Distributed Knowledge Management Architecture and Rule Based
Reasoning for Mobile Machine Operator Performance Assessment
Petri Kannisto
1
, David Hästbacka
1
, Lauri Palmroth
2
and Seppo Kuikka
1
1
Department of Automation Science and Engineering, Tampere University of Technology,
P.O. Box 599, 33101 Tampere, Finland
2
John Deere Forestry, P.O. Box 474, 33101 Tampere, Finland
Keywords: Distributed Knowledge Management, Rule based Reasoning, Operator Performance Assessment, Mobile
Machines.
Abstract: The performance of mobile machine operators has a great impact on productivity that can be translated to,
for example, wasted time or environmental concerns such as fuel consumption. In this paper, solutions for
improving the assessment of mobile machine are studied. Usage data is gathered from machines and utilized
to provide feedback for operators. The feedback is generated with rules that define in what way different
measures indicate performance. The study contributes to developing an architecture to manage both data
collection and inference rules. A prototype is created: rule knowledge is managed with decision tables from
which machine-readable rules are generated. The rules are then distributed to application instances executed
in various locations. The results of the prototype promote several benefits. Rules can be maintained
independent of the actual assessment application, and they can also be distributed from a centrally managed
source. In addition, no IT expertise is required for rule maintenance so the rule administrator can be a pure
domain expert. The results bring the architecture towards a scalable cloud service that combines the benefits
of both centralized knowledge and distributed data management.
1 INTRODUCTION
Even though the degree of automation in industries
is constantly rising, various mobile machines still
require a human operator, no matter how modern
their equipment is. The difficulty of automatization
is due to the irregular manner of not only operating
environment but also the nature of work tasks.
Despite the intelligence of human operators, they are
prone to non-optimal operating of equipment.
Fortunately, as human beings have the ability to
learn, systematic feedback may significantly
improve production performance in the course of
time. For instance, there could be potential to
improve operating speed and reduce fuel
consumption simultaneously. Considering that profit
margins are narrow due to rising costs and a high
degree of competition, even an improvement of a
few percents could bring a substantial advantage.
The information and control system of a modern
mobile machine is readily advanced which provides
a basis for feedback generation. As there are
numerous sensors around the machine that measure
various magnitudes, a lot of data is available for
collection.
The sensor data from the machine level can be
automatically collected and stored by the
manufacturer for example in the case a customer
subscribes performance analysis and targeted
maintenance and tuning for the machine fleet. From
an enterprise wide storage, the data can be accessed
for studying various indicators and comparing the
values between machines of a similar type, age,
working conditions, geographic area etc. Often, the
information needs to be utilizable globally for
analysis but also for generating performance reports
based on e.g. individual machines, fleets, operators,
or market areas.
However, the path from raw sensor data to a
human-readable and analytic feedback report is long.
Several questions arise. Which data is really
relevant? Should all the data be collected just in
case? Where to store the data? What is the
methodology utilized to analyze the data to generate
observations that form the basis of the feedback?
Besides, there may be considerable differencies in
440
Kannisto P., Hästbacka D., Palmroth L. and Kuikka S..
Distributed Knowledge Management Architecture and Rule Based Reasoning for Mobile Machine Operator Performance Assessment.
DOI: 10.5220/0004870004400449
In Proceedings of the 16th International Conference on Enterprise Information Systems (ICEIS-2014), pages 440-449
ISBN: 978-989-758-027-7
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
machine utilization between the countries and
market areas. Therefore, regional differentiation
must be applied.
Among the several interesting aspects of operator
assessment, this paper concentrates on two of them.
The first is the architecture of the information
management and assessment system. The second
aspect is how to formulate and implement the actual
rules applied in the assessment process. The rules
define what level of performance is required for
each measurement. The baseline for the architecture
is how rules are accessed from distributed devices
and how they are managed. The rules should be
separated from the actual application logic so that
rule modifications do not require any part of the
assessment system to be recompiled. Also, no IT
expertise should be required from a person to be able
to create or maintain the rules.
The paper is organized as follows. Section 2
presents the methodology and background to this
research as well as related work. In section 3 the
problem is defined. The solution architecture and the
knowledge management approach is explained in
sections 4 and 5, respectively. Sections 6 and 7
contain discussion of results and conclusions.
