External radiation source radar target detection method and device based on improved ResNet network
By using an improved ResNet network for radar target detection of external radiation sources, and utilizing convolutional modules, residual convolutional networks, and multi-head attention modules, the problem of difficult detection of low, slow, and small targets in traditional methods is solved, achieving higher detection accuracy and robustness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2024-01-19
- Publication Date
- 2026-05-19
AI Technical Summary
Traditional external radiation source radars struggle to effectively suppress clutter when detecting low-speed, small targets and are inaccurate in detecting high-speed moving targets, resulting in low target echo energy and making them difficult to detect.
An improved ResNet network is used for object detection. Initial feature extraction is performed through convolutional modules, deep learning is conducted using residual convolutional networks, feature relationships are captured by multi-head attention modules, and classification is performed using fully connected layers.
It improves the accuracy of target detection under low signal-to-noise ratio, enhances the robustness and real-time performance of external radiation source radar, and solves the difficulty of detecting low, slow and small targets in traditional methods.
Smart Images
Figure CN118091575B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar technology, specifically relating to a method and apparatus for detecting external radiation source radar targets based on an improved ResNet network. Background Technology
[0002] External radiation source radar is a bistatic radar system that uses non-cooperative signals to detect targets. It has significant advantages in anti-jamming, anti-stealth target, anti-anti-radiation missile, and anti-low-altitude penetration. Furthermore, it plays a crucial role in the detection and classification of high-speed maneuvering targets and UAV targets in scenarios such as key area air defense and airport surveillance. Therefore, research on the detection of high and low altitude targets using external radiation source radar is of significant research importance.
[0003] With the gradual opening of low-altitude airspace, the frequent activity of "low, slow, and small" targets such as drones has brought a series of challenges to low-altitude airspace management. External radiation source radar relies on signals from third-party radiation sources, such as digital television broadcasts, to detect these targets. By accumulating and processing these signals over a long period, external radiation source radar can accurately distinguish target speeds, which is beneficial for detecting "low, slow, and small" targets. Because "low, slow, and small" targets are characterized by low flight altitude, slow speed, and small cross-sectional area, the target echo energy of external radiation source radar is low. Furthermore, the complexity of the urban low-altitude detection environment and the presence of a large amount of multipath clutter can easily cause the target echo to be submerged in clutter. Traditional detection methods for external radiation source radar use adaptive filters to suppress a large amount of clutter and treat the residual clutter and noise across the entire range Doppler spectrum as an independent and identically distributed Gaussian distribution model, employing square-law detection combined with various CFAR processing methods to detect targets. However, when suppressing clutter, traditional adaptive filters also filter out slower-moving targets, making them undetectable.
[0004] In recent years, deep learning technology, with neural networks at its core, has achieved revolutionary progress in perception fields such as computer vision. Deep neural networks, by establishing multi-layered representation structures ranging from simple to complex and performing intricate nonlinear processing on training data, enable the network to automatically identify and learn implicit features in the data, achieving a mapping from input to target. In the field of radar target detection and recognition, deep neural networks can automatically extract key features from the data. By constructing an end-to-end network model, it not only improves processing efficiency but also significantly enhances the accuracy of target detection and recognition compared to traditional methods. Therefore, researching deep learning-based methods for detecting "low, slow, and small" targets in external radiation source radar is of great significance.
[0005] Su Ningyuan et al. applied convolutional neural networks (CNNs) to the detection of micro-Doppler signals of maritime targets and intelligent detection of maritime targets by radar in their paper "A Method for Detection and Classification of Maritime Micro-Moving Targets Based on Convolutional Neural Networks [J]. Journal of Radar, 2018, 7(5): 565–574". Gao Xuanhao proposed a target detection algorithm based on residual networks in his paper "Research on Coherent Accumulation Algorithm for Radar Signals of Weak Maneuvering Targets [D]. Information Engineering University of Strategic Support Force, 2023". Most of the above target detection methods are applied to active radars and cannot solve the problem of detecting "low, slow and small" targets of external radiation source radars. Summary of the Invention
[0006] To address the aforementioned problems in the existing technology, this invention provides a method and apparatus for detecting external radiation source radar targets based on an improved ResNet network. The technical problem to be solved by this invention is achieved through the following technical solution:
[0007] In a first aspect, the present invention provides a radar target detection method for external radiation sources based on an improved ResNet network, comprising:
[0008] S100: Obtain the current radar signal from the radar system and use the current radar signal as the signal to be detected;
[0009] S200, the signal to be detected is normalized to obtain a normalized signal;
[0010] S300, the normalized signal is input into a trained network model for detecting external radiation source radar targets based on an improved ResNet network, so that the convolutional module in the network model performs preliminary feature extraction on the normalized signal, the extracted preliminary features are input into a residual convolutional network for residual calculation, and a multi-head attention module is used to capture the interrelationships between different preliminary features and to perform weighted calculations on different preliminary features using the interrelationships to extract important feature information. The feature information is then input into a fully connected layer for classification to obtain the signal category to which the current radar signal belongs.
