Signal identification method of mixed feature sample expansion combined evolutionary LSTM (Long Short Term Memory) network

Abstract

The invention discloses a signal identification method combining mixed feature sample expansion with an evolution LSTM network, and the method comprises the steps: firstly extracting a plurality of mixed features, then carrying out the sample expansion through a sliding window, obtaining a feature sample suitable for an LSTM neural network, and solving a problem that a conventional LSTM neural network hyper-parameter depends on expert experience adjustment, and is difficult to obtain an optimal hyper-parameter. A quantum Arctic wolf mechanism is provided for efficient solving, and the application limitation that a traditional LSTM neural network depends on expert experience and is large in randomness due to the fact that the traditional LSTM neural network depends on manual selection of hyper-parameters is broken through.

Description

Technical Field

[0001] The present invention belongs to the field of communication signal processing, and in particular relates to a signal recognition method combining hybrid feature sample expansion with an evolutionary LSTM network. Background Art

[0002] Automatic modulation identification is a technology that automatically distinguishes the modulation type of the received signal. It is an intermediate step required for signal detection and information recovery. Automatic modulation identification technology for communication signals is one of the key technologies in scenarios such as spectrum allocation, electronic countermeasures, and cognitive radio. In the military field, it is necessary to distinguish the modulation types of various communication signals and radar signals emitted by enemy electronic equipment before the next step of demodulation, even monitoring and interference, can be carried out. In the civilian field, the application of modulation identification technology in the field of cognitive radio can be combined with modules such as parameter estimation and signal demodulation to effectively avoid radio interference and optimize spectrum allocation.

[0003] With the continuous development of machine learning, technologies such as decision trees and artificial neural networks have gradually been applied in many fields such as image processing and signal processing, including product detection, bankruptcy prediction, medical diagnosis, handwritten character recognition, automatic modulation recognition, etc. Since machine learning methods do not require manual design of decision thresholds or calculation of complex likelihood functions, more and more scholars have applied machine learning to modulation signal recognition methods. Evolutionary neural networks have powerful pattern recognition capabilities. Each node automatically and adaptively updates weights and thresholds, which can better handle complex nonlinear problems and achieve better recognition performance. Therefore, evolutionary neural networks can be used as classifiers to recognize the types of modulation signals.

[0004] Modulation signal recognition methods based on evolutionary neural networks generally include three basic steps: preprocessing and feature extraction, evolutionary network parameters or structure, and classification and recognition. Compared with machine learning classifiers such as BP neural networks, K nearest neighbors, decision trees and other algorithms, long short-term memory LSTM neural networks are good at learning the timing information in signals, so they can often achieve better results in modulation signal recognition. It is of great significance and value to study the modulation signal recognition method based on the quantum Arctic wolf mechanism evolutionary LSTM neural network under Gaussian noise.

[0005] Through the search of prior art literature, it was found that the invention patent authorized by Gao Hongyuan et al. in 2023, "A Modulated Signal Recognition Method for Evolved Long Short-Term Memory Neural Network" (Heilongjiang Province: CN202111421628.5, 2023-10-03.) directly used the received signal as the input of the LSTM neural network, and proposed a swordfish predation search mechanism to optimize the three hyperparameters of the LSTM neural network: the number of hidden layer neurons, the learning rate, and the number of training rounds. Good results were achieved, but the lack of feature extraction resulted in a large amount of data, and there were still other hyperparameters in the LSTM neural network that had an important impact on the training results. For example, the small batch size affected the memory consumption and the total number of weight updates during training, and the appropriate discard probability could control the complexity of the model. High network generalization ability; Liu Sheng et al. published "Modulation recognition method based on instantaneous feature parameters" in "Communication Technology" (2022, 55(05): 559-564), extracted three instantaneous features of the signal and used a three-layer fully connected neural network as a classifier to identify the modulated signal, but the type of features used was relatively single, resulting in poor performance under low signal-to-noise ratio; Zhang Wenna et al. published "Digital modulation recognition based on high-order cumulants" in "Foreign Electronic Measurement Technology" (2022, 41(07): 126-132), extracted five high-order cumulant features of the signal and used a decision tree classifier to identify the modulated signal. A single high-order cumulant feature was used, and the performance was not ideal under low signal-to-noise ratio without sample expansion, and it was difficult to achieve intra-class recognition of MFSK signals.

[0006] The search results of existing literature show that most of the existing modulation signal recognition methods based on neural networks require manual adjustment of network hyperparameters, which are highly random and heavily dependent on expert experience. Even if some technologies use optimization algorithms to optimize the hyperparameters of LSTM neural networks, the types of hyperparameters they look for are still insufficient, and some hyperparameters still need to be determined manually. In addition, in order to address the problem that traditional instantaneous or cumulative features do not utilize the timing information between samples and are not convenient for LSTM neural network training, a data preprocessing method using multiple mixed features for sample expansion is proposed to expand the feature samples into a data set suitable for LSTM neural network input. In order to address the problem that the hyperparameters of LSTM neural networks are difficult to determine, a quantum Arctic wolf mechanism is proposed to evolve the hyperparameters of LSTM networks, so as to solve the application limitation that the existing modulation signal recognition methods based on LSTM neural networks rely on manual selection and are difficult to obtain the optimal hyperparameters. Summary of the invention

[0007] The purpose of the present invention is to provide a signal recognition method combining hybrid feature sample expansion with evolutionary LSTM network.

[0008] The purpose of the present invention is achieved through the following technical solutions:

[0009] A signal recognition method combining hybrid feature sample expansion and evolutionary LSTM network, the steps are as follows:

[0010] Step 1: Establish a modulation signal sample library to obtain a modulation signal data set;

[0011] Step 2: Extract instantaneous features and cumulative features from the modulated signal data set obtained in step 1 to construct a mixed feature library;

[0012] Step 3: Preprocess the mixed features extracted in step 2;

[0013] Step 4: Determine the objective function;

[0014] The training set and the set hyperparameters are used to train the LSTM neural network, and the misjudgment rate of the LSTM network on the validation set is used as the objective function, that is, the fitness function F = 1-A c , where A c is the recognition accuracy on the validation set. After the training is completed, the validation set is used to verify the training effect.

