CN107563414B - A kind of complex device degenerate state recognition methods based on Kohonen-SVM - Google Patents
A kind of complex device degenerate state recognition methods based on Kohonen-SVM Download PDFInfo
- Publication number
- CN107563414B CN107563414B CN201710691675.9A CN201710691675A CN107563414B CN 107563414 B CN107563414 B CN 107563414B CN 201710691675 A CN201710691675 A CN 201710691675A CN 107563414 B CN107563414 B CN 107563414B
- Authority
- CN
- China
- Prior art keywords
- data
- svm
- kohonen
- degenerate state
- sample
- 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.)
- Active
Links
Landscapes
- Image Analysis (AREA)
Abstract
A kind of complex device degenerate state recognition methods based on Kohonen SVM, step are as follows：1st, the normal power data of complex device are acquired；2nd, feature extraction selection dimensionality reduction is carried out to the power data after screening；3rd, multidimensional characteristic vectors data are inputted into Kohonen networks and carries out Unsupervised clustering study；4th, division processing is carried out to degeneration sample set data；5th, on the basis of training sample set, optimizing solution is carried out to the parameter of SVM using PSO methods；6th, SVM model learnings are carried out using the parameter and training set sample data of optimization, obtains degenerate state identification model；7th, the accuracy rate of Classification and Identification model is verified using test set sample data；By above-mentioned steps, the present invention realizes the complex device degenerate state recognition methods based on Kohonen SVM, has obtained the typical degradation state sample collection data of equipment and the degenerate state identification model of complex device.
Description
Technical field
The present invention provides a kind of complex device degenerate state recognition methods based on Kohonen-SVM, it is related to a kind of base
In the realization of the complex device degenerate state recognition methods of Kohonen-SVM, belong to complex device reliability, complex device failure
Diagnostic field.
Background technology
With the continuous improvement of industrialized level, the equipment applied to each industrial department is also increasingly sophisticated, these equipment
Reliability directly affects the development of related industry.Complex device is monitored in real time and identifies the failure classes that may occur
Type has great importance to the reliability of equipment.At present, failure is focused primarily upon for the diagnosis of complex device both at home and abroad
The identification of pattern, but there is no account for the degenerate state of equipment for these researchs.Degenerate state image watermarking is in normal shape
In state data, as time increases, degradation failure is eventually evolved into.Wherein, degradation failure refers to that equipment is chronically at Asia
Health status, performance run down caused failure.Therefore, it is very necessary to the Study of recognition of equipment degenerate state.
The present invention proposes a kind of complex device degenerate state recognition methods based on Kohonen-SVM, and this method passes through Ke Helun
(Kohonen) neutral net carries out unsupervised learning to the normal power data of equipment, obtains degenerate state number among typical
According to.Then using support vector machines (Support Vector Machine, abbreviation SVM) model of parameter optimization to status data
Exercise supervision study, degenerate state sorter model is generated, for the identification of complex device degenerate state.
This method is based on Kohonen neutral nets and SVM technologies, the fusion correlation theories such as accident analysis and troubleshooting,
Methods and techniques are realized, while complex device maintainability is improved, reach improve complex device reliability, security,
The purpose of availability.
The content of the invention
(1) the object of the invention
Usually only consider normal and failure two states currently for the method for fault pattern recognition of complex device, not
The intermediate degenerate state of equipment is accounted for, this degenerate state can be evolved into degradation failure with the accumulation of time, pair set
It is standby to cause serious damage.Therefore the present invention will overcome the deficiencies of the prior art and provide a kind of complexity based on Kohonen-SVM
Equipment degenerate state recognition methods.This method is using the degenerate state method for digging of Kohonen neutral nets to the normal of equipment
Power data carries out unsupervised learning, obtains typical degenerate state data.And on the basis of data, parameter optimization is used
SVM models exercise supervision to status data study, degenerate state sorter model are generated, for the identification of equipment degenerate state.
This method can not only excavate the degenerate state of complex device, and classification can also be identified to degenerate state, be
Fault diagnosis field provides a kind of new solution method, and existing method of fault pattern recognition is innovated.
