An unmanned aerial vehicle jamming echo data generation method and system for AI training

Abstract

The application discloses an unmanned aerial vehicle jamming echo data generation method and system for AI training, and belongs to the technical field of radar signal processing and artificial intelligence. The method first constructs a simulation scene and configures global parameters of a radar and a target, then generates a three-dimensional flight trajectory of an unmanned aerial vehicle containing complex maneuvers such as a snake shape and a barrel roll based on a multi-stage kinematic model, configures multiple electromagnetic interference patterns and generates corresponding interference signals, finally superimposes the target echo and the interference signals in time domain, encapsulates standardized multi-dimensional tensor data sets, and synchronously generates corresponding labeling labels such as target positions and interference types. The application can generate high-fidelity unmanned aerial vehicle radar echo samples with interference at low cost and in batches, solves the problems of high data acquisition cost, lack of complex interference scenes and labeling difficulty, and provides standardized data support for radar intelligent target detection and anti-interference algorithm training.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing and artificial intelligence dataset generation technology, specifically relating to a method and system for generating UAV interference echo data for AI training, which can be used to construct training datasets for intelligent detection and anti-interference algorithms of low-altitude security radar. Background Technology

[0002] With the rapid development of the low-altitude economy, consumer and industrial drones are increasingly used in logistics, power line inspection, aerial photography, and agricultural and forestry protection. However, the security threats posed by illegal "black flight" drones to key areas such as airport airspace, nuclear power plants, and large sports venues are becoming increasingly prominent. Drones are characterized by their low flight altitude, slow speed, and small radar cross-section, and often fly in complex urban environments, making it difficult for traditional security methods to achieve effective all-weather, long-range detection. Radar, as an active detection sensor, is not limited by lighting or weather conditions and has become the core equipment for low-altitude drone detection.

[0003] In recent years, deep learning-based intelligent radar signal processing technology has gradually become the mainstream direction for improving the detection performance of low, slow, and small targets. By extracting features from radar echo range-Doppler spectra and time-frequency maps using models such as convolutional neural networks, the accuracy of weak target identification in environments with strong clutter and interference can be significantly improved, effectively distinguishing between drones and non-targets such as birds. However, the training of intelligent algorithms heavily relies on massive amounts of high-quality labeled data, and the current supply of radar drone echo data faces multiple bottlenecks: First, field measurement data collection is costly, time-consuming, and labor-intensive. It is also limited by airspace control, weather conditions, and experimental sites, making it difficult to cover diverse UAV maneuvering attitudes and complex flight scenarios, resulting in insufficient sample richness. Secondly, there are various complex interferences in the real electromagnetic environment, such as suppressive and deceptive interferences. Constructing a real-world interference scenario requires specialized electronic countermeasures equipment and a specific experimental environment, which is difficult and costly to implement. Furthermore, there is a severe shortage of echo samples containing rich interference characteristics. Third, the measured data requires manual annotation, which is not only labor-intensive and inefficient, but also makes it difficult to guarantee the annotation accuracy, and cannot directly meet the supervised training requirements of deep learning models.

[0004] The uniformity, scarcity, and high annotation costs of data samples have become key bottlenecks restricting the performance improvement of radar intelligent detection and anti-jamming algorithms and limiting the generalization ability of models. Therefore, there is an urgent need to develop a UAV echo data generation scheme that can simulate real physical characteristics, cover complex maneuver trajectories and various interference scenarios, and simultaneously generate accurately annotated UAV echo data, providing low-cost, high-fidelity standardized data support for radar intelligent algorithm training. Summary of the Invention

[0005] To address the shortcomings of existing technologies, such as high cost of measured UAV radar echo data, scarcity of complex interference samples, difficulty in labeling, and incompatibility with artificial intelligence training, the present invention aims to provide a method and system for generating UAV interference echo data for AI training. Based on the physical mechanism of radar signals, it generates echo data containing complex maneuver trajectories and multiple types of electromagnetic interference, and directly outputs standardized tensor datasets and synchronously labeled tags, providing high-quality training data support for radar intelligent detection and anti-interference algorithms.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: A method for generating drone interference echo data for AI training includes the following steps: Step 1: Construct a simulation scenario and configure global parameters, including setting radar system parameters and the initial state of the UAV target; Step 2: Generate a three-dimensional complex maneuver flight trajectory of the UAV based on a multi-stage kinematic model. The trajectory includes the position, velocity and attitude information of the UAV in consecutive time frames. Step 3: Configure complex electromagnetic interference patterns and generate corresponding interference signals; Step 4: Generate radar mixed echo data, calculate the target echo based on the flight trajectory, combine it with the interference signal for time-domain superposition, and then encapsulate the mixed echo data into a standardized tensor dataset, and simultaneously generate the corresponding label dataset.

[0007] Furthermore, step 1 specifically includes: Step 1.1: Using a zero-IF complex baseband signal model, the radar transmitted waveform is set as a linear frequency modulated signal. A complex discrete sequence is generated based on the radar bandwidth and sampling rate. The echo amplitude is calculated based on the radar equation. The echo amplitude is determined by the transmitted power, antenna gain, wavelength, target radar cross-section, and target distance, and is used as a scalar multiplier in the baseband echo sequence. The envelope delay term and Doppler phase term of the baseband echo signal are calculated based on the two-way propagation delay of the target distance, and a two-dimensional complex echo matrix is ​​constructed.

