CN110348489A - A kind of partial discharge of transformer mode identification method based on autoencoder network - Google Patents
A kind of partial discharge of transformer mode identification method based on autoencoder network Download PDFInfo
- Publication number
- CN110348489A CN110348489A CN201910532414.1A CN201910532414A CN110348489A CN 110348489 A CN110348489 A CN 110348489A CN 201910532414 A CN201910532414 A CN 201910532414A CN 110348489 A CN110348489 A CN 110348489A
- Authority
- CN
- China
- Prior art keywords
- weight
- hidden
- hidden layer
- data
- network
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 48
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 40
- 238000005457 optimization Methods 0.000 claims abstract description 31
- 238000012549 training Methods 0.000 claims description 28
- 239000013598 vector Substances 0.000 claims description 18
- 239000011159 matrix material Substances 0.000 claims description 10
- 238000012360 testing method Methods 0.000 claims description 9
- 238000011156 evaluation Methods 0.000 claims description 8
- 230000035772 mutation Effects 0.000 claims description 8
- 238000003064 k means clustering Methods 0.000 claims description 5
- 239000005367 kimax Substances 0.000 claims description 5
- 238000009826 distribution Methods 0.000 claims description 4
- 238000001914 filtration Methods 0.000 claims description 3
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 claims description 2
- 238000004364 calculation method Methods 0.000 claims description 2
- 239000000203 mixture Substances 0.000 claims description 2
- 210000002569 neuron Anatomy 0.000 claims description 2
- 230000017105 transposition Effects 0.000 claims description 2
- 239000010410 layer Substances 0.000 description 73
- 238000003909 pattern recognition Methods 0.000 description 8
- 238000009413 insulation Methods 0.000 description 6
- 238000006243 chemical reaction Methods 0.000 description 3
- 230000006378 damage Effects 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 239000002245 particle Substances 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- GRYLNZFGIOXLOG-UHFFFAOYSA-N Nitric acid Chemical compound O[N+]([O-])=O GRYLNZFGIOXLOG-UHFFFAOYSA-N 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 210000004556 brain Anatomy 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 230000006866 deterioration Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000017525 heat dissipation Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 229910017604 nitric acid Inorganic materials 0.000 description 1
- 238000012567 pattern recognition method Methods 0.000 description 1
- 238000003825 pressing Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
- 230000006641 stabilisation Effects 0.000 description 1
- 238000011105 stabilization Methods 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/50—Testing of electric apparatus, lines, cables or components for short-circuits, continuity, leakage current or incorrect line connections
- G01R31/62—Testing of transformers
-
- 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
- G06F18/24—Classification techniques
-
- H02J13/0013—
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J3/00—Circuit arrangements for ac mains or ac distribution networks
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Power Engineering (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of partial discharge of transformer mode identification method based on autoencoder network, step includes the following steps: (1) to handle the local discharge signal data of collection;Step 2 selects autoencoder network as network model;Step 3 is trained the weight of autoencoder network using simulated annealing-brainstorming hybrid optimization algorithm;Step 4 is optimized using hidden layer number and hidden node of the brainstorming optimization algorithm to autoencoder network, to obtain trained network;Data to be identified are input in trained network by step 5, are classified using trained network handles identification data;Step 6, the discrimination for calculating data to be identified.The method of the present invention has preferable accuracy of identification, also saves time and manpower, improve the versatility of network.
Description
Technical field
The invention belongs to power equipment monitoring technical fields, are related to a kind of partial discharge of transformer based on autoencoder network
Mode identification method.
Background technique
Electrical equipment is the chief component of electric system, if electrical equipment breaks down, be will result in huge
Loss.With the development of science and technology, the function of electric system becomes complicated, higher automation is realized.Due to electrical equipment
Function and performance improve, and influence factor increases, therefore a possibility that breaking down is consequently increased.Single trouble unit can draw
Chain reaction is played, is not normally functioning so as to cause electric system, therefore electric system must assure that the normal fortune of electrical equipment
Row.
Transformer is one of electric system key equipment, has the function of electric energy conversion and distribution, and cost is high, structure is multiple
It is miscellaneous, if transformer breaks down, it will lead to being not normally functioning for electric system, thus bring inconvenience to people's life, and
And will also result in huge economic loss, therefore the operation of the normal safety of transformer have to the normal operation of electric system it is important
Effect.The reason of transformer breaks down has very much, such as artificial destruction, the influence of environment, the loss of transformer itself
Etc., wherein insulation degradation caused by loss, that is, longtime running of transformer itself is the main reason for transformer breaks down.
Shelf depreciation endangers the insulation performance of transformer serious, is mainly shown as: charged particle to molecular structure into
Row is hit, and destroys even damage insulation;Since particle generates a large amount of heat in knockout process, so that insulation temperature rises sharply;?
A large amount of oxide is generated in discharge process, and oxide meets water and chemical reaction i.e. generation nitric acid occurs, and causes insulation to occur rotten
Lose phenomenon;Schottky linchpin, which is penetrated, leads to oil decomposition, so that its heat dissipation performance declines, these performances are all long-term slow mistakes
Journey.The reason of transformer insulated deterioration and take the form of shelf depreciation.However shelf depreciation type is different, caused by insulate damage
Bad also different, therefore, the type of shelf depreciation can fast and accurately be judged by carrying out pattern-recognition to shelf depreciation, to transformer
Normal operation have vital meaning.
