CN110084148A - A kind of Mechanical Failure of HV Circuit Breaker diagnostic method - Google Patents
A kind of Mechanical Failure of HV Circuit Breaker diagnostic method Download PDFInfo
- Publication number
- CN110084148A CN110084148A CN201910278356.4A CN201910278356A CN110084148A CN 110084148 A CN110084148 A CN 110084148A CN 201910278356 A CN201910278356 A CN 201910278356A CN 110084148 A CN110084148 A CN 110084148A
- Authority
- CN
- China
- Prior art keywords
- training
- sample
- data
- network
- dbn
- 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
Links
Classifications
-
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Probability & Statistics with Applications (AREA)
- Supply And Distribution Of Alternating Current (AREA)
Abstract
The invention discloses a kind of Mechanical Failure of HV Circuit Breaker diagnostic methods, high-voltage circuitbreaker fault simulation is tested into resulting divide-shut brake coil current as the target data sample of fault diagnosis, by establishing the resulting data of divide-shut brake coil mathematical model simulation as auxiliary data sample, it is realized using deep layer belief network (DBN) to the deep layer excavation of sample data feature and extracted in self-adaptive, and combines the information matches of transfer learning method realization auxiliary data and target data.The method of the present invention combines transfer learning with deep layer belief network, the deep layer excavation and extracted in self-adaptive of data characteristics are carried out to circuit-breaker switching on-off coil current time-domain signal using deep layer belief network, and solve the problems, such as that physical fault training sample data amount is small in conjunction with transfer learning method, improve the generalization ability of fault diagnosis model.
Description
Technical field
The invention belongs to electrical technology field more particularly to a kind of Mechanical Failure of HV Circuit Breaker diagnostic methods.
Background technique
Smart grid be unable to do without the high-voltage circuitbreaker of high reliability, and operating status will directly affect entire electric system
The reliability of stability and power supply.Fault diagnosis technology, can as the intelligentized important content of high-voltage circuitbreaker and developing direction
To provide more reliable diagnostic message for repair based on condition of component, plays the role of prevention apparatus failure, improves overhaul efficiency.
Mainly include at present two parts content for the fault diagnosis of high-voltage circuitbreaker, the height collected is broken first
The signal datas such as road device divide-shut brake coil current, mechanical oscillation, contact displacement carry out feature extraction, calculate then in conjunction with machine learning
Method classify to characteristic quantity realizing the identification to fault state of circuit breaker.
Traditional feature extracting method excessively relies on artificial experience and a large amount of signal processing technology, and faces
All pairs of sensitive features of fault type can not be excavated when the system of complexity.The accurate foundation needs of fault diagnosis model fill
The training data of foot, and the working condition of complexity makes the fault data amount of high-voltage circuitbreaker less and is difficult to obtain in practice,
To influence the generalization ability of conventional machines study fault diagnosis model.
Summary of the invention
Goal of the invention: it in view of the above problems, the present invention proposes a kind of Mechanical Failure of HV Circuit Breaker diagnostic method, solves to pass
Feature extracting method present in Mechanical Failure of HV Circuit Breaker of uniting diagnosis is complicated, extracted characteristic quantity characterization degree is inadequate and
Fault sample data deficiencies and the problem of influence fault identification effect.
Technical solution: to achieve the purpose of the present invention, the technical scheme adopted by the invention is that: a kind of high-voltage circuitbreaker machine
Tool method for diagnosing faults, comprising steps of
(1) the divide-shut brake coil current time domain waveform data of acquisition high-voltage circuitbreaker fault simulation experiment are as fault diagnosis
Target sample data, using the resulting current data of divide-shut brake coil mathematical model simulation as aid sample data;Sample number
According to being pre-processed, target training sample, supplemental training sample and test sample are formed;
(2) pre-training is carried out to target training sample and supplemental training sample data respectively, obtains two DBN feature extractions
Model, and extract target training sample feature samples data corresponding with supplemental training sample;
(3) using the feature samples training BPNN classifier of tape label, by the target signature sample and supplemental characteristic of extraction
Input of the sample as neural network classifier updates sample weights using transfer learning TrAdaboost algorithm and training is classified
Device model parameter;
(4) the good DBN feature extraction network parameter of pre-training and fault grader parameter initialization depth nerve are utilized
Network, and model parameter is reversely finely tuned using initial training data, form Mechanical Failure of HV Circuit Breaker diagnostic model;
(5) accuracy rate of input test sample data verifying model.
