A method for frequency-hopping drone signal recognition based on wireless spectrum analysis

Abstract

The present invention discloses a method for realizing frequency hopping unmanned aerial vehicle signal recognition based on wireless spectrum analysis. The method collects unmanned aerial vehicle image transmission data and frequency hopping remote control data. 50 sets of data are collected from each unmanned aerial vehicle as original data. The collected one-dimensional unmanned aerial vehicle time domain signal is converted into a two-dimensional time-frequency signal by short-time Fourier transform. The data is calibrated. A model framework is constructed. The features of the unmanned aerial vehicle signal are extracted by means of a multi-layer convolutional network of a VGG model. The features are extracted by reasonably designing the convolution kernel and the number of network layers. After training a stable model, the output result of the model is converted into the possible position information of each signal element of the frequency hopping unmanned aerial vehicle signal on the time axis prediction on the time-frequency diagram. The accurate signal element positions obtained by signal matching are detected one by one to ensure the independence of the signal elements. The data verified by the signal are sorted, the statistically redundant and repeatedly detected elements are removed, and the data are classified.

Description

Technical Field

[0001] The present invention relates to the field of mobile communications, and in particular to a method for realizing frequency-hopping unmanned aerial vehicle signal recognition based on wireless spectrum analysis. Background Art

[0002] With the continuous development of the drone industry, the abuse of drones has caused serious damage to public safety, personal privacy, etc. The spectrum signal-based solution can identify the specific drone model according to the differences in spectrum data of different types of drones, and it is easy to obtain the spectrum information of the drone. The jammer can be used to interfere with the frequency band where the drone is located to achieve the purpose of countermeasure. The method based on spectrum feature recognition is simpler and more effective than deep learning to add new models, but as the drone database increases, its design logic tends to be more complicated. At present, the core task of this technology is to solve the problem of simple and fast addition of new drone models while ensuring low false detection rate and high sensitivity. The overall code parameter quantity and speed should not change significantly. The method of detecting frequency hopping signals has the problems of high signal false detection rate, long detection time, more hardware resources used, high method complexity, and a large number of adjustment parameters involved in updating and iteratively adding new drone models. Summary of the invention

[0003] In order to solve the above technical problems, a method for frequency-hopping UAV signal identification based on wireless spectrum analysis is provided. This technical solution solves the problems of high signal false detection rate, long detection time, more hardware resources used, high method complexity, and a large number of adjustment parameters involved in updating and iterating to add new UAV models in the frequency-hopping signal detection method proposed in the above background technology.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is: a method for realizing frequency hopping drone signal recognition based on wireless spectrum analysis, comprising the following steps:

[0005] Step S1: Data collection: Collect the image transmission data and frequency hopping remote control data of the drone. 50 sets of data are collected for each drone as the original data.

[0006] Step S2: Data processing: The collected one-dimensional UAV time domain signal is converted into a two-dimensional time-frequency signal through short-time Fourier transform, and the data is calibrated. After the label is set, the data is divided into training data, verification data, and test data in a ratio of 7:1:2, and the signal is converted to a size that meets the model input, and the data is shuffled;

[0007] Step S3: Framework construction: Use the vgg model to deeply extract signal features;

[0008] Step S4: Application of attention mechanism: Add a small weight coefficient to the frequency region where the signal exists To highlight this part of the signal area, the channel adaptive learning method automatically obtains the importance of each feature channel, and then assigns different weight coefficients to each channel to strengthen important features and suppress unimportant features;

[0009] Step S5 parameter setting: The constructed model framework uses the multi-layer convolutional network of the vgg model to extract the features of the drone signal, and performs feature extraction through the reasonable design of the convolution kernel and the number of network layers;

[0010] Step S6: After training a stable model, the output of the model is converted into the possible position information of each signal element of the frequency-hopping drone signal on the time axis of the time-frequency diagram;

[0011] Step S7 Signal Verification: The accurate signal element positions obtained by signal matching are detected one by one to ensure the independence of the signal elements. The four-position amplitude detection method of each element is adopted. The internal amplitude of the signal element is compared with the upper and lower positions of the signal element on the frequency axis and the left and right positions on the time axis. Only when the amplitude meets the signal slope requirement, it is considered to be an independent signal element.

[0012] Step S8: Signal classification: data verification is sorted to remove statistically redundant and repeatedly detected elements and classify them.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] Stable and fast identification of different types of frequency-hopping drone information data to solve the shortcomings of high false detection rate, long detection time, short monitoring distance and complex database in the current process of identifying frequency-hopping drones based on wireless spectrum analysis;

[0015] Simplification: This method simplifies the parameter space by adaptively adjusting the detection parameters. When adding new drone models, fewer feature parameters need to be provided, and the overall method is less complex.

[0016] High efficiency: This method uses GPU and CPU hybrid computing, with low algorithm coupling and obvious modularization, and is suitable for using a variety of parallel technologies to improve algorithm speed;

[0017] Less resource consumption: This method uses GPU to implement some FPGA operations, so that the hardware platform FPGA can be used to process more complex operations, and the overall hardware resource consumption is less;

[0018] Good scalability. This method provides a general drone detection framework, retains the core modules of the algorithm, and can be used in conjunction with other technologies for detecting drone models. Due to the use of a hybrid computing platform, complex matrix operations and logical operations can be distributed to various hardware platforms for coordinated operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a work flow chart of the present invention;

[0020] Figure 2 It is the overall flow chart of the present invention;

[0021] Figure 3 It is a schematic diagram of the network model structure of the present invention;

[0022] Figure 4 Schematic diagram of the channel attention mechanism of the present invention;

[0023] Figure 5 A schematic diagram of a search core of the present invention;

[0024] Figure 6 It is a schematic diagram of the comparative core of the present invention;

[0025] Figure 7 is a schematic diagram of a stable nucleus of the present invention;

[0026] Figure 8 It is a schematic diagram of a signal verification module of the present invention;

[0027] Fig. 9 A display diagram of a GPU operation result of a frequency hopping signal of the present invention;

[0028] Fig.10 A schematic diagram of a potential location of a signal element for the search core of the present invention;

[0029] Fig.11 It is a signal verification schematic diagram of the present invention;

[0030] Fig.12 Schematic diagram of a priori distance L of a signal according to the present invention;

[0031] Fig.13 It is a schematic diagram of the angle change of adjacent signal elements of the present invention. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] First, use drones to acquire features, which includes data collection and data processing.

[0034] S1 Data collection. Place a common frequency-hopping drone on the market in an open area, keep the device 500m away from the drone horizontally, take off 50m vertically, and collect the drone's image transmission data and frequency-hopping remote control data. Collect 50 sets of data from each drone as raw data. Continue to collect data from various models at distances of 1km, 2km, 3km, and signals containing wifi interference, 50 sets of data for each model, and a total of 10,000 sets of data as training, verification, and test data.

[0035] S2 Data Processing: The collected one-dimensional UAV time domain signal is converted into a two-dimensional time-frequency signal through short-time Fourier transform (STFT).

[0036]

[0037] in, Represents the information after STFT transformation, , Represent the dimensions of time and frequency respectively, Indicates that the length in the time domain is One-dimensional drone RF signal, Represents the window function.

[0038] After performing short-time Fourier transformation on the 10,000 collected data, the data is calibrated, and the frequency range of the frequency-hopping signal elements in the time-frequency graph is observed as the basis for subsequent signal enhancement; the distribution of the frequency-hopping signal elements on the time axis of the time-frequency graph is observed, and the starting position of each signal element is used as the signal label of the time-frequency graph. After the label is set, the data is divided into training data, verification data, and test data in a ratio of 7:1:2, the signal is converted to a size that meets the model input, and the data is shuffled.

[0039] Model training includes framework construction, attention mechanism application, parameter setting, and result application.

[0040] S3 framework is built. In deep learning, convolutional neural networks have a very good effect on feature extraction of image signals, so the vgg model is used to deeply extract signal features. As the number of network layers increases, the model's ability to express features is gradually enhanced.

[0041] S4 application of attention mechanism. In order to achieve better feature extraction effect, the attention mechanism is added to the data processing process of the model. First, a small weight coefficient is added to the frequency region where the signal exists. Highlight this part of the signal area. Secondly, the channel attention mechanism is applied to model the correlation between different channels. The importance of each feature channel is automatically obtained through channel adaptive learning, and then different weight coefficients are assigned to each channel to strengthen important features and suppress unimportant features.

[0042] S5 parameter setting. The constructed model framework uses the multi-layer convolutional network of the vgg model to extract the features of drone signals. The best feature extraction effect is achieved by reasonably designing the convolution kernel and the number of network layers. In training, the batch is set to 256, the momentum is set to 0.9, and the layers except the last layer also use the dropout mechanism, with a dropout rate of 0.5. The learning rate is initially set to 0.01, and the weight decay coefficient is 5×10^(-4). Random initialization is used for the weight layer, and the initialization uses a normal distribution with a mean of 0 and a variance of 0.01. In terms of training content, data expansion is performed by rotating and flipping the data. The Adam optimizer and the cross entropy loss function are used for optimization, and the learning rate is gradually reduced according to the continuous progress of training to find the optimal solution. During the training process, a 3x3 convolution kernel is used, which makes the model have a relatively small receptive field. By superposition, the effect of a large convolution kernel can be achieved, and the number of parameters can be greatly reduced. It not only increases the number of network layers, but also reduces the number of parameters, making the model easy to train.

[0043] S6 Result Application: After training a stable model, the output of the model is converted into the possible position information of each signal element of the frequency-hopping drone signal on the time axis of the time-frequency diagram.

[0044] Signal matching obtains potential signal areas through convolution of three adaptive kernels. Frequency-hopping drone signals belong to prior information. By analyzing the common frequency-hopping signals on the market, the characteristic laws of the model are pre-set as prior information. The potential signal area is determined by searching the potential signal time-frequency graph area using three kernel convolutions of adaptive search kernel, contrast kernel, and stable kernel. Specifically, according to the output results of the model, the adaptive search kernel is used to convolve the signal along the frequency axis to remove the potential time points of the signal, and the time and frequency information of the potential signal elements are determined; when there are multiple potential signal elements at similar positions on the time-frequency graph, the contrast kernel is used to convolve multiple potential areas, and the single potential area with the largest convolution result is retained as the pre-selected area; finally, the stable kernel is used to confirm whether the time axis and frequency axis of the pre-selected area meet the basic requirements of the signal with stable internal amplitude, so as to finally determine the current signal element, and the above three kernels are used to process all potential areas on the time-frequency graph to finally determine the distribution of each element of the frequency-hopping signal.

[0045] S7 signal verification. The accurate signal element positions obtained by signal matching are checked one by one to ensure the independence of signal elements (in terms of Figure 4 As a benchmark), the four-position amplitude detection method of each element is used, requiring that there is a relatively obvious signal amplitude gap between the outer edge of the signal element and the signal itself. In this case, the signal element is considered to be independent and not a certain interference signal that meets the characteristics, and meets the requirements of the frequency hopping signal element. Specifically, the signal element will theoretically have a larger amplitude than the surrounding signal. The internal amplitude of the signal element is compared with the upper and lower positions of the signal element on the frequency axis and the left and right positions on the time axis. It is considered to be an independent signal element only when the amplitude meets certain signal slope requirements.

[0046] S8 signal classification. Data that has been verified by the signal is sorted, statistically redundant and repeatedly detected elements are removed, and classified. According to the a priori signal element distance information and angle change information of the frequency hopping signal, the signal elements of the same category are divided into the same coding set, and finally the classified frequency hopping drone information is obtained, including the number of frequency hopping drones and the signal element set of the corresponding drones.

[0047] Take a frequency hopping drone signal as an example ( Fig. 9 As a benchmark), the drone signal is input into the trained model, and the output is the starting position information of the drone frequency hopping signal element on the time axis. The figure is a time-frequency diagram of a frequency hopping drone signal. The horizontal axis represents time and the vertical axis represents frequency. The yellow regular horizontal bars are the signal elements of the frequency hopping signal. The red arrow indicates the starting position of the possible signal output by the model on the time axis.

[0048] Use the search kernel to remove the potential time points of the signal and perform convolution processing on the signal along the frequency axis. The size of the search kernel is the prior information (in Fig.10The bandwidth length of the frame can be dynamically and adaptively adjusted according to the actual situation. The rectangular frame is convolved from the time point 500 along the frequency axis. The frequency range is (0, 200). The search kernel size is a 4x600 rectangle size, where 4 is the frequency axis width (bandwidth) and 600 is the length on the time axis. After convolution processing, 197 output results can be obtained at t=500. Comparing the 197 convolution results, the top 5% (about 9 candidate areas) with the largest output results are selected as potential areas of signal elements (after search kernel processing, if it is an independent signal element, the convolution result value will be very large; if it is a piece of interference signal, this value will be very small). For these 5% of candidate areas, the comparison kernel is used to compare the signal amplitude, and the convolution results are sorted from large to small to form a decreasing sequence of amplitudes. 20% of them are selected as candidate areas for narrowing the range (at least 1 candidate is retained). On this basis, after stable kernel convolution calculation, the areas with smaller change trends in the time axis and frequency axis are selected as quasi-candidate areas. For each moment on the time axis, according to the calculation of the above three kernels, the specific positions of each signal element of the signal segment (starting position of the time axis, starting position of the frequency axis, span of the time axis, span of the frequency axis) are finally obtained.

