A radar data autonomous identification method and device based on artificial intelligence

Abstract

The present invention provides an artificial intelligence-based autonomous radar data identification method and device, comprising the following steps: Step a: Offline training: Preprocessing historical radar data to construct a training set; Designing a classification model and using the training set to train the classification model; Step b: Online inference: Using the classification model to identify the type of target currently detected by the radar in real time; Step c: Continuous learning: Using human feedback, continuously updating the classification model obtained in Step a using a continuous learning algorithm, and using the updated model for online inference. The artificial intelligence-based autonomous radar data identification method and device provided by the present invention utilizes artificial intelligence methods such as deep learning and continuous learning based on existing radar data to achieve rapid, accurate, and scalable target identification.

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal feature extraction and target recognition and classification, and in particular to a method and device for autonomous recognition of radar data based on artificial intelligence. Background Art

[0002] Radar, as the most crucial detection equipment, is a top priority for development and holds great promise. Currently, radars lack intelligence. Domestic radars are limited to detecting traditional indicators like target presence, altitude, and speed, but are unable to accurately identify and classify targets. Meanwhile, the technology behind new aircraft, such as drones, has matured rapidly across various fields, even finding successful battlefield applications. This presents a significant disadvantage. These drones are difficult for existing radars to distinguish from non-drone targets, such as birds. Therefore, adding new detection capabilities to existing radars and developing technologies for accurately identifying and classifying these new aircraft and other related targets has become an urgent need.

[0003] Patents such as Chinese patents CN201610896005.6, CN202011419393.1, CN201810851549.X, CN202111110784.X, and "Drone Classification Using Convolutional Neural Networks With Merged Doppler Images" by BK Kim, HS Kang, and SO Park all use traditional methods, augmented image information, deep learning, and the micro-Doppler effect to identify small, low-altitude targets such as drones and birds. However, these methods lack feedback to update the model and can only identify two fixed target types. Even if they misidentify, they cannot be corrected. This leads to two problems: first, it is impossible to identify new types of flying objects; second, if one drone is misidentified, all drones can imitate the behavior of the misidentified drone, rendering the recognition system completely ineffective.

[0004] On the other hand, in the data obtained by the radar, information such as RCS is related to the target's posture. Therefore, all data enhancement methods that will cause target translation should not be used, which greatly limits the processing solutions in the case of sample imbalance. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and device for autonomous recognition of radar data based on artificial intelligence, in order to solve the problems existing in the background technology.

[0006] This method is divided into three stages: offline training, online reasoning, and continuous learning. In the offline training stage, a classification model is trained by constructing a training dataset; in the online reasoning stage, the type of radar target is identified in real time through the trained classification model; in the continuous learning stage, the classification model is updated using continuous learning technology by manually finding samples with recognition errors.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] An artificial intelligence-based autonomous radar data recognition method comprises the following steps:

[0009] Step a: Offline training: pre-processing historical data collected by the radar to construct a training set; designing a classification model, and using the training set to train the classification model;

[0010] Step b: Online reasoning: using the classification model to identify the type of target currently detected by the radar in real time;

[0011] Step c: Continuous learning: Using manual feedback, the classification model obtained in step a is continuously updated using a continuous learning algorithm, and the updated model is used for online reasoning.

[0012] As a further optimization, the historical data collected by the radar in step a refers to the track data of different targets with labels; the label refers to the category of the target, such as drones, birds, fixed-wing aircraft, etc.

[0013] As a further optimization, the preprocessing described in step a includes cleaning, interpolation, and framing of trajectory data. Abnormal or too short trajectories are removed through cleaning. Missing data on the trajectory is supplemented through interpolation. The framing operation refers to using a sliding window with a fixed step size k to slide on the trajectory, intercepting a frame of data of length L each time it slides.

[0014] As a further optimization, the design of a classification model described in step a and the use of a training set to train the model means that a skeleton model is pre-built with a neural network. The last layer of the skeleton model is a fully connected classification layer, and the softmax function is used to output the probability of belonging to each class; each frame data obtained by preprocessing is fed into the designed classification model, and the model parameters are trained using labels; the neural network can be ResNet, RNN, LSTM, Transformer, etc., or a network with customized node connection methods and activation functions

[0015] As a further optimization, the use of the classification model obtained in step a in step b to identify in real time the type of the target currently detected by the radar means dividing the track of the target currently detected by the radar into frames according to the framing operation in step a, and then sending it to the classification model obtained in step a to determine the type of each frame; and the type of the target is obtained by combining the judgment results of all frames intercepted from its track according to the voting method.

[0016] As a further optimization, the classification model obtained in step a is continuously updated using a continuous learning algorithm as described in step c, including:

[0017] c1) When the manual feedback classification model incorrectly determines that the target is a type already in the training set of step a, the incorrectly judged samples are collected and pre-processed as described in step a as a difficult sample set, and then fed into the classification model obtained in step a for training. The classification model parameters are fine-tuned using the continuous learning fine-tuning method; finally, the difficult sample set is merged with the training set of step a to update the training set;

[0018] c2) When the manual feedback classification model incorrectly determines that the target is a new type that has not appeared in the training set, a node is added to the last fully connected classification layer of the neural network classification model designed in step a to output the probability of belonging to the new type; at the same time, these incorrectly judged samples are preprocessed as described in step a and then added to the training set through data augmentation to form a new training set, and the classification model is jointly trained;

[0019] c3). The training set in step a is updated as the samples in steps c1 and c2 are merged; the classification model in step a is updated as the fine-tuning and joint training in steps c1 and c2 are performed;

[0020] As a further optimization, the data enhancement in step c2 is to solve the problem of sample imbalance between the new type that has not been seen and the existing types in the training set. The specific approach is to calculate the projection (x, y) of each point on the track of the new type target in the plane rectangular coordinate system and use Indicates information that is independent of the distance from the plane coordinate origin. The track is then rotated along the horizontal plane at angles a1, a2, ..., a according to the coordinate origin. N , get the enhanced data Finally, Gaussian noise α*GaussianNoise(0, 1) with a mean of 0 and a variance of 1 is added to each dimension of the enhanced data, where N is a preset positive integer and α is a preset very small positive number.

[0021] As a further optimization, the fine-tuning method described in step c1 refers to freezing the first T layers of the classification model and further optimizing the parameters of the latter ST layers of the classification model by minimizing the loss function, where S is the total number of layers of the classification model; the joint training described in step c2 refers to initializing the parameters of the classification model and restarting training.

[0022] As a further optimization, the updated model is used for online reasoning in step c, which means that the classification model of step c3 is used to identify the type of target currently detected by the radar in real time;

[0023] The second aspect of the present invention provides an autonomous radar data recognition device based on artificial intelligence, which is composed of a training module, a reasoning module, and an update module; wherein

[0024] The training module is used to construct a training data set and train the classification model; its methods include data interpolation, data cleaning, data framing, model design, and model parameter optimization;

[0025] The inference module uses the classification model to identify the type of radar target in real time; its methods include model inference and result integration;

[0026] The update module continuously updates the classification model based on the results of manual feedback; its methods include data enhancement, parameter fine-tuning, and joint training.

[0027] The present application provides an artificial intelligence-based radar data autonomous recognition method and device with beneficial effects including but not limited to:

[0028] Through continuous learning, the present invention realizes automatic updating of the model and timely correction of errors, thereby preventing targets such as drones from replicating the behavior of misidentified drones and ultimately breaking through the recognition system.

[0029] The present invention uses a unique data enhancement technology to alleviate the sample imbalance problem; at the same time, unlike existing methods, the present invention can identify new categories of flying objects that have never appeared in historical radar data.

[0030] The present invention does not require complex geographical environments, images, or other information, and does not increase any hardware costs. It uses artificial intelligence methods such as deep learning and continuous learning based on existing radar data to achieve rapid, accurate, and scalable target recognition.

[0031] Through continuous learning, the present invention realizes automatic updating of the model and timely correction of errors, thereby preventing targets such as drones from replicating the behavior of misidentified drones and ultimately breaking through the recognition system. At the same time, unlike existing methods, the present invention can identify new categories of flying objects that have never appeared in historical radar data. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flow chart of the radar data autonomous identification method based on artificial intelligence of the present invention.

[0033] Figure 2 A network structure diagram of the classification model in an embodiment of the present invention is given.

[0034] Figure 3 A network structure diagram of the residual block ResBlock(4,8) in an embodiment of the present invention is given.

[0035] Figure 4 It shows that the track T10 is enhanced by 4500 mil and 1500 mil to obtain T11 and T12.

[0036] Figure 5 A schematic diagram of the fine-tuning method in an embodiment of the present invention is given.

[0037] Figure 6 A schematic diagram of joint training in an embodiment of the present invention is given.

[0038] Figure 7 It is a structural diagram of the radar data autonomous identification device based on artificial intelligence of the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0040] On the contrary, this application covers any alternatives, modifications, equivalents, and solutions made within the spirit and scope of this application as defined by the claims. Furthermore, to facilitate a better understanding of this application, certain specific details are described in detail below in the detailed description of this application. Those skilled in the art will be able to fully understand this application without these details.

[0041] The following will be combined Figure 1-7 The following embodiments of the present application are intended to illustrate a method and device for autonomously identifying radar data based on artificial intelligence. It should be noted that the following embodiments are intended only to explain the present application and do not constitute a limitation thereto.

[0042] An artificial intelligence-based autonomous radar data recognition method comprises the following steps:

[0043] Step a: Offline training: pre-processing historical data collected by the radar to construct a training set; designing a classification model, and using the training set to train the classification model;

[0044] Step b: Online reasoning: using the classification model to identify the type of target currently detected by the radar in real time;

[0045] Step c: Continuous learning: Using manual feedback, the classification model obtained in step a is continuously updated using a continuous learning algorithm, and the updated model is used for online reasoning.

[0046] The historical data collected by the radar refers to the track data of different targets that have been labeled; the label refers to the target category, such as drone, bird, fixed-wing aircraft, etc. The track data in this embodiment includes information such as track number, time, distance, direction, pitch, and speed. The historical data includes three types of targets: drone, bird, and fixed-wing aircraft.

[0047] The preprocessing process includes cleaning, interpolation, and framing of trajectory data. Cleaning eliminates abnormal or short trajectories, and interpolation supplements missing data. Framing involves sliding a window across the trajectory with a fixed step size k, capturing a frame of data of length L with each slide. In this embodiment, an abnormality is identified when the calculated velocity of two consecutive points on the trajectory deviates by more than 30% from the velocity of the trajectory data itself, and the trajectory is broken at this point. Linear interpolation is used, with a step size k of 2 and a frame length of L of 6.

[0048] The design of a classification model and the use of a training set to train the model refer to building a skeleton model with a neural network in advance, wherein the last layer of the skeleton model is a fully connected classification layer, and the softmax function is used to output the probability of belonging to each category; each frame data obtained by preprocessing is sent to the designed classification model, and the model parameters are trained using labels.

[0049] In this embodiment, the neural network uses three ResBlock hidden layers, one fully connected layer, and the classification function is the softmax function. The network structure is as follows Figure 2 As shown. The ResBlock layer is a residual block composed of one-dimensional convolution and skip connection. ResBlock(4,8) means that the input channel is 4 and the output channel is 8. The structure is as follows Figure 3 As shown in the figure, Conv[3, 1, 8], padding=same indicates a convolution kernel size of 3, a stride of 1, and 8 channels. Padding=same indicates padding to make the input and output dimensions the same. The loss function chosen is cross-entropy loss. During training, the learning rate upper limit was set to 0.02 and the lower limit was set to 0.01. The Adam optimization algorithm was used, with a batch size of 64. The number of model training iterations was 200.

[0050] The use of the classification model obtained in step a to identify in real time the type of the target currently detected by the radar means dividing the track of the target currently detected by the radar into frames according to the framing operation in step a, and then feeding the classification model obtained in step a to determine the type of each frame; and the type of the target is determined by combining the judgment results of all frames intercepted from its track according to a voting method.

[0051] The voting method of this embodiment is soft voting.

[0052] The method of continuously updating the classification model obtained in step a using a continuous learning algorithm includes:

[0053] c.1) When the manual feedback classification model incorrectly identifies a target that is already a type in the training set in step a, the incorrectly identified samples are collected and preprocessed as described in step a as a difficult sample set. These samples are then fed into the classification model obtained in step a for training, and the classification model parameters are fine-tuned using a continuous learning fine-tuning method. Finally, the difficult sample set is merged with the training set in step a to update the training set.

[0054] c.2) When the manual feedback classification model incorrectly identifies a target of a new type not previously seen in the training set, a node is added to the final fully connected classification layer of the neural network classification model designed in step a to output the probability of belonging to the new type. Simultaneously, these incorrectly identified samples are pre-processed as described in step a and added to the training set to form a new training set, and the classification model is trained jointly.

[0055] c.3) The training set in step a is updated as the samples in steps c.1 and c.2 are merged; the classification model in step a is updated as the fine-tuning and joint training in steps c.1 and c.2 are performed;

[0056] The data augmentation is to solve the problem of sample imbalance between the new type that has not been seen and the existing types in the training set. The specific approach is to calculate the projection (x, y) of each point on the new type target track on the plane rectangular coordinate system and use Indicates information that is independent of the distance from the plane coordinate origin. The track is then rotated along the horizontal plane at angles a1, a2, ..., a according to the coordinate origin. N , and get the enhanced data Finally, Gaussian noise α*GaussianNoise(0, 1) with a mean of 0 and a variance of 1 is added to each dimension of the enhanced data. N is a preset positive integer, and α is a preset small positive number. In this embodiment, the new type of small training aircraft is not included in the historical data, and the number of samples is far smaller than that of drones, birds, and fixed-wing aircraft. Therefore, the data enhancement process is used. Figure 4The enhanced results T11 and T12 of the flight path T10 of a small training aircraft using a1=4500 mils and a2=1500 mils are shown. The noise added in this embodiment is 0.4*GaussianNoise(0,1).

[0057] The fine-tuning method refers to freezing the first T layers of the classification model and further optimizing the parameters of the ST layers of the classification model by minimizing the loss function, where S is the total number of layers of the classification model; the joint training refers to initializing the parameters of the classification model and restarting the training. In this embodiment, the schematic diagram of the fine-tuning method is as follows Figure 5 As shown, the schematic diagram of joint training is as follows Figure 6 shown.

[0058] The use of the updated model for online reasoning refers to using the classification model in step c.3 to identify in real time the type of the target currently detected by the radar.

[0059] It should be noted that the dimension of the track data, the category of the target, the step size and frame length of the frame division, the structure of the neural network, the training method, etc. can all be determined according to actual conditions and are not limited to the specific values ​​listed in this embodiment.

[0060] This embodiment also provides an autonomous radar data recognition device based on artificial intelligence, such as Figure 7 As shown in the figure, it consists of a training module, an inference module, and an update module; the training module is used to construct a training data set and train a classification model, and its methods include data interpolation, data cleaning, data framing, model design, and model parameter optimization; the inference module uses the classification model to identify the type of radar target in real time, and its methods include model inference and result integration; the update module continuously updates the classification model according to the results of manual feedback, and its methods include data enhancement, parameter fine-tuning, and joint training.

[0061] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A radar data autonomous recognition method based on artificial intelligence, characterized in that: The following steps are involved: Step a: Offline training: Preprocess the historical data collected by the radar to construct a training set; Designing a classification model, and training the classification model using the training set; Step b: Online reasoning: using the classification model to identify the type of target currently detected by the radar in real time; Step c: Continuous learning; Through manual feedback, the classification model obtained in step a is continuously updated using a continuous learning algorithm, and the updated model is used for online reasoning; Step c: continuously updating the classification model obtained in step a using a continuous learning algorithm through manual feedback, including: c1) When the manual feedback classification model incorrectly determines that the target is a type already in the training set of step a, the incorrectly judged samples are collected and pre-processed as described in step a as a difficult sample set, and then fed into the classification model obtained in step a for training. The classification model parameters are fine-tuned using the continuous learning fine-tuning method; finally, the difficult sample set is merged with the training set of step a to update the training set; c2) When the manual feedback classification model incorrectly determines that the target is a new type that has not appeared in the training set, a node is added to the last fully connected classification layer of the neural network classification model designed in step a to output the probability of belonging to the new type; at the same time, these incorrectly judged samples are preprocessed as described in step a and then added to the training set through data augmentation to form a new training set, and the classification model is jointly trained; c3). The training set in step a is updated as the samples in steps c1 and c2 are merged; the classification model in step a is updated as the fine-tuning and joint training in steps c1 and c2 are performed; The data enhancement in step c2 includes: calculating the projection of each point on the new type target track on the plane rectangular coordinate system and use Indicates information that is independent of the distance from the plane coordinate origin. The track is then rotated along the horizontal plane according to the coordinate origin. , ,…, , and get the enhanced data 、 ,…, Finally, add Gaussian noise with a mean of 0 and a variance of 1 to each dimension of the enhanced data. ;in is a preset positive integer, is 0.

4.

2. The method for autonomous identification of radar data based on artificial intelligence according to claim 1, characterized in that: The historical data collected by the radar in step a includes: track data of different targets with labels; the labels are the categories of the targets.

3. The method for autonomous recognition of radar data based on artificial intelligence according to claim 1, characterized in that: The preprocessing described in step a includes cleaning, interpolation, and framing of trajectory data.

4. The method for autonomous identification of radar data based on artificial intelligence according to claim 3, characterized in that: The cleaning refers to removing abnormal or too short trajectories; the interpolation refers to supplementing the missing data on the trajectory; the framing operation refers to using a sliding window with a fixed step size k to slide on the trajectory, and each sliding operation intercepts a frame of data with a length of L.

5. The method for autonomous identification of radar data based on artificial intelligence according to claim 1, characterized in that: In step a, the design of the classification model refers to pre-building a skeleton model using a neural network, wherein the last layer of the skeleton model is a fully connected classification layer, and the softmax function is used to output the probability of belonging to each category; the neural network is one of ResNet, RNN, LSTM, Transformer, or a network with customized node connection methods and activation function forms.

6. The method for autonomous identification of radar data based on artificial intelligence according to claim 1, characterized in that: In the step a, the classification model is trained using the training set, including inputting each frame data obtained by preprocessing into the classification model and using labels to train model parameters.

7. The method for autonomous identification of radar data based on artificial intelligence according to claim 1, characterized in that: Step b: using the classification model to identify in real time the type of the target currently detected by the radar; including dividing the track of the target currently detected by the radar into frames according to the framing operation of step a, and then feeding the classification model obtained in step a to determine the type of each frame; wherein the type of the target is obtained by combining the judgment results of all frames intercepted from its track according to a voting method.

8. The method for autonomous identification of radar data based on artificial intelligence according to claim 1, characterized in that: The fine-tuning method in step c1 includes freezing the first T layers of the classification model and optimizing the parameters of the last ST layers of the classification model by minimizing the loss function, where S is the total number of layers of the classification model; the joint training in step c.2 refers to initializing the parameters of the classification model and restarting training.

9. The method for autonomous identification of radar data based on artificial intelligence according to claim 1, characterized in that: The use of the updated model for online reasoning in step c refers to using the classification model of step c3 to identify the type of target currently detected by the radar in real time.

10. An artificial intelligence-based radar data autonomous recognition device, characterized in that: It consists of a training module, an inference module, and an update module; among them: The training module is used to construct a training data set and train the classification model; its methods include data interpolation, data cleaning, data framing, model design, and model parameter optimization; The inference module uses the classification model to identify the type of radar target in real time; its methods include model inference and result integration; The update module continuously updates the classification model based on manual feedback; its methods include data enhancement, parameter fine-tuning, and joint training; The update module uses a continuous learning algorithm to continuously update the classification model obtained by the training module through manual feedback, including: c1) When the manual feedback classification model incorrectly identifies a target that is already in the training set of the training module, the incorrectly identified samples are collected and preprocessed by the training module as a difficult sample set. This is then fed into the classification model obtained by the training module for training, and the classification model parameters are fine-tuned using a continuous learning fine-tuning method. Finally, the difficult sample set is merged with the training set of the training module to update the training set. c2) When the manual feedback classification model incorrectly identifies a target of a new type not previously seen in the training set, a node is added to the final fully connected classification layer of the neural network classification model designed by the training module to output the probability of belonging to the new type. Simultaneously, these incorrectly identified samples are pre-processed by the training module and then subjected to data augmentation before being added to the training set to form a new training set, and the classification model is jointly trained. c3). The training set in the training module is updated as the samples in step c1 and step c2 are merged; the classification model in the training module is updated as the fine-tuning and joint training in step c1 and step c2 are performed; The data enhancement in step c2 includes: calculating the projection of each point on the new type target track on the plane rectangular coordinate system and use Indicates information that is independent of the distance from the plane coordinate origin. The track is then rotated along the horizontal plane according to the coordinate origin. , ,…, , and get the enhanced data 、 ,…, Finally, add Gaussian noise with a mean of 0 and a variance of 1 to each dimension of the enhanced data. ;in is a preset positive integer, is 0.4.

Images