Modulation signal identification method based on continuous wavelet transform and deep learning

Abstract

The invention discloses a modulation signal identification method based on continuous wavelet transform and deep learning, and the method comprises the steps: firstly modulating a to-be-transmitted sequence at a transmitting end, obtaining a wireless continuous time signal at a receiving end, sampling the continuous time signal, carrying out the wavelet decomposition of the sampled data, and filtering out the high-frequency noise; carrying out reconstruction by using wavelet inverse transformation, and carrying out median filtering on the reconstructed signal to obtain a new data set; performing continuous wavelet transform on the new data set to obtain a time-frequency graph, and extracting feature information of a time domain and a frequency domain of the time-frequency graph to obtain a two-dimensional feature graph; constructing a convolutional neural network prediction algorithm model, and training the convolutional neural network prediction algorithm model by using the two-dimensional feature map extracted by continuous wavelet transform; and finally, classifying the modulation modes of the signals by using the trained convolutional neural network prediction algorithm model. According to the method, the modulation signal identification accuracy in a complex environment in reality is improved, the network parameter quantity is reduced, and the modulation signal identification efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of modulation signal recognition, and in particular to a modulation signal recognition method based on continuous wavelet transform and deep learning. Background Art

[0002] The modulation signal classification process is usually performed before signal demodulation and is widely used in military and civilian fields. Early modulation signal recognition was mainly done manually, which was inefficient and difficult to sustain. With the advancement of technology and the increasing demand for frequency perception, Automatic Modulation Recognition (AMR) technology has gradually gained attention. Automatic modulation recognition technology is a key technology for detecting target signals at the receiving end of radio communication and is quite important in the field of signal recognition. Current AMR technology can be divided into three categories: hypothesis testing-based methods, feature extraction-based feature recognition methods, and deep learning-based methods.

[0003] The identification methods based on hypothesis testing mainly include the average likelihood ratio test method and the generalized likelihood ratio test method. The average likelihood ratio test method regards the unknown variable as a random variable and obtains the likelihood function by calculating its average value. The generalized likelihood ratio test method calculates the probability density function of the input signal based on the maximum likelihood estimate of the unknown quantity and determines the modulation mode accordingly. Although the hypothesis testing-based method can theoretically achieve the best classification performance, it not only requires a lot of prior knowledge, but also consumes a lot of computing resources.

[0004] The feature recognition method based on feature extraction mainly relies on the extraction of features such as instantaneous parameters, cyclic spectra and high-order cumulants, combined with machine learning techniques such as support vector machines, decision trees and k-nearest neighbor algorithms. Although the recognition effect of this method is not as good as the hypothesis testing method, the required prior knowledge and computing resources are significantly reduced. However, the recognition rate of the feature recognition method based on feature extraction depends largely on the quality of the manually selected features.

[0005] In response to the problems caused by the above modulation signal recognition methods, deep learning-based methods have been widely used in the field of AMR technology in recent years due to their efficient feature extraction capabilities in image processing. The network architecture commonly used in automatic modulation recognition includes convolutional neural networks (CNN), recurrent neural networks (RNN), or a combination of the two. These architectures stack multiple convolutional layers, recurrent neural network layers, and fully connected layers to ultimately output the probability distribution of the modulation type. Summary of the invention

[0006] The object of the present invention is to provide a modulation signal recognition method based on continuous wavelet transform and deep learning, which has high recognition efficiency, small computational workload, good modulation signal recognition and classification performance, and high robustness.

[0007] The technical solution to achieve the purpose of the present invention is: a modulation signal recognition method based on continuous wavelet transform and deep learning, comprising the following steps:

[0008] Step 1: modulate the sequence to be transmitted at the transmitting end, obtain the wireless continuous time signal at the receiving end, and sample the continuous time signal;

[0009] Step 2: Perform wavelet decomposition on the received signal, select a preset threshold rule to filter out high-frequency noise, reconstruct the obtained wavelet coefficients using inverse wavelet transform, and perform median filtering on the reconstructed signal to obtain a new data set;

[0010] Step 3: Perform continuous wavelet transform on the new data set obtained in step 2 to obtain a time-frequency graph, downsample the effective frequency components of the time-frequency graph, extract feature information in the time domain and frequency domain, and obtain a two-dimensional feature graph;

[0011] Step 4: Construct a convolutional neural network prediction algorithm model;

[0012] Step 5: Using the neural network’s ability to extract signal features, the processed feature map is used as input to train the convolutional neural network prediction algorithm model;

[0013] Step 6: Use the trained convolutional neural network prediction algorithm model and the generated image data test set as input to classify the modulation mode of the signal.

