CN103926526A - Analog circuit fault diagnosis method based on improved RBF neural network - Google Patents

Abstract

The invention discloses an analog circuit fault diagnosis method based on an improved RBF neural network. The analog circuit fault diagnosis method includes the following steps that excitation is exerted on a circuit to be detected, and response signals are processed through improved wavelet packet transformation to extract fault characteristic signals; the extracted candidate characteristic signals are normalized to obtain fault characteristic vectors; the fault characteristic vectors serving as samples are input into the neural network and classified to obtain a result of fault diagnosis. Extraction of the fault characteristic vectors based on wavelet packet transformation is adopted, so that the distinguishability is improved; through normalization and other preprocessing, influences caused by different dimensions and too large numerical value difference on original variables are effectively eliminated; an LMS method in an RBF algorithm is replaced by a genetic optimization algorithm to train parameters of the neural network, so that the performance of the RBF algorithm is improved, an optimizing starting point of a genetic algorithm is set through a K average clustering learning algorithm, the iterations of the algorithm is effectively reduced, errors are reduced, diagnosis speed is increased, and the fault recognition rate is improved.

Description

A kind of analog-circuit fault diagnosis method based on improved RBF neural network

Technical field

The invention belongs to analog circuit fault diagnosing field, relate to a kind of analog-circuit fault diagnosis method based on improved RBF neural network.

Background technology

In the current information age, computing machine, communication, automated system isoelectronic series system and commercial production and daily life are closely bound up, circuit is the hardware foundation of electronic system, once circuit breaks down, electronic system and even whole system all cannot normally be worked.At present, electronic equipment and people's life is closely related, and oneself is widely applied to every field.And its running environment is varied, the severe even environment of very severe that cannot survive to the mankind from the common living environment of the mankind, as environment such as superhigh temperature, ultralow temperature, high humility, nuclear radiation, high electromagnetic fields.Along with electronic equipment extensively uses, people require more and more higher to the reliability index of electronic equipment, and to some special dimension as Aero-Space, require higher.Thereby require people to study new technology and method, further improve the reliability of electronic equipment, this just requires when circuit breaks down, can be in time, identification of defective accurately, so that maintenance and replacing.

At present, the fault diagnosis aspect of mimic channel exists following difficulty:

(1) lack simple fault model;

(2) impact of analog element parameter tolerances;

(3) in mimic channel, extensively there is nonlinear problem;

(4) the actual nodes surveyed is limited;

(5) backfeed loop in practical circuit causes emulation complicated.

Along with the development of artificial intelligence process technology, how artificial intelligence approach is combined with analog circuit fault diagnosing, become popular research direction at that time.After eighties of last century the nineties, neural network, expert system, fuzzy theory etc. are applied in analog circuit fault diagnosing gradually.Analog circuit fault has emerged a lot of methods, but the theory and the method that have proposed still exist many deficiencies, needs further perfect.

In fact analog circuit fault diagnosing is equivalent to a pattern-recognition and classification problem: according to the fault signature decision circuitry state of circuit, belong to which failure classes.Artificial neural network provides a good approach for analog circuit fault diagnosing, compares with conventional BP neural network, and RBF neural network all has greater advantage aspect Function approximation capabilities, classification capacity and pace of learning.But still there are some more difficult problems in the design of network structure.When hidden layer neuron number is less, its diagnostic result is poor, by increasing the number of hidden layer neuron, can improve discrimination.But the possibility of result causes network redundancy, increased the complicacy of network.

Summary of the invention

The above-mentioned technical matters existing for solving analog-circuit fault diagnosis method,, the object of the present invention is to provide a kind of analog-circuit fault diagnosis method based on improved RBF neural network.Utilize K mean cluster learning algorithm and genetic algorithm optimization RBF neural network parameter, set up fault dictionary, thereby realize the fault diagnosis to mimic channel.The method is obviously being better than conventional analog circuits fault neural network class diagnostic method aspect the discrimination of analog circuit fault diagnosing and speed.

