CN111737907A - Transformer fault diagnosis method and device based on deep learning and DGA - Google Patents

Transformer fault diagnosis method and device based on deep learning and DGA Download PDF

Info

Publication number
CN111737907A
CN111737907A CN202010483448.9A CN202010483448A CN111737907A CN 111737907 A CN111737907 A CN 111737907A CN 202010483448 A CN202010483448 A CN 202010483448A CN 111737907 A CN111737907 A CN 111737907A
Authority
CN
China
Prior art keywords
deep learning
fault diagnosis
transformer fault
learning model
layer
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
CN202010483448.9A
Other languages
Chinese (zh)
Inventor
刘力卿
王伟
张鑫
张弛
郗晓光
张春晖
李琳
冯军基
马昊
魏菊芳
姚创
段明辉
文清丰
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.)
State Grid Corp of China SGCC
State Grid Tianjin Electric Power Co Ltd
Electric Power Research Institute of State Grid Tianjin Electric Power Co Ltd
Original Assignee
State Grid Corp of China SGCC
State Grid Tianjin Electric Power Co Ltd
Electric Power Research Institute of State Grid Tianjin 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 State Grid Corp of China SGCC, State Grid Tianjin Electric Power Co Ltd, Electric Power Research Institute of State Grid Tianjin Electric Power Co Ltd filed Critical State Grid Corp of China SGCC
Priority to CN202010483448.9A priority Critical patent/CN111737907A/en
Publication of CN111737907A publication Critical patent/CN111737907A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • 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
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/047—Probabilistic or stochastic networks
    • 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)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Test And Diagnosis Of Digital Computers (AREA)

Abstract

The invention relates to a transformer fault diagnosis method and a device based on deep learning and DGA, which are technically characterized in that: the method comprises the following steps: taking the volume fractions of DGA characteristic gases H2, CH4, C2H6, C2H4 and C2H2 as diagnostic samples; inputting the sample to be diagnosed into the stack type sparse self-coding deep learning model for transformer fault diagnosis after fine adjustment, and outputting a transformer fault diagnosis result. The method can obviously improve the fault diagnosis effect of the transformer and improve the accuracy of the diagnosis result when the training samples are unbalanced.

