Radar unsupervised neural network CFAR detection method based on local clutter power estimation

Abstract

The invention discloses a radar unsupervised neural network CFAR detection method based on local clutter power estimation, relates to the technical field of radar detection, and aims to solve the problem of low CFAR detection accuracy caused by low local clutter power estimation accuracy of an existing adaptive CFAR detector in a complex non-uniform environment. According to the method, the local clutter power at the to-be-detected unit is accurately estimated by using the neural network, and the technical scheme of the invention prevents background information loss after mask shielding caused by too dense CUT (peak point), thereby improving the prediction accuracy of the local clutter power at the mask, and improving the prediction accuracy of the local clutter power at the to-be-detected unit. And finally, the CFAR detection accuracy under the complex non-uniform background is improved. Moreover, the method avoids the training of real manual annotation data, can achieve the training through the large-scale unannotated data in actual use, and improves the generalization of the method for the measured data compared with an intelligent radar target.

Description

Technical Field

[0001] The present invention relates to the field of radar detection technology, and in particular to a radar unsupervised neural network CFAR detection method based on local clutter power estimation. Background Art

[0002] In recent years, high-frequency ground wave radar (HFSWR) has been widely used in ocean monitoring. However, its detection background is interfered by factors such as ionospheric clutter, sea clutter, and ground clutter, which makes target detection face huge challenges.

[0003] Constant False Alarm Rate (CFAR) detectors are widely used in practical radar systems. They use reference cells to estimate the local clutter power (LCPE) at the target cell in real time and dynamically adjust the detection threshold. To adapt to various complex and non-uniform scenarios, adaptive CFAR detectors have been proposed. These detectors achieve more accurate local clutter power estimation in complex and non-uniform environments and dynamically adjust the detection threshold to further improve detection performance. Currently, adaptive CFAR detectors can be divided into two categories based on their local clutter power estimation strategies. The first category, such as ADCCA-CFAR and CML-CFAR, improves LCPE accuracy by eliminating interfering cells in the reference window. The second category, such as VI-CFAR and RWVI-CFAR, improves local clutter power estimation accuracy by selecting the optimal strategy based on differences in the detection background. However, adaptive CFAR detectors still suffer from low local clutter power estimation accuracy in complex and non-uniform environments, which in turn affects CFAR detection accuracy. Summary of the Invention

[0004] The purpose of the present invention is to provide a radar unsupervised neural network CFAR detection method based on local clutter power estimation to address the problem that the existing adaptive CFAR detector has low local clutter power estimation accuracy in complex non-uniform environments, which in turn leads to low CFAR detection accuracy.

[0005] The technical solution adopted by the present invention to solve the above technical problems is:

[0006] The radar unsupervised neural network CFAR detection method based on local clutter power estimation includes the following steps:

[0007] Step 1: Obtain the sea surface reflected echo signal to be identified, and obtain the RD spectrum data matrix X of the sea surface reflected echo signal;

[0008] Step 2: Perform local extreme value detection on the RD spectrum data matrix X of the echo signal reflected from the sea surface to determine all peak points;

[0009] Step 3: Randomly divide all peak points into D non-overlapping groups;

[0010] Step 4: Mask each group to obtain the RD spectrum data mask matrix, and input the RD spectrum data mask matrix into the trained neural network to obtain the prediction result of the output RD spectrum data matrix That is, the prediction result of the local clutter power at the mask;

[0011] Step 5: Use RD spectrum data matrix X and prediction results The difference between them is used to calculate the local signal-to-noise ratio estimation result S at each group of masks ij ;

[0012] Step 6: Repeat steps 4 and 5 to obtain D groups of local SNR estimation results at the mask, and accumulate the estimation results of the D groups to obtain the accumulated result, i.e. the local SNR estimation result of the peak point. Then, the local SNR estimation result of the peak point is compared with the threshold set by the false alarm probability to determine whether there is a target at each peak point, thereby achieving CFAR detection;

[0013] The trained neural network is obtained by the following steps:

[0014] Step A: Acquire the sea surface reflected echo signal and obtain the RD spectrum data matrix of the sea surface reflected echo signal;

[0015] Step B: distinguishing the target area and non-target area in the RD spectrum data matrix;

[0016] Step C: Replace the target area in the RD spectrum data matrix with a non-target area of the same size to obtain a pure clutter RD spectrum data matrix;

[0017] Step D: perform data enhancement on the pure clutter RD spectrum data matrix;

[0018] Step E: Perform random masking on the pure clutter RD spectrum data matrix after data enhancement to obtain the RD spectrum data mask matrix

[0019] Step F: Mask the RD spectrum data to the matrix As input, the pure clutter RD spectrum data matrix after data enhancement is taken as output, and the neural network is trained by fine-tuning the optimal model training strategy to obtain a trained neural network.