[0015] Step 5: Initialize the parameters of the quantum Arctic wolf mechanism;

[0016] Including the number of individuals in the population N q , maximum number of iterations G max , the dimension of the search space M and the upper bound U = [u1,u2,…,u M ] and the lower bound L = [l1,l2,...,l M ], group behavior probability P, hunting attack probability Q in group behavior, persecution attack probability 1-Q in group behavior, update coefficient w1, mutation probability Selection probability Randomly initialize the quantum position of the Arctic wolf between [0,1] And mapped to the corresponding position where i∈{1,2,...,N q}, the symbol “⊙” represents the multiplication of elements at corresponding positions in the matrix or vector, and the superscript 0 represents the number of iterations is 0;

[0017] Step 6: Map the elements of each dimension of the position of the i-th Arctic wolf in the population to the hyperparameters of the LSTM network, i∈{1,2,...,N q}, use the training set and corresponding hyperparameters to train the LSTM network, and use the misjudgment rate of the LSTM network on the validation set as the fitness function of the i-th Arctic wolf;

[0018] Step 7: Enter the iteration loop, set t = t + 1, for the i-th Arctic wolf in the population, calculate the value of the variable m1(t) related to the number of iterations m1(t) = exp(t / Gmax -1), compared with the value of P, if it is less than P, it will enter group behavior; if m1(t) is greater than or equal to P, it will perform scavenging behavior;

[0019] Step 8: For the i-th Arctic wolf in the t-th iteration, if the survival probability Then perform survival behavior and update the quantum rotation angle to Where m2(t) = exp(1-t / G max )-t / G max , where: Indicates the random number of the ith Arctic wolf in the tth iteration that satisfies the uniform distribution between (0,1); otherwise, proceed to step nine;

[0020] Step 9: Update the quantum position of each Arctic wolf according to the quantum rotation angle and quantum gate strategy: Where: represents the dth dimension of the quantum position of the tth iteration of the i-th Arctic wolf, is the corresponding updated quantum rotation angle, is a random number that satisfies a uniform distribution between [0,1]. is the probability of quantum position mutation when the quantum rotation angle is 0;

[0021] Step 10: Calculate the position of the i-th Arctic wolf in the new generation as Convert the elements of each dimension into the corresponding LSTM network hyperparameters. After training the LSTM network, use the misjudgment rate on the validation set as the corresponding fitness function.

[0022] Step 11: Use the retention strategy to retain the updated Arctic wolf: if the updated Arctic wolf has better fitness, retain the corresponding position, quantum position and fitness after the update, otherwise do not retain; update the quantum position corresponding to the optimal individual and location Update the survival probability of each Arctic wolf for

[0023] Step 12: Determine whether the number of iterations t reaches the maximum number of iterations G max , if t=G max , then output the optimal Arctic wolf position As the optimization result, otherwise return to step 7 to continue iteration;

[0024] Step Thirteen: The elements of each dimension are converted into the corresponding optimal LSTM network hyperparameters, and the LSTM neural network with the optimal hyperparameters is trained. After the training is completed, the modulated signal recognition performance of the LSTM network is tested using the test set, and the results are analyzed.

[0025] Furthermore, the digital modulation signal in the step one includes MASK, MPSK, MFSK, MSK and MQAM. In order to reduce the influence of inter-symbol interference, a shaping filter is added to the MASK, MPSK, MFSK and MQAM signals at the transmitting end. The shaping filter uses a square root raised cosine roll-off function to shape the digital baseband signal. The signal is received at the receiving end through a Gaussian white noise channel to obtain a modulated signal data set.

[0026] Furthermore, the step 2 is specifically as follows:

[0027] After the receiver receives the signal, it performs Hilbert transform on it to obtain its analytical form, that is, Where s(t) is the analytical signal of the original signal y(t), j is a complex unit, and is the Hilbert transform of y(t), we have In the formula, represents the convolution operation; its frequency response is

[0028] With sampling frequency f s Sampling the original signal y(t), the total number of points is The discrete sequence y(n) of The instantaneous amplitude is A(n), then The instantaneous phase is θ(n), then

[0029] Since the main value interval of the inverse tangent function is (-π / 2,π / 2), θ(n) may produce a sudden change of ±π at this time, and it is adjusted to obtain a phase value in [0,2π) have

[0030] The actual instantaneous phase ε(n) is related to The relationship is In the formula, mod represents the remainder operation; therefore, There is phase wrapping; since the dewrapped instantaneous phase φ(n) satisfies φ(n)=2πf c T s n+ε(n)+θ, where f c is the carrier frequency, T s Sampling is the period, θ is the initial phase; from the above formula, we know that the deconvolution instantaneous phase is the linear phase component caused by the carrier frequency and the nonlinear component caused by ε(n) and θ; the sequence needs to be Add a correction sequence {c(n)} to achieve deconvolution, defined as At this time, the deconvolution instantaneous phase estimate is When the carrier and code element are completely synchronized, the estimated value of the de-warping instantaneous phase nonlinear component is

[0031] The instantaneous frequency sequence is obtained by differentiating the deconvolved instantaneous phase sequence, that is, In the formula, f s is the sampling frequency;

[0032] On the basis of the instantaneous amplitude, frequency and phase of the signal, multiple characteristic quantities of the instantaneous information of the digital modulation signal are further extracted to obtain six characteristic parameters, including the maximum value of the spectral density of the zero-centered normalized instantaneous amplitude. The standard deviation of the absolute value of the zero-centered normalized instantaneous amplitude σ aa , the standard deviation σ of the instantaneous phase nonlinear component of the zero-centered non-weak signal segment dp , the standard deviation σ of the absolute value of the instantaneous phase nonlinear component of the zero-centered non-weak signal segment ap , the standard deviation of the absolute value of the instantaneous frequency of the zero-centered normalized non-weak signal segment σ af , the variance of the normalized instantaneous frequency

[0033] After the receiver obtains the received baseband signal through down-conversion and matched filtering, five effective cumulant features Z1, Z2, Z3, Z4 and Z5 of the signal are extracted according to the high-order cumulant theory; suppose a random process x(t), whose k-order cumulant is defined as C kx (f1,f2,...,f k )=Cum[x(t),x(t+f1),...,x(t+f k-1 )], where: Cum(·) is the cumulative operation, which is expressed as the calculation of high-order mixing moments; the p-order mixing moment of the random process is M pq =E{x(t) p-q x * (t) q}, where: x*(t) represents the complex conjugate of the random process, q is the number of conjugate multiplications; the cumulants of each order of the random process are as follows: 1. The second-order cumulant is expressed as: C 20 =M 20 , C 21 =M 21 ; 2. The fourth-order cumulant is expressed as: C 41 =M 41 -3M 20 M 21 ; 3. The sixth-order cumulant is expressed as: In order to reduce the influence of signal energy on characteristic parameters, five characteristic parameters with better performance are extracted in the form of ratios as follows:

[0034] The signal is processed in segments, and characteristic parameters are extracted for each segment, eventually obtaining a mixed feature library containing 11 characteristic parameters.

[0035] Furthermore, the preprocessing of the mixed features in step 3 includes normalization, sample expansion, and division into a training set, a validation set, and a test set. The sample expansion uses a sliding rectangular window to treat three consecutive feature samples as input variables of an LSTM neural network.

