Wireless time-frequency domain signal identification method based on wavelet time scattering network

Abstract

The invention provides a wireless time-frequency domain signal identification method based on a wavelet time scattering network, and the method comprises the steps: obtaining a wireless signal, carrying out the preprocessing of the wireless signal, obtaining a scattering coefficient based on the preprocessed wireless signal through wavelet scattering transformation, constructing a scattering coefficient graph, constructing a feature vector based on the scattering coefficient graph, and carrying out the recognition of a wireless time-frequency domain signal. And taking the wavelet scattering coefficient spectrogram as a training data set, constructing a wireless time-frequency domain signal recognition model, training the wireless time-frequency domain signal recognition model based on the training data set, and inputting the image texture features of the wavelet scattering coefficient spectrogram of the wireless signal to be recognized into the trained wireless time-frequency domain signal recognition model for recognition to obtain a recognition result. According to the invention, the method can achieve the efficient recognition of wireless signals, and is good in robustness and accuracy.

Description

Technical Field

[0001] The invention relates to the technical field of wireless time-frequency domain signal recognition, and in particular to a wireless time-frequency domain signal recognition method based on a wavelet time scattering network. Background Art

[0002] With the rapid development of wireless communication technology, the concurrent use of various wireless signals has greatly increased the complexity of signal recognition. Traditional signal recognition methods such as time-frequency analysis based on Fourier transform have certain limitations when dealing with complex and diverse wireless signals, especially when facing non-stationary signals and noisy environments. Therefore, it is necessary to design a wireless time-frequency domain signal recognition method based on wavelet time scattering network. Summary of the invention

[0003] In order to overcome the deficiencies of the prior art, the object of the present invention is to provide a wireless time-frequency domain signal recognition method based on a wavelet time scattering network.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A wireless time-frequency domain signal recognition method based on a wavelet time scattering network, comprising:

[0006] Acquire wireless signals and pre-process them;

[0007] Based on the preprocessed wireless signal, a scattering coefficient is obtained by wavelet scattering transform, and a scattering coefficient map is constructed;

[0008] Construct feature vectors based on the scattering coefficient map and use it as a training data set;

[0009] Build a wireless time-frequency domain signal recognition model and train it based on the training data set;

[0010] The image texture features of the wavelet scattering coefficient spectrum of the wireless signal to be identified are input into the trained wireless time-frequency domain signal recognition model for identification to obtain the identification result.

[0011] Preferably, the wireless signal is obtained and preprocessed, specifically:

[0012] Acquire wireless signals and perform denoising based on the IWOA-VMD algorithm;

[0013] Perform irrelevant frequency band removal processing on the wireless signal after denoising;

[0014] The wireless signal after the irrelevant frequency band is removed is normalized.

[0015] Preferably, the wireless signal is acquired and denoised based on the IWOA-VMD algorithm, specifically:

[0016] Initialize the VMD algorithm parameter optimization range, and find the parameter combination corresponding to the minimum value of the composite fitness function according to the IWOA algorithm;

[0017] Substitute the optimal parameter combination into the VMD decomposition model to decompose the wireless signal into a series of IMFs;

[0018] The multi-scale permutation entropy-variance contribution rate selection principle is used to calculate and identify different IMFs, classify them into noise-dominated components and effective signal components, and perform NLM denoising on the noise-dominated components to obtain the denoised components.

[0019] The denoised component and the effective signal component are reconstructed to obtain the denoised wireless signal.

[0020] Preferably, a scattering coefficient is obtained by wavelet scattering transform based on the preprocessed wireless signal, and a scattering coefficient map is constructed, specifically:

[0021] The pre-processed wireless signal is scattered through the first layer, that is, the signal is averaged through a wavelet low-pass filter to obtain the scattering scale coefficient of the first layer;

[0022] The scattering scale coefficient of the first layer is used as the scattering coefficient of the second layer, and the scattering scale coefficient of the second layer is obtained through the wavelet scattering of the second layer. The scattering scale coefficient of the third layer is obtained through the same modular operation and wavelet low-pass filtering function. The scattering scale coefficient of the third layer is repeated to obtain the scale coefficient of the next layer, and the scattering coefficient map is constructed according to the obtained scattering scale coefficient.

