Distributed radar resource allocation method based on deep learning

Abstract

The application discloses a distributed radar resource allocation method based on deep learning, comprising the following steps: establishing an independent IMM filter for each radar and each tracked target in the radar field of view, and generating a local track of the tracked target; receiving the local track of the tracked target sent by each radar by a central controller, and generating a global fusion track of the tracked target; calculating a core evaluation index of the tracked target based on the global fusion track of the tracked target; converting the core evaluation index of the tracked target into a resource allocation instruction for each radar; taking the global fusion track of the tracked target as input and the predicted track of the tracked target as output, and constructing and training a track prediction model of the tracked target; and the application realizes more comprehensive and intelligent understanding and response to the situation, thereby showing superior tracking accuracy and resource efficiency when coping with high-maneuvering and high-threat targets.

Description

Technical Field

[0001] This invention relates to the field of radar resource management technology, and more specifically to a distributed radar resource allocation method based on deep learning. Background Technology

[0002] In trajectory prediction technology, accurately predicting the position of maneuvering targets is one of the core tasks for improving multi-target tracking performance. Traditional target tracking methods mostly rely on model-based filtering algorithms, such as the Kalman Filter (KF) and its extended forms (EKF, UKF). These methods require pre-assuming that the target's motion follows one or more fixed motion models (e.g., constant velocity (CV), constant acceleration (CA), constant turn (CT) models). However, in reality, the motion patterns of highly maneuverable targets (such as modern fighter jets and drones) are complex and varied, often switching nonlinearly and unexpectedly between different models. This makes it difficult for traditional filters to accurately model these patterns, leading to decreased prediction accuracy and even tracking failure.

[0003] In recent years, with the development of deep learning technology, Recurrent Neural Networks (RNNs) and their advanced variants, especially Long Short-Term Memory Networks (LSTMs), have achieved significant success in the field of time series prediction. LSTMs, through their unique gating mechanisms (input gate, forget gate, output gate), effectively solve the gradient vanishing or exploding problems encountered by traditional RNNs when processing long sequences, enabling them to learn long-term dependencies with large intervals in the sequence data. In the field of target tracking, researchers have begun to apply LSTM networks to learn the motion patterns of targets directly from large amounts of trajectory data. Compared to traditional methods that rely on fixed motion models, LSTMs can autonomously learn more complex nonlinear maneuvering laws in a data-driven manner, providing a powerful and more generalizable new technological approach for predicting the future position of targets and improving tracking accuracy.

[0004] On the other hand, modern radar technology is developing towards intelligence and adaptability, with cognitive radar (CR) being a representative concept. Cognitive radar is defined as a system capable of intelligent closed-loop interaction with its environment. Its core idea lies in a "Perception-Action Cycle": the radar detects the environment by transmitting signals (perception), then analyzes the echo information to understand the scenario (learning and decision-making), and finally dynamically adjusts its next transmission and reception strategy based on this understanding (action) to optimize the preset system objectives.

[0005] Dynamic Resource Management (DRM) is one of the key technologies for realizing cognitive radar. In complex multi-target tracking scenarios, radar resources such as energy and time are limited. DRM aims to optimize the allocation of these limited resources in real time and dynamically based on the priority, importance, or uncertainty of each target in the scenario. This requires the system to have a decision engine capable of evaluating the scenario, quantifying the value of different targets, and making reasonable scheduling accordingly.

[0006] Existing DRM research has proposed some scheduling criteria based on information theory and task, providing a theoretical basis and implementation framework for radar's "cognitive" behavior.

[0007] While some existing cognitive radar resource management schemes attempt to incorporate multi-model filters for resource scheduling, they typically suffer from drawbacks: Most employ a centralized processing architecture, where all radars transmit raw measurement data back to a central hub for unified filtering and decision-making. This not only places extremely high demands on communication bandwidth but also cripples the entire system if the central node fails, resulting in poor robustness. In terms of decision-making, they primarily rely on information theory-based criteria, essentially allocating resources around the single dimension of "uncertainty," lacking sufficient awareness of tactical intentions such as high target mobility and high threat levels. Summary of the Invention

[0008] The purpose of this invention is to provide a deep learning-based distributed radar resource allocation method. It addresses the common problems in existing radar resource management technologies, such as a single evaluation dimension, poor robustness in tracking highly maneuverable targets, and low efficiency of multi-sensor collaborative mechanisms. The invention provides a more intelligent, accurate, and forward-looking cognitive radar resource management system and method.

[0009] To achieve the above functions, this invention designs a distributed radar resource allocation method based on deep learning. For the central controller and the distributed radar network composed of multiple radars, the following steps S1-S5 are executed to complete the resource allocation of each radar in the multi-target tracking task and the trajectory prediction of the tracked target:

[0010] Step S1: For each radar and for the tracked target in the radar's field of view, establish an independent IMM filter. The IMM filter performs preliminary processing on the radar's local observations by running Kalman filters for different motion models in parallel, generating the local trajectory of the tracked target.

[0011] Step S2: The central controller receives the local trajectories of the tracked target sent by each radar, and fuses the local trajectories of the tracked target to obtain the global fused trajectory of the tracked target.

[0012] Step S3: Based on the global fusion trajectory of the tracked target and the trajectory prediction at the current moment, calculate the core evaluation indicators of the tracked target, including uncertainty score, threat score, and distance.

[0013] Step S4: Using decoupled decision-making logic, the core evaluation indicators of the tracked target are transformed into resource allocation instructions for each radar.