[0036] Suppose the feature set composed of the extracted mixed features is M = [m1, m2, ..., m J ], where J is the total number of samples, and the i-th feature sample is Where: superscript T is the transposition symbol, superscript (i) represents the i-th sample;

[0037] The feature set is normalized using the normalization method of zero mean and unit variance: the mean of each feature in the feature set is calculated as μ = [μ1, μ2, ..., μ 11 ] T and standard deviation σ=[σ1,σ2,...,σ 11 ] T , then the normalized feature set is Where: the dot “. / ” means dividing all elements in each row of the matrix by the elements in the corresponding row of the vector;

[0038] Use a rectangular window with a length of 3 and a moving step of 1 to transform the feature set Truncate to data segments, i.e. The kth sequence sample is Where: Since the sliding window length is 3, the last two samples are incomplete, so the last two incomplete samples are discarded. That is, a feature sequence set N for LSTM network recognition is obtained, where the size of each sequence sample is 11×3; then, the feature sequence set N and the corresponding label T are randomly divided into a training set, a validation set, and a test set according to a certain ratio.

[0039] Furthermore, in step 6, the LSTM network hyperparameters to be optimized include the number of neurons in the hidden layer N h , small batch size S b , number of training rounds I m , learning rate l r and the drop probability p d , the learning rate is mapped to the logarithm of the variable ω4 = lg(l r ); Let the row vector composed of all hyperparameters be h = [N h ,S b,I m ,ω4,p d ]=[h1,h2,h3,h4,h5], let its upper bound be and the lower bound is Through the upper and lower bounds, a linear mapping is performed to the upper bound U and lower bound L of the search space of the quantum Arctic wolf mechanism, that is, in Will

[0040] Mapped vector As the position of the ith Arctic wolf in the tth iteration of the quantum Arctic wolf mechanism In the fitness function, the inverse transformation is used to obtain the corresponding LSTM network hyperparameters, that is, Where: i∈{1,2,...,N q},N q is the number of individuals in the population, d∈{1,2,...,M} is the dth hyperparameter, and the independent variable dimension is M=5; then the actual hyperparameters used for training are: the number of neurons in the hidden layer Small batch size Number of training rounds In the formula: round represents rounding off decimals; learning rate Drop probability

[0041] The LSTM network is trained using the hyperparameters corresponding to the i-th Arctic wolf in the t-th iteration, and the accuracy on the validation set is Then its fitness function is Then calculate the corresponding survival probability Right now Assume the number of iterations t=0.

[0042] Furthermore, in step 7, for the i-th Arctic wolf in the population, if m1(t)=exp(t / G max -1)<P, then group behavior is initiated, which includes hunting attack and persecution attack strategies. The hunting attack strategy is performed with probability Q and the quantum rotation angle is updated as Where: r i t is a normally distributed random number with mean 0 and variance 1 for the tth iteration of the i-th Arctic wolf, n a is in [2,N q / 2] represents the number of Arctic wolves to be updated. The quantum position of the kth Arctic wolf that attacks the group in the tth iteration is n randomly selected from the population in each iteration. a A group of different Arctic wolves, is the quantum position of the optimal Arctic wolf, c(t) = cos(πt / 2G max ) is the step coefficient that gradually changes from 1 to 0 as the iteration proceeds; the quantum rotation angle is updated with the probability 1-Q for the persecution attack strategy. And define the update step size Where: is the uniformly distributed random number between (0,2) for the ith Arctic wolf in the tth iteration. are the uniformly distributed random numbers between (0,2π) and (0,π) in the tth iteration, w1 is the update coefficient, v t (m) is the average value of the corresponding dimension of all the Arctic wolf quantum positions in the population at the tth iteration Otherwise m1(t)≥P performs scavenging behavior and updates the quantum rotation angle to Where: is the quantum position of a randomly selected Arctic wolf in the t-th iteration of scavenging behavior, and They represent the random numbers uniformly distributed between [-1,1] for the i-th Arctic wolf in the t-th iteration, p r is a random number that satisfies a uniform distribution between [0,1]. is the set selection probability, is the quantum position of the best Arctic wolf in the t-th iteration population, Represents the quantum position of a randomly selected Arctic wolf from the population in the tth iteration.

[0043] The beneficial effects of the present invention are:

[0044] (1) For the first time, the present invention uses the mixed features of the extracted 11 instantaneous features and cumulative features as a modulation signal recognition feature library after sample expansion, which greatly improves the recognition accuracy compared with other methods based on a single instantaneous feature or a single cumulative feature.

[0045] (2) The modulation signal recognition method of the LSTM neural network evolved by the quantum Arctic wolf mechanism designed by the present invention designs a quantum Arctic wolf mechanism, which can efficiently optimize the hyperparameters of the LSTM neural network and effectively improve the modulation recognition accuracy compared with the LSTM neural network with manually selected hyperparameters.

[0046] (3) Simulation experiments have demonstrated the effectiveness of the modulation signal recognition method of the LSTM neural network evolved by the quantum Arctic wolf mechanism under Gaussian noise, breaking through the application limitations of traditional LSTM neural networks that rely on expert experience and high randomness due to their reliance on artificially selected hyperparameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1Schematic diagram of the LSTM neural network unit structure adopted in the present invention.

[0048] Figure 2 Schematic diagram of the flow chart of the modulation signal recognition method designed by the present invention by expanding the hybrid feature samples and combining the LSTM neural network evolved by the quantum Arctic wolf mechanism.

[0049] Figure 3 The eleven hybrid feature sample expansion methods proposed in the present invention and the LSTM neural network evolved through the designed quantum Arctic wolf mechanism are compared with the single instantaneous feature or cumulative feature recognition accuracy in the comparative literature.

[0050] Figure 4 A comparison chart of the recognition accuracy of the signal recognition method proposed in the present invention using the hybrid feature sample expansion combined with the designed quantum Arctic wolf mechanism evolution LSTM neural network and the LSTM neural network modulation signal recognition method with artificially selected hyperparameters. DETAILED DESCRIPTION

[0051] The present invention is further described below in conjunction with the accompanying drawings.

[0052] according to Figure 1 and Figure 2 The present invention provides a signal recognition method combining hybrid feature sample expansion with evolutionary LSTM network, the steps are as follows:

[0053] Step 1: Establish a modulation signal sample library to obtain a modulation signal data set under Gaussian noise.

