CN116955951A - Transformer fault diagnosis method and device based on self-attention heterogeneous network - Google Patents

Transformer fault diagnosis method and device based on self-attention heterogeneous network Download PDF

Info

Publication number
CN116955951A
CN116955951A CN202310857743.XA CN202310857743A CN116955951A CN 116955951 A CN116955951 A CN 116955951A CN 202310857743 A CN202310857743 A CN 202310857743A CN 116955951 A CN116955951 A CN 116955951A
Authority
CN
China
Prior art keywords
self
attention
transformer
network
fault
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310857743.XA
Other languages
Chinese (zh)
Inventor
李奇越
赵小平
孙伟
李帷韬
葛健
章海斌
郭振宇
刘鑫
马欢
陈龙
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Super High Voltage Branch Of State Grid Anhui Electric Power Co ltd
Hefei University of Technology
State Grid Anhui Electric Power Co Ltd
Original Assignee
Super High Voltage Branch Of State Grid Anhui Electric Power Co ltd
Hefei University of Technology
State Grid Anhui Electric Power Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Super High Voltage Branch Of State Grid Anhui Electric Power Co ltd, Hefei University of Technology, State Grid Anhui Electric Power Co Ltd filed Critical Super High Voltage Branch Of State Grid Anhui Electric Power Co ltd
Priority to CN202310857743.XA priority Critical patent/CN116955951A/en
Publication of CN116955951A publication Critical patent/CN116955951A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
    • G01R31/12—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing
    • G01R31/1227—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing of components, parts or materials
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a transformer fault diagnosis method and device based on a self-attention heterogeneous network, wherein the method comprises the following steps: under the fault state of the transformer, acquiring a thermal infrared image of the transformer and preprocessing to obtain a first training set T 1 The method comprises the steps of carrying out a first treatment on the surface of the Collecting the content of dissolved gas in the oil of the transformer and preprocessing to obtain a second training set T 2 The method comprises the steps of carrying out a first treatment on the surface of the Constructing a heterogeneous fusion network model based on self-attention, wherein the model comprises a self-attention LSTM network focusing on global time scale information, a self-attention residual error network focusing on local space information, a feature fusion layer and Softmax classificationA first training set T 1 Input into the self-attention residual error network, a second training set T 2 Input into a self-attention LSTM network; training a heterogeneous fusion network model based on self-attention, and performing fault diagnosis by using the trained model; the invention has the advantages that: the extracted features are complete, and the final diagnosis result is accurate.

