Modulation recognition method and system based on fusion of numerical features and image features

Abstract

The present invention discloses a modulation recognition method and system based on the fusion of numerical features and image features, belonging to the field of signal recognition technology. It includes receiving a modulated signal; calculating the first characteristic parameter and the second characteristic parameter of the received modulated signal based on the high-order cumulative amount, and obtaining a one-dimensional numerical characteristic vector according to the first characteristic parameter and the second characteristic parameter; using a first transformation to perform time-frequency conversion on the modulated signal to obtain a time-frequency graph; inputting the time-frequency graph into a first neural network to extract a one-dimensional image feature vector; fusing the numerical feature vector and the image feature vector to obtain a fusion feature; sending the fusion feature to at least one fully connected layer and a Softmax layer connected in sequence for classification, and outputting a recognition result. The present invention has a high accuracy rate for modulated signals under low signal-to-noise ratio conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal recognition, and in particular relates to a modulation recognition method and system based on the fusion of numerical features and image features. Background Art

[0002] In order to improve the reliability and transmission efficiency of wireless communication systems, a variety of digital modulation methods with different principles and functions have been proposed. The core function of digital signal modulation recognition is to lay the foundation for signal demodulation analysis by judging the modulation type of the received signal. The information required for digital signal modulation recognition is implicit in the amplitude, phase and frequency of the signal. After linear or nonlinear transformation of the signal, the features that can distinguish the signal modulation type can be extracted. Then the statistical pattern recognition method is used to distinguish the signal. This method does not require prior information and can therefore be effectively applied to non-cooperative communications.

[0003] In the prior art, the type of modulation signal is usually identified by methods such as feature parameter extraction, wavelet transform, high-order cumulant extraction, fast Fourier transform, Gabor transform, Wigner-Ville distribution, etc., but these methods can only have a high recognition rate for a certain modulation type. However, under low signal-to-noise ratio, the recognition rate of the modulation signal is low using the existing methods. Summary of the invention

[0004] Technical problem: The present invention focuses on the problem of low recognition rate of phase modulation orthogonal amplitude modulation signals when using time-frequency image classification under low signal-to-noise ratio conditions, and proposes a modulation recognition method and system based on the fusion of numerical features and image features to improve the recognition rate under low signal-to-noise ratio.

[0005] Technical solution: In the first aspect, the present invention provides a modulation recognition method based on the fusion of numerical features and image features, comprising:

[0006] receiving a modulated signal;

[0007] Calculating a first characteristic parameter and a second characteristic parameter of the received modulated signal based on a high-order cumulant, and obtaining a one-dimensional numerical characteristic vector according to the first characteristic parameter and the second characteristic parameter;

[0008] Performing time-frequency conversion on the modulated signal using the first transformation to obtain a time-frequency diagram;

[0009] Inputting the time-frequency graph into a first neural network to extract a one-dimensional image feature vector;

[0010] Fuse the numerical feature vector and the image feature vector to obtain the fused feature;

[0011] The fused features are sent to at least one fully connected layer and Softmax layer connected in sequence for classification, and the recognition results are output.

[0012] Further, the method for calculating the first characteristic parameter and the second characteristic parameter of the received modulated signal based on the high-order cumulant is:

[0013] Performing Hilbert transform on the received modulated signal to obtain a zero-mean complex random signal;

[0014] Calculate the p+q order mixing matrix of the zero-mean complex random signal. The calculation formula is:

[0015] M p,q =E[X(t) (p-q) X * (t) q ]

[0016] Where X(t) represents a zero-mean complex random signal, and E[·] represents the mean.

[0017] The high-order cumulants of the mixing matrix are calculated using the following formula:

[0018]

[0019] Wherein, x represents the received signal after normalization processing.

[0020] The first characteristic parameter and the second characteristic parameter are calculated according to the cumulative amount of each order, and the calculation formula is:

[0021] f 1 =|C 41 | / |C 42

[0022] f 2 =|C 42 | / |C 21 | 2

[0023] Among them, f 1 is the first characteristic parameter, f 2 is the second characteristic parameter.

[0024] Furthermore, the first transform is a smoothed pseudo-Wigner-Ville distribution, a wavelet transform, a fast Fourier transform, a Gabor transform, or a Wigner-Ville distribution.

