DEVELOPMENTS IN

COMPUTER-AIDED

DRYER SELECTION

C. G. J. Baker and H. M. S. Lababidi

Chemical Engineering Department, Kuwait University,

P. O. Box 5969 – Safat, 13060 Kuwait

ABSTRACT

This paper describes recent advances in the development of
a fuzzy expert system for food dryer selection. An earlier
version, which was restricted to batch dryers, has now
been extended to include continuous dryers. The modular
approach originally proposed by the present authors was
adopted. The current implementation of the system includes
three knowledge bases: the mode (batch-continuous) selector,
the batch-dryer selector, and the continuous-dryer selector. A
blackboard architecture was used to facilitate full data inter-
change between the three knowledge bases. A user interface
and a scheduler were developed to automate the system.
Examples of ancillary programs (design, costing, help, appli-
cations) have also been developed. Satisfactory predictions
were obtained using the selection algorithm. Typical examples
are presented in case studies.

Key Words

:

Dryer selection; Expert systems; Fuzzy logic;

System development; Continuous dryers; Food

1851

Copyright & 2001 by Marcel Dekker, Inc.

www.dekker.com

DRYING TECHNOLOGY, 19(8), 1851–1873 (2001)

1852

BAKER AND LABABIDI

INTRODUCTION

Dryer selection involves an interplay between a relatively large number

of factors, both technical and economic. The chosen dryer must primarily be
capable of performing the required duty in terms of moisture removal,
throughput, and feed-handling capability. In certain applications, such as
in the drying of biomaterials, for example, it may also be required to develop
other desirable product-quality features – e.g. functionality, taste, and color.
These stringent technical requirements should naturally be met at the lowest
possible cost.

Until recently, its complexity has restricted dryer selection to

‘‘experts’’. However, computer-based tools are now available that can be
employed to at least partially de-skill this process. Most are based on con-
ventional expert systems (see Nevenkin and Chavdarov, 1992; Kemp and
Bahu, 1995; Kemp, 1998; Matasov et al., 1998). Other workers, such as
Kraslawski et al. (1999) have employed case-based reasoning as an alter-
native approach.

Baker and Lababidi (1998) presented their initial ideas on the features

that should be incorporated into a computerized dryer selection algorithm.
They proposed the use of a modular fuzzy expert system for this purpose
and outlined a knowledge base that could be used for the selection of batch
dryers for foodstuffs. This aspect of the work was further developed and
refined and was subsequently published elsewhere (Lababidi and Baker,
1999: Baker and Lababidi, 2000). The key objectives of the work described
in the present paper are:

1.

To design a continuous-dryer selection algorithm for food
products.

2.

To integrate the new knowledge base with the earlier batch-dryer
selection algorithm.

3.

To develop a variety of ‘‘foreign’’ programs that will provide the
user with additional information to aid the selection process.

4.

To automate the selection process and to provide a user interface
to aid the data input and output processes.

COMPUTER-BASED DRYER

SELECTION ALGORITHMS

Baker and Lababidi (2000) described a structured technique for dryer

selection. The approach is iterative and involves the following steps: draw-
ing up the process specifications, making a preliminary selection, planning

COMPUTER-AIDED DRYER SELECTION

1853

and conducting bench-scale tests, making an economic comparison of
alternatives, conducting pilot-scale trials, and, finally, selecting the most
appropriate dryer type. As noted by these authors, the principal value of
computer-based algorithms is in the preliminary selection stage in which it is
necessary to screen a relatively large number of dryer types and sub-types.
They cannot, however, provide a substitute for bench- or pilot-scale tests,
which can yield vital information not only on the drying kinetics but also
on the materials handling characteristics, which, in practice, are equally
important.

Computer-based methods that can be used for dryer selection include

expert systems, fuzzy logic, and neural networks. In the case of expert
systems, the selection decisions are based on a series of rules formulated
by the specialist. These are generally quantitative and inflexible. Combining
fuzzy logic with an expert system results in a more flexible knowledge base
reasoning system, in which the selection qualifiers are represented as linguis-
tic variables (e.g. temperature ¼ high, low, very low, as opposed to numeric
values). Internally the system transforms the knowledge into fuzzy repre-
sentation, performs the reasoning process, and finally translates the results
into the appropriate output format. Neural networks vaguely resemble the
vastly more complicated networks of neurons present in the human brain.
They can be ‘‘trained’’ to develop explicit relationships between input and
output variables and update these relationships on the basis of experience.
After an initial assessment of the problem, Baker and Lababidi (2000) con-
cluded that a fuzzy expert system (Zadeh, 1965; Dubois and Prade, 1980)
would provide the most appropriate platform for their selection algorithm.

Their choice was based on the following arguments. Firstly, the selec-

tion process relies on variables that are often difficult to define quantita-
tively. An obvious example is cohesiveness of the feed. For practical
purposes, it is quite adequate to define this on a scale of zero (not sticky)
to one (very sticky). At a later date, it would be relatively easy to convert a
scientific measurement of cohesiveness to an appropriate value on the 0–1
scale for selection purposes. Secondly, as dryer selection is not an exact
science, precise numerical rules are unlikely to be appropriate. Boolean out-
comes (True or False) are rarely adequate. Rather, the terms Possibly or
Probably are more appropriate.

STRUCTURE OF FUZZY EXPERT SYSTEMS

For dryer selection, fuzzy expert systems offer several advantages

over their conventional counterparts. Principal amongst these are their

1854

BAKER AND LABABIDI

ability to handle ill-defined parameters and to deal with the large number of
uncertainties involved in the reasoning process.