Description

Transformer fault diagnosis method and device based on self-attention heterogeneous network
Technical Field
The invention relates to the field of transformer fault diagnosis, in particular to a transformer fault diagnosis method and device of a heterogeneous network based on self-attention.
Background
Transformers are the most costly equipment in power systems, and whether they can operate safely is not only related to the safe production and economic benefits of the power enterprise, but also the social impact is immeasurable. Therefore, it is important to conduct an intensive study on the fault diagnosis of the transformer. The transformer is a power system which is equivalent to the heart and is suitable for human body, and whether the transformer can normally operate determines whether the energy can be effectively transmitted. However, the transformer usually works under a heavy load state, and meanwhile, a series of faults of the transformer can be caused by environmental changes and external force, so that the problems of large-area power failure and the like are caused, and influences are generated to different degrees. Therefore, the problem of safe operation of the transformer becomes an important problem in the power industry of China.
Currently, the dominant method in detecting transformer faults is the dissolved gas analysis method. There are also some disadvantages in detecting the nature of the internal faults of the transformer (overheating or discharging) from only the dissolved gas components in the transformer oil. For example: the transformer faults found by analysis of the dissolved gas method in oil are more than after obvious abnormality of the transformer, but the transformer faults are often serious at the moment; if the content of the dissolved gas is not obvious, the insulation condition of the transformer cannot be accurately judged, and the maintenance of the fault transformer can be delayed, so that more serious faults occur. The infrared imaging technology can accurately acquire the surface temperature distribution condition of the power equipment in operation in a non-contact and long-distance manner, detects latent thermal defect faults existing in a large-scale power transformer, and is widely applied to power equipment fault detection. Therefore, the combination of the infrared imaging technology and the dissolved gas analysis method has important significance for rapid and accurate fault diagnosis of the transformer in practical application.
For example, chinese patent publication No. CN114462508A discloses a method for evaluating the health status of a transformer based on a multi-modal neural network, which collects the thermal infrared image of the transformer and the content of dissolved gas in oil, processes the collected multi-modal information by using wavelet threshold denoising, extracts text features from the gas data in oil by using a one-dimensional convolutional neural network, extracts image features from the infrared image by using a depth residual neural network, and performs fault diagnosis on the transformer by combining an infrared imaging technique with a dissolved gas analysis method. However, the space-time characteristics of the transformer data cannot be extracted, so that the extracted characteristics are not complete enough, and the final diagnosis result is not accurate enough.
Disclosure of Invention
The technical problem to be solved by the invention is that the fault diagnosis method of the transformer in the prior art has inaccurate final diagnosis result due to incomplete extracted characteristics.
The invention solves the technical problems by the following technical means: a method for transformer fault diagnosis of a self-attention based heterogeneous network, the method comprising:
step a: under the fault state of the transformer, acquiring a thermal infrared image of the transformer and preprocessing to obtain a first training set T 1 ；
Step b: under the fault state of the transformer, the content of dissolved gas in the oil of the transformer is collected and preprocessed to obtain a second training set T 2 ；
Step c: constructing a heterogeneous fusion network model based on self-attention, wherein the model comprises a self-attention LSTM network focusing on global time scale information, a self-attention residual error network focusing on local space information, a characteristic fusion layer and a Softmax classifier, and the first training set T 1 Input into the self-attention residual error network, a second training set T 2 The method comprises the steps of inputting the data into a self-attention LSTM network, wherein the output end of the self-attention residual error network and the output end of the self-attention LSTM network are connected with the input end of a feature fusion layer, and the output end of the feature fusion layer is connected with a Softmax classifier;
step d: and training a heterogeneous fusion network model based on self-attention, and performing fault diagnosis by using the trained model.
The beneficial effects are that: the self-attention LSTM network focuses on global time scale information, the self-attention residual error network focuses on local space information, the two information are fused through the feature fusion layer, a complete transformer fault feature vector is constructed, the local space information and the global time scale information are fully developed, space-time information is complete, accurate fault diagnosis results are output, a model is trained, fault diagnosis is carried out by using the trained model, and accuracy of the fault diagnosis results is further improved.
Further, the step a includes:
obtaining L time periods when the transformer operates in h fault states; any first time period of the L time periodsDenoted as T l Will be the first time period T l Dividing into N equally spaced moments; acquiring thermal infrared images of the oil immersed transformer at any nth equidistant moment, preprocessing and normalizing to obtain a transformer fault image sample set containing h multiplied by N multiplied by L as a first training set T 1 。
Still further, the step b includes:
collecting the content of dissolved gas in m oils of the oil immersed transformer at any nth equidistant moment and taking the content as a fault characteristic variable; thus, h transformer fault time sequence samples with the number of rows of N multiplied by L and the number of columns of m are formed and normalized, and the normalized transformer fault time sequence samples are obtained, wherein N is>m; setting the size of a sliding window as w multiplied by m, the step length as 1, and performing longitudinal sliding value taking on the normalized transformer fault time sequence samples to obtain h multiplied by N multiplied by L transformer fault matrixes with w multiplied by m as a second training set T 2 。
Still further, the self-attention residual network comprises three identical self-attention residual modules, which are serially connected in sequence.
Still further, the processing procedure of the self-attention residual error module is as follows:
1) Sequentially inputting two 2D convolutions into a transformer fault image Input1 to carry out residual mapping, and obtaining a feature map f with dimensions of C multiplied by H multiplied by W; transformer fault image Input1 is a first training set T 1 Data in (a);