(2) technical solution
A kind of complex device degenerate state recognition methods based on Kohonen-SVM of the present invention, its step are as follows：
Step 1 is acquired the normal power data of complex device, wherein may be rejected by way of artificial screening
Existing fault data and bad value data；
Step 2 carries out feature extraction-selection-dimensionality reduction to the power data after screening, obtains can be used for the more of input study
Dimensional feature vector data；
Multidimensional characteristic vectors data are inputted into Kohonen networks progress Unsupervised clustering study by step 3, are moved back
Change state sample collection data；
Step 4 carries out division processing to degeneration sample set data, obtains SVM training sample sets data and test sample collection
Data；
Step 5, on the basis of training sample set, using particle group optimizing method (Particle Swarm
Optimization, abbreviation PSO) optimizing solution is carried out to the parameter of SVM；
Step 6 carries out SVM model learnings using the parameter and training set sample data of optimization, obtains effectively identifying and move back
The SVM models of change state, i.e. degenerate state identification model；
Step 7 verifies the accuracy rate of Classification and Identification model using test set sample data.
Wherein, " power data " in step 1, during referring to device action, is set by harvester
Standby real-time current, voltage value, and the characterization equipment state obtained on the basis of these numerical value using computer automatic analysis
Data；" fault data ", when referring to that equipment running process breaks down, the abnormal work(that is detected by comparative analysis
Rate data；" the bad value data ", refer in device action data acquisition, and data value is less than the work(of acquisition regulation threshold
Rate data, the presence of bad value data can influence the accuracy of plant capacity overall data.
Wherein, it is described " feature extraction-selection-dimensionality reduction is carried out to the power data after screening " in step 2, the practice
It is as follows：Time domain data and codomain data are divided into power data, after the completion of division, proceeds by the feature extraction work of data
Make.For power time domain data, this method extraction average, variance, root mean square, kurtosis, inlet and outlet difference, minimax difference,
The parameters such as peak factor, the pulse factor, difference and shape factor are as characteristic；For performance number numeric field data, this method carries
Take the ginsengs such as maximum time value, average value, data points, minimax difference, value median, maximum, time median, mode
Number is as characteristic.After the completion of extraction, using these characteristic construction feature vector spaces, using based on criterion function
Method carry out feature selecting；After feature selecting is completed, by principal component analysis, carry out Feature Dimension Reduction is locally linear embedding into,
The multidimensional characteristic vectors data of acquisition.
Wherein, " multidimensional characteristic vectors data " in step 3, refer to power data by feature extraction-selection-
The data obtained after dimension-reduction treatment, multidimensional characteristic vectors data can be as the input datas of Kohonen networks, for degenerate state
Clustering learning.
Wherein, " the Kohonen networks " in step 3, refers to a kind of Self-organizing Competition type neutral net；The network
Network weight is adjusted by self-organizing feature map, neutral net is made to converge on a kind of expression form, one in this form
Neuron is only to the especially matching or sensitive of certain input pattern；Kohonen networks are comprising before two layers of input layer and competition layer
Present neutral net；Wherein, the 1st layer is input layer, and the number of input layer is consistent with the dimension of input sample vector,
It is m to take input layer number；2nd layer is competition layer, also referred to as output layer, and competition node layer is distributed in two-dimensional array, takes competition layer
Number of nodes is n；It is connected entirely with variable weight between input node and output node, connection weight is Kohonen networks
Study is unsupervised self-organized learning process；By unsupervised learning, different neurons can be to different input moulds in network
Formula is sensitive, realizes that specific neuron can serve as the detector of a certain input pattern in pattern-recognition.After the completion of study, god
Different zones are divided into through member, each region has different response characteristics to input model.
Wherein, in step 3 it is described " multidimensional characteristic vectors data are inputted carried out into Kohonen networks it is unsupervised
Clustering learning ", the practice are as follows：Multidimensional characteristic vectors data are inputted to network, the neuron on network competition layer calculates defeated
Enter the Euclidean distance between sample and competition layer neuron weights, the minimum neuron of distance is triumph neuron.Adjustment
Triumph neuron and adjacent neurons weights make triumph neuron and periphery weights close to the input sample；By repetition training,
The connection weight of final each neuron has certain distribution, and the distribution is all kinds of to representing the similitude tissue between data
Under neuron, make similar neuron that there is similar weight coefficient, inhomogeneous neuron weight coefficient difference is apparent.It should be noted that
It is that in learning process, weights modification learning rate and neuron field are constantly being reduced, so that similar neuron is gradual
It concentrates, completes the automatic cluster of input data, generate different classes of degenerate state sample set data.
Wherein, described " carrying out division processing to degeneration sample set data " in step 4, the practice is as follows：To degenerating
The number of samples of state sample collection data carries out unified, it is specified that the number of samples of each state is consistent, and presses 4:1 ratio will be every
Kind sample is divided into training set and test set；For carrying out SVM modelings, test set data are used for into row degradation shape training set data
The recognition accuracy test of state identification model.