[0014] Step S5: Using the global fusion trajectory of the tracked target obtained in step S2 as input and the trajectory prediction of the tracked target at the next moment as output, construct and train the trajectory prediction model of the tracked target, obtain the trained trajectory prediction model, apply the trained trajectory prediction model to provide a forward-looking decision basis for steps S3 and S4 at the next moment, and complete the trajectory prediction of the tracked target.

[0015] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0016] 1. Significantly improved tracking accuracy and robustness: This invention effectively combines the adaptive capability of the IMM filter to maneuvering targets and the advantages of multi-radar collaborative observation through the architecture of "distributed IMM tracking + centralized CI fusion", overcoming the shortcomings of the single filter model in the existing technology that is difficult to deal with highly maneuvering targets.

[0017] 2. Comprehensiveness and Intelligence of Decision-Making: This invention constructs a three-dimensional cognitive assessment system integrating "uncertainty, threat level, and distance," solving the problem of existing technologies relying solely on the single dimension of "uncertainty" for decision-making. This multi-dimensional assessment enables the system to comprehensively consider its own tracking quality, the target's tactical intentions, and physical observation conditions, resulting in more comprehensive and intelligent decision-making.

[0018] 3. Maximizing resource utilization efficiency: The core advantage of this invention lies in its decoupled "targeted" resource allocation logic, which can accurately allocate the most suitable resources to the "most needed" targets and scenarios, avoiding resource waste on stable, low-threat targets. Attached Figure Description

[0019] Figure 1 This is an overall architecture diagram of a deep learning-based distributed radar resource allocation method provided according to an embodiment of the present invention.

[0020] Figure 2 This is a flowchart of a deep learning-based distributed radar resource allocation method provided according to an embodiment of the present invention;

[0021] Figure 3 This is a graph showing the root mean square error (RMSE) performance of tracking and prediction in a uniform linear motion scenario, provided by an embodiment of the present invention.

[0022] Figure 4 This is a graph showing the root mean square error (RMSE) performance of tracking and prediction in a single turning motion scenario, provided by an embodiment of the present invention.

[0023] Figure 5 This is a graph showing the root mean square error (RMSE) performance of tracking and prediction in a complex S-shaped motion scenario, provided by an embodiment of the present invention.

[0024] Figure 6 This is a combined trajectory diagram from the last Monte Carlo simulation provided according to an embodiment of the present invention;

[0025] Figure 7 This is an average power distribution heatmap provided according to an embodiment of the present invention;

[0026] Figure 8 This is a heatmap of average residence time distribution provided according to an embodiment of the present invention. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0028] The deep learning-based distributed radar resource allocation method provided in this invention proposes a unified, closed-loop integrated multi-radar collaborative target tracking and cognitive resource management technology solution. Its core lies in constructing a highly robust architecture of "distributed interactive multiple model (IMM) filtering tracking - central covariance intersection (CI) fusion," and integrating a three-dimensional cognitive assessment system of "uncertainty-threat level-distance" and decoupled targeted resource allocation logic, achieving an intelligent closed loop of perception, assessment, decision-making, and action. Furthermore, it integrates a data-driven bidirectional long short-term memory network (BiLSTM-Attention) with an attention mechanism, deeply embedding it into the cognitive decision-making loop as a trajectory prediction model for the tracked target, providing forward-looking input.

[0029] Reference Figure 1 The architecture of this method consists of a main cognitive loop integrating prediction and decision-making. This main cognitive loop depicts a complete cognitive cycle of "perception-assessment-decision-action": a distributed radar network generates observation data, which is then processed by a distributed IMM filter to generate local trajectories, which are then aggregated at a central controller for fusion. Based on the output of the central controller, an assessment module calculates threat scores, uncertainty scores, and distance scores, which are then fed into a cognitive decision-making and resource allocation module. This module calculates the optimal resource scheduling command and sends it back to the distributed radar network, completing the loop. A forward-looking trajectory prediction module obtains data from the central controller, and its prediction results serve as one of the core inputs to the cognitive decision-making and resource allocation module for the next time step.

[0030] Reference Figure 2 The specific method is as follows:

[0031] For the central controller and the distributed radar network consisting of multiple radars, the following steps S1-S5 are executed to complete the resource allocation of each radar in the multi-target tracking task and the trajectory prediction of the tracked target:

[0032] Step S1: For each radar and for the tracked target in the radar's field of view, establish an independent IMM filter. The IMM filter performs preliminary processing on the radar's local observations by running Kalman filters for different motion models in parallel, generating the local trajectory of the tracked target.

[0033] Step S1 is executed independently at each node of the radar network. Each radar in the distributed radar network maintains an independent IMM filter for each tracked target within its field of view. The IMM algorithm achieves fast adaptive tracking of highly maneuverable targets by running N Kalman filters for different motion models in parallel (N=3 in this embodiment) and through probabilistic interaction and state fusion between models. The specific steps at time k are as follows:

[0034] Step S1.1: The purpose of this step is to calculate an optimal "initial state" for the upcoming filtering step. It does not directly use the outputs of the Kalman filters from the previous time step. Instead, it assumes that the tracked target may switch from any old motion model to the current new motion model, and weights and fuses the states of all the old motion models, expressed as follows:

[0035] ;

