Frequency hopping signal detection method and system based on time-frequency compressed sampling deep learning

Abstract

The invention relates to a frequency hopping signal detection method and system based on time-frequency compressed sampling deep learning, belongs to the field of signal processing, and solves the problems of low detection probability and high calculation complexity of a frequency hopping signal under a low signal-to-noise ratio. Comprising the following steps: respectively acquiring a frequency hopping signal when the frequency hopping signal exists and noise data when the frequency hopping signal does not exist, and carrying out time-frequency characterization and compressed sampling processing to obtain a frequency hopping time-frequency graph and a noise time-frequency graph; constructing a sample data set based on the frequency hopping time-frequency graph, the noise time-frequency graph and corresponding sample labels; constructing a frequency hopping signal detection network based on a lightweight convolutional neural network ShuffleNet and a convolutional attention module CBAM, and obtaining a trained frequency hopping signal detection network by using the sample data set; acquiring a time-frequency graph corresponding to a to-be-detected signal, and inputting the time-frequency graph into the trained frequency hopping signal detection network to obtain a confidence coefficient of the to-be-detected signal; and performing judgment based on the confidence of the to-be-detected signal to obtain a detection result of the to-be-detected signal. The detection performance is improved; and the reasoning complexity is reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of communication signal processing, and particularly relates to a frequency hopping signal detection method and system based on time-frequency compressed sampling deep learning. BACKGROUND

[0002] Due to the strong anti-interference performance of frequency hopping signals, frequency hopping communication technology is widely used in multiple service fields. The technology enhances the anti-interference ability and confidentiality of the communication system through rapid frequency hopping, and is therefore widely used in modern wireless communication systems. However, due to the fast frequency change and randomness of frequency hopping signals, it is difficult to achieve ideal detection effect in a complex electromagnetic environment. In addition, frequency hopping signals usually require a wide operating bandwidth, resulting in high computational complexity in the detection process. Therefore, how to efficiently, accurately and low-complexity detect frequency hopping signals has become a research hotspot.

[0003] Traditional frequency hopping signal detection algorithms include non-blind detection and blind detection, depending on whether prior knowledge is needed. Non-blind detection uses known signal characteristics and parameter information, can effectively detect frequency hopping signals, and common methods include channelized reception detection, autocorrelation detection and signal decomposition. Blind detection does not require prior information and still shows good detection performance in the case of unknown signal characteristics, and common methods include power spectrum detection and time-frequency detection. Traditional detection methods have certain limitations, need to manually design detection statistics, depend on theoretical knowledge and expert experience, and are highly subjective, so there is an urgent need for more intelligent and effective signal detection methods.

[0004] Deep learning has been widely applied in the field of communication signal processing due to its strong feature learning ability. In the aspect of frequency hopping signal detection, existing methods based on deep learning mainly rely on time-frequency domain analysis and correlation detection technology, but still face high computational complexity and poor real-time performance when processing large-scale data. In the past decade, the theory of compressed sensing has been proposed and continuously developed. Inspired by the theory of compressed sensing, some researchers have proposed frequency hopping signal detection methods based on compressed measurement signals. However, these detection methods combined with compressed sensing still have limitations such as dependence on prior information, need for manual feature design and dependence on specific tasks. Therefore, reducing computational complexity while improving frequency hopping signal detection performance has become a key problem that needs to be solved in the field of frequency hopping signal detection. SUMMARY

[0005] Based on the above analysis, the present application aims to provide a frequency hopping signal detection method and system based on time-frequency compressed sampling deep learning, to solve the technical problems of low detection probability and high computational complexity of frequency hopping signals in existing methods under low signal-to-noise ratio.

[0006] The main purpose of the present application is achieved by the following technical solutions:

[0007] The application provides a frequency hopping signal detection method based on time-frequency compressed sampling deep learning, comprising the following steps:

[0008] Frequency hopping signal data when frequency hopping signals exist and noise data when frequency hopping signals do not exist are collected respectively, a plurality of signal segments are obtained by cutting according to a preset length, time-frequency representation and compressed sampling processing are performed on each signal segment, and frequency hopping time-frequency diagrams and noise time-frequency diagrams are obtained;

[0009] A sample data set is constructed based on the frequency hopping time-frequency diagrams and the noise time-frequency diagrams and corresponding sample labels; wherein the sample labels include signals and noise;

[0010] A frequency hopping signal detection network is constructed based on a light-weight convolutional neural network ShuffleNet and a convolutional attention module CBAM, the sample data set is used, and a trained frequency hopping signal detection network is obtained;

[0011] Time-frequency diagrams corresponding to a plurality of signal segments of a to-be-detected signal are obtained, the trained frequency hopping signal detection network is input, the confidence of each signal segment of the to-be-detected signal is obtained, and a detection result of the to-be-detected signal is obtained based on the confidence of each signal segment of the to-be-detected signal.

[0012] Further, the frequency hopping time-frequency diagrams and the noise time-frequency diagrams are obtained, comprising:

[0013] Short-time Fourier transform is performed on each signal segment of the frequency hopping signal data and the noise data respectively to obtain time-frequency representation, and first frequency hopping time-frequency diagrams and first noise time-frequency diagrams corresponding to the signal segments are obtained;

[0014] Compressed sampling processing is performed on the first frequency hopping time-frequency diagrams and the first noise time-frequency diagrams respectively, and frequency hopping time-frequency diagrams and noise time-frequency diagrams are obtained.