[0011] Secondly, the present invention provides an external radiation source radar target detection device based on an improved ResNet network, comprising:
[0012] The acquisition module is configured to acquire the current radar signal from the radar system and use the current radar signal as the signal to be detected.
[0013] The processing module is configured to perform normalization processing on the signal to be detected to obtain a normalized signal;
[0014] The identification module is configured to input the normalized signal into a trained network model for detecting external radiation source radar targets based on an improved ResNet network, so that the convolutional module in the network model performs preliminary feature extraction on the normalized signal, inputs the extracted preliminary features into a residual convolutional network for residual calculation, and uses a multi-head attention module to capture the interrelationships between different preliminary features and uses the interrelationships to perform weighted calculations on different preliminary features to extract important feature information. The feature information is then input into a fully connected layer for classification to obtain the signal category to which the current radar signal belongs.
[0015] Beneficial effects:
[0016] This invention provides a method and apparatus for detecting external radiation source radar targets based on an improved ResNet network. The method involves acquiring the current radar signal from the radar system as the signal to be detected; normalizing the signal to be detected to obtain a normalized signal; and inputting the normalized signal into a trained network model for external radiation source radar target detection based on an improved ResNet network to obtain the signal category to which the current radar signal belongs. Because this invention uses a network model based on an improved ResNet, which incorporates a multi-head attention mechanism, it solves the problems of traditional external radiation source radar target detection methods, such as difficulty in detecting "low, slow, and small" targets and the need for in-depth statistical analysis of clutter environments. This effectively improves the accuracy of target detection under low signal-to-noise ratio conditions and enhances the robustness and real-time performance of external radiation source radar target detection.
[0017] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating a radar target detection method for external radiation sources based on an improved ResNet network provided by the present invention.
[0019] Figure 2 This is a schematic diagram of the external radiation source radar system model provided by the present invention;
[0020] Figure 3 This is a schematic diagram of the network model for external radiation source radar target detection provided by the present invention using the improved ResNet network. Detailed Implementation
[0021] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.
[0022] Combination Figures 1 to 3 This invention provides a radar target detection method for external radiation sources based on an improved ResNet network, comprising:
[0023] S100: Obtain the current radar signal from the radar system and use the current radar signal as the signal to be detected;
[0024] S200, the signal to be detected is normalized to obtain a normalized signal;
[0025] S300, the normalized signal is input into a trained network model for detecting external radiation source radar targets based on an improved ResNet network, so that the convolutional module in the network model performs preliminary feature extraction on the normalized signal, the extracted preliminary features are input into a residual convolutional network for residual calculation, and a multi-head attention module is used to capture the interrelationships between different preliminary features and to perform weighted calculations on different preliminary features using the interrelationships to extract important feature information. The feature information is then input into a fully connected layer for classification to obtain the signal category to which the current radar signal belongs.
[0026] This invention provides a method for detecting external radiation source radar targets based on an improved ResNet network. The method involves acquiring the current radar signal from the radar system as the signal to be detected; normalizing the signal to be detected to obtain a normalized signal; and inputting the normalized signal into a trained network model for external radiation source radar target detection based on an improved ResNet network to obtain the signal category to which the current radar signal belongs. Because this invention uses a network model based on an improved ResNet, which incorporates a multi-head attention mechanism, it solves the problems of traditional external radiation source radar target detection methods, such as difficulty in detecting "low, slow, and small" targets and the need for in-depth statistical analysis of clutter environments. This effectively improves the accuracy of target detection under low signal-to-noise ratio conditions and enhances the robustness and real-time performance of external radiation source radar target detection.
[0027] In one specific embodiment of the present invention, the network model for detecting external radiation source radar targets based on the improved ResNet network is trained using a training sample set and then tested using a test sample set.