[0014] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the modulation signal recognition method based on continuous wavelet transform and deep learning is implemented.

[0015] Compared with the prior art, the present invention has the following significant advantages: (1) It utilizes wavelet threshold denoising, median filtering and continuous wavelet transform to map the frequency, phase and amplitude information of the signal, thereby generating a two-dimensional feature map, effectively extracting the features of the signal at the data set level, and is applicable to real scenes, effectively improving the accuracy of modulated signal recognition in complex environments in reality; (2) Based on the feature extraction capability of convolutional neural networks for signals, a lightweight deep learning network is designed, which reduces the number of network parameters while ensuring feature learning capabilities, thereby improving the efficiency of modulated signal recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of a modulation signal recognition method based on continuous wavelet transform and deep learning in the present invention.

[0017] Figure 2This is a two-dimensional feature map generated when SNR=20dB in an embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of the structure of the convolutional neural network prediction algorithm model built in an embodiment of the present invention.

[0019] Figure 4 This is a graph showing the convergence speed of the convolutional neural network prediction algorithm model in an embodiment of the present invention.

[0020] Figure 5 This is a recognition performance diagram when the SNR is -10dB to 20dB in an embodiment of the present invention.

[0021] Figure 6 This is a confusion matrix diagram when SNR=-10dB in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to facilitate the understanding of the present invention, the present invention is described in more detail below in conjunction with the accompanying drawings and specific embodiments. Preferred embodiments of the present invention are provided in the accompanying drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in this specification. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0023] It should be noted that, unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0024] like Figure 1 As shown, the present invention provides a modulation signal recognition method based on continuous wavelet transform and deep learning, comprising the following steps:

[0025] Step 1: modulate the sequence to be transmitted at the transmitting end, obtain the wireless continuous time signal at the receiving end, and sample the continuous time signal;

[0026] Step 2: Perform wavelet decomposition on the received signal, select a preset threshold rule to filter out high-frequency noise, reconstruct the obtained wavelet coefficients using inverse wavelet transform, and perform median filtering on the reconstructed signal to obtain a new data set;

[0027] Step 3: Perform continuous wavelet transform on the new data set obtained in step 2 to obtain a time-frequency graph, downsample the effective frequency components of the time-frequency graph, extract feature information in the time domain and frequency domain, and obtain a two-dimensional feature graph;

[0028] Step 4: Construct a convolutional neural network prediction algorithm model;

[0029] Step 5: Using the neural network’s ability to extract signal features, the processed feature map is used as input to train the convolutional neural network prediction algorithm model;

[0030] Step 6: Use the trained convolutional neural network prediction algorithm model and the generated image data test set as input to classify the modulation mode of the signal.

[0031] As a specific example, in step 1, the sequence to be transmitted is modulated at the transmitting end, a wireless continuous time signal is obtained at the receiving end, and the continuous time signal is sampled, as follows:

[0032] Step 1.1: Modulate the sequence to be transmitted at the transmitting end, and let x(t) be the transmitted signal, as shown below:

[0033] x(t)=F(s i )

[0034] Among them, s i is the i-th code element received, t is the time, and F is the modulation function;

[0035] Step 1.2: The signal received by the receiving end is r(t), as shown below:

[0036] r(t)=h(t)*x(t)+n(t)

[0037] Where h(t) is the channel response, n(t) is the additive noise, and * is the convolution operation;

[0038] Given a received signal r(t), the purpose of modulation signal recognition is to extract the frequency, phase and amplitude information of different modulation signals from r(t), so as to identify the modulation type of F;

[0039] Step 1.3, sample the continuous-time signal as follows:

[0040]

[0041] Among them, r[n] is the length of the discrete time series, T s is the sampling interval; during the sampling process, the sampling rate should be higher than the Nyquist sampling rate.

[0042] As a specific example, in step 2, the received signal is subjected to wavelet decomposition, a pre-set threshold rule is selected to filter out high-frequency noise, the obtained wavelet coefficients are reconstructed using inverse wavelet transform, and the reconstructed signal is subjected to median filtering to obtain a new data set, as follows:

[0043] Step 2.1: Perform wavelet decomposition on the received signal, as shown below:

[0044]

[0045] Among them, j is the number of layers of wavelet decomposition, q is the filter length, and a j (k) is the low-frequency wavelet coefficient after decomposition, b j (k) is the high-frequency wavelet coefficient after decomposition, h d (n) represents the low-pass wavelet filter coefficient, g d (kn) represents the high-pass wavelet filter coefficient, Indicates a j-1 Downsampling by a factor of 2 (n);