A fuzzy expert system consists of two principal components: the

knowledge base and the inference engine. The knowledge base contains
the logic upon which a specific decision by the human expert is based.
The inference engine uses the facts and rules in the knowledge base to
arrive at a conclusion.

Baker and Lababidi (2000) described the structure of fuzzy knowledge

bases in some detail. Briefly, they consist of a number of rules formulated as
If-Then statements:

IF condition (hypothesis, antecedent), THEN conclusion (consequent)

It typically states that if we know a fact (the condition, hypothesis or ante-
cedent) then we can infer or, or derive, another fact called a conclusion (or
consequent). A hypothetical example of a rule is as follows.

IF:
Exposure temperature is Low

THEN:
Freeze Dryer

Confidence ¼ 0.5

and

Horizontal Agitated Dryer

Confidence ¼ 1

Here, ‘‘Exposure temperature’’ is a variable whose value is defined by a
linguistic term (e.g. Very Low, Low, Medium, etc.). Fuzzy expert systems
define a linguistic term as a fuzzy set and evaluate the confidence with which
the variable belongs to this set. This is expressed by a value in the range zero
to one, which is determined by a membership function. This must be defined
by the expert for each fuzzified variable employed in the knowledge base.
Typical examples of membership functions are illustrated in Figure 1; the
numerical values of g

ij

employed in this study were summarized by Baker

and Lababidi (2000). Note that any given variable can belong to more than
one set.

Referring again to the hypothetical rule, the terms ‘‘Freeze Dryer’’ and

‘‘Horizontal Agitated Dryer’’ are known as goals. The inference engine
evaluates each rule to determine whether the condition (or conditions) is
satisfied. If it is, the rule will be ‘‘fired’’ and the inference engine proceeds to
assign an appropriate confidence to each goal. This is determined by com-
bining the confidences associated with each condition with the confidences
of the conclusions. Fuzzy expert systems differ in this respect from their
conventional counterparts, which only take the confidence associated with
each goal into account.

COMPUTER-AIDED DRYER SELECTION

1855

The inference engine tests each rule to determine which conditions are

met. In this manner, it produces a set of multiple recommendations arran-
ged in order of likelihood based on the certainty with which the rules are
satisfied.

STRUCTURE OF THE KNOWLEDGE BASES

Baker and Lababidi (2000) argued that a series of modular knowledge

bases was likely to offer advantages in terms of improved accuracy,

Figure 1.

Examples of typical membership functions.

1856

BAKER AND LABABIDI

flexibility, and adaptability over a single all-encompassing set of rules. They
therefore designed their batch-dryer selection algorithm with this in mind.
The present study tests the viability of this arrangement. Three separate
knowledge bases are employed in the current embodiment of system. The
first (KCommon) tests whether a batch and/or a continuous dryer is appro-
priate for the given application. Either or both of these outcomes is possible;
each is evaluated, together with an associated confidence. The second and
third knowledge bases, KBatch and KContinuous, contain selection rules
for batch and continuous dryers, respectively.

KCommon and KBatch were formulated by splitting the original

knowledge base for batch-dryer selection (Baker and Lababidi, 2000) into
two parts. Some minor modifications to the KCommon membership func-
tions and a few of the rules were required to accommodate the continuous-
dryer selection algorithm. However, these did not affect the batch-dryer
predictions. For example, it was necessary to introduce a fourth linguistic
term (Very High) to describe throughputs exceeding 1000 kg/h. This resulted
in a small number of rule changes but enabled a finer distinction between
various types of continuous dryer to be achieved. KCommon now contains
56rules and KBatch 51. Three warning messages have been included in
KCommon to alert the user to possible conflicts. For example, ‘‘Batch
Recognition and Very High Throughput are Incompatible’’ is displayed
when the user inputs these requirements. However, the system does not
automatically terminate the run when this occurs.

Although the above changes to KCommon and KBatch might appear

trivial, this was far from the case. The new system required that data (e.g.
variables, goals and confidences) be transferred from one knowledge base to
the other. This feature was not supported by the commercial expert system
platform employed in this work and required extensive programming to
achieve. Further details are given in the ‘‘System Implementation’’ section
below.

The third knowledge base, KContinuous, is a new development,

which, in its current version, contains 118 rules. Of these, 5 relate to
freeze dryers, 51 to contact dryers, and 56to convective dryers. The remain-
der are either general or specify whether a direct or indirect heater should be
employed. Figure 2 illustrates the types of continuous dryer included in the
knowledge base, classified in the usual manner.

In order to accommodate very high throughputs of difficult-to-process

(e.g. cohesive) materials, the possibility of feedstock modification was also
included, where appropriate. The methods commonly employed include
backmixing dried product with the original feed material in order to make
it less sticky, and diluting liquid feeds to make them pumpable and atomiz-
able. This enables them to be processed in a spray dryer.

COMPUTER-AIDED DRYER SELECTION

1857

Like KBatch, KContinuous shares most of the required selection data

with KCommon. The variables included in this rule base are essentially the
same as those described by Baker and Lababidi (2000) for batch-dryer
selection. Table 1 lists these variables and their corresponding linguistic
terms. As was the case for batch dryers, the principal variables that affect
the choice of continuous dryer are ‘‘Feed-class’’, ‘‘Dryer-class’’ and
‘‘Product fragility’’. The variable ‘‘Feed-class’’ characterizes the feed as a
solid, a paste, or a slurry/solution. Solid feeds are further classified accord-
ing to their size (powders, granules, or pieces) and cohesiveness (free-flowing
or sticky). The latter was defined using a two-set membership function based
on an arbitrary 0–1 scale. ‘‘Dryer-class’’ is determined by the mode of
operation (batch or continuous) and whether freeze, vacuum, atmospheric
(contact) or convective drying is appropriate. ‘‘Product Fragility’’ is defined
by the user in a similar manner to cohesiveness.

