Upload
ngotu
View
235
Download
0
Embed Size (px)
Citation preview
http://repository.osakafu-u.ac.jp/dspace/
TitleSimplified Machine Diagnosis Techniques Techniques Using ARMA Mo
del of Absolute Deterioration Factor with Weight
Author(s) Takeyasu, Kazuhiro; Ishii, Yasuo
Editor(s)
Citation 大阪府立大學經濟研究. 2008, 54(1), p.53-73
Issue Date 2008-06-30
URL http://hdl.handle.net/10466/10999
Rights
Simplified M瀋hine Diagnosis Techniques
Using ARMA Model of Absolute
DeteriorationFactor with Weight
Kazuhiro Takeyasu ・ Yasuo Ishii
Abstract: In order to diagnose machines accurately, the Kurtosis or Bicoherence method was utilized. Calculating system parameter distan鐺 was also utilized to apュ
ply time series dat� to Autoregressive (AR) model or Autoregressive Moving Averュage (ARMA) model. In thi,s paper, a simplified method of calculating autocorrelation
function is introduced and is utilized for ARMA model identification. Furthermore, an absohitedeterioration factor such as Bicoherence is introduced. 恥1achine diagnoュsis can be executed by this simplified method of calculating system parameter disュ
tance with weight. Useful results were obtained.
Keywords: ARMAmodel , autocorrelation function , cumulants , absolute deteriorac
~ion factor with weight
1. INTRODUCTION
53
In mass production industries such as steel making that have largeequipment, sudュ
den stops of production process due to machine fa匀ure can cause severe problems such
as shortage of materials to the later process , delays to the due date and the increasing
idling time.
To prevent such situations, machine diagnosis techniques play important roles. 80
far, Time Based Maintenance (TBM) technique has eonstituted the majority the maュ
chine maintenance. TBM performs checks at previously fixed times, but it performs
checks at scheduled times without taking into account whether theparts.arestill keepュ
ing in good conditions or not. On theother hand, Condit�n Based Maintenance (CBM)
performs maintenance checks by watching the condition of machines. Therefore, if the
parts are still keeping in good condition beyond their expected life span, the cost of
maintenance may be lowered because machines can be used longer than planned~
54 Simplified Machine Diagnosis Techniques Using ARMA Model
of Absolute Det巴rioration Factor with Weight
Therefore the use of CBM has become dominant, since parts maintenance cost less and
also because it has a lower 士atiooffailure. However, it is'mandatory ,to catch sympュ
toms of the failure as soon as possible if a transition 、 from TBM to CBM is to be
made. Many metho司s'have been developed f~c~lsing onthissubjé仇 In this paper, we proュ
pose a method for the early detection � the failure ónr叫ating machìne , which is the
most common theme in the machine failure detection field.
To date, many signal processing methods for machine diagnosis have been proposed
(Bolleter, 1998; Hoffner, 1991).Asfor sensitivityindices, Kurtosis , Bicoherence and Imュ
pact Deterioration Factor (ID Factor) were examined (Yamazaki, 1977; Maekawa et al.,
1997; Shao et al., 2001, Song et al., 1998; Takeyasu, 1989). Calculating system param。
ter distance was also utilized to apply time series data to Autoregressive (AR) model
or Autoregressive MovingAverage (ARMA) model (Yamazaki, 1977).
In thispåper, a simplified method of caleulatirig theautocorrelati�function is introュ
duced and utilized for ARMA model identification. Furthermore, an absQlute deterioraュ
tion factbr such as Bicon鑽ence is introduced. Machine diagnOsis can be executed by
this simplified method of calculating system par疥eter distance.
The rest of the paper isorganized as follows. , In section 2, machine diagnosistech-
niques utilizing system parameter distance are exhibited. The parameter est絈ation
method of ARMA model is stated in section 3. In section 4, moment, Bicoherence and
cumulants are introduced for the fundamental preparation. Utilizing 4th cumulants
and autocorrelation function , the parameter estimation method of ARMA model is ex-
hibited' in' section 5. Utilizing these methods , numerical examples are exhibited in sec-
tion 6. Section 7 summarizes the findings�
2. DETECTION OF FAILURE BY SYSTEM PARAMETER DISTANCE
In the analysis of tim:eseries data, AR model or ARMA model are adoptedfre-
quent1y. In this paper, we adopt ARMA model, because it is a general technique used
in time series analysis ,
Consider the (ρ , q )th order ARMA model expressed as:
p q
Xn+ 害lGzZn t=en+ 豆島en-j (1)
Here
大阪府立大学経済研究 54 ・ 1(223) (2008.6)
Simplified Machine Diagnosis Te鑄niques .UsingARMA' ModeT of Abso1ute Deterioia土ioàFactor'with Weight. 55
{Xj} : SaIhple process � astationary ιergodicGaussian proceSs x(t) ci =,1, 2,, 3;
-・ , 'N, …)
{ι} : Gaussian white noise with mean 0, and variance a; Assume that (1) satisfies the stationary condition.
