Radar automatic target identification method and system based on PBC algorithm and AGB algorithm

Abstract

The invention discloses a radar automatic target identification method and system based on a PBC algorithm and an AGB algorithm. The method comprises the following steps: using a plurality of radars to detect a target at the same time; each radar independently acquires and preprocesses RCS data of a target; using an AGB algorithm to generate a local posterior probability vector used for representing optimization judgment of the radar on the target type; the AGB algorithm comprises the following steps: introducing a single decision tree regression device with multiple vector size outputs, wherein only one decision tree is trained in each iteration; fitting a pseudo residual error by using a least square method, and obtaining an approximate solution by using a Newton-Rafson step; fusing the plurality of local posterior probability vectors by using a Bayesian formula, and calculating and generating a global posterior probability vector used for representing final judgment of the target type; and recursively updating the global posterior probability vector by using a PBC algorithm. According to the method, the accuracy of unmanned aerial vehicle identification can be remarkably improved, the calculation complexity is reduced, the real-time performance is improved, meanwhile, the anti-interference capability is enhanced through a multi-radar cooperation mechanism, and the application requirements in a complex environment are met.

Description

Technical Field

[0001] The present invention relates to the technical field of radar automatic target recognition, and in particular to a method and system for unmanned aerial vehicle recognition based on progressive Bayesian classification and agglomerative gradient boosting algorithm. Background Art

[0002] With the wide application of unmanned aerial vehicles, the resulting safety and privacy issues have become increasingly prominent, and there is an urgent need for efficient and accurate recognition methods. Radar Automatic Target Recognition (RATR) technology has become an important solution for unmanned aerial vehicle recognition due to its all-weather operation ability and cost-effectiveness. However, there are still many problems and deficiencies in the existing RATR technology, making it difficult to meet the requirements of practical applications.

[0003] The defects of the existing technology are as follows: First, the recognition accuracy of traditional single-radar systems in complex noise environments is relatively low, making it difficult to effectively distinguish target signals from interference signals. For example, the RATR method based on support vector machines performs well on small-scale datasets, but its recognition rate drops significantly in high-noise environments. Second, for example, existing methods usually adopt complex algorithms (such as deep learning models), resulting in high consumption of computing resources and difficulty in meeting the requirements of real-time applications. The RATR method based on convolutional neural networks has high recognition accuracy, but its training and inference processes require a large amount of computing resources and time, making it difficult to be deployed on resource-constrained edge devices. Third, existing research mainly focuses on single-radar systems, lacking a multi-radar collaborative recognition mechanism, resulting in limited anti-interference ability and recognition accuracy in complex environments. The RATR framework based on a single radar performs well in small-scale scenarios, but in large-scale and multi-target scenarios, its recognition accuracy and real-time performance drop significantly. Finally, existing technologies often struggle to meet real-time requirements when dealing with large-scale data.

[0004] Therefore, the existing technology has obvious deficiencies in dealing with noise environments, reducing computational complexity, achieving multi-radar collaborative recognition, and meeting real-time requirements, and there is an urgent need for a more efficient and accurate solution. Summary of the Invention

[0005] In view of the above-mentioned defects of the existing technology, the present invention provides a method and system for radar automatic target recognition based on progressive Bayesian classification (PBC) and agglomerative gradient boosting (AGB) algorithm, which can significantly improve the accuracy of unmanned aerial vehicle recognition, reduce computational complexity, and enhance real-time performance. At the same time, the anti-interference ability is enhanced through a multi-radar collaborative mechanism to meet the application requirements in complex environments.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] In a first aspect, a radar automatic target recognition method based on PBC algorithm and AGB algorithm includes the following steps:

[0008] S1. Detect the target using multiple radars simultaneously; each of the radars independently acquires and preprocesses the RCS data of the target;

[0009] S2. Use the AGB algorithm to generate a local posterior probability vector according to the RCS data; the local posterior probability vector is used to represent the optimized judgment of the radar on the target type;

[0010] The AGB algorithm is an improved algorithm based on the Gradient Boosting Trees (GBDT) algorithm, including: introducing a single decision tree regressor with multiple vector size outputs, and only training one decision tree in each iteration; the AGB algorithm first fits the pseudo-residuals by the least squares method, and then obtains an approximate solution using the Newton-Raphson step;

[0011] S3. Use the Bayesian formula to fuse multiple local posterior probability vectors, calculate and generate a global posterior probability vector; the global posterior probability vector is used to represent the final judgment of the target type; use the PBC algorithm to recursively update the global posterior probability vector.

