Low-altitude radar unexpected track suppression method and system based on deep learning

Abstract

The invention discloses a low-altitude radar unexpected track suppression method based on deep learning. According to the method, unexpected tracks are trained and reasoned by using real-time track data, and real tracks and unexpected tracks are identified. According to the method, a conventional method is adopted to carry out track automatic initiation, track initiation data is used as a learning set of an LSTM network to carry out training, and network parameters are determined; in the reasoning stage, the track is used as input for reasoning, the LSTM network outputs an unexpected track mark, if the track is a real track, the target track is established in the radar, and otherwise, the target track is marked as an unexpected track. The method has the advantages of few constraint conditions, online self-learning capability, short learning time and high reasoning speed, and meets the real-time target tracking requirement of the radar. Actual data tests show that the unexpected track recognition probability of the method is larger than 98%, the network generalization ability is high, and the method is an efficient low-altitude radar unexpected track suppression method and can be widely applied to low-altitude radars.

Description

Technical Field

[0001] The present invention relates to the technical field of low-altitude target recognition, and in particular to a method and system for suppressing undesired low-altitude radar tracks based on deep learning. Background Art

[0002] The low-altitude economy is developing rapidly in my country, and an initial industrial chain has been formed. To develop the low-altitude economy, we must first solve the problem of safety support. Radar, as an active detection method, has the advantages of long detection distance, continuous detection and tracking around the clock and in all weather conditions. It can remotely warn and guide various strike weapons such as optoelectronics, lasers, microwave weapons, and net bombs for counterattacks. It is the core and key means of monitoring low-altitude non-cooperative targets and is widely used in various military and civilian anti-satellite systems.

[0003] An important prerequisite for realizing the above radar functions is that the radar can track low-altitude targets in real time and stably, and provide correct, continuous and stable target track data. However, in actual applications, due to the complex working environment, low-altitude radars face serious clutter interference, including ground moving targets (such as people, cars), birds, meteorological clutter (wind, clouds, rain, snow, atmospheric movement), etc. These objective physical entities and phenomena will be detected by low-altitude radars and form tracks. These tracks come from real physical phenomena, and there are huge numbers of them, but they are not the targets that radar detection needs to pay attention to. They are called undesired tracks. The existence of a large number of undesired tracks will have a catastrophic impact on drone countermeasures.

[0004] Traditional radar data processing adopts the idea of ​​clustering and uses the inter-frame motion characteristics to initiate and correlate tracks, thereby establishing and maintaining tracks. Unexpected tracks come from real physical entities and conform to the target motion characteristics. They are almost the same as real UAVs in terms of motion characteristics (altitude, speed, heading, Doppler speed, amplitude). Various target track initiation algorithms (such as logical initiation method, Hough transform, MHT, etc.) do not have the ability to identify and suppress unexpecting targets.

[0005] In order to solve the problem of unexpected tracks, methods based on rule judgment and statistical feature analysis are usually used to judge and suppress unexpected tracks. However, in practical applications, there are common defects such as low recognition and suppression rate of unexpected tracks and poor universality, which has become a bottleneck for low-altitude radar applications. Summary of the invention

[0006] The technical problem to be solved by the present invention is how to suppress the undesired track of low-altitude radar in a lightweight and high-precision manner.

[0007] The present invention solves the above technical problems through the following technical means:

[0008] The low-altitude radar undesired track suppression method based on deep learning includes the following steps:

[0009] S1. Establish an LSTM neural network and initialize the LSTM neural network;

[0010] S2, using conventional radar data processing methods to automatically start the target track;

[0011] S3, manually marking whether the track is an unexpected track, and forming marking data;

[0012] S4, LSTM neural network receives the labeled track data online, performs network training, and determines network parameters;

[0013] S5, LSTM neural network receives the track data to be identified online and performs neural network reasoning;

[0014] S6, LSTM neural network outputs an unexpected track mark; if the LSTM output is an unexpected track, the track is invalidated, if it is a real track, the target track is established, and the unexpected track suppression method is completed.

