You are on page 1of 5

sensors

Artificial intelligence tools


can aid sensor systems
At least seven artificial intelligence (AI) tools can be useful when applied to sensor
systems: knowledge-based systems, fuzzy logic, automatic knowledge acquisition,
neural networks, genetic algorithms, case-based reasoning, and ambient intelligence.

S
even artificial intelligence (AI) tools industrial processes. Research below intends to
Key are reviewed that have proved to be show these concepts can work in practice.
concepts useful with sensor systems. They
are: knowledge-based systems, fuzzy Creating smarter sensor systems
Seven artificial intelligence
(AI) tools—knowledge- logic, automatic knowledge acquisi- Sensor systems can be improved using arti-
based systems, fuzzy tion, neural networks, genetic algorithms, case- ficial intelligence (AI). AI emerged as a com-
logic, automatic knowl- based reasoning, and ambient intelligence. Each puter science discipline in the mid 1950s, and
edge acquisition, neu- AI tool is outlined, together with some examples of it has produced a number of powerful tools that
ral networks, genetic its use with sensor systems. Applications of these are useful in sensor systems for automatically
algorithms, case-based tools within sensor systems have become more solving problems that would normally require
reasoning, and ambient- widespread due to the power and affordability of human intelligence. Seven such tools are here:
intelligence—can be present-day computers. Many new sensor applica- knowledge-based systems, fuzzy logic, inductive
used to improve sensor
tions may emerge, and greater use may be made of learning, neural networks, genetic algorithms,
systems.
hybrid tools that combine the strengths of two or case-based reasoning, and ambient-intelligence.
more of the tools reviewed. AI systems have been improving, and new
The tools and methods reviewed here have advances in machine intelligence are creating
minimal computation seamless interactions between people and digital
complexity and can be sensor systems. Although the introduction of AI
implemented with small into industry has been slow, it promises to bring
sensor systems, sin- improvements in flexibility, reconfigurability,
Case gle sensors, or system and reliability. New machine systems are exceed-
Database
arrays with low-capa- ing human performance in increasing numbers of
bility microcontrollers. tasks. As they merge with us more intimately,
The appropriate deploy- and we combine our brain power with computer
Input Output ment of the new AI tools capacity to deliberate, analyze, deduce, commu-
Interface Engine will contribute to the cre- nicate, and invent, then we may be on the thresh-
ation of more competi- old of a new age of machine intelligence.
tive sensor systems and AI (or machine intelligence) combines a
applications. Other tech- wide variety of advanced technologies to give
Figure 1 shows a case- nological developments in AI that will impact machines an ability to learn, adapt, make deci-
based reasoning (CBR) sensor systems include data mining techniques, sions, and display new behaviors. This is
system. As with rule- multi-agent systems, and distributed self-organiz- achieved using technologies such as neural net-
based systems, CBR ing systems. Ambient sensing involves integrat- works, expert systems, self-organizing maps,
systems are good at ing many microelectronic processors and sensors fuzzy logic, and genetic algorithms, and that
representing knowledge into everyday objects to make them “smart.” machine intelligence technology has been devel-
in a way that is clear to They can explore their environment, communi- oped through its application to many areas where
humans; however, CBR cate with other smart things, and interact with sensor information has needed to be interpreted
systems also have the humans. Advice provided aims to help users cope and processed, for example assembly, biosen-
ability to learn from past with their tasks in intuitive ways, but the reper- sors, building modeling, computer vision, cut-
examples by generating cussion of such integration into our lives is diffi- ting tool diagnosis, environmental engineering,
additional new cases. cult to predict. Using ambient intelligence and a force sensing, health monitoring, human-com-
Figures courtesy: Uni- mix of AI tools uses the best of each technology. puter interaction, Internet use, laser milling,
versity of Portsmouth The concepts are generically applicable across maintenance and inspection, powered assistance,
44 ● DECEMBER 2013 CONTROL ENGINEERING ● www.controleng.com
Input Management E E Advice
Fuzzify Apply E E Improvement
Inputs Rules Design Monitoring Energy Efficiency
Energy Efficiency
Data Analysis Knowledge
Production Problem Solving Services
Management