[0012] Preferably, it further includes step S4. Output the classification result of the target type according to the global posterior probability vector; compare the classification result with the actual target type to verify the accuracy of the classification result.

[0013] Preferably, the preprocessing in step S1 includes performing bilinear interpolation on the RCS data corresponding to the azimuth and elevation angles that have not been collected according to the collected azimuth and elevation angles to generate a complete RCS data set; after generating the complete RCS data set, add Gaussian noise to the RCS data.

[0014] In a second aspect, a radar automatic target recognition system based on PBC algorithm and AGB algorithm includes a radar detection module and a PBC fusion and recursion module; the radar detection module includes an AGB optimization unit and multiple independent radars; the radar detection module is used to acquire and process the RCS data of the target;

[0015] The radar generates the local posterior probability vector through the AGB optimization unit according to the RCS data, and inputs the local posterior probability vector into the PBC fusion and recursion module; the PBC fusion and recursion module generates and recursively updates the global posterior probability vector through the PBC algorithm.

[0016] Compared with the prior art, the beneficial effects of the present invention are reflected in:

[0017] 1. By combining the PBC algorithm and the AGB algorithm, the present invention proposes a RATR framework based on the PBC and AGB algorithms. The classification model is gradually updated by PBC to adapt to dynamic data, improving the robustness and accuracy of RATR. At the same time, AGB is used to reduce the model complexity and enhance the prediction speed. This application can achieve high-precision target recognition in complex noise environments. The PBC algorithm dynamically adapts to changes in target features by gradually updating the prior probability and posterior probability, significantly improving the classification accuracy; the AGB algorithm further optimizes the classification result by integrating multiple decision tree regressors with vector-valued leaves and weighting their outputs, making the overall recognition speed significantly superior to traditional methods.

[0018] 2. Through efficient integration and dynamic weight allocation, the AGB algorithm significantly reduces the computational complexity while ensuring the accuracy. Compared with traditional machine learning (ML) models, this application consumes less computing resources, is suitable for deployment on resource-constrained edge devices, and reduces the hardware cost and energy consumption at the same time.

[0019] 3. The progressive update mechanism of the PBC algorithm and the efficient computing ability of the AGB algorithm enable this application to meet the requirements of real-time applications. In high-dynamic scenarios, this application can quickly process large-scale data and output recognition results, significantly improving the real-time performance of the system.

[0020] 4. Through the multi-radar collaborative recognition mechanism, this application can enhance the anti-interference ability of target recognition. The collaborative work of the multi-radar system not only improves the accuracy of target positioning, but also effectively reduces the influence of noise and interference signals through data fusion, showing higher robustness in complex environments.

[0021] 5. This application is not only applicable to drone recognition, but can also be extended to other radar automatic target recognition scenarios, such as vehicle recognition, ship recognition, aircraft recognition, etc. Its flexibility and versatility make it have broad application prospects in military, civilian and commercial fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a schematic diagram of the multi-radar collaborative target recognition framework in Embodiment 1;

[0023] Figure 2 is a schematic diagram showing the influence of the number of radar base stations on the drone target classification performance in Embodiment 1;

[0024] Figure 3Schematic diagram of the comparison of the decision boundary performance between the GBDT algorithm and the AGB algorithm in Embodiment 1;

[0025] Figure 4 Schematic diagram of the classification accuracy trend with the increase in the number of radar base stations in Embodiment 2;

[0026] Figure 5 Comparison chart of the prediction time between the present invention and the prior art in processing RCS sequences of the same scale;

[0027] Figure 6 Schematic diagram of the anti-noise performance of the present invention under different SNR conditions;

[0028] Figure 7 Trend chart of the target recognition accuracy of the present invention with the time step. Detailed implementation manners

[0029] In order to make the technical means, creative features, achieved purposes and effects of the invention easy to understand, the present invention will be further described below in conjunction with specific drawings. However, the present invention is not limited to the following implemented cases.

