Unmanned aerial vehicle object detection method based on neural network
Patent Information
- Application Number
- MYPI2025003912
- Authority / Receiving Office
- MY · MY
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2022-12-29
- Filing Date
- 2023-10-30
- Publication Date
- 2024-06-29
AI Technical Summary
Existing technology is difficult to accurately identify and distinguish multiple UAV signals in a low signal-to-noise ratio environment, and traditional methods cannot effectively extract the frequency and bandwidth information of UAV signals, resulting in low detection accuracy and consuming a lot of manpower and material resources.
The UAV target detection method based on neural network is used to construct the neural network model through the offline training stage. The three-dimensional time domain matrix, time-frequency matrix and IMF matrix are used for signal preprocessing and feature extraction, and the convolutional neural network and feature pyramid network are combined. Perform multi-scale fusion to achieve effective monitoring and control of UAV signals.
It improves the accuracy of UAV signal detection in low signal-to-noise ratio environments, can automatically identify and classify multiple UAVs, and directly obtain the center frequency, bandwidth and other information of the signal, avoiding errors caused by artificial feature design. , saving resources.
Smart Images

Figure PI2025003912_DRAWING_001
Abstract
Description
A UAV target detection method based on neural network
[0001] CROSS-REFERENCE TO RELATED APPLICATIONS
[0002] This application claims priority to Chinese patent application CN202211710125.4 filed on December 29, 2022, the entire contents of which are incorporated by reference into this application. Technical Field
[0003] The present invention relates to the field of drone detection technology, and in particular to a drone target detection method based on a neural network. Background Art
[0004] In recent years, drones, thanks to their miniaturization and high maneuverability, have been applied in various fields, such as geographic surveying and mapping, emergency rescue, and filming. Technological breakthroughs have also led to a rapid increase in the number of consumer-grade micro-drones. However, the development of drone technology and the industry has also brought difficulties to drone control. Unauthorized drone flights continue to occur, leading to numerous security incidents and confidentiality leaks. Therefore, effective drone monitoring and control is urgently needed.
[0005] Drones have been deployed in various fields, especially in specific areas such as mountainous and urban areas. While interference is less common in mountainous areas, the vast expanse of the landscape makes drone signals weak at long distances. In urban areas, the complex electromagnetic environment creates significant interference, making traditional radio monitoring technology very effective in detecting and identifying small civilian drones. For example, when WiFi signals and drone signals overlap, traditional methods cannot effectively distinguish them. In such low signal-to-noise ratio environments, traditional manual feature engineering methods struggle to accurately extract the characteristic parameters of these signals, resulting in suboptimal accuracy and requiring significant human and material resources, which is not worth the effort.
[0006] In addition, existing conventional neural network-based methods only classify and identify intercepted radio signals, but cannot extract the center frequency, bandwidth, and other information of the drone signal in the radio signal for further signal processing. When dealing with the situation where multiple drones exist at the same time, they cannot classify and identify each drone target separately, which limits their practical application.
[0007] Traditional manual feature engineering methods primarily intercept radio signals in space and analyze them for drone signals, primarily those from drone remote control or image transmission. By analyzing characteristic parameters such as the signal's frequency, bandwidth, frequency hopping period, and duration, they determine whether it belongs to a drone. This is then compared with drone signal features in a feature library to identify the type of drone. However, this method is not ideal in complex electromagnetic environments, such as urban areas. Low signal-to-noise ratios (SNRs) cause drone signals to alias or even drown out interference from Wi-Fi signals and noise, making signal feature extraction difficult or contaminating the extracted features, ultimately reducing detection accuracy. Furthermore, manually extracted features cannot guarantee comprehensive, accurate, and sufficient extraction of drone signal features, limiting further improvements in detection accuracy. Furthermore, manually designing feature parameters is time-consuming and labor-intensive, and can even result in erroneous feature parameters. While conventional neural network-based methods can identify the presence and type of drones, they lack information such as the frequency and bandwidth of the identified drone signals and cannot simultaneously detect multiple drones, limiting their practical application.
[0008] Summary of the Invention
[0009] The technical problem to be solved by the present invention is: in response to the technical problems existing in the prior art, the present invention provides a UAV target detection method based on a neural network with simple principle, easy operation and wide application range.
