A frequency hopping frequency set estimation method and system based on a deep pattern recognition model

Abstract

The application discloses a frequency hopping frequency set estimation method and system based on a deep pattern recognition model, relates to the technical field of frequency hopping communication, and comprises the following steps: collecting in-phase and quadrature signal data containing a frequency hopping signal and preprocessing to obtain a point-like frequency hopping graph and divide a training set and a test set; a frequency hopping cycle pattern deep pattern recognition network model based on an attention mechanism is constructed and trained using the training set; the trained deep pattern recognition network model is used to extract the frequency hopping cycle pattern of the verification set; and the frequency hopping frequency set is extracted after matching and aligning the frequency hopping cycle pattern. The application generates a point-like frequency hopping graph, uses a frequency hopping cycle pattern deep pattern recognition network model based on an attention mechanism to perform pattern analysis and recognition on the point-like frequency hopping graph, extracts the frequency hopping cycle pattern in the point-like frequency hopping graph, extracts the frequency hopping frequency set after matching and aligning the frequency hopping cycle pattern, and significantly improves the accuracy and robustness of frequency hopping frequency set estimation under the condition of low algorithm complexity.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of frequency hopping communication, in particular to a frequency hopping frequency set estimation method and system based on a deep pattern recognition model. BACKGROUND

[0002] There are many kinds of existing technical methods and systems for estimating frequency hopping frequency sets, but from the perspectives of estimation accuracy, complexity and robustness, the existing methods and systems all have their own shortcomings. Traditional frequency hopping frequency set estimation methods and systems are mostly based on time-frequency analysis techniques (such as short-time Fourier transform STFT, Wigner-Ville transform, wavelet transform, etc.) to process the collected IQ signals to obtain time-frequency domain data, detect and extract the position information of the frequency hopping signals in the time and frequency domains based on energy detection, image processing and other methods, and then estimate the frequency hopping frequency table using a probability statistical method. Among them, the AI-based computer vision model technology as an image processing method detects and recognizes the frequency hopping signals in the time-frequency image by using a deep image target detection algorithm (such as the YOLO series), outputs the start and end time and frequency coordinates of the frequency hopping signals, and finally estimates the frequency hopping frequency table. The time-frequency analysis method based on AI deep image target detection has higher estimation accuracy than the traditional non-AI time-frequency analysis method. However, due to the lack of complete training data sets that can cover all complex environmental conditions for model training to capture the characteristics of complex frequency hopping signals and the frequency variation patterns with channel environment and time, the model generalization ability is weak and the robustness is poor in actual conditions. How to achieve high estimation accuracy while ensuring low algorithm complexity and high robustness has become the focus and difficulty of frequency hopping frequency set estimation problems, and deep pattern recognition is an important branch of pattern recognition, which combines deep learning technology to realize the recognition and analysis of complex patterns. Deep pattern recognition uses deep neural networks (such as convolutional neural networks, recurrent neural networks, etc.) to process and analyze these patterns, has strong non-linear modeling ability, and the deep learning model can automatically learn features from data and process complex patterns. Therefore, there is an urgent need for a frequency hopping frequency set estimation method and system based on a deep pattern recognition model. SUMMARY

[0003] The purpose of the present application is to overcome the defects that the existing technology cannot maintain high estimation accuracy while maintaining low algorithm complexity, strong model generalization ability and high robustness in the actual application of the common frequency hopping communication scene, and to provide a frequency hopping frequency set estimation method and system based on a deep pattern recognition model.

[0004] The technical scheme of the present application is as follows:

[0005] To achieve the above object, the application provides a frequency hopping frequency set estimation method based on a deep pattern recognition model, comprising the following steps:

[0006] Collecting in-phase and quadrature signal data containing frequency hopping signals and performing preprocessing to obtain a point-like frequency hopping graph and dividing the point-like frequency hopping graph into a training set and a test set;

[0007] Constructing a frequency hopping cycle pattern deep pattern recognition network model based on an attention mechanism and training the model using the training set;

[0008] Extracting the frequency hopping cycle pattern of the verification set using the trained frequency hopping cycle pattern deep pattern recognition network model based on the attention mechanism;

[0009] Extracting the frequency hopping frequency set after matching and aligning the frequency hopping cycle pattern.

[0010] Preferably, the preprocessing comprises:

[0011] Performing time-frequency transformation processing on the collected in-phase and quadrature signal data containing frequency hopping signals to obtain a two-dimensional amplitude spectrum time-frequency graph:

[0012]

[0013] wherein, T represents the number of time frames, L represents the number of frequency points, STFT represents a short-time Fourier transform, x represents a signal data and is an M-dimensional complex vector, represents a complex field, and M is the vector dimension, represents a real field, represents a real matrix with a dimension of T x L;

[0014] Performing noise reduction on the obtained time-frequency graph Specifically, performing erosion and dilation operations on the time-frequency graph according to the time dimension to remove mixed nonlinear structure signals and noise points;

[0015] wherein, the expression of the erosion operation is as follows:

[0016]

[0017] After the erosion operation is completed, a dilation operation is performed, and the expression of the dilation operation is as follows:

[0018]

[0019] wherein, P represents a positive integer taking the same order of magnitude as the time length of the single-frequency signal in the time-frequency graph,​ represents a set of positive integers;

[0020] obtaining a denoised time-frequency graph after completing the expansion operation , and converting the time-frequency graph into a point-like frequency hopping graph by using edge detection, contour detection and rectangular center extraction in sequence.