Aggregate Output
Defuzzify
Outputs
Repository

Figure 2 shows an architecture for a fuzzy logic- Manufacturing Unit:


based controller. Aml Data Aml
Energy Efficiency
Energy Measurements Processing
Observer
Process Data System
robotics, sensor networks, and teleoperation. User Knowledge
These developments in machine intelligence
are being introduced into ever more complex
sensor systems. The click of a mouse, the flick of Figure 3 shows an experimental system to test the use of ambient intelli-
a switch, or the thought of a brain might convert gence to improve energy efficiency.
almost any sensor data to information and trans-
port it to you. Recent examples of this research
Object
work are provided, which include work at the List
University of Portsmouth. Seven areas follow CAD Model Object Identifier Weld Identifier
Generation
where AI can help sensor systems.

1. Knowledge-based systems Image Capture Weld Start/End


Knowledge-based (or expert) systems are Points
computer programs embodying knowledge New Program
Generator
about a domain for solving problems related to
that domain. An expert system usually has two New Robot Program
main elements, a knowledge base and an infer-
ence mechanism. The knowledge base contains MOTOMAN ROBOT SYSTEM
domain knowledge which may be expressed as a
combination of “IF–THEN” rules, factual state-
ments, frames, objects, procedures, and cases.
An inference mechanism manipulates stored database and can represent the experience of In Figure 4, a proposed
knowledge to produce solutions to problems. human specialists. When a problem occurs that system flow diagram
Knowledge manipulation methods include using a system has not experienced, it compares with shows how a system
inheritance and constraints (in a frame-based previous cases and selects one that is closest to could gather data from
or object-oriented expert system), retrieval and the current problem. It then acts upon the solu- an image sensor. Visual
adaptation of case examples (in case-based sys- tion given and updates the database depending data and CAD model
tems), and the application of inference rules (in upon the success or failure of the action. Case- data will be used in con-
rule-based systems), according to some control based reasoning systems are often considered to junction to determine an
procedure (forward or backward chaining) and be an extension of rule-based systems. They are object list; that object list
search strategy (depth or breadth first). good at representing knowledge in a way that will be passed to a weld
A rule-based system describes knowledge of a is clear to humans, and have the ability to learn identifier module that
system in terms of IF… THEN... ELSE. Specific from past examples by generating new cases. will use AI techniques to
knowledge can be used to make decisions. These determine weld require-
systems are good at representing knowledge 2. Case-based reasoning ments.
and decisions in a way that is understandable Case-based reasoning has been formalized for
to humans. Due to the rigid rule-base structure purposes of computer reasoning as a four-step
they are less good at handling uncertainty and process: 1) Retrieve: Given a target problem,
are poor at handling imprecision. A typical rule- retrieve cases from memory that are relevant to
based system has four basic components: a list of solving it. A case consists of a problem, its solu-
rules or rule base (a specific type of knowledge tion, and, typically, annotations about how the
base); an inference engine or semantic reasoner solution was derived.
(infers information or takes action based on the 2) Reuse: Map the solution from the previ-
interaction of input and the rule base); temporary ous case to the target problem. This may involve
working memory; and a user interface or other adapting as needed to fit the new situation.
connection to the outside world through which 3) Revise: Having mapped the previous solu-
input and output signals are received and sent. tion to the target situation, test the new solution
The concept in case-based reasoning is to in the real world (or a simulation) and, if neces-
adapt solutions from previous problems to cur- sary, revise.
rent problems. These solutions are stored in a 4) Retain: After the solution has been suc-
www.controleng.com ● CONTROL ENGINEERING DECEMBER 2013 ● 45
sensors

