designed at various locations to collect the process data. They play an
based on the physical laws require solid foundations and essential role as a part of the information infrastructure.
insights from natural science. Before their application in While we have seen evident increases in the availability of
engineering systems, signal (energy) transformation and data- physical sensors, the collected yet underexploited raw data will
sampling are usually required. More recently, to improve eventually become data garbage without proper data analysis,
information extraction and knowledge exploitation techniques
Manuscript received September 30, 2020; accepted October 19, 2020. [1]–[3]. Besides, it is neither economic nor beneficial to
Date of publication October 22, 2020; date of current version May 28, install sensors beyond requirement, because calibration and
2021. This work was supported in part by the National Natural Science
Foundation of China under Grant 61873073, in part by the Fok Ying maintenance will lead to unnecessary workload and cost. As an
Tung Education Foundation under Grant 161056, and in part by the alternative to the sensors implemented in the form of physical
National Defense Basic Scientific Research Program of China under entities, soft sensors have become indispensable in the modern
Grant JCKY2017212C005. The associate editor coordinating the review
of this article and approving it for publication was Dr. Edward Sazonov. industry. The model-driven family of soft sensors is most
(Corresponding author: Shen Yin.) commonly based on first principle models [4]. Nevertheless,
Yuchen Jiang is with the Department of Control Science and Engi- when the necessary a priori knowledge is unavailable, or the
neering, Harbin Institute of Technology, Harbin 150001, China, and also
with the Chair of Automatic Control Engineering, Technical University of modelling process is overcomplicated, the construction of soft
Munich, 80333 Munich, Germany (e-mail: [email protected]). sensors also depend on data-driven techniques. Specifically,
Shen Yin is with the Department of Control Science and Engineer- soft sensors make the use of the underlying relations between
ing, Harbin Institute of Technology, Harbin 150001, China (e-mail:
[email protected]). the accessible process variables and provide readouts of the
Jingwei Dong is with the Delft Center for Systems and Control, Delft required physical quantities and/or the performance indices.
University of Technology, 2628 CD Delft, The Netherlands (e-mail: The development of soft sensors has been greatly facilitated
[email protected]).
Okyay Kaynak is with the Department of Electrical and Electron- by the achievements in data science, computing and commu-
ics Engineering, Bogazici University, 34342 Istanbul, Turkey, and also nication technologies, statistical tools, and machine learning
with the School of Automation and Electrical Engineering, University techniques. For instance, networking of large-scale systems
of Science and Technology Beijing, Beijing 100083, China (e-mail:
[email protected]). and system-of-systems allows real-time information exchange.
Digital Object Identifier 10.1109/JSEN.2020.3033153 Geographically dispersed process data can be used to drive the

1558-1748 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Indian Institute of Technology. Downloaded on May 26,2024 at 02:45:55 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: REVIEW ON SOFT SENSORS FOR MONITORING, CONTROL, AND OPTIMIZATION OF INDUSTRIAL PROCESSES 12869

soft sensors, and in turn, the outputs of the soft sensors can be
used to drive the controllers and the process monitors [5]–[7].
Novel communication protocols and techniques help to make
use of the communication bandwidth, and in the meantime,
reduce the probability of package loss and the chance of
external attacks.
In another aspect, to achieve better precision, reliability and
adaptiveness, many complex models are adopted, the training
and the online updating of which require much computing Fig. 1. Different categories of “sensors”.
power. For instance, in a deep neural network, there can
be thousands of parameters to be determined and optimized.
Along with this, a colossal amount of historical data is fed to • The combination of innovative sensor principles with
the construction and the validation of the soft sensors, from modern methods of data analysis and modelling using
heterogeneous sources, most often with different sampling process and product knowledge [13];
rates. Fortunately, the emerging computing techniques (e.g., • The combination of analytical hardware data (from sen-
cloud computing, fog computing, parallel computing, etc.) and sors, analytical devices, instruments and actuators) with
database techniques provide novel solutions to dealing with the mathematical models that create new real-time informa-
big data. tion about the process [14].
This work investigates the development of soft sensors It should be emphasized that there are huge differences
over the past two decades. The significance of soft sens- in the concept of flexible sensors and that of soft sensors
ing, or rather, its irreplaceability, in improving the production discussed in this review. Flexible sensors generally refer to
safety and product quality management is discussed from the non-rigid or deformable physical sensors, such as wearable
perspectives of system monitoring, control and optimization. sensors [15].
For instance, it will be shown that soft sensors are not subject
to the various physical constraints (such as the space for instal- B. Scope and Structure of This Article
lation and the exposure to extreme working conditions). Addi-
In this article, we would like to mainly explore the answers
tionally, it will be argued that many more unique potentials
to the following questions about soft sensors and soft sensing.
of the soft sensors will be magnified in the coming decade,
where virtual instruments and digital avatars are the keys to (Q1) What is the general procedure to construct a soft sensor
elevate full life cycle management of the industrial processes. applicable to the industrial processes?
Before the further use of the buzz words (“soft sensor” (Q2) How can soft sensors be used to help with the monitor-
& “soft sensing”), the related terminologies are clarified and ing, control and optimization of industrial processes?
defined first to avoid confusion. (Q3) Are there any popular areas of soft sensing applica-
tions?
(Q4) What are the major differences in design for different
A. Terminology application scenarios?
The terminologies that are frequently used in literature can (Q5) What advancement have been made during the past
be grouped into two as shown in Fig. 1 (Different categories of twenty years? What challenges are still existing? What
“sensors”): (i) Soft sensor, virtual sensor [8], software sensors are the promising future directions?
[9], soft sensing [10]; (ii) Physical/hardware based sensor [11]. This article is structured as follows. The next section intro-
Soft sensing refers to the approaches and the algorithms duces the related surveys and reviews. Section III elaborates
that are used to estimate or predict certain physical quanti- the general procedures necessary for soft sensor construction.
ties or product quality in the industrial processes based on Section IV summarizes the common problems and advance-
the available measurements and knowledge. Soft sensors are ment in the available solutions. Afterwards, Section V presents
distinguished from the physical sensors in the way that they are a summary of the state-of-the-art applications to the industry,
implemented on computer software-based systems or embed- and VI discusses the open challenges and the future directions.
ded systems. In some limited cases, soft sensors can be
regarded as the digital projections of the hardware sensing II. R ELATED R EVIEW W ORKS
devices in the virtual space, whereas in many other cases, there A few books and review/survey type of articles can be seen
is no physical counterpart. It is to be noted that soft sensors can in literature that discuss soft sensors and the approaches to
overcome the constraint that some physical quantities cannot their design. However, there is no work that presents a com-
be measured accurately in the desired period of time. prehensive review of the progress over the past two decades
In literature, soft sensors are also defined as: (2000–2020) and an outlook for the coming decades or giving
• Mathematical models used to predict the behavior of real answers to the questions proposed in the previous section.
systems [12]; Some of the existing review-type literature is limited by the
• The use of mathematical or statistical models to enrich time-span of investigation. For instance, in the book Soft
the information measured by actual online sensors and Sensors for Monitoring and Control of Industrial Processes
offline sources such as laboratory data [9]; published in 2007 [16], over 60% of its references are before

Authorized licensed use limited to: Indian Institute of Technology. Downloaded on May 26,2024 at 02:45:55 UTC from IEEE Xplore. Restrictions apply.
12870 IEEE SENSORS JOURNAL, VOL. 21, NO. 11, JUNE 1, 2021

Fig. 2. Construction procedure for a soft sensor.

Fig. 3. Data preprocessing tasks.

2000, and only 3% are later than 2005. Many of the approaches
described therein have, since then, been replaced or improved
by novel techniques. In the review [11], only 10% are pub- Fig. 4. Dimensions for categorization.
lished within the past five years. In [17], all of the referred
work is dated before 2015, most of which are the preliminary
results published at conferences. the procedures for soft sensor construction, the application-
In another aspect, some review articles discuss specific cate- related key factors and the performance evaluation indices are
gories of modelling approaches or focus on specific application firstly discussed.
areas. For instance, Chapter 2 of the recently published book
[12] only covers deep learning-based approaches; In [17], A. Key Factors and Performance Evaluators
only the regression-based approaches are investigated. The Although soft sensors can be designed to generate variable
articles [11], [14] and [13] focus only on the applications to estimations or key performance predictors online, the design
polymer processing, the upstream bioprocesses, and the food procedure can hardly be carried out in a goal-oriented manner
& beverage processes, respectively. to directly meet the key requirements and technical specifica-
Based on the fruitful innovative outcomes by senior tions in monitoring, control and optimization (MCO). Instead,
researchers and learning from the existing review work, this it is more of the job of the design engineers of the specific
article presents the current industrial practices and the most sub-tasks to achieve these specifications based on the available
significant advancement over twenty years and introduces the soft sensing data. What they can do is to put forward the
state-of-the-art research where novel techniques inject fresh requirements for the design specifications of the soft sensors.
blood into this field of study. In the meantime, the most In Table I, the differences in the targets of MCO and those
classical and influential papers, published even if over twenty of soft sensor design are highlighted. It can be seen that despite
years ago, are cited whenever necessary. the wide variety of sub-tasks and the application scenarios,
there are some common factors to characterize the design
III. P ROCEDURES FOR S OFT S ENSORS C ONSTRUCTION specifications. Based on the rightmost column of Table I,
One of the major efforts of this work lies in finding the the answer to “What are the significant performance indices
answers to the question “How can soft sensors be used to and criteria used for the evaluation and the validation of the
help with monitoring, control and optimization of industrial soft sensors” is explored.
processes?” Bearing this in mind, we must first clarify what Availability: While any sensing device or measuring instru-
the most significant system specifications are and design ment requires a minimum time to give outputs, it is acceptable
parameters for such purposes. To this end, before diving into as long as two conditions are met: (i) the sampled-data system

