Available online at www.sciencedirect.

com

Computers & Industrial Engineering 55 (2008) 592–605

www.elsevier.com/locate/caie

Ink flow control by multiple models in an offset

lithographic printing process q
Cristofer Englund a,*, Antanas Verikas a,b,1
a
Intellignet Systems Laboratory, School of Information Science, Computer and Electrical Engineering, Box 823, 30118 Halmstad, Sweeden
b
Department of Applied Electronics, Kaunas University of Technology LT-51368, Kaunas, Lithuania

Received 17 November 2006; received in revised form 30 January 2008; accepted 31 January 2008
Available online 7 February 2008

Abstract

A multiple model-based controller has been developed aiming at controlling the ink flow in the offset lithographic print-
ing process. The control system consists of a model pool of four couples of inverse and direct models. Each couple eval-
uates a number of probable control signals and the couple, generating the most suitable control signal is used to control the
printing press, at that moment. The developed system has been tested at a newspaper printing shop during normal pro-
duction. The results show that the developed modelling and control system is able to drive the output of the printing press
to the desired target levels.
Ó 2008 Elsevier Ltd. All rights reserved.

Keywords: Multiple models; Process control; Adaptive modelling; Printing press; Colour printing; Neural networks

1. Introduction

Increased quality demands in the newspaper printing industry have postulated the need to develop tools for
assessing and controlling print quality. Today, print quality assessment and control are two of the few remain-
ing manually handled tasks in the almost completely automated newspaper printing industry.
In offset lithographic printing, a colour print is made of dots of varying size of the four primary colours
cyan (C), magenta (M), yellow (Y), and black (K). Fig. 1 presents an example of such a colour print.
The colour shade perceived by an observer in a small area of a colour print is determined in the halftoning
process where dots of varying size (tonal value) are created, the larger the dots, the darker the colour shade.
After the estimation of the dot sizes the dots are transferred on to the printing plates in the Computer To
Plate (CTP) process. The CTP technique images the dots on thin metal plates. The plates are typically made
out of aluminium that is coated with a thin, approximately 0.1 lm, ink receptive substrate. In the plate devel-
opment process, the substrate is removed from the non-printing areas and the water receptive surface under it

q
This manuscript was processed by Area Editor Gursel A. Suer.
*
Corresponding author. Tel.: +46 35 147261/167693; fax: +46 35 120348.
E-mail addresses: cristofer.enfl[email protected] (C. Englund), [email protected] (A. Verikas).
1
Tel.: +46 35 167140; fax: +46 35 120348.

0360-8352/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cie.2008.01.019
 C. Englund, A. Verikas / Computers & Industrial Engineering 55 (2008) 592–605 593

Fig. 1. (Left): A four colour print. (Right): An enlarged view of a small area of the print depicted to the left. (For interpretation of the
references in colour in this figure legend, the reader is referred to the web version of this article.)

appears. The plates are mounted on the plate cylinders of the inking system of the printing press. Fig. 2 pre-
sents an illustration of the inking system. During printing the coating is worn thin. Fortunately printing plates
wear well for several hundreds of thousands of impressions.
To change the ink feed into the inking system during printing, the ink key openings are adjusted. In the
printing press concerned in this work, the ink keys are approximately 4 cm wide and 36 of them can be found
along the paper web. Each such zone is referred to as an ink zone. The ink is fed from the ink tray into the
inking system by the ink fountain roller. The speed of the fountain roller determines the ink feed over the
entire paper web. The ink is smoothen out to a thin film by the ink rollers. From the ink rollers the ink is
transferred to the image portions of the printing plate. Then, from the plate to the blanket cylinder and finally
onto the paper. Before the application of ink a thin layer of water is applied to the printing plate. The water
prevents the ink from depositing in the non-printing areas.