[0036] In the formula, The mixed probability represents the probability that, given that the tracked target will be in the j-th motion model at time k, it was previously in the ith motion model at time k-1. Let be the posterior probability of the i-th motion model at time k-1, representing the confidence that the tracked target is in the i-th motion model after the previous update; The elements in the motion model transition probability matrix Π represent the prior probability of the tracked target transitioning from the i-th motion model to the j-th motion model, and are fixed design parameters. This is a normalization constant, ensuring that the sum of all probabilities of transitioning from the old motion model to the new motion model is 1;

[0037] The sum of the probabilities of all possible paths to the j-th motion model is used as the normalization constant. It incorporates all possibilities of transitioning from the old motion model (i=1,2,…N) to the new motion model, serving as the denominator when calculating the mixed probability, as follows:

[0038] ;

[0039] In the formula, N is the total number of motion models in the IMM filter;

[0040] Step S1.2: Calculate the initial state of the mixture for each Kalman filter, including the weighted average filter state and the mixture covariance, as the input for prediction of each Kalman filter, as shown in the following formula:

[0041] ;

[0042] ;

[0043] In the formula, This indicates the filter state after weighted averaging. This represents the mixture covariance, used to quantify the degree of dispersion between different motion model states, ensuring that the uncertainty caused by the differences between motion models is not lost during the mixing process; Let represent the posterior state estimate for the i-th motion model at time k-1. Let represent the posterior covariance of the i-th motion model at time k-1;

[0044] Step S1.3: Based on local observations The likelihood function is constructed as follows:

[0045] ;

[0046] This represents the likelihood value of the j-th motion model, which quantifies "the likelihood of the current local observation under the assumption that the tracked target follows the j-th motion model". How likely is it? This represents the Gaussian (normal) probability density function; This represents the actual radar observation value at time k; This represents the prediction of the j-th motion model for the observation at time k; Let represent the covariance of the j-th motion model, representing the uncertainty of the prediction, if the actual local observations... If the prediction of the j-th motion model is very close to that of the j-th motion model, the value of the Gaussian function will be large, meaning that the likelihood value of the j-th motion model is very high. Very high;

[0047] Update the posterior probabilities of the motion model:

[0048] ;

[0049] In the formula, Let represent the posterior probability of the j-th motion model at time k; c represents the constant used for probability summation and normalization.

[0050] The above formula utilizes Bayes' theorem, combined with likelihood values. and Calculate the posterior probability of the j-th motion model. , representing the latest confidence level of the system's judgment that "the tracked target is currently in the j-th motion model" after obtaining the observation data at time k. The higher the likelihood value of the motion model, the higher its updated probability.

[0051] The purpose of this step is to evaluate which motion model best explains the current behavior of the tracked target and update the confidence (probability) of each motion model accordingly.

[0052] Step S1.4: Weight and fuse the results of all parallel Kalman filters to output the radar's local trajectory estimate of the tracked target at time k, as shown in the following formula:

[0053] ;

[0054] ;

[0055] In the formula, This represents the final state estimate, consisting of the states of all independent filters. The weighted average is calculated by taking the weights from the posterior probabilities of the j-th motion model at time k, which are the most recently updated and best reflect the dynamics of the tracked target. ; This represents the state estimate after the j-th motion model is updated at time k; This represents the final covariance; Let represent the covariance after the update of the j-th motion model at time k.

[0056] The following are the simulation environment and parameter settings for the application example:

[0057] (1) Overall simulation and scene setting:

[0058] Monte Carlo simulations: 200 times;

[0059] Number of simulated targets: The scene always contains 3 tracked targets, representing three typical motion modes;

[0060] Trajectory types: The real trajectories used in the simulation are randomly selected from a pre-generated trajectory dataset, ensuring diversity: straight-line motion trajectories, single-turn maneuver trajectories, and complex S-shaped maneuver trajectories;

[0061] Track length: Each track contains 150 time steps;

[0062] Time step: 1.0 second.

[0063] (2) Simulation trajectory dataset:

[0064] Uniform linear trajectory: The entire trajectory is generated using a standard uniform linear motion model;

[0065] Single turning trajectory: The trajectory adopts a uniform turning model for 70-100 time steps, with the turning rate randomly perturbed around +2.0° / s, and a uniform linear motion model for the remaining time.

[0066] Complex S-shaped trajectory: The trajectory makes a positive turn at time steps 39 to 71 (with a random perturbation of the turning rate around +2.0° / s), and a reverse turn at time steps 99 to 131 (with a random perturbation of the turning rate around -3.0° / s). The remaining time adopts a uniform linear motion model.

[0067] (3) Radar network and observation model parameters:

[0068] Radar Quantity and Layout: The system consists of 3 radars, with their Cartesian coordinates (in meters) as follows:

[0069] Radar 1: [-1000, 0];

[0070] Radar 2: [2000, 12000];

[0071] Radar 3: [9000, 6000];

[0072] Standard deviation of base distance noise: 200.0 meters. This value represents the base noise level at a specific reference distance and per unit resource.

[0073] System noise standard deviation: 100.0 meters, representing the fixed noise introduced by internal components such as the radar receiver;

[0074] Angle measurement noise standard deviation: 1.5 degrees;

[0075] Reference distance: 10000.0 meters, used to calculate signal propagation attenuation.

[0076] (4) IMM filter and state evaluation parameters:

[0077] Motion model: Distributed interactive multi-model (IMM) filters operate in parallel with three Kalman filters, corresponding to a uniform linear model.

