CN103729335A - Vibration signal characteristic parameter identification method - Google Patents

Abstract

The invention discloses a vibration signal characteristic parameter identification method. The method includes dividing a sample set into a training sample set and a detecting sample set, and using the testing sample test to train a RBF artificial neural network after optimization of the weights is performed; using the training convergence network to identify samples in the testing sample set. Thus, optionality of training sample selection is guaranteed, dispersion of the training sample among the whole samples is also guaranteed, the experimental sample selecting method is consistent to the statistic theory, and meanwhile, independence between testing results and sample set dividing method is guaranteed; the purpose of improving the accuracy of signal characteristic parameter identification is achieved.

Description

Vibration signal characteristics parameter identification method

Technical field

The present invention relates to security signal process field, particularly, relate to a kind of vibration signal characteristics parameter identification method.

Background technology

At present, artificial neural network has outstanding ability to express to Nonlinear Mapping, 1985, Powell has constructed the RBF function of multivariate interpolation, 1988, Broomhead and Lowe deduce interpolation calculation to calculate for neural, and RBF is applied to Design on Artificial Neural Networks, have constructed RBF Function Network.RBF network is a kind of three layer feedforward neural networks, and hidden layer activation function is radial symmetry kernel function.When input sample propagates into hidden unit space, this group kernel function has formed a group " base " of input sample.

RBF network structure is shown in accompanying drawing 1.Left end is input layer, completes feature vector, X is introduced to network.Centre is hidden layer, and it is connected completely with input layer, and weights are 1, and its effect is equivalent to input pattern to carry out linear transformation, and the pattern input data transformation of low-dimensional, in higher dimensional space, is beneficial to output layer and carries out Classification and Identification.Hidden layer node is chosen basis function as transfer function, and widely used is Gaussian function:

Φ _i(x)＝exp(-||x-c _i|| ²/2σ _i ²),i＝1,2,…,p

In formula, x is n dimension input vector; c _ibe the center of j basis function, there is the vector of same dimension with x; σ _ibe the variable of i perception, it has determined the width of this basis function around central point; P is hidden layer node number.|| x-c _i|| be vector x-c _inorm, it represents x and c _ibetween distance; Gaussian function is at c _ithere is a unique maximal value at place, along with || x-c _i|| increase, function decays to zero gradually.

In existing vibration signal characteristics parameter identification method, conventionally directly artificially vibration signal sample is divided into two groups, one group is used for network training, and another group is for network test.This division methods can not guarantee the randomness that sample set is divided, and the independence between test result and sample set division methods can not be guaranteed.Thereby cause the accuracy of existing signal characteristic parameter identification not high enough.

Summary of the invention

The object of the invention is to, for the problems referred to above, propose a kind of vibration signal characteristics parameter identification method, to realize the advantage that improves signal characteristic parameter identification accuracy.

For achieving the above object, the technical solution used in the present invention is:

A kind of vibration signal characteristics parameter identification method, comprises the following steps:

Step 1, in vibration signal sample set, choose at random a sample as current sample point, and this sample is put into training set;

Step 2, centered by above-mentioned current sample point, take r as radius, do a hypersphere;

Sample point in step 3, above-mentioned vibration signal sample set that above-mentioned hypersphere is covered, does not comprise centre of sphere sample point, puts into test set;

Step 4, in above-mentioned steps three, in vibration signal sample set, in remaining sample, choose at random again a sample point as current sample point, put into training set;

Step 5, repeating step two are to step 4, until the number of samples in training set exceedes the setting number percent of number of samples in sample set first, sample set is divided and finished; If sample is concentrated and there is no sample point when the sample number in training set does not reach sample set number setting number percent, reduce above-mentioned radius of hypersphere r, repeating step rapid two is to step 4, until satisfy condition;

Step 6, above-mentioned steps five is completed to the test sample book collection normalization of division;

The mean value of all samples of vibration signal sample set in step 7, calculation procedure one, i.e. center vector, then calculates the Euclidean distance of each vector and center vector, marks ultimate range;

The multidimensional normal distribution that step 8, structure are put centered by above-mentioned center vector is then chosen random number within the scope of the ultimate range of mark, as the initial weight of RBF artificial neural network in above-mentioned steps seven;

Test sample book set pair after step 9, six normalization of use above-mentioned steps has carried out the RBF artificial neural network of weights optimization trains;

Test sample book collection after step 10, use train the network of restraining to above-mentioned steps six normalization is identified.

According to a preferred embodiment of the invention, in above-mentioned steps six to the normalization of test sample book collection; The test sample book collection obtaining is ${\hat{x}}_i={({\hat{x}}_1,{\hat{x}}_2,...,{\hat{x}}_n)}^T,$ Normalization formula is:

$\hat{X}=\frac{X}{\left|\right|X\left|\right|}=[\frac{x_1}{\sqrt{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{x}_j^2}},...,\frac{x_n}{\sqrt{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{x}_j^2}}]$

Wherein x, x _nand x _jbe the concentrated element of test sample book before normalizing.