[0008] Step 1.2: A statistical model is used to model the instantaneous radar cross section of the UAV. A time-varying radar cross section function containing fluctuation factors is constructed by combining a sine function sequence and a random noise sequence to simulate the radar cross section fluctuation effect caused by the UAV rotor rotation and fuselage sway. Based on the Doppler frequency shift formula, a random range of values ​​for the target's speed and altitude is set to constrain the target to generate low Doppler frequency shift characteristics.

[0009] Furthermore, step 2 specifically includes: Step 2.1: Define the UAV state vector, which includes three-dimensional spatial position, flight speed, pitch angle, yaw angle, and roll angle; use a first-order hysteresis digital filter to smooth the desired state of the UAV, and combine the smoothing coefficient to control the maneuver response sensitivity to update the current speed and attitude angle; use the kinematic numerical integral equation based on discrete time step to update the three-dimensional position coordinates at the next moment, and introduce physical boundary constraint functions to limit the motion state.

[0010] Step 2.2: A preset motion library containing straight-line cruise, serpentine maneuvers, barrel rolls, and dive maneuvers is established, with each motion mapped to the corresponding target attitude and velocity parameters. A time axis is generated using a hierarchical state machine, and motion commands from the motion library are randomly scheduled. When performing a serpentine maneuver, the roll angle is controlled to change sinusoidally, and the resulting lateral acceleration causes the heading angle (yaw angle) to change periodically. When performing a barrel roll, the pitch angle and roll angle are controlled to change periodically according to a preset phase difference. Randomly distributed parameters are introduced during the initialization phase to determine the frequency and amplitude of motion execution.

[0011] Furthermore, the interference patterns configured in step 3 include at least one of the following: Suppressive noise interference: Generate complex Gaussian white noise in the baseband, control the bandwidth through a digital filter to generate broadband blocking noise or narrowband aiming noise, or combine it with a time-domain gating function to generate non-coherent random pulse interference; Deceptive jamming: Based on the principle of digital radio frequency storage, a digital copy of the intercepted radar signal is extracted, and a continuous time delay determined by a time-varying time delay function is applied to generate range dragging, velocity dragging, or multiple false target deception signals; where the range dragging velocity corresponds to the first derivative of the time delay function, the velocity dragging acceleration corresponds to the second derivative, and the Doppler frequency shift is naturally generated by the time delay change, ensuring the physical consistency of range-Doppler; Co-frequency asynchronous interference: Set the distance of the interference source, convert the distance of the interference source into a one-way propagation delay, take the modulus of the radar pulse repetition period to obtain the folded delay, then convert it into the folded apparent range, and superimpose independent random phase factors between pulses to simulate the range folding effect of long-range non-cooperative signals; the modulus operation is based on the time delay domain, which is equivalent to taking the modulus of twice the maximum unambiguous range in the range domain, ensuring that the one-way propagation delay is consistent with the dimensions of the radar two-way ranging.

[0012] Furthermore, step 4 specifically includes: The target echo sequence at each moment is calculated based on the flight trajectory. The target echo sequence is then superimposed in the time domain with one or more interference signal sequences, and complex Gaussian white noise is added to generate a mixed signal. Subsequently, the generated mixed signal is written and encapsulated into a high-dimensional tensor structure according to the sample dimension, time frame dimension, pulse dimension, range gate dimension, and channel dimension. Simultaneously, the target's true range cell index, the Doppler cell index corresponding to the Doppler frequency calculated from the target's radial velocity, and the interference category identifier are extracted as the corresponding label dataset.

[0013] Furthermore, the target echo component, interference signal component, and noise component of the mixed signal are linearly superimposed in the time domain; the complex Gaussian white noise is uncorrelated among its channels; and the channel dimensions of the high-dimensional tensor correspond to the multiple receiving channels of the radar.

[0014] Furthermore, the serpentine maneuver generates lateral acceleration through the sinusoidal change of the roll angle, thereby forming a periodic increment of the heading angle (yaw angle); the barrel roll maneuver causes the UAV to fly in a helical trajectory through periodic coupling of the pitch angle and the roll angle with a 90-degree phase difference; the physical boundary constraint function includes a lower altitude constraint, an upper speed constraint, and a spatial boundary constraint; the random distribution parameters include random values ​​for the duration of the action and random values ​​for the amplitude of the action.

[0015] Furthermore, the suppressive noise interference generates a band-limited noise signal by generating complex Gaussian white noise in the baseband and filtering it with a digital filter; the co-frequency asynchronous interference calculates the folded apparent range by taking the modulus of the single-path propagation delay of the interference source with the radar pulse repetition period, thus folding the long-range interference source into the radar receiving window and ensuring that the dimensions of single-path propagation and radar two-path ranging are consistent; the deceptive interference applies a continuous time delay determined by a time-varying time delay function to the intercepted radar reference signal, with the range dragging velocity corresponding to the first derivative of the time delay function, the velocity dragging acceleration corresponding to the second derivative, and the echo Doppler frequency shift naturally generated by the time delay change and physically consistent with the range change rate, generating a coherent false target echo.

