Unmanned aerial vehicle signal identification system, method and equipment based on deep learning, and medium

Abstract

The invention provides an unmanned aerial vehicle signal identification system, method, device and medium based on deep learning, and the system comprises a signal fusion feature extraction module which is used for obtaining the time-frequency feature, the space-frequency feature and the spatial-temporal feature of an unmanned aerial vehicle signal; the CNN recognition network module is used for analyzing the time-frequency characteristics to obtain possible type information I of the signals and possible frequency hopping information I of the signals, and analyzing the spatial-temporal characteristics to obtain possible type information II of the signals and possible direction information I of the signals; analyzing the space-frequency characteristics to obtain possible direction information 2 of the signal and possible frequency hopping information 2 of the signal; and the fusion network module is used for fusing the possible type information to output the type of the unmanned aerial vehicle signal, fusing the possible direction information to output the direction of the unmanned aerial vehicle signal, and fusing the possible frequency hopping information to output the frequency hopping of the unmanned aerial vehicle signal. According to the invention, the accuracy of unmanned aerial vehicle signal identification is improved by using deep learning.

Description

Technical Field

[0001] The present invention relates to the field of drone signal analysis and recognition, and in particular to a drone signal recognition system, method, device and medium based on deep learning. Background Art

[0002] Drone signal recognition refers to the analysis of drone radio communication signals (such as remote control signals, video transmission signals or navigation signals) to achieve tasks such as detection, classification, positioning and interference of drones. Its core is to extract and identify drone-related signal features from complex electromagnetic environments for monitoring or confrontation. In order to better develop the low-altitude economy, drone signal recognition technology is widely used in airports, military facilities, no-fly zones and other areas to ensure airspace safety and improve security capabilities.

[0003] Some common drone signal recognition methods include: spectrum feature recognition, modulation and protocol recognition, signal recognition based on machine learning, and recognition methods based on array signal processing.

[0004] Signal recognition based on machine learning refers to the use of machine learning algorithms to build models by learning signal features (such as spectral characteristics and time series characteristics) to achieve automatic recognition of drone signals. The signal recognition solution based on machine learning uses labeled data to train classifiers and trained models to classify real-time signals. Compared with traditional methods, it reduces the reliance on manual feature extraction. This method has high recognition performance, but the accuracy of recognition strictly depends on data quality, and it performs poorly in complex scenarios such as multipath effects and has limited generalization capabilities.

[0005] It can be seen that the accuracy of drone signal recognition technology in the existing technology is low. Summary of the invention

[0006] The present invention provides a drone signal recognition system, method, device and medium based on deep learning to solve the problem of low accuracy in drone signal recognition.

[0007] According to a first aspect of the present invention, a drone signal recognition system based on deep learning is provided, comprising:

[0008] The signal fusion feature extraction module is used to perform information fusion and feature extraction on the drone signals received by multiple receiving antennas of the drone signal receiving device to obtain the drone signal time-frequency features, drone signal time-space features and drone signal space-frequency features;

[0009] The CNN recognition network module is used to analyze the time-frequency characteristics of the drone signal to obtain possible type information 1 of the signal and possible frequency hopping information 1 of the signal, analyze the time-space characteristics of the drone signal to obtain possible type information 2 of the signal and possible direction information 1 of the signal, and analyze the space-frequency characteristics of the drone signal to obtain possible direction information 2 of the signal and possible frequency hopping information 2 of the signal;

[0010] The fusion network module is used to fuse the possible type information of the signal one and the possible type information of the signal two to output the type of the drone signal, to fuse the possible direction information of the signal one and the possible direction information of the signal two to output the direction of the drone signal, and to fuse the possible frequency hopping information of the signal one and the possible frequency hopping information of the signal two to output the frequency hopping of the drone signal.

[0011] Optionally, the system also includes an optimization module, which is used to adjust and optimize the CNN recognition network module.

[0012] Optionally, the CNN recognition network module includes:

[0013] The first CNN recognition network unit is used to analyze the time-frequency information of the signal according to the time-frequency characteristics of the drone signal and output it;

[0014] A first type of fully connected layer network unit, used to integrate the time-frequency information of the signal output by the first CNN recognition network unit, and output possible type information of the signal one;

[0015] A first frequency hopping fully connected layer network unit is used to integrate the time-frequency information of the signal output by the first CNN recognition network unit, and output possible frequency hopping information of the signal;

[0016] The second CNN recognition network unit is used to analyze the spatiotemporal information of the signal according to the spatiotemporal characteristics of the drone signal and output it;

[0017] The second type of fully connected layer network unit is used to integrate the spatiotemporal information of the signal output by the second CNN recognition network unit, and output possible type information 2 of the signal;

[0018] The first directional fully connected layer network unit is used to integrate the spatiotemporal information of the signal output by the second CNN recognition network unit, and output possible directional information of the signal one;

[0019] The third CNN recognition network unit is used to analyze the signal's space-frequency information according to the UAV signal's space-frequency characteristics and output it;

[0020] The second directional fully connected layer network unit is used to integrate the space-frequency information of the signal output by the third CNN recognition network unit, and output possible direction information 2 of the signal;

[0021] The second frequency hopping fully connected layer network unit is used to integrate the space-frequency information of the signal output by the third CNN recognition network unit, and output the possible frequency hopping information 2 of the signal.

[0022] Optionally, the system also includes an optimization module, which is used to: adjust and optimize the first CNN recognition network unit and the first type of fully connected layer network unit according to the output of the first type of fully connected layer network unit; adjust and optimize the first CNN recognition network unit and the first frequency hopping fully connected layer network unit according to the output of the first frequency hopping fully connected layer network unit; adjust and optimize the second CNN recognition network unit and the second type of fully connected layer network unit according to the output of the second type of fully connected layer network unit; adjust and optimize the second CNN recognition network unit and the first direction fully connected layer network unit according to the output of the first direction fully connected layer network unit; adjust and optimize the third CNN recognition network unit and the second frequency hopping fully connected layer network unit according to the output of the second frequency hopping fully connected layer network unit; adjust and optimize the third CNN recognition network unit and the second direction fully connected layer network unit according to the output of the second direction fully connected layer network unit.

[0023] Optionally, the optimization module includes:

[0024] A first optimization unit is used to use a first category information determiner to determine possible category information 1 of the signal, obtain category information 1 determined by the determination, input the category information 1 determined by the determination and category label information 1 into a first category loss function, and use the first category loss function to optimize the first CNN recognition network unit and the first category fully connected layer network unit;

[0025] A second optimization unit is used to use the first frequency hopping information determiner to determine the possible frequency hopping information 1 of the signal, obtain the determined frequency hopping information 1, input the determined frequency hopping information 1 and the frequency hopping label information 1 into the first frequency hopping loss function, and use the first frequency hopping loss function to optimize the first CNN recognition network unit and the first frequency hopping fully connected layer network unit;

[0026] A third optimization unit is used to use the second category information determiner to determine the possible category information 2 of the signal, obtain the category information 2 determined by the determination, input the category information 2 determined by the determination and the category label information 2 into the second category loss loss function, and use the second category loss loss function to optimize the second CNN recognition network unit and the second category fully connected layer network unit;

[0027] A fourth optimization unit is used to use the first direction information determiner to determine the possible direction information 1 of the signal, obtain the direction information 1 determined by the determination, input the direction information 1 determined by the determination and the direction label information 1 into the first direction loss function, and use the first direction loss function to optimize the second CNN recognition network unit and the first direction fully connected layer network unit;

[0028] A fifth optimization unit is used to use the second direction information determiner to determine the possible direction information 2 of the signal, obtain the direction information 2 determined by the determination, input the direction information 2 determined by the determination and the direction label information 2 into the second direction loss function, and use the second direction loss function to optimize the third CNN recognition network unit and the second direction fully connected layer network unit;

[0029] The sixth optimization unit is used to use the second frequency hopping information determiner to determine the possible frequency hopping information 2 of the signal, obtain the frequency hopping information 2 determined by the determination, input the frequency hopping information 2 determined by the determination and the frequency hopping label information 2 into the second frequency hopping loss function, and use the second frequency hopping loss function to optimize the third CNN recognition network unit and the second frequency hopping fully connected layer network unit.

[0030] Optionally, the fusion network module includes:

[0031] A type information fusion module is used to fuse possible type information 1 of the signal and possible type information 2 of the signal to output the type recognition result of the drone signal, and is used to adjust and optimize the type information fusion unit according to the type recognition result of the drone signal and the type information fusion related labels;

[0032] A direction information fusion module is used to fuse possible direction information 1 of the signal and possible direction information 2 of the signal to output a direction recognition result of the UAV signal, and is used to adjust and optimize the direction information fusion unit according to the direction recognition result of the UAV signal and the direction information fusion related label;

[0033] The frequency hopping information fusion module is used to fuse the possible frequency hopping information 1 and the possible frequency hopping information 2 of the signal to output the frequency hopping identification result of the drone signal, and is used to adjust and optimize the frequency hopping information fusion unit according to the frequency hopping identification result of the drone signal and the frequency hopping information fusion related labels.

[0034] Optionally, the signal fusion feature extraction module includes:

[0035] The first signal fusion feature extraction unit is used to divide the drone signal received by multiple receiving antennas of the drone signal receiving device into a plurality of groups of frequency domain sub-signals according to time, and perform Fourier transform on the frequency domain sub-signals after the transform according to the time dimension, and after performing a denoising operation on the matrix, fuse the denoised matrix to obtain the time-frequency feature of the drone signal;

[0036] The second signal fusion feature extraction unit is used to divide the drone signal received by multiple receiving antennas of the drone signal receiving device into a plurality of groups of spatial domain sub-signals according to time, perform spatial spectrum fusion on the spatial domain sub-signals of all different receiving antennas in the same time period to obtain a first spatial spectrum signal, and splice the first spatial spectrum signal according to the time dimension to obtain the spatiotemporal features of the drone signal;

[0037] The third signal fusion feature extraction unit is used to divide the drone signal received by multiple receiving antennas of the drone signal receiving device into several groups of frequency domain sub-signals 2 according to the frequency domain sub-carrier bandwidth, perform spatial spectrum fusion on the frequency domain sub-signals 2 of all different receiving antennas in the same frequency band to obtain a second spatial spectrum signal, and splice the second spatial spectrum signal according to the frequency dimension to obtain the drone signal space-frequency feature.

[0038] According to a second aspect of the present invention, a method for identifying drone signals based on deep learning is provided, comprising:

[0039] Obtaining drone signals received by multiple receiving antennas of a drone signal receiving device;

[0040] Information fusion and feature extraction are performed on the drone signals received by multiple receiving antennas to obtain the drone signal time-frequency features, drone signal time-space features, and drone signal space-frequency features;

[0041] The time-frequency characteristics of the UAV signal are analyzed to obtain possible type information of the signal and possible frequency hopping information of the signal;

[0042] The spatiotemporal characteristics of the drone signal are analyzed to obtain possible type information 2 of the signal and possible direction information 1 of the signal;

[0043] The spatial frequency characteristics of the UAV signal are analyzed to obtain possible direction information 2 and possible frequency hopping information 2 of the signal;

[0044] The possible type information 1 of the signal and the possible type information 2 of the signal are merged to output the type of the drone signal;

[0045] The possible direction information 1 of the signal and the possible direction information 2 of the signal are fused to output the direction of the drone signal;

[0046] The possible frequency hopping information one of the signal and the possible frequency hopping information two of the signal are fused to output the frequency hopping of the drone signal.

[0047] Optionally, the analysis of the time-frequency characteristics of the drone signal, the analysis of the time-space characteristics of the drone signal, and the analysis of the space-frequency characteristics of the drone signal are all implemented by a CNN recognition network module, and the method also includes the step of adjusting and optimizing the CNN recognition network module.

[0048] Optionally, the CNN recognition network module includes:

[0049] The first CNN recognition network unit is used to analyze the time-frequency information of the signal according to the time-frequency characteristics of the drone signal and output it;

[0050] A first type of fully connected layer network unit, used to integrate the time-frequency information of the signal output by the first CNN recognition network unit, and output possible type information of the signal one;

[0051] A first frequency hopping fully connected layer network unit is used to integrate the time-frequency information of the signal output by the first CNN recognition network unit, and output possible frequency hopping information of the signal;

[0052] The second CNN recognition network unit is used to analyze the spatiotemporal information of the signal according to the spatiotemporal characteristics of the drone signal and output it;

[0053] The second type of fully connected layer network unit is used to integrate the spatiotemporal information of the signal output by the second CNN recognition network unit, and output possible type information 2 of the signal;

[0054] The first directional fully connected layer network unit is used to integrate the spatiotemporal information of the signal output by the second CNN recognition network unit, and output possible directional information of the signal one;

[0055] The third CNN recognition network unit is used to analyze the signal's space-frequency information according to the UAV signal's space-frequency characteristics and output it;

[0056] The second directional fully connected layer network unit is used to integrate the space-frequency information of the signal output by the third CNN recognition network unit, and output possible direction information 2 of the signal;

[0057] The second frequency hopping fully connected layer network unit is used to integrate the space-frequency information of the signal output by the third CNN recognition network unit, and output the possible frequency hopping information 2 of the signal.

[0058] Optionally, the steps of adjusting and optimizing the CNN recognition network module are specifically as follows:

[0059] Using a first category information determiner to determine possible category information 1 of the signal, obtaining category information 1 determined by the determination, inputting the category information 1 determined by the determination and category label information 1 into a first category loss function, and using the first category loss function to optimize a first CNN recognition network unit and a first category fully connected layer network unit;

[0060] Using a first frequency hopping information determiner to determine possible frequency hopping information 1 of the signal, obtaining determined frequency hopping information 1, inputting the determined frequency hopping information 1 and frequency hopping label information 1 into a first frequency hopping loss function, and using the first frequency hopping loss function to optimize a first CNN recognition network unit and a first frequency hopping fully connected layer network unit;

[0061] Using the second category information determiner to determine the possible category information 2 of the signal, obtaining the category information 2 determined by the determination, inputting the category information 2 determined by the determination and the category label information 2 into the second category loss function, and using the second category loss function to optimize the second CNN recognition network unit and the second category fully connected layer network unit;

[0062] Using the first direction information determiner to determine the possible direction information 1 of the signal, obtaining the direction information 1 determined by the determination, inputting the direction information 1 determined by the determination and the direction label information 1 into the first direction loss function, and using the first direction loss function to optimize the second CNN recognition network unit and the first direction fully connected layer network unit;

[0063] The second direction information determiner is used to determine the possible direction information 2 of the signal to obtain the determined direction information 2, the determined direction information 2 and the direction label information 2 are input into the second direction loss function, and the second direction loss function is used to optimize the third CNN recognition network unit and the second direction fully connected layer network unit;

[0064] The second frequency hopping information determiner is used to determine the possible frequency hopping information 2 of the signal to obtain the determined frequency hopping information 2, the determined frequency hopping information 2 and the frequency hopping label information 2 are input into the second frequency hopping loss function, and the second frequency hopping loss function is used to optimize the third CNN recognition network unit and the second frequency hopping fully connected layer network unit.

[0065] Optionally, the step of fusing possible type information one of the signal and possible type information two of the signal is implemented by a type information fusion module, the step of fusing possible direction information one of the signal and possible direction information two of the signal is implemented by a direction information fusion module, the step of fusing possible frequency hopping information one of the signal and possible frequency hopping information two of the signal is implemented by a frequency hopping information fusion module, and the method also includes the step of adjusting and optimizing the type information fusion module, the direction information fusion module and the frequency hopping information fusion module.

[0066] Optionally, the steps of adjusting and optimizing the type information fusion module, the direction information fusion module and the frequency hopping information fusion module are specifically as follows: according to the type identification result of the UAV signal and the type information fusion related label, adjust and optimize the type information fusion unit; according to the direction identification result of the UAV signal and the direction information fusion related label, adjust and optimize the direction information fusion unit; according to the direction identification result of the UAV signal and the direction information fusion related label, adjust and optimize the direction information fusion unit.