Now , we calculate the~ystem p~rameter distçmce,、Modern con~r91 thE;lory shows that
ρ 豆 q when state space representation is used (Tokumaru et αl., 1977). In this paper,
we setρ== q for sim�licity.
( 2 )
(3 )
'/ b1¥l . " ・・ v
b' =1 v2 I ~ system paraフneters under normal condltion
¥bA/ I 、 pl
I b1¥1
6 弓 l 川 l 飢ma伽1 of system papram仰、s while I I abnormal contion proceeds
¥bp/ J
¥IIIlli--/¥11111111/
1
2
P
1
2
P
αα:
・
αLAαAα:
・
Aα
,frill--Il--
‘
JIlli---\
一一一一
α
《α
Set
Then, following Z[ is utilized as an evaluation function. Set'the system par瀘eters unュ
der normal condition asαi' bi , and th� estimation of system parameters as â i , bi・
ZI= ーよ-,(l = 0, 1, 2, 3, 4, 5,…) 問+1
(4 )
Where
( 5 ) 町二 (2ρ)[x 土l{(dt-U+(6z bz)2}
L= きl{(dtG仲 ( 6 )
Watching the behavior of Z[ , we can judge the system becomes abnormal when Z[ falls
below the certain value (Figure 1). Then we can make failure detection. This evaluation
•abnorma1 system
Z[=l
一一一令 time
Detecting Abnorma1 Conditions by the transition of Z[
Z{=o
Figure 1.
'54 ・ 1(223) (2008,6J 大阪府立大学経済研究
56 Simplified Machine DiagnosisTechniques Using ARMAModel
of Absolute DeteriorationFactorwith Weight
function is an absolute deterioration factor such as Bicoherence. This is 1.0 when the
system is under normal condition, and tends to be 0 when the system b.ecomes abnorュ
mal.
3. ESTIMATION OF SYSTEM PARAMETER BY ARMA MODEL
In estimating ARMA model parameters {的} {bj} , the right hand side of the equaュ
tion is itself a colored noise. Therefore, it should be noted that a biased estimation is obュ
tained by using the least squares method. In order to obtain an unbiased estimate, the
extended least squares, sequential maximum likelihood, instrumental variable, and
pseudo linear regression methods have been developed (Tokumaru et α1., 1982;
Katayama, 1994). The pseudo linear regression method, utilizes the bootstrap method
and makes iterative calculations. However, it has difficulty because there are many
cases where these iterations do not converge (Sagara et αl. , 1994; Katayama, 1994). We
have proposed an improved method for this problem (Takeyasu èt αl. , 2002). By utilizュ
ing this method, the total. calculation time can be reduced.
Introducing the cross correlation function of output and noise, and also using the
theoretical relation, the above improvement was achieved. In this paper, we introduce
a simplified method of calculating autocorrelation function in the cases where shock
waves occur, and propose simplified machine diagnosis techniques utilizing 4th
cumulants and autocorrelation function. Since 4th cumulants are 0 under the normal
distribution , cumulants can be calculated in combination using autocorrelation funcュ
tion and utilizing the formulaic relation of 4 variables.
We have previously proposed a bootstrap algorithm before in order to get an unbi時
ased estimate. This time, ouraim is to propose a new simplified calculation method
for machine diagnosis. Yule-Walker equation for 4th cumulants is introduced. Knowュ
ing that we can obtain an unbiased estimate of AR parameters under the condition
that MA items are white noise under a normal distribution (Inoue, 1992a, 1992b), the
equation for ARparameterestimation is introduced. Furthermore, we introduce a MA
parameter estimation equation which can be estimated analytically in the case of 2nd orュ
der. This is then applied to a simplified calculation method and numerical calculation
is executed. The Proposed method is simple enough to be calculated even on a pocketュ
size calculator.
大阪府立大学経済研究 54 ・1(223) (2008.6J
Simplif�d Machine Diagnosis Techniques Using ARMA Model of Absolute Deterioration Factor with W鑛ght 57
4. FUNDAMENTAL PREPARATION
4.1. About Moment
In cyclic movements such as those of bearings and gears, the vibrations grow larger
whenever the deterioration becomes greater. Also, it is well known that the vibration
grows large when the setting equipment to the ground is unsuitable (Yamazaki, 1977).
Assume the vibration signal is a function of time as, x(t). Also assume that it is a staュ
tionary time series with mean X. Denote the probability density function of these time
series asρ (x).