[0016] Furthermore, the labeled dataset and the tensor dataset correspond one-to-one according to the sample index.

[0017] Correspondingly, the present invention also provides a drone interference echo data generation system for AI training, comprising: The scenario configuration module constructs a simulation scenario and configures global parameters, setting radar system parameters and the initial state of the UAV target. The trajectory generation module generates a three-dimensional complex maneuver flight trajectory of the UAV based on a multi-stage kinematic model. The trajectory includes the position, velocity and attitude information of the UAV in continuous time frames. The interference configuration module configures complex electromagnetic interference patterns and generates corresponding interference signal parameters. The echo synthesis and tensor quantization module generates mixed radar echo data, calculates the target echo based on the flight trajectory, generates an interference signal by combining the interference parameters, and performs time-domain superposition. Before time-domain superposition, the dynamic gain coefficient of each receiving channel is calculated based on the instantaneous azimuth and elevation angles of the UAV using the antenna pattern function, and the echo is weighted in real time. Subsequently, the mixed echo data is encapsulated into a standardized tensor dataset, and a tag dataset containing the target true range cell index, Doppler cell index, and interference category identifier is generated simultaneously.

[0018] Compared with the prior art, the present invention has the following advantages: This invention constructs an echo signal model based on radar equations and the Doppler principle, and combines a time-varying radar cross section fluctuation model to simulate the RCS fluctuation characteristics of UAV rotor modulation. Through first-order hysteresis filtering and kinematic integration, a complex maneuver trajectory conforming to physical inertia is generated. The generated echo data closely approximates the physical characteristics of real radar received signals, which can effectively reduce the domain difference between simulation data and measured data and improve the actual generalization ability of the training model.

[0019] This invention incorporates three typical electromagnetic interference patterns: suppressive noise interference, deceptive interference, and synchronous asynchronous interference. It can flexibly configure single or combined interference scenarios, covering common interference environments in actual radar operation. It solves the problems of difficulty in obtaining measured interference samples and limited scenarios, and can support diversified training of anti-interference algorithms.

[0020] This invention directly outputs a standardized multidimensional tensor structure dataset, simultaneously generating target locations, Doppler values, and interference type labels that correspond one-to-one with the samples. No additional data preprocessing or manual annotation is required. It can be directly connected to the data loading module of a deep learning framework for model training, significantly reducing the time and manpower costs of dataset construction.

[0021] This invention allows for flexible adjustment of factors such as radar system, UAV performance, and interference intensity through parameter configuration, enabling the generation of massive and diverse samples in batches. It eliminates the need for field experiments and dedicated hardware equipment, significantly reducing the data acquisition cost for intelligent algorithm development. Furthermore, it allows for the expansion and addition of new interference patterns and maneuvers as needed. Attached Figure Description

[0022] Figure 1 is a schematic diagram of the range-Doppler spectrum under the multi-false-target interference scenario of the present invention; Figure 2 is a flowchart of the overall processing of the UAV interference echo data generation method for AI training described in this invention. Detailed Implementation

[0023] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that these descriptions are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0024] Example 1 This embodiment provides a specific implementation of a method for generating UAV interference echo data for AI training, focusing on how to achieve radar parameter configuration, UAV state machine-driven trajectory generation, and matrix construction of basic baseband echoes through software design.

[0025] I. Scene Configuration and Radar Signal Modeling During system initialization, the configuration file is read first to generate the radar parameter structure. To balance computational efficiency and simulation fidelity, this embodiment uses a zero-IF complex baseband signal model for simulation instead of directly simulating the radio frequency signal. This reduces the sampling rate requirement while preserving complete Doppler and phase characteristics.

[0026] The radar transmit waveform is set to a linear frequency modulated (LFM) signal. The system first determines the LFM signal based on the configured radar bandwidth. and preset sampling rate The discrete fast time series is calculated, and then a complex discrete array of linear frequency modulated waveforms is generated using vectorization operations:

[0027] in, The pulse width. For frequency modulation slope, The sampling interval is... This is the index for discrete sampling points.

[0028] To accurately reflect the impact of target distance on echo intensity, the system invokes the radar equation module. Let the radar transmit power be... The antenna gain is , wavelength is The radar cross-section of the target is The target distance is Based on radar equations Define the system's overall gain constant. Calculate the first Target voltage amplitude coefficient for each pulse:

[0029] This amplitude coefficient will be applied as a scalar multiplier to the baseband echo array. The echo amplitude is inversely proportional to the square of the target distance.

[0030] The core of radar detection lies in accurately simulating the phase changes of the echo. For the first... Radial distance at each pulse moment Two-way propagation delay is Received baseband echo signal The mathematical expression is:

[0031] in, The term carries the Doppler phase history caused by minute changes in distance.

[0032] The program pre-allocates a two-dimensional complex matrix space in memory. ,in This represents the total number of pulses within the coherent processing interval. This represents the number of fast-time sampling points for a single pulse. The rows of this matrix correspond to the pulse dimension, and the columns correspond to the fast-time sampling dimension. The discretized data is written into this matrix row by row to construct a two-dimensional complex echo matrix.