2. Assessing print quality

In this work, print quality is defined as the disparity between the amount of ink in the sample and the ref-
erence print. To assess the amount of ink printed on paper when using a closed loop ink feed control system,
ink density is usually measured by a spectrophotometer in fulltone target areas (areas fully covered by ink), as
those shown in Fig. 3 (above). Observe that such fulltone areas are printed for the measuring purpose only.
However, in offset lithographic printing a colour picture consists of dots of varying size (halftone printing).
Thus, the amount of ink printed on the paper depends on both the ink density and the dot size. Therefore,
the so called double grey bar, made of halftone dots, illustrated in Fig. 3 (below), is the measuring target uti-
lized in this work. The double grey bar consists of two parts, one part is printed using the black ink and the
other part using the cyan, magenta, and yellow inks. We use a colour charge-coupled device (CCD) camera to
record a Red, Green and Blue, (RGB), image of the printed result. Using the neural networks based technique
(Verikas, Malmqvist, Malmqvist, & Bergman, 1999), the average RGB values recorded from the two parts of
the grey bar are transformed to the estimated amount of C, M, Y, and K inks. The C, M, and Y values are
estimated from the RGB values measured from the ‘‘coloured” part of the grey bar while the K ink is obtained
from the RGB values measured from the ‘‘black” part of the bar. The C, M, Y, or K value equal to 0 corre-
sponds to an area without ink (white paper) while the value 100 means that 100% of the area is covered by the

Distributor Paper path

Inking system rollers
Ink
Ink rollers
fountain Blanket Blanket
roller cylinder cylinder
Plate
Ink-keys cylinder

Dampening
mouthpiece Dampening
rollers
Dampening system

Fig. 2. Illustration of the inking system.

594 C. Englund, A. Verikas / Computers & Industrial Engineering 55 (2008) 592–605

Fig. 3. (Above): An example of the fulltone target areas. (Below): An example of the double grey bar.

ink of a given density. Double grey bars are widely used in printing industry in Scandinavia to visually assess
the ‘‘grey balance” between the two parts of the bars. However, incorporating double grey bars in modelling
and control of the offset printing process, as proposed in this paper, is unique.

3. Related work

In the past 10 years little work has been carried out to build systems for ink feed control in the newspaper
printing industry. Marszalec, Heikkila, Juhola, and Lethonen (1999) review a number of colour measuring
systems used in newspaper printing. The authors conclude that no online system for colour control has been
installed. To our knowledge, still no industrial system has been installed for online ink flow control in the
newspaper printing industry using any other measuring targets than the fulltone areas. Ink density, usually
measured by a densitometer or a spectrophotometer, is the parameter controlled Pope and Sweeney (2000).
Previous works however, have shown that it is possible to determine the ink density from RGB camera mea-
surements (Brydges, Deppner, Kunli, Heuberger, & Hersch, 1998; Seymour, 1995; Södergård & Launonen,
1996; Verikas & Bacauskiene, 2008; Xiaohan, Södergård, & Ylä-Jääski, 1993).
Amongst the advantages of using colour camera-based machine vision for printing process control are: the
ability to automatically find the measuring target area, the ability to store images of the print that can be ana-
lyzed off line, the ability to detect defects of the print, and the possibility, besides the amount of ink, to esti-
mate other quality parameters from the images.
Machine vision has been utilized in several of the aforementioned studies. Colour cameras have been used
for both finding the measuring target areas (Pope & Sweeney, 2000) and measuring the ink density (Brydges
et al., 1998; Seymour, 1995; Södergård & Launonen, 1996; Verikas & Bacauskiene, 2008; Xiaohan et al.,
1993). To our knowledge there have been no attempts to determine the amount of ink in halftone colour pic-
tures online.

4. Process parameters

In this work, we incorporate printing process models, built from historical process data, into a controller
aiming at controlling the ink keys in the offset lithographic printing press. Both inverse models, where the ink
 C. Englund, A. Verikas / Computers & Industrial Engineering 55 (2008) 592–605 595

key opening value constitutes the model output and direct models, where the printed amount of ink constitutes
the output, are built. The process parameters used to model one colour in the offset printing process are given
below. Depending on the modelling task, inverse or direct, different combinations of these parameters are
utilized.

x1 – Printing speed in copies per hour.

x2 – Ink fountain roller speed.
x3 – Ink temperature. The temperature of the ink in the ink tray. The temperature affects the viscosity of the
ink. The higher the temperature the lower the viscosity – the easier does the ink flow through the inking
system.
x4 – Ink level in the ink tray. The ink level affects the ink feed. The higher the level, the higher is the pressure
at the opening where the ink is released.
x5,6,7 – Estimated ink demand for the current, adjacent to the left, and to the right ink zone, respectively.
The ink demand equals to the percentage of the area covered by ink in the corresponding ink zone.
x8,9,10 – Ink key opening for the current, adjacent to the left, and to the right ink zone, respectively – is the
signal controlling the amount of ink dispersed on the paper.
y – Amount of ink of a specific colour estimated from the double grey-bar.