[0021] Preferably, the conversion of the time-frequency graph into the point-like frequency hopping graph by using edge detection, contour detection and rectangular center extraction in sequence specifically comprises the following steps:

[0022] normalizing the denoised time-frequency graph by elements and linearly mapping to [0, 255] to obtain , and performing binaryzation decision on to obtain a binary graph:

[0023]

[0024] wherein U is a preset denoising threshold;

[0025] the possible frequency hopping signals in the time-frequency graph are detected by using edge detection algorithm and contour detection algorithm in sequence to obtain a signal contour image:

[0026]

[0027] wherein, is a Canny edge detection function of an open source computer vision library OpenCV, and are a low threshold and a high threshold of the Canny edge detection function, respectively;

[0028]

[0029] wherein, is a contour detection function of the open source computer vision library OpenCV, the contour retrieval mode adopts , i.e. only the external contour is retrieved, and the contour approximation method adopts , i.e. all contour points are retained;

[0030] rectangular frame analysis and extraction are performed on to obtain a rectangular set :

[0031]

[0032] wherein, is a rectangular contour fitting function in the open source computer vision library OpenCV, which is used for detecting and fitting to obtain a set of circumscribed rectangular contours in the image: ​​​

[0033]

[0034] wherein, is the coordinate of the upper left corner of the rectangle, , are the width and height of the rectangle respectively, and V is the number of detected rectangular frames;

[0035] using calculate the center coordinates of each rectangle to generate a point-like frequency hopping pattern:

[0036] .

[0037] Preferably, the attention mechanism-based frequency hopping period pattern deep mode recognition network model FHPR adopts an encoder-attention module-decoder structure; the encoder includes three convolutional blocks, each of which includes a 3x3 convolutional layer, batch normalization and a ReLU activation function, and a 2x2 max pooling layer is further added after each convolutional block, and the encoder part is further deepened to extract features through two layers of consecutive convolutional layers; the decoder includes three deconvolutional blocks, each of which includes a 3x3 deconvolutional layer, batch normalization and a ReLU activation function, and a 2x2 up-sampling layer corresponding to the 2x2 max pooling layer in the encoder is further added before each deconvolutional block, and each layer in the decoder is connected to the corresponding layer in the encoder through a skip connection; an attention module is further arranged between the encoder and the decoder to strengthen the important features of the point-like frequency hopping pattern.

[0038] Preferably, the input of the model is a point-like frequency hopping pattern with a dimension of First, the multi-scale features are extracted through the encoder part, and then the size of the feature map of the multi-scale features is gradually reduced and the number of channels is gradually increased through convolution and pooling operations;

[0039] The important features of the point-like frequency hopping pattern are strengthened through the attention module between the encoder and the decoder;

[0040] Through the decoder part, the feature map size is gradually up-sampled, the number of channels is gradually reduced, the output is reconstructed, and the features of the encoder are fused using a skip connection to retain detailed information. Finally, a 1x1 convolutional layer is used to map the features of the decoder to the final output, outputting a frequency hopping period pattern mode recognition result with the same dimension as the point-like frequency hopping pattern, and then performing segmentation to independently separate each period pattern mode region in the frequency hopping period pattern.

[0041] Preferably, the attention module divides the input feature map into groups according to the channel dimension, and each group has a size of The input feature map after grouping passes through two parallel branches of the attention module;

[0042] The first branch respectively uses average pooling along the X direction and the Y direction to extract global information, and then performs feature fusion through splicing and a 1x1 convolution layer, and then respectively generates the attention maps of the X direction and the Y direction through a Sigmoid function, and then performs weight adjustment, group normalization and average pooling operation, and then generates the attention map of the first branch through a Softmax function;

[0043] The second branch uses a 3x3 convolution layer to capture local spatial information, and then performs an average pooling operation, and then uses a Softmax function to generate the attention map of the second branch;

[0044] The attention maps generated by the two branches are multiplied by a matrix operation to generate the final feature representation, and then a Sigmoid function is used to generate the final feature map. The final feature map and the input feature map are combined after weight adjustment through addition and multiplication operations to generate the final output feature of the attention module.

[0045] Preferably, the loss function of the frequency hopping period pattern deep pattern recognition network model FHPR based on the attention mechanism is:

[0046]

[0047]

[0048]

[0049] wherein, , is a weight parameter, , usually , is a binary cross-entropy loss, which calculates the cross-entropy between the true label and the predicted value for each pixel, is a Dice loss, which directly measures the overlap between the predicted frequency hopping period pattern region and the true frequency hopping period pattern region to judge the segmentation effect of the model, is the total number of pixels in the image, is the true label of the ith pixel, when the ith pixel belongs to the frequency hopping period pattern region is 1, when the ith pixel belongs to the background region is 0, is the predicted value of the ith pixel, which represents the probability that the pixel belongs to the frequency hopping period pattern region, is a constant to prevent the denominator from being zero.

[0050] Preferably, the matching alignment comprises:

[0051] The first frequency hopping pattern As a template, the remaining frequency hopping patterns The pattern to be matched is used for template matching and alignment. The template matching and alignment rule adopts the normalized squared difference matching method:

[0052]

[0053] Where Q is the number of frequency hopping patterns extracted by the model. In order to make With template Match the aligned displacement.

[0054] Preferably, the step of extracting the frequency hopping set specifically includes:

[0055] Statistical averaging of the matched and aligned frequency hopping patterns by pixels:

[0056]

[0057] Using a non-maximum suppression method, the frequency hopping pattern after template alignment is... To find local maxima, first, filter... The pixel value is greater than a given threshold The pixel coordinates are determined, and the pixel values ​​are sorted from largest to smallest. The pixels are sorted, and then the sorted pixels are traversed sequentially, marking their positions in their surrounding neighborhoods. This ensures that only the coordinates of the pixel with the largest value are retained within the neighborhood, while the coordinates of surrounding pixels with smaller values ​​are suppressed. The final result is the set of pixel coordinates for the processed frequency hopping pattern.

[0058]

[0059] in, The number of pixel coordinates to be retained in the end. Frequency points corresponding to all pixels in set This refers to the extracted set of frequency hopping frequencies, where the estimated frequency hopping values ​​are... , Sampling frequency, The number of FFT points for the Short-Time Fourier Transform (STFT) is given.

[0060] To achieve the objective of the invention, a second aspect of the present invention provides a frequency hopping set estimation system based on a deep pattern recognition model, applying the frequency hopping set estimation method based on a deep pattern recognition model described above. The system includes:

[0061] The data acquisition and preprocessing module is used for collecting and preprocessing in-phase and quadrature signal data containing frequency hopping signals, obtaining a point-like frequency hopping graph, and dividing the point-like frequency hopping graph into a training set and a test set.

[0062] The deep pattern recognition model construction and training module is used for constructing a frequency hopping cycle pattern deep pattern recognition network model based on an attention mechanism and training the model using the training set.

[0063] The frequency hopping cycle pattern extraction module is used for extracting the frequency hopping cycle pattern of the verification set by using the trained frequency hopping cycle pattern deep pattern recognition network model based on the attention mechanism.

[0064] The frequency hopping frequency set extraction module is used for extracting the frequency hopping frequency set after matching and aligning the frequency hopping cycle pattern.

[0065] Compared with the prior art, the present application has the following beneficial effects:

[0066] The present application converts the frequency hopping signal image in the time-frequency graph into a frequency hopping point through image preprocessing to obtain a point-like frequency hopping graph, then performs pattern analysis and recognition on the point-like frequency hopping graph by using a frequency hopping cycle pattern deep pattern recognition network model based on an attention mechanism, extracts the frequency hopping cycle pattern in the point-like frequency hopping graph, and finally extracts the frequency hopping frequency set after matching and aligning the frequency hopping cycle pattern, thereby significantly improving the accuracy and robustness of the frequency hopping frequency set estimation under the condition of low algorithm complexity. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 The step flowchart of the frequency hopping frequency set estimation method based on the deep pattern recognition model of embodiment 1 of the present application;

[0068] Figure 2 The structural schematic diagram of the frequency hopping cycle pattern deep pattern recognition network model based on the attention mechanism of embodiment 1 of the present application;

[0069] Figure 3 The structural schematic diagram of the attention module in the frequency hopping cycle pattern deep pattern recognition network model based on the attention mechanism of embodiment 1 of the present application;

[0070] Figure 4 The pattern recognition training and inference flowchart of the frequency hopping cycle pattern deep pattern recognition network model based on the attention mechanism of embodiment 1 of the present application;

[0071] Figure 5 The average error comparison diagram of the frequency hopping frequency set estimation of different algorithms of embodiment 3 of the present application;

[0072] Figure 6 The graphical visualization result diagram of the frequency hopping signal time-frequency graph of embodiment 3 of the present application;

[0073] Figure 7 a graphical visualization result diagram of a profile detection stage of Embodiment 3 of the present application;

[0074] Figure 8 a graphical visualization result diagram of a rectangular center extraction stage of Embodiment 3 of the present application;

[0075] Figure 9 a graphical visualization result diagram of a frequency hopping cycle pattern depth mode recognition stage of Embodiment 3 of the present application;

[0076] Figure 10 a graphical visualization result diagram of a template matching pixel superposition averaging stage of Embodiment 3 of the present application;

[0077] Figure 11 a graphical visualization result diagram of a frequency hopping frequency set extraction result of Embodiment 3 of the present application. DETAILED DESCRIPTION

[0078] The specific embodiments of the present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. The following embodiments are used to illustrate the present application, but are not used to limit the scope of the present application.

[0079] Embodiment 1

[0080] Please refer to Figure 1 , Figure 1 a step flowchart of a frequency hopping frequency set estimation method based on a depth mode recognition model of Embodiment 1 of the present application, Embodiment 1 provides a frequency hopping frequency set estimation method based on a depth mode recognition model, including the following steps S1-S4:

[0081] Step S1: collecting in-phase quadrature signal data containing frequency hopping signals and performing preprocessing to obtain a point hopping diagram and dividing the point hopping diagram into a training set and a test set.

[0082] First, the collected in-phase quadrature (IQ) data containing frequency hopping signals is preprocessed by short-time Fourier transform (STFT) and filtering and noise reduction, etc., to obtain a point hopping diagram, which is as follows:

[0083] (1) STFT transform and preprocessing

[0084] First, the collected in-phase quadrature signal data containing frequency hopping signals is preprocessed by short-time Fourier transform (STFT) and filtering and noise reduction, etc., to obtain a point hopping diagram, which is as follows:

[0085]

[0086] Among them, ​T represents the number of time frames, L represents the number of frequency points, represents a short-time Fourier transform, represents that the signal data x is an M-dimensional complex vector, represents a complex field, and M is a vector dimension, represents a real field, represents a real matrix with a dimension of T x L.

[0087] (2) Then, the obtained time-frequency graph is denoised, specifically, erosion and dilation operations are performed on the time-frequency graph according to the time dimension, and mixed nonlinear structure signals and noise points are removed;

[0088] The expression of the erosion operation is as follows:

[0089]

[0090] After the erosion operation is completed, the dilation operation is performed, and the expression of the dilation operation is as follows:

[0091]

[0092] wherein, represents a positive integer of the same order of magnitude as the time length of the single-frequency signal in the time-frequency graph, represents a set of positive integers,

[0093] After the dilation operation is completed, the denoised time-frequency graph is obtained, and the time-frequency graph is converted into a point-like frequency hopping graph by using edge detection, contour detection, and rectangular center extraction in sequence.

[0094] (3) Finally, the time-frequency graph is converted into a point-like frequency hopping graph by using edge detection, contour detection, and rectangular center extraction in sequence, specifically including the following steps:

[0095] The denoised time-frequency graph is normalized according to elements and linearly mapped to [0, 255] to obtain , the binary graph is obtained by performing binaryzation decision on :

[0096]

[0097] wherein, U is a preset denoising threshold, and U = 50 in Embodiment 1.

[0098] The edge detection algorithm and the contour detection algorithm are used in sequence to detect the possible frequency hopping signals in the time-frequency graph, and a signal contour image is obtained:

[0099]

[0100] wherein, is a Canny edge detection function of the open source computer vision library OpenCV, and are a low threshold value and a high threshold value of the Canny edge detection function, respectively, and in this embodiment 1, = 50, = 200;

[0101]

[0102] wherein, is a contour detection function of the open source computer vision library OpenCV, and the contour retrieval mode is , i.e. only the external contour is retrieved, and the contour approximation method is , i.e. all contour points are retained;

[0103] rectangular frame analysis is performed on to obtain a rectangular set :

[0104]

[0105] wherein, is a rectangular contour fitting function in the open source computer vision library OpenCV, used to detect and fit the set of external rectangular contours in the image:

[0106]

[0107] wherein, is the upper left corner coordinate of the rectangle, , are the width and height of the rectangle, respectively, and V is the number of detected rectangular frames;

[0108] The center coordinates of each rectangle are calculated using to generate a point-like frequency hopping map:

[0109]

[0110] Finally, the point-like frequency hopping map is divided into a training set and a test set.

[0111] Step S2: Constructing a frequency hopping cycle pattern deep pattern recognition network model based on an attention mechanism and training it using the training set.

[0112] The frequency hopping pattern deep pattern recognition network model FHPR (Frequency Hopping Pattern Recognition) constructed in Example 1 is used to identify frequency hopping patterns and extract the dotted frequency hopping pattern obtained in step S1. The frequency hopping cycle pattern is output as a set of frequency hopping cycle patterns:

[0113]

[0114] Where Q is the number of frequency hopping patterns extracted by the FHPR model.

[0115] Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of the FHPR model constructed in Example 1. The model adopts an encoder-attention module-decoder structure. Figure 2 The convolutional blocks are represented by blue rectangles. The numbers 1, 32, 64, 128, and 256 represent the number of channels. The yellow circle A represents the attention module. The blue rectangle on the left represents the convolutional blocks in the encoder section, and the blue rectangle on the right represents the convolutional blocks in the decoder section. The green arrow pointing upwards represents upsampling, the black arrow pointing to the right represents skip connections, the white arrow pointing to the right represents convolution, and the red arrow pointing downwards represents max pooling.