2 RESEARCH CONTEXT
2.1 Research Methodology
The methodology applied in this work relies on
constructive design science in which new artefacts,
e.g. software, architectures, and algorithms, are
developed to support projections of concepts. The
developed solutions are then evaluated and
compared to existing ones. Through a number of
iterations, the new solutions finally provide new
insights and better theories.
In this case the methodology was applied to
design the conceptual architecture and the rule based
knowledge management. The architecture was
designed based on user requirements and developed
to support the needs of distributed data acquisition,
service-based operation, and rule-based knowledge
management. Concerning the management of rule
based knowledge, experiments with various rule
modeling approaches were performed. Exploiting
the results, the most appropriate solution was chosen
and proposed. The performed requirements analysis
sets the restrictions for the architecture and the
solution proposed in this paper.
2.2 Rule-based Reasoning
Even with simple rules, it might not be an easy task
to implement a well-performing rule engine to apply
reasoning rules to data. According to Schneier and
Matignon, a challenge of applying rules is that even
with a limited set of facts, the fact set will change
once the rules are applied (Figure 1). An altered fact
set means different circumstances so at least some of
the rules must be executed again. The most
straightforward way is to apply each rule after each
fact insertion and start over whenever a rule is fired.
However, the performance of this algorithm is slow
with large data sets. (Schneier 2002; Matignon
2011).
Figure 1: Facts are not only the input but also the output of
rules. Based on (Schneier 2002; Matignon 2011).
The Rete algorithm developed by C.L. Forgy
(1982) is a response to the performance challenge
related to rule-based applications. It has two basic
principles. First, a rule typically affects only a few
facts, and second, rules may be structurally similar
in that some conditions exist in more than one rule.
Using these principles, a graph is built from the
rules. A disadvantage of Rete is that it consumes
relatively lot of memory. (Ingargiola) However,
inference actions are fast because the graph enables
targeting computation to where it is appropriate.
(Schneier 2002; Ingargiola)
While Rete was obviously a significant step
forward it has also received criticism, and several
improvements have been developed. Miranker
(1987) has proposed the TREAT algorithm which has
been inspired by the shortcomings of Rete such as its
memory consumption and some unnecessary
computation it performs. After performing a
comparison of Rete and TREAT, Wang & Hanson
(1992) have stated that TREAT is typically faster
but not always. Charles J. Forgy, the designer of
Rete, has developed Rete II (RETE II) which
showed a significant speed improvement in an
experiment (Benchmarking CLIPS/R2). Further, the
algorithm has been evolved into Rete-NT (Owen
2010). While the principles of the original Rete have
been published, it seems that its successors Rete II
and Rete-NT have been developed specifically for
commercial systems so there is little information
DistributedKnowledgeManagementArchitectureandRuleBasedReasoningforMobileMachineOperatorPerformance
Assessment
441
available about them. That is, while several articles
discuss improving the original Rete, there is no
information available how the suggestions align to
the newest Rete versions. Articles suggesting
extensions and improvements are, for instance,
(Berstel 2002), (Ren & Wang 2008), (Xiao et al.
2009), (Liu et al. 2010) and (Yang et al. 2011). The
Drools rule engine, for example, uses an optimized
Rete version (Drools Expert User Guide 2012).
2.3 Related Work
While there are no publications with a topic similar
to this, several studies have common aspects such as
rules, distribution or knowledge management.
Chen & Wu (2003) introduce OKSA (Open
Knowledge Server Architecture), a framework for
managing knowledge in a network server. The
framework contains service interfaces for retrieving
semantic data as well as for rule-based reasoning.
Jung et al. (2006) take an information
management approach. They propose an architecture
combining knowledge management and business
process management systems. Three types of
information are stored: process template, process
instance and process related. Knowledge storing and
the business process point of view are relevant also
in this study.
Rajasekar et al. (2006) have researched the
application of rules for the management of data
grids. While data grids are distributed by nature, it is
challenging to make sure that the overall state of the
data remains consistent. In that context, a rule-based
approach is advantageous as it raises the flexibility
of maintaining consistency constraints.
Marin-Perianu and Havinga (2007) have
researched fuzzy logic based reasoning applied to
wireless sensor networks. They have developed D-
FLER, a rule-based fuzzy logic engine. Fuzzy output
is generated from sensor readings which are then fed
into an inference engine. The uncertainty of sensors
is taken into account by processing the output of
multiple neighbor sensors simultaneously.