Description

Transformer fault diagnosis method and device based on deep learning and DGA
Technical Field
The invention belongs to the technical field of transformer fault diagnosis, relates to a transformer fault diagnosis method and device based on deep learning and DGA, and particularly relates to a transformer fault diagnosis method and device based on weighted comprehensive loss optimization (DGA).
Background
The power transformer is one of the most important devices of the power system, and the accurate diagnosis and the rapid treatment of the transformer fault have important significance for improving the safe and stable operation level of the power system. However, the structure and the operating environment of the transformer are complex, and how to effectively improve the fault diagnosis effect is still a difficult problem.
Analysis of Dissolved Gas (DGA) in oil is an important basis for realizing transformer fault diagnosis, and simple and convenient transformer fault diagnosis methods such as a three-ratio method, an improved three-ratio method and the like are formed on the basis of the DGA. However, the method has the defects of over-absolute characteristic gas ratio limit, incomplete coding and the like, and the fault diagnosis effect is influenced. In order to solve the above defects, Artificial intelligence and Machine learning methods such as an expert system method, a fuzzy theory method, A Neural Network (ANN) method, a Support Vector Machine (SVM) method, and the like are gradually applied, and a good effect is obtained. However, the diagnosis effect of the transformer fault diagnosis method based on the expert system is greatly influenced by the priori knowledge; the transformer fault diagnosis method based on the fuzzy theory has limited learning capacity, and the fault diagnosis effect is greatly influenced by the initial clustering center; when the ANN processes large sample data, the convergence speed is low, and the ANN is easy to fall into local optimum; the SVM method is essentially a two-classifier method, a plurality of SVM classifiers are needed for realizing the fault diagnosis of the transformer, and the parameter setting of each classifier is complicated, so that the fault diagnosis effect is influenced.
The traditional transformer fault diagnosis method based on artificial intelligence and machine learning is a shallow learning method, has the defects of insufficient learning ability, insufficient feature mining ability and the like, and influences the transformer fault diagnosis effect. Compared with a shallow learning method, the deep learning theory has a multi-hidden-layer structure, and high-efficiency approximation of complex input can be realized. Based on the deep learning theory, the deep mining and analysis of the original characteristics can be completed, and the fault-tolerant and expansion capability is high-efficient. The deep learning has outstanding effects in the aspects of feature learning, classification prediction, intelligent decision and the like, and is rapidly popularized and applied.
However, the traditional fault diagnosis method based on deep learning generally realizes optimization based on a cross entropy loss function, however, the method does not limit the misjudgment probability value, and the fault diagnosis effect is influenced; in addition, due to various limitations of equipment, environment and the like, complete samples are difficult to obtain in the running and monitoring processes of the transformer, and the imbalance of the training samples easily affects the fault diagnosis effect.
Disclosure of Invention
The invention aims to overcome the defects of the prior art, and provides a transformer fault diagnosis method and device based on weighted comprehensive loss optimization deep learning and DGA (differential global analysis), which can obviously improve the transformer fault diagnosis effect and improve the accuracy of diagnosis results when training samples are unbalanced.
The technical problem to be solved by the invention is realized by adopting the following technical scheme:
a transformer fault diagnosis method based on deep learning and DGA is characterized in that:
taking the volume fractions of DGA characteristic gases H2, CH4, C2H6, C2H4 and C2H2 as diagnostic samples;
inputting the sample to be diagnosed into the stack type sparse self-coding deep learning model for transformer fault diagnosis after fine adjustment, and outputting a transformer fault diagnosis result.
And before inputting the sample to be diagnosed into the trimmed stacked sparse self-coding deep learning model for transformer fault diagnosis, the method further comprises the following steps:
constructing a stack type sparse self-coding deep learning model for transformer fault diagnosis;
setting an input layer of a stack type sparse self-coding deep learning model;
setting an output layer of a stack type sparse self-coding deep learning model;
pre-training network parameters of a stack type sparse self-coding deep learning model for transformer fault diagnosis;
and fine-tuning parameters of a stacked sparse self-coding deep learning model for transformer fault diagnosis.
Moreover, the specific steps of constructing the stacked sparse self-coding deep learning model for transformer fault diagnosis include:
(1) let input and output layers have N number of neurons, hidden layer has M number of neurons, and let input x ∈ RNWhich can be expressed as h ∈ R after being encoded and decoded respectivelyMAnd
Figure BDA0002518242150000031
and the encoding and decoding processes can be represented by equation (1) and equation (2), respectively, that is:
h＝f(Wx+b) (1)
Figure BDA0002518242150000032
in the formula, W, W' is encoding and decoding weight matrix respectively; b. b' are respectively encoding and decoding offset vectors; f (-) and g (-) are nonlinear activation functions in the encoding and decoding process, and a Sigmoid function can be generally adopted;
(2) by adjusting network parameters, realize
Figure BDA0002518242150000033
When x reconstruction error is minimum, h is used as the intrinsic characteristic of the original input, and sparsity limitation is performed on the hidden layer, so that a Sparse Auto-Encoder (SAE) is formed, and the SAE cost function is as follows:
Figure BDA0002518242150000034
wherein β is a sparse penalty factor coefficient, which can be generally set to 0.3, KL(. h) is KL divergence; rho0Is a sparsity parameter;
Figure BDA0002518242150000035
is the mean activation of neurons in the hidden layer, and has
Figure BDA0002518242150000036
(3) And stacking the multiple levels of SAEs according to a stacked structure, wherein all SAE input layers are upper SAE output layers, and further constructing a stacked sparse self-coding deep learning model for transformer fault diagnosis.
Moreover, the specific method for setting the input layer of the stacked sparse self-coding deep learning model is as follows:
the input layer of the stacked sparse self-coding deep learning model for transformer fault diagnosis is set to be 5 neurons, and is respectively expressed as x ═ x (x)1,x2,x3,x4,x5)TThe elements thereof each represent hydrogen H2Methane CH4Ethane C2H6Ethylene C2H4Acetylene C2H2Volume fraction of (a).