[0010] In order to solve the above technical problems, the present invention adopts the following technical solutions:
[0011] A method for detecting unmanned aerial vehicle targets based on a neural network, comprising:
[0012] Offline training phase: construct a training data set, build a neural network structure (i.e., neural network model), import the training data set into the neural network for model training, and form a detection model (i.e., drone detection model); the training data set includes several labeled sample data; a single sample data includes a three-dimensional time domain matrix Three-dimensional time-frequency matrix Three-dimensional IMF matrix and complete the labeling;
[0013] Online detection stage: actual detection is performed using the trained drone detection model; during the actual detection process, the collected signals are formed into a feature matrix set, which is sent to the detection model for analysis and processing, and the drone information is output, where the feature matrix set includes three feature matrices: a three-dimensional time domain matrix, a three-dimensional time-frequency matrix, and a three-dimensional IMF matrix.
[0014] As a further improvement of the method of the present invention: the offline training stage includes preparing a training data set: collecting original signal data and preprocessing the original data.
[0015] As a further improvement of the method of the present invention: the collecting of raw signal data includes:
[0016] Raw data of drone image transmission signals: Select different drones to fly within the detection area of the drone detection equipment, and use the detection equipment to collect the raw data of the drone image transmission signals;
[0017] Background environment data: Deploy drone detection equipment in different environments, and then repeatedly collect the raw data of the drone's image transmission signal to form background environment data.
[0018] As a further improvement of the method of the present invention: the raw data preprocessing is to preprocess the collected signal raw data, including:
[0019] Step S10: Analyze the original data and clean the dirty data;
[0020] Step S20: Sliding window interception is performed on the original signal data after analysis and cleaning to obtain H segmented signals, and then the H segmented signals are rearranged according to the time dimension to obtain a two-dimensional time domain signal matrix P of the fast and slow time dimensions. T ;
[0021] Step S30: Using empirical mode decomposition to decompose the two-dimensional time domain signal matrix P T Perform iterative decomposition to obtain the IMF matrix P IMF ;
[0022] Step S40: Two-dimensional time domain data matrix P T (i.e., the two-dimensional time domain signal matrix P T ) is Fourier transformed at K points in the fast time dimension and performed H times, and the matrix P is transformed into T The fast time dimension is converted from the time domain to the frequency domain to extract the frequency domain features of the signal; the K×H time-frequency matrix is obtained, that is, the time-frequency waterfall diagram P of the signal TF ;
[0023] Step S50: Obtain a three-dimensional time domain matrix Three-dimensional time-frequency matrix Three-dimensional IMF matrix It takes a single input sample and labels it, including the actual location of the drone, the type of aircraft, and the frequency range of the device detection.
[0024] As a further improvement of the method of the present invention: Step S30 includes:
[0025] Step S301: Solve the matrix P T The local maximum and minimum extreme points of the fast time dimension;
[0026] Step S302: using the cubic spline interpolation method to interpolate the maximum extreme point, connecting the maximum extreme point and the interpolation point to form an upper envelope; using the cubic spline interpolation method to interpolate the minimum extreme point, connecting the minimum extreme point and the interpolation point to form a lower envelope;
[0027] The upper envelope and the lower envelope are averaged to obtain a mean envelope line;
[0028] Step S303: From the matrix P T The intermediate signal is obtained by differentiating the fast time dimension of the mean envelope. If the intermediate signal satisfies the following conditions: the absolute difference between the number of zero-crossing points and the number of extreme points is less than or equal to 1, and the mean of the local maximum and minimum envelopes is equal to 0 (that is, the mean of the sum of the upper envelope and the lower envelope in the local area is 0), then it is an intrinsic mode function vector (that is, IMF vector);
[0029] Through the matrix P T The fast time dimension of the IMF vector is differentiated, and then step S303 is repeated to obtain the next IMF vector, thereby obtaining the intrinsic mode function vector IMF of the fast time dimension i , where i = 1, 2, 3...N, N is the number of IMF vectors;
[0030] Step S304: Repeat steps S301 to S303 to calculate the matrix P T All fast time dimensions are decomposed, and finally the (S×N)×H IMF matrix P is obtained IMF .