[0015] Furthermore, the specific method of step S1 includes the following sub-steps:

[0016] S1-1. The input of LSTM neural network is a two-dimensional vector of the starting point information of the track, which can be expressed as

[0017]

[0018] in:

[0019] n is the number of track starting points, Ti is the time of the ith track point, R i is the distance of the ith track point, A i is the position of the ith track point, H i is the height of the ith track point, Vi is the speed of the ith track point, C i is the heading of the ith track point, Fd i is the Doppler velocity of the ith track point, Amp i is the amplitude of the ith track point, N i is the background amplitude of the i-th track point, 1≤i≤n;

[0020] S1-2. Refer to the typical LSTM network model structure, use a custom time series network model, and select the LSTM combined with Transformer network model for intelligent undesired track suppression;

[0021] S1-3, Hyperparameter settings: defines input size, hidden layer size, number of classifications, number of training rounds, batch size and learning rate;

[0022] S1-4, data preparation: read radar point data and annotated samples, and define the data loader;

[0023] S1-5, Model definition: Define the deep learning model architecture innovatively designed for this project;

[0024] S1-6, Loss function and optimizer: Define the loss function and optimizer, using the cross entropy loss function and Adam optimizer.

[0025] Furthermore, the specific method of step S2 includes the following sub-steps:

[0026] S2-1. Use the distance, azimuth, altitude, Doppler and other parameters of the radar detection point track to achieve point-to-point and point-to-temporary track association;

[0027] S2-2, judging whether the temporary track meets the initial conditions for forming a stable track;

[0028] S2-3. Perform initial confirmation for the temporary track that meets the initial conditions, where the confirmation conditions include speed, heading, Doppler speed, angular speed, variance of distance differential speed and comprehensive initial quality;

[0029] S2-4. Confirm that the successful temporary track is started as a stable track and used as input for network training.

[0030] Furthermore, the specific method of step S5 includes the following sub-steps:

[0031] S5-1, load the network model, call model loading, load the model file, automatically parse the network structure and load the network weight data;

[0032] S5-2, receive real-time track data from radar, extract characteristic information such as time, distance, azimuth, altitude, amplitude, signal-to-noise ratio, speed, heading, Doppler speed and undesired track marks of the track start data;

[0033] S5-3, perform network forward reasoning;

[0034] S5-4, calculating the probability that the track is an undesirable track;

[0035] S5-5. Determine whether the initial track is an unexpected track according to the probability threshold.

[0036] The present invention also provides a low-altitude radar undesired track suppression system based on deep learning, comprising:

[0037] Network construction module: establish LSTM neural network and initialize LSTM neural network;

[0038] Data processing module: Uses conventional radar data processing methods to automatically start the target track;

[0039] Dataset construction module: manually mark whether the track is an unexpected track to form labeled data;

[0040] Training module: The LSTM neural network receives the labeled track data online, performs network training, and determines network parameters;

[0041] Reasoning module: The LSTM neural network receives the track data to be identified online and performs neural network reasoning;

[0042] Judgment module: The LSTM neural network outputs an unexpected track sign; if the LSTM output is an unexpected track, the track is invalidated; if it is a real track, the target track is established to complete the unexpected track suppression method.

[0043] Furthermore, the specific method of the network construction module includes the following sub-steps:

[0044] S1-1. The input of LSTM neural network is a two-dimensional vector of the starting point information of the track, which can be expressed as

[0045]

[0046] in:

[0047] n is the number of track starting points, Ti is the time of the ith track point, R i is the distance of the ith track point, A i is the position of the ith track point, H i is the height of the ith track point, Vi is the speed of the ith track point, C i is the heading of the ith track point, Fd i is the Doppler velocity of the ith track point, Amp i is the amplitude of the ith track point, N i is the background amplitude of the i-th track point, 1≤i≤n;

[0048] S1-2. Refer to the typical LSTM network model structure, use a custom time series network model, and select the LSTM combined with Transformer network model for intelligent undesired track suppression;

[0049] S1-3, Hyperparameter settings: defines input size, hidden layer size, number of classifications, number of training rounds, batch size and learning rate;

[0050] S1-4, data preparation: read radar point data and annotated samples, and define the data loader;

[0051] S1-5, Model definition: Define the deep learning model architecture innovatively designed for this project;

[0052] S1-6, Loss function and optimizer: Define the loss function and optimizer, using the cross entropy loss function and Adam optimizer.