Terfloth et al. (2007) have also performed rule-
related research for wireless sensor networks. They
have developed an architecture called FACTS which
introduces a middleware layer to facilitate sensor-
related programming by raising abstraction level.
Rules are applied to events that arise from sensors.
Grobelny (2008) suggests a rule-based expert
system to assist the composition of service
architectures. The aim is to raise the abstraction
level to enable composition for domain experts that
are not specialized in software engineering.
Bontchev & Vassileva (2009) have studied the rule-
based approach to create an adaptable e-learning
environment. They have used the Drools rule engine
to enable adaptation based on rules.
Dunkel et al. (2010) introduce an event-driven
decision support system. In their paper, sensors
provide data about the traffic situation in an area to
facilitate traffic management. The data is refined to
events that are processed by a rule engine. In the
traffic management context, real-time constraints are
present and data processing time is limited.
Nalepa et al. (2013) propose an architecture for
business rules modeling. In their approach, the rules
are modeled using BPMN (Business Process Model
and Notation) diagrams. The approach provides
control over in which order rules are applied. In the
flow of a BPMN diagram, each task contains a set of
rules to be executed. The approach aims for easier
ruleset management; the visual representation makes
a complex set of conditions more understandable.
3 PROBLEM DEFINITION
The development approach of this study follows
agile principles. First, a use case analysis was
performed; its results are described in this chapter.
Second, the actual requirements of the architecture
were formulated as given in chapter 4. Finally, a
prototype was implemented as described in chapter
5. The process is analogous to an agile development
process as explained by Douglass (2009): after a set
of requirements has been specified, a prototype is
implemented. As the prototype might not meet all
the known requirements, any shortcomings will be
taken into account in future development. The key is
to control the development process by enabling
proofing and experimenting design choices through
a tangible implementation.
To begin the discovery of the problem, a use case
analysis was performed with the objective to
recognize actors that are involved in utilizing and
maintaining the operator performance assessment
system. This enabled defining requirements on the
architecture and the information management.
Actors are involved in the assessment process in
multiple roles; the most essential actors discovered
in the analysis are illustrated in Figure 2.
Trainers generate feedback reports with the
system. The feedback is then given to machine
operators so that they can improve their way of
working. A feedback report is simply a human-
readable document that shows a comparison of the
operator’s performance to the average in the same
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
442
market area during a certain time period. Typically,
an operator does not even recognize that there is
room for improvement before they receive the
feedback. Naturally, a skilled operator will receive
positive feedback, which indicates that they should
keep working in a similar manner as before.
Figure 2: The most essential actors and the use cases
performed by them that were identified.
Mechanics generate machine reports that show
equipment-related performance data. Modern mobile
machines are often very complex with several
subsystems so there may be multiple components
that need maintenance, parameter tuning or
calibration. These can be detected by investigating
machine performance data. If a machine report
indicates any need for service, the mechanic will
inform the machine operator about it.
While mechanics and trainers are end-users, the
knowledge administrator has an essential role in the
actual assessment process. Based on their domain
knowledge, they will create and maintain the rules
applied to performance data. The knowledge
administrator has qualitative domain expertise to
determine which measures are relevant in the
performance assessment. Whatever methodology is
applied to administer the rules, the ultimate output of
the knowledge administrator is the resulting report.
Otherwise, the knowledge administrator is invisible
to the end-users.
From the assessment system point of view, the
machine itself is an actor as well. It gathers data
from its sensors to be utilized for the assessment. In
addition, the data should include machine type and
the relevant setup that will have an effect in any
forthcoming assessment.
Based on the identified actors and the use case
analysis, the requirements can be summarized as
follows. Whoever modifies the rules, no IT expertise
should be required. That is, preferably no textual
syntax should be utilized for rule definition, or at
least the rule code should be generated based on
some graphical syntax. Also, the rules should be
stored in a location where they can be maintained
separately from the actual application instances. The
interface provided for rule maintenance should be
accessible over the web. Finally, there should be a
method to distribute an updated rule set, i.e. for
generating the assessment reports, to distributed
instances without the need to recompile the rules.