[0049] The accurate signal element positions obtained by signal matching are detected one by one to ensure the independence of the signal elements. The four-position amplitude detection method of each element is used (based on Figure 7 As a benchmark, it is required that there is a relatively obvious signal amplitude difference between the outer edge of the signal element and the signal itself. Compare the amplitude of the signal area (in the red box). When comparing the upper and lower amplitudes of the frequency axis, take the average value of the signal element in the red rectangular box on the frequency axis as the signal amplitude of the signal element at each time point, starting from the upper frequency edge of the red rectangular box, upward Compare the amplitude at each time point within the frequency range, requiring the amplitude of the signal to be 30% higher than the amplitude outside the signal, and more than 70% of the time points in the frequency band meet the requirement, then the frequency point is considered to be a frequency point that is clearly distinguished from the signal itself. If there are two frequency points in the range that meet this requirement, then the signal element is considered to have signal independence in the direction of increasing frequency, and the same comparison is performed in the direction of decreasing frequency to determine the independence in the direction of decreasing frequency; compare the amplitude of the signal area (in the red box). When comparing the amplitudes on the left and right sides of the time axis, take the mean value of the signal element in the red rectangular box on the time axis as the signal amplitude of each frequency point of the signal element, starting from the time position of the red rectangular box, to the left Compare the amplitudes at each frequency point within a time point range, requiring the amplitude of the signal to be 30% higher than the amplitude outside the signal, and if more than 70% of the frequency points at this time point meet the requirement, then the time moment is considered to be a time point that is clearly distinguished from the signal itself. If there are two time points within the range that meet this requirement, then the signal element is considered to have signal independence in the direction of time reduction, and the independence in the direction of time increase is judged in the same way. When the independence requirements are met in the four directions of frequency increase, decrease, time increase, and decrease, then the signal element is independent.

[0050] Signal classification. The set of detected signal element positions is in Indicates the starting coordinate of the signal element on the time axis. Indicates the starting position of the signal element on the frequency axis, Indicates the length of the signal on the time axis, Indicates the length on the frequency axis. The specific set is: according to the type11 signal prior information, the distance between signal elements is L (in Figure 8 As a benchmark), signal elements with uniform spacing of approximately L can be preliminarily classified into one category:

[0051]

[0052] According to the prior information, the angle variation range of the centers of adjacent signal elements is , calculate the angle change between signal elements ( Figure 8 )size ,If the changing angle of the signal element is within the prior angle, it is considered to be the ,signal element of the same UAV and divided into the same set, otherwise it is excluded.

[0053] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for frequency-hopping drone signal recognition based on wireless spectrum analysis, characterized in that: The following steps are involved: Step S1: Data collection: Collect the image transmission data and frequency hopping remote control data of the drone. 50 sets of data are collected for each drone as the original data. Step S2: Data processing: The collected one-dimensional UAV time domain signal is converted into a two-dimensional time-frequency signal through short-time Fourier transform, and the data is calibrated. After the label is set, the data is divided into training data, verification data, and test data in a ratio of 7:1:2, and the signal is converted to a size that meets the model input, and the data is shuffled; Step S3: Framework construction: Use the vgg model to deeply extract signal features; Step S4: Application of attention mechanism: Add a small weight coefficient to the frequency region where the signal exists To highlight this part of the signal area, the channel adaptive learning method automatically obtains the importance of each feature channel, and then assigns different weight coefficients to each channel to strengthen important features and suppress unimportant features; Step S5 parameter setting: The constructed model framework uses the multi-layer convolutional network of the vgg model to extract the features of the drone signal, and performs feature extraction through the reasonable design of the convolution kernel and the number of network layers; Step S6: After training a stable model, the output of the model is converted into the possible position information of each signal element of the frequency-hopping drone signal on the time axis of the time-frequency diagram; Step S7 Signal Verification: The accurate signal element positions obtained by signal matching are detected one by one to ensure the independence of the signal elements. The four-position amplitude detection method of each element is adopted. The internal amplitude of the signal element is compared with the upper and lower positions of the signal element on the frequency axis and the left and right positions on the time axis. Only when the amplitude meets the signal slope requirement, it is considered to be an independent signal element. Step S8: Signal classification: data verification is sorted to remove statistically redundant and repeatedly detected elements and classify them.

Images