CONSISTENT DATA AND DECISION FUSION OF
HETEROGENEOUS INFORMATION DENOISING IN COMPLEX
SYSTEMS DIAGNOSIS
Mincho Hadjiski
Institute of Information and Communication Technologies – BAS;
University of Chemical Technology and Metallurgy,
Lyubka Doukovska
Institute of Information and Communication Technologies – BAS
Keywords: Data Fusion, Data Processing, Decision Making, Denoising, Mill Fan, Vibrodiagnosis.
Abstract:
In the diagnosis of complex industrial systems arise a lot of sever problems to solve due to the
heterogeneous information sources, a large number of directly unmeasurable variables, which should be
replaced by softsensing, big uncertainty of current information, temporal uncoherency of some
measurements because of the very different requirements for the spectral window of corresponding signals
in the different stages of the FDD (Fault Detection and Diagnosis) procedure. In the paper a hybrid
approach of multistep procedure is considered for denoising of diagnostic information in order to achieve
more realistic and more effective decision in a comparison with the conventional statistical approaches
using some techniques from the Computational Intelligence like Neural Networks and Case- Based
Reasoning. The main statements accepted in this investigation are: the different stages of complex
diagnosis could require different information, different methods of partial diagnosis and different methods
of decision making; the main method of hybridization is accepted to be consistent data and decision fusion;
signal processing in particular diagnosis stages should be relevant to the main diagnostic goals in the stage.
In the paper the proposed method for consistent fusion
of data and decisions is implemented for on-
line vibrodiagnosis of mechanical condition of the industrial mill fan of steam boiler in Power
plant.
1 INTRODUCTION
Modern diagnosis significantly extends the scope of
measurements, information and methods used, as
traditional techniques themselves turns out unable to
work in the case of complex problems [10, 16].
Along with approved approaches based on models,
intelligent methods using various techniques of
computational intelligence increasingly enter:
Neural Networks [16], Case Based Reasoning
(CBR) [3,13,15,16], data and decision fusion [3,5,6].
A combination of different intelligent methods is
increasingly observed [5,16,8] for using the
advantages of each component in the hybrid scheme.
Diagnosis of complex technological plants is a key
element in the rapidly evolving field of Condition
Based Monitoring and Maintenance (CBM)
[1,9,10,12,14,16]. The complexity of CBM requires
a new approach in decision making, given the great
variety of possible solutions and the significant
uncertainty [1,9,12]. The intelligent methods for
decision making are an area of intense research in
recent years [1.12].
The present work presents a method for intelligent
decision making, based on consistent fusion of data
and knowledge in conditions of heterogeneous
information, large uncertainty and nonlinearity,
using some ideas of last achievements in the area of
signal denoising some new structures are proposed
in order to overcome the drawbacks in complex
systems diagnosis.
It is recognized that the methods of hybridization in
the field of intelligent decision making are
promising approach, but require specific research to
expand all its features for at least sufficiently wide
class of problems.
163
Hadjiski M. and Doukovska L.
CONSISTENT DATA AND DECISION FUSION OF HETEROGENEOUS INFORMATION DENOISING IN COMPLEX SYSTEMS DIAGNOSIS.
DOI: 10.5220/0005415401630169
In Proceedings of the First International Conference on Telecommunications and Remote Sensing (ICTRS 2012), pages 163-169
ISBN: 978-989-8565-28-0
Copyright
c
 2012 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