The mass data obtained due to monitoring system now, causes partial discharge monitoring to enter " big data " epoch, and passes
The manual features extracting method and shallow-layer neural network of system are extremely difficult to the identification of local discharge signal or even can not realize, because
This studies and using advanced theoretical and method, feature is extracted from shelf depreciation big data, and accurately carries out identification as change
The new problem that depressor shelf depreciation faces.
Summary of the invention
The purpose of the present invention is to provide a kind of Partial Discharge Pattern Recognition Methods based on autoencoder network, get rid of existing
The yoke that only application experiment data carry out pattern-recognition is studied, in the condition for making full use of complicated live Partial Discharge Data
Under, PD Pattern Recognition is carried out, is allowed to be more suitable for the engineering practice that there is mass data now.
The technical scheme is that a kind of partial discharge of transformer mode identification method based on autoencoder network, is pressed
Implement according to following steps:
Step 1 handles the local discharge signal data of collection
Initial data is handled using non-linear filtering method, eliminates the random signal of interference;Determine the survey of network
The type of data and verify data and its classification is tried, i.e. data are expressed as { (x(1),X(1)),...,(x(m),X(m)) or non-band
Data { the x of label(1),x(2),...,x(m), wherein m is data amount check, and i-th of data is x(i), label X(i)∈{1,
2 ..., k }, k is classification number;Determine the characteristic i.e. dimension of data;
Step 2 selects autoencoder network as network model,
2.1) do not have sort feature from code machine due to storehouse, therefore by storehouse from code machine in conjunction with classifier, building
New autoencoder network;
2.2) objective function of autoencoder network is determined;
Step 3 is trained the weight of autoencoder network using simulated annealing-brainstorming hybrid optimization algorithm, tool
Body process is as follows:
3.1) simulated annealing-brainstorming hybrid optimization algorithm parameter is set
Algorithm parameter specifically includes that initial individuals number NP, maximum number of iterations KImax, probability parameter P1、P2、P3、P4, gather
The number n_c of class, initial temperature t0;
3.2) the NP random distributions for meeting constraint condition are generated according to initialization hidden layer number, hidden node and weight formula
Hidden layer number, hidden node and weight;
3.3) it generates and updates NP weight;
Step 4 is optimized using hidden layer number and hidden node of the brainstorming optimization algorithm to autoencoder network, specifically
Process is as follows:
4.1) it generates and updates NP new hidden layer numbers and hidden node;
4.2) it is iterated Optimum search, when reaching the convergence precision or maximum number of iterations that set, is exported optimal
Hidden layer number, hidden node and corresponding weight, to obtain trained network;
Data to be identified are input in trained network by step 5, using trained network handles identify data into
Row classification;
Step 6, the discrimination for calculating data to be identified
In formula (15), A is the correct classification number of data to be identified, and s is the number of overall data to be identified, can be intuitive
The discrimination for calculating local discharge signal data.
The invention has the advantages that feature extraction and classification are carried out to collected Partial Discharge Data, to effective
It solves the problems, such as PD Pattern Recognition, specifically includes:
1) due to the complexity of shelf depreciation scene mass data, Partial Discharge Data is instructed using autoencoder network
Practice, improves the accuracy of identification of shelf depreciation.
2) autoencoder network in the training process, using simulated annealing-brainstorming hybrid optimization algorithm to network hidden layer
Several and hidden node and corresponding weight are trained, so that accuracy of identification is optimal, and also improve the general of network
Property.
Detailed description of the invention
Fig. 1 is the flow chart of the method for the present invention;
Fig. 2 is the autoencoder network structure diagram in the present invention;
Fig. 3 is the single layer automatic coding machine schematic diagram in the present invention;
Fig. 4 is flow chart of the simulated annealing-brainstorming hybrid optimization algorithm in the present invention to Weight Training;
Fig. 5 is the K-means clustering method flow chart in the present invention;
Fig. 6 is the flow chart of the selection weight in the present invention;
Fig. 7 is optimized flow chart of the brainstorming optimization algorithm in the present invention to hidden layer number and hidden node.
Specific embodiment
The following describes the present invention in detail with reference to the accompanying drawings and specific embodiments.
The present invention is based on the partial discharge of transformer mode identification method of autoencoder network, main includes establishing from coding net
Network model simultaneously solves network parameter using simulated annealing-brainstorming hybrid optimization algorithm, as shown in Figure 1, specifically pressing
Implement according to following steps:
Step 1 handles the local discharge signal data of collection,
It include the interference such as noise in the magnanimity shelf depreciation initial data being collected into, in analysis, pre- using before needing to carry out
Processing is to ensure the stabilizations, reliable of data.This step is handled initial data using non-linear filtering method, eliminates interference
Random signal.