Moreover, the specific method for setting the output layer of the stacked sparse self-coding deep learning model is as follows:
the Softmax layer is set as an output layer of the failure diagnosis model, and the layer contains 9 neurons, which are respectively expressed as p ═ p (p)1,p2,p3,p4,p5,p6,p7,p8,p9)TThe elements thereof represent the probability of diagnosing as respective states.
Moreover, the specific steps of pre-training the network parameters of the stacked sparse self-coding deep learning model for transformer fault diagnosis include:
(1) for the training set { (x)i,yi)|i∈1,2,…,N，yi∈ 1,2, … K }, wherein x isiAnd yiRespectively representing the ith training sample feature vector and a state label, and totally having K states; after the state labels of the training set are hidden, the original training set is changed into a label-free data set { xi|i∈1,2,…,N}；
(2) And removing the Softmax layer, adding a decoding layer, and sequentially realizing the training of network parameters of each layer of the SSAE deep learning model by adopting a layer-by-layer greedy method according to an SAE cost function based on label-free sample data.
Moreover, the specific step of fine-tuning the parameters of the stacked sparse self-coding deep learning model for transformer fault diagnosis includes:
(1) removing a traditional output layer of the SSAE deep learning model, adding a Softmax classification layer, and realizing optimization of network parameters of each layer of the SSAE based on a cross entropy loss function gradient value and a BP algorithm;
(2) for training sample (x)i,yi) After SSAE deep learning model, the cross entropy loss value CiCan be expressed as:
Figure BDA0002518242150000041
in the formula 1{k＝yiIs an indicative function, if k is yiIf k is not equal to y, its value is 1iA value of 0;
Figure BDA0002518242150000042
the probability of the state belonging to the kth class is judged for the kth neuron value of the Softmax layer, namely the SSAE deep learning model;
(3) for sample (x)i,yi) Of the combined loss function LiComprises the following steps:
Figure BDA0002518242150000043
in the formula (I), the compound is shown in the specification,
Figure BDA0002518242150000044
to determine the sample as the state k by mistake (k ≠ y)i) Loss values, and have:
Figure BDA0002518242150000045
(4) sample (x)i,yi) Loss weight value when determined to be state k
Figure BDA0002518242150000046
Comprises the following steps:
Figure BDA0002518242150000047
(5) combining the loss weight with the composite loss function to form a weighted composite loss function, as shown in equation (8):
Figure BDA0002518242150000048
(6) adopting a supervised learning mechanism, adding a sample fault type, firstly taking SSAE network parameters obtained by pre-training as initial parameters, and then taking a weighted loss function value between each training sample prediction fault and a real fault thereof as a network convergence cost; and then summing the weighted loss function values of the training samples to obtain the total loss, and performing optimization fine adjustment on the network parameters of the SSAE diagnostic model layer by using a BP algorithm gradient descent method with the aim of minimum total loss.
A transformer fault diagnosis device based on deep learning and DGA (differential global analysis), comprising:
the diagnostic sample input module takes the volume fractions of DGA characteristic gases H2, CH4, C2H6, C2H4 and C2H2 as diagnostic samples;
and the to-be-detected transformer fault diagnosis result output module inputs the to-be-diagnosed sample into the fine-tuned stack type sparse self-coding deep learning model for transformer fault diagnosis and outputs a transformer fault diagnosis result.
Moreover, the transformer fault diagnosis device based on deep learning and DGA further comprises:
the transformer fault diagnosis model building module is used for building a stack type sparse self-coding deep learning model for transformer fault diagnosis;
the input layer setting module is used for setting an input layer of the stack type sparse self-coding deep learning model;
the output layer setting module is used for setting an output layer of the stacked sparse self-coding deep learning model;
the parameter pre-training module is used for pre-training network parameters of a stack type sparse self-coding deep learning model for transformer fault diagnosis;
and the parameter fine-tuning module is used for fine-tuning network parameters of the stacked sparse self-coding deep learning model for transformer fault diagnosis.
Moreover, the transformer fault diagnosis model building module includes:
an encoding and decoding module for setting N neurons in input and output layers and M neurons in hidden layer, and setting input x ∈ RNWhich can be expressed as h ∈ R after being encoded and decoded respectivelyMAnd
Figure BDA0002518242150000051
and the encoding and decoding processes can be represented by equation (1) and equation (2), respectively, that is:
h＝f(Wx+b) (1)
Figure BDA0002518242150000052
in the formula, W, W' is encoding and decoding weight matrix respectively; b. b' are respectively encoding and decoding offset vectors; f (-) and g (-) are nonlinear activation functions in the encoding and decoding process, and a Sigmoid function can be generally adopted;
the sparse self-encoder forming module is used for realizing the adjustment of network parameters
Figure BDA0002518242150000064
When x reconstruction error is minimum, h is used as the intrinsic characteristic of the original input, and sparsity limitation is performed on the hidden layer, so that a sparse auto-Encoder (SAE) is formed, and the SAE cost function is as follows:
Figure BDA0002518242150000061
wherein β is a sparse penalty factor coefficient, which can be generally set to 0.3, KL(. h) is KL divergence; rho0Is a sparsity parameter;
Figure BDA0002518242150000062
is the mean activation of neurons in the hidden layer, and has
Figure BDA0002518242150000063
And the multi-level SAE stacking module is used for stacking the multi-level SAEs according to a stacked structure, and each SAE input layer is an upper SAE output layer, so that a stacked sparse self-coding deep learning model for transformer fault diagnosis is constructed.
The invention has the advantages and positive effects that:
the invention provides a transformer fault diagnosis method based on weighted comprehensive loss improved deep learning and DGA (differential global analysis), which uses DGA (differential global analysis) characteristic gas H2、CH4、C2H6、C2H4、C2H2The volume fraction of the volume fraction is used as input, a transformer fault diagnosis model is constructed by adopting a stack-type sparse self-coding deep learning theory, and meanwhile, network parameters of the deep learning model are finely adjusted by adopting a weighted comprehensive loss function, so that the adverse effect of unbalance of training samples is weakened, and the fault diagnosis effect is improved. The method provided by the invention can obviously improve the transformer fault diagnosis effect and improve the accuracy of the diagnosis result when the training samples are unbalanced.