The average X is stated as follows:
x= よ ~ρ (x)dx (7 )
Variance σ2 is stated as follows:
σ2 = L:(X-X)2p(x)dx ( 8 )
The 3rd moment MT(3) and the 4th moment MT(4) are stated as follows respecュ
tively (Hino, 1977):
MT(3) =に(X-X)3灼)dx
MT(4) = に(X-X)4内)dx
( 9 )
(10)
4.2. About Bicoherence
Bicoherence means the relationship of a function at different points in the frequency
domain and is expressed as:
Bxxx(ん fz)Bic,xxx (iI, fz) =
l' n/ J Sx/iJ) . Sx/五) .Sxx(h十五)、}ノ
唱ai
唱EA
/,‘、
Here
XT(五)・ XT(fz) ・ X;(五十五)Bxxx(iJ, fz) = ---------;;
Tτ
(12)
means Bispectrum. Each item of right hand side of Eq.(12) is a Fourier Transform of
the original time series. See Appendix in detail.
Range of Bicoherence satisfies:
0< Bic ,xxx (ん 12) < 1 (13)
大阪府立大学経済研究 54 ・ 1(223) (2008:6)
58 Simplified Machine Diagnosis T゚chniquesUsing ARMA Model
of Absolute Deterioration Factor with Weight
When there exists a significant relationship between frequencies h and J;, Bicoherence
is near 1. Otherwise, the value of Bicoherence corr冾s close to O.
4.3. About Cumulants
The characteristic function 1JI(u) is definedas follows (Hino, 1977):
1JI(u) ==に e引 (14)
The Taylor expansion � the characteristic function is exhibited in(15). c(n) is th� coefュ
ficient.
lnV(u)=zdn)1Jfun n= ¥ n!
I (-i)ndnln 1JI(u)l c(n) = l' "/ ~
"~" ~ ,~/
I LαU Ju = 0
(15)
(16)
This c(n) is called cumulants. The following relation holds between cumulants and
the moment.
c(l) = x
c(2) = σ2
c(3) = MT(3)
c(4) = MT(4) -3 {MT(2)}2
(17)
(18)
(19)
(20)
It is well-known that when the probability density function is a normal distribuュ
tion , cumulants over the 3rd order are 0 (Inoue, 1992).
4.4. 2nd Order Autocorrelation Function
Using the following formula concerning the relation of Gaussian 4 variables
(Nakamizo, 1988):
E[αbcdJ = E[αbJE[cdJ + E[αcJE[bdJ +E[αdJE[bcJ (21)
We can get following equation for x(t) , x(t+ τ\) , x(t+ τ2) , x(t十月).
E[x(t)x(t+ τ\)x(t+ τ2)x(t+ τ3)J
= R(τ\)R( τ3一 τ2)+R( τ2)R( τ3一 τ\)+R(τ3)R( τ2一 τ\) (22)
Where, RτI(Rτ1 = R(τ\)) is the autocorrelation function of {x(t)}.
Rτ1=E[z(t〉 z(t+τ\)J (23)
Set ττ\, τ3 = 0, then we can obtain
E[X(t)2X (t+τ1)2]=R;+2Rjl(24)
大阪府立大学経済研究 54 ・ 1(223) (2008,6J
Simplified Machine Diagnosis.Techniques Using ARMA Model of Absolute Deterioration Factor with. Weight 59
Set τ1 = -l, τ2 = -l, τ3 = -i, then we can obtain
(25) E[X(t)X(t-l)2X(t-i)J = RoRi+2RIRl-i
This is the correlation function of the correlation function. That is to say that this
is a 2nd order correlation function.
5. ESTIMATION OF SYSTEM PARAMETERS ABOUT ARMA MODEL
First, the AR parameter estimation using Yule-Walker equation concerning 4th
usmg algorithm Next, the ARMA parameter estimation cumulants is exhibited.
autocorrelation function is stated.
5.1. Estimation of AR Parameter
Under a normal distribution , 4th order cumulants becomes O.
(Inoue, 1992), bymultiplying by XnX!-1 As {en} is white under 4th order cumulants
on both sides of (1) and averaging, we can get the following equation:
E[X!X!_IJ + 土lGzE[Znz; 凡 J = O(l ;:::: 1) (26)
Rewriting (26) , we can get the following equation:
C1 + 互角Cl- i = O (lミ1) (27)
where Cl-i is the 4th cumulant.
Cl-i = C(Xn, Xn- l , Xn- l , Xn-)
= E[XnX!-IXn-J
= R( -l)R (lーの +R( ~l)R (l-i)+R( -i)R(O)
= RoRi+2RIRI-i (l ;:::: 1, i = 1 ,…, ρ) (28)
about the 4th (27), following Yule Walker equation 1 toρm Varying 1 from
cumulants can be derived:
C1 αl C1-P C-1 c。
(29) C2
仇μ・
•••
c
α2 -・・ C2-P c。C1
α p c。C~_ , C p-l "p-2
Rewriting this in matrix form , we can get:
(30)
54 ・ 1(223) [2008.6) 大阪府立大学経済研究
cα = -cs
60 Simplified Machine Diagnosis Techniques Using ARMA Model
of Absolute Deterioration Fact� with Weight
Then αis solved as follows:
α=-c ICs (31)
The estimation of Cl-i(l = 1, 2,…, i = 1,…, p) is expressed as �l-i' �l-i is stated as fo1-
10ws:
ê,_, = _1 立ヮ=ム XnX~-IXη-i
l-i N -(Z-i) n= �=i+l (32)
Substituting �l for Cl in each c, cs' ê, �s is obtained. Then the estimate � ofαcan be obュ
tained as the unbiased estimate by the following equation.