[0053] Furthermore, the specific method of the data processing module includes the following sub-steps:

[0054] S2-1. Use the distance, azimuth, altitude, Doppler and other parameters of the radar detection point track to achieve point-to-point and point-to-temporary track association;

[0055] S2-2, judging whether the temporary track meets the initial conditions for forming a stable track;

[0056] S2-3. Perform initial confirmation for the temporary track that meets the initial conditions, where the confirmation conditions include speed, heading, Doppler speed, angular speed, variance of distance differential speed and comprehensive initial quality;

[0057] S2-4. Confirm that the successful temporary track is started as a stable track and used as input for network training.

[0058] Furthermore, the specific method of the reasoning module includes the following sub-steps:

[0059] S5-1, load the network model, call model loading, load the model file, automatically parse the network structure and load the network weight data;

[0060] S5-2, receive real-time track data from radar, extract characteristic information such as time, distance, azimuth, altitude, amplitude, signal-to-noise ratio, speed, heading, Doppler speed and undesired track marks of the track start data;

[0061] S5-3, perform network forward reasoning;

[0062] S5-4, calculating the probability that the track is an undesirable track;

[0063] S5-5. Determine whether the initial track is an unexpected track according to the probability threshold.

[0064] The present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the above method.

[0065] The present invention also provides a processing device, comprising at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the above method by calling the program instructions.

[0066] The advantages of the present invention are:

[0067] The present invention provides a method for suppressing undesired tracks of low-altitude radar based on deep learning. The method transforms the undesired track suppression problem into a binary classification problem, designs an LSTM neural network, uses real-time track data for training and reasoning, and identifies real tracks and undesired tracks. The method has few constraints, has online self-learning capabilities, short learning time, and fast reasoning speed, and meets the real-time target tracking requirements of radar. According to actual data testing, the probability of undesired track recognition of this method is greater than 98%, and the network has strong generalization ability. It is an efficient method for suppressing undesired tracks of low-altitude radar. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a flow chart of a method for suppressing undesired low-altitude radar tracks based on deep learning in an embodiment of the present invention;

[0069] Figure 2 It is a deep learning model architecture for suppressing undesired flight paths of low-altitude radar in an embodiment of the present invention;

[0070] Figure 3 It is an application architecture for suppressing undesired flight paths of low-altitude radar based on deep learning in an embodiment of the present invention;

[0071] Figure 4-1 and Figure 4-2 A comparison chart is shown for the tracks before and after the undesired tracks are suppressed using this method. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0073] Build an LSTM network, collect track training data in real time, combine manual annotation or true value data to form labeled data, and use it as the input of the LSTM network for training to obtain a deep learning model that meets the accuracy requirements.

[0074] like Figure 1 As shown, the low-altitude radar undesired track suppression method based on deep learning includes the following steps:

[0075] S1. Establish an LSTM neural network, determine the number of network input nodes, hidden layers, hidden layer nodes, output nodes and activation function, and initialize the LSTM network;

[0076] S2, using conventional radar data processing methods to automatically start the target track;

[0077] S3, manually marking whether the track is an unexpected track, and forming marking data;

[0078] S4, LSTM neural network receives the labeled track data online, performs network training, and determines network parameters;

[0079] S5, LSTM neural network receives the track data to be identified online and performs neural network reasoning;

[0080] S6, LSTM neural network outputs an unexpected track mark; if the LSTM output is an unexpected track, the track is invalidated, if it is a real track, the target track is established, and the unexpected track suppression method is completed.