[0028] The training sample set and the test sample set are constructed by building an external radiation source radar system model, a target echo model, and a clutter model, and using the external radiation source radar system model, target echo model, and clutter model to receive target echo signals and clutter signals; the target echo signals and clutter signals are preprocessed, and the preprocessed target echo signals and clutter signals are divided into a training sample set and a test sample set.
[0029] The external radiation source radar system model of this invention includes a transmitting station, a target, and a receiving station, and its geometric model schematic diagram is attached. Figure 2 As shown.
[0030] The constructed external radiation source radar system model includes a transmitting station, a target, and a receiving station. Under the target echo model, the reference signal received by the reference channel and the echo signal received by the echo channel are respectively represented as follows:
[0031] S ref_t (t)=As(t)+n ref_t (t);
[0032]
[0033] Where s(t) is the received direct wave signal, and A represents the amplitude of the direct wave signal; B n N represents the amplitude of the target signal in the echo channel. n f is the number of targets. d n represents the Doppler frequency shift of the target. ref_t and n echo_t Indicates noise;
[0034] Under the aforementioned clutter model, the reference signal received by the reference channel and the echo signal received by the echo channel are respectively expressed as:
[0035] S ref_n (t)=As(t)+n ref_n (t);
[0036] S echo_n (t)=n echo_n (t);
[0037] Where s(t) is the received direct wave signal, and A represents the amplitude of the direct wave signal; n ref_n and n echo_n Indicates noise.
[0038] In a specific embodiment of the present invention, the step of preprocessing the target echo signal and clutter signal, and dividing the preprocessed target echo signal and clutter signal into a training sample set and a test sample set includes:
[0039] Sub-step 2a involves performing a Fourier transform on the target echo model to obtain the target echo signal. The Fourier transform process is expressed as follows:
[0040] S r_ti =[FFT(S) ref_ti (n))] i=0,…,M-1;
[0041] S e_ti =[FFT(S) echo_ti (n))] i=0,…,M-1;
[0042] Sub-step 2b involves performing a Fourier transform on the clutter model to obtain the clutter signal; the Fourier transform process is expressed as follows:
[0043] S r_ni =[FFT(S) ref_ni (n))] i=0,…,M-1
[0044] S e_ni =[FFT(S) echo_ni (n))] i=0,…,M-1
[0045] Sub-step 2c involves preprocessing the target echo signal and the clutter signal to obtain a preprocessing result;
[0046] Sub-step 2d: Based on the preprocessing results, use MATLAB simulation to generate M and N sets of data samples. The data sample sets include H0 samples and H1 samples. H0 samples indicate that the echo channel contains only clutter, and H1 samples indicate that the echo channel contains both clutter and the target signal. H0 samples and H1 samples each account for half of the group.
[0047] It is worth noting that: the simulation generates M training set samples, where H0 samples (echo channel contains only clutter) and H1 samples (echo channel contains both clutter and target signal) each account for M / 2; in H0 and H1 samples, the direct wave to noise power ratio JNR in the specified reference channel ∈ {a1,a2,a3} dB, and the echo signal to noise power ratio SNR in the echo channel ∈ [-p,0] dB, where M≥10000, a1,a2,a3≥0, p≥0, and the number of samples corresponding to each SNR and JNR value. The number of samples is randomized. The simulation generates N test set samples, where H0 samples (echo channel contains only clutter) and H1 samples (echo channel contains both clutter and target signal) each account for N / 2. In H0 and H1 samples, the direct wave to noise power ratio JNR ∈ {b1,b2,b3} dB in the specified reference channel, and the echo signal to noise power ratio SNR ∈ [-q,0] dB in the echo channel, where N ≥ 1000, b1,b2,b3 ≥ 0, q ≥ 0, and the number of samples corresponding to each SNR and JNR value is randomized. Each signal has a length L, with its real and imaginary parts being data from the I and Q channels, respectively. The I / Q channels are used as the input channel dimensions of the dataset. L ≥ 20000. In this embodiment, M = 18000, a1 = 25, a1 = 30, a1 = 35, p = 20; N = 3000, b1 = 25, b2 = 30, b3 = 35, q = 20, L = 40000.
[0048] Sub-step 2e generates corresponding labels for samples H0 and H1, where the label y = 0 for sample H0 and the label y = 1 for sample H1.