Drawings
FIG. 1 is a diagram of a transformer fault diagnosis model according to the present invention;
FIG. 2 is a diagram of a basic self-encoder structure according to the present invention;
FIG. 3 is a flow chart of pre-training of network parameters of a transformer fault diagnosis model according to the present invention;
FIG. 4 is a flow chart of fine tuning of network parameters of the transformer fault diagnosis model of the present invention;
FIG. 5 is a flowchart of the transformer fault diagnosis method based on weighted integrated loss optimization deep learning and DGA according to the present invention.
Detailed Description
The embodiments of the invention will be described in further detail below with reference to the accompanying drawings:
example 1:
a transformer fault diagnosis method based on deep learning and DGA, as shown in fig. 1 to 5, comprising the following steps:
step 1, constructing a stack type sparse self-coding deep learning model for transformer fault diagnosis;
as shown in fig. 2, an Auto-encoder (AE) is one of typical structures of deep learning. The AE output layer and the input layer have equal neuron numbers, the process from the input layer to the hidden layer is coding, and the process from the hidden layer to the output layer is decoding.
The specific steps of the step 1 comprise:
(1) let input and output layers have N number of neurons, hidden layer has M number of neurons, and let input x ∈ RNWhich can be expressed as h ∈ R after being encoded and decoded respectivelyMAnd
Figure BDA0002518242150000071
and the encoding and decoding processes can be represented by equation (1) and equation (2), respectively, that is:
h＝f(Wx+b) (1)
Figure BDA0002518242150000072
in the formula, W, W' is encoding and decoding weight matrix respectively; b. b' are respectively encoding and decoding offset vectors; f (-) and g (-) are nonlinear activation functions in the encoding and decoding process, and a Sigmoid function can be generally adopted;
(2) by adjusting network parameters, realize
Figure BDA0002518242150000076
When x has the minimum reconstruction error, h can be used as the intrinsic characteristic of the original input;
when the number of neurons in the hidden layer is large, the AE accuracy is high, but the calculation amount is large. In order to improve the convergence speed, sparsity limitation can be performed on the hidden layer, so as to form a Sparse Auto-Encoder (SAE), where the SAE cost function is:
Figure BDA0002518242150000073
wherein β is a sparse penalty factor coefficient, which can be generally set to 0.3, KL(. h) is KL divergence; rho0The sparsity parameter can be set to 0.05 in general;
Figure BDA0002518242150000074
is the mean activation of neurons in the hidden layer, and has
Figure BDA0002518242150000075
(3) SAE is still a shallow learning model, in order to realize feature depth extraction, multiple levels of SAE can be stacked according to a stacked structure, and each SAE input layer is an upper SAE output layer, so that a stacked Sparse Auto-Encoder (SSAE) deep learning model is constructed. The SSAE deep learning model can obtain more complex and abstract deep feature expression of the original input.
Step 2, setting an input layer of a stack type sparse self-coding deep learning model;
most of large-scale power transformers are oil-immersed, when the state of the transformer is abnormal, the components and the content of dissolved gas in oil can be changed along with the abnormal state, fault diagnosis of the transformer can be realized on the basis of DGA, and H is generally selected2、CH4、C2H6、C2H4、C2H2As the characteristic gas.
Therefore, as shown in fig. 1, the specific method of step 2 is:
the input layer of the stacked sparse self-coding deep learning model for transformer fault diagnosis is set to be 5 neurons (bias neurons are not considered, the same is applied below), and the input layer is respectively expressed as x (x ═ x-1,x2,x3,x4,x5)TThe elements thereof each represent H2、CH4、C2H6、C2H4、C2H2Volume fraction (× 10)-6)。
Step 3, setting an output layer of the stack type sparse self-coding deep learning model;
the common faults of the transformer mainly include low-temperature overheating, medium-temperature overheating, high-temperature overheating, low-energy discharge, high-energy discharge, low-energy discharge and overheating, high-energy discharge and overheating, partial discharge and the like, and the fault state of the transformer can be coded as shown in table 1.
TABLE 1 Transformer Fault status Label and status coding
Figure BDA0002518242150000081
Therefore, as shown in fig. 1, the specific method of step 3 is:
the Softmax layer is set as an output layer of the failure diagnosis model, and the layer contains 9 neurons, which are respectively expressed as p ═ p (p)1,p2,p3,p4,p5,p6,p7,p8,p9)TThe elements thereof represent the probability of diagnosing as respective states.
Step 4, pre-training network parameters of a stack type sparse self-coding deep learning model for transformer fault diagnosis;
the pre-training phase is an unsupervised learning process.
The specific steps of the step 4 comprise:
(1) for the training set { (x)i,yi)|i∈1,2,…,N，yi∈ 1,2, … K }, wherein x isiAnd yiRespectively representing the ith training sample feature vector and a state label, and totally having K states; after the state labels of the training set are hidden, the original training set is changed into a label-free data set { xi|i∈1,2,…,N}。
As shown in fig. 3, the model Softmax layer in fig. 1 is removed, a decoding layer is added, and based on unlabeled sample data, training of network parameters of each layer of the SSAE deep learning model is sequentially achieved by a layer-by-layer greedy method according to a loss function (SAE cost function) shown in formula (3).
Step 5, fine tuning network parameters of a stack type sparse self-coding deep learning model for transformer fault diagnosis;
the fine tuning stage is a supervised learning process.
As shown in fig. 4, the specific steps of step 5 include:
(2) in the stage, the traditional output layer (namely decoding layer) of the SSAE deep learning model is removed, a Softmax classification layer is added, and optimization of network parameters of each layer of the SSAE is achieved based on the cross entropy loss function gradient value and the BP algorithm.
For training sample (x)i,yi) After SSAE deep learning model, the cross entropy loss value CiCan be expressed as:
Figure BDA0002518242150000091
wherein 1{ k ═ yiIs an indicative function, if k is yiIf k is not equal to y, its value is 1iA value of 0;
Figure BDA0002518242150000092
the probability of the state belonging to the kth class is judged for the kth neuron value of the Softmax layer, namely the SSAE deep learning model.