In the direct modelling, y is the model output. However, for inverse modelling, where the modelling task is
to predict the ink key opening, the y value is used as an input parameter, while the parameter x8 constitutes the
model output.
The parameters x8 and y are used from both the current time step (t) and the previous (t  1). Experimental
studies have shown that no further performance gain is achieved by incorporating more time steps e.g. (t  2)
or (t  3). The parameters x5, x6 and x7 describes the ink demand in the current zone (x5) and the two adjacent
(x6, x7) zones. Since ink flows between adjacent zones in the printing press, in particular when the difference
between the parameters is large, the parameters x5, x6, x7 are replaced by their mean x5;6;7 in the models. Fur-
thermore, the level of ink in the ink tray (x4), is automatically kept constant by the printing process control
system and can, therefore, be disregarded in the process models. The process model variables can be found in
Table 1.
Note that the variables given in the Table 1 are used when training the process models. In the operating
mode, when the models are used for control the following changes are made: the inverse model variable xi5
is given by the current measured amount of ink, the variable xi6 is replaced by the desired amount of ink,
the variables xi7 ; xi8 and xi9 come from the current time step (t), and the variable yi is the next control signal
y(t + 1). The direct model predicts the expected amount of ink yd, given the current amount of ink xd9 , the pre-
dicted ink key opening from the inverse model yi (as xd6 ), the current ink key opening xd5 , the current ink key
opening from the adjacent ink zones xd7 and xd8 , and the other variables xd1 ; . . . ; xd4 .

Table 1
Specification of the inverse and direct model variables
Inverse model variables Process parameters variables Direct model Process parameters
(x, y)i (x, y)d
xi1 ! x1(t) xd1 ! x1(t)
xi2 ! x2(t) xd2 ! x2(t)
xi3 ! x3(t) xd3 ! x3(t)
xi4 ! x5;6;7 ðtÞ xd4 ! x5;6;7 ðtÞ
xi5 ! y(t  1) xd5 ! x8(t  1)
xi6 ! y(t) xd6 ! x8(t)
xi7 ! x9(t  1) xd7 ! x9(t  1)
xi8 ! x10(t  1) xd8 ! x10(t  1)
xi9 ! x8(t  1) xd9 ! y(t  1)
yi ! x8(t) yd ! y(t)
596 C. Englund, A. Verikas / Computers & Industrial Engineering 55 (2008) 592–605

5. Data acquisition system

To obtain all parameters necessary for the modelling, a web offset printing press was equipped with an online
press monitoring system, the schematic view of which is shown in Fig. 4. The system consists of a synchronization
computer, a log computer, a colour camera for grey bar image capturing, a linear transmission unit for traversing
the camera across the paper web, a data base, ink temperature sensors, and a bar code reader placed at the reel
stand. The log computer functions as the master unit and has connections to all the monitoring equipment,
whereas the synchronization computer controls and synchronizes the data sampling process and controls the lin-
ear transmission system that traverses the camera taking snapshots of the print over the paper web. The log com-
puter merges the information extracted from the image with the current printing press status, the ink temperature,
the details from the reel stand, and stores all the information in the data base.
There are 36 ink zones along the paper web. The same camera is used to obtain data from all the ink zones.
The time needed to traverse the camera once over the paper web, to take an image of each of the 36 double
grey bars and then return to the initial position is approximately 100 s. It takes approximately 60 s for an ink
key adjustment to come into affect. In the experimental investigations section, we often refer to a sample num-
ber. The time duration between two consecutive samples is approximately 100 s.

6. Methods

Model-based control systems are common in industry because process models have the ability to mimic the
process under consideration, both its direct and inverse behaviour. In this application, we also use the model-
based approach.

6.1. Printing process modelling

Due to wear of the printing press, the process can be classified as having slow variation. Moreover, depend-
ing on a printing job, the time the process stays in a predefined part of the input space may vary significantly,
from minutes to several days. If the process starts to operate in a new region of the input space, different from

Ink
Camera temperature
sensors
Paper path

Motorised
linear guide

Sync Log
Computer Computer

Data base
Data collection system

Printing press
control system Reel stand

Fig. 4. A schematic illustration of the data acquisition system.