[0046] Step 2.2, for the low-frequency and high-frequency wavelet coefficients generated by wavelet decomposition, select the Huarsure threshold rule to filter out the high-frequency noise; the Huarsure threshold rule is composed of the Sqtwolog and Rigrsure threshold rules, and the smaller value of the two is selected as the threshold, and the high-frequency wavelet coefficients are threshold processed to obtain the wavelet coefficients with the high-frequency noise filtered out;

[0047] Step 2.3: Reconstruct the wavelet coefficients after filtering out high-frequency noise using inverse wavelet transform, as shown below:

[0048]

[0049] in, is the reconstructed low-frequency wavelet coefficient, b j (k) is the reconstructed high-frequency wavelet coefficient, h r (n) and g r (n) represent the reconstruction filter coefficients of low frequency and high frequency, respectively, and Express and 2 times upsampling;

[0050] Step 2.4: The median filter takes the median of the window near a point in the discrete signal as the value of the point to make the signal smoother, thereby achieving the purpose of denoising. The signal is filtered using a one-dimensional median filter with a window of 9, as shown below:

[0051]

[0052] in, is the denoised discrete time series, is the sequence after wavelet reconstruction, and the Mid function represents the median of the reconstructed sequence with a window of 9.

[0053] As a specific example, in step 3, a continuous wavelet transform is performed on the new data set obtained in step 2 to obtain a time-frequency graph, and the effective frequency components of the time-frequency graph are downsampled to extract feature information in the time domain and the frequency domain, as follows:

[0054] Step 3.1: Perform continuous wavelet transform on the discrete time series r%[n] as follows:

[0055]

[0056] Among them, W(c m ,d k ) is the time-frequency diagram obtained after continuous wavelet transform, c m is the scale, d k is the translation, ψ(n) is the mother wavelet basis function; the translation refers to moving the wavelets of different scales from the starting position of the signal to the ending position; the scale refers to controlling the expansion and contraction of the wavelet function, indicating the scaling of the signal in time;

[0057] The mother wavelet basis function used is the Morlet function, as shown below:

[0058]

[0059] Among them, ω0 is the center frequency of the wavelet, and the range of ω0 is set to 0 to f by setting the appropriate scale. s / 2; the translation amount increases in intervals of 1;

[0060] Step 3.2: Move the Morlet wavelet functions of different scales and times along the entire signal to generate a coefficient matrix containing scale and translation information; select the frequency number with the largest value on the scale component as the center frequency of the two-dimensional feature map, and downsample the time component to complete the tracking of the effective frequency component.

[0061] As a specific example, in step 4, a convolutional neural network prediction algorithm model is constructed as follows:

[0062] Step 4.1, build a convolutional neural network architecture through convolutional layer, ReLU layer, and maximum pooling layer;

[0063] Step 4.2: Build a deep learning network through a two-layer convolutional neural network architecture;

[0064] Step 4.3: Build a convolutional neural network prediction algorithm model through a deep learning network.

[0065] As a specific example, in step 4.1, a convolutional neural network architecture is constructed through a convolutional layer, a ReLU layer, and a maximum pooling layer, as follows:

[0066] Convolutional neural network architecture is constructed through convolutional layer, ReLU layer, and maximum pooling layer, including convolutional layer-ReLU layer-convolutional layer-ReLU layer-maximum pooling layer set in sequence;

[0067] The convolutional layer learns different features of the two-dimensional image input by the input layer;

[0068] The ReLU layer uses an activation function to remove negative values ​​in the convolution result and keep the positive values ​​unchanged;

[0069] The pooling layer is used to aggregate the two-dimensional graph into a higher dimension to reduce the spatial size of the feature map and remove redundant spatial information.

[0070] As a specific example, in step 4.2, a deep learning network is constructed through a two-layer convolutional neural network architecture, as follows:

[0071] The deep learning network is constructed by using input layer, two-layer convolutional neural network architecture, Flatten layer, fully connected layer, softmax layer, and classification layer, including input layer-two-layer convolutional neural network architecture-Flatten layer-fully connected layer-softmax layer-classification layer set in sequence;

[0072] In the first layer of the convolutional neural network architecture, the size of both convolution kernels is 2×2;

[0073] In the second-layer convolutional neural network architecture, the sizes of the two convolution kernels are 2×6 and 4×2 respectively.

[0074] As a specific example, in step 4.3, a convolutional neural network prediction algorithm model is constructed through a deep learning network, as follows:

[0075] The convolutional neural network prediction algorithm model includes the following arranged in sequence: input layer-first convolutional neural network architecture-second convolutional neural network architecture-Flatten layer-first fully connected layer-fifth ReLU layer-second fully connected layer-sixth ReLU layer-third fully connected layer-softmax layer-classification layer;

[0076] The first convolutional neural network architecture includes the following: first convolutional layer - first ReLU layer - second convolutional layer - second ReLU layer - first maximum pooling layer.