[0031] As a further improvement of the method of the present invention: Step S50 includes:
[0032] The drone detection equipment receives different directions from 0 to 360 degrees, uses X detection channels evenly distributed in the range of 0 to 360 degrees, and obtains X channels of original signal data at the same time Where ch is the channel number, ch = 1, 2..., X;
[0033] According to steps S20 to S40, a two-dimensional time domain signal matrix of X channels is obtained at the same moment. And the size is S×H, IMF matrix And the size is (S×N)×H, and the time-frequency matrix And the size is K×H;
[0034] Time-frequency matrix According to the principle of non-adjacent channel selection, N channels are extracted to form a three-dimensional matrix The matrix size is N×K×H;
[0035] IMF matrix Select according to non-adjacent channels and Based on the principle of channel difference, N channels are extracted to obtain N (S×N)×H matrices, and then the IMF corresponding to these N matrices is calculated. i The vectors are weighted and summed according to a certain weight, and the weight coefficient is a n , and finally get the three-dimensional matrix The matrix size is N×S×H; where i=1,2,3...N, n=1,2,3...N;
[0036] The time domain signal matrix According to the time-frequency matrix The principle of the same channel extracts N channels to form a three-dimensional matrix The matrix size is N×S×H.
[0037] As a further improvement of the method of the present invention: the neural network structure constructed in the offline training stage is a single-stage detection neural network, including an input layer, a backbone network, a connection network, and a detection head network; wherein:
[0038] The input layer is used to receive the data to be detected, and includes three branches, IN1 corresponding to the three-dimensional time domain matrix IN2 corresponds to the three-dimensional time-frequency matrix IN3 corresponds to the three-dimensional IMF matrix IN2 is connected to the backbone network, and IN1 and IN3 are connected to the detection head network across regions.
[0039] The backbone network uses a convolutional neural network to extract features from input data;
[0040] The connection network uses a feature pyramid network to perform multi-scale fusion and reuse of features extracted by the backbone network. The detection head network is used for final feature learning and then outputs detection information of the drone target and radio signal parameter information.
[0041] The detection head network is directly connected to IN1 and IN3, and is downsampled through an L-layer convolutional neural network to make it consistent with the feature map size when IN2 reaches the detection head network; using the channel attention mechanism, it is weighted multiplied with the classification feature map of the detection head network to fuse the extracted signal time domain features and the intrinsic modal features of the signal at different time scales with the features extracted by the main network of IN2 input.
[0042] As a further improvement of the method of the present invention, the process of training with the neural network in the offline training stage includes:
[0043] Send the constructed training sample set, that is, the feature matrix set of the signal, into the constructed neural network for training;
[0044] By fitting the actual data, the constructed validation sample set is input into the neural network for verification after each training cycle, and the accuracy of the network model is monitored in real time;
[0045] After the training is completed, the constructed test sample set is input into the neural network for testing to check the performance of the network model;
[0046] If the network model meets the expected requirements, a weight parameter is obtained through conversion, which is the detection model of the UAV target.
[0047] As a further improvement of the method of the present invention: the online detection stage includes:
[0048] Step S100: Importing the trained drone detection model into the actual detection equipment to obtain all radio signals within the detection range;
[0049] Step S200: pre-processing the intercepted radio signal to obtain a feature matrix set of the signal;
[0050] Step S300: Send the feature matrix set into the detection model for computational reasoning, and output information on whether a drone exists. If so, output information on each drone.
[0051] As a further improvement of the method of the present invention: Step S200 includes:
[0052] Step S2001: intercept, rearrange, and multi-channel weight the original time domain signal data to obtain a three-dimensional time domain matrix
[0053] Step S2002: Use FFT to perform fast and slow time-dimensional time-frequency transformation on the two-dimensional time-domain signal matrix to obtain a three-dimensional time-frequency matrix
[0054] Step S2003: Use empirical mode decomposition to extract the modal features of the signal at different time scales on the two-dimensional time domain signal matrix, and finally obtain a three-dimensional IMF matrix
[0055] Compared with the prior art, the advantages of the present invention are:
[0056] 1. The neural network-based drone target detection method of the present invention has a simple principle, easy operation, and a wide range of applications. It addresses the problems of traditional drone identification methods such as difficulty in feature extraction in low signal-to-noise ratio environments and incomplete feature extraction, resulting in low detection accuracy. At the same time, conventional neural network methods are difficult to obtain the categories and radio signal parameter information of multiple drones at the same time. The present invention preprocesses the intercepted radio signals to obtain the time-frequency spectrum of the signals, and then sends them to the designed detection neural network for target detection. Finally, it outputs whether there are drones in the signal, as well as information such as the types of each drone, the center frequency of the drone, the bandwidth of the drone, and the time width of the drone, to achieve effective monitoring and control of drone signals.