[0025] Furthermore, the first neural network includes a first convolutional layer, a first maximum pooling layer, a second convolutional layer, a second maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer and three fully connected layers arranged in sequence.

[0026] Furthermore, a method for obtaining a one-dimensional numerical feature vector according to the first feature parameter and the second feature parameter is:

[0027] The first characteristic parameter and the second characteristic parameter are respectively taken as 100 bits to form a numerical characteristic vector with a dimension of 200×1.

[0028] Furthermore, the fusion of the numerical feature vector and the image feature vector to obtain the fusion feature includes:

[0029] The 200×1 numerical feature vector is fused with the 256×1 image feature vector to obtain a 456×1 fused feature vector.

[0030] Furthermore, the modulation signal includes BPSK, QPSK, 16QAM, 64QAM, 4ASK, 2FSK and 4FSK.

[0031] Furthermore, the fused features are sent to two sequentially connected fully connected layers and Softmax layers for classification, and the recognition results are output, where the first one is a 256×1 fully connected layer and the second one is a 6×1 fully connected layer.

[0032] Furthermore, the method also includes training the first neural network and the neural network model consisting of the first neural network, at least one fully connected layer and a Softmax layer. During training, the first neural network is first pre-trained, and then the neural network model consisting of the first neural network, at least one fully connected layer and a Softmax layer is trained as a whole; or only the neural network model consisting of the first neural network, at least one fully connected layer and a Softmax layer is trained as a whole.

[0033] In a second aspect, a modulation recognition system based on the fusion of numerical features and image features is provided, wherein a modulation recognition method based on the fusion of numerical features and image features according to any one of the present inventions is used to recognize a modulation signal, including:

[0034] a signal receiving unit configured to receive a modulated signal;

[0035] A numerical feature extraction unit, configured to calculate a first characteristic parameter and a second characteristic parameter of the received modulated signal based on a high-order cumulant, and obtain a one-dimensional numerical feature vector according to the first characteristic parameter and the second characteristic parameter;

[0036] A time-frequency conversion unit configured to perform time-frequency conversion on the modulated signal using the first transformation to obtain a time-frequency graph;

[0037] An image feature extraction unit configured to input the time-frequency graph into a first neural network to extract a one-dimensional image feature vector;

[0038] A feature fusion unit configured to fuse the numerical feature vector and the image feature vector to obtain a fused feature;

[0039] The classification output unit is configured to send the fused features to at least one fully connected layer and a Softmax layer connected in sequence, and output the recognition result.

[0040] The present invention aims at the problem of low recognition rate of phase modulation orthogonal amplitude modulation signals when using time-frequency image classification under low signal-to-noise ratio conditions. The high-order cumulative amount of the received signal is calculated to obtain numerical features, and the received signal is transformed into a time-frequency image at the same time. The image features are extracted through a CNN network, and finally the two types of features are fused. The experimental results show that under the condition of about 0db, the feature fusion method can improve the recognition rate of phase modulation and orthogonal amplitude modulation signals by about 30%, which has a higher recognition accuracy rate than the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flow chart of a modulation recognition method based on the fusion of numerical features and image features in an embodiment of the present invention;

[0042] Figure 2 It is a time-frequency diagram of different signals after SPWVD transformation in an embodiment of the present invention;

[0043] Figure 3 This is a structural diagram of an image feature extraction model in an embodiment of the present invention;

[0044] Figure 4 It is a structural diagram of the feature fusion neural network model of the present invention;

[0045] Figure 5 A comparison chart of the average recognition rates before and after fusion in an embodiment of the present invention;

[0046] Figure 6 A comparison diagram of the confusion matrix before and after fusion when snr=0dB in an embodiment of the present invention;

[0047] Figure 7 4 is a block diagram of a modulation recognition system based on the fusion of numerical features and image features in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be further described below in conjunction with the embodiments and the accompanying drawings.

[0049] The mathematical model of the digital modulation signal of the communication system can be expressed as:

[0050]

[0051] Among them A m(t) represents the baseband signal, g(t) represents the baseband shaping filter, and f c is the carrier frequency, n(t) is Gaussian white noise, and t represents time.

[0052] When modulating the signal, technologies such as MPSK, MFSK, MQAM, and MASK can be used.