Several other parameters, which can also affect continuous-dryer selec-

tion, are listed in Table 1. For example, consider ‘‘Throughput’’. This is
employed within KCommon to determine batch versus continuous opera-
tion. It also appears in KContinuous to distinguish those dryers that can
handle very high feed rates. The more costly processing methods, such as

Figure 2.

Categories of continuous dryer (after Baker, 1997).

1858

BAKER AND LABABIDI

freeze and dielectric drying, can only be considered for high-priced products.
‘‘Product Value’’ is therefore used as a basis for excluding these options
where they are not economically viable.

Four additional variables have been introduced into KContinuous.

Solid feeds are classified by the user as either ‘‘Not Dusty’’ or ‘‘Dusty’’.
In the latter case, a contact dryer is sometimes preferred to a convective
dryer, as particulate emissions, and hence gas-cleaning costs, are often
lower. With liquid feedstocks, ‘‘Product Shape’’ is defined by the user as
either ‘‘Spherical’’ or ‘‘Irregular’’. The system automatically selects a spray
dryer when a spherical product is required. As noted above, it then goes on
to determine whether the feed can be pumped and atomized. If it cannot,
feed modification is necessary. When a pneumatic dryer is selected, the user
is asked to state whether the moisture is restricted to the surface of the
particles or is distributed throughout their bulk. This information is used
to distinguish between a flash dryer and a Ring dryer. In the former case, the
residence time is very short and the dryer is capable of removing only sur-
face moisture. Ring dryers, on the other hand, have much longer residence
times and are therefore considerably more versatile.

Table 1.

Variables Employed in the Continuous-Dryer Selection Algorithm

Variable

Linguistic Terms

Source

Feed-class

Free-flowing powder

Free-flowing pieces

KCommon

Sticky powder

Sticky pieces

Free-flowing granules

Paste

Sticky granules

Solution/Slurry

Dryer-class

Continuous-freeze

Continuous-atmospheric

KCommon

Continuous-vacuum

Continuous-convective

Fragility

Not fragile

Fragile

KCommon

Throughput

High

Very High

KCommon

Not high

Not very high

Product value

Low

High

KCommon

Moisture

content

Very low

Not very low

KCommon

Low

Not low

Material

Not dusty

Dusty

User defined

Product shape

Spherical

Irregular

User defined

Liquid feed

Pumpable/atomizable

Not pumpable/

Atomizable

User defined

Moisture

Confined to surface

Distributed throughout

bulk

User defined

COMPUTER-AIDED DRYER SELECTION

1859

FOREIGN PROGRAMS

Dryer selection is a multi-faceted exercise, involving information from

a variety of different sources. The user must assimilate and weigh all this
information before arriving at a conclusion. The selection algorithm formu-
lated in this work incorporates not only the knowledge bases on which the
primary selection is based but also several ‘‘foreign’’ programs (see below)
that provide supplementary evidence. Given the fact that dryer selection is
qualitative in nature, quantitative procedures are useful in that they provide
additional confidence in the validity of the recommendations. This is ana-
logous to human experts, who use short-cut methods as well as rigorous
calculations to justify their decisions.

At the time of writing, the foreign programs are being run manually,

that is independently of the automated knowledge bases. However, work to
combine all the knowledge sources into a single integrated system is in hand.

Dryer Design Program

The present embodiment of the algorithm incorporates a design pro-

gram for a well-mixed fluidized bed dryer, the performance of which can be
predicted with reasonable accuracy. The program employed was a custo-
mized version of the FORTRAN code described by Baker (1999) and Baker
et al. (2000). It provides the user with estimates of the required drying time,
the bed area, and the specific energy consumption (kJ per kg moisture
evaporated) and checks that these values lie within the accepted norms.
Input data to this program are provided from two sources. Firstly,
material-specific data are entered via a data file for each product. These
include a unique product identifier, a batch drying curve obtained under
specified conditions, equilibrium moisture content data, the minimum flui-
dizing velocity, and the bulk density of the powder. For present purposes,
the model data for ion exchange resin (Baker, 1999) were employed. Process
information, such as throughput, inlet and outlet moisture contents, and
bed temperature are extracted from the ‘‘blackboard’’ (see below). In this
version of the program, the operating air velocity is taken as three-times the
minimum velocity, and the bed height as a constant 0.25 m.

Unlike other dryer selection programs (e.g. Kemp and Bahu, 1995),

the present algorithm does not require a knowledge of the dryer size
(and hence cost) to generate the ranked list of alternatives. Given the com-
plex nature of drying processes, it follows that the accuracy with which
various dryers can be sized will vary from one type to another. Our
approach therefore avoids the uncertainty involved in such comparisons

1860

BAKER AND LABABIDI

and the possibility of eliminating a viable alternative on the basis of inac-
curate sizing data.

Dryer Costing Program

The algorithm includes an example of a simple costing program, again

for well-mixed fluidized bed dryers. This program was also coded in
FORTRAN. The following equation, derived from data provided by
Quinn (1963), may be used to estimate the purchase cost of a basic dryer.
This value excludes the costs of freight, insulation, installation, assembly
and start-up:

C

o

¼

5438D

Bed

þ

110

ð

1Þ

where C

o

is the 1963 cost in U.S. dollars and D

Bed

is the diameter of the bed

in m. Data available to van’t Land (1991) suggest that the cost of ancillaries
(feeder, blower, etc.) is about twice the cost of the basic dryer. Installation
costs are typically 2.5-times the purchased-equipment cost but this does vary
somewhat with the material of construction. The updated installed cost of
the complete drying plant C is therefore

C ¼

5 5438D

Bed

þ

110

½



J=J

o

ð

Þ

ð

2Þ

Here, J is the current value of the Marshall & Swift equipment cost index
( 1106for the 3rd Quarter of 2000) and J

o



400 its value in the

base year, 1963.