Wherein, " SVM " in step 4, refers to supporting vector machine model, which is the base in statistical theory
A kind of sorter model proposed on plinth, available for pattern classification, linear and nonlinear regression analysis；The original of support vector machines
Reason is given training sample, establishes an Optimal Separating Hyperplane as decision-making curved surface, makes the isolated border between sample positive example and counter-example
Edge maximizes, so as to complete the classification of training sample.
Wherein, " particle group optimizing method " in steps of 5, refers to that a kind of optimization based on swarm intelligence theory is calculated
Method, the algorithm have self-teaching and the two-fold advantage learnt to other people, can find optimal solution in less iterations.
Wherein, in steps of 5 it is described " using particle group optimizing method (Particle Swarm Optimization,
Abbreviation PSO) optimizing solution is carried out to the parameter of SVM ", i.e., " optimizing solution is carried out to the parameter of SVM using PSO methods ", made
Method is as follows：Optimizing solution is carried out to the nuclear parameter g and penalty parameter c of SVM, first, generates initialization population, it is each in population
A particle represents a potential optimal solution, and each particle has three index expression its features：Position, speed and fitness
Value；Then, the fitness value of each particle is calculated according to object function, the quality of the value represents the quality of particle, works as particle
When solution space moves, by track individual extreme value and group's extreme value come the position of more new individual；Particle often updates a position,
It needs to recalculate fitness value, and by comparing the fitness value of new particle and individual extreme value, the fitness of group's extreme value
Value comes the more position of new individual extreme value and group's extreme value；Finally, when finding optimal solution good enough or reaching iterations,
PSO algorithms terminate, and export optimal nuclear parameter g and penalty parameter c.
Wherein, it is described in step 6 " to carry out SVM models using the parameter after optimization and training set sample data
Practise ", the practice is as follows：
Step 6.1：If known training set：T={ (x1, y1) ..., (xi, yi)}∈(X×Y)l, wherein, xiBe characterized to
Amount, yiFor corresponding property value, xi∈X∈Rn, yi∈ Y={ -1,1 }, i=1,2 ..., l；
Step 6.2：The kernel function g (x, x ') and appropriate punishment parameter C of PSO optimizations are chosen, constructs and solves optimization
Problem：So that0≤αi≤ C, i=1,2 ..., l, so as to obtain optimal solution
Step 6.3：Choose α*A positive component 0<α*<C, and threshold value is calculated accordingly
Step 6.4：Construct decision function f (x)：
Step 6.5:Classification is exported according to the value of decision function f (x).
By above-mentioned steps, the complex device degenerate state recognition methods based on Kohonen-SVM, this method can be realized
Unsupervised degenerate state method for digging based on Kohonen neutral nets has obtained the typical degradation state sample collection number of equipment
According to；Further, SVM model parameters are optimized using PSO methods on the basis of training sample data collection, by using excellent
The SVM models of change carry out Classification and Identification to degenerate state data, have obtained the degenerate state identification model of complex device.Finally adopt
With test sample data set to the Accuracy Verification of identification disaggregated model, to prove the feasible of this method identification degenerate state
Property.
(3) advantage
The present invention compared with prior art the advantages of be：At present, most of method of fault pattern recognition can only be to complexity
The normal and failure two states of equipment are identified, and cannot have to being in the normally degenerate state between malfunction
The excavation and identification of effect.And the present invention can excavate intermediate degenerate state data from the normal power data of complex device, and
Exercise supervision study on this basis, obtains equipment typically all kinds of degenerate state data.
Description of the drawings
Fig. 1 is the general steps flow of the present invention.
Fig. 2 is the power data feature extraction flow of the present invention.
Fig. 3 is the method flow based on Kohonen networks of the present invention.
Fig. 4 is the particle swarm optimization algorithm flow of the present invention.
Fig. 5 is the degenerate state identification process of the present invention.
Flow is realized in the degenerate state recognition methods that Fig. 6 is the present invention.
Sequence number, symbol, code name are described as follows in figure：
" step 1-7 " in Fig. 1 is step corresponding in technical solution of the present invention；
" PSO " in Fig. 1 is particle group optimizing method, is solved for completing the optimizing of support vector machines parameter；
Fig. 1, the Kohonen networks in 6 are Ke Helun neutral nets, for completing the excacation of degenerate state；
Fig. 1, the SVM in 5,6 is supporting vector machine model, for completing the identification of degenerate state classification work；
Fig. 3, " Y " in 4 represent that Rule of judgment is logic " true ", and " N " represents that Rule of judgment is logic "false".