Determine that the test data of network and the type of verify data and its classification, i.e. data are expressed as { (x(1),X(1)),...,(x(m),X(m)) or non-tape label data { x(1),x(2),...,x(m), wherein m is data amount check, i-th of data
For x(i), label X(i)∈ { 1,2 ..., k }, k are classification number;Determine the characteristic i.e. dimension of data;
Step 2 selects autoencoder network as network model,
2.1) do not have sort feature from code machine due to storehouse, therefore by storehouse from code machine in conjunction with classifier, building
New autoencoder network, as shown in Fig. 2, classifier uses softmax classifier;
2.2) objective function of autoencoder network is determined,
As shown in figure 3, in the training process, input data x is mapped in hidden layer by network model first, to obtain hidden layer
Feature y, this part are known as encoder;Then y is mapped to output layer by next layer network, obtains output data z, this part is known as
Decoder;The two parts are indicated with mathematical formulae are as follows:
In formula (1), W1It is the weight matrix of input layer and hidden layer, W2It is the weight matrix of hidden layer and output layer, b is hidden layer
Bias vector, d be output layer bias vector, Sf、SgIt is sigmoid function shown in formula (2), it may be assumed that
By the weight matrix W between input layer and hidden layer1It is taken as the transposition W ' of the weight matrix of hidden layer and output layer2, thus
Reduce parameter, it may be assumed that
W1=W '2=W (3)
Therefore autoencoder network parameter becomes three, respectively weight W, the bias vector b of hidden layer and the biasing of output layer
Vector d;
Training objective is the difference minimized between output and input, it may be assumed that
In formula (4), z is adjusted in the case where x is given by W, b, d, and c (x, z) is the training objective of each training sample,
Then total training objective are as follows:
In formula (5), C (x, z) is total training objective, and m is the quantity of training sample,
Classifier in this step uses softmax classifier, when training sample set is combined into { (x(1),X(1)),...,(x(m),X(m)), wherein m is data amount check, and i-th of data is x(i), label X(i)∈ { 1,2 ..., k }, k are classification number, then
The hypothesis that softmax is returned are as follows:
Where it is assumed that vector hθ(x(i)) each element p (X(i)=j | x(i);θ) representative sample x(i)Belong to jth class
Probability, the element of vector and be 1, θ1,θ2,...,θkIt is classifier parameters vector, these vectors is write as to the form of matrix:
The then cost function of softmax classifier is defined as:
In formula (8), m is data amount check;1 { } indicated indicator function, when the expression formula in bracket is true, then indicator function
Value is 1, is otherwise 0;The subsequent calculating formula of plus sige is weight attenuation term, is to solve numerical value brought by parameter redundancy and ask
Topic, wherein λ is weight attenuation coefficient,
In order to improve the discrimination of classification, evaluation function is redesigned, so the objective function of autoencoder network is set again
It is set to:
E=η C (x, y)+β J (θ) (9)
In formula (9), η is coefficient of the storehouse from code machine objective function, and β is the coefficient of softmax classifier cost function,
C (x, y) is total training objective of the storehouse from code machine, and J (θ) is the cost function of softmax classifier;
Step 3 is trained the weight of autoencoder network using simulated annealing-brainstorming hybrid optimization algorithm,
This step improves i.e. simulation to brainstorming optimization algorithm and moves back on the basis of brainstorming optimization algorithm
Fire-brainstorming hybrid optimization algorithm recycles simulated annealing-brainstorming hybrid optimization algorithm to the weight of autoencoder network
It is trained, improves the recognition capability of network.
Referring to Fig. 4, simulated annealing-brainstorming hybrid optimization algorithm is trained the weight of autoencoder network, specifically
Process is as follows:
3.1) simulated annealing-brainstorming hybrid optimization algorithm parameter is set,
Algorithm parameter specifically includes that initial individuals number NP, maximum number of iterations KImax, probability parameter P1、P2、P3、P4, gather
The number n_c of class, initial temperature t0;
3.2) the NP random distributions for meeting constraint condition are generated according to initialization hidden layer number, hidden node and weight formula
Hidden layer number, hidden node and weight,
Initialize hidden layer number, hidden node and weight formula are as follows:
3.2.a hidden layer number and hidden node) are initialized,
In formula (10), LiIndicate i-th of hidden layer number, NiIndicate i-th of the number of hidden nodes, max_L=10, max_N=
300, RiIndicate i-th of hidden layer number and the number of hidden nodes composition row vector, randint () indicate within the limits prescribed with
Machine integer,
3.2.b corresponding weight) is initialized according to hidden layer number and hidden node,
In formula (11), n is input layer number, and q is hidden layer neuron number,The scale of wherein every layer of n, q, W, b, d are as shown in table 1, Wi,
bi,diIt is randomly generated using decimal coded mode;riIt is i-th of weight;It is j-th of solution of i-th of weight;Rand () is
(0,1) random number between;
Table 1, the weight scale for initializing every layer
3.3) it generates and updates NP weight, detailed process are as follows:
3.3.a NP weight) is divided into n_c in autoencoder network purpose-function space using K-means clustering algorithm
The step of class, wherein K-means is clustered, is as shown in Figure 5;
3.3.b weight) is selected,
As shown in fig. 6, random value is generated between (0,1), if the value is less than probability parameter P1, then with probability parameter P2
It randomly chooses a cluster centre and realizes right value update, detailed process are as follows: the random value for generating (0,1), if the random value is small
In probability parameter P3, then select cluster centre and add a random value to generate new weight;Otherwise, it is selected at random from this cluster
It selects a weight and adds a random value to generate new weight;
If the value is greater than probability parameter P1, two classes are randomly choosed to generate new weight, renewal process are as follows: generate (0,
1) random value;If it is less than probability parameter P4, the two cluster centres, which merge, adds a random value to generate new weight;It is no
Then, two random weight number combinings of selection add a random value to generate new weight from two clusters of selection;
3.3.c mutation operation) is carried out to weight,
The calculating formula of weight mutation operation is as follows:
In formula (12),D dimension in weight after indicating variation;Indicate d dimension in the weight for being used to update;ξ
