Method and system for fault diagnosis with small samples of power equipment based on virtual and real twin spaces

Patent metadata

Field	Value
Publication number	US-11913854-B2
Application number	US-202017124499-A
Country	US
Kind code	B2
Filing date	Dec 17, 2020
Priority date	Apr 9, 2020
Publication date	Feb 27, 2024
Grant date	Feb 27, 2024

Abstract

A method and a system for fault diagnosis with small samples of power equipment based on virtual and real twin spaces are disclosed, which belong to the field of fault diagnosis of power equipment. The method includes: test samples containing different locations, types and severity levels of fault of power equipment are acquired to form a real physical space; a virtual mirror space is acquired by simulation according to a simulation model of the equipment to be diagnosed; the training set in the real physical space is spatially integrated with the sample set in the virtual mirror space to obtain a training sample set in the twin spaces; the training sample set in the twin spaces serves as the supplement to the training set in the real physical space, and the fault type and fault location serve as diagnostic labels to be input to the deep neural network for training.

First claim

What is claimed is:

1. A method for fault diagnosis with small samples of a power equipment based on virtual and real twin spaces, comprising:
(1) acquiring test samples containing different locations, types and severity levels of fault of a power equipment to form a real physical space;
(2) establishing a simulation model of an equipment to be diagnosed, setting a corresponding fault type, and superimposing random noise on parameters of the simulation model to acquire a required number of samples by simulation to form a virtual mirror space;
(3) dividing samples in the real physical space into a training set and a verification set;
(4) integrating the training set in the real physical space spatially with the sample set in the virtual mirror space to obtain a training sample set in the virtual and real twin spaces;
(5) using the training sample set in the virtual and real twin spaces as a supplement to the training set in the real physical space, and using the fault type and fault location as diagnosis labels to be input to a deep neural network for training, and using the trained deep neural network to perform identification and positioning of fault on the verification set in the real physical space to verify a diagnosis result.

2. The method according to claim 1 , wherein step (1) comprises:
areas are divided according to diagnosis requirements and fault characteristics of the equipment to be diagnosed, and the divided areas are used as different fault locations of the power equipment, the types and severity levels of fault are determined according to the regularity of statistical data of fault of the equipment to be diagnosed, test samples containing different locations, types and severity levels of fault of power equipment are obtained to form the real physical space.

3. The method according to claim 2 , wherein step (2) comprises:
in the simulation model of the equipment to be diagnosed, the simulated fault type and fault location are configured according to the actual fault type and fault location divided in step (1), and the severity level in simulation is randomly set according to a distribution of an actual fault level;
a loop simulation is performed to involve different fault areas, fault types and fault levels, so as to obtain simulated fault samples of the equipment, and the virtual mirror space is formed through the fault samples obtained from all simulations.

4. The method according to claim 1 , wherein step (4) comprises:
a power equipment status label corresponding to the data is set as γ, and the sample with label γ in the real physical space is marked as SR γ ={SR γ1 ; SR γ2 ; SR γ3 ; . . . }, which is called a subset of the real physical space, the subset of the real physical space contains a total of NR γ samples, and NR γ represents a data amount of samples with the label γ in the real physical space;
the samples with the label γ in the virtual mirror space are denoted as SVγ={SVγ1; SVγ2; SVγ3; . . . }, which is called a subset of the virtual mirror space, and the subset of the virtual mirror space contains NVγ samples in total, and NVγ represents the number of samples with the label γ in the virtual mirror space;
fault feature extraction is performed on all samples in the subset in the real physical space and the subset in the virtual mirror space, and the original sample data is replaced with the sample data after feature extraction;
a sample data of the first equipment to be diagnosed with the label γ is taken from the subset in the real physical space, and a sample data of the second equipment to be diagnosed with the label γ is taken from the subset in the virtual mirror space, the sample data of the first equipment to be diagnosed and the sample data of the second equipment to be diagnosed are integrated, so as to complete integration of various samples in the subset in the real physical space and various samples in the subset in the virtual mirror space to obtain a training sample set in the virtual and real twin spaces.

5. The method according to claim 2 , wherein step (4) comprises:
a power equipment status label corresponding to the data is set as γ, and the sample with label γ in the real physical space is marked as SR γ ={SR γ1 ; SR γ2 ; SR γ3 ; . . . }, which is called a subset of the real physical space, the subset of the real physical space contains a total of NR γ samples, and NR γ represents a data amount of samples with the label γ in the real physical space;
the samples with the label γ in the virtual mirror space are denoted as SVγ={SVγ1; SVγ2; SVγ3; . . . }, which is called a subset of the virtual mirror space, and the subset of the virtual mirror space contains NVγ samples in total, and NVγ represents the number of samples with the label γ in the virtual mirror space;
fault feature extraction is performed on all samples in the subset in the real physical space and the subset in the virtual mirror space, and the original sample data is replaced with the sample data after feature extraction;
a sample data of the first equipment to be diagnosed with the label γ is taken from the subset in the real physical space, and a sample data of the second equipment to be diagnosed with the label γ is taken from the subset in the virtual mirror space, the sample data of the first equipment to be diagnosed and the sample data of the second equipment to be diagnosed are integrated, so as to complete integration of various samples in the subset in the real physical space and various samples in the subset in the virtual mirror space to obtain a training sample set in the virtual and real twin spaces.

6. The method according to claim 3 , wherein step (4) comprises:
a power equipment status label corresponding to the data is set as γ, and the sample with label γ in the real physical space is marked as SR γ ={SR γ1 ; SR γ2 ; SR γ3 ; . . . }, which is called a subset of the real physical space, the subset of the real physical space contains a total of NR γ samples, and NR γ represents a data amount of samples with the label γ in the real physical space;
the samples with the label γ in the virtual mirror space are denoted as SVγ={SVγ1; SVγ2; SVγ3; . . . }, which is called a subset of the virtual mirror space, and the subset of the virtual mirror space contains NVγ samples in total, and NVγ represents the number of samples with the label γ in the virtual mirror space;
fault feature extraction is performed on all samples in the subset in the real physical space and the subset in the virtual mirror space, and the original sample data is replaced with the sample data after feature extraction;
a sample data of the first equipment to be diagnosed with the label γ is taken from the subset in the real physical space, and a sample data of the second equipment to be diagnosed with the label γ is taken from the subset in the virtual mirror space, the sample data of the first equipment to be diagnosed and the sample data of the second equipment to be diagnosed are integrated, so as to complete integration of various samples in the subset in the real physical space and various samples in the subset in the virtual mirror space to obtain a training sample set in the virtual and real twin spaces.

7. The method according to claim 4 , wherein the sample integration is performed through φ(R, V)=R⊗V, wherein φ(R, V) represents an integration function, R represents the sample in a subset in the real physical space, and V represents the sample in the subset in the virtual mirror space.

8. The method according to claim 1 , wherein the method of using the deep neural network for fault diagnosis in step (5) is:
a total number of last outputs with a parameter layer in the deep neural network is replaced with n×m+1, so as to perform network trai

Assignees

• Univ Wuhan

Inventors

• He Yigang
• Duan Jiajun
• Wu Xiaoxin
• He Liulu
• Zhang Hui
• Shi Guolong

Classifications

• G06N3/096

  Transfer learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G01M15/02 Primary

  Details or accessories of testing apparatus · CPC title

• G01M15/05

  by combined monitoring of two or more different engine parameters · CPC title