4 SOLUTION ARCHITECTURE
REQUIREMENTS
Due to the requirements of the rule assessment
system, architectural aspects are paramount in the
development. While scalability and any other future
requirements must be addressed, the improved
architecture should also consider the potential
benefits of a previously implemented assessment
system and support a stepwise transition.
The concept of the assessment system is shown
in Figure 3. Various machine types operate in
different geographic areas, and data can be collected
for individual machines as well as an entire machine
fleet of a given entrepreneur. After each cycle of
operation, the collected data is submitted to the
assessment system. The collected data may be
utilized for generating a reference dataset for the
machine type. After the data submission, the
reference dataset is compared to the data collected
from operation and the results are shown to the
operator. The salient aspects of the architecture are
explained in the following sections.
Figure 3: The architecture needs to support data collection
from machines all over the world, provide means for
customization of the assessments based on different
criteria, and deliver the assessments reports based on
centrally managed knowledge.
4.1 Data Acquisition and Storage
Modern mobile machines have sophisticated sensors
that enable versatile application development.
Originally, the sensors have been introduced to serve
the needs of constantly developing machine
automation. Today, they provide new value by
providing the basis of the data collection driven by
assessment needs.
DistributedKnowledgeManagementArchitectureandRuleBasedReasoningforMobileMachineOperatorPerformance
Assessment
443
Despite the possibilities of modern wireless
networking, an off-highway machine cannot be
assumed to have a persistent and stable Internet
connection. Mobile Internet is not available
everywhere and no network is fully reliable. Also,
whichever is the technology utilized to receive and
store the data, occasional downtime is probably
inevitable. Thus, it is required that a machine can
store measurement data locally – if not for long
periods then at least for temporary caching.
Finally, all the data gathered from machines
should be stored remotely for future utilization. To
enable consistent data management easily, a
common storage must be used by all the machines.
To reach the required flexibility, the architecture
should be designed to function on a cloud platform.
Obviously, it would enable both global knowledge
management and storage as well as local
customizations. Moreover, cloud services are
scalable, and as there is no single point of failure,
the availability of the service would be improved.
4.2 Dispersed Information Processing
The measured performance data is processed on the
machine level as well as for groups of machines at
the central data storage. The processing and analysis
of the data differs depending on the purpose for
which the information is produced. The process is
explained in more detail by Palmroth (2011).
On the machine level, the distributed control
system of the machine and sensors produce large
amounts of data. This data must be pre-processed
such that it contains the relevant information needed
for the performance monitoring and assessment
purposes. The data must also be compressed in order
to transfer it from the machine. Naturally, data
compression is always a tradeoff between fidelity
and the amount of data to be transferred.
Quantitative knowledge of the rule based system
is based on statistics of performance data measured
from large machine fleets on separate geographic
and market areas. This data is stored in the central
data storage. It is a very important concept in the
knowledge management architecture that only
qualitative domain knowledge is required from the
knowledge administrator. It is not necessary to
master all the numerical values of performance
measurements from different types of machines etc.,
because the quantitative information is produced by
the statistics of measured performance data.
4.3 Decentralized Knowledge
Management
Domain knowledge has multiple dimensions. Some
qualitative knowledge may be relevant only for a
specific machine type or a set of machine types with
a specific equipment setup. In addition, there could
be considerable differences in machine utilization
between different geographical areas, work methods,
operating temperatures and terrain types. On the
other hand, some knowledge may also be considered
general regardless of areas and machine types. When
there are relevant differences, they should be
incorporated in the domain knowledge.
There is also a lot of variation between operating
locations. The terrain may vary from flat to hilly or
even mountainous. The flatter the land the easier and
less resource-consumptive operation should be
expected. Also, the surface of the land has an effect
as well: it is easier for any machine to move on hard
land. Any obstacles such as plants or stones will also
make operating more challenging. Finally, there may
be variation between the load or material types
processed by the machine: some of them may
require more work for processing than the others.
One of the most essential questions of
knowledge management is the global and distributed
nature of the knowledge and the need to be able to
access it from various locations. With increasing
amounts of data constantly accumulating from
machines all over the world, typical big data
challenges are expected to become important. The
central storage will make it straightforward to access
the data; however, as the knowledge actually
originates from several locations, it will be
challenging to store it in a single-format storage.