Indicate weight coefficient value when generating new weight;N (μ, σ) indicate mean value be μ, the gaussian random function that variance is σ;
In formula (13), logsig () indicates logarithm S type function;KImaxIndicate greatest iteration number;Iter indicates current and changes
Algebra;K indicates to change the slope of logsig () function;Random () indicates the random number between (0,1);
3.3.d) right value update,
The corresponding target function value of new weight generated is solved using the target function type (9) of autoencoder network, evaluation becomes
The weight of different front and back retains optimal weight;
3.3.e) if meeting Metropolis acceptance criterion, i.e., with
Canon of probability, wherein E (rj) it is rjObjective function, taFor a times temperature, 3.3.h is gone to step);Otherwise, it goes to step
3.3.f);
3.3.f) pass through rj=rnew+ rand (1, Dim) generates new weight;Dim is the dimension of weight;rnewIt is mutation operation
Weight afterwards;
3.3.g) ifThen rnew=rj;Otherwise, r is enablednew=
rnew, go to step 3.3.e);
3.3.h t) is updateda, a=a+1;
3.3.i) reach when maximum reaches the number of iterations and export best initial weights;Otherwise, 3.3.a is gone to step);
Step 4 is optimized using hidden layer number and hidden node of the brainstorming optimization algorithm to autoencoder network,
In step 3, simulated annealing-brainstorming hybrid optimization algorithm is the training to network weight, and this step 4 is
Optimization with brainstorming optimization algorithm to network hidden layer number and hidden node, two algorithms do not conflict.It is only true in network
Determine hidden layer number and hidden node, just can determine that weight, but the hidden layer number and hidden node of network become, weight also becomes accordingly
?.But calculating target function has to weight.So in this process, first initializing hidden layer number and hidden node, then
Reinitialize weight, next under the conditions of determining hidden layer number and hidden node, obtains optimal weight, is exactly to utilize simulation
Annealing-brainstorming hybrid optimization algorithm trains to obtain best initial weights.Then optimize hidden layer number and hidden node, be to utilize head
Brain storm optimization algorithm obtains optimal hidden layer number and hidden node, and optimal hidden layer number and hidden node obtain, then again
It trains to obtain the best initial weights under the hidden layer number and hidden node using simulated annealing-brainstorming hybrid optimization algorithm.
This step optimizes the hidden layer number and hidden node of network, obtains on the basis of brainstorming optimization algorithm
To optimal hidden layer number and hidden node and its corresponding best initial weights, flow chart is as shown in fig. 7, detailed process is as follows:
4.1) it generates and updates NP new hidden layer numbers and hidden node, detailed process are as follows:
4.1.a NP hidden layer number and hidden node) are divided into 2 classes using K-means clustering algorithm;
4.1.b hidden layer number and hidden node) are selected,
Random value is generated between (0,1), if the random value is less than probability parameter P1, then with probability parameter P2Random choosing
It selects a cluster centre and realizes that hidden layer number and hidden node update, detailed process are as follows: the random value for generating (0,1), if the value
Less than probability parameter P3, then select cluster centre and add a random value to generate new hidden layer number and hidden node;Otherwise, from this
An individual is randomly choosed in a cluster and adds a random value to generate new hidden layer number and hidden node;
If the value is greater than probability parameter P1, randomly choose two classes to generate new hidden layer number and hidden node, it is updated
Journey are as follows: generate (0,1) random value;If the random value is less than probability parameter P4, the two cluster centres merge plus one random
Value generates new hidden layer number and hidden node;Otherwise, two random hidden layer numbers and hidden layer are selected from two clusters of selection
Node, which merges, adds a random value to generate new hidden layer number and hidden node, the probability parameter P in this step1、P2、P3、P4With
Step 3.3b) selection weight in probability parameter P1、P2、P3、 P4It is consistent;
4.1.c mutation operation) is carried out to hidden layer number and hidden node,
Hidden layer number and the formula of hidden node variation are as follows:
In formula (14),D dimension in hidden layer number and hidden node after indicating variation;It indicates to be used to update
Hidden layer number and hidden node in d dimension;ξ indicates weight coefficient value when generating new hidden layer number and hidden node, calculation method
Same formula (13);
4.1.d) hidden layer number and the corresponding weight of hidden node are initialized, carry out weight using step 3.2.b)
Initialization;
4.1.e hidden layer number and the corresponding weight of hidden node) are updated, carries out right value update using step 3.3);
4.1.f) according to autoencoder network objective function, the hidden layer number and hidden node of evaluation variation front and back retain identification
Rate high hidden layer number and hidden node;
4.2) it is iterated Optimum search, when reaching the convergence precision or maximum number of iterations that set, is exported optimal
Hidden layer number, hidden node and corresponding weight, to obtain trained network;
Data to be identified are input in trained network by step 5, using trained network handles identify data into
Row classification;
Step 6, the discrimination for calculating data to be identified:
In formula (15), A is the correct classification number of data to be identified, and s is the number of overall data to be identified, can be intuitive
The discrimination for calculating local discharge signal data.
Embodiment
For being collected into Partial Discharge Data in live transformer, the implementation process of the method for the present invention is illustrated.
According to the insulation system of inside transformer, shelf depreciation is mainly divided into suspended discharge, needle plate electric discharge, bubble-discharge and puts along face
Electricity.
Step 1, the data of on-site collection are handled, determines that training data and test data wherein put every kind of part
Electric type acquires 700 data, and training data has 2000, and test sample has 500, and sample dimension is 400.
Step 2, selection autoencoder network model to Partial Discharge Data carry out pattern-recognition, storehouse automatic coding machine and
Softmax classifier is combined into autoencoder network, and determines objective function, that is, formula (7) of network.
Step 3, network weight is trained using simulated annealing-brainstorming hybrid optimization algorithm.
