CN103294849A - Alternating-current motor failure diagnosis model building method based on RBF (radial basis function) neutral network - Google Patents
Alternating-current motor failure diagnosis model building method based on RBF (radial basis function) neutral network Download PDFInfo
- Publication number
- CN103294849A CN103294849A CN2013101676623A CN201310167662A CN103294849A CN 103294849 A CN103294849 A CN 103294849A CN 2013101676623 A CN2013101676623 A CN 2013101676623A CN 201310167662 A CN201310167662 A CN 201310167662A CN 103294849 A CN103294849 A CN 103294849A
- Authority
- CN
- China
- Prior art keywords
- submodel
- fault diagnosis
- output
- model
- diagnosis model
- 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
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/80—Technologies aiming to reduce greenhouse gasses emissions common to all road transportation technologies
- Y02T10/82—Elements for improving aerodynamics
Abstract
The invention discloses an alternating-current motor failure diagnosis model building method based on an RBF (radial basis function) neutral network. The method includes the following steps: building a failure diagnosis model; training the failure diagnosis model; testing the failure diagnosis model. A modularized structure is adopted to form modules of the model, and the modules refer to sub-models which are identical in number of input ends, so that the failure diagnosis model can be built only by parallelly connecting the input ends of the well-trained sub-models, and mutual independence among the sub-models is guaranteed. By the connection mode, the failure diagnosis model is simple, convenient and flexible to build. Training of the failure diagnosis model can be realized only by training the sub-models, so that difficulty in training the failure diagnosis model can be greatly lowered and a condition is provided for reducing training time. The sub-models are mutually independent, so that the trained sub-models are more sensitive to learned specific failure states, and failure recognition capability of the failure diagnosis model is improved.
Description
Technical field
The present invention relates to a kind of Fault Diagnosis of AC Electric Machine technology, particularly a kind of Fault Diagnosis of AC Electric Machine model building method based on the RBF neural network.
Background technology
The fault type of induction motor is divided into mechanical fault and electric fault substantially, these faults can be divided into polytypes such as bearing fault, stator failure, rotor fault and eccentric class fault again.Different faults has different fault signatures, and this also is the main foundation of carrying out fault diagnosis.But it should be noted that because induction motor is that collection is mechanical, electrical, the complex apparatus of magnetic one.Therefore, fault signature also is extremely complicated.Under different service conditions, fault of the same type, its feature also can change; Different faults has similar fault signature.These have brought difficulty for the fault diagnosis of induction motor.
At present, most electrical fault diagnostic models based on neural network adopt many input/many export structures usually, namely make up the model that carries out different fault diagnosis with a complicated neural network; Three layers of BP(BackPropagation of the many employings of neural network) neural network or three layers of radially basic (RBF, Radial Basis Function) neural network.RBF neural network and BP neural network all are non-linear multilayer feedforward networks, and they all belong to the general device that approaches.Three-layer neural network is made of input layer, hidden layer and output layer.
Because the transfer function of RBF neural network hidden layer unit is about centrosymmetric radial basis function (as Gaussian function), the hidden layer unit number of such three layers of static RBF feedforward neural network can be according to the particular problem of research, adjust adaptively in the training stage, make that the applicability of network is better.The RBF neural network is adopted nearest neighbor classifier mode training network usually, and the distribution of network hidden layer unit is just only relevant with the width of the distribution of training sample and hidden layer unit like this, and is irrelevant with carrying out of task.On the basis of hidden layer unit distribution, the mapping relations between input and the output realize by the weights of adjusting between hidden layer unit and the output unit.
RBF neural network and BP neural network relatively have characteristics such as calculated amount is little, fast convergence rate, no local minimum.Therefore, the present invention proposes modularized motor fault diagnosis model based on the RBF neural network.About the following document of electrical fault diagnosis problem research based on neural network report is arranged all:
[1] yellow lead, Huang Cailun. based on the electrical fault diagnostic expert system [J] of BP neural network model. instrument and meter for automation, 2003,24 (3): 15-17;
[2] Guo Xijin etc. based on the asynchronous machine turn-to-turn short circuit degree research [J] of BP neural network. coal mine machinery, 2011,32 (10): 276-278;
[3] Zhang Jingzhai .RBF neural network is at hydrogenerator Application in Fault Diagnosis [J]. Computer Simulation, 2011,28 (12): 314-317;
[4] Wang Juan etc. based on the electrical fault Studies on Diagnosis [J] of RBF neural network. system simulation technology, 2009,5 (1): 36-39;
[5] .RBF neural network such as Mu Lijuan is at asynchronous machine Application in Fault Diagnosis [J]. system simulation technology, 2009,5 (1): 148-151;
[6] Bai Chengrong. the application [J] of RBF neural network in the electrical fault diagnosis. Shanxi architecture, 201,37 (17): 213-214.
