CN109086817A - A kind of Fault Diagnosis for HV Circuit Breakers method based on deepness belief network - Google Patents

A kind of Fault Diagnosis for HV Circuit Breakers method based on deepness belief network Download PDF

Info

Publication number
CN109086817A
CN109086817A CN201810827990.4A CN201810827990A CN109086817A CN 109086817 A CN109086817 A CN 109086817A CN 201810827990 A CN201810827990 A CN 201810827990A CN 109086817 A CN109086817 A CN 109086817A
Authority
CN
China
Prior art keywords
layer
training
model
hidden
fault diagnosis
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.)
Granted
Application number
CN201810827990.4A
Other languages
Chinese (zh)
Other versions
CN109086817B (en
Inventor
黄新波
王宁
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.)
Xian Polytechnic University
Original Assignee
Xian Polytechnic University
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 Xian Polytechnic University filed Critical Xian Polytechnic University
Priority to CN201810827990.4A priority Critical patent/CN109086817B/en
Publication of CN109086817A publication Critical patent/CN109086817A/en
Application granted granted Critical
Publication of CN109086817B publication Critical patent/CN109086817B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/24323—Tree-organised classifiers
    • 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
    • G06N3/045—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
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

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

Abstract

The Fault Diagnosis for HV Circuit Breakers method based on deepness belief network that the invention discloses a kind of, the specific steps are step 1: the sample data after unified standardization is divided into test sample and training sample according to special ratios by data sample needed for choosing experiment.Step 2: building and initialize DBN deepness belief network fault diagnosis model.Step 3: by a large amount of unlabeled exemplars of pre-training concentration or exemplar being gone to input from model bottom, using successively unsupervised greedy study to RBM progress pre-training in model；Step 4: entire model being finely adjusted using genetic algorithm, step 5, the Fault Diagnosis for HV Circuit Breakers model that training is obtained carries out failure modes to the test set sample in step 1, obtains failure modes as a result, simultaneously statistical model accuracy rate of diagnosis.A kind of Fault Diagnosis for HV Circuit Breakers method based on deepness belief network of the invention, can be trained big data quantity sample, to realize Fault Diagnosis for HV Circuit Breakers function.