[0078] Uniform speed turning model 1: Turning rate +1.0 degrees / second;

[0079] Uniform speed turning model 2: Turning rate -3.0 degrees / second;

[0080] Motion model transition probability matrix Π: This matrix defines the prior probability of switching between the three motion models within a time step; the matrix setting reflects that the target has a high probability (0.95) of maintaining its current motion state, while the probability of switching to other motion modes is low (0.025). Specific numerical settings are as follows:

[0081] ;

[0082] Process noise covariance: used to describe the uncertainty of the motion model;

[0083] Noise level for the uniform linear model: 0.1;

[0084] Noise during the turning model process: 0.5.

[0085] Step S2: The central controller receives the local trajectories of the tracked target sent by each radar, and fuses the local trajectories of the tracked target to obtain the global fused trajectory of the tracked target.

[0086] This step, performed by the central controller, aims to fuse local trajectories from all distributed radar nodes to generate a more accurate and robust global target state estimate than any single node.

[0087] The central controller calls the `fuse_tracks_ci` function. This function employs the CI algorithm, which is particularly suitable for fusing multiple data sources with unknown or time-varying correlations. It linearly combines the various information matrices (inverses of the covariance matrices) by calculating an optimized weight ω, thereby ensuring the consistency of the covariance after fusion. For the i-th tracked target, at time k, the local trajectories from M radars are fused ( , ), where j=1,2,…M, to obtain the global fusion trajectory ( , ).

[0088] Step S3: Based on the global fusion trajectory of the tracked target and the trajectory prediction at the current moment, calculate the core evaluation indicators of the tracked target, including uncertainty score, threat score, and distance.

[0089] The central controller is based on the global fusion trajectory obtained in step S2. , For each tracked target, three core evaluation metrics are calculated, and the specific steps are as follows:

[0090] Step S3.1: Calculate the uncertainty score, which aims to quantify the system's confidence in estimating the location of the tracked target, as shown in the following formula:

[0091] ;

[0092] In the formula, To integrate the posterior covariance matrix The trace of the mid-position component; This represents the position variance of the i-th tracked target along the X-axis after fusion; This represents the position variance of the i-th tracked target in the Y-axis direction after fusion;

[0093] Will and The summation yields an approximate measure of the area of ​​the uncertainty ellipse of the tracked target on the two-dimensional plane. A larger uncertainty score means that the estimated position of the tracked target may deviate more from its true position, i.e., the tracking quality is lower, and more resources are needed to reduce its uncertainty.

[0094] Step S3.2: Based on the predicted position at time k generated in step S5 at time k-1, the position at time k is... And the global fusion state obtained at time k-1 based on step S2. Calculate the predicted instantaneous velocity and instantaneous acceleration As shown in the following formula:

[0095] ;

[0096] ;

[0097] In the formula, , These represent the position and instantaneous velocity of the tracked target at time k-1, respectively. Indicates time difference;

[0098] The threat score is calculated as follows:

[0099] ;

[0100] In the formula, This represents the threat score of the i-th tracked target. and These represent the magnitudes of the velocity and acceleration vectors of the tracked target at the current moment, respectively. and This represents the preset speed and acceleration weighting coefficients; among them, acceleration is given a higher weight and is enhanced through squaring, because it directly reflects the intensity of the maneuver of the currently tracked target.

[0101] Step S3.3: Based on the predicted position at time k generated at time k-1 in step S5, calculate the distance score as follows:

[0102] ;

[0103] In the formula, Let be the distance score of the i-th tracked target; Let be the predicted position of the i-th tracked target at time k; Let represent the location of the j-th radar in the distributed radar network, and M represent the number of radars. Let be the distance between a single radar and the i-th tracked target.

[0104] According to radar equations, signal strength is inversely proportional to the fourth power of distance; therefore, distance is the most critical factor affecting signal-to-noise ratio (SNR). Selecting the minimum distance represents the physical condition of the node in the network with the greatest observational advantage over the tracked target. A higher distance score means that even the optimal node faces greater difficulty in observing it.

[0105] Step S4: Using decoupled decision-making logic, the core evaluation indicators of the tracked target are transformed into resource allocation instructions for each radar.

[0106] Step S4 transforms the core evaluation indicators from Step S3 into specific radar resource (power, loiter time) allocation instructions, mapping different evaluation dimensions to the radar resources that can most effectively improve that dimension, thus achieving more physically meaningful and targeted management. The specific steps are as follows:

[0107] Step S4.1: Calculate the power allocation priority score for the tracked target. Power allocation is driven by "uncertainty" and "distance," based on the fact that increasing transmission power is the most direct way to overcome signal propagation attenuation (long distance) and improve the signal-to-noise ratio (high uncertainty), as shown in the following formula:

[0108] ;

[0109] In the formula, This represents the power allocation priority score of the i-th tracked target; This represents the normalized distance score; This represents the normalized uncertainty score; , These represent the corresponding weights for distance and uncertainty, respectively.

[0110] Step S4.2: Calculate the time allocation priority score for the tracked target. The time allocation is driven by "threat" and "uncertainty," based on the fact that increasing dwell time is an effective way to capture rapid dynamic changes in the target (high threat) and improve poor tracking status (high uncertainty), as shown in the following formula:

[0111] ;

[0112] In the formula, This represents the time allocation priority score for the i-th tracked target; This represents the normalized threat score; This represents the normalized uncertainty score; , These represent the corresponding weights of the threat score and the uncertainty score, respectively.