[0033] II. Parametric Modeling of "Low, Slow, and Small" Targets To avoid a monotonous simulation scenario due to fixed parameters, the system sets the average radar cross section (RCS), flight speed, and flight altitude of the UAV to configurable parameter ranges. UAVs are small and mostly made of composite materials, resulting in extremely low RCS. To simulate the RCS fluctuation effect caused by UAV attitude changes, the system constructs a time-varying RCS fluctuation function module, implemented using a linear combination of a sine function array and a Gaussian white noise array.

[0034] in, The set average radar cross section, The simulation is based on periodic modulation caused by rotor rotation. The random numbers are standard normally distributed to simulate random disturbances. This time-varying function simulates the fluctuations in radar cross-section caused by the rotor rotation and fuselage sway of the UAV.

[0035] Meanwhile, according to the Doppler frequency shift formula The Doppler frequency shift of the target is constrained at low frequencies, and the target speed is randomly selected within the parameter configuration range (for example, the speed range is set to 5-50m / s or 30-100m / s, both of which belong to the category of low Doppler frequency shift targets), and it is allowed to hover or perform low-speed maneuvers, thereby generating low Doppler frequency shift characteristics that conform to the characteristics of "low, slow and small" targets.

[0036] III. Generation of Complex 3D Maneuvering Flight Trajectories for Unmanned Aerial Vehicles Trajectory generation is encapsulated as a function containing discrete time steps. The loop structure. At each loop step... In the middle, the system maintains and updates the state vector. Where (x, y, z) is the three-dimensional spatial position, v is the flight speed, θ is the pitch angle, and ψ is the yaw angle. This is the roll angle.

[0037] To simulate the inertia of a drone and avoid abrupt changes in the state sequence, the program internally calls a first-order hysteresis digital filter to smooth the desired target state. Time's up The state update formula at time t is:

[0038] in, The smoothing coefficient characterizes the maneuver response sensitivity of the UAV.

[0039] Subsequently, the three-dimensional position coordinates are updated by calling a numerical integration function based on the Euler method and appended to the global trajectory history matrix:

[0040] The system introduces physical boundary constraint functions to limit the motion state, including: lower height constraint (to prevent the trajectory from falling below the ground), upper speed constraint (to prevent exceeding the aerodynamic limits), and spatial boundary constraint (to prevent exceeding the spatial range of the simulation scene), to ensure the physical rationality of the generated trajectory.

[0041] The system incorporates a hierarchical state machine module. The upper-level scheduler generates a timeline using a random segmentation algorithm and randomly binds enumerated values ​​from the action library. The lower-level executor reads the current action enumeration value and activates the corresponding control law equation. The randomly distributed parameters include random values ​​for the action duration and random values ​​for the action amplitude.

[0042] Snake-like maneuver: Simulates the periodic S-shaped swaying of an unmanned aerial vehicle (UAV) in a horizontal plane. The system generates a sinusoidally varying array of target roll angle values ​​and calculates the yaw angle increment caused by the resulting lateral acceleration.

[0043] Among them, A roll For the maximum roll angle, ω snake Let g be the maneuver frequency and g be the acceleration due to gravity. The serpentine maneuver generates lateral acceleration through the sinusoidal change of the roll angle, which in turn creates a periodic increment in the heading angle (yaw angle).

[0044] Barrel roll maneuver: Simulates a drone flying along a spiral trajectory while changing altitude and heading. Pitch angle θ and roll angle There exists periodic coupling with a phase difference of 90°:

[0045] Among them, A roll_max For the maximum roll angle amplitude, A pitch_max ω is the maximum pitch angle amplitude. roll The angular frequency of the roll maneuver. The barrel roll maneuver, through the periodic coupling of the pitch angle and the roll angle with a 90-degree phase difference, causes the UAV to fly in a helical trajectory.

[0046] Dive: Simulates the process of a UAV accelerating downwards after detecting a target, resetting the target pitch angle parameter to a negative value, and applying random acceleration to the velocity scalar.

[0047] Where, θ dive For the dive pitch angle, Δv boost This represents the speed increment.

[0048] During the trajectory simulation initialization phase, random distribution parameters are introduced to determine the frequency and amplitude of the above actions. Specifically, the number of time frames each action is maintained is randomly selected within a preset range, and the amplitude parameters of each maneuver are randomly selected within a preset range.

[0049] Example 2 This embodiment focuses on how to generate multi-mode electromagnetic interference signals through array operations in software engineering, and encapsulate the data into a multi-dimensional tensor structure adapted to deep learning frameworks.

[0050] I. Configuration of Complex Electromagnetic Interference Patterns The system constructs an interference signal generation module independent of the echo, and randomly instantiates an interference parameter structure according to the configuration. The interference patterns include at least one of the following.

