# US 2021/0278478 A1 · Deep parallel fault diagnosis method and system for dissolved gas in transformer oil

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2021/0278478 A1
- Document ID: US20210278478A1
- Status: Published application
- Kind: A1
- Application number: 17/160,415
- Filed: 2021-01-28
- Published: 2021-09-09
- Earliest claimed: 2020-03-02
- Assignees: Wuhan University
- Applicants: Wuhan University
- Inventors: Yigang HE; Xiaoxin WU; Jiajun DUAN; Yuanxin XIONG; Hui ZHANG
- Web page: https://patentgrep.com/patent/US20210278478A1

## CPC

- G01R 31/62: Testing of transformers
- G01N 33/0004: {Gaseous mixtures, e.g. polluted air}
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- G06N 3/096: Transfer learning
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- Y04S 10/50: Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
- G01R: MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES (indicating correct tuning of resonant circuits H03J3/12)
- G01N: INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES (measuring or testing apparatus or processes other than immunoassay, involving enzymes or microorganisms C12M, C12Q)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- Y04S: SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS

## Abstract

The disclosure provides a deep parallel fault diagnosis method and system for dissolved gas in transformer oil, which relate to the field of power transformer fault diagnosis. The deep parallel fault diagnosis method includes: collecting monitoring information of dissolved gas in each transformer substation and performing a normalizing processing on the data; using the dissolved gas in the oil to build feature parameters as the input of the LSTM diagnosis model, and performing image processing on the data as the input of the CNN diagnosis model; building the LSTM diagnosis model and the CNN diagnosis model, respectively, and using the data set to train and verify the diagnosis models according to the proportion; and using the DS evidence theory calculation to perform a deep parallel fusion of the outputs of the softmax layers of the two deep learning models.

## Claims

### Claim 1

A deep parallel fault diagnosis method for dissolved gas in transformer oil, comprising:
step 1: obtaining a plurality of groups of monitoring data of dissolved gas in each transformer oil, analyzing a dissolved gas content in each of the groups of monitoring data, obtaining a corresponding fault type label, performing a normalizing processing on each of the groups of monitoring data, and forming a target data set by combining the normalized groups of monitoring data with the corresponding fault type label;
step 2: dividing the target data set into a first training set and a first verification set, training a long short-term memory (LSTM) diagnosis model with the first training set, and verifying the trained LSTM diagnosis model with the first verification set;
step 3: performing image processing on each group of data in the target data set to obtain an image data set, dividing the image data set into a second training set and a second verification set, training a convolutional neural network (CNN) diagnosis model with the second training set, and verifying the trained CNN diagnosis model with the second verification set; and
step 4: performing a deep parallel fusion on outputs of softmax layers of the trained LSTM diagnosis model and the trained CNN diagnosis model, respectively, and outputting a final diagnosis result according to a maximum confidence principle.

### Claim 2 (depends on 1)

The method according to claim 1, wherein the groups of monitoring data of the dissolved gas in each transformer oil are data, ={ai,1, ai,2, . . . , ai,j, . . . , ai,N, si}i∈[1,K], where K is K groups of monitoring data of the dissolved gas, ai,j is a content of the j(j∈[1,N])-th gas parameter in the i-th group of monitoring data of the dissolved gas, si is a transformer state corresponding to the i-th group of monitoring data of the dissolved gas, and N is the number of gas types, and
data obtained after normalization is datai′={bi,1, bi,2, . . . , bi,j, . . . , bi,N, si}i∈[1,K], where K is K groups of monitoring data of the dissolved gas, bi,j is a normalized value of the content of the j(j∈[1,N])-th gas parameter in the i-th group of monitoring data of the dissolved gas, si is the transformer state corresponding to the i-th group of monitoring data of the dissolved gas, and N is the number of gas types.