[0077] The second convolutional neural network architecture includes the following arranged in sequence: the third convolutional layer - the third ReLU layer - the fourth convolutional layer - the fourth ReLU layer - the second maximum pooling layer;

[0078] In the first convolutional neural network architecture, the convolution layers are all set to traversal sampling with padding; the first convolution layer contains 64 convolution kernels of size 2×2; the second convolution layer contains 128 convolution kernels of size 2×2; the pooling window of the first maximum pooling layer is 1×2;

[0079] In the second convolutional neural network architecture, the convolutional layers are all set to traversal sampling without padding; the third convolutional layer contains 64 convolution kernels of size 2×6; the fourth convolutional layer contains 16 convolution kernels of size 4×2, and the pooling window of the maximum pooling layer is 1×4;

[0080] The first layer of the convolutional neural network architecture is used to increase the dimension of the feature map and reduce the length of the time component in the feature map to half of the original length; the second layer of the convolutional neural network architecture has a larger receptive field and can reduce the length of the time and translation components at the same time;

[0081] The first fully connected layer includes 64 neurons, the second fully connected layer includes 16 neurons, and the third fully connected layer includes 8 neurons; the softmax layer uses the softmax activation function to normalize the output of the third fully connected layer and express it as the probability of the classification layer; the classification layer returns the corresponding classification probability for each input and calculates the loss;

[0082] The convolutional neural network prediction algorithm model uses the Adam optimizer, and the learning rate is set to 0.0001.

[0083] As a specific example, in step 5, the signal feature extraction capability of the neural network is utilized, and the processed feature map is used as input to train the convolutional neural network prediction algorithm model, as follows:

[0084] Step 5.1, load the generated two-dimensional feature data set;

[0085] Step 5.2, the generated two-dimensional feature atlas is divided into an image data training set, a validation set and a test set after data splitting;

[0086] Step 5.3: Use the image data training set and validation set to train the constructed convolutional neural network prediction algorithm model.

[0087] Step 6: Use the trained convolutional neural network prediction algorithm model and the generated image data test set as input to classify the modulation mode of the signal.

[0088] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the modulation signal recognition method based on continuous wavelet transform and deep learning is implemented.

[0089] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0090] Example

[0091] like Figure 1 As shown, the modulation signal recognition method based on continuous wavelet transform and deep learning provided in this embodiment includes the following steps:

[0092] Step 1: modulate the sequence to be transmitted at the transmitting end, obtain the wireless continuous time signal at the receiving end, and sample the continuous time signal, as follows:

[0093] Step 1.1: Modulate the sequence to be transmitted at the transmitting end, and let x(t) be the transmitted signal, as shown below:

[0094] x(t)=F(s i )

[0095] Among them, s i is the i-th code element received, t is the time, and F is the modulation function;

[0096] Step 1.2: The signal received by the receiving end is r(t), as shown below:

[0097] r(t)=h(t)*x(t)+n(t)

[0098] Where h(t) is the channel response, n(t) is the additive noise, and * is the convolution operation. Given the received signal r(t), the purpose of modulation signal recognition is to extract the frequency, phase, and amplitude information of different modulation signals from r(t), so as to identify the modulation type of F.

[0099] Step 1.3, sample the continuous-time signal as follows:

[0100]

[0101] Among them, r[n] is the length of the discrete time series, T s is the sampling interval; during the sampling process, the sampling rate should be higher than the Nyquist sampling rate.

[0102] By using MATLAB software to simulate real channel interference, 8 different types of modulation signals are generated as data sets, including: 2PSK, 2ASK, 2FSK, 4PSK, 4ASK, 4FSK, 16QAM and 64QAM. A total of 11520 samples are generated as training sets, 1280 samples as validation sets and 12800 samples as test sets. Each modulation signal is a real signal, and each instance consists of 100 samples. Each sample selects a continuous time signal with a fixed window of 0.128 seconds, and after sampling, a time series with a length of 2000 is selected as the original data. The signal-to-noise ratio of the samples is uniformly distributed between -10dB and 20dB at intervals of 2dB. Under the condition that the signal-to-noise ratio is unknown, the labels of the generated test set include the modulation type and the true value of the signal-to-noise ratio, while the labels of the training set and validation set only include the modulation type.