[0015] Further, the compressed sampling processing comprises the following steps:

[0016] Two random Gaussian matrices with a mean of 0 and a variance of 1 are generated, then the two random Gaussian matrices are subjected to Gram-Schmidt orthogonalization to eliminate correlation, one is used as a real part, and the other is used as an imaginary part, and an orthogonal complex Gaussian measurement matrix Φ is constructed;

[0017] Based on the orthogonal complex Gaussian measurement matrix Φ, the first frequency hopping time-frequency diagrams and the first noise time-frequency diagrams are measured respectively, and measurement results s = ΦxΦ are obtained T , wherein x represents the first frequency hopping time-frequency diagram or the first noise time-frequency diagram, and s is the corresponding frequency hopping time-frequency diagram or noise time-frequency diagram.

[0018] Further, the frequency hopping signal detection network comprises, in sequence, an input layer, a convolutional layer, a maximum pooling layer, a first UnitB, a first UnitA, a second UnitA, a third UnitA, a second UnitB, a fourth UnitA, a CBAM module, a global pooling layer, a fully connected layer, a Softmax layer, and an output layer.

[0019] The input layer is configured to receive a time-frequency graph.

[0020] The convolutional layer is configured to perform preliminary feature extraction on the input time-frequency graph, capture local features through convolutional operation, and obtain a first time-frequency feature map.

[0021] The maximum pooling layer is configured to reduce the dimensionality of the first time-frequency feature map, retain important features, and reduce the computational load, and obtain a second time-frequency feature map.

[0022] The first UnitB, the first UnitA, the second UnitA, the third UnitA, the second UnitB, and the fourth UnitA are configured to perform multi-scale feature extraction and effective exchange of inter-channel information on the second time-frequency feature map through ShuffleNet unit modules with different configurations, and obtain a third time-frequency feature map.

[0023] The CBAM module is configured to adaptively adjust the channel weight and the spatial weight of the third time-frequency feature map, and obtain a fourth time-frequency feature map.

[0024] The global pooling layer is configured to perform global pooling on the fourth time-frequency feature map, further reduce the dimensionality, and obtain a fifth time-frequency feature map.

[0025] The fully connected layer is configured to map the fifth time-frequency feature map to a category space.

[0026] The Softmax layer is configured to perform normalization processing on the output of the fully connected layer, convert it into a probability distribution, and represent the confidence of the input time-frequency graph belonging to each category label.

[0027] The output layer is configured to make a final classification decision based on the confidence of the input time-frequency graph belonging to each category label, and output a detection result of whether a frequency hopping signal exists.

[0028] Further, the frequency hopping signal detection network is trained using a training set in the sample data set; the training process of the frequency hopping signal detection network is as follows:

[0029] An Adam optimizer based on stochastic gradient descent is used to iteratively update the model parameters and minimize the loss function.

[0030] If the accuracy does not improve in p training cycles, the learning rate in the next training cycle is reduced to q times of the current training cycle;

[0031] When the maximum training round is reached, the training is stopped, the network parameters with the minimum loss function on the validation set during the training are saved as the trained frequency hopping signal detection network; Wherein, p is greater than or equal to 3.

[0032] Further, the confidence of the signal to be detected is based on the decision to obtain the detection result of each signal segment of the signal to be detected, as follows:

[0033]

[0034] Wherein, t signal The confidence of the signal segment to be detected, γ is the decision threshold, H1 indicates that the frequency hopping signal exists, and H0 indicates that the frequency hopping signal does not exist.

[0035] Further, the decision threshold is set by the following method:

[0036] Obtain N pure noise data for time-frequency representation and compressed sampling processing, obtain the corresponding time-frequency graph, input the trained frequency hopping signal detection network, and output the confidence score;

[0037] Then the confidence score is obtained in descending order to obtain the confidence score sequence.

[0038] Based on the preset false alarm probability, the confidence score value on the corresponding quantile point is determined as the final decision threshold.

[0039] Further, if the detection result of each signal segment of the signal to be detected is noise, the signal to be detected is noise without frequency hopping signal, otherwise it is a frequency hopping signal.

[0040] Further, the cross entropy loss function is used in the training process of the frequency hopping signal detection network.

[0041] The present application provides a kind of frequency hopping signal detection system based on time-frequency compressed sampling deep learning, including signal acquisition module, data set construction module, network construction and training module and signal detection module;

[0042] The signal acquisition module is used to collect frequency hopping signal data when frequency hopping signal exists and noise data when frequency hopping signal does not exist respectively, cut to obtain multiple signal segments according to preset length, time-frequency representation and compressed sampling processing are carried out on each signal segment, and frequency hopping time-frequency graph and noise time-frequency graph are obtained.

[0043] The data set construction module is configured to construct a sample data set based on the frequency hopping time-frequency diagram and the noise time-frequency diagram and corresponding sample labels, wherein the sample labels include signals and noise.

[0044] The network construction and training module is configured to construct a frequency hopping signal detection network based on a lightweight convolutional neural network ShuffleNet and a convolutional attention module CBAM, and obtain a trained frequency hopping signal detection network by using the sample data set.

[0045] The signal detection module is configured to obtain time-frequency diagrams corresponding to a plurality of signal segments of a to-be-detected signal, input the trained frequency hopping signal detection network, and obtain confidence degrees of each signal segment of the to-be-detected signal; and make a decision based on the confidence degrees of each signal segment of the to-be-detected signal to obtain a detection result of the to-be-detected signal.

