A simulation data automatic generation method for multi-target direct positioning
By configuring multi-dimensional parameters, constraining the location of radiation sources, and controlling multiple random seeds, combined with time-frequency feature extraction, the problem of inflexible and inconsistent generation of simulation datasets in existing technologies is solved, generating a high-quality, reproducible multi-target localization simulation dataset that meets the feature learning requirements of deep learning models.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2026-03-10
- Publication Date
- 2026-06-16
AI Technical Summary
Existing simulation dataset generation methods for multi-target detection and localization algorithms suffer from problems such as high parameter coupling, lack of physical rationality in parameter settings, unreproducible random processes, and single feature dimensions. These issues result in inflexible and inconsistent dataset generation that differs significantly from real-world scenarios, making it difficult to meet the feature learning needs of deep learning models.
By employing multi-dimensional parameter controllable configuration, radiation source location distance constraint mechanism, independent management of multiple random seeds, and joint modeling of multiple receiving stations and multiple array elements, combined with time-frequency feature extraction, a high-quality, reproducible simulation dataset is generated.
It enables the generation of simulation datasets with flexible and controllable parameters, reasonable physical data, and rich features, supports customized needs for different experimental scenarios, and ensures dataset consistency and improved model resolution.
Smart Images
Figure FT_1
Abstract
Description
Technical Field
[0001] This invention relates to an automated method for generating simulation data for direct multi-target localization, belonging to the fields of radar signal processing and multi-target localization technology. Background Technology
[0002] In research on multi-target detection and localization algorithms in fields such as radar and wireless communication, high-quality simulation datasets are a core prerequisite for algorithm training and validation. Existing dataset generation methods generally suffer from the following shortcomings: The configuration and generation are highly coupled. Simulation parameters such as the number and location of radiation sources and signal-to-noise ratio are deeply bound to the dataset generation code, making it difficult to flexibly adjust experimental conditions and quickly adapt to the testing requirements of different algorithms.
[0003] The parameter settings lack physical rationality, and no distance constraints are set when generating the radiation source location, which easily leads to overlapping target locations; the location of the observation station does not take into account actual measurement errors, and the generated data differs greatly from the real scene.
[0004] The random process is not reproducible, which leads to a lack of systematic seed control mechanism when generating random data. Under the same experimental conditions, it is impossible to generate consistent datasets, resulting in unreliable algorithm performance comparison results.
[0005] The feature dimension is singular, extracting only time-domain or frequency-domain features of the signal without combining time-frequency joint features, which makes it difficult to support the feature learning needs of deep learning models for complex signals.
[0006] To address the above issues, there is an urgent need for a method to generate multi-objective DPD simulation datasets that are parameter-controllable, physically reasonable, reproducible, and feature-rich. Summary of the Invention
[0007] The purpose of this method is to address the shortcomings of existing dataset generation methods by providing a method for generating multi-target direct localization simulation datasets. This method consists of two stages: automatic configuration file generation and batch dataset construction. Through standardized parameter configuration, physical constraint modeling, and joint extraction of time-frequency features, it achieves the generation of high-quality and reproducible simulation datasets.
[0008] The innovative aspects of this invention: 1. Multi-dimensional parameter controllable configuration: Supports flexible setting of parameters such as the number of radiation sources, signal-to-noise ratio range, and bandwidth jitter amplitude to meet the customized needs of different experimental scenarios.
[0009] 2. Radiation source location distance constraint mechanism: Polar coordinate sampling combined with Euclidean distance verification is used to avoid overlapping radiation source locations and ensure the physical rationality of the data.
[0010] 3. Independent control of multiple random seeds: By deriving special seeds from the master seed, the random process of each parameter can be independently controlled, ensuring the complete reproducibility of the dataset.
[0011] 4. Multi-receiving station-multi-array element joint modeling: Supports synchronous generation of signals from three receiving stations, each with 16 array elements, preserving the time-frequency characteristics of all array elements and improving multi-target resolution capability.
[0012] 5. Precise Doppler frequency shift and time-frequency feature extraction: The Doppler frequency shift is calculated based on the theoretical formula of relative velocity and carrier frequency, and the time-frequency joint features are extracted by combining STFT transformation to enrich the data feature dimensions.
[0013] The present invention is achieved using the following technical solution.