2 PROBLEM STATEMENT
In the diagnosis tasks of complex technological and
mechanical plants emerge a set of difficulties. This
paper tries to handle with them using intelligent
techniques to obtain more realistic and more
effective solutions than the existing conventional
technologies.
Some assumptions are stated below:
 The problem of deep diagnosis is decided as
multistage procedure.
 The different steps may require a different
amount of information, different methods for
diagnosis and different methods for decision
making.
 As a basic method for hybridization scheme
with weighted linear combination of solutions
is preferred.
 The weights in fusion are determined by
various metrics of closeness.
 The quality of decision-making procedure of
diagnostic solution is improved by:
 Training to optimize the weights;
 Inclusion of variable number of
components in the procedure for decision
making;
 Using different number of information
resources.
The task becomes the combining of three different
approaches – intelligent filtration with analytical
models, logical analysis based on Case Based
Reasoning and conventional statistical analysis to
achieve better diagnosis decision making quality.
3 DATA AND DECISION FUSION
3.1 Data fusion
In the case of data alternatives solutions for
intelligent data processing include:
 Rejection of corrupted sections of data (non-
stationary, with gaps);
 Replacement of data and/or sections of data
(strings) to preserve the integrity of the
processing sample (training of neural
networks, statistical analysis);
 Selection of parameters of the chosen filters
with predefined structure (moving average,
exponential smoothing);
 Correction of available data sets of
synchronized data, incompatible (data
reconciliation) in respect of the fundamental
requirements (material balance, heat balance).
Decisions are made for particular action in
the signals processing process, and in combining
several actions .
i
DM
1
i
A
1
3.2 Decision fusion at higher diagnosis
level
Figure 1 shows a generalized scheme for the
proposed multi-stage intelligent decision-making for
diagnosis tasks of complex plants and systems. A
three-stage procedure for Intelligent Decision
Making (IDM) is adopted:
 IDM1: solutions to the basic level, mainly
related to faults detection;
 IDM2: solutions to average depth of
diagnosis, most often - without specific
measurements;
 IDM3: solutions for identification of complex
faults or multiple faults.
Generalized scheme for intelligent decision making
diagnosis tasks for complex plants is shown in
Figure 2. In addition to her local feedback, the main
feedback from the plant itself is shown ( Figure 1) to
improve the quality of decisions making. Thus,
assessing both the procedure of decision making and
the actions realizing it( ).
i
A
1
Σ
Σ
Figure 1: Intelligent Decision Making
Combining the solutions multi alternative for each
level of IDM. The quality of decisions is measured
by assessing the effectiveness of diagnosis and is
improved by training based on modifying the
First International Conference on Telecommunications and Remote Sensing
164

methods of decision making at the basic level
their number, as well as the fusion
weights.
,
ij
IDM
Figure 2: Generalized scheme of IDM
3.3 Intelligent diagnosis decision
making by fusion of data,
information and knowledge.
A scheme for consecutive diagnosis through fusion
is shown in Figure 3.
Figure 3: Consistent data and decision fusion
With minor modifications, this scheme retains
its validity in diagnosis of a broad class of complex
plants from similar class.
4 APPLICATION OF THE
PROPOSED METHOD FOR
DIAGNOSIS OF MILL FAN
4.1 Characteristics of mill fans as
objects of diagnosis
Mill fans (MF) are key element in ensuring the
reliable functioning of energy boilers burning low-
grade lignite. The structural scheme of MF with the
necessary signage is shown in Figure 4, but details
are given in [7, 8, 9].
d
n
W
L
Q
СА
G
af
θ
gis
θ
Figure 4: Structure scheme,
where 1 - Row fuel bunker, 2 - Row fuel feeder, 3 -
Controller of row fuel feeder, 4 – Upper side of the
furnace chamber, 5 - Gas intake shaft, 6 - Added cold
air, 7 - Mill fan, 8 – Electric motor, 9 - Separator, 10
- Dust concentrator, 11 - Hot secondary air, 12 -
Main burners, 13 - Discharge burner, 14 -
Synchronized valves of discharge burners, 
af
–
temperature of air-fuel mixture, 
gis
– temperature of
intake drying gases, V – vibration, е – relative
electric energy consumption, B – Throughput
capacity of fuel, G
CA
– Flow rate of added cold air, n
d
– Position of discharge duct valve, – Low fuel
caloricity of working mass.
W
L
Q
Consistent Data and Decision Fusion of Heterogeneous Information Denoising in
Complex Systems Diagnosis
165