[0049] refer to Figure 3 The network model for detecting external radiation source radar targets based on the improved ResNet network includes a backbone network, which is composed of a residual convolutional network ResNet34. A convolutional module is added before the residual convolutional network ResNet34 for preliminary feature extraction, and a multi-head attention module is added after the residual convolutional network ResNet34. The output of the multi-head attention module is connected to a fully connected layer.
[0050] refer to Figure 3 The convolutional module includes a convolutional layer, a ReLU activation layer, a batch normalization layer, and a max pooling layer. The internal connection relationship of the convolutional module is: convolutional layer → batch normalization layer → ReLU activation layer → max pooling layer. The number of convolutional kernels in the convolutional layer is 64, the kernel size is 3, and the stride is 2. The size of the max pooling layer is 2, and the sliding stride is 2.
[0051] The ResNet34 neural network is a deep convolutional neural network containing four residual blocks. The internal connection relationships of the ResNet34 neural network are as follows: the first residual block outputs 64 channels, repeated 3 times; the second residual block outputs 128 channels, repeated 4 times; the third residual block outputs 256 channels, repeated 6 times; and the fourth residual block outputs 512 channels, repeated 3 times. The internal connection relationship of each residual block is: convolutional layer → batch normalization layer → ReLU activation layer → Dropout layer → convolutional layer → batch normalization layer. The convolutional layer in each residual block has a kernel size of 3, a stride of 2, a max pooling layer size of 2, and a sliding stride of 2.
[0052] Multi-head attention involves establishing different projection information in multiple different projection spaces. The input matrix is projected into various ways to obtain many output matrices, which are then concatenated together. Figure 3 As can be seen, Q, K, and V are fixed single values, while the Linear layer has three, and the Scaled Dot-Product Attention layer has three, i.e., three multi-head attention mechanisms. These are then concatenated, and the Linear layer transforms them into a single output value identical to that of a single-head attention mechanism. Multi-head attention is a mechanism that captures the relationships between different features by extracting more important feature information through weighted calculations of different feature vectors. Specifically, the multi-head self-attention layer has six heads, and the query matrix Q, key matrix K, and value matrix V in the multi-head self-attention layer all have a dimension of 70.
[0053] A fully connected layer is set at the output of the multi-head attention module to classify the feature vectors obtained after feature extraction.
[0054] In one specific embodiment of the present invention, the training process of the network model for detecting external radiation source radar targets based on the improved ResNet network is as follows:
[0055] Sub-step 4a: Initialize the network model A for external radiation source radar target detection based on the improved ResNet network with parameters of normally distributed random numbers. t Set the maximum number of iterations T≥200, initialize the number of iterations t=1, and let A=A t ;
[0056] Sub-step 4b: Normalize each training sample in the training sample set;
[0057] Sub-step 4c: The normalized training samples are used as input to the external radiation source radar target detection network based on the improved ResNet network;
[0058] Sub-step 4d involves using a convolution module to perform preliminary feature extraction on the input multi-channel training samples;
[0059] Sub-step 4e inputs the initially extracted features into a deep residual network ResNet34. ResNet allows training very deep neural networks without being affected by vanishing or exploding gradients. Furthermore, through skip connections, even if a layer negatively impacts network performance, it can be skipped through normalization. This allows the deep residual convolutional network ResNet34 to learn deeper target echo features. The data information is obtained by learning deeper target echo features through four residual blocks in the deep residual convolutional network ResNet34.
[0060] The residual block includes a residual portion and a direct mapping portion, expressed by the formula:
[0061]
[0062] Where h(x) i ) represents a direct mapping, w i Let x be the weight matrix. i and y i Let F(·) and f(·) represent the input and output, respectively, and let F(·) and f(·) represent the residual function and ReLU activation function, respectively.
[0063] Sub-step 4f involves feeding the data information from the deep residual network ResNet34 into the multi-head attention machine module, and then using the multi-head attention module to correlate the data information with the weight matrix W. Q W K W VMultiplying these values yields the projections of the data information into three spaces, denoted as Q, K, and V. Passing Q, K, and V through different linear layers yields mappings Q to different subspaces. i ,K i V i For each group Q i ,K i V i Perform attention operations separately to obtain the output Z for each group. i The outputs of all Attention functions are concatenated together and subjected to a linear transformation to obtain the final output Z. The specific calculation expression of the multi-head attention module is as follows:
[0064]
[0065] Z = Concat(Z1, ..., Z) h V i ;
[0066] Where h is the number of attention heads.
