A radar single-multi-target scene classification method based on complex manifold neural network
By using complex manifold neural network to classify radar single-multi-target scenarios in the radar system, the problem that traditional methods are difficult to improve classification accuracy under low signal-to-noise ratio is solved, and higher classification accuracy and calculation pressure relief effect are achieved.
Patent Information
- Application Number
- CN202411815669.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-11
- Publication Date
- 2025-04-04
AI Technical Summary
When the target signal-to-noise is relatively low, traditional radar scene classification methods are difficult to improve classification accuracy, which affects the calculation pressure reduction effect of the radar system.
The radar single-multi-objective scene classification method based on complex manifold neural network is used to determine whether there are multiple targets in the radar echo time series by building complex manifold neural network, building training data sets, training and deploying the network.
It effectively enhances the model's ability to mine information from complex data, improves classification accuracy, and reduces the computing pressure of the radar system.
Smart Images
Figure CN119758282A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of radar information processing and is a method for implementing radar target scene classification. Background Art
[0002] In the application process of new radar, a variety of long-time pulse accumulation algorithms are generally used. For single-target echo and multi-target echo, the complexity of the accumulation algorithm is different. In order to reduce the pressure of subsequent processing, it is necessary to predict whether there are multiple targets in the target echo before using the long-time pulse accumulation algorithm, so as to reduce the calculation pressure of the radar system. However, when the target signal-to-noise ratio is low, it is difficult for traditional scene classification methods to improve classification accuracy, which affects the practical application of the prediction method. Summary of the invention
[0003] In order to solve the problems existing in the prior art, the present invention proposes a radar single-multiple target scene classification method based on complex manifold neural network, which is used to determine the number and type of hidden targets in a specific radar echo sequence under low signal-to-noise ratio conditions. In view of the high complexity of some long-time pulse accumulation algorithms, the present invention pre-judges whether there are multiple targets in the target echo before using the long-time pulse accumulation algorithm. By building a complex manifold neural network, constructing a training data set, training and deploying the network, the complex manifold neural network is used to determine whether there are multiple targets in the radar echo time series, providing a reference for the selection of subsequent algorithms, thereby achieving the purpose of reducing the computational pressure of the radar system.
[0004] The present invention proposes a radar single-multi-target scene classification method based on complex manifold neural network, which uses complex manifold neural network to determine whether there are multiple targets in the radar echo time series. The technical solution adopted includes:
[0005] Step 1: Build a complex manifold neural network;
[0006] Step 2: Construct a training data set through actual testing or simulation methods;
[0007] Step 3: Use the back propagation algorithm to train the complex manifold neural network on the training data set;
[0008] Step 4: Deploy the trained complex manifold neural network model.
[0009] Furthermore, the complex manifold neural network uses the complex manifold distance defined by the following formula instead of the classic Euclidean distance:
[0010]
[0011] Among them, z1 and z2 are complex data, and:
[0012]
[0013] Where F is the complex plane To the manifold space R + ×SO(2) bijective mapping, where R + is a positive real manifold, SO(2) is a plane rotation manifold, θ is the argument principal value of z, and the matrix logarithm logm is defined by expm(·) in the following formula:
[0014]
[0015] Furthermore, the complex manifold neural network uses the weighted Frechet average defined by the following formula to replace the convolution operator in the classical neural network:
[0016]
[0017] In the formula, {w k} is a set of weight factors to be trained, z k is the data for which the weighted Frechet mean is to be calculated, and m is the complex plane. A little bit on top.
[0018] Furthermore, the core components of the complex manifold network include: feature extraction module, distance transformation module, classical feature mining and classification module.
[0019] Furthermore, the feature extraction module is composed of a plurality of weighted Frechet average operator layers or a combination layer of a weighted Frechet average operator layer and an activation function. The steps of the distance transformation module include:
[0020] (1) Calculate the weighted Frechet mean of all eigenvalues of a specific channel, and the weight factor is to be learned;
[0021] (2) Calculate the manifold distance from all eigenvalues of a specific channel to the weighted Frechet mean, and replace the eigenvalue with the distance.
[0022] Furthermore, the radar echo time series in step 1 is radar echo data along the fast time dimension or the slow time dimension, and the data is in the form of radar IQ data containing real and imaginary parts, and there are single or multiple targets and background disturbances in the data. The neural network in step 2 is structurally a one-dimensional convolutional neural network, the input is the real part, imaginary part or complex data of the radar echo time series, and the output is a classifier structure, which gives a classification result of whether the echo belongs to a multi-target scene.
[0023] The beneficial effects of the present invention include:
[0024] (1) Applying complex manifold theory to the classification of single and multiple target scenes in electromagnetic signals can effectively enhance the model's ability to mine information from complex data;
[0025] (2) The size of the neural network used is limited, which can provide support for the selection of subsequent algorithms at a relatively low computing resource cost. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 It is a schematic diagram of the flow chart of the preferred embodiment of the present invention. DETAILED DESCRIPTION
[0027] The present invention is further described below, but the protection scope of the present invention is not limited by the implementing regulations.