Authorized licensed use limited to: Indian Institute of Technology. Downloaded on May 26,2024 at 02:45:55 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: REVIEW ON SOFT SENSORS FOR MONITORING, CONTROL, AND OPTIMIZATION OF INDUSTRIAL PROCESSES 12871

TABLE I
C OMPARISON OF THE S PECIFICATIONS OF C ONVENTIONAL D ESIGN AND S OFT S ENSOR D ESIGN

respects the Nyquist-Shannon sampling theorem, i.e., in order tive or qualitative self-evaluations as the degree of reliability
to restore the analog signal free from distortion, the sampling as a supervisory output, which provides an additional refer-
frequency should be no less than twice the maximum fre- ence for the fault diagnosis systems and the control systems.
quency in the analog signal spectrum; and (ii) the delay is Technically, this requires the models to be designed in the
negligible compared to the system dynamics and will not lead framework of the probability theory. Furthermore, albeit long
to notable performance degradation. time-delay, whenever there are physical sensors available to
As a novel data source for MCO purposes, soft sensors must directly measure the corresponding outputs of the soft sensors,
provide stably available outputs. It should be noted that the corrections and adaptation should be carried out, which is the
availability issue is quite different from that of physical sensors only chance for online calibration.
because, at the online stage, the soft sensors are driven by a The reliability of the soft sensors also indicates the robust-
set of variables which are also measured online, rather than ness against the performance degradation in the data sources.
by the perception of the environment. Therefore, the problems Specifically, when one or more dependent measurements are
of synchronization, multi-rate, and heterogeneous data have biased readouts (lack of precision) or outliers (wrong data),
to be considered. Usually, the rate of giving soft sensing the soft sensor should not be sensitive to them. Besides, soft
outputs cannot exceed the lowest sampling rate among the sensing depends on the well-functioning of all other data
dependent data sources, and the nominal time-delay of the sources. If faults occur at the related physical sensors, it is
soft sensor is determined by the sum of the longest delay of expected that the soft sensor can identify the situation and in
each physical sensor or measuring instrument, the transmission turn help to alert the corresponding faulty data source. This is
(network communication) delay, and the online computational essential for blocking the propagation of faults.
time (algorithm efficiency). Trustworthiness: The trustworthiness issue is raised here
Reliability: It is important that the outputs (estimations and because of the increasing risks of external attacks to the cyber-
predictions) of the soft sensors do not deviate far from the true physical system (CPS) based industrial processes, and the
values. As reported in [16], the performance of some predictive popularized practice of distributed and networked deployment
models degrades promptly due to error propagation, such as of the sensors (sensor networks).
the moving average model and the nonlinear moving average Nowadays, CPS security is not only a task in the ICT
model. For MCO applications, it would lead to disasters if the (Information and Communication Technologies) sector but
accuracy cannot meet the design specifications: there will be also a novel challenge in the systems and control domain.
miss detections in case of faults, and the closed-loop control If the communication channels are compromised, soft sen-
system cannot be stabilized. sors will be driven by malicious external signals, which
Regarding the evaluation of the soft sensors, apart from the can present a colossal problem to the overall safety of the
average accuracy, it is sometimes more important to check monitored or control systems. Especially, concealed attacks
the distribution of the errors, the frequency of unacceptable such as replay attack and false data injection attack must be
errors, as well as whether there are bounding values. In another identified before the soft sensing outputs are controlled by
aspect, it is favorable for the soft sensors to give quantita- illegal parties. Therefore, in order to ensure credibility, soft

Authorized licensed use limited to: Indian Institute of Technology. Downloaded on May 26,2024 at 02:45:55 UTC from IEEE Xplore. Restrictions apply.
12872 IEEE SENSORS JOURNAL, VOL. 21, NO. 11, JUNE 1, 2021

sensors, when using data from external sources, need to verify oriented to specific practical applications. Fig. 3 summarizes
the data source, examine the data integrity and confidentiality the common data preprocessing tasks into three requirements,
(when applicable). namely, for security check, for scaling, and the compensation
Other Factors: During the design of soft sensors, one must of dataset defects. Among them, the security check is empha-
also consider the overheads relating to the computational load, sized due to the trend of using soft sensors for cyber-physical
the memory footprint, the storage space, as well as the required systems. The trustworthiness of the data from external sources,
communication bandwidth. e.g., the third-party devices, must be verified. Besides, in case
that network-based communication is compromised or exposed
B. General Procedure to cyber-attacks, the integrity of data should be examined at the
Since the construction of a soft sensor for practical use data-processing stage to prevent potential misinterpretation.
involves many engineering problems, most of the innovative Furthermore, in the case of confidential data, encryption and
research articles only focus on one or a few parts, rather than decryption techniques are also required.
reporting the complete design procedure in detail. It is also The difference of validation and evaluation lies in the
worth to note the decisive role of the design targets (technical testing items: validation only focuses on the primary design
specifications from the clients, based on practical demands) in specifications while evaluation gives the full report of the
the selection of appropriate approaches for modelling. multi-dimensional performance of the soft sensor, such as
In this part, a generic and complete procedure for soft sensor those introduced in the previous subsection. Based on the
construction is introduced. As shown in Fig. 2 (Construction evaluation results, optimization techniques can be introduced
procedure for a soft sensor) , the construction of a soft sensor to improve (e.g., by fine-tuning) the model parameters, and
includes the stages of data/information collection (information the whole design process forms a close loop until the design
source), design and implementation. specifications can be met.
Information Source: The availability and the quality of the The Stage of Implementation: This stage aims to transform
four categories of information sources constitute the basis of the algorithmic approaches obtained from the design stage into
the subsequent tasks. First of all, there must be a pool of electronics-based solutions, in the form of either executable
historical process data, especially the measurements from other software algorithms (software applications) or dedicated hard-
available hardware sensors. When a subset is considered useful ware implementation. Prototype development and testing are
for soft sensing, the online measurements from the correspond- required for batch applications, such as in distributed power
ing sensors will be used to drive the soft sensor. Associated grids and automotive applications. Furthermore, deployment
with the process data, the metadata describes the dataset (e.g., and commissioning are required to ensure that the soft sensor
the data property) and clarifies how they can be interpreted. is working as expected and the communication channels are
At the design stage, the system/process knowledge, e.g., in the properly established, before using the outputs for monitoring
forms of physical principles and chemical reaction equations, and control purposes.
will greatly contribute to the tasks of variable selection and In the future, when cyber-physical systems become the most
data preprocessing. Furthermore, the operational experience popular industrial practice, people will also see a trend that the
from the field operators and maintenance experts is an addi- stakeholders of the industrial processes tend to contract the soft
tional source of information, which may be valuable for all the sensor construction projects to standalone companies and seek
subsequent steps. As suggested in [16], “Without any expert for third-party technical supports such as dedicated hardware
help or physical insight, a soft sensor design can become an production and software development, as well as consultation
unaffordable task and data can be only partially exploited. The services, rather than build an expert team by themselves. In
task requires the cooperation of soft sensor designer and plant turn, this will facilitate the process of commercialization and
experts, in the form of meetings and interviews.” creating added values of the related techniques.
The Stage of Design: This stage is dedicated to transforming
the available data and knowledge into an applicable tool, IV. C OMMON P ROBLEMS AND A DVANCEMENT IN
i.e., the soft sensor, in the form of algorithmic approaches. S OLUTIONS
In recent years, although some fruitful research outputs are This section first reviews the common problems in soft
reported in the literature that solve the specific sub-tasks (as sensor design. Since many of the solutions suggested in pre-
in the top block of Fig. 2), little change is observed in terms vious review articles reflect the situation a decade ago, in this
of the overall routine for soft sensor design. In fact, the one part we put special focus on presenting the latest techniques
presented here is very much aligned with those reviewed that lead to further advancement to problems encountered at
recently ( [11]), five years ago ( [17]), and even a decade that time. Additionally, emerging technologies and industrial
ago ( [4]). For this reason, in this article, we only highlight demands have led to new challenges and promoted new
some major differences and novel challenges, whereas the research directions.
basic concepts of the widely-recognized steps are not repeated.
Moreover, “Variable selection” and “modelling” are not pre-
sented here, as they will be discussed in Sections IV-B and A. Quality of Data
IV-C, respectively. As have been discussed in Section III-B, data collection
Data preprocessing covers a variety of tasks, corresponding is the basis of the data-driven methods. The quality of the
to different dataset properties. In other words, it is somewhat historical data has a direct influence on the performance of