2) After the feature map f is activated by adopting a 1×1 convolution and Softmax function, an intermediate feature vector with the dimension HW×1×1 can be obtained, and the intermediate feature vector and the feature map f are subjected to dot product to obtain a channel attention feature vector f with the dimension C×1×1 1 ；
3) Global feature vector f 1 After being Input into the two full-connection layers, the transformer is activated by adopting a Sigmoid function, performs dot product with the feature map f, and then performs short-circuit connection with the transformer fault image Input1 to obtain an output Y.
Still further, the self-attentive LSTM network comprises three identical self-attentive LSTM modules, which are serially connected in sequence.
Still further, the self-attention LSTM module comprises the following processing procedures:
1) The transformer fault matrix Input2 is respectively Input to each LSTM unit, and the hidden layer output h= { h at each moment can be obtained 1 ,h 2 ,...,h t And (b) wherein h t Is hs×1, and h is hs×t; the transformer fault matrix Input2 is the second training set T 2 Data in (a);
2) After the hidden layer output h at each moment is activated by adopting a 1D convolution and softmax function, an intermediate feature vector with the dimension of 1 Xhs can be obtained, dot product is carried out on the intermediate feature vector and h, and a time scale attention feature vector with the dimension of 1 Xt is obtained
3) Attention feature vector for time scaleAfter being sequentially input into two full-connection layers, the method is activated by adopting a Sigmoid function, and dot product is carried out on the active signal and H to obtain output H.
Still further, the step d includes:
step d1, defining the current iteration number of the neural network as mu, and initializing mu=1; the maximum iteration number is mu max The method comprises the steps of carrying out a first treatment on the surface of the Carrying out the mu-th random initialization on parameters of each layer in the network;
step d2, initializing i=1;
step d3, from the first training set T 1 Selecting the ith Zhang Bianya device fault image x i Inputting the self-attention residual error network to obtain a feature vector space feature vector F i,μ 1 The dimension is p×1; from the second training set T 2 Selecting the ith transformer fault matrix y i Inputting the self-attention LSTM network to obtain a time sequence feature vector F i,μ 2 The dimension is q multiplied by 1;
step d4, the space feature vector F i,μ 1 And the time sequence characteristic vector F i,μ 2 The input characteristic fusion layer is spliced end to end in sequence to obtain a transformer fault state characteristic vector F i,μ ＝[F i,μ 1 ,F i,μ 2 ] T The dimension is (p+q). Times.1; the fault state characteristic vector F of the transformer i,μ The input Softmax classifier obtains the transformer fault image x of the current network input i And transformer fault matrix y i Fault classification result H i ；
Step d5, after i+1 is assigned to i, judging whether i > h multiplied by N multiplied by L is true or not; if yes, continuing to execute the step d6, otherwise, returning to the step d3;
step d6, calculating root mean square error e of mu iteration output of heterogeneous fusion network based on self-attention μ The method comprises the steps of carrying out a first treatment on the surface of the Judgment e μ <error or mu>μ max If yes, the current network model A is stored μ The method comprises the steps of carrying out a first treatment on the surface of the Otherwise, after assigning mu+1 to mu, returning to the step d2; wherein error is a preset error value.
The present invention also provides a transformer fault diagnosis apparatus for a self-attention based heterogeneous network, the apparatus comprising:
the first training set acquisition module is used for acquiring a thermal infrared image of the transformer to perform preprocessing under the fault state of the transformer to obtain a first training set T 1 ；
The second training set acquisition module is used for acquiring the content of dissolved gas in the oil of the transformer under the fault state of the transformer and preprocessing to obtain a second training set T 2 ；
A model construction module for constructing a heterogeneous fusion network model based on self-attention, wherein the model comprises a self-attention LSTM network focusing on global time scale information, a self-attention residual error network focusing on local space information, a feature fusion layer and a Softmax classifier, and the first training set T 1 Input into the self-attention residual error network, a second training set T 2 The input is input into a self-attention LSTM network, the output end of the self-attention residual error network and the output end of the self-attention LSTM network are connected with the input end of a characteristic fusion layer, and the output end of the characteristic fusion layer is divided from SoftmaxThe class device is connected;
and the model training module is used for training the heterogeneous fusion network model based on the self-attention and carrying out fault diagnosis by using the trained model.
Further, the first training set acquisition module is further configured to:
obtaining L time periods when the transformer operates in h fault states; any first time period of the L time periods is marked as T l Will be the first time period T l Dividing into N equally spaced moments; acquiring thermal infrared images of the oil immersed transformer at any nth equidistant moment, preprocessing and normalizing to obtain a transformer fault image sample set containing h multiplied by N multiplied by L as a first training set T 1 。
Still further, the second training set acquisition module is further configured to:
collecting the content of dissolved gas in m oils of the oil immersed transformer at any nth equidistant moment and taking the content as a fault characteristic variable; thus, h transformer fault time sequence samples with the number of rows of N multiplied by L and the number of columns of m are formed and normalized, and the normalized transformer fault time sequence samples are obtained, wherein N is>m; setting the size of a sliding window as w multiplied by m, the step length as 1, and performing longitudinal sliding value taking on the normalized transformer fault time sequence samples to obtain h multiplied by N multiplied by L transformer fault matrixes with w multiplied by m as a second training set T 2 。
Still further, the self-attention residual network comprises three identical self-attention residual modules, which are serially connected in sequence.
Still further, the processing procedure of the self-attention residual error module is as follows:
1) Sequentially inputting two 2D convolutions into a transformer fault image Input1 to carry out residual mapping, and obtaining a feature map f with dimensions of C multiplied by H multiplied by W; transformer fault image Input1 is a first training set T 1 Data in (a);
2) After the feature map f is activated by adopting 1×1 convolution and Softmax functions, an intermediate feature vector with the dimension HW×1×1 can be obtained, and dot product is carried out on the intermediate feature vector and the feature map f, so that a channel with the dimension Cx1×1 can be obtainedAttention feature vector f 1 ；
3) Global feature vector f 1 After being Input into the two full-connection layers, the transformer is activated by adopting a Sigmoid function, performs dot product with the feature map f, and then performs short-circuit connection with the transformer fault image Input1 to obtain an output Y.