Further, in the step 2, DBN network training includes unsupervised layer-by-layer pre-training and has supervision to finely tune;Without prison
Layer-by-layer pre-training is superintended and directed, using high-voltage circuit-breaker switching on-off coil current time-domain signal as input layer, passes through the side of unsupervised learning
Formula carries out layer-by-layer pre-training to DBN;There is supervision to finely tune, the DBN network that pre-training obtains reversely is stacked and forms reconstructed network, with
The output of reconstructed network and the minimum target of the error function of input data are realized using back-propagation algorithm to DBN network
Fine tuning.
Further, the deep layer belief network obtained with training is reversely stacked and is launched into a deep layer perceptron, deep layer
The weight of perceptron is initialized with biasing using the parameter that successively training obtains, and is joined by back-propagation algorithm to network
Number is adjusted and optimizes.
Further, in the step 2, the dynamic tune of learning rate is carried out according to the performance of front and back moment DBN network performance
It is whole, adjust update mechanism are as follows:
In formula, 1,0 < β < 1 of α >, k is normal number, and t represents the t times iterative process;Represent single RBM
Reconstruct visual vector and the error for being originally inputted visual vector;Δ e=et-et-1Represent the t times iteration and the t-1 times iteration mistake
The difference of error in journey.
When reconstructing visual vector in training process and the former error for inputting visual vector is gradually reduced, then increase study
Rate is corresponding to reduce learning rate if error on the contrary increases larger.
Further, in the step 3, transfer learning algorithm is specifically included:
(3.1) calculating parameterWherein, n represents supplemental training sample number, and N represents iteration time
Number;Initialize the weight vectors of each training sampleWherein,
In formula, preceding n weight corresponds to supplemental training sample, and rear m weight corresponds to target training sample;
(3.2) iterative process, progress sample weights normalization first:
BPNN sorter model is established, according to training sample and normalized sample weights training pattern parameter;Sample is added
This weight, the loss function of BPNN network are as follows:
In formula, u represents the dimension of BPNN classifier output quantity,It is exported for the i-th dimension of j-th of training sample,For jth
The i-th dimension label of each training sample;
(3.3) BPNN network classifier h is calculatedtIn target data set TtOn error rate et:
And constraint condition e need to be mett≤0.5;
(3.4) sample weights are updated:
The utility model has the advantages that the method for the present invention combines deep layer belief network with transfer learning mechanism, and broken applied to height
Road device mechanical fault diagnosis, gives full play to the powerful characteristics extraction ability of DBN network and transfer learning strengthens classifier
Ability improves the generalization ability of fault diagnosis model.
It is adjusted present invention introduces learning rate dynamic and unsupervised reversed fine tuning is carried out by error reconstructed network, it can be into one
Step improves the feature representation ability and separability energy of the extracted feature of DBN network, has certain feasibility and validity.
The method of the present invention combines DBN network with transfer learning algorithm, have stronger feature representation ability with it is extensive
Ability can show good classification performance in the lesser situation of training sample amount, have apparent superiority.
Detailed description of the invention
Fig. 1 is that the present invention is based on the fault diagnosis model structural schematic diagrams of DBN and transfer learning mechanism;
Fig. 2 is method for diagnosing faults implementation process diagram of the invention;
Fig. 3 is the reconstructed error curve under different characteristic extracting method of the present invention;
Fig. 4 is the failure rate change curve under different target sample number of the present invention.
Specific embodiment
Further description of the technical solution of the present invention with reference to the accompanying drawings and examples.
As shown in Figure 1, the fault diagnosis model based on DBN network Yu transfer learning mechanism, is broadly divided into two parts: DBN
Characteristic extraction part and failure modes part.DBN characteristic extraction part is realized using original object and supplemental characteristic sample data
The unsupervised training of DBN network, the extracted in self-adaptive validity feature from high-dimensional initial data.Failure modes part passes through structure
It build on the basis of DBN network structure and increases the BPNN network classifier of label output layer, realize feature extraction and fault diagnosis
Fusion, and using the initialization for the DBN network parameter progress DNN network for completing pre-training, and utilize the training sample of tape label
With corresponding sample weights, carry out reversely thering is supervision to finely tune to network in conjunction with BP algorithm.According to trained network to target data
Classification performance update the weight of training sample using transfer learning algorithm, utilize auxiliary data sample and target data sample
The training for completing entire fault diagnosis model together recently enters target detection data, verifies the generalization ability of model built.