The MF diagnosis is embarrassed due to certain
circumstances:
 A lot of diagnostic parameters are hardly or
are impossible to be measured fuel
consummation, granulometric composition,
coefficient of grindability, coal quality.
 Real operation shows asymmetric wear of
operative wheel blades, variable fan and
grinding capacity between two successive
repairs.
 The measurements of abig number mill fan
variables in the DCS or SCADA system are
rather inaccurate due to the significant
changeability of the conditions for
measurements (wear, slagging, sensor
pollution) and the great amount of external
disturbances (dust and humidity of fuel,
imprecise of coal, stohasticity of temperature
of the intake oven gases due to non-
stationarity of the flame position)
 The mill fan state is multidimensional. The
basic components are grinding productiveness
B [t/h], fan productiveness W [m
3
/h] and
vibration state.
Because of listed circumstances the vibrations of MF
could be considered as nonlinear and extremely
noised with very low relation vibrosignal/noise under
certain assumptions. The MF nonlinear vibration
could be represented with the next equation [9]:
() ()() ((
tqftqftyF
dt
dy
dt
yd
PM 21
2
2
,2 +=++ ξ
))
(1)
where y is amplitude of vibration, the disturbances in
the right side of the equation may be presented as a
function of exciting mechanical disturbances
(damaged bearings, unbalanceness due to wear, etc.)
q
M
and due to operational disturbances (loading,
hydrodynamic instability) q
p
.
The exciting effect of the mode disturbances
q
P
(t) must be eliminated or to be reduced
substantially at the stage of analysis. The exciting
disturbance f
1
(.) is of a deterministic nature and it is
possible to be nonstationary if the fault evolves (e.g.
most often progressive wear leading to debalance).
The operational disturbance f
2
(.) is of a cumulative
nature (due to the co-effect of a variable loading,
change in the coal composition, hydrodynamic
instability) and stochastic. This may be used for
processing of measured vibration signal to separate
the effect from the mechanical excitement f
1
(.) of the
observed vibrations.
4.2 Experimental investigations
To verify the effectiveness of the proposed above
method of cascade data and decision fusion
historical data for 8 months work of MF are used.
The Fig.5 presents raw measurement data from
the Experion DCS system for vibrations amplitude of
the mill fan motor bearing block. The data are
collected with 1 minute interval.
0.5 1 1.5  2  2.5  3
5
0
1
2
3
4
5
6
7
Time, min 
Vibrations amplitude, mm

Figure 5: Vibrations amplitude for entire period of
observations
The following conclusions may be drawn from
these data:
 Because of the large discretization time
T
0
= 1 min these data belong to uncorrelated (due
to the big values of T
0
) random processes.
Therefore these temporal series may be used to
isolate events in the MF vibration state but not
for the detailed MF bearings’ diagnostics
because it is impossible to determine spectra of
MF vibrations in successive time intervals due to
the general non-stationarity of the process as a
result of the wearing-out of the working wheel.
 Vibrosignals demonstrate significant unstability
of the MF oscillations due to series of random
exciting powers q
p
(equation (1)) – a change in
the fuel composition, non-homogeneous filling
of sectors in the working wheel, hydrodynamic
instability due to a change in the flow for the
input and output cross sections of the MF. This
instability is also due to often interrupts and load
changes.
 There is observed non-monotonous rise of
vibrations due to the joint action of leading
factors – erosive wearing-out of the blades
leading to a debalance of the working wheel and
a random combined influence of the enumerated
above exciting the oscillations mode factors (B
B
MF
First International Conference on Telecommunications and Remote Sensing
166

– Throughput capacity of fuel, – Low fuel
caloricity of working mass,
W
L
Q
af
θ
– Temperature
of air-fuel mixture,
gis
θ
– Temperature of intake
drying gases, – Position of discharge duct
valve).
d
n
 The root mean square deviation of the vibration
amplitude
V
σ
is changed during the cycle of the
working wheel from one repair to another.This
could be used an additional symptom for an
isolation of an abnormity and also for a forecast.
 Vibrosignals must be analyzed synchronously
together with the extracts for the operational
parameters (
af
θ
,
gis
θ
, ) due to the high level
of the noise in the causal-effective relations.
d
n
4.3 Implementation of the proposed
method
In order to make denoising of heterogeneous
information from vibrosensors and a various regime
parameters a consistent data and decision fusion is
accomplished according the Fig. 6.
fusion in Mill Fan state estimation
As a main approach for data fusion Neural Networks
(NN) are accepted. The intermediate result of the
proposed intelligent filtration is illustrated in the
Fig.7.
B
ˆ
B
ˆ
V
MB
S
ГЗШ
θ
ас
θ
(
)
БГ
Zf
2
(
)
П
Пf
1
П
П
БГ
Z
ПП
θ
1
NN
П
D
ГЗШ
θ
ГЗШ
θ
СГ
G
ВВ
G
2
NN
Р
д
Q
ˆ
Н
В
Б
Z
3
NN
4
NN
1
ГЗШ
θ
8
ГЗШ
θ
5
NN
MB
S
Figure 6: Scheme of sequential data and decision