That is, to reach an architecture that is both easy to
manage and also adaptable to meet the requirements
of different areas, careful design is required.
In this work, a high level of centralization is
required to make it possible to manage the rule
knowledge. Due to the level of domain expertise
required to manage the rules, there will only be a
few professionals that have the appropriate skills.
That is, by utilizing a centralized architecture, it is
possible for the professionals to have a control of the
rule knowledge as a whole.
4.4 Assessment as a Service
Operator assessment can be seen as a service
provided by the machine manufacturer. In general,
the data is owned by the machine owner but
delivered to the service provider per service
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
444
agreement. The customers will order reports to
improve their efficiency and effectiveness; however,
extra revenue will also be generated for the
manufacturer. To ensure that the operator
assessment system reflects the business
requirements, the system should be implemented
utilizing service thinking.
Compared to a monolithic software application,
service-based design performs better in distributed
business. Maintaining and updating a software
application is easier if instances are not run at
customers’ locations but rather in an environment
controlled by the service provider. Also, principles
such as loose coupling between applications, a high
abstraction level of interfaces and aiming at
reflecting business requirements in design are likely
to ease both development and maintenance.
On the customer side, it is clear that some client
application is required. Due to the capabilities of the
web browsers and the web developers of today, the
client application could actually be a web browser
that accesses a website for report generation.
5 RULE-BASED PERFORMANCE
ASSESSMENT PROTOTYPE
5.1 Implementation
To bring the current assessment system closer to the
above concept, a prototype was implemented. Its
general architecture is illustrated in Figure 4. The
rule base has been implemented with the open
source Drools framework which provides a browser
interface for maintaining the rule knowledge. The
analysis software accesses the rules and applies them
to the machine data to make the assessments. The
prototype has been developed for forestry domain
but can be generalized to any mobile machine type.
Figure
5 presents a more detailed description of
the implementation elaborating the transformation of
rule-based knowledge. For this, there are three major
components: data processor, rule base and rule
converter.
The data processor application is the most
important component. It generates performance
reports according to the rules. The core of the
application, the data processing logic, has been
implemented in Matlab and compiled as a desktop
application. The calculation functionality provided
by Matlab is versatile compared to general-purpose
programming languages. In the prototype, it is
implemented as a desktop application that caches the
rules from the rule base.
Figure 4: The prototype implemented using the Drools
framework includes an interface for the domain expert for
managing the rule knowledge that is used by the analysis
software making the assessments.
Figure 5: Rule based domain knowledge is transformed for
the Matlab based analysis software.
The rule base has been built in a web server
utilizing the Drools rule framework. Drools may be
included as a library to any Java application, but the
rules may also be managed and executed in a web
server application called Guvnor. Guvnor has a
graphical browser interface that is easily accessible
over the Internet. The rule base provides also a rule
engine applying the RETE algorithm for execution.
Drools supports various means for defining rules
such as direct input in a rule language and decision
tables. The rule language, DRL (Drools Rule
Language) (see Drools Expert User Guide 2012),
covers rule needs rather widely. However, while it
may be intuitive for a software engineer, it is
certainly not understandable for a person unfamiliar
with programming. Also, if the number of rules is
large, it may be difficult to manage the rules as a
whole if all of them are in textual form.
To eliminate the burden of learning the DRL
syntax, it is possible to develop domain specific
languages (DSL) over DRL. However, as a DSL can
still lack a graphical representation, decision tables
may also be utilized to model the rules. In a decision
table, each row defines a rule, and each column
represents the value of a condition. The trade-off of
DistributedKnowledgeManagementArchitectureandRuleBasedReasoningforMobileMachineOperatorPerformance
Assessment
445
decision tables is their limited power of expression
compared to DRL, and they also require that the
rules within one table are structurally similar.
In this study, decision tables are utilized due to
their intuitive and easy-to-learn nature. Any
limitations related to the power of expression are not
important as the rules are rather similar by their
structure. Once any decision tables have been
created, the rules defined in them are downloadable
as a DRL file. The rule engine component included
in Guvnor is not utilized as applying inference
locally will raise the flexibility of the solution in
terms of whether there is a persistent Internet
connection or not.
The third component in the architecture is a
converter application that enables the integration of
the rule base and the data processor. The data
processor receives rules in a proprietary XML
format while the output of the rule base is DRL so a
conversion is required between them. The power of
expression provided by DRL is rather extensive, but
only a subset of it is required by the data processor.