[0030] It should be noted that the structures, ratios, sizes, etc. shown in the drawings of this specification are only used to cooperate with the content disclosed in the specification for those who are familiar with this technology to understand and read, and are not used to limit the limited conditions under which the present invention can be implemented. Therefore, they do not have a technical essence. Any modification of the structure, change of the proportional relationship or adjustment of the size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope covered by the technical content disclosed in the present invention.

[0031] Embodiment 1:

[0032] As Figure 1 shown, a radar automatic target recognition method based on the PBC algorithm and the AGB algorithm includes the following steps:

[0033] S1. Detect the target using multiple radars simultaneously; each radar independently acquires and preprocesses the radar cross section (RCS) data of the target; generate a probability distribution vector according to the RCS data; the probability distribution vector is used to represent the preliminary judgment of the radar on the target type.

[0034] Detect the target UAV in an anechoic chamber environment, obtain its RCS data, with a frequency range of 26 - 40 GHz, a resolution of 1 GHz, azimuth coverage of 0° - 180°, elevation coverage of -95° - 95°, and a sampling interval of 1°. Preprocess the collected RCS data, including bilinear interpolation for azimuth and elevation angles that were not collected to generate a complete RCS data set. Meanwhile, to simulate the noise in the real environment, independent Gaussian noise is added to the RCS data to ensure that the data is closer to the actual scenario. Each radar independently calculates the RCS features of the target based on the RCS data at time step t, and generates a probability distribution vector based on these RCS features, representing the preliminary judgment of the radar on the target type.

[0035] Assume that there is a radar target to be classified in the observation space of the radar, and J radars simultaneously detect the target. At each time point t, multiple radars simultaneously transmit pulsed electromagnetic waves towards the radar target, and each radar independently calculates the RCS features of the same radar target under the pulsed electromagnetic wave of this radar. Figure 2 Shows the influence of the number of radar base stations on the classification performance of UAV targets, and verifies the significant advantages of the multi-static radar configuration in a low SNR (SNR = 0 dB) environment. The experimental results show that: as the number of radar base stations increases (4, 9, 16, and 25), the geometric diversity of multi-angle observations significantly improves the discriminative ability of RCS features, and the classification accuracy is improved by 28.646%, 39.961%, 42.355%, and 43.085% respectively compared with the single-radar system.

[0036] S2. Use the Aggregated Gradient Boosting (AGB) algorithm to generate a probability distribution vector, that is, a local posterior probability vector; the local posterior probability vector is used to represent the optimized judgment of the radar on the target type. The AGB algorithm, through a single-model multi-output mechanism and a two-step optimization process, uses a unified vector output decision tree regressor to simultaneously optimize the class prediction vectors (probability distribution vectors) of all radars to ensure the accuracy and efficiency of classification prediction.

[0037] Each radar independently inputs its respective observations into the local discriminant ML model to output multiple independent probability distribution vectors of the radar target type, that is, local posterior probability vectors.

[0038] Each radar node requires an independent machine learning model to generate the local posterior probability vector of the target type. Traditional methods usually use single-output classifiers for multi-class recognition, with their computational complexity being too high, resulting in overly long prediction times and unable to meet the real-time requirements of radar target prediction. To break through this limitation, this application proposes the AGB algorithm as the core supporting component of PBC and achieves efficient collaboration. The AGB algorithm will be described in detail next.

