US 2021/0012232 A1 · FAULT TRANSFER DIAGNOSIS METHOD FOR ROLLING ELEMENT BEARINGS BASED ON POLYNOMIAL KERNEL INDUCED FEATURE DISTRIBUTION ADAPTATION

Abstract

A fault transfer diagnosis method for rolling element bearings based on polynomial kernel induced feature distribution adaptation includes: inputting the data set of the source rolling element bearings and the monitoring data set from the target rolling element bearings into the deep residual network; extracting the transferrable fault features of the source and the transferrable fault features of the target layer by layer; minimizing the distribution discrepancy by the polynomial kernel induced feature adaptation; inputting the transferrable fault features of the target into the Softmax classifier to obtain the probability distribution of the specific state of the target samples; converting the probability distribution into the pseudo labels of the target samples; training the transfer diagnosis model; inputting the monitoring data of the target bearings into the trained diagnostic model, and outputting the label probability distribution corresponding to the data samples.

A fault transfer diagnosis method for rolling element bearings based on polynomial kernel induced feature distribution adaptation, comprising the following steps:

step 1:
obtaining R kinds of sample data set {(xis,yis)}i=1ns with health state labels from source rolling element bearings, wherein, xis∈RN×1 an ith source sample and the ith source sample is composed of N vibration signal data points, health state labels of the rolling element bearings correspond to the ith source sample, and the health state labels of the rolling element bearings are yi∈{1, 2, 3, . . . , R}, and ns is a number of labeled samples;
and obtaining monitoring data set {xit}i=1nt from target rolling element bearings, wherein nt is a number of unlabeled samples;
step 2:
inputting the sample data set of the source rolling element bearings and the monitoring data set of the target rolling element bearings into a domain-sharing deep residual network;
extracting transferrable fault features in the sample data set of the source rolling element bearings and transferrable fault features in the monitoring data set of the target rolling element bearings layer by layer to obtain the transferrable fault features xs,F2=ƒθ(xis) of the source rolling element bearings and the transferrable fault features xt,F2=ƒθ(xit) of the target rolling element bearings;
step 3:
adopting a maximum mean discrepancy measurement method with polynomial kernel induced to measure a distribution discrepancy (xis,F2,xit,F2) of the transferrable fault features in the sample data set of the source rolling element bearings and the transferrable fault features in the monitoring data set of the target rolling element bearings as follows:
D
ℋ
2
(
x
i
s
,
F
2
,
x
i
t
,
F
2
)
=
∑
j
=
1
c

(
c
j
)

a
j

b
c
-
j


E
(
x
i
s
,
F
2
)
j
-
E
(
x
i
t
,
F
2
)
j

ℋ
2
,

wherein
,

{

E
(
x
i
s
,
F
2
)
j
-
E
(
x
i
t
,
F
2
)
j

ℋ
2
=
∑
q
=
0
j

(
j
q
)

(
v
q
T
·
μ
j
-
q
)
v
q
T
=
[
E

[
Λ
xx
-
E

(
Λ
xx
)
]
q
,
E

[
Λ
yy
-
E

(
Λ
yy
)
]
q
,
E

[
Λ
xy
-
E

(
Λ
xy
)
]
q
]
μ
j
-
q
=
[
E

(
Λ
xx
)
j
-
q
,
E

(
Λ
yy
)
j
-
q
,
-
2

E

(
Λ
xx
)
j
-
q
]
T
Λ
xx
=
[
〈
x
1
s
,
F
2
,
x
1
s
,
F
2
〉
,
…

,
〈
x
1
s
,
F
2
,
x
n
s
,
F
2
〉
,
…

,
〈
x
n
s
,
F
2
,
x
1
s
,
F
2
〉
,
…

,
〈
x
n
s
,
F
2
,
x
n
s
,
F
2
〉
]
T
Λ
yy
=
[
〈
x
1
t
,
F
2
,
x
1
t
,
F
2
〉
,
…

,
〈
x
1
t
,
F
2
,
x
n
t
,
F
2
〉
,
…

,
〈
x
n
t
,
F
2
,
x
1
t
,
F
2
〉
,
…

,
〈
x
n
t
,
F
2
,
x
n
t
,
F
2
〉
]
T
Λ
xy
=
[
〈
x
1
s
,
F
2
,
x
1
t
,
F
2
〉
,
…

,
〈
x
1
s
,
F
2
,
x
n
t
,
F
2
〉
,
…

,
〈
x
n
s
,
F
2
,
x
1
t
,
F
2
〉
,
…

,
〈
x
n
s
,
F
2
,
x
n
t
,
F
2
〉
]
T
,
represents a reproducing kernel Hilbert space, and a, b, c represent a slope, an intercept and an order of a polynomial kernel function, respectively;
step 4:
inputting the transferrable fault features of the source rolling element bearings and the transferrable fault features of the target rolling element bearings into an output layer F3 of the domain-sharing deep residual network, and adopting an activation function Softmax to generate a probability distribution of health states of input samples of the sample data set as follows:
=[P(|;θF3)]q=1k,
wherein, a probability calculation formula of a qth sample is as follows:
P
(
y
i

=
q
|
x
i

,
F
2
;
θ
F
3
)
=
exp
(
w
q
F
3
·
x
i

,
F
2
+
b
q
F
3
)
∑
q
=
1
ϕ

exp
(
w
q
F
3
·
x
i

,
F
2
+
b
q
F
3
)
,
wherein, θF3={wF3,bF3} is a parameter to be trained of the output layer F3 and is a bearing data identifier;
converting the probability distribution into pseudo labels ŷit of the target rolling element bearings:
ŷit=[ŷ1t ŷ2t . . . ŷqt . . . ŷkt],
wherein,
y
^
q
t
=
{
1
q
=
arg


max
q


Γ
i
t
0
others
;
step 5:
combining the distribution discrepancy (xis,F2, xit,F2) obtained in step 3 with the pseudo labels r of the target rolling element bearings obtained in step 4 to train a transfer diagnosis model, wherein an objective function is minimized as follows:
J

(
θ
)
=
-
1
n
s

∑
i
=
1
n
s

(
y
i
s
)
T
·
ln

(
Γ
i
s
)
+
β
·
D
ℋ
2
(
x
i
s
,
F
2
,
x
i
t
,
F
2
)
-
α
·
1
n
t

∑
j
=
1
n
t

(
y
^
j
t
)
T
·
ln

(
Γ
j
t
)
,
wherein, α is a tradeoff parameter of a transferrable fault feature distribution adaptation item, β is a tradeoff parameter of a pseudo label training item, and θ is a parameter to be trained;
and
step 6:
inputting the monitoring data set of the target rolling element bearings into a trained transfer diagnosis model, outputting a label probability distribution corresponding to the transferrable fault features in the sample data set of the source rolling element bearings, and taking sample labels corresponding to a maximum probability as the health states {yit}i=1nt of the rolling element bearings.

631 more elements

Fault transfer diagnosis method for rolling element bearings based on polynomial kernel induced feature distribution adaptation

Life of this patent

Fault transfer diagnosis method for rolling element bearings based on polynomial kernel induced feature distribution adaptation

Life of this patent

Abstract

Claims1