α= 一 c -cs (33)
This is the same meaning that an unbiased estimate of AR mode1 parameters can be
obtained by solving Yu1e-Wa1ker equation.
5.1.1. 2ndorder case
¥1It---f
sA9
白
pup
し
//It--1\
、
γ1lj/
-
A
U
一
c
c
c
c
,,/Ili--\
一一
¥1111/
'in
,“
αα
/,
JI--1\
‘
=-(f?γ??で)R~+2R~ 3R~ / \R~+2R;/
1 ((1 +2ρî)(1 -pD\
4(2+ρî)(1一ρî) \1-2ρi+3ρ;-2ρiJ
5.2. Simplified Calculation Method ofAutocorrelation Function
In a discrete time system, samp1ed signa1 data are expressed as follows:
{XJ : i = 1, 2,…, N
ニr which is the mean of {Xi} , is stated as follows:
N
:i= 子:;- I; x, lV i = 1
For the purpose of genera1ization, :i is not set O.
Variance σ2 of {xJ is stated as follows:
N
t= 一二-:;- I; (xi- :i) 辺N-1 i=l
大阪府立大学経済研究 54'・ 1(223) (2008.6J
(34)
(35)
(36)
Simplified Machine Diagnosis Techniques Using ARMA Model of Absolute DeteriorationFactor with Weight
Autocorre1ation function is stated as follows:
1 N-[
Rl = E[x;x;+J = 一一一-;- L; X;X;牛 1 (l = 0, 1, 2,…) ι Z-ZTL.... N-I i-=-l -Z-Z-rl ,~
61
(37)
As {xi} is assumed to be a stationary time series, we can assume E[xJ = 0 without
10ss of genera1ity (Tokumaru et al., 1982). Therefore, (37) can be re-stated as follows:
R[ = E[(xi-.i) (Xi+l-.t)]
N~l
=一二---;- I; (x;-.t) (x山-.t)N-l i::-¥ , -Z けt(38)
When N is sufficient1y 1arge, the Oth autocorre1ation function is near1y equa1 to the
variance. The autocorre1ation coefficient is expressed as follows:
Rl Pl = Ro (39)
When the number of failures on bearings or gears arise, the peak value arise cycliュ
cally. In the early stage of the defect, the peak signal usually appears clearly. Generュ
ally, defects will damage other bearings or gears by contacting the inner covering surュ
face as time passes. Assume that the peak signa1 which has S times impacts from the
norma1 signa1 arises in each m times of samplings. As for determining the sampling inュ
terva1 , the well-known samp1ing theorem can be used (Tokumaru et al., 1982), but in
this paper, we do not pay much attention to this point in order to focus on our pro時
posal. Assume that the peak signa1 which has S times impacts from the norma1 signa1
arises in each m times of samplings , and a1so assume that the mean and variance are
the same except for the case where a special peak signal arises. Expressing 02 of this
case as Õ2, we get the following equation:
2 1 N o 一二-, t (x;-.t)2
N-1 i::-¥ '-z
N N-~ー
N
さ一一一竺Lσ2+ __m . S202 N-1 - , N-1 --
ミ (1+与l)σ2 (40)
For the autocorre1ation function , we express Ri in this case as Rパi = 1, 2, ・ぃ) in the
same way. The Oth autocorre1ation function is nearly equa1 variance, so we get:
Ro ミ õ2 == (1十号l)σ2 (41)
大阪府立大学経済研究 54 ・1(223) (2008.6J
62 Simplified Machine Diagnosis Techniques Using ARMA Model
of Absolute Deterioration Factor with Weight
5.2.1. Autocorrelation function of 1st order lag
In the next, we examine the feature of autocorrelation function of 1st order lag. By
definition, RI is stated as follows:
1 ペニIE 一一
Al-N-12fプ1 内向+1 (42)
When the peak signal of S times impacts from the normal signal arises in each m
times of samplings, the following ^ parts of {xJ may be considered to have peak valュues during the calculation of RI'
1, 2,… , m-2, m-1 , m , m十 1 , m十 2,… , 2m-l , 2m , 2m十 1,・・
^̂ < <
From these, the following products arise:
1 product with peak value and {(m-1)-l} products with ordinary values up to m
3 products with peak values and {(2m -1) -3} products with ordinary values up to 2m
(2k-1) products with peak values and {(km-1)-(2k-1)} products with ordinary
values up to km
2k products with peak values and {(km + 1-1) -2k} products with ordinary values up
to km+1
Therefore when N = km (case 1)
(2(N/m)-1) products with peak values and {(N- 1) 一 (2(N/m)-1)} products with
ordinary values up to N
WhenN= km+α (1 豆 α 三三 (m -1)) (case 2)
(2(N α)/m) products with peak values and {(N- 1) 一 (2(N α)/m)} products with
ordinary values up to N
When S is large, we make the following simplified calculation.