As shown in Fig. 2, Mechanical Failure of HV Circuit Breaker diagnostic method of the present invention, specifically includes step:
Step 1: high-voltage circuitbreaker fault simulation is tested into practical divide-shut brake coil current time domain waveform data collected
As the target sample data of fault diagnosis, divide-shut brake coil mathematical model is established, by divide-shut brake coil mathematical model simulation institute
The current data obtained is as aid sample data;Data sample is pre-processed, original training sample and test sample are formed;
Step 2: it is realized using deep layer belief network (DBN) to the deep layer excavation of sample data feature and extracted in self-adaptive,
Pre-training is carried out respectively using original object sample and aid sample data, obtains two DBN Feature Selection Models, and extract original
The corresponding feature samples data of beginning sample;
DBN characteristic extraction part is passed through using high-voltage circuit-breaker switching on-off coil current time-domain signal as input layer
The mode of unsupervised learning carries out layer-by-layer pre-training to DBN, and the DBN network that pre-training obtains reversely is stacked formation reconstruct net
Network is realized using back-propagation algorithm to DBN net with the minimum target of the error function of the output of reconstructed network and input data
The fine tuning of network further increases the extractability of deep layer belief network characteristic to reduce reconstructed error.
Deep layer belief network can be regarded as to be stacked by multiple limited Boltzmann machines and a classifier, and upper one layer
The output of RBM is the input of next layer of RBM, is mentioned by the stacking realization of multilayer RBM to initial data feature on the middle and senior level
It takes, classifier finally is added in top layer and forms complete network structure, the identification of completion status type.
The training of DBN network is divided into unsupervised layer-by-layer pre-training and has supervision fine tuning two parts.The unsupervised pre-training time-division
It does not train each layer of RBM network unsupervisedly individually, avoids complex calculation brought by whole training.There is supervision fine tuning rank
Section, using label data, carries out network in conjunction with back-propagation algorithm whole by the way that the classifier for having supervision is added in network top
Body fine tuning.
In order to further increase the ability that deep layer belief network extracts characteristic, reduce reconstructed error, the present invention will be with
The deep layer belief network that training obtains, which reversely stacks, is launched into a deep layer perceptron.The weight of deep layer perceptron and biasing use
The parameter that successively training obtains is initialized, and network parameter is adjusted and is optimized by back-propagation algorithm.
Learning rate is the important hyper parameter for influencing the convergence rate and effect of DBN network pre-training, is normally based on experience
It is manually set to fixed value, and fixed learning rate tends not to the entire learning process for being suitable for network well.
In order to realize quickly and effectively learning training process, the present invention is according to the performance of front and back moment network performance
The dynamic of habit rate adjusts, method of adjustment are as follows:
In formula, 1,0 < β < 1 of α >, k is normal number, and t represents the t times iterative process;Represent single RBM
Visual vector and the error for being originally inputted visual vector are reconstructed, calculating can be passed throughWithEuclidean distance obtain;Δ e=et-
et-1Represent the difference of the t times iteration with error in the t-1 times iterative process.
Using above-mentioned mechanism dynamic regularized learning algorithm rate, when in training process when reconstructing visual vector and former input visual vector
Error when being gradually reduced, then increase learning rate, if otherwise error when increasing larger, it is corresponding to reduce learning rate, to realize
The dynamic of learning rate adjusts.
Step 3: using the feature samples training BPNN classifier of tape label, by extracted target signature sample and auxiliary
Input of the feature samples as neural network (NN) classifier updates sample weights simultaneously using transfer learning TrAdaboost algorithm
Training sorter model parameter, realizes the information matches of auxiliary data and target data, to help target signature sample training
The classifier with higher accuracy is obtained, fault diagnosis essence is further increased in the case where failure training sample data amount is small
Degree;
Due to the accuracy for needing a large amount of training datas to guarantee classification that is limited in that of deep learning algorithm, and high pressure
Divide-shut brake coil current data under breaker is nonserviceabled often are difficult to acquire, and the training sample that can be obtained is less.
In order to still ensure that the generalization ability of disaggregated model in the case where small sample target data, present invention introduces migrations
Study mechanism obtains the emulation of a large amount of tape labels under fault state of circuit breaker by constructing the mathematical model of divide-shut brake coil
Data, and the training of classifier is completed jointly in this, as auxiliary data collection and a small amount of target data set, to improve failure
The generalization ability of diagnostic model.