[0046] Compared with the prior art, the present application can achieve the following beneficial effects:

[0047] 1. According to the sparsity of the frequency hopping signal in the time-frequency domain, the present application realizes efficient detection of the frequency hopping signal while reducing the computational complexity by combining the compression downsampling and the lightweight network technology.

[0048] 2. The present application designs an orthogonal complex Gaussian measurement matrix, applies two-dimensional compression sampling and semi-tensor product technology to reduce the dimension of the time-frequency diagram, and reduces the consumption of computing resources.

[0049] 3. The present application designs a frequency hopping signal detection network based on a lightweight neural network and a CBAM attention mechanism, which is beneficial to improve the detection performance.

[0050] Other features and advantages of the present application will be set forth in the specification, and in part will become apparent from the specification, or can be learned by practice of the present application. The objects and other advantages of the present application will be realized and achieved by the specification and the particularly pointed out in the appended claims. BRIEF DESCRIPTION OF DRAWINGS

[0051] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application.

[0052] Figure 1 A frequency hopping signal detection method flowchart based on compression sampling and deep learning in an embodiment of the present application.

[0053] Figure 2 A frequency hopping signal detection network structure schematic diagram in an embodiment of the present application.

[0054] Figure 3 A ShuffleNet unit module structure schematic diagram in an embodiment of the present application.

[0055] Figure 4 A time-frequency compression sampling deep learning-based frequency hopping signal detection system module schematic diagram is provided for the application in the embodiments of the application. DETAILED DESCRIPTION

[0056] The detailed description of the various exemplary embodiments of the application should not be considered to be limiting of the application, but rather a description of certain aspects, features, and embodiments of the application.

[0057] Embodiment one:

[0058] As shown in one specific embodiment of the application, Figure 1 a time-frequency compression sampling deep learning-based frequency hopping signal detection method is disclosed, comprising the following steps:

[0059] Step S1, frequency hopping signal data when frequency hopping signals exist and noise data when frequency hopping signals do not exist are respectively collected, a plurality of signal segments are obtained by cutting according to a preset length, time-frequency representation and compression sampling processing are performed on each signal segment, and frequency hopping time-frequency diagrams and noise time-frequency diagrams are obtained;

[0060] Step S2, a sample data set is constructed based on the frequency hopping time-frequency diagrams and the noise time-frequency diagrams and corresponding sample labels; wherein the sample labels include signals and noise;

[0061] Step S3, a frequency hopping signal detection network is constructed based on a lightweight convolutional neural network ShuffleNet and a convolutional attention module CBAM, and a trained frequency hopping signal detection network is obtained by using the sample data set;

[0062] Step S4, time-frequency diagrams corresponding to a plurality of signal segments of a signal to be detected are obtained, the trained frequency hopping signal detection network is input, and the confidence of each signal segment of the signal to be detected is obtained; a detection result of the signal to be detected is obtained based on the confidence of each signal segment of the signal to be detected.

[0063] Step S1 includes steps S11-S12.

[0064] Step S11, frequency hopping signal data when frequency hopping signals exist and noise data when frequency hopping signals do not exist are collected.

[0065] Signal samples when frequency hopping signals exist and noise samples when frequency hopping signals do not exist are respectively collected, and both are discrete digital signal sequences, and have equal lengths.

[0066] The frequency hopping bandwidth is set to 6MHz, the frequency set size is 256, and the frequency hopping rate is 1000 hops / s. During signal acquisition, sample data is obtained in two cases:

[0067] (1) the frequency hopping signal data collected when the frequency hopping signal exists;

[0068] (2) noise data collected when the frequency hopping signal does not exist.

[0069] Exemplarily, the collected frequency hopping signal data and noise data are cropped into a plurality of fixed-length signal segments, and the length of the signal sequence is 8192 sampling points. In actual application, the cropping length can be adjusted according to specific requirements. Finally, the signal sequence corresponding to the time-frequency diagram after cropping is used for frequency hopping signal detection network training.

[0070] Step S12, time-frequency representation and compressed sampling processing are performed on each signal segment of the collected frequency hopping signal data and noise data to obtain a frequency hopping time-frequency diagram and a noise time-frequency diagram.

[0071] The frequency hopping time-frequency diagram and the noise time-frequency diagram are obtained, including:

[0072] The short-time Fourier transform is performed on each signal segment of the frequency hopping signal data and the noise data respectively to obtain a first frequency hopping time-frequency diagram and a first noise time-frequency diagram.

[0073] The compressed sampling processing is performed on the first frequency hopping time-frequency diagram and the first noise time-frequency diagram respectively to obtain a frequency hopping time-frequency diagram and a noise time-frequency diagram.

[0074] The short-time Fourier transform is performed on each signal segment of the frequency hopping signal data and the noise data respectively, and the TFRSTFT (Time-Frequency Representation based Short-Time Fourier Transform) algorithm is used to discretize the STFT (Short-Time Fourier Transform) and is represented as follows:

[0075]

[0076] Wherein, TFR(m, k) is the short-time Fourier transform result based on time-frequency representation; m is the window analysis time index, representing the current analysis time (center position of the window); k is the frequency domain index, representing the kth frequency component in the Fourier transform; n is the signal sequence index, representing the value at time point n; w * (n-m) is a smoothing window function with a finite length; L is the smoothing window function w *(n-m) is the length of the window; x(n) is a discrete signal sequence of frequency hopping signal data or noise data;

[0077] For example, the window function is selected as a Hanning window or a Gaussian window.