In practice, these and similar equations can only be expected to give a

very approximate estimate of the total cost. Therefore, when comparing
possible alternatives, an accurate quotation should always be obtained
from a reputable dryer manufacturer.

‘‘Help’’ and Case-Based Applications Modules

‘‘Help’’ modules have been written to provide the user with a brief

description of several of the dryers listed in Figure 2. These include
well-mixed and plug flow fluidized bed dryers, vibrofluidized bed dryers,
flash and Ring dryers, rotary dryers, horizontal agitated dryers, and spray
dryers. Each module includes details of the principal features of the dryer,
factors underlying its selection, and published applications.

Finally, an applications database was constructed listing the types

of dryer that have been used in the past to dry specific food products.

COMPUTER-AIDED DRYER SELECTION

1861

The database is searchable and provides further evidence (or otherwise) in
support of a particular selection.

SYSTEM IMPLEMENTATION

The Dryer Selection Expert System (hereinafter known as DrySES)

was implemented as a web-based system. There were two reasons for this.
Firstly, it enabled us to take advantage of the flexibility inherent in the
design of graphical user interfaces, with which all users are undoubtedly
familiar. Secondly, it will enable the authors to publish DrySES on the
Internet, thereby enabling it to be accessed, and hopefully improved, by
experts in the field.

Two principal factors were considered in designing the structure of

DrySES. The first, and more important, was how to handle the large
amount of diverse, qualitative and incomplete knowledge that is involved
in the selection process. The second relates to the need for flexibility
in accommodating future expansions of the system and incorporating
additional features and tasks. Baker and Lababidi (2000) addressed these
considerations and concluded that a knowledge base constructed from a
series of small, compact and independent modules would be superior on
both counts to an all-encompassing knowledge base. In order to use the
various modules effectively, a central mechanism to coordinate communica-
tion between them and to manage the inference process is required. The
blackboard architecture was adopted to fulfil this task. This is based on a
mechanism that allows for the flexible integration of modular sections of
code into a single problem-solving environment. It is analogous to a panel of
experts sitting around a blackboard. Each is able to read everything that is
on the board and add something useful to the debate, when appropriate
(Craig, 1988).

The structure proposed and implemented in this work is shown in

Figure 3. It consists of the knowledge bases, the foreign programs, the
blackboard, the scheduler, and the graphical user interface (GUI). Each is
described briefly below. The main advantage of this modular architecture is
that each constituent can be implemented and tested independently. When
this process is complete, the constituents can be linked together to form an
integrated system capable of performing the specified task. Another impor-
tant feature is flexibility. The system can be expanded or modified with
relative ease simply by adding new knowledge sources or foreign programs.
This is possible with minor modifications to other components of the
system.

1862

BAKER AND LABABIDI

Knowledge Sources

ReSolver (MultiLogic, 1997) was employed as the expert system plat-

form. This is a knowledge base development tool consisting of various
utilities for developing and maintaining the rules, and implementing infe-
rence mechanisms. Its principal attractive feature is that it supports fuzzy
logic amongst other confidence modes.

The knowledge sources consist of the ReSolver knowledge bases and

the ‘‘foreign’’ programs. At the time of writing, the expert system contains
the three knowledge bases, KCommon, KBatch, and KContinuous. As
in the case of the original batch-dryer selection algorithm, backward chain-
ing was employed as the principal inference mechanism to trigger the rules.
Here, goal-driven rules are automatically tested to see if there is another rule
in the knowledge base that will supply the necessary information. Forward
chaining was forced for selected rules to conclude the intermediate facts
necessary to fire the backward-chaining rules. Forward-chaining (data-
driven) rules were tested as they occurred.

The present embodiment of the system incorporates several different

types of foreign program, which were designed to test its capabilities. These

Figure 3.

Structure of the blackboard expert system.

COMPUTER-AIDED DRYER SELECTION

1863

include the well-mixed fluidized bed design and costing programs (coded in
FORTRAN), the HTML-based help module, and the applications database
(HTML and Java). Foreign programs may of course be coded in other
languages, such as Visual Basic and C.

Blackboard

The ‘‘blackboard’’ is a globally accessible working memory. It also acts

as a means of communication between the different knowledge bases and
foreign programs. The items placed on the blackboard are called entries.
These are represented as attribute-value assertion pairs, which are values
assigned either by the user through the graphical user interface or by one or
more of the knowledge sources. Examples of typical blackboard entries are:

THROUGHPUT 1000
DRYERCLASS Continuous-convective (Conf ¼ 1) AND Continuous-
atmospheric (Conf ¼ .533)

The first of these, which indicates that the dryer throughput is 1000 kg/h,
presents data inputted by the user. The second exemplifies information
derived from a knowledge source. It includes not only the outcomes (con-
tinuous-convective and continuous-atmospheric dryers are selected) but the
confidences associated with each.

The main advantage of the proposed blackboard structure is that it

provides an effective means of integrating diverse components of the knowl-
edge sources, without consideration of the programming language or plat-
form. The only requirement is to provide any foreign program with an
interface that transfers the input data from the blackboard and returns
the program results back as entries in the required format.