The present invention uses the conclusion transfer learning mechanism of Case-based Reasoning, and specific implementation algorithm flow is as follows:
1) original auxiliary sample data is obtained by establishing coil electromagnetism iron mathematical model and emulating, passes through fault simulation reality
It tests and collects target domain divide-shut brake coil current original target data, by auxiliary data and target data through data prediction
Form training sample;
2) calculating parameterWherein, n represents supplemental training sample number, and N represents the number of iterations;
Initialize the weight vectors of each training sampleWherein:
In formula, preceding n weight corresponds to supplemental training sample, and rear m weight corresponds to target training sample.
3) iterative process, progress sample weights normalization first:
BPNN sorter model (three-layer neural network) is established, according to training sample and normalized sample weights training mould
Shape parameter;Due to joined sample weights, the loss function modification of BPNN network are as follows:
In formula, u represents the dimension of BPNN classifier output quantity,It is exported for the i-th dimension of j-th of training sample,For jth
The i-th dimension label of each training sample.
4) BPNN network classifier h is calculatedtIn target data set TtOn error rate et:
And constraint condition e need to be mett≤0.5;
5) sample weights are updated:
By algorithm flow it can be found that for source data training sample, when generating erroneous judgement, corresponding sample weights will subtract
It is small, to reduce its influence to global error;For target data training sample, when generating erroneous judgement, corresponding sample power
It will increase again, to reduce error rate of this sample in classification based training next time.After certain the number of iterations, auxiliary data
In sample weights contradictory with target data will become very little, and that target data training will be helped to obtain will be general for the biggish sample of weight
The stronger disaggregated model of change ability.
Step 4: utilizing the good DBN feature extraction network parameter of pre-training and fault grader parameter initialization depth mind
Model parameter is reversely finely tuned through network (DNN), and using initial training data, it is mechanical to form final high-voltage circuitbreaker
Fault diagnosis model;
Step 5: the accuracy rate of input test sample data verifying model.
By learning rate dynamic adjustment complete pre-training DBN network foundation on, using tape label training data into
The entire fault diagnosis model of row has the reversed fine tuning of supervision, improves the feature extraction performance of network.Detailed process is as follows:
1) acquire divide-shut brake coil current entire time domain pulse signal, carry out data prediction, by data normalization to [0,
1];
2) initial data is divided into training set and test set;
3) learning rate, visible elements and the number of hidden nodes of RBM weight and biasing, weight attenuation coefficient and dynamic are initialized
Quantifier determines RBM number (and DBN number of plies), and using training set data, the RBM completed under learning rate dynamic adjusts is successively instructed in advance
Practice;
4) the DBN network after pre-training is reversely stacked and expands into deep layer perceptron, with the output of entire deep layer perceptron
Square error with input realizes the reversed tuning of entire reconstructed network using BP algorithm as loss function;
5) NN network is initialized using the DBN network parameter for completing pre-training and fine tuning;Using in training set
There is label data combination BP algorithm to carry out having for NN network and supervises reversed fine tuning;
6) by the DNN network after test data input training, ability in feature extraction and the fault diagnosis for verifying network are accurate
Rate.
Below by one embodiment, the present invention is described further.
For the SF of certain model 110kV6High-voltage circuitbreaker builds fault simulation experiment porch, and divide-shut brake coil current is made
Acquisition for the key data source of Fault Diagnosis for HV Circuit Breakers, original signal is realized by current clamp and digital oscilloscope.
According to physical condition, on the basis of not destroying circuit breaker internal structure, two quasi-representative failures, respectively division brake cable are devised
Enclose aging and iron core bite.Wherein, it is divided into tri- kinds of fault degrees of A, B, C again under coil aging, respectively represents slight, moderate and again
Spend failure;Iron core bite is divided into two kinds of fault degrees of A, B (slightly and moderate).The aging of divide-shut brake coil passes through in divide-shut brake coil
Circuit seals in adjustable resistor, and carrys out the degree of former-wound coil aging by adjusting different resistance values.Iron core jam faults are logical
It crosses and is simulated in divide-shut brake iron-core coil underhung weight, and the degree of bite is adjusted by changing weight quality.Specifically
Fault type description it is as shown in table 1.