### Claim 3 (depends on 2)

The method according to claim 2, wherein the LSTM diagnosis model comprises:
an input layer;
an LSTM layer;
a fully connected layer;
the softmax layer; and
a classification output layer,
wherein the LSTM layer comprises a plurality of hidden units, and a state activation function of the hidden units is tanh, and a gate activation function of the hidden units is sigmoid, and
the softmax layer takes an output of the fully connected layer in the LSTM diagnosis model as an input vector and obtains a diagnosis support degree of the LSTM diagnosis model for a fault label by
Softmax
⁡
(
x
1
)
=
1
∑
i
=
1
N
⁢
e
θ
i
T
⁢
x
1
⁡
[
e
θ
1
T
⁢
x
1
e
θ
2
T
⁢
x
1
⋮
e
θ
N
T
⁢
x
1
]
,
where x1 represents the output of the fully connected layer in the LSTM diagnosis model, θi, i=1, 2, . . . , N is a weight matrix of the softmax layer in the LSTM diagnosis model, and Softmax is an activation function.

### Claim 4 (depends on 3)

The method according to claim 3, wherein performing image processing on each group of data in the target data set to obtain the image data set in the step 3 comprises:
performing image processing on each group of data in the target data set, presenting differences of the data in image with color to obtain an image data set A and presenting differences of the data in image with height to obtain an image data set B, respectively,
wherein the image data set A is expressed as datai″={ci, si}i∈[1,K], and the image data set B is expressed as datai″′={di,si}i∈[1,K], where K is K groups of monitoring data of the dissolved gas, ci is an image of a parameter conversion of the i-th group of monitoring data of the dissolved gas in the image data set A, di is an image of a parameter conversion of the i-th group of monitoring data of the dissolved gas in the image data set B, and si is the transformer state corresponding to the i-th group of monitoring data of the dissolved gas.

### Claim 5 (depends on 4)

The method according to claim 4, wherein the softmax layer in the CNN diagnosis model takes an output of a fully connected layer in the CNN diagnosis model as an input vector, and obtains a diagnosis support degree for a fault label by
Softmax
⁡
(
x
2
)
=
1
∑
i
=
1
N
⁢
e
θ
i
′
⁢
⁢
T
⁢
x
2
⁡
[
e
θ
1
′
⁢
⁢
T
⁢
x
2
e
θ
2
′
⁢
⁢
T
x
2
⁢
⋮
e
θ
N
′
⁢
⁢
T
⁢
x
2
]
,
where x2 represents the output of the fully connected layer in the CNN diagnosis model, θ′i, i=1, 2, . . . , N is a weight matrix of the softmax layer in the CNN diagnosis model, and Softmax is an activation function.

### Claim 6 (depends on 5)

The method according to claim 5, wherein the step 4 comprises:
step 4.1: obtaining diagnosis support degrees of the softmax layers of the LSTM diagnosis model and the CNN diagnosis model for fault labels corresponding to the same group of monitoring data to be diagnosed: ξk,γ=ξk,1, ξk,2, . . . , ξk,l, where the value of k represents different diagnosis models, I is the total number of fault labels, and γ=1, 2, . . . , l;
step 4.2: taking different diagnosis models k as rows and the diagnosis support degrees ξk,γ as columns to form a support degree matrix {(Hγ, ξk,γ), k=1, 2, γ=1, . . . , l}, where each element of the support degree matrix represents that the support degree of the k-th diagnosis model for a fault label Hγ is ξk,γ;
step 4.3: treating each column in the support degree matrix as an identification framework Θ of a DS evidence theory, thus Θ={Hγ|γ=1, 2, . . . , l}={H1, H2, . . . , Hl};
step 4.4: putting information of diagnosis support degrees of different diagnosis models into the same identification framework Θ to obtain a basic probability assignment; and
step 4.5: obtaining a compound probability assignment from the basic probability assignment to obtain confidence of different fault labels after fusion from the compound probability assignment, and using a fault label corresponding to a maximum confidence as a diagnosis result of the monitoring data to be diagnosed.

### Claim 7 (depends on 6)

The method according to claim 6, wherein the basic probability assignment is obtained by
m
k
,
γ
=
ω
k
⁢
ξ
k
,
γ
m
k
,
H
=
m
k
⁡
(
H
)
=
1
-
∑
γ
=
1
l
⁢
m
k
,
γ
=
1
-
ω
k
⁢
∑
γ
=
1
l
⁢
ξ
k
,
γ
m
•
k
,
H
=
m
•
⁡
(
H
)
=
ω
k
⁡
(
1
-
∑
γ
=
1
l
⁢
ξ
k
,
γ
)
m
_
k
,
H
=
m
_
k
⁡
(
H
)
=
(
1
-
ω
k
)
,
where ωk is a weight value the k-th diagnosis model, mk,γ represents the basic probability assignment of the k-th diagnosis model for the fault label Hγ of a transformer, mk represents a function of calculating the basic probability assignment, and mk,H represents a residual probability assigned to an entire fault label set H.