As can be seen from equation (4), for the training sample (x)i,yi) The deep learning model correctly determines that the state belongs to yiThe larger the probability value of (3), the lower the cross entropy loss value, and the better the effect is considered at this moment. But this approach does not take into account the effect of the probability of misjudging as other states. The ideal deep learning model ensures that the probability of correct judgment is higher, and the probability of misjudgment into other states is lower, so that the discrimination of positive judgment and misjudgment is increased, and the deep learning fault diagnosis effect is improved. In addition, in the actual operation and monitoring process of the transformer, complete samples are difficult to obtain, more training samples (large samples) in normal state are highlighted, and fewer training samples (small samples) in fault state or certain specific type of fault state are highlighted, so that the number of the training samples in various fault states is unbalanced. The deep learning model needs to realize feature extraction based on training samples, and if the number of small samples is too small, the feature extraction capability of deep learning on the state sample is reduced, so that the fault diagnosis effect is influenced. The ideal deep learning transformer fault diagnosis method aims at unbalanced distribution of training samples, increases learning strength of small samples and improves fault diagnosis effect.
(2) In order to improve the discrimination of positive judgment and false judgment of the deep learning model, the invention provides a weighted comprehensive loss function. For sample (x)i,yi) Weighted integral loss function L thereofiComprises the following steps:
Figure BDA0002518242150000101
in the formula, RiCorrectly determines the state y for the sampleiIs the cross entropy loss value, Ri＝Ci；
Figure BDA0002518242150000102
To determine the sample as the state k by mistake (k ≠ y)i) Loss values, and have:
Figure BDA0002518242150000103
as can be seen from the expressions (4) to (6), R is higher as the sample judgment accuracy is higheriThe smaller the value; the lower the sample decision is to be an error class probability, ZiThe lower the value. The comprehensive loss function can enable the probability of correct judgment of the sample to be higher and the probability of misjudgment to be lower, so that the discrimination of the correct judgment and the misjudgment is increased, and the accuracy and the stability of fault diagnosis are improved.
(3) In order to improve the training intensity of the small sample, the invention provides loss weight. Sample (x)i,yi) Loss weight value when determined to be state k
Figure BDA0002518242150000104
Comprises the following steps:
Figure BDA0002518242150000105
from equation (7), the weight is lost
Figure BDA0002518242150000106
Sample correct decision state yiThe higher the probability of (c), the loss weight value
Figure BDA0002518242150000107
The lower. Misjudgment as the state k (k ≠ y)i) The lower the probability of (D), the loss weight value
Figure BDA0002518242150000111
The lower. When the training samples are not balanced, under the deep learning model, the small samples are easier to misjudge, and the loss weight value is larger.
(4) Combining the loss weight with the composite loss function to form a weighted composite loss function, as shown in equation (8):
Figure BDA0002518242150000112
the weighted comprehensive loss function increases the discrimination of positive judgment and false judgment of the sample, and simultaneously improves the proportion of the small sample in the loss function, thereby enhancing the learning strength of the small sample and improving the fault diagnosis effect.
And 6, inputting the sample to be diagnosed into the fine-tuned stack type sparse self-coding deep learning model for transformer fault diagnosis, and outputting a transformer fault diagnosis result.
As shown in FIG. 5, according to the transformer fault diagnosis method based on the weighted integrated loss optimization deep learning and DGA, after the transformer fault diagnosis model is established, according to the sample to be diagnosed, the H value is used as H2、CH4、C2H6、C2H4、C2H2The volume fraction of the transformer is used as input, and the transformer fault diagnosis can be realized according to the output value of the Softmax layer.
Example 2:
a transformer fault diagnosis device based on deep learning and DGA (differential global analysis), comprising:
the transformer fault diagnosis model building module is used for building a stack type sparse self-coding deep learning model for transformer fault diagnosis;
the input layer setting module is used for setting an input layer of the stack type sparse self-coding deep learning model;
the output layer setting module is used for setting an output layer of the stacked sparse self-coding deep learning model;
the parameter pre-training module is used for pre-training network parameters of a stack type sparse self-coding deep learning model for transformer fault diagnosis;
the parameter fine-tuning module is used for fine-tuning network parameters of a stacked sparse self-coding deep learning model for transformer fault diagnosis;
and the transformer fault diagnosis result output module outputs a transformer fault diagnosis result according to the sample to be diagnosed.
The transformer fault diagnosis model building module comprises:
an encoding and decoding module for setting N neurons in input and output layers and M neurons in hidden layer, and setting input x ∈ RNWhich can be expressed as h ∈ R after being encoded and decoded respectivelyMAnd
Figure BDA0002518242150000113
and the encoding and decoding processes can be represented by equation (1) and equation (2), respectively, that is:
h＝f(Wx+b) (1)
Figure BDA0002518242150000121
in the formula, W, W' is encoding and decoding weight matrix respectively; b. b' are respectively encoding and decoding offset vectors; f (-) and g (-) are nonlinear activation functions in the encoding and decoding process, and a Sigmoid function can be generally adopted;
the sparse self-encoder forming module is used for realizing the adjustment of network parameters
Figure BDA0002518242150000122
When x reconstruction error is minimum, h is used as the intrinsic characteristic of the original input, and sparsity limitation is performed on the hidden layer, so that a sparse auto-Encoder (SAE) is formed, and the SAE cost function is as follows:
Figure BDA0002518242150000123
wherein β is a sparse penalty factor coefficient, which can be generally set to 0.3, KL(. h) is KL divergence; rho0Is a sparsity parameter;
Figure BDA0002518242150000124
is the mean activation of neurons in the hidden layer, and has
Figure BDA0002518242150000125
And the multi-level SAE stacking module is used for stacking the multi-level SAEs according to a stacked structure, and each SAE input layer is an upper SAE output layer, so that a stacked sparse self-coding deep learning model for transformer fault diagnosis is constructed.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
It should be emphasized that the embodiments described herein are illustrative rather than restrictive, and thus the present invention is not limited to the embodiments described in the detailed description, but also includes other embodiments that can be derived from the technical solutions of the present invention by those skilled in the art.