[0039] Improved ML algorithm - AGB algorithm

[0040] For the case of multi-class classification, since the base decision tree regressor provides scalar values, the gradient boosting ensemble needs to build a tree for each class and iteration. In this subsection, the extensions required for GBDT are derived to enable multi-class classification using a single decision tree regressor with vector-sized outputs. One of these trees can replace the single-output tree used in the standard GBDT. Also consider the dataset as Select multi-class labels of K class labels k ∈ [1, K] from the D distribution, and use a unique encoding method to convert the multi-class labels into a logical matrix. In this matrix, each row i represents a y {i,k} vector, where if the i-th instance belongs to class k, then y {i,k} equals 1, otherwise 0. The goal is to find K functions to minimize the given loss function in an additive manner

[0041]

[0042] where is a multi-output model with K outputs, and ν is the learning rate hyperparameter. Using the uniquely encoded class labels, in each iteration, a multi-output tree is trained to fit the data, where the values to be learned are a matrix of size N×K that contains the residuals of the previous predictions

[0043]

[0044] For a node j with output vectors of K decision tree regressors, the criterion for minimizing the upcoming split is the average of the mean squared errors of the K outputs, i.e.:

[0045]

[0046] where, is the mean of target k in region j, and N j is the number of samples at node j.

[0047] As mentioned before, when the loss function is different from the loss function when fitting the model to the residual matrix, a second optimization step needs to be performed. In this second step, the K-dimensional vector output by the leaf node needs to be updated to

[0048]

[0049] For multi-class classification, if using logarithmic loss

[0050]

[0051] The Newton-Raphson of the tree top leaves will be updated to:

[0052]

[0053] Using the first-order derivative of the loss function near a given starting point to generate a series of improved approximate values of the root of the objective function. Different from the standard model, in the proposed algorithm, all multiple regression outputs are learned and predicted using one tree and one ensemble per iteration.

[0054] The AGB algorithm model is a two-step optimization algorithm. It follows the original GBDT method, applies gradient descent and Newton-Raphson step approximation to find the optimal solution, and introduces a multi-output decision tree regressor as the additive model in gradient boosting. This method makes the prediction of the model simpler and faster. In subsequent experiments, it is proved that the proposed method enhances the generalization ability, prediction speed, and reduces the model complexity. Figure 3 The decision boundary performance of the traditional GBDT algorithm and the AGB algorithm was compared through synthetic data (a four-class classification problem based on the Madelon dataset, containing 2000 samples and a two-dimensional feature space). The experiment adopted the same hyperparameter configuration (learning rate = 0.1, subsampling rate = 0.75, number of iterations = 100, maximum tree depth = 5), and the relationship between the classification boundary and the training sample distribution of the two algorithms in the feature space was visually presented through the visualization method of color overlay.

[0055] In this embodiment, the AGB algorithm runs independently at each radar node, and generates a local posterior probability vector of the target type based on the RCS feature Its multi-output decision tree structure can efficiently handle multi-class recognition problems and provide a high-precision initial probability distribution for the fusion stage.

[0056] S3. Use Bayes' formula to fuse multiple local posterior probability vectors, calculate and generate a global posterior probability vector; the global posterior probability vector is used to represent the final judgment of the target type; use the PBC algorithm to recursively update the global posterior probability vector. The specific method is as follows:

[0057] Horizontal fusion: Within a single time step, use the optimal fusion rule to integrate the local posterior probability vectors output by the AGB of J radars to form a global posterior probability vector.

[0058] The joint posterior distribution, that is, the global posterior probability vector P(A|Z(t) The optimal fusion rule of ) is:

[0059]

[0060] The optimal fusion rule formula uses Bayesian classification to probabilistically fuse the observation results of multiple radars. Its core idea is to use the independent observation data of each radar to update the posterior probability distribution of the target type. In this formula, A represents the true type of the target, Z( t ) represents the set of observation data from all J radars at time step t. P(A) is the prior probability of the target type, reflecting the initial perception. J is the total number of radars participating in the coordination, whose influence is reflected by the (1-J) exponential term—as the number of radars increases, the influence of the prior probability gradually decreases. Indicates that the jth radar based on its observation The output local posterior probability vector is generated by machine learning models such as AGB. The numerator is obtained by continuous multiplication. The independent observational evidence from all radars is integrated, and the denominator is normalized to ensure that the total probability is 1.