Description

A kind of Fault Diagnosis for HV Circuit Breakers method based on deepness belief network
Technical field
The invention belongs to Fault Diagnosis for HV Circuit Breakers method and technology fields, specifically propose a kind of based on deepness belief network Fault Diagnosis for HV Circuit Breakers method.
Background technique
High-voltage circuitbreaker plays dual parts of control and protection in distribution network system, therefore carries out event to high-voltage circuitbreaker It is great to hinder diagnostic significance.Currently a variety of diagnostic methods are proposed, but there are still some problems and certain limitation, such as expert System needs expertise knowledge abundant, and at this moment compares and be difficult to obtain；Neural network easily falls into local optimum；SVM is two There is the case where classification overlapping and unclassified in sorting algorithm, multi-classification algorithm, such as one-to-one SVM；Although and ELM training speed Degree is than very fast, but the stability of institute's training pattern diagnosis is less better.In addition, existing intelligent failure diagnosis method is mostly right The utilization rate of unlabeled exemplars is lower, and learning ability has limitation, and scalability is poor.And deep learning is as Three generations's neural network has the ability for extracting feature from a large amount of unlabeled exemplars strongly；Simultaneously because the complexity of its multilayer Structure and successively trained stability, can be trained big data quantity sample, comply with the trend of big data era, had Wide application prospect.
In consideration of it, the present invention proposes a kind of high pressure for being based on deepness belief network (deep belief network, DBN) Circuit breaker failure diagnosis algorithm is automatically extracted by 3 layers of RBM (restricted Boltzmann machine, RBM) defeated first Enter the correlation properties between variable, and model tuning is carried out using genetic algorithm, then carries out failure modes using 3 layers of BP network Prediction.This method it is special can to extract information under different variation tendencies using a large amount of unlabeled exemplars data training study of engineering site It seeks peace potential statistical law.
Summary of the invention
The Fault Diagnosis for HV Circuit Breakers method based on deepness belief network that the present invention provides a kind of, can be to big data Amount sample is trained, to realize Fault Diagnosis for HV Circuit Breakers function.
The technical scheme adopted by the invention is that a kind of Fault Diagnosis for HV Circuit Breakers side based on deepness belief network Method is specifically implemented according to the following steps:
Step 1, data sample needed for choosing experiment, the sample data after standardization is divided into according to special ratios Test sample and training sample；
Step 2, high-voltage circuitbreaker DBN deepness belief network fault diagnosis model is constructed,
Step 3, pre-training is carried out to RBM in model using successively unsupervised greedy learning algorithm (CD algorithm),
Step 4, entire model is finely adjusted using genetic algorithm, i.e. model parameter optimizing,
Step 5, by step 4 training obtain Fault Diagnosis for HV Circuit Breakers model to the test set sample in step 1 into Row failure modes obtain failure modes as a result, simultaneously statistical model accuracy rate of diagnosis.
The features of the present invention also characterized in that:
The step 1 specifically,
Step 1.1, the present invention is by most of SF6The I that breaker can all monitor1,I2,I3,t1,t2,t3,t4,t5, with SF6 Pressure, density, micro-water content, decomposition product content, usually SO2、H2Input of the S content as deepness belief network model.
Step 1.2, data collected in step 1.1 are normalized,
Step 1.3, the sample data after standardization is divided into pre-training collection, tuning collection and test according to special ratios Collection；Wherein, pre-training collection is a large amount of unlabeled exemplars composition, and tuning collection and test set are a small amount of sample structure for having class label At.
The specific step of step 2 are as follows:
Step 2.1, building first layer RBM is limited Boltzmann machine unit,
Step 2.2, first layer RBM1 is limited to the hidden layer h of Boltzmann machine unit1Glass is limited as second layer RBM2 The visible layer v of the graceful machine unit of Wurz2, second layer RBM2 is constructed,
Step 2.3, third layer RBM3 is built on the basis of the second layer in an identical manner and is limited Boltzmann machine unit, Three layers of RBM network model building are completed since then,
Step 2.4, the classification layer of DBN network model is constructed.The present invention adds three layers of BP nerve again at the top of RBM3 Network model is used for output category result.By RBM3 hidden layer h3Visible layer input terminal of the node as BP neural network unit,
Step 2.5, the Fault Diagnosis for HV Circuit Breakers model based on deepness belief network is constructed, output layer is by 5 nodes Unit composition, each unit represent a kind of fault category, and data input layer is v1, thus input collected SF6High-voltage circuitbreaker Sample data, mode input are 20 dimensions, i.e. m=20.
The step 2.1 specifically,
RBM network is divided into two layers: i.e. hidden layer h and visible layer v.Hidden layer is made of n implicit stochastic variables: h= (h1, h2, h3, hj..., hn)；Visible layer constitutes v=(v by m stochastic variable1, v2, v3, vi..., vm), for indicating observation Data, it is seen that layer is data input layer,
RBM can be considered that the model based on energy, energy function are defined as
Wherein viFor the input value of visible element；hjFor the output valve of hidden unit, value indicates that the unit is in and swashs when being 1 State living indicates to be in unactivated state when being 0；wijIndicate visible element viWith hidden unit hjBetween connection weight, aiIndicate visible Unit viBiasing, bjIndicate hidden unit hjBiasing, n is hidden layer node quantity, and m is visible layer number of nodes, ai,bj,wij It is real number, θ=(a, b, w) constitutes the model parameter of RBM.
The joint probability distribution P of (v, h)θ(v, h) can be indicated are as follows:
ZθFor normaliztion constant, it is called and does partition function.
For one layer of RBM m visible element for being included and n hidden unit, the condition of hidden unit and visible element is given Probability are as follows:
On the contrary, the conditional probability of hidden unit are as follows:
The activation probability of hidden unit can be derived according to formula (1.1) (1.2) (1.3) are as follows:
It is similar, it is seen that i-th of node of node layer activates probability are as follows:
σ () indicates activation primitive in formula.
It selects and corrects the activation primitive that linear unit function replaces the above traditional neural network, i.e.,
σ (z)=max (0, z) (1.8).
The step 3 specifically:
Step 3.1, unlabeled exemplars are inputted in training network, initializes visible layer location mode v1, imply layer unit Number n, learning rate ε and training maximum cycle T, connection weight w and bias vector a and b randomly select lesser numerical value；