[0078] w(n) is a smoothing window function limited to m=0 and k=0, to ensure that the analysis is focused on the signal near the current time point, which helps to more accurately represent the instantaneous power content of the signal; w * (n-m) is the conjugate of the window function centered at m; e -j2πkn / L is the Fourier transform kernel, used to convert the time domain signal into a frequency domain representation; the time-frequency diagram is obtained by the TFR STFT algorithm, with dimensions [N, N];

[0079] After performing short-time Fourier transform on each signal segment of the frequency hopping signal data and the noise data, the first frequency hopping time-frequency diagram and the first noise time-frequency diagram corresponding to the signal segment are obtained.

[0080] Then, the first frequency hopping time-frequency diagram and the first noise time-frequency diagram are subjected to compression sampling processing.

[0081] The compression sampling processing includes the following steps:

[0082] Two random Gaussian matrices with mean 0 and variance 1 are generated, and then the two random Gaussian matrices are subjected to Gram-Schmidt orthogonalization to eliminate correlation, one is used as the real part and the other is used as the imaginary part, to construct an orthogonal complex Gaussian measurement matrix Φ;

[0083] Based on the orthogonal complex Gaussian measurement matrix Φ, the first frequency hopping time-frequency diagram and the first noise time-frequency diagram are measured respectively, to obtain measurement results s=ΦxΦ T , wherein x represents the first frequency hopping time-frequency diagram or the first noise time-frequency diagram, and s is the corresponding frequency hopping time-frequency diagram or noise time-frequency diagram.

[0084] Designing an orthogonal complex Gaussian measurement matrix: first, generate two random Gaussian matrices A and B with mean 0 and variance 1, represented as

[0085]

[0086] Gram-Schmidt orthogonalization is performed on the two matrices respectively, to obtain Q A and Q B are the orthogonalization results of A and B, respectively, to eliminate the correlation between them.

[0087] Using the random Gaussian matrices A and B, after orthogonalization, a Gaussian measurement matrix Φ is composed.

[0088] Q AQ B is the real part of the complex matrix Q

[0089] Φ = Q A + jQ B Equation (3)

[0090] where j represents the imaginary unit.

[0091] The two-dimensional compressed sensing is used for measurement, and a compressed measurement result y is obtained, which is the down-sampled time-frequency pattern sample data, as follows:

[0092] s = ΦxΦ T Equation (4)

[0093] where s is a frequency hopping time-frequency pattern or a noise time-frequency pattern, Φ, x, and Φ T are a Gaussian measurement matrix, x is a first frequency hopping time-frequency pattern or a first noise time-frequency pattern to be sampled, and Φ T is a transpose matrix of the Gaussian measurement matrix.

[0094] Through the double-matrix operation of Φ and Φ T , efficient compressed sampling of the first frequency hopping time-frequency pattern or the first noise time-frequency pattern is realized.

[0095] The product of the Gaussian measurement matrix and the first frequency hopping time-frequency pattern matrix or the first noise time-frequency pattern matrix uses a semi-tensor product, allowing two-dimensional mismatched matrices to perform multiplication operations, and can realize effective compressed sampling of the time-frequency pattern.

[0096] The function of step S1 is to collect frequency hopping signal data and noise data from the actual environment, and convert them into frequency hopping time-frequency pattern and noise time-frequency pattern samples that can be used for deep learning model training through time-frequency characterization analysis and compressed sampling processing.

[0097] Step S2, in particular.

[0098] All the frequency hopping time-frequency pattern and noise time-frequency pattern samples are randomly divided into 2 sets according to a certain proportion, as a training set and a validation set for lightweight network training.

[0099] Exemplarily, the training set and the validation set are divided according to a proportion of 8:2 or 7:3, and in actual application, the proportion can be changed according to specific needs.

[0100] The sample representation of the training set and the validation set is as follows:

[0101]

[0102] wherein, represents the i-th frequency hopping time-frequency pattern or noise time-frequency pattern sample, represents the i-th frequency hopping time-frequency pattern or noise time-frequency pattern sample, Corresponding real sample label; z is the total number of samples.

[0103] Each sample data in the training set and the validation set includes a frequency hopping time-frequency graph and a noise time-frequency graph sample under different signal-to-noise ratios and a corresponding sample label; the sample label is determined according to whether a frequency hopping signal exists in the corresponding sample, if the frequency hopping signal exists, the label is "signal", otherwise, the label is "noise".

[0104] The step S2 is to randomly divide the frequency hopping time-frequency graph and the noise time-frequency graph sample into the training set and the validation set, which are used for training and verifying the frequency hopping signal detection network in the step S32.

[0105] The step S3 includes the steps S31-S32.

[0106] The step S31 is to construct the frequency hopping signal detection network based on a lightweight convolutional neural network ShuffleNet and a convolutional attention module CBAM.

[0107] The frequency hopping signal detection network sequentially includes an input layer, a convolutional layer, a maximum pooling layer, a first UnitB, a first UnitA, a second UnitA, a third UnitA, a second UnitB, a fourth UnitA, a CBAM module, a global pooling layer, a full connection layer, a Softmax layer and an output layer.

[0108] The input layer is used to receive a time-frequency graph.

[0109] The convolutional layer is used to preliminarily extract features of the input time-frequency graph, capture local features through a convolutional operation, and obtain a first time-frequency feature map.

[0110] The maximum pooling layer is used to reduce the dimension of the first time-frequency feature map, retain important features and reduce the calculation amount, and obtain a second time-frequency feature map.