[0023] Preferably, a feature vector is constructed based on the scattering coefficient map and used as a training data set, specifically:

[0024] Extract the grayscale gradient co-occurrence matrix of the scattering coefficient image and obtain its image eigenvalues;

[0025] A feature vector representing the wireless signal is constructed based on the image feature value and used as a training data set.

[0026] Preferably, a wireless time-frequency domain signal recognition model is constructed and trained based on a training data set, specifically:

[0027] A wireless time-frequency domain signal recognition model is constructed based on the SVM classifier, and the wireless time-frequency domain signal recognition model is trained through a training data set based on the Bayesian SVM hyperparameter optimization algorithm.

[0028] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0029] The present invention provides a wireless time-frequency domain signal recognition method based on a wavelet time scattering network, the method comprising: obtaining a wireless signal, and preprocessing it, obtaining a scattering coefficient based on the preprocessed wireless signal through a wavelet scattering transform, constructing a scattering coefficient map, constructing a feature vector based on the scattering coefficient map, using it as a training data set, constructing a wireless time-frequency domain signal recognition model, and training it based on the training data set, inputting the image texture features of the wavelet scattering coefficient spectrum of the wireless signal to be recognized into the trained wireless time-frequency domain signal recognition model for recognition, and obtaining a recognition result. The present invention can realize efficient recognition of wireless signals, and has good robustness and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0031] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0032] Figure 2 Schematic diagram of denoising process;

[0033] Figure 3 Schematic diagram of the wavelet scattering network structure. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] The purpose of the present invention is to provide a wireless time-frequency domain signal recognition method based on a wavelet time scattering network, which can realize efficient recognition of wireless signals and has good robustness and accuracy.

[0036] Figure 1 A flow chart of a method provided by an embodiment of the present invention, such as Figure 1 As shown, the present invention provides a wireless time-frequency domain signal recognition method based on a wavelet time scattering network, comprising:

[0037] Step 100: Acquire wireless signals and pre-process them;

[0038] Step 200: Obtain a scattering coefficient through wavelet scattering transform based on the preprocessed wireless signal, and construct a scattering coefficient map;

[0039] Step 300: construct a feature vector based on the scattering coefficient map and use it as a training data set;

[0040] Step 400: construct a wireless time-frequency domain signal recognition model, and train it based on a training data set;

[0041] Step 500: inputting the image texture features of the wavelet scattering coefficient spectrum of the wireless signal to be identified into the trained wireless time-frequency domain signal recognition model for identification to obtain the identification result.

[0042] Step 100, obtaining a wireless signal and preprocessing it, specifically:

[0043] Step 101: Acquire a wireless signal and perform denoising on it based on an IWOA-VMD algorithm;

[0044] Step 102: performing irrelevant frequency band removal processing on the wireless signal after the denoising process;

[0045] Step 103: normalize the wireless signal after the irrelevant frequency band is removed.

[0046] like Figure 2 As shown, in step 101, a wireless signal is acquired and denoised based on the IWOA-VMD algorithm, specifically:

[0047] Initialize the VMD algorithm parameter optimization range, and find the parameter combination corresponding to the minimum value of the composite fitness function according to the IWOA algorithm;

[0048] Substitute the optimal parameter combination into the VMD decomposition model to decompose the wireless signal into a series of IMFs;

[0049] The multi-scale permutation entropy-variance contribution rate selection principle is used to calculate and identify different IMFs, classify them into noise-dominated components and effective signal components, and perform NLM denoising on the noise-dominated components to obtain the denoised components.

[0050] The denoised component and the effective signal component are reconstructed to obtain the denoised wireless signal.

[0051] A detailed introduction to the variational mode decomposition (VMD) algorithm;

[0052] As a very active branch in the field of adaptive signal decomposition, VMD algorithm has played a pivotal role in the field of signal processing in recent years. The core idea of ​​VMD algorithm is to adaptively decompose the input signal f(t) into k k nearby modal functions u k (t), while minimizing u k The bandwidth and of (t). The VMD algorithm decomposition process is described as follows:

[0053] Let the kth modal function u k The expression of (t) is:

[0054] u k (t) = A k (t)cos[φ k (t)]

[0055] In the formula, A k (t) is u k (t) and A k (t)≥0,φ k (t) is u k The instantaneous phase of (t) is a non-decreasing function.

[0056] For k modal functions u k (t) Perform Hilbert transform calculation to obtain the corresponding single-sided spectrum. The relevant analytical signal is:

[0057]

[0058] Where δ(t) is the unit pulse function and * represents convolution.