[0067] Optionally, the step of performing information fusion and feature extraction on drone signals received by multiple receiving antennas to obtain drone signal time-frequency features, drone signal space-frequency features, and drone signal time-space features is specifically:

[0068] The number of receiving antennas of the drone signal receiving device is K, where k represents the index of the receiving antenna, k=0, 1, 2, ... K-1. In the same time period, assuming that each receiving antenna Rx of the drone signal receiving device k Each receives a discrete signal r with a total number of samples Q k (q), r k (q) is the receiving antenna Rx k Received drone signal r k (t) discrete sampling, q = 0, 1, ... Q-1;

[0069] r k (0)~r k Each r in (Q-1) k (q) are divided into M1 groups of frequency domain sub-signals according to time, and the length of each frequency domain sub-signal is N1, Q = M1 × N1; the frequency domain sub-signals of M1 groups in different time periods are obtained by Fourier transform; according to the time dimension, the frequency domain sub-signals are spliced ​​into r k (q) The corresponding M1×N1 two-dimensional time-frequency feature matrix By using the noise estimation method, each r k (q) The corresponding time-frequency characteristics The noise floor information is combined with the denoising operation; the time-frequency characteristics of all receiving antennas are integrated through the maximum ratio combining method. Get the time-frequency characteristics S of the drone signal tf :

[0070]

[0071] Among them, (m1,n1) represents the matrix The index of the element in the (m1+1)th row and (n1+1)th column, m=0,1,...M-1, q=0,1,...Q-1, |·| 2 represents the absolute value square operation;

[0072] r k (0)~r k Each r in (Q-1) k (q) are divided into M2 groups of spatial domain sub-signals with a length of N2 according to time, Q = M2 × N2; spatial spectrum fusion is performed on the spatial domain sub-signals of all different receiving antennas in the same time period to obtain a first spatial spectrum signal, and the first spatial spectrum signal is spliced ​​into a M2 × N2 two-dimensional space-time feature matrix S according to the time dimension ts , S ts As the spatiotemporal characteristics of drone signals;

[0073] r k (0)~r k The frequency domain signal of each rk(q) in (Q-1) is divided into M3 groups of frequency domain sub-signals 2 of length N3 according to the frequency domain sub-carrier bandwidth, Q = M3 × N3; the frequency domain sub-signals 2 of all different receiving antennas in the same frequency band are spatially fused to obtain a second spatial spectrum signal, and the second spatial spectrum signal is spliced ​​into a M3 × N3 two-dimensional space-frequency feature matrix S according to the frequency dimension sf , S sf As the air-frequency characteristics of UAV signals.

[0074] According to a third aspect of the present invention, there is provided an electronic device, comprising a processor and a memory.

[0075] The memory is used to store codes and related data;

[0076] The processor is used to execute the code in the memory to implement the method involved in the first aspect and its optional solutions.

[0077] According to a fourth aspect of the present invention, there is provided a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method involved in the first aspect and its optional solutions.

[0078] The drone signal recognition system, method, device and medium based on deep learning provided by the present invention are based on the initial data of drone signal receiving devices with multiple receiving antennas, and information fusion and feature extraction are performed through a signal fusion feature extraction module to obtain drone signal time-frequency features, drone signal space-frequency features and drone signal time-space features. Based on the drone signal time-frequency features, drone signal space-frequency features and drone signal time-space features, the type, direction and frequency hopping are identified, thereby improving the noise resistance of the drone signal recognition system; the drone signal type, direction and frequency hopping recognition network module adopts a CNN recognition network, and the recognition ability of the CNN recognition network is used to ensure the recognition effect; the CNN recognition network module is used to identify the type information based on the drone signal time-frequency features and the drone signal time-space features angle decibel, the direction information is respectively identified based on the drone signal time-space features and the drone signal space-frequency features, and the frequency hopping information is respectively identified based on the drone signal time-frequency features and the drone signal space-frequency features, and the output results are then fused through a fusion network module, thereby greatly improving the drone signal recognition ability, so that the drone signal recognition system can stably and accurately identify the type, direction and frequency hopping information of the drone signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0080] Figure 1 Schematic diagram of a program module of a drone signal recognition system based on deep learning in one embodiment of the present invention Figure 1 ;

[0081] Figure 2 It is a schematic diagram of a UAV signal receiving device acquiring UAV signals and a signal fusion feature extraction module in one embodiment of the present invention;

[0082] Figure 3 It is a structural principle diagram of a CNN recognition network module and an optimization module in one embodiment of the present invention;

[0083] Figure 4 is a structural and principle diagram of a fusion network module in one embodiment of the present invention;

[0084] Figure 5 This is the process of the drone signal recognition method based on deep learning in one embodiment of the present invention Figure 1 ;

[0085] Figure 6This is the process of the drone signal recognition method based on deep learning in one embodiment of the present invention Figure 2 ;

[0086] Figure 7 It is a schematic diagram of the structure of an electronic device in one embodiment of the present invention.

[0087] Numbers in the figure: 10, UAV signal recognition system, 11, signal fusion feature extraction module; 12, CNN recognition network module; 121, first CNN recognition network unit, 122, second CNN recognition network unit, 123, third CNN recognition network unit, 124, first type fully connected layer network unit, 125, first frequency hopping fully connected layer network unit, 126, second type fully connected layer network unit, 127, first direction fully connected layer network unit, 128, second direction fully connected layer network unit, 129, second frequency hopping fully connected layer network unit, 13, fusion network module, 131, type information fusion module, 132, direction information fusion module, 133, frequency hopping information fusion module, 1411, first type information determiner, 1412, first frequency hopping information determiner; 20, electronic device, 21, processor, 22, memory, 23, bus. DETAILED DESCRIPTION

[0088] 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.

[0089] In the specification and claims of the present invention and the above-mentioned drawings (if any), the terms "first", "second", "third", "fourth", etc., and the nouns followed by "one", "two", etc., such as "signal one" and "signal two", are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0090] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0091] The drone signal recognition system 10 based on deep learning of the present invention is based on a drone signal receiving device with multiple receiving antennas and a drone. The drone signal receiving device has multiple receiving antennas with at least three receiving antennas, and the drone signal receiving device can be called a multi-channel monitoring device. For the signal sent by the drone, the drone signal receiving device uses several (usually all) receiving antennas to receive the drone signal at the same time. The number of receiving antennas used to receive drone signals is K, where k represents the index of the receiving antenna, and takes values ​​of 0, 1, 2, ..., K-1; it can be understood as the k+1th receiving antenna, Rx k represents the k+1th receiving antenna.

[0092] Figure 1 Schematic diagram of a module of a drone signal recognition system 10 based on deep learning in one embodiment of the present invention Figure 1 .

[0093] The deep learning-based drone signal recognition system 10 includes:

[0094] The signal fusion feature extraction module 11 is used to perform information fusion and feature extraction on the drone signals received by multiple receiving antennas of the drone signal receiving device to obtain the drone signal time-frequency features, drone signal time-space features and drone signal space-frequency features;

[0095] CNN recognition network module 12, used to analyze the time-frequency characteristics of the drone signal to obtain possible type information 1 of the signal and possible frequency hopping information 1 of the signal, analyze the time-space characteristics of the drone signal to obtain possible type information 2 of the signal and possible direction information 1 of the signal, and analyze the space-frequency characteristics of the drone signal to obtain possible direction information 2 of the signal and possible frequency hopping information 2 of the signal;

[0096] The fusion network module 13 is used to fuse the possible type information 1 of the signal and the possible type information 2 of the signal to output the type of the drone signal, to fuse the possible direction information 1 of the signal and the possible direction information 2 of the signal to output the direction of the drone signal, and to fuse the possible frequency hopping information 1 of the signal and the possible frequency hopping information 2 of the signal to output the frequency hopping of the drone signal.

[0097] Here, the order and specific method of information fusion and feature extraction described in the signal fusion feature extraction module 11 are not limited. Information fusion and feature extraction can be performed in sequence or crosswise. As an example but not a limitation, for each receiving antenna, the drone signal rk (t) feature extraction (Rx k The received drone signal is r k (t)), and then fuse the results of the feature extraction, and then perform feature extraction on the fusion results to obtain the time-frequency characteristics of the UAV signal. The space-frequency characteristics and time-space characteristics of the UAV signal can refer to the process of obtaining the time-frequency characteristics of the UAV signal.

[0098] The deep learning-based drone signal recognition system 10 of this embodiment is based on the initial data of drone signal receiving devices with multiple receiving antennas, and performs information fusion and feature extraction through the signal fusion feature extraction module 11 to obtain the drone signal time-frequency characteristics, drone signal space-frequency characteristics and drone signal time-space characteristics. Based on the drone signal time-frequency characteristics, drone signal space-frequency characteristics and drone signal time-space characteristics, the type, direction and frequency hopping are identified, thereby improving the noise resistance of the drone signal recognition system 10; the drone signal type, direction and frequency hopping recognition network module adopts a CNN recognition network, and uses the recognition ability of the CNN recognition network to ensure the recognition effect. The CNN recognition network module 12 identifies the type information based on the time-frequency characteristics of the drone signal and the time-space characteristics of the drone signal, and then fuses the possible type information 1 of the signal and the possible type information 2 of the signal through the fusion network module 13 to obtain the final type recognition result. The CNN recognition network module 12 identifies the direction information based on the time-space characteristics of the drone signal and the space-frequency characteristics of the drone signal, and then fuses the possible direction information 1 of the signal and the possible direction information 2 of the signal through the fusion network module 13 to obtain the final direction recognition result. The CNN recognition network module 12 identifies the frequency hopping information based on the time-frequency characteristics of the drone signal and the space-frequency characteristics of the drone signal, and then fuses the possible frequency hopping information 1 of the signal and the possible frequency hopping information 2 of the signal through the fusion network module 13 to obtain the final frequency hopping recognition result. Through such a multi-angle data basis and fusion recognition method, the drone signal recognition capability is greatly improved, and the accuracy and stability of the drone signal recognition system 10 are improved, so that the drone signal recognition system 10 can stably and accurately identify the type, direction and frequency hopping information of the drone signal.

[0099] Figure 2 , Figure 3 and Figure 4 It is a schematic diagram of the specific structure and principle of the modules of the drone signal recognition system 10 based on deep learning in one embodiment of the present invention.

[0100] In one implementation, please refer to Figure 1 ,exist Figure 1On the basis of, the system also includes an optimization module for adjusting and optimizing the CNN recognition network module 12; and specifically introduces the specific structure, function and principle of the modules involved in the system.

[0101] Figure 2 The schematic diagram of the module 11 for extracting the drone signal and signal fusion features is provided for the drone signal receiving device.

[0102] The drone signal recognition system 10 may include a signal fusion feature extraction module 11. The signal fusion feature extraction module 11 may be located in a drone signal receiving device, and the specific location is not limited.

[0103] The signal fusion feature extraction module 11 is used to perform information fusion and feature extraction on the drone signals received by multiple receiving antennas of the drone signal receiving device to obtain the drone signal time-frequency features, drone signal space-frequency features and drone signal time-space features.

[0104] The signal fusion feature extraction module 11 includes a first signal fusion feature extraction unit, a second signal fusion feature extraction unit and a third signal fusion feature extraction unit.

[0105] In the same time period, assuming that each receiving antenna Rx of the drone signal receiving device k A discrete signal r with a total number of samples Q can be received k (q), where r k (q) is r k (t), q = 0, 1, ... Q-1, Q is the total number of drone signals received by a receiving antenna in the same time period. It can be understood that, usually, the total number of drone signals received by different receiving antennas is equal.

[0106] The first signal fusion feature extraction unit is used to perform information fusion and feature extraction on the drone signals received by the multiple receiving antennas of the drone signal receiving device to obtain the drone signal time-frequency features. Specifically, it is used to divide the drone signals received by the multiple receiving antennas of the drone signal receiving device into several groups of frequency domain sub-signals according to time, perform Fourier transform together, splice the transformed frequency domain sub-signals into a matrix according to the time dimension, perform denoising operation on the matrix, and fuse the denoised matrix to obtain the drone signal time-frequency features. Specifically, in order to obtain Rx k Receive signal at r k (q) time-frequency characteristics, we need to convert r k (0)~r k Each r in (Q-1) k(q) are divided into M1 groups of frequency domain signals (called frequency domain sub-signals 1) according to time, and the length of each signal segment is N1, Q = M1 × N1. Secondly, the frequency domain sub-signals 1 are subjected to Fourier transform to obtain M1 groups of frequency domain signals of different time periods (called transformed frequency domain sub-signals 1). According to the dimension of time, these transformed frequency domain sub-signals 1 are spliced ​​into r k (q) The corresponding M1×N1 two-dimensional time-frequency feature matrix Then, the noise estimation method is used to obtain each r k (q) The corresponding time-frequency characteristics The noise floor information is combined with the denoising operation to reduce the impact of the noise floor. Finally, the time-frequency characteristics of all receiving antennas are fused through the maximum ratio combining method. Get the time-frequency characteristics S of the drone signal tf The final result:

[0107]

[0108] Among them, (m1,n1) represents the matrix The index of the element in the (m1+1)th row and (n1+1)th column, m=0,1,…M-1, q=0,1,…Q-1, |·| 2 Represents the absolute value square operation.

[0109] The UAV signal time-frequency characteristics S tf As the input of CNN recognition network module 12.

[0110] The second signal fusion feature extraction unit is used to perform information fusion and feature extraction on the drone signals received by multiple receiving antennas of the drone signal receiving device to obtain the drone signal space-frequency features. Specifically, it is used to divide the drone signals received by multiple receiving antennas of the drone signal receiving device into several groups of spatial domain sub-signals according to time, perform spatial spectrum fusion on the spatial domain sub-signals of all different receiving antennas in the same time period, obtain the first spatial spectrum signal, and splice the first spatial spectrum signal according to the time dimension to obtain the drone signal spatiotemporal features. Specifically, the drone signal spatiotemporal features are obtained in a similar way to the drone signal time-frequency features. k (0)~r k Each rk(q) in (Q-1) is divided into M2 groups of signals of length N2 (called spatial domain sub-signals) according to time, Q = M2×N2. On this basis, the spatial domain sub-signals of all different receiving antennas in the same time period are spatially fused to obtain the first spatial spectrum signal, that is, the spatiotemporal features in a specific time period. Finally, these first spatial spectrum signals are spliced ​​into a two-dimensional spatiotemporal feature matrix S of M2×N2 according to the time dimension. ts , that is, output the spatiotemporal characteristics S of the drone signalts , as the input of CNN recognition network module 12.

[0111] The third signal fusion feature extraction unit is used to perform information fusion and feature extraction on the drone signals received by multiple receiving antennas of the drone signal receiving device to obtain the spatiotemporal features of the drone signals. Specifically, it is used to divide the drone signals received by multiple receiving antennas of the drone signal receiving device into several groups of frequency domain sub-signals 2 according to the frequency domain sub-carrier bandwidth, perform spatial spectrum fusion on the frequency domain sub-signals 2 of all different receiving antennas in the same frequency band to obtain the second spatial spectrum signal, and splice the second spatial spectrum signal according to the frequency dimension to obtain the spatial-frequency features of the drone signal. Specifically, compared with the method of obtaining the spatiotemporal features of the drone signal, the calculation of the spatial-frequency features requires first obtaining r k (0)~r k The frequency domain signal of each rk(q) in (Q-1) is then k (0)~r k The frequency domain signal of each rk(q) in (Q-1) is divided into M3 groups of frequency domain signals of length N3 (called frequency domain sub-signal 2) according to the frequency domain sub-carrier bandwidth, Q = M3 × N3. Then, the frequency domain sub-signal 2 of all different receiving antennas in the same frequency band is spatially fused to obtain the second spatial spectrum signal, that is, the space-frequency feature in a specific frequency band. Finally, these second spatial spectrum signals are spliced ​​into a two-dimensional space-frequency feature matrix S of M3 × N3 according to the frequency dimension. sf , that is, output the UAV signal space-frequency characteristics S sf , as the input of CNN recognition network module 12.

[0112] Figure 3 The structure and schematic diagram of the CNN identification network module 12 and the optimization module.

[0113] The input of the CNN recognition network module 12 is the information fused and enhanced by the signal fusion feature extraction module 11. The CNN recognition network module 12 outputs all possible types, frequency hopping and directions. The optimization module optimizes the CNN recognition network module 12 using the judgment result of the judge. The judgment result is usually one type among all possible types, one frequency hopping information among all possible frequency hopping information, and one direction among all possible directions. For example, in a time-frequency diagram (characterizing time-frequency characteristics), the type of signal can be determined by the law of how the bandwidth and signal frequency change over time. For example, the signal type can be a broadband signal, a narrowband signal, or a frequency-hopping signal. The frequency-hopping information related to the signal can be determined, and the frequency-hopping information includes: center frequency information; in a time-space diagram (characterizing time-space characteristics), the law of how the spatial intensity changes over time can be used to determine the signal propagation direction and speed through the slope of the signal, where the direction is the propagation direction, such as forward propagation and reverse propagation, and the type of signal can be determined through the waveform shape, such as periodic waves, pulse signals, and random signals; in a space-frequency diagram (characterizing space-frequency characteristics), the type of signal can be determined by observing the frequency distribution through spatial position, such as AM modulation and FM modulation, and the frequency-hopping information of the signal can be determined through the frequency attenuation in space, such as the attenuation of different frequency signals as they propagate in space.

[0114] The CNN recognition network module 12 includes: a first CNN recognition network unit 121, a first type of fully connected layer network unit 124, a first frequency hopping fully connected layer network unit 125, a second CNN recognition network unit 122, a second type of fully connected layer network unit 126, a first direction fully connected layer network unit 127, a third CNN recognition network unit 123, a second direction fully connected layer network unit 128 and a second frequency hopping fully connected layer network unit 129.

[0115] The first CNN recognition network unit 121 inputs the time-frequency characteristics of the drone signal, and the first CNN recognition network unit 121 is used to analyze the time-frequency information of the signal according to the time-frequency characteristics of the drone signal and output it. The first category fully connected layer network unit 124 inputs the time-frequency information of the signal output by the first CNN recognition network unit 121, and is used to integrate the time-frequency information, output the possible category information of the signal, and output it to the optimization module and the fusion network module 13. The first frequency hopping fully connected layer network unit 125 inputs the time-frequency information of the signal output by the first CNN recognition network unit 121, and is used to integrate the time-frequency information, output the possible frequency hopping information of the signal, and output it to the optimization module and the fusion network module 13.

[0116] Figure 3The example in which the first CNN recognition network unit 121 and the first category fully connected layer network unit 124 estimate 10 possible categories, and the first CNN recognition network unit 121 and the first frequency hopping fully connected layer network unit 125 estimate 8 possible frequency hopping information, also illustrates all possible results estimated by other recognition network units and fully connected layer network units.

[0117] The second CNN recognition network unit 122 inputs the spatiotemporal features of the drone signal, and is used to analyze the spatiotemporal information of the signal according to the spatiotemporal features of the drone signal and output it. The second type fully connected layer network unit 126 inputs the spatiotemporal information of the signal output by the second CNN recognition network unit 122, and is used to integrate the spatiotemporal information of the signal output by the second CNN recognition network unit 122, and output possible type information 2 of the signal, and output it to the optimization module and the fusion network module 13. The first direction fully connected layer network unit 127 inputs the spatiotemporal information of the signal output by the second CNN recognition network unit 122, and is used to integrate the spatiotemporal information of the signal output by the second CNN recognition network unit 122, and output possible direction information 1 of the signal, and output it to the optimization module and the fusion network module 13.

[0118] The third CNN recognition network unit 123 inputs the drone signal space-frequency feature analysis signal, which is used to analyze the space-frequency information of the signal according to the drone signal space-frequency feature and output it.

[0119] The second-direction fully connected layer network unit 128 inputs the space-frequency information of the signal output by the third CNN recognition network unit 123, and is used to integrate the space-frequency information of the signal output by the third CNN recognition network unit 123, and output the possible direction information 2 of the output signal to the optimization module and the fusion network module 13.

[0120] The second frequency hopping fully connected layer network unit 129 inputs the space-frequency information of the signal output by the third CNN identification network unit 123, and is used to integrate the space-frequency information of the signal output by the third CNN identification network unit 123, and output the possible frequency hopping information 2 of the output signal to the optimization module and the fusion network module 13.

[0121] The optimization module is specifically used to: adjust and optimize the first CNN recognition network unit 121 and the first type fully connected layer network unit 124 according to the output of the first type fully connected layer network unit 124; adjust and optimize the first CNN recognition network unit 121 and the first frequency hopping fully connected layer network unit 125 according to the output of the first frequency hopping fully connected layer network unit 125; adjust and optimize the second CNN recognition network unit 122 and the second type fully connected layer network unit 126 according to the output of the second type fully connected layer network unit 126; adjust and optimize the second CNN recognition network unit 122 and the first direction fully connected layer network unit 127 according to the output of the first direction fully connected layer network unit 127; adjust and optimize the third CNN recognition network unit 123 and the second frequency hopping fully connected layer network unit 129 according to the output of the second frequency hopping fully connected layer network unit 129; adjust and optimize the third CNN recognition network unit 123 and the second direction fully connected layer network unit 128 according to the output of the second direction fully connected layer network unit 128.

[0122] In this embodiment, in order to optimize the CNN recognition network module 12, it is necessary to select unique label sample information related to the decision estimation result of the judge, and transmit this unique label sample information and the decision result as input information to the loss function. Afterwards, the loss function will adjust the CNN recognition network unit and the fully connected layer network unit according to this information, so that the fully connected layer network unit re-outputs all possible results of a certain dimension, thereby achieving the purpose of iterative optimization of the network. After each iteration of training the network, the CNN recognition network unit will re-output all possible results of a certain feature information in a certain dimension through the fully connected layer network unit (such as information based on the time-frequency feature output type and frequency hopping information in two dimensions; based on the time-space feature output type and direction information in two dimensions, based on the space-frequency feature output direction and frequency hopping information information in two dimensions).

[0123] Specifically, the optimization module includes:

[0124] The first optimization unit is used to use the first category information judger 1411 to judge the possible category information 1 of the signal, obtain the category information 1 determined by the judgment, input the category information 1 determined by the judgment and the category label information 1 into the first category loss loss function, and use the first category loss loss function to optimize the CNN recognition network module 12. That is, the first optimization unit includes the first category information judger 1411, the first category loss loss function calculation unit and the first model update unit. The first category information judger 1411 uses the first category information judger 1411 to judge the possible category information 1 of the signal, and obtains the category information 1 determined by the judgment. The judge is used to calculate the average difference between the predicted value and the true value. Here, the output result of the first category information judger 1411 is a category. The input of the first category loss loss function calculation unit is the category information 1 determined by the judgment and the category label information 1, and the first loss function value is output. The first model update unit uses the first loss function value to adjust the parameters of the first CNN recognition network unit 121 and the parameters of the first category fully connected layer network unit 124.

[0125] The second optimization unit is used to use the first frequency hopping information determiner 1412 to determine the possible frequency hopping information 1 of the signal, obtain the frequency hopping information 1 determined by the determination, input the frequency hopping information 1 determined by the determination and the frequency hopping label information 1 into the first frequency hopping loss loss function, use the first frequency hopping loss function to optimize the CNN recognition network module 12, adjust and optimize the first CNN recognition network unit 121 and the first frequency hopping fully connected layer network unit 125, that is, adjust the parameters of the first CNN recognition network unit 121 and the first frequency hopping fully connected layer network unit 125.

[0126] The third optimization unit is used to use the second category information determiner to determine the possible category information 2 of the signal, obtain the category information 2 determined by the determination, input the category information 2 determined by the determination and the category label information 2 into the second category loss function, and use the second category loss function to optimize the CNN recognition network module 12, that is, adjust the parameters of the second CNN recognition network unit 122 and the second category fully connected layer network unit 126.

[0127] The fourth optimization unit is used to use the first direction information determiner to determine the possible direction information 1 of the signal, obtain the direction information 1 determined by the determination, input the direction information 1 determined by the determination and the direction label information 1 into the first direction loss function, and use the first direction loss function to optimize the CNN recognition network module 12, that is, adjust the parameters of the second CNN recognition network unit 122 and the first direction fully connected layer network unit 127.

