CN102262211A - Analog circuit failure diagnosis method based on dynamic classification of echo state network - Google Patents

Abstract

An analog circuit failure diagnosis method based on dynamic classification of an echo state network relates to the analog circuit failure diagnosis method, which solves the problem that precision for analog circuit failure diagnosis by adopting the conventional neural network is relatively low. The method comprises the following steps of: exciting an analog circuit to work by using a unit pulse signal to acquire a response signal of the circuit to be diagnosed; acquiring a unit pulse response output signal of the analog circuit and taking the unit pulse response output signal as a failure data sample; inputting the failure data sample into the echo state network, training, and establishing an analog circuit failure diagnosis model according to a training result; and taking the acquired response signal of the circuit to be diagnosed as failure data and inputting the failure data into the analog circuit failure diagnosis model to obtain and output a failure diagnosis result. The method is applicable to the analog circuit failure diagnosis.

Description

Analog-circuit fault diagnosis method based on echo state network dynamic cataloging

Technical field

The present invention relates to a kind of analog-circuit fault diagnosis method.

Background technology

In electronic equipment, mimic channel is the weak link that the most easily breaks down, and mimic channel is carried out the maintainability that fault diagnosis can improve electronic equipment.Because mimic channel lacks good fault model, exist complicated nonlinear relationship and measuring point limited in number etc. between circuit response and component parameters, analog circuit fault diagnosing is studied prematurity still.In this case, be introduced in the analog circuit fault diagnosing based on the method for artificial intelligence, these class methods are regarded analog circuit fault diagnosing as pattern recognition problem.Owing to have good non-linear mapping capability, self study adaptive faculty etc., neural network is the most commonly used in the mimic channel intelligent diagnosing method.But traditional neural network is as adopting the multilayer perceptron of BP back-propagation algorithm training, exist problems such as easily being absorbed in local minimum, training algorithm complexity.

In intelligent diagnosing method, at first need the information that can be characterized circuit characteristic from obtaining the diagnostic circuit, promptly obtain the feature that circuit is showed under various duties.Usually, select value to change circuit output influence device is greatly injected the unit as fault, be the characteristic that abundant research circuit is shown under different more or less terms, be provided with that resistance and electric capacity are operated within the scope that allows tolerance in the circuit, be generally ± 5% or ± 10%.In the time of in the components and parts in the circuit all are operated in the permission tolerance, circuit belongs to unfaulty conditions; When any one of the device that injects the unit as fault is higher or lower than the certain limit of its normal value, and other devices are worked in allowing tolerance, think that then circuit breaks down.In order to obtain the job information of circuit under various states, generally to circuit input end input unit-pulse signal, and the unit impulse response signal of Acquisition Circuit.

For reflect the duty of circuit comprehensively, the sampling interval operated by rotary motion of output signal is less, sampling number is more, if adopt the static classification device as fault diagnosis model, signal processing methods such as general needs employing wavelet transformation carry out feature extraction to fault data, to reduce dimension fault data is converted to static nature, but characteristic extraction step have certain influence to the fault diagnosis performance, if feature extracting method is selected or improper use, will cause the fault diagnosis precision lower.

Summary of the invention

The present invention adopts traditional neural network to carry out the lower problem of diagnostic accuracy of analog circuit fault diagnosing in order to solve, thereby a kind of analog-circuit fault diagnosis method based on echo state network dynamic cataloging is provided.

Based on the analog-circuit fault diagnosis method of echo state network dynamic cataloging, it is realized by following steps:

Step 1, employing unit-pulse signal excitation simulation circuit working obtain circuit response signal to be diagnosed; Gather the unit impulse response output signal of mimic channel, and with described unit impulse response output signal as the fault data sample;

Step 2, the fault data sample that step 1 is obtained input in the echo state network and train, and set up the analog circuit fault diagnosing model according to training result;

Step 3, the circuit that step 1 is obtained wait to diagnose response signal as fault data, and input in the analog circuit fault diagnosing model of setting up in the step 2, obtain also output fault diagnosis result.

Setting up the analog circuit fault diagnosing model in the step 2 is to adopt the method for echo state network dynamic cataloging to realize.

The fault data sample that described in the step 2 step 1 is obtained inputs to the concrete grammar of training in the echo state network and is:

Steps A, parameter is set, described parameter comprises deposit pond processing unit number, inner connection weight spectral radius, the flexible yardstick of input and the sparse degree in deposit pond;

Step B, initialization echo state network, input connection weight matrix W ^InAnd inner connection weight matrix W;

Step C, the fault data sample is imported in the initialized echo state network, collected the state variable and the output variable of echo state network; Wherein, for state variable, only collect last state variable of each fault data sample;

Step D, find the solution output weight matrix W ^Out, obtain training result.