Claims (10)

1. A transformer fault diagnosis method based on deep learning and DGA is characterized in that: the method comprises the following steps:
gas H characterized by DGA2、CH4、C2H6、C2H4、C2H2The volume fraction of (a) is a diagnostic sample;
inputting the sample to be diagnosed into the stack type sparse self-coding deep learning model for transformer fault diagnosis after fine adjustment, and outputting a transformer fault diagnosis result.
2. The transformer fault diagnosis method based on deep learning and DGA as claimed in claim 1, wherein: before inputting a sample to be diagnosed into the trimmed stack type sparse self-coding deep learning model for transformer fault diagnosis, the method further comprises the following steps:
constructing a stack type sparse self-coding deep learning model for transformer fault diagnosis;
setting an input layer of a stack type sparse self-coding deep learning model;
setting an output layer of a stack type sparse self-coding deep learning model;
pre-training network parameters of a stack type sparse self-coding deep learning model for transformer fault diagnosis;
and fine-tuning parameters of a stacked sparse self-coding deep learning model for transformer fault diagnosis.
3. The transformer fault diagnosis method based on deep learning and DGA as claimed in claim 2, wherein: the specific steps of constructing the stacked sparse self-coding deep learning model for transformer fault diagnosis comprise:
(1) let input and output layers have N number of neurons, hidden layer has M number of neurons, and let input x ∈ RNWhich can be expressed as h ∈ R after being encoded and decoded respectivelyMAnd
Figure FDA0002518242140000011
and the encoding and decoding processes can be represented by equation (1) and equation (2), respectively, that is:
h＝f(Wx+b) (1)
Figure FDA0002518242140000012
in the formula, W, W' is encoding and decoding weight matrix respectively; b. b' are respectively encoding and decoding offset vectors; f (-) and g (-) are nonlinear activation functions in the encoding and decoding process, and a Sigmoid function can be generally adopted;
(2) by adjusting network parameters, realize
Figure FDA0002518242140000021
When x reconstruction error is minimum, h is used as the intrinsic characteristic of the original input, and sparsity limitation is performed on the hidden layer, so that a Sparse Auto-Encoder (SAE) is formed, and the SAE cost function is as follows:
Figure FDA0002518242140000022
wherein β is a sparse penalty factor coefficient, which can be generally set to 0.3, KL(. h) is KL divergence; rho0Is a sparsity parameter;
Figure FDA0002518242140000023
is the mean activation of neurons in the hidden layer, and has
Figure FDA0002518242140000024
(3) And stacking the multiple levels of SAEs according to a stacked structure, wherein all SAE input layers are upper SAE output layers, and further constructing a stacked sparse self-coding deep learning model for transformer fault diagnosis.
4. The transformer fault diagnosis method based on deep learning and DGA as claimed in claim 2, wherein: the specific method for setting the input layer of the stacked sparse self-coding deep learning model comprises the following steps:
the input layer of the stacked sparse self-coding deep learning model for transformer fault diagnosis is set to be 5 neurons, and is respectively expressed as x ═ x (x)1,x2,x3,x4,x5)TThe elements thereof each represent hydrogen H2Methane CH4Ethane C2H6Ethylene C2H4Acetylene C2H2Volume fraction of (a).
5. The transformer fault diagnosis method based on deep learning and DGA as claimed in claim 2, wherein: the specific method for setting the output layer of the stacked sparse self-coding deep learning model comprises the following steps:
the Softmax layer is set as an output layer of the failure diagnosis model, and the layer contains 9 neurons, which are respectively expressed as p ═ p (p)1,p2,p3,p4,p5,p6,p7,p8,p9)TThe elements thereof represent the probability of diagnosing as respective states.
6. The transformer fault diagnosis method based on deep learning and DGA as claimed in claim 2, wherein: the specific steps of pre-training the network parameters of the stacked sparse self-coding deep learning model for transformer fault diagnosis comprise:
(1) for the training set { (x)i,yi)|i∈1,2,…,N，yi∈ 1,2, … K }, wherein x isiAnd yiAre respectively provided withThe feature vector and the state label of the ith training sample are provided, and K states are provided in total; after the state labels of the training set are hidden, the original training set is changed into a label-free data set { xi|i∈1,2,…,N}；
(2) And removing the Softmax layer, adding a decoding layer, and sequentially realizing the training of network parameters of each layer of the SSAE deep learning model by adopting a layer-by-layer greedy method according to an SAE cost function based on label-free sample data.
7. The transformer fault diagnosis method based on deep learning and DGA as claimed in claim 2, wherein: the specific steps of finely adjusting the parameters of the stacked sparse self-coding deep learning model for transformer fault diagnosis comprise:
(1) removing a traditional output layer of the SSAE deep learning model, adding a Softmax classification layer, and realizing optimization of network parameters of each layer of the SSAE based on a cross entropy loss function gradient value and a BP algorithm;
(2) for training sample (x)i,yi) After SSAE deep learning model, the cross entropy loss value CiCan be expressed as:
Figure FDA0002518242140000031
wherein 1{ k ═ yiIs an indicative function, if k is yiIf k is not equal to y, its value is 1iA value of 0;
Figure FDA0002518242140000032
the probability of the state belonging to the kth class is judged for the kth neuron value of the Softmax layer, namely the SSAE deep learning model;
(3) for sample (x)i,yi) Of the combined loss function LiComprises the following steps:
Figure FDA0002518242140000033
in the formula (I), the compound is shown in the specification,
Figure FDA0002518242140000034
to determine the sample as the state k by mistake (k ≠ y)i) Loss values, and have:
Figure FDA0002518242140000035
(4) sample (x)i,yi) Loss weight value when determined to be state k
Figure FDA0002518242140000036
Comprises the following steps:
Figure FDA0002518242140000037
(5) combining the loss weight with the composite loss function to form a weighted composite loss function, as shown in equation (8):
Figure FDA0002518242140000038
(6) adopting a supervised learning mechanism, adding a sample fault type, firstly taking SSAE network parameters obtained by pre-training as initial parameters, and then taking a weighted loss function value between each training sample prediction fault and a real fault thereof as a network convergence cost; and then summing the weighted loss function values of the training samples to obtain the total loss, and performing optimization fine adjustment on the network parameters of the SSAE diagnostic model layer by using a BP algorithm gradient descent method with the aim of minimum total loss.
8. A transformer fault diagnosis device based on deep learning and DGA is characterized in that: the method comprises the following steps:
the diagnostic sample input module takes the volume fractions of DGA characteristic gases H2, CH4, C2H6, C2H4 and C2H2 as diagnostic samples;
and the to-be-detected transformer fault diagnosis result output module inputs the to-be-diagnosed sample into the fine-tuned stack type sparse self-coding deep learning model for transformer fault diagnosis and outputs a transformer fault diagnosis result.
9. A transformer fault diagnosis device based on deep learning and DGA is characterized in that: further comprising:
the transformer fault diagnosis model building module is used for building a stack type sparse self-coding deep learning model for transformer fault diagnosis;
the input layer setting module is used for setting an input layer of the stack type sparse self-coding deep learning model;
the output layer setting module is used for setting an output layer of the stacked sparse self-coding deep learning model;
the parameter pre-training module is used for pre-training network parameters of a stack type sparse self-coding deep learning model for transformer fault diagnosis;
and the parameter fine-tuning module is used for fine-tuning network parameters of the stacked sparse self-coding deep learning model for transformer fault diagnosis.
10. The deep learning and DGA-based transformer fault diagnosis apparatus according to claim 9, wherein: the transformer fault diagnosis model building module comprises:
an encoding and decoding module for setting N neurons in input and output layers and M neurons in hidden layer, and setting input x ∈ RNWhich can be expressed as h ∈ R after being encoded and decoded respectivelyMAnd
Figure FDA0002518242140000041
and the encoding and decoding processes can be represented by equation (1) and equation (2), respectively, that is:
h＝f(Wx+b) (1)
Figure FDA0002518242140000042
in the formula, W, W' is encoding and decoding weight matrix respectively; b. b' are respectively encoding and decoding offset vectors; f (-) and g (-) are nonlinear activation functions in the encoding and decoding process, and a Sigmoid function can be generally adopted;
the sparse self-encoder forming module is used for realizing the adjustment of network parameters
Figure FDA0002518242140000043
When x reconstruction error is minimum, h is used as the intrinsic characteristic of the original input, and sparsity limitation is performed on the hidden layer, so that a Sparse Auto-Encoder (SAE) is formed, and the SAE cost function is as follows:
Figure FDA0002518242140000051
wherein β is a sparse penalty factor coefficient, which can be generally set to 0.3, KL(. h) is KL divergence; rho0Is a sparsity parameter;
Figure FDA0002518242140000052
is the mean activation of neurons in the hidden layer, and has
Figure FDA0002518242140000053
And the multi-level SAE stacking module is used for stacking the multi-level SAEs according to a stacked structure, and each SAE input layer is an upper SAE output layer, so that a stacked sparse self-coding deep learning model for transformer fault diagnosis is constructed.
CN202010483448.9A 2020-06-01 2020-06-01 Transformer fault diagnosis method and device based on deep learning and DGA Pending CN111737907A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010483448.9A CN111737907A (en) 2020-06-01 2020-06-01 Transformer fault diagnosis method and device based on deep learning and DGA

