44
3.5 Formulation of the Relationship between the Observed Data and
the Hidden State
In order to specify the relationship between the observed data
i
y
and the hidden state
i
x , we  propose  to  use  the  method  of  floating  scale  parameters.  This  can  be  achieved  by
assuming that
i
y
will follow a two-parameter family of distributions, such as gamma or Weibull,  denoted  by  a  probability  density  function  pdf.,
, ;
 
i
y f
,  where   is  the
scale and   is the shape parameter respectively. Then, by letting
 
 
 
2 ,
; 1
, ;
|
2 2
1 1
X if
y f
X if
y f
x y
p
i i
i i
 
 
3-17
the relationship between
i
y
and
i
x  can be established. The details of this establishment of the relationship will be shown when presenting the development of the model using
simulated and real data.
3.6 Modelling Development
We started the development of the model by observing the data in  Wang 2002, which shows  a  typical  stochastic  nature  of  a  bearing  life  distribution  with  vibration
information. See Figure 3-4 below.
Figure 3-4: Vibration data for six bearings
45 Figure  3-4
shows  the  data  of  the  overall  vibration  level  obtained  from  six  rolling element  bearings,  which  is  conducted  in  a  laboratory  fatigue  experiment  several  years
ago. Generally, the overall vibration level is a measure of the total vibration amplitude over  a  wide  range  of  frequencies.  The  velocity  amplitude  in  this  case  is  expressed  in
term  of  root  mean  square  rms  value,  which  tells  us  the  energy  of  the  vibration.  It  is noted,  this  is  the  only  condition  monitoring  data  that  are  available  to  us,  since  other
associated  information  about  the  experiment  such  as  size  and  type  of  the  bearings,
running speed and loading have been lost. Figure 3-4 shows that for every bearing, the
overall  vibration  level  was  fairly  flat  until  a  time  where  it  started  to  increase  rapidly. This  illustrates  the  idea  of  the  two-stage  failure  process  stated  by  Christer  and  Waller
1984 as discussed in Chapter 2.
Since this is an early developmental stage of model building, we are reluctant to use this dataset  to  develop  our  model.  Hence,  a  simulation  approach  was  purposely  used  to
approximate  the  pattern  of  the  data  and  test  the  feasibility  of  our  model.  By  using simulation,  we  know  the  true  parameters  and  the  model  used,  so  it  helps  us  to  check
whether  we  can  successfully  recover  the  model  parameters  from  simulated  data.  Also, using these simulated
i
y ~
, we tested our model to see whether it was able to predict the underlying state accurately. Generally, by simulation it is easy to validate a model under
development,  by  comparing  the  results  produced by  the  model,  since  we  know  all  the
1
l ‘s and
2
l
‘s.  Figure 3-5 below shows the set-up algorithm for the simulation.
46
Figure 3-5: Set-up algorithm to generate simulated pattern data
We now demonstrate the modelling development through a numerical example based on our simulated data. For this example, to calculate
|
i i
x P
  we assume:
1 1
l f
=
1 1
1 l
e
3-18
2 2
l f
=
2 2
2 l
e
3-19
1 1
1
1 1
1 1
l
e l
L P
l F
 
 
3-20
2 2
1
2 2
2 2
l
e l
L P
l F
 
 
3-21
47 with
025 .
1
 
and 05
.
2
 
purposely  selected  in  order  to  repeat  the  pattern  in
Figure  3-4
. As the relationship between
i
y
and
i
x  can be treated through a conditional distribution
|
i i
x y
p
,  a  Weibull  distribution  is  chosen  for  the  time  being.  The distribution of
i
y
given
i
x  is given below:
2 1
ˆ ˆ
|
ˆ 1
2 2
1 1
1
2 1
 
 
 
 
 
i i
y i
y i
i i
x x
if if
e y
e y
x y
p
i i
 
 
 
 
 
 
where
2
ˆ
 
 
i i
t t
i
dl l
f dl
l f
l t
b
1 1
1 1
1 1
1 1
1
1
 
3-22
Equation 3-22 requires explanation, particularly when
2 
i
X
. It is known that during the normal working period of
1
l , the measured condition  information data may display no  particular  trend  and  fluctuate  around  a  constant  mean;  then  by  letting  the  scale
parameter  be
1
  while  the  shape  parameter  is  fixed,  we  have
1
1 1
|
 
i i
X Y
E
. During the second stage, the readings may start to increase and display a trend, in which
case  we  assume
2 |
i i
X Y
E
2
1 
 1
1 1
l t
b
i
 
 
 ; that  is, the  expected  value
of the observed data when
2 
i
X
is approximately a  linear  function of the time since the  start  of
2
l ,  with  the  intercept  approximately  equal  to  the  expected  value  of  the observed data in stage 1. It is straightforward to show that
2
 1
1 1
1
l t
b
i
 
 
 3-23
Since  we  don‘t  know  exactly  when
1
L  finishes,  we  use  an  expected  value  for
2
 conditional on
i
t L
1
, which  is given by  equation  3-22. Figure  3-6 demonstrates the
working principle of the above modelling idea.
48
Figure 3-6: Relationship between
i
y
and
i
x
The above is just one of many possible ideas to model the relationship between
i
y
and
i
x  motivated by the vibration-based monitoring cases and the real-life data we have. It is  possible  to  have  an  exponential  function  instead of  a  linear  relationship  between
i
y
and
i
x , but let us use the linear case first.
3.7 Numerical Examples