According to a preferred embodiment of the invention, in above-mentioned steps one, the probability density distribution of vibration signal sample set is: $f\left(x\right)=\frac{1}{\mathrm{\σ}\sqrt{2\mathrm{\π}}}\exp \{-\frac{x-\mathrm{\μ}}{2\mathrm{\σ}}\}.$

Technical scheme of the present invention has following beneficial effect:

Technical scheme of the present invention, sample set is divided into training sample set and test sample book collection, guaranteed the randomness that training sample is chosen, guaranteed again the dispersiveness of training sample in all samples, the test sample choosing method of coincidence statistics principle has guaranteed the independence between test result and sample set division methods simultaneously.On a database that has 1175 vibration signal samples, test, the recognition correct rate average out to 92.87% that uses classic method to obtain, the recognition correct rate average out to 97.04% that uses the method for the technical program to obtain, has improved 4.17%.Reached the object that improves signal characteristic parameter identification accuracy.

Below by drawings and Examples, technical scheme of the present invention is described in further detail.

Accompanying drawing explanation

Fig. 1 is existing RBF neural network structure schematic diagram;

Fig. 2 is RBF neural network model characteristic parameter identification process figure;

Fig. 3 is the process flow diagram of the vibration signal characteristics parameter identification method described in the embodiment of the present invention.

Embodiment

Below in conjunction with accompanying drawing, the preferred embodiments of the present invention are described, should be appreciated that preferred embodiment described herein, only for description and interpretation the present invention, is not intended to limit the present invention.

Embodiment mono-,

As shown in Figure 3, a kind of vibration signal characteristics parameter identification method, comprises the following steps:

Step 1, in vibration signal sample set, choose at random a sample as current sample point, and this sample is put into training set;

Step 2, centered by above-mentioned current sample point, take r as radius, do a hypersphere;

Sample point in step 3, above-mentioned vibration signal sample set that above-mentioned hypersphere is covered, does not comprise centre of sphere sample point, puts into test set;

Step 4, in above-mentioned steps three, in vibration signal sample set, in remaining sample, choose at random again a sample point as current sample point, put into training set;

Step 5, repeating step two are to step 4, until the number of samples in training set exceedes the setting number percent of number of samples in sample set first, sample set is divided and finished; If sample is concentrated and there is no sample point when the sample number in training set does not reach sample set number setting number percent, reduce above-mentioned radius of hypersphere r, repeating step rapid two is to step 4, until satisfy condition;

Step 6, above-mentioned steps five is completed to the test sample book collection normalization of division;

The mean value of all samples of vibration signal sample set in step 7, calculation procedure one, i.e. center vector, then calculates the Euclidean distance of each vector and center vector, marks ultimate range;

The multidimensional normal distribution that step 8, structure are put centered by above-mentioned center vector is then chosen random number within the scope of the ultimate range of mark, as the initial weight of RBF artificial neural network in above-mentioned steps seven;

Test sample book set pair after step 9, six normalization of use above-mentioned steps has carried out the RBF artificial neural network of weights optimization trains;

Test sample book collection after step 10, use train the network of restraining to above-mentioned steps six normalization is identified.

First in vibration signal sample set, choose at random a sample as current sample point, and this sample is put into training set; Centered by current sample point, take r as radius, do a hypersphere, like this, the sample point that is less than r with the space length of current sample point will be covered by hypersphere; The sample point (not comprising centre of sphere sample point) that hypersphere is covered is put into test set; In the residue sample of sample set, choose at random again a sample point as current sample point, put into training set; Take r as radius, do again a hypersphere, the sample point (not comprising centre of sphere sample point) that hypersphere is covered is put into test set, and then get back in sample set and get a sample point from residue sample, centered by it, r covers again as radius, so repeatedly until the number of samples in training set exceedes certain number percent of number of samples in sample set first, as 50%, sample set is divided and is finished.If sample is concentrated when the sample number in training set does not reach sample set number predetermined percentage, there is no sample point, illustrate that to distribute to the number of samples of test set on the high side, reduced radius of hypersphere r, repetition above-mentioned steps is until satisfy condition.

Then to the normalization of test sample book collection; Obtain ${\hat{x}}_i={({\hat{x}}_1,{\hat{x}}_2,...,{\hat{x}}_n)}^T,$ Be formulated as:

$\hat{X}=\frac{X}{\left|\right|X\left|\right|}=[\frac{x_1}{\sqrt{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{x}_j^2}},...,\frac{x_n}{\sqrt{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{x}_j^2}}]$

Calculate all

mean value, i.e. center vector, and calculate the Euclidean distance of each vector and center vector, marks ultimate range d _max; Suppose between each attribute of sample independently, probability density distribution is point centered by center vector, builds multidimensional normal distribution, then at d _maxin scope, choose the initial weight of random number as RBF artificial neural network.

Finally, use test sample set is trained the RBF artificial neural network that carries out weights optimization; Use the network of training convergence to identify the sample in test set.