Step 3.2, it will be seen that unit vector is mapped to hidden unit.All hidden units of hidden layer are calculated according to (1.7), from Condition is distributed P (hj|v1) extract hj∈{0,1}。
Step 3.3, to visible layer according to (1.6) formula, from P (vi|h1) in extract vi∈{0,1}；
Step 3.4, according to formula (1.7), all hidden units of hidden layer are calculated；
Step 3.5, weight and its each bias are updated:
W←W+ε[(P(h1=1 | v1)v1 T-P(h2=1 | v2)v2 T]
a←a+ε(v1-v2)
b←b+ε(P(h1=1 | v1)-P(h2=1 | v2))
It repeats step 3.2- step 3.5 and stops the layer when reaching the number of iterations maximum value or satisfactory reconstructed error The training of RBM.
In RBM1, it is seen that unit number m=20, hidden unit number are set in advance as n=50, and cycle of training, T=50, was learned Practise rate ε=0.02.
After step 3.6 has trained RBM1, the parameter of fixed RBM1, the state that RBM1 training is completed, i.e. h1Hidden layer State repeats above-mentioned training process as the input of next layer of RBM2, the training up to completing all RBM numbers of plies, then pre-training Terminate.
The step 4 specifically,
Step 4.1, after pre-training, whole network is equivalent to BP neural network, at this point, this BP neural network is first Beginning weight and biasing are provided in pre-training.GA parameter: Population Size M=100 is set, termination algebra of evolving is T=200, initial crossing-over rate are Pc0=0.7, initial aberration rate is Pm0=0.001；
Step 4.2, after step 3 pre-training, all parameter weights and threshold vector that model is generated, i.e. wij, ai,bj And BP classification layer parameter is generated as the chromosome vector of genetic algorithm, and in its restriction range using compiling method at random Initial population；
Step 4.3, the tuning collection sample obtained in step 1 is input to the input layer of BP neural network after normalizing, That is the visible layer of model RBM1 calculates fitness letter corresponding under haunting a population vector by BP neural network forwards algorithms Numerical value.Then, optimal population is searched for, keeps following adaptive value index E minimum:
In formula: N is number of training；yg dIt is the idea output of g group sample；ygIt is the reality output of g group sample Value, E are the adaptive value in M population.
Step 4.4, if having reached desired output, model training is completed, and all model parameters are updated to correspondence Database in.If failing to have reached desired output, connection weight existing in model and threshold value are extracted, with GA pairs Existing connection weight and threshold value are encoded；
Step 4.5, it carries out a GA operation and generates new group, then decoding generates new connection weight and threshold value to replace Value originally；
Step 4.6, step 4.2 step 4.5 is repeated, it, will most until output reaches expectation or the number of iterations reaches Substitution model parameter after excellent chromosome decoding, so far the Fault Diagnosis for HV Circuit Breakers model training based on deepness belief network is complete Finish.
The beneficial effects of the present invention are:
1. high-voltage circuitbreaker tape label fault sample is often rare, engineering site can be obtained largely by on-Line Monitor Device Unlabeled exemplars, DBN can make full use of these unlabeled exemplars to carry out fault signature study, substantially increase data sample Utilization efficiency.
2. carrying out tuning to entire DBN model using genetic algorithm, that is, it ensure that the higher judging nicety rate of model, and not Easily fall into local optimum.
3.DBN can realize to the analog case of any complicated function have by building multitiered network structural model Stronger learning ability is applied to Fault Diagnosis for HV Circuit Breakers, is expected to further increase performance of fault diagnosis.
Detailed description of the invention
Fig. 1 is the flow chart of the method for the present invention；
Fig. 2 is the structure chart of the RBM model in the method for the present invention；
DBN deepness belief network model structure in Fig. 3 the method for the present invention.
Specific embodiment
The following describes the present invention in detail with reference to the accompanying drawings and specific embodiments.
The present invention is a kind of Fault Diagnosis for HV Circuit Breakers method based on deepness belief network, as shown in Figure 1, specifically pressing Implement according to following steps:
Step 1, data sample needed for choosing experiment, the sample data after standardization is divided into according to special ratios Test sample and training sample, the specific steps are,
Step 1.1, the present invention is by most of SF6The I that breaker can all monitor1,I2,I3,t1,t2,t3,t4,t5(to divide, Current waveform when combined floodgate on coil extracts data), with SF6Pressure, density, micro-water content, decomposition product content (usually SO2、H2S content) input as deepness belief network model proposed in this paper.Breaking-closing operating brownout, iron core are run There are bite, operating mechanism to have bite, auxiliary switch movement poor contact, the normal permanent fault result of mechanism defeated as model Out.
Specific fault type and coding are as shown in table 1:
Table 1
Step 1.2, data collected in step 1.1 are normalized.
Step 1.3, the sample data after standardization is divided into pre-training collection, tuning collection and test according to special ratios Collection；Wherein, pre-training collection is a large amount of unlabeled exemplars composition, and tuning collection and test set are a small amount of sample structure for having class label At.
It to avoid sample set deflection, while can guarantee to obtain enough samples again, multiple engineering site records can be chosen Same model breaker online monitoring data, these data are no label data, may be used as pre-training sample.
Step 2, high-voltage circuitbreaker DBN deepness belief network fault diagnosis model, specific step are constructed are as follows:
Step 2.1, building first layer RBM is limited Boltzmann machine unit.
RBM is a kind of undirected graph model of bipartite structure, and structure is as shown in Figure 2.RBM network is divided into two layers: i.e. hidden layer h With visible layer v.Hidden layer is made of n implicit stochastic variables: h=(h1, h2, h3, hj..., hn)；Visible layer is random by m Variable constitutes v=(v1,v2,v3,vi,…,vm), for indicating the data of observation, it is seen that layer is data input layer.Network connection It only exists between two layers, it is connectionless between same layer built-in variable.
RBM can be considered that the model based on energy, energy function are defined as
Wherein viFor the input value of visible element；hjFor the output valve of hidden unit, value indicates that the unit is in and swashs when being 1 State living indicates to be in unactivated state when being 0；wijIndicate visible element viWith hidden unit hjBetween connection weight, aiIndicate visible Unit viBiasing, bjIndicate hidden unit hjBiasing, n is hidden layer node quantity, and m is visible layer number of nodes, ai,bj,wij It is real number, θ=(a, b, w) constitutes the model parameter of RBM.The joint probability distribution P of (v, h)θ(v, h) can be indicated are as follows:
ZθFor normaliztion constant, it is called and does partition function (partition function).