Observation period data
0 50 100 150 200 250 300 350 400 450
2 
2.5 
3 
3.5 
4 
4.5 
5 
5.5 
6 
Maximal density value

Before
replacement

A
fter
replacement

Figure 7: Maximal density values of vibrations amplitude
Consistent Data and Decision Fusion of Heterogeneous Information Denoising in
Complex Systems Diagnosis
167

4.4. Application for Condition Based
Maintenance of MF
The MF vibration state may become a rather useful
component of their diagnostics to determine their
affiliation to some zone of efficiency - S
1
– the
normal one; S
2
– partial damages; still possible
exploitation with lowered mode parameters (e.g.
loading) and measures for current maintenance
(lubrication, jamming bolt joints of MF to the
bearers, technological adjustments (angles of
rotation of valves, jalousie); S
3
– zone of serious
damages, requiring immediate stopping at the first
opportunity (stop the unit).
Each of the diagnostic states S
j
is related to a
given discrete moment of time k and it also
possesses a structure of the “attribute-value” type.
() ()()(
3,2,1,
)
=
= ikHGkS
j
(2)
The current state
(
)
kS
MB
of a mill fan is related
to some diagnostic state S
j
( k ) using a classifier of
the “comparison-with defined-thresholds” type
based on the values h
i
( k ) using a system of N rules
R
i
, for ( i=1,N ).
According the proposed above method (Fig.3) a
multistage procedure is accepted to estimate the
mill fan vibrational state – , where the defined
limits h
V
MB
S
i
t
are changed adaptively depending on the
estimate of the root-mean-square value for the
reduced noise in the registered vibrations.
The actions for technical support M are
represented as a multiset:
()
4321
,,, MMMMM =
(3)
The components
(
)
4,1
=
iM
i
are subsets with
the following components:
M
1
– change in the mode parameters in cases of
conditionally allowed diagnostic state, e.g. with 3
elements.
M
2
– current repair, e.g. with 5 elements.
M
3
– replacing elements without big breaks of
the mill fan operation, e.g. with 4 elements.
M
4
– stopping for repair, e.g. with 7 elements.
It is accepted in the present paper that the basic
part of the attributes in the problem section P and
the solution S are presented by the simplest type of
data: “number” and “symbol”. Still for some
attributes such representation by pairs “attribute-
value” is incomplete and they (especially in the
portion for the supporting activities M (13)) may
include free text or they may contain links to other
related external information. Part of this
information may not be directly used in the CBR
algorithm but it gives the operators an additional
knowledge for secondary using of archived results
from the mill fan exploitation.
Independently on the presented significant
difficulties during the determination of the vibration
state of MF - , it is advisable to include it as an
important component in the assessment of the
overall technical state of mill fan. The assessment
of the mill fan vibration state is a complex problem
due to the exceptionally big uncertainty in the
measurements which follows from the temporally
re-covered changes of multimode factors. The mill
fan vibration state ( ) is a valuable integral
indicator for its working capacity. The
determination and the usage of mill fan vibration
state indicators are realistic and profitable for the
operative staff because vibrosensors are obligatory
for contemporary decentralized control DCS
systems.
V
MB
S
V
MB
S
5 CONCLUSIONS
1. The proposed method allows resolving
complex diagnostic tasks in multistage hierarchical
sequence, which cannot be done in single stage
procedure.
2. Heterogeneous data from different resources
could be used – direct and indirect measurements,
data bases, case bases and knowledge bases.
3. At each stage of decision making process the
most appropriate method for fusion, FFD and
decision making could be used. This allows
improvement of decision making quality, based on
the specific characteristics of the considered
diagnostic problem.
4. The method allows the fusion of particular
procedures of decision making, which have
different time evaluation scope, because of the
different time characteristics.
5. This method allows estimation of the decision
quality at each hierarchical level, and based on this
to improve the particular and the common
procedure of decision making.
First International Conference on Telecommunications and Remote Sensing
168