[0054] The modulation signals to be identified are 9 modulation signals, including 2ASK, 4ASK, 2PSK, 4PSK, 2FSK, 4FSK, MSK, 8QAM and 16QAM, and are not limited to these 9 modulation modes. d =38400baud, sampling frequency f s =3.072MHz, the number of sampling points per symbol is 80. Carrier frequency f c =408kHz, the carrier frequencies of 2FSK are 306kHz and 510kHz, respectively, with a frequency interval of 204kHz; the carrier frequencies of 4FSK are 255kHz, 357kHz, 459kHz and 561kHz, respectively, with a frequency interval of 102kHz. A shaping filter is added to the transmitter for MASK, MPSK, MFSK and MQAM signals. The shaping filter uses a square root raised cosine roll-off function to shape the digital baseband signal. The expression is: Where t is the sampling time, τ is the raised cosine roll-off coefficient, and T sis the symbol period. Set the roll-off factor to τ = 0.4 and the duration to 4 symbol periods. The signal passes through a Gaussian white noise channel with a signal-to-noise ratio range of -10dB to 10dB and an interval of 2dB. After receiving the signal, each modulated signal is divided into 40 symbol-length Multiple data segments for a point.

[0055] Step 2: Extract instantaneous features and cumulative features from the modulated signal data set to construct a hybrid feature library.

[0056] After the receiver receives the signal, it performs Hilbert transform on it to obtain its analytical form, that is, Where s(t) is the analytical signal of the original signal y(t), j is a complex unit, and is the Hilbert transform of y(t), we have In the formula, represents the convolution operation. Its frequency response is

[0057] With sampling frequency f s Sampling the original signal y(t) yields a discrete sequence y(n) with a total number of points N, whose analytical form is The instantaneous amplitude is A(n), then The instantaneous phase is θ(n), then

[0058] Since the main value interval of the inverse tangent function is (-π / 2,π / 2), θ(n) may produce a sudden change of ±π at this time, and it is adjusted to obtain a phase value in [0,2π) have

[0059] The actual instantaneous phase ε(n) is related to The relationship is Here, mod represents the remainder operation. Therefore, There is phase wrapping. Since the dewrapped instantaneous phase φ(n) satisfies φ(n)=2πf c T s n+ε(n)+θ, where f c is the carrier frequency, T s Sampling is the period, and θ is the initial phase. From the above formula, we can see that the deconvolution instantaneous phase is the linear phase component caused by the carrier frequency and the nonlinear component caused by ε(n) and θ. Add a correction sequence {c(n)} to achieve deconvolution, defined as At this time, the deconvolution instantaneous phase estimate is When the carrier and code element are completely synchronized, the estimated value of the de-warping instantaneous phase nonlinear component is

[0060] The instantaneous frequency sequence can be obtained by differentiating the deconvolved instantaneous phase sequence, that is, In the formula, f s is the sampling frequency.

[0061] On the basis of the instantaneous amplitude, frequency and phase of the signal, multiple characteristic quantities of the instantaneous information of the digital modulation signal are further extracted to obtain six characteristic parameters, including the maximum value of the spectral density of the zero-centered normalized instantaneous amplitude. The standard deviation of the absolute value of the zero-centered normalized instantaneous amplitude σ aa , the standard deviation σ of the instantaneous phase nonlinear component of the zero-centered non-weak signal segment dp , the standard deviation σ of the absolute value of the instantaneous phase nonlinear component of the zero-centered non-weak signal segment ap , the standard deviation of the absolute value of the instantaneous frequency of the zero-centered normalized non-weak signal segment σ af , the variance of the normalized instantaneous frequency

[0062] Characteristic parameter 1. Maximum value of the spectral density of the zero-centered normalized instantaneous amplitude In the formula, a cn (i) is the zero-centered normalized instantaneous amplitude, and a cn (i) = a n (i) -1, Where a(i) is the instantaneous amplitude of the signal, is the length of each signal segment. It reflects the changing characteristics of the instantaneous amplitude of the signal.

[0063] Characteristic parameter 2. Standard deviation σ of the absolute value of the zero-centered normalized instantaneous amplitude aa , σ aa Reflects the absolute amplitude information of the signal.

[0064] Characteristic parameter 3. Standard deviation σ of the instantaneous phase nonlinear component of the zero-centered non-weak signal segment dp , In the formula, a t It is the amplitude threshold to be set, which is usually 1. is the zero-centered instantaneous phase nonlinear component, Where: is the instantaneous phase of the signal. dp It reflects the changing characteristics of the signal's instantaneous phase.

[0065] Characteristic parameter 4. Standard deviation σ of the absolute value of the instantaneous phase nonlinear component of the zero-centered non-weak signal segment ap , σ ap Reflects the changing characteristics of the signal's instantaneous absolute phase.

[0066] Characteristic parameter 5. Standard deviation σ of the absolute value of the instantaneous frequency of the zero-centered normalized non-weak signal segment af , In the formula, f cn (i) is the zero-centered normalized instantaneous frequency, and f cn (i) = f n (i) -1, is the normalized instantaneous frequency, and f(i) is the instantaneous frequency. af It reflects the changing characteristics of the instantaneous absolute frequency.

[0067] Characteristic parameter 6. Variance of normalized instantaneous frequency It reflects the changing characteristics of instantaneous frequency.

[0068] After the receiver obtains the received baseband signal through down-conversion and matched filtering, five effective cumulant features Z1, Z2, Z3, Z4 and Z5 of the signal are extracted according to the high-order cumulant theory. Suppose a random process x(t), whose k-order cumulant can be defined as C kx (f1,f2,...,f k )=Cum[x(t),x(t+f1),...,x(t+f k-1 )], where Cum(·) is the cumulative operation, which can be expressed as the calculation of high-order mixing moments. The p-order mixing moment of a random process is M pq =E{x(t) p-q x * (t) q}, where: x*(t) represents the complex conjugate of the random process, and q is the number of times the conjugate is multiplied. The cumulants of each order of the random process are as follows: 1. The second-order cumulant can be expressed as: C 20 =M 20 , C 21 =M 21 2. The fourth-order cumulant can be expressed as: C 41 =M 41 -3M 20 M 21 The sixth-order cumulant can be expressed as: In order to reduce the influence of signal energy on characteristic parameters, five characteristic parameters with better performance are extracted in the form of ratios as follows:

[0069] The signal is processed in segments, and characteristic parameters are extracted for each segment, eventually obtaining a mixed feature library containing 11 characteristic parameters.

[0070] Step 3: Preprocessing of mixed features.

[0071] The preprocessing of mixed features includes normalization, sample expansion, and division into three parts: training set, validation set, and test set. Sample expansion uses a sliding rectangular window to take three consecutive feature samples as input variables of an LSTM neural network.

[0072] Suppose the feature set composed of the extracted mixed features is M = [m1, m2, ..., m J ], where J is the total number of samples, and the i-th feature sample is Where: superscript T is the transposition symbol, and superscript (i) represents the i-th sample.

[0073] The feature set is normalized using the normalization method of zero mean and unit variance: the mean of each feature in the feature set is calculated as μ = [μ1, μ2, ..., μ 11 ] T and standard deviation σ=[σ1,σ2,...,σ 11 ] T , then the normalized feature set is Where: The dot division “. / ” means dividing all elements in each row of the matrix by the elements in the corresponding row of the vector.

[0074] Use a rectangular window with a length of 3 and a moving step of 1 to transform the feature set Truncate to data segments, i.e. The kth sequence sample is Where: Since the sliding window length is 3, the last two samples are incomplete, so the last two incomplete samples are discarded. At this point, the feature sequence set N for LSTM network recognition can be obtained, where the size of each sequence sample is 11 × 3. Subsequently, the feature sequence set N and the corresponding label T are randomly divided into a training set, a validation set, and a test set according to a certain ratio.

[0075] Step 4: Determine the objective function.

[0076] The LSTM neural network is trained using the training set and the set hyperparameters. The misjudgment rate of the LSTM network on the validation set is used as the objective function. After the training is completed, the training effect is verified by the validation set to obtain the recognition accuracy A on the validation set. c , the misjudgment rate of the LSTM network on the validation set is used as the objective function, that is, the fitness function F = 1-A c .

[0077] The principle of LSTM network is as follows:

[0078] Figure 1This is a schematic diagram of the structure of a unit of the LSTM neural network. It contains three gates. Its gating mechanism mainly includes three parts: forget gate, input gate and output gate. The forget gate mainly controls what kind of historical information is remembered or forgotten, retaining information with output values ​​close to 1 and forgetting information with output values ​​close to 0. The output of the forget gate of the kth LSTM unit is f k =sigmoid(W f ⊙x k +W f ⊙h k-1 +b f ), where: sigmoid(x) = 1 / (1+e -x ) is an S-shaped function that limits the output to between (0,1), W f is the forget gate weight matrix, x k is the input of the current LSTM unit, “⊙” represents the multiplication of the elements at the corresponding positions of the matrix or vector, and h k-1 is the output of the previous LSTM unit, b f is the forget gate threshold matrix. The output of the input gate is i k =sigmoid(W i ⊙x k +W i ⊙h k-1 +b i ), where: W i is the input gate weight matrix, b i is the input gate threshold matrix. The input gate determines what information should be added to the memory cell. The output of the output gate is o k =sigmoid(W o ⊙x k +W o ⊙h k-1 +b o ), where: W o is the output gate weight matrix, b o is the output gate threshold matrix. The temporary memory state of the current LSTM unit is Where: tanh(x)=(e x -e -x ) / (e x +e -x ) is the hyperbolic tangent function, which limits the output to between -1 and 1, W c is the weight matrix of the temporary memory state, b c is the threshold matrix. The current LSTM unit hidden state is updated to Finally, the LSTM unit output is h k =o k ⊙tanh(c k ).

[0079] In the classification task, an output layer is required after the LSTM hidden layer. Suppose the output of the current sample at the tth iteration is Where k∈{1,2,...,K}, K is the total number of categories that the network output needs to be classified. The activation function is the softmax function, so the probability of the kth category of the actual output is According to the label y corresponding to the sample, y = [y1, y2, ..., y K ], the loss function used to train the LSTM network is cross entropy: Where: represents the actual output of the current sample at the tth iteration, W t-1 is the network weight and threshold parameter matrix for the current iteration, indicating that the loss function is related to the weight matrix and the expected output.

[0080] The Adam optimizer is used to update the weights of the LSTM neural network. Assume that the current gradient is Where: represents the partial derivative of the loss function with respect to the weight parameter W. The biased first moment estimate can be updated as u t =β1·u t-1 +(1-β1)·g t , where the exponential decay rate is defined as β1, which mainly controls the update weight of the first moment estimate and the gradient, and is generally taken as 0.9. The biased second moment estimate can be updated as v t =β2·v t-1 +(1-β2)·(g t ⊙g t ), where: β2 is the exponential decay rate, which is mainly used to control the update weight of the second moment estimate and the gradient square, and is generally set to 0.999 by default. The first moment estimate is corrected for deviations: Where: is the t-th power of β1. The second moment estimate is bias-corrected as Where: is the t-th power of β2. Therefore, the weight parameter of the t-th iteration is updated as Where: η is the learning rate, μ is a very small constant to ensure that the denominator is not zero, generally defaulted to 10 -8 .

[0081] Step 5: Initialize the parameters of the quantum Arctic wolf mechanism.

[0082] Including the number of individuals in the population N q , maximum number of iterations G max , the search space dimension M and the upper bound U = [u1,u2,...,u M ] and the lower bound L = [l1,l2,...,l M], group behavior probability P, hunting attack probability Q in group behavior, persecution attack probability 1-Q in group behavior, update coefficient w1, mutation probability Selection probability Randomly initialize the quantum position of the Arctic wolf between [0,1] And mapped to the corresponding position where i∈{1,2,...,N q}, the symbol “⊙” represents the multiplication of elements at corresponding positions in the matrix or vector, and the superscript 0 represents the number of iterations is 0.

[0083] Step 6: Map the elements of each dimension of the position of the i-th Arctic wolf in the population to the hyperparameters of the LSTM network, i∈{1,2,...,N q}, use the training set and corresponding hyperparameters to train the LSTM network, and use the misjudgment rate of the LSTM network on the validation set as the fitness function of the i-th Arctic wolf.

[0084] Assume that the LSTM network hyperparameters to be optimized include the number of hidden layer neurons N h , small batch size S b , number of training rounds I m , learning rate l r and the drop probability p d , the learning rate is mapped to the logarithm of the variable ω4 = lg(l r ). Let the row vector of all hyperparameters be h = [N h ,S b ,I m ,ω4,p d ]=[h1,h2,h3,h4,h5], let its upper bound be and the lower bound is Through the upper and lower bounds, a linear mapping is performed to the upper bound U and lower bound L of the search space of the quantum Arctic wolf mechanism, that is, in After mapping, the vector As the position of the ith Arctic wolf in the tth iteration of the quantum Arctic wolf mechanism In the fitness function, the inverse transformation is used to obtain the corresponding LSTM network hyperparameters, that is, Where: i∈{1,2,...,N q},N q is the number of individuals in the population, d∈{1,2,...,M} is the dth hyperparameter, and the independent variable dimension is M=5. The actual hyperparameters used for training are: the number of neurons in the hidden layer Small batch size Number of training rounds In the formula: round represents rounding off decimals. Learning rate Drop probability

[0085] The LSTM network is trained using the hyperparameters corresponding to the i-th Arctic wolf in the t-th iteration, and the accuracy on the validation set is Then its fitness function is Then calculate the corresponding survival probability Right now Assume the number of iterations t=0.

[0086] Step 7, enter the iteration loop, let t = t + 1, for the i-th Arctic wolf in the population, calculate the value of the variable m1(t) related to the number of iterations m1(t) = exp(t / G max -1), and compared with the value of P, if it is less than P, it will enter the group behavior, and the hunting attack strategy will be updated with probability Q to the quantum rotation angle Where: r i t is a normally distributed random number with mean 0 and variance 1 for the tth iteration of the i-th Arctic wolf, n a is in [2,N q / 2] represents the number of Arctic wolves to be updated. The quantum position of the kth Arctic wolf that attacks the group in the tth iteration is n randomly selected from the population in each iteration. a A group of different Arctic wolves, is the quantum position of the optimal Arctic wolf, c(t) = cos(πt / 2G max ) is the step coefficient that gradually changes from 1 to 0 as the iteration proceeds; the quantum rotation angle is updated with the probability 1-Q for the persecution attack strategy. And define the update step size Where: is the uniformly distributed random number between (0,2) for the ith Arctic wolf in the tth iteration. are the uniformly distributed random numbers between (0,2π) and (0,π) in the tth iteration, w1 is the update coefficient, v t (m) is the average value of the corresponding dimension of all the Arctic wolf quantum positions in the population at the tth iteration Otherwise, m1(t) is greater than or equal to P and the scavenging behavior is performed to update the quantum rotation angle to Where: is the quantum position of a randomly selected Arctic wolf in the t-th iteration of scavenging behavior, and They represent the random numbers uniformly distributed between [-1,1] for the i-th Arctic wolf in the t-th iteration, p r is a random number that satisfies a uniform distribution between [0,1]. is the set selection probability, is the quantum position of the best Arctic wolf in the t-th iteration population, Represents the quantum position of a randomly selected Arctic wolf from the population in the tth iteration.

[0087] Step 8: For the i-th Arctic wolf in the t-th iteration, if the survival probability Then perform survival behavior and update the quantum rotation angle to Where m2(t) = exp(1-t / G max )-t / G max , where: Indicates that the random number of the i-th Arctic wolf in the t-th iteration satisfies the uniform distribution between (0,1); otherwise, go to step nine.

[0088] Step 9: Update the quantum position of each Arctic wolf according to the quantum rotation angle and quantum gate strategy: Where: represents the dth dimension of the quantum position of the tth iteration of the i-th Arctic wolf, is the corresponding updated quantum rotation angle, is a random number that satisfies a uniform distribution between [0,1]. is the probability of mutation of the quantum position when the quantum rotation angle is 0.

[0089] Step 10: Calculate the position of the i-th Arctic wolf in the new generation as Convert the elements of each dimension into the corresponding LSTM network hyperparameters. After training the LSTM network, use the misjudgment rate on the validation set as the corresponding fitness function.

[0090] Step 11: Use the retention strategy to retain the updated Arctic wolf: if the updated Arctic wolf has better fitness, retain the corresponding position, quantum position and fitness after the update; otherwise, do not retain them. Update the quantum position corresponding to the optimal individual and location Update the survival probability of each Arctic wolf for

[0091] Step 12: Determine whether the number of iterations t reaches the maximum number of iterations G max , if t=G max , then output the optimal Arctic wolf position As the optimization result, otherwise return to step 7 to continue iterating.

[0092] Step 13: The elements of each dimension are converted into the corresponding optimal LSTM network hyperparameters, and the LSTM neural network with the optimal hyperparameters is trained. After the training is completed, the modulated signal recognition performance of the LSTM network is tested using the test set, and the results are analyzed.

[0093] exist Figure 3 In the present invention, the hybrid feature sample expansion method proposed in the present invention and the LSTM neural network evolution method through the designed quantum Arctic wolf mechanism are recorded as MF-QAW-LSTM, the three instantaneous features described in the "Modulation recognition method based on instantaneous feature parameters" published by Liu Sheng et al. in "Communication Technology" (2022, 55(05): 559-564) and the modulation signal recognition method combined with a three-layer fully connected neural network are recorded as IF-TFCNN, and the five high-order cumulant features described in the "Digital modulation recognition based on high-order cumulants" published by Zhang Wenna et al. in "Foreign Electronic Measurement Technology" (2022, 41(07): 126-132) and the modulation signal recognition method combined with a decision tree are recorded as CF-DT.

[0094] exist Figure 4 In the present invention, the hybrid feature sample expansion method proposed in the present invention and the LSTM neural network evolution method through the designed quantum Arctic wolf mechanism are denoted as MF-QAW-LSTM, and the LSTM neural network modulation signal recognition method with artificially selected hyperparameters is denoted as MS-LSTM.

[0095] The simulation experiment parameters of the modulation signal recognition method of hybrid feature sample expansion combined with quantum Arctic wolf mechanism evolution LSTM neural network are set as follows:

[0096] The modulation signals to be identified are 9 modulation signals, including 2ASK, 4ASK, 2PSK, 4PSK, 2FSK, 4FSK, MSK, 8QAM and 16QAM, and are not limited to these 9 modulation modes. d =38400baud, sampling frequency f s =3.072MHz, the number of sampling points per symbol is 80. Carrier frequency f c =408kHz, the carrier frequencies of 2FSK are 306kHz and 510kHz, with a frequency interval of 204kHz; the carrier frequencies of 4FSK are 255kHz, 357kHz, 459kHz and 561kHz, with a frequency interval of 102kHz. The modulated signal is subjected to root raised cosine shaping filtering with a roll-off factor of 0.4 and a duration of 4 symbol periods. The signal-to-noise ratio ranges from -10dB to 10dB with an interval of 2dB. Each modulated signal is divided into multiple data segments of 40 symbols in length and a total of 3200 points.

[0097] Six instantaneous features and five cumulative features are extracted, and each signal contains 502 feature samples at each signal-to-noise ratio to form the original feature set. It is normalized to zero mean and unit variance, and 500 feature sequence samples are obtained for each signal at each signal-to-noise ratio through a sliding rectangular window of length 3. The training set, validation set, and test set are randomly divided according to the ratio of 3:1:1. The training set has a total of 29,700 samples, and the validation set and test set have 9,900 samples each.

[0098] The hyperparameter range of the LSTM network optimized by the quantum Arctic Wolf mechanism is set to: the number of neurons in the hidden layer, the size of the mini-batch, the number of training rounds, the logarithm of the learning rate, and the upper bound of the drop probability are set to The corresponding lower bound is

[0099] The parameters of the quantum Arctic wolf mechanism are set as follows: group behavior probability P = 0.5, hunting attack probability Q = 0.7 in group behavior, persecution attack probability 1-Q = 0.3 in group behavior, dimension M = 5, search space upper bound U = [1,1,1,1,1], lower bound L = [-1,-1,-1,-1,-1], update coefficient u1 = 0.74 in persecution attack, and mutation probability when the quantum rotation angle is 0 Selection probability The number of Arctic wolves in the population N q =20, maximum number of iterations G max =20.

[0100] from Figure 3 It can be seen that the MF-QAW-LSTM method proposed in the present invention extracts richer signal features to form 11 mixed feature libraries compared with the comparative methods IF-TFCNN and CF-DT, and expands samples through a sliding rectangular window. The expanded samples containing time series information are used to train the LSTM neural network, and a quantum Arctic wolf mechanism is proposed to optimize the five key hyperparameters of the LSTM neural network. Therefore, the recognition accuracy is greatly improved, which expands the application limitations of the traditional modulation signal recognition method based on the LSTM neural network under low signal-to-noise ratio.

[0101] from Figure 4 It can be seen that the MF-QAW-LSTM method proposed in the present invention is able to automatically optimize the LSTM neural network hyperparameters compared with the LSTM neural network method MS-LSTM in which hyperparameters are manually selected, thus avoiding the need for expert experience and the difficulty in obtaining the optimal hyperparameters caused by manual determination of hyperparameters. The recognition accuracy is greatly improved under low signal-to-noise ratio, which expands the application limitations of traditional modulation signal recognition methods based on LSTM neural networks under low signal-to-noise ratios.

[0102] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A signal recognition method combining hybrid feature sample expansion and evolutionary LSTM network, characterized by: Here are the steps: Step 1: Establish a modulation signal sample library to obtain a modulation signal data set; Step 2: Extract instantaneous features and cumulative features from the modulated signal data set obtained in step 1 to construct a mixed feature library; Step 3: Preprocess the mixed features extracted in step 2; Step 4: Determine the objective function; The training set and the set hyperparameters are used to train the LSTM neural network, and the misjudgment rate of the LSTM network on the validation set is used as the objective function, that is, the fitness function F = 1-A c , where A c is the recognition accuracy on the validation set. After the training is completed, the validation set is used to verify the training effect. Step 5: Initialize the parameters of the quantum Arctic wolf mechanism; Including the number of individuals in the population N q , maximum number of iterations G max , the search space dimension M and the upper bound U = [u1,u2,...,u M ] and the lower bound L = [l1,l2,…,l M ], group behavior probability P, hunting attack probability Q in group behavior, persecution attack probability 1-Q in group behavior, update coefficient w1, mutation probability Selection probability Randomly initialize the quantum position of the Arctic wolf between [0,1] And mapped to the corresponding position where i∈{1,2,…,N q }, the symbol "⊙" represents the multiplication of elements at corresponding positions in the matrix or vector, and the superscript 0 represents the number of iterations is 0; Step 6: Map the elements of each dimension of the position of the i-th Arctic wolf in the population to the hyperparameters of the LSTM network, i∈{1,2,…,N q }, use the training set and corresponding hyperparameters to train the LSTM network, and use the misjudgment rate of the LSTM network on the validation set as the fitness function of the i-th Arctic wolf; Step 7: Enter the iteration loop, set t = t + 1, for the i-th Arctic wolf in the population, calculate the value of the variable m1(t) related to the number of iterations m1(t) = exp(t / G max -1), compared with the value of P, if it is less than P, it will enter group behavior; if m1(t) is greater than or equal to P, it will perform scavenging behavior; Step 8: For the i-th Arctic wolf in the t-th iteration, if the survival probability Then perform survival behavior and update the quantum rotation angle to Where m2(t) = exp(1-t / G max )-t / G max , where: Indicates the random number of the ith Arctic wolf in the tth iteration that satisfies the uniform distribution between (0,1); otherwise, proceed to step nine; Step 9: Update the quantum position of each Arctic wolf according to the quantum rotation angle and quantum gate strategy: Where: represents the dth dimension of the quantum position of the tth iteration of the i-th Arctic wolf, is the corresponding updated quantum rotation angle, is a random number that satisfies a uniform distribution between [0,1]. is the probability of quantum position mutation when the quantum rotation angle is 0; Step 10: Calculate the position of the i-th Arctic wolf in the new generation as Convert the elements of each dimension into the corresponding LSTM network hyperparameters. After training the LSTM network, use the misjudgment rate on the validation set as the corresponding fitness function. Step 11: Use the retention strategy to retain the updated Arctic wolf: if the updated Arctic wolf has better fitness, retain the corresponding position, quantum position and fitness after the update, otherwise do not retain; update the quantum position corresponding to the optimal individual and location Update the survival probability of each Arctic wolf for Step 12: Determine whether the number of iterations t reaches the maximum number of iterations G max , if t=G max , then output the optimal Arctic wolf position As the optimization result, otherwise return to step 7 to continue iteration; Step Thirteen: The elements of each dimension are converted into the corresponding optimal LSTM network hyperparameters, and the LSTM neural network with the optimal hyperparameters is trained. After the training is completed, the modulated signal recognition performance of the LSTM network is tested using the test set, and the results are analyzed.

2. The signal recognition method of a hybrid feature sample expansion combined with an evolved LSTM network according to claim 1, characterized in that: The digital modulation signals in step 1 include MASK, MPSK, MFSK, MSK and MQAM. In order to reduce the influence of inter-symbol interference, a shaping filter is added to the MASK, MPSK, MFSK and MQAM signals at the transmitting end. The shaping filter uses a square root raised cosine roll-off function to shape the digital baseband signal. The signal is received at the receiving end through a Gaussian white noise channel to obtain a modulated signal data set.