Since RBM is the connectionless two-value composition of interlayer, so its hidden unit and visible element are independent from each other unit, For one layer of RBM m visible element for being included and n hidden unit, the conditional probability of hidden unit and visible element is given are as follows:
On the contrary, the conditional probability of hidden unit are as follows:
The activation probability of hidden unit can be derived according to formula (1.1) (1.2) (1.3) are as follows:
It is similar, it is seen that i-th of node of node layer activates probability are as follows:
σ () indicates activation primitive in formula.Common activation primitive is sigmoid function and tanh function, and expression sees below formula.
There are the characteristics that derivative and saturation value scale for above-mentioned activation primitive, when carrying out the backpropagation of stepping type multilayer, Gradient error can constantly decay, so that learning efficiency reduces.In addition, by ReLU realize it is sparse after model can preferably dig Correlated characteristic is dug, training data is fitted.Therefore the linear unit function of selection amendment herein (rectified linear units, ReLU) function replaces the activation primitive of the above traditional neural network.
σ (z)=max (0, z) (1.8)
Step 2.2, first layer RBM1 is limited to the hidden layer h of Boltzmann machine unit1Glass is limited as second layer RBM2 The visible layer v of the graceful machine unit of Wurz2, as shown in Figure 2.Construct second layer RBM2.
Step 2.3, third layer RBM3 is built on the basis of the second layer in an identical manner and is limited Boltzmann machine unit, Three layers of RBM network model building are completed since then.
Step 2.4, the classification layer of DBN network model is constructed.The present invention adds three layers of BP nerve again at the top of RBM3 Network model is used for output category result.By RBM3 hidden layer h3Visible layer input terminal of the node as BP neural network unit.
Step 2.5, the Fault Diagnosis for HV Circuit Breakers model as shown in Figure 3 based on deepness belief network is constructed.Output Layer is made of 5 node units, and each unit represents a kind of fault category, and data input layer is v1, thus input collected SF6High-voltage circuitbreaker sample data, mode input are 20 dimensions (i.e. m=20).
Step 3, pre-training is carried out to RBM in model using successively unsupervised greedy learning algorithm (CD algorithm).
Step 3.1, unlabeled exemplars are inputted in training network, initializes visible layer location mode v1, imply layer unit Number n, learning rate ε and training maximum cycle T, connection weight w and bias vector a and b randomly select lesser numerical value.
Step 3.2, it will be seen that unit vector is mapped to hidden unit.All hidden units of hidden layer are calculated according to (1.7), from Condition is distributed P (hj|v1) extract hj∈{0,1}。
Step 3.3, to visible layer according to (1.6) formula, from P (vi|h1) in extract vi∈{0,1}；
Step 3.4, according to formula (1.7), all hidden units of hidden layer are calculated；
Step 3.5, weight and its each bias are updated:
W←W+ε[(P(h1=1 | v1)v1 T-P(h2=1 | v2)v2 T]
a←a+ε(v1-v2)
b←b+ε(P(h1=1 | v1)-P(h2=1 | v2))
It repeats step 3.2- step 3.5 and stops the layer when reaching the number of iterations maximum value or satisfactory reconstructed error The training of RBM.
Because of the presence (formula 1.3) of normalization factor Z, joint probability distribution P solves complex.Therefore the present invention proposes Using to sdpecific dispersion (contrastive divergence, CD) algorithm, in the training process, aobvious first DUAL PROBLEMS OF VECTOR MAPPING is arrived first Then hidden member shows first vector with hidden member reconstruct, then aobvious first DUAL PROBLEMS OF VECTOR MAPPING is repeated above step T times to hidden member, final real Now the quick training of RBM is learnt.In RBM1, it is seen that unit number m=20, hidden unit number are set in advance as n=50, instruction Practice cycle T=50, learning rate ε=0.02.
After step 3.6 has trained RBM1, the parameter of fixed RBM1, state (the i.e. h that RBM1 training is completed1Hidden layer State) input as next layer of RBM2, above-mentioned training process is repeated, until complete the training of all RBM numbers of plies, then pre-training Terminate.Learning process of every layer of RBM is mutually indepedent during this, enormously simplifies the training process of model.
Step 4, entire model is finely adjusted using genetic algorithm, i.e. model parameter optimizing.
Step 4.1, after pre-training, whole network is equivalent to BP neural network, at this point, this BP neural network is first Beginning weight and biasing are provided in pre-training.GA parameter: Population Size M=100 is set, termination algebra of evolving is T=200, initial crossing-over rate are Pc0=0.7, initial aberration rate is Pm0=0.001.
Step 4.2, after step 3 pre-training, all parameter weights and threshold vector (i.e. w that model is generatedij, ai,bj And BP classifies layer parameter) chromosome vector as genetic algorithm, and in its restriction range, it is generated at random using compiling method Initial population.
Step 4.3, the tuning collection sample obtained in step 1 is input to the input layer of BP neural network after normalizing, That is the visible layer of model RBM1 calculates fitness letter corresponding under haunting a population vector by BP neural network forwards algorithms Numerical value.Then, optimal population is searched for, keeps following adaptive value index E minimum:
In formula: N is number of training；yg dIt is the idea output of g group sample；ygIt is the reality output of g group sample Value, E are the adaptive value in M population.
Step 4.4, if having reached desired output, model training is completed, and all model parameters are updated to correspondence Database in.If failing to have reached desired output, connection weight existing in model and threshold value are extracted, with GA pairs Existing connection weight and threshold value are encoded；
Step 4.5, it carries out a GA operation and generates new group, then decoding generates new connection weight and threshold value to replace Value originally；
Step 4.6, step 4.2 step 4.5 is repeated, it, will most until output reaches expectation or the number of iterations reaches Substitution model parameter after excellent chromosome decoding, so far the Fault Diagnosis for HV Circuit Breakers model training based on deepness belief network is complete Finish.
Step 5, by step 4 training obtain Fault Diagnosis for HV Circuit Breakers model to the test set sample in step 1 into Row failure modes obtain failure modes as a result, simultaneously statistical model accuracy rate of diagnosis.
It is discovered by experiment that the network structure number of plies and cycle of training and predictablity rate have certain relationship.When RBM is increased to At 3 layers, prediction accuracy is substantially improved, and increases to 7 layers by 3 layers, accuracy rate promotes low effort.When network structure determines, As cycle of training increases, predictablity rate is in rising trend, and trend gradually slows down.Collective model prediction effect and operation effect Rate determines that RBM network structure is 3 layers, cycle of training 250 herein.
Show that the algorithm can make full use of no label data by test result, using genetic algorithm, it is easier to find Globally optimal solution can effectively promote the accuracy rate of failure modes.