 C. Englund, A. Verikas / Computers & Industrial Engineering 55 (2008) 592–605 597

the training region, the model performance may deteriorate significantly. To handle such situations, an adap-
tive data mining and modelling approach (Englund & Verikas, 2007) was recently proposed. The data mining
technique monitors the process data and keeps an up to date data set of a reasonable size characterizing the
process. This data set is used to build process models. The adaptive modelling is aiming at building models of
optimal complexity. Starting with a linear model, a number of models of increasing complexity (MLP with an
increasing number of hidden units) are built and the model with the lowest generalization error is selected.
Since during the process run the data set is automatically updated, there can arise a need to update the process
models as well. The need to update the models is automatically detected and the models are retrained. In this
work, we use this technique to create and update the process models.

6.2. Types of models used

Models of four types are used in this work for modelling the printing process.

 A model specific for each ink key/zone. These models are called specialized, since they have specific knowl-
edge about a certain ink key/zone. Each specialized model is trained using data from a specific ink zone.
The complexity of the specialized model is dependent on the data and the model providing the lowest gen-
eralization error is automatically selected from the model pool, as described in Section 6.3. The model can
be linear or non-linear (neural network based).
 A committee of specific (specialized) models – models chosen for each ink key according to the description
given in the previous paragraph. Specialized models implementing similar functions are aggregated into a
committee. We use the approach developed in (Englund & Verikas, 2005) for building committees of mod-
els, where both the number of members and the aggregation weights of the members are data dependent.
Thus, specialized models included into the committee are specific for each data point.
 A non-linear general model (a neural network) built using data from all the K ink-zones. The general model
is built using more data than the specialized one and therefore it generalizes better than the specialized
model. The complexity of the general model is dependent on the data and the most appropriate model
is automatically selected from the model pool, as described in Section 6.3.
 A linear general model built using data from all the K ink zones.

The specialized models and committees of the models provide the highest modelling accuracy. However,
due to the limited training data set used, the models may run into generalization problems. In such situations,
general models are used instead, which are built using much more data points than the specialized ones. How
the generalization performance is assessed will be described in the next section. Since the complexity of the
models is determined automatically, the general model may be linear or non-linear. If a non-linear general
model is automatically selected, a linear general model is also built. The linear general model exhibits the low-
est modelling accuracy, however the best generalization ability.

6.3. Selecting models from a model pool

The linear model is given by

Y ¼ Xh þ  ð1Þ
where X is the input data matrix,  is the vector of random errors assumed to be identically distributed with
zero mean and unknown variance r2, h is the vector of regression parameters, and Y is the predicted N  1
output vector obtained from the regression model. Given the input xi, the well known Tukey’s ‘‘hat” matrix
H = X(XTX)1XT can be used to determine the variance
s2 ¼ r
b 2 hii ð2Þ
of the model output f(xi, h), where hii is the diagonal element of the matrix H and r b 2 is the estimate of r2.
For a non-linear model, we calculate the hii values from the orthogonal projection matrix Z(ZTZ)1 ZT,
namely hii ¼ zTi ðZT ZÞ1 zi , i = 1, . . ., N, where Z = [z1 z2, . . ., zN]T is the Jacobian matrix with
598 C. Englund, A. Verikas / Computers & Industrial Engineering 55 (2008) 592–605

of ðxi ; hÞ
zi ¼ ð3Þ
oh h¼hLS
where N is the number of data points, hLS is the least-squares estimate of the parameter vector h of the process
model network, and the N-vector f(X,h) = [f(x1,h), . . ., f(xN,h)]T stands for the process model outputs.
To find the appropriate models, we exploited the analysis of leverages as suggested in (Monari & Dreyfus,
2002). We trained a series of models (one hidden layer perceptrons) – starting from a linear one – with an increas-
ing number of hidden units until the condition 0 6 hii < 0.9 was steadily violated. For each model, training was
repeated 10 times starting from different initial weights. Models not satisfying the relation 0 6 hii < 0.9 were dis-
carded and the approved ones were evaluated by calculating the approximate leave-one-out error
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N  2
u1 X ri
ELOO ¼ t ð4Þ
N i¼1 1  hii

where ri = yi  f(xi,hLS) is the ith residual, and yi is the target.

The distribution of hii values gives an indication on how well the training was made. The more even the
distribution, the less over-fitting was made during the training. It was suggested to use the following quality
measure to assess the distribution (Monari & Dreyfus, 2002).
sffiffiffiffiffiffiffiffiffi
1 X N
N
l¼ hii ð5Þ
N i¼1 q

where q is the number of adjustable model parameters. Since (Monari & Dreyfus, 2002)
(
l61
ð6Þ
l ¼ 1 () ðhii ¼ Nq Þ 8i 2 ½1 . . . N 

the closer l is to unity, the more even is the influence of the data points on the model. Thus, amongst the mod-
els exhibiting approximately the same values of ELOO, the model with the largest l value is selected.

