Roffel, B., & Betlem, B. (2006). Process Dynamics and Control: Modeling for Control and Prediction.

West Sussex, England: John Wiley & Sons Ltd.

Chapter 1: Introduction to Process Modeling
1.2 Dynamic Systems Modeling
Modeling is the procedure to formulate the dynamic effects of the system that will be considered
into the mathematical equations.
The outputs: dependent variables that characterize and describe the response of the system
The manipulated inputs: adjustable independent variables that are not influenced by the system
The disturbances: external changes that cannot be influenced, but have an impact on the behavior
of the system
Manipulated inputs and disturbances both have an impact on the system
The former can be manipulated; the latter cannot.
A model of a system is a representation of the essential aspects of the system in a suitable
mathematical form that can be experimentally verified in order to clarify questions about the
system.
Black box model: Experimental model, such as fuzzy models or neural network models
When modeling, one should try to develop the simplest form of the model that is required to
achieve the set goal
Limit complexity
The purpose of the model is to provide insights
A useful addition to the modeling steps is a sensitivity analysis, which can give an indication which
relationships determine the result of the model
Level of detail: time, space and function
Models can encompass large time frame/ time constants.
For spatial description: relevant to know whether the system can be considered lumped or not.
Lumped: all variables are independent of the location.
Distributed systems: variables are location dependent. Most variables change with time or location.
1.3 Modeling Steps
Modelling steps: System analysis->Model design->Model analysis
System Analysis
o Focuses on the system itself, in order to investigate which physical phenomena take place
and are relevant with respect to the modeling goal
 o Input and output model
Model Design
o Determine the assumptions
o Model structure, equations, parameters
o Verification and validation
Model Analysis
o Model will have to be solved mathematically, simulated by using a computer or be
transformed to a mathematical form that is suitable for the analysis
o Suitable forms: transfer function and state models
o Transfer function: linearized input-output relationships
o This form of model will show how the output signal changes for a given change in the input
signal
o State models are differential equations in matrix format
1.5 Types of model
White box versus black box models
o Based on input/output data and only describes the relationship between the measure input
and output data of the process
o Mathematical representations include time series models (such as ARMA, ARX, Box and
Jenkins models, recurrent neural network models, recurrent fuzzy models)
o The models are usually discrete linear transfer function type models, which provide a
representation of the dynamic behavior of the process at discrete sampling intervals
o Simpler to develop and both their structure and parameters can empirically be identified
from plant data
Linear versus non-linear models
o Linear models often provide an accurate description of the reality provided the operating
range is limited
o A first order Taylor series approximation is appropriate, where the first-order derivative is
used to describe the behavior around the operating point
o Much easier to handle mathematically and easier to interpret as the relationship between
input and output is explicit
o Sometimes, linear techniques can be used to describe non-linear process behavior, such as
fuzzy model, which is a combination of local linear models in distinct operating areas
 o Application of empirical techniques for modeling nonlinear process behavior has therefore
become very popular, such as application of neural networks
Continuous vs Discrete models
o Discrete time models describe the state of the system at a given time intervals and are
therefore useful for efficient computation
o When using computers for calculation of the model output, continuous models need to be
discretized, since numerical solution methods require discrete models
o Discretisation of model equation refers to the approximation of the first and higher order
derivatives in the model: forward difference, backward difference & central diff approx..
o For process control, it is a good practice to start with linear model

Chapter 5 Transformation Techniques

5.2 Laplace Transform
To analyse the behavior or to control complex systems, linear methods are more suitable
Nonlinear: continuous/ discrete time domain
Linear: Laplace/z-domain
Laplace transformation is a transformation from the time domain to the s-domain, where s is
variable in the complex s-plane: s=a+jb
Y: process output; x: process input; tau: process time constant; K: process gain
Laplace transform is an elegant way of writing the process input-output relationships in the form of
a transfer function
5.4 Transfer functions
A transfer function is defined as the ratio between the output and input in the Laplace domain
Gp = y(s)/u(s)
Laplace transformation can be used to analyse system behavior for control purposes and to solve a
system analytically for different kinds input signals

Chapter 19: Introduction to Black Box Modeling

19.1 Need for Different Model Types
Process behavior is usually nonlinear
If the process is controlled and the operating range is small, a linear process model may be an
adequate approximation of reality
 For control and prediction type applications, models are usually dynamic
If process conditions vary over a wide range, there may be a need for a nonlinear empirical model
Dynamic neural network or fuzzy model: dynamic nonlinear model
Wiener model: process dynamics are represented by linear model, such as state space model
The static characteristics of the process are then modeled by a polynomial, able to represent the
nonlinearity
If model is linear and static: use partial least squares modeling
Neural network: true nonlinear model
Fuzzy model: Linear model

Chapter 24: Time-series Identification

24.4 Cross-correlation and Autocorrelation
The data that are or will be collected should be appropriate data for model identification
The data set should contain sufficient dynamic and static information, such that the model can be
identified and reflects the true process behavior
Often normal process operating data are not fit for dynamic model identification
It does not contain sufficient information, resulting in a poor model
The process input is usually perturbed (influencing a system by altering its path), for example using a
Pseudo-Random Binary Sequence (PRBS)
The PRBS signal provides a sequence of upward and downward steps
Once a proper PRBS signal has been identified and process data has been collected, one should
determine the cross correlation of the process input-output signals and the autocorrelation of the
process input.
Cross correlation can be used to check whether there is sufficient impact of the process input on the
process output
24.7 Design of plant experiments
24.7.2 Step type input changes
Effective way of determining the relationship between an input and an output
Once the data has been collected, it can be analysed by assuming a simple first-order plus dead time
model (process gain, time constant and time delay) and fitting model to the process data
24.7.3 PRBS Type Input Change
In parametric model identification
24.7.4 Type of Experiment
The best method of collecting meaningful data is to perturb the process input
One could perform an open loop test, in which the process is not controlled.
The disadvantage of this method is that the process may drift away from its normal operating point

Chan, R. W., Yuen, J. K., Lee, E. W., & Arashpour, M. (2014). Application of Nonlinear-Autoregressive
Exogenous model to predict the hysteretic behaviour of passive control systems. Engineering
Structures, 85(2015), 1-10.
Explains about NARX and justification of using it to detect hysteresis in control systems.
Not that relevant
Folder no. 3

Srinivasan, B., Spinner, T., & Rengaswamy, R. (2012). A reliability measure for model based stiction
detection approaches. 8th IFAC Symposium on Advanced Control of Chemical Processes. Singapore: The
International Federation of Automatic Control.
Stiction in control valves is one of the long-standing problems in the process industries which lead to
oscillations in closed loop systems.
Numerous methods have been developed to detection stiction in linear closed loop systems.
Almost all of these methods utilize the fact that the presence of stiction introduces nonlinearities in
the closed loop control system.
However, there exists no measure of reliability for the results provided by these techniques
A measure of reliability is developed for model based stiction detection approaches using frequency
domain analysis of closed loop systems.
Introduction
Plant oscillations are a common cause for variations in product quality and reduction in overall profit
Stiction hinders valve movement, and therefore control action is not implemented properly in the
control loop
This improper movement causes oscillatory behavior in the control loop by introducing a
nonlinearity between the outputs of controller and process
Several methods have been developed for identification of stiction in linear closed-loop systems
 Most stiction detection techniques for linear closed loop systems can be classified into the
following:
o Shape based
o Model based
o Frequency domain based
Rengaswamy et al developed shape-based stiction detection technique in which OP data is fitted, in
piecewise fashion, to both triangular and sinusoidal waves, using least squares estimation which is
used for valve stiction detection.
Choudhury proposed a frequency domain based method for detecting and quantifying stiction in
control valves.
Srinivasan et al presented a Hammerstein based model identification approach for the diagnosis and
quantification of valve stiction.
o The idea behind is to fit a Hammerstein model (linear +nonlinear) between the OP and PV
data sets
o The identified nonlinear element is used for detection and quantification of stiction
Later, several variant of Hammerstein models were proposed for stiction detection in linear closed-
loop systems
Though exist several frameworks to identify stiction in linear closed loop systems, a measure of
reliability for the results provided by these approaches has not been reported.
This work provides a reliability measure for model based stiction detection approaches using
frequency domain analysis of closed loop systems.
Background
1. Problem Definition
A typical closed loop system with stiction is shown
Stiction is explicitly shown as a component preceding the valve though it is a nonlinear phenomenon
associated with the control valve.
**Explains each of the closed loop system variables
Given this typical closed-loop system, the problem addressed in this work is to develop a realiability
measure for model based stction detection technique that
identify stiction by building a data-based model between the controller and process outputs
2. Hammerstein model based stiction detection technique
 Hammerstein-based stiction detection approaches identify a nonlinear stiction element along with a
linear dynamical model between the process output (PV) and controller (OP).
The differences between the measured process outputs and those predicted by model are used to
obtain the Root Mean Squared Error (RMSE)
Stiction parameters correspond to minimum RSME are used for detection and quantification of
stiction
Some of the data-driven models used for the approaches include:
o One parameter stiction model (Srinivasan et al)
o Two parameter stiction (Choudhury et al)
One parameter presents the past and present stem movements
The two parameter namely,
o S, provides information about deadband plus stickband, and
o the slip-jump J, takes into account the offset between valve input and output signals and
represents the jump start of the control valve after it overcomes the stiction and dead band
Using Hammerstein approach, the best linear is compute for each set of sticition parameters based
on Akaikes Information Criterion (AIC)
o AIC is an estimator of the relative quality of statistical models for a given set of data. Given a
collection of models for the data, AIC estimates the quality of each model, relative to each
of the other models. Thus, AIC provides a means for model selection.
The fitting of best linear model is repeated at several stiction model parameters
The combination of the stiction parameters (one or two) and the best linear model at those
parameters that provides the minimum RSME is considered to be the best model for the system
And the stiction parameters from this combination is used to identy the presence of valve stiction in
control loop
Model based stiction detection methods build a linear model along with a nonlinear element
between the controller and process outputs

Kano, M., Maruta, H., Kugemoto, H., & Shimizu, K. (2004). Practical Model and Detection Algorithm for
Valve Stiction. IFAC Dynamics and Control of Process Systems, 859-864.
Stiction is the most common problem in pneumatic control valves.
Stiction causes fluctuation of process variables, which lowers productivity.
 Therefore, developing a method to detect stiction and distinguish it from other causes is crucial to
help operators take an appropriate action for improving control performance.
The present work proposes a valve siction model and new stiction detection algorithms
The proposed model can describe the stiction phenomenon with sufficient accuracy using two
parameters
The usefulness of the proposed detection are demonstrated by comparing with conventional
method
The proposed methods are shown to successfully detect valve stiction, distinguish it from bad tuning
or disturbances, and quantify the degree of stiction by using simulation data sets and real operation
data sets of several chemical processes.
Introduction
The variability of process variables makes it difficult to keep operating conditions close to their
bound and thus causes excessive energy consumption.
Control systems should achieve their best performance to maximize productivity
The poor control performance is caused not only by bad controller but also by undesired
characteristics of control valves
Stiction is the most common in process industry
Important to develop a practical method to detect and distinguish it from other causes, include
inadequate controller tuninig
Two types of valve stiction model:
o A detailed physical model that formulates stiction phenomenon as precisely as possible
Unknown physical parameters; time consuming; not practical to simulate
o A data-driven model that describes the relationship between a controller output and a valve
position
Only has a few parameters; easy to identify and understand; However,
Disadvantage of data-driven model by Choudhury:
o Assumption that input signal is deterministic (only investigate sinusoidal inputs)
o When the input signal is stochastic, the model does not function well
o A valve stiction model should be able to cope with both deterministic and stochastic model
o In deterministic models, the output of the model is fully determined by the parameter values
and the initial conditions initial conditions.
o Stochastic models possess some inherent randomness.
 o Accuracy of model is validated by comparing simulation with real operation data results
Valve stiction model
1. Pheumatic control valve and stiction
General structure of a pneumatic control valve is shown
The valve is closed by elastic force and opened by air pressure
Flowrate is changed according to plug position, which is determine by the balance betweem elastic
force and air pressure
The plug is connected to the valve stem
The stem is moved against the static or kinetic frictional force caused by packing, which is a sealing
device to prevent leakage of process fluid
Smooth movement of stem is impeded by excessive static friction
The valve position cannot be changed until the controller output overcomes static friction
And it is suddenly changed when the difference between elastic force and air pressure exceeds the
maximum static frictional force
2. Stiction model
To model relationship between controller output and valve position of a pneumatic control valve,
need to balance elastic force, air pressure and frictional force
Relationship can be described in figure
The dashed line denotes the states where elastic force and air pressure are balanced
The controller output and the valve position change along this line in an ideal situation without any
friction
The valve is resting at (a) where elastic force and air pressure are balanced
The valve position cannot be changed due to static friction even if the controller output (air
pressure) is increased.
The valve begins to open at (b), where the difference between air pressure and elastic force exceeds
the maximum static frictional force.
Since the frictional force changes from static fs to kinetic fd when the valve starts to move at (b), the
slip-jump of the size J= fs-fd happens and the valve state changes from (b) to (c).
The valve state changes along the line l2, which deviates from the ideal line by fd because the
difference between air pressure and elastic force is equal to fd.
 When the valve stops at (d), the difference between air pressure elastic force needs to exceed fs
again for the valve to open further
Since the difference between them is fd at (d), air pressure must increase by J to open the valve
Once air pressure exceeds elastic force by fd, the valve state changes to (e) and then follows l2
Air pressure begins to decrease when the controller orders the valve to close at (f).
Valve changes its direction and comes to rest momentarily
Valve position does not change until the difference between elastic force and air pressure exceeds
the maximum static frictional force fs.
The valve state (h) is just point-symmetric to (b).
The difference of air pressure between (f) and (h): S=fs+fd
Valve position decreases as it follows line l1
Controller output u; valve position y
Shown flowchart of valve stiction model
o Check if the ipper and lower bounds of the controller output are satisfied
o Two states of valve are explicitly distinguished:
A moving state (stp =0)
A resting state (stp =1)
o us: the controller output at the moment the valve state changes from moving to resting
o us is updated and the state is changed to resting state (stp=1) only when the valve stops or
changes its direction (du(t)du(t-1)<0) while its state is moving (stp=0)
o Two conditions concerning the difference between u(t) and u (s) are checked unless the valve
is in moving state
o The first condition judges whether the valve changes its direction and overcomes the
maximum static friction (corresponding to b and h)
o d=+-1 denotes the direction of frictional force
o The second condition judges whether the valve moves in the same direction and overcomes
friction
o If one of these two conditions is satisfied or the valve is in moving state, the valve position is
updated via the following equation y(t)=u(t)-d/2(S-J)
o The valve position remains unchanged if the valve remains in a resting state
Pf(f) =1/0.2s+1
Advantages of model by Kano: able to cope with stochastic input as well as deterministic input
 us can be updated at appropriate timings by introducing the valve state stp
can change the degree of stiction according to direction of valve movement
Stiction detection methods
There are sections where the valve position does not change even though the controller output
changes. Stiction is stronger as such sections are longer.
The relationship between the controller output and the valve position takes the shape of a
parallelogram if slip-jump J is neglected. Stiction is stronger as the distance between l1 and l2 is
longer.
Choudhury, M., Thornhill, N. F., & Shah, S. L. (2003). A Data-Driven Model for Valve Stiction. IFAC
Advanced Control of Chemical Processes, 245-250.
The presence of nonlinearities, e.g., stiction, hysteresis and backlash in a control valve limits control
loop performance
Stiction is most common in spring-diaphragm type valves that is widely used in the process industry
Though many attempts to understand the stiction phenomena and model it,
There is a lack of proper model which can be understood and related directly to the practical
situation as observed in real valves in the process industry
This study focuses on the understanding , from industrial data, of the mechanism that causes
stiction and proposes a new data-driven model of stiction, which can be directly related to real
valves
Compares simulation results generated using the proposed model with industrial data
Introduction
A typical chemical plant has thousands of control loops
Control performance is very important to ensure tight product quality and low cost of the product in
such plants.
The presence of oscillation in a control loop increases the variability of the process variables, thus
causing inferior quality products, larger rejection rates, increased energy consumption and reduced
profitability
..reported that about 30% control loops are oscillatory due to control valve problems
The only moving part in a control loop is the control valve
If the control valve contains static nonlinearities (stiction, backlash, deadband), the valve output
may be oscillatory which in turn can cause oscillations in the process output.
 Many studies have been conducted to define and detect static friction or stiction.
However, there is a lack of a unique definition and description of the mechanism of stiction.
Many previous studies are based on some physical model of valve friction, where parameters of
physical model such as mass of moving parts of the valve, spring constants and forces are not
explicitly known.
These parameters need to be tuned properly to produce desired responsDe of the valve.
Sticiton and other related problems are identified in terms of % of the valve travel or span of the
valve input signal
The purpose paper: develop an empirical data-driven model of stiction that is useful for simulation
and diagnosis
What is stiction
Instrument Society of America
o Stiction is the resistance to the start of motion, usually measured as the difference between
the driving values required to overcome static friction upscale and downscale
Horch 2000
o The control valve is stuck in a certain position due to high static friction. The integrating
controller then increases the set point to the valve until the static friction can be overcome.
Then the valve breaks off and moves to a new position (slip phase) where it sticks again. The
new position is usually on the other side of the desired set point such that the process starts
in the opposite direction again
o On the contrary, once valve overcomes stiction, it might travel smoothly for some time and
then stick again when the velocity of the valve is close to zero
No clear definition, etc etc
Proposed figure that investigates the phase plot of the valve input-output behavior of a valve
suffering from stiction
It consists of four components: deadband, stickband, slip jump and moving phase
When the valve comes to a rest or changes the direction (point A in figure 1), the valve sticks
After the controller output overcomes the deadband (AB) plus the stickband (BC) of the valve, the
valve jumps to a new position (D) and continues to move.
The deadband and stickband: the behavior of the valve when it is not moving even though the input
to the valve is changing.
 Slip jump: the abrupt release of potential energy stored in the actuator chamber as kinetic energy
due to high static friction,as the valve starts to move.
The magnitude of slip jump is important in determining the limit cyclic behavior introduced by
stiction
Once the valve moves, it continues to move until it sticks again ( E)
In this moving phase, dynamic friction may be much lower than static friction
Practical examples of valve stiction
This section shows the effects of stiction from investigation of industrial control loops data.
First data set: power plant; Second data set: petroleum refinery
Loop 1: level control loop, which controls the level of condensate in the outlet of a turbine by
manipulating the flow rate of the liquid condensate
**Figure 2 shows time domain data**
pv: condensate flow rate
op: controller output
mv: valve position
Loop 2: slave control loop cascaded with a master level control loop
Look at pv-sp plot if it is a cascaded loop and the slave loop is operating under proportional control
Data driven model of valve
Useful because the parameters are easy to choose and the effect of these parameter change is
simple to understand
The proposed data driven model has parameters that can be directly determined from plant data
Model needs: input signal and specification of deadband plus stickband and slip jump parameters
1. Model formulation
Valve might be sticking only when it is at rest or it is changing its direction
When valve changes direction, it comes to a rest momentarily
Once valve overcomes stiction, it starts moving and may keep moving on for sometime depending
on how much stiction is rpesent in the valve
In this moving phase, it suffers only dynamic friction, which is much smaller than the static friction
It continues to move until its velocity is again very close to zero or it changes its direction
Stiction is generally measured as a % of the valve travel or the span of the control signal
 A 2 % stiction means that when the valve gets stuck, it will start moving only after the cumulative
change of its control signal is greater than or equal to 2%
If the range of the control signal is 4 to 20 mA, then 2% stiction means that a change of the control
signal is less than 0.32 mA in magnitude will not be able to move the valve.
This measure includes deadband and stickband. No info about slip jump.
In this modeling approach, the control signal has been translated to the percentage of valve travel
with the help of a linear look-up table.
The model consists of two parameters: deadband+stickband (s) and slip jump (j)
**Figure 4: Flow chart for algorithm of data-driven stiction model**
o The controller output (mA) is provided to the look-up table where It is converted to valve
travel %
o If this is less than 0 or more than 100, the valve is saturated
o If the signal is within 0 to 100% range, it calculates the slope of the controller output signal
o The change of the direction of the slope of the input signal is taken into consideration
o If the sign of the slope changes or remains zero for two consecutive instants, the valve is
assumed to be stuck and does not move
o When the cumulative change of the input signal is more than the amount of the stickband,
the valve slips and starts moving
o The output is again converted back to a mA signal using a look-up table based on the valve
characteristics.
The parameter j signifies the slip jump at the start of the control valve immediately after it
overcomes the deadband plus stickband.
It accounts for the offset between the valve input and output signals.
2. Open loop response of the model under a sinusoidal input
**Figure 5 shows open loop behavior of the new data-driven stiction model in presence of various
types of stiction.
o Second row corresponds to the pure deadband without any slip jump (j=0). The magnitude
of stickband is zero
o Third row shows the undershoot case of a sticky valve, where j < s/2. The valve output can
never reach the valve input. There is always some offset.
 o Fourth row shows pure stick-slip behavior, where j=s/2. No offset between input and
output. Once the valve overcomes stiction, the valve output tracks the valve input
accurately.
o Fifth row shows the valve output overshoots the desired set point or the valve input due to
excessive stiction. j > s/2.
3. Closed loop behavior of the data-driven model
Concentration loop: has slow dynamics with large dead time
Level loop: only has an integrator
**Table 1 shows the transfer functions, controllers and parameters used in simulation.
Concentration loop
o The triangular shape of the time trend controller output is one of the characteristics of
stiction.
o Presence of stiction causes limit cycling of the process output
o In the absence of stiction, there are no limit cycles, which is shown in first row of figure 6
o The presence of pure deadband cannot produce any limit cycle. It only adds dead time to
the process
o Presence of deadband with an integrator produces limit cycle
o If valve position data is available, one should plot valve position (mv) against controller
output (op)
Level control loop
o Second row in Figure 8 shows the deadband can produce oscillations
o With the presence of integrator, a pure deadband can produce limit cycles, otherwise cycle
decays to zero

