US 11,630,034 B2 · Method for diagnosing and predicting operation conditions of large-scale equipment based on feature fusion and conversion

Abstract

A method for diagnosing and predicting operation conditions of large-scale equipment based on feature fusion and conversion, including: collecting a vibration signal of each operating condition of the equipment, and establishing an original vibration acceleration data set of the vibration signal; performing noise reduction on the original vibration acceleration data set, and calculating a time domain parameter; performing EMD on a de-noised vibration acceleration and calculating a frequency domain parameter; constructing a training sample data set through the time domain parameter and the frequency domain parameter; establishing a GBDT model, and inputting the training sample data set into the GBDT model; extracting a leaf node number set from a trained GBDT model; performing one-hot encoding on the leaf node number set to obtain a sparse matrix; and inputting the sparse matrix into a factorization machine to obtain a prediction result.

A method for diagnosing and predicting operation conditions of large-scale equipment, comprising:

(S1) collecting a vibration signal of individual operating conditions of the large-scale equipment, and establishing an original vibration acceleration data set of the vibration signal;
(S2) de-noising the original vibration acceleration data set to obtain a de-noised vibration acceleration;
(S3) calculating a time domain parameter based on the de-noised vibration acceleration;
(S4) performing empirical mode decomposition (EMD) on the de-noised vibration acceleration to obtain a plurality of components;
(S5) calculating a frequency domain parameter based on the plurality of components;
(S6) constructing a training sample data set based on the time domain parameter and the frequency domain parameter, dividing the training sample data set into a first training subset and a second training subset;
(S7) establishing a gradient boosting decision tree (GBDT) model, and training the GBDT model with the first training subset;
(S8) extracting a leaf node number set from the trained GBDT model;
(S9) performing one-hot encoding on the leaf node number set to obtain a sparse matrix;
(S10) inputting the sparse matrix into a factorization machine to obtain a prediction result, wherein the factorization machine is trained with the second training subset;
and
(S11) identifying early failure of the large-scale equipment based on the prediction result and performing maintenance on the large-scale equipment based on the identified early failure;
wherein the step (S2) comprises:
(S2-1) removing a direct-current (DC) component of a hth original vibration acceleration gh(i) in the original vibration acceleration data set according to formula (1) to obtain a hth first vibration acceleration g′h(i):
g
h
⁡
(
i
)
′
=
g
h
⁡
(
i
)
+
∑
i
=
1
n
g
h
⁡
(
i
)
n
;
(
1
)
wherein n is a data length of the hth original vibration acceleration;
and
(S2-2) removing a trend of the hth first vibration acceleration according to formulas (2)-(4) to obtain a hth de-noised vibration acceleration:
g
h
⁡
(
i
)
″
=
g
h
⁡
(
i
)
′
-
z
;
(
2
)
z
=
α
′
+
β
′
⁢
t
;
(
3
)
(
α
′
,
β
′
)
=
arg
⁢
min
(
α
,
β
)
[
∑
i
=
1
n
(
α
+
β
⁢
t
-
g
h
⁡
(
i
)
′
)
2
]
;
(
4
)
wherein z is the fitting trend linear function when α=α′ and β=β′;
α and β are parameters of a fitting trend linear function;
(α′,β′) represents values of α and β corresponding to a minimum value of
∑
i
=
1
n
(
α
+
β
⁢
t
-
g
h
⁡
(
i
)
′
)
2
;
arg min[•] indicates minimization operation;
and t represents time within a range of one sample;
a sampling time ti corresponding to the hth de-noised vibration acceleration is expressed as:
ti=e+f×i i≤n (5);
wherein i is an integer starting from 0 for numbering each element in the training sample data set, e is an origin time of vibration acceleration;