6.4. Process controller

Since we use multiple models to model the printing process, the multiple model approach has also been
adopted for controlling the process.
Multiple models-based approach has shown a great success in several industrial control applications. Chen
and Narendra (2001) suggest to combine linear and non-linear models in order to assure stability of the non-
linear dynamic system, by assuring a bounded control signal, and to gain performance in the system. Exper-
iments show that both stability and improved performance of the controlled system can be achieved simulta-
neously by using multiple models-based controller (Chen & Narendra, 2001). In Ravindranathan and Leitch
(1999), how different models incorporate different amount of scope, precision and generality and how the use
of multiple models can improve performance of the control system are discussed. Yu (2006) has recently pro-
posed a multiple model-based controller to deal with processes that operates in different environments and
where conditions change frequently. It is proposed to select among a number of different models both for
plant identification and for the inverse controller. The selection is based on a performance index that is found
from the identification error for the inverse model.
Fig. 5 illustrates the proposed controller configuration, where IM stands for inverse model and DM means
direct model. Four aforementioned model types are used and the model showing the best performance index is
selected. Models incorporated in the control configuration are:

Sing – A single specialized model.

Com – A committee of specialized models.
LGen – A single general linear model.
NLGen – A single general non-linear model.
 C. Englund, A. Verikas / Computers & Industrial Engineering 55 (2008) 592–605 599

xi IM Sing DM Sing
S xd S
xi IM Com i d DM Com
y x
yd
x i
IM LGen xd DM LGen
xd
xi IMNLGen DM NLGen

Fig. 5. The proposed control configuration.

The control configuration functions as follows. The control signal u is given by the output of one of the
inverse models. We assume that the inverse model output is normally distributed with the mean given by
the model output and the standard deviation s. A large standard deviation of the predicted control signal indi-
cates model uncertainty.
By sampling from the distribution of the inverse model output, as suggested in (Herzallah & Lowe, 2004),
we produce a set of control samples U = [u1u2, . . ., uD] that are then evaluated using the direct model, see
Fig. 5. The number of samples D is determined by the model standard deviation s. The larger the s the more
samples are generated.
The output of the inverse model yi and the direct model yd are given by
y i ¼ f i ðxi ; hi Þ ð7Þ
d d d d
y ¼ f ðx ; h Þ ð8Þ
where h is the model parameter vector and the functions f are either linear or non-linear.
The control signals ui1ui2, . . ., uiD generated by each of the inverse models, Sing, Com, LGen, and NLGen
(i = 1, . . ., 4) are used to calculate the direct model outputs y d11 y d21 ; . . . y d41 ; . . . ; y d4D . The output y dij is given by
y dij ¼ fid ðxdij ; hd Þ ð9Þ
where, i = 1, . . ., 4 refers to a model. The model selected is that minimizing the error eij, the difference between
the output of the direct model y dij and the target (the desired amount of ink) y: eij ¼ ky dij  yk. Having all eij, the
indices p, q of the control signal upq sent to the plant are found as follows:
p; q ¼ arg min eij ð10Þ
i;j

where p = 1, 2, 3, 4 refers to a specialized model, a committee, a non-linear general model, and a linear
general model, respectively. For example, if the committee model exhibits the lowest error amongst all the
four models and the error is achieved for the third element in the set of control samples U, then p = 2 and
q = 3.
If for a given x epq > b and p – 3, the linear general model is used to avoid using non-linear models with a
large prediction error.

7. Experimental investigations

The experiments have been made during normal production at the offset printing shop. Since noise is
always present in the process output, to avoid oscillatory control actions, no adjustments were made if the
discrepancy between the desired and measured amount of ink was less than the standard deviation of the mea-
sured amount of ink evaluated for constant process parameter values. This criterion was found experimentally
and worked well in all the tests. A graph showing the measured amount of magenta ink where the ink key
opening, the printing speed, and the ink temperature were constant is given in Fig. 6. We found that the stan-
dard deviation of the measured amount of ink was 2.
There are a number of typical scenarios to test in order to verify the effectiveness of the proposed technique.
To show how the controller adjusts the ink keys to achieve the desired amount of ink in the printed result, the
printing speed, the temperature, the ink demand, and the desired amount of ink was varied.
600 C. Englund, A. Verikas / Computers & Industrial Engineering 55 (2008) 592–605

50

40

Relative units
30

20
Measured amount of magenta ink
10 Ink key opening
Printing speed x (1/1000)
Temperature
0
10 20 30 40 50 60
Sample number

Fig. 6. The variation of the amount of magenta ink, while the other process parameters are kept constant. (For interpretation of the
references in colour in this figure legend, the reader is referred to the web version of this article.)