Algorithm relative parameters setting are as follows: population scale NP=30;Cluster number n_c=2;And probability parameter is set as P1=
0.2;P2=0.8;P3=0.4;P4=0.5;Maximum number of iterations is 50;The maximum number of iterations of Weight Training is 2000 times,η=0.00002;β=20.
Initial hidden layer number, hidden node and corresponding weight are generated, mainly according to the step in specific embodiment
3.2) principle described in is configured.
Weight is updated, is mainly trained according to the step 3.3) in specific embodiment is described, in hidden layer number and hidden layer
Node obtains best initial weights in the case where determining.
Hidden layer number and hidden node are updated, mainly optimizes, obtains according to the step 4) in specific embodiment is described
Optimal hidden layer number, hidden node and corresponding weight, so that it is determined that the parameter of network.
Step 4, test data is brought into trained network and carries out pattern-recognition.
More clearly to verify original autoencoder network (AE), BSO (brainstorming optimization algorithm) autoencoder network
(BSO-AE) and the performance of SABSO (simulated annealing-brainstorming hybrid optimization algorithm) autoencoder network (SABSO-AE),
AE, BSO-AE and SABSO-AE are with the evaluation function as BSO-AE and SABSO-AE only by the error letter of storehouse automatic coding machine
BSO-AE1 and SABSO-AE1 is compared when number is constituted.
As can be seen that discrimination ratio BSO-AE, the BSO-AE1 of AE in training data and test data from table 2 and table 3
Discrimination it is high, the discrimination of SABSO-AE1 on the training data is lower than the discrimination of AE, but AE is above in test data
Discrimination, and discrimination of the SABSO-AE in training data and test data is all better than other methods, so SABSO-AE
It can preferably identify Partial Discharge Data, improve the discrimination of Partial Discharge Data, show method proposed by the present invention
It is effective.
The operation result of table 2, training data on different evaluation function
The operation result of table 3, test data on different evaluation function
By the above results as it can be seen that the method for the present invention, getting rid of existing research, only application experiment data carry out pattern-recognition
Yoke carries out PD Pattern Recognition, is allowed to more applicable under conditions of making full use of complicated live Partial Discharge Data
In the engineering practice that there is mass data now;And it is adjusted using simulated annealing-brainstorming hybrid optimization algorithm intelligence
The parameter of autoencoder network.Compared with other algorithms, the method for the present invention has preferable accuracy of identification, and saves time and people
Power improves the versatility of network.
Claims (5)
1. a kind of partial discharge of transformer mode identification method based on autoencoder network, which is characterized in that according to the following steps
Implement:
Step 1 handles the local discharge signal data of collection,
Initial data is handled using non-linear filtering method, eliminates the random signal of interference;Determine the test number of network
According to and verify data and its classification type, i.e. data are expressed as { (x(1),X(1)),...,(x(m),X(m)) or non-tape label
Data { x(1),x(2),...,x(m), wherein m is data amount check, and i-th of data is x(i), label X(i)∈{1,2,...,
K }, k is classification number;Determine the characteristic i.e. dimension of data;
Step 2 selects autoencoder network as network model,
2.1) do not have sort feature from code machine due to storehouse, therefore by storehouse from code machine in conjunction with classifier, it constructs new
Autoencoder network;
2.2) objective function of autoencoder network is determined;
Step 3 is trained the weight of autoencoder network using simulated annealing-brainstorming hybrid optimization algorithm, specific mistake
Journey is as follows:
3.1) simulated annealing-brainstorming hybrid optimization algorithm parameter is set,
Algorithm parameter specifically includes that initial individuals number NP, maximum number of iterations KImax, probability parameter P1、P2、P3、P4, of cluster
Number n_c, initial temperature t0;
3.2) the NP random distribution hidden layers for meeting constraint condition are generated according to initialization hidden layer number, hidden node and weight formula
Number, hidden node and weight;
3.3) it generates and updates NP weight;
Step 4 is optimized, detailed process using hidden layer number and hidden node of the brainstorming optimization algorithm to autoencoder network
It is as follows:
4.1) it generates and updates NP new hidden layer numbers and hidden node;
4.2) it is iterated Optimum search, when reaching the convergence precision or maximum number of iterations that set, is exported optimal hidden
The number of plies, hidden node and corresponding weight, to obtain trained network;
Data to be identified are input in trained network by step 5, are divided using trained network handles identification data
Class;
Step 6, the discrimination for calculating data to be identified:
In formula (15), A is the correct classification number of data to be identified, and s is the number of overall data to be identified, can intuitively be counted
Calculate the discrimination of local discharge signal data.