The connection weight matrix of input described in step B W ^InAnd the method for inner connection weight matrix W is: generate at random according to even distribution.

Among the step C, the state variable of described collection echo state network and the concrete grammar of output variable are: state variable and output variable are inputed to respectively in echo state network deposit pond processing unit activation function and the output unit activation function handle, the activation function that described echo state network deposit pond processing unit adopts is a hyperbolic tangent function, and the activation function that output unit adopts is an identity function.

The method that the state variable of described echo state network is collected is:

To each fault data sample, u (n)=(u ₁(n) ..., u _m(n)), with each data point u of u (n) ₁(n) ..., u _m(n) import formula successively:

x _i(n)＝tanh(W ⁱⁿu _i(n)+Wx _i-1(n))

The computing mode variable; In the formula, m is a fault data sample sequence length;

When i=m, calculate corresponding state variable x (n)=(x of u (n) ₁(n) ..., x _N(n)) ^T,, only collect last state variable of each fault data sample promptly for state variable; I=1,2 ..., K, K is an integer;

If the fault data sample size is a, the state variable x (n) of echo state network is collected into matrix M:

$M=\left[\begin{array}{c}{x}_1\left(1\right),x_2\left(1\right),...,x_N\left(1\right)\\ x_1\left(2\right),x_2\left(2\right),...,x_N\left(2\right)\\ ...\\ x_1\left(a\right),x_2\left(a\right),...,x_N\left(a\right)\end{array}\right]\∈R^{a\×N}$

The state variable that obtains the echo state network is collected the result; N is a positive integer, and R is a real number space.

The collection method of the output variable of described echo state network is:

Output variable is collected into matrix T:

T＝[d ₁，d ₂，...，d _a] ^T∈R ^a×L

In the formula, d ₁, d ₂..., d _aBe respectively the fault category sign of a fault data sample correspondence; L is the output unit number.

Find the solution output weight matrix W among the step D ^OutThe pseudo-algorithm for inversion of employing realize, promptly

W ^out＝(M ^-1T) ^T。

Beneficial effect: the present invention is an analytic target with the uniformly-spaced fault data that collects directly, adopt echo state network dynamic cataloging method to set up the analog circuit fault diagnosing model, than adopting traditional neural network to carry out analog-circuit fault diagnosis method, diagnostic accuracy of the present invention is higher.

Description of drawings

Fig. 1 is a schematic flow sheet of the present invention.

Embodiment

Embodiment one, in conjunction with Fig. 1 this embodiment is described, the described analog-circuit fault diagnosis method based on echo state network dynamic cataloging of present embodiment is realized by following steps:

Step 1, employing unit-pulse signal excitation simulation circuit working obtain circuit response signal to be diagnosed; Gather the unit impulse response output signal of mimic channel, and with described unit impulse response output signal as the fault data sample;

Step 2, the fault data sample that step 1 is obtained input in the echo state network and train, and set up the analog circuit fault diagnosing model according to training result;

Step 3, the circuit that step 1 is obtained wait to diagnose response signal as fault data, and input in the analog circuit fault diagnosing model of setting up in the step 2, obtain also output fault diagnosis result.

In the present embodiment, set up the analog circuit fault diagnosing model in the step 2 and can adopt the method for echo state network dynamic cataloging to realize.

Embodiment two, this embodiment are to do further qualification to embodiment one is described based on the step 2 in the analog-circuit fault diagnosis method of echo state network dynamic cataloging, in the described step 2, the fault data sample that step 1 is obtained inputs to the concrete grammar of training in the echo state network and is:

Steps A, parameter is set, described parameter comprises deposit pond processing unit number, inner connection weight spectral radius, the flexible yardstick of input and the sparse degree in deposit pond;

Step B, initialization echo state network, input connection weight matrix W ^InAnd inner connection weight matrix W;

Step C, the fault data sample is imported in the initialized echo state network collection status variable and output variable; Wherein, for state variable, only collect last state variable of each fault data sample;

Step D, find the solution output weight matrix W ^Out, obtain training result.

The connection weight matrix of input described in step B W ^InAnd the method for inner connection weight matrix W is: generate at random according to even distribution.

