A method for calculating frequency hopping signal parameters based on image segmentation

Abstract

The present invention discloses a method for calculating frequency hopping signal parameters based on image segmentation. The present invention uses an image segmentation method based on deep learning to directly extract features from the signal spectrum graph, uses a large number of random samples containing frequency hopping signals for learning, has strong anti-noise capability, can effectively extract features even in the presence of interference, and has a good effect on weak signals. The present invention uses a deep learning method for denoising and anti-interference, and uses frequency hopping prior knowledge for processing and supplementation. The anti-noise and anti-interference processing time can be controlled within 1 second under various complex conditions.

Description

Technical Field

[0001] The invention relates to the technical field of signal calculation, and in particular to a frequency hopping signal parameter calculation method based on image segmentation. Background Art

[0002] Since its invention, radio communication has played an important role in many fields. It is now inseparable from human society. More and more scientific and technological personnel are engaged in the research and development and optimization of radio communication technology. Due to the characteristics of radio itself, it will be affected by factors such as weather, obstacles, and electromagnetic fields during the transmission process, causing the signal to attenuate or interfere during transmission. These factors will seriously affect the effect of signal detection. In traditional radio signal frequency hopping detection, technicians usually need to manually confirm whether there is a frequency hopping signal in the frequency band, and use traditional signal analysis methods to confirm the signal's frequency set, frequency hopping pattern and other information. This method is not very real-time, and in the case of weak signals, interference or strong noise, inaccurate parameter estimation will occur.

[0003] In recent years, deep learning technology has matured in the field of image processing and has been applied to more and more industries, achieving good results. Combining deep learning technology with radio signal detection has become a trend, but frequency hopping signals themselves have certain a priori rules, such as equal duration of single hops and periodicity. If the general deep learning signal detection algorithm is directly used for the detection of radio frequency hopping signals, it will directly conflict with the a priori rules due to false detection and incompleteness. Therefore, we need to use the strong feature extraction ability of deep learning and combine it with the characteristics of the signal itself to calculate the signal parameters more reasonably and effectively. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method for calculating frequency hopping signal parameters based on image segmentation to solve the problem of inaccurate parameter estimation.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a method for calculating frequency hopping signal parameters based on image segmentation, comprising the following steps:

[0006] S1. Obtain N frequency hopping signals, and record and save the frequency, start time and end time of the frequency hopping signals;

[0007] S2. Perform Fourier transform on the generated frequency hopping signal IQ data to obtain a signal spectrum diagram, map and calculate the frequency, start time and end time of the frequency hopping signal to obtain the position information of the frequency hopping signal in the spectrum diagram;

[0008] S3, constructing a two-dimensional feature matrix with the same height and width as the spectrum diagram, assigning 1 to the same position in the feature matrix according to the position information of the frequency hopping signal in the spectrum diagram, assigning 0 to other positions, and storing the feature matrix;

[0009] S4, uniformly slicing the spectrum graph and the feature matrix corresponding thereto to obtain spectrum graph samples and the feature matrix samples corresponding thereto, and making them into a data set;

[0010] S5. Use the deep learning image segmentation model to train the data set. After the model converges, obtain the deep learning image segmentation weight file;

[0011] S6. Perform Fourier transform on the predicted signal IQ data to obtain a predicted signal spectrum diagram, and perform uniform slicing to obtain M signal spectrum diagram samples;

[0012] S7, using the deep learning image segmentation model and the trained weight file, predicting M signal spectrum samples to obtain M two-dimensional feature matrices;

[0013] S8, concatenating the M two-dimensional feature matrices in the order of segmentation to obtain a signal spectrum feature matrix, and setting values ​​in the matrix greater than 0.5 to 1 and values ​​less than 0.5 to 0;

[0014] S9, treating the predicted signal corresponding to the position with a value of 1 in the signal spectrum feature matrix as a frequency hopping signal, performing a time continuity analysis on each row in the signal spectrum feature matrix to obtain a linear signal, and performing a frequency continuity analysis on the linear signal to obtain a rectangular signal;

[0015] S10, performing periodic analysis on all obtained rectangular signals to obtain a frequency hopping pattern of the frequency hopping signal;

[0016] S11. Perform frequency and time mapping calculations based on the position information of the signal in the frequency hopping pattern in the signal spectrum feature matrix to obtain the frequency hopping pattern parameter information.

[0017] Furthermore, the uniform slicing process is as follows: the length of each slice is 512, and when the length of a slice is less than 512, zero padding is performed.

[0018] Further: the specific steps of the time continuity analysis are:

[0019] A1. If four signals appear continuously in each row of the signal spectrum feature matrix, the row matrix is ​​retained, otherwise the row matrix is ​​set to 0, and the continuous signal is processed for time continuity to obtain a linear signal;

[0020] A2. Analyze adjacent linear signals in the same line. If the interval between adjacent signals is less than 3 or one tenth of the total length of the signal, merge the linear signals.

[0021] Further: The specific steps of the frequency continuity analysis are:

[0022] B1. Perform cluster analysis on the linear signal lengths to obtain the signal length with the most concentrated distribution as the benchmark length;

[0023] B2. Screen the linear signal. When the ratio of the difference between its length and the reference length to the reference length is less than 0.7, the linear signal is removed.

[0024] B3. Randomly find a linear signal to initialize the rectangular signal;

[0025] B4, determine whether the intersection-and-combination ratio of the rectangular signal and the adjacent linear signal is greater than 0.7, if so, merge the rectangular signal with the adjacent linear signal;

[0026] B5, return to step B4, until the rectangular signal can no longer be merged;

[0027] B6. Return to step B3 until no linear signal is found and a set of rectangular signals is obtained.

[0028] Further: The specific steps of the periodic analysis are:

[0029] C1. Extract the rectangular signal of the same frequency from the rectangular signal;

[0030] C2, calculate the hop counts of adjacent intervals of the same-frequency rectangular signal and put them into the hop count set;

[0031] C3. Perform cluster analysis on the hop count set, and the hop count with the most concentrated distribution is the hop count of a cycle;

[0032] C4. Perform cluster analysis on the starting height and ending height of the frequency hopping signals at the same position in different cycles to obtain the location information of the starting height and ending height of each hop with the most concentrated distribution;

[0033] C5. Perform cluster analysis on the duration of all rectangular signals to obtain the duration of a single frequency hopping signal. The analysis can obtain the position information of all frequency hopping signals within one cycle.

[0034] Further: the input of the deep learning image segmentation model is 512*512*1. First, four scales of convolution kernels of 1*1, 3*3, 5*5, and 7*7 are used for feature extraction and standardization. Then, the feature matrix is ​​superimposed and pooled to obtain a 256*256*256 feature matrix. The pooling result uses 256 3*3 convolution kernels for feature extraction.

[0035] The obtained feature matrix is ​​extracted and standardized using convolution kernels of four scales: 1*1, 3*3, 5*5, and 7*7. The feature matrix is ​​then superimposed and pooled. The pooling result is extracted using 512 3*3 convolution kernels to obtain a feature matrix of 128*128*1024.

[0036] The obtained feature matrix is ​​extracted and standardized using convolution kernels of four scales: 1*1, 3*3, 5*5, and 7*7. The feature matrix is ​​then superimposed and pooled to obtain a 64*64*2048 feature matrix. The pooled result is extracted using 256 3*3 convolution kernels.

[0037] Use 128 convolution kernels to deconvolve the feature matrix to obtain a 128*128*128 feature matrix; use 64 convolution kernels to deconvolve the feature matrix to obtain a 256*256*64 feature matrix; finally, use 1 convolution kernel to deconvolve the feature matrix to obtain a 512*512*1 feature matrix.

[0038] Furthermore: the frequency hopping pattern parameter information includes the start time, end time, center frequency and bandwidth information of each hop.

[0039] The beneficial effects of the present invention are:

[0040] 1. In the traditional calculation of radio frequency hopping signal parameters, it is inevitable to process the noise and interference in the signal, and the randomness of noise and interference often causes the traditional calculation method to be less effective when encountering new situations; the present invention uses an image segmentation method based on deep learning to directly extract features from the signal spectrum graph, uses a large number of random samples containing frequency hopping signals for learning, has strong anti-noise ability, can perform effective feature extraction even in the presence of interference, and has a good effect on weak signals.

[0041] 2. The present invention has a faster and more stable processing speed. In the traditional radio frequency hopping signal parameter calculation, noise and interference need to be processed, and a single anti-noise processing or anti-interference method is difficult to achieve satisfactory results, so generally more complex anti-noise and anti-interference processing is used, but this method directly leads to poor timeliness of calculation, and the calculation time is different when processing different situations. The present invention uses deep learning methods for denoising and anti-interference, and uses frequency hopping prior knowledge for processing and supplementation. The anti-noise and anti-interference processing time can be controlled within 1 second in various complex situations.

[0042] 3. The present invention is simple to update and maintain, has strong feasibility, and has a controllable iteration cycle. When the traditional frequency-hopping radio frequency-hopping signal parameter calculation encounters an emergency, a strategy will be added for the specific noise or interference encountered, but this method has high maintenance costs, unstable iteration cycles, and the calculation code will become more and more complicated, and the maintainability is not high; while in the present invention, only the encountered burst signal needs to be added to the data set for training, and the obtained model can be replaced, which is convenient for maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of the present invention;

[0044] Figure 2 This is a schematic diagram of the training of the deep learning image segmentation model for frequency hopping signals;

[0045] Figure 3 It is a schematic diagram of the calculation of radio frequency hopping signal parameters of the present invention;

[0046] Figure 4 This is a schematic diagram of the structure of the deep learning image segmentation neural network of the present invention;

[0047] Figure 5 The spectrum diagram shows the calculation results of the radio frequency hopping signal with a low signal-to-noise ratio according to the present invention. DETAILED DESCRIPTION

[0048] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0049] like Figure 1 As shown, a method for realizing radio signal detection based on deep learning, the specific method steps are:

[0050] S1. Access N (N>=10000) frequency hopping signals from the sensor, and record and save the frequency, start time, and end time information of the frequency hopping signals;

[0051] S2. Perform Fourier transform on the generated frequency hopping signal IQ data to obtain the signal spectrum diagram, calculate the corresponding frequency hopping signal frequency, start time and end time, and obtain the position information of the signal in the spectrum diagram (the coordinates of the upper left corner and the lower right corner of the signal).

[0052] S3, constructing a two-dimensional all-zero feature matrix with the same height and width as the spectrum graph, assigning 1 to the same position in the matrix according to the position information of the frequency hopping signal in the spectrum graph, and performing matrix storage;

[0053] S4. Divide the signal spectrogram and the corresponding two-dimensional matrix into equal parts of length 512. Pad with zeros if the length is less than 512 to obtain the spectrogram samples and their corresponding 01 feature matrix samples, and make them into a dataset.

[0054] S5. Use a deep learning convolutional image segmentation network to train the dataset. Stop training after the loss value converges to obtain the model weight file.

[0055] S6. Perform Fourier transform on the original IQ data of the predicted signal to obtain the signal spectrogram, and perform a slicing process with a width of 512. Pad with zeros if the width is less than 512 to obtain M signal spectrogram samples.

[0056] S7. Use the deep learning image segmentation network and the trained model weight file to predict the M signal spectrogram samples to obtain M two-dimensional feature matrices.

[0057] S8. Concatenate the M two-dimensional feature matrices in order to obtain the spectrogram feature matrix, and set the values greater than 0.5 to 1 and the values less than 0.5 to 0 in the matrix.

[0058] S9. Consider the positions where 1 is set in the matrix as the frequency hopping signals. The X-axis of the matrix represents the time of the signal, and the Y-axis represents the frequency of the signal. Analyze the signal continuity for each row in the spectrogram feature matrix to obtain the linear signals, and then analyze the frequency continuity of the linear signals to obtain the rectangular signals.