[0081] Furthermore, the specific method of step S1 includes the following sub-steps:

[0082] S1-1. The input of LSTM neural network is a two-dimensional vector of the starting point information of the track, which can be expressed as

[0083]

[0084] in:

[0085] n is the number of track starting points (usually 3 to 5), Ti is the time of the ith track point, R i is the distance of the ith track point, A i is the position of the ith track point, H i is the height of the ith track point, Vi is the speed of the ith track point, C i is the heading of the ith track point, Fd i is the Doppler velocity of the ith track point, Amp i is the amplitude of the ith track point, N i is the background amplitude of the i-th track point, 1≤i≤n.

[0086] S1-2. Refer to the typical LSTM network model structure, use a custom time series network model, and select the LSTM combined with Transformer network model for intelligent unexpected track suppression

[0087] S1-3, Hyperparameter settings: defines input size, hidden layer size, number of classifications, number of training rounds, batch size and learning rate;

[0088] S1-4, data preparation: read radar point data and annotated samples, and define the data loader;

[0089] S1-5, Model definition: Define the deep learning model architecture innovatively designed for this project;

[0090] S1-6, Loss function and optimizer: Define the loss function and optimizer, using the cross entropy loss function and Adam optimizer.

[0091] The 9 data of the track start data, including time, distance, azimuth, altitude, amplitude, signal-to-noise ratio, speed, heading, and Doppler speed, are used as input for training and reasoning. Assuming that the target start frame number is n, the input of the deep learning model is n*9. For each set of track start data, the probability that the track corresponding to the set of start data is the expected track is output. Therefore, the output node is 1, and the network model is as follows: Figure 2 shown.

[0092] Furthermore, the specific method of step S2 includes the following sub-steps:

[0093] S2-1. Use the distance, azimuth, altitude, Doppler and other parameters of the radar detection point track to achieve point-to-point and point-to-temporary track association;

[0094] S2-2, judging whether the temporary track meets the initial conditions for forming a stable track;

[0095] S2-3. Perform initial confirmation for the temporary track that meets the initial conditions, the confirmation conditions include speed, heading, Doppler speed, angular speed, variance of distance differential speed and comprehensive initial quality, etc.;

[0096] S2-4. Confirm that the successful temporary track is started as a stable track and used as input for network training.

[0097] Furthermore, the specific method of step S5 includes the following sub-steps:

[0098] S5-1, load the network model, call model loading, load the model file, automatically parse the network structure and load the network weight data;

[0099] S5-2, receive real-time track data from radar, extract characteristic information such as time, distance, azimuth, altitude, amplitude, signal-to-noise ratio, speed, heading, Doppler speed and undesired track marks of the track start data;

[0100] S5-3, perform network forward reasoning;

[0101] S5-4, calculating the probability that the track is an undesirable track;

[0102] S5-5. Determine whether the initial track is an unexpected track according to the probability threshold.

[0103] In the specific implementation process, this method uses a Linux system based on Centos7.0, the programming language used is C++, and the open source software framework tensorflow framework is used to build a deep learning model for training. In this implementation process, this method uses actual detection data of a certain type of low-altitude radar to test the accuracy of the undesired track suppression method.

[0104] Using a 5-point track as the starting point, each point includes 9 parameters, and the input parameters of the deep learning network are 5*9=45;

[0105] 10 hidden layers, 1000 neurons in each layer;

[0106] 1 output node;

[0107] 1000 training samples;

[0108] The number of training times is 20,000;

[0109] There are 100 test samples, including 50 expected tracks and 50 unexpected tracks.

[0110] The test results of the unexpected track suppression method of the present invention are as follows: 51 batches of expected tracks, 49 batches of unexpected tracks, and an unexpected track suppression rate of 98%.