For neural network can being learnt preferably and distinguishing the various faults feature, this network should have good complex nonlinear characteristic mapping ability.Certainly will require the hidden layer of this network that more node should be arranged like this.Although existing document has pointed out that three-layer neural network is to approach any nonlinear characteristic in theory, to a network that the hidden node number is bigger, obviously can cause the difficult problem of aspects such as convergence, training and retraining.The deficiency that has following several respects for the fault diagnosis model that is constituted by a multiple-input and multiple-output network at least:
1) fault type that will identify is more many, and the network of component model is just more complicated;
2) complicated network structure, then network convergence and the corresponding increase of training difficulty;
3) in case the newly-increased fault type that will identify or certain fault signature had the additional sample data just need carry out the network retraining, increase and train cost.
Summary of the invention
For solving the problems referred to above that prior art exists, the present invention will design a kind of Fault Diagnosis of AC Electric Machine model building method based on the RBF neural network that can realize following purpose:
1, have conveniently, flexibly, the modularization model structure that can conveniently make up;
2, reduce the convergence and training difficulty of model;
3, model modification is convenient and swift.
To achieve these goals, technical scheme of the present invention is as follows: a kind of Fault Diagnosis of AC Electric Machine model building method based on the RBF neural network may further comprise the steps:
A, set up fault diagnosis model
A1, determine the structure of fault diagnosis model
At first determine to constitute the submodel number of electrical fault diagnostic model according to the number of intending diagnosis alternating current generator failure mode; Each submodel is the model of input more than-single output, and namely each submodel has a plurality of input ends and single output terminal, and the input end number of each submodel is identical; The input end of each submodel is connected in parallel namely constitutes the electrical fault diagnostic model, and the output terminal number of all submodels namely constitutes the output terminal number of electrical fault diagnostic model, that is the number of submodel is exactly the output terminal number of electrical fault diagnostic model;
Described submodel is to be made of input more than-single output three-layer neural network; The input end number of the corresponding submodel of the input layer number of described neural network, output layer node number is 1, the output terminal number of corresponding submodel, each submodel is used for representing a kind of malfunction of motor;
A2, determine the input end number of fault diagnosis model
If submodel has m input end, that is fault diagnosis model has m input end, and the input signal of establishing m input end constitutes an input vector x, is expressed as follows:
x＝(x
1x
2…x
m)
In the formula, x
iBe i input signal of submodel, i=1,2 ..., m;
Described input signal is the radial vibration acceleration signal of alternating current generator stator current signal or alternating current arbor or the axial vibration acceleration signal of alternating current arbor; Described input vector is the electrical fault energy ratio vector that input signal obtains through WAVELET PACKET DECOMPOSITION; Described energy is known as proper vector than vector, and each component correspondence of proper vector the signal energy of a certain frequency band and the gross energy ratio of full range band signal; The element number of proper vector is determined by the signal band number of decomposing;
The element number of proper vector is the input number m of fault diagnosis model, just the input number m of each submodel;
A3, determine the output terminal number of fault diagnosis model
If fault diagnosis model is made of n submodel, then its output vector y is made of the output signal of its all submodels that comprise, that is:
y=(y
1y
2…y
n)
In the formula, y
jBe j output signal of electrical fault diagnostic model, the output signal of j submodel just, j=1,2 ..., n, the output terminal number of each submodel is fixed as 1, and expression is by a kind of malfunction of diagnosing motor;
A4, determine the latent node number of submodel
For the sake of simplicity, earlier with default value 2 or the 4 latent node numbers as submodel, finally adjust the latent node number of submodel according to training result;
B, training fault diagnosis model
The output state of B1, definition submodel
For each submodel defines the fault type that it will be represented, any one submodel after the training so has only when importing its corresponding fault type energy when vectorial, and the output signal of this submodel is 1, otherwise output signal is 0;
B2, preparation submodel training sample set
To each fault type, the energy that adopts the fault data be not less than 10 groups to generate constitutes the training sample set of each submodel than vector;
B3, training submodel
Training sample set with resulting each submodel is trained each submodel respectively;
B4, synthetic fault diagnosis model
After all submodel training finished, the input end parallel connection with each submodel after the training had just constituted the Fault Diagnosis of AC Electric Machine model that will set up;
C, test failure diagnostic model
C1, preparation fault diagnosis model test sample book collection
Energy with the fault type that comprises all submodel correspondences that is different from training sample set constitutes the test sample book collection than vector, according to each fault energy of test sample book collection putting in order than vector, corresponding motor status output table is referred to herein as the ideal output table of fault diagnosis model; The corresponding test sample book of each row in the table is concentrated the represented malfunction output of corresponding line, the ideal of fault diagnosis model output just;
The fault diagnosis performance of C2, test failure diagnostic model
Fault energy in the continuous input test sample set is than vector successively, the output vector of record cast, obtain the actual output table of model, ideal output table and the actual output table of model compared, come the fault diagnosis performance of fault diagnosis model is carried out test and evaluation; If the matching degree of the actual output table of fault diagnosis model and desirable output table thinks namely to meet the demands that greater than 80% then this fault diagnosis model just can come into operation.