cessfully adapted to the and system configuring. Specific tasks under-


Object List target problem, store the taken by expert systems are assembly, automat-
resulting experience as a ic programming, controlling intelligent complex
new case in memory. vehicles, planning inspection, predicting risk of
Critics argue that it is disease, selecting tools and machining strategies,
Case Base
an approach that accepts sequence planning, and controlling plant growth.
Rule Base Fuzzy Logic anecdotal evidence as its
Reasoning
main operating principle. 3. Fuzzy logic
Without statistically rele- A disadvantage of ordinary rule-based expert
vant data for backing and systems is that they cannot handle new situa-
implicit generalization, tions not covered explicitly in their knowledge
Decision System
there is no guarantee that bases (that is, situations not fitting exactly those
the generalization is cor- described in the “IF” parts of the rules). These
rect. All inductive rea- rule-based systems are unable to produce con-
soning where data is too clusions when such situations are encountered.
scarce for statistical rele- They are shallow systems which fail in a “brittle”
In Figure 5, the sen- vance is based on anecdotal evidence. manner, rather than exhibit a gradual reduction in
sor systems make weld The concept in case-based reasoning (CBR) performance when faced with increasingly unfa-
requirement sugges- is to adapt solutions from previous problems to miliar problems, as human experts would.
tions to a weld identifier current problems. These solutions are stored in a The use of fuzzy logic, which reflects the
module shown here. database and represent the experience of human qualitative and inexact nature of human reason-
This module evaluates specialists. When a problem occurs that a sys- ing, can enable expert systems to be more resil-
suggestions and deter- tem has not experienced, it compares with pre- ient. With fuzzy logic, the precise value of a
mines the optimum weld vious cases and selects one closest to the current variable is replaced by a linguistic description,
path. Suggestions will problem. It then acts upon the solution given and the meaning of which is represented by a fuzzy
then be passed to the updates the database depending upon the success set, and inferencing is carried out based on this
existing robot program or failure of the action. representation. For example, an input from a sen-
generator. CBR systems are often considered to be an sor system of 20 can be replaced by “normal” as
extension of rule-based systems. As with rule- the linguistic description of the variable “sen-
based systems, CBR systems are good at rep- sor input.” A fuzzy set defining the term “nor-
resenting knowledge in a way clear to humans; mal sensor input” might be: normal sensor input
however, CBR systems also have the ability to = 0.0/below 10 widgets per minute +0.5/10−15
learn from past examples by generating addition- widgets per minute +1.0/15−25 widgets per min-
al new cases. Figure 1 shows a CBR system. ute +0.5/25−30 widgets per minute +0.0/above
Many expert systems are developed using 30 widgets per minute.
programs known as “shells,” which are ready- The values 0.0, 0.5, and 1.0 are the degrees
made expert systems complete with inferencing or grades of membership of the sensor rang-
and knowledge storage facilities but without the es below 10 (above 30), 10−15 (25−30), and
domain knowledge. Some sophisticated expert 15−25 to the given fuzzy set. A grade of mem-
systems are constructed with help of “develop- bership equal to 1 indicates full membership, and
ment environments.” The latter are more flexible a null grade of membership corresponds to total
than shells, providing means for users to imple- non-membership.
ment their own inferencing and knowledge rep- Expert system knowledge employing fuzzy
resentation methods. logic can be expressed as qualitative statements
Expert systems are probably the most mature (or fuzzy rules), such as, “If the input from the
among tools here, with many commercial shells room temperature sensor is normal, then set the
and development tools available to facilitate heat input to normal.” A reasoning procedure
their construction. Once the domain knowledge known as the compositional rule of inference
to be incorporated in an expert system has been (the equivalent of the modus-ponens rule in rule-
extracted, the process of building the system is based expert systems) enables conclusions to be
relatively simple. The ease with which expert drawn by generalization (extrapolation or inter-
systems can be developed has led to a large num- polation) from the qualitative information stored
ber of applications of the tool. In sensor sys- in the knowledge base. When a sensor input is
tems, applications can be found for a variety of detected to be “slightly below normal,” a con-
tasks, including selection of sensor inputs, inter- trolling fuzzy expert system might deduce that
preting signals, condition monitoring, fault diag- the sensor inputs should be set to “slightly above
nosis, machine and process control, machine normal.” (This conclusion might not have been
design, process planning, production scheduling, contained in any fuzzy rule stored in the system.)
46 ● DECEMBER 2013 CONTROL ENGINEERING ● www.controleng.com
Fuzzy expert systems (FES) use fuzzy logic Another approach is to use logic program-
to handle the uncertainties generated by incom- ming instead of propositional logic to describe
plete or partially corrupt data. The technique uses examples and represent new concepts. That
the mathematical theory of fuzzy sets to simulate approach employs the more powerful predicate
human reasoning. Humans can easily deal with logic to represent training examples and back-
ambiguity (areas of gray) when decision making, ground knowledge and to express new concepts.
yet machines find it difficult. Figure 2 shows an Predicate logic permits the use of different forms
architecture for a fuzzy logic-based controller. of training examples and background knowledge.
Fuzzy logic has many applications in sen- It enables the results of the induction process
sor systems where the domain knowledge can (the induced concepts) to be described as gener-
be imprecise. Fuzzy logic is well suited where al first-order clauses with variables and not just
imprecision is inherent due to imprecise limits as zero-order propositional clauses made up of
between structures or objects, limited resolution, attribute-value pairs. There are two main types
numerical reconstruction methods, and image of these systems, the first based on the top-down