Specific embodiment
To make the technical problem to be solved in the present invention, technical solution and advantage clearer, carried out below in conjunction with attached drawing
It is described in detail.
The present invention proposes a kind of complex device degenerate state recognition methods based on Kohonen-SVM, and this method passes through
Clustering processing is carried out to the normal power data of complex device, obtains the typical degradation state sample collection data of equipment.Further
On the basis of training sample data collection, SVM model parameters are optimized using PSO methods, by using parameter optimization
SVM models exercise supervision to degenerate state data study, obtain the degenerate state identification model of equipment.Finally use test sample
Data are to identifying that the accuracy rate of disaggregated model is verified, it was demonstrated that feasibility of this method in terms of degenerate state is identified.
A kind of complex device degenerate state recognition methods based on Kohonen-SVM of the present invention, as shown in Figure 1, it is specific
Construction step is as follows：
Step 1：The normal power data of complex device are carried out with real-time collection and continual collection, data should use " when m- numerical value "
Form be indicated.After the completion of data acquisition, fault data wherein that may be present and bad is rejected by artificial screening method
Value Data；
Step 2：" feature extraction-selection-dimensionality reduction " processing is carried out to the power data after screening, as Kohonen nerves
The learning data of network model.Wherein, the processing method of " feature extraction-selection-dimensionality reduction " is as shown in Figure 2.This method is right first
Power data is divided, and is typically divided into time domain data and codomain data.After the completion of division, the feature for proceeding by data carries
Take work.For power time domain data, this method extraction average, variance, root mean square, kurtosis, inlet and outlet difference, minimax are poor
The parameters such as value, peak factor, the pulse factor, difference and shape factor are as characteristic；For performance number numeric field data, the party
Method extraction maximum time value, average value, data points, minimax difference, value median, maximum, time median, mode
Etc. parameters as characteristic.After the completion of extraction, using these characteristic construction feature vector spaces, using based on criterion
The method of function carries out feature selecting.After feature selecting is completed, by principal component analysis, progress feature drop is locally linear embedding into
Dimension, the multidimensional characteristic vectors data of acquisition；
Step 3：Multidimensional characteristic vectors data are inputted into Kohonen networks and carry out Unsupervised clustering study, are moved back
Change state sample collection data.Degenerate state clustering method flow based on Kohonen networks is as shown in figure 3, power by extraction
Characteristic after the input data of model as, it is necessary to initialized according to the characteristics of data to network, determining network inputs
Output node number and learning parameter.Network inputs number of nodes should be consistent with the dimension of input feature value, output node
It is several then to be determined by the classification number of degenerate state.Kohonen network output layer node on behalf input data is potentially degenerated shape
State classification, general output layer number of nodes are more than the number of input data concrete class.For complex device, the present invention to be identified
Degenerate state classification be less than 10 kinds, so only select 36 to compete output classification of the node layers as network here, and provide
These competition node layers are arranged with the box formation that 6 rows 6 arrange.
For the learning parameter of network, since field radius and learning rate can taper into during evolution,
It needs to optimize tune to the node weights in the winning node field radius r of network according to the field formula of Kohonen networks
Whole, such input data, so as to fulfill the function of convergence of network, can obtain typical gradually into certain several output layer set of node
Degenerate state sample set data；
Step 4：Degenerate state sample set data are pre-processed, obtain training sample set data and test sample collection
Data.Here, it is necessary to be carried out to the number of samples of degenerate state sample set data unified, it is specified that the number of samples one of each state
It causes, and by 4:Each sample is divided into training set and test set by 1 ratio.Wherein, training set data is built for carrying out SVM
Mould, test set data are used to carry out the recognition accuracy test of degenerate state identification model；
Step 5：On the basis of training set sample data, using particle swarm optimization algorithm (Particle Swarm
Optimization, PSO) optimizing solution is carried out to the nuclear parameter g and penalty parameter c of SVM, it obtains optimal nuclear parameter g and punishes
Penalty parameter c, the two parameters can generate conclusive influence to the optimal performance of grader.Using PSO algorithms to SVM nuclear parameters
With punishment parameter optimizing flow as shown in figure 4, first, generating initialization population, each particle represents one and dives in population
Optimal solution, each particle has three index expression its features：Position, speed and fitness value.Then, according to target letter
Number calculates the fitness values of each particles, and the quality of the value represents the quality of particle, when particle is when solution space moves, passes through
Track individual extreme value and group's extreme value come the position of more new individual.Particle often updates a position, it is necessary to recalculate fitness
Value, and by comparing new particle fitness value and individual extreme value, the fitness value of group's extreme value come more new individual extreme value and
The position of group's extreme value.Finally, when finding optimal solution good enough or reaching iterations, PSO algorithms terminate, and output is most
Excellent nuclear parameter g and penalty parameter c；
Step 6：SVM model learnings are carried out using the parameter after optimization and training set sample data, obtain effectively identifying
The SVM models of degenerate state, i.e. degenerate state identification model.Because SVM models have very big in Small Sample Database classification
Superiority, and the number of samples that the degenerate state of Kohonen neutral nets is excavated not is very much.Therefore the present invention adopts
By the use of SVM as the recognition methods of degenerate state.Overall flow such as Fig. 5 of complex device degenerate state recognition methods based on SVM
Shown, " selected training set and the test set data " and " data set pretreatment " in figure are in " step 2 " before to " step
It is completed in four ".This step is substantially carried out " training SVM models " and " generation identification disaggregated model " two parts, the training process of SVM
It is the supervised learning process using classification accuracy rate as index, that is, finds the process of optimal classifying face.In training, penalty
C and nuclear parameter g can generate the optimal performance of SVM classifier conclusive influence, and the present invention employs PSO in " step 5 "
Method completes the Optimization Work of both parameters, therefore combines the parameter of optimization and training set sample data, carries out SVM supervision and learns
The practice of habit is as follows：
Step 6.1：If known training set：T={ (x1, y1) ..., (xi, yi)}∈(X×Y)l, wherein, xiBe characterized to
Amount, yiFor corresponding property value, xi∈X∈Rn, yi∈ Y={ -1,1 }, i=1,2 ..., l；
Step 6.2：The kernel function g (x, x ') and appropriate punishment parameter C of PSO optimizations are chosen, constructs and solves optimization
Problem：So that0≤αi≤ C, i=1,2 ..., l, so as to obtain optimal solution
Step 6.3：Choose α*A positive component 0<α*<C, and threshold value b is calculated accordingly*：
Step 6.4：Construct decision function f (x)：
Step 6.5:Classification is exported according to the value of decision function f (x).
After the completion of SVM training, point of the degenerate state recognition classifier model available for equipment degenerate state data is obtained
Class.And " the identification model verification " in Fig. 5 can " step 7 " middle completion below；
Step 7：The accuracy rate of Classification and Identification model is verified using test set sample data, passes through test set sample
The mode of notebook data injection verifies the accuracy rate of sorter model, comparison prediction classification results and actual classification as a result,
Show that degenerate state identifies classification accuracy, so as to judge that can grader meet field demand.
Wherein, the field formula of the Kohonen networks described in step 3 is as follows：
Ni(t)=(t | find (norm (posc,post)<R)) t=1,2 ..., n
In formula, posc、postThe position of neuron c and t are represented respectively, and find is that distance searches function, and norm is used to calculate
Euclidean distance between two neurons, r are field radius, and η is learning rate.R, η generally with the increase of evolution number and
Linear decline；
R described in step 6nIt is n dimension real number spaces；
Sgn described in step 6 is a kind of sign function, for return an integer variable, it is indicated that parameter it is positive and negative
Number.The grammer of the function is sgn (number), and number parameters are any effective numerical expressions.If return value is digital
More than 0, then sgn returns to 1, and number is equal to 0, then returns to 0, number is less than 0, then returns to -1, the symbol of digital parameters determines sgn
The return value of function；
By above-mentioned steps, the complex device degenerate state recognition methods based on Kohonen-SVM, entirety can be completed
Flow is realized as shown in fig. 6, this method is obtained complexity and set by the unsupervised degenerate state excavation of Kohonen neutral nets
Standby typical degradation state sample collection data.Further, particle cluster algorithm is used on the basis of training sample data collection to SVM
Model parameter optimizes, and carries out Classification and Identification to degenerate state data by using the SVM models of optimization, has obtained complexity and set
Standby degenerate state identification model.Finally using test sample data to identifying that the accuracy of disaggregated model is verified, it was demonstrated that
This method identifies the feasibility of degenerate state.
Non-elaborated part of the present invention belongs to techniques well known.
The above is only part specific embodiment of the present invention, but protection scope of the present invention is not limited thereto, and is appointed
In the technical scope disclosed by the present invention, the change or replacement that can be readily occurred in should all be covered what those skilled in the art
Within protection scope of the present invention.
Claims (9)
1. a kind of complex device degenerate state recognition methods based on Kohonen-SVM, it is characterised in that：Its step are as follows：
Step 1 is acquired the normal power data of complex device, rejected by way of artificial screening wherein there may be
Fault data and bad value data；
Step 2 carries out feature extraction-selection-dimensionality reduction to the power data after screening, and the multidimensional for obtaining to be used for input study is special
Levy vector data；
Multidimensional characteristic vectors data are inputted into Kohonen networks progress Unsupervised clustering study by step 3, obtain degeneration shape
State sample set data；
Step 4 carries out division processing to degeneration sample set data, obtains SVM training sample sets data and test sample collection data；
Step 5, on the basis of training sample set, optimizing is carried out to the parameter of SVM using particle group optimizing method, that is, PSO and is asked
Solution；
Step 6 carries out SVM model learnings using the parameter and training set sample data of optimization, obtains effectively identifying degeneration shape
The SVM models of state, i.e. degenerate state identification model；
Step 7 verifies the accuracy rate of Classification and Identification model using test set sample data；
Wherein, it is described in step 3 " multidimensional characteristic vectors data to be inputted, Unsupervised clustering is carried out into Kohonen networks
Study ", the practice is as follows：Multidimensional characteristic vectors data are inputted to network, the neuron on network competition layer calculates input sample
Originally the Euclidean distance between competition layer neuron weights, the minimum neuron of distance is triumph neuron；Adjustment is won
Neuron and adjacent neurons weights make triumph neuron and periphery weights close to the input sample；By repetition training, finally
The connection weight of each neuron has certain distribution, and the distribution is the similitude tissue between data to representing all kinds of nerves
Under member, make similar neuron that there is similar weight coefficient, inhomogeneous neuron weight coefficient difference is apparent；It should be noted that
In learning process, weights modification learning rate and neuron field are constantly being reduced, so that similar neuron gradually collects
In, the automatic cluster of input data is completed, generates different classes of degenerate state sample set data；
By above-mentioned steps, the complex device degenerate state recognition methods based on Kohonen-SVM can be realized, this method is based on
The unsupervised degenerate state method for digging of Kohonen neutral nets has obtained the typical degradation state sample collection data of equipment；Into
One step optimizes SVM model parameters using PSO methods on the basis of training sample data collection, by using optimization
SVM models carry out Classification and Identification to degenerate state data, have obtained the degenerate state identification model of complex device；Finally using survey
Sample notebook data set pair identifies the Accuracy Verification of disaggregated model, to prove the feasibility of this method identification degenerate state.
2. a kind of complex device degenerate state recognition methods based on Kohonen-SVM according to claim 1, feature
It is：
" power data " in step 1 during referring to device action, the real-time electricity of equipment is obtained by harvester
Stream, voltage value, and the data of the characterization equipment state obtained on the basis of these numerical value using computer automatic analysis；It is described
" fault data ", when referring to that equipment running process breaks down, the abnormal power data that are detected by comparative analysis；It is described
" bad value data ", refer in device action data acquisition, data value be less than acquisition defined threshold power data, bad value
The presence of data can influence the accuracy of plant capacity overall data.
3. a kind of complex device degenerate state recognition methods based on Kohonen-SVM according to claim 1, feature
It is：
Described " carrying out feature extraction-selection-dimensionality reduction to the power data after screening " in step 2, the practice is as follows：To work(
Rate data are divided into time domain data and codomain data, after the completion of division, proceed by the feature extraction work of data；For work(
Rate time domain data, this method extraction average, variance, root mean square, kurtosis, inlet and outlet difference, minimax difference, peak factor,
The pulse factor, difference and shape factor parameter are as characteristic；For performance number numeric field data, this method extraction maximum time
Value, average value, data points, minimax difference, value median, maximum, time median and mode parameter are as feature
Data；After the completion of extraction, using these characteristic construction feature vector spaces, carried out using the method based on criterion function
Feature selecting；After feature selecting is completed, by principal component analysis, carry out Feature Dimension Reduction, the multidimensional of acquisition are locally linear embedding into
Characteristic vector data.
4. a kind of complex device degenerate state recognition methods based on Kohonen-SVM according to claim 1, feature
It is：
" the multidimensional characteristic vectors data " in step 3 refer to that power data passes through feature extraction-selection-dimension-reduction treatment
The data obtained afterwards, input data of the multidimensional characteristic vectors data as Kohonen networks, for the cluster of degenerate state
It practises；" the Kohonen networks ", refers to a kind of Self-organizing Competition type neutral net；The network passes through self-organizing feature map
Network weight is adjusted, neutral net is made to converge on a kind of expression form, a neuron is only to certain input in this form
Pattern especially matching or sensitive；Kohonen networks are the feedforward neural networks comprising two layers of input layer and competition layer；Wherein, the 1st
Layer is input layer, and the number of input layer is consistent with the dimension of input sample vector, and it is m to take input layer number；
2nd layer is competition layer, also referred to as output layer, and competition node layer is distributed in two-dimensional array, and it is n to take competition layer number of nodes；Input node
It is connected entirely with variable weight between output node, connection weight is unsupervised from group for the study of Kohonen networks
Knit learning process；By unsupervised learning, different neurons can be sensitive to different input patterns in network, realizes that neuron exists
The detector of a certain input pattern is served as in pattern-recognition；After the completion of study, neuron is divided into different zones, each region pair
Input model has different response characteristics.
5. a kind of complex device degenerate state recognition methods based on Kohonen-SVM according to claim 1, feature
It is：
Described " carrying out division processing to degeneration sample set data " in step 4, the practice is as follows：To degenerate state sample set
The number of samples of data carries out unified, it is specified that the number of samples of each state is consistent, and presses 4:1 ratio divides each sample
For training set and test set；For carrying out SVM modelings, test set data are used to carry out degenerate state identification model training set data
Recognition accuracy test.
6. a kind of complex device degenerate state recognition methods based on Kohonen-SVM according to claim 1, feature
It is：
" SVM " in step 4, refers to supporting vector machine model, which proposed on the basis of statistical theory
A kind of sorter model, for pattern classification, linear and nonlinear regression analysis；The method of support vector machines is given instruction
Practice sample, establish an Optimal Separating Hyperplane as decision-making curved surface, maximize the isolation edge between sample positive example and counter-example, from
And complete the classification of training sample.
7. a kind of complex device degenerate state recognition methods based on Kohonen-SVM according to claim 1, feature
It is：
" particle group optimizing method " in steps of 5 refers to a kind of optimization algorithm based on swarm intelligence theory, the algorithm
With self-teaching and to the two-fold advantage that other people learn, optimal solution can be found in less iterations.
8. a kind of complex device degenerate state recognition methods based on Kohonen-SVM according to claim 1, feature
It is：
It is described in steps of 5 " right using particle group optimizing method (Particle Swarm Optimization, abbreviation PSO)
The parameter of SVM carries out optimizing solution ", i.e., " optimizing solution is carried out to the parameter of SVM using PSO methods ", the practice is as follows：It is right
The nuclear parameter g and penalty parameter c of SVM carries out optimizing solution, first, generates initialization population, each particle generation in population
One potential optimal solution of table, each particle have three index expression its features：Position, speed and fitness value；Then, root
The fitness value of each particle is calculated according to object function, the quality of the value represents the quality of particle, when particle is transported in solution space
When dynamic, by track individual extreme value and group's extreme value come the position of more new individual；Particle often updates a position, it is necessary to count again
Fitness value is calculated, and it is a to update by comparing the fitness value of new particle and individual extreme value, the fitness value of group's extreme value
The position of body extreme value and group's extreme value；Finally, when finding optimal solution good enough or reaching iterations, PSO algorithm knots
Beam exports optimal nuclear parameter g and penalty parameter c.
9. a kind of complex device degenerate state recognition methods based on Kohonen-SVM according to claim 1, feature
It is：
Described " carrying out SVM model learnings using the parameter after optimization and training set sample data " in step 6, the practice is such as
Under：
Step 6.1：If known training set：T={ (x1,y1),...,(xi,yi)}∈(X×Y)l, wherein, xiFor feature vector, yi
For corresponding property value, xi∈X∈Rn, yi∈ Y={ -1,1 }, i=1,2 ..., l；
Step 6.2：The kernel function g and punishment parameter C of PSO optimizations are chosen, constructs and solves optimization problem：So thatSo as to obtain
Optimal solution
Step 6.3：Choose α*A positive component 0<α*<C, and threshold value b is calculated accordingly*：
Step 6.4：Construct decision function f (x)：
Step 6.5:Classification is exported according to the value of decision function f (x).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710691675.9A CN107563414B (en) | 2017-08-14 | 2017-08-14 | A kind of complex device degenerate state recognition methods based on Kohonen-SVM |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710691675.9A CN107563414B (en) | 2017-08-14 | 2017-08-14 | A kind of complex device degenerate state recognition methods based on Kohonen-SVM |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107563414A CN107563414A (en) | 2018-01-09 |
| CN107563414B true CN107563414B (en) | 2018-05-29 |
Family
ID=60975504
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710691675.9A Active CN107563414B (en) | 2017-08-14 | 2017-08-14 | A kind of complex device degenerate state recognition methods based on Kohonen-SVM |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107563414B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109597315B (en) * | 2018-10-31 | 2020-08-18 | 华中科技大学 | Method, equipment and system for identifying health degradation state of mechanical equipment |
| CN109856494A (en) * | 2019-01-02 | 2019-06-07 | 广东工业大学 | A kind of Diagnosis Method of Transformer Faults based on support vector machines |
| CN110033082B (en) * | 2019-03-19 | 2021-05-18 | 浙江工业大学 | Method for identifying deep learning model in AI (Artificial intelligence) equipment |
| CN112418317A (en) * | 2020-11-24 | 2021-02-26 | 西南交通大学 | Method for identifying and classifying precision machining structural part based on PSO-SVM |
| CN113255795B (en) * | 2021-06-02 | 2021-10-26 | 杭州安脉盛智能技术有限公司 | Equipment state monitoring method based on multi-index cluster analysis |
| CN114397521A (en) * | 2021-12-24 | 2022-04-26 | 中国人民解放军海军航空大学 | Fault diagnosis method and system for electronic equipment |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102609764A (en) * | 2012-02-01 | 2012-07-25 | 上海电力学院 | CPN neural network-based fault diagnosis method for stream-turbine generator set |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE60130742T2 (en) * | 2001-05-28 | 2008-07-17 | Honda Research Institute Europe Gmbh | Pattern recognition with hierarchical networks |
-
2017
- 2017-08-14 CN CN201710691675.9A patent/CN107563414B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102609764A (en) * | 2012-02-01 | 2012-07-25 | 上海电力学院 | CPN neural network-based fault diagnosis method for stream-turbine generator set |
Non-Patent Citations (2)
| Title |
|---|
| APPLICATION OF TIIE KOHONEN’CSOMPETITIVELAYERIN THE IMPLEMENTATION OF THE SPACE VECTOR MODULATIO;A. Bakhshai 等;《IEEE》;19971231;全文 * |
| 贝叶斯网络理论及其在设备故障诊断中的应用;李俭川 等;《中国机械工程》;20030531(第10期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107563414A (en) | 2018-01-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107563414B (en) | A kind of complex device degenerate state recognition methods based on Kohonen-SVM | |
| CN109669087A (en) | A kind of method for diagnosing fault of power transformer based on Multi-source Information Fusion | |
| Chang et al. | Enforced self‐organizing map neural networks for river flood forecasting | |
| CN102024179A (en) | Genetic algorithm-self-organization map (GA-SOM) clustering method based on semi-supervised learning | |
| Chen et al. | AI‐enhanced soil management and smart farming | |
| Bong et al. | Multiobjective clustering with metaheuristic: current trends and methods in image segmentation | |
| Swarup et al. | ANN approach assesses system security | |
| CN106647272A (en) | Robot route planning method by employing improved convolutional neural network based on K mean value | |
| CN116738339A (en) | Multi-classification deep learning recognition detection method for small-sample electric signals | |
| Cruz-Vega et al. | A comparison of intelligent classifiers of thermal patterns in diabetic foot | |
| Su et al. | Generative adversarial networks for gearbox of wind turbine with unbalanced data sets in fault diagnosis | |
| Tsang et al. | Refinement of generated fuzzy production rules by using a fuzzy neural network | |
| Baruque et al. | Hybrid classification ensemble using topology-preserving clustering | |
| Ismail et al. | River flow forecasting: a hybrid model of self organizing maps and least square support vector machine | |
| Antonelli et al. | Multi-objective evolutionary generation of Mamdani fuzzy rule-based systems based on rule and condition selection | |
| CN112465253B (en) | Method and device for predicting links in urban road network | |
| Meschino et al. | Using SOM as a tool for automated design of clustering systems based on fuzzy predicates | |
| Wang et al. | Fault diagnosis of ship ballast water system based on support vector machine optimized by improved sparrow search algorithm | |
| Bao et al. | Back Propagation Optimization of Convolutional Neural Network Based on the left and the right hands Identification | |
| El-Gamal | Genetically evolved neural networks for fault classification in analog circuits | |
| Chen et al. | Hierarchical fuzzy systems | |
| Gonsalves et al. | Data clustering through particle swarm optimization driven self-organizing maps | |
| Castaño et al. | Classification by evolutionary generalised radial basis functions | |
| Iyoda et al. | Evolutionary design of neurofuzzy networks for pattern classification | |
| Pan et al. | Use probabilistic neural network to construct early warning model for business financial distress |
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 |