A Feature Extraction Method for Frequency-Hopping Signals with Multilevel Feature Compression

Abstract

The present invention discloses a method for extracting the characteristics of a frequency-hopping signal with multi-level feature compression. The present invention has strong anti-noise and anti-interference capabilities and still has good effects when the signal is weak. Since the frequency-hopping signals within one period often have different modulation methods and different signal strengths, and in a complex electromagnetic environment, noise and interference will seriously affect the calculation of signal parameters, the traditional single-threshold feature extraction method is not applicable. If the threshold is too low, noise will be introduced, and if the threshold is too high, the signal will be incomplete or missing. The present invention uses a segmented filtering method to perform feature extraction at multiple levels. While ensuring the complete extraction of signal features, the features are cleaned and integrated, and parameter calculations are performed, resulting in stronger robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and particularly to a method for extracting hopping signal features with multi-level feature compression. Background Art

[0002] Since the invention of radio communication, it has played an important role in many fields and is now inseparable from human society. More and more scientific and technological personnel are engaged in the research and optimization of radio communication technology. Due to the characteristics of radio itself, the signal will be affected by factors such as weather, obstacles, and electromagnetic fields during transmission, resulting in signal attenuation or interference. These factors will seriously affect the effect of signal detection. In traditional radio signal detection, technicians usually need to manually confirm whether there is a hopping signal in the frequency band and use traditional signal analysis methods to confirm information such as the frequency set and hopping pattern of the signal. In the case of weak signals, interference, or strong noise, the parameter estimation will be inaccurate.

[0003] In recent years, computer vision technology has developed maturely and has been applied to more and more industries with good results. Combining computer vision-related technologies with radio signal detection has become the choice of more and more signal researchers. However, due to the characteristics of hopping signals, the hopping signals within the same hopping pattern may have different modulation methods and different signal strengths, which will greatly increase the complexity of calculation and processing. Traditional single feature extraction methods often lead to signal omission, incompleteness, etc. Therefore, feature extraction needs to be carried out at different levels to ensure the correctness of signal parameter calculation. Summary of the Invention

[0004] Aiming at the above deficiencies in the prior art, the method for extracting hopping signal features with multi-level feature compression provided by the present invention solves the problems of signal omission and incompleteness caused by traditional single feature extraction methods.

[0005] In order to achieve the above invention purpose, the technical solution adopted by the present invention is: a method for extracting hopping signal features with multi-level feature compression, characterized by including the following steps:

[0006] S1. Perform Fourier transform on the input signal IQ data to obtain a signal spectrogram;

[0007] S2. Perform clustering analysis on the signal spectrogram with respect to amplitude to obtain the amplitude value with concentrated noise distribution, and calculate the maximum amplitude value of the spectrogram;

[0008] S3. Set the step size step, construct N high-pass filters with the amplitude value with concentrated noise distribution as the starting value and the maximum amplitude value of the spectrogram as the ending value;

[0009] S4. Use high-pass filters with N different thresholds to process the signal spectrogram, setting the amplitude values greater than or equal to the threshold in the signal spectrogram to 1 and the others to 0, obtaining N two-dimensional feature matrices;

[0010] S5. Default the positions with amplitude value 1 in the feature matrix as signals, perform time continuity analysis on the feature matrix to obtain a set of linear signals;

[0011] S6. Perform clustering analysis on the lengths of the linear signals in the set of linear signals to obtain the linear signal with the most concentrated length distribution, and take this linear signal as the reference signal, with the length of this reference signal being L1;

[0012] S7. Select the linear signals in the set of linear signals with lengths in the range of 0.8*L1 to 1.2*L1. If the overall duty cycle is less than the threshold, set the length of the reference signal to 0;

[0013] S8. Perform clustering analysis on the lengths of all reference signals in the N feature matrices to obtain the signal length L2 with the most concentrated length distribution. Combine the sets of linear signals in the N feature matrices, and remove and clean the linear signals outside the range of 0.8*L2 to 1.2*L2;

[0014] S9. Perform frequency continuity analysis on the cleaned linear signals to obtain a set of rectangular signals, and the position information of the rectangular signals is the position information of the frequency hopping signal in the spectrogram;

[0015] S10. Output the feature matrix of the set of rectangular signals.

[0016] Further: The threshold of the high-pass filter is: starting from the amplitude value where the noise distribution is concentrated as the starting value, using the maximum amplitude value of the spectrogram as the ending value, setting the step size as step, and taking the single-step result as the threshold of the high-pass filter.

[0017] Further: The height and width of the two-dimensional feature matrix are both the same as the height and width of the signal spectrogram.

[0018] Further: The specific steps of the time continuity analysis in step S5 are as follows:

[0019] S51. When four consecutive signals appear in each row of the feature matrix, retain the feature matrix; otherwise, set the feature matrix to 0. Perform time continuous signal processing on the retained feature matrices to obtain linear signals;

[0020] S52. Analyze the adjacent linear signals in the same row. If the signal adjacent interval is less than 3 or one-tenth of the total signal length, perform linear signal merging.

[0021] Furthermore, the specific steps of the frequency continuity analysis in step S9 are as follows:

[0022] S91. Randomly find a linear signal for rectangular signal initialization;

[0023] S92. Determine whether the union-intersection ratio of the rectangular signal and the adjacent row linear signal is greater than 0.7. If so, merge the rectangular signal and the linear signal;

[0024] S93. Return to step S92 until the rectangular signal cannot be merged further, and then enter step S94;

[0025] S94. Return to step S91 until there is no linear signal, and obtain a set of rectangular signals.

[0026] Furthermore, the elements of the feature matrix are the starting frequency position, ending frequency position, starting time position, and ending time position of the rectangular signal.

[0027] The beneficial effects of the present invention are as follows:

[0028] 1. The present invention has strong anti-noise and anti-interference capabilities and still has good effects when the signal is weak. Since the frequency-hopping signals within one period often have different modulation methods and different signal strengths, and in a complex electromagnetic environment, noise and interference will seriously affect the calculation of signal parameters. The traditional single-threshold feature extraction method is not applicable. If the threshold is too low, noise will be introduced, and if the threshold is too high, the signal will be incomplete or missing; while the present invention uses a segmented filtering method to perform feature extraction at multiple levels. While ensuring the complete extraction of signal features, the features are cleaned and integrated, and parameter calculations are performed, with stronger robustness.

[0029] 2. The present invention has a fast calculation speed. When the signal sampling rate is high or the signal duration is too long, the number of parameters in the signal spectrogram will be extremely large, easily reaching a scale of hundreds of millions of parameters, which will greatly increase the calculation time overhead during calculation and analysis; the present invention uses a matrix processing method, with faster execution speed and efficiency. At the same time, the matrix is subjected to phased feature compression, from the initial spectral Figure 2 dimensional matrix to the two-dimensional matrix of linear signals, and then to the two-dimensional matrix of rectangular signals at the end. The matrix parameters decrease exponentially, and the calculation efficiency is higher.

[0030] 3. The parameter calculation of the present invention has high accuracy and credibility. Due to the influence of a complex electromagnetic environment, there are often certain deviations in signal parameters within different frequency-hopping periods, such as starting frequency, ending frequency, duration, etc. In this frequency-hopping signal parameter calculation method, density clustering is widely used, which can find the position where the parameter distribution is the most concentrated, making the parameter calculation more accurate and reasonable. Description of the Drawings

[0031] Figure 1 This is the overall calculation flow diagram of the present invention;

[0032] Figure 2 This is the structural diagram of the segmented filter of the present invention;

[0033] Figure 3 This is the schematic diagram of cleaning linear signal data of the present invention;

[0034] Figure 4 This is the schematic diagram of hierarchical feature extraction and cleaning and merging of the present invention;

[0035] Figure 5 This is the schematic diagram of data flow of multi-level feature compression of the present invention. Specific embodiments

[0036] The following describes the specific embodiments of the present invention to facilitate those skilled in the art of the present technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art of the present technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

[0037] As Figure 1 shown, a method for extracting the features of a frequency-hopping signal with multi-level feature compression, characterized by comprising the following steps:

[0038] S1. Perform Fourier transform on the input signal IQ data to obtain a signal spectrogram;

[0039] S2. Perform clustering analysis on the signal spectrogram with respect to amplitude to obtain the amplitude value where the noise distribution is concentrated, and calculate the maximum amplitude value of the spectrogram;

[0040] S3. Set the step size step, and construct N high-pass filters with the amplitude value where the noise distribution is concentrated as the starting value and the maximum amplitude value of the spectrogram as the ending value;

[0041] S4. Use N high-pass filters with different thresholds to perform feature processing on the signal spectrogram, so that the amplitude values greater than or equal to the threshold in the signal spectrogram are set to 1, otherwise set to 0, to obtain N two-dimensional feature matrices; the height and width of the matrix are the same as those of the spectrogram, the X-axis represents time, and the Y-axis represents frequency;

[0042] S5. Default the positions with amplitude value 1 in the feature matrix as signals, and perform time continuity analysis on the feature matrix to obtain a set of linear signals;

[0043] S6. Perform clustering analysis on the lengths of the linear signals in the linear signal set to obtain the linear signal with the most concentrated length distribution, and use this linear signal as the reference signal, where the length of the reference signal is L1;

[0044] S7. Screen out the linear signals in the linear signal set with lengths in the range of 0.8*L1 to 1.2*L1. If the overall duty cycle is less than 0.5, set the reference signal length of this linear signal set to 0; otherwise, keep the reference signal length as L1;

[0045] S8. Perform clustering analysis on all the reference signal lengths in the N feature matrices to obtain the signal length L2 (greater than 0) with the most concentrated length distribution. Merge the linear signal sets in the N feature matrices, and remove and clean the linear signals outside the range of 0.8*L2 to 1.2*L2;

[0046] S9. Perform frequency continuity analysis on the cleaned linear signals to obtain a rectangular signal set, and the position information of the rectangular signals is the position information of the frequency hopping signals in the spectrogram;

[0047] S10. Output the feature matrix of the rectangular signal set.

[0048] In this specific embodiment, as Figure 1 shown, it is a schematic diagram of the overall process of the detection method of the present invention. First, perform Fourier transform on the input signal IQ data to obtain a signal spectrogram, perform clustering analysis on the signal spectrogram to obtain the concentrated amplitude of the noise distribution, construct N segmented filters and perform feature extraction. The method for constructing the segmented filters is as Figure 2 shown. Starting from the concentrated amplitude value of the noise distribution and ending with the maximum amplitude of the spectrogram, set the step size step, and use the single-step result as the threshold of the segmented filter. The result of the filter acting on the signal spectrogram is that if the amplitude in the spectrogram is greater than or equal to the threshold, it is 1, and if it is less than the threshold, it is set to 0. Finally, obtain N two-dimensional feature matrices. The height and width of the matrix are the same as those of the spectrogram. The X-axis represents time, and the Y-axis represents frequency; perform time continuity analysis on the feature matrices to obtain a linear signal set, and perform duty cycle analysis on it. The method is as Figure 3As shown in the figure, the linear signal set uses length clustering analysis to obtain the most concentrated length L1, which is used as the linear reference signal. The signals within the set are screened for lengths within the range of 0.8*L1 to 1.2*L1, and it is determined whether their overall duty cycle is greater than 0.5. If it is less than the threshold, the linear reference length is set to 0; length clustering analysis is performed on all reference signals greater than 0 to obtain the estimated signal length L2 (greater than 0). The linear signals of N feature matrices are merged, and at the same time, the linear signals outside the range of 0.8*L2 to 1.2*L2 are removed and cleaned; frequency continuity analysis is performed on the cleaned linear signals to obtain a rectangular signal set, and the position information of the rectangular signals is the position information of the frequency hopping signal in the spectrogram; finally, the feature matrix of the frequency hopping signal position is returned.

[0049] Among them, the processing method of the time continuity analysis algorithm is as follows:

[0050] Step 1: Process the feature matrix. If four signals are continuously connected in each row, they are retained; otherwise, they are set to 0, and linear signals are obtained by processing the continuous signals.

[0051] Step 2: Analyze adjacent linear signals in the same row. If the signal adjacent interval is less than 3 or one-tenth of the total signal length, the linear signals are merged.

[0052] Among them, the processing method of the frequency continuity analysis algorithm is as follows:

[0053] Step 1: Randomly find a linear signal for rectangular signal initialization.

[0054] Step 2: Determine whether the intersection-over-union ratio of the rectangular signal and the adjacent row linear signal is greater than 0.7. If the condition is satisfied, they are merged.

[0055] Step 3: Return to Step 2 until the rectangular signal cannot be merged continuously.

[0056] Step 4: Return to Step 1 until there are no linear signals, and a rectangular signal set is obtained.

[0057] The hierarchical feature extraction, cleaning, and merging process of this calculation method is as Figure 4 shown. The IQ data is transformed by Fourier transform to obtain a spectrogram matrix of size h*w. It is segmented and filtered using n filters to obtain n h*w feature matrices. After binarization and linear analysis, an n*k*3 feature matrix is obtained (k is the maximum value of the linear signals extracted from the binarized matrix, and zeros are filled for those with less than k extractions). Through frequency continuity analysis, a t*4 feature matrix is finally obtained. The elements of the feature matrix are the start frequency position, end frequency position, start time position, and end time position of the rectangular signal. The feature compression process is as Figure 5 shown.

Claims

1. A method for extracting the features of a frequency-hopping signal with multi-level feature compression, characterized in that The following steps are involved: S1. Perform Fourier transform on the input signal IQ data to obtain a signal spectrum diagram; S2. Performing a cluster analysis on the amplitude of the signal spectrum to obtain the concentrated amplitude value of the noise distribution and calculating the maximum amplitude value of the spectrum; S3, taking the noise distribution concentrated amplitude value as the starting value and the maximum amplitude value of the spectrum graph as the ending value, setting the step size step, and constructing N high-pass filters; S4, using N high-pass filters with different thresholds to perform feature processing on the signal spectrum graph, so that the amplitude values ​​in the signal spectrum graph that are greater than or equal to the threshold are set to 1, otherwise they are set to 0, to obtain N two-dimensional feature matrices; S5, taking the position with an amplitude value of 1 in the feature matrix as a signal by default, performing a time continuity analysis on the feature matrix, and obtaining a linear signal set; S6, performing cluster analysis on the lengths of the linear signals in the linear signal set to obtain a linear signal with the most concentrated length distribution, and using the linear signal as a reference signal, the length of the reference signal being L1; S7, filter out the linear signals with lengths ranging from 0.8*L1 to 1.2*L1 in the linear signal set, and if the overall time duty cycle thereof is less than the threshold, set the reference signal length to 0; S8, performing cluster analysis on all reference signal lengths in N feature matrices to obtain the signal length L2 with the most concentrated length distribution, merging the linear signal sets in the N feature matrices, and removing and cleaning the linear signals with lengths outside the range of 0.8*L2 to 1.2*L2; S9, performing frequency continuity analysis on the cleaned linear signal to obtain a rectangular signal set, wherein the position information of the rectangular signal is the position information of the frequency hopping signal in the spectrum diagram; S10. Output the characteristic matrix of the rectangular signal set.

2. The method for extracting the characteristics of a frequency-hopping signal with multi-level feature compression according to claim 1, wherein The threshold of the high-pass filter is: the noise distribution concentration amplitude value is used as the starting value, the maximum amplitude value of the spectrum graph is used as the ending value, the step length step is set, and the single-step result is used as the threshold of the high-pass filter.

3. The method for extracting the characteristics of a frequency-hopping signal with multi-level feature compression according to claim 1, characterized in that, The height and width of the two-dimensional feature matrix are respectively the same as the height and width of the signal spectrum graph.

4. The method for extracting the characteristics of a frequency-hopping signal with multi-level feature compression according to claim 1, wherein The specific steps of the time continuity analysis in step S5 are: S51, when four continuous signals appear in each row of the feature matrix, the feature matrix is ​​retained, otherwise the feature matrix is ​​set to 0, and time continuous signal processing is performed on the retained feature matrix to obtain a linear signal; S52, analyzing adjacent linear signals in the same line, and merging the linear signals if the interval between adjacent signals is less than 3 or one tenth of the total length of the signals.

5. The method for extracting the characteristics of a frequency-hopping signal with multi-level feature compression according to claim 1, wherein The specific steps of the frequency continuity analysis in step S9 are: S91, randomly find a linear signal to initialize the rectangular signal; S92, determining whether the parallel-crossing ratio of the rectangular signal and the adjacent row linear signal is greater than 0.7, if so, merging the rectangular signal with the linear signal; S93, return to step S92, until the rectangular signal can no longer be merged, and proceed to step S94; S94, return to step S91, until there is no linear signal, and obtain a rectangular signal set.

6. The method for extracting the characteristics of a frequency-hopping signal with multi-level feature compression according to claim 1, wherein The elements of the characteristic matrix are the starting frequency position, the ending frequency position, the starting time position and the ending time position of the rectangular signal.

Images