[0067] Multi-head attention mechanisms allow each head to focus on different parts of the input, thereby capturing more complex feature dimensional relationships and obtaining more accurate classification results.
[0068] Sub-step 4g involves extracting the feature vector from the multi-head attention module. The input is processed through a fully connected layer for object detection, resulting in predicted labels.
[0069] Sub-step 4h involves calculating the loss function for the current iteration number based on the predicted labels and their corresponding true labels, and then applying the stochastic gradient descent algorithm to the network model A for external radiation source radar target detection based on the improved ResNet network. t The parameters are updated to obtain the updated network A′ for external radiation source radar target detection based on the improved ResNet network; the cross-entropy loss function is used, and A′ is calculated by using the predicted labels and their corresponding ground truth labels. t Loss value L loss Then, the backpropagation algorithm is used, and through L... loss For A t The parameters are updated; among them, the loss value L... loss The calculation formula is:
[0070]
[0071] Among them, y iq The indicator variable y represents the value of the i-th sample if the predicted label matches the true label. iq The value is 1 otherwise the value is 0. iq∑ represents the predicted probability that the i-th sample belongs to the q-th class, and ∑ represents the summation operation.
[0072] Sub-step 4i: Determine whether t = T is true. If yes, obtain the trained convolutional neural network A′. Otherwise, let t = t + 1 and execute sub-step 4a.
[0073] This invention uses a trained, improved ResNet network to identify targets for each sample in the test set and output the corresponding category. Specifically, the normalized test set is used as the network input for target detection of external radiation source radar based on the trained, improved ResNet network to obtain the corresponding category.
[0074] Secondly, the present invention provides an external radiation source radar target detection device based on an improved ResNet network, comprising:
[0075] The acquisition module is configured to acquire the current radar signal from the radar system and use the current radar signal as the signal to be detected.
[0076] The processing module is configured to perform normalization processing on the signal to be detected to obtain a normalized signal;
[0077] The identification module is configured to input the normalized signal into a trained network model for detecting external radiation source radar targets based on an improved ResNet network, so that the convolutional module in the network model performs preliminary feature extraction on the normalized signal, inputs the extracted preliminary features into a residual convolutional network for residual calculation, and uses a multi-head attention module to capture the interrelationships between different preliminary features and uses the interrelationships to perform weighted calculations on different preliminary features to extract important feature information. The feature information is then input into a fully connected layer for classification to obtain the signal category to which the current radar signal belongs.
[0078] This invention provides an external radiation source radar target detection device based on an improved ResNet network. The device acquires the current radar signal from the radar system as the signal to be detected; normalizes the signal to be detected to obtain a normalized signal; and inputs the normalized signal into a trained network model for external radiation source radar target detection based on an improved ResNet network to obtain the signal category to which the current radar signal belongs. Because this invention uses a network model based on an improved ResNet, which incorporates a multi-head attention mechanism, it solves the problems of traditional external radiation source radar target detection methods, such as difficulty in detecting "low, slow, and small" targets and the need for in-depth statistical analysis of clutter environments. This effectively improves the accuracy of target detection under low signal-to-noise ratio conditions and enhances the robustness and real-time performance of external radiation source radar target detection.
[0079] The effects of this invention can be further illustrated by the following simulation experiments.
[0080] (1) Experimental conditions
[0081] The hardware platform for the simulation experiment of this invention is as follows: GPU is NVIDIA RTX A5500 with 16GB of video memory, CPU is i9-12900H with 16 cores and a clock speed of 2.3GHz, and memory size is 16GB.
[0082] The software platform for the simulation experiment of this invention is Windows 11.
[0083] The training and test sample sets for the simulation experiments of this invention use QPSK simulation signals, with a signal accumulation time of 0.2s, a sampling rate of 15.12MHz, a target delay range of k = 1000–5000 bits, and a target echo signal Doppler frequency shift f. d ∈[-3,3]H z .
[0084] (2) Simulation content and result analysis
[0085] Simulation Experiment: Based on the training and test sample sets of the direct wave to noise power ratio (JNR) and echo signal to noise power ratio (SNR) in different reference channels obtained from the simulation of this invention, the results of the external radiation source radar target detection method using the improved ResNet network are shown below.
[0086] Table 1 shows the recognition accuracy of the improved ResNet network proposed in this invention.