[0051] 1. Suppressive Noise Interference Based on Complex White Gaussian Noise and Digital Filtering The system allocates a complex array of the length of the receiving window to generate independent and identically distributed complex Gaussian white noise sequences. Based on the required interference bandwidth (broadband blocking or narrowband targeting), design the corresponding digital low-pass or band-pass filter (FIR / IIR), and compare the noise sequence with the filter impulse response. Perform convolution to obtain the band-limited baseband noise signal:

[0052] in The amplitude of the interference signal is represented by the filter passband covering the entire bandwidth of the radar receiver. Narrowband aiming noise is achieved by setting the filter passband to cover only a narrow band near the center frequency of the radar signal. Scattered pulse interference is further enhanced with time-domain gating logic, outputting the signal only within a randomly generated pulse width index interval, simulating a non-coherent pulse flow. The suppressive noise generated by this scheme exhibits a Gaussian amplitude distribution and random phase characteristics, consistent with the statistical characteristics of real-world environmental noise interference.

[0053] 2. Co-frequency asynchronous interference based on the distance folding principle The program sets the distance to the interference source. Calculate its one-way propagation delay The modulo operator is used to determine the radar pulse repetition period. Folding is performed to obtain the folding delay:

[0054] This is then converted to the folded apparent distance (equivalent to the distance domain pair). (Modulation) (equivalent to) ) Consider folding as distance Converted to array index offset. In each pulse cycle, the system generates a random scalar [0, 2π] as a phase factor multiplied into the interference array, making it non-coherent in Doppler processing. Co-frequency asynchronous interference achieves range folding through time delay domain modulo operation. That is, the interference source range is first converted into a one-way propagation time delay, and then the modulo operation of the radar pulse repetition period is taken to obtain the folded time delay and converted into apparent range. This is equivalent to taking the modulo operation of twice the maximum unambiguous range in the range domain, ensuring that the over-range interference source is correctly mapped into the radar receiving window.

[0055] 3. Deceptive interference based on digital radio frequency storage principles The system allocates a buffer to extract digital copies of the echo. A time-varying delay function is introduced. ,in For an initial fixed delay, This represents a dynamically changing time delay increment. The decoy target echo is generated directly by time-domain shifting the intercepted radar reference signal using a digital delay line.

[0056] in, The intercepted radar reference signal. Range and towing speed are determined by... (The first derivative of the time delay function) represents the velocity-induced acceleration. The second derivative of the time delay function represents the Doppler frequency shift of the echo signal due to the time delay. The phase modulation is introduced naturally, without the need for an additional, independent Doppler phase rotation term, thus ensuring the physical consistency of the distance-Doppler coupling relationship.

[0057] II. Construction of Multi-channel Receiver Antenna Gain Model To simulate the radar's ability to sense target location and meet the requirements of multi-channel signal processing, this embodiment constructs an antenna pattern model containing K receiving channels. Within each pulse repetition period, the system first calculates the instantaneous azimuth angle of the target relative to the radar line of sight based on the current three-dimensional spatial coordinates of the UAV and the radar site coordinates. With pitch angle Then, the antenna pattern function is called. Calculate the first The gain coefficients of each receiving channel at the current angle form an instantaneous gain vector. In this embodiment, the three receiving channels are exemplarily set as a sum beam channel, a difference beam channel, and a guard beam channel, and their radiation pattern functions correspond to... , and protection beam The normalized voltage gain. This instantaneous gain vector is used for pulse-by-pulse dynamic weighting of the spatial dimension of the subsequent tensor. Correspondingly, the channel dimension of the high-dimensional tensor corresponds to the multiple receiving channels of the radar.

[0058] III. Synthesis and Tensor Quantization of Echo Data The data synthesis process is completed through a double-nested "frame-pulse" loop. First, the instantaneous azimuth and elevation angles are calculated based on the UAV's spatial position at the current pulse moment. Then, the antenna pattern function is called to calculate the instantaneous gain vector at the current angle. Subsequently, within the inner pulse loop, the program independently calculates the pure target complex array. Complex array of interference signals And add them together. Then combine this one-dimensional array with the instantaneous gain vector. An outer product operation is performed to expand the channel dimension, which dynamically changes with the target angle. To simulate thermal noise, the program calls a complex Gaussian white noise generation function to generate a noise matrix. And superimposed. The target echo component, interference signal component, and noise component are linearly superimposed in the time domain, and the complex Gaussian white noise is uncorrelated between the channels. The final mixed signal matrix is ​​expressed as:

[0059] in, and These are the target azimuth and elevation angles as they change over time. This is the instantaneous gain vector at the corresponding moment, and thus the gain of each receiving channel is dynamically updated as the target's flight attitude and spatial position change; The noise components in the signal follow a complex Gaussian distribution CN(0,σn2), and the noise in each channel is uncorrelated.

[0060] To seamlessly integrate with the data loader of the artificial intelligence framework, the program predeclares a 5-dimensional tensor object in memory, whose dimensional structure is strictly defined as follows:

[0061] Where, N sample N is the total number of samples. frame N represents the number of frames per sample. pulse N represents the number of pulses per frame. range N represents the number of fast-time sampling points for a single pulse. channel This represents the number of channels.

[0062] Each piece of data generated in the nested loop is directly written into the corresponding slice of the tensor. While generating the high-dimensional tensor, the software module simultaneously extracts and saves the target's true distance cell index, Doppler cell index, and interference category identifier as a label dataset. The label dataset and the tensor dataset correspond one-to-one according to the sample index.