Unavoidably, certain limitations and difficulties were encountered

when using ReSolver in the present context. The main problem was asso-
ciated with communication between the knowledge sources and the black-
board. This is not a standard ReSolver feature and was achieved by
programming command files using the low-level language employed in the
expert system.

Scheduler

The third component is the ‘‘scheduler’’, which is responsible

for guiding the system-reasoning activities. Its two basic functions are

1864

BAKER AND LABABIDI

event-scheduling and knowledge-scheduling. In the first of these, the sche-
duler selects the triggering events from the blackboard entries and sorts
them according to a given priority list. For example, KCommon will
always be run first. This will be followed by either KBatch or
KContinuous, as appropriate, and a hierarchy of foreign programs. In its
knowledge-scheduling role, it will act upon entries on the blackboard to
identify the most appropriate knowledge source to execute. Hence, the
main function of the scheduler is to monitor the blackboard for changes,
such as the addition of a new entry or the modification of an existing one.

The scheduler has been fully implemented using Java. It performs the

following actions during a typical consultation session:

1.

It initiates the session by presenting the data entry form for the
user to complete.

2.

It activates the KCommon knowledge base, which determines the
Dryer-class (e.g. Batch-vacuum, Continuous-convective). This
decision is translated as a blackboard entry.

3.

It monitors the blackboard entries and initiates the necessary
action. For example, if batch operation is selected, the scheduler
activates KBatch. Similarly, if continuous operation is indicated,
it activates KContinuous. If both modes are selected, both knowl-
edge bases are activated simultaneously, sharing the information
stored on the blackboard.

4.

It captures the decisions made by the knowledge bases and triggers
the Java programs used to create the HTML code, which gener-
ates the various output screens.

5.

If a well-mixed fluidized-bed dryer is selected, it provides the user
with the option of running the design and costing programs.

Graphical User Interface (GUI)

The graphical user interface provides a friendly interaction with the

user. It includes the data entry screens, results and conclusion screens and
other screens to help the user in understanding and interpreting the results.
The GUI is programmed in HTML and JavaScript. The latter is mainly
used to validate data entries.

CASE STUDIES

At the time of writing, the continuous-dryer selection algorithm has

undergone extensive testing; some 65 test runs with satisfactory outcomes

COMPUTER-AIDED DRYER SELECTION

1865

have been performed. Despite this, if the authors’ experience with the batch-
dryer selection algorithm development is repeated, additional modifica-
tions will be required. However, these are likely to be minor and will
probably affect only the confidence values rather than the particular choices
of dryer type.

We will now consider two hypothetical examples illustrating the use of

the selection algorithms. The cases cited involved all three knowledge bases,
KCommon, KContinuous, and KBatch. In addition, the second example
calls upon the design and costing modules.

Example 1

We wish to dry 1000 kg/h of cooked cereal grains. Experience suggests

that they will be moderately sticky when they leave the cooker and that
it is undesirable to dry them above 80



C. The product (at 8% moisture)

is reasonably robust. The grains are ellipsoidal in shape, with a nominal
size of 6 3 mm. The dryer is to form part of a dedicated continuous
production line.

As is frequently the case, the algorithm predicts a number of different

options. In order to gain more confidence in the accuracy of the selection, it
is always worthwhile undertaking a sensitivity analysis before arriving at a
final judgement. This was done in the present case. The following param-
eters were considered likely to have the most influence on dryer selection:
throughput, feed size (because of the ellipsoidal shape of the grains), cohe-
siveness and fragility. Values of the principal input parameters employed are
summarized in Table 2; here Run 1 is the base case.

In all cases, atmospheric operation was selected exclusively with a

confidence of 1. Continuous-mode was also predicted with a confidence of
1. Batch-mode (confidence 0.25) was predicted in Runs 1, 4, 5, and 6. In
Runs 2 and 3, the corresponding confidences were 0 and 0.48, respectively.
No suitable batch dryers were selected by KBatch, however, in any run.

Table 2.

Input Variables for Example 1

Parameter

Run 1 Run 2 Run 3 Run 4 Run 5 Run 6

1.

Throughput, kg/h

1000

1200

800

1000

1000

1000

2.

Feed Size, mm

6663

66

3.

Cohesiveness (Value in range 0–1)

1

1

1

1

0.5

1

3.

Fragility (Value in range 0–1)

0

0

0

0

0

0.5

1866

BAKER AND LABABIDI

The continuous-dryer selections are summarized in Table 3. Dryers

1–4 and 7 are selected in all runs and with reasonable levels of confidence.
The other dryers appear only occasionally and can be disregarded. The
vibrofluidized bed, through-flow band dryer and the cascading rotary
dryer appear the most promising. The vibrofluidized bed dryer might be
eliminated if significant future increases in production are envisaged; the
cascading rotary dryer is probably more appropriate than the band dryer
under these circumstances. However, further technical and economic
evaluation of all three types is recommended.

Example 2

In this example, we wish to select a dryer capable of drying 3000 kg/h

of a relatively ‘‘well behaved’’ food ingredient from a moisture content of
20% down to 6% or less. The powder in question has a mean size of 0.5 mm
and is ‘‘not regarded as being particularly fragile’’. Because of production
difficulties, the feedrate tends to fluctuate by 10% about the design value
and at times the powder entering the dryer can be ‘‘slightly sticky’’. Like
most food products, it is thermally sensitive and the solids should not exceed
70



C during drying. It is not dusty and is wetted by surface moisture only.

Table 3.

System Recommendations for Example 1

Dryer Type

Confidence

Run No.

1

2

3

4

5

6

1.

Vibrofluidized bed dryer

1

0.4

1

0.8

0.75

0.75

2.

