Transcript Saint-Petersburg State Technological Institute

SAINT-PETERSBURG STATE INSTITUTE OF
TECHNOLOGY (TECHNICAL UNIVERSITY)
Online diagnosis of incipient faults
at technological processes on the
basis of nonlinear PCA and linear
discriminant analysis
M.R.Galiaskarov
V.V.Kurkina
L.A.Rusinov
MOTIVATION
It was necessary to
develop a diagnostic system
for potentially dangerous
process of the pyrolysis of
gasoline.
The feature of this
process was the presence of
two distinct classes of faults,
varying
in
speed
of
development by orders of
magnitude.
2
THE OBJECT OF DIAGNOSTICS
•
•
•
•
straight-run gasoline
ethane
propane
butane-propane
Pyrolysis gas
The main component
aromatic hydrocarbons
3
THE OBJECT OF DIAGNOSTICS
The process of thermal
decomposition of hydrocarbons
takes place in tubular pyrolysis
furnace.
Furnaces are the most
critical equipment in relation to
dangerous
factors
of
the
production process of pyrolysis.
Structurally, the furnace is
made of two sections: convection
and radiant. The pyrolysis process
is carried out in the pipe coils of
the radiant section.
4
THE OBJECT OF DIAGNOSTICS
800-8550С
The presence of open
flame and pipes with
gasoline inside the
furnace
Dangerous
factors
The pipelines with fuel
gas around each
furnace
The rapid cooling of the
pyrogas in quenchingvaporizing devices of the
furnaces
the
unwanted
reactions
5
THE OBJECT OF DIAGNOSTICS
stage
decoking
stage of
pyrolysis
Operating cycle
of the furnace is
700-1440 hours
stage of
pyrolysis
Solid film from carbon
adsorbed on the inner
surface of the radiant
coil, called pyrolysis
coke.
stage
decoking
The criterion for the pyrolysis furnace to change to the
mode of decoking now is the increase of the wall temperature
of the radiant coils to the permissible maximum. But it is very
fuzzy criterion not reflected the situation correctly
6
THE OBJECT OF DIAGNOSTICS
A preliminary analysis
of problematics of
trouble-free operation of
the furnace with the
maximum work interval
Uneven coke
deposition in the pipes
of the radiant coils
The most of the
process faults
lead to early coke
formation
Contributes to uneven
premature coking of
the pipes
HAZOP analysis
The abrupt and
"fast" incipient
faults
"fast"
faults
"slow”
faults
many
common
symptoms
very
slowly
The reduction
of the work
interval
75% of
process
faults
The
remaining
faults
very
unevenly
cause the
need for
periodic
shutdown
of the
furnace
poor
controlled
In addition, a general analysis of the process
showed that it is significantly nonlinear.
7
MOVING KPCA
“jumping”
moving KPCA
kernel PCA
To detect “fast”
faults of nonlinear
process
sliding moving
KPCA
wavelet transform
In this case, KPCA model for normal process is constructed
and the thresholds for T2 and Q statistics are determined traditionally
and the monitoring in a window of W samples is carried out but with
storage of new data samples from the process.
If a fault does not occur, on the basis of the stored W samples
a new KPCA model is constructed, new threshold values of statistics is
calculated and the monitoring process repeats.
8
SHEWHART CARD
Because there are no direct measurements of the degree of
coking it is proposed to determine its value by setting the corresponding
threshold of statistics T2, i.e. actually by implementing the multivariate
Shewhart card with KPCA for reducing the dimension of the task.
For slow faults
the task is in the
determining the
critical degree of
their
development,
when the
furnace must be
stopped.
multivariate
Shewhart
card
is proposed
to monitor
slow faults of
the pyrolysis
process.
KPCA with
moving
window
9
The
singular
value
decomp
osition
a. Measure new vector of
observations xk+1, center it
by previous mean and
scale on previous MSE.
Evaluate statistics Q and
T2 and compare with
previous threshold values.
are calculated
Building
the
initial
KPCA
model
is used
begins with
MONITORING ALGORITHM
-matrix of
scores
-statistics
T2 and Q
-thresholds
values
The number W is chosen experimentally
and depends on the dynamics of both "fast"
and "slow" faults. In our case it was chosen
equal to W=1200.
Operations
b. If during r (1<r<W) sequential
steps any of statistics exceeds the
threshold values, the detection of
fault is stated.
If during W steps such exceeding of
threshold values do not happen, the
new matrix Xk+1 is formed from W
stored vectors {xk+1, ..., xk+W}.
c. Construct
new PCAmodel, compute
new values of
thresholds CQ
and CT. Return
to step a.
10
MONITORING ALGORITHM
For slow
faults
(caused by
coke
deposits )
normal data
The initial
model
data corresponding to
faults caused by
coke deposits
The
Thus, the use of KPCA with moving
window for detecting faults of the first
class, and usual KPCA or multivariate
Shewhart card to detecting a critical
value of faults caused by coking is
proposed to monitor the pyrolysis
process.
statistics
T2 is
defined
11
MONITORING / IDENTIFICATION
However, to identify faults (this applies to faults of
the first class) the use of contributions into faulty statistics
proved unreliable due to the lack in most cases distinct
changes in the contributions of the variables, especially for
faults associated with the process
, %
CONTRIGPE
ТINP BUTIONS
IE/D
OF
ТE/D VARIABGIn
LES IN Q
ΔР
(NORMAL
Т
OPERATION)
80
60
40
20
0
20
40
60
80
n
12
IDENTIFICATION ALGORITHM
end
the expert
system
with frameproduction
model and
fuzzy rules
later
first
Previously for identification we used the expert system with frameproduction model and fuzzy rules. But later we decided to apply КFDA,
especially because its first stage - the mapping of initial data into the
feature space that actually implements KPCA. However, the traditional
КFDA did not give satisfactory results because of the significant blurring
of fault clusters. Therefore, for prevent blurring it is necessary to start
FDA only after fault detection
КFDA
KPCA
+FDA
13
IDENTIFICATION ALGORITHM
14
CASE STUDY
For case study
three pairs of
competing faults
were selected
S2. The
burnout of
the coil
S1. The ingress
of water in the
coil with dilution
steam
f
a
u
l
t
s
S3. The coking
in the coil
S4. Reducing the
amount of gasoline
supplied to the coil
S5. Increased coke
deposition in the
pipe of annealingvaporizing
apparatus
S6. Decrease of
water level in the
drum of boilerutilizer
15
CASE STUDY
Given that the type of kernel affects the quality of work NPCA, we studied
polynomial and sigmoidal kernels as well as linear PCA. The results is as follows
(in order of increasing the rate of fault detection):
sigmoidal kernel - polynomial kernel - linear PCA
Thus, NPCA with polynomial kernel provided the detection rate close to
linear PCA, but with significantly fewer false alarms and significantly better fault
discrimination.
16
CASE STUDY
Methods of diagnostics
-
% of correct
definition
Nonlinear DA
(polynomial kernel)
35,2%
NPCA+LDA *)
83,6%
Proposed method
99,8%
*) LDA starts from the moment of fault detection,
but continues until full development of faults
17
18