[0063] Finally, the system serializes and saves the 5D tensor as a standard data file (such as NumPy's .npy format or HDF5 format), and outputs a trajectory label array containing the target's true position and Doppler parameters, as well as an interference label array containing the interference type.

[0064] IV. Signal Processing and Visual Verification The system can also extract tensor slices for verification through the signal processing visualization module. Specifically, by performing pulse compression (matched filtering) on ​​a frame of data in the tensor along the fast time dimension and then performing a fast Fourier transform along the slow time dimension, the distance-Doppler spectrum can be rendered. Figure 1 The diagram illustrates the range-Doppler spectrum of the multi-false target interference echoes generated by the method described in this invention, in which the real target echoes and multiple false target deception signals can be clearly distinguished. Figure 2 The overall processing flowchart of the method described in this invention is presented, fully demonstrating the entire process from scene configuration, trajectory generation, interference configuration to echo synthesis and tensor output. Verification results show that the generated data conforms to the theoretical expectations of radar signal processing in terms of time, frequency, and spatial characteristics, proving the effectiveness of the method described in this invention.

[0065] Example 3 This embodiment provides a specific implementation of a drone interference echo data generation system for AI training.

[0066] The system's overall architecture comprises four core modules, which interact with each other through standard data interfaces: I. Scene Configuration Module The scenario configuration module is responsible for building the simulation scenario and configuring global parameters. This module includes a radar signal modeling unit and a target parameter modeling unit. The radar signal modeling unit employs a zero-IF complex baseband signal model, setting the radar transmitted waveform as a linear frequency modulated signal. A complex discrete sequence is generated based on the radar bandwidth and sampling rate. The echo amplitude is calculated based on the radar equations. This echo amplitude is determined by the transmitted power, antenna gain, wavelength, target radar cross-section, and target range, and is used as a scalar multiplier in the baseband echo sequence. The envelope delay and Doppler phase terms of the baseband echo signal are calculated based on the two-way propagation delay of the target range, constructing a two-dimensional complex echo matrix. The echo amplitude is inversely proportional to the square of the target range. The rows of the two-dimensional complex echo matrix correspond to the pulse dimension, and the columns correspond to the fast-time sampling dimension.

[0067] The target parameter modeling unit performs parameterized modeling of the UAV target: it uses a statistical model to model the instantaneous radar cross section of the UAV, and uses a combination of sine function sequences and random noise sequences to construct a time-varying radar cross section function containing fluctuation factors to simulate the radar cross section fluctuation effect caused by the UAV rotor rotation and fuselage sway; at the same time, it sets random value ranges for the target's speed and altitude based on the Doppler frequency shift formula to constrain the target to generate low Doppler frequency shift characteristics.

[0068] II. Trajectory Generation Module The trajectory generation module generates complex 3D maneuvering flight trajectories for UAVs based on a multi-stage kinematic model. This module includes a motion rule modeling unit and an action mapping and scheduling unit. The motion rule modeling unit defines a UAV state vector containing three-dimensional spatial position, flight speed, pitch angle, yaw angle, and roll angle. A first-order hysteresis digital filter is used to smooth the UAV's desired state, and a smoothing coefficient is used to control the maneuver response sensitivity to update the current speed and attitude angles. Kinematic numerical integral equations based on discrete time steps are used to update the three-dimensional position coordinates for the next moment, and physical boundary constraint functions are introduced to restrict the motion state. These physical boundary constraint functions include a lower altitude constraint, an upper velocity constraint, and a spatial boundary constraint.

[0069] The motion mapping and scheduling unit pre-defines a motion library containing straight-line cruise, serpentine maneuvers, barrel rolls, and dive maneuvers, with target attitude and velocity parameters corresponding to each motion mapping. It uses a hierarchical state machine to generate a timeline and randomly schedules motion commands from the motion library. When performing a serpentine maneuver, the roll angle is controlled to change sinusoidally, resulting in lateral acceleration that causes periodic changes in the heading angle (yaw angle). When performing a barrel roll, the pitch and roll angles are controlled to change periodically according to a preset phase difference. During the initialization phase, randomly distributed parameters are introduced to determine the frequency and amplitude of motion execution. These random distributed parameters include random values ​​for the motion duration and random values ​​for the motion amplitude.

[0070] III. Interference Configuration Module The interference configuration module configures complex electromagnetic interference patterns and generates corresponding interference signal parameters. The module's preset interference type library includes at least one of the following: The suppressive noise interference unit generates complex Gaussian white noise in the baseband and controls the bandwidth through a digital filter to generate broadband blocking noise or narrowband aiming noise, or combines a time-domain gating function to generate non-coherent cluttered pulse interference.

[0071] The deceptive jamming unit is based on the principle of digital radio frequency storage. It extracts a digital copy of the intercepted radar signal and applies a continuous time delay determined by a time-varying time delay function. The range drag velocity corresponds to the first derivative of the time delay function, the velocity drag acceleration corresponds to the second derivative, and the Doppler frequency shift is naturally generated by the time delay change, thus generating range drag, velocity drag, or multiple false target deception signals.