The actual output table of the described fault diagnosis model of step C2 of the present invention and the desirable acquisition methods of exporting the matching degree of table may further comprise the steps:
Numerical value in the fault diagnosis model test process in the actual output table of resulting fault diagnosis model is one [0,1] interval real number; Each value in each value in the actual output table of fault diagnosis model is more shown close to ideal output represents that then the fault diagnosis performance of fault diagnosis model is more good; In order to obtain the actual output of fault diagnosis model and the desirable matching degree of exporting, earlier each numerical value in the actual output table of fault diagnosis model being carried out round handles, obtain a correspondence by 0, the 1 new table that constitutes, be called the actual output of fault diagnosis model here and round table; Respectively be worth when in full accord with desirable output table corresponding row when the actual output of fault diagnosis model rounds table, claim the corresponding row coupling, if actual output rounds in table and the desirable output table all couplings of all row, then claim actual output to round table and desirable output table 100% coupling.
Compared with prior art, the present invention has following beneficial effect:
1, the fault diagnosis model structure is easy, flexible.
The present invention proposes a kind of modularization Fault Diagnosis of AC Electric Machine model based on radial basis function (RBF) neural network.Because this electrical fault diagnostic model has adopted modular construction, constitute each module of this model, namely submodel has identical input number, the input end of each submodel that the structure of fault diagnosis model only need will train is connected in parallel and gets final product, and has guaranteed between each submodel separate.This connected mode makes that the fault diagnosis model structure is easy, flexible.Three layers of RBF neural network of the corresponding input more than of each submodel-single output are used for identifying a kind of motor status.This design has guaranteed that the RBF neural network structure of formation submodel is simple relatively, for acceleration disturbance diagnostic model speed of convergence, shortening training time provide condition.In actual applications, can determine required submodel number according to concrete needs and condition.The input number of fault diagnosis model is exactly the input number of submodel, is determined by practical situations, is exactly the number of determining decomposition frequency band according to actual conditions specifically; Can be set at 2 or 4 under the default situation of latent node number of submodel, occurrence can increase and decrease according to the convergence situation in the training process.
2, fault diagnosis model fast convergence rate, training time weak point.
Therefore the present invention, greatly reduces the difficulty of fault diagnosis model training because the fault diagnosis model training only needs submodel is carried out, and provides condition for shortening the training time.
The RBF radial basis function neural network that the present invention adopts is a kind of neural network of feed forward type efficiently, and it has the best that other feedforward networks do not have and approaches performance and global optimum's characteristic, and simple in structure, and training speed is fast.The RBF network is called local acceptance domain network again, and namely only when input drops in the input space in the very little appointed area, hidden unit is just made significant non-zero response.Characteristic is accepted in the part of network makes its decision-making the time imply the concept of distance, namely has only when the acceptance domain imported near the RBF network, and network just can be made response to it.This has just been avoided BP network lineoid to cut apart any division characteristic of bringing.
For the RBF neural network, input layer is 1 to the equal value of the weights between the hidden layer usually.Center and the radius of hidden layer unit radial basis function also pre-determine usually, have only hidden layer adjustable to the weights between the output layer.The hidden layer of RBF network is carried out a kind of changeless nonlinear transformation, and the input space is mapped to a new hidden layer space, and output layer is realized linear combination in this new space.Obviously owing to the linear characteristic of output unit, its parameter is regulated very simple, and does not have the local minimum problem.In addition, there are some researches show that the non-linear activation function form that general RBF network utilizes not is most important to the influence of network performance, key factor is choosing of basis function center.Therefore, belong to the RBF network of feedforward neural network together with the BP network, overcome the more existing problems of BP network to a certain extent.
For these reasons, the present invention has better convergence property and shorter training time.
3, has better Fault Identification ability.
Since separate between each submodel of the present invention, so the submodel after the process training is more responsive to the specific fault conditions of learning, improved the Fault Identification ability of fault diagnosis model like this.
4, can recombinate easily and the model retraining according to its application conditions.