For achieving the above object, the invention provides following technical scheme:

An analog-circuit fault diagnosis method based on improved RBF neural network, comprises the following steps:

Step 1: apply excitation to circuit under test, test node at circuit is measured exciter response signal, by the response signal (characteristics of low-frequency of fault-signal and high-frequency characteristic) of measuring, after making wavelet packet and improving the wavelet package transforms denoising Processing of energy, extract candidate's fault characteristic signals;

Step 2: extracted candidate feature vector is normalized, obtains fault feature vector;

Step 3: input in the RBF neural network after genetic algorithm optimization using fault feature vector as training sample, Training RBF Neural Network, make its error sum of squares be less than anticipation error, the hidden layer node number of RBF neural network is thought of as variable, optimize and select hidden layer node number, consider that performance of genetic algorithms is subject to initial point impact, utilizes K mean cluster learning algorithm that optimizing starting point is set;

Step 4: utilize and optimize the rear RBF neural network of training, combined training sample, sets up analog circuit fault dictionary, the network weight of the neural network after storage training, threshold value and error;

Step 5: when diagnosis is implemented, circuit-under-test is applied to pumping signal, extract corresponding fault signature, be input in the RBF neural network having trained, to by various diagnosing malfunctions and the identification of diagnostic circuit;

Step 6: the fault signature for not identifying in test, determine whether it is new fault signature (pattern), add training sample set.

Useful technique effect of the present invention is: in the present invention, adopt the extraction of the candidate's fault feature vector based on wavelet package transforms, improved the resolution of fault; By pre-service such as normalization, form fault signature, effectively eliminated the impact that former variable is different because of dimension and numerical value difference is brought too greatly, realized the extraction of fault signature; By substitute LMS method (minimum mean square error method) in RBF algorithm by genetic Optimization Algorithm, carry out the parameter (weights and threshold value) of neural network training, can improve the performance of RBF algorithm, utilize K mean cluster learning algorithm that the optimizing starting point of genetic algorithm is set simultaneously, effectively reduced the iterations of algorithm, reduce error, improved the discrimination of diagnosis speed and fault.

Accompanying drawing explanation

In order to make object of the present invention, technical scheme and beneficial effect clearer, the invention provides following accompanying drawing and describe:

Fig. 1 is the process flow diagram of method for diagnosing faults of the present invention

Fig. 2 is the structural drawing of RBF neural network

Fig. 3 is the block scheme of Training RBF Neural Network

Fig. 4 is the process flow diagram of the RBF neural network and genetic algorithm optimization method based on K mean cluster

Fig. 5 is the schematic diagram of real coding mode

Embodiment

Below in conjunction with accompanying drawing, the preferred embodiments of the present invention are described in detail.

In invention, adopt the extraction of the candidate's fault feature vector based on wavelet package transforms, improved the resolution of fault; By pre-service such as normalization, form fault signature, effectively eliminated the impact that former variable is different because of dimension and numerical value difference is brought too greatly, realized the extraction of fault signature; By substitute LMS method (minimum mean square error method) in RBF algorithm by genetic Optimization Algorithm, carry out the parameter (weights and threshold value etc.) of neural network training, can improve the performance of RBF algorithm, utilize K mean cluster learning algorithm that the optimizing starting point of genetic algorithm is set simultaneously, effectively reduced the iterations of algorithm, reduced error, improved diagnosis speed and the discrimination of fault.

Fig. 1 is the process flow diagram of method for diagnosing faults of the present invention.As shown in the figure, this RBF neuron network simulation circuit fault diagnosis comprises two courses of work, the implementation phase of setting up fault dictionary stage (learning process) and diagnosis.It is as follows that the present invention's employing realizes analog circuit fault diagnosing process based on genetic algorithm optimization RBF neural network:

1) mimic channel to be measured is applied to pumping signal, can survey node measurement exciter response signal;

2) the exciter response signal of measurement is made to the wavelet package transforms denoising Processing of wavelet packet and improvement energy, extracted the energy feature of each frequency inband signaling as proper vector;

3) proper vector of obtaining is input in the neural network through genetic algorithm optimization, the output of neural network is by the diagnostic result of diagnostic circuit.