### Claim 8 (depends on 7)

The according to claim 7, wherein the compound probability assignment is obtained by
{
H
γ
}
⁢
:
⁢
m
γ
=
K
⁡
[
∏
k
=
1
2
⁢
(
m
k
,
γ
+
m
_
k
,
H
+
m
•
k
,
H
)
-
∏
k
=
1
2
⁢
(
m
_
k
,
H
+
m
•
k
,
H
)
]
,
where mγ represents the compound probability assignment of the fault label Hγ, K is an intermediate variable, and
a normalized probability assignment ξγ of the fault label Hγ is calculated by
ξ
γ
=
m
γ
1
-
m
_
H
,
γ
=
1
,
…
⁢
⁢
l
m
_
H
=
K
⁡
[
∏
k
=
1
2
⁢
(
m
_
k
,
H
)
]
,
and a label corresponding to a maximum probability is output as the final diagnosis result according to the maximum confidence principle.

### Claim 9

A deep parallel fault diagnosis system for dissolved gas in transformer oil, comprising:
a data processing module, configured to obtain a plurality of groups of monitoring data of dissolved gas in each transformer oil, analyze a dissolved gas content in each of the groups of monitoring data, obtain a corresponding fault type label, perform a normalizing processing on each of the groups of monitoring data, and form a target data set by combining the normalized groups of monitoring data with the corresponding fault type label;
a first training module, configured to divide the target data set into a first training set and a first verification set, train a long short-term memory (LSTM) diagnosis model with the first training set, and verify the trained LSTM diagnosis model with the first verification set;
a second training module, configured to perform image processing on each group of data in the target data set to obtain an image data set, divide the image data set into a second training set and a second verification set, train a convolutional neural network (CNN) diagnosis model with the second training set, and verify the trained CNN diagnosis model with the second verification set; and
a diagnosis module, configured to perform a deep parallel fusion on outputs of softmax layers of the trained LSTM diagnosis model and the trained CNN diagnosis model, respectively, and output a final diagnosis result according to a maximum confidence principle.

### Claim 10 (depends on 1)

A computer-readable storage medium with program instructions stored thereon, wherein the deep parallel fault diagnosis method of dissolved gas in transformer oil according to claim 1 is implemented by executing the program instructions by a processor.

### Claim 11 (depends on 2)

A computer-readable storage medium with program instructions stored thereon, wherein the deep parallel fault diagnosis method of dissolved gas in transformer oil according to claim 2 is implemented by executing the program instructions by a processor.

### Claim 12 (depends on 3)

A computer-readable storage medium with program instructions stored thereon, wherein the deep parallel fault diagnosis method of dissolved gas in transformer oil according to claim 3 is implemented by executing the program instructions by a processor.

### Claim 13 (depends on 4)

A computer-readable storage medium with program instructions stored thereon, wherein the deep parallel fault diagnosis method of dissolved gas in transformer oil according to claim 4 is implemented by executing the program instructions by a processor.

### Claim 14 (depends on 5)

A computer-readable storage medium with program instructions stored thereon, wherein the deep parallel fault diagnosis method of dissolved gas in transformer oil according to claim 5 is implemented by executing the program instructions by a processor.

### Claim 15 (depends on 6)

A computer-readable storage medium with program instructions stored thereon, wherein the deep parallel fault diagnosis method of dissolved gas in transformer oil according to claim 6 is implemented by executing the program instructions by a processor.

### Claim 16 (depends on 7)

A computer-readable storage medium with program instructions stored thereon, wherein the deep parallel fault diagnosis method of dissolved gas in transformer oil according to claim 7 is implemented by executing the program instructions by a processor.

### Claim 17 (depends on 8)

A computer-readable storage medium with program instructions stored thereon, wherein the deep parallel fault diagnosis method of dissolved gas in transformer oil according to claim 8 is implemented by executing the program instructions by a processor.