[0061] The fusion method combines the classification probability vectors of radar target types of each radar in a probabilistic way, where the key assumption is that for a given radar target type, the observations of multiple single radars within time step t are conditionally independent, that is, there exists a formula By dynamically balancing the weights of prior knowledge and observational evidence, the fused global posterior probability P(A|Z( t ), thus achieving highly robust target recognition under multi-radar collaboration. As the amount of observation data increases, the classification results will quickly converge to the true target type.

[0062] Vertical iteration: When crossing time steps, the PBC algorithm uses the global posterior probability vector of the current fusion result as the new prior, and continuously updates the global posterior probability vector in combination with subsequent observations to achieve progressive accuracy improvement.

[0063] As the radar's dwell time on a single target increases, the uncertainty and accuracy of the radar target type decrease and increase, respectively, due to the increase in the number of data observations. Furthermore, the Bayesian method exhibits excellent robustness when dealing with noisy data, making it highly advantageous in radar-related applications. This is especially true in common and complex operating conditions such as high-clutter environments, multiple scattering points, and multipath propagation. The Bayesian method can effectively reduce noise interference, ensuring stable and accurate radar system operation, and providing reliable support for critical tasks such as target detection and recognition.

[0064] The PBC posterior probability of the radar target type is derived as follows:

[0065]

[0066] Among them, the marginal distribution P(A)=∫P(A|Z (t) )(Z (t) )dZ (t) is approximately the prior distribution 1 / K (where K is the number of classification categories). The PBC based on the optimal fusion strategy constructs a principled framework with a solid theoretical basis for the classification of RCS time series data. Under this framework, by means of the progressive Bayesian classification method, accurate classification of radar targets is achieved and the robustness under low Signal-to-Noise Ratio (SNR) is improved, providing strong support for decision-making in the RATR field.

[0067] Using the PBC method, the global posterior probability vector is recursively updated. As the radar continuously tracks the target, the amount of observed data increases, the uncertainty of target type classification decreases, and the classification accuracy gradually improves. The PBC method gradually approaches the true type of the target through continuous iterative updates.

[0068] S4. Output the classification result of the target type according to the global posterior probability vector; compare the classification result with the actual target type to verify the accuracy of the classification result. Evaluate the performance of the radar automatic target recognition framework under different SNRs and radar configurations through experiments to ensure its robustness and stability in complex environments.

[0069] S5. Performance optimization and iteration: According to the verification results, optimize the performance of the radar automatic target recognition framework, such as adjusting algorithm parameters, optimizing data preprocessing steps, or improving the multi-radar fusion strategy. Further improve the accuracy and efficiency of target recognition through continuous iterative optimization.

[0070] The efficient classification of AGB provides reliable probability input for PBC, while the multi-radar / multi-temporal fusion mechanism of PBC makes up for the limitations of single-radar AGB in a noisy environment. The combination of the two ensures both real-time performance (fast inference of AGB) and robustness (noise-resistant optimization of PBC). The AGB algorithm and the PBC algorithm work together through a multi-level probability fusion mechanism in the technical solution. Figure 5 Shows the comparison of the prediction times of each machine learning model under different radar station configurations. The experimental results show that the unified RATR framework proposed in this application has a significantly lower prediction time when processing RCS sequences of the same scale compared to other algorithms. Specifically, as the number of radar stations increases, the prediction time of traditional algorithms shows a linear growth trend, while the unified RATR framework benefits from its optimized design of progressive Bayesian classification and aggregated gradient boosting and maintains relatively stable low-latency characteristics. Figure 6The experimental results show that the unified RATR framework proposed in this application demonstrates significant advantages under different SNR conditions. Compared with the single-radar configuration, this framework can still maintain an accuracy improvement of more than 30% even in a low SNR environment, fully demonstrating its excellent anti-noise performance. Figure 7 It shows the variation trend of target recognition accuracy with time steps under different radar station configurations and SNR conditions, indicating that increasing the radar deployment density within the monitoring area can achieve two key advantages simultaneously: one is to significantly improve the classification accuracy, and the other is to effectively reduce the average classification delay. Specifically: 1) At the same time node, the recognition accuracy of the multi-radar station configuration is always better than that of the single-radar station system; 2) In a low SNR working environment, the performance improvement brought by multi-radar collaborative observation is particularly significant.

[0071] Example 2:

[0072] Suppose there are 9 randomly distributed radar stations in a space, and each radar station is independent and has no influence on each other. And each radar is equipped with an AGB algorithm model trained with 5 types of radar target RCS data, that is, each radar can independently observe the radar target and predict the type of the radar target accordingly.

[0073] Within the common observation area of the 9 radar stations, there is an unknown flying target. The 9 radar stations simultaneously observe the target, independently obtain the RCS time series of the radar target respectively, and the time resolution Δt = 0.1 second, that is, the RCS sequence of the radar target is obtained every 0.1 second. Denote the RCS at time t (t is an integer multiple of Δt) as Collectively denoted as Z( t ), and since each radar is equipped with an AGB algorithm model, they will each process based on the obtained RCS data.

[0074] The AGB algorithm model in each radar station will first take the RCS sequence data received by this radar station as input. Taking radar station 1 as an example, the data obtained at time t will be input into the corresponding AGB model. Inside the AGB model, it is processed according to its unique two-step optimization algorithm. First step, the base learner is fitted to the pseudo-residual by the least squares method. In this process, the model will extract and analyze the features in the data. In the AGB algorithm, for multi-class classification tasks, the data set Process it, and convert multiple types of labels into logical matrices using one-hot encoding. In the feature extraction stage, the decision tree will sort the values of each attribute, and after sorting, traverse these values to calculate the gain of each possible split point, and repeat the above operation for all P attributes to extract feature information from the data. In terms of finding potential rules and patterns in the data, the decision tree in the AGB algorithm starts from the root node and recursively partitions the data according to the criterion of minimizing the mean squared error (MSE), and calculates the to determine the splitting method. In the second step, use the Newton-Raphson step approximation solution according to the original GBDT strategy to further optimize the parameters of the model, so that the model can more accurately classify and predict the target. In the multi-class classification task, the terminal leaves of the multi-output tree update the output vector through Newton-Raphson, and the formula is:

[0075]

[0076] Generate a series of improved approximations near the given starting point using the first derivative of the loss function, and gradually approach the root of the objective function, so that the model can fit the data more accurately. After these two optimizations, the AGB model of Radar Station 1 outputs a classification probability vector of a radar target type. For example, the probabilities of predicting that the target belongs to 5 radar target types are P 11 、P 12 、P 13 、P 14 and P 15 , that is

[0077]

[0078] Similarly, the other 8 radar stations will also process the RCS data they obtain in the same way. Radar Stations 2 to 9 are in turn, and respectively output their respective prediction probability vectors …、

[0079]

[0080] Next, it is necessary to fuse the classification probability vectors output by these 9 radar stations. According to the optimal fusion rule, the fused probability of the radar target at time t can be obtained as Since the radar target appears randomly, so P(A) = [1 / 5, 1 / 5, 1 / 5, 1 / 5, 1 / 5] = [0.2, 0.2, 0.2, 0.2, 0.2].

[0081] As time t increases continuously, the fused probability of the radar target at different times t, P(A|Z( t)) is also constantly changing. Therefore, a fusion method based on PBC is adopted to further improve the radar target recognition accuracy. This method combines the classification probability vectors of the radar target types of each radar in a probabilistic manner. According to the optimal fusion rule of the joint posterior distribution, it is assumed that for a given radar target type, the observations of multiple single radars within the time step are conditionally independent. Based on this, the fused probability distribution is updated according to Bayes' principle, that is, by using the posterior classification probability distribution of the radar target type at time node t and all previous times, combined with the classification probability distribution of the radar target type fused from 9 radar stations at time t for updating.