Embodiment three, this embodiment is to do further qualification to embodiment two is described based on the step C in the analog-circuit fault diagnosis method of echo state network dynamic cataloging, among the described step C, collecting the state variable of echo state network and the concrete grammar of output variable is: collection status variable and output variable are inputed to respectively in echo state network deposit pond processing unit activation function and the output unit activation function handle: the activation function that described echo state network deposit pond processing unit adopts is a hyperbolic tangent function, the activation function that output unit adopts is an identity function, does not adopt the connection of input layer to output layer, output layer is to the connection in deposit pond, output layer is to the connection of output layer.

Embodiment four, this embodiment are that embodiment three described methods of collecting based on the state variable of the echo state network described in the analog-circuit fault diagnosis method of echo state network dynamic cataloging are done further qualification, and the method that the state variable of described echo state network is collected is:

To each fault data sample, u (n)=(u ₁(n) ..., u _m(n)), with each data point u of u (n) ₁(n) ..., u _m(n) import formula successively:

x _i(n)＝tanh(W ⁱⁿu _i(n)+Wx _i-1(n))

The computing mode variable; In the formula, m is a fault data sample sequence length;

When i=m, calculate corresponding state variable x (n)=(x of u (n) ₁(n) ..., x _N(n)) ^T,, only collect last state variable of each fault data sample promptly for state variable; I=1,2 ..., K, K is an integer;

If the fault data sample size is a, the state variable x (n) of echo state network is collected into matrix M:

$M=\left[\begin{array}{c}{x}_1\left(1\right),x_2\left(1\right),...,x_N\left(1\right)\\ x_1\left(2\right),x_2\left(2\right),...,x_N\left(2\right)\\ ...\\ x_1\left(a\right),x_2\left(a\right),...,x_N\left(a\right)\end{array}\right]\∈R^{a\×N}$

The state variable that obtains the echo state network is collected the result; N is a positive integer, and R is a real number space.

Embodiment five, this embodiment are that the embodiment three described collection methods of exporting based on the state variable of the echo state network described in the analog-circuit fault diagnosis method of echo state network dynamic cataloging are done further qualification, and the collection method of the output variable of described echo state network is:

Output variable is collected into matrix T:

T＝[d ₁，d ₂，...，d _a] ^T∈R ^a×L

In the formula, d ₁, d ₂..., d _aBe respectively the fault category sign of a fault data sample correspondence; L is the output unit number.

Embodiment six, this embodiment are to do further qualification to embodiment two is described based on step D in the analog-circuit fault diagnosis method of echo state network dynamic cataloging, find the solution output weight matrix W among the described step D ^OutThe pseudo-algorithm for inversion of employing realize, promptly

W ^out＝(M ^-1T) ^T。

The principle of the echo state network described in the present invention is: the echo state network is the improvement to traditional recurrent neural network training algorithm.Be characterized in adopting the deposit pond of forming by the neuron of a large amount of sparse connections as hidden layer, in order to input is carried out higher-dimension, nonlinear expression, and only need train the weights of deposit pond to output layer, the training process of network is simplified, solved problems such as training algorithm complexity, network structure that traditional recurrent neural network exists are difficult to determine.

The typical structure of echo state network is made up of input layer, deposit pond and output layer.For by K input block, N deposit pond processing unit and L the echo state network that output unit is formed, its fundamental equation is as follows:

x(n+1)＝f(W ⁱⁿu(n+1)+Wx(n)+W ^backy(n)) (1)

y(n+1)＝f ^out(W ^out(u(n+1)，x(n+1)，y(n))) (2)

Wherein, x (n)=(x ₁(n) ..., x _N(n)) ^T, the state variable of expression echo state network.

Y (n)=(y ₁(n) ..., y _L(n)) ^T, the output variable of expression echo state network.

U (n)=(u ₁(n) ..., u _K(n)) ^TThe input variable of expression echo state network.

F=(f ₁..., f _N) for laying in the activation function vector of pond processing unit.

Activation function vector for deposit pond output unit.

Input block passes through

$W^{\mathrm{in}}=\left(w_{\mathrm{ij}}^{\mathrm{in}}\right)\∈R^{N\×K}$

Be connected W=(W with deposit pond processing unit _Ij) ∈ R ^{N * N}Be the connection weights between the processing unit of deposit pond,

$W^{\mathrm{back}}=\left(w_{\mathrm{ij}}^{\mathrm{back}}\right)\∈R^{N\×L}$

Be the connection weights of output layer to the deposit pond, the deposit pond is passed through

$W^{\mathrm{out}}=\left(w_{\mathrm{ij}}^{\mathrm{out}}\right)\∈R^{L\×(K+N+L)}$

Be connected with output unit.Wherein, W ^In, W and W ^BackNeed not training, after initial given, remain unchanged.