Table 1
Fault type | Fault degree | Analogy method | Label |
Normally | - | - | F11 |
Coil aging | A | Circuit seals in adjustable resistance (resistance value be adjusted to coil resistance value 20%) | F21 |
Coil aging | B | Circuit seals in adjustable resistance (resistance value be adjusted to coil resistance value 50%) | F22 |
Coil aging | C | Circuit seals in adjustable resistance (resistance value be adjusted to coil resistance value 80%) | F23 |
Iron core bite | A | Iron core underhung weight | F31 |
Iron core bite | B | Iron core underhung weight | F32 |
In addition, for the electromagnet of the spring operating mechanism of 110kV SF6 high-voltage circuitbreaker in physical fault simulated experiment
Part constructs its dynamic mathematical models using Simulink, while by changing loop resistance resistance value R and iron core quality in model
The parameter equivalents former-wound coil aging such as m and iron core bite these two types failure.According to the working principle of electromagnet system, in conjunction with electric current
Divide-shut brake coil former is reduced to 3 rank mathematical models by circuit and kinetics equation:
In order to verify the ability in feature extraction of DBN network, 900 groups of closing coils obtained will be tested by fault simulation
Current data is as data source.Each group of primary data sample is after pretreatment containing 800 dimension strong points and as DBN network
Input.The DBN number of plies for feature extraction is set as 5 layers, and wherein input layer number is 800, feature output layer neuron
Number is that the neuron number of 10,3 hidden layers is respectively 300-100-50.The initial learning rate of RBM is respectively set to initial momentum item
0.01 and 0.02, network weight attenuation coefficient is set as 0.5, RBM pre-training and is all provided with reversed fine tuning stage maximum number of iterations
It is set to 1000.Training sample is inputted into DBN network and carries out unsupervised training and reversed fine tuning, learning rate dynamic change when training
The reconstructed error curve of the characteristic value of process and extracted characteristic value and traditional method for extracting is as shown in Figure 3.
It can be seen from reconstructed error curve traditional characteristic extracting method in an iterative process reconstructed error fluctuation it is larger and
Convergence rate is slower, and lower reconstructed error and faster convergence will be obtained by carrying out current characteristic extraction using DBN network
Speed.Wherein, on traditional DBN network foundation, feature reconstruction will be substantially reduced by carrying out reversed fine tuning by construction reconstructed network
Error reduces convergence rate.And adjusted by the dynamic to learning rate in training process, reconstructed error can be further decreased,
To obtain more preferably feature extracted in self-adaptive network.Table 2, which is calculated, to be mentioned using the mentioned method of the application with conventional method
Euclidean distance of the characteristic value taken in fault type of the same race and between different faults type.As it can be seen that with traditional characteristic extraction side
Formula is compared, and has smaller inter- object distance and bigger between class distance using the characteristic value that DBN network extracts, this illustrates to utilize
The feature that DBN network extracts will have stronger separability.
Table 2
The 180 groups of target signature samples and 540 groups of supplemental characteristic samples for selecting DBN network to extract are as BPNN classifier
Training data, wherein the dimension of feature samples be 10.The number of iterations of transfer learning algorithm is set as 100, utilizes feature extraction
The model parameter of part and failure modes part initializes DNN network, and carries out small parameter perturbations using BP algorithm, to formed
Whole fault diagnosis model.Table 3 show the corresponding fault identification accuracy rate of different models.As it can be seen that the mentioned method of the application will
DBN network is combined with transfer learning algorithm, is trained resulting failure by the weighted value of continuous adjusting training sample and is examined
Disconnected model, accuracy rate of diagnosis with higher;If not adjusting training sample weights, directly target data and auxiliary data are closed
It is trained together, then no matter which kind of fault diagnosis algorithm is used, accuracy rate reduces nearly 30%.This is because auxiliary data
In training effect with target data unmatched some effects entirety, thus greatly reduce the accuracy rate of model, therefore
It only selects to ensure last fault diagnosis effect with the stronger part of target data matching in supplemental training data.
