A new method of pictorial analysis and representation is proposed for
decision making and control applications in complex industrial
systems. The pictorial images generated by the proposed approach are
capable of repesentating the complex interactions of thousands of
dynamically related parameters of the system. The images can be used
by a human operator to detect incipient equipment malfunctions, faults
and process abnormalities and also can be used to specify the root
causes of these events in ways that are more precise and reliable than
conventional methods. The ideas discussed in this proposal are the
culmination of 30 years of experience of the P.I. in the theory and
application of pictorial analysis and representations in diverse
branches of engineering and science.


0. Abstract

It has often been said that a picture is worth a thousand words, or a
thousand numbers as the case may be. This is the basic premise behind
the approach outlined in this proposal for presenting vast amounts of
process information to operators that must solve complex decision
making problems associated with detection, diagnosis and control tasks
in large-scale industrial systems. With the complexity of industrial
systems increasing, human operators are often overwhelmed by
information transmitted from the process to the control room. The
theory of pictorial analysis, developed and applied by the P.I., is a
valuable methodology for data reduction, representation and
discrimination. In this research project we will further develop the
idea of pictorial representations of technological objects and apply
this methodology to industrial process decision making and control
problems that require operator interpretation or intervention.
Conventional approaches to operator decision making during complex
opersting conditions often include the use of operational rules,
guidelines and procedures manuals. The major objective of this Phase I
proposal is to demonstrate via a PC implementation of the software and
graphics the feasibility and unique characteristics of pictorial
anlysis and representations analysis applied to industrial decision
making and control problems. In this approach, we incorporate both a
design and exploration phase.  Using an iterative procedure in the
design phase, plant data is transformed into a set of visual images
that will help the operator identify different operating states of the
process. In the exploration phase, the operator uses the human vision
system's unique abilites of feature extraction to determine the
current operating state of the process from real-time images generated
from the plant data. This information forms the basis for operator
decision making; fault detection, diagnosos and control.


1. Introduction

Automation and control of large-scale industrial systems, for example
fossil fuel or nuclear power generation and distribution systems, is a
complex problem. Current implementations include geographically
distributed sensors, actuators and control processors. Decision making
and control tasks are very often organized in a hierarchical
configuration that includes direct level and supervisory level
decision and control functions. Direct level control functions are
intended to respond to disturbances and commands in the natural
time-scale of the process to maintain safe and acceptable system
performance. Human operators are usually not involved at the direct
level of the automation and control hierarchy. The decision making and
control tasks in the supervisory level can include planning for
operation, fault detection and isolation, and adaptive and
self-organizing control functions. Human operators are an integral
part of the supervisory automation and control functions, especially
during abnormal system operation, for example during large transients
resulting from disturbances, equipment malfunctions and emergency
operating situations. The sequential and modulating control systems
operating in the direct level are designed to maintain safe operation
of the process during severe transients to provide operators with the
time necessary to diagnose the problem and respond. Appropriate
operator training on realistic simulations of the process is necessary
to provide operators with experience on abnormal process operating
characteristics. In the training environment the operator learns to
associate certain characteristics of the process response with
critical operating situations and to effectively respond to mitigate a
dangerous operating situation.

Operator control rooms include many operator stations with graphical
and numerical displays and enunciator panels. Information from
sensors, actuators and local process controllers in the pant is
brought to the control room for display to the operator. One of the
most challenging engineering problems is the design of the information
management and operator display system.  In this research project we
propose to develop a method based on pictorial analysis and
representation of information as a technique for transforming raw
plant data into visual images that can be effectively used by plant
operators for fault detection, diagnosis and control in large-scale
complex industrial processes.


2. Objectives of the Proposed Research Study

There are a variety of complicated decision and control tasks in
industrial process systems that are not adequately solved by
conventional methods. These include:

		
i. Early diagnostics, forecasting and troubleshooting of different
types of faults in process equipment (e.g. small leaks, vibrations
induced by mechanical failures in rotating equipment, actuator or
sensor failures) and control system malfunctions;


ii. Preventing equipment failures and emergency situations through
proper recognition of abnormal operating conditions and the timely
initiation of appropriate decision and control strategies;


iii. Process monitoring to determine if the current operating
situation is normal or if it necessary to modify the plant's current
operating policy to guarantee the continued safe and reliable
operation of the system;


iv. Control of the system during severe transients or during start-up
and shut down of the plant, including the proper initiation of manual
and automatic operating control modes in the process.