7.1. Varying the printing speed

The ink transfer rate in the inking system depends on the printing speed. The higher the printing speed the
more ink is required to maintain the desired amount of ink printed on the paper. The revolution speed of the
cylinders in the inking system is the same as the printing speed, except for the ink fountain roller. The revo-
lution speed of this roller is unique for each colour and is found from a lookup table. The lookup table values
for the press used in the experiments are plotted in Fig. 7. However, the revolution speed of the ink fountain
roller may not always compensate for the higher ink consumption caused by the increased printing speed.
Therefore, there is a need to adjust the ink key opening when varying the printing speed. Fig. 8 shows the
results of the test for cyan and yellow inks. The task is to maintain the amount of cyan, magenta, and yellow
inks on the paper at the target level when the printing speed is changed. The results for magenta ink were very
similar two those obtained for yellow and, therefore, are not shown in Fig. 8. During the test, the printing
speed is changed from 17,000 copies/hour to 20,000 and then to 26,000. The ink key level can vary between
0 (ink key opening is closed) to 100. The initial ink key level depends mainly on the ink demand. Therefore, the
ink demand values are also shown in Fig. 8. As it can be seen, the controller is able to maintain the amount of
ink at the desired level.

7.2. Varying the temperature

Ink temperature highly affects the printing process. The viscosity of the ink decreases as the temperature
increases. Hence, the higher the temperature the easier does the ink flow through the printing press. Typically,
50
[%] of maximum revolution speed

Cyan
Magenta
40 Yellow
Black
30

20

10

0
0 5000 10000 15000 20000 25000 30000
Printing speed [copies/hour]

Fig. 7. The revolution speed of the ink fountain rollers, expressed in percent of the maximum ink fountain roller speed, and as a function
of the printing speed.
 C. Englund, A. Verikas / Computers & Industrial Engineering 55 (2008) 592–605 601

100 Cyan ink key level Ink demand = 50% 100 Yellow ink key level Ink demand = 38%
Printing speed x (1/1000) Printing speed x (1/1000)
50 50

0 0
5 10 15 20 5 10 15 20
Sample number Sample number

60 60
Measured amount of cyan ink Measured amount of yellow ink
50 Target amount of ink 50 Target amount of ink

40 40

30 30
5 10 15 20 5 10 15 20
Sample number Sample number

Fig. 8. Results from controlling the amount of cyan (left) and yellow ink (right) as the printing speed changes. (For interpretation of the
references in colour in this figure legend, the reader is referred to the web version of this article.)

the ink key opening needs to be adjusted as the temperature changes during the process run. Given that all the
process parameters are constant, except the temperature, the printed amount of ink will increase along with
the temperature.
An experiment was conducted, where the printing press has not been used for some time before starting the
experiment. Thus, the printing process started at low ink temperature, which increased during the process run.
The results of the test for cyan and magenta inks controlled in two chosen ink zones are shown in Fig. 9. As
can be seen in Fig. 9, the initial ink key openings are too low and are raised by the controller until the desired
amount of ink is achieved. The initial ink key openings values are usually calculated for a ‘‘warm” press.
Therefore, the values were too low for the ‘‘cold” press. After some time, the ink temperature increases,
the ink flows more easily, and consequently, the ink key openings are lowered. The temperature of the ink
in the ink tray increased approximately 3 degrees during the experiment.

7.3. Varying ink demand

It is important to utilize proper control actions at various ink demands. The ink demand can vary signif-
icantly between adjacent ink zones and between printing jobs. This variable has a high positive correlation
with the ink key opening. A large value of the ink demand means that a large area of paper is to be covered
with ink in the corresponding ink zone. It is common that the initial ink key settings are not optimal and need
Ink key opening

Ink key opening

Temperature

Temperature

50 Cyan ink key Ink temperature 26 50 Magenta ink key Ink temperature 26

40 24 40 24
22 22
30 30
20 20
20 20
5 10 15 20 25 30 35 40 5 10 15 20 25 30 35 40
Sample number Sample number