3. The signal recognition method of a hybrid feature sample expansion combined with an evolved LSTM network according to claim 1, characterized in that: The step 2 is specifically as follows: After the receiver receives the signal, it performs Hilbert transform on it to obtain its analytical form, that is, Where s(t) is the analytical signal of the original signal y(t), j is a complex unit, and is the Hilbert transform of y(t), we have In the formula, represents the convolution operation; Its frequency response is With sampling frequency f s Sampling the original signal y(t), the total number of points is The discrete sequence y(n) of The instantaneous amplitude is A(n), then The instantaneous phase is θ(n), then Since the main value interval of the inverse tangent function is (-π / 2,π / 2), θ(n) may produce a sudden change of ±π at this time, and it is adjusted to obtain a phase value in [0,2π) have The actual instantaneous phase ε(n) is related to The relationship is In the formula, mod represents the remainder operation; therefore, There is phase wrapping; since the dewrapped instantaneous phase φ(n) satisfies φ(n)=2πf c T s n+ε(n)+θ, where f c is the carrier frequency, T s Sampling is the period, θ is the initial phase; From the above formula, we know that the de-wrapping instantaneous phase is the linear phase component caused by the carrier frequency and the nonlinear component caused by ε(n) and θ. Add a correction sequence {c(n)} to achieve deconvolution, defined as At this time, the deconvolution instantaneous phase estimate is When the carrier and code element are completely synchronized, the estimated value of the de-warping instantaneous phase nonlinear component is The instantaneous frequency sequence is obtained by differentiating the deconvolved instantaneous phase sequence, that is, In the formula, f s is the sampling frequency; On the basis of the instantaneous amplitude, frequency and phase of the signal, multiple characteristic quantities of the instantaneous information of the digital modulation signal are further extracted to obtain six characteristic parameters, including the maximum value of the spectral density of the zero-centered normalized instantaneous amplitude. The standard deviation of the absolute value of the zero-centered normalized instantaneous amplitude σ aa , the standard deviation σ of the instantaneous phase nonlinear component of the zero-centered non-weak signal segment dp , the standard deviation σ of the absolute value of the instantaneous phase nonlinear component of the zero-centered non-weak signal segment ap , the standard deviation of the absolute value of the instantaneous frequency of the zero-centered normalized non-weak signal segment σ af , the variance of the normalized instantaneous frequency After the receiver obtains the received baseband signal through down-conversion and matched filtering, five effective cumulant features Z1, Z2, Z3, Z4 and Z5 of the signal are extracted according to the high-order cumulant theory; suppose a random process x(t), whose k-order cumulant is defined as C kx (f1,f2,…,f k )=Cum[x(t),x(t+f1),…,x(t+f k-1 )], where: Cum(·) is the cumulative operation, which is expressed as the calculation of high-order mixing moments; the p-order mixing moment of the random process is M pq =E{x(t) p-q x * (t) q }, where: x*(t) represents the complex conjugate of the random process, q is the number of conjugate multiplications; the cumulants of each order of the random process are as follows:

1. The second-order cumulant is expressed as: C 20 =M 20 , C 21 =M 21 ; 2. The fourth-order cumulant is expressed as: C 41 =M 41 -3M 20 M 21 ; 3. The sixth-order cumulant is expressed as: In order to reduce the influence of signal energy on characteristic parameters, five characteristic parameters with better performance are extracted in the form of ratios as follows: The signal is processed in segments, and characteristic parameters are extracted for each segment, eventually obtaining a mixed feature library containing 11 characteristic parameters.

4. The signal recognition method of a hybrid feature sample expansion combined with an evolved LSTM network according to claim 1, characterized in that: The step 3 preprocesses the mixed features, including normalization, sample expansion, and division into training set, validation set, and test set. The sample expansion uses a sliding rectangular window to take three consecutive feature samples as input variables of an LSTM neural network. Assume that the feature set composed of the extracted mixed features is M = [m1, m2, ..., m J ], where J is the total number of samples, and the i-th feature sample is Where: superscript T is the transposition symbol, superscript (i) represents the i-th sample; The feature set is normalized using the normalization method of zero mean and unit variance: the mean μ of each feature in the feature set is calculated as [μ1, μ2, …, μ 11 ] T and standard deviation σ=[σ1,σ2,…,σ 11 ] T , then the normalized feature set is In the formula: ". / " means dividing all elements in each row of the matrix by the elements in the corresponding row of the vector; Use a rectangular window with a length of 3 and a moving step of 1 to transform the feature set Truncate to data segments, i.e. The kth sequence sample is Where: Since the sliding window length is 3, the last two samples are incomplete, so the last two incomplete samples are discarded. That is, a feature sequence set N for LSTM network recognition is obtained, where the size of each sequence sample is 11×3; then, the feature sequence set N and the corresponding label T are randomly divided into a training set, a validation set, and a test set according to a certain ratio.

5. The signal recognition method of a hybrid feature sample expansion combined with an evolved LSTM network according to claim 1, characterized in that: In step 6, the LSTM network hyperparameters to be optimized include the number of neurons in the hidden layer N h , small batch size S b , number of training rounds I m , learning rate l r and the drop probability p d , the learning rate is mapped to the logarithm of the variable ω4 = lg(l r ); Let the row vector composed of all hyperparameters be h = [N h ,S b ,I m ,ω4,p d ]=[h1,h2,h3,h4,h5], let its upper bound be and the lower bound is Through the upper and lower bounds, a linear mapping is performed to the upper bound U and lower bound L of the search space of the quantum Arctic wolf mechanism, that is, in After mapping, the vector As the position of the ith Arctic wolf in the tth iteration of the quantum Arctic wolf mechanism In the fitness function, the inverse transformation is used to obtain the corresponding LSTM network hyperparameters, that is, Where: i∈{1,2,…,N q },N q is the number of individuals in the population, d∈{1,2,…,M} is the dth hyperparameter, and the independent variable dimension is M=5; then the actual hyperparameters used for training are: the number of neurons in the hidden layer Small batch size Number of training rounds In the formula: round represents rounding off decimals; learning rate Drop probability The LSTM network is trained using the hyperparameters corresponding to the i-th Arctic wolf in the t-th iteration, and the accuracy on the validation set is Then its fitness function is Then calculate the corresponding survival probability Right now Assume the number of iterations t=0.

6. The signal recognition method of a hybrid feature sample expansion combined with an evolved LSTM network according to claim 1, characterized in that: In step 7, for the i-th Arctic wolf in the population, if m1(t)=exp(t / G max -1)<P, then group behavior is initiated, which includes hunting attack and persecution attack strategies. The hunting attack strategy is performed with probability Q and the quantum rotation angle is updated as Where: r i t is a normally distributed random number with mean 0 and variance 1 for the tth iteration of the i-th Arctic wolf, n a is in [2,N q / 2] represents the number of Arctic wolves to be updated. The quantum position of the kth Arctic wolf that attacks the group in the tth iteration is n randomly selected from the population in each iteration. a A group of different Arctic wolves, is the quantum position of the optimal Arctic wolf, c(t) = cos(πt / 2G max ) is the step coefficient that gradually changes from 1 to 0 as the iteration proceeds; the quantum rotation angle is updated with the probability 1-Q for the persecution attack strategy. And define the update step size Where: is a random number uniformly distributed between (0,2) for the ith Arctic wolf in the tth iteration. are the uniformly distributed random numbers between (0,2π) and (0,π) in the tth iteration, w1 is the update coefficient, and v t (m) is the average value of the corresponding dimension of all the Arctic wolf quantum positions in the population at the tth iteration Otherwise m1(t)≥P performs scavenging behavior and updates the quantum rotation angle to Where: is the quantum position of a randomly selected Arctic wolf in the t-th iteration of scavenging behavior, and They represent the random numbers uniformly distributed between [-1,1] for the i-th Arctic wolf in the t-th iteration, p r is a random number that satisfies a uniform distribution between [0,1]. is the set selection probability, is the quantum position of the best Arctic wolf in the t-th iteration population, Represents the quantum position of a randomly selected Arctic wolf from the population in the tth iteration.

Images