2. the partial discharge of transformer mode identification method according to claim 1 based on autoencoder network, feature exist
In, in the step 2.2),
In the training process, input data x is mapped in hidden layer by network model first, and to obtain hidden layer feature y, this part claims
For encoder;Then y is mapped to output layer by next layer network, obtains output data z, this part is known as decoder;The two
Part is indicated with mathematical formulae are as follows:
In formula (1), W1It is the weight matrix of input layer and hidden layer, W2It is the weight matrix of hidden layer and output layer, b is the inclined of hidden layer
Vector is set, d is the bias vector of output layer, Sf、SgIt is sigmoid function shown in formula (2), it may be assumed that
By the weight matrix W between input layer and hidden layer1It is taken as the transposition W ' of the weight matrix of hidden layer and output layer2, to reduce
Parameter, it may be assumed that
W1=W '2=W (3)
Therefore autoencoder network parameter becomes three, respectively weight W, the bias vector b of hidden layer and the bias vector of output layer
d;
Training objective is the difference minimized between output and input, it may be assumed that
In formula (4), z is adjusted in the case where x is given by W, b, d, and c (x, z) is the training objective of each training sample, then always
Training objective are as follows:
In formula (5), C (x, z) is total training objective, and m is the quantity of training sample,
Classifier in this step uses softmax classifier, when training sample set is combined into { (x(1),X(1)),...,(x(m),X(m)), wherein m is data amount check, and i-th of data is x(i), label X(i)∈ { 1,2 ..., k }, k are classification number, then
The hypothesis that softmax is returned are as follows:
Where it is assumed that vector hθ(x(i)) each element p (X(i)=j | x(i);θ) representative sample x(i)Belong to the probability of jth class,
The element of vector and be 1, θ1,θ2,...,θkIt is classifier parameters vector, these vectors is write as to the form of matrix:
The then cost function of softmax classifier is defined as:
In formula (8), m is data amount check;1 { } indicated indicator function, and when the expression formula in bracket is very, then indicator function value is
1, it is otherwise 0;The subsequent calculating formula of plus sige is weight attenuation term, and wherein λ is weight attenuation coefficient,
The objective function of autoencoder network is reset are as follows:
E=η C (x, y)+β J (θ) (9)
In formula (9), η is coefficient of the storehouse from code machine objective function, and β is the coefficient of softmax classifier cost function, C (x,
It y) is total training objective of the storehouse from code machine, J (θ) is the cost function of softmax classifier.
3. the partial discharge of transformer mode identification method according to claim 1 based on autoencoder network, feature exist
In, in the step 3.2), initialization hidden layer number, hidden node and weight formula are as follows:
3.2.a hidden layer number and hidden node) are initialized
In formula (10), LiIndicate i-th of hidden layer number, NiIndicate i-th of the number of hidden nodes, max_L=10, max_N=300, RiTable
Show the row vector of i-th of hidden layer number and the number of hidden nodes composition, randint () indicates random integers within the limits prescribed;
3.2.b corresponding weight) is initialized according to hidden layer number and hidden node,
In formula (11), n is input layer number, and q is hidden layer neuron number,The scale of wherein every layer of n, q, W, b, d are as shown in table 1, Wi,
bi,diIt is randomly generated using decimal coded mode;riIt is i-th of weight;It is j-th of solution of i-th of weight;Rand () is
(0,1) random number between;
Table 1, the weight scale for initializing every layer
4. the partial discharge of transformer mode identification method according to claim 1 based on autoencoder network, feature exist
In, in the step 3.3), detailed process are as follows:
3.3.a NP weight) is divided into n_c class in autoencoder network purpose-function space using K-means clustering algorithm;
3.3.b weight) is selected,
Random value is generated between (0,1), if the value is less than probability parameter P1, then with probability parameter P2Random selection one poly-
Realize right value update, detailed process in class center are as follows: the random value of (0,1) is generated, if the random value is less than probability parameter P3, then
Selection cluster centre simultaneously adds a random value to generate new weight;Otherwise, a weight is randomly choosed from this cluster and is added
One random value generates new weight;
If the value is greater than probability parameter P1, two classes are randomly choosed to generate new weight, renewal process are as follows: it is random to generate (0,1)
Value;If it is less than probability parameter P4, the two cluster centres, which merge, adds a random value to generate new weight;Otherwise, from choosing
Two random weight number combinings of selection add a random value to generate new weight in two clusters selected;
3.3.c mutation operation) is carried out to weight, the calculating formula of weight mutation operation is as follows:
In formula (12),D dimension in weight after indicating variation;Indicate d dimension in the weight for being used to update;ξ is indicated
Generate weight coefficient value when new weight;N (μ, σ) indicate mean value be μ, the gaussian random function that variance is σ;
In formula (13), logsig () indicates logarithm S type function;KImaxIndicate greatest iteration number;Iter indicates current number of iterations;
K indicates to change the slope of logsig () function;Random () indicates the random number between (0,1);
3.3.d) right value update,
The corresponding target function value of new weight generated is solved using the target function type (9) of autoencoder network, before evaluation variation
Weight afterwards retains optimal weight;
3.3.e) if meeting Metropolis acceptance criterion, i.e., withCanon of probability,
Wherein E (rj) it is rjObjective function, taFor a times temperature, 3.3.h is gone to step);Otherwise, 3.3.f is gone to step);
3.3.f) pass through rj=rnew+ rand (1, Dim) generates new weight;Dim is the dimension of weight;rnewAfter being mutation operation
Weight;
3.3.g) ifThen rnew=rj;Otherwise, r is enablednew=rnew, turn
Step 3.3.e);
3.3.h t) is updateda, a=a+1;
3.3.i) reach when maximum reaches the number of iterations and export best initial weights;Otherwise, 3.3.a is gone to step).