ACKNOWLEDGEMENTS
This work has been supported by the Bulgarian
Scientific Fund to Ministry of Education, Youth and
Science under the contract No 10-0267/10.
REFERENCES
1.Advances in Intelligent Decision Tehnologies,
(G.Phillips-Wren, Ed.), Springer, 2010.
2.Belsak, A., and J. Prezelj, Analysis of Vibrations and
Noise to Determine the Condition of Gear Unit,
Advances in Vibration Analysis Research, No. 2,
2007.
3.Das,S., High Level Data Fusion, Artech House, 2008.
4.Donoho,D., Denoising by Soft-Thresholding, IEEE
Trans. On Information Theory, V.41, 1995.
5.Dong,J., D.Zhang, Y.Huang, and J.Fu, Advances in
Multi-Sensor Data Fusion: Algorithms and
Applications, Sensors, 9, 2009.
6.Elmenreich,W., A Review on System Architectures for
Sensor Fusion Application, LNCS 476, Springer,
2007.
7.Hadjiski M., L. Doukovska, CBR approach for Technical
Diagnostic on Mill Fan System, Comptes rendus de
l’Academie bulgare des Sciences, ISSN 0861-1459,
2012 (to be published). IF – 0.219.
8.Hadjiski M., L. Doukovska, St. Kojnov, Nonlinear Trend
Analysis of Mill Fan System Vibrations for
Diagnostics and Predictive Maintenance, International
Journal of Electronics and Telecommunications
(JET), Versita, Warsaw, Poland, ISSN 0867-6747,
2012, (to be published).
9.Hadjiski M., L. Doukovska, P. Koprinkova-Hristova,
Intelligent Diagnostic on Mill Fan System, Proc. of
the 6th IEEE International Conference on Intelligent
Systems IS'12, 6-8 September 2012, Sofia, Bulgaria
(to be published).
10.Iserman,R., Fault Diagnosis Systems, Springer, 2006.
11.Krishnan,S. and R.Rangayyan, Denoising Knee Joint
Vibration Signals Using Adaptive Time-Frequency
representation, Canadian Conf. on Electrical and
Computer Engineering, V.3, 1999.
12.Li Hai Cheng, Qi Zhi, Application of Evidence Fusion
Theory in Water Turbine Model, In: Intelligent
Decision Technologies (J.Vatada, G.Phillips-Wren,
C.Lakhimi, R.Howlett, Eds.), Springer, 2011.
13.Rasovska,I., B.Chebel-Mollero and N.Zerhouni, A Case
Elaboration Methodology for a Diagnostic and Repair
Help System Based on Case-Based Reasoning, Proc.
of AAAI, 2007 (www.aaai.org).
14.Rafiee, J., F. Arvani, A. Narifi and M. H. Saheghi,
Intelligent Condition Monitoring of a Gearbox Using
Artificial Neural Networks, Mechanical Systems and
Signal Processing, No. 21, 2007.
15.Recèo-Garcia, J.A., B.Diaz-Agudo, A.A.Sanches-Ruiz,
and P.A.Gonzales-Calero, Lessons Learned in the
Development of a CBR Framework, Expert Update,
Vol.10, №1, 2010.
16.Vachtsevanos,G., F.Lewis, M.Roemer, A.Hess and B.Wu,
Intelligent Fault Diagnosis and Prognosis for
Engineering Systems, John Wiley, 2006.
Consistent Data and Decision Fusion of Heterogeneous Information Denoising in
Complex Systems Diagnosis
169