[0057] 2. The neural network-based drone target detection method of the present invention adopts neural network technology, which includes key technologies such as sample labeling method, neural network structure, and drone end-to-end target detection solution. It can also better extract drone signal features in low signal-to-noise ratio environments or weak signal environments, thereby improving detection accuracy.
[0058] 3. The neural network-based drone target detection method of the present invention is based on a method that cannot guarantee the comprehensive, correct and sufficient extraction of drone signal features by traditional manual feature engineering methods. The present invention uses a method of time-frequency transformation + multimodal decomposition + neural network to improve detection performance while avoiding the impact of errors caused by manual feature design on the final detection performance, avoiding the waste of manpower and material resources, and saving resources.
[0059] 4. The neural network-based drone target detection method of this invention not only identifies the presence of drones and classifies their types in intercepted radio signals, but also automatically and directly acquires information such as the center frequency and bandwidth of each drone signal when multiple drones are present. Through offline training and online detection, it ultimately outputs information such as the drone's model and radio parameters, ultimately enabling robust detection even in harsh environments. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] FIG1 is a schematic diagram showing the principle of the method of the present invention;
[0061] FIG2 is a diagram showing the architecture of the neural network of the present invention. DETAILED DESCRIPTION
[0062] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0063] As shown in FIG1 , a neural network-based drone target detection method of the present invention includes:
[0064] Offline training phase: construct a training data set, build a neural network structure (i.e., neural network model), import the training data set into the neural network for model training, and form a detection model (i.e., drone detection model); the training data set includes several labeled sample data; a single sample data includes a three-dimensional time domain matrix Three-dimensional time-frequency matrix Three-dimensional IMF matrix and complete the labeling;
[0065] Online detection stage: The trained drone detection model is used for actual detection. During the actual detection process, the collected signals are formed into a feature matrix set, which is sent to the detection model for analysis and processing, and the drone information is output. The feature matrix set includes three feature matrices: a three-dimensional time domain matrix, a three-dimensional time-frequency matrix, and a three-dimensional IMF matrix.
[0066] In a specific application example, during the online detection stage, analysis and processing are used to determine whether drone information exists. If so, the model, frequency, bandwidth, time width and other information of each drone are output. Ultimately, this solves the problem of effective detection of targets when the drone signal is weak or the signal-to-noise ratio is low, and multiple drones may exist at the same time.
[0067] In a specific application example, during the offline training phase, the training data set is first prepared, including collecting raw signal data and preprocessing the raw data.
[0068] The collected signal raw data may include but is not limited to the following:
[0069] Raw data of drone image transmission signals: Within the detection area of the drone detection equipment, different drones are selected to fly at different directions from 0 to 360 degrees and different distances from 0 to R kilometers, and the raw data of the drone image transmission signals are collected through the detection equipment.
[0070] Background environment data: Deploy drone detection equipment in different environments, including but not limited to typical scenes such as suburban fields, mountain forests, and urban high-rise buildings. Then repeatedly collect raw data from drone image transmission signals to enrich the background environment data.
[0071] The raw data preprocessing is to preprocess the collected signal raw data, which may include but is not limited to:
[0072] Step S10: Analyze the original data and clean the dirty data;
[0073] Eliminate signal data such as signal clutter caused by hardware issues of the receiving device, abnormally high background noise, abnormal drone signals, and strong horizontal interference of short-range remote control signals;
[0074] Clean useless data and filter out signal data such as the drone signal being completely submerged due to excessive interference, the drone flying distance exceeding the distance limit of the receiving device, the drone signal being unable to be received, the drone signal being completely overlapped with the WiFi signal, and the drone signal being completely overlapped with the base station signal.