[0053] MPSK (Multi-level Frequency Shift Keying) is a digital modulation technology that controls the carrier phase based on signal data. Its modulation signal can be expressed as:

[0054]

[0055] Where m=1, 2…M.

[0056] MFSK (Multi-Frequency Shift Keying) is a digital modulation technology that controls the carrier frequency based on signal data. Its modulation signal can be expressed as:

[0057]

[0058] where f m =(2m-1-M)Δf.

[0059] MQAM (Multi-ary Quadrature Amplitude Modulation) is a technique of amplitude modulation and phase modulation. Its modulation signal can be expressed as:

[0060]

[0061] in

[0062] MASK (multi-level frequency shift keying), its modulation signal can be expressed as:

[0063]

[0064] where a m =0,1...M-1, the M mentioned above represents a base number.

[0065] In a first aspect, an embodiment of the present invention provides a modulation recognition method based on the fusion of numerical features and image features. Figure 1 The flowchart of the method in the embodiment of the present invention is shown. Figure 1 As shown, the method comprises the following steps:

[0066] Step S100: Receive a modulated signal. In the method of the embodiment of the present invention, a modulated signal from a Gaussian channel under different signal-to-noise ratios can be received. For example, in an example of the present invention, seven modulated signals including BPSK, QPSK, 16QAM, 64QAM, 4ASK, 2FSK, and 4FSK are included, the carrier frequency of the signal is 10KH, the sampling rate is 40KH, and the signal length is 256.

[0067] Step S200: Calculate the first characteristic parameter and the second characteristic parameter of the received modulated signal based on the high-order cumulative amount, and obtain a one-dimensional numerical characteristic vector according to the first characteristic parameter and the second characteristic parameter. This step is to extract the HOC (high-order cumulative amount) numerical feature. The specific method is:

[0068] First, the received modulated signal is Hilbert transformed to obtain a zero-mean complex random signal;

[0069] Secondly, calculate the p+q order mixing matrix of the zero-mean complex random signal, and the calculation formula is:

[0070] M p,q =E[X(t) (p-q) X * (t) q ]

[0071] Then, the high-order cumulant of the mixing matrix is ​​calculated, and the calculation formula is:

[0072]

[0073] For example:

[0074] C 21 =M 21

[0075]

[0076] C 41 =M 41 -3*M 20 *M 21

[0077] By calculating the high-order cumulants of the ideal modulation signal, we can get its theoretical value:

[0078] Table 1 Theoretical values ​​of high-order cumulants

[0079] C21 C40 C41 C42 BPSK E <![CDATA[2E 2 ]]> <![CDATA[2E 2 ]]> <![CDATA[2E 2 ]]> 4PSK E <![CDATA[E 2 ]]> 0 <![CDATA[E 2 ]]> 16QAM E <![CDATA[0.68E 2 ]]> 0 <![CDATA[0.68E 2 ]]> 64QAM E <![CDATA[0.619E 2 ]]> 0 <![CDATA[0.619E 2 ]]>

[0080] Finally, the first characteristic parameter f is calculated based on the cumulative amount of each order 1 and the second characteristic parameter f 2In the example of the present invention, the HOC characteristic parameters are extracted as auxiliary information for identifying 16QAM, 64QAM, QPSK, and BPSK, so the characteristic parameters are taken as follows:

[0081] f 1 =|C 41 | / |C 42 |

[0082] f 2 =|C 42 | / |C 21 | 2

[0083] The ideal BPSK signal f 1 =1, f of ideal QPSK signal 1 = 0, f of ideal 16QAM signal 2 =0.68, f of ideal 64QAM signal 2 =0.619.

[0084] In the embodiment of the present invention, in order to increase the weight of the numerical feature parameter, the first feature parameter and the second feature parameter are respectively taken as 100 bits to form a 200×1 numerical feature vector.

[0085] Step S300: Perform time-frequency conversion on the modulated signal using the first transformation to obtain a time-frequency diagram. In the embodiment of the present invention, smooth pseudo-Wigner-Ville distribution, wavelet transform, fast Fourier transform, Gabor transform, or Wigner-Ville distribution can be used for time-frequency conversion.