The present invention is when needing to increase model to the identification kind time-like of electrical fault, or certain malfunction there is new characteristic, in the time of need training corresponding submodel again, only need train or retraining corresponding submodel, be recombined into then in the original model and get final product.And for the electrical fault diagnostic model that is constituted by input-many output nerves network more than, then in said case, must train again whole network.Model structure is more complicated, and then convergence is just more difficult with training.
Description of drawings
5 of drawings attacheds of the present invention, wherein:
Fig. 1 is structural representation of the present invention;
Fig. 2 is process flow diagram of the present invention;
Fig. 3 is training process flow diagram of the present invention;
Fig. 4 is test flow chart of the present invention;
Fig. 5 is an explanation use-case structural representation of the present invention.
Embodiment
Below in conjunction with accompanying drawing the present invention is described further.A kind of Fault Diagnosis of AC Electric Machine model building method based on the RBF neural network may further comprise the steps:
A, by the fault diagnosis model of setting up shown in Figure 2
A1, determine the structure of fault diagnosis model
At first determine to constitute the submodel number of electrical fault diagnostic model according to the number of intending diagnosis alternating current generator failure mode; Each submodel is the model of input more than-single output, and namely each submodel has a plurality of input ends and single output terminal, and the input end number of each submodel is identical; The input of each submodel is connected in parallel has just constituted the electrical fault diagnostic model that is made of n submodel, as shown in Figure 1.As shown in Figure 1, the input of model is made of each submodel input parallel connection, and the output of model is made of the output of all submodels, that is to say, the number of submodel is the output number of model just.
Described submodel as part in the solid box among Fig. 1, is to be made of input more than-single output three-layer neural network.The input number of the corresponding submodel of the input layer number of neural network, output layer node number is 1, the output number of corresponding submodel, each submodel is used for representing a kind of malfunction of motor;
A2, determine the input end number of fault diagnosis model
If submodel has m input end, that is fault diagnosis model has m input end, and the input signal of establishing m input end constitutes an input vector x, is expressed as follows:
x＝(x
1x
2…x
m)
In the formula, x
iBe i input signal of submodel, i=1,2 ..., m;
Described input signal is the radial vibration acceleration signal of alternating current generator stator current signal or alternating current arbor or the axial vibration acceleration signal of alternating current arbor; Described input vector is the electrical fault energy ratio vector that input signal obtains through WAVELET PACKET DECOMPOSITION; Described energy is known as proper vector than vector, and each component correspondence of proper vector the signal energy of a certain frequency band and the gross energy ratio of full range band signal; The element number of proper vector is determined by the signal band number of decomposing; Establishing the signal band number that will decompose here is 4, and like this, the sample set form that prepare is the form shown in the table 1.What each row in the table 1 provided is an energy ratio vector, because the signal band number of decomposing is 4, so this energy has four elements than vector, has also determined the input number m=4 of fault diagnosis model simultaneously.
The element number of proper vector is the input number m of fault diagnosis model, just the input number m of each submodel;
Table 1 training and test sample book sample table
Table 1 is submodel training and test sample book collection sample table.In order to say something simply, only provided the two states of motor in the table, i.e. unfaulty conditions and malfunction.Each electrical fault proper vector is made of four components (element), represent four different sub-band signal energy respectively with the ratio of full range band signal energy.
A3, determine the output terminal number of fault diagnosis model
If fault diagnosis model is made of n submodel, then its output vector y is made of the output signal of its all submodels that comprise, that is:
y=(y
1y
2…y
n)
In the formula, y
jBe j output signal of electrical fault diagnostic model, the output signal of j submodel just, j=1,2 ..., n, the output terminal number of each submodel is fixed as 1, and expression is by a kind of malfunction of diagnosing motor;
Simple for illustrating, establish and will identify two kinds of electrical fault states, unfaulty conditions and malfunction; This two states is indicated with a submodel respectively.That is to say that the electrical fault diagnostic model that will construct is made of two submodels, might as well identify the unfaulty conditions of indication motor in other words with submodel 1 here, identify the malfunction of motor with submodel 2.At this moment, the output number n=2 of electrical fault diagnostic model.
