Prediction of Paperboard Properties with Virtual OnLine
Analyzers
By Osvaldo Vieira, Klabin S.A.; Enrique L. Lima, Professor, COPPE/UFRJ,
Brazil; Ivo Neitzel, Professor, Universidade Estadual de Maringá,
Brazil; Mário Gerhard, Director, Pavilion Technologies Inc.
Under a program to reduce production costs and increase quality,
Telêmaco Borba’s Klabin S.A. mill installed a system using
mathematical models to predict, online and in realtime, several properties
of the products. The installed virtual online analyzers (VOA) run every
minute, providing values of several paperboard properties. In addition
to an increase in quality, the system has boosted operator confidence,
allowing them to operate closer to the process constraints and the minimum
product specification. Making product closer to the specification reduces
the amount of fiber giveaway. It is anticipated that this reduction in
fiber consumption will result in savings of over US$800,000 at current
market pulp prices.
Mathematical models can be used to represent phenomena that occur within
a paper machine. Those cause and effect relationships between input variables
(stock flows, headbox pressures, basis weight and others) and the output
sheet properties (MD and CD bending stiffness and consequently the geometric
average, latic acid and peroxide absorption, edge wicking index, and plybond)
are often not well known by engineers and scientists, making the use of
traditional first principles modeling techniques impractical.
The mathematical models utilized in this system were developed using
a gain constrained neural network empirical technique, as described in
reference [1]. Neural networks are well suited to define complex, nonlinear
cause and effect relationships between input and output variables, eliminating
the need for first principles equations. The operating data of a process,
in this case a paper machine, is used to build these models. Accurate
paper machine sheet property predictions developed from neural networks
are used as soft sensors, replacements or enhancements for lab testing
or feedback to closed control and optimization systems.
A simplified schema of a neural network is shown in Figure 1.
The nodes represent transfer functions, which are highly interconnected.
The degree of interconnectivity is defined by weights, shown as P in the
figure. In order to build models, the data history series is presented
sequentially to the network. For each set of input data at a certain instant,
the outputs calculated by the neural net are compared to the real values
(the lab analysis of the sheet properties), and an algorithm adjusts the
weights of the network in order to minimize the difference between the
real values and the calculated ones. The neural network is therefore a
mathematical structure that adapts to the data. This process of adaptation
is called training the neural net. After training, the relationships between
inputs and outputs of the process are established.
Figure 1. Neural Network Structure
Theoretically, the construction of models using neural networks is simple,
and the technique is well known. However in practice, this must be done
carefully. The fundamental requirement is that the historical data is
of good quality. As there are always shutdowns, instrument failures and
other problems, the data must be carefully cleaned. Such cleaning needs
the use of several techniques as well as common sense and knowledge of
the process operation. In this application, several variables were moderately
filtered while others were mathematically created inputs for the models.
Input variables that have a linear correlation coefficient higher than
0.8 present serious problems for the training of neural networks because
they carry almost the same information; the algorithm can’t distinguish
appropriately the difference of the influence of each variable, producing
wrong process gains. The gain constrained training technique solves this
problem, but must be used in conjunction with knowledge of the process.
The resulting model is of very high quality as it is based on all of the
information available about the process, both data and the engineer’s
knowledge.
Any model that is to be used online is carefully analyzed for precision,
coverage of full operational region of the process and the correct gains,
among other features. For each model, validation is partially verified
through the use of additional data that was not used for training.
Once an adequate model is obtained, it is useful for two primary purposes
on a paper machine: virtual online analyzers (VOA) or control. To achieve
this, specific software was employed, that was developed for ease of use
at a reasonable cost. Specifically for VOA’s, it is important that
the software has the additional functionality such as validation of realtime
input data, alarming in case of abnormalities and calculations using the
inferred values.
Finally, no model is perfect. Even if many input variables are used,
the possibility always exists that disturbances not accounted for in the
model will alter the paperboard properties. Typical examples are changes
in the stock quality or slow drifts in the instruments that collect data
for the model. These types of disturbances can be compensated for by a
key feature of the virtual analyzers – they can be calibrated online
during normal operation. Laboratory analysis can be used, regularly or
at random, and a correction factor can be generated that compensates for
the effects of the unmeasured disturbances, as presented in Figure
2 and detailed in reference [2].
Figure 2. The output of the virtual analyzer is calibrated online.
For this online application, once the historical process data from the
paperboard machine was formatted and preprocessed, model development was
performed offline. The models were then transferred to an online computer
that communicates with the mill’s distributed control system (DCS).
A specific software module collects variable values, which are input to
the online model every minute. The models generate a set of predicted
values and the results are written back to the DCS for display to the
operators, according to Figure 3.