Still further, the self-attentive LSTM network comprises three identical self-attentive LSTM modules, which are serially connected in sequence.
Still further, the self-attention LSTM module comprises the following processing procedures:
1) The transformer fault matrix Input2 is respectively Input to each LSTM unit, and the hidden layer output h= { h at each moment can be obtained 1 ,h 2 ,...,h t And (b) wherein h t Is hs×1, and h is hs×t; the transformer fault matrix Input2 is the second training set T 2 Data in (a);
2) After the hidden layer output h at each moment is activated by adopting a 1D convolution and softmax function, an intermediate feature vector with the dimension of 1 Xhs can be obtained, dot product is carried out on the intermediate feature vector and h, and a time scale attention feature vector with the dimension of 1 Xt is obtained
3) Attention feature vector for time scaleAfter being sequentially input into two full-connection layers, the method is activated by adopting a Sigmoid function, and dot product is carried out on the active signal and H to obtain output H.
Still further, the model training module is further configured to:
step d1, defining the current iteration number of the neural network as mu, and initializing mu=1; the maximum iteration number is mu max The method comprises the steps of carrying out a first treatment on the surface of the Carrying out the mu-th random initialization on parameters of each layer in the network;
step d2, initializing i=1;
step d3, from the first training set T 1 Selecting the ith Zhang Bianya device fault image x i Transport and deliverObtaining a feature vector space feature vector F by entering the self-attention residual error network i,μ 1 The dimension is p×1; from the second training set T 2 Selecting the ith transformer fault matrix y i Inputting the self-attention LSTM network to obtain a time sequence feature vector F i,μ 2 The dimension is q multiplied by 1;
step d4, the space feature vector F i,μ 1 And the time sequence characteristic vector F i,μ 2 The input characteristic fusion layer is spliced end to end in sequence to obtain a transformer fault state characteristic vector F i,μ ＝[F i,μ 1 ,F i,μ 2 ] T The dimension is (p+q). Times.1; the fault state characteristic vector F of the transformer i,μ The input Softmax classifier obtains the transformer fault image x of the current network input i And transformer fault matrix y i Fault classification result H i ；
Step d5, after i+1 is assigned to i, judging whether i > h multiplied by N multiplied by L is true or not; if yes, continuing to execute the step d6, otherwise, returning to the step d3;
step d6, calculating root mean square error e of mu iteration output of heterogeneous fusion network based on self-attention μ The method comprises the steps of carrying out a first treatment on the surface of the Judgment e μ <error or mu>μ max If yes, the current network model A is stored μ The method comprises the steps of carrying out a first treatment on the surface of the Otherwise, after assigning mu+1 to mu, returning to the step d2; wherein error is a preset error value.
The invention has the advantages that:
(1) The self-attention LSTM network focuses on global time scale information, the self-attention residual error network focuses on local space information, the two information are fused through the feature fusion layer, a complete transformer fault feature vector is constructed, the local space information and the global time scale information are fully developed, space-time information is complete, accurate fault diagnosis results are output, a model is trained, fault diagnosis is carried out by using the trained model, and accuracy of the fault diagnosis results is further improved.
(2) The invention learns self information by constructing a self-attention LSTM network, and applies attention with different magnitudes to different time scale information; the characteristic diagrams of different channels are processed by constructing the self-attention residual error network, so that the fault characteristic information of the transformer is highlighted, the interference of redundant information is reduced, and the fault diagnosis accuracy is improved.
Drawings
Fig. 1 is a flowchart of a method for diagnosing a transformer fault of a self-attention based heterogeneous network disclosed in embodiment 1 of the present invention;
fig. 2 is a schematic structural diagram of a self-attention residual module in the method for diagnosing a transformer fault based on a self-attention heterogeneous network disclosed in embodiment 1 of the present invention;
fig. 3 is a schematic structural diagram of a self-attention LSTM module in the method for diagnosing a transformer fault based on a self-attention heterogeneous network disclosed in embodiment 1 of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions in the embodiments of the present invention will be clearly and completely described in the following in conjunction with the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Example 1
As shown in fig. 1, the present invention provides a transformer fault diagnosis method of a self-attention based heterogeneous network, the method comprising the steps of:
step a, obtaining L time periods when the transformer operates in h fault states; any first time period of the L time periods is marked as T l Will be the first time period T l Dividing into N equally spaced moments; acquiring thermal infrared images of the oil immersed transformer at any nth equidistant moment, preprocessing and normalizing to obtain a transformer fault image sample set containing h multiplied by N multiplied by L as a first training set T 1 ；
Step b, collecting the content of dissolved gas in m oils of the oil immersed transformer at any nth equidistant moment and taking the content as a fault characteristic variable; thus, h transformer fault time sequence samples with the number of rows of NxL and the number of columns of m are formed, normalization operation is carried out, and the transformer fault time sequence samples after normalization operation are obtained, wherein N is greater than m;
setting the size of a sliding window as w multiplied by m, the step length as 1, and performing longitudinal sliding value taking on the normalized transformer fault time sequence samples to obtain h multiplied by N multiplied by L transformer fault matrixes with w multiplied by m as a second training set T 2 ；
Step c, constructing a heterogeneous fusion network model based on self-attention;
the self-attention heterogeneous fusion network model comprises a self-attention LSTM network, a self-attention residual network, a characteristic fusion layer and a Softmax classifier; the self-attention LSTM network has an output dimension of p multiplied by 1, and the self-attention residual network has an output dimension of q multiplied by 1;
in step c1, the self-attention residual error network comprises three identical self-attention residual error modules which are sequentially connected in series, wherein the content of the self-attention residual error modules comprises:
1) Sequentially inputting two 2D convolutions into a transformer fault image Input1 to carry out residual mapping, and obtaining a feature map f with dimensions of C multiplied by H multiplied by W;