Table 3
Further to analyze influence of the quantity of target sample data to transfer learning performance, target sample data volume is adjusted
And draw corresponding accuracy rate variation diagram, such as Fig. 4.As it can be seen that under the premise of auxiliary data amount and target data amount ratio are certain,
With the increase of target data amount, either conventional machines learning method, DBN network or the DBN+ transfer learning mentioned herein
Fault diagnosis model, accuracy rate of diagnosis all step up;After target amount of training data reaches 120 groups, DBN model and DBN+
The fault diagnosis accuracy rate of transfer learning model is very close and all 99% or more, illustrates that target amount of training data is at this time
It is used to train high performance diagnostic model, necessity of non-migratory study enough;When only 15 groups of target amount of training data, due to instruction
Practice that data volume is too small, and the accuracy rate of conventional machines learning method substantially reduces, though DBN network accuracy rate is promoted but still not
To 80%, and the fault diagnosis model accuracy rate of DBN+ transfer learning remains to be maintained at 90% or so, this has fully demonstrated this mould
Powerful advantages of the type in small sample fault diagnosis.
Claims (5)
1. a kind of Mechanical Failure of HV Circuit Breaker diagnostic method, which is characterized in that comprising steps of
(1) mesh of the divide-shut brake coil current time domain waveform data of acquisition high-voltage circuitbreaker fault simulation experiment as fault diagnosis
Standard specimen notebook data, using the resulting current data of divide-shut brake coil mathematical model simulation as aid sample data;Sample data into
Row pretreatment, forms target training sample, supplemental training sample and test sample;
(2) pre-training is carried out to target training sample and supplemental training sample data respectively, obtains two DBN feature extraction moulds
Type, and extract target training sample feature samples data corresponding with supplemental training sample;
(3) using the feature samples training BPNN classifier of tape label, by the target signature sample of extraction and supplemental characteristic sample
As the input of neural network classifier, sample weights and training classifier mould are updated using transfer learning TrAdaboost algorithm
Shape parameter;
(4) the DBN feature extraction network parameter and fault grader parameter initialization deep neural network good using pre-training,
And model parameter is reversely finely tuned using initial training data, form Mechanical Failure of HV Circuit Breaker diagnostic model;
(5) accuracy rate of input test sample data verifying model.
2. Mechanical Failure of HV Circuit Breaker diagnostic method according to claim 1, which is characterized in that in the step 2,
DBN network training includes unsupervised layer-by-layer pre-training and has supervision to finely tune;
Unsupervised layer-by-layer pre-training, using high-voltage circuit-breaker switching on-off coil current time-domain signal as input layer, by unsupervised
The mode of study carries out layer-by-layer pre-training to DBN;
Have supervision finely tune, the DBN network that pre-training obtains reversely is stacked and forms reconstructed network, with the output of reconstructed network with it is defeated
The minimum target of error function for entering data realizes the fine tuning to DBN network using back-propagation algorithm.
3. Mechanical Failure of HV Circuit Breaker diagnostic method according to claim 2, which is characterized in that, will in the step 2
The deep layer belief network obtained with training, which reversely stacks, is launched into a deep layer perceptron, and weight and the biasing of deep layer perceptron are adopted
It is initialized with the parameter that layer-by-layer training obtains, and network parameter is adjusted and is optimized by back-propagation algorithm.
4. Mechanical Failure of HV Circuit Breaker diagnostic method according to claim 1, which is characterized in that in the step 2, root
The dynamic adjustment that learning rate is carried out according to the performance of front and back moment DBN network performance, adjusts update mechanism are as follows:
In formula, 1,0 < β < 1 of α >, k is normal number, and t represents the t times iterative process;Representing single RBM reconstruct can
Visual direction amount and the error for being originally inputted visual vector;Δ e=et-et-1It represents and is missed in the t times iteration and the t-1 times iterative process
The difference of difference;
When reconstructing visual vector in training process and the former error for inputting visual vector is gradually reduced, then increase learning rate, instead
If error when increasing larger, it is corresponding to reduce learning rate.