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010483448.9A CN111737907A (en) 2020-06-01 2020-06-01 Transformer fault diagnosis method and device based on deep learning and DGA

Publications (1)

Publication Number Publication Date
CN111737907A true CN111737907A (en) 2020-10-02

Family

ID=72646563

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010483448.9A Pending CN111737907A (en) 2020-06-01 2020-06-01 Transformer fault diagnosis method and device based on deep learning and DGA

Country Status (1)

Country Link
CN (1) CN111737907A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112365009A (en) * 2020-10-28 2021-02-12 国网山东省电力公司电力科学研究院 Secondary equipment abnormity diagnosis method based on deep learning network
CN114416972A (en) * 2021-12-10 2022-04-29 厦门市世纪网通网络服务有限公司 DGA domain name detection method based on density improvement unbalance sample
CN114545294A (en) * 2022-01-14 2022-05-27 国电南瑞科技股份有限公司 Transformer fault diagnosis method and system, storage medium and computing device
CN117571901A (en) * 2023-11-17 2024-02-20 承德神源太阳能发电有限公司 Method, system and equipment for early warning and overhauling faults of photovoltaic power station transformer
CN117571901B (en) * 2023-11-17 2024-06-11 承德神源太阳能发电有限公司 Method, system and equipment for early warning and overhauling faults of photovoltaic power station transformer

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106323636A (en) * 2016-08-16 2017-01-11 重庆交通大学 Adaptive extraction and diagnosis method for degree features of mechanical fault through stack-type sparse automatic coding depth neural network

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106323636A (en) * 2016-08-16 2017-01-11 重庆交通大学 Adaptive extraction and diagnosis method for degree features of mechanical fault through stack-type sparse automatic coding depth neural network

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
王伟;唐庆华;刘力卿;李敏;谢军;: "基于加权综合损失优化深度学习和DGA的变压器故障诊断方法", 南方电网技术, no. 03, pages 29 - 34 *
许倩文;吉兴全;张玉振;李军;于永进;: "基于栈式稀疏自编码器的矿用变压器故障诊断", 工矿自动化, no. 10 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112365009A (en) * 2020-10-28 2021-02-12 国网山东省电力公司电力科学研究院 Secondary equipment abnormity diagnosis method based on deep learning network
CN114416972A (en) * 2021-12-10 2022-04-29 厦门市世纪网通网络服务有限公司 DGA domain name detection method based on density improvement unbalance sample
CN114545294A (en) * 2022-01-14 2022-05-27 国电南瑞科技股份有限公司 Transformer fault diagnosis method and system, storage medium and computing device
CN114545294B (en) * 2022-01-14 2023-06-16 国电南瑞科技股份有限公司 Transformer fault diagnosis method, system, storage medium and computing device
CN117571901A (en) * 2023-11-17 2024-02-20 承德神源太阳能发电有限公司 Method, system and equipment for early warning and overhauling faults of photovoltaic power station transformer
CN117571901B (en) * 2023-11-17 2024-06-11 承德神源太阳能发电有限公司 Method, system and equipment for early warning and overhauling faults of photovoltaic power station transformer

Similar Documents

Publication Publication Date Title
CN112070128B (en) Transformer fault diagnosis method based on deep learning
Li et al. Interpretation of DGA for transformer fault diagnosis with complementary SaE-ELM and arctangent transform
Dai et al. Dissolved gas analysis of insulating oil for power transformer fault diagnosis with deep belief network
CN111737907A (en) Transformer fault diagnosis method and device based on deep learning and DGA
CN110175386B (en) Method for predicting temperature of electrical equipment of transformer substation
CN109711483B (en) Spark Autoencoder-based power system operation mode clustering method
CN110879373B (en) Oil-immersed transformer fault diagnosis method with neural network and decision fusion
CN108170994A (en) A kind of oil-immersed electric reactor method for diagnosing faults based on two-way depth network
CN110689068B (en) Transformer fault type diagnosis method based on semi-supervised SVM
CN108717149A (en) Diagnosis Method of Transformer Faults based on M-RVM fusion dynamic weightings AdaBoost
CN115563563A (en) Fault diagnosis method and device based on transformer oil chromatographic analysis
CN110689069A (en) Transformer fault type diagnosis method based on semi-supervised BP network
CN113240011A (en) Deep learning driven abnormity identification and repair method and intelligent system
CN112183591A (en) Transformer fault diagnosis method based on stack sparse noise reduction automatic coding network
Yin et al. Deep learning based feature reduction for power system transient stability assessment
CN112731137A (en) Cage type asynchronous motor stator and rotor fault joint diagnosis method based on stack type self-coding and light gradient elevator algorithm
CN111539486A (en) Transformer fault diagnosis method based on Dropout deep confidence network
CN116562114A (en) Power transformer fault diagnosis method based on graph convolution neural network
CN115018512A (en) Electricity stealing detection method and device based on Transformer neural network
CN112633315A (en) Electric power system disturbance classification method
Li et al. Transformer fault diagnosis based on improved deep coupled dense convolutional neural network
CN116562121A (en) XGBoost and FocalLoss combined cable aging state assessment method
CN115712871A (en) Power electronic system fault diagnosis method combining resampling and integrated learning
CN111275204A (en) Transformer state identification method based on hybrid sampling and ensemble learning
CN111831955A (en) Lithium ion battery residual life prediction method and system

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