CN110263837A - A kind of circuit breaker failure diagnostic method based on multilayer DBN model - Google Patents
A kind of circuit breaker failure diagnostic method based on multilayer DBN model Download PDFInfo
- Publication number
- CN110263837A CN110263837A CN201910510449.5A CN201910510449A CN110263837A CN 110263837 A CN110263837 A CN 110263837A CN 201910510449 A CN201910510449 A CN 201910510449A CN 110263837 A CN110263837 A CN 110263837A
- Authority
- CN
- China
- Prior art keywords
- dbn
- data
- sample
- cluster
- submodel
- 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
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
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)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The present invention discloses a kind of circuit breaker failure diagnostic method based on multilayer DBN model, mainly comprises the steps that and the high-voltage circuitbreaker fault data sample after standardization is classified using k-means algorithm, optimal cluster number of clusters is determined by silhouette coefficient;It establishes multilayer DBN model and is trained, first layer includes all training datas, and label is the cluster label that k mean cluster obtains;The second layer be DBN submodel, if cluster number of clusters be n, establish n DBN submodel;Submodel training data is the same cluster data after k-means cluster;Test data inputs training pattern in two steps: the first step inputs in first layer DBN model, obtains submodel selection label;Test data is inputted corresponding DBN submodel, obtains final class label, i.e. diagnostic result by second step according to the selection label of the first step.The verifying of data experiment has been obtained in precision of this method under the conditions of data are unbalanced.
Description
Technical field
The invention belongs to fault diagnosis technology, specifically a kind of circuit breaker failure based on multilayer DBN model
Diagnostic method.
Background technique
High-voltage circuitbreaker carries in the power system cut-offs and is closed the important work such as power circuit, line fault protection
With being the important equipment in power grid.The failure of high-voltage circuitbreaker can bring heavy losses to electric system, and health status is direct
Affect the safe and stable operation of power grid.Discovery failure and identification failure are to improve overhaul efficiency in time, extend equipment life, keep away
Exempt from the key of loss of outage.Fault diagnosis technology can submit necessary information for identifying and positioning for failure, be maintenance model
The premise developed from periodic inspection to repair based on condition of component.Electric equipment intelligent technology based on on-line monitoring and fault diagonosing will
It can play an important role in the construction and development of smart grid.As the important component of smart grid, electrical equipment
Fault diagnosis be always research hotspot.
Traditional method for diagnosing faults includes method based on model, time-domain and frequency-domain analysis method signal-based, is based on
The method of knowledge and Hybrid approaches of diagnosis etc..In fault data selection, mainly there are vibration signal, division brake current and contact
Displacement or angle of eccentricity etc..Machine learning is the one kind quickly grown in Knowledge based engineering method, with clustering, artificial neuron
Network (ANN) and support vector machines (SVM) are representative, more prominent in fault diagnosis field performance.The basic principle is that utilizing
Circuit breaker failure characteristic is established diagnostic model by learning and training, is assessed the state of breaker, its essence is
There is the classification of supervision.New mixing innovatory algorithm continuously emerges in recent years, improves the performance of diagnostic model.Wan Shuting
Et al. propose multiple features entropy fusion (MFEF) and hybrid classifer Mechanical Failure of HV Circuit Breaker diagnostic method.Li
Bing proposes the time-frequency entropy of improvement experience wavelet transformation (EWT) and optimization generalized regression nerve networks (GRNN) classifier is used for
Fault Diagnosis for HV Circuit Breakers.Ma Suliang is proposed based on WAVELET PACKET DECOMPOSITION technology and the open circuit of the high pressure of random forests algorithm
Device method for diagnosing faults.Zhu Kedong is proposed based on particle group optimizing-support vector domain description and particle group optimizing-mould
Paste the Fault Diagnosis for HV Circuit Breakers method of c- mean value kernel clustering.
However, circuit breaker failure diagnosis engineering application still has some problems.It is mainly reflected in: first is that machine learning
The fault signature data with label are needed to carry out training pattern in most cases, but most of fault data of on-site collection is all
It is the initial data that do not classify, data processing difficulty is big;Second, the fault type of breaker is numerous, and different faults occur
Probability it is also inconsistent, the precision of fault data sample imbalance, diagnostic model is severely impacted.