[0103] Step 2: Perform wavelet decomposition on the received signal, select a preset threshold rule to filter out high-frequency noise, reconstruct the obtained wavelet coefficients using inverse wavelet transform, and perform median filtering on the reconstructed signal to obtain a new data set, as follows:

[0104] Step 2.1: Perform wavelet decomposition on the received signal, as shown below:

[0105]

[0106] Among them, j is the number of layers of wavelet decomposition, q is the filter length, and a j (k) is the low-frequency wavelet coefficient after decomposition, b j (k) is the high-frequency wavelet coefficient after decomposition, h d (n) represents the low-pass wavelet filter coefficient, g d (kn) represents the high-pass wavelet filter coefficient, Indicates a j-1 Downsampling by a factor of 2 (n);

[0107] Step 2.2, for the low-frequency and high-frequency wavelet coefficients generated by wavelet decomposition, select the Huarsure threshold rule to filter out the high-frequency noise; the Huarsure threshold rule is composed of the Sqtwolog and Rigrsure threshold rules, and the smaller value of the two is selected as the threshold, and the high-frequency wavelet coefficients are threshold processed to obtain the wavelet coefficients with the high-frequency noise filtered out;

[0108] Step 2.3: Reconstruct the wavelet coefficients after filtering out high-frequency noise using inverse wavelet transform, as shown below:

[0109]

[0110] in, is the reconstructed low-frequency wavelet coefficient, b j (k) is the reconstructed high-frequency wavelet coefficient, h r (n) and g r (n) represent the reconstruction filter coefficients of low frequency and high frequency, respectively, and Express and 2 times upsampling;

[0111] Step 2.4: The median filter takes the median of the window near a point in the discrete signal as the value of the point to make the signal smoother, thereby achieving the purpose of denoising. The signal is filtered using a one-dimensional median filter with a window of 9, as shown below:

[0112]

[0113] in, is the denoised discrete time series, is the sequence after wavelet reconstruction, and the Mid function represents the median of the reconstructed sequence with a window of 9.

[0114] Step 3: Perform continuous wavelet transform on the new data set obtained in step 2 to obtain a time-frequency graph, and downsample the effective frequency components of the time-frequency graph to extract feature information in the time domain and frequency domain, as follows:

[0115] Step 3.1: Discrete time series Perform a continuous wavelet transform as follows:

[0116]

[0117] Among them, W(c m ,d k ) is the time-frequency diagram obtained after continuous wavelet transform, c m is the scale, d k is the translation, ψ(n) is the mother wavelet basis function;

[0118] Translation refers to moving wavelets of different scales from the starting position of the signal to the ending position; scale refers to controlling the expansion and contraction of the wavelet function, which represents the scaling of the signal in time;

[0119] The mother wavelet basis function used is the Morlet function, as shown below:

[0120]

[0121] Among them, ω0 is the center frequency of the wavelet, and the range of ω0 is set to 0 to f by setting the appropriate scale. s / 2; the translation amount increases in intervals of 1;

[0122] Step 3.2: Move the Morlet wavelet functions of different scales and times along the entire signal to generate a coefficient matrix containing scale and translation information; select the frequency number with the largest value on the scale component in the 2D time-frequency diagram as the center frequency of the two-dimensional feature map, and downsample the time component by 20 times to complete the tracking of the effective frequency component, and obtain a 15×100 two-dimensional image.

[0123] like Figure 2 As shown, the feature graphs of each modulation method are obviously different, and a suitable algorithm model is built based on this to classify and identify them.

[0124] Step 4: Construct a convolutional neural network prediction algorithm model, as follows:

[0125] Step 4.1: Construct a convolutional neural network architecture through convolutional layer, ReLU layer, and maximum pooling layer, as follows:

[0126] Convolutional neural network architecture is constructed through convolutional layer, ReLU layer, and maximum pooling layer, including convolutional layer-ReLU layer-convolutional layer-ReLU layer-maximum pooling layer set in sequence;

[0127] The convolutional layer learns different features of the two-dimensional image input by the input layer;

[0128] The ReLU layer uses an activation function to remove negative values ​​in the convolution result and keep the positive values ​​unchanged;

[0129] The pooling layer is used to aggregate the two-dimensional graph into a higher dimension to reduce the spatial size of the feature map and remove redundant spatial information.

[0130] Step 4.2: Build a deep learning network through a two-layer convolutional neural network architecture, as follows:

[0131] The deep learning network is constructed by using input layer, two-layer convolutional neural network architecture, Flatten layer, fully connected layer, softmax layer, and classification layer, including input layer-two-layer convolutional neural network architecture-Flatten layer-fully connected layer-softmax layer-classification layer set in sequence;

[0132] In the first layer of the convolutional neural network architecture, the size of both convolution kernels is 2×2;

[0133] In the second-layer convolutional neural network architecture, the sizes of the two convolution kernels are 2×6 and 4×2 respectively.