[0111] The overall software architecture of this method is as follows Figure 3 shown. Figure 3 It is a deep learning-based low-altitude radar undesired track suppression application architecture. Figure 4-1 and Figure 4-2 The track comparison diagram before and after using this method to suppress undesired tracks. The orange in the figure is the original point track detected by the radar. Figure 4-1 The white track in the middle is the car track after comparison with the road map, and the blue track includes the expected target track and the unexpected track; Figure 4-2 It is the target track processed by the method of the present invention (compared with the true value data, it is the real flying UAV target). After being processed by the method of the present invention, all the undesired targets such as flying birds, ground cars, weather and fairy waves are suppressed.

[0112] In summary, a method for suppressing undesired tracks of low-altitude radar based on deep learning transforms the problem of suppressing undesired tracks into a binary classification problem, designs an LSTM neural network, uses real-time track data for training and reasoning, and identifies real tracks and undesired tracks. The method is divided into two stages: training and reasoning. In the training stage, conventional methods are used to automatically start the track, and the track start data is used as the learning set of the LSTM network for training to determine the network parameters; in the reasoning stage, the track is used as input for reasoning, and the LSTM network outputs the undesired track mark. If it is a real track, the target track is established in the radar, otherwise it is marked as an undesired track. The method has few constraints, has online self-learning ability, short learning time, fast reasoning speed, and meets the real-time target tracking needs of radar. According to actual data testing, the probability of undesired track recognition of this method is greater than 98%, and the network has strong generalization ability. It is an efficient method for suppressing undesired tracks of low-altitude radars and can be widely used in low-altitude radars.

[0113] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for suppressing undesired low-altitude radar tracks based on deep learning, characterized in that: The following steps are involved: S1. Establish an LSTM neural network and initialize the LSTM neural network; S2, using conventional radar data processing methods to automatically start the target track; S3, manually marking whether the track is an unexpected track, and forming marking data; S4, LSTM neural network receives the labeled track data online, performs network training, and determines network parameters; S5, LSTM neural network receives the track data to be identified online and performs neural network reasoning; S6, LSTM neural network outputs an unexpected track mark; if the LSTM output is an unexpected track, the track is invalidated, if it is a real track, the target track is established, and the unexpected track suppression method is completed.

2. The method for suppressing undesired low-altitude radar tracks based on deep learning according to claim 1 is characterized in that: The specific method of step S1 includes the following sub-steps: S1-1. The input of LSTM neural network is a two-dimensional vector of the starting point information of the track, which can be expressed as in: n is the number of starting points of the track, T i is the time of the ith track point, R i is the distance of the ith track point, A i is the position of the ith track point, H i is the height of the ith track point, V i is the velocity of the ith track point, C i is the heading of the ith track point, Fd i is the Doppler velocity of the ith track point, Amp i is the amplitude of the ith track point, N i is the background amplitude of the i-th track point, 1≤i≤n; S1-2. Refer to the typical LSTM network model structure, use a custom time series network model, and select the LSTM combined with Transformer network model for intelligent undesired track suppression; S1-3, Hyperparameter settings: defines input size, hidden layer size, number of classifications, number of training rounds, batch size and learning rate; S1-4, data preparation: read radar point data and annotated samples, and define the data loader; S1-5, Model definition: Define the deep learning model architecture innovatively designed for this project; S1-6, Loss function and optimizer: Define the loss function and optimizer, using the cross entropy loss function and Adam optimizer.

3. The method for suppressing undesired low-altitude radar tracks based on deep learning according to claim 1 is characterized in that: The specific method of step S2 includes the following sub-steps: S2-1. Use the distance, azimuth, altitude, Doppler and other parameters of the radar detection point track to achieve point-to-point and point-to-temporary track association; S2-2, judging whether the temporary track meets the initial conditions for forming a stable track; S2-3. Perform initial confirmation for the temporary track that meets the initial conditions, where the confirmation conditions include speed, heading, Doppler speed, angular speed, variance of distance differential speed and comprehensive initial quality; S2-4. Confirm that the successful temporary track is started as a stable track and used as input for network training.

4. The method for suppressing undesired low-altitude radar tracks based on deep learning according to claim 1, characterized in that: The specific method of step S5 includes the following sub-steps: S5-1, load the network model, call model loading, load the model file, automatically parse the network structure and load the network weight data; S5-2, receive real-time track data from radar, extract characteristic information such as time, distance, azimuth, altitude, amplitude, signal-to-noise ratio, speed, heading, Doppler speed and undesired track marks of the track start data; S5-3, perform network forward reasoning; S5-4, calculating the probability that the track is an undesirable track; S5-5. Determine whether the initial track is an unexpected track according to the probability threshold.

5. The low-altitude radar undesired track suppression system based on deep learning is characterized by: include: Network construction module: establish LSTM neural network and initialize LSTM neural network; Data processing module: Use conventional radar data processing methods to automatically start the target track; Dataset construction module: manually mark whether the track is an unexpected track to form labeled data; Training module: The LSTM neural network receives the labeled track data online, performs network training, and determines network parameters; Reasoning module: The LSTM neural network receives the track data to be identified online and performs neural network reasoning; Judgment module: The LSTM neural network outputs an unexpected track sign; if the LSTM output is an unexpected track, the track is invalidated; if it is a real track, the target track is established to complete the unexpected track suppression method.

6. The low-altitude radar undesired track suppression system based on deep learning according to claim 5 is characterized in that: The specific method of the network construction module includes the following sub-steps: S1-1. The input of LSTM neural network is a two-dimensional vector of the starting point information of the track, which can be expressed as in: n is the number of starting points of the track, T i is the time of the ith track point, R i is the distance of the ith track point, A i is the position of the ith track point, H i is the height of the ith track point, V i is the velocity of the ith track point, C i is the heading of the ith track point, Fd i is the Doppler velocity of the ith track point, Amp i is the amplitude of the ith track point, N i is the background amplitude of the i-th track point, 1≤i≤n; S1-2. Refer to the typical LSTM network model structure, use a custom time series network model, and select the LSTM combined with Transformer network model for intelligent undesired track suppression; S1-3, Hyperparameter settings: defines input size, hidden layer size, number of classifications, number of training rounds, batch size and learning rate; S1-4, data preparation: read radar point data and annotated samples, and define the data loader; S1-5, Model definition: Define the deep learning model architecture innovatively designed for this project; S1-6, Loss function and optimizer: Define the loss function and optimizer, using the cross entropy loss function and Adam optimizer.

7. The low-altitude radar undesired track suppression system based on deep learning according to claim 5 is characterized in that: The specific method of the data processing module includes the following sub-steps: S2-1. Use the distance, azimuth, altitude, Doppler and other parameters of the radar detection point track to achieve point-to-point and point-to-temporary track association; S2-2, judging whether the temporary track meets the initial conditions for forming a stable track; S2-3. Perform initial confirmation for the temporary track that meets the initial conditions, where the confirmation conditions include speed, heading, Doppler speed, angular speed, variance of distance differential speed and comprehensive initial quality; S2-4. Confirm that the successful temporary track is started as a stable track and used as input for network training.

8. The low-altitude radar undesired track suppression system based on deep learning according to claim 5 is characterized in that: The specific method of the reasoning module includes the following sub-steps: S5-1, load the network model, call model loading, load the model file, automatically parse the network structure and load the network weight data; S5-2, receive real-time track data from radar, extract characteristic information such as time, distance, azimuth, altitude, amplitude, signal-to-noise ratio, speed, heading, Doppler speed and undesired track marks of the track start data; S5-3, perform network forward reasoning; S5-4, calculating the probability that the track is an undesirable track; S5-5. Determine whether the initial track is an unexpected track according to the probability threshold.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the method according to any one of claims 1 to 4.

10. A processing device, characterized in that: The method comprises at least one processor and at least one memory in communication with the processor, wherein the memory stores program instructions executable by the processor, and the processor can execute the method according to any one of claims 1 to 4 by calling the program instructions.

Images