50 50
Measured amount of cyan ink Measured amount of magenta ink
40 Target amount of ink 40 Target amount of ink

30 30

20 20
5 10 15 20 25 30 35 40 5 10 15 20 25 30 35 40
Sample number Sample number

Fig. 9. Result from controlling the amount of cyan (left) and magenta ink (right) as the temperature increases. (For interpretation of the
references in colour in this figure legend, the reader is referred to the web version of this article.)
602 C. Englund, A. Verikas / Computers & Industrial Engineering 55 (2008) 592–605

80 80
Cyan ink key level Ink demand = 7% Cyan ink key level Ink demand = 24%
60 60
40 40
20 20
0 0
10 20 30 40 10 20 30 40
Sample number Sample number

60 60
Measured amount of cyan ink Measured amount of cyan ink
Target amount of ink Target amount of ink
40 40

20 20
10 20 30 40 10 20 30 40
Sample number Sample number

80 80
Magenta ink key level Ink demand = 45% Magenta ink key level Ink demand = 65%
60 60
40 40
20 20
0 0
5 10 15 20 25 5 10 15 20 25
Sample number Sample number

60 60
Measured amount of magenta ink Measured amount of magenta ink
Target amount of ink Target amount of ink
40 40

20 20
5 10 15 20 25 5 10 15 20 25
Sample number Sample number

Fig. 10. Results from controlling the cyan and magenta ink keys for two different ink demand values. (For interpretation of the references
in colour in this figure legend, the reader is referred to the web version of this article.)

to be adjusted during the process run. The results of controlling the cyan and the magenta ink keys of two
different ink zones are found in Fig. 10. There is a different ink demand in all the four cases and the controller
is able to drive the output to the desired level. The control actions start when the line of the target amount of
ink appears.

7.4. Varying the desired amount of ink

During normal production at the printing shop, numerous printing jobs are run each day. Different custom-
ers may demand different basic inking level – different amount of ink in the printed result. This implies that the
controller is to be able to drive the output to a different amount of ink. Fig. 11 presents the results of the exper-
iment where the target amount of ink is changed after some time during a print job. Normally, the target
amount of ink varies from approximately 30 to 45 depending on the job. The target amount ink is shown
by a solid line. In this experiment, the target is changed from 38 to 43 for the cyan and magenta inks and from
41 to 46 for the yellow ink. As it can be seen, the controller is able to drive the output to the new targets.

8. Discussion

One can wonder, if models of all the four types are utilized by the control system. In Fig. 12, the employ-
ment frequency of the different models is presented. The histogram is made for the ink data extracted from two
different printing jobs. For each sample processed the model exhibiting the lowest error between the predicted
and the target amount of ink is selected and, as it can be seen from Fig. 12, models of all the four types are
 C. Englund, A. Verikas / Computers & Industrial Engineering 55 (2008) 592–605 603

80 80
Cyan ink key level Ink demand = 19% Cyan ink key level Ink demand = 22%
60 60
40 40
20 20
0 0
5 10 15 20 25 5 10 15 20 25
Sample number Sample number

60 Measured amount of cyan ink 60 Measured amount of cyan ink

Target amount of ink Target amount of ink
50 50
40 40
30 30
5 10 15 20 25 5 10 15 20 25
Sample number Sample number

80 80
Magenta ink key level Ink demand = 17% Magenta ink key level Ink demand = 12%
60 60
40 40
20 20
0 0
5 10 15 20 25 30 5 10 15 20 25 30
Sample number Sample number

60 Measured amount of magenta ink 60 Measured amount of magenta ink

Target amount of ink Target amount of ink
50 50
40 40
30 30
5 10 15 20 25 30 5 10 15 20 25 30
Sample number Sample number

80 80
Yellow ink key level Ink demand = 18% Yellow ink key level Ink demand = 11%
60 60
40 40
20 20
0 0
5 10 15 20 25 30 5 10 15 20 25 30
Sample number Sample number

60 Measured amount of yellow ink 60 Measured amount of yellow ink

Target amount of ink Target amount of ink
50 50
40 40
30 30
5 10 15 20 25 30 5 10 15 20 25 30
Sample number Sample number

Fig. 11. Results of controlling the cyan, magenta, and yellow ink keys for various ink demand and target levels. (For interpretation of the
references in colour in this figure legend, the reader is referred to the web version of this article.)