5. Mechanical Failure of HV Circuit Breaker diagnostic method according to claim 1, which is characterized in that in the step 3, move
Learning algorithm is moved to specifically include:
(3.1) calculating parameterWherein, n represents supplemental training sample number, and N represents the number of iterations;Just
The weight vectors of each training sample of beginningizationWherein,
In formula, preceding n weight corresponds to supplemental training sample, and rear m weight corresponds to target training sample;
(3.2) iterative process, progress sample weights normalization first:
BPNN sorter model is established, according to training sample and normalized sample weights training pattern parameter;Sample power is added
Weight, the loss function of BPNN network are as follows:
In formula, u represents the dimension of BPNN classifier output quantity,It is exported for the i-th dimension of j-th of training sample,It is respectively instructed for jth
Practice the i-th dimension label of sample;
(3.3) BPNN network classifier h is calculatedtIn target data set TtOn error rate et:
And constraint condition e need to be mett≤0.5;
(3.4) sample weights are updated:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910278356.4A CN110084148A (en) | 2019-04-09 | 2019-04-09 | A kind of Mechanical Failure of HV Circuit Breaker diagnostic method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910278356.4A CN110084148A (en) | 2019-04-09 | 2019-04-09 | A kind of Mechanical Failure of HV Circuit Breaker diagnostic method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110084148A true CN110084148A (en) | 2019-08-02 |
Family
ID=67414533
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910278356.4A Pending CN110084148A (en) | 2019-04-09 | 2019-04-09 | A kind of Mechanical Failure of HV Circuit Breaker diagnostic method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110084148A (en) |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110376519A (en) * | 2019-08-30 | 2019-10-25 | 国家电网有限公司 | Fault Diagnosis for HV Circuit Breakers method, apparatus and terminal device |
CN110705114A (en) * | 2019-10-10 | 2020-01-17 | 辽宁工程技术大学 | Ventilation fault diagnosis method without training sample |
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 |
CN111314113A (en) * | 2020-01-19 | 2020-06-19 | 赣江新区智慧物联研究院有限公司 | Internet of things node fault detection method and device, storage medium and computer equipment |
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 |
CN111738286A (en) * | 2020-03-17 | 2020-10-02 | 北京京东乾石科技有限公司 | Fault determination and model training method, device, equipment and storage medium thereof |
CN112200032A (en) * | 2020-09-28 | 2021-01-08 | 辽宁石油化工大学 | Attention mechanism-based high-voltage circuit breaker mechanical state online monitoring method |
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 |
CN112884093A (en) * | 2021-04-29 | 2021-06-01 | 四川大学 | Rotary machine fault diagnosis method and equipment based on DSCRN model and storage medium |
CN113255213A (en) * | 2021-05-19 | 2021-08-13 | 南京工程学院 | Power converter composite fault identification method |
CN113506245A (en) * | 2021-06-11 | 2021-10-15 | 国网浙江省电力有限公司嘉兴供电公司 | High-voltage circuit breaker mechanical performance evaluation method based on deep residual error network |
NL2032669A (en) * | 2022-08-03 | 2022-09-20 | Honghe Power Supply Bureau Of Yunnan Power Grid Co Ltd | Intelligent method for diagnosing mechanical fault of high-voltage isolating switch based on transfer learning |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105891707A (en) * | 2016-05-05 | 2016-08-24 | 河北工业大学 | Opening-closing fault diagnosis method for air circuit breaker based on vibration signals |
CN106769048A (en) * | 2017-01-17 | 2017-05-31 | 苏州大学 | Adaptive deep confidence network bearing fault diagnosis method based on Nesterov momentum method |
CN107247231A (en) * | 2017-07-28 | 2017-10-13 | 南京航空航天大学 | A kind of aerogenerator fault signature extracting method based on OBLGWO DBN models |
-
2019
- 2019-04-09 CN CN201910278356.4A patent/CN110084148A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105891707A (en) * | 2016-05-05 | 2016-08-24 | 河北工业大学 | Opening-closing fault diagnosis method for air circuit breaker based on vibration signals |
CN106769048A (en) * | 2017-01-17 | 2017-05-31 | 苏州大学 | Adaptive deep confidence network bearing fault diagnosis method based on Nesterov momentum method |
CN107247231A (en) * | 2017-07-28 | 2017-10-13 | 南京航空航天大学 | A kind of aerogenerator fault signature extracting method based on OBLGWO DBN models |
Non-Patent Citations (1)
Title |
---|
YI PAN等: "An Approach for HVCB Mechanical Fault Diagnosis Based on a Deep Belief Network and a Transfer Learning Strategy", 《JOURNAL OF ELECTRICAL ENGINEERING & TECHNOLOGY》 * |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110376519A (en) * | 2019-08-30 | 2019-10-25 | 国家电网有限公司 | Fault Diagnosis for HV Circuit Breakers method, apparatus and terminal device |