As shown in Figure 2, the initial weight of RBF artificial neural network is extremely important to follow-up training, good initial weight can accelerating network the speed of convergence of training, poor initial weight can cause learning number of times to be increased, not even convergence.While initial weight being set, the method conventionally adopting is to give the random number in certain limit to it in the past.Although in most cases this method can obtain the training result of convergence, but its defect is initial weight and input sample non-correlation, different samples to be difficult to obtain optimum network training result, to make often to need in order obtaining a good result a large amount of revision tests.This method adopts the weights choosing method relevant to input sample to optimize weights, this initial weight choosing method guarantees that initial weight is distributed in the center of sample, guarantee again its discreteness in sample, overcome the defect of above-mentioned random selection initial weight.

Embodiment bis-,

The present embodiment is tested on the net at the iron of having laid optical fiber vibration sensing system.Tester has 9, wherein 3 women of 6 male sex.These 9 testers are brought into test environment one by one, independent iron net is carried out the free-hand climbing, extruding, beating of free way, back and forth rocks the invasion of totally 4 kinds of patterns or destroy action.And record everyone every kind pattern to 10 of iron net behaviors, and like this, the invasion of every kind of pattern or destruction action just have 9 × 10=90 time, and 4 kinds of patterns have (total sample number) 10 × 4 × 9=360 time.

In 10 behaviors of every kind of pattern at everyone to iron net, take out at random 3 times as training sample, all the other 7 times as recognition sample.Characteristic parameter adopts Mel frequency cepstral coefficient.Under above-mentioned the same terms, first adopt vector quantization and traditional RBF neural network recognition method, when using vector quantization method to identify, climbing, extruding, pat, back and forth rock these 4 kinds of recognition correct rates corresponding to pattern and be respectively 89.62%, 91.71%, 95.02%, 93.27%, use the corresponding result of traditional RBF neural network recognization to be respectively 90.01%, 91.98%, 96.81%, 94.53%.And characteristic parameter when identification that adopts the present invention to propose, accuracy is respectively 94.87%, 95.92%, 99.31%, 98.04%, than classic method, has on average improved 4.17%.

Finally it should be noted that: the foregoing is only the preferred embodiments of the present invention, be not limited to the present invention, although the present invention is had been described in detail with reference to previous embodiment, for a person skilled in the art, its technical scheme that still can record aforementioned each embodiment is modified, or part technical characterictic is wherein equal to replacement.Within the spirit and principles in the present invention all, any modification of doing, be equal to replacement, improvement etc., within all should being included in protection scope of the present invention.

Claims (3)

1. a vibration signal characteristics parameter identification method, is characterized in that, comprises the following steps:

Step 1, in vibration signal sample set, choose at random a sample as current sample point, and this sample is put into training set;

Step 2, centered by above-mentioned current sample point, take r as radius, do a hypersphere;

Sample point in step 3, above-mentioned vibration signal sample set that above-mentioned hypersphere is covered, does not comprise centre of sphere sample point, puts into test set;

Step 4, in above-mentioned steps three, in vibration signal sample set, in remaining sample, choose at random again a sample point as current sample point, put into training set;

Step 5, repeating step two are to step 4, until the number of samples in training set exceedes the setting number percent of number of samples in sample set first, sample set is divided and finished; If sample is concentrated and there is no sample point when the sample number in training set does not reach sample set number setting number percent, reduce above-mentioned radius of hypersphere r, repeating step rapid two is to step 4, until satisfy condition;

Step 6, above-mentioned steps five is completed to the test sample book collection normalization of division;

The mean value of all samples of vibration signal sample set in step 7, calculation procedure one, i.e. center vector, then calculates the Euclidean distance of each vector and center vector, marks ultimate range;

The multidimensional normal distribution that step 8, structure are put centered by above-mentioned center vector is then chosen random number within the scope of the ultimate range of mark, as the initial weight of RBF artificial neural network in above-mentioned steps seven;

Test sample book set pair after step 9, six normalization of use above-mentioned steps has carried out the RBF artificial neural network of weights optimization trains;

Test sample book collection after step 10, use train the network of restraining to above-mentioned steps six normalization is identified.

2. vibration signal characteristics parameter identification method according to claim 1, is characterized in that, in above-mentioned steps six to the normalization of test sample book collection; The test sample book collection obtaining is

${\hat{x}}_i={({\hat{x}}_1,{\hat{x}}_2,...,{\hat{x}}_n)}^T,$ Normalization formula is:

$\hat{X}=\frac{X}{\left|\right|X\left|\right|}=[\frac{x_1}{\sqrt{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{x}_j^2}},...,\frac{x_n}{\sqrt{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{x}_j^2}}]$

Wherein x, x _nand x _jbe the concentrated element of test sample book before normalizing.

3. vibration signal characteristics parameter identification method according to claim 1 and 2, is characterized in that, in above-mentioned steps one, the probability density distribution of vibration signal sample set is:

$f\left(x\right)=\frac{1}{\mathrm{\σ}\sqrt{2\mathrm{\π}}}\exp \{-\frac{x-\mathrm{\μ}}{2\mathrm{\σ}}\}.$

Images