[0082] That is, it is assumed that before time t, there is already a certain posterior probability distribution P(A|Z (1:t-1) ). After fusing the radar target fusion probability P(A|Z (t) ) of 9 radar stations at time t, the new posterior probability distribution will be calculated according to the formula of PBC, that is

[0083]

[0084] In this calculation process, the observation information of each radar station at each time point is fully utilized, and the good robustness of the Bayesian method in dealing with noisy data is considered, effectively reducing the interference of environmental noise and clutter on the target RCS data, making the classification decision more accurate.

[0085] After fusion and update, the final classification result of the radar target type is obtained. Suppose the fused probability distribution shows that the probability that the target belongs to the 3rd radar target type is the highest, then it is determined that the unknown flying target is of the 3rd radar target type.

[0086] Embodiment 3: A radar automatic target recognition system based on the PBC algorithm and the AGB algorithm, which is used to execute the radar automatic target recognition method based on the PBC algorithm and the AGB algorithm; the radar automatic target recognition system based on the PBC algorithm and the AGB algorithm includes a radar detection module and a PBC fusion recursion module; the radar detection module includes an AGB optimization unit; the radar detection module also includes a plurality of independent radars, and each radar is equipped with an AGB optimization unit; the radar detection module is used to obtain the RCS data of the target; the radar uses the AGB algorithm in the AGB optimization unit to generate a probability distribution vector according to the RCS data, that is, a local posterior probability vector, and inputs the local posterior probability vector into the PBC fusion recursion module; the PBC fusion recursion module is used to generate and recursively update the global posterior probability vector.

[0087] Figure 4The experimental results show that as the number of radar base stations increases, the classification accuracy first increases rapidly and then gradually saturates. When the number of base stations increases from 1 to 4, the accuracy increases significantly by 28.6%; when it continues to increase to 9, the increase is 39.96%, and then the increase rate slows down significantly. This indicates that 9 base stations can fully capture the target features, and the benefit of continuing to increase the number of base stations is limited. Considering performance and cost comprehensively, the configuration of 9 base stations shows the best cost performance in the experiment.

Claims

1. A radar automatic target recognition method based on the PBC algorithm and the AGB algorithm, characterized in that It includes the following steps: S1. Detect the target using multiple radars simultaneously; each of the radars independently acquires and preprocesses the RCS data of the target; S2. Use the AGB algorithm to generate a local posterior probability vector based on the RCS data; the local posterior probability vector is used to represent the optimized judgment of the radar on the target type; The AGB algorithm is an improved algorithm based on the gradient boosting tree algorithm, including: introducing a single decision tree regressor with multiple vector size outputs, and only training one decision tree in each iteration; the AGB algorithm first fits the pseudo-residuals by the least squares method, and then obtains an approximate solution using the Newton-Raphson step; S3. Use the Bayesian formula to fuse multiple local posterior probability vectors, and calculate and generate a global posterior probability vector; the global posterior probability vector is used to represent the final judgment of the target type; use the PBC algorithm to recursively update the global posterior probability vector.

2. The radar automatic target recognition method based on the PBC algorithm and the AGB algorithm according to claim 1, characterized in that, It further includes step S4. Output the classification result of the target type according to the global posterior probability vector; compare the classification result with the actual target type to verify the accuracy of the classification result.

3. The radar automatic target recognition method based on the PBC algorithm and the AGB algorithm according to claim 2, characterized in that, The preprocessing in step S1 includes performing bilinear interpolation on the RCS data corresponding to the azimuth and elevation angles that have not been collected according to the collected azimuth and elevation angles to generate a complete RCS data set; after generating the complete RCS data set, Gaussian noise is added to the RCS data.

4. A radar automatic target recognition system based on the PBC algorithm and the AGB algorithm, which is used to execute the radar automatic target recognition method based on the PBC algorithm and the AGB algorithm as claimed in claim 1, characterized in that it includes a radar detection module and a PBC fusion recursive module; the radar detection module includes an AGB optimization unit and multiple independent radars; the radar detection module is used to acquire and process the RCS data of the target; the radar generates the local posterior probability vector through the AGB optimization unit according to the RCS data, and inputs the local posterior probability vector into the PBC fusion recursive module; the PBC fusion recursive module generates and recursively updates the global posterior probability vector through the PBC algorithm.

Images