CN102662142A - Prediction method for multi-parameter identification fault of power electronic circuit based on RVM-QNN - Google Patents

Abstract

The invention discloses a prediction method for multi-parameter identification faults of a power electronic circuit based on RVM-QNN, comprising the following steps: (1) setting different parameter values for various components in the power electronic circuit and acquiring circuit performance parameters; (2) constructing a QNN neural network multi-parameter identifier by taking the circuit performance parameters and component parameters as a training sample; (3) obtaining time series of the circuit performance parameters at intervals; (4) predicting the circuit performance parameters by using a RVM algorithm and acquiring circuit performance parameters at a time in the future; (5) identifying and acquiring parameters of various components in the circuit by taking the predicted circuit performance parameters as input; and (6) evaluating fault conditions of a system according to the predicted parameters of various components in the circuit. The prediction method is capable of monitoring operation status of a circuit in real time, predicting parameters of various components in the power electronic circuit at a time in the future, evaluating fault conditions of the circuit in the future, and figuring out causes of faults.

Description

A kind of Power Electronic Circuit multiparameter identification of defective Forecasting Methodology based on RVM-QNN

Technical field

The present invention relates to a kind of Power Electronic Circuit failure prediction method, especially a kind of Power Electronic Circuit multiparameter identification of defective Forecasting Methodology based on RVM-QNN.

Background technology

Along with the continuous expansion of engineering real system scale, the complicacy of system, pathosis and non-linear also progressively raising, therefore higher to security of system and reliability requirement.Failure prediction is meant the past and present monitoring information according to system, and whether the following some time etching system of time that the prediction fault takes place or judgement breaks down.Modern electronic equipment has been penetrated into national defense and military, industry, traffic, IT, agricultural, communication, commerce, medicine is made so that in each neighborhood system such as household electrical appliance; The loss that causes because of electronic failure also increases greatly; Therefore, more important for the power-supply system effect of power electronic equipment.Power Electronic Circuit is the core of power-supply system; Study its failure prediction method; Guarantee that it is safe, lasting, reliably working is the part of electronic equipment health control; Become failure prediction hot research fields and difficult point, at huge, the complex structure of investment, fields such as the Aero-Space that reliability requirement is high, nuclear energy have great realistic meaning.

The fault of Power Electronic Circuit is mainly caused by the degradation failure of respectively forming components and parts in the circuit; Therefore; Respectively form following variation tendency of components and parts in the circuit if can dope, confirm the following parameter value of each components and parts constantly, can assess out the fault situation.

On the other hand, quantum nerve network is the product that quantum calculation combines with neural network, and it has better stability and validity, also has quick learning ability and very high information processing rate, can be used in the identification of Power Electronic Circuit real-time parameter.

Associated vector machine (RVM) is a kind of similar with SVMs (SVM) the sparse probability model that Michael E.Tipping proposes, and the training of RVM is carried out under Bayesian frame, can carry out time series and return the estimation prediction.The advantage of RVM is can utilize in the training process Bayes's evidence process to confirm the ultra parameter of model automatically; And need not use crosscheck; Use support vector seldom, and reduce the computing time of export target amount predicted value, can be used for the real-time estimate of Power Electronic Circuit.

Summary of the invention

The object of the invention; Be to provide a kind of Power Electronic Circuit multiparameter identification of defective Forecasting Methodology based on RVM-QNN; It is the observation circuit operation conditions in real time; Each component parameter in following certain moment Power Electronic Circuit of prediction, and then assess following fault situation, and can judge failure cause.

In order to reach above-mentioned purpose, solution of the present invention is:

A kind of Power Electronic Circuit multiparameter identification of defective Forecasting Methodology based on RVM-QNN comprises the steps:

(1) each components and parts in the Power Electronic Circuit is provided with different parameter values, monitors the circuit measuring point signal under the corresponding situation, and calculate the acquisition cuicuit performance parameter;

(2) the QNN neural network is trained as training sample with the circuit performance parameters that obtained, component parameter in the step (1), make up QNN neural network multiparameter identifier;

(3) in the circuit working process, every certain interval of time, monitoring system measuring point signal, counting circuit performance parameter, the time series of acquisition cuicuit performance parameter;

(4) utilize the RVM algorithm that circuit performance parameters is predicted, obtain following certain circuit performance parameters constantly;

(5) be input with the circuit performance parameters of being predicted, utilize the QNN neural network identifier after the training in the step (2), following certain each component parameter of circuit is constantly obtained in identification;

(6), system is carried out the fault state assessment according to each component parameter of circuit of being predicted.

In the above-mentioned steps (2), when making up QNN neural network multiparameter identifier, suppose that institute predict that the components and parts number is N in the Power Electronic Circuit, the component parameter number also is that N is individual, and circuit performance parameters is that M is individual, and then the building process of identifier is:

(21) the QNN neural network structure is three layers; Wherein input layer is M; The output layer node is N, and middle layer node is

round numbers;

(22) produce the initialization weights by random function;

(23) use training sample to network training, with the actual output computational grid output error of desired output and network, by the weights and the quantum interval of each layer of the minimum adjustment of network output error;

(24) will accomplish the training after the QNN network as the multiparameter identifier.

