Time-Frequency Spectrum Reconstruction Method Based on UAV Communication Signal Data Enhancement

Abstract

The present invention discloses a time-frequency spectrum reconstruction method based on data enhancement of UAV communication signals, which includes: collecting various types of UAV communication signals; determining the variance of additive white Gaussian noise based on multiple first short sampling sequences obtained by segmenting the UAV communication signals, and reconstructing the first short sampling sequences based on the variance of the additive white Gaussian noise to obtain second short sampling sequences; determining the time-frequency matrix corresponding to the second short sampling sequences; obtaining a time-frequency matrix data set after constructing data enhancement according to all the time-frequency matrices; inputting the time-frequency matrices in the time-frequency matrix data set into a trained ResNet network, and outputting the type serial number of the UAV. The reconstructed UAV time-frequency spectrum of the present invention has the advantages of controllable aliasing interference, high quality, and rich samples, solves the problem that data enhancement is not convenient to be directly applied to UAV communication signals, and broadens the applicable range of data enhancement.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing, and relates to a time-frequency spectrum reconstruction method based on data enhancement of UAV communication signals. Background Art

[0002] The time-frequency spectrum of the UAV downlink video transmission signal usually contains the time-domain information and frequency-domain information of the UAV radio frequency signal, and the time-frequency spectrum textures of different types of UAVs have different characteristics. Therefore, the time-frequency spectrum of the UAV communication signal is often used in the UAV target recognition and classification tasks. At present, due to various urban interferences, the UAV time-frequency spectrum is severely aliased, and the neural network has poor recognition performance for the UAV time-frequency spectrum. Common data enhancement methods include flipping, rotating, scaling images or adding noise to images. These methods will damage the texture structure of the original image and thus affect the recognition performance of the classification system. However, the time and frequency distributions of the UAV communication signal follow fixed formats and transmission rules, so traditional data enhancement is not convenient to be directly applied to the UAV communication signal.

[0003] Although there is currently a UAV spectrum recognition method based on the short-time Fourier transform. However, firstly, this method collects the UAV signals of all operating channels by manually controlling the opening and closing of the signal acquisition equipment. Secondly, the short-time Fourier transform is performed on the UAV data to generate a time-frequency spectrum. Then, the training data set and the test data set are divided for the UAV spectrum. Finally, the region-based convolutional neural network R-CNN (Regions with CNN features) is used for target recognition to obtain the recognition result. Although deep learning has a significant effect on image classification, making the recognition accuracy of the time-frequency spectrum generated by the short-time Fourier transform higher. However, the disadvantage of this method is still that due to various urban interferences, the UAV time-frequency spectrum is severely aliased, and the neural network has poor recognition performance for the UAV time-frequency spectrum.

[0004] Currently, there is also a method for identifying unmanned aerial vehicles (UAVs) based on UAV measurement and control signals. The implementation steps of this method are as follows: 1) Obtain the measurement and control signals and generate a time-frequency spectrum diagram using the short-time Fourier transform; 2) Label the types of the time-frequency spectrum diagram and divide it into a training data set and a test data set; 3) Use a UAV data set and perform target recognition using a convolutional neural network (CNN) to obtain the recognition result. Although the measurement and control signals of different models of UAVs are collected and the generated UAV time-frequency spectrum diagrams after the short-time Fourier transform can be used for UAV target recognition. However, the deficiencies of this method are still that the traditional time-frequency spectrum aliasing is relatively serious, resulting in poor recognition performance of the classification system. Using image data augmentation will damage the texture structure of the time-frequency spectrum diagram. However, UAV communication signals follow fixed formats and transmission rules, and data augmentation is not convenient to be directly applied to UAV communication signals. Therefore, how to solve the problems that the traditional time-frequency spectrum diagram has poor data set quality due to spectrum aliasing and data augmentation is not convenient to be directly applied to UAV communication signals. Summary of the Invention

[0005] To solve the above problems existing in the prior art, the present invention provides a method for reconstructing a time-frequency spectrum diagram based on data augmentation of UAV communication signals. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0006] A method for reconstructing a time-frequency spectrum diagram based on data augmentation of UAV communication signals, the time-frequency spectrum diagram reconstruction method includes:

[0007] Step 1: Collect various types of UAV communication signals;

[0008] Step 2: Determine the variance of additive white Gaussian noise based on a plurality of first short sampling sequences obtained by segmenting the UAV communication signals, and reconstruct the first short sampling sequences based on the variance of the additive white Gaussian noise to obtain second short sampling sequences;

[0009] Step 3: Determine the time-frequency matrix corresponding to the second short sampling sequences;

[0010] Step 4: Obtain a time-frequency matrix data set after constructing data augmentation according to all the time-frequency matrices, and the time-frequency matrix data set includes a plurality of uniquely named time-frequency matrices;

[0011] Step 5: Input the time-frequency matrices in the time-frequency matrix data set into a trained ResNet network and output the type serial numbers of the UAVs.

[0012] In an embodiment of the present invention, the step 2 includes:

[0013] Step 2.1: Uniformly divide each segment of the UAV communication signal at an interval of L to obtain x segments of first short sampling sequences;

[0014] Step 2.2: Obtain the power spectral density according to the first short sampling sequence;

[0015] Step 2.3: Obtain the variance of additive white Gaussian noise according to the signal-to-noise ratio and the power spectral density;

[0016] Step 2.4: Generate a signal of additive white Gaussian noise according to the variance of the additive white Gaussian noise;

[0017] Step 2.5: Superimpose the signal of the white Gaussian noise on the first short sampling sequence to obtain a second short sampling sequence.

[0018] In an embodiment of the present invention, the power spectral density is:

[0019]

[0020] where, represents the power spectral density of the first short sampling sequence, m(i) represents the i-th sampling point in the first short sampling sequence, 1 ≤ i ≤ L, and |·| represents the absolute value operation.

[0021] In an embodiment of the present invention, the i-th sampling point in the second short sampling sequence is:

[0022]

[0023] where, y(i) represents the i-th sampling point in the second short sampling sequence, m(i) represents the i-th sampling point in the first short sampling sequence, α represents the weighting coefficient, f c represents the random frequency deviation, f s represents the system sampling rate, B(i) represents the signal of additive white Gaussian noise, and j represents the imaginary unit symbol, 1 ≤ i ≤ L.

[0024] In an embodiment of the present invention, the step 3 includes:

[0025] Perform STFT transformation on the second short sampling sequence to obtain the time-frequency matrix corresponding to the second short sampling sequence.

[0026] In an embodiment of the present invention, the time-frequency matrix is:

[0027]

[0028] where, Denote the time-frequency matrix with m rows and n columns generated after the STFT transformation of the f-th second-shortest sampling sequence, where 1 ≤ f ≤ x, 0 ≤ m ≤ M - 1, 0 ≤ n ≤ N - 1, M represents the dimension of the time domain in the STFT transformation, N represents the dimension of the frequency domain in the STFT transformation, Σ represents the summation operation, and y f (i) represents the i-th element in the f-th second-shortest sampling sequence, w(·) represents the Hamming window function, d represents the length of the STFT transformation sliding, d = N, and e (·) Denote the exponential operation with the natural constant e as the base.

[0029] In an embodiment of the present invention, step 4 includes:

[0030] Obtain the named time-frequency matrix according to the type of the UAV to which the time-frequency matrix belongs, the channel where it is located, and the location where it is located, so as to construct a time-frequency matrix dataset through all the named time-frequency matrices.

[0031] In an embodiment of the present invention, the training method of the ResNet network includes:

[0032] S1. Obtain a training dataset, where the training dataset includes several time-frequency spectra obtained by performing STFT transformation on UAV communication signals;

[0033] S2. Input the time-frequency spectra into the ResNet network, and use the loss value of the loss function to iteratively update the parameters of the ResNet network until the loss function converges, so as to obtain a trained ResNet network.

[0034] In an embodiment of the present invention, the loss function is:

[0035]

[0036] where Loss represents the loss value of the loss function, Nk represents the total number of samples in the training dataset, ii and MM both represent the sample numbers in the training dataset, cc represents the number of the UAV type label output by the ResNet network, 1 ≤ ii ≤ Nk, 1 ≤ MM ≤ Nk, and y iicc Denote the sign function. When the number of the true type label of the sample number ii is equal to the number of the UAV type label cc output by the ResNet network, y iicc is 1, otherwise, y iicc is 0, and p iicc represents the predicted probability of belonging to the number cc of the type label when the sample number is ii.

[0037] Advantages of the present invention:

[0038] First, since the present invention reconstructs the time-frequency spectrum samples of unmanned aerial vehicles (UAVs), the reconstructed time-frequency spectrum of UAVs contains at least five types of UAVs, and each type contains multiple samples. This overcomes the problem in the prior art that the traditional time-frequency spectrum has poor data set quality due to spectral aliasing. As a result, the time-frequency spectrum of UAVs reconstructed by the present invention has the advantages of controllable aliasing interference, high quality, and rich samples, and can be directly applied to the system for detecting and identifying based on the time-frequency spectrum of UAVs.

[0039] Second, since the present invention enhances the UAV communication signal through data augmentation, it overcomes the problem that when enhancing the quality of the time-frequency spectrum by enhancing the existing image data, the original texture structure of the data will be damaged, and the UAV communication signal follows a fixed format and transmission rule, making it inconvenient to directly apply data augmentation to the UAV communication signal. The present invention solves the problem that data augmentation is not convenient to directly apply to UAV communication signals and broadens the applicable range of data augmentation.

[0040] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. Description of the Drawings

[0041] Figure 1 It is a schematic flowchart of a method for reconstructing a time-frequency spectrum based on data augmentation of UAV communication signals provided by an embodiment of the present invention;

[0042] Figure 2 It is a scene graph provided by an embodiment of the present invention;

[0043] Figure 3 It is a simulation diagram of the recognition performance of the UAV time-frequency spectrum without data augmentation provided by an embodiment of the present invention;

[0044] Figure 4 It is a simulation diagram of the recognition performance of the UAV time-frequency spectrum with data augmentation provided by an embodiment of the present invention. Detailed Embodiments

[0045] The present invention will be further described in detail below in conjunction with specific embodiments, but the embodiments of the present invention are not limited thereto.

[0046] Embodiment 1

[0047] Please refer to Figure 1 , Figure 1 It is a schematic flowchart of a method for reconstructing a time-frequency spectrum based on data augmentation of UAV communication signals provided by an embodiment of the present invention. This embodiment provides a method for reconstructing a time-frequency spectrum based on data augmentation of UAV communication signals. The time-frequency spectrum reconstruction method includes steps 1-step 5, where:

[0048] Step 1: Collect various types of UAV communication signals.

[0049] Specifically, at least five types of UAV communication signals are collected respectively, with at least one UAV for each type. Each UAV collects a long sampling sequence of length K under each channel, that is, the UAV communication signal.

[0050] Preferably, K = 3×10 8 .

[0051] In a specific embodiment, step 1 may include step 1.1 - step 1.2, where:

[0052] Step 1.1, set up the UAV working environment and signal acquisition equipment to prepare for subsequent signal acquisition work.

[0053] For example, in this embodiment, 5 UAVs of 5 different types are selected, with one UAV corresponding to each type. Each UAV works on two channels, namely channel 153 and channel 161 under 5.8 GHz, as shown in Table 1 specifically.

[0054] Table 1 Sample information table in the UAV time - frequency database

[0055]

[0056] Please refer to Figure 2 for a further description of the signal acquisition equipment and the working principle of the UAV in the embodiment. Under the same layout in the microwave anechoic chamber and complex environment, each UAV is placed 100 meters above the signal acquisition equipment. Figure 2 The Universal Software Radio Peripheral (USRP) X310 in [] is connected to the computer via optical fiber. The GNU radio platform in the computer controls the USRP X310 and stores the received communication signal of each UAV in the computer memory.

[0057] Step 1.2, collect the long sampling sequences of each UAV under channel 153 and channel 161 respectively.

[0058] First step, use GNU Radio to control the USRP, with a sampling rate of 100 Msa / s, collect the UAV communication signals of each UAV under channel 153 for 5 minutes, obtain long sampling sequences with a sampling length of K for all 5 UAVs, and store them in the computer memory.

[0059] Second step, use GNU Radio to control the USRP, with a sampling rate of 100 Msa / s, collect the UAV communication signals of each UAV under channel 161 for 5 minutes, obtain long sampling sequences with a sampling length of K for all 5 UAVs, and store them in the computer memory.

[0060] Step 2: Determine the variance of additive white Gaussian noise based on multiple first short sampling sequences obtained by splitting the UAV communication signal, and reconstruct the first short sampling sequences based on the variance of the additive white Gaussian noise to obtain second short sampling sequences.

[0061] In a specific embodiment, Step 2 may include Steps 2.1 - 2.5, where:

[0062] Step 2.1: Uniformly split each segment of the UAV communication signal at an interval of L to obtain x first short sampling sequences.

[0063] Specifically, uniformly split the UAV communication signal of each long sampling sequence at an interval of L to obtain x first short sampling sequences after splitting each long sampling sequence.

[0064] For example, in the embodiment of the present invention, according to Calculate the number of segments x of each sampling point sequence, where represents rounding down. Since K = 3×10 10 、L = 131072 elements, and x = 228881, it can be calculated that each long sampling point sequence can be split into 228881 first short sampling sequences.

[0065] Step 2.2: Obtain the power spectral density according to the first short sampling sequence, where the power spectral density is:

[0066]

[0067] where represents the power spectral density of the first short sampling sequence, m(i) represents the i-th sampling point in the first short sampling sequence, 1 ≤ i ≤ L, and |·| represents the absolute value operation.

[0068] Step 2.3: Obtain the variance of the additive white Gaussian noise according to the signal-to-noise ratio and the power spectral density.

[0069] Specifically, according to the following formula for calculating the SNR (signal-to-noise ratio) and the prior condition within a 100 MHz receiving bandwidth obtain

[0070]

[0071] where ∫· represents the integral operation, B s represents the UAV signal bandwidth, log represents the logarithm operation with the natural constant 10 as the base, represents the power spectral density of AWGN (Additive White Gaussian Noise), σ 2 represents the variance of the AWGN signal.

[0072] Step 2.4: Generate the signal of additive white Gaussian noise according to the variance of the additive white Gaussian noise.

[0073] Specifically, according to the variance σ of the additive white Gaussian noise 2 the AWGN can be generated using the MATLAB software platform.

[0074] Step 2.5: Superimpose the signal of the white Gaussian noise on the first short sampling sequence to obtain the second short sampling sequence. Use the following formula to superimpose the sampling points of the AWGN signal into the first short sampling sequence. The i-th sampling point in the obtained second short sampling sequence is:

[0075]

[0076] where y(i) represents the i-th sampling point in the second short sampling sequence, m(i) represents the i-th sampling point in the first short sampling sequence, α represents the weighting coefficient, f c represents the random frequency deviation, f s represents the system sampling rate, B(i) represents the signal of the additive white Gaussian noise, j represents the imaginary unit symbol, and 1 ≤ i ≤ L.

[0077] Step 3: Determine the time-frequency matrix corresponding to the second short sampling sequence.

[0078] Specifically, perform the STFT (Short-Time Fourier Transform) transform on each second short sampling sequence to obtain the time-frequency matrix corresponding to each second short sampling sequence.

[0079] In the embodiment of the present invention, L = 131072 elements, x = 228881. According to the following formula, perform the STFT transform on each short sampling sequence:

[0080]

[0081] where represents the time-frequency matrix with m rows and n columns generated after the STFT transform of the f-th second short sampling sequence, 1 ≤ f ≤ x, that is, 1 ≤ f ≤ 228881, 0 ≤ m ≤ M - 1, 0 ≤ n ≤ N - 1, M represents the dimension of the time domain in the STFT transform, M = 256, N represents the dimension of the frequency domain in the STFT transform, N = 512, Σ represents the summation operation, y f(i) represents the i-th element in the f-th second-shortest sampling sequence, w(·) represents the Hamming window function, d represents the length of the STFT transform sliding, d = N, e (·) represents the exponential operation with the natural constant e as the base, and j represents the imaginary unit symbol.

[0082] Step 4: Obtain the time-frequency matrix dataset after constructing data augmentation based on all time-frequency matrices. The time-frequency matrix dataset includes multiple uniquely named time-frequency matrices.

[0083] Specifically, obtain the named time-frequency matrices according to the type of the drone to which the time-frequency matrix belongs, the channel where it is located, and the position where it is located, so as to construct the time-frequency matrix dataset through all the named time-frequency matrices.

[0084] In this embodiment, according to the type V of the drone to which the time-frequency matrix belongs, the channel C where it is located, and the position of the time-frequency matrix (i.e., the O-th time-frequency matrix), each time-frequency matrix is named in the format of V_C_O. Each named time-frequency matrix is stored in the folder corresponding to it with V as the name according to the number of V in the name. The address of the time-frequency matrix in the folder and the recognition label are stored in the time-frequency matrix dataset.

[0085] In the embodiment of the present invention, each time-frequency spectrum is named in the format of V_C_O according to the type of the drone, the channel information, and the number of time-frequency matrices, so as to distinguish each time-frequency spectrum of different drone models under different channels. Where V represents the recognition label of 5 drones, 0 ≤ V ≤ 4, C represents the 153 and 161 channels of the drone. Where C is 0 represents the 149 channel, C is 1 represents the 153 channel, C is 2 represents the 157 channel, C is 3 represents the 161 channel, C is 4 represents the 165 channel, O represents the position of a drone in a total of 228881 time-frequency matrices under a certain channel, 1 ≤ O ≤ 228881. Each named time-frequency matrix is stored in the folder with V as the name according to the number corresponding to V in the name. The absolute path address of each time-frequency matrix in the folder and the folder name are stored in the time-frequency matrix dataset

[0086] Taking "Mavic Mini" in Table 1 as an example, the naming of the corresponding time-frequency spectrum in the format of V_C_O is described as follows: Define the "Mavic Mini" drone as the first type of drone, and mark the recognition label V as 1. Since the 161 channel where the drone is located belongs to the third channel, mark C as 3. The time-frequency matrix of the drone is the eighth time-frequency matrix, and mark O as 8. Therefore, the naming of the eighth time-frequency matrix of Mavic Mini under the 161 channel is 2_3_8.

[0087] In this embodiment, the training method of the ResNet network includes:

[0088] S1. Obtain a training data set, which includes a number of time-frequency spectrograms obtained by performing STFT transformation on UAV communication signals.

[0089] Specifically, randomly select unreconstructed UAV communication signals (i.e., original UAV communication signals), use STFT to transform the unreconstructed UAV communication signals into time-frequency spectrograms and make them into an initial time-frequency spectrogram data set (i.e., the training data set). Among them, the UAV works on channel 161, and 80% of the total number of samples with type labels 0 - 4, with a total of 183,104 samples, are composed of the absolute path addresses and folder names to form the training data set.

[0090] S2. Input the time-frequency spectrogram into the ResNet network, and use the loss value of the loss function to iteratively update the parameters of the ResNet network until the loss function converges, obtaining a trained ResNet network. Among them, the loss function is:

[0091]

[0092] Among them, Loss represents the loss value of the loss function, Nk represents the total number of samples in the training data set, ii and MM both represent the sample numbers in the training data set, cc represents the serial number of the UAV type label output by the ResNet network, 1 ≤ ii ≤ Nk, 1 ≤ MM ≤ Nk. For example, 1 ≤ ii ≤ 183,104, 1 ≤ MM ≤ 183,104, 0 ≤ cc ≤ 4, y iicc represents the sign function. When the serial number of the true type label of sample number ii is equal to the serial number cc of the UAV type label output by the ResNet network, y iicc is 1, otherwise, y iicc is 0, p iicc represents the predicted probability of belonging to the serial number cc of the type label when the sample number is ii.

[0093] In addition, the absolute path addresses and folder names of the remaining 20% of the samples in the sample library are composed of a test data set, which is input into the trained ResNet network for testing, and the recognition accuracy obtained is made into a confusion matrix.

[0094] In an embodiment of the present invention, using the time-frequency spectrograms with type tags 0-4 under channel 161, a training data set is composed of absolute path addresses and folder names that account for 80% of the total 228,881, and a test data set is composed of absolute path addresses and folder names that account for 20% of the total. All the time-frequency matrices (i.e., time-frequency spectrograms) in the training type set are input into the ResNet network in the Python software database, and the parameters of the neurons are iteratively updated. When the maximum training cycle max_epoch = 100 is reached, the training is completed. The test data set is input into the trained ResNet network, the number of the serial numbers of the output type tags that are equal to the serial numbers of the true type tags of the drones is counted, the result after counting is divided by the total number of samples in the test data set to obtain the recognition rate, and the recognition rate of each class is made into a confusion matrix. The structure of the confusion matrix is as Figure 3 shown, where the values of row w and column q from 1 to 5 represent 5 types of drones, and H wq represents the recognition rate when the test is type w and is recognized as q.

[0095] Step 5: Input the time-frequency matrices in the time-frequency matrix data set into the trained ResNet (Residual Network) network to output the type serial numbers of the drones.

[0096] Specifically, in the time-frequency matrix data set generated after data augmentation of the drone communication signal, 20% of the absolute path addresses and folder names of the samples are randomly selected to form a test set, which is input into the trained ResNet network to output the serial numbers of the type tags of each type of drone. Using the formula calculate the recognition rate of each type of drone, and make the recognition rates of all types of drones into a confusion matrix, where P V represents the total number of the serial numbers of all drone type tags output by the ResNet network that belong to the V-th type of drone type, Q represents the total number of samples in the test set, and T V represents the accuracy rate of being recognized as the V-th type of drone type.

[0097] In an embodiment of the present invention, in the time-frequency matrix data set generated after data augmentation of the drone communication signal, 20% of the absolute path addresses and folder names of the total are randomly selected to form a test set. The test set is input into the trained ResNet network, the number of the serial numbers of the output type tags that are equal to the true serial numbers of the drones is counted, the result after counting each class is divided by the total number of samples in the test set to obtain the recognition rate, and the recognition rate is made into a confusion matrix. The results are as Figure 4 shown.

[0098] The reconstructed time-frequency spectrum of the UAV in the present invention contains at least five types of UAV time-frequency spectra, and each type contains 10,000 samples. When performing the UAV recognition task, it eliminates the problem that the traditional UAV time-frequency spectrum is severely aliased due to various urban interferences and the poor recognition performance of the neural network for the UAV time-frequency spectrum. Secondly, through the communication signal data enhancement technology of the present invention, background noise with controllable noise power is added to the UAV communication signal, which can not only eliminate the hidden danger of aliasing interference, but also increase the diversity of samples from the source of the UAV signal, improve the recognition performance of the neural network for the UAV time-frequency spectrum, and solve the problem that traditional data enhancement is not convenient to be directly applied to UAV communication signals due to the limitation that UAV communication signals follow fixed formats and transmission rules.

[0099] The following further illustrates the effect of the present invention through simulation experiments:

[0100] 1. Simulation conditions:

[0101] The hardware platform for the simulation experiment of the present invention is: Intel(R) Core(TM) i7-8700k CPU, the main frequency of the processor is 3.20 GHz, the memory is 16 GB, one USRP X310 and one optical fiber.

[0102] The software platforms for the simulation experiment of the present invention are: GNUradio software platform, MATLAB R2020a software platform, and Python 3.7 software platform.

[0103] 2. Simulation content and result analysis:

[0104] To verify the quality of the UAV time-frequency spectrum database constructed by the present invention, the residual neural network ResNet in the Python 3.7 software platform database is used as a classifier to classify the types of UAVs. Randomly select the absolute path addresses and folder names of 80% of the samples of 6 different types of UAVs in the initial time-frequency spectrum database to form a training set and 20% of the samples to form a test set. Input the training set into the ResNet classifier, and after 100 iterations of training, obtain the trained ResNet classifier. Input the test set into the trained ResNet classifier, output the predictions of the ResNet classifier for the types of each sample in the test set, and calculate the ratio of the number of test samples whose predicted types by the ResNet classifier match the types of the samples to the total number of test samples to obtain the recognition accuracies of 6 UAVs as Figure 3As shown in the confusion matrix of the initial time-frequency database. From the time-frequency map dataset generated after data augmentation of UAV communication signals, randomly select the absolute path addresses and folder names of 20% of the samples of 6 different types of UAVs in the time-frequency map database to form a test set, and input the test set into the trained ResNet classifier. Calculate the ratio of the number of test samples whose predicted type by the ResNet classifier matches the type of the sample to the total number of test samples in the test set, and obtain the recognition accuracy of 6 types of UAVs as Figure 4 As shown in the confusion matrix of the augmented time-frequency database. The higher the target recognition accuracy, the better the quality of the constructed UAV time-frequency map database.

[0105] Figure 3 and Figure 4 The abscissas in , from left to right, represent the predicted types of 6 types of UAVs and 7 types predicted as having no UAV signals respectively, and the ordinates, from top to bottom, represent the true types of 6 types of UAVs and 7 types having no UAV signals respectively. Figure 3 and Figure 4 The values of the elements in the matrix of represent the probabilities that each type on the ordinate is predicted as each type on the abscissa respectively, Figure 3 and Figure 4 The values on the diagonal marked in the black background area in represent the probabilities of correct prediction of 7 types respectively, and the recognition probabilities are expressed as ratio decimals.

[0106] By comparing the confusion matrices in Figure 3 and Figure 4 it is found that Figure 4 in the confusion matrix, the diagonal accuracy is higher than Figure 3 , so the classification accuracy of generating time-frequency maps by data augmentation of UAV communication signals is higher than that without data augmentation. Therefore, the method of the present invention can achieve effective estimation and meet the engineering requirements of UAV communication signal recognition. Thus, compared with the prior art, the method of the present invention solves the problems that the traditional time-frequency map has poor dataset quality due to spectral aliasing, and when improving the quality of the time-frequency map by image data augmentation, the original data texture structure will be damaged, and the UAV communication signal follows a fixed format and transmission rule, so data augmentation is not convenient to be directly applied to UAV communication signals, resulting in the inability to generate a dataset with high quality and poor recognition performance of the classification system.

[0107] The method of the present invention can be used in the recognition scenarios of micro UAVs, light UAVs, and small UAV targets.

[0108] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality of" means two or more unless otherwise specifically defined.

[0109] In the description of this specification, the description with reference to terms such as "an embodiment", "some embodiments", "examples", "specific examples", or "some examples", etc. means that the specific features, structures, materials, or specific data points described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or specific data points described may be combined in a suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0110] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited only to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A time-frequency spectrum reconstruction method based on the enhancement of UAV communication signal data, characterized in that The time-frequency spectrum reconstruction method includes the following steps: Step 1: Collect various types of UAV communication signals; Step 2: Determine the variance of additive white Gaussian noise based on multiple first short sampling sequences obtained by segmenting the UAV communication signals, and reconstruct the first short sampling sequences based on the variance of the additive white Gaussian noise to obtain second short sampling sequences; Step 3: Determine the time-frequency matrix corresponding to the second short sampling sequences; Step 4: Obtain a time-frequency matrix dataset after constructing data augmentation according to all the time-frequency matrices, and the time-frequency matrix dataset includes multiple uniquely named time-frequency matrices; Step 5: Input the time-frequency matrices in the time-frequency matrix dataset into the trained ResNet network, and output the type serial number of the UAV. The said Step 2 includes: Step 2.1: Uniformly divide each segment of the UAV communication signal at intervals of L to obtain x segments of the first short sampling sequence; Step 2.2: Obtain the power spectral density according to the first short sampling sequences; Step 2.3: Obtain the variance of additive white Gaussian noise according to the signal-to-noise ratio and the power spectral density; Step 2.4: Generate a signal of additive white Gaussian noise according to the variance of the additive white Gaussian noise; Step 2.5: Superimpose the signal of the Gaussian white noise on the first short sampling sequences to obtain second short sampling sequences.

2. The time-frequency spectrum reconstruction method based on the enhancement of drone communication signal data according to claim 1, wherein The power spectral density is: Among them, represents the power spectral density of the first short sampling sequence, represents the i -th sampling point in the first short sampling sequence, 1 ≤ i ≤ L, represents the absolute value operation.

3. The method for reconstructing a time-frequency spectrum map based on enhancing drone communication signal data according to claim 2, wherein The i th sampling point in the second short sampling sequence is: Among them, represents the i th sampling point in the second shortest sampling sequence, represents the i th sampling point in the first shortest sampling sequence, α represents the weighting coefficient, represents the random frequency deviation, represents the system sampling rate, represents the signal of additive white Gaussian noise, represents the imaginary unit symbol, 1 ≤ i ≤ L.

4. The time-frequency spectrum reconstruction method based on the enhancement of UAV communication signal data according to claim 1, wherein The said Step 3 includes: Perform STFT transformation on the second short sampling sequences to obtain the time-frequency matrix corresponding to the second short sampling sequences.

5. The method for reconstructing a time-frequency spectrum map based on enhancing drone communication signal data according to claim 4, wherein The time-frequency matrix is: Among them, represents the th row and m th column time-frequency matrix generated after the STFT transformation of the n th second-shortest sampling sequence, , , M represents the dimension of the time domain in the STFT transformation, N represents the dimension of the frequency domain in the STFT transformation, represents the summation operation, represents the th element in the i th second-shortest sampling sequence, represents the Hamming window function, represents the length of the STFT transformation sliding, , represents the exponential operation with the natural constant as the base.

6. The method for reconstructing a time-frequency spectrogram based on enhancing drone communication signal data according to claim 1, wherein The said Step 4 includes: Obtain the named time-frequency matrices according to the type of UAV to which the time-frequency matrices belong, the channels where they are located, and the positions where they are located, and construct a time-frequency matrix dataset through all the named time-frequency matrices.

7. The method for reconstructing a time-frequency spectrum based on enhancing drone communication signal data according to claim 1, wherein The training method of the ResNet network includes: S1: Obtain a training dataset, and the training dataset includes several time-frequency spectra obtained by performing STFT transformation on UAV communication signals; S2: Input the time-frequency spectra into the ResNet network, and use the loss value of the loss function to iteratively update the parameters of the ResNet network until the loss function converges, and obtain the trained ResNet network.

8. The time-frequency spectrum reconstruction method based on enhancing drone communication signal data according to claim 7, wherein The loss function is: Among them, Loss represents the loss value of the loss function, Nk represents the total number of samples in the training dataset, ii and both represent the serial numbers of samples in the training dataset, cc represents the serial number of the UAV type label output by the ResNet network, Nk, Nk, represents the sign function. When the serial number of the true type label of the sample serial number ii is equal to the serial number cc of the UAV type label output by the ResNet network, it is 1, otherwise, it is 0, represents the predicted probability of belonging to the serial number cc of the type label when the sample serial number is ii.

Images