DBN is a kind of deep learning neural network being used widely in recent years, is to carry out having for fault diagnosis modeling
Effect tool.There are two clear advantages for the mechanism tool of successively greedy training: first, the step of carrying out feature extraction is not needed, and
It is by the training most suitable sample characteristics of extracted in self-adaptive.Second, only partial data mark is just needed in the reversed fine tuning stage
Label, do not need whole data labels.These features determine that DBN is suitable for establishing circuit breaker failure diagnostic model, Ke Yiqu
Obtain preferable learning effect.However, in practical engineering applications, due to the missing of data sample type, often diagnostic model face
Face the unbalanced problem of data, a small amount of unbalanced sample outlier easy to form in a model influences diagnostic accuracy.Therefore,
It invents one kind and the high pressure of advanced deep learning algorithm is utilized for fault data sample imbalance problem based on fault data
Circuit breaker failure diagnostic method becomes the problem of urgent need to resolve.
Summary of the invention
In response to the problems existing in the prior art, the purpose of the present invention is to provide a kind of breakers based on multilayer DBN model
Method for diagnosing faults influences diagnostic accuracy for solving the problem of that sample data is unbalanced.
To achieve the above object, the technical solution adopted by the present invention is that:
A kind of circuit breaker failure diagnostic method based on multilayer DBN model, diagnostic method are broadly divided into three steps:
One step is classified the high-voltage circuitbreaker fault data sample after standardization using k mean cluster, is determined by silhouette coefficient
Optimal cluster number of clusters;Second step is established multilayer DBN model and is trained, and first layer includes all training datas, mark
Label are the obtained cluster label of k mean cluster, and the second layer is that DBN submodel if number of clusters is n establishes n DBN submodel, submodule
Type training data is the same cluster data after k mean cluster;Third step, test data input training pattern, the first step in two steps
It inputs in first layer DBN model, obtains submodel selection label, second step is according to first conclusion, again according to selection label
Test data is inputted into corresponding DBN submodel, obtains final class label, i.e. diagnosis;The diagnostic method specifically wraps
Include following steps:
The fault data of breaker is carried out resampling by S1, data preparation, and each sample retains 800 data, establishes event
Hinder database;Fault data in the Mishap Database is divided into training set and test set, and by the training set and test
The fault data of concentration is standardized;
S2 presets cluster numbers C, is repeatedly clustered respectively to training data, determine optimum clustering number Copt, obtain optimal
The cluster label of cluster result and each sample;
S3 establishes DBN model and training using whole fault datas of training set and the corresponding cluster label of each data;
S4 establishes a DBN submodel and training, the DBN submodule using the sample in each cluster of training set respectively
Type shares CoptIt is a;
Sample in test set is inputted the DBN model, obtains the cluster label of test sample by S5;
S6 selects corresponding DBN submodel according to the cluster label of the test sample, by the test sample input with
The corresponding DBN submodel of cluster label, obtains diagnostic result.
Specifically, in step S1, method that the fault data in the training set and test set is standardized are as follows: adopt
It is standardized with Min-max, original sample value is mapped between [0,1], standardization formula is as follows:
Wherein, x ' is the data sample after standardization, and x is primary data sample, xmaxAnd xminFor the maximum of sample data
Value and minimum value.
Specifically, in step S2, the preset range of the cluster numbers C is 2~10.
Further, in step S2, the method that determines the optimum clustering number Copt are as follows: use k mean cluster (k-
Means) algorithm repeatedly clusters training data respectively, and the silhouette coefficient clustered each time, choosing are calculated according to cluster result
The corresponding cluster numbers of maximum silhouette coefficient are selected as optimum clustering number Copt.
Specifically, in step S5, the data that use are the sample data and often of training set whole when DBN model training
The corresponding cluster cluster number of a sample.
Specifically, in step S6, the data that the DBN submodel uses when training are in cluster corresponding with DBN submodel
Sample data and classification number;The classification number corresponds to the type of circuit breaker failure.
Specifically, test sample data are first inputted the DBN model, are tested by the diagnostic method in test
The cluster number recognition result of sample, i.e. the selection label of DBN submodel;The test sample is inputted again corresponding with the label
DBN submodel obtains the last diagnostic result of the test sample.
Specifically, the multilayer DBN model shares two layers, and first layer is 1 DBN model, and the second layer has n DBN submodule
Type;The DBN model is named as DBN-FL;The DBN submodel is named as DBN-SL1~DBN-SLn;Wherein, n=Copt.