Suppose
zε < xi < .i+E
ε> 0, ci = 1, 2,… , N)
except for the case when peak signals arise. In that case we get the following relaュ
tions:
<Case 1>
When.i > 0
大阪府立大学経済研究 54 ・ 1(223) (2008.6J
Simplified Machine Diagnosis Techniques Using ARMA Model of Absolute Deterioration Factor with Weight
合{s(づ l)+(N パづト e)2 < Rj
<合{s(寸小(N- l)-(づ寸山)2When:i < 0
N~l {s(づ小(N- l) 一べ(づ一 1)ト)ドい(ω( :i -e
>tτ1{ド似仲Sベや(←2づすす一 1)ド+刊(N-1ο) 一-(2すす一 1)}山)2 (44)
Rj
ρ 豆J
~ N~l {やす-1) + (N -1) -(づ l)) ;2ー l
= {(去-tTポョ)(s 山}. mぷ-1<Case 2> When:i > 0
N~l {S(2号旦)+山)-(2 N:竺)}(:i-e)2< 昆l
<長τ{S(2 N:旦)+ω- 1) -(2 N:旦)}(:i +e)2When:i < 0
tIK44+山)_(2 N:a )}臼一ε)2> R
>tTK2t旦)+(N-1) -( 2 N:a)}(日)2In the same method as < Case 1> , we get the following equation:
Rj
ρ Ro
=長τ{や号旦)+仰ーパ
= {(去(詰サ)(S一山}. mA1
大阪府立大学経済研究 54 ・ 1(223) (2008.6)
(45)
(46)
(47)
(48)
63
Simplified Machine Diagnosis Techniques Using ARMA Model of Absolute Deterioration Factor with Weight 64
(45) and (48) are the simplified calculations of coefficient of autocorrelation function of
1st order lag. Under normal conditions , ρ1 = lwhenS = 1. WhenN→∞, S →∞, /5 1 • O.
Generally, αis very small compared with N. Therefore we make simplified calcula-
tion using (45). Under the assumption of 5.1, set m = 12. Considering the case 8=2, 4,
6, 8, 10 and the case N = 100, we obtain Table 1 from the calculation of (45).
10 0.2621
8
0.3372
with varying S
6
0.4573
'rable 1.戸l
|4
|0.6554 I 2
0.9266 S
f!J
/5 1, can be utilized as a deterioration factor. We name this as “Deterioration Factor
of the Autocorrelation Function Type of 1st Order Lag'¥
5.2.2. Autocorrelation coefficients of the 2nd order lag through 4th order lag
Then, we can get the autocorrelation function of 2nd order lag through 4th order
lag using the same method.
10 0.2620 0.2619 0.2618
Table 2. P2 through P4 with varying S
s=ρ二ρ=ρ
5.3. Estimation of MA Parameters about ARMA Model
Generally, the MA part of the ARMA model becomes a non-linearequation as fol-
lows:
P
Xn = Xn + I; aixn-i
8et
(49)
then we can get the following equation from (1):
q
tn=en+FI 鳥en-j (50)
From (49) , (50) , we set 九 as the autocorrelation function of Xn , and set bo = 1. Then we
can get the following relations which are well known (Tokumaru et al., 1982).
54 ・1(223) (2008.6J 大阪府立大学経済研究
Simplified Machine Diagnosis Techniques Using ARMA Model of Absolute Deterioration Factor with Weight
、,EBEEtrEEEEEJ
¥/
唱EA
、1JJq
+
<一
q
k>
一
fkL
凡/l¥
.,J +
ル凡
AU
b
nu
t
o
「Z=
E-y
qLe
o
rEEEEE屯EEEEI
、
一一
九
q
久 = a: I; b・j=O
65
(51)
(52)
A recursive algorithm has been developed as for this. The only parameters to be estiュ
mated are b¥ and b2・ Therefore they can be solved using the following method.
The case of q = 2;
ち =0
ち =σ';b 2
円 =σ;(b\+bん)
九 = a;(1 + bi+ bD
When a; • 1, then we get the following equations:
九= 1 +bi+b~
角 = b¥(1+b2)
ち = b2
Therefore we can obtain b¥ and b2 in the following equations.