Authorized licensed use limited to: Indian Institute of Technology. Downloaded on May 26,2024 at 02:45:55 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: REVIEW ON SOFT SENSORS FOR MONITORING, CONTROL, AND OPTIMIZATION OF INDUSTRIAL PROCESSES 12873

the soft sensors. Data collection in the industrial processes clusters. Density-based approaches estimate the distribution
sometimes suffers from problems like sampling time, missing and identify outliers that locate at the low-density regions.
data, outliers, and so forth. As a result, the data are not directly Examples include the Parzen window, C-means and the
applicable to soft sensor modelling. Gaussian mixture model [26]. The Parzen window approach is
1) Sampling Rate: In industrial systems, it is common that easy to implement whilst sensitive to the choice of the initial
heterogeneous types of variables have different sampling rates. cell volume. When applying the C-means approach, the cluster
Some physical quantities can be measured by highly sensitive number need to be defined based on the knowledge of the
devices, thus having high sample rates. Some critical variables experts. The Gaussian mixture model approach is able to deal
are difficult to measure, thus having low sampling rates. with non-Gaussian data. However, its performance deteriorates
To deal with this, the traditional approach is resampling, in case of very high dimensional data. As for the proximity-
including down-sampling and up-sampling. based approaches, the similarity in data with their neighbors
The former may lose some of the useful process informa- is considered. Popular solutions include k-nearest neighbor
tion while the latter may lead to poor model performance [27], local outlier factor [28] and angle-based outlier factor
due to the interpolation of the sparsely available quality [29]. The K-nearest neighbor approach removes a proportion
data. An alternative is to model the multi-rate data by the of observations of large distances, thus losing the information
finite impulse response model [18] and use the output to therein. The performance of the local outlier factor approach
approximate the unmeasurable variables. This approach avoids is heavily dependent on the number of nearest neighbors, and
excluding the samples. More recently, data fusion techniques the angle-based outlier factor approach is not suitable for large
are adopted. In [19], an extension of the modified track-to- dataset.
track fusion approach is proposed. The state estimations of two Although there are many available solutions, there are some
independent Kalman filters are fused optimally, one of which limitations as well. First, strong assumptions to the dataset
deals with the slow sampling rate and the varying delay. make them poor in the generalization capability. Second, most
2) Missing Data: This happens when no value is stored solutions are time-consuming due to the need to calculate
for one or more variables in an observation. Under the metrics between all data. This is especially notable in the case
condition that the proportion of the missing values is small, of high dimensional data. Third, the results still have to be
the time instances containing missing data can be removed validated by process experts to reduce false detection.
with listwise or pairwise deletion approaches. Alternatively,
the missing values can be imputed employing mean substitu- B. Variable and Feature Selection
tion, hot-deck substitution, regression substitution, conditional There is a colossal number of variables in modern large-
distribution-based substitution and multiple imputations [2]. scale industrial processes. Some variables are highly correlated
It is worth to note that the more recently developed multiple with the target variables while the others are useless for
imputation approach has better performance than single impu- soft sensor construction. An appropriate selection procedure
tation in most cases. The most studied multiple imputation will reduce the model development effort, simplify the model
approach is based on probabilistic principal analysis [20]. structure, and improve the soft sensing performance. With the
Nevertheless, it has a heavy computational load. aid of system experts and some reliable automatic tools, two
Two other approaches employed to deal with missing data aspects are usually considered: determine a criterion to assess
are expectation maximization [21] and maximum likelihood the most useful feature subset; employ a search strategy for a
[22], which assume the missing data are subject to a cer- suitable variable subset. Three different approaches that try to
tain distribution. Additionally, an autoencoder is employed to achieve this are briefly reviewed below.
reconstruct the missing data in [23]. Based on this, several 1) Ranking-Based Approaches: Ranking-based approaches
modern approaches have been developed, such as extreme calculate the relevance of each variable to the target variable
learning machine autoencoder [24] and supervised variational using, e.g., the correlation coefficient or mutual information
autoencoders [25]. The limitation of using the autoencoders is [30]. Then, a subset of variables is selected by comparing the
in the size of the auto-associative neural networks. scores to a threshold. In [31], an efficient feature selection
3) Outliers: Outliers refer to the measurements that deviate algorithm is proposed based on normalized mutual informa-
from the typical and the meaningful range of the variables. tion. However, the estimation of multidimensional probability
Many popular outlier detection approaches are based on the density function leads to heavy computational load. Most
statistical properties of the historical data. The 3σ rule is one recently, improvement is achieved by the integration of Tabu
of the simplest solutions, which assumes the variables subject Search into the approach. Falling into a local minimum is thus
to normal distribution. However, its performance is unsatisfac- avoided [32].
tory in the case of multivariate outliers. The correlation among 2) Wrapper Approaches: Wrapper approaches make selec-
variables needs to be considered by multivariate approaches. tions according to the generalization performance of the mod-
Multivariate approaches can be categorized into the els that are trained using different subsets of variables. The
distance-based, the density-based and the proximity-based key is to reduce the computational burden (unacceptably large
ones [2]. Distance-based approaches treat the samples that number of subsets) with a search procedure. Sequential feature
are far from the center of data distribution as outliers using, selection and the genetic algorithm are the potential choices
for instance, Mahalanobis metric. However, the Mahalanobis in this respect [33]. It is interesting to notice that, recently,
distance-based approach can only be used to describe Gaussian efforts are made in combining the wrapper approaches and

Authorized licensed use limited to: Indian Institute of Technology. Downloaded on May 26,2024 at 02:45:55 UTC from IEEE Xplore. Restrictions apply.
12874 IEEE SENSORS JOURNAL, VOL. 21, NO. 11, JUNE 1, 2021

TABLE II
C OMPARISON OF S INGLE AND M ULTIPLE I MPUTATION

the ranking-based approaches. Good performance in specific The strategies to deal with variations in the operational
applications has been reported in [34], [35]. The drawback lies point and the external environmental factors can be grouped
in that each time a subset of variables needs to be evaluated, into two. Pre-determined switching laws based on pattern
a regression model has to be constructed. recognition have been extensively studied. It is simple and
3) Embedded Approaches: Embedded approaches incorpo- effective when the variations to be considered are finite (thus
rate feature selection as a part of the training process. A typical the switching regions are finite) [48]. Otherwise, online adap-
class of algorithms that follow this idea is the regularization- tive strategies tend to be more robust in the case of continuous
based ones with a penalty term [36]. In [37], regularization is (infinite) and unexpected variations.
applied to the neural networks. A penalty function is defined to In the following, soft sensing model construction will be
weaken the influence of the useless variables. Another solution categorized according to different basic assumptions.
first uses all the inputs to train the model. Then, the irrelevant
1) Multivariate Statistical Analysis: The basic assumption is
variables are removed sequentially according to the result of a
the existence of some consistent statistical features, such as
sensitivity analysis [38]. Similar to the wrapper approaches,
constant variance, covariance and some latent structures. Many
the model is retained each time the variables need to be
basic approaches have been reviewed in [11], [17], [49]. For
evaluated.
instance, a robust improvement of PLS was proposed in [50],
In addition to the above, unsupervised approaches have
[51]. In [52], recursive PLS model was fused with wavelet
received increasing attention in recent years [39]–[41]. Com-
coefficient matrices to represent different frequency. Recently,
parisons are made in [42], [43] on the partial least squares
an open-source toolbox was published [42] where many tools
(PLS) related approaches.
are available for basic multivariate statistical analysis.

C. Model Selection and Model Construction 2) Data-Driven and Identification Based Observers: The basic
assumption is the existence of the system states. Model-based
There are various categories of approaches for soft sen-
observers have been well-studied in modern control theory,
sor modelling. Each class is based on different assumptions
where system matrices are known (calculated using accurate
and applies to specific system types. Fig. 4 shows a three-
models based on physical principles. For complex industrial
dimensional space that can cover a wide range of soft sensors,
processes, this is sometimes infeasible due to the need for
designed during the past twenty years.
comprehensive knowledge. Black-box system identification
According to linearity, the model type can be categorized
provides an alternative. System matrices are solved offline
into linear models, piecewise linear models and nonlinear
using dedicatedly designed excitation signals and the corre-
models (e.g. nonlinear autoregressive with exogenous input
sponding outputs and will remain unchanged at the online
model, neural networks, Bayesian network, fuzzy model and
stage [53]. In contrast, data-driven observers rely only on
wavelet model) [44]. Generally, linear models are considered
process data and are more easily extended for online update
first. If the performance is unsatisfactory, piecewise linear
and adaptation [54], [55].
models shall be tried before seeking for solutions by the
nonlinear ones. Based on the Monte Carlo study, it was shown 3) Machine Learning: The basic assumption is that the
in [45] that reliable criterion for model selection includes the outputs are uniquely determined by the nonlinear projection of
Bayesian information criteria and the Geweke-Meese criteria. the inputs, or the dynamics of the systems mimic some natural
Moreover, nonlinear function estimation is another strategy to behavior. In most scenarios, the quality and the quantity of the
quantify the nonlinearity degree in the process variables [46]. training dataset governs the modelling performance.
In another aspect, the selection of the soft sensor model In recent years, intelligent algorithms, especially the deep
should also be appropriate to the system dynamics. In some learning techniques, have boosted the development of soft
industrial processes, the variables of interest show a static rela- sensors. To capture the abstract features, deep structures based
tionship with the available measurement variables. However, on the convolutional neural network, deep brief network and
the dynamic characteristics must be considered when there autoencoders have been extensively studied [56], [57]. For
is a non-negligible correlation in time (time series) and in time series data, long short-term memory networks and its
states (state-dependent) [47]. That is one of the major issues to modifications (such as gated recurrent unit) are the most
be considered during model selection, between, for instance, popular tools [58], [59]. As for the traditional machine learning
the multivariate statistical analysis-based model construction approaches (e.g., support vector regression), for soft sensor
approaches and the observer/filter-based approaches. applications, they are usually integrated with the optimization