[0086] In a preferred example of the present invention, a smooth pseudo-Wigner-Viller distribution is selected for time-frequency conversion. As the most basic type of Cohen-class bilinear time-frequency distribution, the Wigner-Viller distribution (WVD) for a certain deterministic time continuous signal x(t) is defined as:

[0087]

[0088] Where v represents the angular frequency and x represents the time domain signal

[0089] This transformation can be regarded as multiplying the signal at a specific moment in the past with the signal at a specific moment in the future, and taking the Fourier transform of the time difference τ. This type of transformation can effectively extract the modulation information such as amplitude and frequency of digital signals, and can be used in signal modulation recognition. Its disadvantage is that it cannot guarantee non-negativity, will produce serious cross-term interference, and blur the original characteristics of the signal.

[0090] Pseudo Wigner-Ville distribution (pseudo WVD, PWVD) is a windowing process of the basic WVD, and its distribution definition is:

[0091]

[0092] Where h(t) is a regular window. The result of windowing compresses the cross terms of multi-component signals to some extent, but at the same time destroys some edge characteristics of WVD. In order to realize the positive energy distribution, the WVD can be convolved with a smoothing function to obtain a smooth pseudo Wigner-Ville distribution (SPWVD), whose distribution definition is:

[0093]

[0094] Among them, g() and h() represent window functions

[0095] SPWVD distribution has a good effect of eliminating cross-terms, and its time-frequency characteristics and aggregation performance are also maintained well. In the example of the present invention, SPWVD is used as an estimate of time-frequency analysis, and the time-frequency diagram of the signal is obtained through the tftb toolbox of MATALB. The time-frequency diagrams of different modulated signals are shown in FIG. Figure 2 shown.

[0096] Step S400: Input the time-frequency graph into the first neural network to extract a one-dimensional image feature vector. In an embodiment of the present invention, a convolutional neural network is used to extract and identify signal features. In a preferred example of the present invention, the first neural network is a convolutional neural network, such as Figure 3 As shown, the convolutional neural network includes a first convolutional layer, a first maximum pooling layer, a second convolutional layer, a second maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer and three fully connected layers arranged in sequence. After the last fully connected layer, a 256×1 image feature vector is obtained. In the neural network constructed in this embodiment, all pooling layers use maximum pooling to avoid the blurring effect of average pooling.

[0097] It should be noted that, in other embodiments, steps S200 to S400 are not limited to the above order, and S300 to S400 may be performed first, and then S200.

[0098] Step S500: Fusing the numerical feature vector and the image feature vector to obtain a fused feature. In the embodiment of the present invention, the numerical feature vector is a vector with a dimension of 200×1, and the image feature vector is a vector with a dimension of 256×1. By fusing in the vector dimension, a 456×1 fused feature is obtained after fusion. The feature extraction and fusion process is as follows: Figure 4 shown.

[0099] Step S600: Send the fused features to at least one fully connected layer and Softmax layer connected in sequence, and output the recognition result. In a preferred embodiment of the present invention, the fused features are sent to two fully connected layers connected in sequence, wherein the first is a 256×1 fully connected layer and the second is a 6×1 fully connected layer. The fused features are classified by passing through a 256×1 and a 6×1 fully connected layer and a Softmax layer to obtain the output, and the type of modulation confidence can be identified.

[0100] It is further explained that when using a convolutional neural network for operations such as feature extraction, the convolutional neural network must be trained. In an embodiment of the present invention, the training can be performed in the following manner. First, pre-train the first neural network, and then train the neural network model consisting of the first neural network, at least one fully connected layer and a Softmax layer as a whole. The form of overall training can also be adopted, that is, only the neural network model consisting of the first neural network, the fully connected layer and the Softmax layer is trained as a whole. During training, the cross entropy function is used to calculate the loss, which represents the difference between the true prediction probability and the expected prediction probability. Its mathematical expression is:

[0101]

[0102] Among them, p(x) represents the true probability distribution, q(x) represents the predicted probability distribution, and the smaller the cross entropy value, the higher the prediction accuracy of the model. At the same time, in order to minimize the loss value, the Adam (Adaptive Moment Estimation) algorithm is used in network training, and the learning rate of each parameter is dynamically adjusted using the first-order moment estimation and second-order moment estimation of the gradient. Its update algorithm can be expressed as:

[0103] g t =▽ θ f t (θ t-1 )

[0104] m t =β 1 *m t-1 +(1-β 1 )*g t

[0105]

[0106] where g t is the gradient at time step t, θ is the parameter to be updated, f(θ) is the random objective function, β 1 , β 2 are the exponential decay rates of the first-order moment and the second-order moment, respectively, m tis the first-order moment estimate of the gradient, n t is the second-order moment estimate of the gradient, is the calibrated parameter, α is the learning rate, and ε is the constant added to maintain numerical stability. The advantage of this algorithm is that after bias correction, the stability of the parameters is guaranteed by limiting the range of the learning rate at each iteration.