[0059] By multiplying the analytical signal by Get the baseband signal:

[0060]

[0061] In order to obtain the bandwidth of each modal function, signal demodulation is achieved through Gaussian smoothing. The constrained variational model is as follows:

[0062]

[0063] Where K is the preset number of decomposed modes.

[0064] In order to solve the above constrained variational model, the Lagrange multiplier λ and the quadratic penalty factor α are introduced, and the above formula is transformed into an unconstrained variational model as shown below:

[0065]

[0066] The modal function {u k}、Center frequency {ω k} and the Lagrange multiplier λ are continuously updated and iterated, and the formula is as follows:

[0067]

[0068]

[0069] Where τ is the time step.

[0070] When the following conditions are met, the ADMM algorithm converges and the above update iteration process terminates:

[0071]

[0072] Where c is the convergence parameter, usually ε = 1 × 10 -6 .

[0073] A detailed introduction to the whale optimization algorithm (WOA):

[0074] The whale optimization algorithm is a new swarm intelligence optimization algorithm based on iterative search of humpback whale populations. It models its predation behavior to achieve optimization of complex optimization problems. According to its algorithm characteristics, it is divided into three stages: searching for prey, surrounding prey, and spiral bubble net predation.

[0075] (1) Searching for prey

[0076] In this stage, according to the random walk mechanism, each whale participating in the hunt iteratively updates its position based on each other's current position. According to the t-th search position, when the coefficient |A|>1, the whale group randomly searches for high-quality prey in the global range, and its t+1-th position update expression is:

[0077] D=|C·x rand (t)-x(t)|

[0078] x(t+1)=x rand (t)-A·D

[0079] Among them, t is the current iteration number, x rand (t) is the arbitrary position of the whale, x(t) is the current position, and:

[0080] A=2ar1-a

[0081] C=2r2

[0082] Among them, r1,r2∈rand[0,1],a=2-2t / T, and T is the number of selections. According to the linear decreasing property of a, A gradually decreases with a. When |A|≤1, the algorithm then enters the stage of surrounding the prey.

[0083] (2) Encirclement of prey

[0084] After discovering the target prey, the whale group will gradually narrow its search range and update the search position according to the best individual in the current population. The corresponding expression is:

[0085] D1=|C·x best (t)-x(t)|

[0086] x(t+1)=x best (t)-A·D

[0087] In the formula, x best (t) is the best individual position in the current whale group.

[0088] (3) Spiral bubble net predation stage

[0089] Whales have a unique hunting method, using the bubble net they spit out to spiral upward and shrink to surround and approach the target. The mathematical model is as follows:

[0090] D2=|x best (t)-x(t)|

[0091] x(t+1)=D2e bl cos(2πl)+x best (t)

[0092] Where b is a constant coefficient and l is a random number in the interval [-1, 1].

[0093] The hunting phase is divided by setting random probability. When the random probability p<0.5, the whale updates the best individual position; when p≥0.5, the whale gradually approaches the target. During the approach process, |A| decreases continuously, and the optimal solution is obtained when |A|=0. The overall hunting mechanism is described as follows:

[0094]

[0095] Next, we introduce the improvement strategy of the whale optimization algorithm:

[0096] The traditional whale optimization algorithm adjusts global search and local development through coefficient A. The value of A is mainly related to a. However, the linearly changing convergence factor a has disadvantages in early search capability and later iteration speed. In order to improve its impact on algorithm performance, a piecewise nonlinear convergence factor m is proposed without affecting the overall change trend. Its expression is as follows:

[0097]

[0098] During the entire iteration process, m shows a nonlinear decreasing trend. In the early stage, the larger parameter m gradually decreases, which can improve the global exploration capability; in the later stage, the smaller parameter m can increase the convergence speed of the algorithm.

[0099] In addition, considering the impact of inertia weight on the algorithm's optimization ability, an adaptive weight strategy is proposed here. The formula is as follows:

[0100]

[0101] So the position update formula becomes:

[0102]

[0103] The traditional WOA algorithm is prone to fall into local optimality. Here, the Cauchy operator is introduced to give the optimal individual Cauchy perturbation. The Cauchy operator can generate larger perturbations to make the algorithm escape from the local optimality, and can also generate smaller perturbations to increase the convergence speed.

