A high-precision and robust radar missing signal completion system
By using an adaptive autoencoder neural network for radar signal completion, the problems of high computational cost, insufficient robustness, and insufficient accuracy in traditional methods are solved, achieving high-precision and high-robust signal completion while reducing computational and time costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2022-12-21
- Publication Date
- 2026-05-26
AI Technical Summary
Traditional radar signal completion methods suffer from high computational and time overhead, as well as insufficient robustness and accuracy, making it difficult to effectively improve performance.
An adaptive autoencoder neural network is used for radar signal completion. Through data preprocessing, autoencoder neural network training, and result display, combined with L1 regularization and relative entropy loss function to suppress overfitting, adaptive adjustment of the encoded data dimension is achieved.
It improves the accuracy and robustness of radar missing signal completion, reduces computational and time overhead, and achieves faster missing signal completion speed.
Smart Images
Figure CN116070076B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of digital signal processing and deep learning algorithms, and in particular to a radar missing signal completion system that combines deep learning. Background Technology
[0002] Radar missing signal completion is a significant and frequently researched topic in signal processing, with wide-ranging applications. Increasingly, international organizations are paying attention to and promoting its development. Traditional radar missing signal completion techniques are based on interpolation algorithms and traditional machine learning algorithms, using information from before and after the missing signal location to estimate the value of the missing signal.
[0003] In traditional radar missing signal completion tasks, the position of the missing signal within the overall radar signal and the values of the signals before and after that position have a crucial impact. However, traditional radar missing signal completion methods require learning for each radar signal, resulting in significant computational and time overhead. Furthermore, interpolation algorithms and traditional machine learning algorithms face bottlenecks in accuracy improvement. Simultaneously, the number of missing signal positions within a radar signal segment also significantly affects the effectiveness of traditional radar missing signal completion methods. Therefore, the main research objective for radar missing signal completion tasks is to improve the robustness of algorithms to different missing signals and the accuracy of signal completion, while simultaneously reducing the computational and time overhead. This is not only of significant theoretical and practical importance but also represents a challenging and hotly debated area in international missing signal completion tasks. Summary of the Invention
[0004] To overcome the problems of poor robustness, low accuracy, and high computational and time costs in traditional radar missing signal completion tasks, the present invention aims to provide a high-precision and robust radar missing signal completion system that can complete different radar missing signals.
[0005] The technical solution adopted by this invention to solve its technical problem is: a high-precision and robust radar missing signal completion system, including a database and a host computer, the database and the host computer being connected sequentially, characterized in that: the database contains pre-collected radar signals, and the host computer includes:
[0006] The data preprocessing module is used to preprocess radar signals stored in the database.
[0007] The missing signal completion module is used to build an adaptive autoencoder neural network to complete radar signals that have been processed by the data preprocessing module and have missing signals.
[0008] Furthermore, the data preprocessing module is completed using the following process:
[0009] Radar signals are standardized using the following formula:
[0010]
[0011] Where, x ij Let x represent the j-th sampling point of the i-th radar signal. s_ij x represents ij The standardized form, μ i σ represents the average value of all sampling points of the i-th radar signal. i Let x represent the standard deviation of all sampling points of the i-th radar signal. s To represent the standardized radar signal.
[0012] Furthermore, the missing signal completion module is completed using the following process:
[0013] Building an autoencoder neural network:
[0014] The basic form of an autoencoder neural network is shown below:
[0015]
[0016] f(x)=σ(W f ·x+b f )
[0017] g(h)=σ(W g ·h+b h )
[0018] Where f(·) represents the encoder function, g(·) represents the decoder function, x represents the input data, h represents the encoded data, x′ represents the reconstructed data, σ represents the activation function, and W f W represents the weight matrix in the encoding function. g b represents the weight matrix in the decoding function. f b represents the bias in the encoding function. h This represents the bias in the decoding function. The dimension of the encoded data h is manually set in the experiment based on the specific data.
[0019] An autoencoder neural network is trained using gradient descent and backpropagation, with the following constraints applied during training:
[0020] x≈x′
[0021] Where x′=g(f(x)). The goal of training is to make the data reconstructed by the decoder as close as possible to the input data.
[0022] Mean squared error is used to measure the similarity between the input data and the reconstructed data:
[0023] L(x, g(f(x)))=E x～x′ ||xg(f(x))|| 2
[0024] Where L(·) represents the loss function, E x～x′ (·) represents the expected value of the squared error between the input data x and the reconstructed data x′.