Figure 3. The VOA software reads the input variables, performs the
calculations and writes the inferred results back in the DCS.
Process description and data used
Paper machine # 7 at Telemaco Borba uses three different pulps: unbleachead
for the bottom and middle layers, and bleached at the top layer. The machine
has three stock chests and three white water tanks with their associated
fan pumps. At each of these tanks the chemicals are added: sizing agent,
retention agents and starch, among others. The bottom and middle pulps
are launched in the same fourdrinier through a single headbox. The top
layer has its own specific headbox and uses another fourdrinier. The two
wet sheets are joined at the presses and raw starch is used to glue them.
The machine has seven groups of drying cylinders, after which a scanning
head (frame # 3) measures basis weight, caliper and moisture. Two coaters
are used, and after each one there is an additional scanning head. The
machine has conventional controls for basis weight and moisture.
Due to the complexity of the machine, more than 150 variables were collected,
at a 1minute frequency. This represents either snapshots or a 1minute
average of the variables, without any other treatment such as filtering
or averaging. Additionally, laboratory data was used from samples collected at the end of each reel. Eight months
of data, corresponding to 6700 reels where used for the construction of
the models.
Figure 4 shows a select group of variables over a fiveday
period. The paperboard properties and other analysis appear as steps,
as the data system records the values of the variables and treats them
as constant until the next analysis. For building the models, only the
values of the variables at the moment of the changes were used. These
variables are referred to/identified with the suffix “pontual”
(instantaneous) in the graphics and tables that follow.
Figure 4. Some process variables and paperboard properties used
in modeling.
The data shown has already been cleaned, taking out periods of shutdowns,
sheet breaks, and problems with instrumentation and laboratory errors.
Bending stiffness models
The bending stiffness models were initially constructed using a large
number of input variables, which were later reduced by applying the engineer’s
knowledge of the paper machine and from analysis of results obtained from
preliminary models. All the models are of the quasistatic type; they
take into account the specific time delay of each input variable to the
output. The properties of the sheet are referred to at instant zero while
the input variables are delayed (the time delay, in minutes, is indicated
in the figures and tables with a negative sign in brackets, after the
name of the variable).
Stiffness is mainly related to caliper and basis weight of the sheet.
However, other variables also influence stiffness, such as moisture, degree
of pulp refining, jet to wire ratio, sizing agent, and retention agents.
Basis weight and caliper have different influences over MD and CD stiffness.
The detection of this difference as well as the influence of some of the
less important variables was only possible from empirical modeling. This
represents additional benefit – increased process knowledge for
the plant engineers and operators.
Figure 5 shows some of the variables that have greater
influence on MD and CD stiffness, ranked by the normalized sensitivity.
Figure 6 includes the response curves of these properties
(in the yaxis) to the input variables (in the xaxis, scaled from 0 to
100% of each variable range). Some nonlinearities of the process can
be seen by the curvature of the lines in this figure.
Figure 5. Bending stiffness input variables ranked by sensitivities.
Figure 6. Response curves of bending stiffness to some input variables.
Figure 7 shows the actual value (lab analysis) of the
CD stiffness on the xaxis and the value calculated by the model on the
yaxis for all paperboard grades produced by machine MP7. The standard
deviation of the residue (difference between predicted and actual) is
1.9% of the range of CD stiffness. The distribution of this residue is
shown in Figure 8 (the average of the error is almost
zero). Using a timeline for the xaxis, the actual and the predicted values
are plotted in Figure 9, during a period when several
grades of paperboard were produced.
Figure 7. CD bending stiffness real and predicted values.
Figure 8. Residue distribution for all kinds of paperboards.
Figure 9. Laboratory and prediction of CD bending stiffness of various reels.
Both the MD and CD bending stiffness models are of high quality, this
being due mainly to the strong influence of the basis weight and caliper
over these properties.
Peroxide and lactic acid absorption index models
These two properties are similar in nature, primarily influenced by the
chemical additives injected in the stock flow, particularly the sizing
agent and starch. However, other variables such as moisture and density
also influence the properties.
Figure 10 shows that, for lactic acid, the influence
of sizing agent is positive, reducing absorption. On the other side, moisture
and retention agents negatively affect the property. The high nonlinearity
of the absorption as a function of the glue content can also be seen.
Figure 10. Response of lactic acid to some input variables.
In Figure 11, the actual peroxide absorption (lab measurement)
and the values calculated by the model are plotted for a selected period
of time. The standard deviation of the residue is 4.1% of the range of
the property. It also can be seen that the laboratory analysis of this
property presents high variability.
Figure 11. Laboratory and prediction of peroxide absorption for
various reels.
Plybond model
This property represents the energy necessary to separate the plies of
the paperboard. The operational objective is that it must not be less
than a certain minimum specification value.
Contrary to other properties, plybond depends on many input variables.
The prediction results are reasonable, but more reliable models are still
under development. The standard deviation of the residue is 6.3% of the
range of the property (Figure 12).
Figure 12. Plybond real and predicted values.
Online monitoring of the properties
Using the virtual online analyzers, the MP7 operators can now monitor
in realtime (every 1 minute) the key properties of the paperboard being
manufactured. This helps in two main quality aspects:
a) It is possible to take corrective actions on the properties and see
the results during the production of the reel. Figure 13
shows a period of time in which five reels of the same paperboard grade
were produced resulting in small variations in stiffness. The variable
in steps is the CD bending stiffness as measured in the lab, while StiffCD_VOA
is the result of the prediction. Each row represents one minute. It can
be seen that stiffness can vary significantly during the production of
a reel; the main cause is due to moisture variability. Based on the 1minute
information, operators can act so as to compensate for the most significant
trends in the properties, stabilizing the values and avoiding violations
of the specification limits.
Figure 13. CD bending stiffness monitoring at every minute.
b) It is possible to reach the specification for a new grade faster than
previously. Figure 14 shows a situation before installation
of the VOA’s, when there was a transition from a 190 g/m2 coated
grade to a 266 g/m2 uncoated board. As can be seen, almost two reels were
out of specification, with the target finally achieved at the third reel.
The specification during a transition is now reached at the beginning
of the first reel, because of the continuous monitoring using the VOA’s.
Figure 14. Grade transition before the use of VOAs.
Economic results
Using the virtual online analyzers, the improvement of the paperboard
quality is significant, expressed as the result of less variability.
However, the most significant measurable economic benefits come from
the possibility of producing paperboard closer to the limits of the specifications,
reducing give away. Previously, with information only after the end of
the reel, operators relied only on their experience and a too broad margin
of safety was used. With the VOA’s as instantaneous and continuous
guides, the operators are confident in reducing the operating margins.
In a review after an eightmonth period of production, the mill staff
was able to operate the process closer to minimum quality specifications.
Bending stiffness was formerly maintained at a level 1.5% above the minimum
specifications; the current typical operation is 0.7% above the minimum
specification. This reduction in fiber give away, reflected in reduced
basis weight has resulted in a savings of more than US$800,000/year based
on current market pulp prices.
Next steps
Each quality property of paperboard depends on several process variables.
Some can be manipulated; others are a consequence of factors that are
beyond the operator’s actions (the type of furnish, etc.). Such
disturbances are many and they act in different ways in the properties.
The control of MP7 machine is therefore a multivariable problem, which
is difficult for the operators to manage. Klabin plans on implementing
a nonlinear modelbased predictive controller to directly control the
properties referred to in this paper. Better moisture control is also
in the scope of the planned project.
In addition, there is the possibility of minimizing the overall cost
of chemicals. For instance, sizing agent has a positive effect on the
peroxide absorption but other additives have a negative effect. There
are tradeoffs that can be considered in the use of these multiple chemicals
– a classic economic optimization problem. Offline optimization
has already been tested and Klabin plans to implement online optimization
using technology that is already available in the market [4].
References:
 Training Feedforward Neural Networks with Gain Constraints; E. Hartman;
Neural Computation, 1999.
 Analisadores Virtuais; M. Gerhard; Revista Intech Brazil, Feb/02.
 Prediction and Control of Paper Machine Parameters Using Adaptive
Technologies in Process Modeling; C.A. Schweiger, J. B. Rudd; TAPPI 1994
Process Control Symposium
 Neural Network Based Model Predictive Control; S. Piché, J.
Keeler, G. Martin, G. Boe, D. Johnson, M. Gerules; Neural Information
Processing Systems (NIPS)’99 Conference.
About the authors:
Osvaldo Vieira is process engineer at Klabin S.A., Brazil, and can be
reached at osvaldov@klabinpr.com.br;
Enrique L. Lima is a professor at COPPE/UFRJ, Brazil, and can be reched
at enrique@peq.coppe.ufrj.br;
Ivo Neitzel, is a professor at Universidade Estadual de Maringá,
Brazil, and can be reached at ivo@deq.uem.br;
and Mário Gerhard is director, Pavilion Technologies, Inc., Brazil,
and can be reached at mgerhard@pav.com.
For more information on Pavilion’s products in the pulp and paper industry,contact David Garcia at dgarcia@pav.com or 1 512 4381525.