When circuit breaks down, the customizing messages of its fault-signal is included in low frequency and high band part as slight change and distortion part, while extracting fault signature, with small echo and wavelet packet analysis, wavelet coefficient is walked around and is changed into energy as fault feature vector; While extracting fault by the wavelet package transforms that improves energy, the wavelet coefficient of each node is walked around and changed into energy as fault feature vector.

Fig. 2 is the RBF neural network structure of n input, a h hidden node, a m export structure.X=(x ₁, x ₂..., x _n) ^t∈ R ⁿfor neural network input vector, W ∈ R ^{h * m}for output weight matrix, the activation function of i hidden node is Φ _i(*), the Σ of output layer represents that neuronic activation function is linear function.H _idata center's vector of i hidden node in network, || * || represent 2-norm.The input of the hidden node of BP neural network adopts linear function, and activation function adopts the nonlinear functions such as Sigmoid function.And different from BP neural network are and are also that the outstanding feature of RBF neural network is that the input of the hidden node of RBF neural network adopts distance function, activation function is used radial basis function (as Gaussian function).Compare with BP neural network, RBF neural network not only has the nearly excitement of some biological neuron, the far away physiological Foundations suppressing, and structure it is simpler, pace of learning is also faster.

Fig. 3 is the process flow diagram of Training RBF Neural Network, using the proper vector extracting as training sample, be input in RBF neural network, utilize K mean cluster and genetic algorithm to improve RBF neural network, specific practice is: adopt K means clustering method to produce primary data center and width, then utilize genetic algorithm optimization data center and width, weights and skew are obtained by LMS method; In genetic algorithm, the generation of initial population adopts the method for K mean cluster to produce, rather than in scope, produces at random.In order to improve fault recognition rate, the number of the hidden layer node of neural network is thought of as variable, optimizes and selects hidden layer node number.

Fig. 4 is the RBF neural network and genetic algorithm optimization method process flow diagram based on K mean cluster, adopts following steps:

1) K means clustering method produces primary data center h _iwith width δ _i;

2) by the h of data center _iwith width δ _ireal coding, and produce initial population; The mode of the real coding that this method adopts, as shown in Figure 5, wherein the length of each coded strings is MH+H, wherein M is input number of nodes.

3) calculate fitness, utilize LMS (least mean-square error) to calculate weights, calculate MSE and fitness; Fitness function is taken as the inverse of square error, and to i chromosome, its fitness is:

$\mathrm{fitness}\left(i\right)=\frac{1}{\mathrm{mse}\left(e\right)}=\frac{1}{\frac{1}{\mathrm{PN}}\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(d(j,p)-y(j,p))}^2}---\left(1\right)$

Wherein, P is number of training, and N is output layer neuron number.

4) judge whether MSE (square error) meets error requirements, satisfied finish, otherwise continue;

5) adopt roulette wheel selection method, according to the size of fitness, select corresponding individual; If the fitness of i individuality is fitness (i), its selected probability is:

$P\left(i\right)=\frac{\mathrm{fitness}\left(i\right)}{\underset{j=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{fitness}\left(j\right)},i=\mathrm{1,2},\·\·\·,M---\left(2\right)$

In formula, M is Population Size.