[0134] Step 4.3: Use the deep learning network to build a convolutional neural network prediction algorithm model, such as Figure 3 As shown, the details are as follows:

[0135] The convolutional neural network prediction algorithm model includes the following arranged in sequence: input layer-first convolutional neural network architecture-second convolutional neural network architecture-Flatten layer-first fully connected layer-fifth ReLU layer-second fully connected layer-sixth ReLU layer-third fully connected layer-softmax layer-classification layer;

[0136] The first convolutional neural network architecture includes the following: first convolutional layer - first ReLU layer - second convolutional layer - second ReLU layer - first maximum pooling layer.

[0137] The second convolutional neural network architecture includes the following arranged in sequence: the third convolutional layer - the third ReLU layer - the fourth convolutional layer - the fourth ReLU layer - the second maximum pooling layer;

[0138] In the first convolutional neural network architecture, the convolution layers are all set to traversal sampling with padding; the first convolution layer contains 64 convolution kernels of size 2×2; the second convolution layer contains 128 convolution kernels of size 2×2; the pooling window of the first maximum pooling layer is 1×2;

[0139] In the second convolutional neural network architecture, the convolutional layers are all set to traversal sampling without padding; the third convolutional layer contains 64 convolution kernels of size 2×6; the fourth convolutional layer contains 16 convolution kernels of size 4×2, and the pooling window of the maximum pooling layer is 1×4;

[0140] The first layer of the convolutional neural network architecture is used to increase the dimension of the feature map and reduce the length of the time component in the feature map to half of the original length; the second layer of the convolutional neural network architecture has a larger receptive field and can reduce the length of the time and translation components at the same time;

[0141] The first fully connected layer includes 64 neurons, the second fully connected layer includes 16 neurons, and the third fully connected layer includes 8 neurons; the softmax layer uses the softmax activation function to normalize the output of the fully connected layer and express it as the probability of the classification layer; the classification layer returns the corresponding classification probability for each input and calculates the loss;

[0142] The convolutional neural network prediction algorithm model uses the Adam optimizer, and the learning rate is set to 0.0001. After a large number of experimental tests, the classification accuracy of the above parameter settings is optimal.

[0143] Step 5: Using the neural network’s ability to extract signal features, the processed feature map is used as input to train the convolutional neural network prediction algorithm model, as follows:

[0144] Step 5.1, load the generated two-dimensional feature map dataset;

[0145] Step 5.2, the generated two-dimensional feature atlas is divided into an image data training set, a validation set and a test set after data splitting;

[0146] Step 5.3: Use the image data training set and validation set to train the constructed convolutional neural network prediction algorithm model.

[0147] Step 6: Use the trained convolutional neural network prediction algorithm model, take the generated image data test set as input, and classify the modulation mode of the signal after sufficient iterative training. The convergence speed of the convolutional neural network prediction algorithm model is as follows: Figure 4 As shown. Figure 5 As shown, the convolutional neural network prediction algorithm model provided in this embodiment has a recognition accuracy of 98.2% in a signal-to-noise ratio of -10dB to 20dB.

[0148] like Figure 6 As shown in the figure, the confusion matrix is ​​used to observe the classification accuracy of each modulation. When SNR = -10dB, 4FSK and 4PSK still maintain a recognition rate of more than 90%, due to their better distinction in the scale component of the feature map. In the case of low signal-to-noise ratio, the model as a whole can still show high recognition performance.

[0149] Therefore, the present invention adopts the above-mentioned modulation signal recognition method combining wavelet threshold denoising, median filtering, continuous wavelet transform and deep learning, uses wavelet threshold denoising and median filtering to process the sampled signal, obtains a two-dimensional feature map through continuous wavelet transform, extracts feature information in the time domain and frequency domain, and inputs it into the deep learning prediction algorithm model for training and testing, thereby improving the accuracy of modulation signal recognition.

[0150] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A modulation signal recognition method based on continuous wavelet transform and deep learning, characterized in that: The following steps are involved: Step 1: modulate the sequence to be transmitted at the transmitting end, obtain the wireless continuous time signal at the receiving end, and sample the continuous time signal; Step 2: Perform wavelet decomposition on the received signal, select a preset threshold rule to filter out high-frequency noise, reconstruct the obtained wavelet coefficients using inverse wavelet transform, and perform median filtering on the reconstructed signal to obtain a new data set; Step 3: Perform continuous wavelet transform on the new data set obtained in step 2 to obtain a time-frequency graph, downsample the effective frequency components of the time-frequency graph, extract feature information in the time domain and frequency domain, and obtain a two-dimensional feature graph; Step 4: Construct a convolutional neural network prediction algorithm model; Step 5: Using the neural network’s ability to extract signal features, the processed feature map is used as input to train the convolutional neural network prediction algorithm model; Step 6: Use the trained convolutional neural network prediction algorithm model and the generated image data test set as input to classify the modulation mode of the signal.