[0072] The co-frequency asynchronous interference unit sets the interference source distance, converts the interference source distance into a one-way propagation delay, takes the modulus of the radar pulse repetition period to obtain the folded delay, and then converts it into the folded apparent distance (equivalent to taking the modulus of twice the maximum unambiguous distance in the range domain). Independent random phase factors are superimposed between pulses to simulate the range folding effect of long-range non-cooperative signals.

[0073] IV. Echo Synthesis and Tensor Quantization Module The echo synthesis and tensor quantization module generates mixed radar echo data, calculates the target echo based on the flight trajectory, generates interference signals by combining interference parameters, and performs temporal superposition. The mixed echo data is then encapsulated into a standardized tensor dataset, and a labeled dataset containing target true range cell indices, Doppler cell indices, and interference category identifiers is generated simultaneously. This module includes an antenna gain modeling unit and a tensor quantization construction unit. The antenna gain modeling unit constructs a multi-channel receiving antenna pattern model based on instantaneous angles. Within each pulse repetition period, it calculates the instantaneous azimuth and elevation angles relative to the radar line of sight based on the UAV's current spatial coordinates. It then calls the antenna pattern function to calculate the gain coefficient of each receiving channel at the current angle, forming a dynamically changing instantaneous gain vector over time. The antenna gain coefficients correspond to the normalized voltage gain values ​​of the sum beam, difference beam, and guard beam, respectively, providing amplitude weighting for the echo channel dimension that is updated in real time with the target angle. Correspondingly, the channel dimension of the high-dimensional tensor corresponds to the radar's multiple receiving channels.

[0074] In the inner loop of the pulse period, the tensor quantization construction unit first calculates the instantaneous azimuth and elevation angles of the UAV relative to the radar at the current pulse moment. It then obtains the corresponding instantaneous gain vector through the antenna pattern function, independently calculates the clean target echo sequence and the selected interference signal sequence, and performs time-domain superposition. The superimposed one-dimensional sequence is then subjected to an outer product operation or broadcast operation with the instantaneous gain coefficient vector to expand the channel dimension, which is dynamically weighted according to the real-time target angle. An independent and identically distributed complex Gaussian white noise matrix is ​​then superimposed to generate a multi-channel mixed signal. Subsequently, the generated mixed signal is written and encapsulated into a high-dimensional tensor structure according to the sample dimension, time frame dimension, pulse dimension, range gate dimension, and channel dimension. Simultaneously, the target's true range cell index, Doppler cell index, and interference category identifier are extracted and saved as a label dataset. The label dataset and the tensor dataset correspond one-to-one according to the sample index.

[0075] The modules described above can be implemented using software programs, hardware circuits, or a combination of both. Data transfer between modules is accomplished through standardized data interfaces, ensuring the system's scalability and maintainability.

[0076] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A method for generating UAV interference echo data for AI training, characterized in that, Includes the following steps: Step 1: Construct a simulation scenario and configure global parameters, including setting radar system parameters and the initial state of the UAV target; Step 2: Generate a three-dimensional complex maneuver flight trajectory of the UAV based on a multi-stage kinematic model. The trajectory includes the position, velocity and attitude information of the UAV in consecutive time frames. Step 3: Configure complex electromagnetic interference patterns and generate corresponding interference signals; Step 4: Generate radar mixed echo data, calculate the target echo based on the flight trajectory, combine it with the interference signal for time-domain superposition, and then encapsulate the mixed echo data into a standardized tensor dataset, and simultaneously generate the corresponding label dataset.

2. The method according to claim 1, characterized in that, Step 1 specifically includes: Step 1.1: Using a zero-IF complex baseband signal model, the radar transmitted waveform is set as a linear frequency modulated signal. A complex discrete sequence is generated based on the radar bandwidth and sampling rate. The echo amplitude is calculated based on the radar equation. The echo amplitude is determined by the transmitted power, antenna gain, wavelength, target radar cross-section, and target distance, and is used as a scalar multiplier in the baseband echo sequence. The envelope delay term and Doppler phase term of the baseband echo signal are calculated based on the two-way propagation delay of the target distance, and a two-dimensional complex echo matrix is ​​constructed. Step 1.2: A statistical model is used to model the instantaneous radar cross section of the UAV. A time-varying radar cross section function containing fluctuation factors is constructed by combining a sine function sequence and a random noise sequence to simulate the radar cross section fluctuation effect caused by the UAV rotor rotation and fuselage sway. Based on the Doppler frequency shift formula, a random range of values ​​for the target's speed and altitude is set to constrain the target to generate low Doppler frequency shift characteristics.