Compared with prior art, the beneficial effects of the present invention are: the present invention is real by k means clustering algorithm and silhouette coefficient
The Rational Classification of existing fault sample, determines optimum clustering number, establishes multilayer DBN model according to optimum clustering number, and determine second
The quantity of layer DBN submodel;Compared to traditional DBN model, sample is more nearly in the submodel that the present invention refines, Ke Yiyou
Effect avoids the appearance of wild extreme point, and preferable diagnosis effect still can be obtained under conditions of data nonbalance, is greatly improved
Diagnostic accuracy.
Detailed description of the invention
Fig. 1 is a kind of flow diagram of the circuit breaker failure diagnostic method based on multilayer DBN model of the embodiment of the present invention;
Fig. 2 is the silhouette coefficient line chart being calculated in the embodiment of the present invention using k means clustering algorithm.
Specific embodiment
Below in conjunction with the attached drawing in the present invention, technical solution of the present invention is clearly and completely described, it is clear that
Described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.Based on the implementation in the present invention
Example, those of ordinary skill in the art's all other embodiment obtained under the conditions of not making creative work belong to
The scope of protection of the invention.
As shown in Figure 1, present embodiments providing a kind of circuit breaker failure diagnostic method based on multilayer DBN model, specifically
The following steps are included:
The fault data of breaker is carried out resampling by S1, data preparation, and each sample retains 800 data, establishes event
Hinder database;Fault data in the Mishap Database is divided into training set and test set, and by the training set and test
The fault data of concentration is standardized;
The method that fault data in the training set and test set is standardized are as follows: use Min-max standard
Change, original sample value be mapped between [0,1], standardization formula is as follows:
Wherein, x ' is the data sample after standardization, and x is primary data sample, xmaxAnd xminFor the maximum of sample data
Value and minimum value;
S2 is preset cluster numbers C (C ∈ 2~10), is carried out respectively using k mean cluster (k-means) algorithm to training data
Repeatedly cluster calculates the silhouette coefficient clustered each time according to cluster result, selects the corresponding cluster numbers of maximum silhouette coefficient
As optimum clustering number Copt;Obtain optimum cluster result and the cluster label of each sample;
S3 establishes first layer DBN model simultaneously using whole fault datas of training set and the corresponding cluster label of each data
Training, the first layer DBN model are named as DBN-FL;
S4 establishes a DBN submodel and training, the DBN submodule using the sample in each cluster of training set respectively
Type shares CoptIt is a, CopA DBN submodel is respectively designated as DBN-SL1~DBN-SLn, wherein n=Copt;
Sample in test set is inputted the DBN model, obtains the cluster label of test sample by S5;
S6 selects corresponding DBN submodel according to the cluster label of the test sample, by the test sample input with
The corresponding DBN submodel of cluster label, obtains diagnostic result.
Specifically, in step S2, as a kind of hard clustering algorithm, k-means main thought is: determining sample data first
Cluster numbers C;Then optional C data are as initial cluster center;Then each data according to Euclidean distance size be placed in
In its most like class;The average value of each new class is recalculated, and using this average value as new cluster centre;It changes repeatedly
In generation, until meeting the condition of convergence, i.e. objective function reaches minimum value;Euclidean distance is defined as follows:
Wherein, dijIndicate i-th of sample and j-th of sample xjkBetween distance.dijSmaller, two samples are closer;Target
Function generallys use square error criterion:
Wherein, E indicates the square error of all clustering objects, xqFor clustering object, miIt is class CiEach clustering object
Average value, calculation formula are as follows:
Wherein: | Ci| indicate class CiClustering object number;
K-means method needs the determining class number C before cluster, selects different C values, therefrom optimal value is selected to be must
Must;The present embodiment determines optimum clustering number using mean profile coefficient;The silhouette coefficient of some sample is defined as:
Wherein, a is sample xiWith the average distance of other samples of same cluster, referred to as condensation degree, b is xiWith the institute of nearest cluster
There are the average distance of sample, referred to as separating degree;Nearest cluster is defined as:
Calculate average distance distance as xi to the cluster of the xi to all samples of some cluster, selection minimum range correspondence
Cluster as nearest cluster, its silhouette coefficient can be calculated for each sample, the wheel for taking its average value S to cluster as this
The value of wide coefficient, S is bigger, and Clustering Effect is better, and therefore, the corresponding C value of maximum S is exactly optimum clustering number.
In step S3, in general, DBN fault diagnosis model uses multiple limited Boltzmann machines (RBM) and one
Top-level categories device stacks;RBM is the life forming model based on energy, is a kind of Boltzmann machine of specific type.