[0075] Then the cleaned data is sorted to get x t According to the sampling frequency f0, the total sampling time T, and the data magnitude M, a fixed time step S is set for the original data to perform sliding window interception. The sliding window width is equal to the step S. The intercepted H segmented signals are rearranged according to the time dimension to obtain the two-dimensional time domain signal matrix P of the fast and slow time dimensions. T , that is, the two-dimensional time domain signal matrix of a single channel ch is the channel number, where width is S and height is H;
[0076] Step S20: Using empirical mode decomposition to decompose the two-dimensional time domain signal matrix P T Perform iterative decomposition;
[0077] The detailed process includes:
[0078] Step S201: Solve the matrix P T The two upper and lower envelopes of the fast time dimension, that is, first solve the local maximum extreme point and the minimum extreme point;
[0079] Step S202: using the cubic spline interpolation method to interpolate the maximum extreme point, connecting the maximum extreme point and the interpolation point to form an upper envelope; using the cubic spline interpolation method to interpolate the minimum extreme point, connecting the minimum extreme point and the interpolation point to form a lower envelope;
[0080] The upper envelope and the lower envelope are averaged to obtain a mean envelope line;
[0081] Step S203: From the matrix P T The intermediate signal is obtained by differentiating the fast time dimension from the mean envelope. If the difference between the number of zero-crossing points and the number of extreme points in the intermediate signal is less than or equal to 1 and the mean of the local maximum and minimum envelopes is equal to 0 (that is, the mean of the sum of the upper and lower envelopes in the local area is 0), then it is an Intrinsic Mode Function (IMF) vector. The zero-crossing point refers to the point where the intermediate signal is equal to 0.
[0082] Through the matrix P T The fast time dimension is differentiated from the IMF vector and the above steps are continued to obtain the IMF vector; the intrinsic mode function vector IMF of the fast time dimension is obtained. i (i=1,2,3...N), N is the number of IMF vectors;
[0083] Step S204: Repeat steps S201 to S203 to calculate the matrix P T All fast time dimensions are decomposed, and finally the (S×N)×H IMF matrix P is obtained IMF (i.e., the IMF matrix of a single channel ), where S is the matrix P T The number of fast time dimensions.
[0084] Step S30: Two-dimensional time domain signal matrix P T The fast time dimension of the matrix is transformed from the time domain to the frequency domain by performing a K-point Fourier transform and performing H times, so as to extract the frequency domain features of the signal and finally obtain a K×H time-frequency matrix (i.e., a two-dimensional time-frequency matrix of a single channel). ), that is, the time-frequency waterfall diagram of the signal P TF , where K is the number of Fourier transform points and H is the number of times Fourier transform is performed;
[0085] Step S40: Obtain a three-dimensional time domain matrix Three-dimensional time-frequency matrix Three-dimensional IMF matrix It is used as a single input sample (i.e., single sample data) and labeled, mainly marking the actual location of the drone, the type of aircraft, and the frequency range of the equipment detection.
[0086] The drone detection equipment receives signals from different directions from 0 to 360 degrees. Therefore, X detection channels are evenly distributed in the range of 0 to 360 degrees, and the original signal data of X channels can be obtained at the same time. (ch=0,1,2...X). Therefore, the above steps S20 to S40 can further obtain the two-dimensional time domain signal matrix of X channels at the same time. And the size is S×H, IMF matrix And the size is (S×N)×H, and the time-frequency matrix And the size is K×H.
[0087] Time-frequency matrix According to the principle of non-adjacent channel selection, N channels are extracted to form a three-dimensional matrix The matrix size is N×K×H.
[0088] IMF matrix Select according to non-adjacent channels and Based on the principle of channel difference, N channels are extracted to obtain N (S×N)×H matrices, and then the IMF corresponding to these N matrices is calculated. i(i=1,2,3...N) vectors are weighted and summed according to certain weights, and the weight coefficient is a n (n=1,2,3...N), and finally get a three-dimensional matrix The matrix size is N×S×H.