[0104] Therefore, the present invention uses the following Cauchy variation formula to update the best individual position as follows:

[0105] x new (t) = x best (t)·[1+Cauchy(0,1)]

[0106] In the formula, x new (t) is the new value of the current optimal value after disturbance, and Cauchy(0,1) is the Cauchy operator.

[0107] The fitness function of the IWOA-VMD algorithm is:

[0108] Energy entropy can characterize the energy distribution of a signal in the frequency domain. The smaller its value, the higher the energy proportion of the corresponding component. The energy calculation formula for each component is:

[0109]

[0110] Where N is the signal length. The energy probability density formula obtained by normalizing it is:

[0111]

[0112] The energy entropy calculation formula is:

[0113]

[0114] In practical applications, if energy entropy is used as a single fitness function, it is easy to decompose the signal into simple harmonic waves. This is because the signal is over-decomposed in pursuit of energy concentration. In order to improve the over-decomposition of the signal, kurtosis is introduced to describe the waveform characteristics of the signal. Kurtosis is a dimensionless parameter that characterizes the degree of data spikes and is often used for fault diagnosis of nonlinear signals. Its expression is as follows:

[0115]

[0116] In the formula, μ and σ represent the mean and variance of the signal respectively; E(x-μ) 4 represents the fourth-order mathematical expectation.

[0117] According to the characteristics of wireless signals, a composite fitness function is constructed by combining energy entropy and kurtosis. By iteratively searching for the minimum value of the composite fitness function, the (K, α) parameter combination with the best decomposition effect is found. The fitness function calculation formula is as follows:

[0118] F=1 / Ku+0.1×H E

[0119] The principle of selecting the optimal modal component of multi-scale permutation entropy-variance contribution rate is introduced:

[0120] Multi-scale permutation entropy (MPE) describes the randomness and complexity of signals by introducing permutation entropies of different scales and comprehensively considering the signal characteristics at multiple scales in a weighted average manner. The basic principle of MPE is as follows:

[0121] 1. For the time series x={x1,x2,…,x L Multi-scale coarse-graining

[0122]

[0123] Where s is the scale factor, is a multi-scale time series. When s is set to 1, For the original time series, the permutation entropy can be calculated.

[0124] 2. Time series Reconstructing the phase space

[0125]

[0126] Where τ is the time delay factor, m is the embedding dimension, To reconstruct the sequence.

[0127] 3. Reconstruction Arrange in ascending order, and you can get m! permutations and combinations, and the number of times each permutation occurs is N t , the corresponding probability is Right now:

[0128]

[0129] 4. Scale factor s Permutation entropy calculation and normalization

[0130]

[0131] As a data statistics theory based on the Gaussian hypothesis, principal component analysis can reduce the redundancy and invalid features of data components, eliminate data correlation to highlight the advantages and useful information of the original sample data. Variance contribution rate is a commonly used method for selecting principal components in principal component analysis. It mainly represents the proportion of IMF energy in the signal energy before decomposition. The larger the variance contribution rate, the stronger the IMF's ability to explain the original signal. Therefore, the IMF can be ranked according to the size of the variance contribution rate to evaluate the importance of the IMF. The calculation formula of variance contribution rate is as follows:

[0132]

[0133] In the formula, D(IMF k ) represents the variance of the kth IMF, V k represents the variance contribution rate of the kth IMF.

[0134] A detailed introduction to NLM algorithm denoising:

[0135] Based on the weighted average operation of similar structures in the image to improve the signal-to-noise ratio, the non-local mean filtering algorithm achieves the purpose of image denoising. The NLM algorithm screens all the information in the image, retains the obvious main features, and removes obviously irrelevant redundant components. By analyzing the similarity characteristics, it is found that the idea of ​​the NLM algorithm can also be applied to the field of signal filtering. Using the NLM algorithm for filtering, find as many similar values ​​as possible within the search range, and use the weighted coefficient to obtain the estimated value for weighted averaging to make it close to the true value in the signal, thereby achieving the purpose of noise reduction. The principle of the NLM algorithm is as follows:

[0136] Assume that the noisy signal y(t) is the superposition of the original signal u(t) and the external interference noise n(t), as shown in the following formula:

[0137] y(t)=u(t)+n(t)

[0138] The NLM algorithm estimates the original signal u(t) by calculating the weighted average K(t) of all similar blocks in y(t). The calculation process is as follows:

[0139]

[0140] In the formula, ω(t, s) refers to the similarity between two search blocks centered on t and s respectively. N(t) represents the set of all points in the target search area centered on the target sample point t. Z(t) is the normalization constant, representing the sum of the similarities of all search blocks in the search domain. ω(t, s) also needs to satisfy and These two basic conditions, and then calculate ω(t, s) as:

[0141]

[0142] In the formula, L Δ is the neighborhood block centered on s, L Δ = 2P + 1. The parameter P affects the number of similar structural blocks found. △ is the search block centered on t. Take K as half of the length of the △ region. The parameter K affects the amount of calculation and calculation time. λ is the filter bandwidth parameter, which affects the smoothness of the filtered signal.

[0143] The selection of these three key parameters P, K, and λ greatly affects the denoising effect of the algorithm on one-dimensional signals. If the similarity block parameter P is too small, it is easy to find more similar blocks within the search range. If P is too large, no similar blocks can be found. In the present invention, P is set to 10 < P < 20. Theoretically, the search block parameter K can search the entire domain, but this will increase the calculation difficulty of the algorithm and the calculation efficiency is low. The literature points out that when the signal length N > 4000, take 0.25N < K < 0.35N. The search range contains more signal information and reaches a balance in terms of the amount of calculation and calculation time. When N < 4000, take K = 0.5N. The search range can cover the entire domain and most of the signal information can be obtained. If the bandwidth parameter A is too small, it will cause noise fluctuations, which will interfere with the weighted average of similar blocks and the filtering effect is poor; if λ is too large, it will misjudge the similarity of search blocks and the signal is too smooth, and the filtering effect is too strong. In the present invention, the filtering effect is better when 0.3σ < λ < 0.6σ (σ is the standard deviation of the noisy signal).

[0144] In step 102, the processed wireless signal after denoising is subjected to irrelevant frequency band removal processing, specifically:

[0145] The present invention provides an embodiment. For different signal types (such as Wi-Fi, LTE, Bluetooth, etc.), the signal within a specific frequency range is retained through a band-pass filter, and the irrelevant frequency bands are filtered out to highlight the characteristics of the signal.

[0146] In step 103, the wireless signal after the irrelevant frequency band is removed is normalized, specifically:

[0147] The signal amplitude is normalized to ensure that the feature extraction process is not affected by signal strength fluctuations and to improve the stability of the system.

[0148] In step 200, a scattering coefficient is obtained by wavelet scattering transform based on the preprocessed wireless signal, and a scattering coefficient map is constructed, specifically:

[0149] First, the wavelet scattering network is introduced:

[0150] The wavelet scattering transform has a computationally translation-invariant image representation that is stable to deformation and can retain high-frequency information for classification. It cascades wavelet inverse convolutions with nonlinear modulus and average operators. The first network layer plays the role of the scale-invariant feature transform matching algorithm (SIFT), while the next layer provides complementary invariant information to improve classification. Therefore, wavelet scattering is a deep convolutional network cascaded by wavelet transform and modulus operator. The main advantages of wavelet scattering feature extraction are reduced requirements for data volume and model complexity; automatic extraction of relevant compact frameworks.

[0151] The wavelet scattering network is mainly composed of three parts: signal input, deep feature extraction, and feature output. The deep feature extraction part is mainly composed of multiple submodules, each of which is composed of three parts: wavelet convolution, nonlinearity, and averaging. The network structure is as follows: Figure 3 shown.

[0152] The pre-processed wireless signal is scattered through the first layer, that is, the signal is averaged through a wavelet low-pass filter to obtain the scattering scale coefficient of the first layer;

[0153] The scattering scale coefficient of the first layer is used as the scattering coefficient of the second layer, and the scattering scale coefficient of the second layer is obtained through the wavelet scattering of the second layer. The scattering scale coefficient of the third layer is obtained through the same modular operation and wavelet low-pass filter function. The scale coefficient of the next layer of the scattering scale coefficient of the third layer is obtained repeatedly, and a scattering coefficient map is constructed according to the obtained scattering scale coefficient.

[0154] Considering practical applications, as the number of layers increases, each iterative operation consumes the energy of the wireless signal, so the number of scattering layers in the present invention is selected as 3 layers. Among them, the first layer performs an averaging operation and loses the high-frequency details of the wireless signal. The second layer captures the details and realizes the scale-invariant feature transformation function. The third layer provides supplementary wireless signal information to improve the classification effect.