2) After the feature map f is activated by adopting a 1×1 convolution and Softmax function, an intermediate feature vector with the dimension HW×1×1 can be obtained, and the intermediate feature vector and the feature map f are subjected to dot product to obtain a channel attention feature vector f with the dimension C×1×1 1 ；
3) Global feature vector f 1 After being Input into the two full-connection layers, the transformer is activated by adopting a Sigmoid function, performs dot product with the feature map f, and then performs short-circuit connection with the transformer fault image Input1 to obtain an output Y. The self-attention residual module structure is shown in fig. 2.
Step c2, the self-attention LSTM network includes three identical self-attention LSTM modules connected in series, where the self-attention LSTM module content includes:
1) The transformer fault matrix Input2 is respectively Input into each LSTM unit to obtain each unitHidden layer output h= { h at each instant 1 ,h 2 ,...,h t And (b) wherein h t Is hs×1, and h is hs×t;
2) After the h is activated by adopting a 1D convolution and softmax function, an intermediate feature vector with the dimension of 1 Xhs can be obtained, dot product is carried out on the intermediate feature vector and h, and a time scale attention feature vector with the dimension of 1 Xt is obtained
3) Will beAfter being sequentially input into two full-connection layers, the method is activated by adopting a Sigmoid function, and dot product is carried out on the active signal and H to obtain output H. The self-attention LSTM module structure is shown in fig. 3.
Step d: training a heterogeneous fusion network model based on self-attention, and performing fault diagnosis by using the trained model, wherein the specific process is as follows:
step d1, defining the current iteration number of the neural network as mu, and initializing mu=1; the maximum iteration number is mu max The method comprises the steps of carrying out a first treatment on the surface of the Carrying out the mu-th random initialization on parameters of each layer in the disturbance neural network so as to obtain a self-attention-based heterogeneous fusion network of the mu-th iteration;
step d2, initializing i=1;
step d3, from the first training set T 1 Selecting the ith Zhang Bianya device fault image x i Inputting the self-attention residual error network to obtain a feature vector space feature vector F i,μ 1 The dimension is p×1; from the second training set T 2 Selecting the ith transformer fault matrix y i Inputting the self-attention LSTM network to obtain a time sequence feature vector F i,μ 2 The dimension is q multiplied by 1;
step d4, the space feature vector F i,μ 1 And the time sequence characteristic vector F i,μ 2 The input characteristic fusion layer is spliced end to end in sequence to obtain a transformer fault state characteristic vector F i,μ ＝[F i,μ 1 ,F i,μ 2 ] T The dimension is (p+q). Times.1; the fault state characteristic vector F of the transformer i,μ The input Softmax classifier obtains the transformer fault image x of the current network input i And transformer fault matrix y i Fault classification result H i 。
Step d5, after i+1 is assigned to i, judging whether i > h multiplied by N multiplied by L is true or not; if yes, continuing to execute the step d6, otherwise, returning to the step d3;
step d6, calculating root mean square error e of mu iteration output of heterogeneous fusion network based on self-attention μ The method comprises the steps of carrying out a first treatment on the surface of the Judgment e μ <error or mu>μ max If yes, the current network model A is stored μ The method comprises the steps of carrying out a first treatment on the surface of the Otherwise, after assigning mu+1 to mu, returning to the step d2; wherein error is a preset error value.
Through the technical scheme, the intelligent fault diagnosis method for the transformer based on the self-attention heterogeneous fusion network is provided, so that the fault diagnosis of the transformer can be rapidly and accurately carried out, and the safety requirement of power equipment is met. Learning self information by constructing a self-attention LSTM network, and applying attention with different magnitudes to different time mark information; the characteristic diagrams of different channels are processed by constructing the self-attention residual error network, so that the fault characteristic information of the transformer is highlighted, the interference of redundant information is reduced, and the fault diagnosis accuracy is improved. The complete fault feature of the transformer is obtained by adopting a double-branch fusion network, the self-attention LSTM network focuses on global time scale information, the self-attention residual error network focuses on local space information, and a complete fault feature vector of the transformer is constructed through a feature fusion layer, so that the method is closer to a human local-global cognition method, the local space information and the global time scale information are fully explored, and the model has excellent generalization capability.
Example 2
Based on embodiment 1, embodiment 2 of the present invention further provides a transformer fault diagnosis apparatus for a self-attention based heterogeneous network, the apparatus comprising:
a first training set acquisition module for use in a transformerUnder the fault state, acquiring a thermal infrared image of a transformer and preprocessing to obtain a first training set T 1 ；
The second training set acquisition module is used for acquiring the content of dissolved gas in the oil of the transformer under the fault state of the transformer and preprocessing to obtain a second training set T 2 ；
A model construction module for constructing a heterogeneous fusion network model based on self-attention, wherein the model comprises a self-attention LSTM network focusing on global time scale information, a self-attention residual error network focusing on local space information, a feature fusion layer and a Softmax classifier, and the first training set T 1 Input into the self-attention residual error network, a second training set T 2 The method comprises the steps of inputting the data into a self-attention LSTM network, wherein the output end of the self-attention residual error network and the output end of the self-attention LSTM network are connected with the input end of a feature fusion layer, and the output end of the feature fusion layer is connected with a Softmax classifier;
and the model training module is used for training the heterogeneous fusion network model based on the self-attention and carrying out fault diagnosis by using the trained model.
Specifically, the first training set acquisition module is further configured to:
obtaining L time periods when the transformer operates in h fault states; any first time period of the L time periods is marked as T l Will be the first time period T l Dividing into N equally spaced moments; acquiring thermal infrared images of the oil immersed transformer at any nth equidistant moment, preprocessing and normalizing to obtain a transformer fault image sample set containing h multiplied by N multiplied by L as a first training set T 1 。
More specifically, the second training set acquisition module is further configured to:
collecting the content of dissolved gas in m oils of the oil immersed transformer at any nth equidistant moment and taking the content as a fault characteristic variable; thus, h transformer fault time sequence samples with the number of rows of N multiplied by L and the number of columns of m are formed and normalized, and the normalized transformer fault time sequence samples are obtained, wherein N is>m; setting the size of the sliding window as w multiplied by m, the step length as 1, and normalizing the sliding windowPerforming longitudinal sliding value taking on the transformer fault time sequence samples after chemical operation to obtain h multiplied by N multiplied by L multiplied by w multiplied by m transformer fault matrixes as a second training set T 2 。
More specifically, the self-attention residual network comprises three identical self-attention residual modules which are sequentially connected in series.
More specifically, the processing procedure of the self-attention residual error module is as follows:
1) Sequentially inputting two 2D convolutions into a transformer fault image Input1 to carry out residual mapping, and obtaining a feature map f with dimensions of C multiplied by H multiplied by W; transformer fault image Input1 is a first training set T 1 Data in (a);
2) After the feature map f is activated by adopting a 1×1 convolution and Softmax function, an intermediate feature vector with the dimension HW×1×1 can be obtained, and the intermediate feature vector and the feature map f are subjected to dot product to obtain a channel attention feature vector f with the dimension C×1×1 1 ；
3) Global feature vector f 1 After being Input into the two full-connection layers, the transformer is activated by adopting a Sigmoid function, performs dot product with the feature map f, and then performs short-circuit connection with the transformer fault image Input1 to obtain an output Y.
More specifically, the self-attentive LSTM network includes three identical self-attentive LSTM modules, which are serially connected in sequence.
More specifically, the self-attention LSTM module comprises the following processing procedures:
1) The transformer fault matrix Input2 is respectively Input to each LSTM unit, and the hidden layer output h= { h at each moment can be obtained 1 ,h 2 ,...,h t And (b) wherein h t Is hs×1, and h is hs×t; the transformer fault matrix Input2 is the second training set T 2 Data in (a);
2) After the hidden layer output h at each moment is activated by adopting a 1D convolution and softmax function, an intermediate feature vector with the dimension of 1 Xhs can be obtained, dot product is carried out on the intermediate feature vector and h, and a time scale attention feature vector with the dimension of 1 Xt is obtained
3) Attention feature vector for time scaleAfter being sequentially input into two full-connection layers, the method is activated by adopting a Sigmoid function, and dot product is carried out on the active signal and H to obtain output H.
More specifically, the model training module is further configured to:
step d1, defining the current iteration number of the neural network as mu, and initializing mu=1; the maximum iteration number is mu max The method comprises the steps of carrying out a first treatment on the surface of the Carrying out the mu-th random initialization on parameters of each layer in the network;
step d2, initializing i=1;
step d3, from the first training set T 1 Selecting the ith Zhang Bianya device fault image x i Inputting the self-attention residual error network to obtain a feature vector space feature vector F i,μ 1 The dimension is p×1; from the second training set T 2 Selecting the ith transformer fault matrix y i Inputting the self-attention LSTM network to obtain a time sequence feature vector F i,μ 2 The dimension is q multiplied by 1;
step d4, the space feature vector F i,μ 1 And the time sequence characteristic vector F i,μ 2 The input characteristic fusion layer is spliced end to end in sequence to obtain a transformer fault state characteristic vector F i,μ ＝[F i,μ 1 ,F i,μ 2 ] T The dimension is (p+q). Times.1; the fault state characteristic vector F of the transformer i,μ The input Softmax classifier obtains the transformer fault image x of the current network input i And transformer fault matrix y i Fault classification result H i ；
Step d5, after i+1 is assigned to i, judging whether i > h multiplied by N multiplied by L is true or not; if yes, continuing to execute the step d6, otherwise, returning to the step d3;
step d6, calculating root mean square error e of mu iteration output of heterogeneous fusion network based on self-attention μ The method comprises the steps of carrying out a first treatment on the surface of the Judginge μ <error or mu>μ max If yes, the current network model A is stored μ The method comprises the steps of carrying out a first treatment on the surface of the Otherwise, after assigning mu+1 to mu, returning to the step d2; wherein error is a preset error value.
The above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. A method for diagnosing a transformer fault in a self-care heterogeneous network, the method comprising:
step a: under the fault state of the transformer, acquiring a thermal infrared image of the transformer and preprocessing to obtain a first training set T 1 ；
Step b: under the fault state of the transformer, the content of dissolved gas in the oil of the transformer is collected and preprocessed to obtain a second training set T 2 ；
Step c: constructing a heterogeneous fusion network model based on self-attention, wherein the model comprises a self-attention LSTM network focusing on global time scale information, a self-attention residual error network focusing on local space information, a characteristic fusion layer and a Softmax classifier, and the first training set T 1 Input into the self-attention residual error network, a second training set T 2 The method comprises the steps of inputting the data into a self-attention LSTM network, wherein the output end of the self-attention residual error network and the output end of the self-attention LSTM network are connected with the input end of a feature fusion layer, and the output end of the feature fusion layer is connected with a Softmax classifier;
step d: and training a heterogeneous fusion network model based on self-attention, and performing fault diagnosis by using the trained model.
2. The method for diagnosing a transformer fault in a self-attention based heterogeneous network according to claim 1, wherein the step a includes:
obtaining L time periods when the transformer operates in h fault states; any first time period of the L time periods is marked as T l Will be the first time period T l Dividing into N equally spaced moments; acquiring thermal infrared images of the oil immersed transformer at any nth equidistant moment, preprocessing and normalizing to obtain a transformer fault image sample set containing h multiplied by N multiplied by L as a first training set T 1 。
3. The method for diagnosing a transformer fault in a self-attention based heterogeneous network according to claim 2, wherein the step b comprises:
collecting the content of dissolved gas in m oils of the oil immersed transformer at any nth equidistant moment and taking the content as a fault characteristic variable; thus, h transformer fault time sequence samples with the number of rows of N multiplied by L and the number of columns of m are formed and normalized, and the normalized transformer fault time sequence samples are obtained, wherein N is>m; setting the size of a sliding window as w multiplied by m, the step length as 1, and performing longitudinal sliding value taking on the normalized transformer fault time sequence samples to obtain h multiplied by N multiplied by L transformer fault matrixes with w multiplied by m as a second training set T 2 。
