A radar scene-based deep neural network target recognition method

Abstract

The application discloses a radar scene-based deep neural network target identification method and belongs to the technical field of radar target identification.The application comprises the following steps: obtaining identification attribute information of a target, obtaining trajectory data of an attack execution object, establishing a static RCS database based on electromagnetic simulation of a target model, establishing a radar echo database, pre-processing radar echo data, training neural network parameters based on the pre-processed data, and constructing a neural network model for target identification.In the application, when a target is attacked, the radar echo sequence of the target is obtained through target electromagnetic simulation and radar distance factors, and the target is identified by using a deep neural network based on echo sequence data samples, so that the identification problem of an electrically large target under near-field conditions is solved.

Description

Technical Field

[0001] This invention belongs to the field of radar target recognition technology, specifically relating to a deep neural network target recognition method based on radar scenes. Background Technology

[0002] Throughout the missile's attack on land and sea targets (such as electrically large targets like surface ships), the radar signals received by the missile's seeker radar are complex and variable, influenced by numerous factors including the target's own electromagnetic characteristics, the relative motion between the missile and the target, and radar parameters. This poses a challenge to the missile's success rate in terminal target engagement. Existing radar target identification mainly relies on radar imaging, the target's micro-motion characteristics during motion, and dynamic RCS (radar cross-section) time series. These methods, after undergoing radar signal processing algorithms, often lose some of the target's electromagnetic characteristics. Deep learning neural networks (such as CNN convolutional neural networks) possess powerful feature learning capabilities, autonomously learning and extracting radar echo features without manual feature extraction, thus reducing feature information loss. Currently, most neural network target identification technologies based on dynamic RCS time series do not consider the impact of radar distance under different missile trajectory states. Throughout the missile's attack on the target, the distance between the seeker and the target has a significant impact on the radar echo. Therefore, it is necessary to propose an artificial intelligence-based radar identification technology, which starts from the target radar echo signal and, based on artificial intelligence technology, establishes a target identification technology that addresses the complex electromagnetic characteristics of missiles attacking (environmental) targets under different motion trajectories throughout the entire process. Summary of the Invention

[0003] This invention provides a deep neural network target recognition method based on radar scenes, which can be used to improve the recognition performance of targets in near-field conditions.

[0004] The technical solution adopted in this invention is as follows:

[0005] A deep neural network target recognition method based on radar scenes, the method comprising:

[0006] Step 1: Obtain the identification attribute information of the attack target, including the target's location information, the operating parameters of the target detection radar, and the location information of the attack execution object and the target; wherein, the location information of the attack execution object and the target includes: the elevation angle, azimuth angle and distance relative to the target;

[0007] Step 2: Based on the relative motion information between the attack execution object and the attack target, simulate multiple trajectory data r of the attack execution object. i (t), where t represents time and i represents the trajectory separator;

[0008] Step 3: Establish a static RCS (Radar Cross Section) database based on electromagnetic simulation of the target model:

[0009] Different types of target models to be identified are imported into modeling software for mesh generation.

[0010] Electromagnetic scattering simulation tools were used to simulate and calculate the target RCS in each direction under far-field conditions for each type of target model: within the range of pitch angle θ and azimuth angle φ, the pitch angle θ and azimuth angle φ were discretized into multiple discrete angles according to the specified angle step, and the discrete pitch angle θ and azimuth angle φ were combined in pairs, and the RCS value corresponding to each group (θ,φ) was simulated to establish a database of static RCS of each target under far-field conditions;

[0011] Step 4, establish a radar echo database:

[0012] For each trajectory data r i (t) Discrete sampling is performed at certain time steps to obtain the position sequence r of the attack target. i (n), and according to the position sequence r i (n) Calculate the corresponding radar range R(n), where n = 1, 2, 3...j, and j represents the number of sampling points;

[0013] Based on the static RCS database established in step 3, the position sequence r is estimated through interpolation. i The RCS value of each sampling point of (n) is used to obtain each position sequence r. i The RCS time series σ(n) of (n);

[0014] according to The target radar echo sequence X(n) at different radar distances is obtained, and a radar echo database is obtained based on the radar echo sequences of different types of targets;

[0015] Step 5: Perform data preprocessing on the radar echo database to obtain the sample dataset:

[0016] Set the target type label y for each target radar echo sequence X(n) in the radar echo database;

[0017] The radar echo sequence X(n) of each target is standardized separately;

[0018] The target type label y of each target radar echo sequence X(n) is normalized, and then the normalized target type label is one-hot encoded.