r, b , = 一
1+ち
b2 = ち
From (49) , we get the following equations:
内 = E[xnxn+¥]
= E[(xn +a\xnー\+α2Xn-2) (Xn十 I 十 α \Xη+α2Xn-\)]
=αJ1+α2) (Ro+ R2) +α/R\十R3)+ (1 +αi+必 )R\
ち = E [XnXn+2]
= E[(九+α\Xnー I 十 α 2Xn-2) (xn + 2 +a\xn十 l 十 α 2Xn)]
=α\ (1 +α2) (R¥ + R3) +α2(Ro + R4) + (1 +αi+αDR2
b , = -:-丘一 α\ (1 +α2) (Ro+ R2) +α2(R\+R)+ (1 +αi+αDR\ 一一
1+ち 1+α\ (1 +α2) (R¥+ R3) +α2(Ro+ R4) + (1 +αi+α~)R2
b2 = ち =α\ (1 +α2) (R\ 十R3)+α2(Ro十R4)+ (1十αi+必 )R2
(53)
(54)
(55)
(56)
(57)
(58)
(59)
(60)
(61)
(62)
(63)
(64)
As mentioned above, the (2 ,2) of the ARMA model can be solved analytically. We
(65)
大阪府立大学経済研究 54 ・ 1(223) (2008:6J
66 Simplified Machine Diagnosis Techniques Using ARMA MQdel
of Absolute Deterioration Factor with Weight
can solve ARMA mode1 on1y by calcu1ating the autocorre1ation function from the 1st
through 4th order.
6. NUMERICAL EXAMPLES
6.1. Estimation of AR Parameters
Under the assumption of 5.2 , set m = 12. Considering the case S = 2, 4, 6, 8, 10 and
the case N = 100 , we obtain Tab1e3 from the calcu1ation of (34) using the va1ues of
ρ] , ρ2'ρ3'ρ4 in the Tab1e 1 ~2 of 5.2.
S
。
'rable 3. Estimation of AR parameters with varying S
2
ー0.2380
-0.2373
10
ー 0.1375
-0.1375
We get the estimations â] , �2 of the AR parametersα], a 2 ・ Ca1cu1ation mode1 data
chart is exhibited in Figure 2.
Figure2. Calculation model data chart
6.2. Estimation of MA Parameters
Under the same assumption of 5.1 , we solve b], b2, which are an estimate of b] , b2 of
MA parameters respective1y , using ρ] , ρ2'ρ3'ρ4 , and â], �2 as mentioned above.
As {xJ is assumed to be a stationary time series , we can assume:i = 0, a2 = 1 with日
out 10ss of genera1ity (Tokumaru et al., 1982).
Ro= σ2 = 1
Therefore, from (34) we can get following equation:
R[= ρ[ , Ro = ρl (66)
Then
大阪府立大学経済研究 54 ・ 1(223) [2008.6)
Simplified Machine Diagnosis Techniques Using ARMA Model ofAbsolute Deterioration Factor with Weight 67
ρ。 主主 l (67)
Therefore we get the estimates b1, b2 of the MA parameters b1, b2 using (64) and (65) (Taュ
ble 4).
Table 4. Estimation of MA paramet目s with varying S
10 'OアO
0.0485
0.0363
6.3. System Parameter Distance
6.3.1. AR parame白m
When S = 1 ,出e system is considered to be under normal conditions. Therefore we
get the following equation by (34).
(1+2ρî)C1 ρD α=
4(2+ρî)C1 -ρî) (68)
l-2ρ~+3ρi-2ρi ao
4(2+ρî)C1 -ρî)
When S = 1, from (45) , we get
P1= ρ1, βρ2 = 1
(69)
When
lim j(x) = 0, lim g(x) = 0 Z • x • O
(70)
utilizing the following relation:
g(x) g'(x) 一… 一一…z• ô j(x) ;• ô j'(x)
we obtain the following results:
日 1-ρa , 11m 一一一一-0 = u Aρl→ 1 ,ρ2 → 12+4ρ;
(71)
(72)
1+2ρil α。 11口1 一一一一一一一uρ1→ 1 ,ρ2 → 1 -2-4ρ; 乙
(73)
6.3.2. MA parameters
When S = 1, the system is considered to be under normal conditions. Therefore we
can calculate the following equation by substituting a 1, a2 of (72) , (73) into (62) , (63).
大阪府立大学経済研究 54 ・ 1(223) (2008.6)
68
bj =
Simplified Machine Diagnosis Techniques Using ARMAModel of Absolute Deterioration Factor with We�ht
3Rj-2R3
4- 2Ro+5R2ー 2R4
1 _ .5_ 1 b? = --::-Rn+~ R?--::-R,
話 2 り 4 乙 2 ~..
(74)
(75)
From (66), (67) and (45) we get
Ro= ρ。 = 1, Rj = ρj = 1, R2 = ρ2 = 1,
R3 = ρ3 = 1, R4 = ρ4 = 1
From (74) and (75) , we obtain the following results.
1 _ 1 b, = -::-. b? = -=:-
5' -" 4
6.3.3. Calculation of the evaluation function
Calculating Zz using (4) , we obtain Table 5 and Figure 3 to 6.
As S grows larger, the evaluation function Zz becomes smaller. As l grows larger Zz beュ
comes smaller. Therefore we can utilize this index as a machine diagnosis index. Z2
has relatively suitable value for machine diagnosis which is easy to catch a symptom
of machine failure. These results are useful for machine diagnosis in early stage.