4. A method of diagnosing a transformer failure in a self-attention based heterogeneous network as claimed in claim 3, wherein said self-attention residual network includes three identical self-attention residual modules connected in series in sequence.
5. The transformer fault diagnosis method based on the self-attention heterogeneous network according to claim 4, wherein the processing procedure of the self-attention residual error module is as follows:
1) Sequentially inputting two 2D convolutions into a transformer fault image Input1 to carry out residual mapping, and obtaining a feature map f with dimensions of C multiplied by H multiplied by W; transformer fault image Input1 is the first trainingTraining set T 1 Data in (a);
2) After the feature map f is activated by adopting a 1×1 convolution and Softmax function, an intermediate feature vector with the dimension HW×1×1 can be obtained, and the intermediate feature vector and the feature map f are subjected to dot product to obtain a channel attention feature vector f with the dimension C×1×1 1 ；
3) Global feature vector f 1 After being Input into the two full-connection layers, the transformer is activated by adopting a Sigmoid function, performs dot product with the feature map f, and then performs short-circuit connection with the transformer fault image Input1 to obtain an output Y.
6. The method of claim 4, wherein the self-attentive LSTM network includes three identical self-attentive LSTM modules connected in series.
7. The method for diagnosing a transformer fault in a self-focusing heterogeneous network according to claim 6, wherein the self-focusing LSTM module is processed as follows:
1) The transformer fault matrix Input2 is respectively Input to each LSTM unit, and the hidden layer output h= { h at each moment can be obtained 1 ,h 2 ,...,h t And (b) wherein h t Is hs×1, and h is hs×t; the transformer fault matrix Input2 is the second training set T 2 Data in (a);
2) After the hidden layer output h at each moment is activated by adopting a 1D convolution and softmax function, an intermediate feature vector with the dimension of 1 Xhs can be obtained, dot product is carried out on the intermediate feature vector and h, and a time scale attention feature vector with the dimension of 1 Xt is obtained
3) Attention feature vector for time scaleSequentially inputting to two full-connection layers, and exciting with Sigmoid functionAnd (3) carrying out dot product on the obtained product with H to obtain an output H.
8. The method for diagnosing a transformer failure in a self-attention based heterogeneous network as recited in claim 7, wherein said step d includes:
step d1, defining the current iteration number of the neural network as mu, and initializing mu=1; the maximum iteration number is mu max The method comprises the steps of carrying out a first treatment on the surface of the Carrying out the mu-th random initialization on parameters of each layer in the network;
step d2, initializing i=1;
step d3, from the first training set T 1 Selecting the ith Zhang Bianya device fault image x i Inputting the self-attention residual error network to obtain a feature vector space feature vector F i,μ 1 The dimension is p×1; from the second training set T 2 Selecting the ith transformer fault matrix y i Inputting the self-attention LSTM network to obtain a time sequence feature vector F i,μ 2 The dimension is q multiplied by 1;
step d4, the space feature vector F i,μ 1 And the time sequence characteristic vector F i,μ 2 The input characteristic fusion layer is spliced end to end in sequence to obtain a transformer fault state characteristic vector F i,μ ＝[F i,μ 1 ,F i,μ 2 ] T The dimension is (p+q). Times.1; the fault state characteristic vector F of the transformer i,μ The input Softmax classifier obtains the transformer fault image x of the current network input i And transformer fault matrix y i Fault classification result H i ；
Step d5, after i+1 is assigned to i, judging whether i > h multiplied by N multiplied by L is true or not; if yes, continuing to execute the step d6, otherwise, returning to the step d3;
step d6, calculating root mean square error e of mu iteration output of heterogeneous fusion network based on self-attention μ The method comprises the steps of carrying out a first treatment on the surface of the Judgment e μ <error or mu>μ max If yes, the current network model A is stored μ The method comprises the steps of carrying out a first treatment on the surface of the Otherwise, after assigning mu+1 to mu, returning to the step d2; wherein error is presetIs a function of the error value of (a).
9. A transformer fault diagnosis apparatus for a self-attention based heterogeneous network, the apparatus comprising:
the first training set acquisition module is used for acquiring a thermal infrared image of the transformer to perform preprocessing under the fault state of the transformer to obtain a first training set T 1 ；
The second training set acquisition module is used for acquiring the content of dissolved gas in the oil of the transformer under the fault state of the transformer and preprocessing to obtain a second training set T 2 ；
A model construction module for constructing a heterogeneous fusion network model based on self-attention, wherein the model comprises a self-attention LSTM network focusing on global time scale information, a self-attention residual error network focusing on local space information, a feature fusion layer and a Softmax classifier, and the first training set T 1 Input into the self-attention residual error network, a second training set T 2 The method comprises the steps of inputting the data into a self-attention LSTM network, wherein the output end of the self-attention residual error network and the output end of the self-attention LSTM network are connected with the input end of a feature fusion layer, and the output end of the feature fusion layer is connected with a Softmax classifier;
and the model training module is used for training the heterogeneous fusion network model based on the self-attention and carrying out fault diagnosis by using the trained model.
10. The self-attention based heterogeneous network transformer fault diagnosis device of claim 9, wherein the first training set acquisition module is further configured to:
obtaining L time periods when the transformer operates in h fault states; any first time period of the L time periods is marked as T l Will be the first time period T l Dividing into N equally spaced moments; acquiring thermal infrared images of the oil immersed transformer at any nth equidistant moment, preprocessing and normalizing to obtain a transformer fault image sample set containing h multiplied by N multiplied by L as a first training set T 1 。
CN202310857743.XA 2023-07-12 2023-07-12 Transformer fault diagnosis method and device based on self-attention heterogeneous network Pending CN116955951A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310857743.XA CN116955951A (en) 2023-07-12 2023-07-12 Transformer fault diagnosis method and device based on self-attention heterogeneous network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310857743.XA CN116955951A (en) 2023-07-12 2023-07-12 Transformer fault diagnosis method and device based on self-attention heterogeneous network