6) adopt self-adapted genetic algorithm, adaptively selected crossing-over rate and aberration rate;

Crossing-over rate Pc and aberration rate Pm carry out self-adaptation adjustment by following formula:

$\mathrm{Pc}=\left\{\begin{array}{c}\mathrm{Pc}\max -\frac{(\mathrm{Pc}\max -\mathrm{Pc}\min )(f^{\&prime;}-\mathrm{favg})}{f\max -\mathrm{favg}},f^{\&prime;}\&GreaterEqual;\mathrm{favg}\\ \mathrm{Pc}\max ,f^{\&prime;}<\mathrm{favg}\end{array}\right.---\left(3\right)$

$\mathrm{Pm}=\left\{\begin{array}{c}\mathrm{Pm}\max -\frac{(\mathrm{Pm}\max -\mathrm{Pm}\min )(f^{\&prime;}-\mathrm{favg})}{f\max -\mathrm{favg}},f^{\&prime;}\&GreaterEqual;\mathrm{favg}\\ \mathrm{Pm}\max ,f^{\&prime;}<\mathrm{favg}\end{array}\right.---\left(4\right)$

Pcmax and Pcmin are respectively the upper and lower bound of crossing-over rate, Pmmax and Pmmin are respectively the upper and lower bound of aberration rate, f ' is larger fitness value in 2 individualities that will intersect, the average fitness that favg is population, and fmax is fitness value maximum in population.

7) adopt improved elitism system of selection, to guarantee the high-quality evolution of population.

By above 7 steps, can utilize genetic algorithm to realize the optimization to the RBF neural net method of K mean cluster.In the end in a step, adopt improved elitism system of selection, when the fitness that only has contemporary optimum solution is less than to previous generation, the optimum solution of previous generation is destroyed when to be population evolve toward " harm ", just by motionless the copying in the next generation of the optimum deblocking of current population; And the fitness of optimum solution in the present age is when population is evolved toward " benefit " while being more than or equal to previous generation, need not copy.With the optimum solution that prevents from producing in evolutionary process, by crossover and mutation, destroyed, guarantee that the high-quality of population is evolved.

Consider that performance of genetic algorithms is subject to initial point impact, utilize K mean cluster learning algorithm that optimizing starting point is set, as shown in Figure 4, the concrete steps that produce primary data center and width are:

1) algorithm initialization: select h different initial cluster center, and make k=1.H initial cluster center can be chosen at random, but is necessary for different values.

2) calculate the Euclidean distance of sample input and cluster centre:

||X _j-h _i(k)||,i＝1,2,…,h,j＝1,2,…,n ((5)

3) to sample input X _jby minimal distance principle, it is classified: work as

I(X _j)=min||X _j-h _i(k) ||, i=1,2 ..., during h, Xj is classified as i class,

X _j∈w _i(k) (6)

4) recalculate the cluster centre of each hidden node:

$h_i(k+1)=\frac{1}{N_i}\underset{x\&Element;w_i\left(k\right)}{\mathrm{\&Sigma;}}x,i=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;h---\left(7\right)$

In formula, N _ibe i Clustering Domain w _i(k) sample number comprising in.

5) if h _i(k+1) ≠ h _i(k), forward step 2 to), otherwise cluster finishes, and forwards step 6 to) be cluster centre while no longer changing, stop cluster.

6), according to the distance between each cluster centre, determine each primary data center and width.

δ _i＝κd _i (8)

D wherein _ibe the minimum distance between i data center and other data centers,

$d_i=\underset{j\&NotEqual;i}{\min }\left|\right|h_j-h_i\left(k\right)\left|\right|,$ κ is overlap coefficient.

When running into the fault signature that can not identify, determine whether new fault signature (pattern), add training sample to concentrate, constantly to improve fault dictionary.

Finally explanation is, above preferred embodiment is only unrestricted in order to technical scheme of the present invention to be described, although the present invention is described in detail by above preferred embodiment, but those skilled in the art are to be understood that, can to it, make various changes in the form and details, and not depart from the claims in the present invention book limited range.

