Frequency Hopping Signal Parameter Estimation Method Based on Maximum Entropy Binarized Time-Spectrum Map and ABF-YOLOv5s

Abstract

The present invention belongs to the field of frequency hopping signal parameter estimation and provides a frequency hopping signal parameter estimation method based on maximum entropy binarized time-frequency graph and ABF-YOLOv5s. The method comprises the following steps: first obtaining the frequency hopping signal, performing short-time Fourier transform and maximum entropy binarization processing respectively to obtain a maximum entropy binarized time-frequency graph; then inputting the maximum entropy binarized time-frequency graph into the ABF-Yolov5s model to obtain the frequency hopping signal position information; and finally, performing period and frequency estimation of the frequency hopping signal based on the correlation between the frequency hopping signal position information and the time-frequency parameters. The present invention introduces the ASPP module and the BOT3 module into the ABF-Backbone portion and the BiFPN module into the ABF-Neck portion. This method can improve the accuracy of frequency hopping signal period and frequency estimation under low signal-to-noise ratio conditions and has good robustness.

Description

Technical Field

[0001] The present invention belongs to the field of frequency hopping signal parameter estimation and relates to a frequency hopping signal period estimation and frequency estimation method based on maximum entropy binarization time-frequency diagram and ABF-YOLOv5s. Background Art

[0002] Frequency hopping (FH) communication is a technology that rapidly switches between different frequencies to transmit data. It offers strong resistance to interference, interception, and fading. Obtaining the parameters of the FH signal is a prerequisite for implementing FH jamming. However, with the increasing complexity of the electromagnetic environment, obtaining accurate FH signal parameters has become a challenge.

[0003] Because the spectrum of a frequency-hopping signal varies over time, it is a typical non-stationary signal. Time-frequency analysis can characterize different frequencies and signal energy intensities under different time conditions. Therefore, it is the main method for frequency-hopping signal parameter estimation. Hui DI et al., in their paper "Research on parameter blind estimation algorithm of costas frequency-hopping signals," used the short-time Fourier transform (STFT) to represent frequency-hopping signals in time and frequency. However, due to the influence of the window function, this method suffers from time-frequency uncertainty and cannot simultaneously balance time resolution and frequency resolution, resulting in suboptimal accuracy for both period and frequency estimation of frequency-hopping signals. Wang Zhongren et al. proposed using the Wigner-Ville distribution (WVD) for parameter estimation of frequency-hopping signals. Compared to the single linear transform of the STFT, the WVD is a bilinear time-frequency transform method that solves the problem of the STFT window function and improves the accuracy of frequency-hopping signal parameter estimation. However, for frequency-hopping signals, which have multiple components with different frequencies in the spectrum, the WVD transform is prone to generate cross terms, which interfere with the frequency estimation of the frequency-hopping signal. Chen T proposed using the smoothed pseudo-Wigner-Vieri distribution (SPWD) to analyze frequency-hopping signals. SPWD uses smoothing by windowing in both the time and frequency domains to suppress cross-interference in WVD and reduce the frequency estimation error of the frequency-hopping signal, but this also reduces the time-frequency resolution. Chen Chao, Fu Weihong et al. combined STFT with WVD or SPWVD, suppressing cross-interference while maintaining high time-frequency resolution. This further improved the parameter estimation accuracy of frequency-hopping signals at high signal-to-noise ratios. However, at low signal-to-noise ratios, the time-frequency distribution is significantly affected by noise, resulting in low parameter estimation accuracy for frequency-hopping signals. To improve parameter estimation accuracy for frequency-hopping signals at low signal-to-noise ratios, Xi Youyou et al. combined binarization with morphological filtering. Using a global thresholding method for binarization, they effectively removed noise. However, the global thresholding method only considers maximizing background and object segmentation when segmenting the image, which can easily lose edge information in the frequency-hopping signal. Morphological filtering, on the other hand, can partially restore edge information in the frequency-hopping signal. With the development of deep learning, applying deep learning to the field of wireless communications has become a research hotspot. Wang Yuyang and others used the SSD model to estimate the parameters of frequency hopping signals. Experimental results show that deep learning can achieve high accuracy in hopping rate and frequency estimation of frequency hopping signals. However, the SSD model uses a single network on multiple feature maps, and the representation ability of shallow feature maps is not strong, which makes it easy to miss or misdetect frequency hopping signals, resulting in large errors in the estimation of the frequency hopping signal period. Summary of the Invention

[0004] In order to improve the accuracy of frequency hopping signal period estimation and frequency estimation under low signal-to-noise ratio, this paper proposes a frequency hopping signal parameter estimation method based on maximum entropy binarization time-frequency graph and ABF-YOLOv5s, which includes the following steps:

[0005] S1. Obtain a frequency hopping signal and perform short-time Fourier transform on the frequency hopping signal to obtain a time-frequency graph;

[0006] The time-frequency graph is then subjected to maximum entropy binarization processing; multi-scale morphological filtering is then used to fill the holes in the frequency-hopping signal after maximum entropy binarization processing, remove the points or lines corresponding to the noise signal, better highlight the outline and edge of the frequency-hopping signal, thereby eliminating the influence of noise and obtaining a clear maximum entropy binarized time-frequency graph;

[0007] S2. Input the maximum entropy binarized time-frequency graph into the ABF-Yolov5s model to obtain the frequency hopping signal position information:

[0008] The ABF-Yolov5s model includes an input part, an ABF-Backbone part, an ABF-Neck part, and a Detection part; the ABF-Yolov5s model is improved on the basis of the Yolov5s model;

[0009] The ABF-Backbone part and the ABF-Neck part correspond to the Backbone part and the Neck part of the Yolov5s model respectively; the Detection part of the ABF-Yolov5s model has the same structure as the Detection part of the Yolov5s model; the input part of the ABF-Yolov5s model has the same structure as the input part of the Yolov5s model;

[0010] The SPP module of the Backbone part of the Yolov5s model is replaced with an ASPP module; the last C3 module of the Backbone part of the Yolov5s model is replaced with a BOT3 module to obtain the ABF-Backbone part; the BOT3 module is set after the ASPP module; the four PANET modules of the Neck part of the Yolov5s model are respectively replaced with BiFPN modules to obtain the ABF-Neck part;

[0011] The Detection part obtains the frequency hopping signal position information through three 1×1 convolutions;

[0012] S3. Estimating the period and frequency of the frequency hopping signal based on the relationship between the frequency hopping signal position information and the time-frequency parameters, wherein the time-frequency parameters are extracted from the time-frequency diagram obtained in step S1.

[0013] Preferably, in S1, the maximum entropy binarization process includes: performing binarization process on the time-frequency graph using maximum entropy threshold segmentation; the maximum entropy threshold segmentation refers to obtaining an optimal threshold value to maximize the sum of entropies of the background and the object.

[0014] Preferably, in S1, the mathematical model of the frequency hopping signal is:

[0015]

[0016] Among them, T represents the frequency hopping period, rect T represents a rectangular window with a width of T, τ represents the jump time, and f k represents the hopping frequency, M represents the total number of hopping frequencies within the observation time, j is the imaginary unit, and v(t) represents Gaussian white noise;

[0017] The method of performing short-time Fourier transform on the frequency hopping signal to obtain a time-frequency graph comprises the following steps:

[0018] The STFT spectrum is obtained according to the following formula:

[0019]

[0020] Where h(n) represents the window function, k = 1, 2, 3, ..., K represents the discrete frequency point, n = 1, 2, 3, ..., N represents the discrete time point, and s is the frequency hopping signal;

[0021] The time-frequency diagram is obtained from the STFT spectrum diagram according to the following formula:

[0022] SPCE s (k,n)=|STFT s (k,n)| 2

[0023] Preferably, in S1, the calculation of the maximum entropy threshold comprises the following steps:

[0024] The entropy of a digital image is defined as:

[0025]

[0026] Where I represents an image with N pixels; u represents the average grayscale value of pixels below the threshold t, corresponding to the grayscale value of the background area; v represents the average grayscale value of pixels above the threshold t, corresponding to the grayscale value of the object area; p represents the probability; g = I(u,v); log b(x) represents the logarithm with base b. Assuming the threshold is t, A is the grayscale distribution of {0, 1, 2, ..., t}, and B is the grayscale distribution of {t+1, t+2, ..., L-1}, then the entropies of A and B are H(A) and H(B), respectively, which are calculated as follows;

[0027]

[0028]

[0029] Among them, p n and 1-p n denote the cumulative probabilities of background and object pixels segmented by t threshold, respectively;

[0030] Define the objective function is the sum of H(A) and H(B):

[0031]

[0032] The maximum entropy threshold is calculated as follows:

[0033]

[0034] Preferably, in S2, the ASPP module is used to increase the receptive field while keeping the resolution of the maximum entropy binarized time-frequency graph unchanged, and fully extract the global features and local features of the maximum entropy binarized time-frequency graph; the global features include the shape contour features of the maximum entropy binarized time-frequency graph; the local features include the corner points and edge features of the maximum entropy binarized time-frequency graph.

[0035] Preferably, in S2, the BOT3 module is used to focus the ABF-Yolov5s model on the region where the frequency hopping signal is located in the maximum entropy binarized time-frequency graph through a multi-head self-attention mechanism. Paying attention to information at different locations of the frequency hopping signal in the maximum entropy binarized time-frequency graph helps the model focus on the region where the frequency hopping signal is located, thereby improving the accuracy of frequency hopping signal positioning.

[0036] Preferably, in S2, the BiFPN module introduces a bidirectional connection to fuse the multi-scale shallow graphic features output by the c3 module in the ABF-Backbone and the deep semantic features output by the BOT3 module for distinguishing the frequency hopping signal and the background to obtain a feature map, thereby improving the edge detection accuracy of the frequency hopping signal while improving the positioning accuracy of the frequency hopping signal and reducing the errors in the period estimation and frequency estimation of the frequency hopping signal.

[0037] Preferably, the step S3 specifically includes the following steps:

[0038] The frequency hopping signal position information is specifically the normalized coordinates (x1, y1) (x2, y2) of the frequency hopping signal in the maximum entropy binarized time-frequency diagram;

[0039] Where x1 is the starting position of the frequency hopping signal, x2 is the ending position of the frequency hopping signal; y1 is the upper limit of the frequency hopping signal, and y2 is the lower limit of the frequency hopping signal. The period estimate T and frequency estimate F of the frequency hopping signal are calculated as follows:

[0040] T = (x3 - x1) × t_size

[0041] F=(y1+y2) / 2×f_size

[0042] Here, x3 refers to the starting position of the next hop signal, t_size refers to the maximum value of the time axis in the time-frequency graph, and f_size refers to the maximum value of the frequency axis in the time-frequency graph.

[0043] Then, the trimmed mean of the results calculated in each time-frequency diagram is calculated, and finally the mean of the time-frequency diagram results under the same signal-to-noise ratio is calculated again to obtain the period and frequency of the frequency hopping signal estimated by the ABF-Yolov5s network.

[0044] Preferably, the value of t_size is 0.02s, and the value of f_size is 2.5×10 5 Hz.

[0045] Preferably, the value of b is 2.

[0046] The beneficial effects of the present invention are as follows:

[0047] 1. This paper proposes the ABF-Yolov5s model, which uses deep learning to solve the problem of large errors in period estimation and frequency estimation of frequency hopping signals in low signal-to-noise ratio environments.

[0048] 2. The time-frequency graph is binarized using the maximum entropy threshold segmentation method and combined with morphological filtering to eliminate the problem of reduced resolution of the time-frequency graph caused by noise.

[0049] 3. The ASPP module is introduced into the ABF-Backbone part to increase the receptive field while maintaining the resolution of the maximum entropy binarized time-frequency map, fully extracting global features such as the shape and contour of the frequency-hopping signal and local features such as corners and edges; the BOT3 module is introduced to focus on the information of the frequency-hopping signal at different positions in the maximum entropy binarized time-frequency map through the multi-head self-attention mechanism, helping the model focus on the area where the frequency-hopping signal is located, thereby improving the positioning accuracy of the frequency-hopping signal.

[0050] 4. The BiFPN module is introduced into the ABF-Neck part. Through bidirectional connections, it better fuses multi-scale shallow graphic features and deep semantic features used to distinguish frequency-hopping signals from backgrounds. Learnable weights are introduced to determine the importance of different input features, thereby enhancing the effect of frequency-hopping signal feature fusion. This improves the accuracy of frequency-hopping signal positioning while improving the edge detection accuracy of frequency-hopping signals, and reduces the errors in frequency-hopping signal period estimation and frequency estimation.

[0051] In summary, the present invention can effectively reduce the errors in frequency hopping signal period estimation and frequency estimation under low signal-to-noise ratio and has good robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 The images are processed with time-frequency graph (SNR = -1dB). In the figure, (a) is the original time-frequency graph, (b) is the grayscale image, (c) is the maximum entropy binarization image, and (d) is the image after morphological filtering.

[0053] Figure 2 The network structure of ABF-Yolov5s;

[0054] Figure 3 It is the ASPP structure;

[0055] Figure 4 It is the BOT3 structure and Bottleneck Transformer structure;

[0056] Figure 5 Comparison diagram of BiFPN and PANet structures;

[0057] Figure 6 This is the frequency hopping signal detection result based on Yolov5s;

[0058] Figure 7 This is the frequency hopping signal detection result based on ABF-Yolov5s;

[0059] Figure 8The experimental results of frequency hopping signal period estimation and frequency estimation are shown in the figure. In the figure, (a) is the frequency hopping signal period estimation error, and (b) is the frequency hopping signal frequency estimation error. From the experimental results, it can be seen that under low signal-to-noise ratio conditions, the frequency hopping signal period estimation and frequency estimation method based on maximum entropy binarization time-frequency diagram and ABF-Yolov5s proposed in this paper can effectively reduce the error. At -5dB, the error of the frequency hopping signal period estimation and frequency estimation method based on STFT is reduced by about 2 orders of magnitude; at -1dB, the error is also significantly reduced compared with the frequency hopping signal period estimation and frequency estimation method based on SPWVD; compared with the frequency hopping signal period estimation and frequency estimation method based on SSD, the method proposed in this paper also has a significant reduction in the frequency hopping signal period estimation error; compared with the original Yolov5s model, the ABF-Yolov5s proposed in this paper has higher detection accuracy and smaller frequency hopping signal period estimation and frequency estimation errors. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0061] The present invention first uses the maximum entropy threshold segmentation method to binarize the time-frequency graph and combines it with morphological filtering. This method eliminates noise while better highlighting the edge information of the frequency-hopping signal, thereby obtaining a clear maximum entropy binarized time-frequency graph. Secondly, this paper proposes the ABF-Yolov5s model. In ABF-Backbone, the ASPP module is used to increase the receptive field while maintaining the resolution of the maximum entropy binarized time-frequency graph. It fully extracts global features such as the shape and contour of the frequency-hopping signal and local features such as corners and edges. The BOT3 module introduces a multi-head self-attention mechanism, focusing on the information of the frequency-hopping signal at different locations in the maximum entropy binarized time-frequency graph, helping the model focus on the area where the frequency-hopping signal is located, thereby improving the accuracy of frequency-hopping signal positioning. In ABF-Neck, the BiFPN module introduces bidirectional connections, better integrating multi-scale shallow graphic features with deep semantic features used to distinguish the frequency-hopping signal from the background. This improves the accuracy of frequency-hopping signal positioning while improving the edge detection accuracy of the frequency-hopping signal and reduces the errors in the frequency-hopping signal period estimation and frequency estimation. Finally, the period estimation and frequency estimation of the frequency hopping signal are completed by comparing the signal coordinate position obtained by ABF-Yolov5s with the coordinates in the maximum entropy binarized time-frequency diagram.

[0062] The technical solution adopted by the present invention to solve the technical problem includes the following steps:

[0063] Step 1: Perform short-time Fourier transform on the frequency-hopping signal to obtain a time-frequency diagram, then perform maximum entropy binarization on the time-frequency diagram, and use multi-scale morphological filtering to fill the holes inside the binarized frequency-hopping signal, remove points or lines isolated from the useful signal, better highlight the outline and edge of the frequency-hopping signal, thereby eliminating the influence of noise and obtaining a clear maximum entropy binarized time-frequency diagram.

[0064] Step 2: Input the clear maximum entropy binarized time-frequency map into the ABF-Yolov5s model to estimate the frequency-hopping signal parameters. The ASPP module is used in the ABF-Backbone to increase the receptive field while maintaining the resolution of the maximum entropy binarized time-frequency map. This allows the model to fully extract global features such as the shape and contour of the frequency-hopping signal, as well as local features such as corners and edges. At the end of the ABF-Backbone, the BOT3 module uses a multi-head self-attention mechanism to focus on information at different locations of the frequency-hopping signal in the maximum entropy binarized time-frequency map, helping the model focus on the region where the frequency-hopping signal resides, thereby improving the accuracy of frequency-hopping signal positioning.

[0065] Step 3: Use the BiFPN module in the ABF-Neck part and introduce bidirectional connections to better integrate multi-scale shallow graphic features with deep semantic features used to distinguish between frequency hopping signals and backgrounds. This improves the accuracy of frequency hopping signal positioning while improving the edge detection accuracy of frequency hopping signals and reduces the errors in frequency hopping signal period estimation and frequency estimation.

[0066] Step 4: Finally, the frequency hopping signal position information is obtained at the head end through three 1×1 convolutions. According to the correlation between the signal coordinate position and the time-frequency parameters, the period estimation and frequency estimation of the frequency hopping signal can be obtained.

[0067] Step 1: Frequency hopping signal processing process is as follows:

[0068] Frequency hopping signal is a non-stationary signal whose frequency keeps changing under the control of pseudo-random sequence. Its mathematical model is:

[0069]

[0070] Among them, T represents the frequency hopping period, rect T represents a rectangular window with a width of T, τ represents the jump time, and f k represents the hopping frequency, M represents the total number of hopping frequencies within the observation time, j is the imaginary unit, and v(t) represents Gaussian white noise.

[0071] By performing short-time Fourier transform on the signal in different time windows, the change of the signal in time and frequency can be obtained. To facilitate computer calculation, discretization can be performed to obtain the STFT spectrum:

[0072]

[0073] Wherein, h(n) represents a window function, k=1, 2, 3, ..., K represents a discrete frequency point, n=1, 2, 3, ..., N represents a discrete time point, and s is a frequency hopping signal.

[0074] The time-frequency graph is defined as the square of the STFT modulus value, which has good time-frequency aggregation and its expression is:

[0075] SPCE s (k,n)=|STFT s (k,n)| 2 (3)

[0076] The time-frequency image is binarized using the maximum entropy threshold segmentation. The principle of maximum entropy thresholding is similar to the Otsu method. The Otsu method is to find the optimal threshold to divide the image into background and object to obtain the maximum variance, while the maximum entropy is to find the optimal threshold to maximize the sum of the entropy of the background and object. The definition of entropy in digital images is:

[0077]

[0078] Where I represents an image with N pixels; u represents the average grayscale value of pixels below the threshold t, which usually corresponds to the grayscale value of the background area; v represents the average grayscale value of pixels above the threshold t, which usually corresponds to the grayscale value of the object area; p represents the probability; g = I(u,v); log b (x) represents the logarithm with base b, where b is usually 2. Assuming the threshold is t, A is the grayscale distribution of {0, 1, 2, ..., t}, and B is the grayscale distribution of {t+1, t+2, ..., L-1}, then the entropies associated with the two probability densities are H(A) and H(B), respectively, as shown in Equations 6 and 5.

[0079]

[0080]

[0081] Among them, p n and 1-p n denote the cumulative probabilities of background and object pixels segmented by t threshold, respectively.

[0082] Define the function as the sum of H(A) and H(B):

[0083]

[0084] Then the maximum entropy threshold is:

[0085]

[0086] By binarizing the time-frequency graph using this threshold, the edge information of the frequency hopping signal can be retained as much as possible while removing noise. However, due to the influence of noise, the binarized time-frequency graph still has some problems. The influence of noise can be further removed through multi-scale morphological filtering.

[0087] Due to the influence of noise, after the time-frequency graph is converted into a binary image, there may be points or lines that are isolated from the useful signal image; or there may be burrs around the useful signal image; or there may be holes inside the useful signal image. Therefore, morphological filtering is needed to repair the useful signal image. Erosion and dilation operations are the most basic operations in morphology. Erosion operations are mainly used to remove small isolated points and broken lines in the image; dilation operations are mainly used to fill holes in the image and connect broken lines. The operation of first erosion and then dilation is called a closing operation, which is defined as follows: where X represents the image to be processed, B represents the structural element of the image, is the expansion operation, Θ is the corrosion operation:

[0088]

[0089] Morphological filtering extracts local image features by performing operations on an image using a set structuring element. Therefore, the most important aspect of morphological filtering is the selection of the structuring element, including its shape and size. The shape of the element affects the recognition and segmentation of objects of varying shapes within the image. Choosing an appropriate element shape can better highlight the outlines or edges of the object. The size of the element determines the granularity of the filter. Smaller elements are typically used to refine the image and better preserve details, while larger elements are typically used to smooth the image, emphasize object outlines, or fill holes. Because frequency-hopping signals are generally rectangular, this paper first uses a 10×10 rectangular element for closing operations. This closes the holes left by noise during the binarization of the frequency-hopping signal and better highlights the edges of the signal, thereby improving the accuracy of ABF-Yolov5s localization of the signal. A 5-pixel horizontal line element is then created to remove points or lines that are isolated from the useful signal image after the time-frequency graph is converted to a binary image, preserving image details. After binarization and morphological filtering, the influence of noise can be eliminated, a clear time-frequency image can be obtained, and the characteristics of the frequency hopping signal can be retained, which is conducive to the subsequent use of ABF-Yolov5s to estimate the period and frequency of the frequency hopping signal.

[0090] Step 2: The ABF-Backbone part extracts the frequency hopping signal features as follows:

[0091] In the ABF-Backbone network structure, the Focus module generates a 320×320×32 frequency hopping signal feature map. Then, through the deep convolutional network constructed by the C3 module and the Conv module, the feature map will become 1 / 2 of its original size after each combination of Conv and C3 modules, thereby obtaining frequency hopping signal graphic features of different scales. Then, the ASPP module uses parallel sampling with dilated convolution at different sampling rates to fully extract the global and local features of the frequency hopping signal and perform preliminary interaction. Finally, the BOT3 module is used to extract the semantic features of the frequency hopping signal in the deep network.

[0092] ASPP module:

[0093] ASPP uses dilated convolutions with different dilation rates for parallel sampling, including two 1×1 convolutional layers, multiple 3×3 convolutional layers with different dilation rates, and a pooling layer. The pooled feature information is restored to the expected size of the maximum entropy binarized time-frequency map through 1×1 convolution and upsampling. Finally, all the obtained results are concatenated to expand the number of channels, and then the number of channels is reduced to the expected value through 1×1 convolution for output, thereby increasing the receptive field while maintaining the resolution of the feature map. Multiple parallel convolutional layers are used to capture frequency hopping signal features of different scales, fully extract the global and local features of the frequency hopping signal, and perform preliminary interaction. Since global features can reflect the overall properties of the frequency hopping signal, such as the shape and contour, while local features focus more on specific key points such as corners or edges, their interaction can complement each other's advantages, improve the accuracy of frequency hopping signal detection, and thus reduce the errors in frequency hopping signal period estimation and frequency estimation.

[0094] BOT3 module:

[0095] This paper uses BOT3 at the Backbone end because the frequency-hopping signal feature map here is of lower resolution and contains more high-level semantic information. The stronger the semantic information, the better the model's ability to distinguish the frequency-hopping signal from the background. However, the deeper the network, the more ambiguous the location information of the frequency-hopping signal. The most important structure in the BOT3 module is the multi-head self-attention mechanism (MHSA). It uses multiple different attention heads to perform parallel calculations on the input image to obtain spatially sensitive similar features, making it easier for the model to focus on the area where the frequency-hopping signal is located, avoiding the loss of frequency-hopping signal location information. At the same time, the semantic features in the deep network are used to maximize the distinction between the frequency-hopping signal and the background. They can also be integrated with the graphical features output by the shallow network to improve the model's detection and location accuracy of the frequency-hopping signal and reduce the errors in the frequency-hopping signal period and frequency estimation.

[0096] Step 3: The ABF-Neck part integrates the frequency hopping signal characteristics as follows:

[0097] The ABF-Neck network is primarily responsible for fusing frequency-hopping signal feature maps at different scales. It consists of a C3 module, an upsampling module, a BiFPN module, and a Conv module. In the left-hand path, features at different levels are integrated through upsampling and fusion with a coarser-grained frequency-hopping signal feature map. In the right-hand path, features at different levels are integrated using convolutional layers. The resulting frequency-hopping signal feature map is used to detect and locate the frequency-hopping signal and estimate its period and frequency. In the ABF-Neck network, BiFPN is used to implement bidirectional feature propagation and multi-resolution feature processing, promoting more thorough and detailed cross-level information transfer and fusion. Specifically, the equivariant features (graphic features) used in the shallow network to identify frequency-hopping signal position and scale variations are better integrated with the invariant features (semantic features) used in the deep network to separate the frequency-hopping signal from the background. Learnable weights are introduced to determine the importance of different input features, enhancing the effectiveness of frequency-hopping signal feature fusion, improving the accuracy of frequency-hopping signal detection and localization, and reducing errors in frequency-hopping signal period and frequency estimation.

[0098] BiFPN:

[0099] In neural network models, low-level frequency-hopping signal features often contain less semantic information but accurate signal locations, while high-level features contain rich semantic information but coarse signal locations. To improve the accuracy of frequency-hopping signal target detection and recognition, it is necessary to fuse these two features. BiFPN first removes input edge nodes that contribute less to the performance of fusing different frequency-hopping signal features. Secondly, it adds a skip connection between input and output nodes of the same scale to integrate more frequency-hopping signal image features. Unlike PANet, which has only one top-down and one bottom-up path, BiFPN treats each bidirectional (top-down and bottom-up) path as a feature network layer, and repeats the same layer multiple times to achieve higher-level frequency-hopping signal feature fusion. Finally, it introduces learnable weights to determine the importance of different input features, enhancing the effectiveness of frequency-hopping signal feature fusion, improving the accuracy of frequency-hopping signal positioning, and reducing errors in frequency-hopping signal period and frequency estimation.

[0100] Step 4: The frequency hopping signal period estimation and frequency estimation method are as follows:

[0101] After running the ABF-Yolov5s network, we can obtain the model confidence level and the normalized coordinates (x1, y1)(x2, y2) of the frequency-hopping signal. Here, x1 is the starting position of the frequency-hopping signal, and x2 is the ending position; y1 is the upper frequency limit of the frequency-hopping signal, and y2 is the lower frequency limit of the frequency-hopping signal. By calculating the correlation between the signal coordinates and the time-frequency parameters, we can obtain the estimated period and frequency of the frequency-hopping signal.

[0102] T=(x3-x1)×t_size (9)

[0103] F=(y1+y2) / 2×f_size (10)

[0104] In formula 9, x3 refers to the starting position of the next hop signal, t_size refers to the maximum value of the time axis in the time-frequency diagram, which is 0.02s in this paper, and f_size refers to the maximum value of the frequency in the time-frequency diagram, which is 2.5×10 5 Hz. Then, the trimmed mean of the results calculated in each image is calculated, and finally the mean of the results of the images under the same signal-to-noise ratio is calculated again to obtain the period and frequency of the frequency hopping signal estimated by the ABF-Yolov5s network.

[0105] Example:

[0106] 1. Dataset

[0107] A frequency-hopping signal was generated using MATLAB. Frequency-shift keying (FSK) modulation was performed using a spread spectrum code sequence. Gaussian noise was added, resulting in signal-to-noise ratios of -5dB, -3dB, -1dB, 1dB, 3dB, and 5dB, respectively. The pseudocode length was 31, the code rate was 1k, the sampling rate was 248kHz, the FH period was 1ms, and the frequency set was [160, 200, 148, 208, 172, 116, 100, 180, 204, 192, 124, 140, 104, 136, 188, 164, 156, 184, 112, 176, 152, 108, 220, 144, 132, 128, 212, 120, 216, 196, 168]kHz.

[0108] First, a short-time Fourier transform (SFT) was used to generate a time-frequency plot of the frequency-hopping signal. 600 images were generated for each signal-to-noise ratio (SNR), totaling 3,600 images for the six SNRs. These images were then binarized and morphologically filtered in batches. The images were divided into training, test, and validation sets in a 4:1:1 ratio, with 400 images in the training set, 100 images in the test set, and 100 images in the validation set for each SNR. Labels were then created using Labelme, and the .png images were converted to .json format. Finally, code was used to convert them to .text format.

[0109] 2. Experimental Environment

[0110] The experimental server GPU was an NVIDIA TITAN RTX 3090, and the deep learning framework used was PyTorch. The model was implemented in Python and programmed using Sublime Text 3. Because model training on a CPU is relatively slow, we trained the network framework on a GPU. The GPU used was an NVIDIA TITAN RTX 3090, the model was trained for 100 epochs, and the cross-entropy loss function was used.

[0111] 3. Performance simulation

[0112] Figure 6 and Figure 7 The detection results of the original Yolov5s and ABF-Yolov5s for frequency hopping signals under the same signal-to-noise ratio are shown. Compared with the original Yolov5s model, the average confidence value is 0.66, and the average confidence value of ABF-Yolov5s is 0.92. The recognition and detection capabilities of frequency hopping signals have been greatly improved, thereby effectively reducing the period estimation and frequency estimation errors of the frequency hopping signal.

[0113] This paper compares the frequency hopping signal period estimation and frequency estimation based on ABF-Yolov5s with the frequency hopping signal period estimation and frequency estimation based on STFT, the frequency hopping signal period estimation and frequency estimation based on SPWVD, the frequency hopping signal period estimation and frequency estimation based on Yolov5s and the frequency hopping signal period estimation and frequency estimation based on SSD. The experimental results are as follows: Figure 8 shown.

[0114] Finally, it should be noted that the purpose of disclosing the embodiments is to facilitate a further understanding of the present invention. However, those skilled in the art will appreciate that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments; the scope of protection claimed by the present invention shall be determined by the scope defined in the claims.

Claims

1. A frequency hopping signal parameter estimation method based on maximum entropy binarized time-frequency graph and ABF-YOLOv5s, characterized in that: The following steps are involved: S1. Obtain a frequency hopping signal and perform short-time Fourier transform on the frequency hopping signal to obtain a time-frequency graph; The time-frequency graph is then subjected to maximum entropy binarization processing; multi-scale morphological filtering is then used to fill the holes inside the frequency-hopping signal after maximum entropy binarization processing, remove the points or lines corresponding to the noise signal, and obtain a clear maximum entropy binarization time-frequency graph; S2. Input the maximum entropy binarized time-frequency graph into the ABF-Yolov5s model to obtain the frequency hopping signal position information: The ABF-Yolov5s model includes an input part, an ABF-Backbone part, an ABF-Neck part, and a Detection part; The ABF-Backbone part and the ABF-Neck part correspond to the Backbone part and the Neck part of the Yolov5s model respectively; the Detection part of the ABF-Yolov5s model has the same structure as the Detection part of the Yolov5s model; the input part of the ABF-Yolov5s model has the same structure as the input part of the Yolov5s model; The SPP module of the Backbone part of the Yolov5s model is replaced with an ASPP module; the last C3 module of the Backbone part of the Yolov5s model is replaced with a BOT3 module to obtain the ABF-Backbone part; the BOT3 module is set after the ASPP module; the four PANET modules of the Neck part of the Yolov5s model are respectively replaced with BiFPN modules to obtain the ABF-Neck part; The Detection part obtains the frequency hopping signal position information through three 1×1 convolutions; S3. Estimating the period and frequency of the frequency hopping signal based on the relationship between the frequency hopping signal position information and the time-frequency parameters, wherein the time-frequency parameters are extracted from the time-frequency diagram obtained in step S1.

2. The method for estimating frequency hopping signal parameters based on maximum entropy binarization time-frequency graph and ABF-YOLOv5s according to claim 1, wherein: In S1, the maximum entropy binarization process includes: performing binarization process on the time-frequency graph using maximum entropy threshold segmentation; the maximum entropy threshold segmentation refers to finding the optimal threshold to maximize the sum of entropies of the background and the object.