Band dryer (through-flow)

1

0.61

0.4

0.7

0.7

3.

Band dryer (cross-flow)

0.4

0.1

0.4

0.60.36

0.36

4.

Band dryer (air impingement)

0.60.2

0.60.2

0.51

0.44

5.

Cascading rotary dryer

0.60.8

0.60.60.58

0.3

6.

Cascading rotary dryer (Short-fall)

0

0

0

0

0

0.1

7.

Rotary steam-tube dryer

0.48

0.4

0.48

0.24

0.42

0.3

8.

Horizontal agitated dryer

0

0

0

0.560

0

9.

Fluidized bed dryer (well-mixed)

0

0

0

0.1

0

0

10.

Pneumatic (ring) dryer,

after feed modification

0

0

0

0.2

0

0

11.

Plate dryer

0

0

0

0

0.30

0

12.

Rotary louvre dryer

0

0

0

0.2

0.3

13.

Tunnel dryer

0

0

0

0

0.2

0

COMPUTER-AIDED DRYER SELECTION

1867

As in the previous example, it is sensible to conduct a sensitivity analy-

sis here too. The following variables should be examined: feedrate (since this
fluctuates), cohesiveness (the feed is sometimes sticky), and fragility (there is
little information about this property). In addition, as it may be necessary to
dry the ingredient down to less than 6% moisture, the effect of this para-
meter on dryer selection should also be considered. Values of the input
variables are summarized in Table 4; Run 1 is again the base case.

Continuous operation is selected in all cases, with convective dryers

marginally preferred to atmospheric contact dryers (confidences 1.00 and
0.80, respectively). Table 5 summarizes the recommendations for the seven
selection runs performed. These include fluidized bed, flash, and rotary
dryers. In each case, more than one possible option is proposed. In Run
1, the base case, the leading contenders are the well-mixed fluidized bed
dryer and the flash dryer. In both cases, the confidence associated with
these selections is unity. The plug flow and vibrofluidized bed dryers are
also selected, but the confidence is much lower. This is the result of the
relatively high values of the feed rate and product moisture content speci-
fied. A flash-, rather than a Ring, dryer is selected because there is no need
to remove bulk moisture. Of the rotary dryers, the steam-tube dryer (con-
fidence ¼ 0.8) is preferred to the other types in which the air velocity would
be much higher. The low air flow requirement is imposed by the need to
maintain the small particles in the drum for a sufficient length of time to
enable them to dry properly.

Runs 2 and 3 indicate that, within the specified range, feed rate has

no effect on dryer selection. However, as shown by Run 4, increasing the
cohesiveness index I

c

from 0 to 0.5 does have a marked effect on the selec-

tion. In line with expectations, the confidences associated with the selection
of the well-mixed fluidized bed and the steam-tube rotary dryers are both
considerably reduced. The well-mixed fluidized bed dryer is more likely to
experience solids-handling problems with sticky feedstocks; in the extreme

Table 4.

Input Variables for Example 2

Parameter

Run 1 Run 2 Run 3 Run 4 Run 5 Run 6Run 7

1.

Throughput, kg/h

3000

3300

2700

3000

3000

3000

3300

2.

Feed size, mm

0.5

0.5

0.5

0.5

0.5

0.5

0.5

3.

Cohesiveness

(value in range 0–1)

0

0

0

0.5

0

0

0.5

4.

Fragility

(value in range 0–1)

0

0

0

0

0.5

0

0.5

5.

Outlet moisture content, %

666665

5

1868

BAKER AND LABABIDI

(I

c

¼

1), the bed may become defluidized. In some cases, however, this prob-

lem can be overcome by mechanical agitation of the bed. The steam-tube
rotary will possibly experience caking on the heat-transfer surfaces leading
to reduced efficiency. Under these conditions, a cascading rotary dryer may
be more suitable provided a very low air velocity can be employed.

The reason underlying the relatively low confidence associated with

the flash-dryer selections (Dryers 6and 7 in Table 5) is less obvious, how-
ever. It is due to the fact that, when the cohesiveness index I

c

¼

0.5, ‘‘Feed-

class’’ belongs to two sets, namely ‘‘Free-flowing Powder’’ and ‘‘Sticky
Powder’’. The confidence associated with membership of each of these
sets is only 0.5. This has the effect of reducing the overall confidence asso-
ciated with each selection. At the particular value of I

c

chosen to test the

effect of cohesiveness, the confidence values for both options are equal and
relatively low, 0.37. Table 6shows the effect of cohesiveness index on the
confidences associated with the selection of Dryers 6and 7. When I

c

¼

0, the

feed can be regarded as a free-flowing powder, and the confidence associated

Table 5.

System Recommendations for Example 2

Dryer Type

Confidence

Run No.

1

2

3

4

5

67

1.

Fluidized bed dryer (well-mixed)

1

1

1

0.5

0.63

0.7

0.24

2.

Fluidized bed dryer (well-mixed),

0

0

0

0

0

0

0.14

after Feed Modification

3.

Fluidized bed dryer (plug flow)

0.3

0.3

0.3

0.15

0.19

0.3

0.1

5.

Vibrofluidized bed dryer

0.2

0.2

0.2

0.28

0.28

0.2

0.27

6.

Pneumatic (flash) dryer

1

1

1

0.38

0

1

0

7.

Pneumatic (flash) dryer, after feed

0

0

0

0.38

0

0

0

Modification

8.

Rotary steam-tube dryer

0.8

0.8

0.8

0.4

0.57

0.8

0.31

9.

Rotary steam-tube dryer,

after feed modification

0

0

0

0

0

0

0.33

10.