[0028] The present invention proposes a radar single-multi-target scene classification method based on complex manifold neural network to realize radar small target scene classification. The preferred implementation process includes:
[0029] Step 1: Build a complex manifold neural network;
[0030] Step 2: Construct a training data set through actual testing or simulation methods;
[0031] Step 3: Use the back propagation algorithm to train the complex manifold neural network on the training data set;
[0032] Step 4: Deploy the trained complex manifold neural network model.
[0033] Complex manifold neural networks use the complex manifold distance defined as follows instead of the classic Euclidean distance:
[0034]
[0035] Among them, z1 and z2 are complex data, and:
[0036]
[0037] Where F is the complex plane To the manifold space R + ×SO(2) bijective mapping, where R + is a positive real manifold, SO(2) is a plane rotation manifold, θ is the argument principal value of z, and the matrix logarithm logm is defined by expm(·) in the following formula:
[0038]
[0039] The complex manifold network replaces the convolution operator in the classical neural network with a weighted Frechet average defined as follows:
[0040]
[0041] In the formula, {w k} is a set of weight factors to be trained, z k is the data for which the weighted Frechet mean is to be calculated, and m is the complex plane. A little bit on top.
[0042] The core components of the complex manifold network include: feature extraction module, distance transformation module, classical feature mining and classification module. Among them, the feature extraction module is composed of multiple weighted Frechet average operator layers or a combination of weighted Frechet average operator layers and activation functions. The steps of the distance transformation module are as follows:
[0043] (1) Calculate the weighted Frechet mean of all eigenvalues of a specific channel, and the weight factor is to be learned;
[0044] (2) Calculate the manifold distance from all eigenvalues of a specific channel to the weighted Frechet mean, and replace the eigenvalue with the distance.
[0045] The above radar echo time series is radar echo data along the fast time dimension or the slow time dimension, and the data is in the form of radar IQ data including real and imaginary parts, and there are single or multiple targets and background disturbances in the data.
[0046] The neural network is structurally a one-dimensional convolutional neural network. The input is the real part, imaginary part or complex data of the radar echo time series, and the output is a classifier structure that gives the classification result of whether the echo belongs to a multi-target scene.
[0047] According to the above description: the classification network structure used in this embodiment includes:
[0048] (1) Three-layer weighted Frechet average;
[0049] (2) a distance transformation;
[0050] (3) The classification module consists of two layers of real convolution and two layers of fully connected layers.
[0051] After testing, on simulation data with a signal-to-noise ratio lower than -5dB before compression, the classification accuracy of the traditional real network is only 50%, indicating that the real network cannot learn effective information from the data; by using the network described in this embodiment, the classification accuracy can be increased to more than 80%.
Claims
1. A radar single-multi-target scene classification method based on complex manifold neural network, characterized by: The complex manifold neural network is used to determine whether there are multiple targets in the radar echo time series, including: Step 1: Build a complex manifold neural network; Step 2: Construct a training data set through actual testing or simulation methods; Step 3: Use the back propagation algorithm to train the complex manifold neural network on the training data set; Step 4: Deploy the trained complex manifold neural network model.
2. The radar single-multiple target scene classification method based on complex manifold neural network according to claim 1, characterized in that: Complex manifold neural networks use the complex manifold distance defined as follows instead of the classic Euclidean distance: Among them, z1 and z2 are complex data, and: Where F is the complex plane To the manifold space R + ×SO(2) bijective mapping, where R + is a positive real manifold, SO(2) is a plane rotation manifold, θ is the argument principal value of z, and the matrix logarithm logm is defined by expm(·) in the following formula:
3. The radar single-multiple target scene classification method based on complex manifold neural network according to claim 2, characterized in that: The complex manifold neural network uses the weighted Frechet average defined by the following formula to replace the convolution operator in the classical neural network: In the formula, {w k } is a set of weight factors to be trained, z k is the data for which the weighted Frechet mean is to be calculated, and m is the complex plane. A little bit on top.
4. The radar single-multiple target scene classification method based on complex manifold neural network according to claim 3, characterized in that: The core components of the complex manifold neural network include: feature extraction module, distance transformation module, and classic feature mining and classification module.
5. The radar single-multiple target scene classification method based on complex manifold neural network according to claim 4, characterized in that: The feature extraction module is composed of a plurality of weighted Frechet average operator layers or a combination layer of a weighted Frechet average operator layer and an activation function.
6. The radar single-multiple target scene classification method based on complex manifold neural network according to claim 4, characterized in that: The steps of the distance transformation module include: (1) Calculate the weighted Frechet mean of all eigenvalues of a specific channel, and the weight factor is to be learned; (2) Calculate the manifold distance from all eigenvalues of a specific channel to the weighted Frechet mean, and replace the eigenvalue with the distance.
7. The radar single-multiple target scene classification method based on complex manifold neural network according to claim 1, characterized in that: The radar echo time series in step 1 is radar echo data along the fast time dimension or the slow time dimension, and the data is in the form of radar IQ data including real and imaginary parts, and the radar IQ data contains single or multiple targets and background disturbances.
8. The radar single-multiple target scene classification method based on complex manifold neural network according to claim 1, characterized in that: The neural network in step 2 is structurally a one-dimensional convolutional neural network, the input is the real part, imaginary part or complex data of the radar echo time series, and the output is a classifier structure that gives a classification result of whether the echo belongs to a multi-target scene.