3. The method for estimating frequency hopping signal parameters based on maximum entropy binarization time-frequency graph and ABF-YOLOv5s according to claim 2, wherein: In S1, the mathematical model of the frequency hopping signal is: Among them, T represents the frequency hopping period, rect T represents a rectangular window with a width of T, τ represents the jump time, and f k represents the hopping frequency, M represents the total number of hopping frequencies within the observation time, j is the imaginary unit, and v(t) represents Gaussian white noise; The method of performing short-time Fourier transform on the frequency hopping signal to obtain a time-frequency graph comprises the following steps: The STFT spectrum is obtained according to the following formula: Where h(n) represents the window function, k = 1, 2, 3, ..., K represents the discrete frequency point, n = 1, 2, 3, ..., N represents the discrete time point, and s is the frequency hopping signal; The time-frequency diagram is obtained from the STFT spectrum diagram according to the following formula: SPCEs(k,n)=|STFTs(k,n)| 2 。 4. The method for estimating frequency hopping signal parameters based on maximum entropy binarization time-frequency graph and ABF-YOLOv5s according to claim 3, wherein: In S1, the calculation of the maximum entropy threshold comprises the following steps: The entropy of a digital image is defined as: Where I represents an image with N pixels; u represents the average grayscale value of pixels below the threshold t, corresponding to the grayscale value of the background area; v represents the average grayscale value of pixels above the threshold t, corresponding to the grayscale value of the object area; p represents the probability; g = I(u,v); log b (x) represents the logarithm with base b. Assuming the threshold is t, A is the grayscale distribution of {0, 1, 2, ..., t}, and B is the grayscale distribution of {t+1, t+2, ..., L-1}, then the entropies of A and B are H(A) and H(B), respectively, which are calculated as follows; Among them, p n and 1-p n denote the cumulative probabilities of background and object pixels segmented by t threshold, respectively; Define the objective function is the sum of H(A) and H(B): The maximum entropy threshold is calculated as follows:

5. The method for estimating frequency hopping signal parameters based on maximum entropy binarization time-frequency graph and ABF-YOLOv5s according to claim 1, wherein: In S2, the ASPP module is used to extract global features and local features of the maximum entropy binarized time-frequency graph; the global features include shape contour features of the maximum entropy binarized time-frequency graph; the local features include corner points and edge features of the maximum entropy binarized time-frequency graph.

6. The method for estimating frequency hopping signal parameters based on maximum entropy binarization time-frequency graph and ABF-YOLOv5s according to claim 1, wherein: In S2, the BOT3 module is used to focus the ABF-Yolov5s model on the area where the frequency hopping signal is located in the maximum entropy binarized time-frequency graph through a multi-head self-attention mechanism.

7. The method for estimating frequency hopping signal parameters based on maximum entropy binarized time-frequency graph and ABF-YOLOv5s according to claim 1, wherein: In S2, the BiFPN module is used to fuse the multi-scale shallow graphic features output by the c3 module in the ABF-Backbone and the deep semantic features output by the BOT3 module for distinguishing the frequency hopping signal and the background to obtain a feature map.

8. The method for estimating frequency hopping signal parameters based on maximum entropy binarization time-frequency graph and ABF-YOLOv5s according to claim 1, wherein: The S3 specifically includes the following steps: The frequency hopping signal position information is specifically the normalized coordinates (x1, y1) (x2, y2) of the frequency hopping signal in the maximum entropy binarized time-frequency diagram; Where x1 is the starting position of the frequency hopping signal, x2 is the ending position of the frequency hopping signal; y1 is the upper limit of the frequency hopping signal, and y2 is the lower limit of the frequency hopping signal. The period estimate T and frequency estimate F of the frequency hopping signal are calculated as follows: T = (x3 - x1) × t_size F=(y1+y2) / 2×f_size Here, x3 refers to the starting position of the next hop signal, t_size refers to the maximum value of the time axis in the time-frequency graph, and f_size refers to the maximum value of the frequency axis in the time-frequency graph. Then, the trimmed mean of the results calculated in each time-frequency diagram is calculated, and finally the mean of the time-frequency diagram results under the same signal-to-noise ratio is calculated again to obtain the period and frequency of the frequency hopping signal estimated by the ABF-Yolov5s network.

9. The method for estimating frequency hopping signal parameters based on maximum entropy binarization time-frequency graph and ABF-YOLOv5s according to claim 8, wherein: The value of t_size is 0.02s, and the value of f_size is 2.5×10 5 Hz.

10. The method for estimating frequency hopping signal parameters based on maximum entropy binarized time-frequency graph and ABF-YOLOv5s according to claim 4, wherein: The value of b is 2.

Images