CN105136138A - X-ray pulsar photon signal identification method based on nuclear extreme learning machine - Google Patents

Abstract

The present invention discloses an X-ray pulsar photon signal identification method based on an extreme learning machine. A purpose of the present invention is mainly to solve the problems of long time consuming, large calculation amount and poor practicality in the existing pulsar signal identification algorithm. The achieving steps of the method comprise: 1, constructing training sample data and test sample data according to a pulsar photon signal probability distribution function; 2, training the training sample to obtain the classifier output function of the extreme learning machine; and 3, substituting the test sample data into the classifier output function of the extreme learning machine to obtain the classification label of the test sample so as to complete the pulsar signal identification. According to the present invention, the extreme learning machine method is used to replace a large number of the high-order spectrum calculations, such that the calculation amount is reduced, the pulsar photon signal identification speed is increased, and the method can be used for the pulsar navigation system.

Description

Based on the X-ray pulsar photon signal discrimination method of core extreme learning machine

Technical field

The invention belongs to navigation signal processing technology field, be specifically related to a kind of pulsar photon signal discrimination method, can be used for pulsar navigation system.

Background technology

Pulsar has good stability of period, and outside radiation signal antijamming capability is strong, it comprises the signal of each wave band, and what these features made X-ray pulsar navigate is embodied as in order to possibility, simultaneously for celestial navigation provides vast potential for future development.X-ray pulsar navigational system has the advantages that reliability is high, independence is strong, applied widely, can provide the information such as attitude, position, time, speed, also can provide unified space-time datum for the satellite navigation system such as GPS, the Big Dipper for spacecraft.

Utilizing the X-ray pulsar photon received to carry out pulsar signal identification time of arrival is the prerequisite of carrying out pulsar navigation system, and therefore set up one identification algorithm fast and effectively, the research of paired pulses star navigational system has huge promotion meaning.Conventional signal recognition method extracts the higher-order spectrum of pile-up pulse profile, then obtains proper vector, carry out relevant matches.Two spectrum is a kind of conventional Higher Order Cumulants spectrum, Chinese scholars have studied the method much utilizing double-spectrum analysis to carry out pulsar signal identification, these methods all achieve reasonable effect in pulsar signal identification, but the higher-order spectrum computing of random signal needs very high computation complexity, identification can not be carried out by paired pulses star signal fast, this makes the real-time of navigational system be greatly affected, and has significant limitation in actual applications.

Summary of the invention

The object of the invention is to the deficiency for above-mentioned prior art, a kind of X-ray pulsar photon signal discrimination method based on extreme learning machine is proposed, to reduce computation complexity greatly, improve System Discrimination speed, realize the photon signal identification of paired pulses star navigation positioning system.

Thinking of the present invention is: directly start with from the pulsar photon signal received, paired pulses star signal carries out identification, training sample data and test sample book data are built according to pulsar photon signal probability distribution function, by training the sorter output function obtaining extreme learning machine to training sample, sorter output function test sample book data being substituted into extreme learning machine obtains the class label of test sample book, completes the identification of paired pulses star signal.Its technical scheme comprises the steps:

(1) 11 pulsar outline datas are chosen from EPN website, extract the temporal information of each pulsar and amplitude information as profile information, carry fitting tool paired pulses star profile information with matlab and carry out Gauss curve fitting, obtain probability distribution function P (k) of pulsar photon signal in different time sections, k indicating impulse star photon number variable quantity;

(2) N number of training sample and N number of test sample book is built according to probability distribution function P (k) of pulsar photon signal wherein, x _i=[x _i1, x _i2..., x _in] ^t∈ R ⁿrepresent the input vector of extreme learning machine, t _brepresent the time interval, λ (t) represents total flow density, t _irepresent input vector x _icorresponding sample label;

(3) bring N number of training sample data into extreme learning machine to train, obtain output function f (x) of extreme learning machine sorter;

(4) input vector of test sample book data is substituted into output function f (x), try to achieve class label label (x) of test sample book data, complete the identification of paired pulses starlight subsignal.

The present invention starts with due to the direct pulsar photon signal from receiving, paired pulses star signal carries out identification, thus not only avoid a large amount of higher-order spectrums compared with prior art to calculate, and while obtaining similar identification accuracy rate, significantly reduce computation complexity, improve System Discrimination speed.

Accompanying drawing explanation

Fig. 1 is realization flow figure of the present invention;

Fig. 2 is with the classification experiments result figure of the present invention to 1100 pieces of noise-free samples;

Fig. 3 is the classification experiments result figure with the present invention, 1100 pieces being had to noise sample.

Embodiment

With reference to Fig. 1, the X-ray pulsar signal recognition method based on extreme learning machine of the present invention, comprises the steps:

Step 1: set up X-ray pulsar photon signal probability distribution function.

1a) choose 11 pulsar outline datas from EPN website, in matlab, extract the temporal information of each pulsar and amplitude information respectively as profile information;

1b) carry fitting tool with matlab and Gauss curve fitting is carried out to profile information, obtain the Gauss curve fitting coefficient of each pulsar signal profile wherein φ ₀represent initial phase, f _st () represents X-ray pulsar frequency, f _dt () represents Doppler shift;

1c) calculate effective noise flux density: λ _b=b η A, wherein, b represents background density, and η represents detector speed, and A represents concentrated area;

1d) calculate effective X-ray pulsar photon stream metric density: λ _s=s η A, wherein, s represents the light stream intensity of X-ray pulsar signal;

1e) by 1b)-1d) result substitute into following formula, obtain total flow density λ (t) of pulsar photon signal:

$\mathrm{\λ}\left(t\right)=b\·\mathrm{\η}\·A+(s\·\mathrm{\η}\·A)h({\mathrm{\φ}}_0+f_s(t-t_0)+{\∫}_{t_0}^t{f}_d(t\left)d\mathrm{\τ}\right);$

1f) according to total flow density λ (t), obtain the probability distribution function of X-ray pulsar photon signal based on nonhomogeneous Poisson process:

$P\left(k\right)=\frac{({\∫}_{t_0}^{t_0+\mathrm{\Δ}t}\mathrm{\λ}\left(t\right)dt{)}^k}{k!}\exp \{-{\∫}_{t_0}^{t_0+\mathrm{\Δ}t}\mathrm{\λ}\left(t\right)dt\}.$

Step 2 builds extreme learning machine training sample and test sample book according to pulsar photon signal probability distribution function.

The training sample of N number of extreme learning machine 2a) is built according to probability distribution function P (k) of pulsar photon signal wherein, x _i=[x _i1, x _i2..., x _in] ^t∈ R ⁿrepresent the input vector of training sample, t _brepresent the time interval, λ (t) represents total flow density, t _irepresent training sample input vector x _icorresponding sample label;

The test sample book of N number of extreme learning machine 2b) is built according to probability distribution function P (k) of pulsar photon signal wherein, x _a=[x _a1, x _a2..., x _an] ^t∈ R ⁿrepresent the input vector of test sample book, t _arepresent test sample book input vector x _acorresponding sample label.

Step 3: training sample is updated to the sorter output function that extreme learning machine obtains extreme learning machine.

3a) stochastic generation extreme learning machine input weight w _jwith biased b _j, j ∈ { 1,2 ... L}, by the input vector x of training sample _i, i ∈ 1,2 ... N} substitutes into following formula, obtains the hidden layer output matrix H of extreme learning machine:

$H=\left[\begin{array}{ccccc}g(w_1\·x_1+b_1)& ...& g(w_j\·x_1+b_j)& ...& g(w_L\·x_1+b_L)\\ \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}\\ g(w_1\·x_i+b_1)& ...& g(w_j\·x_i+b_j)& ...& g(w_L\·x_i+b_L)\\ \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}\\ g(w_1\·x_N+b_1)& ...& g(w_j\·x_N+b_j)& ...& g(w_L\·x_N+b_L)\end{array}\right]$

3b) according to optimization constraint principle and the KKT constraint condition of multi output ELM sorter, obtain the concealed nodes of extreme learning machine and the weight vector β of output node:

$\mathrm{\β}=H^T{(\frac{1}{C}+{\mathrm{HH}}^T)}^{-1}T$

Wherein, H ^trepresent the transposition of H, C represents constant factor, and T represents sample label t _ithe matrix of composition;

3c) weight vector β is substituted into following formula, obtains the output function of extreme learning machine sorter:

$f\left(x\right)=h\left(x\right)\mathrm{\β}=h\left(x\right)H^T{(\frac{1}{C}+{\mathrm{HH}}^T)}^{-1}T$

Wherein, h (x)=[g (w ₁x+b ₁) ... g (w _lx+b _l)], x represents input vector.

Step 4: test data is substituted into extreme learning machine sorter output function, obtains pulsar photon test sample book class label.

4a) by the input vector x of test data _asubstitute into the sorter output function of extreme learning machine, obtain functional value f (x _a)=[f ₁(x) ... f _l(x) ... f _m(x)] ^t, wherein, f _lx () represents the output valve of l output node of extreme learning machine, l ∈ { 1 ... m};

4b) using output valve maximum in output node as test sample book class label label (x), that is:

$label\left(x\right)=\underset{l\∈\{1,\mathrm{...},m\}}{max}{f}_l\left(x\right).$

Effect of the present invention can be carried out emulation experiment by Matlab paired pulses starlight sub stream signals and be further illustrated:

1, emulation experiment condition

Emulation of the present invention is at windows7, CPUIntel (R) core (TM) i5-2400, basic frequency 3.10GHZ, and software platform is that Matlab7.0.1 runs.Emulation experiment data choose 11 pulsar outline datas such as B0531+21, B1706-44, B1937+21, B1951+32 and B0045+33 from EPN website.

2, content and result is emulated

Emulation one, classify to 1100 pieces of noise-free samples with the present invention, result is as Fig. 2.

As can be seen from Figure 2, the present invention has good identification effect to muting pulsar photon signal, and only have 16 to be divided by mistake in 1100 samples, identification accuracy rate reaches 98.4560%.

Emulation two, have noise sample to classify with the present invention to 1100 pieces, result is as Fig. 3.

As can be seen from Figure 3, the present invention has good identification effect to noisy pulsar photon signal, has 22 to be divided by mistake in 1100 samples, and identification accuracy rate is 98.0018%.

Claims (5)

1., based on an X-ray pulsar photon signal discrimination method for extreme learning machine, comprise the steps:

(1) 11 pulsar outline datas are chosen from EPN website, extract the temporal information of each pulsar and amplitude information as profile information, carry fitting tool paired pulses star profile information with matlab and carry out Gauss curve fitting, obtain probability distribution function P (k) of pulsar photon signal in different time sections, k indicating impulse star photon number variable quantity;

(2) N number of training sample and N number of test sample book is built according to probability distribution function P (k) of pulsar photon signal;

(3) bring N number of training sample data into extreme learning machine to train, obtain output function f (x) of extreme learning machine sorter;

(4) input vector of test sample book data is substituted into output function f (x), try to achieve class label label (x) of test sample book data, complete the identification of paired pulses starlight subsignal.

2. the method according to claims 1, carries fitting tool paired pulses star profile information with matlab in wherein said step (1) and carries out Gauss curve fitting, carry out as follows:

1a) choose 11 pulsar outline datas from EPN website, in matlab, extract the temporal information of each pulsar and amplitude information respectively as profile information;

1b) carry fitting tool with matlab and Gauss curve fitting is carried out to profile information, obtain the Gauss curve fitting coefficient of each pulsar signal profile wherein φ ₀represent initial phase, f _st () represents X-ray pulsar frequency, f _dt () represents Doppler shift;

1c) calculate effective noise flux density: λ _b=b η A, wherein, b represents background density, and η represents detector speed, and A represents concentrated area;

1d) calculate effective X-ray pulsar photon stream metric density: λ _s=s η A, wherein, s represents the light stream intensity of X-ray pulsar signal;

1e) by 1b)-1d) result substitute into following formula, obtain total flow density λ (t) of pulsar photon signal:

$\mathrm{\λ}\left(t\right)=b\·\mathrm{\η}\·A+(s\·\mathrm{\η}\·A)h({\mathrm{\φ}}_0+f_s(t-t_0)+{\∫}_{t_0}^t{f}_d(t\left)d\mathrm{\τ}\right);$

1f) according to total flow density λ (t), obtain the probability distribution function of X-ray pulsar photon signal based on nonhomogeneous Poisson process:

$P\left(k\right)=\frac{({\∫}_{t_0}^{t_0+\mathrm{\Δ}t}\mathrm{\λ}\left(t\right)dt{)}^k}{k!}\exp \{-{\∫}_{t_0}^{t_0+\mathrm{\Δ}t}\mathrm{\λ}\left(t\right)dt\}.$

3. the method according to claims 1, builds N number of training sample and N number of test sample book according to probability distribution function P (k) of pulsar photon signal in wherein said step (2), carries out as follows

The training sample of N number of extreme learning machine 2a) is built according to probability distribution function P (k) of pulsar photon signal wherein, x _i=[x _i1, x _i2..., x _in] ^t∈ R ⁿrepresent the input vector of training sample, t _brepresent the time interval, λ (t) represents total flow density, t _irepresent training sample input vector x _icorresponding sample label;

The test sample book of N number of extreme learning machine 2b) is built according to probability distribution function P (k) of pulsar photon signal wherein, x _a=[x _a1, x _a2..., x _an] ^t∈ R ⁿrepresent the input vector of test sample book, t _arepresent test sample book input vector x _acorresponding sample label.

4. the method according to claims 1, bring N number of training sample data into extreme learning machine in wherein said step (3) and train, concrete steps are as follows:

3a) stochastic generation extreme learning machine input weight w _jwith biased b _j, j ∈ { 1,2 ... L}, obtains the hidden layer output matrix H of extreme learning machine:

$H=\left[\begin{array}{ccccc}g(w_1\·x_1+b_1)& ...& g(w_j\·x_1+b_j)& ...& g(w_L\·x_1+b_L)\\ \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}\\ g(w_1\·x_i+b_1)& ...& g(w_j\·x_i+b_j)& ...& g(w_L\·x_i+b_L)\\ \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}& \begin{array}{c}\·\\ \·\\ \·\end{array}\\ g(w_1\·x_N+b_1)& ...& g(w_j\·x_N+b_j)& ...& g(w_L\·x_N+b_L)\end{array}\right]$

Wherein, g (w _jx _i+ b _j) represent hidden layer activation function, x _irepresent the input vector of training sample, i ∈ 1,2 ... N};

3b) according to optimization constraint principle and the KKT constraint condition of extreme learning machine multi output sorter, obtain the concealed nodes of extreme learning machine and the weight vector β of output node:

$\mathrm{\β}=H^T{(\frac{1}{C}+{\mathrm{HH}}^T)}^{-1}T,$

Wherein, H ^trepresent the transposition of H, C represents constant factor, and T represents sample label t _ithe matrix of composition;

3c) weight vector β is substituted into following formula, obtains the output function of extreme learning machine sorter:

$f\left(x\right)=h\left(x\right)\mathrm{\β}=h\left(x\right)H^T{(\frac{1}{C}+{\mathrm{HH}}^T)}^{-1}T$

Wherein, h (x)=[g (w ₁x+b ₁) ... g (w _lx+b _l)], x represents input vector.

5. the method according to claims 1, in wherein said step (4), the input vector of test sample book data is substituted into output function f (x), try to achieve class label label (x) of test sample book data, carry out as follows:

4a) by the input vector x of test data _asubstitute into the sorter output function of extreme learning machine, obtain result f (x _a)=[f ₁(x) ... f _l(x) ... f _m(x)] ^t, wherein, f _lx () represents the output valve of l output node of extreme learning machine, l ∈ { 1 ... m};

4b) using output valve maximum in output node as test sample book class label label (x), that is:

$label\left(x\right)=\underset{l\∈\{1,\mathrm{...},m\}}{max}{f}_l\left(x\right).$