CN110376519B (en) * | 2019-08-30 | 2021-11-16 | 国家电网有限公司 | High-voltage circuit breaker fault diagnosis method and device and terminal equipment |
CN110705114A (en) * | 2019-10-10 | 2020-01-17 | 辽宁工程技术大学 | Ventilation fault diagnosis method without training sample |
CN110705114B (en) * | 2019-10-10 | 2023-04-07 | 辽宁工程技术大学 | Ventilation fault diagnosis method without training sample |
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 |
CN111314113A (en) * | 2020-01-19 | 2020-06-19 | 赣江新区智慧物联研究院有限公司 | Internet of things node fault detection method and device, storage medium and computer equipment |
CN111738286A (en) * | 2020-03-17 | 2020-10-02 | 北京京东乾石科技有限公司 | Fault determination and model training method, device, equipment and storage medium thereof |
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 |
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 |
CN112200032A (en) * | 2020-09-28 | 2021-01-08 | 辽宁石油化工大学 | Attention mechanism-based high-voltage circuit breaker mechanical state online monitoring method |
CN112200032B (en) * | 2020-09-28 | 2023-05-30 | 辽宁石油化工大学 | Online monitoring method for mechanical state of high-voltage circuit breaker based on attention mechanism |
CN112884093A (en) * | 2021-04-29 | 2021-06-01 | 四川大学 | Rotary machine fault diagnosis method and equipment based on DSCRN model and storage medium |
CN113255213A (en) * | 2021-05-19 | 2021-08-13 | 南京工程学院 | Power converter composite fault identification method |
CN113255213B (en) * | 2021-05-19 | 2023-08-22 | 南京工程学院 | Power converter composite fault identification method |
CN113506245A (en) * | 2021-06-11 | 2021-10-15 | 国网浙江省电力有限公司嘉兴供电公司 | High-voltage circuit breaker mechanical performance evaluation method based on deep residual error network |
NL2032669A (en) * | 2022-08-03 | 2022-09-20 | Honghe Power Supply Bureau Of Yunnan Power Grid Co Ltd | Intelligent method for diagnosing mechanical fault of high-voltage isolating switch based on transfer learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110084148A (en) | A kind of Mechanical Failure of HV Circuit Breaker diagnostic method | |
Shao et al. | Generative adversarial networks for data augmentation in machine fault diagnosis | |
CN109086817B (en) | High-voltage circuit breaker fault diagnosis method based on deep belief network | |
Guo et al. | Deep-learning-based earth fault detection using continuous wavelet transform and convolutional neural network in resonant grounding distribution systems | |
Pan et al. | An approach for HVCB mechanical fault diagnosis based on a deep belief network and a transfer learning strategy | |
CN107784293B (en) | A kind of Human bodys' response method classified based on global characteristics and rarefaction representation | |
CN107194433B (en) | A kind of Radar range profile's target identification method based on depth autoencoder network | |
CN109765333A (en) | A kind of Diagnosis Method of Transformer Faults based on GoogleNet model | |
CN110542819B (en) | Transformer fault type diagnosis method based on semi-supervised DBNC | |
CN109635928A (en) | A kind of voltage sag reason recognition methods based on deep learning Model Fusion | |
CN103728551B (en) | A kind of analog-circuit fault diagnosis method based on cascade integrated classifier | |
CN110534132A (en) | A kind of speech-emotion recognition method of the parallel-convolution Recognition with Recurrent Neural Network based on chromatogram characteristic | |
CN108921285A (en) | Single-element classification method in sequence based on bidirectional valve controlled Recognition with Recurrent Neural Network | |
CN109669087A (en) | A kind of method for diagnosing fault of power transformer based on Multi-source Information Fusion | |
CN109034054B (en) | Harmonic multi-label classification method based on LSTM | |
CN103926526A (en) | Analog circuit fault diagnosis method based on improved RBF neural network | |
CN107766794A (en) | The image, semantic dividing method that a kind of Fusion Features coefficient can learn | |
CN108520301A (en) | A kind of circuit intermittent fault diagnostic method based on depth confidence network | |
CN109443382A (en) | Vision SLAM closed loop detection method based on feature extraction Yu dimensionality reduction neural network | |
CN107272644B (en) | The DBN network fault diagnosis method of latent oil reciprocating oil pumping unit | |
CN110070172A (en) | The method for building up of sequential forecasting models based on two-way independent loops neural network | |
CN106980073A (en) | A kind of two ends of electric transmission line fault recorder data matching process based on convolutional neural networks | |
CN102279358B (en) | MCSKPCA based neural network fault diagnosis method for analog circuits | |
CN104155574A (en) | Power distribution network fault classification method based on adaptive neuro-fuzzy inference system | |
CN105528516A (en) | Clinic pathology data classification method based on combination of principal component analysis and extreme learning machine |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190802 |