RBM model is made of visual layers v=(v1, v2, v3 ..., vn) and hidden layer h=(h1, h2, h3 ..., hm), visual layers and hidden
It is connectionless between internal node containing layer, while allowing the two-way full connection of node between visual layers and hidden layer.As one
System, the energy function of RBM visual layers and hidden layer unit associations state (v, h) is defined as:
Wherein, θ=(wij,aj,bi) it is RBM parameter, wijIndicate the weight between visible elements j and hidden layer node i, aj
Indicate the biasing of visual node layer j, biIndicate the biasing of hidden layer node i, m and n respectively indicate the section of visual layers and hidden layer
Point number.According to Boltzmann distribution function, the joint probability distribution of (v, h) are as follows:
Wherein, Z (θ) is normalization factor.
But interlayer this for RBM network has a connection and structure connectionless in layer, the state of different nodes in network
As long as distribution is independent from each other that is state that gives certain node layer, the state of adjacent another node layer is just
It can be expressed as:
Thus, it is supposed that oneself knows the state of visible layer or each node of hidden layer, so that it may respectively obtain hidden layer node or can
See the conditional probability function of node layer:
p(hj=1 | v)=δ (bj+∑jviWij)
p(vi=1 | h)=δ (ai+∑ihjWij)
Wherein δ (x) is sigmoid function, is defined as δ (x)=1/ (1+e-x);
To sdpecific dispersion (Contrastive Divergence, CD-k) algorithm, can quickly be improved while keeping precision
Calculating speed.It may be used to determine parameter θ=(wij, aj, bi) value.Greedy algorithm, i.e. successively trained mechanism are used to train
RBM.Only individually some RBM layers of visible layer and hidden layer are trained every time, obtain the optimized parameter of this layer, while
The feature extraction result of current layer is obtained;Then using the output data of this layer as next RBM layers of input sample, continue
Individually train this RBM layers;With this recursion, entire DBN network is wolfishly successively trained.
In order to further illustrate the diagnostic method of the present embodiment, the present embodiment by the processing and modeling to experimental data,
Verify the validity of the fault diagnosis model based on k-means and DBN proposed;For breaker normal condition, divide-shut brake
Control loop failure, operating mechanism jam faults have carried out 27 groups of simulated experiments, each group of experiment according to the difference of severity
Acquire 100 samples.These samples pass through resampling, are normalized to every 800 point data of sample.10 are selected in every kind of fault type
A sample is as test data, other 90 samples are as training data.Therefore, test set includes 270 samples, training set packet
Containing 2430 samples.The fault type for testing simulation is as shown in table 1:
The fault type of 1 fault simulation of table experiment simulation
Modeling tool uses the tool box DeepLearnToolbox-master of MATLAB.By 2430 in training set
Sample is clustered, and sets cluster numbers as 2~10.Its silhouette coefficient is calculated separately, as shown in Figure 2.As can be seen that working as K=6
When, silhouette coefficient reaches maximum value, so optimum clustering number Copt=6.Table 2 show corresponding optimum cluster result.It is each
It include several groups in cluster, these groups number is as shown in table 1.Such as cluster 1 includes the data of 11 groups: 1~8,10,11,
13 total sample numbers are 990.
2 optimum cluster result of table
Cluster number | Classification number | Sample size |
1 | 1~8,10,11,13 | 990 |
2 | 16 | 90 |
3 | 19,20,22,23,25 | 450 |
4 | 21,24,27 | 270 |
5 | 9,12 | 180 |
6 | 14,15,17,18,26 | 450 |
Diagnostic accuracy is tested under the conditions of carrying out data balancing first.Process according to Fig. 1, layering DBN will be carried out two layers
Model training, the first layer model DBN-FL training sample be 2430, sample label is cluster number.Pass through cluster, training sample
It is divided into 6 clusters, therefore, the quantity of the second straton model is 6 (DBN-SL1~DBN-SL6).Training sample is the sample in cluster
This, model parameter difference is as shown in table 3.Wherein, due to data only a kind of in the second cluster, DBN-SL2 does not need to model.
270 test samples are sequentially input into first layer and the second layer model again, it is as shown in table 4 to obtain test result.As can be seen that the
One layer of accuracy of identification has reached 100%, and the erroneous judgement quantity of the second layer is 3.Therefore, total accuracy rate has reached 98.9%.
It reuses identical data and establishes not stratified overall model, input layer is 800 nodes, 27 node of output layer.Middle layer is divided into two
Layer respectively includes 100 nodes.Learning rate is α=0.1, and the number of iterations is 500 times.Accuracy of identification is 96.3%.Visible delamination DBN
Precision improve.
3 data balancing lower leaf DBN model parameter of table
Model | Parameter | Wrong diagnosis quantity |
DBN-FL | Model structure: 800-100-100-6, learning rate: 0.01 | 0 |
DBN-SL1 | Model structure: 800-100-50-11, learning rate: 0.1 | 1 |
DBN-SL2 | Nothing | |
DBN-SL3 | Model structure: 800-150-100-10-5, learning rate: 0.0001 | 0 |
DBN-SL4 | Model structure: 800-150-100-10-3, learning rate: 0.0001 | 2 |
DBN-SL5 | Model structure: 800-100-20-5-2, learning rate: 0.001 | 0 |
DBN-SL6 | Model structure: 800-150-100-10-5, learning rate: 0.01 | 0 |
To test diagnostic accuracy of the model of the present invention under sample imbalance state, 4 kinds of uneven scenes are devised.Scene
1 and scene 2 be separating brake data nonbalance, scene 3 and scene 4 are combined floodgate data nonbalances.There is training data under every kind of scene
Excalation establishes layering DBN model according to process according to Fig. 1.The diagnostic accuracy of basic DBN model simultaneously carries out pair
Than diagnostic test results are as shown in table 4- table 7.
Scene (1)
Under the scene, in separating brake fault data, normal sample and slight sub-gate circuit fault sample missing are serious, and serious point
Lock loop fault sample and drive failure sample then largely retain.Specifically, first kind sample retains 10,2-7
Class sample retains 30, and 8-12 class sample retains 60, and other kinds of sample number is 90.The diagnostic error of multilayered model
Total sample number is 10, accuracy rate 96.3%.The basic DBN model established using same data, accuracy rate are
77.78%.
4 data nonbalance lower leaf DBN model parameter of table-scene (1)
Model | Parameter | Wrong diagnosis quantity |
DBN-FL | Model structure: 800-100-100-6, learning rate: 0.001 | 0 |
DBN-SL1 | Model structure: 800-100-50-11, learning rate: 0.1 | 10 |
DBN-SL2 | Nothing | |
DBN-SL3 | Model structure: 800-100-50-10-5, learning rate: 0.001 | 0 |
DBN-SL4 | Model structure: 800-100-50-10-3, learning rate: 0.0001 | 0 |
DBN-SL5 | Model structure: 800-100-10-2-2, learning rate: 0.0002 | 0 |
DBN-SL6 | Model structure: 800-150-50-10-5, learning rate: 0.0001 | 0 |
Scene (2)
Under the scene, in separating brake fault data, normal sample and slight sub-gate circuit fault sample largely retain, seriously
Sub-gate circuit fault sample and drive failure sample then lack seriously.Specifically, 2-7 class sample retains 60,8-12
Class sample retains 30, and 13-14 class sample retains 10, and other kinds of sample number is 90.The diagnostic error of multilayered model
Total sample number be 20, accuracy rate 92.59%.The basic DBN model established using same data, accuracy rate are
77.41%.
5 data nonbalance lower leaf DBN model parameter of table-scene (2)
Model | Parameter | Wrong diagnosis quantity |
DBN-FL | Model structure: 800-100-50-10-6, learning rate: 0.0001 | 0 |
DBN-SL1 | Model structure: 800-100-50-11, learning rate: 0.1 | 20 |
DBN-SL2 | Nothing | |
DBN-SL3 | Model structure: 800-100-50-10-5, learning rate: 0.001 | 0 |
DBN-SL4 | Model structure: 800-100-50-10-3, learning rate: 0.0001 | 0 |
DBN-SL5 | Model structure: 800-100-10-2-2, learning rate: 0.0002 | 0 |
DBN-SL6 | Model structure: 800-150-100-10-5, learning rate: 0.0001 | 0 |
Scene (3)
Under the scene, in Closing fault data, normal sample and slight sub-gate circuit fault sample missing are serious, and serious point
Lock loop fault sample and drive failure sample then largely retain.Specifically, the 15th class sample retains 10,16-
20 class samples retain 60, and 21-25 class sample retains 30, and other kinds of sample number is 90.The diagnosis of multilayered model is wrong
Total sample number accidentally is 14, accuracy rate 94.81%.The basic DBN model established using same data, accuracy rate are
74.81%.
6 data nonbalance lower leaf DBN model parameter of table-scene (3)
Scene (4)
Under the scene, in Closing fault data, normal sample and slight sub-gate circuit fault sample largely retain, seriously
Sub-gate circuit fault sample and drive failure sample then lack seriously.Specifically, 16-20 class sample retains 60,21-
25 class samples retain 30, and 26-27 class sample retains 10, and other kinds of sample number is 90.The diagnosis of multilayered model is wrong
Total sample number accidentally is 24, accuracy rate 90.37%;The basic DBN model established using same data, accuracy rate are
81.48%.
7 data nonbalance lower leaf DBN model parameter of table-scene (4)
Model | Parameter | Wrong diagnosis quantity |
DBN-FL | Model structure: 800-100-50-10-6, learning rate: 0.0001 | 0 |
DBN-SL1 | Model structure: 800-100-45-11, learning rate: 0.01 | 14 |
DBN-SL2 | Nothing | |
DBN-SL3 | Model structure: 800-100-50-10-5, learning rate: 0.00001 | 0 |
DBN-SL4 | Model structure: 800-100-50-10-3, learning rate: 0.001 | 2 |
DBN-SL5 | Model structure: 800-100-10-2-2, learning rate: 0.0002 | 0 |
DBN-SL6 | Model structure: 800-150-100-10-5, learning rate: 0.00001 | 10 |
In conclusion the present invention weakens the influence of data nonbalance by establishing two layers of DBN model first, diagnosis is improved
Precision, performance are confirmed by the result of data test.
It although an embodiment of the present invention has been shown and described, for the ordinary skill in the art, can be with
A variety of variations, modification, replacement can be carried out to these embodiments without departing from the principles and spirit of the present invention by understanding
And modification, the scope of the present invention is defined by the appended.
Claims (7)
1. a kind of circuit breaker failure diagnostic method based on multilayer DBN model, which comprises the following steps:
The fault data of breaker is carried out resampling by S1, data preparation, and each sample retains 800 data, establishes number of faults
According to library;Fault data in the Mishap Database is divided into training set and test set, and will be in the training set and test set
Fault data be standardized;
S2 presets cluster numbers C, is repeatedly clustered respectively to training data, determine optimum clustering number Copt, obtain optimum cluster
And the cluster label of each sample as a result;
S3 establishes DBN model and training using whole fault datas of training set and the corresponding cluster label of each data;
S4 establishes a DBN submodel and training using the sample in each cluster of training set respectively, and the DBN submodel is total
There is CoptIt is a;
Sample in test set is inputted the DBN model, obtains the cluster label of test sample by S5;
S6 selects corresponding DBN submodel according to the cluster label of the test sample, by test sample input and cluster mark
Corresponding DBN submodel is signed, diagnostic result is obtained.
2. a kind of circuit breaker failure diagnostic method based on multilayer DBN model according to claim 1, which is characterized in that
In step S1, method that the fault data in the training set and test set is standardized are as follows: use Min-max mark
Original sample value is mapped between [0,1] by standardization, and standardization formula is as follows:
Wherein, x ' is the data sample after standardization, and x is primary data sample, xmaxAnd xminFor sample data maximum value and
Minimum value.
3. a kind of circuit breaker failure diagnostic method based on multilayer DBN model according to claim 1, which is characterized in that
In step S2, the preset range of the cluster numbers C is 2~10.
4. a kind of circuit breaker failure diagnostic method based on multilayer DBN model according to claim 1, which is characterized in that
In step S2, the optimum clustering number C is determinedoptMethod are as follows: using k means clustering algorithm training data is carried out respectively it is more
Secondary cluster calculates the silhouette coefficient clustered each time, selects the corresponding cluster numbers of maximum silhouette coefficient as optimum clustering number
Copt。
5. a kind of circuit breaker failure diagnostic method based on multilayer DBN model according to claim 1, which is characterized in that
In step S5, the data that the DBN model uses when training are corresponding poly- for the sample data of training set whole and each sample
Class cluster number.
6. a kind of circuit breaker failure diagnostic method based on multilayer DBN model according to claim 1, which is characterized in that
In step S6, the data that the DBN submodel uses when training are sample data and classification in cluster corresponding with DBN submodel
Number;The classification number corresponds to the type of circuit breaker failure.