Utilize the method for RVM algorithm predicts circuit performance parameters following in the above-mentioned steps (4):

(41) performance parameter of circuit is X _i, i=1,2, L, M, the performance parameter time series of circuit is X _{I, 1}, X _{I, 2}, X _{I, 3}..., then obtain the performance parameter X of institute's prediction circuit _iP performance parameter sequence constantly is { X continuously _{I, j}, X _{I, j+1}, L, X _{I, j+p}, j=1,2,3, L, p=1,2,3, L;

(42) utilize RVM to { X _{I, j}, X _{I, j+1}, L, X _{I, j+p}Carry out forward direction q step prediction, obtain (j+p+q) performance parameter X of circuit constantly _{I, j+p+q}, q=1,2,3, L.

In the above-mentioned steps (5), the method that following certain each parameter of the moment is obtained in identification is: with X _{I, j+p+q}Be input, i=1,2, L, M picks out (j+p+q) each component parameter θ constantly _{R, j+p+q}, r=1,2, L, N.

After adopting such scheme; The present invention is through obtaining the circuit performance parameters of Power Electronic Circuit under different component parameters; As training sample, off-line training QNN neural network makes up QNN neural network multiparameter identifier with circuit performance parameters, component parameter; Through real time on-line monitoring Circuits System signal, the circuit performance parameters time series in the acquisition cuicuit practical work process utilizes the RVM algorithm predicts to go out following certain circuit performance parameters constantly; At last, be input with following certain circuit performance parameters constantly, adopt the QNN neural network identifier, obtain following certain each component parameter value of circuit constantly.According to each component parameter of the moment in future that picks out, system failure situation is assessed, and can be judged failure cause.The present invention need not to set up the mathematical model of Power Electronic Circuit, can realize the online failure prediction of Power Electronic Circuit.

Description of drawings

Fig. 1 is a process flow diagram of the present invention.

Embodiment

Below will combine accompanying drawing, technical scheme of the present invention will be elaborated.

As shown in Figure 1, the present invention provides a kind of Power Electronic Circuit multiparameter identification of defective Forecasting Methodology based on RVM-QNN (Relevant Vector Machine-Quantum Neural Networks), comprises the steps:

(1) each components and parts in the Power Electronic Circuit is provided with different parameter values, monitors the circuit measuring point signal under the corresponding situation, and calculate the acquisition cuicuit performance parameter;

(2) the QNN neural network is trained as training sample with the circuit performance parameters that obtained, component parameter in the step (1), make up QNN neural network multiparameter identifier;

When making up QNN neural network multiparameter identifier, suppose that institute predict that the components and parts number is N in the Power Electronic Circuit, the component parameter number also is a N system, and circuit performance parameters is M individual (as exporting average voltage, electric current etc.), and then the building process of identifier is:

(21) the QNN neural network structure is three layers; Wherein input layer is M; The output layer node is N, and middle layer node is

(round numbers);

(22) produce the initialization weights by random function.Random function is the function that produces number, is used to produce random number, is prior art, can adopt the rand function among the Matlab to produce by equally distributed random number between (0,1) among the present invention.

(23) use training sample to network training, with the actual output computational grid output error of desired output and network, by the weights and the quantum interval of each layer of the minimum adjustment of network output error;

(24) will accomplish the training after the QNN network as the multiparameter identifier.

(3) in the circuit working process, every certain interval of time, monitoring system measuring point signal, counting circuit performance parameter, the time series of acquisition cuicuit performance parameter;

(4) utilize the RVM algorithm that circuit performance parameters is predicted, obtain following certain circuit performance parameters constantly;

The relevant of said RVM algorithm is provided with as follows:

A) with Gaussian wavelet basis function structure RVM kernel function:

The Gaussian wavelet basis function φ (t) that constitutes by the second derivative of Gaussian function, formula as follows.

$\mathrm{\φ}\left(t\right)=-\frac{1}{{\mathrm{\σ\π}}^{1/4}}(1-\frac{t^2}{{\mathrm{\σ}}^2})\exp (-\frac{t^2}{2{\mathrm{\σ}}^2})$

Wherein, t is a time variable.

With following formula Gaussian wavelet basis function structure RVM kernel function, formula as follows.