The method of echo state network training: the weight matrix W of input and output training sample data by generating at random ^InAnd W ^BackExcitation deposit pond processing unit, the sparse connection weight value matrix W between the processing unit of deposit pond also generates at random, adopts linear regression to make the minimized method of training square error promptly obtain W ^Out

Claims (8)

1. based on the analog-circuit fault diagnosis method of echo state network dynamic cataloging, it is characterized in that: it is realized by following steps:

Step 1, employing unit-pulse signal excitation simulation circuit working obtain circuit response signal to be diagnosed; Gather the unit impulse response output signal of mimic channel, and with described unit impulse response output signal as the fault data sample;

Step 2, the fault data sample that step 1 is obtained input in the echo state network and train, and set up the analog circuit fault diagnosing model according to training result;

Step 3, the circuit that step 1 is obtained wait to diagnose response signal as fault data, and input in the analog circuit fault diagnosing model of setting up in the step 2, obtain also output fault diagnosis result.

2. the analog-circuit fault diagnosis method based on echo state network dynamic cataloging according to claim 1, it is characterized in that setting up in the step 2 analog circuit fault diagnosing model is to adopt the method for echo state network dynamic cataloging to realize.

3. the analog-circuit fault diagnosis method based on echo state network dynamic cataloging according to claim 1 is characterized in that the fault data sample that described in the step 2 step 1 is obtained inputs to the concrete grammar of training in the echo state network and is:

Steps A, parameter is set, described parameter comprises deposit pond processing unit number, inner connection weight spectral radius, the flexible yardstick of input and the sparse degree in deposit pond;

Step B, initialization echo state network, input connection weight matrix W ^InAnd inner connection weight matrix W;

Step C, the fault data sample is imported in the initialized echo state network, collected the state variable and the output variable of echo state network; Wherein, for state variable, only collect last state variable of each fault data sample;

Step D, find the solution output weight matrix W ^Out, obtain training result.

4. the analog-circuit fault diagnosis method based on echo state network dynamic cataloging according to claim 3 is characterized in that the connection weight matrix of input described in step B W ^InAnd the method for inner connection weight matrix W is: generate at random according to even distribution.

5. the analog-circuit fault diagnosis method based on echo state network dynamic cataloging according to claim 3, it is characterized in that among the step C, the state variable of described collection echo state network and the concrete grammar of output variable are: state variable and output variable are inputed to respectively in echo state network deposit pond processing unit activation function and the output unit activation function handle, the activation function that described echo state network deposit pond processing unit adopts is a hyperbolic tangent function, and the activation function that output unit adopts is an identity function.

6. the analog-circuit fault diagnosis method based on echo state network dynamic cataloging according to claim 5 is characterized in that the method for the state variable collection of described echo state network is:

To each fault data sample, u (n)=(u ₁(n) ..., u _m(n)), with each data point u of u (n) ₁(n) ..., u _m(n)

Import formula successively:

x _i(n)＝tanh(W ⁱⁿu _i(n)+Wx _i-1(n))

The computing mode variable; In the formula, m is a fault data sample sequence length;

When i=m, calculate corresponding state variable x (n)=(x of u (n) ₁(n) ..., x _N(n)) ^T,, only collect last state variable of each fault data sample promptly for state variable; I=1,2 ..., K, K is an integer;

If the fault data sample size is a, the state variable x (n) of echo state network is collected into matrix M:

$M=\left[\begin{array}{c}{x}_1\left(1\right),x_2\left(1\right),...,x_N\left(1\right)\\ x_1\left(2\right),x_2\left(2\right),...,x_N\left(2\right)\\ ...\\ x_1\left(a\right),x_2\left(a\right),...,x_N\left(a\right)\end{array}\right]\∈R^{a\×N}$

The state variable that obtains the echo state network is collected the result; N is a positive integer, and R is a real number space.

7. the analog-circuit fault diagnosis method based on echo state network dynamic cataloging according to claim 5 is characterized in that the collection method of the output variable of described echo state network is:

Output variable is collected into matrix T:

T＝[d ₁，d ₂，...，d _a] ^T∈R ^a×L

In the formula, d ₁, d ₂..., d _aBe respectively the fault category sign of a fault data sample correspondence; L is the output unit number.

8. the analog-circuit fault diagnosis method based on echo state network dynamic cataloging according to claim 5 is characterized in that finding the solution among the step D output weight matrix W ^OutThe pseudo-algorithm for inversion of employing realize, promptly

W ^out＝(M ^-1T) ^T。

Images