Article

Cite This: Ind. Eng. Chem. Res. 2019, 58, 14180−14192 pubs.acs.org/IECR

Oscillation Detection in Process Industries by a Machine Learning-

Based Approach
Jônathan W. V. Dambros,*,†,‡ Jorge O. Trierweiler,† Marcelo Farenzena,† and Marius Kloft‡,§
†
Department of Chemical Engineering, Federal University of Rio Grande do Sul, R. Eng. Luiz Englert, s/n, Campus Central, Porto
Alegre, Rio Grande do Sul, Brazil
‡
Department of Computer Science, University of Kaiserslautern, 67663 Kaiserslautern, Germany
§
Department of Computer Science, University of Southern California, 1002 Childs Way, Los Angeles, California, United States
*
S Supporting Information
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.
Downloaded via UNIV GADJAH MADA on February 20, 2021 at 02:19:40 (UTC).

ABSTRACT: Oscillatory control loop is a frequent problem in process industries. Its incidence highly degrades the plant
profitability, which means oscillation detection and removal is fundamental. For detection, many automatic techniques have
been proposed. These are usually based on rules compiled into an algorithm. For industrial application, in which the time series
have very distinct properties and are subject to interferences such as noise and disturbances, the algorithm must include rules
covering all possible time series structures. Since the development of this algorithm is near impractical, it is reasonable to say
that current rule-based techniques are subject to incorrect detection. This work presents a machine learning-based approach for
automatic oscillation detection in process industries. Rather than being rule-based, the technique learns the features of
oscillatory and nonoscillatory loops by examples. A model based on deep feedforward network is trained with artificial data for
oscillation detection. Additionally, two other models are trained for the quantification of the number of periods and oscillation
amplitude. The evaluation of the technique using industrial data with different features reveals its robustness.

1. INTRODUCTION threshold, a counter is increased; if the counter exceeds a given

Oscillation in process industries is a common problem that value, the presence of oscillation is confirmed. Miao and Seborg8
affects between 30% and 41% of the control loops.1−3 The developed a method that evaluates the decay ratio in the
removal of such oscillatory loops is of great interest since a autocorrelation function. If the decay is higher than a certain
decrease in variability means that the process variables are held threshold, oscillation is detected. Thornhill et al.9 proposed a
closer to their desired conditions, resulting in financial benefits. method in which the regularity of the period of oscillation in the
The first step in oscillation removal is its detection. This task autocorrelation function is evaluated. If the period is regular,
can be driven by individual visual inspection of each time series. oscillation is detected.
Unfortunately, this approach is unfeasible when a full-plant These three methods are a brief overview of more than 30
diagnosis is required. Typically, process industries have between different approaches (some of which will be presented in the
500 and 5000 control loops.4 Visual inspection would consume following section). The techniques are usually data-driven
the full personal resources and limit the investigation to part of methods based on rules founded on criteria similar to those used
these loops,5 resulting in unnoticed oscillations.6 To overcome in visual inspection and/or mathematical concepts. These rules
this limitation, automatic oscillation detection techniques are are usually the computation of a parameter (IAE, decay ratio,
required.
Over the last 25 years, researchers have been working on Received: March 15, 2019
automatic methods. The technique proposed by Hägglund7 Revised: June 24, 2019
computes the integral absolute error (IAE) for each segment Accepted: July 5, 2019
limited by zero crossings. If the IAE value is larger than a certain Published: July 5, 2019

© 2019 American Chemical Society 14180 DOI: 10.1021/acs.iecr.9b01456

Ind. Eng. Chem. Res. 2019, 58, 14180−14192
Industrial & Engineering Chemistry Research Article

regularity) that is later evaluated by an if/else statement. These oscillation detection (STSOD) techniques, in which the focus is
approaches may work adequately for well-behaved oscillations. the detection in individual loops/variables, and plant-wide
Unfortunately, industrial data are usually corrupted by noise and oscillation detection (PWOD) techniques, in which the
disturbance, the frequency and amplitude of oscillation may be methods apply the detection directly to a set of loops/variables.
irregular, the oscillation may be intermittent, and multiple Among the PWOD techniques, the method proposed by
oscillations may be present in the same time series. Despite this, Thornhill et al.17,18 is highlighted. The technique decomposes
the incoming data present a large frequency and amplitude the spectrum of the set in basis, each corresponding to a
range. For a rule-based method to work correctly, all of these dominant oscillation. Then, each spectrum is compared to each
influences must be incorporated into the algorithm, making it basis for the grouping of time series with similar features.
complex and extensive. STSOD techniques are subclassified into time domain, ACF-
Machine learning (ML) techniques and applications have based, frequency domain, continuous wavelet transform
been strongly explored over the last two decades. ML applied to (CWD), and decomposition methods. Time domain methods
computer vision, speech recognition, and robot control, for are usually simple. In this group is included the method
example, have been broadly explored by several research groups proposed by Hägglund,7 reviewed in section 1. Also is included
worldwide; however, many areas remain underexplored10 and the technique proposed by Forsman and Stattin,19 which
therefore are excellent research and innovation opportunities. evaluates both the IAE magnitude and the interval between zero
In the field of oscillation detection and diagnosis, few works crossings. If both values are kept close to constant over time,
based on ML have been published. This includes, for example, oscillation is detected. Zakharov and co-workers20,21 designed
the works proposed by Zabiri et al.,11 Venceslau et al.,12 and an algorithm that evaluates the correlation between two periods
Farenzena and Trierweiler13 for stiction quantification, in which of oscillation. If the correlation is high, oscillation is detected.
stiction is a phenomenon that causes oscillation. In the big field Since noise is a frequent problem for time domain methods, it
of fault detection, ML has been extensively exploited in the last is convenient to transform the time series into the ACF domain,
20 years. The number of publications has been increasing faster where the resulting signal has the same frequency of oscillation
with the larger amount of data and greater computation power. with attenuated noise. This strategy forms the second group
An overview and discussion on these methods are found in the named ACF-based, which includes the works by Miao and
recent works by Ge et al. and Wuest et al.14,15 Seborg8 and Thornhill et al.9
ML techniques, by definition, do not require rules. Instead, an In most cases, the oscillatory time series can easily be detected
ML model adapts itself according to given examples to improve visually by a peak in the f requency domain. The technique
its performance to a specific task. Applied to oscillation proposed by Zhang et al.22 tries to capture this peak by setting a
detection, this means that specific rules for each of the influences threshold. If the magnitude of any component in the frequency
are not required, making the approach simpler and more robust. domain crosses the threshold, oscillation is detected.
This work presents a new technique for oscillation detection Matsuo and co-workers23−25 presented a series of works in
based on ML. Here, data with distinct features are generated which continuous wavelet transforms (CWT) are applied.
artificially. A deep feedforward network (DFN) is trained using Oscillation is detected by the identification of high values in
the magnitude in the frequency domain of the data as input and the wavelet plot by visual inspection. This approach is similar to
the labels nonoscillatory, oscillatory with regular and nonregular the visual detection in the frequency domain with the advantage
oscillation as output. Finally, the trained network is applied for of time information on the oscillation occurrence.
industrial data for oscillation detection. Empirical mode decomposition (EMD) is a technique that
Furthermore, because the number of oscillatory loops in an decomposes the time series into components of different
industrial plant is usually high, it is also important to rank them frequencies. Srinivasan et al.26 applied a modified EMD
based on the strength of the oscillation to isolate loops that technique for later evaluation of the oscillation presence in
require immediate maintenance from those that do not influence each component. In a different approach, Li et al.27 used the
plant performance. In this work, the quantification of the discrete cosine transform (DCT) to isolate different compo-
number of periods and oscillation amplitude is performed by a nents in the frequency domain. Each component is transformed
second and third network. Finally, this work also aims to back to the time domain, and its regularity is evaluated for
introduce the application of machine learning on control loop oscillation detection. These methods are included in the
performance monitoring (CPM), showing its potential and decomposition subgroup, in which the oscillatory components
influencing new works in this area. are isolated for further analysis by a simpler approach (mostly by
Following the introduction, this work presents a short review the regularity index proposed by Thornhill et al.9). Many other
of oscillation detection and machine learning in section 2. The recent methods are included in this subgroup.27−32
proposed oscillation detection technique is introduced in As seen, these methods are also based on the evaluation of a
section 3 and evaluated for artificial and real industrial data in property followed by an if/else statement. Since they are based
section 4. Finally, the conclusions are presented in section 5. on rules, they occasionally return unreliable results due to the
presence of features not considered during their implementation
2. REVIEW OF OSCILLATION DETECTION AND but present in industrial data.
MACHINE LEARNING According to Thornhill et al.9 and Karra et al.,33 a good
This section presents a brief review of oscillation detection oscillation detection method has the following features:
methods followed by a short elucidation on machine learning • Evaluates the period and magnitude of the oscillations;
and, more specifically, deep learning. • Requires only time series data;
2.1. Oscillation Detection. In addition to the oscillation
• is robustness to noise and disturbances;
detection methods described in the introduction, many others
have been published. According to Dambros et al.,16 the • is able to handle multiple and intermittent oscillations;
techniques are classified into two groups: single time series • is completely automatic.
14181 DOI: 10.1021/acs.iecr.9b01456
Ind. Eng. Chem. Res. 2019, 58, 14180−14192
Industrial & Engineering Chemistry Research Article