7. a kind of circuit breaker failure diagnostic method based on multilayer DBN model according to claim 1, which is characterized in that
Test sample data are first inputted the DBN model in test by the diagnostic method, obtain the cluster number identification knot of test sample
Fruit, i.e. the selection label of DBN submodel;The test sample is inputted into DBN submodel corresponding with the label again, is somebody's turn to do
The last diagnostic result of test sample.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910510449.5A CN110263837A (en) | 2019-06-13 | 2019-06-13 | A kind of circuit breaker failure diagnostic method based on multilayer DBN model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910510449.5A CN110263837A (en) | 2019-06-13 | 2019-06-13 | A kind of circuit breaker failure diagnostic method based on multilayer DBN model |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110263837A true CN110263837A (en) | 2019-09-20 |
Family
ID=67918056
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910510449.5A Pending CN110263837A (en) | 2019-06-13 | 2019-06-13 | A kind of circuit breaker failure diagnostic method based on multilayer DBN model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110263837A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110880369A (en) * | 2019-10-08 | 2020-03-13 | 中国石油大学(华东) | Gas marker detection method based on radial basis function neural network and application |
CN110940518A (en) * | 2019-11-27 | 2020-03-31 | 中北大学 | Aerospace transmission mechanism analysis method based on fault data |
CN111191687A (en) * | 2019-12-14 | 2020-05-22 | 贵州电网有限责任公司 | Power communication data clustering method based on improved K-means algorithm |
CN112083328A (en) * | 2020-09-08 | 2020-12-15 | 广东电网有限责任公司广州供电局 | Fault diagnosis method, system and device for high-voltage circuit breaker |
WO2021175278A1 (en) * | 2020-03-04 | 2021-09-10 | 华为技术有限公司 | Model updating method and related device |
CN113434317A (en) * | 2021-03-11 | 2021-09-24 | 同济大学 | Fault diagnosis method and system based on unbalanced data |
CN113567844A (en) * | 2021-06-11 | 2021-10-29 | 国网浙江省电力有限公司嘉兴供电公司 | High-voltage circuit breaker state diagnosis method and vacuum arc-extinguishing chamber structure |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105759201A (en) * | 2016-03-11 | 2016-07-13 | 江苏镇安电力设备有限公司 | High voltage circuit breaker self-diagnosis method based on abnormal sample identification |
CN105930861A (en) * | 2016-04-13 | 2016-09-07 | 西安西拓电气股份有限公司 | Adaboost algorithm based transformer fault diagnosis method |
CN108020781A (en) * | 2017-12-19 | 2018-05-11 | 上海电机学院 | A kind of circuit breaker failure diagnostic method |
CN108828441A (en) * | 2018-06-12 | 2018-11-16 | 江苏镇安电力设备有限公司 | Fault Diagnosis for HV Circuit Breakers method |
CN109031114A (en) * | 2018-09-29 | 2018-12-18 | 华南理工大学 | A kind of modeling of spring actuator mechanism circuit-breaker and method for diagnosing faults |
CN109086817A (en) * | 2018-07-25 | 2018-12-25 | 西安工程大学 | A kind of Fault Diagnosis for HV Circuit Breakers method based on deepness belief network |
CN109100646A (en) * | 2018-08-17 | 2018-12-28 | 国网江苏省电力有限公司检修分公司 | A kind of Fault Diagnosis for HV Circuit Breakers method |
CN109359668A (en) * | 2018-09-07 | 2019-02-19 | 南京航空航天大学 | A kind of concurrent diagnostic method of aero-engine multiple faults |
-
2019
- 2019-06-13 CN CN201910510449.5A patent/CN110263837A/en active Pending
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105759201A (en) * | 2016-03-11 | 2016-07-13 | 江苏镇安电力设备有限公司 | High voltage circuit breaker self-diagnosis method based on abnormal sample identification |
CN105930861A (en) * | 2016-04-13 | 2016-09-07 | 西安西拓电气股份有限公司 | Adaboost algorithm based transformer fault diagnosis method |
CN108020781A (en) * | 2017-12-19 | 2018-05-11 | 上海电机学院 | A kind of circuit breaker failure diagnostic method |
CN108828441A (en) * | 2018-06-12 | 2018-11-16 | 江苏镇安电力设备有限公司 | Fault Diagnosis for HV Circuit Breakers method |
CN109086817A (en) * | 2018-07-25 | 2018-12-25 | 西安工程大学 | A kind of Fault Diagnosis for HV Circuit Breakers method based on deepness belief network |
CN109100646A (en) * | 2018-08-17 | 2018-12-28 | 国网江苏省电力有限公司检修分公司 | A kind of Fault Diagnosis for HV Circuit Breakers method |
CN109359668A (en) * | 2018-09-07 | 2019-02-19 | 南京航空航天大学 | A kind of concurrent diagnostic method of aero-engine multiple faults |
CN109031114A (en) * | 2018-09-29 | 2018-12-18 | 华南理工大学 | A kind of modeling of spring actuator mechanism circuit-breaker and method for diagnosing faults |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110880369A (en) * | 2019-10-08 | 2020-03-13 | 中国石油大学(华东) | Gas marker detection method based on radial basis function neural network and application |
CN110940518A (en) * | 2019-11-27 | 2020-03-31 | 中北大学 | Aerospace transmission mechanism analysis method based on fault data |
CN110940518B (en) * | 2019-11-27 | 2021-08-24 | 中北大学 | Aerospace transmission mechanism analysis method based on fault data |
CN111191687A (en) * | 2019-12-14 | 2020-05-22 | 贵州电网有限责任公司 | Power communication data clustering method based on improved K-means algorithm |
CN111191687B (en) * | 2019-12-14 | 2023-02-10 | 贵州电网有限责任公司 | Power communication data clustering method based on improved K-means algorithm |
WO2021175278A1 (en) * | 2020-03-04 | 2021-09-10 | 华为技术有限公司 | Model updating method and related device |
CN112083328A (en) * | 2020-09-08 | 2020-12-15 | 广东电网有限责任公司广州供电局 | Fault diagnosis method, system and device for high-voltage circuit breaker |
CN113434317A (en) * | 2021-03-11 | 2021-09-24 | 同济大学 | Fault diagnosis method and system based on unbalanced data |
CN113567844A (en) * | 2021-06-11 | 2021-10-29 | 国网浙江省电力有限公司嘉兴供电公司 | High-voltage circuit breaker state diagnosis method and vacuum arc-extinguishing chamber structure |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110263837A (en) | A kind of circuit breaker failure diagnostic method based on multilayer DBN model | |
CN110929918B (en) | 10kV feeder fault prediction method based on CNN and LightGBM | |
CN109165504B (en) | Power system false data attack identification method based on anti-generation network | |
CN109669087A (en) | A kind of method for diagnosing fault of power transformer based on Multi-source Information Fusion | |
WO2015158198A1 (en) | Fault recognition method and system based on neural network self-learning | |
CN109842373A (en) | Diagnosing failure of photovoltaic array method and device based on spatial and temporal distributions characteristic | |
CN109141847A (en) | A kind of aircraft system faults diagnostic method based on MSCNN deep learning | |
CN105117602A (en) | Metering apparatus operation state early warning method | |
CN108847686B (en) | Photovoltaic inverter fault prediction method | |
CN108133225A (en) | A kind of icing flashover fault early warning method based on support vector machines | |
CN113141008B (en) | Data-driven power distribution network distributed new energy consumption capability assessment method | |
CN108304931A (en) | A kind of Condition-based Maintenance of Substation Equipment method for diagnosing faults | |
CN110118928A (en) | A kind of circuit breaker failure diagnostic method based on Back Propagation Algorithm | |
CN110530650A (en) | Heavy duty gas turbine performance state monitoring method based on generalized regression nerve networks Yu box map analysis | |
CN114169445A (en) | Day-ahead photovoltaic power prediction method, device and system based on CAE and GAN hybrid network | |
CN116562114A (en) | Power transformer fault diagnosis method based on graph convolution neural network | |
Yan et al. | Deep learning based total transfer capability calculation model | |
CN110569888A (en) | transformer fault diagnosis method and device based on directed acyclic graph support vector machine | |
CN112257335B (en) | Oil-immersed transformer fault diagnosis method combining PNN and SVM | |
CN110348489B (en) | Transformer partial discharge mode identification method based on self-coding network | |
CN105701265A (en) | Double-fed wind generator modeling method and apparatus | |
CN115659258A (en) | Power distribution network fault detection method based on multi-scale graph convolution twin network | |
CN109188162A (en) | It is a kind of based on the Transformer condition evaluation that can open up radial base neural net | |
CN105205492B (en) | A kind of fuzzy arest neighbors fusion diagnosis method of therrmodynamic system failure | |
CN114861792A (en) | Complex power grid key node identification method based on deep reinforcement learning |
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 |
Application publication date: 20190920 |
|
RJ01 | Rejection of invention patent application after publication |