2. The modulation signal recognition method based on continuous wavelet transform and deep learning according to claim 1 is characterized in that: In step 1, the sequence to be transmitted is modulated at the transmitting end, a wireless continuous time signal is obtained at the receiving end, and the continuous time signal is sampled, as follows: Step 1.1: Modulate the sequence to be transmitted at the transmitting end, and let x(t) be the transmitted signal, as shown below: x(t)=F(s i ) Among them, s i is the i-th code element received, t is the time, and F is the modulation function; Step 1.2: The signal received by the receiving end is r(t), as shown below: r(t)=h(t)*x(t)+n(t) Where h(t) is the channel response, n(t) is the additive noise, and * is the convolution operation; Given a received signal r(t), the purpose of modulation signal recognition is to extract the frequency, phase and amplitude information of different modulation signals from r(t), so as to identify the modulation type of F; Step 1.3, sample the continuous-time signal as follows: Among them, r[n] is the length of the discrete time series, T s is the sampling interval; during the sampling process, the sampling rate should be higher than the Nyquist sampling rate.

3. The modulation signal recognition method based on continuous wavelet transform and deep learning according to claim 1 is characterized in that: In step 2, the received signal is decomposed by wavelet, a pre-set threshold rule is selected to filter out high-frequency noise, the obtained wavelet coefficients are reconstructed by inverse wavelet transform, and the reconstructed signal is median filtered to obtain a new data set, as follows: Step 2.1: Perform wavelet decomposition on the received signal, as shown below: Among them, j is the number of layers of wavelet decomposition, q is the filter length, and a j (k) is the low-frequency wavelet coefficient after decomposition, b j (k) is the high-frequency wavelet coefficient after decomposition, h d (n) represents the low-pass wavelet filter coefficient, g d (kn) represents the high-pass wavelet filter coefficient, Indicates a j-1 Downsampling by a factor of 2 (n); Step 2.2, for the low-frequency and high-frequency wavelet coefficients generated by wavelet decomposition, select the Huarsure threshold rule to filter out the high-frequency noise; the Huarsure threshold rule is composed of the Sqtwolog and Rigrsure threshold rules, and the smaller value of the two is selected as the threshold, and the high-frequency wavelet coefficients are threshold processed to obtain the wavelet coefficients with the high-frequency noise filtered out; Step 2.3: Reconstruct the wavelet coefficients after filtering out high-frequency noise using inverse wavelet transform, as shown below: in, is the reconstructed low-frequency wavelet coefficient, b j (k) is the reconstructed high-frequency wavelet coefficient, h r (n) and g r (n) represent the reconstruction filter coefficients of low frequency and high frequency, respectively, and Express and 2 times upsampling; Step 2.4: Filter the signal using a one-dimensional median filter with a window of 9, as shown below: in, is the denoised discrete time series, is the sequence after wavelet reconstruction, and the Mid function represents the median of the reconstructed sequence with a window of 9.

4. The modulation signal recognition method based on continuous wavelet transform and deep learning according to claim 1 is characterized in that: In step 3, the new data set obtained in step 2 is subjected to continuous wavelet transform to obtain a time-frequency graph, and the effective frequency components of the time-frequency graph are downsampled to extract feature information in the time domain and frequency domain to obtain a two-dimensional feature graph, as follows: Step 3.1: Discrete time series Perform a continuous wavelet transform as follows: Among them, W(c m ,d k ) is the time-frequency diagram obtained after continuous wavelet transform, c m is the scale, d k is the translation, ψ(n) is the mother wavelet basis function; the translation refers to moving the wavelets of different scales from the starting position of the signal to the ending position; the scale refers to controlling the expansion and contraction of the wavelet function, indicating the scaling of the signal in time; The mother wavelet basis function is the Morlet function, as shown below: Among them, ω0 is the center frequency of the wavelet, and the range of ω0 is set to 0 to f by setting the appropriate scale. s / 2; the translation amount increases in intervals of 1; Step 3.2: Move the Morlet wavelet functions of different scales and times along the entire signal to generate a coefficient matrix containing scale and translation information; select the frequency number with the largest value on the scale component as the center frequency of the two-dimensional feature map, and downsample the time component to complete the tracking of the effective frequency component.