Authorized licensed use limited to: Indian Institute of Technology. Downloaded on May 26,2024 at 02:45:55 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: REVIEW ON SOFT SENSORS FOR MONITORING, CONTROL, AND OPTIMIZATION OF INDUSTRIAL PROCESSES 12875

procedures, such as random forest and generic algorithm [60], A. Application Scenarios
[61]. 1) Virtual Redundancy of Physical Sensors: While the out-
4) Filtering (Kalman Filter): The condition for the adoption of puts of the soft sensors can be used for the monitoring, control
Kalman filter and extended Kalman filter is that the knowledge and optimization of the whole process, soft sensor itself can be
about certain stochastic properties of measurements and noises regarded as a nominal system or a fault-free digital twin of a
are available. As optimal state estimators, they are usually real sensing device (when exists). In this context, soft sensors
derived from the models constructed based on the practical are also called observers, filters, state estimators, or predictors
process to deal with noises [62], [63]. In [18], the Kalman in the disciplines of control and fault diagnosis. By examining
filter was adopted as the core of a data fusion algorithm to deal the difference in the outputs of the physical entities and their
with multi-rate sampling problem. More recently, an unscented virtual counterparts, the healthy status of the physical sensors
Kalman filter was proposed for fast-sampled measurements including well-functioning, material ageing, software errors
while a modified form (delay-dependent variable step-ahead and hardware failure can be determined with the aid of fault
prediction) was used for slow-sampled measurements [64]. detection, fault classification and fault identification techniques
5) First Principle Models: The condition that this type of [8], [16], [54]. Through this capability, periodic maintenance
models is feasible is to have sufficient knowledge about the can be cancelled or reduced, which can be consequential
underlying principles of physics, chemistry and even biology. for the case of remotely and distributed installed sensors.
The construction activities must be case-oriented. In complex Therefore, the soft sensors can be applied in the following
scenarios, finite element analysis is usually necessary. Typical scenarios:
first principle models are based on the principle of Conser-
i) Online diagnosis of the hardware-based sensors.
vation of Energy [8], [64], the principle of material balance
ii) Temporary backups for physical sensors during routine
[65], as well as kinematics and system dynamics [66]. In most
maintenance.
soft sensor publications, first principle models are rarely used
iii) Digital twin of a physical sensor, characterizing the full
directly, but rather are usually integrated with other data-driven
life cycle and covering the stages of design, manufactur-
models or calibration procedures [67].
ing, commissioning, deployment, operation, maintenance,
and disposal.
D. Soft Sensor Maintenance
2) Estimator for Unmeasurable Physical Quantities: Each
The performance of soft sensors degrades over time when physical sensing device has a designed range, precision and
drifting and variations occur in the industrial processes. There sampling rate for measurements. For instance, the clinical ther-
is limited information contained in the historical data (at least mometer can usually give readouts between 35 ◦ C and 42 ◦ C.
with a limited period of time) that can be used to construct However, there are practical demands for many industrial
soft sensors. Considering this, the online operational data processes to operate in unusual conditions; such as extremely
and the contextual data that can characterize the evolution high pressure, extremely high temperature or extremely low
of the processes are the necessary sources of information. temperature. It is also not rare to see some processes working
Unlike physical sensors which usually go through scheduled under a wide range of and frequent variations. In another
calibration, frequent examination and periodic replacement, aspect, the large-scale systems like dams and power grids
there still is no consensus on how to maintain soft sensors must also depend on mathematical tools to compensate for
for long-term functioning. the unreachable order of magnitude of the physical sensors.
Continuous adaptation methods can avoid the problems As a result, soft sensors become the only possibility to gain
caused by fixed models. To achieve this, sample selection process knowledge and carry out online monitoring and control
can be applied to select the most relevant data samples in these cases.
while sample weighting contributes to changing the degree In a drilling process, the primary task is to perceive the
of importance of the samples dynamically according to, for downhole circumstances and detect, in a timely manner,
instance, the time of measurement. These strategies can easily whether a part of the mechanical structure, most importantly
be combined with the modelling methods based on, e.g., the drill, is broken. However, high pressure and high temper-
principal component analysis and artificial neural networks ature pose severe challenges to the use of downhole sensors.
[68]. Typical solutions proposed recently include just-in-time In today’s engineering practice, the flow rate of the volume
learning [69], [70], iterative learning [71], incremental learning of the drilling fluid that enters and returns from the wellbore
[72], and ensemble learning [52], [73]–[76]. Adaptive data- is used as an alternative for monitoring and control purposes.
driven soft sensing approaches based on various machine Due to the existence of the mixed cuttings and gas squeezed
learning techniques have been compared in [68]. in the drilling fluid, the performance of the physical flow
meters can be degraded. To deal with this, soft sensors for
V. S TATE - OF - THE -A RT IN S OFT S ENSOR A PPLICATIONS the flow rate estimation are designed based on the upstream
In this section, a categorization of the soft sensor application level data collected by ultrasonic equipment the and the
scenarios is firstly presented. It is shown that the applications corresponding reference data [77]. The necessity of soft sensor
fall into at least 10 different purposes. After that, the appli- application due to the similar reason is also demonstrated
cations to the practical areas, as well as several real-world in the blast furnace iron making process [74], shaft furnace
deployment cases are introduced. roasting process [10], etc.

Authorized licensed use limited to: Indian Institute of Technology. Downloaded on May 26,2024 at 02:45:55 UTC from IEEE Xplore. Restrictions apply.
12876 IEEE SENSORS JOURNAL, VOL. 21, NO. 11, JUNE 1, 2021

In the waste water treatment processes, the ammonia con- ix) Give early reports before time-consuming laboratory
centration is a key variable for the control and evaluation of the analysis (for early diagnosis and predictive maintenance).
removal performance of the harmful substances. While ammo- x) Monitor online/in real-time the quality and quantity of
nia concentration can be determined by some commercial-off- the final products and by-products.
the-shelf devices such as in-situ ion selective electrodes and
ex-situ analyzers, these high-end yet error-prone instruments B. Application Areas
require frequent maintenance by skilled staff. A more econom- Table III summarizes the soft sensor applications in indus-
ical alternative was reported in [78]. The authors explained trial processes into several common areas. These examples are
the possibility of using cheaper sensors (pH and oxidation- based on the articles published within the past decade (2010–
reduction potential) to construct soft sensors to achieve the 2020), and can therefore reflect the latest research interests and
same goal. More recently, the reliability of the soft sensors the urgent practical needs. Corresponding to each industrial
is improved in a way that compensate for the drift of the pH process, the specific physical quantities of sensing are listed.
sensors: in [79], qualitative features (the minima and maxima The symbols O2 , CO, CO2 , H2 S, and SO2 are chemical
of the pH difference signal) are suggested to be included in symbols while the others denote surrogate symbols that are
the design. defined in the related references.
In summary, the soft sensors play a significant role in the A major focus of study is on the chemical industry. To a
following scenarios: great extent, this is due to the lack of applicable mechanism
models in the chemical reactions—although some processes
iv) In extreme operation condition that exceeds the limit of
can be described by the chemical reaction equations, there
the physical sensors.
are always side reactions accompanying the main reaction.
v) When implementation of some physical sensors is eco-
Many of the side reactions are highly-complicated or even
nomically unacceptable or undesirable.
unknown. Furthermore, many chemical reactions are sensitive
Furthermore, it is also easy to see the value of soft sensors in to the variation of external environmental factors, such as
the following cases: temperature, pressure, and the availability of the catalysts.
As a result, solving the chemical equilibrium online in real-
vi) Infrequent and/or delayed measurement by hardware
time is usually infeasible. On the other hand, the correlation
sensing devices ( [64]).
among the reaction processes and among the process vari-
vii) Requirement for non-invasive measurement.
ables (including the reactants, the feed rates, the flow rates,
viii) There is no space or load for the installation of the
temperature, pressure, the pH values, etc.) provides a solid
physical sensors (e.g., spacecraft).
theoretical foundation to the development of the soft sensors.
3) Product Quality Prediction: The concept of soft sensor In other words, the basic assumption on the dependency of
also covers the prediction of abstract features as well as the variables can be nicely satisfied, which is required by the
qualitative or quantitative assessment indices in the industrial multivariate analysis models, the regression models, and some
processes, for instance, the key performance indicators that (reduced-order) observers.
characterize the final product quality. In wine production Among the soft sensor applications to the chemical industry,
process, historical data and expert knowledge are required the most popular ones are the Tennessee Eastman process
to adjust the timing and the quantity of the feeding of the (TEP) and the debutanizer distillation column. TEP is an open
components. However, this is an open-loop process. The final and challenging chemical model simulation platform devel-
quality of the red wine cannot be learnt until the wine tasters oped by the Eastman Chemical Company in the United States.
give professional comments. Nevertheless, there are actually It was designed to reflect the characteristics of the real-world
early hints and signs of potential quality degradation from chemical reaction process, and is now widely recognized as a
the large amounts of process variables at different stages [80]. benchmark in the areas of process control and fault diagnosis
Following this assumption, soft sensors can be designed to give of complex industrial processes [93]. As shown in Table III,
early predictions, using the multivariate statistical tools or the soft sensors were designed for the online estimation of differ-
machine learning techniques. An interesting attempt was made ent components in several TEP units. The extensively studied
in [81] where four-dimensional gradings made by the wine debutanizer distillation column was initially presented in [94]
tasters were collected, namely, the presentation 15 points, fra- and soon afterwards included in the book [16]. It is a part
grance 30 points, mouth feel 44 points, and subjective feeling of a desulfuring and naphtha splitter plant that is located
11 points, such that the overall wine quality is 100 points. at the Mediterranean Oil Refinery Company (ERG Raffineria
Based on these evaluations and the corresponding process data, Mediterranea S.R.L.) in Syracuse, Italy. In order to maximize
the soft sensors are trained as an alternative to human profes- the stabilized gasoline content in the liquefied petroleum gas
sionals, and beyond this, to give frequent online evaluations. splitter feed and to minimize the butane content in the naphtha
Soft sensors for other key performance indicators were stud- splitter feed, soft sensors are required to provide real-time
ied in [50], [65], [70], [82], [83] (chemical production), [10], estimates of the butane concentration in gasoline and the
[52], [79], [84]–[86] (chemical processing), [47], [52], [70], gasoline concentration in butane [94].
[87]–[89] (petroleum refining), [67], [90], [91] (mechanical In addition to the application to the chemical processes,
industry), [92] (drug industry), etc. In these examples, soft there are also a number of practical needs in the mechanical
sensors: sector. The most notable difference from the chemical industry

Authorized licensed use limited to: Indian Institute of Technology. Downloaded on May 26,2024 at 02:45:55 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: REVIEW ON SOFT SENSORS FOR MONITORING, CONTROL, AND OPTIMIZATION OF INDUSTRIAL PROCESSES 12877

TABLE III
A PPLICATION E XAMPLES OF S OFT S ENSORS IN I NDUSTRIAL P ROCESSES

lies in the feature of the physical quantities of sensing: in Case 1: (2007, data-driven) A soft sensing scheme was
most cases, soft sensors are constructed to meet the design proposed in [9] where an automated stepwise linear regression
requirement of high precision, which cannot (presently) be approach was used to find the relevant predictor variables. The
achieved by the physical sensors. These physical quantities soft sensor was adopted by the largest alumina refinery in
include the volumetric flow, the strip thickness, the particle Europe, RAAL (RUSAL Aughinish Alumina Ltd., located in
size, etc. southern Ireland). It replaced the previous statistical forecast-
In the past decade, a few applications have been reported ing tool of RAAL, and was deployed to predict the strength
in the domains of nuclear industry, smart buildings, drug of alumina crystal conglomerates—a critical alumina quality
production, and food industry. However, it should be noted parameter in the Bayer process, and to forecast deviations off
that the application domains (sub-domains) and the value of the normal operating conditions.
the soft sensing techniques are more than what have been Case 2: (2010, model & data integrated) In [10], the authors
summarized here. Beyond the scope of this study, soft sensors proposed to combine the fuzzy mechanism model with a
have good potentials in medical care systems, automotive neural network compensator to online estimate the magnetic
systems, and robotic systems. tube recovery ratio in the mineral processing. Data clustering
techniques and a fast training algorithm were employed. The
soft sensing approach was adopted by a metal company in
C. Real-World Deployment Cases Lanzhou, China, and plays a supervisory role as a reference
Most of the existing research activities have been conducted value to monitor the shaft furnace roasting process.
as performance evaluation of soft sensors in laboratory setups Case 3: (2017, model-driven) In [91], the P80 particle size
(based on simulations) or with limited amounts of experimen- in a horizontally stirred mill (an industrial IsaMill) in Western
tal data acquired from real plants. Only a few have been tested Australia was estimated online. The proposed soft sensor is
and verified on real systems and real plants. In this part, several based on the random forest model. However, according to
real-world deployment cases are introduced. the observation over two months, it was pointed out that

Authorized licensed use limited to: Indian Institute of Technology. Downloaded on May 26,2024 at 02:45:55 UTC from IEEE Xplore. Restrictions apply.
12878 IEEE SENSORS JOURNAL, VOL. 21, NO. 11, JUNE 1, 2021

recalibration is required to compensate for the frequent process purposes. Compared with measurements from the sensors,
drift in the mineral processing operations, owing to large the challenge of adopting soft sensing outputs lies in the
variations in the feed. strategy to ensure reliability in case of faults in the dependent
Case 4: (2018, model & data integrated) The authors of variables, and to cut off the fault propagation path.
[8] exploited the physical relationship inside the building • Exploiting the potential of the deep learning techniques
ventilation units and constructed soft sensors based on simple and other intelligent approaches: Deep learning is powerful
linear and nonlinear regression models. The soft sensors are in modelling highly complex nonlinear processes [12], [61].
used to estimate the temperature, air flow, fan speed, and to However, several challenges still hinder its application to
diagnose faults in the heater energy meter. The developed soft sensor design, for instance, the interpretability of the
soft sensors were tested on a real building for teaching—the features extracted from data and the relationship with the
Odense undervisning building 44 (OU44) — at the University outputs, the sensitivity to the network hyper-parameters, and
of Southern Denmark, campus Odense, built in 2015. the influence from the size of available training dataset.

VI. O PEN C HALLENGES AND F UTURE D IRECTIONS VII. C ONCLUSION

Based on the above discussions, this section summarizes the This review focuses on the key scientific problems in the
open challenges to overcome and the future directions that are design and implementation of soft sensors in the context of
promising in the coming decade. the modern industry. Based on an exhaustive investigation of
• Improving the autonomous capability (decision-making the literature, it is dedicated to associating the studies reported
in design and self-validation): At present, several steps at the over the past twenty years, and giving answers to the research
design and implementation stages rely on the assistance of questions proposed in Section I-B. While these questions are
human experts, such as model selection and feature selection. discussed whenever relevant, they are addressed mainly in
The performance of soft sensors has better guaranteed with the corresponding sections as follows: Q1 (III-B); Q2 (III-
the aid of the experts’ knowledge and experience. However, A, IV); Q3 (V); Q4 (IV, V-A); Q5 (Advancement: Discussed
this can be optimized in the future, to reduce the dependency throughout the paper; Challenges and future directions: VI).
on the designer skills. Although this direction has not been In the coming decade and beyond, apart from the efforts
approached systematically and extensively today, it is inspiring to overcome the aforementioned challenges, it is expected
to notice that there are some attempts reported. In [95], to witness more novel ideas, more advanced techniques and
a probabilistic self-validating approach was proposed where broader platforms where soft sensing plays an indispensable
unreliable data are identified and corrected. To deal with time- role in the monitoring, control and optimization of the indus-
varying environment and operating conditions, authors in [96] trial processes, not to mention the use in Digital Twins. The
proposed an online variable prediction approach where models future of the soft sensing techniques is destined to be inspiring
are added or removed adaptively for improved ensemble and exciting, no less that today looking back at the past.
performance; authors in [97] proposed to conduct automated
validation to improve the accuracy and reliability of soft R EFERENCES
sensors, by integrating just-in-time models and relevant vector [1] A. W. Colombo, S. Karnouskos, O. Kaynak, Y. Shi, and S. Yin,
“Industrial cyberphysical systems: A backbone of the fourth industrial
machine. revolution,” IEEE Ind. Electron. Mag., vol. 11, no. 1, pp. 6–16,
• Filling the gap between the laboratory outcome and Mar. 2017.
the industrial practice: Due to the desire for good perfor- [2] J. Zhu, Z. Ge, Z. Song, and F. Gao, “Review and big data perspectives
on robust data mining approaches for industrial process modeling with
mance, algorithms developed for soft sensor design are usually outliers and missing data,” Annu. Rev. Control, vol. 46, pp. 107–133,
complicated. However, a compromise in performance against 2018.
computational efficiency is found to be unfavorable to meet [3] S. Yin and O. Kaynak, “Big data for modern industry: Challenges and
trends [Point of View],” Proc. IEEE, vol. 103, no. 2, pp. 143–146,
the practical needs. As discussed in [13], although proven Feb. 2015.
to be very effective in the academic domain, there are only [4] P. Kadlec, B. Gabrys, and S. Strandt, “Data-driven soft sensors in
few practical applications of evolutionary algorithms in the the process industry,” Comput. Chem. Eng., vol. 33, pp. 759–814,
Apr. 2009.
field of food and beverage production due to the heavy [5] D. Ciuonzo, G. Romano, and P. S. Rossi, “Channel-aware decision
computational burden, especially the requirement of massive fusion in distributed MIMO wireless sensor networks: Decode-and-
parallel implementation. In addition to the difficulties in online fuse vs. decode-then-fuse,” IEEE Trans. Wireless Commun., vol. 11,
no. 8, pp. 2976–2985, Aug. 2012.
implementation, there are quite few practical cases reported [6] Q. Cheng, P. K. Varshney, J. H. Michels, and C. M. Belcastro,
as deployments in real industrial plants. As discussed in “Distributed fault detection with correlated decision fusion,” IEEE
the previous section, most works are limited to simulations Trans. Aerosp. Electron. Syst., vol. 45, no. 4, pp. 1448–1465, Oct. 2009.
[7] D. Ciuonzo and P. Salvo Rossi, “Distributed detection of a non-
and laboratory-scale experiments. To improve this, further in- cooperative target via generalized locally-optimum approaches,” Inf.
depth collaboration between the academia and the industry is Fusion, vol. 36, pp. 261–274, Jul. 2017.
required. [8] C. G. Mattera, J. Quevedo, T. Escobet, H. R. Shaker, and M. Jradi,
“A method for fault detection and diagnostics in ventilation units using
• Embedded in the integrated framework of control and virtual sensors,” Sensors, vol. 18, no. 3931, pp. 1–21, 2018.
optimization: Soft sensing is driven by process data online and [9] V. Cregan, W. T. Lee, and L. Clune, “A soft sensor for the Bayer
in turn provides key information about the system. It is mean- process,” J. Math. Ind., vol. 7, no. 1, pp. 1–6, Dec. 2017.
[10] F. Wu and T. Chai, “Soft sensing method for magnetic tube recovery
ingful to expand the role if these estimations and predictions ratio via fuzzy systems and neural networks,” Neurocomputing, vol. 73,
can be used in the close-loop for control and optimization nos. 13–15, pp. 2489–2497, Aug. 2010.