[0116] The encoder consists of three convolutional blocks, each containing a 3x3 convolutional layer, batch normalization, and a ReLU activation function. A 2x2 max-pooling layer is added after each convolutional block to reduce the size of the feature map while preserving important information. The encoder part is then passed through two consecutive convolutional layers. Figure 2 The blue rectangle at the bottom center and the white arrow pointing to the right further extract features to enhance the network's ability to represent features in the input data. The decoder includes three deconvolutional blocks, each consisting of a 3x3 deconvolutional layer, batch normalization, and a ReLU activation function. Before each deconvolutional block, a 2x2 upsampling layer corresponding to the 2x2 max-pooling layer in the encoder is added to increase the size of the feature map. Each layer in the decoder is connected to the corresponding layer in the encoder via skip connections. An attention module is also set between the encoder and decoder to enhance important features of the dotted frequency hopping pattern.

[0117] The input to the FHPR model is a dimension of The dotted frequency hopping map first passes through the encoder part to extract multi-scale features, and then the size of the feature map of the multi-scale features is gradually reduced and the number of channels is increased through convolution and pooling operations.

[0118] Then, an attention module is used between the encoder and decoder to enhance key features of the dotted frequency hopping mode. See also Figure 3 , Figure 3Fig. 1 is a structural schematic diagram of an attention module of an FHPR model of Embodiment 1, Figure 3 In the formula, 1xHxW, 1 / G, 1 / Gx1xW, etc. represent dimension sizes. The attention module divides the input feature map into G groups (G is usually taken as 8, 16, 32, 64) according to the channel dimension, and each group has a size of The input feature map after grouping passes through two parallel branches of the attention module.

[0119] The first branch uses average pooling along the X direction and the Y direction to extract global information, and then passes through a concatenation and a 1x1 convolution layer and performs feature fusion (concatenating feature maps at different levels to realize multi-scale feature fusion), and then generates attention maps in the X direction and the Y direction through Sigmoid activation functions, and then performs weight adjustment, group normalization operation and average pooling operation, and then generates an attention map of the first branch through a Softmax activation function.

[0120] The second branch uses a 3x3 convolution layer to capture local spatial information, and then performs an average pooling operation and uses a Softmax activation function to standardize to generate an attention map of the second branch.

[0121] The attention maps generated by the two branches are multiplied by a matrix operation to generate the final feature representation, and then a Sigmoid function is used to generate the final feature map. The final feature map is combined with the input feature map after weight adjustment to generate the final output feature of the attention module, i.e., the attention map. In this process, the feature map is fused through multiple steps to generate a global attention map, which captures the pixel-level pairing relationship.

[0122] Then, through the decoder part, the feature map size is gradually upsampled, the channel number is reduced, the output is reconstructed, and the features of the encoder are fused using a skip connection to retain detailed information. Finally, a 1x1 convolution layer is used to map the features of the decoder to the final output, and an output of the same dimension as the point-like frequency hopping pattern is obtained. The frequency hopping period pattern recognition result is then segmented to separate each period pattern region in the frequency hopping period pattern.

[0123] The loss function L of the FHPR model of Embodiment 1 is:

[0124]

[0125]

[0126]

[0127] wherein, , is a weight parameter, , and is usually taken as , is the binary cross-entropy loss, for each pixel, it calculates the cross-entropy between the true label and the predicted value, is the Dice loss, by directly measuring the overlap between the predicted frequency hopping period pattern region and the true frequency hopping period pattern region, to judge the segmentation effect of the model, is the total number of pixels in the image, is the true label of the ith pixel, when the ith pixel belongs to the frequency hopping period pattern region is 1, when the ith pixel belongs to the background region is 0, is the predicted value of the ith pixel, indicating the probability that the pixel belongs to the frequency hopping period pattern region, is a small constant to prevent the denominator from being zero.

[0128] After the FHPR model is constructed, it needs to be trained using the training set obtained in S1. After the model training is completed, it can be used for subsequent inference of the test set.

[0129] Please refer to Figure 4 , Figure 4 is the pattern recognition training and inference process diagram of the attention mechanism based frequency hopping period pattern deep pattern recognition network model of embodiment 1 of the present application.

[0130] In the training phase, the IQ signal samples in the training set are subjected to STFT transformation and preprocessing to obtain point-shaped frequency hopping pattern sample set, and are used as input to train the attention mechanism based frequency hopping period pattern deep pattern recognition network model. During the learning process, the model is trained using the Adam optimizer in a supervised learning manner, and the label data is used to guide the model to learn the mapping relationship between the frequency cycle hopping rule and the frequency hopping period pattern mode. Subsequently, in the inference phase, the trained model is used to recognize the frequency hopping pattern mode of the test set samples to obtain the frequency hopping period pattern set.

[0131] Step S3: Extract the frequency hopping period pattern of the verification set using the trained attention mechanism based frequency hopping period pattern deep pattern recognition network model.

[0132] Please refer to Figure 4 , in the inference phase, the trained model is used to recognize the frequency hopping pattern deep pattern of the test set samples to obtain the frequency hopping period pattern set.

[0133] Step S4: Extract the frequency hopping frequency set after matching and aligning the frequency hopping period pattern.

[0134] Matching and alignment includes:

[0135] The first frequency hopping pattern As a template, the remaining frequency hopping patterns The pattern to be matched is used for template matching and alignment. The template matching and alignment rule adopts the normalized squared difference matching method:

[0136]

[0137] Where Q is the number of frequency hopping patterns extracted by the FHPR model. In order to make With template Match the aligned displacement.

[0138] Statistical averaging of the matched and aligned frequency hopping patterns by pixels:

[0139]

[0140] Using a non-maximum suppression method, the frequency hopping pattern after template alignment is... To find local maxima, first, filter... The pixel value is greater than a given threshold (In this embodiment 1) The pixel coordinates are obtained, and the pixel values ​​are sorted from largest to smallest. The pixels are sorted, and then the sorted pixels are traversed sequentially, marking their positions in their surrounding neighborhoods. This ensures that only the coordinates of the pixel with the largest value are retained within the neighborhood, while the coordinates of surrounding pixels with smaller values ​​are suppressed. The final result is the set of pixel coordinates for the processed frequency hopping pattern.

[0141]

[0142] in, The number of pixel coordinates to be retained in the end. Frequency points corresponding to all pixels in set This refers to the extracted set of frequency hopping frequencies, where the estimated frequency hopping values ​​are... , Sampling frequency, The number of FFT points for the Short-Time Fourier Transform (STFT) is given.

[0143] Example 2

[0144] This embodiment 2, based on embodiment 1, applies a frequency hopping set estimation method based on a deep pattern recognition model from embodiment 1 to provide a frequency hopping set estimation system based on a deep pattern recognition model. The system includes:

[0145] A data acquisition and preprocessing module is configured to collect in-phase and quadrature signal data containing frequency hopping signals and perform preprocessing to obtain a point-like frequency hopping graph and divide the point-like frequency hopping graph into a training set and a test set.

[0146] A deep pattern recognition model construction and training module is configured to construct a frequency hopping cycle pattern deep pattern recognition network model based on an attention mechanism and train the model using the training set.

[0147] A frequency hopping cycle pattern extraction module is configured to extract frequency hopping cycle patterns of the test set using the trained frequency hopping cycle pattern deep pattern recognition network model based on the attention mechanism.

[0148] A frequency hopping frequency set extraction module is configured to extract a frequency hopping frequency set after matching and aligning the frequency hopping cycle patterns.

[0149] The other steps or technical details of this embodiment 2 are the same as those of embodiment 1, which will not be described here again.

[0150] Embodiment 3

[0151] To verify the effectiveness and progressiveness of the present application, this embodiment 3 is based on the simulation results and visual analysis of embodiments 1 and 2.

[0152] A frequency hopping sequence and a random bit stream are generated by simulation, wherein the frequency hopping signal adopts MSK modulation, the number of frequency hopping points is 62, the frequency hopping interval is 200 Hz, the frequency hopping rate is 500 Hz, and the signal-to-noise ratio range is [-12 dB, 6 dB]. The receiving sampling rate is 10 MHz, and the number of FFT points of STFT . Compared with the traditional power spectrum peak detection method and the time-frequency diagram YOLO-v5 frequency hopping signal detection method, the average estimation error is taken as the performance evaluation index:

[0153]

[0154] wherein, is the kth frequency real value, K is the number of real frequency sets, is the frequency estimation value corresponding to the kth frequency real value in the nth experiment, N is the number of simulation experiments. The performance comparison results of different frequency hopping frequency set estimations can be found in Figure 5 , Figure 5 is the average error comparison chart of the frequency hopping frequency set estimations of different algorithms of embodiment 3 of the present application, and it can be known from Figure 5 that the average error of the frequency hopping frequency set estimation of the frequency hopping frequency set estimation method based on the deep pattern recognition model of the present application is the lowest among several methods under the same signal-to-noise ratio, which reflects the progressiveness of the method. The visual process of the simulation experiment can be found in Figures 6-11, respectively, are the visual simulation results of the output of each stage of the frequency hopping frequency set estimation method, wherein, Figure 6 is a graphical visualization result of a time-frequency diagram of a frequency hopping signal, Figure 7 is a graphical visualization result of a contour detection stage, Figure 8 is a graphical visualization result of a rectangle center extraction stage, Figure 9 is a graphical visualization result of a frequency hopping period pattern depth mode recognition stage, Figure 10 is a graphical visualization result of a template matching pixel superposition average stage, Figure 11 is a graphical visualization result of a frequency hopping frequency set extraction result.

[0155] In summary, the embodiment described above, the present application is aimed at the common frequency hopping communication scene in practical application, a frequency hopping frequency set estimation method of frequency hopping signal is proposed. First, the image preprocessing is used to convert the frequency hopping signal image in the time-frequency diagram into a frequency hopping point, and a point-like frequency hopping diagram is obtained. Then, a deep network model based on attention mechanism is used to analyze and identify the pattern of the point-like frequency hopping diagram, and the frequency hopping period pattern in the point-like frequency hopping diagram is extracted. Finally, the frequency hopping frequency set is extracted after matching and aligning the frequency hopping period pattern. The method significantly improves the accuracy and robustness of the frequency hopping frequency set estimation under the condition of low algorithm complexity.

[0156] The above is the preferred embodiment of the present application. It should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should also be considered as the protection scope of the present application.

Claims

1. A frequency hopping set estimation method based on a deep pattern recognition model, characterized in that, Includes the following steps: Acquire in-phase quadrature signal data containing frequency hopping signals and perform preprocessing, the preprocessing including: For the acquired in-phase quadrature signal data containing frequency hopping signals Perform time-frequency transformation to obtain a two-dimensional amplitude spectrum time-frequency diagram: in, T represents the number of time frames. L represents the number of frequency points. Represents the short-time Fourier transform. This indicates that the signal data x is an M-dimensional complex vector. Represents the complex field, where M is the vector dimension. Represents the real number field. Represents a real matrix of dimension T×L; The obtained time-frequency diagram To reduce noise, the time-frequency graph is subjected to erosion and dilation operations along the time dimension to remove mixed nonlinear structural signals and noise. The expression for the erosion operation is as follows: After the erosion operation is completed, the dilation operation is performed. The dilation operation expression is as follows: in, This indicates that P is a positive integer of the same order of magnitude as the time length of the single-frequency signal in the time-frequency diagram. Represents the set of positive integers; After performing the dilation operation, the denoised time-frequency graph is obtained. The time-frequency map is processed sequentially using edge detection, contour detection, and rectangle center extraction. Convert to a dotted frequency hopping plot; The dotted frequency hopping map is divided into a training set and a test set; A frequency-hopping periodic pattern deep pattern recognition network model based on an attention mechanism is constructed and trained using the training set. The frequency-hopping periodic pattern deep pattern recognition network model based on an attention mechanism adopts an encoder-attention module-decoder structure. Each layer in the decoder is connected to the corresponding layer in the encoder via skip connections. An attention module is also set between the encoder and the decoder. The attention module groups the feature maps extracted by the encoder according to the channel dimension. The grouped input feature maps are processed by two parallel branches of the attention module and the final result is obtained by matrix multiplication, sigmoid function and weighted adjustment. The result is then input into the decoder to generate a frequency-hopping periodic pattern. The frequency hopping periodic patterns of the validation set are extracted using the trained attention-based frequency hopping periodic pattern deep pattern recognition network model. The frequency hopping periodic pattern is matched and aligned. The matching and alignment rule adopts the normalized squared difference matching method. The non-maximum suppression method is used to extract the frequency hopping frequency set by statistically averaging the matched and aligned frequency hopping pattern by pixels.

2. The frequency hopping set estimation method based on a deep pattern recognition model according to claim 1, characterized in that, The time-frequency map is sequentially processed using edge detection, contour detection, and rectangle center extraction. The conversion to a dotted frequency hopping graph involves the following steps: Denoising time-frequency graph Element-normalized and linearly mapped to [0, 255] yields ,right Binarization decision is performed to obtain a binary image: Where U is the preset noise reduction threshold; The edge detection algorithm and the contour detection algorithm are used sequentially to detect possible frequency hopping signals in the time-frequency map, and the signal contour image is obtained: in, This refers to the Canny edge detection function in the open-source computer vision library OpenCV. and These are the low and high thresholds for the Canny edge detection function, respectively. in, This is a contour detection function from the open-source computer vision library OpenCV. The contour retrieval mode adopts... That is, only the outer contour is retrieved, and the contour approximation method adopts... That is, all contour points are preserved; right Perform bounding box analysis and extraction to obtain a set of rectangles. : in, This is a rectangular contour fitting function from the open-source computer vision library OpenCV, used to detect and fit a set of bounding rectangular contours in an image. in, Let be the coordinates of the top-left corner of the rectangle. , These represent the width and height of the rectangle, respectively, and V is the number of detected rectangles. use Calculate the center coordinates of each rectangle to generate a dotted frequency hopping map: 。 3. The frequency hopping set estimation method based on a deep pattern recognition model according to claim 1, characterized in that, The frequency-hopping periodic pattern deep pattern recognition network model FHPR based on the attention mechanism adopts an encoder-attention module-decoder structure. The encoder includes three convolutional blocks, each of which includes a 3x3 convolutional layer, batch normalization, and ReLU activation function. A 2x2 max pooling layer is added after each convolutional block. The encoder part is then further processed by two consecutive convolutional layers to extract features. The decoder includes three deconvolutional blocks, each of which includes a 3x3 deconvolutional layer, batch normalization, and ReLU activation function. A 2x2 upsampling layer corresponding to the 2x2 max pooling layer in the encoder is added before each deconvolutional block. Each layer in the decoder is connected to the corresponding layer in the encoder via skip connections. An attention module is also set between the encoder and decoder to enhance the important features of the dotted frequency-hopping pattern.

4. The frequency hopping set estimation method based on a deep pattern recognition model according to claim 3, characterized in that, The input to the model is of dimension . The dotted frequency hopping map first passes through the encoder part to extract multi-scale features, and then gradually reduces the size of the feature map of multi-scale features and increases the number of channels through convolution and pooling operations; An attention module is used between the encoder and decoder to enhance key features of the dotted frequency hopping mode; In the decoder section, the feature map size is progressively upsampled to reduce the number of channels and reconstruct the output. At the same time, skip connections are used to fuse the encoder features to preserve detailed information. Finally, a 1x1 convolutional layer maps the decoder features to the final output, which outputs a frequency hopping periodic pattern recognition result with the same dimension as the dotted frequency hopping map. Then, segmentation is performed to independently separate the periodic pattern regions in the frequency hopping periodic pattern.

5. The frequency hopping set estimation method based on a deep pattern recognition model according to claim 4, characterized in that, The attention module divides the input feature map into G groups according to the channel dimension, with each group having a size of [missing information]. The grouped input feature maps pass through two parallel branches of the attention module; The first branch uses average pooling along the X and Y directions to extract global information, then performs feature fusion through concatenation and a 1×1 convolutional layer. After generating attention maps in the X and Y directions using the Sigmoid function, weight adjustment, grouping normalization, and average pooling are performed. Finally, the attention map of the first branch is generated using the Softmax function. The second branch uses a 3×3 convolutional layer to capture local spatial information, followed by average pooling and then normalization using the Softmax function to generate the attention map for the second branch. The attention maps generated from these two branches are multiplied by matrix to generate the final feature representation. Then, the final feature map is generated by the sigmoid function. The final feature map is combined with the input feature map after addition and multiplication operations, i.e., weight adjustment, to generate the final output feature of the attention module.

6. The frequency hopping set estimation method based on a deep pattern recognition model according to claim 4, characterized in that, The loss function of the frequency-hopping periodic pattern deep pattern recognition network model FHPR based on the attention mechanism is: in, , For weight parameters, , For the binary cross-entropy loss, for each pixel, it calculates the cross-entropy between the true label and the predicted value. The Dice loss is used to judge the segmentation performance of the model by directly measuring the overlap between the predicted frequency-hopping periodic pattern region and the actual frequency-hopping periodic pattern region. It is the total number of pixels in the image. It is the true label of the i-th pixel, when the i-th pixel belongs to the frequency hopping periodic pattern region. The value is 1 when the i-th pixel belongs to the background region. =0, It is the predicted value of the i-th pixel, representing the probability that the pixel belongs to the frequency hopping periodic pattern region. It is a constant to prevent the denominator from being zero.

7. The frequency hopping set estimation method based on a deep pattern recognition model according to claim 1, characterized in that, The matching alignment includes: Using the first frequency hopping pattern As a template, the remaining frequency hopping patterns The pattern to be matched is used for template matching and alignment. The template matching and alignment rule adopts the normalized squared difference matching method: Where Q is the number of frequency hopping patterns extracted by the model. To make the remaining frequency hopping patterns With template Match the aligned displacement.

8. The frequency hopping set estimation method based on a deep pattern recognition model according to claim 7, characterized in that, The specific steps for extracting the frequency hopping set are as follows: Statistical averaging of the matched and aligned frequency hopping patterns by pixels: Using a non-maximum suppression method, the frequency hopping pattern after template alignment is... To find local maxima, first, filter... The pixel value is greater than a given threshold The pixel coordinates are determined, and the pixel values ​​are sorted from largest to smallest. The pixels are sorted, and then the sorted pixels are traversed sequentially, marking their positions in their surrounding neighborhoods. This ensures that only the coordinates of the pixel with the largest value are retained within the neighborhood, while the coordinates of surrounding pixels with smaller values ​​are suppressed. The final result is the set of pixel coordinates for the processed frequency hopping pattern. in, The number of pixel coordinates to be retained in the end. Frequency points corresponding to all pixels in set This refers to the extracted set of frequency hopping frequencies, where the estimated frequency hopping values ​​are... , Sampling frequency, The number of FFT points for the Short-Time Fourier Transform (STFT) is given.

9. A frequency hopping set estimation system based on a deep pattern recognition model, employing the frequency hopping set estimation method based on a deep pattern recognition model as described in any one of claims 1-8, characterized in that, The system includes: The data acquisition and preprocessing module is used to collect in-phase orthogonal signal data containing frequency hopping signals and perform preprocessing to obtain a dotted frequency hopping map and divide the dotted frequency hopping map into a training set and a test set; A deep pattern recognition model construction and training module is used to construct a frequency-hopping periodic pattern deep pattern recognition network model based on an attention mechanism and train it using the training set. The frequency hopping period pattern extraction module is used to extract the frequency hopping period pattern of the validation set using the trained attention-based frequency hopping period pattern deep pattern recognition network model. The frequency hopping frequency set extraction module is used to extract the frequency hopping frequency set after matching and aligning the frequency hopping period pattern.

Images