5. the partial discharge of transformer mode identification method according to claim 1 based on autoencoder network, feature exist
In, in the step 4.1), detailed process are as follows:
4.1.a NP hidden layer number and hidden node) are divided into 2 classes using K-means clustering algorithm;
4.1.b hidden layer number and hidden node) are selected,
Random value is generated between (0,1), if the random value is less than probability parameter P1, then with probability parameter P2Random selection one
A cluster centre realizes that hidden layer number and hidden node update, detailed process are as follows: the random value for generating (0,1), if the value is less than
Probability parameter P3, then select cluster centre and add a random value to generate new hidden layer number and hidden node;Otherwise, poly- from this
An individual is randomly choosed in class and adds a random value to generate new hidden layer number and hidden node;
If the value is greater than probability parameter P1, two classes are randomly choosed to generate new hidden layer number and hidden node, renewal process are as follows:
Generate (0,1) random value;If the random value is less than probability parameter P4, the two cluster centres, which merge, adds a random value to produce
Raw new hidden layer number and hidden node;Otherwise, two random hidden layer numbers and hidden node are selected to close from two clusters of selection
And add a random value to generate new hidden layer number and hidden node;
4.1.c mutation operation) is carried out to hidden layer number and hidden node,
Hidden layer number and the formula of hidden node variation are as follows:
In formula (14),D dimension in hidden layer number and hidden node after indicating variation;Indicate the hidden layer for being used to update
D dimension in several and hidden node;ξ indicates weight coefficient value when generating new hidden layer number and hidden node, the same formula of calculation method
(13);
4.1.d) hidden layer number and the corresponding weight of hidden node are initialized, carry out the initial of weight using step 3.2.b)
Change;
4.1.e hidden layer number and the corresponding weight of hidden node) are updated, carries out right value update using step 3.3);
4.1.f) according to autoencoder network objective function, it is high to retain discrimination for the hidden layer number and hidden node of evaluation variation front and back
Hidden layer number and hidden node.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910532414.1A CN110348489B (en) | 2019-06-19 | 2019-06-19 | Transformer partial discharge mode identification method based on self-coding network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910532414.1A CN110348489B (en) | 2019-06-19 | 2019-06-19 | Transformer partial discharge mode identification method based on self-coding network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110348489A true CN110348489A (en) | 2019-10-18 |
CN110348489B CN110348489B (en) | 2021-04-06 |
Family
ID=68182398
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910532414.1A Active CN110348489B (en) | 2019-06-19 | 2019-06-19 | Transformer partial discharge mode identification method based on self-coding network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110348489B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111142001A (en) * | 2020-01-10 | 2020-05-12 | 三峡大学 | Transformer multi-source partial discharge mode identification method based on parallel characteristic domain |
CN112327219A (en) * | 2020-10-29 | 2021-02-05 | 国网福建省电力有限公司南平供电公司 | Distribution transformer fault diagnosis method with automatic feature mining and automatic parameter optimization |
Citations (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102156245A (en) * | 2011-03-11 | 2011-08-17 | 太原理工大学 | On-line fault diagnosis and prewarning method of mine high-voltage cable |
CN103323755A (en) * | 2013-06-17 | 2013-09-25 | 广东电网公司电力科学研究院 | Method and system for recognition of GIS ultrahigh frequency partial discharge signal |
CN104037757A (en) * | 2014-05-20 | 2014-09-10 | 西安理工大学 | Brainstorming-based thermal power plant economic environment scheduling method |
CN105334436A (en) * | 2015-10-30 | 2016-02-17 | 山东电力研究院 | Cross-linked cable partial discharge mode identification method based on SOM-BP combined neural network |
CN106503689A (en) * | 2016-11-14 | 2017-03-15 | 哈尔滨理工大学 | Neutral net local discharge signal mode identification method based on particle cluster algorithm |
CN108040324A (en) * | 2017-11-16 | 2018-05-15 | 南方科技大学 | A kind of localization method and alignment system of survival capsule robot |
CN108399105A (en) * | 2018-02-27 | 2018-08-14 | 天津大学 | A kind of Method for HW/SW partitioning based on improvement brainstorming algorithm |
CN108573225A (en) * | 2018-03-30 | 2018-09-25 | 国网天津市电力公司电力科学研究院 | A kind of local discharge signal mode identification method and system |
CN108694473A (en) * | 2018-06-15 | 2018-10-23 | 常州瑞信电子科技有限公司 | Building energy consumption prediction technique based on RBF neural |
CN108957251A (en) * | 2018-05-18 | 2018-12-07 | 深圳供电局有限公司 | A kind of cable connector Partial Discharge Pattern Recognition Method |
CN109102012A (en) * | 2018-07-30 | 2018-12-28 | 上海交通大学 | A kind of defect identification method and system of local discharge signal |
CN109188211A (en) * | 2018-07-30 | 2019-01-11 | 上海交通大学 | A kind of high voltage equipment insulation method for diagnosing faults and system |
CN109375116A (en) * | 2018-08-09 | 2019-02-22 | 上海国际汽车城(集团)有限公司 | A kind of battery system exception battery discriminating method based on self-encoding encoder |
CN109388858A (en) * | 2018-09-17 | 2019-02-26 | 西安航空电子科技有限公司 | Nonlinear transducer bearing calibration based on brainstorming optimization algorithm |
-
2019
- 2019-06-19 CN CN201910532414.1A patent/CN110348489B/en active Active
Patent Citations (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102156245A (en) * | 2011-03-11 | 2011-08-17 | 太原理工大学 | On-line fault diagnosis and prewarning method of mine high-voltage cable |
CN103323755A (en) * | 2013-06-17 | 2013-09-25 | 广东电网公司电力科学研究院 | Method and system for recognition of GIS ultrahigh frequency partial discharge signal |