[0155] In step 300, a feature vector is constructed based on the scattering coefficient map and used as a training data set, specifically:

[0156] Extract the grayscale gradient co-occurrence matrix of the scattering coefficient image and obtain its image eigenvalues;

[0157] Construct a feature vector representing the wireless signal based on the image feature value and use it as a training data set;

[0158] A detailed description of this process:

[0159] The grayscale-gradient co-occurrence matrix can reflect the relationship between the grayscale and gradient of each pixel in the image. The grayscale of the image is the basis of the image, and the gradient of the image is the element that constitutes the edge contour of the image.

[0160] Therefore, the comprehensive information of grayscale and gradient can effectively characterize the texture characteristics of the image. The grayscale gradient space clearly depicts the grayscale and gradient resolution of each pixel in the image, and also gives the spatial relationship between each pixel and its domain pixel. It can well depict the texture of the image, and directional texture can be reflected from the direction of the gradient. The specific extraction steps are as follows:

[0161] 1. Assuming that the grayscale matrix of the image is f(i, j), calculate the gradient value of each pixel. The calculation formula is as follows:

[0162] g(i,j)=4f(i,j)-f(i+1,j)-f(i-1,j)-f(i,j+1)-f(i,j-1)

[0163] 2. Normalize the gradient value and grayscale value. The specific processing formula is as follows:

[0164]

[0165] In the formula, INT means round down, f max is the maximum gray value in the image, N f is the maximum grayscale value after normalization, which is 64. +1 is to ensure that the new pixel value is not 0. The final pixel value range is [0, N f ];

[0166]

[0167] In the formula, G(i, j) is the maximum gradient value within the range of the (i, j) pixel, g max is the maximum gradient value of the image; N g is the maximum gradient value after normalization, which is 64.

[0168] 3. Calculate the number of pixels that make F(i,j) = x and G(i,j) = y, which is the (x, y)th element of the grayscale gradient co-occurrence matrix H(x, y). Normalize it, and the formula is:

[0169]

[0170] There are 15 commonly used parameters of grayscale gradient co-occurrence matrix. Through comparative analysis, it is found that the grayscale distribution inhomogeneity, gradient distribution inhomogeneity, difference moment and inverse difference moment in the parameters of the grayscale gradient co-occurrence matrix of the extracted wavelet scattering coefficient map are similar in value, and the difference is not large, which is not conducive to the recognition and classification of the recognition model. Therefore, the final selection is: small gradient advantage, large gradient advantage, energy, grayscale average, gradient average, grayscale mean square error, gradient mean square error, correlation, grayscale entropy, gradient entropy, mixed entropy, a total of 11 eigenvalues, forming an eleven-dimensional eigenvector.

[0171] In step 400, a wireless time-frequency domain signal recognition model is constructed and trained based on a training data set, specifically:

[0172] The wireless time-frequency domain signal recognition model is constructed based on the SVM classifier. The specific steps are as follows:

[0173] 1. Input 11 parameters as feature parameters into the SVM classifier for training. The parameter expression is:

[0174] x i =(T1, T2, T3, T4, T5, T6, T7, T8, T9, T 10 , T 11 )

[0175] Where, T1 is the small gradient advantage; T2 is the large gradient advantage; T3 is the energy; T4 is the average grayscale; T5 is the average gradient; T6 is the grayscale mean square error; T7 is the gradient mean square error; T8 is the correlation; T9 is the grayscale entropy; T 10 is the gradient entropy; T 11 is the mixing entropy.

[0176] 2. Using the established recognition model as the main classification basis, establish a linear SVM and seek the optimal hyperplane, which can be transformed into

[0177]

[0178] Where ω is the normal vector of the hyperplane and b is the error.

[0179] Among them, subject to:

[0180] y i (ω T x i +b)≥1,i=1,2,…,p

[0181] In the formula, y i is the wireless signal category label; p is the total number of signals to be identified.