Table 5. Zl with varying S
S 1 2 4 6 8 10
1=0 Zo 1.0000 0.8883 0.8791 0.8598 0.8389 0,8204
1 = 1 Zj 1.0000 0.6652 0.6452 0.6053 0.5655 0.5332
1 =2 Z2 1.0000 0.3319 0.3125 0.2772 0.2455 0~2221
1=3 Z3 1.0000 0.1105 0.1020 0.0875 0.07522 0.0666
0.8
0.6
0.4 直函
0.2
。
8=1 8=2 8=4 8=6 8=8 8=10
Figure 3. The Transitionof Evaluation Function Zo
大阪府立大学経済研究 54 ・ 1(223) (2008.6J
Simplified Machine Diagnosis Techniques Using ARMA Model of Absolute Deterioration Factor with Weight
0.8
0.6
0.4
0.2
0.8
0.6
0.4
0.2
。
。
匝ヨ
8=1 8=2 8=4 8=6 8=8 8=10
Figure 4. The Transition of Evaluation Function z�.
亙回
8=1 8=2 8=4 8=6 8=8 8=10
Figure 5. The τ'ransition of Evaluation Function Z2
11!1.l1=31
8=1 8=2 8=4 8=6 8=8 8=10
Figure 6. The 'I旨ansition of Evaluation Function Z3
大阪府立大学経済研究 54 ・ 1(223) (2008.6J
69
70 Simplified Machine Diagnosis Techniques Using ARMA Model
of Absolute Deterioration Factor with Weight
Now , we compare this new index with Bicoherence. We made experiment in the past
[15, 16]. Summary of the experiment is as follows. Pitching defects are pressed on the
gears of small testing machine.
Small defect condition: Pitching defects pressed on 1/3 gears of the total gear.
Middle defect condition: Pitching defects pressed on 2/3 gears of the total gear.
Big defect condition: Pitching defects pressed on whole gears of the total gear.
We examined several cases for the .t;, Jz, in Eq.(l1). We got best-fit result in the follow-
mg case.
l川…y……五: 2五
We obtained the following Bicoherence values in this case (Table 6).
Condition Bicoherence
0.8
0.6
Table 6. Transition of Bicoherence value
Normal
0.99
Small defect
0.38
Middle defect
0.09
Big defect
0.02
0.4 厄白己ohe…ー10.2
。
Normal Small Middle Big defect defect defect
Figure 7 . The Transition of Bicoherence
Thus , Bicoherence proved to be a very sensitive good index. Bicoherence is an absoュ
lute index of which range is 1 to O. Therefore it can be said that it is a universal inュ
dex. In those experiment, smal1 defect condition is generally assumed to be S = 2 and
big defect condition is generally assumed to be S = 6 [10].
In the case of l = 2 (Figure. 5) in Table 5, the value is 0.3319 at small defect
大阪府立大学経済研究 54 ・ 1(223) (2008.6J
Simplified Machine Diagnosis Techniques Using ARMA Model of Absolute Deterioration Factor with Weight 71
condition and 0.3125 at middle defect condition and 0.2772 at big defect condition
which show nearly sensitive behavior in the early stage of Bicoherence. The case of
l = 3 would be too much sensitive. It could besaid that the case l = 2 would be sensiュ
tive enough for the practical use. Therefore, the case l = 2 would be recommended in
this new method. This calculation method is simple enough to be executed even on a
pocketsize calculator. Compared with Bicoherence which has to be calculated by Eq.
(11) , (12) , (76) and (77) , proposed method is by far a simple one and easy to handle on
the field defection.
7. CONCLUSION
In order to make machine diagnosis , the method of calculating Kurtosis or
Bicoherence was utilized in the past. Calculating system parameter distance was a1so
uti1ized to app1y time series data to Autoregressive (AR) model or Autoregressive Movュ
ing A verage (ARMA) model.
In this paper, a simplified calculation method of autocorrelation function was introュ
duced and it was utilized for ARMA mode1 identification. Furthermore , an absolute deュ
terioration factor such as Bicoherence is introduced. Compared with the results obュ
tained so far , the results of numerical examples of this paper are reasonable and
nearly sensitive of another method such as Bicoherence. Judging from these results ,
our method is properly considered to be effective for the early stage failure detection esュ
pecially. This calcu1ation method is very simp1e and is very practica1 at the factory of
maintenance site. This can be installed in microcomputer chips and uti1ized as the too1
for the early stage detection of the failure.
Machine diagnosis could be executed by this simplified calculation method of system
parameter distance with weight. Proposed method proved to be a practical index for maュ
chine diagnosis by numerical examples. The effectiveness of this method should be examュ
ined in various cases.
APPENDIX
XT(f) is a Fourier Transform stated as follows.