[0113] Step S4.3: Calculate the proportional allocation parameters with minimum guarantee as follows:

[0114] ;

[0115] ;

[0116] In the formula, This represents the proportion of resources ultimately obtained by the i-th tracked target; The minimum guaranteed resource ratio set for each tracked target; This represents the total number of tracked targets in the scene. The remaining dynamically allocable resource proportion after removing the minimum guaranteed resource proportion; The total priority score of the i-th tracked target;

[0117] Step S4.4: Perform a two-stage update of resource allocation for each radar:

[0118] Phase One: Basic Support; for all Each of the tracked targets is assigned a fixed minimum guaranteed resource ratio. This ensures that even the lowest priority tracked targets receive the "safety net" resources needed to maintain basic tracking, avoiding trajectory loss due to a complete lack of resource allocation.

[0119] Phase Two: On-Demand Allocation; Allocating the remaining dynamically allocable resources. According to the total priority score of each tracked target The proportion of resources allocated to each tracked target within the total score is dynamically determined. Ultimately, each tracked target receives a specific proportion of resources. It is the sum of its "guaranteed minimum" and "on-demand" portions.

[0120] The following are the decision logic and resource allocation parameter settings in the application example:

[0121] Decision-making logic: A "decoupled" decision-making approach is adopted, where uncertainty, threat level, and distance drive the allocation of different resources respectively;

[0122] Threat weighting: When calculating the "threat score," the weighting coefficients for velocity and acceleration are as follows:

[0123] Speed ​​weight: 0.4;

[0124] Acceleration weight: 0.6;

[0125] Resource allocation priority weight:

[0126] Power allocation: distance weight 0.7, uncertainty weight 0.3;

[0127] Time allocation: Threat weight 0.7, Uncertainty weight 0.3;

[0128] Total resource budget:

[0129] Total power budget: 50.0;

[0130] Total stay budget: 1.0;

[0131] It should be specifically noted that the total resource budget value set in this embodiment of the invention is an abstract unit defined for this simulation experiment, aiming to construct a stable and analyzable verification environment, rather than representing the physical parameters of any specific radar hardware. The threat weight and resource allocation priority weight given are optimal combinations obtained through optimization in numerous Monte Carlo simulation experiments, with the goal of maximizing system tracking accuracy and resource utilization efficiency. These specific values ​​aim to verify the feasibility and superiority of the technical solution of this invention.

[0132] Minimum resource guarantee ratio: To ensure the continuity of tracking for all targets, a minimum of 2% of the resources is allocated to each target as a guarantee.

[0133] Step S5: Using the global fusion trajectory of the tracked target obtained in step S2 as input and the trajectory prediction of the tracked target at the next moment as output, construct and train the trajectory prediction model of the tracked target, obtain the trained trajectory prediction model, apply the trained trajectory prediction model to provide a forward-looking decision basis for steps S3 and S4 at the next moment, and complete the trajectory prediction of the tracked target.

[0134] The specific steps of step S5 are as follows:

[0135] Step S5.1: Perform differential calculation on the historical position sequence of the global fusion trajectory of the tracked target with a length of 16 obtained in step S2 to construct a 6-dimensional feature vector containing position, velocity, and acceleration. This provides the trajectory prediction model with information about the momentum and maneuvering trends of the tracked target. As shown in the following formula:

[0136] ;

[0137] In the formula, This represents the 6-dimensional eigenvector at time k; Indicates the position at time k; Represents the velocity at time k; Represents the acceleration at time k;

[0138] 6-dimensional feature vectors Scaling to a uniform numerical range improves the stability and efficiency of trajectory prediction model training;

[0139] Step S5.2: Based on the Bidirectional Long Short-Term Memory Network (BiLSTM-Attention) with attention mechanism, construct the trajectory prediction model of the tracked target. Take the 6-dimensional feature vector at time k obtained in step S5.1 as input and the position of the tracked target at time k+1 as output to train the trajectory prediction model of the tracked target.

[0140] At the heart of the trajectory prediction model is a BiLSTM-Attention network, whose architecture is designed to maximize the extraction of key information for prediction from time-series data.

[0141] The BiLSTM layer, serving as the foundation of the model, processes the input sequence from both positive and negative directions. This bidirectional mechanism allows it to simultaneously acquire past and future contextual information when analyzing any point in the sequence, thus enabling a deeper understanding of the complete maneuver process and effectively capturing long-term dependencies in the trajectory. Connected to the BiLSTM layer, an attention mechanism dynamically assigns "importance" weights to each time step in the sequence through autonomous learning. This allows the trajectory prediction model to focus on the most critical time points for prediction (the turning point where the maneuver begins), while ignoring relatively unimportant information, resulting in a more intelligent and efficient generalization of the entire sequence.

[0142] The output of the attention mechanism (a weighted and fused context vector) is ultimately fed into a fully connected layer, which is responsible for mapping it from the high-dimensional feature space to the final output, i.e., for the next time step (pos). x , pos y Precise prediction of two-dimensional location.

[0143] The trajectory prediction model is trained offline using supervised learning. Its core objective is to learn an accurate mapping from the historical trajectory after IMM filtering to the next moment's position of the ideal noise-free trajectory. This training strategy enables the trajectory prediction model to learn not only prediction but also "denoising." The training process uses Mean Squared Error (MSE) as the loss function and combines regularization techniques such as Dropout and early stopping to prevent overfitting, thus ensuring its ability to generalize to unseen new trajectories.

