prognosis of gear health using stochastic dynamical models with online parameter estimation
DESCRIPTION
Prognosis of gear health using stochastic dynamical models with online parameter estimation. Matej Ga š perin, Pavle Bo š koski, Đani Juričić “ Jožef Stefan ” Institute, Ljubljana, Slovenia. 10th International PhD Workshop on Systems and Control a Young Generation Viewpoint - PowerPoint PPT PresentationTRANSCRIPT
Prognosis of gear health using stochastic dynamical models with
online parameter estimation
10th International PhD Workshop on Systems and Controla Young Generation Viewpoint
Hluboka nad Vltavou, Czech Republic, September 22-26, 2009
Matej Gašperin, Pavle Boškoski, Đani Juričić“Jožef Stefan” Institute, Ljubljana, Slovenia
Motivation An estimated 95% of installed drives belong to older generation
- no embedded diagnostics functionality- poorly or not monitored
These machines will still be in operation for some time!
Goal: to design a low cost, intelligent condition monitoring module
Prognosis of gear health using stochastic dynamical models with online parameter estimation
Outline Introduction Experimental setup Feature extraction State-space model of the time series Parameter estimation using EM algorithm Experimental results Conclusion
Prognosis of gear health using stochastic dynamical models with online parameter estimation
Problem setup Gear health prognosis using feature values
from vibration sensors Model the time series using discrete-time
stochastic model Online parameter estimation Time series prediction using the estimated
model Prediction of first passage time (FPT)
Prognosis of gear health using stochastic dynamical models with online parameter estimation
Experimental setup Experimental test bed with motor-generator
pair and single stage gearbox
Prognosis of gear health using stochastic dynamical models with online parameter estimation
Experimental Setup Experiment description• 65 hours• 390 samples• constant torque (82.5Nm)• constant speed (990rpm)• accelerated damage mechanism (decreased
surface area)
Prognosis of gear health using stochastic dynamical models with online parameter estimation
Experimental Setup Vibration sensors
Signal acquisition
Prognosis of gear health using stochastic dynamical models with online parameter estimation
Feature Extraction
For each sensor, a time series of feature value evolution is obtained
Prognosis of gear health using stochastic dynamical models with online parameter estimation
State space model of a time series Consider a time series as an output of a
SDSSM:
Assuming that functions f and g are linear and that the system has no measurable inputs leads to the following form:
)(
)(
Θeuxy
Θwuxx
ttt
tt1t
,,,g
,,,f
ttt
tt1t
eCxy
wAxx
Prognosis of gear health using stochastic dynamical models with online parameter estimation
DSSM estimation via EM algorithm Expectation – Maximization is applied as an
iterative method to calculate a maximum likelihood estimate of unknown parameters (Θ).
And the iterative scheme is given as:
Algorithm alternates between maximizing the likelihood function with respect to hidden system states (M-step) and with respect to unknown parameters (E-step)
}{ 00 Σ,xR,Q,C,A,Θ
ΘXYΘ ΘY,|XΘ
|,logmaxarg1 pEkk
Prognosis of gear health using stochastic dynamical models with online parameter estimation
E – step of EM algorithm In dynamic systems with hidden states, E-step
directly corresponds to solving the smoothing problem.
Given an estimate of parameter values
an optimum estimate of system states, can be obtained by employing the Rauch-Tung-Striebel (RTS) smoother
}{ kkkkkkk 00 Σ,x,R,Q,C,AΘ
)'(
)(
ttk
t
E
Ek
xx
x
ΘY,|X
ΘY,|X
Prognosis of gear health using stochastic dynamical models with online parameter estimation
M – step of EM algorithm Maximization of the objective function l(Θ,
Θ’), rather than data log-likelihood function
l(Θ, Θ’) is bounded from above by L(Θ) increasing the value of l(Θ, Θ’) will increase
the value of L(Θ)
)',()',()( ΘΘΘΘΘ VlL
Prognosis of gear health using stochastic dynamical models with online parameter estimation
M – step of EM algorithm Complete data log-likelihood l(Θ, Θ’):
)()'(||log
)()'(||log
)()'(||ln)',(2
1
1
11
11
0010000
tt
N
ttt
tt
N
ttt
n
n
l
CxyRCxyR
AxxQAxxQ
μxΣμxΣΘΘ
N
tkkk
N
tkkkt
NtNtNtNtNt
xypxxpxp
xxpxxyypxxyyp
)|()|()(
),...,(),...,|,...,(),...,,,...,(
11
111
ΘΘΘ
ΘΘΘ
Prognosis of gear health using stochastic dynamical models with online parameter estimation
M – step of EM algorithm Taking the expected value of l(Θ, Θ’) with
respect to the current parameter estimate and complete observed data:
Derivatives of the function with respect to all parameters can be calculated analytically
')'()(')'()(||ln
')'()'(')'()(||ln
))((||ln=)',(
1=1=1=1=
1
111=
11=
11=1=
1
00100
CxxCyxCCxyyyRR
AxxAxxAAxxxxQQ
μxμxPΣΣΘΘ 00n0
tt
n
ttt
n
ttt
n
ttt
n
t
tt
n
ttt
n
ttt
n
ttt
n
t
nn
EEEEtrn
EEEEtrn
trl
Prognosis of gear health using stochastic dynamical models with online parameter estimation
Estimation setup Model structure
Model parameters are estimated using a time window of 100 samples
Convergence is achieved when log-likelihood increase is less than 10-4
)()()(
)()()1(
)()()()1(
11
22222
12211111
tetxcty
twtxatx
twtxatxatx
Prognosis of gear health using stochastic dynamical models with online parameter estimation
Online tracking of model parameters
Time series
Eigenvalues of system
matrix A
Diagonal elements of noise
covariance matrix Q
Prognosis of gear health using stochastic dynamical models with online parameter estimation
Prediction of first passage time
Prognosis of gear health using stochastic dynamical models with online parameter estimation
The goal of prediction was to determine when the extracted feature will exceed a certain value
For each time window, system parameters were estimated
Using estimated model, MC simulation was performed and time when the predicted value becomes greater than a certain threshold was obtained
For each time window, the simulation was performed 1000-times, so the result is a distribution of time
Prediction of first passage time
Prognosis of gear health using stochastic dynamical models with online parameter estimation
Prediction of first passage time
Prognosis of gear health using stochastic dynamical models with online parameter estimation
Conclusion Second order dynamical system is sufficient to
model the dynamical behavior of vibration feature value
Using the ML estimate of the parameters, accurate prediction of first passage time can be made 15-20 hours in advance
Comparison to linear regression:
Prognosis of gear health using stochastic dynamical models with online parameter estimation