[0089] The time domain signal matrix According to the time-frequency matrix The principle of the same channel extracts N channels to form a three-dimensional matrix The matrix size is N×S×H.
[0090] In a specific application example, during the offline training phase, the neural network structure constructed is a single-stage detection neural network, which consists of four parts, including the input layer, the backbone network, the connection network, and the detection head network, as shown in Figure 2.
[0091] The input layer is used to receive the data to be detected (ie, sample data), and includes three branches, IN1 corresponding to the three-dimensional time domain matrix IN2 corresponds to the three-dimensional time-frequency matrix IN3 corresponds to the three-dimensional IMF matrix IN2 is connected to the backbone network, and IN1 and IN3 are connected to the detection head network across regions.
[0092] The backbone network uses a convolutional neural network to process the input data (i.e., the three-dimensional time-frequency matrix corresponding to IN2). ) to perform feature extraction;
[0093] The connection network uses a feature pyramid network to perform multi-scale fusion and reuse of features extracted by the backbone network. The detection head network is used for final feature learning and then outputs detection information of the drone target and radio signal parameter information.
[0094] Among them, in addition to the input of the backbone network feature data (ie, the three-dimensional time-frequency matrix corresponding to IN2 ), IN1 and IN3 are directly connected to the detection head network, and an L-layer convolutional neural network is designed for downsampling to make it consistent with the feature map size when IN2 reaches the detection head network. Finally, the channel attention mechanism is used to perform weighted multiplication with the classification feature map of the detection head network, and the extracted signal time domain features and the intrinsic modal features of the signal at different time scales are fused with the features extracted by the main network of IN2 input to improve the final detection accuracy.
[0095] In a specific application example, during the offline training phase, the process of training using a neural network includes:
[0096] The training sample set to be constructed is the characteristic matrix set of the signal (i.e. the three-dimensional time domain matrix Three-dimensional time-frequency matrix Three-dimensional IMF matrix ), and sent it to the constructed neural network for training;
[0097] By fitting the actual data, the constructed validation sample set is input into the neural network for verification after each training cycle, and the accuracy of the network model (i.e., the neural network model) is monitored in real time;
[0098] After the training is completed, the constructed test sample set is input into the neural network for testing to check the performance of the network model (i.e., neural network model), such as accuracy, recall rate, generalization, etc.
[0099] If the network model meets the expected requirements, a weight parameter can be obtained through conversion, which is the detection model of the drone target.
[0100] In a specific application example, the online detection stage includes the following steps:
[0101] Step S100: Importing the trained drone detection model into the actual detection equipment to obtain all radio signals within the detection range;
[0102] That is, using drone detection equipment to monitor detection areas in open areas such as mountainous and forested areas, or low signal-to-noise ratio environments such as urban areas, and intercepting all radio signals within the device's pre-set frequency range;
[0103] Step S200: pre-processing the intercepted radio signal to obtain a feature matrix set of the signal;
[0104] Step S2001: intercept, rearrange, and multi-channel weight the original time domain signal data (i.e., radio signal) to obtain a three-dimensional time domain matrix
[0105] Step S2002: Use FFT to perform fast and slow time-dimensional time-frequency transformation on the two-dimensional time-domain signal matrix to obtain a three-dimensional time-frequency matrix
[0106] Step S2003: Use empirical mode decomposition to extract the modal features of the signal at different time scales on the two-dimensional time domain signal matrix, and finally obtain a three-dimensional IMF matrix
[0107] Step S300: The feature matrix set is sent to the detection model for computational reasoning, and information on whether a drone exists is output. If so, information such as the model, frequency, bandwidth, and time width of each drone is output. Ultimately, the effective detection of targets is achieved in environments with weak drone signals or low signal-to-noise ratios, where multiple drones may exist simultaneously.
[0108] Step S400: The signals obtained by the detection are displayed on a host computer or the data is further processed.
[0109] From the above, it can be seen that the neural network-based drone target detection method of the present invention can automatically find the most favorable features for detection through automatic learning of the actual data itself, without excessive human intervention, avoiding the incompleteness, insufficiency or artificial errors of manually designed features, saving manpower and material resources, and releasing resources. At the same time, it can also obtain better robust characteristics when the environmental interference is large.