[0059] S10. Perform periodic analysis on all the obtained rectangular signals to obtain the frequency hopping pattern (the position information of the frequency hopping signals in a complete cycle).

[0060] S11. Through the position information of the signals in the matrix in the frequency hopping pattern, the start time, end time, center frequency, and bandwidth information of each hop in the frequency hopping pattern can be calculated.

[0061] In this specific embodiment, as Figure 2The figure shows a schematic diagram of the training process of the deep learning image segmentation model of the frequency hopping signal of the present invention. First, the IQ data of the frequency hopping signal is accessed, and the frequency and start and end time of the frequency hopping signal are saved; the Fourier transform of the accessed IQ data is performed to obtain the spectrum diagram of the signal, and the frequency of the frequency hopping signal and the start and end time are mapped and calculated to obtain the position information of the frequency hopping signal in the spectrum diagram; according to the spectrum diagram and the position information of the signal in the spectrum diagram, a 01 matrix is ​​established, and the height and width of the matrix are the same as the spectrum diagram, the position with the signal is set to 1, and the others are 0; the spectrum diagram and the 01 feature matrix are uniformly sliced, each slice length is 512, and zero is added if it is less than 512, and they are matched and a data set is established; a custom deep learning image segmentation model is used for training, and after the model converges, a deep learning image segmentation weight file is obtained.

[0062] In this specific embodiment, Figure 3 The figure shows a flow chart of the frequency hopping signal parameter calculation process of the present invention. First, the prediction signal is Fourier transformed to obtain the prediction signal spectrum, which is sliced ​​to obtain the spectrum sample, and then input into the trained deep learning image segmentation model for prediction to obtain its feature matrix, and the feature matrix is ​​spliced ​​in the order of slicing to obtain the signal spectrum feature matrix, and the matrix is ​​set to 1 if it is greater than 0.5, and set to 0 if it is less than 0.5. The X-axis of the matrix represents time, and the Y-axis represents frequency. The position of 1 represents the presence of a signal; the matrix is ​​analyzed for time continuity to obtain a linear signal, and the linear signal is analyzed for frequency continuity to obtain a rectangular signal, and the rectangular signal is analyzed for periodicity to obtain the frequency hopping pattern position information of the frequency hopping signal, and the frequency and time mapping calculation is performed on its position to obtain the frequency hopping pattern parameter information, Figure 5 This is a diagram showing the effect of calculating the parameters of a frequency hopping signal with a low signal-to-noise ratio. The white box is the position information of the frequency hopping pattern in the spectrum diagram.

[0063] The time continuity analysis algorithm processing method is:

[0064] Step 1: Process the feature matrix. If four signals are continuously connected in each row, they are retained. Otherwise, they are set to 0. The continuous signals are processed to obtain linear signals.

[0065] Step 2: Analyze adjacent linear signals in the same line. If the interval between adjacent signals is less than 3 or one tenth of the total length of the signals, merge the linear signals.

[0066] The frequency continuity analysis algorithm processing method is:

[0067] Step 1: Perform cluster analysis on the length of linear signals to obtain the length of the signal with the most concentrated distribution as the reference length of the signal;

[0068] Step 2: Screen the linear signals, and remove them when the ratio of the difference between their length and the reference length is less than 0.7;

[0069] Step 3, randomly find a linear signal to initialize the rectangular signal;

[0070] Step 4: determine whether the intersection-and-combination ratio of the rectangular signal and the adjacent row linear signal is greater than 0.7, and merge them if the condition is met;

[0071] Step 5: Return to step 4 until the rectangular signal can no longer be merged;

[0072] Step 6: Return to step 3 until there is no linear signal and a rectangular signal set is obtained;

[0073] The periodic analysis algorithm processing method is:

[0074] Step 1, extract the rectangular signal with the same frequency (frequency intersection and conjunction ratio is greater than 0.5) from the rectangular signal;

[0075] Step 2: Calculate the number of hops between adjacent intervals of the same-frequency rectangular signal and put them into a set;

[0076] Step 3: Perform cluster analysis on the hop count set, and obtain the hop count with the most concentrated distribution, which is the hop count of a cycle;

[0077] Step 4: Perform cluster analysis on the starting height and ending height of the frequency hopping signal at the same position in different periods to obtain the most concentrated starting height and ending height position information of each hop.

[0078] Step 5: Perform cluster analysis on the duration of all rectangular signals to obtain the duration of a single frequency hopping signal. Through the above information, the position information of all frequency hopping signals within a cycle can be obtained through analysis.

[0079] In this specific embodiment, the present invention designs an image segmentation network based on deep learning. The network structure is as follows: Figure 4As shown in the figure, the input of the network is 512*512*1. The network first uses convolution kernels of four scales, 1*1, 3*3, 5*5, and 7*7, to extract and standardize features, and then superimposes and pools the feature matrix to obtain a 256*256*256 feature matrix. The pooling result uses 256 3*3 convolution kernels for feature extraction; the obtained feature matrix uses convolution kernels of four scales, 1*1, 3*3, 5*5, and 7*7, to extract and standardize features, and then superimposes and pools the feature matrix. The pooling result uses 512 3*3 convolution kernels for feature extraction, and obtains a 128*128*1024 feature matrix; The obtained feature matrix is ​​extracted and standardized using convolution kernels of four scales: 1*1, 3*3, 5*5, and 7*7. The feature matrix is ​​then superimposed and pooled to obtain a 64*64*2048 feature matrix. The pooling result is extracted using 256 3*3 convolution kernels. The feature matrix is ​​deconvolved using 128 convolution kernels to obtain a 128*128*128 feature matrix. The feature matrix is ​​deconvolved using 64 convolution kernels to obtain a 256*256*64 feature matrix. Finally, the feature matrix is ​​deconvolved using 1 convolution kernel to obtain a 512*512*1 feature matrix.

Claims

1. A method for calculating frequency hopping signal parameters based on image segmentation. It is characterized in that The following steps are involved: S1. Obtain N frequency hopping signals, and record and save the frequency, start time and end time of the frequency hopping signals; S2. Perform Fourier transform on the generated frequency hopping signal IQ data to obtain a signal spectrum diagram, map and calculate the frequency, start time and end time of the frequency hopping signal to obtain the position information of the frequency hopping signal in the spectrum diagram; S3, constructing a two-dimensional feature matrix with the same height and width as the spectrum diagram, assigning 1 to the same position in the feature matrix according to the position information of the frequency hopping signal in the spectrum diagram, assigning 0 to other positions, and storing the feature matrix; S4, uniformly slicing the spectrum graph and the feature matrix corresponding thereto to obtain spectrum graph samples and the feature matrix samples corresponding thereto, and making them into a data set; S5. Use the deep learning image segmentation model to train the data set. After the model converges, obtain the deep learning image segmentation weight file; S6. Perform Fourier transform on the predicted signal IQ data to obtain a predicted signal spectrum diagram, and perform uniform slicing to obtain M signal spectrum diagram samples; S7, using the deep learning image segmentation model and the trained weight file, predicting M signal spectrum samples to obtain M two-dimensional feature matrices; S8, concatenating the M two-dimensional feature matrices according to the segmentation order to obtain a signal spectrum feature matrix, and setting the values ​​in the matrix greater than 0.5 to 1 and those less than 0.5 to 0; S9, treating the predicted signal corresponding to the position with a value of 1 in the signal spectrum feature matrix as a frequency hopping signal, performing a time continuity analysis on each row in the signal spectrum feature matrix to obtain a linear signal, and performing a frequency continuity analysis on the linear signal to obtain a rectangular signal; S10, performing periodic analysis on all obtained rectangular signals to obtain a frequency hopping pattern of the frequency hopping signal; S11. Perform frequency and time mapping calculations based on the position information of the signal in the frequency hopping pattern in the signal spectrum feature matrix to obtain the frequency hopping pattern parameter information.

2. The method for calculating frequency hopping signal parameters based on image segmentation according to claim 1, It is characterized in that The uniform slicing process is as follows: the length of each slice is 512, and when the length of a slice is less than 512, zero padding is performed.

3. The method for calculating frequency hopping signal parameters based on image segmentation according to claim 1, It is characterized in that The specific steps of the time continuity analysis are: A1. If four signals appear continuously in each row of the signal spectrum feature matrix, the row matrix is ​​retained, otherwise the row matrix is ​​set to 0, and the continuous signal is processed for time continuity to obtain a linear signal; A2. Analyze adjacent linear signals in the same line. If the interval between adjacent signals is less than 3 or one tenth of the total length of the signal, merge the linear signals.

4. The method for calculating frequency hopping signal parameters based on image segmentation according to claim 1, It is characterized in that The specific steps of the frequency continuity analysis are: B1. Perform cluster analysis on the linear signal lengths to obtain the signal length with the most concentrated distribution as the benchmark length; B2. Screen the linear signal. When the ratio of the difference between its length and the reference length to the reference length is less than 0.7, the linear signal is removed. B3. Randomly find a linear signal to initialize the rectangular signal; B4, determine whether the intersection-and-combination ratio of the rectangular signal and the adjacent linear signal is greater than 0.7, if so, merge the rectangular signal with the adjacent linear signal; B5, return to step B4, until the rectangular signal can no longer be merged; B6. Return to step B3 until no linear signal is found and a set of rectangular signals is obtained.

5. The method for calculating frequency hopping signal parameters based on image segmentation according to claim 1, It is characterized in that The specific steps of the periodic analysis are: C1. Extract the rectangular signal of the same frequency from the rectangular signal; C2, calculate the hop counts of adjacent intervals of the same-frequency rectangular signal and put them into the hop count set; C3. Perform cluster analysis on the hop count set, and the hop count with the most concentrated distribution is the hop count of a cycle; C4. Perform cluster analysis on the starting height and ending height of the frequency hopping signals at the same position in different cycles to obtain the location information of the starting height and ending height of each hop with the most concentrated distribution; C5. Perform cluster analysis on the duration of all rectangular signals to obtain the duration of a single frequency hopping signal. The analysis can obtain the position information of all frequency hopping signals within one cycle.

6. The method for calculating frequency hopping signal parameters based on image segmentation according to claim 1, It is characterized in that The input of the deep learning image segmentation model is 512*512*1. First, four convolution kernels of 1*1, 3*3, 5*5, and 7*7 are used to extract features and standardize them. Then, the feature matrix is ​​superimposed and pooled to obtain a 256*256*256 feature matrix. The pooling result uses 256 3*3 convolution kernels for feature extraction. The obtained feature matrix is ​​extracted and standardized using convolution kernels of four scales: 1*1, 3*3, 5*5, and 7*7. The feature matrix is ​​then superimposed and pooled. The pooling result is extracted using 512 3*3 convolution kernels to obtain a feature matrix of 128*128*1024. The obtained feature matrix is ​​extracted and standardized using convolution kernels of four scales: 1*1, 3*3, 5*5, and 7*7. The feature matrix is ​​then superimposed and pooled to obtain a 64*64*2048 feature matrix. The pooled result is extracted using 256 3*3 convolution kernels. Use 128 convolution kernels to deconvolve the feature matrix to obtain a 128*128*128 feature matrix; use 64 convolution kernels to deconvolve the feature matrix to obtain a 256*256*64 feature matrix; finally, use 1 convolution kernel to deconvolve the feature matrix to obtain a 512*512*1 feature matrix.

7. The method for calculating frequency hopping signal parameters based on image segmentation according to claim 1, It is characterized in that The frequency hopping pattern parameter information includes the start time, end time, center frequency and bandwidth information of each hop.

Images