[0025] Building an adaptive autoencoder neural network:
[0026] 1) Set the initial dimension of the encoded data h to K, and preprocess the radar signal x. s The input is used for training in an autoencoder neural network.
[0027] 2) After the autoencoder neural network is trained, calculate the standard deviation δ = (δ1, δ2, ..., δ) of each dimension of the encoded data h. K ).
[0028] 3) If δ exists i If (1≤i≤K) is less than the threshold 0.1, then K=K-1, proceed to step 1); otherwise, end.
[0029] The novel autoencoder neural network described above achieves adaptive adjustment of the dimension of the encoded data h, and has higher accuracy in missing signal completion compared to traditional missing signal completion methods.
[0030] As a preferred approach, to suppress overfitting during the training process of autoencoder neural networks, a loss function with L1 regularization and a loss function with relative entropy can be used, as follows:
[0031]
[0032] L KLD (x, g(f(x)))=E x～x′ ||xg(f(x))|| 2 +KLD
[0033]
[0034] Among them, L λ (·) denotes the loss function with L1 regularization, where L… KLD (·) denotes a loss function with relative entropy, λ and β represent the penalty coefficients in the two loss functions, M represents the number of neurons, wi represents the weight matrix of each neuron in the neural network, KLD represents the relative entropy, and ρ represents the neuron density. This represents the average activation level of the j-th neuron.
[0035] As a preferred embodiment, the host computer further includes a result display module, used to display the completion results of the missing signal completion module on the host computer.
[0036] The beneficial effects of this invention are mainly reflected in the following aspects: using an adaptive autoencoder neural network for radar missing signal completion has better robustness and higher accuracy compared to traditional missing signal completion methods; the radar missing signal completion system based on the adaptive autoencoder neural network completes missing signals faster than traditional missing signal completion methods after the model is trained. Attached Figure Description
[0037] Figure 1 This is a functional structure diagram of the radar missing signal completion system proposed in this invention;
[0038] Figure 2 This is a detailed illustration of each module of the radar missing signal completion system proposed in this invention. Detailed Implementation
[0039] The present invention will be further described below with reference to the accompanying drawings. The embodiments of the present invention are used to explain and illustrate the invention, but not to limit it. Any modifications and alterations made to the present invention within the spirit and scope of the claims fall within the protection scope of the present invention.
[0040] Example 1
[0041] Reference Figure 1 and Figure 2 A high-precision and robust radar missing signal completion system includes a database 1 and a host computer 2, which are connected sequentially. The database 1 contains existing and acquired radar signals, and the host computer includes:
[0042] Data preprocessing module 3 is used to preprocess the radar data stored in the database, and the process is as follows:
[0043] Radar signals are standardized using the following formula:
[0044]
[0045] Where, x ij Let x represent the j-th sampling point of the i-th radar signal. s_ij x represents ij The standardized form, μ i σ represents the average value of all sampling points of the i-th radar signal. i Let x represent the standard deviation of all sampling points of the i-th radar signal. s To represent the standardized radar signal.
[0046] Missing signal completion module 4 is used to build an adaptive autoencoder neural network to complete radar signals with missing signals processed by the data preprocessing module. It is accomplished through the following process:
[0047] Building an autoencoder neural network:
[0048] The basic form of an autoencoder neural network is shown below:
[0049]
[0050] f(x)=σ(W f ·x+b f )
[0051] g(h)=σ(W g ·h+b h )
[0052] Where f(·) represents the encoder function, g(·) represents the decoder function, x represents the input data, h represents the encoded data, x′ represents the reconstructed data, σ represents the activation function, and W f W represents the weight matrix in the encoding function. g b represents the weight matrix in the decoding function. f b represents the bias in the encoding function. h This represents the bias in the decoding function. The dimension of the encoded data h is manually set in the experiment based on the specific data.
[0053] An autoencoder neural network is trained using gradient descent and backpropagation, with the following constraints applied during training:
[0054] x≈x′
[0055] Where x′=g(f(x)). The goal of training is to make the data reconstructed by the decoder as close as possible to the input data.
[0056] Mean squared error is used to measure the similarity between the input data and the reconstructed data:
[0057] L(x, g(f(x)))=E x～x′ ||xg(f(x))|| 2
[0058] Where L(·) represents the loss function, E x～x′ (·) represents the expected value of the squared error between the input data x and the reconstructed data x′.
[0059] Building an adaptive autoencoder neural network:
[0060] 1) Set the initial dimension of the encoded data h to K, and preprocess the radar signal x. s The input is used for training in an autoencoder neural network.