Publications (1)

Publication Number Publication Date
CN116955951A true CN116955951A (en) 2023-10-27

Family

ID=88452240

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310857743.XA Pending CN116955951A (en) 2023-07-12 2023-07-12 Transformer fault diagnosis method and device based on self-attention heterogeneous network

Country Status (1)

Country Link
CN (1) CN116955951A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117725529A (en) * 2024-02-18 2024-03-19 南京邮电大学 Transformer fault diagnosis method based on multi-mode self-attention mechanism

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117725529A (en) * 2024-02-18 2024-03-19 南京邮电大学 Transformer fault diagnosis method based on multi-mode self-attention mechanism
CN117725529B (en) * 2024-02-18 2024-05-24 南京邮电大学 Transformer fault diagnosis method based on multi-mode self-attention mechanism

Similar Documents

Publication Publication Date Title
CN109612708B (en) Power transformer on-line detection system and method based on improved convolutional neural network
Islam et al. Automated bearing fault diagnosis scheme using 2D representation of wavelet packet transform and deep convolutional neural network
Wang et al. A novel convolutional neural network based fault recognition method via image fusion of multi-vibration-signals
CN116955951A (en) Transformer fault diagnosis method and device based on self-attention heterogeneous network
Jin et al. A multi-scale convolutional neural network for bearing compound fault diagnosis under various noise conditions
CN112200244A (en) Intelligent detection method for anomaly of aerospace engine based on hierarchical countermeasure training
CN111881627A (en) Nuclear power device fault diagnosis method and system
Zhao et al. Fault diagnosis of shaft misalignment and crack in rotor system based on MI-CNN
CN111272886A (en) Pressure vessel damage positioning method based on convolutional neural network
CN114022715A (en) Method and device for detecting lead sealing defect of high-voltage cable terminal
CN116842379A (en) Mechanical bearing residual service life prediction method based on DRSN-CS and BiGRU+MLP models
Di et al. Fault diagnosis of rotating machinery based on domain adversarial training of neural networks
CN113837994B (en) Photovoltaic panel defect diagnosis method based on edge detection convolutional neural network
Bao et al. A transformer model-based approach to bearing fault diagnosis
Zhang et al. CarNet: A dual correlation method for health perception of rotating machinery
CN107689015A (en) A kind of improved power system bad data recognition method
Hejazi et al. Novel preprocessing of multimodal condition monitoring data for classifying induction motor faults using deep learning methods
CN117056865B (en) Method and device for diagnosing operation faults of machine pump equipment based on feature fusion
CN108761263B (en) Fault diagnosis system based on evidence theory
CN113112039A (en) Active power distribution system initial fault identification method based on time-frequency memory recurrent neural network
CN116503354A (en) Method and device for detecting and evaluating hot spots of photovoltaic cells based on multi-mode fusion
CN115019129B (en) Double-active-bridge converter fault diagnosis method and system based on time sequence imaging and image fusion
Wang et al. Few‐shot multiscene fault diagnosis of rolling bearing under compound variable working conditions
Yao et al. Remaining useful life estimation by empirical mode decomposition and ensemble deep convolution neural networks
Shi et al. Using long short term memory based approaches for carbon steel fatigue remaining useful life prediction

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination