A method for constructing a recognition model of drone signals and a corresponding recognition method and system

Through time-frequency graph processing and machine learning model, the feature vectors of the drone signal are extracted, and the identification and classification of non-cooperative drone signals are realized, the problem of insufficient recognition capabilities in the existing technology is solved, and multiple units are realized Individual identification of similar drones.

CN114580476BActive Publication Date: 2024-11-08BEIJING QINGRUI WEIHANG TECH DEV CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202210199577.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-02
Publication Date
2024-11-08
Estimated Expiration
2042-03-02

AI Technical Summary

Technical Problem

The prior art cannot effectively identify and classify frequency hopping control signals of non-cooperative drones, and it is particularly difficult to achieve individual identification of multiple drones of the same type.

Method used

By obtaining the time-frequency diagram of the drone signal, performing connectivity domain recognition and clustering analysis, building a convolutional neural network for signal classification, and combining the machine learning classifier model, feature vectors are extracted for identification and classification of drone signals.

Benefits of technology

Effective detection, classification and identification of signals of multiple drone types and similar types has been achieved, and the problem of insufficient recognition capabilities in the prior art has been solved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114580476B_ABST
    Figure CN114580476B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for constructing an identification model of a drone and a corresponding identification method and system. The present invention constructs a neural network model for drone identification using radio signals, trains the neural network model, and applies it to the identification of drone signals. The present invention converts radio signals into time-frequency graphs including at least time, frequency, and power information, respectively, performs feature extraction and substitutes them into a convolutional neural network model, the model is first trained using training data, and then substitutes test data to identify the type of the target drone; connective domain identification is performed on the time-frequency graph, and then category similarity judgment is performed based on the time-frequency graph, and then the ID part of the radio signal is determined, and feature extraction is performed and substituted into a machine learning classifier model, the model is first trained using training data, and then substitutes test data to identify the individual of the target drone.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of radio signal reconnaissance, and in particular to a method for constructing a recognition model of unmanned aerial vehicle signals and a method and system for recognizing unmanned aerial vehicle control signals. Background Art

[0002] In recent years, due to its wide application in many fields and the potential safety hazards it causes, the detection and identification technology for non-cooperative drones has received widespread attention. The perception and analysis of drone remote control signals is an important way to achieve non-cooperative drone identification. In order to improve the anti-interference ability of communication between the controller and the drone, small and micro drones mostly use frequency hopping signals as control signals: by periodically changing the carrier frequency, the frequency of interference in a specific frequency band is reduced, and the difficulty of detection and identification in non-cooperative situations is increased. There is no technology in the prior art that can effectively detect and identify non-cooperative frequency hopping control signals.

[0003] Traditional radio signal detection and recognition methods mainly follow the following two steps: First, the target signal is detected from the background noise and other interference. This step mainly uses methods such as matched filtering and energy detection. Then, the signal is classified by methods such as signal demodulation and manual selection of signal features. However, the above methods usually have the following problems:

[0004] 1. Traditional signal detection and identification methods require the design of dedicated detection and identification methods for specific types of radio signals. They have poor detection capabilities for non-cooperative signals and are difficult to expand.

[0005] 2. Existing signal detection methods based on machine learning transform the signal detection problem into an image recognition problem and use the traditional convolutional neural network structure for image recognition tasks, which fails to fully utilize the characteristics of the signal itself.

[0006] 3. For drone signal identification scenarios that use frequency hopping signals as control signals, the use of time-frequency diagrams cannot achieve individual identification of multiple drones of the same type with the same signal format (frequency hopping pattern and modulation mode).

[0007] Therefore, existing signal detection and recognition techniques cannot solve the detection, classification and identification problems of non-cooperative UAV individuals. Summary of the invention

[0008] Aiming at the problems existing in the detection and identification of non-cooperative UAVs, the present invention proposes an identification method for individual UAVs, which can realize the detection, classification and identification of control signals of multiple non-cooperative UAVs of different and same types.

[0009] Specifically, on the one hand, the present invention provides a method for constructing a recognition model of a drone, characterized in that the method comprises the following steps:

[0010] (1) Acquire radio signals of multiple drones, and convert the radio signals of each drone into a time-frequency graph including at least time, frequency, and power information;

[0011] (2) Identifying connected domains on the time-frequency graphs to obtain connected regions in each of the time-frequency graphs;

[0012] (3) Perform cluster analysis on the connected regions in each time-frequency graph, wherein the categories of the connected regions in each time-frequency graph are determined, and the frequency hopping signal of the corresponding time-frequency graph is determined for each connected region category, and the frequency components of the category signal are retained to obtain the time-frequency graph for the category frequency hopping signal;

[0013] (4) constructing a convolutional neural network for drone identification based on the time-frequency diagram of radio signals, processing the sample data sets representing the labels according to steps (1)-(3) to obtain the corresponding time-frequency diagrams, and bringing them into the convolutional neural network, training the convolutional neural network model, and using the trained convolutional neural network to classify drone signals based on the time-frequency diagram;

[0014] (5) For multiple pulses of the same type of signal, use their signal envelopes to perform cross-correlation operations with the sample signal envelopes, perform peak detection on the cross-correlation results, estimate the radio signal pulse position corresponding to the current time-frequency diagram, and estimate the ID part of the radio pulse signal, and perform feature extraction on this part;

[0015] (6) Constructing a machine learning classifier model based on feature vectors, and training the machine learning classifier model using sample data.

[0016] In a preferred implementation, the convolutional neural network includes a first convolutional neural sub-network, a second convolutional neural sub-network and a fully connected network, the first convolutional neural sub-network includes multiple convolutional layers, the second convolutional neural sub-network includes spaced convolutional layers and downsampling layers, and the outputs of the first convolutional neural sub-network and the second convolutional neural sub-network are flattened and concatenated as input to the fully connected network.

[0017] In another preferred implementation, the method also includes estimating the signal structure of pulses from similar drone signals, determining the synchronization signal, ID signal and control signal of the drone signal based on the degree of similarity of different signal parts of multiple pulses of multiple drones and the same drone, and performing feature extraction on the ID signal.

[0018] In another preferred implementation, the signal structure estimation includes: determining the similarity of any two signal segments based on the following formula:

[0019] Among them, s it [n]、s jt [n] respectively represent the signals from s i [n] and s j [n] fragments at the same position, μ it , μ jt Respectively represent s it [n]、s jt The mean of [n].

[0020] In another preferred implementation, if the similarities of a certain signal part of multiple pulses from multiple drones and the same drone are both higher than or lower than a threshold, then the signal part is preliminarily determined to be a synchronization signal; if the similarities between the signal segments corresponding to this part of multiple pulses from multiple drones and the same drone are both lower than the second threshold, then the signal part is preliminarily determined to be a control signal; if the similarities between the signal segments corresponding to a certain signal part from multiple drones are lower than the third threshold, and the similarities between the signal segments corresponding to this part of multiple pulses from the same drone are higher than the fourth threshold, then the signal part is determined to be an ID signal.

[0021] In another preferred implementation, feature extraction is performed on the radio signal segment determined to be an ID signal to generate a feature vector, which is then substituted into a machine learning classifier model.

[0022] On the other hand, the present invention provides a method for identifying drones using a model constructed by the method, characterized in that the method comprises: obtaining a radio signal of a drone to be identified, processing the radio signal according to steps (1) to (3) to obtain a time-frequency graph of the current signal, and substituting the time-frequency graph into a trained convolutional neural network to classify the drone signal;

[0023] Performing feature extraction on the classified drone signal according to step (5);

[0024] The extracted features are input into the trained machine learning classifier model for unanswered signal recognition.

[0025] On the other hand, the present invention provides a drone signal recognition system, characterized in that the system comprises: a signal receiving and collecting module, a signal processing module, a signal type recognition module and a classification recognition module, wherein:

[0026] The signal receiving and collecting module is used to receive and collect training data and measured data of drone signals;

[0027] The signal processing module is used to convert the radio signal into a time-frequency graph including at least time, frequency and power information, obtain the connected area in each of the time-frequency graphs, perform cluster analysis on the connected areas in each of the time-frequency graphs to determine the category of the connected areas in each of the time-frequency graphs, determine the frequency hopping signal of the corresponding time-frequency graph according to the category similarity of the connected areas, retain the frequency components of the same type of signals, and obtain the time-frequency graph for the current frequency hopping signal;

[0028] The signal type recognition module is integrated with a convolutional neural network for identifying the type of drone based on the time-frequency diagram of the radio signal, and is used to identify the type of the drone signal based on the time-frequency diagram;

[0029] The signal individual identification module is used to perform a cross-correlation operation on multiple pulses of the same type of signal using their signal envelopes and the sample signal envelope, perform peak detection on the cross-correlation results, estimate the radio signal pulse position corresponding to the current time-frequency diagram, determine the ID part of the radio signal, extract features from the part, and use a feature vector-based machine learning classifier model to perform drone signal identification based on the extracted features.

[0030] Aiming at the complex scenario of non-cooperative UAV identification, the present invention proposes an identification method for UAV control signals, solves the three problems existing in the prior art, and effectively realizes the classification and identification of different types of UAVs and the individual identification of multiple UAVs of the same type. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 : A schematic diagram of the flow of the drone individual identification method provided by the present invention.

[0032] Figure 2 This is a sample signal with obvious interference signal.

[0033] Figure 3 is the time-frequency diagram of the example signal obtained by short-time Fourier transform.

[0034] Figure 4 It is the result of clustering the connected areas in the time-frequency diagram according to their similarity. After clustering, the signals corresponding to other clusters are removed, and only the frequency hopping signal (i.e., category 1 in the figure) is retained to obtain the time-frequency diagram corresponding to the current frequency hopping signal.

[0035] Figure 5 Schematic diagram of the convolutional neural network structure of step 2.2 in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0037] Generally speaking, the method of the present invention includes three main stages: 1. signal conversion stage, 2. type recognition stage, and 3. individual recognition stage.

[0038] The signal mentioned in the present invention refers to the radio signal of a drone or other target identification device. The radio signal can be received and collected using existing conventional radio receiving equipment. The process of receiving and collecting it will not be described in detail here. The radio signals mentioned below are all digital radio signals that have been converted from analog to digital, and the originally acquired radio signals are time domain signals.

[0039] Below, the signal processing, model construction and use process are described in detail in stages.

[0040] 1. Signal conversion stage

[0041] The signal conversion phase mainly includes the following steps:

[0042] 1.1. The time domain radio signal is converted into a two-dimensional time-frequency diagram through STFT transformation.

[0043] Among them, STFT transform (Short-Time Fourier Transform) is a mathematical transformation that converts the time domain signal containing time and signal strength information into a time-frequency diagram containing time, frequency and power information by moving the window function w[n] in the time domain signal s[n] and calculating the power spectrum at different times. rec [n], its time-frequency diagram can be expressed as:

[0044] SG rec [ω,n]=|∑ n‘ s rec [n′]w[n′-n]e -jωn′ | 2 .

[0045] Where n and ω represent time and frequency respectively, n′ is the summation variable with respect to time, and j is an imaginary number symbol.

[0046] 1.2. The frequency hopping pulse information in the time-frequency diagram is obtained through frequency hopping pattern recognition.

[0047] The frequency hopping pattern recognition includes improving the signal-to-noise ratio, generating connected regions, and unsupervised clustering steps. In this embodiment, for the time-frequency diagram, the effect of improving the signal-to-noise ratio can be obtained by cutting off the background noise below the threshold, wherein the threshold can be automatically obtained by the characteristics of the energy distribution curve in the time-frequency diagram. A connected region refers to an image region composed of pixels whose pixel values ​​in the image are greater than zero and whose positions are adjacent (four-connected or eight-connected). Connected region generation refers to obtaining the connected regions in the time-frequency diagram by traversing the image (i.e., the time-frequency diagram) and stacking them in sequence (or other methods). Unsupervised clustering refers to clustering the connected regions extracted from the time-frequency diagram into several categories according to their similarities based on the shape and other features of the connected regions; signals clustered into the same category and having similar time intervals are regarded as frequency hopping signals; retaining the frequency components of the same type of signals, removing other frequency components, and obtaining a time-frequency diagram corresponding to the current frequency hopping signal (i.e., eliminating interference and other non-similar signals).

[0048] 1.3. Obtain high-precision position information of frequency-hopping pulses through fine pulse positioning. Specifically, the synchronization of signals is achieved through the cross-correlation information between multiple pulses of the same signal. Furthermore, for frequency-hopping pulses with changing carrier frequencies, the envelope is selected to replace the original waveform of the signal for cross-correlation to achieve synchronization to avoid zero correlation between pulses with different carrier frequencies due to orthogonal frequencies. For the time domain IQ (In-phase, Quadrature, i.e., in-phase, orthogonal) signal s[n] = s I [n]+js Q [n], where s I [n] and s Q [n] represents the in-phase part and the orthogonal part of the signal respectively, and its envelope e[n] can be expressed as:

[0049] e[n]=|s I [n]+js Q [n]|

[0050] For a single-channel real signal s[n], its envelope e[n] can be obtained by Hilbert transform as follows:

[0051] e[n]=|s[n]+j∑ n′ h[nn′]s[n′]|

[0052] Where n′ is the summation variable with respect to time, and h[n] is the Hilbert filter:

[0053]

[0054] Then, by finding the peak value of the correlation value between the sample signal envelope (known signal or reference signal) and the target signal envelope, the fine pulse position estimation of the target signal is achieved:

[0055]

[0056] 2. Type identification stage

[0057] The type identification phase includes:

[0058] 2.1. Training the convolutional neural network model. Before classification, the two-dimensional time-frequency graph generated by the sample data is imported into the convolutional neural network model for training, and the trained convolutional neural network model is obtained as a classifier. The sample data can be specified by the user; it can also be new data encountered during system operation, which is used to identify the drone the next time it appears.

[0059] 2.2. Use the trained convolutional neural network model to classify the signal. In the signal classification and recognition stage, the time-frequency diagram obtained by STFT conversion of the time domain signal collected by the signal acquisition device is used as input, and the trained convolutional neural network model is imported for classification and recognition to obtain the classification result of the current signal type.

[0060] The convolutional neural network model mentioned in the present embodiment 2.1 and 2.2 adopts the two-stream convolutional neural network model (Two-Stream Convolutional Neural Network) specially designed for the time-frequency graph features of the present invention. It consists of two convolutional neural networks with three layers and five layers (excluding the input layer) respectively and a three-layer (excluding the input layer) fully connected network. The first convolutional neural network stream includes input, the first convolutional layer C1, the second convolutional layer C2 and the third convolutional layer C3 along the input-to-output direction; the second convolutional neural network stream includes input, the fourth convolutional layer C4, the fifth convolutional layer C5, the first downsampling layer P1, the sixth convolutional layer C6 and the second downsampling layer P2 along the input-to-output direction. The outputs of the two convolutional neural networks are flattened and spliced ​​as the input of the fully connected network. The fully connected network includes input, the first fully connected layer F1, the second fully connected F2 and output respectively along the input-to-output direction. Compared with the traditional neural network structure, this neural network structure has the following innovations and advantages: (1) Through two convolutional neural networks with different scales and steps (in convolutional neural networks, scale refers to the size of the convolution kernel, which represents the size of the receptive field of the convolution layer; step size refers to the step size of the convolution operation, which represents the accuracy of extraction. The "two convolutional neural networks with different scales and steps" here refer to the "three-layer and five-layer" convolutional neural networks mentioned above), full use is made of the different scale features on the time-frequency diagram caused by the frequency hopping pattern and signal format of different frequency hopping signals; (2) By utilizing the computational properties of the convolutional layer and the linear rectifier function (Rectified Linear Unit, ReLU), the denoising process of the spectrum diagram is implicitly realized by using the convolutional layer; (3) The number of layers of the convolutional neural network is reduced through the dual-stream structure, and a random zeroing operation is added to the fully connected network to reduce the possibility of overfitting.

[0061] Through the above steps, pulses from the same type of drone information are screened out. Combining the pulse positions obtained in the preprocessing process with the classification information obtained by the neural network structure, the pulses from the same type of drone are output to the individual recognition stage for individual recognition.

[0062] The individual identification phase includes:

[0063] 1. In the training stage, the signal structure of the pulses from the same type of drone signals is estimated by estimating the signal structure. The inventors noticed that the same type of drones use the same standard of control signals. For the signal part used for synchronization, the similarity between the signal segments corresponding to this part of multiple pulse signals from multiple drones and the same drone (here refers to the part corresponding to the synchronization signal in a single frequency hopping pulse) is high; for the control signaling part, the similarity between the signal segments corresponding to this part of multiple pulses from multiple drones and the same drone is low; for the ID part, the similarity between the signal segments corresponding to this part from multiple drones is low, and the similarity between the signal segments corresponding to this part from multiple pulses from the same drone is high. This difference in similarity is used to automatically obtain a rough estimate of the signal structure, and further extract features from the ID part of the signal. During the training process, the similarity threshold can be set according to the model training situation, and the threshold can be adjusted based on the model feedback situation.

[0064] The similarity in step 1 can be obtained by the following coefficient. Its value range is -1 to +1. For two sets of variables X and Y, the larger the absolute value of the coefficient between them, the more obvious the linear relationship between variables X and Y; the smaller the absolute value, the weaker the linear relationship between variables X and Y. i [n] and s j [n] fragments at the same position it [n]、s jt [n], its similarity corr ij,t It can be expressed as:

[0065]

[0066] where μ it , μ jt Respectively represent s it [n]、s jt The mean of [n].

[0067] 2. Extract features from the most likely ID fragment of a single pulse in the radio signal to generate a feature vector. For signals with amplitude modulation, the fractal dimension can be used to extract features from its envelope.

[0068] The fractal dimension mentioned in step 2 is a measure of the irregularity of complex shapes, reflecting the effectiveness of the space occupied by complex shapes. For the time domain signal s, its fractal dimension D can be expressed as:

[0069]

[0070] Where ε represents the side length of the small cube, and N(s,ε) represents the number of small cubes covering the measured shape.

[0071] 3.1 Training the machine learning classifier model. Before classification, the feature vector generated by the sample data is imported into the machine learning classifier model for training, and the trained machine learning classifier model is obtained as the classifier. The sample data can be specified by the user; it can also be new data encountered during system operation, which is used to identify the drone the next time it appears.

[0072] 3.2 Use the trained machine learning classifier model to classify the signal. In the signal individual recognition stage, the features of similar signal pulses extracted by the feature extraction module are used as input, and the trained machine learning classifier model is imported for classification and recognition to obtain the recognition result of the current signal individual.

[0073] In a preferred implementation, the machine learning classifier model in steps 3.1 and 3.2 uses a support vector machine based on radial basis kernel function and one-versus-one (OVO) method to implement multi-classification. For a given training set X={(x 1 ,y 1 ),...(x N ,y N )}, where x n ∈R K is a K-dimensional vector, class label y n ={1,2,...,M}, first train the i-th and j-th binary support vector machines {w ij ,b ij}, where index i,j∈{1,2,...,M}, i<j:

[0074]

[0075]

[0076]

[0077]

[0078] The subscript t represents the index of the sample in the union of the i-th class and the j-th class. is the slack variable, C is the penalty term for the slack variable, and φ(x) represents the nonlinear mapping from the input space to the feature space. In this example, the radial basis function is used to construct the kernel function to achieve this mapping. have:

[0079]

[0080] Where σ is the parameter of the radial basis kernel function. For each pair of indexes {i, j}, we train them separately and obtain a total of A binary support vector machine. Then, for the new data sample x new , from the above The two-class support vector machines predict the classification one by one, and use the voting results as the classification results; take the two-class support vector machines of the i-th and j-th categories as an example, if the prediction of the two-class support vector machine pair is the i-th category, then the i-th category gets +1 vote, otherwise the j-th category gets +1 vote; then, the votes of each category are counted, and the category with the most votes is the category of this group of support vector machines for the new data sample x new prediction; in particular, if there is a tie, select the class with the smaller index as the new data sample x new prediction.