Rotary louvre dryer

0.60.60.60.3

0.6

5

0.60.36

11.

Rotary louvre dryer,

after feed modification

0

0

0

0.2

0

0

0.28

12.

Cascading rotary dryer

0.4

0.4

0.4

0.52

0.2

0.4

0.1

13.

Cascading rotary dryer,

after Feed modification

0

0

0

0.3

0

0

0.15

14.

Cascading rotary dryer (short-fall)

0

0

0

0

0.3

0

0.1

15.

Cascading rotary dryer (short-fall),

after feed modifications

0

0

0

0

0

0

0.15

COMPUTER-AIDED DRYER SELECTION

1869

with Option 6(no backmixing) is unity whereas that associated with Option
7 (with backmixing) is zero. As I

c

increases to 1.0 (sticky powder), the

reverse is true. Thus, if there is any real possibility of material-handling
problems as a result of the feed being cohesive, Option 7 will always be a
safe selection. However, backmixing does add complexity (and cost) to the
dryer and should therefore be avoided if possible.

As shown in Run 5 (Table 5), fragility also affects the selection. When

the fragility index I

f

is increased from 0.0 to 0.5, the confidence associated

with the well-mixed fluidized bed dryer decreases from 1.00 in the base case
to 0.63; that associated with the steam-tube dryer also decreases, from 0.8 to
0.57. In contrast, the vibrofluidized bed and rotary louvre dryers, both of
which are well-suited to handling fragile feeds, start to look more attractive.
The vibrofluidized bed cannot be regarded as being a real contender but the
rotary louvre is now preferred to the steam-tube dryer. The major difference
between the results of Runs 1 and 5 is, however, the elimination of the flash
drying options. Because of the high air velocities employed in pneumatic
dryers, they are not suitable for handling fragile powders.

The effect of reducing the product moisture content from 6% to 5% is

illustrated in Run 6. The flash and steam-tube dryer selections are unaf-
fected. However, the confidence associated with the well-mixed fluidized bed
dryer decreases from 1.00 to 0.70. This reduction is to be expected as well-
mixed fluidized beds are not well suited to drying product down to low
average moisture contents.

Finally, Run 7 represents the ‘‘worst case’’ scenario in which the feed

rate is at its maximum, the feedstock is regarded as being both somewhat
cohesive and fragile, and the outlet moisture content is at its lowest envi-
saged value. The confidences associated with most viable options are

Table 6.

Effect of Cohesiveness on Flash Dryer Selection

Confidence

Cohesiveness
Index, I

c

Dryer 6: Pneumatic

Dryer (Flash)

Dryer 7: Pneumatic Dryer

(Flash), after Feed Modification

0.0

1

0

0.2

1

0

0.4

0.61

0.61

0.5

0.38

0.38

0.60.2

0.6

1

0.8

0

1

1.0

0

1

1870

BAKER AND LABABIDI

considerably lower than those obtained in the base case (Run 1). The flash
dryer has again been eliminated.

The results obtained for this case study are not as clear-cut as those

obtained in Example 1. Run 1 displays two clear favorites, the well-mixed
fluidized bed dryer and the flash dryer; both are selected with a confidence of
unity. The steam-tube dryer (confidence ¼ 0.80) is a close runner up. The
sensitivity analysis reveals a somewhat different picture, however, and high-
lights the need for more detailed testing. The effects of possible variations in
feed rate and product moisture content are relatively minor. However,
uncertainty in the physical characteristics of the feed and product are of
much greater concern. If the material is indeed found to be fragile, and the
resulting product is unacceptable, then a flash dryer cannot be used.
However, backmixing can probably be employed to offset the effects of
cohesiveness, not only in the case of the flash dryer, but also with the
fluidized bed and steam tube dryers in the unlikely event that this is necess-
ary. Consequently, bench- and pilot-scale trials are vital in this case.

The final step in any dryer selection process is an economic evaluation

of acceptable alternatives. Programs similar to that employed in the present
algorithm for the design and costing of a well-mixed fluidized bed dryer can
be employed to obtain very approximate costs (see Table 7 for the predic-
tions for the present case study). However, as noted above, definitive pur-
chase costs should always be obtained from a reputable dryer manufacturer,
who will also guarantee the performance of his equipment.

ONGOING DEVELOPMENTS

Two further aspects of this work are currently in progress and will be

described in future publications. The first involves total integration of all the

Table 7.

Summary of Design and Costing Calculations for the Well-Mixed

Fluidized Bed Dryer

Table 5!

Run 1

Run 7

Drying time, s

880

1026

Bed area, m

2

2.77

3.22

Specific energy consumption, E

s

, kJ/kg

5480

5700

Cost, basic dryer, $

28,500

30,700

Cost, including ancillaries, $

85,600

92,300

Cost, installed, $

214,100

230,700

COMPUTER-AIDED DRYER SELECTION

1871

knowledge sources into a single automated system. This will enable the
selection algorithm to be run in a user-friendly manner over the Internet.

In the second, which is designed to test further the advantages of a

modular knowledge-base structure, the feasibility of incorporating an
additional, subsidiary, knowledge base is also being explored. Spray
dryers have been chosen for this purpose as they are available in a variety
of different designs to suit particular feed and product characteristics.
Features that may be varied include nozzle type (primarily wheel and pres-
sure nozzles), and flow characteristics (cocurrent-, mixed- and, occasionally,
countercurrent-flow). In addition, there is also the possibility of using multi-
stage dryers in which the spray dryer is combined with one or more fluidized
bed dryer. The latter may form an integral part of the drying chamber or be
an external device. Two-stage spray dryers with an integral drying belt are
also available.