[0144] The following are the trajectory prediction model parameters and training hyperparameter settings in the application example:

[0145] Model architecture: Bidirectional long short-term memory network with attention mechanism;

[0146] Input feature dimensions: 6 (x / y position, x / y velocity, x / y acceleration);

[0147] Input sequence length: 16;

[0148] Number of hidden units in LSTM layer: 256;

[0149] Number of LSTM layers: 2;

[0150] Output dimension: 2 (x / y position at the next time step);

[0151] Regularization: 0.2 (applied to the LSTM layer and before the final output);

[0152] The training hyperparameters are set as follows:

[0153] Optimizer: Adam;

[0154] Initial learning rate: 0.001;

[0155] Batch size: 128;

[0156] Training cycles: 200;

[0157] Early stopping mechanism: Training is terminated early when the validation set loss does not improve for 20 consecutive epochs;

[0158] Learning rate scheduler: When the validation set loss does not improve for 10 consecutive epochs, the learning rate is multiplied by 0.5.

[0159] This invention implements the deep learning-based distributed radar resource allocation method by constructing a distributed tracking and fusion module, an evaluation module, a cognitive decision-making and resource allocation module, and a trajectory prediction module.

[0160] Distributed tracking and fusion module: For each radar and for the tracked target in the radar's field of view, an independent IMM filter is established. The IMM filter performs preliminary processing on the radar's local observations by running Kalman filters for different motion models in parallel, and generates the local trajectory of the tracked target.

[0161] The central controller receives the local trajectories of the tracked targets from each radar, fuses the local trajectories of the tracked targets, and obtains the global fused trajectory of the tracked targets; this module realizes the perception link in the main cognitive closed loop.

[0162] The following are the specific settings of the distributed tracking and fusion module in the application example:

[0163] The system consists of three radars, each radar node instantiating a RadarTracker class. Internally, this class maintains an independent IMM state for each target. The IMM filter runs three models in parallel: a uniform straight-line (CV) model and two uniform turning (CT) models with different turning rates (1° / s and -3° / s). The transition probability matrix between models is set to have a higher probability of state preservation. At each time step, the local trajectory outputs of all RadarTrackers are fed into a central fuse_tracks_ci function, which calculates the global fused trajectory using a covariance crossover algorithm.

[0164] Evaluation Module: Based on the global fusion trajectory of the tracked target and the trajectory prediction at the current moment, this module calculates the core evaluation indicators of the tracked target, including uncertainty score, threat score, and distance. This module implements the evaluation link in the main cognitive loop.

[0165] Cognitive Decision-Making and Resource Allocation Module: This module employs decoupled decision-making logic to transform the core evaluation indicators of the tracked target into resource allocation instructions for each radar, specifically including power and dwell time resource allocation. This module implements the decision-making and action phases in the main cognitive closed loop. The resource allocation instructions generated by this module for each radar are sent back to the distributed radar network to complete the closed loop.

[0166] The following are the specific settings of the cognitive decision-making and resource allocation module in the application example:

[0167] This module is implemented by the `CognitiveSystemController` class. Its core method, `allocate_resources`, receives a global fused trajectory dictionary for all targets. This method first calculates the uncertainty, threat level, and range score for each target. Then, it calculates the priority of power and loiter time. Finally, using a strategy with a minimum guarantee ratio of 0.02, it proportionally allocates the total power budget (50.0) and the total loiter time budget (1.0). The total power allocated to each target is further decomposed among the three radars based on the inverse fourth power of the range of each radar.

[0168] Trajectory Prediction Module: Taking the global fusion trajectory of the tracked target as input and the predicted trajectory of the tracked target at the next moment as output, this module constructs and trains a trajectory prediction model for the tracked target. This module is the forward-looking core of the cognitive loop. Its prediction results will serve as the core input of the evaluation module (step S3) and the cognitive decision-making and resource allocation module (step S4) at the next moment, realizing prediction-based resource allocation.

[0169] The following are the specific settings of the trajectory prediction module in the application example:

[0170] This module is implemented using a BiLSTMAttention class, which inherits from torch.nn.Module. Its internal structure includes a two-layer bidirectional LSTM layer with Dropout, a linear attention mechanism layer, and a fully connected layer for outputting the final 2D position prediction. The model's input dimensions are 6 (x / y position, velocity, acceleration), and the sequence length is 16. The model is trained offline on a large-scale trajectory dataset. The training objective is to learn the mapping from the IMM-filtered smoothed trajectory to the real noise-free trajectory. The loss function is MSE, and the optimizer is Adam. In the simulation, this module receives the latest 16 fused trajectory points, standardizes them, and feeds them into the model, outputting a prediction of the position at the next time step.

[0171] This method and system possess inherent robustness in their architecture. The main cognitive closed loop, based on "distributed IMM tracking + centralized CI fusion," forms the foundation for the system's stability and reliability. The IMM filter itself has the ability to rapidly adapt to target maneuver patterns, ensuring that the system can maintain effective tracking even when faced with novel and complex maneuver patterns not included in the training dataset.

[0172] The tightly coupled BiLSTM-based prediction module serves as a high-performance enhancement engine for the system. It leverages the powerful nonlinear modeling capabilities of deep learning to provide higher-precision trajectory predictions based on learned patterns. Ideally, its predictions can provide more forward-looking auxiliary information for resource allocation. However, in extreme cases, even if the module's prediction accuracy decreases due to encountering unknown maneuvering patterns, the IMM-based main cognitive loop can still revert to a reactive mode relying solely on S2 to complete the tracking and resource allocation tasks.

