A multi-dimensional feature false alarm elimination method based on ensemble learning

Abstract

This invention relates to a multi-dimensional feature false alarm rejection method based on ensemble learning, comprising: a) processing downlink AD data from a radar to form an intensity map; b) performing low-threshold constant false alarm detection (CFAD) based on the processed intensity map; c) performing intra-frame aggregation on the CFAD detection results; d) deblurring the aggregated suspected target points and extracting multi-dimensional fuzzy features; e) using an ensemble learning false alarm rejection model to predict the input multi-dimensional fuzzy features, discarding suspected targets predicted as false alarms, thus achieving false alarm rejection; f) performing inter-frame aggregation on the targets after false alarm rejection to obtain the final point trace result. This invention provides a method for constructing a multi-dimensional feature space for targets, and based on the constructed feature space, employs a data-driven approach and an ensemble learning method to further discriminate suspected targets pre-detected at low thresholds, thereby improving the detection capability of the radar system.

Description

Technical Field

[0001] This invention relates to the fields of radar signal processing and machine learning, and in particular to a multidimensional feature false alarm removal method based on ensemble learning. Background Technology

[0002] Radar is widely used in traditional military fields such as detection, early warning, guidance, remote sensing, and navigation. It also has broad application prospects in civilian fields such as weather forecasting, resource exploration, airport monitoring, coastal security, driver assistance, human detection, landslide monitoring, and long-range search and rescue. The continuous development of intelligent applications such as autonomous driving and smart meteorology requires radar to possess high-precision and high-reliability detection capabilities in complex environments. More importantly, in future modern warfare, the uncertain and highly dynamic combat environment, the rapidly increasing complexity of missions, and the rapid development of target mobility and stealth capabilities place even higher demands on radar detection and perception performance.

[0003] Currently, the common method for target detection in cluttered environments is the energy detector, which utilizes the first-order (amplitude) or second-order (power, power spectrum) statistical characteristics of accumulated echo data for target detection. However, in complex environments with diverse, highly variable, high-intensity, and non-stationary clutter, coupled with unintentional interference, clutter and targets are difficult to distinguish based on energy dimensions such as amplitude and power. Relying solely on energy-based threshold detection severely impacts the performance of current radar target detection, specifically manifesting as poor detection capabilities for weak and small targets, numerous false alarms, significantly affected target tracking quality, and the generation of numerous short and false tracks.

[0004] Unlike traditional knowledge-driven target detection methods based on statistical signal models, data-driven machine learning methods offer a completely new technological path for radar information processing. Fields such as computer vision, natural language processing, and content recommendation have all achieved rapid progress thanks to the widespread application of machine learning technology. Applying machine learning theories and methods to radar target detection, combined with knowledge from traditional radar signal processing models, and further developing machine learning methods and technologies suitable for this field, holds promise for achieving more accurate differentiation between targets and false alarms, resulting in better false alarm rejection and weak target detection compared to traditional methods.

[0005] Currently, some studies have used machine learning methods such as convolutional neural networks (CNN) [1-3], support vector machines (SVM) [4,5], and clustering [6] to distinguish between targets and clutter based on time-frequency feature maps or fractal features, achieving more accurate target detection than traditional threshold detection methods. However, in general wide-area search scenarios, due to the limitation of dwell time, high-resolution time-frequency feature maps are often difficult to obtain, leading to a decrease in the performance or even failure of machine learning methods based on time-frequency feature maps.

[0006] [1]Liu Ningbo,Xu Yanan,Ding Hao,Xue Yonghua,Guan Jian,“High-Dimensional Feature Extraction of Sea Clutter and Target Signal forIntelligent Maritime Monitoring Network”,in Computer Communications,147:76-84,2019

[0007] [2]Bjorklund,S.;Wadstromer,N.,“Target Detection and Classification ofSmall Drones by Deep Learning on Radar Micro-Doppler”,in 2019InternationalRadar Conference(RADAR2019),TOULON,France,2019