That is, only certain structures of DRL are
recognized and converted to the XML format. As
long as simple decision tables are used, no
compatibility problems are expected as the DRL
code generated from decision tables is uniform. The
conversion of rule formats is illustrated in
Figure 6.
Figure 6: A simplified example of representing a decision
table row as DRL and converting it to an XML node.
5.2 Example of Use
To clarify the proposal in this paper, this section
provides an example of use of the prototype in
mechanized forestry. The example is about
forwarder operations in Brazil and in Denmark.
Forwarder is a forest machine type that is used to
collect felled trees from forest and to transport them
to roadside for further transportation.
The entire performance assessment process is
based on the rules created by knowledge
administrators. Rules are created separately for
different machine types such as harvesters or
forwarders, including the various models offered by
the manufacturer. The knowledge administrators
with global expertise create the global rules that
form the basis of all the rules of a machine model.
To maintain the rules, they log into Drools Guvnor
using a web browser. The rules are defined as
decision tables using a graphical user interface.
Both Brazil and Denmark have their local
characteristics that are considered by the regional
knowledge administrators of the countries, e.g. an
area sales support manager with the required domain
knowledge. For example, Denmark is a relatively
flat country which makes driving the forwarder fast
and easy. In contrast, there are hilly forests in Brazil
which makes driving more demanding, raising fuel
consumption and lengthening driving times. On the
other hand, the trees in Brazil are almost entirely
eucalyptus that is easy to collect, whereas Denmark
has a variety of different tree species which need to
be sorted separately. Therefore, longer loading times
are expected in Denmark. Both Danish and Brazilian
knowledge administrators make local customizations
to the global rules. Like global knowledge
administrators, the local ones maintain rules as
decision tables with Drools Guvnor.
After local customizations, the assessment rules
are ready for distribution. In the prototype, each
local knowledge administrator downloads a DRL
representation of the decision tables manually from
Drools Guvnor. Then, with the converter
application, the DRL representation is converted to
XML so it can be utilized by the analysis software.
Finally, the XML rule sets are distributed to trainers.
Once customized, transformed and distributed,
the rules can be utilized in assessment. The Brazilian
and Danish operators use the forwarders in forests.
After some time period, trainers retrieve raw usage
data from each machine with their laptops and feed
the data to the analysis software. Once the analysis
software has generated performance data from the
raw data, it applies the localized XML rule set to the
performance values. As the result, a textual feedback
document is generated for a specific operator and a
specific time period. It is the final output of the
process; the results are given to operators so they
can improve their way of working.
As the understanding of performance assessment
improves over time, the assessment rules are likely
to evolve. This will require modifications in not only
global but also in local rule sets. Whenever the
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
446
global rule set is modified, it will be communicated
to local admins that consider the modifications.
To conclude, the assessment process is as
follows:
1. Global knowledge admin: create global rules
2. Local knowledge admin: localize rules
3. Local knowledge admin: convert and distribute
rules
4. Trainer: get raw data from machine and run
assessment with the analysis software
5. Operator: learn from assessment results
6 RESULTS AND DISCUSSION
During the evaluation of the prototype, it was found
out that the chosen approach has several benefits.
Whenever rules are applied in the data processor, no
Internet connectivity is required which is often the
case with off-highway vehicles. The rule-based
knowledge can also be centrally managed by a
domain expert without ICT skills. The architecture is
also relatively straightforward to implement.
In the prototype, however, manual work is still
required to distribute rules – whenever the rules are
changed, a new instance of them must be made
available to the analysis application instances. In
addition, having a separate conversion application
adds an extra component and a new level of
complexity – the entire distribution process should
be made automatic to maximize the ease of use and
management. However, even though complexity is
not desirable, a multistep conversion process makes
it possible to change one step independently as long
as interfaces remain as they are. The ultimate goal is
to cache the entire rule distribution service for the
analysis software: if there is an Internet connection,
check for recent changes, if not, utilize a cached
version.
From the point of view of this work, Drools has
both advantages and disadvantages. Its capabilities
for rule modeling are well beyond the complexity
required for the current system. So, there is no
shortcoming from this point of view, but Drools also
makes it possible to model rules so complex that
they cannot be converted for utilization in analysis.