[0111] The first UnitB, the first UnitA, the second UnitA, the third UnitA, the second UnitB and the fourth UnitA are used to realize multi-scale feature extraction and effective exchange of information between channels of the second time-frequency feature map through ShuffleNet unit modules with different configurations, and obtain a third time-frequency feature map.

[0112] The CBAM module is used to adaptively adjust the channel weight and the spatial weight of the third time-frequency feature map, and obtain a fourth time-frequency feature map.

[0113] The global pooling layer is used to globally pool the fourth time-frequency feature map, further reduce the dimension, and obtain a fifth time-frequency feature map.

[0114] The full connection layer is configured to map the fifth time-frequency feature map to a category space.

[0115] The Softmax layer is configured to perform normalization processing on the output of the full connection layer, convert the output into a probability distribution, and represent the confidence of the input time-frequency map belonging to each category label.

[0116] The output layer is configured to perform a final classification decision based on the confidence of the input time-frequency map belonging to each category label, and output a detection result of whether a frequency hopping signal exists.

[0117] A light network for frequency hopping signal detection is constructed, and a network structure diagram of the frequency hopping signal detection network is as shown in Figure 2 .

[0118] The frequency hopping signal detection network mainly consists of three parts:

[0119] (1) The feature extraction part is composed of a convolution layer, a maximum pooling layer, a first UnitB, a first UnitA, a second UnitA, a third UnitA, a second UnitB, and a fourth UnitA, and is configured to extract high-dimensional features of a time-frequency map;

[0120] (2) The attention part is composed of a CBAM (Convolutional Block Attention Module) module. The CBAM is a mechanism module that combines channel attention and spatial attention, and can adaptively adjust the channel weight and spatial weight of a feature map, thereby highlighting key information in the input features and suppressing redundant or invalid features;

[0121] (3) The confidence mapping part is composed of a global pooling layer, a full connection layer, a Softmax layer, and an output layer, and is configured to convert the extracted feature information into the confidence of a corresponding label.

[0122] As shown in Figure 2 , "3x3Conv" is a convolution layer, which is a traditional convolution layer with a convolution kernel size of 3x3;

[0123] "Maxpool" represents a maximum pooling layer;

[0124] Then, the first UnitB (ShuffleNet UnitB), the first UnitA (ShuffleNet UnitA), the second ShuffleNet UnitA (ShuffleNet UnitA), the third ShuffleNet UnitA (ShuffleNet UnitA), the second ShuffleNet UnitB (ShuffleNet UnitB), and the fourth ShuffleNet UnitA (ShuffleNet UnitA) are sequentially arranged.

[0125] ShuffleNetUnitA and ShuffleNetUnitB are two basic units in the lightweight network ShuffleNet network, which are based on residual blocks and channel rearrangement design, mainly composed of deep convolution, pointwise group convolution and pointwise group convolution. ShuffleNetUnitA is usually used in the case of step 1, while ShuffleNetUnitB is used in the case of step 2.

[0126] The first, second, third and fourth UnitA structures are the same, and the ShuffleNet UnitA network structure is shown in Fig. 1(a) and includes a first group convolution layer of UnitA, a batch processing and ReLU activation layer, a channel shuffle layer, a deep convolution layer, a first batch processing layer, a second group convolution layer, a skip connection from the input of the first group convolution to the output of the second batch processing, and a ReLU activation layer. Figure 3

[0127] (1) The "1x1 GConv" is the first group convolution layer, which is used to reduce the number of input feature channels, reduce the computational complexity and parameter quantity, and divide the input channels into multiple groups, each group performing convolution operation independently;

[0128] (2) "BN ReLU" is a batch processing and ReLU activation layer. Batch Normalization (Batch Normalization) is used to accelerate the training process and improve the stability of the model, by normalizing the input of each batch of data; ReLU (Rectified Linear Unit) activation function is used to introduce nonlinearity, so that the model can learn more complex features;

[0129] (3) "Channel Shuffle" is a channel shuffle layer that uniformly rearranges the channels after group ReLU activation, solving the problem of information isolation between groups;

[0130] (4) 3x3 DWConv deep convolution layer, which extracts spatial features from the channel shuffled feature map, using deep separable convolution to greatly reduce the parameter quantity;

[0131] (5) "BN" first batch processing layer, which is used; batch normalization is applied again to stabilize the training process and accelerate convergence;

[0132] (6) "1x1 GConv" second group convolution layer, which is used to restore the number of channels to provide sufficient feature dimensions for subsequent operations;

[0133] (7) "BN" second batch processing, which is used to batch normalize the output of the second group convolution, further stabilizing the training process; ​

[0134] (8) A skip connection and ReLU activation layer are used from the first group convolution input to the second batch output to add the input directly to the output, forming a residual connection, which helps alleviate the gradient vanishing problem and promotes information flow. The ReLU activation function introduces nonlinearity again, increasing the expressive power of the model.

[0135] The structure of ShuffleNet UnitA achieves effective feature extraction and fusion while maintaining computational efficiency through a combination of grouped convolution and depthwise convolution, as well as the design of channel shuffling and residual connection.

[0136] The first and second UnitB structures are the same, and the ShuffleNet UnitB network structure is as follows: Figure 3 (b) shows the network structure of UnitB, which includes the first group convolution layer, batch processing and ReLU activation layer, channel shuffling layer, depthwise convolution layer with stride 2, the first batch processing layer, the second group convolution layer, the second batch processing, and the average pooling layer with stride 2. The input of UnitB passes through the average pooling layer with stride 2, and is then connected to the output of the second batch processing layer and activated with ReLU.

[0137] UnitB's first grouped convolutional layer, batch processing and ReLU activation layer, channel shuffling layer, depthwise convolutional layer with a stride of 2, first batch processing layer, second grouped convolutional layer, second batch processing, and average pooling layer with a stride of 2 are consistent with the structure and function of UnitA;

[0138] After the input of UnitB passes through the average pooling layer with a stride of 2, the connection between it and the output of the second batch processing layer is concatenated in the channel dimension instead of directly adding them to achieve feature fusion.

[0139] like Figure 2 As shown in the figure, the CBAM (Convolutional Block Attention Module) module includes a cascaded channel attention submodule (Channel Attention) and a spatial attention submodule (SpatialAttention);

[0140] For the channel attention submodule:

[0141] (1) Avgpool and Maxpool, parallel average pooling and maximum pooling:

[0142] Global average pooling and global maximum pooling are performed on the input third time-frequency feature map respectively to obtain two different feature descriptions.

[0143] (2) Two 1x1 Convs in parallel, with a 1x1 convolution kernel:

[0144] The two pooling results of average pooling and max pooling are respectively passed through a 1x1 convolution layer to adjust the feature dimension.

[0145] (3) Addition (+): Add the outputs of two parallel 1x1 convolution layers.

[0146] (4) Sigmoid: Generate channel attention weights through the Sigmoid activation function, representing the importance of each channel.

[0147] (5) Multiplication (x): Multiply the channel attention weights with the third time-frequency feature map channel by channel, weight the third time-frequency feature map, highlight important channels, and suppress unimportant channels.

[0148] For the spatial attention sub-module:

[0149] (1) Avgpool and Maxpool: Average pooling and max pooling are performed on the feature maps output by the channel attention sub-module, respectively, to obtain two different feature descriptions.

[0150] (2) Concat: Concatenate the two pooling results of average pooling and max pooling in the channel dimension.

[0151] (3) 3x1 Conv, a convolution layer with a 3x1 convolution kernel, used to extract spatial features of the concatenated features.

[0152] (4) Sigmoid: Generate spatial attention weights through the Sigmoid activation function, with the weights representing the importance of spatial features at each spatial position.

[0153] (5) Multiplication (x): Multiply the spatial attention weights with the output feature map of the channel attention sub-module element by element, weight the output feature map of the channel attention sub-module, highlight important spatial positions, and suppress unimportant spatial positions.

[0154] The channel attention and spatial attention mechanisms are integrated into the frequency hopping signal detection network to enhance the ability of the frequency hopping signal detection network to capture important features and improve the performance of the frequency hopping signal detection network.

[0155] Channel attention focuses on the importance of different channels, while spatial attention focuses on the importance of different spatial positions. By combining these two attention mechanisms, the frequency hopping signal detection network can more comprehensively understand the features of the input data.

[0156] The fourth time-frequency feature map output by the CBAM module is sequentially subjected to a global pooling layer (Gapool), a fully connected layer (Fc), a Softamx layer and an output layer to obtain a detection result of whether the sample data contains a frequency hopping signal.

[0157] The fully connected layer (Fc) is a fully connected layer with two neurons, and the softmax layer outputs a confidence (0-1).

[0158] Step S32, using the sample data set, training the frequency hopping signal detection network to obtain a trained frequency hopping signal detection network.

[0159] The training set in the sample data set is used to train the frequency hopping signal detection network; the training process of the frequency hopping signal detection network is as follows:

[0160] The model parameters are iteratively updated using an Adam optimizer based on stochastic gradient descent, and the loss function is minimized.

[0161] If the accuracy does not improve within p training cycles, the learning rate in the next training cycle is reduced to q times the current training cycle;

[0162] When the maximum number of training rounds is reached, the training is stopped, the network parameters with the minimum loss function on the validation set during the training are saved as the trained frequency hopping signal detection network; wherein, p> 3.

[0163] If the verification performance does not improve after 3 iterations (here, the verification performance represents the detection performance of the frequency hopping signal detection network on the validation set, i.e. the accuracy, if the detection accuracy does not improve, the learning rate will be modified); for example, q is 0.1, and the learning rate in the next training cycle is reduced to 0.1 times the learning rate in the current training cycle.

[0164] For example, the maximum number of training rounds is set to 10, which can be adjusted according to specific needs in actual application.

[0165] The cross-entropy loss function is used in the training process of the frequency hopping signal detection network.

[0166] When the maximum number of rounds is reached, the training is completed, and a trained frequency hopping signal network is obtained, which is a lightweight network model.

[0167] The function of step S3 is to build and train a lightweight frequency hopping signal detection network containing ShuffleNet and CBAM module, optimize the network parameters using the sample data set, and obtain a trained frequency hopping signal detection network to realize efficient signal detection.

[0168] Step S4 includes steps S41-S42.

[0169] Step S41, obtain the time-frequency graph corresponding to the to-be-detected signal.

[0170] Collect the to-be-detected signal, and intercept the length of the to-be-detected signal to be the same as the length of the signal in step S11; obtain a plurality of signal segments;

[0171] Perform short-time Fourier transform processing and compressed sampling processing on the signal segments of the to-be-detected signal to obtain the corresponding time-frequency graph.

[0172] Step S42, input the time-frequency graphs of the plurality of signal segments of the to-be-detected signal obtained into the trained frequency hopping signal detection network to obtain the confidence of each signal segment of the to-be-detected signal; and make a decision based on the confidence of each signal segment of the to-be-detected signal to obtain the detection result of the to-be-detected signal.

[0173] The decision based on the confidence of the to-be-detected signal to obtain the detection result of the to-be-detected signal is as follows:

[0174]

[0175] Where, t signal is the confidence of the to-be-detected signal, γ is the decision threshold, H1 represents that the frequency hopping signal is detected to exist, and H0 represents that the frequency hopping signal does not exist.

[0176] If the detection result of each signal segment of the to-be-detected signal is noise, then the to-be-detected signal is noise and does not exist frequency hopping signal, otherwise it is a frequency hopping signal.

[0177] The decision threshold is set by the following method:

[0178] Obtain N pure noise data for time-frequency representation and compressed sampling processing to obtain the corresponding time-frequency graph, input the trained frequency hopping signal detection network, and output the confidence score;

[0179] Then, the confidence score is arranged in descending order to obtain a confidence score sequence;

[0180] Determine the confidence score value on the corresponding quantile point based on the preset false alarm probability as the final decision threshold.

[0181] The preset false alarm probability is the allowable misjudgment rate.

[0182] The function of step S4 is to input the time-frequency graph of the to-be-detected signal into the trained frequency hopping signal detection network, calculate the confidence thereof, and determine whether the frequency hopping signal exists in the signal according to the preset decision threshold, so as to obtain the detection result.

[0183] Embodiment two:

[0184] Another specific embodiment of the present application discloses a time-frequency compression sampling deep learning-based frequency hopping signal detection system, thereby realizing the time-frequency compression sampling deep learning-based frequency hopping signal detection method in embodiment one. The specific implementation of each module is described in the corresponding description in embodiment one.

[0185] As shown in Figure 4 , the system comprises a signal acquisition module M1, a data set construction module M2, a network construction M3 and a training module, and a signal detection module M4.

[0186] The signal acquisition module M1 is configured to acquire frequency hopping signal data when the frequency hopping signal exists and noise data when the frequency hopping signal does not exist, respectively, to obtain a plurality of signal segments by clipping according to a preset length, to perform time-frequency characterization and compression sampling processing on each signal segment, and to obtain frequency hopping time-frequency diagrams and noise time-frequency diagrams.

[0187] The data set construction module M2 is configured to construct a sample data set based on the frequency hopping time-frequency diagrams and the noise time-frequency diagrams and corresponding sample labels; wherein the sample labels comprise signals and noise.

[0188] The network construction and training module M3 is configured to construct a frequency hopping signal detection network based on a lightweight convolutional neural network ShuffleNet and a convolutional attention module CBAM, and to obtain a trained frequency hopping signal detection network by using the sample data set.

[0189] The signal detection module M4 is configured to obtain time-frequency diagrams corresponding to a plurality of signal segments of a signal to be detected, to input the trained frequency hopping signal detection network, to obtain the confidence of each signal segment of the signal to be detected, and to determine the detection result of the signal to be detected based on the confidence of each signal segment of the signal to be detected.

[0190] In summary, the time-frequency compression sampling deep learning-based frequency hopping signal detection method and system of the present application has the following beneficial effects:

[0191] 1. According to the sparsity of the frequency hopping signal in the time-frequency domain, the present application realizes efficient detection of the frequency hopping signal while reducing the computational complexity by combining compression downsampling and lightweight network technology.

[0192] 2. The present application designs an orthogonal complex Gaussian measurement matrix, applies two-dimensional compression sampling and semi-tensor product technology to reduce the dimension of the time-frequency diagram, and reduces the consumption of computing resources.

[0193] 3. The present application designs a frequency hopping signal detection network based on a lightweight neural network and a CBAM attention mechanism, which is conducive to improving the detection performance.

[0194] Those skilled in the art can understand that all or part of the processes of the above-mentioned embodiment methods can be completed by instructing the relevant hardware by a computer program, and the program can be stored in a computer readable storage medium. The computer readable storage medium is a disk, an optical disk, a read-only memory, a random access memory, etc.

[0195] The above description is merely preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A frequency hopping signal detection method based on time-frequency compression sampling deep learning, characterized in that: include: Collect frequency hopping signal data when the frequency hopping signal exists and noise data when the frequency hopping signal does not exist, cut them into multiple signal segments according to the preset length, perform time-frequency representation and compressive sampling processing on each signal segment, and obtain frequency hopping time-frequency graph and noise time-frequency graph; Constructing a sample data set based on the frequency hopping time-frequency graph and the noise time-frequency graph and corresponding sample labels; wherein the sample labels include signal and noise; A frequency hopping signal detection network is constructed based on a lightweight convolutional neural network ShuffleNet and a convolutional attention module CBAM, and a trained frequency hopping signal detection network is obtained using the sample dataset; Obtain time-frequency diagrams corresponding to multiple signal segments of the signal to be detected, input the trained frequency hopping signal detection network, and obtain the confidence of each signal segment of the signal to be detected; make a judgment based on the confidence of each signal segment of the signal to be detected, and obtain a detection result of the signal to be detected.

2. The method according to claim 1, characterized in that The obtaining of the frequency hopping time-frequency graph and the noise time-frequency graph includes: Performing short-time Fourier transform on each signal segment of the frequency-hopping signal data and the noise data to perform time-frequency characterization, and obtaining a first frequency-hopping time-frequency graph and a first noise time-frequency graph corresponding to the signal segment; Compressed sampling is performed on the first frequency-hopping time-frequency graph and the first noise time-frequency graph to obtain a frequency-hopping time-frequency graph and a noise time-frequency graph.

3. The method according to claim 2, characterized in that The compressed sampling process includes the following steps: Generate two random Gaussian matrices with a mean of 0 and a variance of 1, then perform Gram-Schmidt orthogonalization on these two random Gaussian matrices to eliminate correlation, use one as the real part and the other as the imaginary part to construct an orthogonal complex Gaussian measurement matrix Φ; Based on the orthogonal complex Gaussian measurement matrix Φ, the first frequency hopping time-frequency diagram and the first noise time-frequency diagram are measured respectively to obtain a measurement result s=ΦxΦ T , where x represents the first frequency hopping time-frequency diagram or the first noise time-frequency diagram, and s represents the corresponding frequency hopping time-frequency diagram or noise time-frequency diagram.

4. The method according to claim 1, characterized in that The frequency hopping signal detection network includes an input layer, a convolutional layer, a maximum pooling layer, a first UnitB, a first UnitA, a second UnitA, a third UnitA, a second UnitB, a fourth UnitA, a CBAM module, a global pooling layer, a fully connected layer, a Softmax layer and an output layer in sequence; The input layer is used to receive the time-frequency graph; The convolution layer is used to perform preliminary feature extraction on the input time-frequency map, capture local features through convolution operation, and obtain a first time-frequency feature map; The maximum pooling layer is used to reduce the dimension of the first time-frequency feature map, retaining important features while reducing the amount of calculation, to obtain a second time-frequency feature map; The first UnitB, the first UnitA, the second UnitA, the third UnitA, the second UnitB and the fourth UnitA are used to implement multi-scale feature extraction and effective exchange of inter-channel information on the second time-frequency feature map through ShuffleNet unit modules with different configurations, so as to obtain a third time-frequency feature map; The CBAM module is used to adaptively adjust the channel weight and spatial weight of the third time-frequency feature map to obtain a fourth time-frequency feature map; The global pooling layer is used to perform global pooling on the fourth time-frequency feature map to further reduce the dimension and obtain a fifth time-frequency feature map; The fully connected layer is used to map the fifth time-frequency feature map to a category space; The Softmax layer is used to normalize the output of the fully connected layer and convert it into a probability distribution, indicating the confidence that the input time-frequency graph belongs to each category label; The output layer is used to make a final classification decision based on the confidence that the input time-frequency graph belongs to each category label, and output a detection result of whether the signal contains a frequency hopping signal.

5. The method according to any one of claims 1 to 4, characterized in that: The frequency hopping signal detection network is trained using a training set in the sample data set. The training process of the frequency hopping signal detection network is as follows: Use the Adam optimizer based on stochastic gradient descent to iteratively update the model parameters and minimize the loss function; If the accuracy does not improve within p training cycles, the learning rate in the next training cycle is reduced to q times that of the current training cycle; When the maximum number of training rounds is reached, the training is stopped and the network parameters with the minimum loss function on the validation set during the training process are saved as the trained frequency hopping signal detection network; where p≥3.

6. The method according to claim 5, characterized in that The determination is made based on the confidence level of the signal to be detected to obtain the detection result of each signal segment of the signal to be detected, as follows: Among them, t signal is the signal segment confidence of the signal to be detected, γ is the decision threshold, H1 indicates that the frequency hopping signal is detected, and H0 indicates that the frequency hopping signal does not exist.

7. The method according to claim 6, characterized in that The decision threshold is set as follows: Obtain N pure noise data for time-frequency representation and compressed sampling processing to obtain corresponding time-frequency graphs, input the trained frequency hopping signal detection network, and output a confidence score; Then, the confidence scores are arranged in descending order to obtain a confidence score sequence; The confidence score value at the corresponding quantile is determined based on the preset false alarm probability as the final decision threshold.

8. The method according to claim 7, characterized in that: If the detection results of each signal segment of the signal to be detected are noise, then the signal to be detected is noise and there is no frequency hopping signal; otherwise, there is a frequency hopping signal.

9. The method according to claim 5, characterized in that: The cross entropy loss function is used in the training process of the frequency hopping signal detection network.

10. A frequency hopping signal detection system based on time-frequency compression sampling deep learning, characterized in that: It includes signal acquisition module, data set construction module, network construction and training module and signal detection module; The signal acquisition module is used to respectively collect frequency hopping signal data when the frequency hopping signal exists and noise data when the frequency hopping signal does not exist, cut them according to preset lengths to obtain multiple signal segments, perform time-frequency characterization and compressed sampling processing on each signal segment, and obtain a frequency hopping time-frequency graph and a noise time-frequency graph; The data set construction module is used to construct a sample data set based on the frequency hopping time-frequency graph and the noise time-frequency graph and the corresponding sample labels; wherein the sample labels include signal and noise; The network construction and training module is used to build a frequency hopping signal detection network based on the lightweight convolutional neural network ShuffleNet and the convolutional attention module CBAM, and obtain a trained frequency hopping signal detection network using the sample data set; The signal detection module is used to obtain time-frequency diagrams corresponding to multiple signal segments of the signal to be detected, input the trained frequency hopping signal detection network, and obtain the confidence level of each signal segment of the signal to be detected; make a judgment based on the confidence level of each signal segment of the signal to be detected, and obtain a detection result of the signal to be detected.

Images