Each of the tasks given above requires that a human operator is
present in the information and control loop of the process. Although
ruled based systems have been effective for many applications,
AI/Expert systems have not lived up to their early expectations and it
is not realistic at this point in time to consider autonomous plant
control.  In order to help operators deal with the immense data sets
that are available in most large-scale industrial process systems, we
are proposing to develop an information/operator display system that
will help the operate correlate and interpret process information from
various parts of the plant. The information processing and operator
display will depend on the particular function, for example fault
detection and diagnosis, emergency plant control, start-up and shut
down, etc.

The main objectives of the first phase of this research project are:

i. To select the most important unsolved operational decision and
control problems that involve a human operator in industrial
processes. Typical examples include incipient fault detection,
e.g. early detection of leaks in a feedwater heater string in a power
generating plant, correlation of various sources of vibration and
operating data to detect mechanical problems in large rotating turbo
machinery, early detection of turbine vane cracks, etc.;

ii. To develop computer models for selected applications as determined
in i. above. These computer models will generate representative
operating data for different operating states of the process,
e.g. normal versus faulty operation (e.g. tube leaks in a heat
exchanger process);

iii. To develop methods, algorithms and preliminary software for
pictorial man-computer analysis of the data generated by the computer
models. The intent of the data processing is to find the dependencies
between selected operating states of the process and the measured data
available for decision making and control functions;

iv. To develop representational forms of the plant data so that human
operators can identify the important operating situations in a control
room environment.


REMARK: Although in the Phase I research effort we are focusing on the
application of pictorial analysis to operator based decision making in
industrial systems, the approach has considerable promise in the area
of total plant automation. In the automation problem, the role of the
operator in recognition and decision making is replaced by an
automatic (algorithmic and rule-based) system. We will address this in
detail in the Phase II and Phase III research programs.


The models that we are referring to above are intended to only
generate simulated plant data for the particular applications
chosen. Such simulation experiments are necessary in the Phase I
effort where we will be demonstrating the feasibility of the proposed
approach. In the later phases of this work, Phases II and III,
implementation on commercial distributed control hardware will be
undertaken and real plant data or realistic training simulators will
be used to evaluate the concept.


3. Proposed Modeling Approach for Pictorial Representations
	
For the types of decision making and control problems to be addressed
in this study, each state of a technological object is specified by a
data vector X(n)={xn(1), xn(2), ..., xn(D)} where n denotes the sample
path or realization of the data vector and for example xn(1) is the
temperature, xn(2) is the pressure, etc. Here D denotes the
dimensionality of the data vector and D can be on the order of
thousands to tens of thousands of variables in many applications,
e.g. a nuclear power plant. If each value xn(i) is represented along
an independent orthogonal axis, the data vectors generate a
D-dimensional space E(D) with separate points corresponding to
different realizations X(n), n=1,2,... .

Even if only a few parameters xn(i) are available for process
monitoring, the representation problem may still be high dimensional
because of the dynamics (non-stationary) characteristics of the
process. That is, several realizations of the data xn(i), n=1,2,..., N
are needed for an effective solution. In a typical application, for
example engine diagnostics [ ], several hundreds of samples of 8 to 12
parameters are checked. This requires that thousands of pieces of data
are simultaneously analyzed (each failure state of the engine is
characterized by a very complicated trajectory in E( D ) ).

There are a variety of technical issues that need to be addressed: 
         
3.1 Pattern Recognition Learning: A learning sample is made of
K classes (collections) of vectors { X(n,k) } where k=1,2..., K, where
K denotes the sequential number of the class and n denotes the
specific realization of the sample. Every class consists of data
vectors which are known to be in similar operating states for a
particular operating condition that is to be monitored and
diagnosed. The operating state of the process, for example normal or
faulted, can be verified by experts, by direct possibly invasive
inspection of process equipment, or it can be determined from
historical data from the process. In the later instance it is
necessary to learn how to distinguish between various operating states
of the process, i.e. the corresponding regions of E(D) in our
representation of process data. Fault detection depends on the
classification and recognition of patterns or features in the data
that correspond to process conditions to be detected.

Every state that is to be monitored for diagnostic or analysis
purposes is represented in the learning sample by a set of its
realizations. Collectively, these realizations are used to account for
the variability of the process data.

		
3.2 Pattern Recognition, Self-Learning, Taxonomy and Clustering 	
	
If there are new classes of states, unknown in advance and
indistinguishable by available methods, it is necessary to
automatically identify coherent classes for these states. The learning
sample is not divided into classes apriori and the analysis is
expected to indicate those decompositions that follow from the
structure of the data and make possible a meaningful interpretation of
the data in terms of the features of the process data. Once such a
decomposition has been determined, the process model can be chosen or
updated, the process dynamics can be described in detail, and
monitoring, detection and diagnosis functions can be improved.
	
In the next section we discuss the limitations of conventional
approaches to deal with these and other technical issues. Usually the
conventional approaches discribed in the next secion are used in an
automatic operating mode. They can alsobe used in preprocessor plant
data in an operator assisted mode. This is the context in which we are
discussing these methods.


4. Limitations of the Conventional Approaches

4.1 Discrimination Methods of Pattern Recognition and Clustering
	
In this approach a solution is searched for as some combination of N
hyperplanes or hypersurfaces. The computational complexity (C) of
searching for these hypersurfaces is directly proportional to the
problem dimension D, a nonlinearity measure for the problem- N and the
number of classes K, i.e. C = DN*(K -1).

Experience with this approach has shown that only problems with 
C < 2,000-5,000 are solvable on modern medium sized computers, and, at
the same time, that there exists real world problems with complexity
much greater than 5,000.  However, to achieve discrimination in
problems with complexity measures < 10,000, an expert with experience
and knowledge of the problem domain has to choose the informative
features for the patterns, choose the solution algorithm, solution
class, the number N, initial positions of the separating
hypersurfaces, optimization criteria for discrimination, etc.

What may be even more important is that these methods are usually
based on an optimization approach where a local extremum in the
feature space is sought. Consequently, a screening of all of the local
extremum is required and this is a computationally complex procedure.
Also, it is often very difficult for a human operator to interpret the
results of the analysis. Also according to specialists, most methods
of computer pattern recognition are approximately equal, and the
overall performance of a particular approach depends most critically
on the expertise of the problem domain experts.

However, because of the lack of effective visual representations of
complicated time series of data, the unique pattern recognition
abilities of the human vision system are almost never used in these
conventional methods. Overcoming this drawback of most pattern
recognition approaches is one of the unique features of the work
proposed here. By beginning with the assumption that feature
discrimination will be accomplished by the human operator, we will be
able to develop pictorial representations for non-stationary process
data that are designed to take advantage of the unique processing and
feature extraction capabilities of the human vision system. This may
simultaneously lead to automatic recognition algorithms that are much
different than those that are in common use today.

Artificial Neural Networks (ANNs) have found numerous applications in
pattern recognition. Through training, usually formulated as a
mathematical programming problem, ANNs can learn a nonlinear
discrimination rule to separate features in the data. However, ANNs
are plagued by the same problems as described above. In particular,
for large data sets with strong nonlinear dependencies in the data,
the networks are very difficult to train and computational time can be
prohibitively large. Another drawback of the ANN approach is that it
is very difficult, if not impossible, to gain insight into the
clustering mechanism by examination of the network parameters, for
example the weights and thresholds of the network. Therefore, it is
particularly difficult for an uninitiated user, like the system
operator, to interpret the results of the network solution and use
this information to gain insight into the operation of the process.
For this reason, it is very unlikely that operators will have
confidence in the results generated by such a technique.

4.2 Projection  Pursuit and Visualiztion Methods	

The conventional problems encountered when using automatic pattern
recognition techniques have stimulated the development of new methods
that are based on a visual analysis of multidimensional spatial
structures of data samples. This is accomplished by constructing
appropriate mappings for two- and three-dimensional structures [20].

There are several possible approaches:
	
1. One method involves projecting the initial sample onto various
planes of its main (statistically optimal) components ( this is
commonly referred to as principal component analysis). Methods of this
type are linear and can only be applied to problems with 
D < 40-50. The primary restriction is that a particular data matrix is
of full rank and an inverse must be computed.

2. The "multidimensional scaling" type of methods automatically
arrange a set of L data points on a plane.  This can be accomplished
by optimizing some performance functional, e.g. a standard linear
least squares problem, that captures the important characteristics of
a sample. This reduces to the problem of multiextremal optimization
for L variables [27].

3. Other methods are based on systems of the Peano-Hilbert type with
imbedded developments (scans) which generate, for example, the density
curves of the sample vectors distributed along regular
trajectories. Evidently, it is very difficult to estimate visually the
features of multidimensional, even two- or three-dimensional
structures by such a curve. One has to use the logical synthesis of
the observed picture with the knowledge of the development trajectory
[27]. There are other visualization methods that are also used, but
all of them use only a small part of the overall capability of the
human vision systems [6,16, 24, 25, 35].

4.4. Statistical Analysis of Time Series Data	

Methods based on the statistical analysis of time series can only be
applied to the analysis of stationary processes, unfortunately for the
applications of interest in this proposed work, e.g. fault detection
in industrial processes, the most informative data is obtained in the
transient regimes of operation ( see section 7.), where time series
mehtods can not be applied.


4.5. Expert Systems

Expert systems primarily use verbalized expert knowledge and
limited-subconscious expert skills and "know-how." Difficulties often
encountered include differences of opinions from two experts
commenting on the same aspect of a behavior and the difficulty for a
novice in the problem domain, the expert system builder for example,
to ask the correct questions to obtain the necessary information to
construct an appropriate knowledge base. Pictorial methods of data
representation, recognition and analysis as proposed in this work can
be used as a tool to interact with domain experts and also to
incorporate the history of past operational data into the knowledge
base of the expert system. Such an enhancement to the knowledge base
of an expert system may help to increase the eventual practical uses
of such systems in real-time industrial decision making and control
problems.


4.6. Decision Making of the Human Operator

Some problem solving for control and diagnostics would be impossible,
ineffective or inexpedient without the human operator. Today, the
design of information displays (a combination of tables, text, hard
panels, etc.) requires an operator to solve many complex formal-logic
type problems in order to reach the desired conclusion. For example,
consider the problem of an operator attempting to estimate the
operating state of an industrial process from real-time plant data in
order to make a control decision. Information and decisions might
include : Valve A is open, parameter B is in some definite range,
variable C is decreasing, etc.; this means that variable T needs to be
decreased to maintain acceptable operating performance [1]. This form
of decision making is referred to as logical decision making. Here,
inferences on the state of the system are made using plant data. This
is often cumbersome and difficult for human operators, especially
during abnormal operating conditions of a complex nonstaionary
process. An alternative is suggested in this work where plant data is
transformed via pictorial analysis and representation such that
decision making by the operator is accomplished in the image domain.

Psychological experiments [] and other results from the literature []
have shown that operators can not simultaneously handle more than 7-12
variables for decision making tasks. For more complicated problem
solving either a hierarchy (tree) of windows and/or an expert systems
are required. The first enhancement to the operator
information/decision making system requires that the operator look
through many levels of windows to obtain the relevant information. The
operator is required to search through various branches of a tree
structured data base. This increases the decision making time and very
detailed operator training is required. Some of the restrictions and
limitations for the use of expert systems have been given above.


	
5. The Potential of Pictorial Representations 

The problems of fault detection, diagnosis and decision/control in
complex industrial systems potentially requires the estimation of
hundreds of dynamic parameters with interrelations. We propose to
develop methods incorporating integral images that can be realized by
pictorial representations (contours, matrices, structural
representations, etc.). Examination of previous results [34] reveals
that pictorial representations can essentially decrease errors in
system diagnostic functions and also can reduce the stress on an
operator induced by excessive information in text or tabular form,
thereby increasing their operating efficiency and effectiveness. Next
we discuss how human visual processing of images can be used in
conjunction with pictorial representations of data.


Multidimensional structures and objects can be visually analyzed
without using spatial relationships. Humans are capable of
distinguishing and comparing several images in parallel with hundreds
of local attributes and features relating to the form, color, or
brightness of the images. These attributes and features are then
organized into a multilevel hierarchy that can be partially
verbalized. Rapid movement through this hierarchy, searching for more
details or generalizing the attributes allows for the simultaneous
examination of many facets of the image in terms of a variety of
attributes (or features). Comparatively, conventional computer
implementations of algorithms requires a sequential search. Refer to
Fig.1 [2,3] .

The main idea of the pictorial analysis of data described below is the
analysis of multidimensional relations through the maximum use of the
human operator's potential for transforming the initial data
(heterogeneous arrays, signals and fields ) into pictures without
information compression (degeneracy), i.e. initial data visualization
[26].

Although the notion of an oscillogram and diagram have been previously
used to analyze multidimensional data [7,9], this analysis is usually
empirical, and in fact does not take advantage of the human vision
processing capabilities. In the scheme proposed in this work, we will
capitalize on the ability of the human vision processing system to
distinguish between images using feature discrimination.  A
transformation of the process data into a pictorial representation is
used in conjunction with a human expert to discriminate features in
the pictorial representation of the plant data. With this approach we
can effectively deal with complex data structures and human operator
decision making for fault detection, diagnosis and control
applications in real-time applications.


5. The General Scheme of Pictorial Analysis for Contorl System Design
(refer to Fig. 2 )[ 27] 

An initial data sample W ={ X(n) }, with the cardinality of W = L, is
entered into the computer. The descriptions of the vectors X(n)
contained in W are transformed ( for example, from spectral, or
statistical representations, see Fig. 3) to a pictorial representation
of X(n), or groups of vectors in W, in the form of a separate picture
P(n), i.e.  X(n) -> P(n), refer to Fig. 4. The human expert, or
possibly an expert system provides a visual comparison of the various
pictures P(n) and determines the attributes {q} that can be used to
distinguish the various vectors or groups of vectors in W. The
descriptions {Q}, selected from the attributes {q}, provide a set of
distinguishing classes or informative clusters for the pictorial
representations of the data P(n). Through an iterative approach, the
most effective pictorial representation, i.e. the one that leads to
the most effect set of discriminants, is selected.

This approach has the distinguishing feature that the computer only
transforms initial data into the pictorial representation of the data
and the human expert, making the maximum use of human vision
capabilities for problem solving, determines the appropriate
representation and the corresponding features for discrimination for
the intended application.

Indeed, any reliable apriori knowledge should be used to facilitate
the choice of the above transformations and the corresponding
features. It has been shown that the human-computer interactions
described above is very effective because human perception and
cognition abilities that are capable of solving complex problems are
used in the approach [32,33]. Here, the computer is relegated to
numerical and graphical processing tasks that it can perform in an
efficient manner [30,31].

The technique we are proposing can be decomposed into four basic
steps, steps 1-3 are the design phase of the method and step 4 is the
exploration phase of the method. The procedure can be outlined as
follows:


        Step 1: Data obtained from the plant is transformed into pictorial
representations;


	Step 2: Using this pictorial representation of the plant data,
distinguishing features and attributes are identified by problem
domain experts based on the application being considered;


	Step 3: Iteration between steps 2 and 3 continues until a set
of pictorial images that capture the necessary information have been
obtained;


	Step 4: The pictorial images that result from step 3 are
templates for pattern matching by the operators. During real- time
operation of the plant, data is transformed into pictorial images
using the same transformations that were used to generate the
templates for step 3. Then, trained operators match real-time patterns
with the templates for fault detection, diagnosis or control tasks.


6. Main Advantages of Pictorial Methods
	
For many years the P.I. has been developing a principally new
methodology for the solution of complex multidimensional decision
making problems using the pictorial representation and analysis of
data. The objective is to effectively use a human expert to analyze
pictorial representations of multidimensional data which are initially
patterns of heterogeneous data, arrays, signals, and fields.

The technique is based on a computer-aided dialogue incorporating a
pictorial representation of multidimensional data that allows a human
expert through visual examination to detect complex nonlinear
multidimensional dependencies in the original data. The major
advantages of the pictorial method are that the technique:

	
	1.  can handle high dimensional problems,

	2. can choose from a number of extremal nonlinear
	solutions for the representation,


	3. the results can be interpreted at any stage of the
        procedure,

	4. at any stage of the procedure new classes of  
  	states (defects) can be introduced and the initial set of
        features are preserved,


	5. individual objects can be modified to enhance recognition
	and discrimination, and

	6. can handle non-stationary time series of data, this is
	particularly important for failure detection, diagnosis and
	control applications.


Many years of experience is available for the theory and application
of pictorial methods in different fields. This work has revealed the
advantages of this approach when compared with other approaches
[15,27].  The major reason for the success of pictorial methods is
that we can deal with non-stationary and essentially nonlinear
processes with dependencies.


6.1 Types of Pictorial Displays: 

Among the pictorial representations that can be used for detection,
diagnosis and control problems associated with technological objects,
we have found that bargraphs and diagrams in rectangular and polar
coordinate systems are effective for classifying complex interactions
that include as many as two dozen parameters [27]. Three dimensional
and color-luminous matrix pictorial representations can be used to
capture the dynamical interactions between thousands of
parameters. [28,33]

For a particular application, the choice of the appropriate pictorial
representation depends on the capabilities of the human expert and
operator.  In dealing with pictorial representations, information
should be uniformly divided among all sub modalities (perception of
brightness, color, form, dynamics, texture, space, etc.) and their
"sub channels" (for example, among integral, local, position and other
properties of the form).  Each image characteristic should have the
minimal number (not exceeding 5-7) of grades, making sure that the
grades used are sufficient to produces variations within the range of
human visual perception making discrimination by a human operator more
reliable [27].


7. ILLUSTRATIVE APPLICATION: PICTORIAL FUNCTIONAL DIAGNOSTICS FOR
GAS-TURBINE ENGINES (GTE)                   

This application was developed and implemented for a Russian fighter
under the scientific leadership of Dr. V.G. Grishin.

Purpose: Early preventive functional diagnostics and forecasting of
GTE technical state during flight operation.


Basic Method: 
1. Pictorial representation of GTE thermal, gas and dynamic parameters
(operating procces parameters) in transient mode of flight; current
fuel consumption; oil system parameters; positions of passage section
controlled elements and engine controls.

2. Visual analysis and comparison of pictorial representations of
these process data by trained human-operator on the basis of special
methodology and supporting expert system.

3.  Decision making, checking and verification by the operator using
an expert system and knowledge base.


Advantages: 
     Pictorial visualization allows a human operator to effectively use the
human vision system capabilites to analyze the transitient modes of
the engine in repeated flights. As a result, it is possible to observe
the effect of incipeint failures through observable symptoms of the
malfunctions in the transitient modes of engine operation. Therfore,
improvements in the reliability of diagnostic conclusions was achieved
as compared to convential stationary analysis techniqies.


General Conclusions:
1. Average time of obtaining a diagnostic conclusion: 3-5 minutes.
2. Diagnostic period:1-3 flights.
3. Reliability of correct classification using deviations from the
norm which are expressed by recorded parameters (probability of
correct detection of symptoms) better than 0.99.
4. Systematic reliability of early detection of malfunction  0.8.-0.95
(probability of detecting a malfunction no less than a flight before
the critical event). 
5. Forecasting interval:1-5 flights.
   Using the technology of a pictorial dialog allowed for the early
diagnosis of engine malfunctions with almost 100% reliability 
resulting in timely maintenance of the engine which signifiacntly
increased to engine's availability. 


Practical Results:     
     The diagnostic system was implemented on one type of by-pass GTE
that had fuel control equipment with an electronic corrector, an
adjustable compressor and an adjustable nozzle. Real-time measurements
were recorded for ten parameters of the process. After the development
and implementaion of the pictorial functional diagnostic system for
interflight processing of the recorded data in a ground-based unit,
the following results were obtained:

 - in the early stages, before 1-20 engine hours-to-failure, 80% of 
the operating malfunctions were detected;
 - preprocessing of the data and the use of correlations practically
eliminated false detections: the probability of a false detection was
decreased by 15 times;
 - the average flight time till inflight failure was increased by 40%;
 - seven new types of malfunctions were detected and classified during
operation; this establishes the high adaptability of the system to any
deviations from the norm;
 - 70% of the detected malfunctions were correctly forecasted for 3-7 
flights.



8. Phase I Objectives

The scientific (psychological, engineering, mathematical, etc.)
fundamentals of pictorial analysis and representations have been
developed by the P.I. over a period of time. The knowledge and
expertise of the P.I., accumulated over these many years, has resulted
in the know-how for solving a variety of complex decision problems
from diverse areas of engineering and science and a wealth of
experience in using the unique potentials of pictorial analysis and
representation methods. Based on this experience we are proposing the
development of a method of data analysis and representation to enhance
the decision making capabilities of human operators in com-plex
industrial systems. The approach is based on a design and exploration
phase as outlined in section 5. The major objective of the Phase I
research program is to demostrate the feasibility of the proposed
research by developing a PC based pictorial representation software
system to illustrate the potential applications of this methodology to
fault detection, diagnosis and control in industrial process
systems. This demonstration system will be used to present these ideas
to several distributed control system vendors, for example, Bailey
Controls, Honeywell and Foxboro. The objective will be to interest
them in using this technology in their operator display stations. We
believe that the techniques of pictorial analysis and representation
will be invaluable in helping operators cope with the information
explosion that has occurred in industrial control applications. Also,
the pictorial representation approach will play an important role in
complete plant automation functions as it can provide a bridge between
the quantitative methods of engineering analysis and the qualitative
methods of rule-based systems. The results of the discussions with
distributed control system vendors will provide important information
for our Phase II proposal. We expect that if we are successful in our
Phase II efforts, this technology will find immediate applications in
the process control industry.


9. References

1. S. Cote, J. Gallagher, J. Mariani, T. Rodden, A. Scott, and
D. Shepherd, "Graphical and iconic programming languages for
distributed process control, "Proceedings of the 1988 IEEE Workshop on
Visual Languages (October 1988), pp.183-190.

2. Humphreys, G.W., Quinlan, P.T.,& Riddoch, M.J.(1989)."Grouping
procceses in visual search: Effects with single-and combined-feature
targets".  Journal of Experimental Psychology :General , 118,258-279.

3. Biederman, I.(1987)."Recognition-by-components :A theory of human
image understanding. Psychological Review", 94,115-147.

4. Humphreys, G.W., Riddoch, M. J.,& Quinlan, P.T.(1988)." Cascade
processes in picture identification".Cognitive Neuropsychology, 5,
67-104.

5. Marr, D. (1982), "Vision: A Computational Investigation into the
Human Representation and Processing of Visual Information",
W. H. Freeman and Co., San Fransisco.

6. Pickett, R. M., & Grinestein, G. G.(1988). "Iconographic displays
for visializing multidimensional data". Proceedings of the IEEE
International Conference on Systems, Man, and Cybernetics.Beijing and
Shenyang, PRC.

7. Richard A. Becker, William S. Cleveland and Allan R.Wilks, "Dynamic
Graphics for Data Analysis", Statistical Science, Vol. 2, No. 4,
pp.355-395, November 1987.

8. Levine, R., "Visualisation Barriers," Computer Graphics World,
pp.28-35, (Aug.1988).

9. Tufte, E. R., (1983), "The Visual Display of Quantitative
Information", Graphics Press, Chesire, Connecticut.

10. R. N. Haber, (1987), "Visual Perception," Annual Review of
Psychology, Vol.29, pp.31-59.

11. S. Pinker, (1984), "Visual Cognition," Cognition, Vol.18, pp.1-63.

12. H. T. Fisher (1982), Mapping Information, Cambridge, MA, Abt
Books.

13. S. Wehrend, C. Lewis, A Problem-Oriented Classification of
Visualization Techniques (1990). Proceedings"Visualization-90",IEEE
Computer Society Press, pp.139-143.

14. J. LeBlanc, M. O. Ward, N. Wittels, (1990), Exploring
N-Dimensional Databases, Proceedings "Visualization-90",IEEE Computer
Society Press, pp.230-237.

15. Farrel, E. J., (1987), "Visual Interpretation of Complex Data",IBM
Systems Journal, 26, No.2.

16. J. Beddow, (1990), Shape coding of multidimensional data on a
microcomputer display, Proceedings "Visualization-90", IEEE Computer
Society Press, pp.238-246.

17. H.Chernoff (1973). The Use of Faces to Represent Points in
k-Dimensional Space Graphically, J. Amer. Stat. Assoc. 68:361-368.

18. S. M. Cohan, D. C. H. Yang (1986), Mobility Analysis of Planar
Four-Bar Mechanisms through the Parallel Coordinate System, J. of
Mech.& Mach.Th.21: 63-

19. P. J. Hubert (1985). Projection Pursuit, Ann. Statistics,
13:435-475.

20. W. Siedlecki, K.Siedlecka and J.Sklansky (1988). An overview of
mapping techniques for exploratory data analysis. Pattern Recognition
21,411-429.

21. Cleveland,W.S. (1985). The Elements of Graphing Data, Monterey,
CA:Wadsworth.

22. Tukey, P. (1986). "A Data Analyst's View of Statistical Plots", in
1986 Proceedings of the Section on Statistical Graphics of the
Amer. Stat. Assoc., pp.21-29.

23. Squire, L. R. (1987). "Memory and brain". New York: Oxford
University Press.

24. Wang P.C.C.  (1978), editor, "Graphical Representation of
Multivariate Data", Academic Press, New York.

25. Mihalisin, T., " Visualizing Maltivariate Functions, Data and
Distributions," IEEE Computer Graphics and Applications, Vol. 2,
No. 3, 28-35, (1991).

26. V. G. Grishin (1977).Man-machine interaction technique for
control, diagnosis and classification algorithms search. Measurement,
Check, Automation. No.2(10), Moscow, pp.3-10.

27. V. G. Grishin (1982) "Pictorial Representation of Experimental
Data" Nauka Publishing, Moscow, pp.238.

28. V. G. Grishin with co. (1991) The Color Coding Models For Data
Matrix and Scalar Fields. Institute of Control Sciences Publishing,
pp.48-

29. V. G. Grishin (1975)."Errors of analog methods of dynamic
spectroscopy." Automation & Remote Control (Transl.from Russian),
Plenum Publ. Corp., No.10 pp.
 
30. L. Verin, V. G. Grishin (1986) Matrix representation of
multidimensional binary data samples , No.1 pp.97-105.

31. V. G. Grishin Algorithm for interactive forming matrix data
represantation and estimation of its efficiency. Pattern Recognition
Letters (1986),4, pp.193-200, North-Holland.

32. V. G. Grishin (1982) Pictorial analysis of data structures and
objects. in Preprints IFAC/IFIP/IFORS Int. Conference"Analysis,Design
& Evaluation of Man-Machine Systems", Pergamon Press: pp.319-327.

33. V. G. Grishin (1985) Complex visualization of measurement
fields. in Preprints IFAC/IFIP/IFORS Int.Conference "Analysis,Design &
Evaluation of Man-Machine Systems", Pergamon Press: pp.57-65.

34. V. G. Grishin (1988) CAM operator functions and pictorial
representation of information, in Preprints IFAC/IFIP/IFORS
Int.Conference "Analysis,Design & Evaluation of Man-Machine Systems",
Pergamon Press: vol.2,pp.422-425.

35. B. Kleiner, J. A. Hartigan (1981).  Representing Points in Many
Dimensions by Trees and Castles J.Amer.Stat.Assoc. 76:260-269.