A4, determine the latent node number of submodel
For the sake of simplicity, earlier with default value 2 or the 4 latent node numbers as submodel, finally adjust the latent node number of submodel according to training result;
B, by training fault diagnosis model shown in Figure 3
The output state of B1, definition submodel
For each submodel defines the fault type that it will be represented, any one submodel after the training so has only when importing its corresponding fault type energy when vectorial, and the output signal of this submodel is 1, otherwise output signal is 0;
B2, preparation submodel training sample set
To each fault type, the energy that adopts the fault data be not less than 10 groups to generate constitutes the training sample set of each submodel than vector; When the electrical fault number of types that will identify is 2, the sub-band number that decompose is 4 o'clock, and the submodel training sample set of preparing is as shown in table 1.Table 1 training sample is concentrated preceding 5 groups, and what provide is the energy ratio of four sub-frequency bands correspondences under the motor unfaulty conditions, and each row constitutes an energy than vector, the input vector when training submodel; Corresponding these 5 energy are than vector, and the output of the ideal of submodel 1 is 1; What table 1 training sample concentrated that 5 groups of backs provide is the energy ratio of four sub-frequency bands correspondences under the electrical fault state, and each row constitutes an energy than vector, the input vector during as the training submodel; Corresponding these 5 energy are than vector, and the output of the ideal of submodel 1 is 0.In like manner, when with these 10 groups of sample datas submodel 2 being trained, preceding 5 energy are 0 than the ideal output of vectorial corresponding submodel 2; The ideal output of the submodel 2 that 5 energy in back are more corresponding than vector is 1.
B3, training submodel
Training sample set with resulting each submodel is trained each submodel respectively; This to submodel 1 and submodel 2 all the training sample set in the employing table 1 train.
B4, synthetic fault diagnosis model
After all submodel training finished, the input end parallel connection with each submodel after the training had just constituted the Fault Diagnosis of AC Electric Machine model that will set up, as shown in Figure 5;
C, by test failure diagnostic model shown in Figure 4
C1, preparation fault diagnosis model test sample book collection
Energy with the fault type that comprises all submodel correspondences that is different from training sample set constitutes the test sample book collection than vector, and is as shown in table 1.Than vectorial putting in order, corresponding motor status output is shown, and is as shown in table 2 according to each fault energy of test sample book collection, shows referred to herein as the ideal output of fault diagnosis model; The corresponding test sample book of each row in the table is concentrated the represented malfunction output of corresponding line, the ideal of fault diagnosis model output just;
The fault diagnosis performance of C2, test failure diagnostic model
Fault energy in the continuous input test sample set is than vector successively, the output vector of record trouble diagnostic model, obtain the actual output table of fault diagnosis model, ideal output table and the actual output table of fault diagnosis model are compared, come the fault diagnosis performance of fault diagnosis model is carried out test and evaluation; If the matching degree of the actual output table of fault diagnosis model and desirable output table thinks namely to meet the demands that greater than 80% then this fault diagnosis model just can come into operation.
The actual output table of the described fault diagnosis model of step C2 of the present invention and the desirable acquisition methods of exporting the matching degree of table may further comprise the steps:
Numerical value in the fault diagnosis model test process in the actual output table of resulting fault diagnosis model is one [0,1] interval real number; Each value in each value in the actual output table of fault diagnosis model is more shown close to ideal output represents that then the fault diagnosis performance of fault diagnosis model is more good; In order to obtain the actual output of fault diagnosis model and the desirable matching degree of exporting, earlier each numerical value in the actual output table of fault diagnosis model being carried out round handles, obtain a correspondence by 0, the 1 new table that constitutes, be called the actual output of fault diagnosis model here and round table; Respectively be worth when in full accord with desirable output table corresponding row when the actual output of fault diagnosis model rounds table, claim the corresponding row coupling, if actual output rounds in table and the desirable output table all couplings of all row, then claim actual output to round table and desirable output table 100% coupling.
In the present embodiment, with the calculating that rounds up of the real output value in the table 3, obtain table 4, compare with table 2 line by line then.Have only the third line of the third line and table 2 inconsistent in this routine table 4, so its matching degree is (1-1/12) * 100%=92%.Illustrate that this fault diagnosis model meets the demands, and can come into operation.
The model ideal output table of table 2 corresponding tables 1 test sample book collection
Table 2 is the desirable output tables of fault diagnostic model.Table 2 is at test sample book in the table 1, at hypothesis submodel 1 output y
1The indication unfaulty conditions, submodel 2 output y
2The desirable output table of fault diagnosis model under the indicating fault status situation.
The actual output table of the model of table 3 corresponding tables 1 test sample book collection
Table 3 is the actual output tables of fault diagnostic model.Table 3 is that test sample book order successively is input to the actual output table of model that the back generates in the fault diagnosis model.
The actual output of table 4 model rounds table
Table 4 is the tables that round of the actual output of fault diagnostic model.This is shown from table 3, obtains after each element in the table 3 is rounded according to the principle that rounds up.
For the fault diagnosis model that can come into operation, if some components are y among the actual output vector y of fault diagnosis model
j=1, show that then motor is in j the malfunction that submodel is indicated; If y
j=0, show that then this fault does not appear in motor; If 0＜y
j＜1, show that then motor is with y
jDegree of confidence this fault may appear.
Claims (2)
1. Fault Diagnosis of AC Electric Machine model building method based on the RBF neural network is characterized in that: may further comprise the steps:
A, set up fault diagnosis model
A1, determine the structure of fault diagnosis model
At first determine to constitute the submodel number of electrical fault diagnostic model according to the number of intending diagnosis alternating current generator failure mode; Each submodel is the model of input more than-single output, and namely each submodel has a plurality of input ends and single output terminal, and the input end number of each submodel is identical; The input end of each submodel is connected in parallel namely constitutes the electrical fault diagnostic model, and the output terminal number of all submodels namely constitutes the output terminal number of electrical fault diagnostic model, that is the number of submodel is exactly the output terminal number of electrical fault diagnostic model;
Described submodel is to be made of input more than-single output three-layer neural network; The input end number of the corresponding submodel of the input layer number of described neural network, output layer node number is 1, the output terminal number of corresponding submodel, each submodel is used for representing a kind of malfunction of motor;
A2, determine the input end number of fault diagnosis model
If submodel has m input end, that is fault diagnosis model has m input end, and the input signal of establishing m input end constitutes an input vector x, is expressed as follows:
x＝(x
1x
2…x
m)
In the formula, x
iBe i input signal of submodel, i=1,2 ..., m;
Described input signal is the radial vibration acceleration signal of alternating current generator stator current signal or alternating current arbor or the axial vibration acceleration signal of alternating current arbor; Described input vector is the electrical fault energy ratio vector that input signal obtains through WAVELET PACKET DECOMPOSITION; Described energy is known as proper vector than vector, and each component correspondence of proper vector the signal energy of a certain frequency band and the gross energy ratio of full range band signal; The element number of proper vector is determined by the signal band number of decomposing;
The element number of proper vector is the input number m of fault diagnosis model, just the input number m of each submodel;
A3, determine the output terminal number of fault diagnosis model
If fault diagnosis model is made of n submodel, then its output vector y is made of the output signal of its all submodels that comprise, that is:
y=(y
1y
2…y
n)
In the formula, y
jBe j output signal of electrical fault diagnostic model, the output signal of j submodel just, j=1,2 ..., n, the output terminal number of each submodel is fixed as 1, and expression is by a kind of malfunction of diagnosing motor;
A4, determine the latent node number of submodel
For the sake of simplicity, earlier with default value 2 or the 4 latent node numbers as submodel, finally adjust the latent node number of submodel according to training result;
B, training fault diagnosis model
The output state of B1, definition submodel
For each submodel defines the fault type that it will be represented, any one submodel after the training so has only when importing its corresponding fault type energy when vectorial, and the output signal of this submodel is 1, otherwise output signal is 0;
B2, preparation submodel training sample set
To each fault type, the energy that adopts the fault data be not less than 10 groups to generate constitutes the training sample set of each submodel than vector;
B3, training submodel
Training sample set with resulting each submodel is trained each submodel respectively;
B4, synthetic fault diagnosis model
After all submodel training finished, the input end parallel connection with each submodel after the training had just constituted the Fault Diagnosis of AC Electric Machine model that will set up;
C, test failure diagnostic model
C1, preparation fault diagnosis model test sample book collection
Energy with the fault type that comprises all submodel correspondences that is different from training sample set constitutes the test sample book collection than vector, according to each fault energy of test sample book collection putting in order than vector, corresponding motor status output table is referred to herein as the ideal output table of fault diagnosis model; The corresponding test sample book of each row in the table is concentrated the represented malfunction output of corresponding line, the ideal of fault diagnosis model output just;
The fault diagnosis performance of C2, test failure diagnostic model
Fault energy in the continuous input test sample set is than vector successively, the output vector of record cast, obtain the actual output table of model, ideal output table and the actual output table of model compared, come the fault diagnosis performance of fault diagnosis model is carried out test and evaluation; If the matching degree of the actual output table of fault diagnosis model and desirable output table thinks namely to meet the demands that greater than 80% then this fault diagnosis model just can come into operation.
2. a kind of Fault Diagnosis of AC Electric Machine model building method based on the RBF neural network according to claim 1, it is characterized in that: the actual output table of the described fault diagnosis model of step C2 and the desirable acquisition methods of exporting the matching degree of table may further comprise the steps:
Numerical value in the fault diagnosis model test process in the actual output table of resulting fault diagnosis model is one [0,1] interval real number; Each value in each value in the actual output table of fault diagnosis model is more shown close to ideal output represents that then the fault diagnosis performance of fault diagnosis model is more good; In order to obtain the actual output of fault diagnosis model and the desirable matching degree of exporting, earlier each numerical value in the actual output table of fault diagnosis model being carried out round handles, obtain a correspondence by 0, the 1 new table that constitutes, be called the actual output of fault diagnosis model here and round table; Respectively be worth when in full accord with desirable output table corresponding row when the actual output of fault diagnosis model rounds table, claim the corresponding row coupling, if actual output rounds in table and the desirable output table all couplings of all row, then claim actual output to round table and desirable output table 100% coupling.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310167662.3A CN103294849B (en) | 2013-05-08 | 2013-05-08 | Based on the Fault Diagnosis of AC Electric Machine model construction method of RBF neural |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310167662.3A CN103294849B (en) | 2013-05-08 | 2013-05-08 | Based on the Fault Diagnosis of AC Electric Machine model construction method of RBF neural |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103294849A true CN103294849A (en) | 2013-09-11 |
| CN103294849B CN103294849B (en) | 2015-11-18 |
Family
ID=49095706
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310167662.3A Expired - Fee Related CN103294849B (en) | 2013-05-08 | 2013-05-08 | Based on the Fault Diagnosis of AC Electric Machine model construction method of RBF neural |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103294849B (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105116323A (en) * | 2015-08-14 | 2015-12-02 | 江苏科技大学 | Motor fault detection method based on RBF |
| CN105487009A (en) * | 2015-11-19 | 2016-04-13 | 上海电机学院 | Motor fault diagnosis method based on k-means RBF neural network algorithm |
| CN106796579A (en) * | 2014-04-24 | 2017-05-31 | 阿尔斯通运输技术公司 | For the method and system of the defect in automatic detection rotary shaft |
| CN106872894A (en) * | 2017-03-03 | 2017-06-20 | 南方科技大学 | The fault detection method and device of a kind of three phase electric machine |
| CN108106846A (en) * | 2017-12-21 | 2018-06-01 | 大连交通大学 | A kind of rolling bearing fault damage extent identification method |
| CN108931724A (en) * | 2018-07-30 | 2018-12-04 | 袁小芳 | A kind of servo motor method for diagnosing faults |
| CN110135021A (en) * | 2019-04-24 | 2019-08-16 | 南京航空航天大学 | ATRU system failure grading diagnosis method based on source signal and RBF neural |
| CN110954827A (en) * | 2019-12-17 | 2020-04-03 | 北京昊鹏智能技术有限公司 | Fault diagnosis method and device, electronic equipment and system |
| CN111002832A (en) * | 2019-12-24 | 2020-04-14 | 东风电子科技股份有限公司 | Method for realizing configurable processing of motor controller fault diagnosis system in electric automobile based on computer software |
| CN112346941A (en) * | 2019-08-08 | 2021-02-09 | 北京国双科技有限公司 | Fault diagnosis method and device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101661075A (en) * | 2009-06-08 | 2010-03-03 | 浙江大学 | Power system failure diagnostic method based on neural network and fuzzy integral |
| CN101872165A (en) * | 2010-06-13 | 2010-10-27 | 西安交通大学 | Method for fault diagnosis of wind turbines on basis of genetic neural network |
| TW201144840A (en) * | 2010-06-07 | 2011-12-16 | Nat Univ Chin Yi Technology | Diagnosis method and device for dysfunctional motor using RBF information integration technique |
| CN102707228A (en) * | 2011-11-12 | 2012-10-03 | 江苏兴航智控科技股份有限公司 | Neural network expert system-based electric machine fault intelligent diagnosis system |
-
2013
- 2013-05-08 CN CN201310167662.3A patent/CN103294849B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101661075A (en) * | 2009-06-08 | 2010-03-03 | 浙江大学 | Power system failure diagnostic method based on neural network and fuzzy integral |
| TW201144840A (en) * | 2010-06-07 | 2011-12-16 | Nat Univ Chin Yi Technology | Diagnosis method and device for dysfunctional motor using RBF information integration technique |
| CN101872165A (en) * | 2010-06-13 | 2010-10-27 | 西安交通大学 | Method for fault diagnosis of wind turbines on basis of genetic neural network |
| CN102707228A (en) * | 2011-11-12 | 2012-10-03 | 江苏兴航智控科技股份有限公司 | Neural network expert system-based electric machine fault intelligent diagnosis system |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106796579A (en) * | 2014-04-24 | 2017-05-31 | 阿尔斯通运输技术公司 | For the method and system of the defect in automatic detection rotary shaft |
| CN105116323B (en) * | 2015-08-14 | 2017-10-17 | 江苏科技大学 | A kind of electrical fault detection method based on RBF |
| CN105116323A (en) * | 2015-08-14 | 2015-12-02 | 江苏科技大学 | Motor fault detection method based on RBF |
| CN105487009A (en) * | 2015-11-19 | 2016-04-13 | 上海电机学院 | Motor fault diagnosis method based on k-means RBF neural network algorithm |
| CN106872894B (en) * | 2017-03-03 | 2020-01-17 | 南方科技大学 | Fault detection method and device for three-phase motor |
| CN106872894A (en) * | 2017-03-03 | 2017-06-20 | 南方科技大学 | The fault detection method and device of a kind of three phase electric machine |
| CN108106846A (en) * | 2017-12-21 | 2018-06-01 | 大连交通大学 | A kind of rolling bearing fault damage extent identification method |
| CN108106846B (en) * | 2017-12-21 | 2019-09-27 | 大连交通大学 | A kind of rolling bearing fault damage extent identification method |
| CN108931724A (en) * | 2018-07-30 | 2018-12-04 | 袁小芳 | A kind of servo motor method for diagnosing faults |
| CN110135021A (en) * | 2019-04-24 | 2019-08-16 | 南京航空航天大学 | ATRU system failure grading diagnosis method based on source signal and RBF neural |
| CN110135021B (en) * | 2019-04-24 | 2023-07-14 | 南京航空航天大学 | ATRU system fault grading diagnosis method based on multi-source signals and RBF neural network |
| CN112346941A (en) * | 2019-08-08 | 2021-02-09 | 北京国双科技有限公司 | Fault diagnosis method and device |
| CN110954827A (en) * | 2019-12-17 | 2020-04-03 | 北京昊鹏智能技术有限公司 | Fault diagnosis method and device, electronic equipment and system |
| CN111002832A (en) * | 2019-12-24 | 2020-04-14 | 东风电子科技股份有限公司 | Method for realizing configurable processing of motor controller fault diagnosis system in electric automobile based on computer software |
| CN111002832B (en) * | 2019-12-24 | 2022-12-23 | 东风电子科技股份有限公司 | Method for realizing configurable processing of motor controller fault diagnosis system in electric automobile based on computer software |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103294849B (en) | 2015-11-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103294849A (en) | Alternating-current motor failure diagnosis model building method based on RBF (radial basis function) neutral network | |
| CN106184215B (en) | The active damping controls device and method of hybrid vehicle | |
| CN112041693B (en) | Power distribution network fault positioning system based on mixed wave recording | |
| CN108732528A (en) | A kind of digitalized electrical energy meter method for diagnosing faults based on depth confidence network | |
| CN100539354C (en) | Power system separation decision space screening method | |
| CN101975910A (en) | Intelligent fault classification and location method for ultra-high voltage direct current transmission line | |
| CN106329516A (en) | Typical scene recognition based dynamic reconstruction method of power distribution network | |
| CN107909118A (en) | A kind of power distribution network operating mode recording sorting technique based on deep neural network | |
| CN108154223B (en) | Power distribution network working condition wave recording classification method based on network topology and long time sequence information | |
| CN103207950A (en) | Intelligent transformer fault diagnostic method based on RBF (radial basis function) neural network | |
| CN106779066A (en) | A kind of radar circuit plate method for diagnosing faults | |
| CN113255458A (en) | Bearing fault diagnosis method based on multi-view associated feature learning | |
| CN108418242B (en) | Doubly-fed wind turbine dynamic equivalence method based on similarity coherence | |
| CN103455658A (en) | Weighted grey target theory based fault-tolerant motor health status assessment method | |
| CN115641283A (en) | Transformer fault diagnosis method and system based on multi-sensor information fusion | |
| Guolian et al. | Research on fault diagnosis of wind turbine control system based on artificial neural network | |
| CN107566007A (en) | Chromacoder, processing unit, communication system and signal conversion method | |
| Vishwakarma et al. | Application of genetic algorithm trained masterslave Neural Network for differential protection of power transformer | |
| Hussein et al. | Induction motors stator fault analysis based on artificial intelligence | |
| CN104915470A (en) | Generalized state space average modeling method based on aircraft electric power system dynamic characteristics | |
| CN104090228A (en) | Analog circuit fuzzy group identification method | |
| CN104502754A (en) | Fault diagnosis method for pure electric vehicle power system | |
| CN104699994A (en) | FBF (fuzzy basis function) neural network based motor air gap eccentricity fault diagnosis method | |
| Hung et al. | Fault diagnosis of steam turbine-generator sets using CMAC neural network approach and portable diagnosis apparatus implementation | |
| Lehtoranta et al. | Fault diagnosis of induction motors with dynamical neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20151118 Termination date: 20160508 |