[0110] Furthermore, the present invention can not only identify the presence of drones and classify the types of drones in intercepted radio signals, but also automatically and directly obtain information such as the center frequency and bandwidth of the drone signals in the intercepted radio signals. If there are multiple drones at the same time, it can also better detect them and obtain information about each drone signal separately, which is beneficial to practical applications and enhances the intensity of drone monitoring.
[0111] Those skilled in the art will appreciate that the above-mentioned embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present application is described with reference to the flow chart and / or block diagram of the method, device (system), and computer program product according to the embodiment of the present application. It should be understood that each flow and / or box in the flow chart and / or block diagram and the combination of the flow chart and / or box in the flow chart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for realizing the function specified in one flow chart flow or multiple flows and / or one block or multiple blocks of the block diagram. These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device that implements the functions specified in one or more processes of the flowchart and / or one or more blocks of the block diagram. These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes of the flowchart and / or one or more blocks of the block diagram.
[0112] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should be considered within the scope of protection of the present invention.
Claims
1. A UAV target detection method based on neural network, It is characterized in that include: Offline training phase: construct a training data set, build a neural network structure, import the training data set into the neural network for model training, and form a detection model; The training data set includes several labeled sample data; a single sample data includes a three-dimensional time domain matrix Three-dimensional time-frequency matrix Three-dimensional IMF matrix and label them; Online detection stage: use the trained drone detection model to perform actual detection; during the actual detection process, the collected signals are formed into three feature matrix sets, the three feature matrix sets are sent to the detection model for analysis and processing, and the drone information is output.
2. The method for detecting unmanned aerial vehicle targets based on a neural network according to claim 1, It is characterized in that The offline training stage includes preparing a training data set: collecting original signal data and preprocessing the original data.
3. The method for detecting unmanned aerial vehicle targets based on a neural network according to claim 2, It is characterized in that The collected signal raw data includes: Raw data of drone image transmission signals: Select different drones to fly in the detection area of the drone detection equipment, and collect the raw data of drone image transmission signals through the detection equipment; Background environment data: Deploy drone detection equipment in different environments, and then repeatedly collect the raw data of drone image transmission signals to form background environment data.
4. The method for detecting unmanned aerial vehicle targets based on a neural network according to claim 2, It is characterized in that The raw data preprocessing is to preprocess the collected signal raw data, including: Step S10: Analyze the original data and clean the dirty data; Step S20: Using empirical mode decomposition to decompose the two-dimensional time domain signal matrix P T Perform iterative decomposition; Step S30: Calculate the two-dimensional time domain data matrix P T The fast time dimension of the signal is transformed by performing a K-point Fourier transform H times to convert the fast time dimension of the matrix from the time domain to the frequency domain to extract the frequency domain features of the signal; a K×H time-frequency matrix is obtained, i.e., the time-frequency waterfall diagram P of the signal. TF ; Step S40: Obtain a three-dimensional time domain matrix Three-dimensional time-frequency matrix Three-dimensional IMF matrix It is used as a single input sample and annotated with the actual location of the drone, the type of aircraft, and the frequency range of the device detection.
5. The method for detecting unmanned aerial vehicle targets based on a neural network according to claim 4, It is characterized in that The step S20 comprises: Step S201: Solve the matrix P T The two upper and lower envelopes of the fast time dimension, that is, solving the local maximum and minimum extreme points; Step S202: Connect the maximum extreme value points to form an upper envelope, connect the minimum extreme value points to form a lower envelope, and then use the cubic spline interpolation method to fit the envelope line; obtain the matrix P T The upper and lower envelopes are averaged to obtain a mean envelope line; Step S203: From the matrix P T The fast time dimension is differentiated from the mean envelope to obtain the intermediate signal. If the intermediate signal satisfies zero crossing If the absolute difference between the number of points and the number of extreme points is less than or equal to 1 and the mean of the local maximum and minimum envelopes is equal to 0, then it is an eigenmode function vector; at the same time, through the matrix P T Make a difference between the fast time dimension and the IMF vector and then continue the above steps to obtain the IMF vector; get the intrinsic mode function vector IMF of the fast time dimension i , where i = 1, 2, 3...N; Step S204: Matrix P T All fast time dimensions are decomposed to finally obtain the (S×N)×H IMF matrix P IMF .