3. The method according to claim 1, characterized in that, Step 2 specifically includes: Step 2.1: Define the UAV state vector containing three-dimensional spatial position, flight speed, pitch angle, yaw angle, and roll angle; use a first-order hysteresis digital filter to smooth the UAV's desired state, and combine the smoothing coefficient to control the maneuver response sensitivity to update the current speed and attitude angle; use the kinematic numerical integral equation based on discrete time step to update the three-dimensional position coordinates of the next moment, and introduce physical boundary constraint functions to restrict the motion state. Step 2.2: A preset motion library containing straight-line cruise, serpentine maneuvers, barrel rolls, and dive maneuvers is established, with each motion mapped to the corresponding target attitude and velocity parameters. A time axis is generated using a hierarchical state machine, and motion commands from the motion library are randomly scheduled. When performing a serpentine maneuver, the roll angle is controlled to change sinusoidally, and the resulting lateral acceleration causes the heading angle to change periodically. When performing a barrel roll, the pitch and roll angles are controlled to change periodically according to a preset phase difference. Randomly distributed parameters are introduced during the initialization phase to determine the frequency and amplitude of motion execution.

4. The method according to claim 1, characterized in that, The interference patterns configured in step 3 include at least one of the following: Suppressive noise interference: Generate complex Gaussian white noise in the baseband, control the bandwidth through a digital filter to generate broadband blocking noise or narrowband aiming noise, or combine it with a time-domain gating function to generate non-coherent random pulse interference; Deceptive jamming: Based on the principle of digital radio frequency storage, a digital copy of the intercepted radar signal is extracted, and a continuous time delay determined by a time-varying time delay function is applied to generate range dragging, velocity dragging, or multiple false target deception signals; where the range dragging velocity corresponds to the first derivative of the time delay function, the velocity dragging acceleration corresponds to the second derivative, and the Doppler frequency shift is naturally generated by the time delay change; Co-frequency asynchronous interference: Set the distance of the interference source, convert the distance of the interference source into a one-way propagation delay, take the modulus of the radar pulse repetition period to obtain the folded delay, convert it into the folded apparent distance, and superimpose independent random phase factors between pulses to simulate the range folding effect of long-distance non-cooperative signals. The modulo operation is performed in the time delay domain, which is equivalent to taking the modulo of twice the maximum unambiguous distance in the distance domain.

5. The method according to claim 1, characterized in that, Step 4 specifically includes: The target echo sequence at each moment is calculated based on the flight trajectory. The target echo sequence is then superimposed in the time domain with one or more interference signal sequences, and complex Gaussian white noise is added to generate a mixed signal. Subsequently, the generated mixed signal is written and encapsulated into a high-dimensional tensor structure according to the sample dimension, time frame dimension, pulse dimension, range gate dimension and channel dimension. At the same time, the target true range cell index, the Doppler cell index corresponding to the Doppler frequency calculated from the target radial velocity and the interference category label are extracted as the corresponding label dataset.

6. The method according to claim 5, characterized in that, The target echo component, interference signal component, and noise component of the mixed signal are linearly superimposed in the time domain; the complex Gaussian white noise is uncorrelated among its channels; and the channel dimensions of the high-dimensional tensor correspond to the multiple receiving channels of the radar.

7. The method according to claim 3, characterized in that, The serpentine maneuver generates lateral acceleration through the sinusoidal change of the roll angle, thereby forming a periodic increment in the heading angle; the barrel roll maneuver causes the UAV to fly in a helical trajectory through periodic coupling of the pitch and roll angles with a 90-degree phase difference; the dive maneuver simulates the rapid descent of the UAV by setting the target pitch angle to a negative value and applying a random velocity increment; the physical boundary constraint function includes a lower altitude constraint, a higher velocity constraint, and a spatial boundary constraint; the random distribution parameters include random values ​​for the duration and amplitude of the maneuver.

8. The method according to claim 4, characterized in that, The suppressive noise interference generates a band-limited noise signal by generating complex Gaussian white noise in the baseband and filtering it with a digital filter; the co-frequency asynchronous interference folds the long-range interference source into the radar receiving window by taking the pulse repetition period modulus of the one-way propagation delay of the interference source in the time delay domain; the deceptive interference applies a continuous time delay determined by a time-varying time delay function to the intercepted radar reference signal, with the range drag velocity corresponding to the first derivative of the time delay function, the velocity drag acceleration corresponding to the second derivative, and the echo Doppler frequency shift naturally generated by the time delay change and physically consistent with the range change rate, generating a coherent false target echo.

9. The method according to claim 5, characterized in that, The labeled dataset and the tensor dataset correspond one-to-one according to the sample index.

10. A drone interference echo data generation system for AI training, characterized in that, include: The scenario configuration module constructs a simulation scenario and configures global parameters, setting radar system parameters and the initial state of the UAV target. The trajectory generation module generates a three-dimensional complex maneuver flight trajectory of the UAV based on a multi-stage kinematic model. The trajectory includes the position, velocity and attitude information of the UAV in continuous time frames. The interference configuration module configures complex electromagnetic interference patterns and generates corresponding interference signal parameters. The echo synthesis and tensor quantization module generates mixed radar echo data, calculates the target echo based on the flight trajectory, and performs time-domain superposition of the signals. Before time-domain superposition, the dynamic gain coefficients of each receiving channel are calculated using the antenna pattern function based on the instantaneous azimuth and elevation angles of the UAV, and the echoes are weighted in real time. Subsequently, the mixed echo data is encapsulated into a standardized tensor dataset, and a tag dataset containing the target true range cell index, Doppler cell index, and interference category identifier is generated simultaneously.

Images