[0014] A method for generating a multi-objective DPD simulation dataset includes two stages: a configuration file generation stage and a dataset construction stage. The specific steps are as follows: Step 1: Configuration file generation stage This stage generates a CSV format configuration file containing all parameters of the samples, providing standardized input for dataset construction. It includes the following sub-steps: Step 1.1: Random Seed Management A master random seed, master_seed, is set, and seven dedicated random seeds are derived from the master seed to control the random generation process of parameters such as position jitter, signal power, bandwidth jitter, radiation source position, observation station noise, and radiation source velocity, ensuring that the random processes of each parameter are independent and reproducible.
[0015] Step 1.2: Signal Parameter Generation Signal power, based on a reference value of 1.0, with ±2dB of random jitter added. The bandwidth parameter is based on the base bandwidth value base_bandwidths. A certain amount of random jitter is added to the bandwidth of each radiation source. If the bandwidth of radiation sources is less than the preset number, it is automatically padded with 0 to ensure the consistency of the data structure.
[0016] Step 1.3: Observation Station Location Processing Initial position jitter: Random jitter is added to the X and Y directions of the original observation station position, while the Z direction remains unchanged; Center of gravity calculation: The average position of the three observation stations after the shaking is used as the center of gravity; Coordinate transformation: Converting the location of the observation station into a local coordinate system relative to the center of gravity; Label generation: Add normally distributed noise with a mean of 0 and a certain standard deviation to the actual location to generate noisy label locations, simulating the actual measurement error.
[0017] Step 1.4: Generating the location and velocity of the radiation source Radiation source location: The location is generated using a polar coordinate random sampling + distance constraint verification mechanism.
[0018] Radiation source velocity: The velocities in the X and Y directions are uniformly and randomly distributed, while the velocities in the Z direction are fixed at 0. If the number of radiation sources is less than the preset number, the velocities will be automatically filled with 0.
[0019] Step 1.5: Data Assembly and Output The sample ID, signal parameters, observation station parameters, and radiation source parameters are assembled into row data according to a preset field structure. A CSV configuration file is generated using pandas, with fields including Sample_ID, Num_Sources, SNR_dB, St1~St3_True_X / Y / Z, Src1~Src3_X / Y / Z / Vx / Vy / Vz, etc.
[0020] Step 2: Dataset Construction Phase This stage reads the CSV configuration file and, through signal modeling, time-frequency analysis, noise addition, and data normalization, generates a .pkl format dataset containing time-frequency plots, time-domain plots, and label information. Specifically, it includes the following sub-steps: Step 2.1: Initialize the dataset generator Set the number of receiving stations and the number of array elements for each receiving station; Initialize placeholders for receiver position, center of gravity, and velocity; receiver velocity is set to default. Step 2.2: Doppler frequency shift calculation Calculate the Doppler frequency shift for each array element based on the carrier frequency of the radiation source and the relative velocity. Step 2.3: Single-sample signal modeling and feature extraction Chirp signal generation, signal delay and Doppler modulation: The signal delay is calculated based on the distance between the array element and the radiation source, and the delay is achieved through time-domain rolling; the modulation factor is generated by combining Doppler frequency shift, and the delayed signal is modulated. Noise addition: Noise power is generated based on the signal-to-noise ratio (SNR) _dB, complex Gaussian noise is generated and superimposed on the signal.
[0021] Time-domain feature extraction: Extract the real and imaginary parts of the signal at time t=0 and construct a time-domain graph; Time-frequency feature extraction: STFT transformation is performed using the Hanning window, specifying the window length, overlap rate, and number of Fourier transform points; the amplitude spectrum is extracted from the STFT result and resampled into a time-frequency plot; Data normalization: Divide the time-frequency plot by the global maximum value, and divide the real and imaginary parts of the time-domain plot by their respective maximum absolute values to ensure that the numerical range is [0,1] and [-1,1].
[0022] Step 2.4: Tag Parameter Assembly Information such as the receiving station status, the normalized position of the array elements, and the true position and velocity of the radiation source are assembled into tag data.
[0023] Step 2.5: Batch Generation and Saving Iterate through all samples in the CSV configuration file, create subdirectories based on "number of radiation sources / signal-to-noise ratio", and save the sample data as .pkl format files; at the same time, generate a log file to record the sample ID, save path, generation status, and error information. Beneficial effects
[0024] The method of the present invention has the following advantages compared with the prior art: 1. Flexible and controllable parameters: Supports customization of multiple parameters such as the number of radiation sources, signal-to-noise ratio range, and bandwidth jitter amplitude. It can quickly generate datasets that meet different experimental scenarios and adapt to the training and validation needs of algorithms such as multi-target detection and localization.
[0025] 2. Physically sound data: The target location overlap is avoided by using the radiation source distance constraint mechanism, and normal distribution noise is added to the location of the observation station tag to simulate measurement error, so that the generated data is closer to the signal characteristics of real radar / communication systems.