[0087]
[0088] The H1 test set is used to estimate the detection probability P. D The H0 test set is used to estimate the false alarm probability P. FA Both probabilities were estimated using the Monte Carlo method. As can be seen from the experimental results in Table 1, the network of this invention can perform target detection for external radiation source radar and has good target detection accuracy.
[0089] Table 2 shows the recognition accuracy of different SNRs in the JNR=25dB echo channel of the reference channel.
[0090]
[0091] Table 3 shows the recognition accuracy of different SNRs in the echo channel with JNR=30dB in the reference channel.
[0092]
[0093] Table 4 shows the recognition accuracy of different SNRs in the JNR=35dB echo channel of the reference channel.
[0094]
[0095] As can be seen from the experimental results in Tables 2, 3, and 4, the network of the present invention can complete target detection of external radiation source radar under different signal-to-noise ratios, and the target detection performance is good.
[0096] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.
[0097] Although this application has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.
[0098] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.
Claims
1. A radar target detection method for external radiation sources based on an improved ResNet network, characterized in that, include: S100: Obtain the current radar signal from the radar system and use the current radar signal as the signal to be detected; S200, the signal to be detected is normalized to obtain a normalized signal; S300, the normalized signal is input into a trained network model for detecting external radiation source radar targets based on an improved ResNet network, so that the convolutional module in the network model performs preliminary feature extraction on the normalized signal, the extracted preliminary features are input into a residual convolutional network for residual calculation, and a multi-head attention module is used to capture the interrelationships between different preliminary features and to perform weighted calculations on different preliminary features using the interrelationships to extract important feature information. The feature information is then input into a fully connected layer for classification to obtain the signal category to which the current radar signal belongs.
2. The method for detecting external radiation source radar targets based on an improved ResNet network according to claim 1, characterized in that, The network model for detecting external radiation source radar targets based on the improved ResNet network is trained using a training sample set and then tested using a test sample set. The training sample set and the test sample set are constructed by building an external radiation source radar system model, a target echo model, and a clutter model, and using the external radiation source radar system model, target echo model, and clutter model to receive target echo signals and clutter signals; the target echo signals and clutter signals are preprocessed, and the preprocessed target echo signals and clutter signals are divided into a training sample set and a test sample set.
3. The method for detecting external radiation source radar targets based on an improved ResNet network according to claim 2, characterized in that, The step of preprocessing the target echo signal and clutter signal, and dividing the preprocessed target echo signal and clutter signal into a training sample set and a test sample set includes: Sub-step 2a: Perform a Fourier transform on the target echo model to obtain the target echo signal; Sub-step 2b: Perform a Fourier transform on the clutter model to obtain the clutter signal; Sub-step 2c involves preprocessing the target echo signal and the clutter signal to obtain a preprocessing result; Sub-step 2d: Based on the preprocessing results, generate using MATLAB simulation. Groups and Group data sample set, in which The dataset is used as the training dataset. The data sample set is used as the test sample set; the data sample set includes Samples and sample, The sample indicates that the echo channel contains only clutter. The sample indicates that the echo channel contains clutter and the target signal. Samples and Each sample comprises half of the group; Sub-step 2e is Samples and The sample generates corresponding labels, where, Sample corresponding label , Sample corresponding label .
4. The method for detecting external radiation source radar targets based on an improved ResNet network according to claim 1, characterized in that, The network model for detecting external radiation source radar targets based on the improved ResNet network includes a backbone network, which is composed of a residual convolutional network ResNet34. A convolutional module is added before the residual convolutional network ResNet34 for preliminary feature extraction, and a multi-head attention module is added after the residual convolutional network ResNet34. The output of the multi-head attention module is connected to a fully connected layer.
5. The method for detecting external radiation source radar targets based on an improved ResNet network according to claim 4, characterized in that, The convolutional module contains convolutional layers, ReLU activation layers, batch normalization layers, and max pooling layers. The internal connection relationship of the convolutional module is: convolutional layer → batch normalization layer → ReLU activation layer → max pooling layer. The number of convolutional kernels in the convolutional layer is 64, the kernel size is 3, and the stride is 2. The size of the max pooling layer is 2, and the sliding stride is 2. The ResNet34 residual convolutional network comprises four residual blocks. The internal connection relationships of the ResNet34 residual convolutional network are as follows: the first residual block outputs 64 channels, repeated 3 times; the second residual block outputs 128 channels, repeated 4 times; the third residual block outputs 256 channels, repeated 6 times; and the fourth residual block outputs 512 channels, repeated 3 times. The internal connection relationship of each residual block is: convolutional layer → batch normalization layer → ReLU activation layer → Dropout layer → convolutional layer → batch normalization layer. The convolutional layer within each residual block has a kernel size of 3 and a stride of 2, and the max pooling layer has a size of 2 and a sliding stride of 2. The multi-head attention module has 6 sub-heads, and the dimensions of the query matrix Q, key matrix K, and value matrix V in the multi-head attention module are all 70. A fully connected layer is set at the output of the multi-head attention module to classify the feature vectors output by the multi-head attention module.