6. The method for detecting unmanned aerial vehicle targets based on a neural network according to claim 4, It is characterized in that The step S40 comprises: The drone detection equipment receives different directions from 0 to 360 degrees, uses X detection channels evenly distributed in the range of 0 to 360 degrees, and obtains X channels of original signal data at the same time Where ch=0,1,2...,X; get the two-dimensional time domain signal matrix of X channels at the same time The size is S×H, IMF matrix The size is (S×N)×H, and the time-frequency matrix The size is K×H; Time-frequency matrix According to the principle of non-adjacent channel selection, N channels are extracted to form a three-dimensional matrix The matrix size is N×K×H; IMF Matrix Select according to non-adjacent channels and Based on the principle of channel difference, N channels are extracted to obtain N (S×N)×H matrices, and then the IMF corresponding to these N matrices is calculated. i The vectors are weighted and summed according to a certain weight, and the weight coefficient is a n , and finally get the three-dimensional matrix The matrix size is N×S×H; where i=1,2,3...N, n=1,2,3...N; The time domain signal matrix According to the time-frequency matrix The same channel principle extracts N channels to form a three-dimensional matrix The matrix size is N×S×H.
7. The method for detecting unmanned aerial vehicle targets based on a neural network according to any one of claims 1 to 6, It is characterized in that The neural network structure constructed in the offline training stage is a single-stage detection neural network, including an input layer, a backbone network, a connection network, and a detection head network; wherein: The input layer is used to receive the data to be detected, and includes three branches: IN 1 Corresponding to the three-dimensional time domain matrix IN 2 Corresponding to the three-dimensional time-frequency matrix IN 3 Corresponding three-dimensional IMF matrix Where IN 2 Connect to the backbone network, IN 1 with IN 3 Connect to the detection head network across regions; The backbone network uses a convolutional neural network to extract features from input data; The connection network uses a feature pyramid network to perform multi-scale fusion and reuse of features extracted by the backbone network, and the detection head network is used for the final feature learning and then outputs the detection information of the drone target and the radio signal parameter information; Detection head network, directly connected to IN 1 with IN 3 , downsampled by L layers of convolutional neural network to make it consistent with IN 2 The feature map size is consistent when it reaches the detection head network; the weighted multiplication is performed with the classification feature map of the detection head network in the form of a channel attention mechanism, and the extracted signal time domain features and the intrinsic modal features of the signal at different time scales are combined with IN 2 The features extracted by the input main network are fused.
8. The method for detecting unmanned aerial vehicle targets based on a neural network according to any one of claims 1 to 6, It is characterized in that The process of training using a neural network in the offline training phase includes: Send the constructed training sample set, i.e. the three feature matrix sets of the signal, into the constructed neural network for training; By fitting the actual data, the constructed validation sample set is input into the neural network for validation after each training cycle, and the accuracy of the network model is monitored in real time; After the training is completed, the constructed test sample set is input into the neural network for testing to check the performance of the network model; If the network model meets the expected requirements, a weight parameter is obtained through conversion, which is the detection model of the UAV target.
9. The method for detecting unmanned aerial vehicle targets based on a neural network according to any one of claims 1 to 6, It is characterized in that The online detection stage includes: Step S100: importing the trained drone detection model into the actual detection equipment to obtain all radio signals within the detection range; Step S200: pre-processing the intercepted radio signal to obtain three characteristic matrix sets of the signal; Step S300: Send the three feature matrix sets into the detection model for calculation and reasoning, and output whether the drone exists. If so, output the information of each drone.
10. The method for detecting unmanned aerial vehicle targets based on a neural network according to claim 9, It is characterized in that The step S200 includes: Step S2001: intercept, rearrange, and multi-channel weight the original time domain signal data to obtain a three-dimensional time domain matrix Step S2002: Use FFT to perform fast and slow time-dimensional time-frequency transformation on the two-dimensional time-domain signal matrix to obtain a three-dimensional time-frequency matrix Step S2003: Use empirical mode decomposition to extract the modal features of the signal at different time scales on the two-dimensional time domain signal matrix, and finally obtain a three-dimensional IMF matrix