$K(x,x_i)=\underset{i=1}{\overset{d}{\mathrm{\Π}}}\mathrm{\φ}\left(\frac{x-x_i}{a}\right)=\underset{i=1}{\overset{d}{\mathrm{\Π}}}-\frac{1}{{\mathrm{\σ\π}}^{1/4}}(2-\frac{{\left|\right|x-x_i\left|\right|}^2}{a^2{\mathrm{\σ}}^2})\exp (-\frac{{\left|\right|x-x_i\left|\right|}^2}{2a^2{\mathrm{\σ}}^2})$

Wherein, x is all training samples, x _iBe i training sample, d is the training sample number, and α, σ are the kernel function parameter, and α is the kernel function width.

B) kernel function parameter alpha=0.6, σ=1.

Utilize the method for RVM algorithm predicts circuit performance parameters following:

(41) performance parameter of circuit is X _i, i=1,2, L, M, the performance parameter time series of circuit is X _{I, 1}, X _{I, 2}, X _{I, 3}..., then obtain the performance parameter X of institute's prediction circuit _iP performance parameter sequence constantly is { X continuously _{I, j}, X _{I, j+1}, L, X _{I, j+p}, j=1,2,3, L, p=1,2,3, L;

(42) utilize RVM to { X _{I, j}, X _{I, j+1}, L, X _{I, j+p}Carry out forward direction q step prediction, obtain (j+p+q) performance parameter X of circuit constantly _{I, j+p+q}, q=1,2,3, L.

(5) be input with the circuit performance parameters of being predicted, utilize the QNN neural network identifier after the training in the step (2), following certain each component parameter of circuit is constantly obtained in identification;

Specifically, with X _{I, j+p+q}For input (i=1,2, L M), picks out (j+p+q) each component parameter θ constantly _{R, j+p+q}, r=1,2, L, N.

(6) according to each component parameter of circuit of being predicted, system is carried out the valuation condition evaluation.

Above embodiment is merely explanation technological thought of the present invention, can not limit protection scope of the present invention with this, every technological thought that proposes according to the present invention, and any change of on the technical scheme basis, being done all falls within the protection domain of the present invention.

Claims (4)

1. the Power Electronic Circuit multiparameter identification of defective Forecasting Methodology based on RVM-QNN is characterized in that comprising the steps:

(1) each components and parts in the Power Electronic Circuit is provided with different parameter values, monitors the circuit measuring point signal under the corresponding situation, and calculate the acquisition cuicuit performance parameter;

(2) the QNN neural network is trained as training sample with the circuit performance parameters that obtained, component parameter in the step (1), make up QNN neural network multiparameter identifier;

(3) in the circuit working process, every certain interval of time, monitoring system measuring point signal, counting circuit performance parameter, the time series of acquisition cuicuit performance parameter;

(4) utilize the RVM algorithm that circuit performance parameters is predicted, obtain following certain circuit performance parameters constantly;

(5) be input with the circuit performance parameters of being predicted, utilize the QNN neural network identifier after the training in the step (2), following certain each component parameter of circuit is constantly obtained in identification;

(6), system is carried out the fault state assessment according to each component parameter of circuit of being predicted.

2. a kind of Power Electronic Circuit multiparameter identification of defective Forecasting Methodology as claimed in claim 1 based on RVM-QNN; It is characterized in that: in the said step (2); When making up QNN neural network multiparameter identifier, suppose that institute predict that the components and parts number is that N is individual in the Power Electronic Circuit, the component parameter number also is that N is individual; Circuit performance parameters is M, and then the building process of identifier is:

(21) the QNN neural network structure is three layers; Wherein input layer is M; The output layer node is N, and middle layer node is

round numbers;

(22) produce the initialization weights by random function;

(23) use training sample to network training, with the actual output computational grid output error of desired output and network, by the weights and the quantum interval of each layer of the minimum adjustment of network output error;

(24) will accomplish the training after the QNN network as the multiparameter identifier.

3. a kind of Power Electronic Circuit multiparameter identification of defective Forecasting Methodology based on RVM-QNN as claimed in claim 2 is characterized in that: utilize the method for RVM algorithm predicts circuit performance parameters following in the said step (4):

(41) performance parameter of circuit is X _i, i=1,2, L, M, the performance parameter time series of circuit is X _{I, 1}, X _{I, 2}, X _{I, 3}..., then obtain the performance parameter X of institute's prediction circuit _iP performance parameter sequence constantly is { X continuously _{I, j}, X _{I, j+1}, L, X _{I, j+p}, j=1,2,3, L, p=1,2,3, L;

(42) utilize RVM to { X _{I, j}, X _{I, j+1}, L, X _{I, j+p}Carry out forward direction q step prediction, obtain (j+p+q) performance parameter X of circuit constantly _{I, j+p+q}, q=1,2,3, L.

4. a kind of Power Electronic Circuit multiparameter identification of defective Forecasting Methodology based on RVM-QNN as claimed in claim 3 is characterized in that: in the said step (5), the method that following certain each component parameter of the moment is obtained in identification is: with X _{I, j+p+q}Be input, i=1,2, L, M picks out (j+p+q) each component parameter θ constantly _{R, j+p+q}, r=1,2, L, N.

Images