[0019] Step 6: Construct and train the neural network model for target recognition:

[0020] The neural network model for target recognition sequentially comprises: a first stacked network consisting of four convolutional pooling blocks, a second stacked network consisting of two convolutional blocks, a convolutional layer, a third stacked network consisting of two convolutional blocks, a global pooling layer, a Dropout layer, and a fully connected layer with a Softmax function; wherein, the convolutional pooling block sequentially comprises a convolutional layer, a batch normalization layer, an activation function mapping, and a max pooling layer; the convolutional block sequentially comprises a convolutional layer, a batch normalization layer, and an activation function mapping;

[0021] The neural network model for target recognition is trained based on the sample dataset and a preset loss function. When the preset training termination condition is met, the target type is identified based on the radar echo data of the target to be identified using the trained neural network model.

[0022] The technical solution provided by this invention brings at least the following beneficial effects:

[0023] In this invention, when a target is attacked, the radar echo sequence of the target is obtained through target electromagnetic simulation and radar distance factors. Based on these echo sequence data samples, a deep neural network is used to identify the target, thereby solving the identification problem of electrically large targets (the physical size of the target is much larger than the wavelength of the electromagnetic waves emitted by the radar) in the near field. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of a deep neural network target recognition method based on radar scenes provided by an embodiment of the present invention;

[0026] Figure 2 This is a schematic diagram of the target in an embodiment of the present invention;

[0027] Figure 3 These are schematic diagrams of target models for different target types used in embodiments of the present invention;

[0028] Figure 4 This is a schematic diagram of the neural network model structure used in the embodiments of the present invention;

[0029] Figure 5 This refers to the recognition accuracy of the method in the embodiments of the present invention on the training set and the validation set. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0031] like Figure 1 As shown, the deep neural network target recognition method based on radar scenes provided by this embodiment of the invention includes the following steps:

[0032] Step 1, obtain the target's identification attribute information:

[0033] The acquired identification attribute information includes the target's position information, the target detection radar's operating parameters, and the position information of the attack execution object and the target, including: the elevation angle, horizontal azimuth angle, and distance relative to the target;

[0034] like Figure 2 As shown, the ground surface of the target (e.g., the target of a missile attack, or the sea surface of the target for a maritime target) is located in the XOY coordinate plane, with its average height at Z=0. Here, θ represents the pitch angle of the attacking object relative to the target, and φ represents the azimuth angle (horizontal azimuth) of the attacking object relative to the target. The target is located above a rough surface; generally, both the rough surface and the target's electrical properties can be considered as good metal. Typical scenarios include: surface ships, ground tanks, etc. That is, in this embodiment of the invention, it is assumed that the target is on... Figure 2 In the scenario shown, the target is located at the origin. The attack target (missile) can attack the target from all sides along different trajectories. The seeker radar emits electromagnetic waves into space and receives the radar echo reflected from the target. The radar operates at frequency f and wavelength λ. The seeker's elevation angle relative to the target is θ(t), its horizontal azimuth angle is φ(t), and its range is R(t), where t is a specific moment. Based on this scenario, the radar echo data of the target is simulated.

[0035] Step 2, obtain the trajectory data of the attack execution target:

[0036] The seeker's trajectory is obtained based on the relative motion information between the missile and the target. i (t), i = 1, 2, ..., L, with L ballistic trajectories, r i (t) represents the position of the seeker at different times t in the i-th trajectory. In this embodiment, a total of 600 ballistic trajectories attacking the target from different directions were simulated.

[0037] Step 3: Establish a static RCS database based on electromagnetic simulation of the target model:

[0038] First, different target models (in this example, such as...) Figure 3 The six ships were imported into modeling software for mesh generation;

[0039] Secondly, electromagnetic scattering simulation tools (such as ray tracing) are used to simulate and calculate the target RCS (Radar Cross Section) for each target model under far-field conditions in various directions, i.e., a set of (θ, φ) corresponding to one RCS value (Radar Cross Section). The elevation angle θ ∈ [0°, 90°], with a value interval of 0.1°; the azimuth angle φ ∈ [0°, 360°], with a value interval of 1°. This establishes a database of the static RCS for each target under far-field conditions.

[0040] Within the range of pitch and azimuth angles, the pitch and azimuth angles are discretized into multiple discrete angles according to a specified angular step, and then combined in pairs. The RCS value corresponding to each pair (θ, φ) is simulated to establish a database of the static RCS of each target under far-field conditions.

[0041] Step 4, establish a radar echo database:

[0042] First, the trajectory r of the seeker. i (t) is discretely sampled at certain time steps, and the ballistic trajectory is obtained as sequence r after discrete sampling. i (n), calculate the corresponding radar range R(n), where n=1,2,3...j, j is the number of sampling points;

[0043] Secondly, interpolation is performed using a static RCS database under far-field conditions to obtain the RCS time series. The interpolation process involves filling in missing data using approximate data; that is, the target RCS corresponding to a point on the trajectory is approximated by data from the database. This yields the RCS at each position r on each trajectory. i The σ(n) corresponding to (n) is the RCS time series σ(n);

[0044] Finally, from the formula The target radar echo sequence X(n) at different radar distances is obtained. In this embodiment, the number of sampling points j for each trajectory is 400, that is, the radar echo sequence has a sequence length of 400. A radar echo database is obtained based on the radar echo sequences of different targets.

[0045] Step 5, Radar echo data preprocessing:

[0046] First, different labels (X(n), y) are used for the echo sequences of different targets, where y = 1, 2, 3...k (k is the number of target categories). That is, the echo dataset contains the mapping relationship between echo sequence data and their corresponding target categories. The sequence data with the mapping relationship and their corresponding target types are used as a group of samples.

[0047] Secondly, the echo data in the radar echo database is standardized using the following formula:

[0048]

[0049] Among them, X std This is the standardized data. mean(X) is the average value of X, std(X) is the standard deviation of X, and X is the value of the echo sequence X(n). The label y-values ​​are then normalized using the following formula:

[0050]

[0051] Where Y is the normalized value, y min y max These are the minimum and maximum values ​​of y, respectively.

[0052] Then Y is converted to one-hot encoding. One-hot encoding, also known as one-bit valid encoding, uses an m-bit state register to encode m states. Each state has its own independent register bit, and at any given time, only one bit is valid.

[0053] After the above preprocessing of the radar echo data, a sample dataset is obtained. The data samples in the sample dataset are divided into training set, validation set, and test set according to a certain ratio. The training set and validation set are used for network training, and the test set is used for testing.

[0054] In this embodiment, the echo sequences of the six targets are labeled with corresponding categories y = 1, 2, 3... 6. The 600 (number of ballistic trajectories) * 6 (number of target models) = 3600 data samples of the six targets are standardized and normalized, and randomly divided into training set, validation set and test set in a ratio of 8:1:1.

[0055] Step 6: Train the neural network parameters and construct a neural network model for target recognition:

[0056] This embodiment uses a convolutional neural network (CNN), such as Figure 4 As shown in the embodiment of the present invention, the constructed target recognition neural network model sequentially includes: a first stacked network composed of four convolutional pooling blocks, a second stacked network composed of two convolutional blocks, a convolutional layer, a third stacked network composed of two convolutional blocks, a global pooling layer (Global Average Pooling), a Dropout layer, and a fully connected layer with a Softmax function; wherein, the convolutional pooling block sequentially includes a convolutional layer, a batch normalization layer, an activation function mapping, and a max pooling layer; the convolutional block sequentially includes a convolutional layer, a batch normalization layer, and an activation function mapping.

[0057] In this invention, the target recognition neural network comprises nine convolutional layers. Following the forward propagation direction, the number of kernels and kernel sizes for each layer are as follows: 64 and 32×1, 128 and 24×1, 256 and 16×1, 512 and 16×1, 1024 and 12×1, 512 and 8×1, 256 and 6×1, 128 and 4×1, and 64 and 2×1. The pooling window sizes for each max-pooling layer are set to 16, 12, 12, and 12, respectively.

[0058] The neural network model for target recognition is trained based on the sample dataset and a preset loss function. When the preset training termination condition is met, the target type is identified based on the radar echo data of the target to be identified using the trained neural network model.

[0059] In this embodiment of the invention, a Dropout layer is added to prevent overfitting, while the Global Average Pooling layer averages the output of the feature maps globally, transforming a W*H*D tensor into a 1*1*D tensor. For example, if the tensor output before average pooling is 8*8*1024, performing average pooling (taking an average value) on each 8*8 feature map yields 1024 scalars. The purpose is to minimize overfitting by reducing the total number of parameters in the model. Except for the last fully connected layer, which uses the Softmax function, the ReLU function is used after each convolutional layer. The loss function is set to the Cross Entropy Loss Function. The function expression is as follows:

[0060]

[0061] Where N is the number of training samples, and y is the label of the real sample. This refers to the category predicted by the network. The optimizer is set to Adam, and an initial learning rate is set. The training set data is input into the convolutional neural network to train it. The backpropagation algorithm is used to continuously reduce the loss function value, making the network output value increasingly closer to the actual value. Then, by training networks with different numbers of layers and structural parameters such as convolutional kernels, the impact of these structural parameters on the neural network performance is analyzed to obtain the optimal network structure, i.e., the optimal number of neurons (convolutional kernels) and the optimal number of layers in each layer of the convolutional neural network. In this example, the final training optimization is as follows: Figure 4 Network structure model. For example... Figure 5 As shown, after training 800 times, the recognition accuracy of this convolutional neural network model is 99% on the training set, approximately 93.5% on the validation set, and 93.2% on the test set.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0063] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A deep neural network target recognition method based on radar scenes, characterized in that, Includes the following steps: Step 1: Obtain the identification attribute information of the attack target, including the target's location information, the operating parameters of the target detection radar, and the location information of the attack execution object and the target; wherein, the location information of the attack execution object and the target includes: the elevation angle, azimuth angle and distance relative to the target; Step 2: Based on the relative motion information between the attack execution object and the attack target, simulate multiple trajectory data of the attack execution object. Where t represents time and i represents the trajectory separator; Step 3: Establish a static radar cross section (RCS) database based on electromagnetic simulation of the target model: Different types of target models to be identified are imported into modeling software for mesh generation. Electromagnetic scattering simulation tools were used to simulate and calculate the target RCS in each direction under far-field conditions: at elevation angles... and azimuth Within the range of values, the pitch angle is adjusted according to the specified angle steps. and azimuth Discretized into multiple discrete angles, for each discrete pitch angle and azimuth Perform pairwise combinations and simulate each pair ( The corresponding RCS value is used to establish a database of static RCS for each target under far-field conditions; Step 4, establish a radar echo database: For each trajectory data Discrete sampling is performed at certain time intervals to obtain the location sequence of the attack target. And according to the position sequence Calculate the corresponding radar range ,in , Indicates the number of sampling points; Based on the static RCS database established in step 3, the position sequence is estimated through interpolation. The RCS value of each sampling point is used to obtain each position sequence. RCS time series ; according to Obtain target radar echo sequences at different radar distances A radar echo database is constructed based on radar echo sequences of different types of targets; Step 5: Perform data preprocessing on the radar echo database to obtain the sample dataset: Set the radar echo sequence of each target in the radar echo database. Target type label ; Radar echo sequences of each target Standardization processing is required; radar echo sequences of each target Target type label Normalization is performed, and then one-hot encoding is performed on the normalized target type labels; Step 6: Construct and train the neural network model for target recognition: The neural network model for target recognition sequentially comprises: a first stacked network consisting of four convolutional pooling blocks, a second stacked network consisting of two convolutional blocks, a convolutional layer, a third stacked network consisting of two convolutional blocks, a global pooling layer, a Dropout layer, and a fully connected layer with a Softmax function; wherein, the convolutional pooling block sequentially comprises a convolutional layer, a batch normalization layer, an activation function mapping, and a max pooling layer; the convolutional block sequentially comprises a convolutional layer, a batch normalization layer, and an activation function mapping; The neural network model for target recognition is trained based on the sample dataset and a preset loss function. When the preset training termination condition is met, the target type is identified based on the radar echo data of the target to be identified using the trained neural network model.

2. The method as described in claim 1, characterized in that, In step 3, the pitch angle The angle step is set to 0.1°, and the azimuth angle is... The angle step is set to 1°.

3. The method as described in claim 1, characterized in that, In step 5, the target radar echo sequence The standardization process specifically involves: ； in, Represents the target radar echo sequence Data before standardization This represents the standardized data. Represents a sequence The average value, Represents a sequence The standard deviation.

4. The method as described in claim 1, characterized in that, In step 5, the radar echo sequences of each target Target type label The normalization process is as follows: ； in, This represents the normalized value. , These are the target type labels. The minimum and maximum values.

5. The method as described in claim 1, characterized in that, In step 6, the cross-entropy loss function is used.

Images