[0173] In summary, this method adopts a centralized decision-making logic of "first integration, then evaluation, and finally allocation," and the overall execution process is as follows: Figure 2 The process starts at "Start (Time k)," which represents the beginning of a processing cycle. First, the "Cognitive Controller" module starts, which calculates the optimal overall resource allocation scheme for the current time k based on the global fusion trajectory stored at time k-1 and the trajectory prediction generated at time k-1.

[0174] Next, the system proceeds to the step of "decomposing the total resources for each target based on the efficiency of each radar". In this step, the system further decomposes the total resource budget (total power, total time) generated in the previous step for each target into refined resource instructions allocated to each specific radar, based on the observation efficiency of each node in the radar network for the tracked target.

[0175] Next, the system enters a parallel processing loop that "traverses the radar network." In this loop, each radar in the network will perform the following series of operations:

[0176] (1) Receive radar resource scheduling instructions: Each radar node receives the specific power and dwell time parameters assigned to it;

[0177] (2) Simulate the generation of observations of each target: The radar node performs the detection task according to the received resource instructions and obtains the raw observation data of the target;

[0178] (3) Update the radar’s local IMM filter: Using the new observation data, each node updates its internally maintained IMM filter for each target;

[0179] (4) Output radar local trajectory: After completing the filter update, each node outputs its latest local estimate of the target state, i.e., the local trajectory.

[0180] After all radar nodes have completed the above operations, the loop ends. The local trajectories output by all nodes are aggregated and sent to the CI module. This module, acting as a central fusion unit, fuses information from multiple local trajectories from different nodes to generate a globally unified trajectory.

[0181] The fused result is used to "generate and store the global fused trajectory at time k". The result of this step is the optimal estimate of all target states at the current time k, which will be stored as the basis for the cognitive controller to make decisions in the next cycle (time k+1).

[0182] After generating and storing the global fusion trajectory at time t, the BiLSTM-Attention module extracts a historical sequence as input from the newly generated global trajectory and uses its deep learning model to predict the target's position at time k+1. The prediction result is stored and serves as the core input to the cognitive controller module at the start of the loop at time k+1.

[0183] Finally, the main process reaches "end (time k)," marking the completion of the current processing cycle, and the system then prepares to enter the next time step k+1 of the loop.

[0184] The following are the effect experiments and result analysis of the application examples:

[0185] Figure 3 , Figure 4 and Figure 5The figures demonstrate the system's root mean square error (RMSE) performance in uniform straight-line motion, single-turn, and complex S-shaped motion scenarios. As shown in the figures, the RMSE of the fused trajectory output by the core tracking framework of this invention (red line) is significantly lower than the fusion observation error (orange dashed line), which represents the input noise level, in all scenarios, and it effectively handles target maneuvers (such as...). Figure 4 , Figure 5 The peak error in the data demonstrates its strong filtering capability and robustness. The parallel LSTM prediction RMSE (blue line) performs better in most cases, especially during violent maneuvers, demonstrating the potential of data-driven models in capturing complex patterns.

[0186] Figure 6 The results visually demonstrate that in a distributed network containing three radars, the fused trajectory (dashed line) output by the system closely matches the actual trajectory (solid line), proving the superior tracking performance of this invention in multi-radar collaborative scenarios.

[0187] Figure 7 and Figure 8 Clearly verify the effectiveness of decoupled cognitive decision-making. Figure 7 The system shows that it allocated the most power resources to the farthest target 1 to counteract signal attenuation. Figure 8 The data shows that the system precisely allocated the most dwell time resources to targets 2 and 3 that were maneuvering, and the peak allocation was consistent with... Figure 4 and Figure 5 The peak times of RMSE in the data are highly consistent.