filtering. Applications in structural object recog- generalization/specialization method, and the
nition and scene interpretation have been devel- second on the principle of inverse resolution. Fuzzy logic
oped using fuzzy sets within expert systems. A number of learning programs have been
Fuzzy expert systems are suitable for applica- developed, for example ID3, which is a divide-
has many
tions that require an ability to handle uncertain and-conquer program; the AQ program, which applications
and imprecise situations. They do not have the follows the covering approach; the FOIL pro-
ability to learn as the values within the system gram, which is an ILP system adopting the in sensor
are preset and cannot be changed. generalization/specialization method; and the systems where
Fuzzy logic successes have been achieved in GOLEM program, which is an ILP system based
the areas of cooperative robots, mobile robots, on inverse resolution. Although most programs the domain
prediction of sensory properties, supply chain only generate crisp decision rules, algorithms knowledge
management, and welding. also have been developed to produce fuzzy rules.
The requirement for a set of examples in a can be
4. Automatic knowledge acquisition

rigid format (with known attributes and of known imprecise.
Getting domain knowledge to build into a classes) has been easily satisfied by requirements
knowledge base can be complicated and time in sensor systems and networks so that automatic
consuming. It can be a bottleneck in constructing learning has been widely used in sensor systems.
an expert system. Automatic knowledge acquisi- This sort of learning is most suitable for prob-
tion techniques were developed to address this, lems where attributes have discrete or symbolic
for example, in the form of IF–THEN rules (or values rather than those with continuous-valued
an equivalent decision tree). This sort of learning attributes as in many sensor systems problems.
program usually requires a set of examples as a Examples of inductive learning applications
learning input. Each example is characterized by are laser cutting, mine detection, and robotics.
the values of a number of attributes and the class
to which they belong. One approach is through 5. Neural networks
a process of “dividing-and-conquering,” where Neural networks can also capture domain
attributes are selected according to some strat- knowledge from examples. They do not archive
egy (for example, to maximize the information the acquired knowledge in an explicit form such
gain) to divide the original example set into sub- as rules or decision trees, and they can readily
sets, and the inductive learning program builds handle continuous and discrete data. They also
a decision tree that correctly classifies the given have a good generalization capability as with
example set. The tree represents the knowledge fuzzy expert systems. A neural network is a com-
generalized from the specific examples in the set. putational model of the brain. Neural network
This can subsequently be used to handle situa- models usually assume that computation is dis-
tions not explicitly covered by the example set. tributed over several simple units called neurons,
In another approach, called the “cover- which are interconnected and operate in parallel.
ing approach,” the inductive learning program (Neural networks also are called parallel-distribut-
attempts to find groups of attributes uniquely ed-processing systems or connectionist systems.)
shared by examples in given classes and forms The most popular neural network is the multi-
rules with the IF part as conjunctions of those layer perceptron, which is a feedforward net-
attributes and the THEN part as the classes. The work: all signals flow in one direction from the
program removes correctly classified examples input to the output of the network. Feedforward
from consideration and stops when rules have networks can perform static mapping between an
been formed to classify examples in the set. input space and an output space: the output at a
www.controleng.com ● CONTROL ENGINEERING DECEMBER 2013 ● 47
sensors