大阪府立大学経済研究 54 ・ 1(223) (2008.6J
72 Simplified Machine Diagnosis Techniques Using ARMA Model
of Absolute Deterioration Factor with Weight
r x(t) (0 < t < T) Xr(t) = ~
品 L 0 (else)
T: Basic Frequency Interval
XTC!) = 工二XT(t)e-j2ψdt
SxxC!) is a power spectrum stated as follows.
Sxxc!) = ~ XTC!)X;ω *: conjugate
REFERENCES
(76)
(77)
[ 1 J Bolleter, U.,“'Blade Passage Tones of Centrifugal Pumpsて Vibrαtion 4 (3) , 8-13, 1988
[2 J Box. Jenkins, "Time Series Anαlysis", Third Edition, PRENTICE HALL, 1994
[3 J Hino, M,“Spectrum Anαlysii;" (in JIαpαnese) Asakura Publishing, 1977
[4 J Hoffner, J , "Preventive mαintenαnce for No-Twist rod mills using vibration signαture
anαlysis", Iron αnd Steel Engineer 68 (1), 55-61 , 1991
[5 J Inoue鳥, Y. , and T. Ma抗tsui ,
Non-minimum Phαωse", Proc of the Workshop on Higher-Order Spectral Analysis, 1989
[ 6 J Inoue, Y. , "Signαl processing-1 by Using Higher Order Cumulants", the Institute of Sysュ
tems, Control and Information Engineers, Vol. 36, No. 2, 1992
[7 J Inoue, Y.,“Signal processing四 n by Using Higher Order Cumulants", the Institute of Sysュ
tems, Control and Information Engineers, Vol. 36, No. 5, 1992
[8 J Katayama, T. , "Introduction to System Identificαtion" (in JIαpαnese), Asakura Publishing,
1994
[9 J Krishanamurthy, V. , and H. Vincent Poor,“'Asymptotic Anαlysis of an Algorithm for Paュ
rαmeter Estimαtion αnd Identification of l-b Quαntized AR Time Series", IEEE Trans. Sigュ
nalProcess. Vol. 44, No. 1, 1996
口OJ Maekawa, K., S. Nakajima, and T. Toyoda, "New Severity Index for Failures of MαchineEle司
ments by Impαct Vibration" (in Japanese). J. SOPE Japαn 9 (3) , 163-168, 1997
[l1J Nakamizo T.,“Signal Anαlysis αnd System Identification" (in J,α:panese), CORNA PUBュ
LISHING, 1988
口2J Sagara, S. , K. Akizuki, T. Nakamizo , T. Katayama, "System Identification", The Society
of Instrument and Control Engineers, 1987
口3J Shao , Y. , K. Nezu, T. Matsuura, Y. Hasegawa, and N. Kansawa,“Bearing F,αult Diαgno
sis Using αnAdα:ptive Filter" (in Japanese) , J. SOPE Japan 12 (3) , 71-77, 2001
大阪府立大学経済研究 54 ・ 1(223) (2008.6)
Simplified Machine Diagnosis Techniques Using ARMA Model of Absolute Deterioration Factor with Weight 73
口4J Swami, A. , and J. M. Mendel,“'ARMA Pαrαmeter Estimαtion Using Only Output
Cumulants", IEEE Trans. Acoust, Speech Signal Processing, Vol. ASSP-38, No. 7, 1990
口5J Takeyasu, K., "Wαtching Method of Circulating Moving Object" (in Japanese) , Certified
Patent by Japanese Patent Agency (1987)
口6J Takeyasu, K.,“Wαtching Method of Circulating Moving Object" (in Japanese), Certified
Patent by Japanese Patent Agency (1989)
口7J Takeyasu, K., T. Amemiya and H. Goto,“'Robust System Identification Algorithm of Bootュ
strα:p Type", IEMS Vol. 1, No. 1, 2002
日8J Takeyasu, K., Y. Ishii,匂0'stem p,αrameter Distance for Mαchine Diagnosis", IFORS, 2005
口9J Tokumaru,H. , K. Takeyasu,“The expectation of the Spectral Density by the Discrete Time
Lineαr Model", The Society of Instrument and Control Engineers, Vol. 13, No. 2, pp. 148-
153, 1977
[20J Tokumaru, H., T. Soeda, T. Nakamizo , and K. Akizuki,“'Meαsurement αnd Cαlculαtion"
(in J;α:panese), Baifukan Publishing, 1982
[21J Toresten Sりderstörm , "An efficient αnd Versalile Algorithm for Computing the Covαriance
Function of αn ARMA Process", IEEE Trans on Signal Processing, Vol. 46 , No. 6, 1994
[22J Xian-Da Zhang, H. Takeda "An Approαch to Time Series Analysis and ARMA Spectral Esュ
timation", IEEE Trans on Acoustics, Speech & Signal Processing, Vol. ASSP-35 No. 9,
1987
[23J Yamazaki, H,“Failure Detection αnd Prediction" (In Japanese), Kogyo Chosakai Publish>
ing, 1977
大阪府立大学経済研究 54 ・ 1(223) (2008.6J