6. The method for detecting external radiation source radar targets based on an improved ResNet network according to claim 2, characterized in that, The training process of the network model for detecting external radiation source radar targets based on the improved ResNet network is as follows: Sub-step 4a: Initialize the network model for external radiation source radar target detection based on an improved ResNet network with parameters set to normally distributed random numbers. Set the maximum number of iterations. Initialize the number of iterations and order ; Sub-step 4b: Normalize each training sample in the training sample set; Sub-step 4c: The normalized training samples are used as input to the external radiation source radar target detection network based on the improved ResNet network; Sub-step 4d involves using a convolution module to perform preliminary feature extraction on the input multi-channel training samples; In sub-step 4e, the initially extracted features are input into the deep residual network ResNet34. Through the four residual blocks in the deep residual convolutional network ResNet34, deeper target echo features are learned to obtain data information. Sub-step 4f involves feeding the data information from the deep residual network ResNet34 into the multi-head attention machine module, and then using the multi-head attention module to correlate the data information with the weight matrix. Multiplying these components yields the projections of the data information into the three spaces, denoted as... ;Will Different subspaces are mapped through different linear layers. For each group Perform attention operations separately to obtain the output for each group. The outputs of all Attention functions are concatenated together and subjected to a linear transformation to obtain the final output. ; Sub-step 4g involves extracting the feature vector from the multi-head attention module. The input is processed through a fully connected layer for object detection, resulting in predicted labels. Sub-step 4h calculates the loss function for the current iteration based on the predicted labels and their corresponding ground truth labels, and applies the stochastic gradient descent algorithm to the network model for external radiation source radar target detection based on the improved ResNet network. The parameters are updated to obtain the updated network for external radiation source radar target detection based on the improved ResNet network. The cross-entropy loss function is used, and the result is calculated by predicting the label and its corresponding true label. Loss value Then, the backpropagation algorithm is used, and through right Update the parameters; Sub-step 4i, determine If true, then a trained convolutional neural network is obtained. Otherwise Then execute sub-step 4a.
7. The method for detecting external radiation source radar targets based on an improved ResNet network according to claim 6, characterized in that, The residual block includes a residual portion and a direct mapping portion, expressed by the formula: ； in, For direct mapping, This is the weight matrix. and These represent the input and output, respectively. and These represent the residual function and the ReLU activation function, respectively.
8. The method for detecting external radiation source radar targets based on an improved ResNet network according to claim 7, characterized in that, The specific calculation expression for the multi-head attention module in sub-step 4f is as follows: ； ； in, The number of attention points.
9. The method for detecting external radiation source radar targets based on an improved ResNet network according to claim 8, characterized in that, In the sub-step 4h, the loss value The calculation formula is: ； in, Represents an indicator variable, if the first... If the predicted label of a sample is the same as the true label, then The value is 1 otherwise the value is 0. Representing the The sample belongs to the first Predicted probability of class This indicates a summation operation.
10. A radar target detection device for external radiation sources based on an improved ResNet network, characterized in that, include: The acquisition module is configured to acquire the current radar signal from the radar system and use the current radar signal as the signal to be detected. The processing module is configured to perform normalization processing on the signal to be detected to obtain a normalized signal; The identification module is configured to input the normalized signal into a trained network model for detecting external radiation source radar targets based on an improved ResNet network, so that the convolutional module in the network model performs preliminary feature extraction on the normalized signal, inputs the extracted preliminary features into a residual convolutional network for residual calculation, and uses a multi-head attention module to capture the interrelationships between different preliminary features and uses the interrelationships to perform weighted calculations on different preliminary features to extract important feature information. The feature information is then input into a fully connected layer for classification to obtain the signal category to which the current radar signal belongs.