However, a support for rules more complex than the
current ones might be required in the future. In
addition, the current decision table format sets an
effective restriction to rule complexity so there is
little danger that “too complex” rules are modeled.
On the other hand, a more flexible modeling method
than decision tables might be needed in the future –
any other methods provided by the current Drools
require too much expertise for most non-ICT users.
One advantage of Drools is its wide utilization in
both research and practical applications so further
development and support for the framework are
expected. To conclude, while Drools is not a perfect
solution, its features provide a solid foundation.
A global service-oriented approach would,
perhaps, suit better for the requirements than
running multiple instances of the application in
various locations. The approach of the prototype
allows for realizing the concept to provide the
operator assessment also as a centralized service
over the Internet, as envisioned in section 4.4. The
performance assessment system should, however, be
later implemented as a cloud service to enable not
only centralized management but also distributed
one where it is necessary. As centralization and
distribution can in general be considered conflicting
goals, it is an important design question which
functions should be centralized and which ones
distributed or local.
The current requirement to have a trainer on-site
to get raw machine data and generate feedback for
operators is a limitation. In the vision, raw machine
data would be stored in the cloud. Then, operators
could utilize the analysis service to get feedback
whenever they want to. Of course, the expertise of a
trainer could still be beneficial for an operator when
interpreting feedback.
7 CONCLUSIONS
The paper introduces a concept of a distributed
knowledge management solution. A prototype has
been developed for maintaining rule-based domain
knowledge that is used in operator performance
assessment. The developed functionality allows
domain experts to maintain the knowledge without
ICT skills. A system architecture has also been
developed that supports centralized management and
customization of this knowledge and the globally
distributed utilization when generating assessment
reports. The solution has been applied in the
assessment of forest machine operators’ work
technique and performance.
The implemented rule management system
enables rule modeling and distribution for non-ICT
users. However, there is room for improvement in
ease of modelling as well as adaptability and
flexibility – simply put, how to have control over the
entire assessment process with less human effort.
While rule-based architectures and distributed
knowledge management systems have been
DistributedKnowledgeManagementArchitectureandRuleBasedReasoningforMobileMachineOperatorPerformance
Assessment
447
implemented before, this work is novel in the
domain of mobile machines. In addition, the support
for not only global distribution and management but
also the consideration of local customizations is an
advance in this field of science. Moreover, different
usage roles are also considered in this paper. The
approach can be adapted to other applications for
mobile machines or settings where similar
information management challenges are present.
In the future, more development work is needed
to improve flexibility, scalability and reachability of
the system. Transition towards a cloud based
approach could improve this by facilitating central
management of shared resources as well as
utilization and integration of data sources, i.e. global
machine data and machine fleets. Another
interesting goal is applying the approach towards
condition monitoring and diagnostics for detecting
events from sensor data for reactive or preventive
maintenance.
REFERENCES
Benchmarking CLIPS/R2. Production Systems
Technologies, Inc. Available at: http://www.pst.com/
benchcr2.htm (Last visited 9th Oct. 2013).
Berstel, B., 2002. Extending the RETE Algorithm for
Event Management. In TIME 2002, Proceedings of the
Ninth International Symposium on Temporal
Representation and Reasoning. IEEE. pp. 49-51.
Bontchev, B., Vassileva, D., 2009. Rule-driven approach
for adaptable e-learning content delivery. In EISTA
2009, Proceedings of 7th International Conference on
Education and Inf. Systems, Technologies and
Applications. IEEE. Pp. 10-13.
Chen, H., Wu, Z., 2003. OKSA: an open knowledge
service architecture for building large scale knowledge
system in semantic Web. In International Conference
on Systems, Man and Cybernetics. IEEE. Vol. 5, pp.
4858-4863.
Douglass, D., 2009. Real-Time Agility, Addison-Wesley.
1
st
edition.
Drools Expert User Guide, 2012. Available at:
http://docs.jboss.org/drools/release/5.5.0.Final/drools-
expert-docs/html_single/
Drools Guvnor. Available at: http://www.jboss.org/drools/
drools-guvnor.html (Last visited 9th Oct. 2013).
Dunkel, J., Fernández, A., Ortiz, R., Ossowski, S., 2011.