‘ Many new
sensor systems
given instant is a function only of the input at
that instant. Recurrent networks, where the out-
puts of some neurons are fed back to the same
neurons or to neurons in layers before them, are
found applications in sensor systems involv-
ing complex combinatorial or multi-parameter
optimization, including assembly, assembly line
balancing, fault diagnosis, health monitoring,
applications
said to have a dynamic memory: the output of and power steering.
may emerge, such networks at a given instant reflects the cur-
and greater rent input as well as previous inputs and outputs. 7. Ambient intelligence
Implicit “knowledge” is built into a neural Ambient intelligence has been promoted
use may be network by training it. Some neural networks can for the last decade as a vision of people work-
be trained by being presented with typical input ing easily in digitally controlled environments
made of hybrid patterns and the corresponding expected output in which the electronics can anticipate their
tools that patterns. Error between the actual and expect- behavior and respond to their presence. The
ed outputs is used to modify the strengths, or concept of ambient intelligence is for seamless
combine the weights, of the connections between the neurons. interaction between people and sensor systems
strengths of This method is known as supervised training. In to meet actual and anticipated needs. Use in
a multi-layer perceptron, the back-propagation industry has been limited, but new, more intel-
two or more of algorithm for supervised training is often adopt- ligent and more interactive systems are at the


these tools. ed to propagate the error from the output neu- research stage.
rons and compute the weight modifications for
the neurons in the hidden layers. Expanding systems, intelligence
Some neural networks are trained in an unsu- AI can increase effective communication,
pervised mode, where only input patterns are reduce mistakes, minimize errors, and extend
provided in training and the networks learn auto- sensor life. Over the past 40 years, artificial
matically to make groups with similar features. intelligence has produced a number of powerful
Artificial neural networks (ANNs) typically tools, including those here. Use of these tools in
have inputs and outputs, with processing with- sensor systems applications have become more
in hidden layers in between. Inputs are indepen- widespread due to the power and affordability of
dent variables and outputs are dependent. ANNs present-day computers. Many new sensor sys-
are flexible mathematical functions with config- tems applications may emerge, and greater use
urable internal parameters. To accurately repre- may be made of hybrid tools that combine the
sent complicated relationships, these parameters strengths of two or more of these tools. Other
are adjusted through a learning algorithm. In developments in AI that will impact sensor sys-
“supervised” learning, examples of inputs and tems include data mining, multi-agent systems,
corresponding desired outputs are simultaneous- and distributed self-organizing systems. Appro-
ly presented to networks, which iteratively self- priate deployment of new AI tools will contribute
adjust to accurately represent as many examples to more competitive sensor systems. It may take
as possible. Once trained, ANNs can accept new another decade for engineers to recognize the
inputs and attempt to predict accurate outputs. To benefits given the current lack of familiarity and
Author produce an output, the network simply performs the technical barriers associated with using these
David Sanders is function evaluation. The only assumption is that tools. The field of study is expanding. Tools and
research coordinator, there exists some continuous functional rela- methods have minimal computation complex-
School of Engineering, tionship between input and output data. Neural ity and can be implemented on small assembly
Reader in Systems & networks can be employed as mapping devices, lines, single robots, or systems with low-capabil-
Knowledge Engineer- pattern classifiers, or pattern completers (auto- ity microcontrollers. These approaches proposed
ing, University of Ports- associative content addressable memories and use ambient intelligence and the mixing of dif-
mouth, U.K. pattern associators). ferent AI tools in an effort to use the best of each
Edited by Mark T. Recent applications are feature recognition, across many processes. ce
Hoske, content manag- heat exchangers, inspection of soldering joints,
er, Control Engineering, optimizing spot welding parameters, power, tac-
mhoske@cfemedia.com. tile displays, and vehicle sensor systems. Go Online
Consider this... 6. Genetic algorithms www.controleng.com/archives December has more on
areas 5, 6, and 7, mixed systems, applications, added
A genetic algorithm is a stochastic optimiza- research, 99 references and citations, along with more
How can you apply arti- tion procedure inspired by natural evolution. A author information.
ficial intelligence tools to
genetic algorithm can yield the global optimum Or, if you’re reading the digital edition, click the headline
help other areas of control,
solution in a complex multi-modal search space to read the full version online.
automation, and instru-
mentation? without requiring specific knowledge about the www.port.ac.uk/departments/academic/eng/
problem to be solved. Genetic algorithms have
48 ● DECEMBER 2013 CONTROL ENGINEERING ● www.controleng.com

You might also like