and f is a sampling frequency;
in step (S3), the time domain parameter comprises an effective value, kurtosis, variance and crest factor of the de-noised vibration acceleration;
the step (S5) comprises:
(S5-1) obtaining an energy value Pk of a kth component according to formula (6):
P
k
=
∑
ξ
=
1
N
1
h
k
2
(
t
*
)
×
dT
;
(
6
)
wherein N1 is the total number of discrete data points in the kth component hk(t*);
dT represents a sampling frequency;
and i1 represents a serial number of the discrete data points in the kth component hk(t*);
(S5-2) sorting energy values of the plurality of components;
selecting X components with the largest energy value;
and
subjecting the X components to fast Fourier transform to obtain X initial spectra;
(S5-3) extracting a maximum amplitude P0X of each of the X initial spectra;
(S5-4) setting a threshold;
and
selecting frequencies respectively corresponding to an amplitude higher than the threshold to establish X frequency-amplitude data sets;
(S5-5) calculating an average amplitude P1X of the X frequency-amplitude data sets according to formula (7):
P
1
X
=
∑
i
2
=
1
N
2
❘
"\[LeftBracketingBar]"
f
i
2
❘
"\[RightBracketingBar]"
N
2
⁢
N
2
≤
N
1
;
(
7
)
wherein N2 is the number of selected frequencies;
and fi2 is an amplitude corresponding to a i2th frequency;
(S5-6) calculating a root mean square value P2X of the X frequency-amplitude data sets according to formula (8):
P
2
X
=
∑
i
2
=
1
N
2
❘
"\[LeftBracketingBar]"
f
i
2
2
❘
"\[RightBracketingBar]"
4
⁢
π
2
⁢
f
i
2
2
;
(
8
)
wherein π represents a 180-degree angle;
and
(S5-7) selecting a frequency P3X corresponding to a maximum amplitude in the X frequency-amplitude data sets;
in step (S7), the GBDT model is established through steps of:
(S7-1) initializing a first weak learner G0(D) according to formula (9):
G
0
(
D
)
=
log
⁢
P
⁡
(
y
k
=
0
|
D
)
1
-
P
⁡
(
y
k
=
0
|
D
)
⁢
h
∈
{
1
,
…
,
q
}
;
(
9
)
wherein P(yh=0|D) represents a probability that a hth tagged value yh in the training sample data set D is marked as 0 in an integrated feature real number set;
and log is a base-10 logarithmic function;
(S7-2) establishing S trees, and calculating a pseudo-residual rsss,h of a sth tree G′s(xh) by using a log-likelihood function, expressed as:
YSS
s
,
h
=
y
h
-
1
1
+
e
-
G
s
′
(
x
h
)
⁢
s
∈
{
0
,
1
,
…
⁢
S
-
1
}
;
(
10
)
wherein xh is a hth integrated feature real number in the training sample data set D;
and e represents natural logarithm;
(S7-3) fitting data (xh,rsss,h) using a classification and regression tree (Cart) to obtain a sth regression tree Gs(xh) as a sth weak learner;
(S7-4) calculating an optimal negative gradient fitted value cs,ks of each leaf node in the sth regression tree according to formula (11):
c
s
,
ks
=
❘
"\[LeftBracketingBar]"
∑
x
k
∈
R
s
,
ks
rss
s
,
h
∑
x
j
∈
R
s
,
ks
(
y
k
-
rss
s
,
h
)
⁢
(
1
-
y
h
+
rss
s
,
h
)
❘
"\[RightBracketingBar]"
;
(
11
)
wherein Rs,ks is a leaf node region of the sth regression tree;
and ks represents a kth leaf node of the sth regression tree;
(S7-5) updating a strong learner Gs+1(xh) according to formula (12):
G
s
+
1
(
x
h
)
=
G
s
(
x
h
)
+
∑
kz
=
1
K
s
c
z
,
ks
⁢
I
⁡
(
x
k
∈
R
s
,
ks
)
;
(
12
)
wherein I(•) is a weighted count of xh within the leaf node region;
and Ks is the total number of leaf nodes of the sth regression tree;
and
(S7-6) obtaining a final strong learner Gs+1′(xin) as the trained GBDT model based on an output cs+1,ks of an updated strong learner according to formula (13):
G
s
+
1
′
(
x
k
)
=
G
0
(
x
h
)
+
∑
s
=
2
S
∑
k
=
1
K
s
c
s
-
1
,
k
⁢
I
⁡
(
x
∈
R
s
,
ks
)
.
(
13
)

501 more elements

Method for diagnosing and predicting operation conditions of large-scale equipment based on feature fusion and conversion

Life of this patent

Method for diagnosing and predicting operation conditions of large-scale equipment based on feature fusion and conversion

Life of this patent

Abstract

Claims6