[0081] In addition, this embodiment also provides a non-cooperative UAV control signal recognition system based on radio frequency machine learning, including a signal receiving and acquisition module, a signal preprocessing module, a signal type recognition module, a signal individual recognition module, and a classification recognition result display module.

[0082] The signal receiving and collecting module includes an antenna, a down converter and a digital sampling device, which is used to sample the radio signal and obtain the time domain signal.

[0083] The signal sample preprocessing module includes a time-frequency diagram conversion module, a frequency hopping pattern recognition module, and a fine frequency hopping position estimation module, which are used to convert the time domain data of the radio signal into two-dimensional time-frequency diagram data through STFT transformation, and obtain the pulse position estimation by combining the time domain data sample and the time-frequency domain data sample information.

[0084] The signal type recognition module includes a signal type recognition classifier module and a pulse screening module; the signal type recognition classifier module is used to take the time-frequency diagram information output by the time-frequency diagram conversion module as input, import the trained neural network model to classify and recognize the signal, and output the classification and recognition results to the classification and recognition result display module and the pulse screening module; the pulse screening module combines the classification and recognition results of the signal, and inputs the pulses in the time domain data into the signal individual recognition module according to the category for individual recognition of the signal.

[0085] The signal individual identification module includes a pulse structure estimation module, a feature extraction module, and a signal individual identification classifier module; the pulse structure estimation module is used to perform statistical analysis on similar signal pulses output by the pulse screening module to give a rough estimate of the signal structure; the feature extraction module is used to extract features from the most likely ID part obtained by the pulse structure estimation to achieve data dimensionality reduction; the signal individual identification classifier module uses the feature vector output by the feature extraction module as input, imports the trained machine learning classifier model for signal individual identification, and outputs the individual identification results to the classification identification result display module. The classification identification result display module is used to display the classification identification results and individual identification results of the signal.

[0086] Although the principles of the present invention are described in detail above in conjunction with the preferred embodiments of the present invention, those skilled in the art should understand that the above embodiments are merely explanations of the exemplary implementations of the present invention, and are not intended to limit the scope of the present invention. The details in the embodiments do not constitute limitations on the scope of the present invention, and any obvious changes such as equivalent transformations, simple replacements, etc. based on the technical solution of the present invention, without departing from the spirit and scope of the present invention, fall within the protection scope of the present invention.

Claims

1. A method for constructing a recognition model of a drone, characterized in that: The method comprises the following steps: (1) Acquire radio signals of multiple drones, and convert the radio signals of each drone into a time-frequency graph including at least time, frequency, and power information; (2) Identifying connected domains on the time-frequency graphs to obtain connected regions in each of the time-frequency graphs; (3) Perform cluster analysis on the connected regions in each time-frequency graph, wherein the categories of the connected regions in each time-frequency graph are determined, and the frequency hopping signal of the corresponding time-frequency graph is determined for each connected region category, and the frequency components of the category signal are retained to obtain the time-frequency graph for the category frequency hopping signal; (4) constructing a convolutional neural network for drone identification based on the time-frequency diagram of radio signals, processing the sample data sets representing the labels according to steps (1)-(3) to obtain the corresponding time-frequency diagrams, and bringing them into the convolutional neural network, training the convolutional neural network model, and using the trained convolutional neural network to classify drone signals based on the time-frequency diagram; (5) For multiple pulses of the same type of signal, use their signal envelopes to perform cross-correlation operations with the sample signal envelopes, perform peak detection on the cross-correlation results, estimate the radio signal pulse position corresponding to the current time-frequency diagram, and estimate the ID part of the radio signal, and perform feature extraction on this part; (6) constructing a machine learning classifier model based on feature vectors, and training the machine learning classifier model using sample data; The method also includes estimating the signal structure of pulses from similar drone signals, determining the segments of the synchronization signal, ID signal and control signal of the drone signal based on the similarity of different signal parts of multiple pulses of multiple drones and the same drone, and extracting features from the ID signal. The similarity of any two signal segments is determined based on the following formula: Among them, s it [n]、s jt [n] respectively represent the signals from s i [n] and s j [n] fragments in the same position, ω it ,ω jt Respectively represent s it [n]、s jt The mean of [n].