[0061] 2) After the autoencoder neural network is trained, calculate the standard deviation δ = (δ1, δ2, ..., δ) of each dimension of the encoded data h. K ).
[0062] 3) If δ exists i If (1≤i≤K) is less than the threshold 0.1, then K=K-1, proceed to step 1); otherwise, end.
[0063] The novel autoencoder neural network described above achieves adaptive adjustment of the dimension of the encoded data h. Compared with traditional missing signal completion methods, it has better adaptability to different radar signal data and also has higher accuracy in missing signal completion.
[0064] In the missing signal completion module 4, to suppress overfitting of the autoencoder neural network during training, a loss function with L1 regularization and a loss function with relative entropy can be used, as follows:
[0065]
[0066] L KLD (x, g(f(x)))=E x～x′ ||xg(f(x))|| 2 +KLD
[0067]
[0068] Among them, L λ (·) denotes the loss function with L1 regularization, where L… KLD (·) denotes a loss function with relative entropy, λ and β represent the penalty coefficients in the two loss functions, M represents the number of neurons, wi represents the weight matrix of each neuron in the neural network, KLD represents the relative entropy, and ρ represents the neuron density. This represents the average activation level of the j-th neuron.
[0069] The host computer 2 also includes a result display module 5, which displays the classification results of the classification module on the host computer.
[0070] The hardware of the host computer 2 includes: I / O components for data acquisition and information transmission; a data storage device for storing data samples and operating parameters required for operation; a program storage device for storing software programs that implement functional modules; an arithmetic unit for executing programs and implementing specified functions; and a display module for displaying set parameters and detection results.
Claims
1. A high-precision and robust radar missing signal completion system, comprising a database and a host computer, wherein the database and the host computer are connected sequentially, characterized in that: The database contains pre-collected radar signals, and the host computer includes: The data preprocessing module is used to preprocess radar signals stored in the database; The missing signal completion module is used to build an adaptive autoencoder neural network to complete radar signals that have missing signals after being processed by the data preprocessing module; The missing signal completion module is completed using the following process: (3.1) Constructing an autoencoder neural network: The basic form of an autoencoder neural network is shown below: Where f(·) represents the encoder function and g(·) represents the decoder function. h represents the input data, and h represents the encoded data. Indicates data reconstruction. This represents the activation function. This represents the weight matrix in the encoding function. This represents the weight matrix in the decoding function. This represents the bias in the encoding function. This represents the bias in the decoding function; the dimension of the encoded data h is manually set in the experiment based on the specific data. An autoencoder neural network is trained using gradient descent and backpropagation, with the following constraints applied during training: in, The goal of training is to make the data reconstructed by the decoder as close as possible to the input data. Mean squared error is used to measure the similarity between the input data and the reconstructed data. Among them, L · Represents the loss function. · This represents the expected value of the squared error between the input data 𝑥 and the reconstructed data 𝑥'; (3.2) Construct an adaptive autoencoder neural network, including the following sub-steps: (3.2.1) The initial value of the dimension of the encoded data h is set to K, and the radar signal after data preprocessing is used. The input is used for training in an autoencoder neural network; (3.2.2) After the autoencoder neural network is trained, calculate the standard deviation of each dimension of the encoded data h. ; (3.2.3) If it exists Less than the threshold of 0.1 If the result is positive, then K = K-1, proceed to step 3.2.1; otherwise, end.
2. The high-precision and robust radar missing signal completion system according to claim 1, characterized in that: The data preprocessing module is completed using the following process: Radar signals are standardized using the following formula: in, This represents the j-th sampling point of the i-th radar signal. express Standardized form, This represents the average value of all sampling points of the i-th radar signal. This represents the standard deviation of all sampling points of the i-th radar signal.
3. The high-precision and robust radar missing signal completion system according to claim 1, characterized in that: The loss function has the following form: To suppress overfitting during the training process of autoencoder neural networks, a loss function with L1 regularization and a loss function with relative entropy are used, as follows: in, (·) represents the loss function with L1 regularization. (·) represents the loss function with relative entropy. and Let represent the penalty coefficients in the two loss functions, and M represent the number of neurons. This represents the weight matrix of each neuron in the neural network. Represents relative entropy. Indicates neuron density, This represents the average activation level of the j-th neuron.
4. The high-precision and robust radar missing signal completion system according to claim 1, characterized in that: The host computer also includes a result display module, which displays the signal completion results of the missing signal completion module on the host computer.