[0107] In order to verify the effectiveness of the method in the embodiment of the present invention, a simulation comparison was conducted on the classification effect using time-frequency features before feature fusion and after fusion. The results are as follows: Figure 5 As shown in the figure, the average recognition rate comparison before and after fusion under different signal-to-noise ratios is obtained. The blue solid line represents the recognition rate before fusion, and the red dotted line represents the recognition rate after fusion. It can be seen that the average recognition rate is increased by about 30% after feature fusion. The confusion matrix comparison at 0dB is further given, as shown in Figure 6 As shown in the figure, the feature fusion method can not only maintain the good classification performance of time-frequency features for ASK and FSK, but also improve the recognition rate of PSK and QAM signals to a higher level.

[0108] In a second aspect, the present invention provides a modulation recognition system based on the fusion of numerical features and image features. The modulation recognition method based on the fusion of numerical features and image features proposed in the present invention recognizes the modulation signal, such as Figure 7 As shown, the system includes:

[0109] a signal receiving unit configured to receive a modulated signal;

[0110] A numerical feature extraction unit, configured to calculate a first characteristic parameter and a second characteristic parameter of the received modulated signal based on a high-order cumulant, and obtain a one-dimensional numerical feature vector according to the first characteristic parameter and the second characteristic parameter;

[0111] A time-frequency conversion unit configured to perform time-frequency conversion on the modulated signal using the first transformation to obtain a time-frequency graph;

[0112] An image feature extraction unit configured to input the time-frequency graph into a first neural network to extract a one-dimensional image feature vector;

[0113] A feature fusion unit configured to fuse the numerical feature vector and the image feature vector to obtain a fused feature;

[0114] The classification output unit is configured to send the fused features to at least one fully connected layer and a Softmax layer connected in sequence, and output the recognition result.

[0115] The specific functional implementation of the above functional modules corresponds to the corresponding steps of the modulation recognition method based on the fusion of numerical features and image features provided in the embodiments of the present invention, and will not be repeated here.

[0116] The above embodiments are only preferred implementation modes of the present invention. It should be pointed out that ordinary technicians in this technical field can make several improvements and equivalent substitutions without departing from the principles of the present invention. These technical solutions after improvements and equivalent substitutions to the claims of the present invention all fall within the protection scope of the present invention.