At the time of writing, a draft stand-alone version of the knowledge

base (KSpray) has been written and is currently being tested and refined.
Once this process is complete, it will be adapted for incorporation into the
integrated system. If KContinuous selects a spray dryer as a feasible alter-
native, the user will be given the option of calling upon KSpray to provide
further details of the probable design.

CONCLUSIONS

The fuzzy expert system for batch-dryer selection described in previous

publications has been extended to include continuous dryers. The modular
structure proposed earlier was successfully adopted in this study. Three
inter-linked knowledge bases with data transfer between them were deve-
loped using the blackboard architecture. The first two, the mode (batch-
continuous) selector and the batch-dryer selector, were constructed with
relative ease from the earlier batch-dryer selection algorithm. The third,
the continuous-dryer selector, is a new development. A user interface and
scheduler have also been developed to automate the running of the three
knowledge bases.

Foreign programs have been written to provide additional information

to aid the selection process. These include design and costing programs for a
well-mixed fluidized bed dryer, a help module, and a searchable applications
database. Case studies are included to demonstrate the use of the system.

Work is in progress to integrate the foreign programs into the auto-

mated system. A subsidiary knowledge base relating to spray dryer selection
is also being prepared.

1872

BAKER AND LABABIDI

NOMENCLATURE

C

Capital cost of basic dryer, current year

$

C

o

Capital cost of basic dryer, 1963

$

D

Bed

Bed diameter

m

E

s

Specific energy consumption

kJ/kg

I

c

Cohesiveness Index

–

I

f

Fragility Index

–

J

Marshall & Swift

equipment cost index, current year

–

J

o

Marshall & Swift equipment cost index, 1963

–

Greek Letters

g

ij

Membership function

–

m

Fuzzy variable (Various

–

ACKNOWLEDGMENTS

The authors gratefully acknowledge the financial support of Kuwait

University Research Administration (Grant No. EC 089), which made this
work possible.

Thanks are also due to Dr Reena George for preparing the Help files

and the Applications database and to Eng. Khalid Damyar for undertaking
much of the detailed programming. Finally, the authors acknowledge with
thanks

the

valuable

advice

received

from

Dr

Keith

Masters

of

SprayDryConsult, Charlottenlund, Denmark, on spray dryer selection.

REFERENCES

Baker, C.G.J. (1997), Dryer selection, in Industrial drying of foods,

ed. Baker, C.G.J., Blackie Academic & Professional, London,
pp. 242–271.

Baker, C.G.J. (1999), Predicting the energy consumption of continuous well-

mixed fluidized bed dryers from drying kinetic data, Drying Technol.,
Vol. 17, nos. 7 & 8, pp. 1533–1555.

Baker, C.G.J., Al-Roomi, Y.M., Al-Sharrah, G. (2000), Optimal design

of well-mixed fluidized bed dryers, Drying Technol., Vol. 18, no. 7,
1509–1535.

Baker, C.G.J., Lababidi, H.M.S. (1998), Application of fuzzy expert systems

in dryer selection, Drying ’98, ed. Akritidis, C.B., Marinos-Kouris, D.
and Saravakos, G.D., Ziti Publ., Thessaloniki, Vol. A, pp. 448–455.

COMPUTER-AIDED DRYER SELECTION

1873

Baker, C.G.J., Lababidi, H.M.S. (2000), Development of a fuzzy expert

system for the selection of batch dryers for foodstuffs, Drying
Technol., Vol. 18, nos. 1 & 2, pp. 117–141.

Craig, I.D. (1988), Blackboard systems, Artificial Intelligence Rev., Vol. 2,

pp. 103–118.

Dubois, D., Prade, H. (1980), Fuzzy sets and systems – Theory and applica-

tions, Academic Press, New York.

Kemp, I.C. (1998), Progress in dryer selection techniques, Drying ’98, ed.

Akritidis, C.B., Marinos-Kouris, D. and Saravakos, G.D., Ziti Publ.,
Thessaloniki, Vol. A, pp. 668–675.

Kemp, I.C., Bahu, R.E. (1995), A new algorithm for dryer selection, Drying

Technol., Vol. 13, nos. 5–7, pp. 1563–1578.

Kraslawski, A., Lyssov, I., Kudra, T., Nystrom, L. (1999), Case-based rea-

soning for equipment selection using rough sets analysis in adaptation
phase, Computers in Chemical Engineering Supplement, pp. S707–
S710.

Lababidi, H.M.S., Baker, C.G.J. (1999), An expert system for dryer selec-

tion using fuzzy logic, Computers in Chemical Engineering Supplement,
pp. S691–S694.

Matasov, A., Menshutina, N. and Kudra, T. (1998), Information system for

the selection of dryers, Drying ’98, ed. Akritidis, C.B., Marinos-Kouris,
D. and Saravakos, G.D., Ziti Publ., Thessaloniki, Vol. A, pp. 624–629.

MultiLogic, Inc. (1997), ReSolver: User Manual.
Nevenkin, S.L., Chavdarov, G.I. (1992), Expert system for synthesis

of structures of drying installations for dispersive materials, Drying
’92, ed. Mujumdar, A.S., Elsevier Science Publ., Amsterdam, pp.
1745–1751.

Quinn, M.F. (1963), Fluidized bed dryers, Ind. Eng. Chem., Vol. 55, no. 7,

July, pp. 19–24.

van t’Land, C.M. (1991), Industrial drying equipment – Selection and appli-

cation, Marcel Dekker, New York.

Zadeh, L.A. (1965), Fuzzy sets, Inform. and Control, Vol. 80, pp. 338–353.