[0128] The fifth optimization unit is used to use the second direction information determiner to determine the possible direction information 2 of the signal, obtain the direction information 2 determined by the determination, input the direction information 2 determined by the determination and the direction label information 2 into the second direction loss loss function, and use the second direction loss loss function to optimize the CNN recognition network module 12, that is, adjust the parameters of the third CNN recognition network unit 123 and the second direction fully connected layer network unit 128.

[0129] The sixth optimization unit is used to use the second frequency hopping information determiner to determine the possible frequency hopping information 2 of the signal, obtain the frequency hopping information 2 determined by the determination, input the frequency hopping information 2 determined by the determination and the frequency hopping label information 2 into the second frequency hopping loss function, and use the second frequency hopping loss function to optimize the CNN recognition network module 12, that is, adjust the parameters of the third CNN recognition network unit 123 and the second frequency hopping fully connected layer network unit 129.

[0130] The specific structures of the second to sixth optimization units can all refer to the first optimization unit.

[0131] Figure 4 It is a structure and principle diagram of the fusion network module 13.

[0132] See also Figure 4 , the fusion network module 13 includes:

[0133] The type information fusion module 131 is used to fuse the possible type information 1 of the signal and the possible type information 2 of the signal to output the type recognition result of the drone signal;

[0134] The direction information fusion module 132 is used to fuse the possible direction information 1 of the signal and the possible direction information 2 of the signal to output the direction recognition result of the drone signal;

[0135] The frequency hopping information fusion module 133 is used to fuse the possible frequency hopping information 1 of the signal and the possible frequency hopping information 2 of the signal to output the frequency hopping identification result of the drone signal.

[0136] Here, the fusion network module 13 is retrained through the possible results of the fully connected layer network unit and the corresponding fusion network training labels (the relevant labels), and obtains the direction, type, and frequency hopping information, and finally outputs the only final result in a specific dimension (direction, type, and frequency hopping information dimension) to realize the signal recognition of the drone. By fusing the relevant results of different features, the accuracy and robustness of signal recognition are improved.

[0137] Correspondingly, the fusion network module 13 has the function of adjusting and optimizing itself according to the relevant labels.

[0138] The type information fusion module 131 includes a type information fusion unit for fusing possible type information 1 of the signal and possible type information 2 of the signal to output the type identification result of the drone signal. The direction information fusion module 132 includes a direction information fusion unit for fusing possible direction information 1 of the signal and possible direction information 2 of the signal to output the direction identification result of the drone signal. The frequency hopping information fusion module 133 includes a frequency hopping information fusion unit for fusing possible frequency hopping information 1 of the signal and possible frequency hopping information 2 of the signal to output the frequency hopping identification result of the drone signal.

[0139] The type information fusion module 131 includes a first adjustment and optimization unit, which is used to adjust and optimize the type information fusion unit according to the type recognition result of the drone signal and the type information fusion related tags. That is, the type information fusion module 131 has the function of adjusting and optimizing itself.

[0140] The direction information fusion module 132 includes a second adjustment and optimization unit, which is used to adjust and optimize the direction information fusion unit according to the direction recognition result of the drone signal and the direction information fusion related label. That is, the direction information fusion module 132 has the function of adjusting and optimizing itself.

[0141] The frequency hopping information fusion module 133 includes a third adjustment and optimization unit, which is used to adjust and optimize the frequency hopping information fusion unit according to the frequency hopping identification result of the drone signal and the frequency hopping information fusion related label. That is, the frequency hopping information fusion module 133 has the function of adjusting and optimizing itself.

[0142] The above-mentioned second adjustment and optimization unit, the second adjustment and optimization unit, and the third adjustment and optimization unit can adjust the parameters of the fusion module based on the fusion result of the corresponding fusion module and the loss function calculated based on the relevant labels.

[0143] This embodiment makes full use of the receiving diversity characteristics of multi-channel multi-antenna equipment. First, by performing deep fusion and feature extraction on drone signals received by multiple receiving antennas, feature information with higher gain can be obtained, thereby improving the noise resistance of the drone signal recognition system 10. Then, by fusing the output information of different feature information in the same dimension, the scheme can fully enjoy the performance gain brought by the diversity of different feature information, and improve the accuracy and robustness of drone signal recognition results. Finally, the advanced CNN deep learning technology and the diversity characteristics of fused information are comprehensively used, and a good recognition effect is achieved without the need for additional hardware equipment. This scheme selects the information fusion method for recognition, and at the cost of slightly increasing the complexity of the system algorithm, achieves the purpose of greatly improving the drone signal recognition capability, and uses deep learning to improve the accuracy of drone signal recognition, and improves the accuracy and robustness of the drone signal recognition system 10, so that the drone signal recognition system 10 can stably and accurately identify the type, direction and frequency hopping information of the drone signal.

[0144] Figure 5 The process of drone signal recognition method based on deep learning Figure 1 .

[0145] See also Figure 5 ,The drone signal recognition method includes:

[0146] Obtaining drone signals received by multiple receiving antennas of a drone signal receiving device;

[0147] Information fusion and feature extraction are performed on the drone signals received by multiple receiving antennas to obtain the drone signal time-frequency features, drone signal space-frequency features, and drone signal time-space features;

[0148] The time-frequency characteristics of the UAV signal are analyzed to obtain possible type information of the signal and possible frequency hopping information of the signal;

[0149] The spatiotemporal characteristics of the drone signal are analyzed to obtain possible type information 2 of the signal and possible direction information 1 of the signal;

[0150] The spatial frequency characteristics of the UAV signal are analyzed to obtain possible direction information 2 and possible frequency hopping information 2 of the signal;

[0151] The possible type information 1 of the signal and the possible type information 2 of the signal are merged to output the type of the drone signal;

[0152] The possible direction information 1 of the signal and the possible direction information 2 of the signal are fused to output the direction of the drone signal;

[0153] The possible frequency hopping information one of the signal and the possible frequency hopping information two of the signal are fused to output the frequency hopping of the drone signal.

[0154] It can be understood that the execution order of the steps in this embodiment is not limited, and is based on being reasonable and feasible, and other processing flows and functions other than the above examples are not excluded.

[0155] Figure 6 The process of drone signal recognition method based on deep learning Figure 2 .

[0156] See also Figure 6 , compared to Figure 5 In this embodiment, the step of analyzing the time-frequency characteristics of the drone signal to obtain possible type information 1 of the signal and possible frequency hopping information 1 of the signal, the step of analyzing the time-space characteristics of the drone signal to obtain possible type information 2 of the signal and possible direction information 1 of the signal, and the step of analyzing the space-frequency characteristics of the drone signal to obtain possible direction information 2 of the signal and possible frequency hopping information 2 of the signal are all implemented by the CNN recognition network module 12; the drone signal recognition method also includes the step of adjusting and optimizing the CNN recognition network module 12.

[0157] In this embodiment, the step of fusing the possible type information 1 of the signal and the possible type information 2 of the signal to output the type of the drone signal, the step of fusing the possible direction information 1 of the signal and the possible direction information 2 of the signal to output the direction of the drone signal, and the step of fusing the possible frequency hopping information 1 of the signal and the possible frequency hopping information 2 of the signal to output the frequency hopping of the drone signal are all implemented by the fusion network module 13; the drone signal identification method also includes the step of adjusting and optimizing the fusion network module 13.

[0158] It should be noted that the embodiments of the drone signal recognition system 10 may also correspond to embodiments applicable to the drone signal recognition method, so they are not described in detail in the embodiments of the drone signal recognition method.

[0159] In summary, the drone signal identification method provided in this embodiment can stably and accurately identify the type, direction and frequency hopping information of the drone signal.

[0160] Figure 7 It is a schematic diagram of the structure of an electronic device in one embodiment of the present invention.

[0161] Please refer to Figure 7 , an electronic device 20 is provided, comprising:

[0162] a processor 21; and

[0163] A memory 22, used to store executable instructions of the processor;

[0164] The processor 21 is configured to execute the above-mentioned method by executing the executable instructions.

[0165] The processor 21 can communicate with the memory 22 via a bus 23 .

[0166] This embodiment further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the above-mentioned method is implemented.

[0167] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The UAV signal recognition system based on deep learning is characterized by: include: The signal fusion feature extraction module is used to perform information fusion and feature extraction on the drone signals received by multiple receiving antennas of the drone signal receiving device to obtain the drone signal time-frequency features, drone signal time-space features and drone signal space-frequency features; The CNN recognition network module is used to analyze the time-frequency characteristics of the drone signal to obtain possible type information 1 of the signal and possible frequency hopping information 1 of the signal, analyze the time-space characteristics of the drone signal to obtain possible type information 2 of the signal and possible direction information 1 of the signal, and analyze the space-frequency characteristics of the drone signal to obtain possible direction information 2 of the signal and possible frequency hopping information 2 of the signal; The fusion network module is used to fuse the possible type information of the signal one and the possible type information of the signal two to output the type of the drone signal, to fuse the possible direction information of the signal one and the possible direction information of the signal two to output the direction of the drone signal, and to fuse the possible frequency hopping information of the signal one and the possible frequency hopping information of the signal two to output the frequency hopping of the drone signal.

2. The deep learning-based drone signal recognition system according to claim 1, characterized in that: The system also includes an optimization module for adjusting and optimizing the CNN recognition network module.

3. The UAV signal recognition system based on deep learning according to claim 1, characterized in that: The CNN recognition network module includes: The first CNN recognition network unit is used to analyze the time-frequency information of the signal according to the time-frequency characteristics of the drone signal and output it; A first type of fully connected layer network unit, used to integrate the time-frequency information of the signal output by the first CNN recognition network unit, and output possible type information of the signal one; A first frequency hopping fully connected layer network unit is used to integrate the time-frequency information of the signal output by the first CNN recognition network unit, and output possible frequency hopping information of the signal; The second CNN recognition network unit is used to analyze the spatiotemporal information of the signal according to the spatiotemporal characteristics of the drone signal and output it; The second type of fully connected layer network unit is used to integrate the spatiotemporal information of the signal output by the second CNN recognition network unit, and output possible type information 2 of the signal; The first directional fully connected layer network unit is used to integrate the spatiotemporal information of the signal output by the second CNN recognition network unit, and output possible directional information of the signal one; The third CNN recognition network unit is used to analyze the signal's space-frequency information according to the UAV signal's space-frequency characteristics and output it; The second directional fully connected layer network unit is used to integrate the space-frequency information of the signal output by the third CNN recognition network unit, and output possible direction information 2 of the signal; The second frequency hopping fully connected layer network unit is used to integrate the space-frequency information of the signal output by the third CNN recognition network unit, and output the possible frequency hopping information 2 of the signal.

4. The UAV signal recognition system based on deep learning according to claim 2, characterized in that: The system also includes an optimization module, which is used to: adjust and optimize the first CNN recognition network unit and the first type of fully connected layer network unit according to the output of the first type of fully connected layer network unit; adjust and optimize the first CNN recognition network unit and the first frequency hopping fully connected layer network unit according to the output of the first frequency hopping fully connected layer network unit; adjust and optimize the second CNN recognition network unit and the second type of fully connected layer network unit according to the output of the second type of fully connected layer network unit; adjust and optimize the second CNN recognition network unit and the first direction fully connected layer network unit according to the output of the first direction fully connected layer network unit; adjust and optimize the third CNN recognition network unit and the second frequency hopping fully connected layer network unit according to the output of the second frequency hopping fully connected layer network unit; adjust and optimize the third CNN recognition network unit and the second direction fully connected layer network unit according to the output of the second direction fully connected layer network unit.

5. The deep learning-based drone signal recognition system according to claim 4, characterized in that: The optimization module includes: A first optimization unit is used to use a first category information determiner to determine possible category information 1 of the signal, obtain category information 1 determined by the determination, input the category information 1 determined by the determination and category label information 1 into a first category loss function, and use the first category loss function to optimize the first CNN recognition network unit and the first category fully connected layer network unit; A second optimization unit is used to use the first frequency hopping information determiner to determine the possible frequency hopping information 1 of the signal, obtain the determined frequency hopping information 1, input the determined frequency hopping information 1 and the frequency hopping label information 1 into the first frequency hopping loss function, and use the first frequency hopping loss function to optimize the first CNN recognition network unit and the first frequency hopping fully connected layer network unit; A third optimization unit is used to use the second category information determiner to determine the possible category information 2 of the signal, obtain the category information 2 determined by the determination, input the category information 2 determined by the determination and the category label information 2 into the second category loss function, and use the second category loss function to optimize the second CNN recognition network unit and the second category fully connected layer network unit; A fourth optimization unit is used to use the first direction information determiner to determine the possible direction information 1 of the signal, obtain the direction information 1 determined by the determination, input the direction information 1 determined by the determination and the direction label information 1 into the first direction loss function, and use the first direction loss function to optimize the second CNN recognition network unit and the first direction fully connected layer network unit; A fifth optimization unit is used to use the second direction information determiner to determine the possible direction information 2 of the signal, obtain the direction information 2 determined by the determination, input the direction information 2 determined by the determination and the direction label information 2 into the second direction loss function, and use the second direction loss function to optimize the third CNN recognition network unit and the second direction fully connected layer network unit; The sixth optimization unit is used to use the second frequency hopping information determiner to determine the possible frequency hopping information 2 of the signal, obtain the frequency hopping information 2 determined by the determination, input the frequency hopping information 2 determined by the determination and the frequency hopping label information 2 into the second frequency hopping loss function, and use the second frequency hopping loss function to optimize the third CNN recognition network unit and the second frequency hopping fully connected layer network unit.

6. The UAV signal recognition system based on deep learning according to claim 1, characterized in that: The fusion network module includes: A type information fusion module is used to fuse possible type information 1 of the signal and possible type information 2 of the signal to output the type recognition result of the drone signal, and is used to adjust and optimize the type information fusion unit according to the type recognition result of the drone signal and the type information fusion related labels; A direction information fusion module is used to fuse possible direction information 1 of the signal and possible direction information 2 of the signal to output a direction recognition result of the UAV signal, and is used to adjust and optimize the direction information fusion unit according to the direction recognition result of the UAV signal and the direction information fusion related label; The frequency hopping information fusion module is used to fuse the possible frequency hopping information 1 and the possible frequency hopping information 2 of the signal to output the frequency hopping identification result of the drone signal, and is used to adjust and optimize the frequency hopping information fusion unit according to the frequency hopping identification result of the drone signal and the frequency hopping information fusion related labels.

7. The deep learning-based drone signal recognition system according to claim 1, characterized in that: The signal fusion feature extraction module comprises: The first signal fusion feature extraction unit is used to divide the drone signal received by multiple receiving antennas of the drone signal receiving device into a plurality of groups of frequency domain sub-signals according to time, and perform Fourier transform on the frequency domain sub-signals after the transform according to the time dimension, and after performing a denoising operation on the matrix, fuse the denoised matrix to obtain the time-frequency feature of the drone signal; The second signal fusion feature extraction unit is used to divide the drone signal received by multiple receiving antennas of the drone signal receiving device into a plurality of groups of spatial domain sub-signals according to time, perform spatial spectrum fusion on the spatial domain sub-signals of all different receiving antennas in the same time period to obtain a first spatial spectrum signal, and splice the first spatial spectrum signal according to the time dimension to obtain the spatiotemporal features of the drone signal; The third signal fusion feature extraction unit is used to divide the drone signal received by multiple receiving antennas of the drone signal receiving device into several groups of frequency domain sub-signals 2 according to the frequency domain sub-carrier bandwidth, perform spatial spectrum fusion on the frequency domain sub-signals 2 of all different receiving antennas in the same frequency band to obtain a second spatial spectrum signal, and splice the second spatial spectrum signal according to the frequency dimension to obtain the drone signal space-frequency feature.

8. A drone signal recognition method based on deep learning, characterized in that: include: Obtaining drone signals received by multiple receiving antennas of a drone signal receiving device; Information fusion and feature extraction are performed on the drone signals received by multiple receiving antennas to obtain the drone signal time-frequency features, drone signal time-space features, and drone signal space-frequency features; The time-frequency characteristics of the UAV signal are analyzed to obtain possible type information of the signal and possible frequency hopping information of the signal; The spatiotemporal characteristics of the drone signal are analyzed to obtain possible type information 2 of the signal and possible direction information 1 of the signal; The spatial frequency characteristics of the UAV signal are analyzed to obtain possible direction information 2 and possible frequency hopping information 2 of the signal; The possible type information 1 of the signal and the possible type information 2 of the signal are merged to output the type of the drone signal; The possible direction information 1 of the signal and the possible direction information 2 of the signal are fused to output the direction of the drone signal; The possible frequency hopping information one of the signal and the possible frequency hopping information two of the signal are fused to output the frequency hopping of the drone signal.

9. The method for identifying drone signals based on deep learning according to claim 8, characterized in that: The analysis of the time-frequency characteristics of the drone signal, the analysis of the time-space characteristics of the drone signal, and the analysis of the space-frequency characteristics of the drone signal are all implemented by a CNN recognition network module. The method also includes the step of adjusting and optimizing the CNN recognition network module.

10. The method for identifying drone signals based on deep learning according to claim 9, characterized in that: The CNN recognition network module includes: The first CNN recognition network unit is used to analyze the time-frequency information of the signal according to the time-frequency characteristics of the drone signal and output it; A first type of fully connected layer network unit, used to integrate the time-frequency information of the signal output by the first CNN recognition network unit, and output possible type information of the signal one; A first frequency hopping fully connected layer network unit is used to integrate the time-frequency information of the signal output by the first CNN recognition network unit, and output possible frequency hopping information of the signal; The second CNN recognition network unit is used to analyze the spatiotemporal information of the signal according to the spatiotemporal characteristics of the drone signal and output it; The second type of fully connected layer network unit is used to integrate the spatiotemporal information of the signal output by the second CNN recognition network unit, and output possible type information 2 of the signal; The first directional fully connected layer network unit is used to integrate the spatiotemporal information of the signal output by the second CNN recognition network unit, and output possible directional information of the signal one; The third CNN recognition network unit is used to analyze the signal's space-frequency information according to the UAV signal's space-frequency characteristics and output it; The second directional fully connected layer network unit is used to integrate the space-frequency information of the signal output by the third CNN recognition network unit, and output possible direction information 2 of the signal; The second frequency hopping fully connected layer network unit is used to integrate the space-frequency information of the signal output by the third CNN recognition network unit, and output the possible frequency hopping information 2 of the signal.

11. The method for identifying drone signals based on deep learning according to claim 10, characterized in that: The steps of adjusting and optimizing the CNN recognition network module are specifically as follows: Using a first category information determiner to determine possible category information 1 of the signal, obtaining category information 1 determined by the determination, inputting the category information 1 determined by the determination and category label information 1 into a first category loss function, and using the first category loss function to optimize a first CNN recognition network unit and a first category fully connected layer network unit; Using a first frequency hopping information determiner to determine possible frequency hopping information 1 of the signal, obtaining determined frequency hopping information 1, inputting the determined frequency hopping information 1 and frequency hopping label information 1 into a first frequency hopping loss function, and using the first frequency hopping loss function to optimize a first CNN recognition network unit and a first frequency hopping fully connected layer network unit; Using the second category information determiner to determine the possible category information 2 of the signal, obtaining the category information 2 determined by the determination, inputting the category information 2 determined by the determination and the category label information 2 into the second category loss function, and using the second category loss function to optimize the second CNN recognition network unit and the second category fully connected layer network unit; Using the first direction information determiner to determine the possible direction information 1 of the signal, obtaining the direction information 1 determined by the determination, inputting the direction information 1 determined by the determination and the direction label information 1 into the first direction loss function, and using the first direction loss function to optimize the second CNN recognition network unit and the first direction fully connected layer network unit; The second direction information determiner is used to determine the possible direction information 2 of the signal to obtain the determined direction information 2, the determined direction information 2 and the direction label information 2 are input into the second direction loss function, and the second direction loss function is used to optimize the third CNN recognition network unit and the second direction fully connected layer network unit; The second frequency hopping information determiner is used to determine the possible frequency hopping information 2 of the signal to obtain the determined frequency hopping information 2, the determined frequency hopping information 2 and the frequency hopping label information 2 are input into the second frequency hopping loss function, and the second frequency hopping loss function is used to optimize the third CNN recognition network unit and the second frequency hopping fully connected layer network unit.

12. The method for identifying drone signals based on deep learning according to claim 8, characterized in that: The step of fusing possible type information 1 of the signal with possible type information 2 of the signal is implemented by a type information fusion module, the step of fusing possible direction information 1 of the signal with possible direction information 2 of the signal is implemented by a direction information fusion module, the step of fusing possible frequency hopping information 1 of the signal with possible frequency hopping information 2 of the signal is implemented by a frequency hopping information fusion module, and the method also includes the step of adjusting and optimizing the type information fusion module, the direction information fusion module and the frequency hopping information fusion module.

13. The method for identifying drone signals based on deep learning according to claim 12, characterized in that: The steps of adjusting and optimizing the type information fusion module, the direction information fusion module and the frequency hopping information fusion module are specifically as follows: according to the type recognition result of the drone signal and the type information fusion related label, adjust and optimize the type information fusion unit; according to the direction recognition result of the drone signal and the direction information fusion related label, adjust and optimize the direction information fusion unit; according to the direction recognition result of the drone signal and the direction information fusion related label, adjust and optimize the direction information fusion unit.

14. The method for identifying drone signals based on deep learning according to claim 8, characterized in that: The steps of performing information fusion and feature extraction on the drone signals received by multiple receiving antennas to obtain the drone signal time-frequency features, drone signal space-frequency features, and drone signal time-space features are specifically as follows: The number of receiving antennas of the drone signal receiving device is K, where k represents the index of the receiving antenna, k=0, 1, 2, ... K-1. In the same time period, assuming that each receiving antenna Rx of the drone signal receiving device k Each receives a discrete signal r with a total number of samples Q k (q), r k (q) is the receiving antenna Rx k Received drone signal r k (t) discrete sampling, q = 0, 1, ... Q-1; r k (0)~r k Each r in (Q-1) k (q) are divided into M1 groups of frequency domain sub-signals according to time, and the length of each frequency domain sub-signal is N1, Q = M1 × N1; the frequency domain sub-signals of M1 groups in different time periods are obtained by Fourier transform; according to the time dimension, the frequency domain sub-signals are spliced ​​into r k (q) The corresponding M1×N1 two-dimensional time-frequency feature matrix By using the noise estimation method, each r k (q) The corresponding time-frequency characteristics The noise floor information is combined with the denoising operation; the time-frequency characteristics of all receiving antennas are integrated through the maximum ratio combining method. Get the time-frequency characteristics S of the drone signal tf : Among them, (m1,n1) represents the matrix The index of the element in the (m1+1)th row and (n1+1)th column, m=0,1,...M-1, q=0,1,...Q-1, |·| 2 represents the absolute value square operation; r k (0)~r k Each r in (Q-1) k (q) are divided into M2 groups of spatial domain sub-signals with a length of N2 according to time, Q = M2 × N2; spatial spectrum fusion is performed on the spatial domain sub-signals of all different receiving antennas in the same time period to obtain a first spatial spectrum signal, and the first spatial spectrum signal is spliced ​​into a M2 × N2 two-dimensional space-time feature matrix S according to the time dimension ts , S ts As the spatiotemporal characteristics of drone signals; r k (0)~r k Each r in (Q-1) k The frequency domain signals of (q) are divided into M3 groups of frequency domain sub-signals 2 with a length of N3 according to the frequency domain sub-carrier bandwidth, Q = M3 × N3; the frequency domain sub-signals 2 of all different receiving antennas in the same frequency band are spatially fused to obtain a second spatial spectrum signal, and the second spatial spectrum signal is spliced ​​into a M3 × N3 two-dimensional space-frequency feature matrix S according to the frequency dimension sf , S sf As the air-frequency characteristics of UAV signals.

15. An electronic device, characterized in that: Including processor and memory, The memory is used to store codes and related data; The processor is used to execute the code in the memory to implement the method according to any one of claims 8 to 14.

16. A storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the method according to any one of claims 8 to 14 is implemented.

Images