utilized to control the process. Thus, if the models have zero bias and approximately the same variance, we can
only expect worse performance of the controller if one or several of the models are eliminated.
As it was mentioned in Section 5, the sampling time is approximately 100 s and the time for a control action
to come into affect is approximately 60 s. The low sampling frequency is due to the large number of ink zones
(36) sampled by the same camera. The computation time needed for the multiple models-based controller to
calculate the control signals for the four inks is approximately 1 second. Consequently, there is no lack of time
to estimate the control signal.
604 C. Englund, A. Verikas / Computers & Industrial Engineering 55 (2008) 592–605

50

40

[%] of data samples

30

20

10

0
Sing Com LGen NLGen

Fig. 12. The employment frequency of the different models during cyan, magenta, and yellow ink feed control. (For interpretation of the
references in colour in this figure legend, the reader is referred to the web version of this article.)

9. Conclusions

In this work an approach to utilizing colour camera measurements made on halftone colour pictures along
with printing process parameters to build a control system for ink flow control in the offset lithographic print-
ing process was developed. The multiple model-based controller for ink flow control was presented. Four cou-
ples of direct and inverse process models are used to find the most suitable control signal at the time moment.
The approach has been tested during normal production at the printing shop in Halmstad, Sweden. The exper-
imental results from a number of typical scenarios, where the printing speed, the temperature, the ink demand,
and the target amount of ink printed on paper are varied, show that the ink keys can be appropriately adjusted
to achieve the desired amount of ink. Furthermore, the employment frequency of the different models dem-
onstrates that all the models are utilized to control the process.

References

Brydges, D., Deppner, F., Kunli, H., Heuberger, K., Hersch, R. (1998). Application of a 3-CCD color camera for colorimetric and
densitometric measures. In: SPIE Proceedings. Vol. 3300. pp. 292–301.
Chen, L., & Narendra, K. S. (2001). Nonlinear adaptive control using neural networks and multiple models. Proceedings of the American
control conference (Vol. 6, pp. 4199–4203). Chicago, Illinois: IEEE.
Englund, C., & Verikas, A. (2005). A SOM based model combination strategy. In J. Wang, X. Liao, & Z. Yi (Eds.). Lecture notes in
computer science (Vol. 3496 of Part 1, pp. 461–466). Springer-Verlag.
Englund, C., & Verikas, A. (2007). A SOM based data mining strategy for adaptive modelling of the offset lithographic printing process.
Engineering Applications of Artificial Intelligence, 20(3), 391–400.
Herzallah, R., & Lowe, D. (2004). A mixture density network approach to modelling and exploiting uncertainty in nonlinear control
problems. Engineering Applications of Artificial Intelligence, 17, 145–158.
Marszalec, E., Heikkila, I., Juhola, H., & Lethonen, T. (1999). On-line devices and measuring systems for the automatic control of
newspaper printing. SPIE, 3826, 304–312.
Monari, G., & Dreyfus, B. (2002). Local overfitting control via leverages. Neural Computation, 14(6), 1481–1506.
Pope, B., Sweeney, J. (2000). Performance of an online closed-loop color control system. In: TAGA 2000 Proceedings. San Diego, USA,
pp. 417–431.
Ravindranathan, M., & Leitch, R. (1999). Model switching in intelligent control systems. Artificial Intelligence in Engineering(13),
175–187.
Södergård, C., & Launonen, R. (1996). Inspection of colour printing quality. International Journal of Pattern Recognition and Artificial
Intelligence, 10(2), 115–128.
Seymour, J. (1995). The why and the how of video-based on-line densitometry. In Proceedings of the IS& T’s 4th technical symposium on
prepress, proofing & printing. pp. 23–28.
Verikas, A., & Bacauskiene, M. (2008). Estimating ink density from colour camera RGB values by the local kernel ridge regression.
Engineering Applications of Artificial Intelligence, 21(1), 35–42.
Verikas, A., Malmqvist, K., Malmqvist, L., & Bergman, L. (1999). A new method for colour measurements in graphic arts. Color Research
& Application, 24(3), 185–196.
 C. Englund, A. Verikas / Computers & Industrial Engineering 55 (2008) 592–605 605

Xiaohan, Y., Södergård, C., & Ylä-Jääski, J. (1993). On-line control of the colour printing quality by image processing. In Proceedings of
the 1993 IEEE region 10 conference on computer, communication, control and power engineering (TENCON’93) (pp. 1039–1041).
Beijing, China: IEEE.
Yu, W. (2006). Multiple recurrent neural networks for stable adaptive control. Neurocomputing, 70, 430–444.