[0026] 3. The experiment is fully reproducible: The multi-random seed control mechanism can generate completely consistent datasets under the same input parameters, ensuring the reliability of performance comparison results of different algorithms and different experimental batches.
[0027] 4. Rich feature dimensions: Simultaneously extracts time-domain features at time t=0 and STFT time-frequency joint features, providing more comprehensive signal features for deep learning models and improving the model's ability to distinguish complex multi-target scenes.
[0028] 5. Batch Automated and Efficient: Supports sample ID range filtering, subdirectory classification and saving, and generation of status records, enabling automated generation and management of large-scale datasets and reducing manual operation costs. Attached Figure Description
[0029] Figure 1 This is a schematic diagram of the process in the method and embodiments of the present invention; Detailed Implementation
[0030] The technical solution of the method of the present invention will be clearly and completely described below with reference to specific parameter configurations.
[0031] Example 1: Configuration file generation 1. Set input parameters Output filename: sim_config.csv; Combinations of radiation sources: [1,2,3]; Signal-to-noise ratio range: np.arange(-15,-4,1) (in dB); Number of samples per configuration: 10; Basic bandwidth: [8000, 16000, 8000] (in Hz); Original observation station location: [[500,250,600],[400,76.79492,600],[600,76.79492,600]] (unit: km); Master random seed: 100; Minimum distance between radiation sources: 10km.
[0032] 2. Execute the configuration file to generate code. Run the configuration file generation script to generate a CSV configuration file based on the above parameters. The total number of samples is 3×11×10=330. Each sample contains complete fields such as sample ID, signal parameters, observation station parameters, and radiation source parameters.
[0033] Example 2: Dataset Construction 1. Initialize the dataset generator Create a MultiDPDDatasetGenerator instance with a default number of 3 receiving stations, 16 array elements, and a sampling rate of 50kHz.
[0034] 2. Set dataset generation parameters CSV configuration file path: sim_config.csv; Output root directory: . / generated_datasets; Sample ID range: (0, 329).
[0035] 3. Execute the dataset generation code The dataset generation script is run to batch read sample parameters from the CSV configuration file and sequentially perform operations such as signal modeling, delay and Doppler modulation, noise addition, time-frequency feature extraction and normalization, and label assembly. The generated .pkl files are saved according to "number of radiation sources / signal-to-noise ratio," with the file naming format sample_${Sample_ID}_${num_sources}_${SNR}_${local_index}.pkl. Simultaneously, a generation_record.txt file is generated to record the generation status of each sample.
Claims
1. A method for automatically generating simulation data for direct localization of multiple targets, characterized in that, This includes the configuration file generation phase and the dataset construction phase; The configuration file generation stage is used to generate a standardized configuration file containing all parameters of the sample, specifically including: A master random seed is set, and multiple dedicated random seeds are derived based on the master random seed, each independently controlling the random generation process of position, power, bandwidth and speed parameters; Generate signal parameters, including signal power and bandwidth parameters with random jitter; Processing the location of the observation station includes randomly jittering the original location, coordinate transformation, and generating noisy labeled locations; The location and velocity of the radiation source are generated. The location of the radiation source is generated using a distance constraint mechanism that combines polar coordinate sampling with Euclidean distance verification to avoid location overlap. The above parameters are assembled into structured row data and output as a configuration file; The dataset construction phase is used to read the configuration file and generate simulation datasets in batches, specifically including: Initialize the dataset generator, and set the number of receiving stations and array elements; Doppler frequency shift is calculated based on the carrier frequency of the radiation source and the relative velocity. Single-sample signal modeling and feature extraction are performed, including generating Chirp signals and performing time-domain delay and Doppler modulation, superimposing complex Gaussian noise, extracting time-domain features at t=0 and time-frequency joint features based on short-time Fourier transform, and normalizing the extracted features. Assemble tag parameters, including receiver station status, array element normalized position, and actual motion state of radiation source; Iterate through all samples in the configuration file, classify them according to preset rules, save them as dataset files, and generate record files.
2. The method according to claim 1, characterized in that, During the configuration file generation phase, the process of deriving multiple dedicated random seeds based on the master random seed specifically includes: Set a master random seed master_seed; Based on the master random seed, seven dedicated random seeds are derived through deterministic hashing or offset operations; The seven dedicated random seeds are used independently to control the randomness of observation station position jitter, signal power generation, bandwidth jitter generation, radiation source position sampling, observation station noise generation, radiation source velocity generation, and sample ID allocation, ensuring that the random processes of each parameter are independent of each other and completely reproducible under the same master seed.