2. The method according to claim 1, characterized in that: The convolutional neural network includes a first convolutional neural sub-network, a second convolutional neural sub-network and a fully connected network. The first convolutional neural sub-network includes multiple convolutional layers, the second convolutional neural sub-network includes spaced convolutional layers and downsampling layers, and the outputs of the first convolutional neural sub-network and the second convolutional neural sub-network are flattened and concatenated as inputs to the fully connected network.

3. The method according to claim 1, characterized in that If the similarities of a certain signal part of multiple pulses from multiple drones and the same drone are all higher than the first threshold, then the signal part is preliminarily determined to be a synchronization signal; if the similarities between the signal segments corresponding to this part of multiple pulses from multiple drones and the same drone are all lower than the second threshold, then the signal part is preliminarily determined to be a control signal; if the similarities between the signal segments corresponding to a certain signal part from multiple drones are lower than the third threshold, and the similarities between the signal segments corresponding to this part of multiple pulses from the same drone are higher than the fourth threshold, then the signal part is determined to be an ID signal.

4. The method according to claim 3, characterized in that The features of the radio signal segments determined to be ID signals are extracted to generate feature vectors, which are then substituted into the machine learning classifier model.

5. A method for identifying drones using the model constructed by the method of claim 1, characterized in that: The method comprises: obtaining a radio signal of a drone to be identified, processing the radio signal according to steps (1) to (3) to obtain a time-frequency diagram of the current signal, and substituting the time-frequency diagram into a trained convolutional neural network to classify the drone signal; Performing feature extraction on the classified drone signal according to step (5); The extracted features are input into the trained machine learning classifier model for unanswered signal recognition.

6. A drone signal recognition system, based on the method according to any one of claims 1 to 5, characterized in that: The system includes: a signal receiving and collecting module, a signal processing module, a signal type identification module, a signal individual identification module and a classification identification module, wherein: The signal receiving and collecting module is used to receive and collect training data and measured data of drone signals; The signal processing module is used to convert the radio signal into a time-frequency graph including at least time, frequency and power information, obtain the connected area in each of the time-frequency graphs, perform cluster analysis on the connected areas in each time-frequency graph to determine the category of the connected areas in each time-frequency graph, determine the frequency hopping signal of the corresponding time-frequency graph according to the category similarity of the connected areas, retain the frequency components of the same type of signal, and obtain the time-frequency graph for the current frequency hopping signal; the signal type recognition module is integrated with a convolutional neural network for drone type recognition based on the time-frequency graph of the radio signal, and is used to recognize the type of the drone signal based on the time-frequency graph; The signal individual identification module is used to perform a cross-correlation operation on multiple pulses of the same type of signal using their signal envelopes and the sample signal envelope, perform peak detection on the cross-correlation results, estimate the radio signal pulse position corresponding to the current time-frequency diagram, determine the ID part of the radio signal, extract features from the part, and use a feature vector-based machine learning classifier model to perform drone signal identification based on the extracted features.

Citation Information

Cited By

  • Unmanned aerial vehicle frequency system real-time identification method and system fusing transfer learning

    CN122365197A

  • An unmanned aerial vehicle frequency system real-time identification method and system based on fusion of transfer learning

    CN122365197B