Authorized licensed use limited to: Indian Institute of Technology. Downloaded on May 26,2024 at 02:45:55 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: REVIEW ON SOFT SENSORS FOR MONITORING, CONTROL, AND OPTIMIZATION OF INDUSTRIAL PROCESSES 12879

[11] C. Abeykoon, “Design and applications of soft sensors in polymer [37] T. Similä and J. Tikka, “Combined input variable selection and model
processing: A review,” IEEE Sensors J., vol. 19, no. 8, pp. 2801–2813, complexity control for nonlinear regression,” Pattern Recognit. Lett.,
Apr. 2019. vol. 30, no. 3, pp. 231–236, Feb. 2009.
[12] S. Graziani and M. G. Xibilia, Development and Analysis of Deep [38] I.-C. Yeh and W.-L. Cheng, “First and second order sensitivity analy-
Learning Architectures (Deep Learning for Soft Sensor Design). Cham, sis of MLP,” Neurocomputing, vol. 73, nos. 10–12, pp. 2225–2233,
Switzerland: Springer, ch. 2, 2020. Jun. 2010.
[13] S. Birle, M. A. Hussein, and T. Becker, “Fuzzy logic control and soft [39] S. Salcedo-Sanz, L. Cornejo-Bueno, L. Prieto, D. Paredes, and
sensing applications in food and beverage processes,” Food Control, R. García-Herrera, “Feature selection in machine learning prediction
vol. 29, no. 1, pp. 254–269, Jan. 2013. systems for renewable energy applications,” Renew. Sustain. Energy
[14] J. Randek and C.-F. Mandenius, “On-line soft sensing in upstream Rev., vol. 90, pp. 728–741, Jul. 2018.
bioprocessing,” Crit. Rev. Biotechnol., vol. 38, no. 1, pp. 106–121, [40] Y. Jiang, S. Yin, and O. Kaynak, “Data-driven monitoring and safety
Jan. 2018, doi: 10.1080/07388551.2017.1312271. control of industrial cyber-physical systems: Basics and beyond,” IEEE
[15] R. Dahiya et al., “Large-area soft e-skin: The challenges beyond sensor Access, vol. 6, pp. 47374–47384, 2018.
designs,” Proc. IEEE, vol. 107, no. 10, pp. 2016–2033, Oct. 2019. [41] P. Zhu, W. Zuo, L. Zhang, Q. Hu, and S. C. K. Shiu, “Unsupervised
[16] L. Fortuna, S. Graziani, A. Rizzo, and M. G. Xibilia, Soft Sensors feature selection by regularized self-representation,” Pattern Recognit.,
for Monitoring and Control of Industrial Processes. London, U.K.: vol. 48, no. 2, pp. 438–446, Feb. 2015.
Springer-Verlag, 2007. [42] Y. Jiang and S. Yin, “Recent advances in key-performance-indicator
[17] F. Souza, R. Araujo, and J. Mendes, “Review of soft sensor methods oriented prognosis and diagnosis with a MATLAB toolbox: DB-KIT,”
for regression applications,” Chemom. Intell. Lab. Syst., vol. 152, IEEE Trans. Ind. Informat., vol. 15, no. 5, pp. 2849–2858, May 2019.
pp. 69–79, Mar. 2016. [43] Z. X. Wang, Q. P. He, and J. Wang, “Comparison of variable selection
[18] L. Xie, H. Yang, and B. Huang, “FIR model identification of multirate methods for PLS-based soft sensor modeling,” J. Process Control,
processes with random delays using EM algorithm,” AIChE J., vol. 59, vol. 26, pp. 56–72, Feb. 2015.
no. 11, pp. 4124–4132, Nov. 2013. [44] Y. Gu and H.-L. Wei, “A robust model structure selection method
[19] A. Fatehi and B. Huang, “Kalman filtering approach to multi-rate for small sample size and multiple datasets problems,” Inf. Sci.,
information fusion in the presence of irregular sampling rate and vols. 451–452, pp. 195–209, Jul. 2018.
variable measurement delay,” J. Process Control, vol. 53, pp. 15–25, [45] P. N. Sharma et al., “PLS-based model selection: The role of alternative
May 2017. explanations in information systems research,” J. Assoc. for Inf. Syst.,
[20] S. Dray and J. Josse, “Principal component analysis with missing vol. 20, no. 4, pp. 346–397, 2019.
values: A comparative survey of methods,” Plant Ecol., vol. 216, no. 5, [46] Y. Jiang, S. Yin, and O. Kaynak, “Optimized design of parity
pp. 657–667, May 2015. relation based residual generator for fault detection: Data-driven
[21] C. K. Enders, “A primer on maximum likelihood algorithms available approaches,” IEEE Trans. Ind. Informat., early access, Apr. 16, 2020,
for use with missing data,” Structural Equation Model., A Multidisci- doi: 10.1109/TII.2020.2987840.
plinary J., vol. 8, no. 1, pp. 128–141, Jan. 2001. [47] B. Bidar, J. Sadeghi, F. Shahraki, and M. M. Khalilipour, “Data-
[22] J. M. Jerez et al., “Missing data imputation using statistical and driven soft sensor approach for online quality prediction using state
machine learning methods in a real breast cancer problem,” Artif. Intell. dependent parameter models,” Chemometric Intell. Lab. Syst., vol. 162,
Med., vol. 50, no. 2, pp. 105–115, Oct. 2010. pp. 130–141, Mar. 2017.
[23] V. Miranda, J. Krstulovic, H. Keko, C. Moreira, and J. Pereira, [48] F. A. A. Souza and R. Araújo, “Mixture of partial least squares experts
“Reconstructing missing data in state estimation with autoencoders,” and application in prediction settings with multiple operating modes,”
IEEE Trans. Power Syst., vol. 27, no. 2, pp. 604–611, May 2012. Chemometric Intell. Lab. Syst., vol. 130, pp. 192–202, Jan. 2014.
[24] C.-B. Lu and Y. Mei, “An imputation method for missing data based [49] S. Yin, S. X. Ding, X. Xie, and H. Luo, “A review on basic data-
on an extreme learning machine auto-encoder,” IEEE Access, vol. 6, driven approaches for industrial process monitoring,” IEEE Trans. Ind.
pp. 52930–52935, 2018. Electron., vol. 61, no. 11, pp. 6414–6428, Jan. 2014.
[25] R. Xie, N. M. Jan, K. Hao, L. Chen, and B. Huang, “Supervised [50] S. Yin, G. Wang, and X. Yang, “Robust PLS approach for KPI-related
variational autoencoders for soft sensor modeling with missing data,” prediction and diagnosis against outliers and missing data,” Int. J. Syst.
IEEE Trans. Ind. Informat., vol. 16, no. 4, pp. 2820–2828, Apr. 2020. Sci., vol. 45, no. 7, pp. 1375–1382, Jul. 2014.
[26] V. Hodge and J. Austin, “A survey of outlier detection methodologies,” [51] S. Yin and G. Wang, “A modified partial robust m-regression to
Artif. Intell. Rev., vol. 22, no. 2, pp. 85–126, Oct. 2004. improve prediction performance for data with outliers,” in Proc. IEEE
[27] M.-L. Zhang and Z.-H. Zhou, “ML-KNN: A lazy learning approach to Int. Symp. Ind. Electron., May 2013, pp. 1–6.
multi-label learning,” Pattern Recognit., vol. 40, no. 7, pp. 2038–2048, [52] D. V. Poerio and S. D. Brown, “A frequency-localized recursive partial
Jul. 2007. least squares ensemble for soft sensing,” J. Chemometrics, vol. 32,
[28] M. M. Breunig, H.-P. Kriegel, R. T. Ng, and J. Sander, “LOF: no. 5, p. e2999, 2018.
Identifying density-based local outliers,” in Proc. ACM SIGMOD, 2000, [53] H. Li, Y. Gao, P. Shi, and H.-K. Lam, “Observer-based fault detection
pp. 93–104. for nonlinear systems with sensor fault and limited communication
[29] H.-P. Kriegel, M. Schubert, and A. Zimek, “Angle-based outlier capacity,” IEEE Trans. Autom. Control, vol. 61, no. 9, pp. 2745–2751,
detection in high-dimensional data,” in Proc. ACM SIGKDD, 2008, Sep. 2016.
pp. 444–452. [54] S. X. Ding, “Data-driven design of monitoring and diagnosis sys-
[30] I. Guyon and A. Elisseeff, “An introduction to variable and feature tems for dynamic processes: A review of subspace technique based
selection,” J. Mach. Learn. Res., vol. 3, pp. 1157–1182, Jan. 2003. schemes and some recent results,” J. Process Control, vol. 24, no. 2,
[31] P. A. Estevez, M. Tesmer, C. A. Perez, and J. M. Zurada, “Normalized pp. 431–449, Feb. 2014.
mutual information feature selection,” IEEE Trans. Neural Netw., [55] S. Yin, X. Yang, and H. R. Karimi, “Data-driven adaptive observer for
vol. 20, no. 2, pp. 189–201, Feb. 2009. fault diagnosis,” Math. Problems Eng., vol. 2012, pp. 1–21, Aug. 2012.
[32] K. Sun, P. Tian, H. Qi, F. Ma, and G. Yang, “An improved normalized [56] W. Yan, D. Tang, and Y. Lin, “A data-driven soft sensor modeling
mutual information variable selection algorithm for neural network- method based on deep learning and its application,” IEEE Trans. Ind.
based soft sensors,” Sensors, vol. 19, no. 24, p. 5368, Dec. 2019. Electron., vol. 64, no. 5, pp. 4237–4245, May 2017.
[33] G. Chandrashekar and F. Sahin, “A survey on feature selection meth- [57] X. Wu, J. Chen, L. Xie, L. L. T. Chan, and C.-I. Chen, “Development
ods,” Comput. Electr. Eng., vol. 40, no. 1, pp. 16–28, Jan. 2014. of convolutional neural network based Gaussian process regression
[34] S. Solorio-Fernández, J. A. Carrasco-Ochoa, and to construct a novel probabilistic virtual metrology in multi-stage
J. F. Martínez-Trinidad, “A new hybrid filter–wrapper feature selection semiconductor processes,” Control Eng. Pract., vol. 96, Mar. 2020,
method for clustering based on ranking,” Neurocomputing, vol. 214, Art. no. 104262.
pp. 866–880, Nov. 2016. [58] C.-S. Hsu and J.-R. Jiang, “Remaining useful life estimation
[35] S. Huda, J. Abawajy, M. Alazab, M. Abdollalihian, R. Islam, and using long short-term memory deep learning,” in Proc. IEEE Int.
J. Yearwood, “Hybrids of support vector machine wrapper and filter Conf. Appl. Syst. Invention (ICASI), Apr. 2018, pp. 58–61, doi:
based framework for malware detection,” Future Gener. Comput. Syst., 10.1109/ICASI.2018.8394326.
vol. 55, pp. 376–390, Feb. 2016. [59] J. Chen, H. Jing, Y. Chang, and Q. Liu, “Gated recurrent unit based
[36] H. Zou and T. Hastie, “Regularization and variable selection via the recurrent neural network for remaining useful life prediction of nonlin-
elastic net,” J. Roy. Stat. Soc. B, Stat. Methodol., vol. 67, no. 2, ear deterioration process,” Rel. Eng. Syst. Saf., vol. 185, pp. 372–382,
pp. 301–320, Apr. 2005. May 2019.