Figure 1. A generic deep feedforward network.

Moreover, according to Bauer et al.,34 simple methods are The architecture in Figure 1 is called feedforward methods
preferred by control engineers. because there are no feedback connections where one or more
The literature on oscillation detection is extensive, in which outputs of the network are fed back to itself. Here, information
more than 30 works have been published. In this section, a few flows from the input x, through the defined mapping f to the
works on each STSOD subgroup are reviewed. For further output y. The represented feedforward network has M inputs,
reading, the recent and complete review by Dambros et al.16 is predicts Mout outputs, and has D hidden layers, each one with Md
recommended. For a condensed review, the works by Thornhill neurons, where d is a specific layer, z(d)
j is evaluated according to
and Horch35 and Duan et al.36 are suggested. the following equation:
2.2. Machine Learning. From the statistical perspective,
machine learning is an approach in which a given model is z(j d) = h(d)(a(j d)) (1)
adapted by examples to perform a particular task. The type of
task a machine learning technique performs is divided, broadly, where j is a specific neuron in layer d, h(d) is the activation
into two groups: supervised and unsupervised learning. function, and a(d)
j is equal to

Supervised learning is the group in which the model adapts Md − 1

itself to a known output, while for unsupervised learning, no a(j d) = ∑ θji(d)zi(d − 1) + θj(0d)
output is given, leaving the model to find the best structure by i=1 (2)
itself. Since for oscillation detection, the task is the classification (d − 1)
of a time series in two known classes (oscillatory or where θ(d)
ji are the weights. For d = 1, Md−1 = M and zi = xi.38
nonoscillatory), supervised learning is the natural approach. The weights θ are the adaptable parameters in the feedforward
Formally, supervised learning involves observing several network. These parameters are optimized to minimize a
specified loss function E(θ). A common loss function is the
examples of an input vector x and an associated output value
sum-of-squares error function, where
or vector y, then learning to predict y from x.37 After training, the
model is used to predict the unknown y corresponding to a new 1
N
2
x. Among the supervised machine learning techniques, deep E (θ ) = ∑ f (xn , θ ) − yn
learning, supporting vector machines, and decision trees are 2 n=1 (3)
broadly explored.
Deep learning models are mathematical structures mimicking During the optimization, each round in which all the training
examples are evaluated is called an epoch. Usually, the weights
the function of a biological brain. Figure 1 shows a generic
are updated once at the end of each epoch. However, especially
representation of a deep feedforward network (DFN), a simple
for long data sets, in which the evaluation of all the examples
and frequently applied deep learning architecture. consumes considerable computational effort, the weights are
The goal of a feedforward network is the approximation of a updated more than once in each epoch. The update occurs after
highly nonlinear function f * such that y = f *(x) maps an input x a certain number of examples (batch) have been evaluated. The
to an output y. The network defines a mapping y = f(x; θ) and optimization is then run for a given number of epochs or until
learns the value of the parameters θ that result in the best the loss function stops decreasing.
approximation. Usually, a feedforward network is the The initial weight values are also of great importance since
composition of many different functions. For example, f(x) = they influence the optimization convergence and speed and can
f(3)(f(2)(f(1)(x))) consists of three functions, where f(1) is the significantly reduce the vanishing/exploding gradients prob-
computation of the first layer of the network, f(2) is the second lem.39 Many approaches for their initialization are available.
layer, and so on. As seen, the feedforward network requires a large number of
The last layer of the network is the output layer, and its output specified parameters: activation functions, number of layers and
values are required to be approximately equal to y for the neurons, optimization algorithm, batch size, and weight
training points x. The behavior of the other layers, named initialization approach, for example. These parameters are
hidden layers, is not directly specified by the training data, and it defined before the training and are called hyperparameters. The
is the task of the algorithm to adapt these layers to approximate selection of the ideal hyperparameter is a fundamental and time-
f *.37 consuming task.
14182 DOI: 10.1021/acs.iecr.9b01456
Ind. Eng. Chem. Res. 2019, 58, 14180−14192
Industrial & Engineering Chemistry Research Article

One common approach for the selection, called grid search, • If available, the amount may not be sufficient to train a
selects potential hyperparameters and then tests for all possible model with good generalization;
combinations of hyperparameters. This approach is sometimes • If the amount is suitable, labeling is required. In addition
unfeasible due to the high number of combinations. In this case, to being a time-consuming task, labeling by visual
the random search approach is applied, in which random sets are inspection is subject to user interpretation about what is
tested. and what is not an oscillatory loop;
The means the DFN uses to correlate the input to the output • Assuming that the data are correctly labeled, it does not
data is usually unknown or hardly noticeable. This is a weak guarantee that the data cover a sufficiently broad range of
point on using the DFN or most of the other ML approaches. processes and parameters.
This section is not long enough to present all the aspects of To avoid these problems, artificial data are generated for
deep learning and DFNs in detail. For the interested reader, we training and validation of the network. Later, the network
recommend the book by Goodfellow et al.37 The term deep performance is tested not only on the artificial data but also on
learning adopted in this work can interchangeably be the industrial data.
denominated artificial neural networks. The first was chosen Two main rules are strictly followed for data generation:
since it has been preferred in recent works.
(1). Artificial data must be as similar as possible to the
3. PROPOSED OSCILLATION DETECTION AND industrial data;
QUANTIFICATION TECHNIQUE (2) Artificial data must have examples from processes with
different dynamics, configurations, and characteristics.
The proposed oscillation detection technique follows the
desired features presented in section 2.1. The technique is Following these two rules, the generated data have the
applied individually to each time series, and thus it is classified as following features:
a single time series oscillation detection (STSOD) technique. • oscillatory and nonoscillatory examples of different
To maintain the simplicity and guarantee the reproducibility of lengths,
the technique, an already implemented machine learning library • noise and disturbance with different amplitudes,
is used and recommended. In this work, the Keras40 library is • oscillatory time series with sinusoidal, triangular, or
used with the TensorFlow41 backend. square waveform with different numbers of periods of
This section is divided into four subsections. First, the oscillation,
procedure for data generation is presented, which is followed by • waveform smoothed with different intensities to approx-
a simple data processing approach. Subsections 3 and 4 show the imate an oscillatory time series filtered by the process, and
description of the oscillation detection and quantification • part of the oscillatory time series with variable frequency
approach, respectively. The overview of the technique is with different intensities.
presented in Figure 2. Finally, the distributions of the variables used to generate the
time series are presented in Table 1.

Table 1. Distributions of the Variables for the Artificial Data

Generation
variable distribution parameters
Parameters for All Time Series
signal length (points) linear-exponential 200, 2000, 2000
noise variance exponential 0.1
disturbance amplitude exponential 1
oscillation probability Bernoulli 0.6
Parameters for Oscillatory Time Series Only
number of periods linear-exponential 2, 20, 20
waveform uniform
smoothing factor exponential 0.2
variable freq prob. Bernoulli 0.5
Parameters for Oscillatory Time Series with Variable Frequency Only
frequency change factor linear-exponential 0.5, 0.25, 0.5

In Table 1, noise variance is the variance of a random vector

Figure 2. Overview of the proposed technique. with a Gaussian distribution and zero mean. Disturbance
amplitude is the difference between the maximum and minimum
value of a random vector with a Gaussian distribution and zero
3.1. Data Generation. In an ideal scenario, industrial mean smoothed by the following transfer function:
labeled data (and not artificial data) are used to train the model
since the final application uses industrial data. This approach is 1
D(s) =
unlikely to be feasible due to the following reasons: 100s 2 + 10s + 1 (4)
• Industrial data are hardly available due to confidential and The oscillatory time series are smoothed by the following
strategic issues; equation:
14183 DOI: 10.1021/acs.iecr.9b01456
Ind. Eng. Chem. Res. 2019, 58, 14180−14192
Industrial & Engineering Chemistry Research Article

Figure 3. Distinct examples of generated time series.

Figure 4. Distribution of the variables for the generated time series.

1 on the distribution are required. First, the distribution cannot

S( s ) =
SF*s + 1 (5) start at zero; second, it is preferred to have a better resolution in
the middle region because extreme values are infrequent in a real
where SF is the smoothing factor. Finally, half of the oscillatory application. Once no distribution with these features is found, a
time series has a variable frequency, in which the frequency new distribution is presented in this work. The distribution,
change factor is the difference between the highest and lowest named linear-exponential distribution, is the combination of a
frequency divided by its mean frequency. If a nonregular
linear followed by an exponential distribution and is described in
oscillation has a mean frequency equal to 0.1 and a frequency
Appendix A.
change factor equal to 0.5, the lowest and highest frequency
Finally, 106 time series are generated to train, 105 to validate,
would be about 0.075 and 0.125, for example.
and 105 to test the DFN model. Here, the validation data set is
Also in Table 1, four different distributions are presented. The
used to select the optimum set of hyperparameters, while the test
uniform distribution represents the distributions where all the
data set is used to evaluate the final model. All the oscillatory
elements have the same probability of being selected. The
time series have oscillation amplitude equal to 1, which, in this
Bernoulli distribution is the discrete distribution where the given
work, is the distance between the maximum and minimum value
parameter is the probability of the variable taking the value 1
in the magnitude.
(true). The exponential distribution is represented by the
Nine distinct examples with the following features are

i 1y
following equation:

f jjjjx ; zzzz = e−x/ β

presented in Figure 3: (1) high noise level, (2) high disturbance
level, (3) sinusoidal waveform, (4) triangular waveform, (5)
k {
1
β β square waveform, (6) high-frequency oscillation, (7) low-
(6)
frequency oscillation, (8) smoothed square waveform, and (9)
This distribution requires only one parameter (β), which is square waveform with variable frequency. Figure 3 shows the
the value of x where the probability decreases by a factor of e. widely distinct features of the time series, which is an important
This distribution prioritizes small values, which is a required characteristic to make the technique robust to distinct processes.
feature to improve the resolution in this region. To demonstrate The distributions of the random variables selected according
the importance of this better resolution, take two time series to Table 1 are presented in Figure 4.
with a noise variance equal to 0.01 and 0.1. An increase of 0.01 in The large data set provides two benefits. First, it guarantees no
both variances results in 100% and 10% increases, respectively. model overfitting.37 Second, it makes the detection performance
In other words, small values are more sensitive to changes. almost independent of the ML technique.42 This makes tests
For the three other variables (signal length, number of with different techniques unnecessary, such as supporting vector
periods, and frequency change factor), other different features machines and decision trees.
14184 DOI: 10.1021/acs.iecr.9b01456
Ind. Eng. Chem. Res. 2019, 58, 14180−14192
Industrial & Engineering Chemistry Research Article

The range of the parameters in Table 1 was made large in library. For the sake of brevity, the mathematical foundation
order to include examples of the most distinct process behind each of the values is not presented in this work. For
conditions. Thus, it virtually guarantees that any time series further information, Keras documentation40 or any good book
extracted from the industry have a similar pair in the artificial on deep learning may be reviewed (Goodfellow’s book37 is
data set. In the rare cases in which the time series has different always a good choice).
features or features outside the range in Table 1, the precision in The hyperparameter tuning can be performed by a grid
the detection and quantification is not guaranteed until the search. This approach requires the network to be trained to all
model is retrained with data that encompass the new features. 11 907 combinations of hyperparameter values, which is
The authors are aware of the extensive and detailed procedure unfeasible. Instead, 100 sets are randomly generated and
for the generation of artificial data. Thus, the authors make evaluated. The set that returns the best performance to the
themselves available to make the data accessible, by request, to validation data set is selected for the final network.
those interested. Beyond the hyperparameter tuning by random search, the
3.2. Data Processing. Before training the network with the softmax function is the activation function of the output layer,
generated artificial data, two simple data processing procedures the sum-of-squares error (as presented in eq 3) is the loss
are still required. The first is the normalization of the magnitude function, and the accuracy is used to evaluate the performance of
of all the time series to a value equal to 1. It guarantees that the the network.
detection is not influenced by the source of the time series, The proposed technique is intended to be applied in offline
where its magnitude may vary largely if it is a pressure or mode. This follows the trend of most of the works on oscillation
concentration measurement, for example. detection. Online detection is usually not required since
The second step is the transformation of the time series from detection speed is not usually crucial in terms of minutes or
the time to the frequency domain and the isolation of the hours.44 If fast detection is required, the technique must be
magnitude information. The magnitude in the frequency applied on a moving window, in which a data window with a
domain is preferred for the following reasons: fixed length is updated whenever new data are collected.
• The time series have different lengths. Transforming these 3.4. Oscillation Quantification Models. Two new DFNs
data by discrete Fourier transform (DFT) with a fixed are trained for the quantification of the number of periods and
number of components guarantees that the data in the the oscillation amplitude. The training procedure is similar to
frequency domain have a constant length. that presented in section 3.3 for oscillation detection with few
exceptions. Since it is not a classification but a regression
• The magnitude of the transformed data is independent of
problem, the output layer has a linear activation function, and
the oscillation phase; furthermore, the training step is
the network performance is measured by the mean absolute
faster and less sensitive to noise, as noted by Nanopoulos
error (MAE). Additionally, it makes sense to quantify only the
et al.43
oscillatory loops, so that the training data set is limited to the
The magnitude in the frequency domain can be computed by oscillatory time series.
the absolute value of the complex vector obtained by fast Fourier The number of periods can easily be converted to the
transform. oscillation frequency or period, which is essential information
In this work, DFT with a fixed length equal to 213 is applied to for plant-wide oscillation detection. When different loops have
the time series. The first half of the DFT magnitude is saved for oscillations with similar frequencies, it is likely that the
later use; the second half is discarded since it is a mirrored oscillation generated in one loop is propagated to the others.
representation of the first. The output of the network y is the value obtained by the
3.3. Oscillation Detection Model. The oscillation equation
i n per zyz
= log10jjj
detection model is, broadly, the deep feedforward network
z
k len*NFFT {
(presented previously in Figure 1) trained with the training data
yn
set (generated in section 3.1), where the input x is the magnitude per (7)
of the DFT, and the output y is labeled as nonoscillatory, regular
oscillation, or nonregular oscillation. As discussed in section 2.2, where nper is the number of periods, len is the length of time
the network requires the tuning of a large number of series, and NDFT is the length of the DFT. The ratio inside the
hyperparameters. For this application, the values in Table 2 logarithm function transforms nper into a value close to the peak
are preselected. in the DFT, and the logarithm function is used to increase the
The preselected values for the layers hyperparameter are resolution of small values.
vectors where each value is the number of neurons in each layer. The oscillation amplitude is the parameter that identifies the
The values’ names are the same as those used in the Keras seriousness of an oscillation. This information can be used to
isolate oscillations that require immediate maintenance from
Table 2. Preselected Values for the Hyperparameters those that do not affect plant performance. Here, the output
value y is
hyperparameter values
1
batch size 3000, 10000, 30000 yamp =
max − min (8)
optimization SGD, RMSprop, Adagrad, Adadelta, Adam, Adamax, Nadam
algorithm
where max and min are the maximum and minimum values of
initialization uniform, lecun_uniform, normal, glorot_normal,
glorot_uniform, he_normal, he_uniform the time series, respectively, before the normalization in section
activation softmax, softplus, softsign, relu, tanh, sigmoid, hard_sigmoid, 3.2. The quantified yamp is the ratio of the amplitude that
function linear, elu corresponds to the oscillation, disregarding the noise and
layers [400, 60], [200, 40], [100, 20], [50, 10], [50, 5], [400, 100, disturbance. After quantification, the ratio must be multiplied
20], [200, 50, 10], [100, 25, 10], [100, 25, 5] back to (max − min) to obtain the true oscillation amplitude.
14185 DOI: 10.1021/acs.iecr.9b01456
Ind. Eng. Chem. Res. 2019, 58, 14180−14192
Industrial & Engineering Chemistry Research Article

In addition to the two quantified parameters presented in this is highly probable that each of the time series in the testing data
section, a similar procedure can be applied to the quantification has a similar pair in the training data.
of the other features present in Table 1. Figure 6 presents eight selected misclassified time series,
whose labels are expected (predicted) and 0, 1, and 2 are
4. TECHNIQUE EVALUATION nonoscillation, regular oscillation, and nonregular oscillation,
The oscillation detection, number of periods quantification, and respectively.
oscillation amplitude quantification networks are evaluated in The classification of a nonoscillatory time series as oscillatory
this section on the artificial testing data set and two industrial is due to strong disturbances that randomly generate a pattern
data sets. similar to an oscillation. The misclassification of the oscillatory
The most computationally expensive step on the technique examples is usually due to the low number of periods and short
application is the search for the model’s hyperparameters. Even length, where the oscillation is easily hidden by a strong variance.
then, the computational time at this stage was less than 1 h for Noise and disturbance are the influences of higher concern on
the search in a high-performance computer in which the training oscillation detection. In Figure 7, the examples with the
was performed in an Nvidia TITAN Xp Graphics Card. For a strongest noise and disturbance effects are presented (0.1%
simple home computer with the training running in a single strongest), in which, again, the labels follow the notation
processor core, the computational time may reach 1 day. For expected (predicted). The accuracy of this analysis was
simplicity, the hyperparameters found in this work can be used, degraded to 70%. The misclassified examples are due to
and thus hyperparameter tuning is not necessary. With the oscillations hidden in strong noise or disturbance. Note that the
hyperparameters in hand, the training of the final models is fast, classification of these examples is difficult even by visual
in the order of a few minutes. inspection. For the analyses of the time series with 1% strongest
4.1. Evaluation on Artificial Data. First, the sets of noise and disturbance influence, the accuracy was higher than
hyperparameters that returned the best performance to the 90%.
validation data set are presented in Table 3. As seen, the Also from Figure 5, it is seen that the misclassification is
hyperparameters differ according to the network task, which slightly more frequent in nonregular oscillations. Nonregularity
reflects the significance of tuning the hyperparameters in time domain spreads the oscillation power in the frequency
individually to each network. domain, which results in a larger oscillatory band with lower
amplitude.46 Thus, the peak in the frequency domain is less
Table 3. Hyperparameter Sets Obtained by a Random Search apparent, harming the detection. In industrial data, nonregular
Technique oscillations usually have the frequency slightly altered around a
mean. This knowledge was incorporated into the detection
number of per oscil amplitude model, which is able to deal with most of the nonregular
hyperparameter oscil detection quant quant
oscillatory signals.
batch size 10000 3000 3000 For the quantification networks, Figure 8 and Figure 9 show
optim. algorithm Adam Adamax Adam the distribution of the examples that returned the worst results
initialization he_uniform uniform glorot_uniform when compared to the expected values. As seen in Figure 8, the
activ. function hard_sigmoid hard_sigmoid sigmoid worst results seem to have a random distribution. The
layers [400, 100, 20] [200, 50, 10] [100, 25, 5] distribution for the number of periods is an exception, where a
region from 100 to 120 inexplicably returned the worst results.
The performance of the networks is then evaluated to the Figure 9 shows the distribution for the amplitude
training, validation, and test data sets generated in Section 3.1, quantification network. Here, the time series with a low number
and the returned results are shown in Table 4. The performance of periods returned the worst results. Also, the returned results
is lower to the validation and test data set, as expected, but the are worse for nonregular oscillations.
small differences do not mean critical overfitting. 4.2. Evaluation of Industrial Data. The final goal of the
technique is the industrial application; thus, the real practical
Table 4. Performance of the Three Networks to the Training, results are the evolution of the technique on industrial data. In
Validation, and Test Datasets this section, the proposed technique is evaluated on two data
oscil detection number of per oscil amplitude
sets. The first data set presents examples borrowed from Jelali
hyperparameter (accuracy) quant (MAE) quant (MAE) and Huang,45 while the second data set exhibits time series
train 0.9961 0.0592 0.0261 extracted from a Brazilian refinery. Some time series are
validation 0.9771 0.0637 0.0530 evaluated in more detail and compared to the results of six
test 0.9746 0.0628 0.0526 other oscillation detection methods extracted from a recent
comparative work by Dambros et al.46
4.2.1. Data Borrowed from Literature. The data borrowed
As seen in Table 4, the accuracy of the oscillation detection from Jelali and Huang45 is a benchmark data set on oscillation
network is higher than 97% for the test data set, which is a good detection and diagnosis. The data set contains the controller
result. Figure 5 shows the distribution of misclassified examples. output (OP) and process output (PV) measurements from 93
The figure indicates that the misclassification usually occurs due control loops. The measurements are mostly from chemical
to a short length, a low number of periods, and a high industries and loops in which the flow is the controlled variable.
disturbance variance. Additionally, the misclassified examples The data are very distinct, with lengths ranging from 200 to
are usually oscillatory time series classified as nonoscillatory. 277 115 points, amplitudes ranging from 0.0184 to 2303.4, and
The accuracy over 97% on the testing data is achievable oscillations in different waveforms and frequencies. The
because the data set has a large number of examples, and both complete results obtained by the application of the proposed
the training and testing data have a similar distribution. Thus, it technique to the 93 control loops are presented in Supporting
14186 DOI: 10.1021/acs.iecr.9b01456
Ind. Eng. Chem. Res. 2019, 58, 14180−14192
Industrial & Engineering Chemistry Research Article

Figure 5. Distribution of the misclassified time series.

Figure 6. Examples of misclassified time series.

Figure 7. Examples in the testing data set with the 0.1% strongest noise and disturbance.

Information. The overall results for the OP and PV time series The results for the oscillation detection by the proposed
are shown in Table 5. technique are presented in Table 6. Also in this table, the results
The technique detects the oscillations in almost all the extracted from Dambros et al.46 for six well-established
provided time series, and most of these oscillations are detection techniques.
nonregular. Additionally, a slight difference is observed between As seen, the proposed technique returned the expected results
the results for the OP and PV time series. It is known that most for all the time series, while the other techniques returned two to
of the loops in the data set are oscillatory,47 but, usually, the five correct results. By the analysis, it is possible to conclude that
conventional techniques identify a lower number of oscillatory
loops, which indicates that the proposed technique is susceptible • Loop CHEM01 shows a typical case of oscillation caused
to weak oscillations. by stiction. The waveform is not a problem for any of the
For further analyses, consider the selected loops with distinct methods, but the mean nonstationarity, even weak, causes
features shown in Figure 10. misclassification by the Thornhill method.

14187 DOI: 10.1021/acs.iecr.9b01456

Ind. Eng. Chem. Res. 2019, 58, 14180−14192
Industrial & Engineering Chemistry Research Article

Figure 8. Distribution of the 5% worst results for the quantification of the number of periods.

Figure 9. Distribution of the 5% worst results for the quantification of the oscillation amplitude.

Table 5. Summary of the Results on the Dataset Borrowed In Figure 10, the efficiency of the quantification is easily
from Jelali and Huang45 verified. In loop CHEM01, for example, the time series clearly
has approximately 10 periods of oscillation. The magnitude of
nonoscillatory regular oscillation nonregular oscillation
the OP and PV time series ranges from approximately 36 to 37
OP 9 29 55
and from 56 to 58, which gives the amplitude of oscillation equal
PV 4 31 58
to approximately 1 and 2. These approximated values match
those presented in Table 7. All other time series can be verified
• Loop CHEM09 is a clear nonoscillatory loop incorrectly
classified as oscillatory by the Zakharov method. in the same way.
• Loops CHEM17 and CHEM21 present nonregular Finally, two important details are observed. First, when the
oscillations, and the time series are also affected by strong oscillation amplitude is variable, the quantified value is
mean nonstationarity. Both features contribute to approximately the mean value. Second, when the time series
misclassification by many methods. have multiple oscillations, an intermediary value for the number
• The time series in loop CHEM20 shows multiple low- of periods is returned. This intermediary value, seen in the OP
frequency, nonregular oscillations (more evident in the measurement in loop CHEM20, for example, does not represent
PV measurement), which are most likely the cause of any useful information. Thus, the quantification of the number
misclassification by three methods. of periods of oscillations in a time series with multiple
• Loop CHEM31 shows a time series with a high number of oscillations by the technique is not efficient.
periods of oscillation and a long length. Again, the This behavior is expected since time series with multiple
proposed technique is not affected by these different oscillations were not included in the training data set. The
features, which is not true for the other 2 evaluated inclusion, however, would be an exhausting task, in which many
methods. different combinations of oscillation frequencies with a different
These results show the capacity of the proposed technique for number of oscillatory components would be required.
the detection of time series with the different features usually The evaluation of the time series by the proposed technique is
found in industrial data. not a time-consuming task. With the trained networks, the
The oscillation quantification of the selected time series by the complete analysis is performed in only a few seconds, even on a
proposed technique is presented in Table 7. simple home computer.
14188 DOI: 10.1021/acs.iecr.9b01456
Ind. Eng. Chem. Res. 2019, 58, 14180−14192
Industrial & Engineering Chemistry Research Article

Figure 10. Examples of time series present in the borrowed data set.

Table 6. Evaluation of the Proposed Technique on the Borrowed Dataset, Where C Represents a Correct Result, and W
Represents an Incorrect Result
loop expected proposed Thornhill Miao Forsman Zakharov Srinivasan Li
CHEM01 1 C W C C C C C
CHEM09 0 C C C C W C C
CHEM17 1 C W C C C W C
CHEM20 1 C C C W C W W
CHEM21 1 C W W W W W C
CHEM31 1 C W C W C C C

Table 7. Quantification of the Selected Examples Borrowed 4.2.2. Data Extracted from Industry. Since the first data set
from the Literature consists almost exclusively of oscillatory time series, a second
data set is also analyzed. This presents 602 measurements (OP,
number of periods amplitude
PV, and/or MV measurements, among others) from 191 loops
loop OP PV OP PV from a refinery extracted over a period of 2 days. In this section,
CHEM01 9.927 9.927 0.962 1.663 selected time series and fragments are selected and analyzed in
CHEM17 66.612 66.612 1.140 1.678 detail.
CHEM20 4.404 12.164 4.223 2.572 The selected fragments are presented in Figure 11, where the
CHEM21 42.551 42.551 4.640 0.859 time series on the left and on the right correspond to time
CHEM31 1019.874 1019.194 1.293 1558.549 windows of 1 and 48 h, respectively. The time series were
selected to test the techniques on time series with different
features, such as low and high-frequency oscillation, nonregular

Figure 11. Examples of time series extract from a refinery (normalized due to confidential issues).

14189 DOI: 10.1021/acs.iecr.9b01456

Ind. Eng. Chem. Res. 2019, 58, 14180−14192
Industrial & Engineering Chemistry Research Article

Table 8. Oscillation Detection Applied to the Dataset Extracted from a Refinery, Where C Represents a Correct Result, and W
Represents an Incorrect Result
time expected proposed Thornhill Miao Forsman Zakharov Srinivasan Li
01H01 1 C C C W C C C
01H02 1 C C C C C C C
01H03 1 C C C C C C C
01H04 1 C W W W C C W
01H05 0 C C C C C C C
01H06 0 C C C C C C C
48H01 1 C W C C C W C
48H02 1 C W C C C C C
48H03 1 C W C W W C C
48H04 1 C W C W W W C
48H05 0 C C C C C C W
48H06 0 C C C C C W C

Table 9. Quantification of the Selected Examples (Normalized Due to Confidential Issues)

time series 01H01 01H02 01H03 01H04 48H01 48H02 48H03 48H04
number of period 14.235 32.561 48.709 9.909 40.513 41.291 26.049 102.453
amplitude 0.506 0.320 0.673 0.506 0.478 0.409 0.437 0.301

oscillation, noise, and disturbance, intermittence, and saturation. to increase the frequency range of coverage of the proposed
The time series present an amplitude ranging from 0.01 to 200. technique.
Here, the time series were normalized to amplitude equal to 1 (1) Evaluate the full-length data set by the method;
and a mean equal to 0 (after the application of the techniques)
(2) Downsample the data by a factor of 10;
due to confidential issues.
The results of the oscillation detection by the proposed (3) Evaluate the sampled data;
technique and the six other selected techniques are presented in (4) If the length of the data is greater than 2000 (10 times the
Table 8. Again, the proposed technique correctly detects the lower bound in the training data set), go to step 2.
oscillation in all time series, while the efficiency of the other Even with an increase in coverage, the detection in time series
techniques ranges from 7 to 11 correct detections over the 12 with less than two periods of oscillation or less than three points
time series. per cycle may be incorrect.
The quantification of the time series by the proposed
technique is presented in Table 9. Compared to Figure 11, the 5. CONCLUSION
results are close to the expected, with the exception of the This work presents a new oscillation detection and quantifica-
number of periods quantified by loop 01H03. The incorrect tion technique based on machine learning. More specifically,
result was also caused by the presence of multiple oscillations. three deep feedforward networks are trained with artificial data
As seen through the analyses of both data sets, the technique is with distinct features aimed at oscillation detection, the number
very sensitive to weak oscillations. This could cause alarm floods of periods quantification, and the oscillation amplitude
in real industrial application. However, since the oscillation quantification.
amplitude is quantified, weak oscillations can be dismissed by The technique was tested using artificial and industrial data.
the selection of a threshold that defines the minimum amplitude Even being based on the DFT, the technique was able to learn
for the detection. Oscillations, even weak, often hide a problem the different features found on industrial time series, which
whose detection may be relevant. Therefore, it is preferred to includes data with noise, disturbances (mean nonstationarity),
report any obtained information about the oscillation intermittence, saturation, and nonregularity in oscillations. In
independent of its magnitude.48 addition, the technique is robust for triangular and square
Also from both data sets, Miao and Li methods reported waveforms, which are common shapes of nonlinearity induced
worse but similar results compared to the proposed technique. It oscillations. These waveforms generate harmonics in the
is known from Dambros et al.46 that these methods have high frequency domain, which were disregarded.
As a disadvantage, the number of periods quantified in the
sensitivity and low specificity, which means they tend to report
case of multiple oscillations is unreliable. For future works, the
oscillation in nonoscillatory signals. Since most of the signals training of models for the specific detection and quantification of
tested are oscillatory, the methods’ accuracy is high, which does multiple oscillations is recommended.
not mean good performance. The reader can check the work by In this work, only the number of periods and the oscillation
Dambros et al.46 for deeper analyses. amplitude were quantified. Also, other properties could be
Note that the original data window size is large (2 days). The measured by a similar approach, such as the noise variance,
direct application of the technique to only the raw data favors the disturbance amplitude and level of oscillation nonregularity.
detection of low-frequency oscillations, while high-frequency Additionally, similar procedures could be applied to different
oscillations can mostly be misclassified as noise. The window fields in control loop performance monitoring (CPM), such as
size decision is a known problem on oscillation detection whose controller performance assessment, root cause analysis, stiction
solution is still open. The following procedure is recommended detection, and stiction quantification.
14190 DOI: 10.1021/acs.iecr.9b01456
Ind. Eng. Chem. Res. 2019, 58, 14180−14192
Industrial & Engineering Chemistry Research Article

Figure 12. Proposed distributions for m, c, and M are equal to 0, 1, and 1, respectively.

The data generation is the most extensive procedure required

for the application of the proposed technique. Even using the
■
*
ASSOCIATED CONTENT
S Supporting Information
data available (by request) by the authors, it is strongly The Supporting Information is available free of charge on the
suggested that future works use industrial data combined with ACS Publications website at DOI: 10.1021/acs.iecr.9b01456.
data augmentation techniques to fulfill the data requirements of Table of the complete results obtained by the proposed
the machine learning technique. Thus, there would be a greater technique applied to the dataset borrowed from Jelali and
guarantee that the data collected from the real industrial Huang, 2010 (PDF)
application are closer to those used for training.

■ APPENDIX A: LINEAR-EXPONENTIAL
■ AUTHOR INFORMATION
Corresponding Author
DISTRIBUTION *E-mail: [email protected].
The probability density function (pdf) of the linear-exponential ORCID
distribution is Jônathan W. V. Dambros: 0000-0002-4584-2952

l
o
o
o
Notes

o
o
0 x < m, The authors declare no competing financial interest.
f (x ; m , c , M ) = m
o
o
o
o
o
o
n λe
λ
−(x − c)/ M
m ≤ x < c,
x ≥ c, (9)
■ ACKNOWLEDGMENTS
The development of this work was only possible thanks to the
support of many contributors. The authors are very grateful for
where m is the beginning point of the linear part, c is the the scholarship provided by CAPES under Edital CAPES no.
transition point, M is the length required for the exponential part 15/2017, the Nvidia TITAN Xp Graphics Card provided by
to decrease by e, and Nvidia under the GPU Grant Program, the space provided by
the University of Kaiserslautern, the industrial data provided by
1 Jelali and Huang, and all the support provided by Petrobras and
λ= Trisolutions. Additionally, this work was partly funded by the
M+c−m (10)
German Research Foundation (DFG) Awards KL 2698/2-1 and
The cumulative distribution function is given by GRK1589/2 and by the Federal Ministry of Science and
Education (BMBF) Awards 031L0023A and 01IS18051A.

l
o
F (x ; m , c , M )
o
o
o
■ REFERENCES

o
=o
0 x < m,
m
o
(1) Bialkowski, W. L. Dreams vs. Reality: A View from Both Sides of

o
o
o
the Gap. Pulp and Paper, Canada 1994, 19−27.

o
λ (x − m) m ≤ x < c,
o
n λ(c − m) + λ(M − Me
(2) Ender, D. B. Process Control Performance: Not as Good as You
−(x − c)/ M Think. Control Eng. 1993, 40 (10), 180−190.
) x≥c
(3) Torres, B. S.; Carvalho, F. B.; Fonseca, M. O.; Filho, C. S.
(11) Performance Assessment of Control Loops − Case Studies. In Proc.
IFAC ADCHEM; Gramado, Brasil, 2006.
The linear-exponential distributions for m, c, and M equal 0, 1, (4) Desborough, L.; Miller, R. Increasing Customer Value of
and 1, respectively, are exemplified in Figure 12. The Industrial Control Performance Monitoring  Honeywell’s Experi-
distribution is simple to implement and has intuitive parameters. ence. AIChE Symp. Ser. 2002, 153−186.

14191 DOI: 10.1021/acs.iecr.9b01456

Ind. Eng. Chem. Res. 2019, 58, 14180−14192
Industrial & Engineering Chemistry Research Article

(5) Paulonis, M. A.; Cox, J. W. A Practical Approach for Large-Scale (28) Ullah, M. F.; Das, L.; Parmar, S.; Rengaswamy, R.; Srinivasan, B.
Controller Performance Assessment, Diagnosis, and Improvement. J. On Developing a Framework for Detection of Oscillations in Data. ISA
Process Control 2003, 13 (2), 155−168. Trans. 2019, 89, 96.
(6) Choudhury, M. A. A. S. Automatic Detection and Estimation of (29) Wang, J.; Huang, B.; Lu, S. Improved DCT-Based Method for
Amplitudes and Frequencies of Multiple Oscillations in Process Data. Online Detection of Oscillations in Univariate Time Series. Control Eng.
In ADCONIP 2014; Hiroshima, Japan, 2014; pp 514−519. Pract. 2013, 21 (5), 622−630.
(7) Hägglund, T. A Control-Loop Performance Monitor. Control Eng. (30) Xie, L.; Lang, X.; Horch, A.; Yang, Y. Online Oscillation
Pract. 1995, 3 (11), 1543−1551. Detection in the Presence of Signal Intermittency. Control Eng. Pract.
(8) Miao, T.; Seborg, D. E. Automatic Detection of Excessively 2016, 55, 91−100.
Oscillatory Feedback Control Loops. In Proceedings of the 1999 IEEE (31) Aftab, M. F.; Hovd, M.; Sivalingam, S. Improved Oscillation
International Conference on Control Applications (Cat. No.99CH36328); Detection via Noise-Assisted Data Analysis. Control Eng. Pract. 2018,
IEEE, 1999; Vol. 1, pp 359−364. 81, 162−171.
(9) Thornhill, N. F.; Huang, B.; Zhang, H. Detection of Multiple (32) Xie, L.; Lang, X.; Chen, J.; Horch, A.; Su, H. Time-Varying
Oscillations in Control Loops. J. Process Control 2003, 13 (1), 91−100. Oscillation Detector Based on Improved LMD and Robust Lempel−
(10) Jordan, M. I.; Mitchell, T. M. Machine Learning: Trends, Ziv Complexity. Control Eng. Pract. 2016, 51, 48−57.
Perspectives, and Prospects. Science (Washington, DC, U. S.) 2015, 349 (33) Karra, S.; Jelali, M.; Karim, M. N.; Horch, A. Detection of
(6245), 255−260. Oscillating Control Loops. In Detection and Diagnosis of Stiction in
(11) Zabiri, H.; Maulud, A.; Omar, N. NN-Based Algorithm for Control Loops; Springer-Verlag: London, 2010; pp 61−100.
Control Valve Stiction Quantification. WSEAS Trans. Syst. Control (34) Bauer, M.; Horch, A.; Xie, L.; Jelali, M.; Thornhill, N. The
2009, 4 (2), 88−97. Current State of Control Loop Performance Monitoring − A Survey of
(12) Venceslau, A. R. S.; Guedes, L. A.; Silva, D. R. C. Artificial Neural Application in Industry. J. Process Control 2016, 38, 1−10.
Network Approach for Detection and Diagnosis of Valve Stiction. In (35) Thornhill, N. F.; Horch, A. Advances and New Directions in
Proceedings of 2012 IEEE 17th International Conference on Emerging Plant-Wide Disturbance Detection and Diagnosis. Control Eng. Pract.
Technologies & Factory Automation (ETFA 2012); IEEE, 2012; pp 1−4. 2007, 15 (10), 1196−1206.
(13) Farenzena, M.; Trierweiler, J. O. A Novel Technique to Estimate (36) Duan, P.; Chen, T.; Shah, S. L.; Yang, F. Methods for Root Cause
Valve Stiction Based on Pattern Recognition; Elsevier Inc., 2009; Vol. 27. Diagnosis of Plant-Wide Oscillations. AIChE J. 2014, 60 (6), 2019−
(14) Ge, Z.; Song, Z.; Ding, S. X.; Huang, B. Data Mining and 2034.
Analytics in the Process Industry: The Role of Machine Learning. IEEE (37) Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT
Access 2017, 5, 20590−20616. Press, 2016.
(15) Wuest, T.; Weimer, D.; Irgens, C.; Thoben, K. D. Machine (38) Bishop, C. M. Pattern Recognition and Machine Learning
Learning in Manufacturing: Advantages, Challenges, and Applications. (Information Science and Statistics); Springer-Verlag: Berlin, 2006.
Prod. Manuf. Res. 2016, 4 (1), 23−45. (39) Géron, A. Hands-On Machine Learning with Scikit-Learn and
(16) Dambros, J. W. V; Trierweiler, J. O.; Farenzena, M. Oscillation TensorFlow; O’Reilly Media, Inc., 2017.
Detection in Process Industries–Part I: Review of the Detection (40) Chollet, F.; et al. Keras, 2015, https://keras.io.
Methods. J. Process Control 2019, 78, 108−123. (41) Abadi, M.; Barham, P.; Chen, J.; Chen, Z.; Davis, A.; Dean, J.;
(17) Thornhill, N. F.; Shah, S. L.; Huang, B. Detection and Diagnosis Devin, M.; Ghemawat, S.; Irving, G.; Isard, M.; et al. TensorFlow: Large-
of Unit-Wide Oscillations. In Process Control and Instrumentation 2000 Scale Machine Learning on Heterogeneous Systems, 2015.
(PCI2000); 2000; pp 26−28. (42) Banko, M.; Brill, E. Scaling to Very Very Large Corpora for
(18) Thornhill, N. F.; Shah, S. L.; Huang, B. Detection of Distributed Natural Language Disambiguation. In Proceedings of the 39th Annual
Oscillations and Root-Cause Diagnosis. In Proceedings of CHEMFAS 4; Meeting on Association for Computational Linguistics; 2001; pp 26−33.
Jejudo Island, Korea, 2001; pp 167−172. (43) Nanopoulos, A.; Alcock, R.; Manolopoulos, Y. Information
(19) Forsman, K.; Stattin, A. A New Criterion for Detecting Processing and Technology; Mastorakis, N., Nikolopoulos, S. D., Eds.;
Oscillations in Control Loops. In Control Conference (ECC), 1999 Nova Science Publishers, Inc.: Commack, NY, USA, 2001; pp 49−61.
European; Karlruhe, Germany, 1999; pp 2313−2316. (44) Horch, A. Oscillation Diagnosis in Control Loops ∼ Stiction and
(20) Zakharov, A.; Zattoni, E.; Xie, L.; Garcia, O.; Jämsä-Jounela, S. Other Causes. In 2006 American Control Conference; IEEE: Minneap-
An Autonomous Valve Stiction Detection System Based on Data olis, MN, USA, 2006; pp 2086−2096.
Characterization. Control Eng. Pract. 2013, 21 (11), 1507−1518. (45) Jelali, M.; Huang, B. Detection and Diagnosis of Stiction in Control
(21) Zakharov, A.; Jämsä-Jounela, S. Robust Oscillation Detection Loops; Jelali, M., Huang, B., Eds.; Advances in Industrial Control;
Index and Characterization of Oscillating Signals for Valve Stiction Springer: London, 2010.
Detection. Ind. Eng. Chem. Res. 2014, 53 (14), 5973−5981. (46) Dambros, J. W. V; Trierweiler, J. O.; Farenzena, M.; Kempf, A.;
(22) Zhang, K.; Huang, B.; Ji, G. Multiple Oscillations Detection in Longhi, L. G. S.; Teixeira, H. C. G. Oscillation Detection in Process
Control Loops by Using the DFT and Raleigh Distribution. IFAC- Industries–Part II: Industrial Application. J. Process Control 2019, 78,
139−154.
PapersOnLine 2015, 48 (21), 529−534.
(47) Jelali, M.; Scali, C. Comparative Study of Valve-Stiction-
(23) Matsuo, T.; Sasaoka, H.; Yamashita, Y. Detection and Diagnosis
Detection Methods. In Detection and Diagnosis of Stiction in Control
of Oscillations in Process Plants. In Knowledge-Based Intelligent
Loops: State of the Art and Advanced Methods; Jelali, M., Huang, B., Eds.;
Information and Engineering Systems; 2003; pp 1258−1264.
Springer: London, 2010; pp 295−358.
(24) Matsuo, T.; Tadakuma, I.; Thornhill, N. F. Diagnosis of a Unit-
(48) Horch, A. Benchmarking Control Loops with Oscillations and
Wide Disturbance Caused by Saturation in a Manipulated Variable. In
Stiction. In Process Control Performance Assessment; Ordys, A., Uduehi,
IEEE Advanced Process Control Applications for Industry Workshop;
D., Johnson, M., Eds.; Advances in Industrial Control; Springer:
IEEE: Vancouver, Canada, 2004; pp 1−9.
London, 2007; pp 227−257.
(25) Matsuo, T. Application of Wavelet Transform to Control System
Diagnosis. In IEE Seminar on Control Loop Assessment and Diagnosis;
IEE: London, 2005; Vol. 2005, pp 81−88.
(26) Srinivasan, R.; Rengaswamy, R.; Miller, R. A Modified Empirical
Mode Decomposition (EMD) Process for Oscillation Characterization
in Control Loops. Control Eng. Pract. 2007, 15 (9), 1135−1148.
(27) Li, X.; Wang, J.; Huang, B.; Lu, S. The DCT-Based Oscillation
Detection Method for a Single Time Series. J. Process Control 2010, 20
(5), 609−617.

14192 DOI: 10.1021/acs.iecr.9b01456

Ind. Eng. Chem. Res. 2019, 58, 14180−14192