Capaci, R. B., Scali, C., & Pannocchia, G. (2015). Identification techniques for stiction quantification in the
presence of nonstationary disturbances. Journal of Process Control, 46, 629-634
References
Capaci, R. B., Scali, C., & Pannocchia, G. (2015). Identification techniques for stiction quantification in the
presence of nonstationary disturbances. Journal of Process Control, 46, 629-634.
Chan, R. W., Yuen, J. K., Lee, E. W., & Arashpour, M. (2014). Application of Nonlinear-Autoregressive
Exogenous model to predict the hysteretic behaviour of passive control systems. Engineering
Structures, 85(2015), 1-10.
Choudhury, M., Thornhill, N. F., & Shah, S. L. (2003). A Data-Driven Model for Valve Stiction. IFAC
Advanced Control of Chemical Processes, 245-250.
Kano, M., Maruta, H., Kugemoto, H., & Shimizu, K. (2004). Practical Model and Detection Algorithm for
Valve Stiction. IFAC Dynamics and Control of Process Systems, 859-864.
Roffel, B., & Betlem, B. (2006). Process Dynamics and Control: Modeling for Control and Prediction. West
Sussex, England: John Wiley & Sons Ltd.
Srinivasan, B., Spinner, T., & Rengaswamy, R. (2012). A reliability measure for model based stiction
detection approaches. 8th IFAC Symposium on Advanced Control of Chemical Processess.
Singapore: The International Federation of Automatic Control.