Claims (6)

1. a kind of Fault Diagnosis for HV Circuit Breakers method based on deepness belief network, which is characterized in that specifically according to following step It is rapid to implement:
Step 1, data sample needed for choosing experiment, is divided into test according to special ratios for the sample data after standardization Sample and training sample；
Step 2, high-voltage circuitbreaker DBN deepness belief network fault diagnosis model is constructed,
Step 3, pre-training is carried out to RBM in model using successively unsupervised greedy learning algorithm (CD algorithm),
Step 4, entire model is finely adjusted using genetic algorithm, i.e. model parameter optimizing,
Step 5, the Fault Diagnosis for HV Circuit Breakers model that step 4 training obtains is subjected to event to the test set sample in step 1 Barrier classification obtains failure modes as a result, simultaneously statistical model accuracy rate of diagnosis.
2. a kind of Fault Diagnosis for HV Circuit Breakers method based on deepness belief network according to claim 1, feature Be, the step 1 specifically,
Step 1.1, the present invention is by most of SF6The I that breaker can all monitor1,I2,I3,t1,t2,t3,t4,t5, with SF6Pressure Power, density, micro-water content, decomposition product content, usually SO2、H2Input of the S content as deepness belief network model,
Step 1.2, data collected in step 1.1 are normalized,
Step 1.3, the sample data after standardization is divided into pre-training collection, tuning collection and test set according to special ratios；Its In, pre-training collection is a large amount of unlabeled exemplars composition, and tuning collection and test set are that a small amount of sample with class label is constituted.
3. a kind of Fault Diagnosis for HV Circuit Breakers method based on deepness belief network according to claim 1, feature It is, the specific step of the step 2 are as follows:
Step 2.1, building first layer RBM is limited Boltzmann machine unit,
Step 2.2, first layer RBM1 is limited to the hidden layer h of Boltzmann machine unit1Boltzmann is limited as second layer RBM2 The visible layer v of machine unit2, second layer RBM2 is constructed,
Step 2.3, third layer RBM3 is built on the basis of the second layer in an identical manner and is limited Boltzmann machine unit, since then Three layers of RBM network model building are completed,
Step 2.4, the classification layer of DBN network model is constructed, the present invention adds three layers of BP neural network again at the top of RBM3 Model is used for output category result, by RBM3 hidden layer h3Visible layer input terminal of the node as BP neural network unit,
Step 2.5, the Fault Diagnosis for HV Circuit Breakers model based on deepness belief network is constructed, output layer is by 5 node units Composition, each unit represent a kind of fault category, and data input layer is v1, thus input collected SF6High-voltage circuitbreaker sample Data, mode input are 20 dimensions, i.e. m=20.
4. a kind of Fault Diagnosis for HV Circuit Breakers method based on deepness belief network according to claim 3, feature Be, the step 2.1 specifically,
RBM network is divided into two layers: i.e. hidden layer h and visible layer v, and hidden layer is made of n implicit stochastic variables: h (h1, h2, h3, hj..., hn)；Visible layer constitutes v=(v by m stochastic variable1, v2,v3, vi..., vm) for indicating the data of observation, it can See that layer is data input layer,
RBM can be considered that the model based on energy, energy function are defined as
Wherein viFor the input value of visible element；hjFor the output valve of hidden unit, value indicates that the unit is in activated state when being 1, It indicates to be in unactivated state when being 0；wijIndicate visible element viWith hidden unit hjBetween connection weight, aiIndicate visible element viBiasing, bjIndicate hidden unit hjBiasing, n is hidden layer node quantity, and m is visible layer number of nodes, ai,bj,wijIt is Real number, θ=(a, b, w) constitute the model parameter of RBM,
The joint probability distribution P of (v, h)θ(v, h) can be indicated are as follows:
ZθFor normaliztion constant, it is called and does partition function,
For one layer of RBM m visible element for being included and n hidden unit, the conditional probability of hidden unit and visible element is given Are as follows:
On the contrary, the conditional probability of hidden unit are as follows:
The activation probability of hidden unit can be derived according to formula (1.1) (1.2) (1.3) are as follows:
It is similar, it is seen that i-th of node of node layer activates probability are as follows:
σ () indicates activation primitive in formula,
It selects and corrects the activation primitive that linear unit function replaces the above traditional neural network, i.e.,
σ (z)=max (0, z) (1.8).
5. a kind of Fault Diagnosis for HV Circuit Breakers method based on deepness belief network according to claim 1, feature Be, the step 3 specifically,
Step 3.1, unlabeled exemplars are inputted in training network, initializes visible layer location mode v1, hidden layer unit number n, Rate ε and training maximum cycle T is practised, connection weight w and bias vector a and b randomly select lesser numerical value；
Step 3.2, it will be seen that unit vector is mapped to hidden unit, is calculated according to (1.7) all hidden units of hidden layer, from condition It is distributed P (hj|v1) extract hj∈ { 0,1 },
Step 3.3, to visible layer according to (1.6) formula, from P (vi|h1) in extract vi∈{0,1}；
Step 3.4, according to formula (1.7), all hidden units of hidden layer are calculated；
Step 3.5, weight and its each bias are updated:
W←W+ε[(P(h1=1 | v1)v1 T-P(h2=1 | v2)v2 T]
a←a+ε(v1-v2)
b←b+ε(P(h1=1 | v1)-P(h2=1 | v2))
It repeats step 3.2- step 3.5 and stops this layer of RBM when reaching the number of iterations maximum value or satisfactory reconstructed error Training,
In RBM1, it is seen that unit number m=20, hidden unit number are set in advance as n=50, cycle of training T=50, study speed Rate ε=0.02,
After step 3.6 has trained RBM1, the parameter of fixed RBM1, the state that RBM1 training is completed, i.e. h1The state of hidden layer, As the input of next layer of RBM2, above-mentioned training process is repeated, until completing the training of all RBM numbers of plies, then pre-training terminates.
6. a kind of Fault Diagnosis for HV Circuit Breakers method based on deepness belief network according to claim 1, feature Be, the step 4 specifically,
Step 4.1, after pre-training, whole network is equivalent to BP neural network, at this point, the initial power of this BP neural network Value and biasing are provided in pre-training, set GA parameter: Population Size M=100, and evolving and terminating algebra is T= 200, initial crossing-over rate is Pc0=0.7, initial aberration rate is Pm0=0.001；
Step 4.2, after step 3 pre-training, all parameter weights and threshold vector that model is generated, i.e. wij, ai,bjAnd BP classification layer parameter is generated initial at random as the chromosome vector of genetic algorithm, and in its restriction range using compiling method Population；
Step 4.3, the tuning collection sample obtained in step 1 is input to the input layer of BP neural network, i.e. mould after normalizing The visible layer of type RBM1 calculates fitness function corresponding under haunting a population vector by BP neural network forwards algorithms Value, then, searches for optimal population, keeps following adaptive value index E minimum:
In formula: N is number of training；yg dIt is the idea output of g group sample；ygIt is the real output value of g group sample, E For the adaptive value in M population,
Step 4.4, if having reached desired output, model training is completed, and all model parameters are updated to corresponding number According in library, if failing to have reached desired output, connection weight existing in model and threshold value are extracted, with GA to existing Connection weight and threshold value encoded；
Step 4.5, it carries out a GA operation and generates new group, then it is original to replace to generate new connection weight and threshold value for decoding Value；
Step 4.6, step 4.2 step 4.5 is repeated, until output reaches expectation or the number of iterations reaches, by optimal dye Substitution model parameter after colour solid decoding, so far the Fault Diagnosis for HV Circuit Breakers model training based on deepness belief network finishes.
CN201810827990.4A 2018-07-25 2018-07-25 High-voltage circuit breaker fault diagnosis method based on deep belief network Active CN109086817B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810827990.4A CN109086817B (en) 2018-07-25 2018-07-25 High-voltage circuit breaker fault diagnosis method based on deep belief network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810827990.4A CN109086817B (en) 2018-07-25 2018-07-25 High-voltage circuit breaker fault diagnosis method based on deep belief network

Publications (2)

Publication Number Publication Date
CN109086817A true CN109086817A (en) 2018-12-25
CN109086817B CN109086817B (en) 2022-04-26

Family

ID=64838620

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810827990.4A Active CN109086817B (en) 2018-07-25 2018-07-25 High-voltage circuit breaker fault diagnosis method based on deep belief network

Country Status (1)

Country Link
CN (1) CN109086817B (en)

Cited By (37)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109669087A (en) * 2019-01-31 2019-04-23 国网河南省电力公司 A kind of method for diagnosing fault of power transformer based on Multi-source Information Fusion
CN109784276A (en) * 2019-01-16 2019-05-21 东南大学 A kind of voltage dip feature extraction based on DBN and temporarily drop source discrimination method
CN109885948A (en) * 2019-02-28 2019-06-14 沈阳天眼智云信息科技有限公司 Method for diagnosing fault of power transformer
CN109948589A (en) * 2019-03-31 2019-06-28 西安电子科技大学 Facial expression recognizing method based on quantum deepness belief network
CN110132627A (en) * 2019-05-28 2019-08-16 上海海事大学 A kind of method for diagnosing faults of propeller
CN110175386A (en) * 2019-05-21 2019-08-27 陕西科技大学 Substation Electric Equipment temperature predicting method
CN110231165A (en) * 2019-05-31 2019-09-13 西安交通大学 A kind of Trouble Diagnostic Method of Machinery Equipment based on desired difference constraint confidence network
CN110263837A (en) * 2019-06-13 2019-09-20 河海大学 A kind of circuit breaker failure diagnostic method based on multilayer DBN model
CN110297480A (en) * 2019-06-17 2019-10-01 重庆仲澜科技有限公司 The TE process failure diagnosis method of deepness belief network model based on parameter optimization
CN110309586A (en) * 2019-06-28 2019-10-08 西安工程大学 GIS based on MA-DBN decomposes gas content prediction method
CN110516757A (en) * 2019-09-02 2019-11-29 广东工业大学 A kind of transformer fault detection method and relevant apparatus
CN110737998A (en) * 2019-09-25 2020-01-31 中国电力科学研究院有限公司 equalizing ring optimization design method based on finite element and depth belief network
CN110763997A (en) * 2019-11-04 2020-02-07 华北电力大学（保定） Early fault early warning method for synchronous motor stator
CN110879377A (en) * 2019-11-22 2020-03-13 国网新疆电力有限公司电力科学研究院 Metering device fault tracing method based on deep belief network
CN111060813A (en) * 2019-12-09 2020-04-24 国网北京市电力公司 Fault diagnosis method and device for high-voltage circuit breaker operating mechanism and electronic equipment
CN111105877A (en) * 2019-12-24 2020-05-05 郑州科技学院 Chronic disease accurate intervention method and system based on deep belief network
CN111178506A (en) * 2019-12-26 2020-05-19 哈尔滨工业大学 Large-scale high-speed rotation equipment deflection and inclination eliminating method based on deep belief neural network
CN111368888A (en) * 2020-02-25 2020-07-03 重庆邮电大学 Service function chain fault diagnosis method based on deep dynamic Bayesian network
CN111562248A (en) * 2020-05-15 2020-08-21 云南电网有限责任公司电力科学研究院 Based on SF6Internal standard GIS fault diagnosis method
CN111612029A (en) * 2020-03-30 2020-09-01 西南电子技术研究所（中国电子科技集团公司第十研究所） Airborne electronic product fault prediction method
CN111653272A (en) * 2020-06-01 2020-09-11 重庆科技学院 Vehicle-mounted voice enhancement algorithm based on deep belief network
CN111678679A (en) * 2020-05-06 2020-09-18 内蒙古电力(集团)有限责任公司电力调度控制分公司 Circuit breaker fault diagnosis method based on PCA-BPNN
CN111814878A (en) * 2020-07-09 2020-10-23 仲恺农业工程学院 SSDA-HELM-SOFTMAX-based real-time classification prediction method for agricultural input products
CN111827982A (en) * 2019-04-17 2020-10-27 中国石油天然气集团有限公司 Method and device for predicting overflow and leakage working conditions of drilling well
CN111898650A (en) * 2020-07-08 2020-11-06 国网浙江省电力有限公司杭州供电公司 Marketing and distribution data automatic clustering analysis equipment and method based on deep learning
CN112001066A (en) * 2020-07-30 2020-11-27 四川大学 Deep learning-based method for calculating limit transmission capacity
CN112287813A (en) * 2020-10-27 2021-01-29 西安科技大学 Intelligent ferrographic image identification method based on deep belief network
CN112308208A (en) * 2020-09-23 2021-02-02 云南电网有限责任公司普洱供电局 Transformer fault diagnosis method based on deep learning model
CN112307851A (en) * 2019-08-02 2021-02-02 上海交通大学烟台信息技术研究院 Method and system for identifying bird nest on electric power iron tower
CN112633558A (en) * 2020-12-08 2021-04-09 武汉理工大学 Fan blade icing state detection method and device and readable storage medium
CN112686340A (en) * 2021-03-12 2021-04-20 成都点泽智能科技有限公司 Dense small target detection method based on deep neural network
CN113030717A (en) * 2021-03-12 2021-06-25 广东电网有限责任公司 Circuit breaker fault diagnosis method and circuit based on electric signal deep learning
CN113033833A (en) * 2021-02-25 2021-06-25 江苏大学 Method for diagnosing harvesting state fault of threshing cylinder of combine harvester
CN113486598A (en) * 2021-07-27 2021-10-08 西安科技大学 ECT filling pipeline flow pattern prediction method based on PSO-DBN-ELM
CN113671361A (en) * 2021-06-25 2021-11-19 海南电网有限责任公司海南输变电检修分公司 High-voltage circuit breaker characteristic parameter prediction method and system based on multi-source signal fusion
CN113858218A (en) * 2021-12-06 2021-12-31 湖南工商大学 Fault diagnosis method for mechanical arm actuator
CN116471198A (en) * 2023-06-19 2023-07-21 南京典格通信科技有限公司 Power amplifier fault prediction method based on limited Boltzmann machine

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105973594A (en) * 2016-04-25 2016-09-28 西北工业大学 Rolling bearing fault prediction method based on continuous deep belief network
CN106897744A (en) * 2017-02-27 2017-06-27 郑州云海信息技术有限公司 A kind of self adaptation sets the method and system of depth confidence network parameter

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105973594A (en) * 2016-04-25 2016-09-28 西北工业大学 Rolling bearing fault prediction method based on continuous deep belief network
CN106897744A (en) * 2017-02-27 2017-06-27 郑州云海信息技术有限公司 A kind of self adaptation sets the method and system of depth confidence network parameter

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
JUN HE ET AL.: "Unsupervised Fault Diagnosis of a Gear Transmission Chain Using a Deep Belief Network", 《 SENSORS》 *
李巍华 等: "基于深度信念网络的轴承故障分类识别", 《振动工程学报》 *
母卓元: "基于ReLU深度信念网络的35kV开关柜故障诊断", 《电气传动自动化》 *

Cited By (48)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109784276A (en) * 2019-01-16 2019-05-21 东南大学 A kind of voltage dip feature extraction based on DBN and temporarily drop source discrimination method
CN109669087A (en) * 2019-01-31 2019-04-23 国网河南省电力公司 A kind of method for diagnosing fault of power transformer based on Multi-source Information Fusion
CN109885948A (en) * 2019-02-28 2019-06-14 沈阳天眼智云信息科技有限公司 Method for diagnosing fault of power transformer
CN109948589A (en) * 2019-03-31 2019-06-28 西安电子科技大学 Facial expression recognizing method based on quantum deepness belief network
CN111827982A (en) * 2019-04-17 2020-10-27 中国石油天然气集团有限公司 Method and device for predicting overflow and leakage working conditions of drilling well
CN110175386A (en) * 2019-05-21 2019-08-27 陕西科技大学 Substation Electric Equipment temperature predicting method
CN110175386B (en) * 2019-05-21 2022-11-25 陕西科技大学 Method for predicting temperature of electrical equipment of transformer substation
CN110132627A (en) * 2019-05-28 2019-08-16 上海海事大学 A kind of method for diagnosing faults of propeller
CN110231165A (en) * 2019-05-31 2019-09-13 西安交通大学 A kind of Trouble Diagnostic Method of Machinery Equipment based on desired difference constraint confidence network
CN110231165B (en) * 2019-05-31 2020-03-31 西安交通大学 Mechanical equipment fault diagnosis method based on expectation difference constraint confidence network
CN110263837A (en) * 2019-06-13 2019-09-20 河海大学 A kind of circuit breaker failure diagnostic method based on multilayer DBN model
CN110297480A (en) * 2019-06-17 2019-10-01 重庆仲澜科技有限公司 The TE process failure diagnosis method of deepness belief network model based on parameter optimization
CN110297480B (en) * 2019-06-17 2022-07-01 重庆仲澜科技有限公司 TE process fault diagnosis method of deep belief network model based on parameter optimization
CN110309586A (en) * 2019-06-28 2019-10-08 西安工程大学 GIS based on MA-DBN decomposes gas content prediction method
CN112307851A (en) * 2019-08-02 2021-02-02 上海交通大学烟台信息技术研究院 Method and system for identifying bird nest on electric power iron tower
CN110516757A (en) * 2019-09-02 2019-11-29 广东工业大学 A kind of transformer fault detection method and relevant apparatus
CN110737998A (en) * 2019-09-25 2020-01-31 中国电力科学研究院有限公司 equalizing ring optimization design method based on finite element and depth belief network
CN110737998B (en) * 2019-09-25 2022-07-26 中国电力科学研究院有限公司 Grading ring optimization design method based on finite element and deep belief network
CN110763997A (en) * 2019-11-04 2020-02-07 华北电力大学（保定） Early fault early warning method for synchronous motor stator
CN110879377A (en) * 2019-11-22 2020-03-13 国网新疆电力有限公司电力科学研究院 Metering device fault tracing method based on deep belief network
CN111060813A (en) * 2019-12-09 2020-04-24 国网北京市电力公司 Fault diagnosis method and device for high-voltage circuit breaker operating mechanism and electronic equipment
CN111105877A (en) * 2019-12-24 2020-05-05 郑州科技学院 Chronic disease accurate intervention method and system based on deep belief network
CN111178506A (en) * 2019-12-26 2020-05-19 哈尔滨工业大学 Large-scale high-speed rotation equipment deflection and inclination eliminating method based on deep belief neural network
CN111368888B (en) * 2020-02-25 2022-07-01 重庆邮电大学 Service function chain fault diagnosis method based on deep dynamic Bayesian network
CN111368888A (en) * 2020-02-25 2020-07-03 重庆邮电大学 Service function chain fault diagnosis method based on deep dynamic Bayesian network
CN111612029A (en) * 2020-03-30 2020-09-01 西南电子技术研究所（中国电子科技集团公司第十研究所） Airborne electronic product fault prediction method
CN111678679A (en) * 2020-05-06 2020-09-18 内蒙古电力(集团)有限责任公司电力调度控制分公司 Circuit breaker fault diagnosis method based on PCA-BPNN
CN111562248A (en) * 2020-05-15 2020-08-21 云南电网有限责任公司电力科学研究院 Based on SF6Internal standard GIS fault diagnosis method
CN111653272A (en) * 2020-06-01 2020-09-11 重庆科技学院 Vehicle-mounted voice enhancement algorithm based on deep belief network
CN111898650A (en) * 2020-07-08 2020-11-06 国网浙江省电力有限公司杭州供电公司 Marketing and distribution data automatic clustering analysis equipment and method based on deep learning
CN111814878B (en) * 2020-07-09 2023-06-06 仲恺农业工程学院 Agricultural input product real-time classification prediction method based on SSDA-HELM-SOFTMAX
CN111814878A (en) * 2020-07-09 2020-10-23 仲恺农业工程学院 SSDA-HELM-SOFTMAX-based real-time classification prediction method for agricultural input products
CN112001066A (en) * 2020-07-30 2020-11-27 四川大学 Deep learning-based method for calculating limit transmission capacity
CN112001066B (en) * 2020-07-30 2022-11-04 四川大学 Deep learning-based method for calculating limit transmission capacity
CN112308208B (en) * 2020-09-23 2023-01-24 云南电网有限责任公司普洱供电局 Transformer fault diagnosis method based on deep learning model
CN112308208A (en) * 2020-09-23 2021-02-02 云南电网有限责任公司普洱供电局 Transformer fault diagnosis method based on deep learning model
CN112287813B (en) * 2020-10-27 2023-06-23 西安科技大学 Ferrograph image intelligent recognition method based on deep belief network
CN112287813A (en) * 2020-10-27 2021-01-29 西安科技大学 Intelligent ferrographic image identification method based on deep belief network
CN112633558A (en) * 2020-12-08 2021-04-09 武汉理工大学 Fan blade icing state detection method and device and readable storage medium
CN113033833A (en) * 2021-02-25 2021-06-25 江苏大学 Method for diagnosing harvesting state fault of threshing cylinder of combine harvester
CN112686340A (en) * 2021-03-12 2021-04-20 成都点泽智能科技有限公司 Dense small target detection method based on deep neural network
CN113030717A (en) * 2021-03-12 2021-06-25 广东电网有限责任公司 Circuit breaker fault diagnosis method and circuit based on electric signal deep learning
CN112686340B (en) * 2021-03-12 2021-07-13 成都点泽智能科技有限公司 Dense small target detection method based on deep neural network
CN113671361A (en) * 2021-06-25 2021-11-19 海南电网有限责任公司海南输变电检修分公司 High-voltage circuit breaker characteristic parameter prediction method and system based on multi-source signal fusion
CN113486598A (en) * 2021-07-27 2021-10-08 西安科技大学 ECT filling pipeline flow pattern prediction method based on PSO-DBN-ELM
CN113858218A (en) * 2021-12-06 2021-12-31 湖南工商大学 Fault diagnosis method for mechanical arm actuator
CN116471198A (en) * 2023-06-19 2023-07-21 南京典格通信科技有限公司 Power amplifier fault prediction method based on limited Boltzmann machine
CN116471198B (en) * 2023-06-19 2023-10-03 南京典格通信科技有限公司 Power amplifier fault prediction method based on limited Boltzmann machine

Also Published As

Publication number Publication date
CN109086817B (en) 2022-04-26

Similar Documents

Publication Publication Date Title
CN109086817A (en) A kind of Fault Diagnosis for HV Circuit Breakers method based on deepness belief network
Dai et al. Dissolved gas analysis of insulating oil for power transformer fault diagnosis with deep belief network
Omlin et al. Extraction of rules from discrete-time recurrent neural networks
CN105930901B (en) A kind of Diagnosis Method of Transformer Faults based on RBPNN
Li et al. Siamese contrastive embedding network for compositional zero-shot learning
Cho et al. Classifying and visualizing motion capture sequences using deep neural networks
CN108805188A (en) A kind of feature based recalibration generates the image classification method of confrontation network
CN108734202A (en) A kind of Fault Diagnosis for HV Circuit Breakers method based on improved BP
CN109271975A (en) A kind of electrical energy power quality disturbance recognition methods based on big data multi-feature extraction synergetic classification
Esmin et al. Hybrid evolutionary algorithm based on PSO and GA mutation
CN109711483B (en) Spark Autoencoder-based power system operation mode clustering method
CN106874963B (en) A kind of Fault Diagnosis Method for Distribution Networks and system based on big data technology
CN109298330A (en) Fault Diagnosis for HV Circuit Breakers method based on GHPSO-BP
CN108717149A (en) Diagnosis Method of Transformer Faults based on M-RVM fusion dynamic weightings AdaBoost
CN108170994A (en) A kind of oil-immersed electric reactor method for diagnosing faults based on two-way depth network
CN116759100B (en) Method for constructing chronic cardiovascular disease large model based on federal learning
Parvin et al. A novel classifier ensemble method based on class weightening in huge dataset
CN111737907A (en) Transformer fault diagnosis method and device based on deep learning and DGA
CN114661905A (en) Power grid fault diagnosis method based on BERT
Waqas et al. Robust bag classification approach for multi-instance learning via subspace fuzzy clustering
CN109901064A (en) Fault Diagnosis for HV Circuit Breakers method based on ICA-LVQ
Huang et al. Heterogeneous graph neural network for identifying hadronically decayed tau leptons at the high luminosity LHC
Qian et al. Fault diagnosis for generator unit based on RBF neural network optimized by GA-PSO
CN114861792A (en) Complex power grid key node identification method based on deep reinforcement learning
CN115168864A (en) Intelligent cross contract vulnerability detection method based on feature cross

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
GR01 Patent grant
GR01 Patent grant