3. The method according to claim 1, characterized in that, During the configuration file generation phase, the generated signal parameters specifically include: Signal power generation: Based on a reference power value, random jitter values within a preset range are superimposed; Bandwidth parameter generation: Based on the base bandwidth value, random jitter is added to the bandwidth of each radiation source; if the number of radiation sources in the current sample is less than the preset maximum number, the bandwidth parameters of the remaining radiation sources are automatically padded with zeros to ensure the consistency of the data structure dimensions.
4. The method according to claim 1, characterized in that, During the configuration file generation phase, the processing of the observation station location specifically includes: Initial position jitter: Random jitter is added to the X-axis and Y-axis coordinates of the original observation station position, while the Z-axis coordinate remains unchanged; Center of gravity calculation and coordinate transformation: Calculate the geometric center of gravity of all observation stations after the shaking, and transform the coordinates of each observation station into local coordinate system coordinates relative to the center of gravity; Tag generation: Based on the actual location of the observation station, a normal distribution noise with a mean of 0 and a preset standard deviation is superimposed to generate a noisy tag location that simulates the actual measurement error.
5. The method according to claim 1, characterized in that, During the configuration file generation phase, the distance constraint mechanism for generating the radiation source position and velocity specifically includes: Candidate radiation source locations are generated using a polar coordinate random sampling method; Calculate the Euclidean distance between the candidate location and the location of the generated radiation source; If the Euclidean distance is less than the preset minimum spacing threshold, the candidate position is discarded and resampled until the distance constraint condition is met or the maximum number of attempts is reached. Radiation source velocity generation: Velocity components are generated in a uniform random distribution in the X and Y axes, while the velocity component in the Z axis is fixed at zero; if the number of radiation sources is less than the preset maximum number, the velocity parameters of the remaining radiation sources are automatically filled with zeros.
6. The method according to claim 1, characterized in that, In the dataset construction phase, the single-sample signal modeling and feature extraction specifically include: Chirp signal generation and modulation: The signal propagation delay is calculated based on the distance between the array element and the radiation source, and the delay is processed by time-domain rolling operation; the phase modulation factor is generated by combining the calculated Doppler frequency shift, and the delayed signal is Doppler modulated. Noise addition: Calculate the noise power based on the signal-to-noise ratio (SNR) _dB specified in the configuration file, generate complex Gaussian white noise of the corresponding power, and superimpose it onto the modulated signal; Time-domain feature extraction: Extract the complex signal snapshots of all array elements at time t=0, separate the real and imaginary parts to construct a time-domain graph; Time-frequency feature extraction: Perform short-time Fourier transform (STFT) on the signal, specify the window function type, window length and overlap rate, extract the amplitude spectrum and resample it into a time-frequency graph of fixed size.
7. The method according to claim 6, characterized in that, During the dataset construction phase, the normalization process for the extracted features specifically includes: For the time-frequency plot features, divide by the global maximum value to compress the numerical range to the [0, 1] interval; For the time-domain plot features, the real and imaginary data are divided by their respective maximum absolute values to compress the numerical range to the interval [-1, 1].
8. The method according to claim 1, characterized in that, The process of classifying and saving the dataset according to preset rules during the dataset construction phase specifically includes: In the output root directory, create subdirectories based on the two levels of "number of radiation sources" and "signal-to-noise ratio"; The time-frequency plot, time-domain plot and label information of each sample are serialized into binary files and saved to the corresponding subdirectory according to the naming format including sample ID, number of radiation sources, signal-to-noise ratio and local index; Generate a text-formatted log file that records the ID, save path, generation status, and error information of each sample in real time.
9. An automated simulation data generation device for direct positioning of multiple targets, characterized in that, include: A configuration generation module is used to perform the configuration file generation stage steps as described in any one of claims 1 to 5, generating a standardized configuration file that includes multiple random seed control, distance-constrained location, and noisy labels; The dataset construction module is used to perform the dataset construction phase steps as described in any one of claims 1, 6 to 8, read the configuration file and complete signal modeling, Doppler modulation, time-frequency feature extraction and batch saving; The storage management module is used to classify and manage the generated dataset files and generation status record files according to the number of radiation sources and the signal-to-noise ratio.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the simulation data automatic generation method for direct localization of multiple targets as described in any one of claims 1 to 8.
11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method for automatically generating simulation data for direct localization of multiple targets as described in any one of claims 1 to 8.