[0188] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A distributed radar resource allocation method based on deep learning, characterized in that, For the central controller and the distributed radar network consisting of multiple radars, the following steps S1-S5 are executed to complete the resource allocation of each radar in the multi-target tracking task and the trajectory prediction of the tracked target: Step S1: For each radar and for the tracked target in the radar's field of view, establish an independent IMM filter. The IMM filter performs preliminary processing on the radar's local observations by running Kalman filters for different motion models in parallel, generating the local trajectory of the tracked target. The specific steps of step S1 are as follows: Step S1.1: Assuming the tracked target may switch from any old motion model to the current new motion model, the states of all old motion models are weighted and fused, as shown in the following formula: ； In the formula, The mixed probability represents the probability that, given that the tracked target will be in the j-th motion model at time k, it was previously in the ith motion model at time k-1. Let be the posterior probability of the i-th motion model at time k-1, representing the confidence that the tracked target is in the i-th motion model after the previous update; The elements in the motion model transition probability matrix Π represent the prior probability of the tracked target transitioning from the i-th motion model to the j-th motion model. This is the normalization constant; The sum of the probabilities of all possible paths to the j-th motion model is used as the normalization constant. Its calculation is as follows: ； In the formula, N is the total number of motion models in the IMM filter; Step S1.2: Calculate the initial state of the mixture for each Kalman filter, including the weighted average filter state and the mixture covariance, as the input for prediction of each Kalman filter, as shown in the following formula: ； ； In the formula, This indicates the filter state after weighted averaging. Indicates mixed covariance. Let represent the posterior state estimate for the i-th motion model at time k-1. Let represent the posterior covariance of the i-th motion model at time k-1; Step S1.3: Based on local observations The likelihood function is constructed as follows: ； This represents the likelihood value of the j-th motion model; Represents the Gaussian probability density function; This represents the actual radar observation value at time k; This represents the prediction of the j-th motion model for the observation at time k; Represents the covariance of the j-th motion model; Update the posterior probabilities of the motion model: ； In the formula, Let represent the posterior probability of the j-th motion model at time k; c represents the constant used for probability summation and normalization. Step S1.4: Weight and fuse the results of all parallel Kalman filters to output the radar's local trajectory estimate of the tracked target at time k, as shown in the following formula: ； ； In the formula, This represents the final state estimate; This represents the state estimate after the j-th motion model is updated at time k; This represents the final covariance; This represents the covariance of the j-th motion model update at time k; Step S2: The central controller receives the local trajectories of the tracked target sent by each radar, and fuses the local trajectories of the tracked target to obtain the global fused trajectory of the tracked target. The specific method for step S2 is as follows: At the central controller, for the i-th tracked target, at time k, the CI algorithm is used to fuse the local trajectories from M radars ( , ), where j=1,2,…M, to obtain the global fusion trajectory ( , ); Step S3: Based on the global fusion trajectory of the tracked target and the trajectory prediction at the current moment, calculate the core evaluation indicators of the tracked target, including uncertainty score, threat score, and distance. In step S3, the central controller, based on the global fusion trajectory obtained in step S2, , For each tracked target, three core evaluation metrics are calculated, and the specific steps are as follows: Step S3.1: Calculate the uncertainty fraction as follows: ； In the formula, To integrate the posterior covariance matrix The trace of the mid-position component; This represents the position variance of the i-th tracked target along the X-axis after fusion; This represents the position variance of the i-th tracked target in the Y-axis direction after fusion; Step S3.2: Based on the predicted position at time k generated in step S5 at time k-1, the position at time k is... And the global fusion state obtained at time k-1 based on step S2. Calculate the predicted instantaneous velocity and instantaneous acceleration As shown in the following formula: ； ； In the formula, , These represent the position and instantaneous velocity of the tracked target at time k-1, respectively. Indicates time difference; The threat score is calculated as follows: ； In the formula, This represents the threat score of the i-th tracked target. and These represent the magnitudes of the velocity and acceleration vectors of the tracked target at the current moment, respectively. and This indicates the preset velocity and acceleration weighting coefficients; Step S3.3: Based on the predicted position at time k generated at time k-1 in step S5, calculate the distance score as follows: ； In the formula, Let be the distance score of the i-th tracked target; Let be the predicted position of the i-th tracked target at time k; Let represent the location of the j-th radar in the distributed radar network, and M represent the number of radars. The distance between a single radar and the i-th tracked target; Step S4: Using decoupled decision-making logic, the core evaluation indicators of the tracked target are transformed into resource allocation instructions for each radar. Step S5: Using the global fusion trajectory of the tracked target obtained in step S2 as input and the trajectory prediction of the tracked target at the next moment as output, construct and train the trajectory prediction model of the tracked target, obtain the trained trajectory prediction model, apply the trained trajectory prediction model to provide forward-looking decisions for steps S3 and S4 at the next moment, and complete the trajectory prediction of the tracked target.

2. The deep learning-based distributed radar resource allocation method according to claim 1, characterized in that, The IMM filter in step S1 is achieved by running three Kalman filters for different motion models in parallel.

3. The deep learning-based distributed radar resource allocation method according to claim 1, characterized in that, The specific steps of step S4 are as follows: Step S4.1: Calculate the power allocation priority score of the tracked target as follows: ； In the formula, This represents the power allocation priority score of the i-th tracked target; This represents the normalized distance score; This represents the normalized uncertainty score; , These represent the corresponding weights for distance and uncertainty, respectively. Step S4.2: Calculate the time allocation priority score of the tracked target as follows: ； In the formula, This represents the time allocation priority score for the i-th tracked target; This represents the normalized threat score; This represents the normalized uncertainty score; , These represent the corresponding weights of the threat score and the uncertainty score, respectively. Step S4.3: Calculate the proportional allocation parameters with minimum guarantee as follows: ； ； In the formula, This represents the proportion of resources ultimately obtained by the i-th tracked target; The minimum guaranteed resource ratio set for each tracked target; This represents the total number of tracked targets in the scene. The remaining dynamically allocable resource proportion after removing the minimum guaranteed resource proportion; The total priority score of the i-th tracked target; Step S4.4: Perform a two-stage update of resource allocation for each radar: Phase One: For All Each of the tracked targets is assigned a fixed minimum guaranteed resource ratio. ; Phase Two: Allocate the remaining proportion of dynamically describable resources According to the total priority score of each tracked target The proportion of the total score is dynamically allocated.

4. The deep learning-based distributed radar resource allocation method according to claim 1, characterized in that, The specific steps of step S5 are as follows: Step S5.1: The global fusion trajectory of the tracked target obtained in step S2 is a historical position sequence, in the form of a 6-dimensional feature vector containing position, velocity, and acceleration. As shown in the following formula: ； In the formula, This represents the 6-dimensional eigenvector at time k; Indicates the position at time k; Represents the velocity at time k; Represents the acceleration at time k; 6-dimensional feature vectors Scale to a uniform numerical range; Step S5.2: Based on a bidirectional long short-term memory network with an attention mechanism, construct a trajectory prediction model for the tracked target. Use the 6-dimensional feature vector at time k obtained in step S5.1 as input and the position of the tracked target at time k+1 as output to train the trajectory prediction model for the tracked target.

Images