CN104037757A (en) * | 2014-05-20 | 2014-09-10 | 西安理工大学 | Brainstorming-based thermal power plant economic environment scheduling method |
CN105334436A (en) * | 2015-10-30 | 2016-02-17 | 山东电力研究院 | Cross-linked cable partial discharge mode identification method based on SOM-BP combined neural network |
CN106503689A (en) * | 2016-11-14 | 2017-03-15 | 哈尔滨理工大学 | Neutral net local discharge signal mode identification method based on particle cluster algorithm |
CN108040324A (en) * | 2017-11-16 | 2018-05-15 | 南方科技大学 | A kind of localization method and alignment system of survival capsule robot |
CN108399105A (en) * | 2018-02-27 | 2018-08-14 | 天津大学 | A kind of Method for HW/SW partitioning based on improvement brainstorming algorithm |
CN108573225A (en) * | 2018-03-30 | 2018-09-25 | 国网天津市电力公司电力科学研究院 | A kind of local discharge signal mode identification method and system |
CN108957251A (en) * | 2018-05-18 | 2018-12-07 | 深圳供电局有限公司 | A kind of cable connector Partial Discharge Pattern Recognition Method |
CN108694473A (en) * | 2018-06-15 | 2018-10-23 | 常州瑞信电子科技有限公司 | Building energy consumption prediction technique based on RBF neural |
CN109102012A (en) * | 2018-07-30 | 2018-12-28 | 上海交通大学 | A kind of defect identification method and system of local discharge signal |
CN109188211A (en) * | 2018-07-30 | 2019-01-11 | 上海交通大学 | A kind of high voltage equipment insulation method for diagnosing faults and system |
CN109375116A (en) * | 2018-08-09 | 2019-02-22 | 上海国际汽车城(集团)有限公司 | A kind of battery system exception battery discriminating method based on self-encoding encoder |
CN109388858A (en) * | 2018-09-17 | 2019-02-26 | 西安航空电子科技有限公司 | Nonlinear transducer bearing calibration based on brainstorming optimization algorithm |
Non-Patent Citations (2)
Title |
---|
ZHENGXUAN JIA ET AL.: ""Hybrid brain storm optimisation and simulated annealing algorithm for continuous optimisation problems"", 《INTERNATIONAL JOURNAL OF BIO-INSPIRED COMPUTATION》 * |
ZIJIAN CAO ET AL.: ""An Improved Brain Storm Optimization with Differential Evolution Strategy for Applications of ANNs"", 《MATHEMATICAL PROBLEMS IN ENGINEERING》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111142001A (en) * | 2020-01-10 | 2020-05-12 | 三峡大学 | Transformer multi-source partial discharge mode identification method based on parallel characteristic domain |
CN111142001B (en) * | 2020-01-10 | 2022-04-22 | 三峡大学 | Transformer multi-source partial discharge mode identification method based on parallel characteristic domain |
CN112327219A (en) * | 2020-10-29 | 2021-02-05 | 国网福建省电力有限公司南平供电公司 | Distribution transformer fault diagnosis method with automatic feature mining and automatic parameter optimization |
CN112327219B (en) * | 2020-10-29 | 2024-03-12 | 国网福建省电力有限公司南平供电公司 | Distribution transformer fault diagnosis method with automatic feature mining and parameter automatic optimizing functions |
Also Published As
Publication number | Publication date |
---|---|
CN110348489B (en) | 2021-04-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110929763A (en) | Multi-source data fusion-based mechanical fault diagnosis method for medium-voltage vacuum circuit breaker | |
CN111401755B (en) | Multi-new-energy output scene generation method, device and system based on Markov chain | |
CN105388402A (en) | Cross-linked cable partial discharge mode recognition method based on parameter optimization SVM (Support Vector Machine) algorithm | |
CN106649479A (en) | Probability graph-based transformer state association rule mining method | |
CN113344288B (en) | Cascade hydropower station group water level prediction method and device and computer readable storage medium | |
CN113792754A (en) | Method for processing DGA (differential global alignment) online monitoring data of converter transformer by removing different elements and then repairing | |
CN112485622A (en) | Partial discharge pattern recognition method based on GA-BPNN | |
CN110569888A (en) | transformer fault diagnosis method and device based on directed acyclic graph support vector machine | |
CN110348489A (en) | A kind of partial discharge of transformer mode identification method based on autoencoder network | |
CN111859283A (en) | Scene generation method considering multi-energy charge-time sequence and correlation | |
CN116484299A (en) | Charging pile fault diagnosis method based on integration of gradient lifting tree and multi-layer perceptron | |
CN113762591B (en) | Short-term electric quantity prediction method and system based on GRU and multi-core SVM countermeasure learning | |
CN111428766A (en) | Power consumption mode classification method for high-dimensional mass measurement data | |
Kim et al. | Anomaly detection using clustered deep one-class classification | |
CN112327190B (en) | Method for identifying health state of energy storage battery | |
Pei et al. | The real‐time state identification of the electricity‐heat system based on Borderline‐SMOTE and XGBoost | |
CN113884807A (en) | Power distribution network fault prediction method based on random forest and multi-layer architecture clustering | |
CN116736133A (en) | Early prediction method for capacity degradation track of lithium ion battery in full life cycle | |
CN117034762A (en) | Composite model lithium battery life prediction method based on multi-algorithm weighted sum | |
CN116167465A (en) | Solar irradiance prediction method based on multivariate time series ensemble learning | |
CN115630311A (en) | Wind, photovoltaic and hydrogen scene reduction method considering correlation of uncertain factors | |
Oprea et al. | Electricity load profile calculation using self-organizing maps | |
Shuqi et al. | Research on Abnormal Electrical Behavior Identification Technology Based on Cost Sensitive RAE-AdaBoost Algorithm | |
CN113435628B (en) | Medium-long-term runoff prediction method and system based on linear discriminant analysis and IALO-ELM | |
Wang et al. | A Non-Intrusive Load Decomposition Method Based on Generative Adversarial Networks and Asynchronous Federated 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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |