CN114172770B

CN114172770B - Modulation signal identification method of quantum root tree mechanism evolution extreme learning machine

Info

Publication number: CN114172770B
Application number: CN202111423647.1A
Authority: CN
Inventors: 高洪元; 郭瑞晨; 崔志华; 程建华; 杜亚男; 陈梦晗; 刘亚鹏; 赵立帅; 武文道
Original assignee: Harbin Engineering University
Current assignee: Harbin Engineering University
Priority date: 2021-11-26
Filing date: 2021-11-26
Publication date: 2023-05-02
Anticipated expiration: 2041-11-26
Also published as: CN114172770A

Abstract

The invention provides a modulation signal identification method of a quantum root tree mechanism evolution extreme learning machine, which utilizes a weighted Myriad filter to inhibit impact noise, provides a quantum root tree mechanism for carrying out efficient solution, and breaks through some application limitations of the existing modulation signal identification method based on the evolution extreme learning machine. The modulation signal identification method of the quantum root tree mechanism evolution extreme learning machine designs the quantum root tree mechanism, can carry out high-precision solution on the weight and the threshold value of the extreme learning machine under impact noise, and effectively improves the modulation identification rate. Simulation experiments prove that the effectiveness of the modulation signal identification method of the quantum root tree mechanism evolution extreme learning machine under impact noise breaks through the application limitation of the traditional method that the performance is deteriorated or even fails under the impact noise and low signal-to-noise ratio environment, and compared with the traditional method, the identification rate is greatly improved.

Description

Modulation signal identification method of quantum root tree mechanism evolution extreme learning machine

Technical Field

The invention relates to a modulation signal identification method based on a quantum root tree mechanism in an impact noise environment, and belongs to the field of communication signal processing.

Background

In recent years, the automatic modulation and identification technology of communication signals is widely applied to the scenes of spectrum allocation, electronic countermeasure, cognitive radio and the like. In the military field, it is necessary to distinguish between various communication signals sent by enemy electronic equipment and modulation types of radar signals, and then the next demodulation, even monitoring and interference, can be performed. In the civil field, the modulation recognition technology is applied to the cognitive radio field, and can be matched with modules such as parameter estimation, signal demodulation and the like, so that radio interference is effectively avoided, and spectrum allocation is optimized.

With technological advancement and social development, the electromagnetic environment of wireless communication is gradually becoming more and more complex, the variety and frequency spectrum of signals are gradually expanding upwards, and various interference and noise exist in the environment, such as receiving noise in radar and satellite communication, and impulse noise generated by electrocardiosignals and interinterplanar gravitational fields and the like. This presents a significant challenge for the task of identifying the modulated signal due to the spike nature and thicker probability density function tail of the impact noise that is actually present, as well as the complexity of the electromagnetic environment. Most of the traditional modulation signal identification methods need to obtain priori information of signals in advance, and factors such as frequency offset, noise and the like can cause inaccurate parameter estimation or feature extraction, so that the identification accuracy of the modulation signal under low signal-to-noise ratio is not ideal. In addition, the traditional modulation recognition method mainly researches the problem of modulation recognition in Gaussian noise environment, and the original method performance is drastically reduced or even fails in impact noise environment.

With the continuous development of machine learning, technologies such as an artificial neural network and a support vector machine are gradually applied to a plurality of fields such as image processing and signal processing. Since the machine learning method does not need to manually design decision threshold or calculate complex likelihood functions, more and more students have applied machine learning to the modulation signal recognition method. The neural network has strong pattern recognition capability, and each node automatically and adaptively updates the weight and the threshold value, so that the complex nonlinear problem can be well processed. Thus, the identification of the modulated signal species may be achieved using a neural network as a classifier.

The modulation signal recognition method based on the neural network generally comprises three basic steps of feature extraction, network training and classification recognition. In the back propagation training process of the BP neural network, the training result and convergence condition of the network are highly dependent on the initial weight, the threshold value and the network structure. Therefore, how to determine the appropriate initial weights and thresholds becomes an important issue. The general BP neural network adopts a method of randomly initializing weights and thresholds before training, is easy to sink into local convergence in the training process, and has unsatisfactory recognition accuracy in a low signal-to-noise ratio or impact noise environment.

The BP neural network needs to iteratively adjust the weight and the threshold value of the network step by step in the training process, the calculated amount is large, and the BP neural network is easy to fall into local optimum. The method is favorable for realizing the real-time processing of the modulation signal identification, and has greater engineering application value.

However, the feature of randomly initializing hidden layer parameters by the extreme learning machine makes it necessary to have more hidden layer nodes than the BP neural network to achieve similar classification effects, which can lead to more complex networks and reduce the generalization capability of the network. Therefore, after determining the objective function, the intelligent optimization algorithm can be used for evolving the weights and the thresholds of the hidden layers of the input layer and the hidden layer of the extreme learning machine, so that the recognition accuracy of the modulation signal is improved. Therefore, the modulation signal identification method based on the quantum root tree mechanism evolution extreme learning machine under the research impact noise has important significance and value.

Through the search of the prior art documents, song Lihui and the like in the 'digital communication modulation identification based on extreme learning machine' published in laser journal (2016,37 (3): 119-122), the extreme learning machine is utilized to realize the type identification of 7 digital modulation signals under Gaussian noise, but the influence of impact noise is not considered, the extreme learning machine is not evolved, and the optimal parameters are difficult to obtain; zhang Hui in the "research of communication signal modulation recognition method based on machine learning" published by the university of Harbin engineering, 2016, the particle swarm algorithm and principal component analysis evolution extreme learning machine parameters and structures are utilized, so that the recognition rate is improved under Gaussian noise, but the influence of the impact noise environment is not considered.

The retrieval results of the existing documents show that the existing modulation signal identification method based on the evolution extreme learning machine is mostly realized in a Gaussian noise environment, and performance is deteriorated in an impact noise environment, so that the modulation signal identification method based on the quantum-root-tree mechanism evolution extreme learning machine is provided, the impact noise is restrained by a weighted Myriad filter in the overall process, feature extraction is performed on the basis, and then parameters of the quantum-root-tree mechanism evolution extreme learning machine are used for solving the problem that the performance of the existing modulation signal identification method based on the evolution extreme learning machine is reduced in the impact noise environment.

Disclosure of Invention

Aiming at the defects and shortcomings of the existing modulation signal identification method based on the evolution extreme learning machine, the invention designs the modulation signal identification method based on the quantum root tree mechanism under impact noise, the method utilizes a weighted Myriad filter to inhibit impact noise, and a quantum root tree mechanism is provided for carrying out efficient solution, so that some application limitations of the existing modulation signal identification method based on the evolution extreme learning machine are broken through.

The purpose of the invention is realized in the following way: the method comprises the following steps:

step one, obtaining communication modulation signals and signal preprocessing, and obtaining a modulation signal preprocessing data set under an impact noise background. Signal preprocessing includes shaping filtering, power normalization, adding impulse noise and suppression.

And secondly, adopting a weighted Myriad filter to inhibit impact noise, and obtaining a modulated signal preprocessing data set through segmentation processing.

And thirdly, extracting instantaneous characteristic parameters from the modulated signal pretreatment data set to obtain a characteristic data set for training an extreme learning machine of the extreme learning machine.

And step four, determining an objective function of the optimal parameters of the extreme learning machine.

And fifthly, initializing quantum root tree mechanism parameters.

And step six, calculating the fitness and the wettability of all roots in the population, and arranging the population according to the ascending order of the wettability.

And seventhly, updating different individuals in the population by adopting the simulated quantum revolving doors.

And step eight, calculating the fitness and the wettability of the new generation of roots, updating the globally optimal roots, and arranging the populations according to the ascending order of the wettability.

Step nine, judging the iteration times

Whether or not the maximum number of iterations G is reached _max If the maximum iteration times are reached, ending the iteration and outputting an optimal weight and a threshold vector; otherwise, returning to the step seven.

And step ten, using an extreme learning machine with optimal weight and threshold as a classifier to identify the modulation signal under the impact noise background. The method comprises the steps of obtaining an optimal weight and a threshold value through a quantum root tree mechanism evolution extreme learning machine, taking the optimal weight and the threshold value as an initial weight and the threshold value of the extreme learning machine, training by utilizing training set data, taking the trained extreme learning machine with the optimal weight and the threshold value as a classifier for modulating signal recognition under an impact noise background, and finally outputting a modulating recognition result by adopting a test set or collected data.

Further, the first step specifically includes: adding a shaping filter at the transmitting end, wherein the shaping filter adopts a raised cosine roll-off function to shape the digital baseband signal, and the expression is

-3T < 3T, where T is the sampling time, μ is the raised cosine roll off coefficient, and T is the symbol period. And then carrying out power normalization, and normalizing the average power of the signals of each modulation mode to be 1.

Adopting Alpha stable distribution theory S _α (beta, gamma, delta) constructing an impact noise simulation model, wherein alphaThe characteristic index is called as characteristic index, and represents the impact degree of Alpha stable distribution, the value range is 0 < Alpha less than or equal to 2, and the impact degree is larger when Alpha is smaller. Beta is called a symmetrical parameter or a skew factor, represents the skew degree of Alpha stable distribution relative to the center, the value range is-1, and if beta is less than 0, the Alpha stable distribution is negative skew distribution; if β > 0, the distribution is forward skew. Gamma is called a scale parameter and represents the deviation degree of Alpha stable distribution relative to the center, the value range is gamma > 0, and the larger gamma is, the larger the deviation degree of the Alpha from the center is. α1 is called a position parameter, representing the position of Alpha stable distribution, the value range is- +_delta < +_infinity, and when 1+_0 is less than or equal to 2, delta represents the average value; delta represents the median value when 0 < alpha.ltoreq.1.

Further, the second step specifically includes: given a set of N observation samples

And weight set->

Define input vector x= [ x ] ₁ ,x ₂ ,...,x _N ] ^T And weight vector->

For a given nonlinearity parameter K > 0, assume the random variable +.>

Independent of each other and subject to the position parameter θ and the scale parameter +.>

Can be obtained with a probability density function of +.>

Definitions->

Weighting Myriad causes likelihood function +.>

Maximum, can obtain

Definition of the definition

Introducing a function ρ (v) =ln (1+v ² ) Where v is an argument, then the weighted Myriad output is +.>

Let Q (θ) be the weighted Myriad objective function. Defining the derivative of the function ρ (v)>

Where v is an argument, weighted Myriad output +.>

Is a local minimum of Q (θ) and therefore +.>

Definition of the function->

Where v is an argument, introducing a positive function

Where i=1, 2, once again, N, then ∈>

The following conclusions are thus drawn: comprises->

The local minimum points of all objective functions Q (θ) within can be represented as paired inputs x _i Form of weighted mean, i.e. +.>

Definition map->

The local minimum points of Q (θ), i.e., the root of Q' (θ) =0, can be considered as the points of T (θ), which are calculated using a fixed point iterative algorithm, i.e. +.>

Where m is the fixed point iteration number.

The segmentation process divides the modulation signal of each modulation mode into a plurality of data segments with equal length and a set form of labels corresponding to each data segment.

Further, the third step specifically includes: after the receiver receives the signal, it is subjected to Hilbert transform to obtain its resolved form, i.e

Wherein s (t) is the analysis signal of the original signal y (t), and +.>

Is the Hilbert transform of y (t), there is +.>

In (1) the->

Representing a convolution operation. Its frequency response is +.>

By sampling frequency f _s Sampling the original signal y (t) to obtain the total point number of

In the form of the discrete sequence y (n) of +.>

Instantaneous amplitude A (n)) Then

The instantaneous phase is θ (n), with +.>

Because the main value interval of the arctangent function is (-pi/2, pi/2), the theta (n) can generate + -pi mutation, and the phase with the value of [0,2 pi ] is obtained by adjusting the mutation

There is->

Actual instantaneous phase ε (n) and

the relation of (2) is->

Where mod represents the remainder operation. Thus (S)>

There is a phase wrap. Since the unwind-fold instantaneous phase phi (n) satisfies phi (n) =2pi f _c T _s n+ε (n) +θ, where f _c Is the carrier frequency, T _s The samples are periods and θ is the initial phase. From the above equation, the deconvolution instantaneous phase is a linear phase component caused by the carrier frequency and a nonlinear component caused by epsilon (n) and θ. The sequence is required->

The deconvolution is achieved by adding a correction sequence { c (n) }, defined as +.>

At this time, the deconvolution instantaneous phase estimation value is +.>

Under the condition of complete synchronization of carrier and code element, the estimated value of the non-linear component of the deconvolution instantaneous phase is +.>

The instantaneous frequency sequence can be obtained by differential deconvolution of instantaneous phase sequences, i.e

Wherein f _s Is the sampling frequency.

On the basis of obtaining the instantaneous amplitude, frequency and phase of the signal in the impact noise environment, further extracting a plurality of characteristic quantities of the instantaneous information of the digital modulation signal to obtain six characteristic parameters including the maximum value of the spectral density of the zero-center normalized instantaneous amplitude

Standard deviation sigma of zero-center normalized instantaneous amplitude absolute value _aa Standard deviation sigma of instantaneous phase nonlinear component of zero center non-weak signal segment _dp Standard deviation sigma of absolute value of instantaneous phase nonlinear component of zero center non-weak signal segment _ap Standard deviation sigma of zero center normalized non-weak signal segment instantaneous frequency absolute value _af Normalized instantaneous frequency variance +.>

The method comprises the steps of extracting characteristic parameters to obtain a data set containing six characteristic parameters, dividing the characteristic parameter data set into a training set and a testing set according to a certain proportion, and training an extreme learning machine for identifying digital modulation signals by using the training set.

Further, the fourth step specifically includes: the learning process of the extreme learning machine is as follows:

let the number of input layer nodes

The number of hidden layer nodes is l, and the output layer nodesThe number is m, and the weight matrix of the input layer and the hidden layer is +.>

There is->

In (1) the->

Is the connection weight of the i node of the input layer and the k node of the hidden layer. Setting the weight matrix of the hidden layer and the output layer as +.>

There is->

In (1) the->

Is the connection weight of the i node of the hidden layer and the k node of the output layer. Let the hidden layer threshold be b= [ b ] ₁ ,b ₂ ,...,b _l ] ^T Input feature matrix of extreme learning machine>

Comprising q samples, the corresponding desired output matrix is +.>

Has the following components

And->

The input vector corresponding to the ith sample is

The desired output vector is +.>

Let hidden layer node activation function be g (x), H be hidden layer output matrix, then

In (1) the->

Let the network output matrix be O, then o= [ O ] ₁ ,o ₂ ,...,o _q ] _m×q The output vector of the kth input sample is o _k There is

Under the conditions that the weight and the threshold are real numbers, the number of hidden layer nodes is the same as the number of samples, and the activation function is wireless and micro, the single hidden layer feedforward neural network can approximate the q sample feature vectors with 0 error, and the single hidden layer feedforward neural network comprises

In (1) the->

Thus, there is a single hidden layer feed-forward neural network such that

The q equations can be expressed as +.>

For a given sample

And->

The number of hidden layer nodes i of the single hidden layer feed-forward neural network is typically much smaller than the number of input samples q, so there may not be a single hidden layer feed-forward neural network that satisfies the above equation. For the traditional learning algorithm, a set of +.>

And b, minimizing errors, i.e

Wherein->

And->

Representing the input layer and hidden layer weights, the output layer threshold and the hidden layer and output layer weights, respectively, when the error is minimal, which equates to a minimization of the loss function +.>

In the extreme learning machine, the weights of the input layer and the hidden layer can be randomly initialized

And an implied layer threshold b, at a given sample +.>

In the case of (2), the hidden layer output matrix H is uniquely determined. Thus, extreme learning machine learning is equivalent to finding a linear system +.>

Least squares solution of (i.e.)

Its least-norm least-squares solution is +.>

In (1) the->

Is the Moore-Penrose generalized inverse of H.

And training the output of the prediction system after the extreme learning machine by using the characteristic training set, and taking the average absolute error between the predicted output and the expected output as an objective function. Let the number of input layer joints

The number of hidden layer nodes is l, the number of output layer nodes is m, and the number of samples is q, so that the optimal solution equation can be described as +.>

In (1) the->

For the expected output of the kth sample of the ith node of the network output layer, o _ik For the predicted output of the kth sample at the ith node of the output layer,

for the neural network weight and threshold vector, +.>

For the optimal network weight and threshold vector, M is the total number of the weight and threshold of the extreme learning machine, and is +.>

Further, the fifth step specifically includes: quantum root tree mechanismThe parameters were set as follows: population size of roots of

The quantum position dimension of each root is M, and the upper bound is set to u= [ U ] ₁ ,U ₂ ,...,U _M ]The lower bound is set to l= [ L ₁ ,L ₂ ,...,L _M ]Setting the maximum iteration number as G _max Iteration number->

Setting the ratio of the three update equations as R _r 、R _n And R is _c The adjustable parameters in the updated formula are c respectively ₁ 、c ₂ And c ₃ The variation probabilities when the quantum rotation angle is 0 are e respectively ₁ 、e ₂ And e ₃ . Randomly generating quantum positions of each root in a quantum position definition domain, wherein each dimension of quantum positions is limited to [0,1 ]]First->

The quantum position of the ith root of the next iteration is +.>

The corresponding position is +.>

And is also provided with

In (1) the->

L _d Is the lower bound of the d-th dimension position, U _d Is the upper bound of the d-th dimension position.

Further, the sixth step specifically includes: evaluation of the first

Adaptation of the ith root of the next iteration +.>

Mean absolute error is taken as a fitness function, therefore +.>

Wherein (1)>

Is->

Predictive output of kth sample of ith node of next iteration output layer,/th node of next iteration output layer>

Is the expected output of the kth sample of the ith node of the output layer. According to

Calculate->

The corresponding wettability of the ith root is iterated for a second time, and then the corresponding wettability is adjusted according to +.>

All roots in the population are arranged in ascending order, and the position of the globally optimal root is marked as +.>

The corresponding quantum position is marked as->

Further, the seventh step specifically includes: update Process 1, for the first less humid of the population

Root, th->

The quantum position d-th dimension updating formula of the ith root of the iteration is that

In the formula e ₁ The variation probability is 0,1/M]Constant between->

Is a random number with a value range between (0, 1), and is +>

D-th dimensional quantum position, which is the root of the previous generation random selection,>

is the corresponding quantum rotation angle. The d-th dimensional update formula of the quantum rotation angle is

Wherein randn is a value ranging from [ -1,1]A gaussian distributed random number in between.

Update process 2, for the first less humid of the population

Root, th->

In the formula e ₂ The variation probability is 0,1/M]Constant between->

D-th dimension quantum position, which is globally optimal root,>

is the corresponding quantum rotation angle. The d-th dimensional update formula of the quantum rotation angle is +.>

Update process 3, for the first less humid in the population

Root, th->

In the formula e ₃ The variation probability is 0,1/M]Constant between->

In (1) the->

Is the d-th dimension position of the root with global optimum fitness, rand represents [0,1 ]]A uniform random number therebetween.

Further, the step eight specifically includes: after the quantum positions of all the roots are updated, defining 'x' as multiplication of elements in corresponding dimensions of the front vector and the rear vector, and mapping the quantum position of each root into a position, wherein the mapping relation is that

Wherein the method comprises the steps of

First->

Amount of secondary iterationThe position of the ith root after the child position update is

Let the weight between the input layer and the hidden layer be +.>

Wherein f=n×l, the hidden layer threshold is +.>

Where m=n×l+l. Calculate->

The fitness of the ith root after the iterative update is +.>

Wherein->

Is->

Is the expected output of the kth sample of the ith node of the output layer. Calculating updated wettability according to wettability definition, there is +.>

Updating the location of a globally optimal root

And the corresponding quantum position->

The population is arranged according to the ascending order of wettability, and the iteration times are +.>

Compared with the prior art, the invention has the beneficial effects that:

(1) Aiming at the problem that the performance of the existing modulating signal recognition method based on the extreme learning machine is deteriorated in the impact noise environment, the modulating signal recognition method of the quantum root tree mechanism evolution extreme learning machine capable of effectively inhibiting the impact noise is designed, and the weighting Myriad filter is used for inhibiting the impact noise.

(2) The modulation signal identification method of the quantum root tree mechanism evolution extreme learning machine designs the quantum root tree mechanism, can carry out high-precision solution on the weight and the threshold value of the extreme learning machine under impact noise, and effectively improves the modulation identification rate.

(3) Simulation experiments prove that the effectiveness of the modulation signal identification method of the quantum root tree mechanism evolution extreme learning machine under impact noise breaks through the application limitation of the traditional method that the performance is deteriorated or even fails under the impact noise and low signal-to-noise ratio environment, and compared with the traditional method, the identification rate is greatly improved.

Drawings

Fig. 1 is a schematic flow diagram of a modulation signal recognition method of a quantum root tree mechanism evolution extreme learning machine designed by the invention.

FIG. 2 is a graph showing the comparison of the recognition accuracy of the method according to the present invention and the recognition accuracy of the original method.

Detailed Description

The invention is described in further detail below with reference to the drawings and the detailed description.

Referring to fig. 1 to 2, the steps of the present invention are as follows:

step one, obtaining communication modulation signals and signal preprocessing, and constructing a modulation signal data set under the impact noise background.

The types of communication modulation signals used in the present invention are 2ASK, 4ASK, 2PSK, 4PSK, 2FSK, 4FSK and MSK, respectively, and are not limited to these modulation schemes. Symbol rate f _d Carrier frequency f=38400 bit/s _c =408 kHz, carrier frequencies are 204kHz and 408kHz for 2FSK, respectively, and carrier frequencies are 4FSK, respectively102kHz, 204kHz, 306kHz and 408kHz. Sampling frequency f _s Time of sampling t = 3.264MHz ₀ =0.25 s, the number of samples per symbol is 85.

Adding a shaping filter at the transmitting end, wherein the shaping filter adopts a raised cosine roll-off function to shape the digital baseband signal, and the expression is

-3T < 3T, where T is the sampling time, μ is the raised cosine roll off coefficient, and T is the symbol period. And then carrying out power normalization to average power of signals of each modulation mode to be 1.

Adopting Alpha stable distribution theory S _α And (beta, gamma, delta) constructing an impact noise simulation model, wherein delta 1 is called a characteristic index and represents the impact degree of Alpha stable distribution, the value range is 0 < delta 3 < 2, and the impact degree is larger as Alpha is smaller. Beta is called a symmetrical parameter or a skew factor, represents the skew degree of Alpha stable distribution relative to the center, the value range is-1, and if beta is less than 0, the Alpha stable distribution is negative skew distribution; if β > 0, the distribution is forward skew. Gamma is called a scale parameter and represents the deviation degree of Alpha stable distribution relative to the center, the value range is gamma > 0, and the larger gamma is, the larger the deviation degree of the Alpha from the center is. δ0 is called a position parameter, representing the position of Alpha steady distribution, the value range is- +_δ2 < +_infinity, and when 1+_α is less than or equal to 2, δ4 represents the average value; delta represents the median value when 0 < alpha.ltoreq.1.

The impact noise parameter is set as: α=1.5, β=0, γ=1, δ=0. Measuring the relation between signal and noise intensity by using generalized signal-to-noise ratio GSNR, i.e

In (1) the->

Is the variance of the signal, γ is the scale parameter of the Alpha stability distribution, and GSNR ranges from-10 dB to 20dB,5dB spacing.

Given a set of N observation samples

And weight set->

Define input vector x= [ x ] ₁ ,x ₂ ,...,x _N ] ^T Sum weight vector w= [ w ] ₁ ,w ₂ ,...,w _N ] ^T For a given nonlinearity parameter K > 0, a random variable is assumed

Is distributed in Cauchy, and the probability density function is obtained

Definitions->

Weighting Myriad causes likelihood functions

Max, available->

Definitions->

Where v is an argument, weighted Myriad output +.>

Is a local minimum of Q (θ) and therefore +.>

Definition of the function->

Where v is an argument, introducing a positive function

Where i=1, 2, once again, N, then ∈>

The following conclusions are thus drawn: comprises->

Definition map->

Where m is the fixed point iteration number.

On the basis of obtaining the instantaneous amplitude, frequency and phase of the signal in the impact noise environment, a plurality of characteristic quantities of the instantaneous information of the digital modulation signal are further extracted, and 6 characteristic parameters are obtained.

Characteristic parameter 1. Spectral Density maximum of zero center normalized instantaneous amplitude

Wherein a is _cn (i) Normalizing the instantaneous amplitude for zero center, with a _cn (i)＝a _n (i)-1，

Wherein a (i) is the instantaneous amplitude of the signal, < >>

Is the length of the signal. Gamma ray _max The change characteristic of the instantaneous amplitude of the signal is reflected.

Characteristic parameter 2 standard deviation sigma of zero center normalized instantaneous amplitude absolute value _aa ，

σ _aa Reflecting the absolute amplitude information of the signal.

Characteristic parameter 3 standard deviation sigma of instantaneous phase nonlinear component of zero center non-weak signal segment _dp ，

Wherein a is _t For a set amplitude threshold, 1 is typically taken. />

Is zero center instantaneous phase nonlinear component with +.>

In (1) the->

Is the instantaneous phase of the signal. Sigma (sigma) _dp The characteristic of the change of the instantaneous phase of the signal is reflected.

Characteristic parameter 4. Standard deviation sigma of absolute value of instantaneous phase nonlinear component of zero center non-weak signal segment _ap ，

σ _ap The change characteristic of the instantaneous absolute phase of the signal is reflected.

Characteristic parameter 5. Standard deviation sigma of zero center normalized non-weak signal segment instantaneous frequency absolute value _af ，

Wherein f _cn (i) Is zero center normalized instantaneous frequency, has f _cn (i)＝f(i)/m _f -1，/>

f (i) is the instantaneous frequency. Sigma (sigma) _af The change characteristic of the instantaneous absolute frequency is reflected. Characteristic parameter 6. Variance of normalized instantaneous frequency +.>

Wherein f _n (i) Is the normalized instantaneous frequency, f _n (i)＝f(i)/m _f 。/>

The change characteristic of the instantaneous frequency is reflected.

The method comprises the steps of extracting characteristic parameters to obtain a data set containing six characteristic parameters, dividing the characteristic parameter data set into a training set and a testing set according to a ratio of 3:1, and training an extreme learning machine for identifying digital modulation signals by using the training set.

And training the output of the prediction system after the extreme learning machine by using the characteristic training set, and taking the average absolute error between the predicted output and the expected output as an objective function. Let the number of input layer nodes be n, the number of hidden layer nodes be l, the number of output layer nodes be m, the number of samples be q, the optimal solution equation can be described as

In (1) the->

for the neural network weight and threshold vector, +.>

For the optimal network weight and threshold vector, M is the total number of the weight and threshold of the extreme learning machine, and is the dimension of the quantum root tree mechanism, and there is +.>

And fifthly, initializing quantum root tree mechanism parameters.

The extreme learning machine parameters were set as follows: the number of nodes of the input layer is

The hidden layer node number l=20, the output layer node number m=7, and the hidden layer activation function is sigmoid.

The quantum root tree mechanism parameters are set as follows: population number

The dimension m=140 is calculated, and the upper bound is set to u= [ U ] ₁ ,U ₂ ,...,U _M ]＝[1,1,...,1]The lower bound is set to l= [ L ₁ ,L ₂ ,...,L _M ]＝[-1,-1,...,-1]Maximum iteration number G _max =100, iteration number->

The ratio of the three updating strategies is R _r ＝0.3、R _n =0.1 and R _c =0.6; adjustable parameters respectively positive c ₁ ＝30，c ₂ ＝c ₃ =10; variation probability e when quantum rotation angle is 0 ₁ ＝e ₂ ＝e ₃ =1/M. Randomly generating quantum positions of each root in a quantum position definition domain, wherein each dimension of quantum positions is limited to [0,1 ]]First->

The quantum position of the ith root of the next iteration is +.>

The corresponding position is +.>

And is also provided with

In (1) the->

L _d Is the d-th dimension lower bound, U _d Is the upper bound of dimension d.

Evaluation of the first

Adaptation of the ith root of the next iterationDegree->

The average absolute error is taken as a fitness function, thus

Wherein (1)>

Is->

The predicted output of the kth sample at the ith node of the output layer is iterated for a time,

is the expected output of the kth sample of the ith node of the output layer. According to->

Calculate->

The corresponding quantum position is marked as->

Update Process 1, for the first less humid of the population

Root, the first%>

The quantum position d-th dimensional update formula of the ith root of the next iteration is +.>

In the formula e ₁ The variation probability is 0,1/M]Constant between->

Is a random number with a value range between (0, 1), and is +>

Update process 2, for the first less humid of the population

Root, th->

In the formula e ₂ The variation probability is 0,1/M]Constant between->

D-th dimension quantum position, which is globally optimal root,>

Update process 3, for the first less humid in the population

Root, th->

In the formula e ₃ The variation probability is 0,1/M]Constant between->

In (1) the->

Is the d-th dimension position of the root with global optimum fitness, rand is [0,1]A uniform random number therebetween.

After the quantum positions of all the roots are updated, the quantum positions of each root are mapped into positions, and the mapping relation is that

Wherein the method comprises the steps of

"x" means the multiplication of elements in the corresponding dimension of the front and back vectors. First->

The position of the ith root after the secondary iteration quantum position is +.>

Let the weight between the input layer and the hidden layer be +.>

Where f=n×l, the hidden layer threshold is

Where m=n×l+l. Calculate->

The fitness of the ith root after the iterative updating is that

Wherein (1)>

Is->

Updating the position of the globally optimal root->

And correspond toQuantum position->

Step nine, judging the iteration times

In fig. 2, the modulation signal recognition method of the quantum root tree mechanism evolution extreme learning machine designed by the invention is named as WMy-QRTO-ELM, and the modulation signal recognition method of the original extreme learning machine is named as ELM.

The simulation experiment parameters of the modulation signal identification method of the quantum root tree mechanism evolution extreme learning machine are set as follows: the input layer node number of the extreme learning machine is 6, the hidden layer node number is 20, the output layer node number is 7, and the hidden layer activation function is sigmoid. Population size of quantum root tree mechanism

The search upper and lower bounds are [ -1,1]Calculating the total number M=140 of weights and thresholds, and the maximum iteration number G _max Ratio R of three update strategies =100 _r ＝0.3，R _n ＝0.1，R _c ＝0.6，c ₁ ＝30，c ₂ ＝c ₃ Variation probability e when quantum rotation angle is 0 =10 ₁ ＝e ₂ ＝e ₃ ＝1/M。

From fig. 2, it can be seen that after weighted Myriad filtering and quantum root tree mechanism evolution, the recognition rate under the condition of low generalized signal-to-noise ratio is greatly improved, and the application limit of the traditional method is broken through.

Claims

1. The modulation signal identification method of the quantum root tree mechanism evolution extreme learning machine is characterized by comprising the following steps:

step one: acquiring a communication modulation signal, and performing signal preprocessing to obtain a modulation signal preprocessing data set under an impact noise background; the signal preprocessing comprises shaping filtering, power normalization, impact noise addition and suppression;

step two: adopting a weighted Myriad filter to inhibit impact noise, and obtaining a modulated signal preprocessing data set through segmentation processing;

given a set of N observation samples

And weight set->

Define input vector x= [ x ] ₁ ,x ₂ ,...,x _N ] ^T And weight vector->

For a given nonlinearity parameter K > 0, assume the random variable +.>

Is distributed in Cauchy, and the probability density function is obtained

Definitions->

Weighting Myriad causes likelihood functions

Max, available->

Definition of the definition

Q (θ) is called the weighted Myriad objective function; defining the derivative of the function ρ (v)>

Where v is an argument, weighted Myriad output +.>

Is a local minimum of Q (θ) and therefore +.>

Definition of the function->

Where v is an argument, introducing a positive function

Where i=1, 2, once again, N, then ∈>

The following conclusions are thus drawn: comprises->

Definition map->

Where m is the number of fixed-point iterations;

the segmentation processing divides the modulation signal of each modulation mode into a plurality of data segments with equal length and a set form of labels corresponding to each data segment;

step three: extracting instantaneous characteristic parameters from the modulated signal preprocessing data set to obtain a characteristic data set for training an extreme learning machine;

after the receiver receives the signal, it is subjected to Hilbert transform to obtain its resolved form, i.e

Wherein s (t) is the analysis signal of the original signal y (t), and +.>

Is the Hilbert transform of y (t), there are

In (1) the->

Representing convolution operations with a frequency response of

In the form of its resolution is

Instantaneous amplitude A (n), then +.>

The instantaneous phase is θ (n), with +.>

There is->

Actual instantaneous phase ε (n) and

the relation of (2) is->

Where mod represents the remainder operation,

there is a phase wrap, since the unwrapped instantaneous phase φ (n) satisfies φ (n) =2πf _c T _s n+ε (n) +θ, where f _c Is the carrier frequency, T _s The sampling is the period, θ is the initial phase, and from the above equation, the deconvolution instantaneous phase is the linear phase component due to the carrier frequency and the nonlinear component due to ε (n) and θ, the sequence is required +.>

At this time, the deconvolution instantaneous phase estimation value is +.>

Wherein f _s Is the sampling frequency; />

Step four: determining an objective function of the optimal parameters of the extreme learning machine;

training the output of the prediction system after the extreme learning machine by using the feature training set, taking the average absolute error between the predicted output and the expected output as an objective function, and setting the number of nodes at the input layer as

In (1) the->

for the neural network weight and threshold vector, +.>

Step five: initializing a quantum root tree mechanism parameter;

the quantum root tree mechanism parameters are set as follows: population size of roots of

The quantum position dimension of each root is M, and the upper bound is set to u= [ U ] ₁ ,U ₂ ,...,U _M ]The lower bound is set to l= [ L ₁ ,L ₂ ,...,L _M ]Setting the maximum iteration number as G _max Number of iterations

Setting the ratio of the three update equations as R _r 、R _n And R is _c The adjustable parameters in the updated formula are c respectively ₁ 、c ₂ And c ₃ The variation probabilities when the quantum rotation angle is 0 are e respectively ₁ 、e ₂ And e ₃ The method comprises the steps of carrying out a first treatment on the surface of the Randomly generating a quantum position of each root in a quantum position definition domain, each dimension of quantum positionAre limited to [0,1 ]]First->

The quantum position of the ith root of the next iteration is

The corresponding position is +.>

And->

In (1) the->

L _d Is the lower bound of the d-th dimension position, U _d Is the upper bound of the d-th dimension position;

step six: calculating the fitness and wettability of all roots in the population, and arranging the population according to the ascending order of the wettability;

evaluation of the first

Adaptation of the ith root of the next iteration +.>

The average absolute error is taken as a fitness function, thus

Wherein (1)>

Is->

is the expected output of the kth sample of the ith node of the output layer; according to->

Calculate->

The corresponding quantum position is marked as->

Step seven: adopting a simulated quantum revolving door to update different individuals in the population respectively;

update Process 1, for the first less humid of the population

Root, th->

In the formula e ₁ The variation probability is 0,1/M]Constant between->

Is a random number with a value range between (0, 1), and is +>

is the corresponding quantum rotation angle; the d-th dimensional update formula of the quantum rotation angle is

Wherein randn is a value ranging from [ -1,1]A gaussian distributed random number in between;

update process 2, for the first less humid of the population

Root, th->

In the formula e ₂ The variation probability is 0,1/M]Constant between->

D-th dimension quantum position, which is globally optimal root,>

is the corresponding quantum rotation angle; the d-th dimensional update formula of the quantum rotation angle is +.>

Update process 3, for the first less humid in the population

Root, first

In the formula e ₃ The variation probability is 0,1/M]Constant between->

In (1) the->

Is the d-th dimension position of the root with global optimum fitness, rand represents [0,1 ]]A uniform random number therebetween;

step eight: calculating the fitness and the wettability of the new generation of roots, updating the globally optimal roots, and arranging the populations according to the ascending order of the wettability;

after the quantum positions of all the roots are updated, defining 'x' as multiplication of elements in corresponding dimensions of the front vector and the rear vector, and mapping the quantum position of each root into a position, wherein the mapping relation is that

Wherein the method comprises the steps of

First->

Let the weight between the input layer and the hidden layer be +.>

Wherein f=n×l, the hidden layer threshold is +.>

Wherein m=n×l+l; calculate->

The fitness of the ith root after the iterative update is +.>

Wherein->

Is->

Is the expected output of the kth sample of the ith node of the output layer; calculating updated wettability according to wettability definition, there is +.>

Updating the location of a globally optimal root

And the corresponding quantum position->

Step nine: judging the iteration times

Whether or not the maximum number of iterations G is reached _max If the maximum iteration times are reached, ending the iteration and outputting an optimal weight and a threshold vector; otherwise, returning to the step seven;

step ten: and identifying the modulation signal under the impact noise background by using an extreme learning machine with an optimal weight and a threshold value as a classifier, evolving the extreme learning machine through a quantum root tree mechanism to obtain the optimal weight and the threshold value, using the optimal weight and the threshold value as initial weight and threshold values of the extreme learning machine, training by using training set data, using the trained extreme learning machine with the optimal weight and threshold value as the classifier for identifying the modulation signal under the impact noise background, and finally outputting a modulation identification result by using a test set or collected data.