[0008] [3]Xiaolong Chen;Ningyuan Su;Jian Guan;Xiaoqian Mou;Yonghua Xue,“Integrated Processing of Radar Detection and Classification for MovingTarget via Time-frequency Graph and CNN Learning”,in 2019URSI Asia-PacificRadio Science Conference(AP-RASC),March 2019,New Delhi,India

[0009] [4]Xiaolong Chen;Ningyuan Su;Yong Huang;Jian Guan,“False-Alarm-Controllable Radar Detection for Marine Target Based on Multi Features Fusionvia CNNs”,in IEEE Sensors Journal,21(7):9099-111,2021

[0010] [5] D.Callaghan; J.Burger; Amit K Mishra, "A Machine Learning Approach toRadar Sea Clutter Suppression", in 2017 IEEE Radar Conference (RadarConf), 2017, Seattle, WA, USA

[0011] [6] Hu Wen, Li Mengxia, Di Jiaying, Wang Weiguang, Wang Yadong, Chen Jie, “A Radar Clutter Suppression Method Based on Machine Learning”, Publication No.: CN109444840B Summary of the Invention

[0012] To address the existing technical problems, this invention provides a multi-dimensional feature false alarm removal method based on ensemble learning.

[0013] The specific content of this invention is as follows: A multi-dimensional feature false alarm removal method based on ensemble learning, comprising the following steps:

[0014] a) Perform signal processing on the radar downlink AD data to generate an intensity map;

[0015] b) Low-threshold constant false alarm rate detection based on the intensity map after signal processing;

[0016] c) Perform intra-frame aggregation on low-threshold constant false alarm rate (CFAR) detection results;

[0017] d) Defuzzify the suspected target points after aggregation and extract multidimensional fuzzy features;

[0018] e) Use an ensemble learning false alarm elimination model to predict the multidimensional fuzzy features of the input, and discard the suspected targets that are predicted to be false alarms to achieve false alarm elimination;

[0019] f) Perform inter-frame aggregation on the targets after false alarm removal to obtain the final spot results;

[0020] Some of the above steps can be omitted. For example, if the radar downlink data is directly the intensity map accumulated after each frame, step a can be omitted; if the radar downlink data is the result of suspected target points after threshold detection, steps a and c can be omitted; even if the radar downlink data is AD data, steps c (intra-frame aggregation) and f (inter-frame aggregation) can also be omitted.

[0021] Furthermore, the processing of AD data in step a includes, but is not limited to, pulse compression, digital beamforming, pulse Doppler, and spatiotemporal adaptive processing, and the order of each processing step can be adjusted.

[0022] Furthermore, the extracted multidimensional fuzzy features are used to construct a multidimensional feature space, including:

[0023] A multidimensional feature space is constructed using the properties measured by the target, including but not limited to the target's range, azimuth, pitch, Doppler velocity, SNR, SCR, ACE, RCS, number of extended range gates, and number of extended Doppler values.

[0024] A multidimensional feature space is constructed using the distances of the target's measured attributes across multiple unblurred frames. This space includes, but is not limited to, the target's distance dimension, azimuth dimension, pitch dimension, Doppler velocity dimension, SNR dimension, SCR dimension, ACE dimension, and RCS dimension. The distance calculation for these attributes can employ, but is not limited to, Manhattan distance, Euclidean distance, Mahalanobis distance, Minkowski distance, correlation coefficient distance, and cosine similarity distance.

[0025] Furthermore, in step e, the method for constructing the ensemble learning false alarm removal model includes:

[0026] (11) Select the input point:

[0027] The deblurred suspected point targets can be used as input, or the suspected target points before deblurring can be selected as input. After the deblurring operation, the deblurred suspected point targets are obtained. When selecting the suspected target points before deblurring as input, the suspected targets that have not been aggregated after CFAR detection can be selected for deblurring, or the aggregated suspected targets can be used for deblurring.

[0028] (12) Error correction:

[0029] By comparing the external information, which serves as the true value, with the suspected target points detected by radar, the systematic bias between the external information and radar detection is corrected.

[0030] (13) Matching:

[0031] The deblurred suspected target point results are matched with the external information that serves as the ground truth. Matched suspected targets are marked as real targets, while those that cannot be matched are marked as false targets.

[0032] (14) Data labeling:

[0033] The dataset is constructed by combining the multidimensional features of the deblurred suspected target with the corresponding annotations;

[0034] (15) Database construction:

[0035] A portion of the dataset is used as training data to train a multidimensional feature false alarm removal model based on ensemble learning, while the remainder is used as validation or test data, where the validation and test datasets can be omitted.

[0036] Furthermore, in step (12), the system bias between external information and radar detection is corrected, including but not limited to time bias, range bias, azimuth bias, pitch bias and velocity bias; the matching in step (13) can be performed on some or all of the dimensions of time-range-azimuth-pitch-velocity, and the distance calculation in the matching includes but is not limited to Manhattan distance, Euclidean distance, Mahalanobis distance, Minkowski distance, correlation coefficient distance and cosine similarity distance.

[0037] Furthermore, the ensemble learning false alarm removal model includes multiple base learners. The output of the ensemble learning false alarm removal model is generated by merging the outputs of each base learner. The model training process includes the following steps:

[0038] (21) The training is scheduled to run for a total of T0 rounds, and it can be set to terminate early;

[0039] (22) In each iteration, m samples are randomly selected from the original training dataset to train and build each base learner. The number of selected samples m is less than or equal to the number of samples m0 in the entire original training dataset.

[0040] (23) After generating the base learner in each round, evaluate the validation loss using the validation dataset. If the early termination condition is met in the Tth iteration, then terminate the training, where T≤T0.

[0041] (24) After repeated independent sampling T times, T different base learners are generated. The classification results of these base learners are combined by the fusion and convergence strategy to obtain the final classification result.

[0042] Furthermore, in step (21), an early termination strategy can be set based on, but not limited to, the method of evaluating the validation loss using the validation dataset in each iteration.

[0043] In step (22), each base learner can be a decision tree, support vector machine or convolutional neural network, and the base learners can be the same or different. In each iteration, samples are randomly selected from the original training dataset. The selection strategy can be selection with replacement or selection without replacement. The random selection strategy can be uniform random selection or weighted probability selection.

[0044] In step (23), the loss corresponding to each sample can be weighted or unweighted when training each base learner;

[0045] In step (24), the convergence and fusion strategy of the base learner output can be either a voting strategy or a weighted average method;

[0046] The sample selection strategy, base learner training method, and base learner output convergence and fusion method correspond to algorithms including but not limited to Bagging, Boosting, Stacking, GBDT, xgboost, and random forest and their variants.

[0047] Furthermore, the ensemble learning false alarm elimination model performs inference and prediction on the input. It can directly select the class with high confidence as the output class, or it can select a certain threshold. When the confidence of a certain class is higher than the threshold, the input is determined to be of that class.

[0048] Furthermore, only the corresponding suspected target points that are determined to be real targets are retained as the output point trace results.

[0049] Furthermore, K can be set as needed when building ensemble learning models. F Different categories of fake targets and K T Different real target categories are used to determine the K corresponding to the multidimensional fuzzy features of the input target during inference. T If any of the different real target categories are selected, that target will be retained as the output point trace.

[0050] Unlike traditional methods that rely solely on energy threshold detection, this invention provides a novel method for constructing a multidimensional feature space for targets. Based on this constructed feature space, an ensemble learning approach driven by data is used to further discriminate suspected targets detected with low thresholds, thereby enhancing the detection capabilities of radar systems. Attached Figure Description

[0051] The specific embodiments of the present invention will be further explained below with reference to the accompanying drawings.

[0052] Figure 1 This is a flowchart of the false alarm rejection process based on ensemble learning.

[0053] Figure 2 Flowchart for constructing a false alarm rejection model based on ensemble learning;

[0054] Figure 3 Flowchart for training a false alarm removal model based on ensemble learning;

[0055] Figure 4 An example of false alarm removal based on multidimensional features of Bagged-trees is constructed;

[0056] Figure 5 The range-azimuth result of a radar detection point is given, where (a) is the result of the point detection after normal threshold detection; and (b) is the result of the point detection after ensemble learning of multidimensional features to eliminate false alarms and combining low threshold detection. Detailed Implementation

[0057] This invention discloses a multi-dimensional feature false alarm removal method based on ensemble learning. The invention establishes a novel multi-dimensional feature space for targets and constructs an ensemble learning-based false alarm removal model to effectively process the established multi-dimensional features, thereby achieving effective suppression of false targets. Combined with low-threshold detection, the constructed ensemble learning-based multi-dimensional feature false alarm removal model is used to remove suspected points obtained from low-threshold detection, which can reduce the false alarm rate while improving the target detection rate, thus effectively enhancing radar detection and sensing performance.

[0058] like Figure 1 As shown, the multidimensional feature false alarm removal method based on ensemble learning includes the following steps:

[0059] a) Perform signal processing on the radar downlink AD data to generate an intensity map;

[0060] The accumulated intensity map is obtained by performing processing steps, including but not limited to pulse compression, digital beamforming, pulse Doppler (PD) processing, and space-time adaptive processing (STAP), on the downlink AD data of the radar at each wave position. Some processing steps can be omitted, and the order can be adjusted. In this embodiment, the accumulated intensity map is obtained after pulse compression, digital beamforming, and pulse Doppler (PD) processing, and its dimension is range-Doppler-beam.

[0061] b) Low-threshold constant false alarm rate (CFAR) detection based on the intensity map after signal processing;

[0062] In CFAR detection, the cross region is selected as the reference unit, and the adjacent units of the test unit are selected as the protection unit; a low detection threshold is selected, such as 10dB; after CFAR detection, information such as the distance gate and Doppler gate of the suspected target point is obtained.

[0063] c) Perform intra-frame aggregation on low-threshold constant false alarm rate (CFAR) detection results;

[0064] Calculate the distance gate difference and Doppler gate difference between suspected target points detected by CFAR. For any two suspected target points whose distance gate difference and Doppler gate difference are within the set threshold range, select the suspected target point with the larger SNR (signal-to-noise ratio) and discard the suspected target point with the smaller SNR.

[0065] d) Defuzzify the suspected target points after aggregation and extract multidimensional fuzzy features;

[0066] Before defuzzing, angles can be measured on each target point after agglomeration, either by sum-and-sum angle measurement or sum-and-difference angle measurement, as needed. At the same time, the corresponding ACE values ​​are calculated. Then, the suspected target points after agglomeration angle measurement are defuzzified, and the multidimensional fuzzy features of the suspected targets are extracted.

[0067] Match the suspected target points in the current frame with the suspected target points in the i-th previous frame (i < N, where N is the number of frames used for deblurring) in terms of distance, Doppler velocity, and orientation. Output the matched suspected target point traces as the deblurred suspected target point traces. Record the multidimensional blurring features such as distance difference, orientation difference, Doppler velocity, SNR, SCR, ACE, and RCS measured in the current frame and the i-th previous frame for each deblurred suspected target, as shown in Table 1.

[0068] Table 1 Multi-position fuzzy features

[0069] symbol explain <![CDATA[v1]]> Speed ​​of the suspected target in the first frame <![CDATA[SNR1]]> Signal-to-noise ratio (SNR) of the suspected target in the first frame. <![CDATA[SCR1]]> Signal-to-noise ratio (SCR) of a suspected target in the first frame. <![CDATA[ACE1]]> Adaptive coherent estimation (ACE) of the suspected target in the first frame. <![CDATA[RCS1]]> Radar cross-section (RCS) of a suspected target in the first frame. <![CDATA[v2]]> Speed ​​of the suspected target in the second frame <![CDATA[SNR2]]> SNR of the suspected target in the second frame <![CDATA[SCR2]]> The suspected target's SCR in the second frame <![CDATA[ACE2]]> The suspected target's ACE in the second frame <![CDATA[RCS2]]> RCS of the suspected target in the second frame ΔR Distance difference between two measurements of the suspected target Δθ Azimuth difference between two measurements of the suspected target Δφ Pitch difference between two measurements of the suspected target

[0070] e) Perform ensemble learning to eliminate false alarms using multi-dimensional features;

[0071] The defuzzed suspected target points and their corresponding fuzzy features are input into the constructed ensemble learning false alarm elimination model. The ensemble learning model predicts whether the suspected target is a real target and discards the suspected target points that are predicted to be false targets, thereby achieving the suppression of false alarms.

[0072] f) Perform inter-frame aggregation on the targets after false alarm removal to obtain the final spot results;

[0073] Calculate the distance difference and Doppler difference between target points in adjacent n frames after false alarm rejection (n is a preset value, such as 3). For any two target points whose distance difference and Doppler difference are within the set threshold range, select the target point with the larger SNR and discard the target point with the smaller SNR.

[0074] Some steps af above can be omitted. For example, if the radar downlink data is directly the intensity map accumulated after each frame, step a can be omitted; if the radar downlink data is the result of suspected target points after threshold detection, steps a and c can be omitted; even if the radar downlink data is AD data, steps c (intra-frame aggregation) and f (inter-frame aggregation) can be omitted.

[0075] In this preferred embodiment, step d involves extracting multidimensional features of the target to construct a multidimensional feature space. This multidimensional feature space is constructed using target-measured attributes, including but not limited to target range, azimuth, pitch, Doppler velocity, SNR, SCR, ACE, RCS, number of extended range gates, and number of extended Doppler values.

[0076] A multidimensional feature space is constructed using the distances of the target's measured attributes across multiple unblurred frames. This space includes, but is not limited to, the target's distance dimension, azimuth dimension, pitch dimension, Doppler velocity dimension, SNR dimension, SCR dimension, ACE dimension, and RCS dimension. The distance calculation for these attributes can employ, but is not limited to, Manhattan distance, Euclidean distance, Mahalanobis distance, Minkowski distance, correlation coefficient distance, and cosine similarity distance.

[0077] In this preferred embodiment, the construction process of the multidimensional feature false alarm removal model based on ensemble learning in step e is as follows: Figure 2 As shown, the steps are as follows:

[0078] (11) Input the dots before deblurring after CFAR detection.

[0079] (12) Perform conventional defuzzification on these dots to obtain the defuzzified suspected dot results and the set of multiple measurement attributes generated during the defuzzification process, i.e. the above-mentioned target multidimensional defuzzification features.

[0080] In the above steps, either the deblurred suspected point targets can be used as input, or the unblurred suspected target points can be selected as input and then the deblurred suspected point targets are obtained after the deblurring operation. When selecting the unblurred suspected target points as input, either unagglomerated suspected targets after CFAR detection can be used for deblurring, or agglomerated suspected targets can be used for deblurring.

[0081] (13) Match the defuzzified point results with external information such as ADS-B or secondary tracks as the true values, mark the matched suspected targets as real targets, and mark the unmatched targets as false targets.

[0082] Before comparing external information such as ADS-B or secondary tracks, which serve as true values, with radar-detected suspected target points, the system bias between the external information and radar detection is corrected, including but not limited to time bias, range bias, azimuth bias, pitch bias, and velocity bias.

[0083] Matching can be performed on some or all of the dimensions of time, distance, orientation, pitch, and velocity. Distance calculations during matching include, but are not limited to, Manhattan distance, Euclidean distance, Mahalanobis distance, Minkowski distance, correlation coefficient distance, and cosine similarity distance.

[0084] (14) A dataset is constructed by combining the extracted fuzzy features and annotation information.

[0085] (15) Use a portion of the dataset as training data to train a multidimensional feature false alarm removal model based on ensemble learning, and use the remainder as validation or test data. The validation dataset and test dataset can be omitted.

[0086] In this embodiment, the ensemble learning false alarm removal model consists of multiple base learners. Each base learner can be of the same or different types, and can take forms such as decision trees, support vector machines (SVM), or convolutional neural networks (CNN). The output of the ensemble learning model is generated by merging the outputs of each base learner. The training process can be described as follows:

[0087] (21) The training is scheduled to run for a total of T0 rounds, and it can be set to terminate early;

[0088] (22) In each iteration, m samples are randomly selected from the original training dataset to train and build each base learner. The number of selected samples m is less than or equal to the number of samples m0 in the entire original training dataset.

[0089] (23) After generating base learners in each round, evaluate the validation loss by using the validation dataset. If the early termination condition is met in the Tth iteration, terminate the training. Where T≤T0, the termination condition is related to the type of base learner. Select the termination condition according to the type of base learner.

[0090] (24) After repeated independent sampling T times, T different base learners are generated. The classification results of these base learners are combined by the fusion and convergence strategy to obtain the final classification result.

[0091] In this preferred embodiment, in step (21), an early termination strategy can be set according to, but not limited to, the method of evaluating the verification loss using the verification dataset in each iteration;

[0092] In step (22), each base learner can be a decision tree, a support vector machine (SVM), or a convolutional neural network (CNN), and the base learners can be the same or different. In each iteration, samples are randomly selected from the original training dataset. The selection strategy can be selection with replacement or selection without replacement. The random selection strategy can be uniform random selection or weighted probability selection. The corresponding training method is used when generating each base learner. For example, if the base learner is a decision tree, classification and regression trees (CART), ID3, and C4.5 can be used; if the base learner is a CNN, optimizers such as Adam and SDG are selected to optimize the loss function using gradient descent to train and generate the CNN model.

[0093] In step (23), the loss corresponding to each sample can be weighted or unweighted when training each base learner;

[0094] In step (24), the convergence and fusion strategy for the base learner outputs can be a voting strategy, that is, using the output of a certain classification of the base learners with more base learners as the output of the ensemble learning false alarm removal model; or a weighted average method can be used, that is, assigning different weights to different base learners.

[0095] The sample selection strategy, base learner training method, and base learner output convergence and fusion method correspond to algorithms including but not limited to Bagging, Boosting, Stacking, GBDT, xgboost, and random forest and their variants.

[0096] This embodiment uses an ensemble decision tree model, Bagged-trees, as an example to describe the construction of the ensemble learning multidimensional feature false alarm removal model proposed in this invention. Its pseudocode is as follows: Figure 4 As shown. Where X m and y m These are the fuzzy features and labels of the m-th suspected target, respectively, and each base learner is a decision tree. The main steps of its model training are as follows:

[0097] The training is set to run for a total of T rounds, without validation or early termination.

[0098] In each iteration, m samples are randomly selected with replacement from the original training dataset to serve as the training set for training the decision tree. In this embodiment, m = 0.8 × N, where N is the number of samples in the training dataset.

[0099] The Classification and Regression Tree (CART) algorithm is used to construct each decision tree based on the selected m samples. CART is a binary tree that uses a binary splitting method, dividing the data into two parts each time, which are then sent to the left and right subtrees respectively. Furthermore, each non-leaf node has two children, so CART has one more leaf node than non-leaf node.

[0100] In CART classification, the Gini index is used to select the best feature for data splitting. The Gini coefficient is decreased with each iteration of CART, and its calculation formula is shown below.

[0101]

[0102] Where the dataset is D, and the proportion of samples of class k in the dataset is Let Gini(D) be the set of samples from the k-th class. Gini(D) reflects the probability that two randomly selected samples from dataset D are of different classes. Therefore, the smaller the Gini(D) value, the higher the purity of dataset D.

[0103] The expression for the Gini coefficient Gini(D|a) of a certain feature a in a multidimensional fuzzy feature space A is as follows.

[0104]

[0105] Where V is the set of possible values ​​for feature a. If feature a is used to partition dataset D, V branch nodes may be generated, where the v-th branch node contains all nodes in D that take the value a for feature A. v The sample is D v .

[0106] CART selects the attribute that minimizes the Gini coefficient after partitioning in the multidimensional fuzzy feature space A as the partitioning attribute.

[0107] The splitting stops when the number of leaf nodes in the split decision tree reaches the preset minimum or the decision tree depth reaches the preset maximum, meaning that a base learner has been generated.

[0108] (4) Employing a voting strategy, the outputs of the T decision trees generated in the above steps are aggregated and merged. The expression for this is:

[0109]

[0110] Among them, h t (x) is the output of the t-th decision tree given the input feature x, when h t When (x) matches the given label y otherwise Y is the set of all label categories. In this example, Y = {0, 1}, where 0 represents a fake target and 1 represents a real target.

[0111] Figure 5 A comparison of range-azimuth results based on radar detection points is presented. Figure (a) shows the results of point detection using traditional CFAR energy threshold detection with a 13.5dB threshold; Figure (b) shows the results of point detection using the proposed integrated learning multidimensional feature false alarm elimination combined with low-threshold CFAR detection. Comparing the two figures shows that low-threshold detection combined with integrated learning multidimensional feature false alarm elimination significantly improves radar detection capability and effectively controls the false alarm rate. Compared to the normal threshold results, at least three more obvious target trajectories are detected, while the total number of targets is less, resulting in a lower false alarm rate.

[0112] This invention constructs a novel multidimensional feature space by fully utilizing the similarity of measured attributes of real targets across multiple frames used for deblurring. Based on a data-driven approach, it automatically and finely partitions the feature space and employs an ensemble learning model to accurately distinguish between real and false targets, effectively suppressing false alarms. Furthermore, by combining this with low-threshold detection, the detection rate can be improved while reducing the false alarm rate. This invention offers the following advantages: a lower false alarm rate compared to traditional energy threshold detection methods; improved detection rate achieved by combining with low-threshold pre-detection; compatibility with traditional signal processing and detection workflows, allowing for performance enhancement through embedding the technology; high computational efficiency, meeting real-time requirements; and a model with fewer parameters, whose performance is insensitive to non-critical parameter values, significantly reducing the burden of manual parameter tuning.

[0113] Many specific details have been set forth in the foregoing description to provide a thorough understanding of the present invention. However, the above description is merely a preferred embodiment of the present invention, and the present invention can be implemented in many other ways different from those described herein. Therefore, the present invention is not limited to the specific embodiments disclosed above. Furthermore, any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention, or modify them into equivalent embodiments, using the methods and techniques disclosed above, without departing from the scope of the present invention. Any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention, without departing from the content of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A multi-dimensional feature false alarm removal method based on ensemble learning, characterized in that: Includes the following steps: a) Perform signal processing on the radar downlink AD data to generate an intensity map; b) Low-threshold constant false alarm rate detection based on the intensity map after signal processing; c) Perform intra-frame aggregation on low-threshold constant false alarm rate (CFAR) detection results; d) Defuzzify the suspected target points after aggregation and extract multidimensional fuzzy features; The extracted multidimensional fuzzy features are used to construct a multidimensional feature space, including: A multidimensional feature space is constructed using the properties measured by the target, including but not limited to the target's range, azimuth, pitch, Doppler velocity, SNR, SCR, ACE, RCS, number of extended range gates, and number of extended Doppler values. A multidimensional feature space is constructed using the distances of the target's measured attributes across multiple unblurred frames. This space includes, but is not limited to, the target's distance dimension, azimuth dimension, pitch dimension, Doppler velocity dimension, SNR dimension, SCR dimension, ACE dimension, and RCS dimension. The distance calculation for these attributes can be, but is not limited to, Manhattan distance, Euclidean distance, Mahalanobis distance, Minkowski distance, correlation coefficient distance, and cosine similarity distance. e) Utilize an ensemble learning false alarm elimination model to predict the multidimensional fuzzy features of the input, discard the suspected targets that are predicted as false alarms, and achieve false alarm elimination. Methods for constructing ensemble learning false alarm rejection models include: (11) Select the input point: The deblurred suspected point targets can be used as input, or the suspected target points before deblurring can be selected as input. After the deblurring operation, the deblurred suspected point targets are obtained. When selecting the suspected target points before deblurring as input, the suspected targets that have not been aggregated after CFAR detection can be selected for deblurring, or the aggregated suspected targets can be used for deblurring. (12) Error correction: By comparing the external information, which serves as the true value, with the suspected target points detected by radar, the systematic bias between the external information and radar detection is corrected. (13) Matching: The deblurred suspected target point results are matched with the external information that serves as the ground truth. Matched suspected targets are marked as real targets, while those that cannot be matched are marked as false targets. (14) Data labeling: The dataset is constructed by combining the multidimensional features of the deblurred suspected target with the corresponding annotations; (15) Database construction: A portion of the dataset is used as training data to train a multidimensional feature false alarm removal model based on ensemble learning, and the remainder is used as validation or test data, where the validation dataset and test dataset can be omitted. f) Perform inter-frame aggregation on the targets after false alarm removal to obtain the final spot results.

2. The multidimensional feature false alarm removal method based on ensemble learning according to claim 1, characterized in that: Step a involves processing the AD data, including but not limited to pulse compression, digital beamforming, pulse Doppler, and spatiotemporal adaptive processing. The order of these processes can be adjusted.

3. The multidimensional feature false alarm removal method based on ensemble learning according to claim 1, characterized in that: In step (12), the system bias between external information and radar detection is corrected, including but not limited to time bias, range bias, azimuth bias, pitch bias and velocity bias; the matching in step (13) can be performed on some or all of the dimensions of time-range-azimuth-pitch-velocity, and the distance calculation in the matching includes but is not limited to Manhattan distance, Euclidean distance, Mahalanobis distance, Minkowski distance, correlation coefficient distance and cosine similarity distance.

4. The multidimensional feature false alarm removal method based on ensemble learning according to claim 1, characterized in that: The ensemble learning false alarm removal model consists of multiple base learners. The output of the ensemble learning false alarm removal model is generated by merging the outputs of each base learner. The model training process includes the following steps: (21) The training is scheduled to be carried out in total. The wheel can be set to terminate early; (22) Each iteration randomly selects from the original training dataset. For each base learner, a sample is selected to train and construct the base learner. Less than or equal to the number of samples in the entire original training dataset ; (23) After generating the base learner in each round, evaluate the validation loss using the validation dataset. If in the first round... Training terminates if the early termination condition is met during each iteration. ; (24) Repetition Generated after independent sampling The classification results of these different base learners are combined through a fusion and convergence strategy to obtain the final classification result.

5. The multidimensional feature false alarm removal method based on ensemble learning according to claim 4, characterized in that: In step (21), an early termination strategy can be set based on, but not limited to, the method of evaluating the validation loss using the validation dataset in each iteration. In step (22), each base learner can be a decision tree, support vector machine or convolutional neural network, and the base learners can be the same or different. In each iteration, samples are randomly selected from the original training dataset. The selection strategy can be selection with replacement or selection without replacement. The random selection strategy can be uniform random selection or weighted probability selection. In step (23), the loss corresponding to each sample can be weighted or unweighted when training each base learner; In step (24), the convergence and fusion strategy of the base learner output can be either a voting strategy or a weighted average method; The sample selection strategy, base learner training method, and base learner output convergence and fusion method correspond to algorithms including but not limited to Bagging, Boosting, Stacking, GBDT, xgboost, and random forest and their variants.

6. The multidimensional feature false alarm removal method based on ensemble learning according to claim 5, characterized in that: Ensemble learning false alarm rejection models perform inference and prediction on inputs. They can directly select the class with high confidence as the output class, or they can select a certain threshold. When the confidence of a certain class is higher than the threshold, the input is determined to be that class.

7. The multidimensional feature false alarm removal method based on ensemble learning according to claim 5, characterized in that: Only the corresponding suspected target points that are determined to be real targets are retained as the output point results.

8. The multidimensional feature false alarm removal method based on ensemble learning according to claim 7, characterized in that: When building an ensemble learning model, settings can be configured as needed. Different categories of fake targets and Different real target categories are used to determine the correspondence of multidimensional fuzzy features of the input target during inference. If any of the different real target categories are selected, that target will be retained as the output point trace.

Images