Authorized licensed use limited to: Indian Institute of Technology. Downloaded on May 26,2024 at 02:45:55 UTC from IEEE Xplore. Restrictions apply.
12880 IEEE SENSORS JOURNAL, VOL. 21, NO. 11, JUNE 1, 2021

[60] W. Yan, H. Shao, and X. Wang, “Soft sensing modeling based on [82] L. Bao, X. Yuan, and Z. Ge, “Co-training partial least squares model
support vector machine and Bayesian model selection,” Comput. Chem. for semi-supervised soft sensor development,” Chemometric Intell. Lab.
Eng., vol. 28, no. 8, pp. 1489–1498, Jul. 2004. Syst., vol. 147, pp. 75–85, Oct. 2015.
[61] M. Vallejo, C. de la Espriella, J. Gómez-Santamaría, [83] J. Zhu, Z. Ge, and Z. Song, “Robust semi-supervised mixture
A. F. Ramírez-Barrera, and E. Delgado-Trejos, “Soft metrology probabilistic principal component regression model development and
based on machine learning: A review,” Meas. Sci. Tech., vol. 3, no. 3, application to soft sensors,” J. Process Control, vol. 32, pp. 25–37,
pp. 1–16, 2019. Aug. 2015.
[62] A. Gryzlov, W. Schiferli, and R. F. Mudde, “Soft-sensors: Model-based [84] B. Bidar, F. Shahraki, J. Sadeghi, and M. M. Khalilipour, “Soft
estimation of inflow in horizontal wells using the extended Kalman sensor modeling based on multi-state-dependent parameter models and
filter,” Flow Meas. Instrum., vol. 34, pp. 91–104, Dec. 2013. application for quality monitoring in industrial sulfur recovery process,”
[63] A. M. Nair, A. Fanta, F. A. Haugen, and H. Ratnaweera, “Implementing IEEE Sensors J., vol. 18, no. 11, pp. 4583–4591, Jun. 2018.
an extended Kalman filter for estimating nutrient composition in a [85] M. Krueger, H. Luo, S. X. Ding, S. Dominic, and S. Yin, “Data-
sequential batch MBBR pilot plant,” Water Sci. Technol., vol. 80, no. 2, driven approach of KPI monitoring and prediction with application
pp. 317–328, Jul. 2019. to wastewater treatment process,” IFAC-PapersOnLine, vol. 48, no. 21,
[64] X. Yang, Y. Zhang, Y. A. W. Shardt, X. Li, J. Cui, and C. Tong, pp. 627–632, 2015.
“A KPI-based soft sensor development approach incorporating infre- [86] L. Yiqi, H. Daoping, and L. Zhifu, “A SEVA soft sensor method
quent, variable time delayed measurements,” IEEE Trans. Control based on self-calibration model and uncertainty description algorithm,”
Syst. Technol., vol. 28, no. 6, pp. 2523–2531, Nov. 2020, doi: Chemometric Intell. Lab. Syst., vol. 126, pp. 38–49, Jul. 2013.
10.1109/TCST.2019.2929478. [87] W. Shao, X. Tian, P. Wang, X. Deng, and S. Chen, “Online soft sensor
[65] Z. Ge, F. Gao, and Z. Song, “Mixture probabilistic PCR model for design using local partial least squares models with adaptive process
soft sensing of multimode processes,” Chemometric Intell. Lab. Syst., state partition,” Chemometric Intell. Lab. Syst., vol. 144, pp. 108–121,
vol. 105, no. 1, pp. 91–105, Jan. 2011. May 2015.
[66] I. S. Choi, J. A. Rossiter, and P. J. Fleming, “Looper and tension [88] Z. Ge, B. Huang, and Z. Song, “Nonlinear semisupervised principal
control in hot rolling mills: A survey,” J. Process Control, vol. 17, component regression for soft sensor modeling and its mixture form,”
no. 6, pp. 509–521, Jul. 2007. J. Chemometrics, vol. 28, pp. 793–804, Nov. 2014.
[67] S. X. Ding, S. Yin, K. Peng, H. Hao, and B. Shen, “A novel scheme for [89] Z. Ge, B. Huang, and Z. Song, “Mixture semisupervised principal
key performance indicator prediction and diagnosis with application to component regression model and soft sensor application,” AIChE J.,
an industrial hot strip mill,” IEEE Trans. Ind. Informat., vol. 9, no. 4, vol. 60, no. 2, pp. 533–545, Feb. 2014.
pp. 2239–2247, Nov. 2013. [90] Z. Ge, Z. Song, L. Zhao, and F. Gao, “Two-level PLS model for quality
[68] P. Kadlec, R. Grbic, and B. Gabrys, “Review of adaptation mechanisms prediction of multiphase batch processes,” Chemometric Intell. Lab.
for data-driven soft sensors,” Comput. Chem. Eng., vol. 35, no. 1, Syst., vol. 130, pp. 29–36, Jan. 2014.
pp. 1–24, Jan. 2011. [91] L. F. A. Napier and C. Aldrich, “An IsaMill soft sensor based on ran-
[69] Y. Liu and J. Chen, “Integrated soft sensor using just-in-time support dom forests and principal component analysis,” IFAC-PapersOnLine,
vector regression and probabilistic analysis for quality prediction of vol. 50, no. 1, pp. 1175–1180, Jul. 2017.
multi-grade processes,” J. Process Control, vol. 23, no. 6, pp. 793–804, [92] J. Ji, H. Wang, K. Chen, Y. Liu, N. Zhang, and J. Yan, “Recursive
Jul. 2013. weighted kernel regression for semi-supervised soft-sensing modeling
of fed-batch processes,” J. Taiwan Inst. Chem. Engineers, vol. 43, no. 1,
[70] Z. Ge and Z. Song, “A comparative study of just-in-time-learning based
pp. 67–76, Jan. 2012.
methods for online soft sensor modeling,” Chemometric Intell. Lab.
[93] J. J. Downs and E. F. Vogel, “A plant-wide industrial process control
Syst., vol. 104, no. 2, pp. 306–317, Dec. 2010.
problem,” Comput. Chem. Eng., vol. 17, no. 3, pp. 245–255, Mar. 1993.
[71] Y. Jiang and S. Yin, “Recursive total principle component regression
[94] L. Fortuna, S. Graziani, and M. G. Xibilia, “Soft sensors for product
based fault detection and its application to vehicular cyber-physical
quality monitoring in debutanizer distillation columns,” Control Eng.
systems,” IEEE Trans. Ind. Informat., vol. 14, no. 4, pp. 1415–1423,
Pract., vol. 13, no. 4, pp. 499–508, Apr. 2005.
Apr. 2018.
[95] Y. Liu, J. Chen, Z. Sun, Y. Li, and D. Huang, “A probabilistic
[72] S. Yin, X. Xie, J. Lam, K. C. Cheung, and H. Gao, “An improved self-validating soft-sensor with application to wastewater treatment,”
incremental learning approach for KPI prognosis of dynamic fuel
Comput. Chem. Eng., vol. 71, pp. 263–280, Dec. 2014.
cell system,” IEEE Trans. Cybern., vol. 46, no. 12, pp. 3135–3144,
[96] S. G. Soares and R. Araújo, “An adaptive ensemble of on-line extreme
Dec. 2016, doi: 10.1109/TCYB.2015.2498194.
learning machines with variable forgetting factor for dynamic system
[73] Q. Sun and Z. Ge, “Deep learning for industrial KPI prediction: prediction,” Neurocomputing, vol. 171, pp. 693–707, Jan. 2016.
When ensemble learning meets semi-supervised data,” IEEE Trans. Ind. [97] Y. Liu, “Adaptive just-in-time and relevant vector machine based soft-
Informat., early access, Jan. 27, 2020, doi: 10.1109/TII.2020.2969709. sensors with adaptive differential evolution algorithms for parameter
[74] X. Yuan, J. Zhou, Y. Wang, and C. Yang, “Multi-similarity measure- optimization,” Chem. Eng. Sci., vol. 172, pp. 571–584, Nov. 2017.
ment driven ensemble just-in-time learning for soft sensing of industrial [98] E. Frauendorfer and W.-D. Hergeth, “Soft sensor applications in indus-
processes,” J. Chemometrics, vol. 32, Sep. 2018, Art. no. e3040. trial vinylacetate-ethylene (VAE) polymerization processes,” Macro-
[75] Z. Ge and Z. Song, “Ensemble independent component regression molecular Reaction Eng., vol. 11, Aug. 2017, Art. no. 1700008.
models and soft sensing application,” Chemometric Intell. Lab. Syst., [99] W. Shao, Z. Ge, Z. Song, and K. Wang, “Nonlinear industrial soft
vol. 130, pp. 115–122, Jan. 2014. sensor development based on semi-supervised probabilistic mixture of
[76] P. Kadlec and B. Gabrys, “Local learning-based adaptive soft sensor for extreme learning machines,” Control Eng. Pract., vol. 91, Oct. 2019,
catalyst activation prediction,” AIChE J., vol. 57, no. 5, pp. 1288–1301, Art. no. 104098.
May 2011. [100] Y. Jiang, H. Yu, J. Yin, and C. Yang, “Study on KPI-related subspace
[77] K. Chhantyal, M. H. Jondahl, H. Viumdal, and S. Mylvaganam, decomposition for fault detection and robust KPI prediction against
“Upstream ultrasonic level based soft sensing of volumetric flow of abnormal data,” in Proc. IEEE 25th Int. Symp. Ind. Electron. (ISIE),
non-newtonian fluids in open venturi channels,” IEEE Sensors J., Jun. 2016, pp. 100–105.
vol. 18, no. 12, pp. 5002–5013, Jun. 2018. [101] Z. Ge, “Mixture Bayesian regularization of PCR model and soft sensing
[78] M. V. Ruano, J. Ribes, A. Seco, and J. Ferrer, “Low cost-sensors as application,” IEEE Trans. Ind. Electron., vol. 62, no. 7, pp. 4336–4343,
a real alternative to on-line nitrogen analysers in continuous systems,” Jul. 2015.
Water Sci. Technol., vol. 60, no. 12, pp. 3261–3268, Dec. 2009. [102] A. Rogina, I. Šiško, I. Mohler, Ž. Ujevic, and N. Bolf, “Soft sensor
[79] C. M. Thürlimann, D. J. Dürrenmatt, and K. Villez, “Soft-sensing for continuous product quality estimation (in crude distillation unit),”
with qualitative trend analysis for wastewater treatment plant control,” Chem. Eng. Res. Design, vol. 89, no. 10, pp. 2070–2077, Oct. 2011.
Control Eng. Pract., vol. 70, pp. 121–133, Jan. 2018. [103] A. Rani, V. Singh, and J. R. P. Gupta, “Development of soft sensor
[80] S. Yin, L. Liu, and J. Hou, “A multivariate statistical combination fore- for neural network based control of distillation column,” ISA Trans.,
casting method for product quality evaluation,” Inf. Sci., vols. 355–356, vol. 52, no. 3, pp. 438–449, May 2013.
pp. 229–236, Aug. 2016. [104] J. Ma, H. Jiang, K. Huang, Z. Bi, and K. L. Man, “Novel field-support
[81] S. Yin, X. Zhu, and H. R. Karimi, “Quality evaluation based on vector regression-based soft sensor for accurate estimation of solar
multivariate statistical methods,” Math. Problems Eng., vol. 2013, irradiance,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 64, no. 12,
pp. 1–10, Dec. 2013. pp. 3181–3191, Dec. 2017.