5. The modulation signal recognition method based on continuous wavelet transform and deep learning according to claim 1 is characterized in that: In step 4, a convolutional neural network prediction algorithm model is constructed as follows: Step 4.1, build a convolutional neural network architecture through convolutional layer, ReLU layer, and maximum pooling layer; Step 4.2: Build a deep learning network through a two-layer convolutional neural network architecture; Step 4.3: Build a convolutional neural network prediction algorithm model through a deep learning network.

6. The modulation signal recognition method based on continuous wavelet transform and deep learning according to claim 5 is characterized in that: In step 4.1, the convolutional neural network architecture is constructed through the convolutional layer, ReLU layer, and maximum pooling layer, as follows: Convolutional neural network architecture is constructed through convolutional layer, ReLU layer, and maximum pooling layer, including convolutional layer-ReLU layer-convolutional layer-ReLU layer-maximum pooling layer set in sequence; The convolutional layer learns different features of the two-dimensional image input by the input layer; The ReLU layer uses an activation function to remove negative values ​​in the convolution result and keep the positive values ​​unchanged; The pooling layer is used to aggregate the two-dimensional graph into a higher dimension to reduce the spatial size of the feature map and remove redundant spatial information.

7. The modulation signal recognition method based on continuous wavelet transform and deep learning according to claim 5 is characterized in that: In step 4.2, a deep learning network is constructed through a two-layer convolutional neural network architecture, as follows: The deep learning network is constructed by using input layer, two-layer convolutional neural network architecture, Flatten layer, fully connected layer, softmax layer, and classification layer, including input layer-two-layer convolutional neural network architecture-Flatten layer-fully connected layer-softmax layer-classification layer set in sequence; In the first layer of the convolutional neural network architecture, the size of both convolution kernels is 2×2; In the second-layer convolutional neural network architecture, the sizes of the two convolution kernels are 2×6 and 4×2 respectively.

8. The modulation signal recognition method based on continuous wavelet transform and deep learning according to claim 5 is characterized in that: In step 4.3, a convolutional neural network prediction algorithm model is constructed through a deep learning network, as follows: The convolutional neural network prediction algorithm model includes the following arranged in sequence: input layer-first convolutional neural network architecture-second convolutional neural network architecture-Flatten layer-first fully connected layer-fifth ReLU layer-second fully connected layer-sixth ReLU layer-third fully connected layer-softmax layer-classification layer; The first convolutional neural network architecture includes the following: first convolutional layer - first ReLU layer - second convolutional layer - second ReLU layer - first maximum pooling layer. The second convolutional neural network architecture includes the following arranged in sequence: the third convolutional layer - the third ReLU layer - the fourth convolutional layer - the fourth ReLU layer - the second maximum pooling layer; In the first convolutional neural network architecture, the convolution layers are all set to traversal sampling with padding; the first convolution layer contains 64 convolution kernels of size 2×2; the second convolution layer contains 128 convolution kernels of size 2×2; the pooling window of the first maximum pooling layer is 1×2; In the second convolutional neural network architecture, the convolutional layers are all set to traversal sampling without padding; the third convolutional layer contains 64 convolution kernels of size 2×6; the fourth convolutional layer contains 16 convolution kernels of size 4×2, and the pooling window of the maximum pooling layer is 1×4; The first layer of the convolutional neural network architecture is used to increase the dimension of the feature map and reduce the length of the time component in the feature map to half of the original length; the second layer of the convolutional neural network architecture has a larger receptive field and can reduce the length of the time and translation components at the same time; The first fully connected layer includes 64 neurons, the second fully connected layer includes 16 neurons, and the third fully connected layer includes 8 neurons; the softmax layer uses the softmax activation function to normalize the output of the third fully connected layer and express it as the probability of the classification layer; the classification layer returns the corresponding classification probability for each input and calculates the loss; The convolutional neural network prediction algorithm model uses the Adam optimizer, and the learning rate is set to 0.0001.

9. The modulation signal recognition method based on continuous wavelet transform and deep learning according to claim 1 is characterized in that: In step 5, the neural network’s ability to extract signal features is used to train the convolutional neural network prediction algorithm model using the processed feature map as input, as follows: Step 5.1, load the generated two-dimensional feature map dataset; Step 5.2, the generated two-dimensional feature atlas is divided into an image data training set, a validation set and a test set after data splitting; Step 5.3: Use the image data training set and validation set to train the constructed convolutional neural network prediction algorithm model.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the modulation signal recognition method based on continuous wavelet transform and deep learning as described in any one of claims 1 to 9.

Images