[0182] The wireless time-frequency domain signal recognition model is trained through the training data set based on the Bayesian SVM hyperparameter optimization algorithm. The specific process is as follows:

[0183] In order to select appropriate hyperparameters to complete the training of the SVM classifier and improve the recognition accuracy of SVM classification, practical Bayesian optimization is selected to solve this problem, that is, by establishing a probability model based on the past evaluation results of the objective function to find the value that minimizes the objective function, thereby achieving the purpose of optimization. The hyperparameters used for Bayesian optimization are the regularization coefficient and the gamma kernel parameter.

[0184] The goal of the hyperparameter optimization algorithm is to find the local optimal solution in the fastest way. To achieve this goal, the Bayesian optimization steps are:

[0185] 1. Determine the input quantity. The 11 parameters obtained by calculating the gray-level co-occurrence matrix of the wavelet scattering coefficient map are used as the calculation quantity, and the type label is used as the verification value.

[0186] 2. Determine the probability model. The probability model is used to replace the complex objective function of the original evaluation. The present invention selects Gaussian process regression.

[0187] 3. Train and verify the network. The set of hyperparameters with the smallest error during the optimization process is used as the optimized hyperparameters, and the training ends when the verification error meets the requirements.

[0188] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0189] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A wireless time-frequency domain signal recognition method based on wavelet time scattering network, characterized in that: include: Acquire wireless signals and pre-process them; Based on the preprocessed wireless signal, a scattering coefficient is obtained by wavelet scattering transform, and a scattering coefficient map is constructed; Construct feature vectors based on the scattering coefficient map and use it as a training data set; Build a wireless time-frequency domain signal recognition model and train it based on the training data set; The image texture features of the wavelet scattering coefficient spectrum of the wireless signal to be identified are input into the trained wireless time-frequency domain signal recognition model for identification to obtain the identification result.

2. The wireless time-frequency domain signal recognition method based on wavelet time scattering network according to claim 1 is characterized in that: Acquire wireless signals and preprocess them, specifically: Acquire wireless signals and perform denoising based on the IWOA-VMD algorithm; Perform irrelevant frequency band removal processing on the wireless signal after denoising; The wireless signal after the irrelevant frequency band is removed is normalized.

3. The wireless time-frequency domain signal recognition method based on wavelet time scattering network according to claim 2 is characterized in that: The wireless signal is acquired and denoised based on the IWOA-VMD algorithm. Specifically: Initialize the VMD algorithm parameter optimization range, and find the parameter combination corresponding to the minimum value of the composite fitness function according to the IWOA algorithm; Substitute the optimal parameter combination into the VMD decomposition model to decompose the wireless signal into a series of IMFs; The multi-scale permutation entropy-variance contribution rate selection principle is used to calculate and identify different IMFs, classify them into noise-dominated components and effective signal components, and perform NLM denoising on the noise-dominated components to obtain the denoised components. The denoised component and the effective signal component are reconstructed to obtain the denoised wireless signal.

4. The wireless time-frequency domain signal recognition method based on wavelet time scattering network according to claim 1 is characterized in that: Based on the preprocessed wireless signal, the scattering coefficient is obtained by wavelet scattering transform, and a scattering coefficient map is constructed, which is as follows: The pre-processed wireless signal is scattered through the first layer, that is, the signal is averaged through a wavelet low-pass filter to obtain the scattering scale coefficient of the first layer; The scattering scale coefficient of the first layer is used as the scattering coefficient of the second layer, and the scattering scale coefficient of the second layer is obtained through the wavelet scattering of the second layer. The scattering scale coefficient of the third layer is obtained through the same modular operation and wavelet low-pass filtering function. The scattering scale coefficient of the third layer is repeated to obtain the scale coefficient of the next layer, and the scattering coefficient map is constructed according to the obtained scattering scale coefficient.

5. The wireless time-frequency domain signal recognition method based on wavelet time scattering network according to claim 1 is characterized in that: Based on the scattering coefficient map, a feature vector is constructed and used as a training data set, specifically: Extract the grayscale gradient co-occurrence matrix of the scattering coefficient image and obtain its image eigenvalues; A feature vector representing the wireless signal is constructed based on the image feature value and used as a training data set.

6. The wireless time-frequency domain signal recognition method based on wavelet time scattering network according to claim 1 is characterized in that: Construct a wireless time-frequency domain signal recognition model and train it based on the training data set, specifically: A wireless time-frequency domain signal recognition model is constructed based on the SVM classifier, and the wireless time-frequency domain signal recognition model is trained through a training data set based on the Bayesian SVM hyperparameter optimization algorithm.

Images