Authorized licensed use limited to: Indian Institute of Technology. Downloaded on May 26,2024 at 02:45:55 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: REVIEW ON SOFT SENSORS FOR MONITORING, CONTROL, AND OPTIMIZATION OF INDUSTRIAL PROCESSES 12881

[105] D. H. Lim, S. H. Lee, and M. G. Na, “Smart soft-sensing for the Jingwei Dong received the B.E. degree in
feedwater flowrate at PWRs using a GMDH algorithm,” IEEE Trans. automation and the M.E. degree in control sci-
Nucl. Sci., vol. 57, no. 1, pp. 340–347, Feb. 2010. ence and engineering from the Harbin Institute
[106] M. K. Al-Sharman, Y. Zweiri, M. A. K. Jaradat, R. Al-Husari, D. Gan, of Technology, Harbin, China, in 2016 and 2018,
and L. D. Seneviratne, “Deep-learning-based neural network training respectively. He is currently pursuing the Ph.D.
for state estimation enhancement: Application to attitude estimation,” degree with the Delft Center for Systems and
IEEE Trans. Instrum. Meas., vol. 69, no. 1, pp. 24–34, Jan. 2020. Control, Delft University of Technology, Delft, The
[107] H. Yang, Y. Chen, Y. Sun, and L. Hao, “A novel pneumatic soft sensor Netherlands.
for measuring contact force and curvature of a soft gripper,” Sens. His research interests include fault diagnosis,
Actuators A, Phys., vol. 266, pp. 318–327, Oct. 2017. security of large-scale systems, and optimization
[108] S. Han, T. Kim, D. Kim, Y.-L. Park, and S. Jo, “Use of deep learning methods.
for characterization of microfluidic soft sensors,” IEEE Robot. Autom.
Lett., vol. 3, no. 2, pp. 873–880, Apr. 2018.

Yuchen Jiang (Student Member, IEEE) received

the B.E. (Hons.) degree in automation from the
Harbin Institute of Technology, Harbin, China,
in 2016, where he is currently pursuing the Ph.D.
degree with the Department of Control Science
and Engineering.
From 2019 to 2020, he was a Visiting
Student Researcher with Technische Universität Okyay Kaynak (Life Fellow, IEEE) received the
München, Lehrstuhl für Steuerungs- und B.Sc. (Hons.) and Ph.D. degrees in electronic
Regelungstechnik, Munich, Germany. His and electrical engineering from the University
research interests include data-driven process of Birmingham, Birmingham, U.K., in 1969 and
monitoring, fault diagnosis and prognosis, industrial cyber-physical 1972, respectively.
systems, and artificial intelligence. From 1972 to 1979, he held various positions
within the industry. In 1979, he joined the Depart-
ment of Electrical and Electronics Engineering,
Bogazici University, Istanbul, Turkey, where he
Shen Yin (Senior Member, IEEE) received the is currently a Professor Emeritus, holding the
B.E. degree in automation from the Harbin Insti- UNESCO Chair on Mechatronics and also a
tute of Technology, Harbin, China, in 2004, and 1000 Talents Program Professor with the University of Science and
the M.Sc. degree in control and information sys- Technology Beijing, China. He has held long-term (near to or more than a
tem and the Ph.D. degree in electrical engineer- year) visiting professor/scholar positions at various institutions in Japan,
ing and information technology from the Univer- Germany, U.S.A., Singapore, and China. His current research interests
sity of Duisburg-Essen, Germany, in 2007 and include the fields of intelligent control and CPS. He has authored three
2012, respectively. books and edited five and authored or coauthored more than 400 articles
He is currently a Professor with the School that have appeared in various journals, books, and conference proceed-
of Astronautics, Harbin Institute of Technology. ings.
His research interests include data-driven con- Dr. Kaynak has been on many committees of the IEEE and was the
trol and diagnosis, machine learning and artificial intelligence, safety, President of IEEE Industrial Electronics Society from 2002 to 2003. He
reliability and security of complicated systems, CPS, and autonomous has recently received the Chinese Government’s Friendship Award and
systems. the Humboldt Research Prize.

Authorized licensed use limited to: Indian Institute of Technology. Downloaded on May 26,2024 at 02:45:55 UTC from IEEE Xplore. Restrictions apply.