Claims (6)

1. the analog-circuit fault diagnosis method based on improved RBF neural network, applies excitation to circuit under test, at the test node of circuit, measures exciter response signal; After the response signal of measurement being made to wavelet packet and improving the wavelet package transforms denoising Processing of energy, extract candidate's fault characteristic signals, then be normalized to extract fault characteristic information, using fault characteristic information as sample, the neural network of input after genetic algorithm optimization carried out failure modes again, consider that performance of genetic algorithms is subject to initial point impact, utilizes K mean cluster learning algorithm that optimizing starting point is set; Specifically comprise the following steps:

Step 1: apply excitation to circuit under test, test node at circuit is measured exciter response signal, by the response signal (characteristics of low-frequency of fault-signal and high-frequency characteristic) of measuring, after making wavelet packet and improving the wavelet package transforms denoising Processing of energy, extract candidate's fault characteristic signals;

Step 2: extracted candidate feature vector is normalized, obtains fault feature vector;

Step 3: input in the RBF neural network after genetic algorithm optimization using fault feature vector as training sample, Training RBF Neural Network, make its error sum of squares be less than anticipation error, the hidden layer node number of RBF neural network is thought of as variable, optimize and select hidden layer node number, consider that performance of genetic algorithms is subject to initial point impact, utilizes K mean cluster learning algorithm that optimizing starting point is set;

Step 4: utilize and optimize the rear RBF neural network of training, combined training sample, sets up analog circuit fault dictionary, for storing the network weight of the neural network after training, threshold value and error;

Step 5: when diagnosis is implemented, circuit-under-test is applied to pumping signal, extract corresponding fault signature, be input in the RBF neural network having trained, the output of RBF neural network is fault type;

Step 6: the fault signature for not identifying in test, determine whether it is new fault signature (pattern), add training sample set.

2. the analog-circuit fault diagnosis method based on improved RBF neural network according to claim 1, it is characterized in that: in step 1, when circuit breaks down, the customizing messages of its fault-signal is included in low frequency and high band part as slight change and distortion part, after the response signal of measurement being made to wavelet packet and improving the wavelet package transforms denoising Processing of energy, extracts candidate's fault characteristic signals.

3. the analog-circuit fault diagnosis method based on improved RBF neural network according to claim 1, it is characterized in that: the RBF neural network of the training described in step 3, in order to improve fault recognition rate, the number of the hidden layer node of neural network is thought of as variable, optimizes and selects hidden layer node number.

4. the analog-circuit fault diagnosis method based on improved RBF neural network according to claim 1, is characterized in that: described in step 3, utilize genetic algorithm optimization RBF neural net method, concrete steps are:

1) K means clustering method produces primary data center and width;

2) by the h of data center _iwith width δ _ireal coding, and produce initial population;

3) calculate fitness, utilize LMS to calculate weights, calculate MSE and fitness;

4) judge whether MSE meets error requirements, satisfied finish, otherwise continue;

5) adopt roulette wheel selection method, according to the size of fitness, select corresponding individual;

6) adopt self-adapted genetic algorithm, adaptively selected crossing-over rate and aberration rate;

7) adopt improved elitism system of selection, to guarantee the high-quality evolution of population.

5. the analog-circuit fault diagnosis method based on improved RBF neural network according to claim 1, it is characterized in that: the K mean cluster learning algorithm that utilizes described in step 3 arranges optimizing starting point, the concrete steps that produce primary data center and width K are:

1) algorithm initialization, chooses the cluster centre of h different value;

2) calculate sample input X _jwith cluster centre h _i(k) Euclidean distance;

3) to sample input X _jby minimal distance principle, it is classified;

4) recalculate the cluster centre h of each hidden node _i(k+1);

5) if h _i(k+1) ≠ h _i(k), forward step (2) to, otherwise cluster finishes, when forwarding step (6) to and be cluster centre and no longer changing, stop cluster;

6), according to the distance between each cluster centre, determine each primary data center and width.

6. an application rights requires the analog-circuit fault diagnosis method based on improved RBF neural network described in any one in 1 to 5.