Claims

1. A modulation recognition method based on the fusion of numerical features and image features, characterized in that: include: receiving a modulated signal; Calculating a first characteristic parameter and a second characteristic parameter of the received modulated signal based on a high-order cumulant, and obtaining a one-dimensional numerical characteristic vector according to the first characteristic parameter and the second characteristic parameter; The method for calculating the first characteristic parameter and the second characteristic parameter of the received modulated signal based on the high-order cumulant is: Performing Hilbert transform on the received modulated signal to obtain a zero-mean complex random signal; Calculate the p+q order mixing matrix of the zero-mean complex random signal. The calculation formula is: M p,q =E[X(t) (p-q) X * (t) q ] Where X(t) represents a zero-mean complex random signal, and E[·] represents the mean value. The high-order cumulants of the mixing matrix are calculated using the following formula: Wherein, x represents the received signal after normalization processing; The first characteristic parameter and the second characteristic parameter are calculated according to the cumulative amount of each order, and the calculation formula is: f1=|C 41 | / |C 42 | f2=|C 42 / |C 21 | 2 Among them, f1 is the first characteristic parameter, f2 is the second characteristic parameter; Performing time-frequency conversion on the modulated signal using the first transformation to obtain a time-frequency diagram; Smoothed pseudo-Wigner-Williams distribution was used for time-frequency conversion; The Wigner-Willi distribution for a deterministic time-continuous signal x(t) is defined as: Where v represents the angular frequency and x represents the time domain signal; The pseudo-Wigner-Willi distribution is a windowed version of the basic Wigner-Willi distribution, and its distribution definition is: Where h(t) is the regular window; Convolving the Wigner Willy distribution with the smoothing function yields a smoothed pseudo-Wigner Willy distribution, whose distribution definition is: Among them, g() and h() represent window functions; Inputting the time-frequency graph into a first neural network to extract a one-dimensional image feature vector; Fuse the numerical feature vector and the image feature vector to obtain the fused feature; Send the fused features to at least one fully connected layer and Softmax layer connected in sequence for classification, and output the recognition result; Convolutional neural network is used to extract and identify signal features, and the convolutional neural network is trained. The training method is: Pre-training the first neural network, and then training the neural network model consisting of the first neural network, at least one fully connected layer and a Softmax layer as a whole; or, only training the neural network model consisting of the first neural network, the fully connected layer and the Softmax layer as a whole; During training, the cross entropy function is used to calculate the loss, which represents the difference between the true prediction probability and the expected prediction probability. Its mathematical expression is: Among them, p(x) represents the true probability distribution, and q(x) represents the predicted probability distribution; The Adam algorithm is used in network training, and the first-order moment estimation and second-order moment estimation of the gradient are used to dynamically adjust the learning rate of each parameter. The update algorithm is: m t =β1*m t-1 +(1-β1)*g t Among them, g t is the gradient at time step t, θ is the parameter to be updated, f(θ) is the random target parameter, β1 and β2 are the exponential decay rates of the first-order moment and the second-order moment, respectively, and m t is the first-order moment estimate of the gradient, nt is the second-order moment estimate of the gradient, is the calibrated parameter, α is the learning rate, and ε is a constant added to maintain numerical stability; The method of obtaining a one-dimensional numerical eigenvector according to the first eigenvalue and the second eigenvalue is: The first characteristic parameter and the second characteristic parameter are respectively taken as 100 bits to form a numerical characteristic vector with a dimension of 200×1.

2. The method according to claim 1, characterized in that The first neural network includes a first convolutional layer, a first maximum pooling layer, a second convolutional layer, a second maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer and three fully connected layers, which are arranged in sequence.

3. The method according to claim 2, characterized in that The fusion of the numerical feature vector and the image feature vector to obtain the fusion feature includes: The 200×1 numerical feature vector is fused with the 256×1 image feature vector to obtain a 456×1 fused feature vector.

4. The method according to any one of claims 1 to 3, characterized in that: The modulation signals include BPSK, QPSK, 16QAM, 64QAM, 4ASK, 2FSK, and 4FSK.

5. The method according to claim 4, characterized in that The fused features are sent to two fully connected layers and Softmax layers connected in sequence for classification, and the recognition results are output. The first one is a 256×1 fully connected layer and the second one is a 6×1 fully connected layer.

6. The method according to claim 5, characterized in that The method further includes training the first neural network and the neural network model consisting of the first neural network, at least one fully connected layer and a Softmax layer, wherein during the training, the first neural network is first pre-trained, and then the neural network model consisting of the first neural network, at least one fully connected layer and a Softmax layer is trained as a whole; Or only the neural network model consisting of the first neural network, at least one fully connected layer and a Softmax layer is trained as a whole.

7. A modulation recognition system based on the fusion of numerical features and image features, identifying a modulation signal according to the modulation recognition method based on the fusion of numerical features and image features according to any one of claims 1 to 6, characterized in that: include: a signal receiving unit configured to receive a modulated signal; A numerical feature extraction unit, configured to calculate a first characteristic parameter and a second characteristic parameter of the received modulated signal based on a high-order cumulant, and obtain a one-dimensional numerical feature vector according to the first characteristic parameter and the second characteristic parameter; A time-frequency conversion unit configured to perform time-frequency conversion on the modulated signal using the first transformation to obtain a time-frequency graph; An image feature extraction unit configured to input the time-frequency graph into a first neural network to extract a one-dimensional image feature vector; A feature fusion unit configured to fuse the numerical feature vector and the image feature vector to obtain a fused feature; The classification output unit is configured to send the fused features to at least one fully connected layer and a Softmax layer connected in sequence, and output the recognition result.

Images