Event-driven architecture for decision support in
traffic management systems. In Expert Systems with
Applications, 38(6), 6530-6539.
Forgy, C. L., 1982. Rete: A fast algorithm for the many
pattern/many object pattern match problem. In
Artificial Intelligence, Volume 19, Issue 1, September
1982, Pages 17-37, DOI: 10.1016/0004-
3702(82)90020-0.
Grobelny, P., 2008. Knowledge representation in services
oriented architecture. In Przeglad Telekomunikacyjny,
6, 793-796.
Ingargiola, G. The RETE Algorithm. Available at: http://
www.cis.temple.edu/~giorgio/cis587/readings/rete.htm
l (Last visited 8th Oct. 2013).
Jung, J., Choi, I., Song, M., 2006. An integration
architecture for knowledge management systems and
business process management systems. In Computers
in Industry, 58(1), pp. 21-34.
Liu, D., Gu, T., Xue, J.-P., 2010. Rule Engine Based on
Improvement Rete Algorithm. In ICACIA 2010,
International Conference on Apperceiving Computing
and Intelligence Analysis. IEEE. pp. 346-349.
Marin-Perianu, M., Havinga, P., 2007. D-FLER – a
distributed fuzzy logic engine for rule-based wireless
sensor networks. In Ubiquitous Computing Systems
(pp. 86-101). Springer Berlin Heidelberg.
Matignon, C.-A., 2011. Rete Algorithm Demystified! —
Part 2. Available at: http://techondec.wordpress.com/
2011/03/14/rete-algorithm-demystified-part-2/
Miranker, D.P., 1987. TREAT: A Better Match Algorithm
for AI Production Systems; Long Version. Technical
report AI TR87-58.
Nalepa, G. J., Kluza, K., Kaczor, K., 2013. Proposal of an
Inference Engine Architecture for Business Rules and
Processes. In Artificial Intelligence and Soft
Computing (pp. 453-464). Springer Berlin Heidelberg.
Owen, J., 2010. World's fastest rules engine. InfoWorld.
Available at: http://www.infoworld.com/t/business-
rule-management-systems/worlds-fastest-rules-engine-
822.
Palmroth, L., 2011. Performance Monitoring and
Operator Assistance Systems in Mobile Machines.
Doctoral dissertation, Department of Automation
Science and Engineering, Tampere University of
Technology, Tampere, Finland.
Rajasekar, A., Wan, M., Moore, R., Schroeder, W., 2006.
A prototype rule-based distributed data management
system. In HPDC workshop on Next Generation
Distributed Data Management (Vol. 102).
Ren, Z., Wang, D., 2008. The Improvement Research on
Rule Matching Algorithm Rete in Electronic
Commerce Application Systems. In WiCOM 2008, 4th
International Conference on Wireless
Communications, Networking and Mobile Computing.
IEEE. pp. 1-4.
RETE II. Production Systems Technologies, Inc.
Available at: http://www.pst.com/rete2.htm (Last
visited 9th Oct. 2013).
Schneier, B., 2002. The Rete Matching Algorithm.
Available at: http://www.drdobbs.com/architecture-
and-design/the-rete-matching-algorithm/184405218.
Terfloth, K., Wittenburg, G., Schiller, J., 2006. FACTS – a
rule-based middleware architecture for wireless sensor
networks. In Comsware 2006, First International
Conference on Communication System Software and
Middleware. Pp. 1-8. IEEE.
Wang, Y.-W., Hanson, E.N., 1992. A Performance
Comparison of the Rete and TREAT Algorithms for
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
448
Testing Database Rule Conditions. In ICDE 1992,
Eighth International Conference on Data Engineering.
IEEE. pp. 88-97.
Xiao, D., Tong, Y., Yang, H., Cao, M., 2009. The
Improvement for Rete Algorithm. In ICISE 2009, 1st
International Conference on Information Science and
Engineering. IEEE. Pp. 5222-5225.
Yang, P., Yang, Y., Wang, N., 2011. IRETE: An
Improved RETE Multi-entity Match Algorithm. In
ICECC 2011, International Conference on
Electronics, Communications and Control. IEEE. Pp.
4363-4366.
DistributedKnowledgeManagementArchitectureandRuleBasedReasoningforMobileMachineOperatorPerformance
Assessment
449