[0020] Furthermore, in step B, the target area and the non-target area in the RD spectrum data matrix are distinguished by obtaining the positive and negative Doppler frequencies corresponding to the Bragg peaks in the first-order sea clutter. The positive and negative Doppler frequencies corresponding to the Bragg peaks in the first-order sea clutter are expressed as:

[0021]

[0022] Where f is the radar carrier frequency, g is the acceleration due to gravity, and c is the speed of light.

[0023] Furthermore, the pure clutter RD spectrum data matrix is expressed as:

[0024]

[0025] Y1=y1+Δy / 2

[0026] Y2=y2-Δy / 2

[0027]

[0028] Among them, N d =(Y2-Y1) / 2, X i Represents a single RD spectrum data matrix, N d , Y1, y1, Y2 and y2 represent intermediate variables, Δy represents the number of Doppler units occupied by the unilateral first-order sea clutter, f single It represents the frequency resolution corresponding to a single Doppler unit, and N2 represents the total number of Doppler units in the RD spectrum.

[0029] Furthermore, the masking in step E is a random masking, specifically:

[0030] Step E1: Obtain the pure clutter RD spectrum data matrix after data enhancement Where N1 is the total number of range cells of the RD spectrum, and N2 is the total number of Doppler cells of the RD spectrum;

[0031] Step E2: Randomly select N in the range of N1×N2 mask coordinates As the center position of the random mask, its coordinate range is a j ∈[1,N1],b j ∈[1,N2];

[0032] Step E3: Generate a basic matrix with the same dimension as the pure clutter RD spectrum data matrix after data enhancement All elements in the basic matrix are 1;

[0033] Step E4: In the basic matrix I, for each coordinate (a j ,b j ) Generate a q×q all-0 mask for the center;

[0034] Step E5: Repeat step E4 until all coordinates are processed and a random mask matrix is obtained. Random mask matrix M i Expressed as:

[0035]

[0036] Where q is the preset mask size;

[0037] Step E6: Random mask matrix M i Perform dot multiplication with the pure clutter RD spectrum data matrix after data enhancement to generate the RD spectrum data mask matrix.

[0038] Furthermore, the masking in step 3 is group masking, specifically:

[0039] Step 31: Generate a basic matrix with the same dimension as the pure clutter RD spectrum data matrix after data enhancement All elements in the basic matrix are 1;

[0040] Step 32: In the basic matrix I, the position of each peak point Generate a q×q all-0 mask for the center;

[0041] Step 33: Repeat step 32 until all peak points are processed and the mask matrix is obtained. M ij Expressed as:

[0042]

[0043] Step 34: Repeat steps 32 and 33 to obtain a set of mask matrices for D groups.

[0044] Step 35: Set the mask matrix Perform dot multiplication with the pure clutter RD spectrum data matrix after data enhancement to generate the RD spectrum data mask matrix.

[0045] Furthermore, the local signal-to-noise ratio estimation results S at each group of masks are ij Expressed as:

[0046]

[0047] Where k = (p-1) / 2, p is the preset size, usually smaller than the mask size q, k, a, b represent intermediate variables, P ij Represents the position label matrix of all peak points in group j, E ij Represents the prediction results of group j Each mask and RD spectrum data matrix X i The difference between X i represents the RD spectrum matrix to be detected, represents the j-th group prediction result, M ij represents the j-th group mask matrix.

[0048] Furthermore, the local signal-to-noise ratio estimation result of the peak point is expressed as:

[0049]

[0050] Furthermore, the loss function of the neural network is:

[0051]

[0052] Among them, N train represents the number of RD spectra to be trained, Represents the pure clutter RD spectrum data matrix after data enhancement, Y i pred (m,n) represents the prediction result.

[0053] Furthermore, the training neural network is trained by fine-tuning the optimal model training strategy, specifically: first, through a fixed number of basic training stages, the current optimal model is saved as the basic model using an artificial verification set, then the optimal model obtained in the basic training stage is loaded and the batch normalization layer parameters are fixed, and the unfrozen layers are fine-tuned iteratively trained, and the final network model with the best detection performance is obtained through multiple rounds of parameter optimization;

[0054] The artificial validation set is constructed by injecting the real target in the RD spectrum of the HFSWR into the target area of the validation set RD spectrum by adjusting the signal-to-noise ratio;

[0055] The optimal model is obtained by the following steps:

[0056] False alarm rate of passing non-target area Under the condition of keeping constant, select the probability of detecting the simulated target The highest model is taken as the best performing model, that is, the trained neural network:

[0057]

[0058] in, Indicates the number of simulated targets detected, Indicates the total number of Doppler cells in the non-target area, Indicates the number of false alarm points in the non-target area.

[0059] Furthermore, the data enhancement includes horizontal translation and horizontal flipping.

[0060] The beneficial effects of the present invention are:

[0061] This application uses a neural network to accurately estimate the local clutter power at the unit to be detected, and the technical solution of this application prevents the loss of background information after the mask is blocked due to the CUT (peak point) being too dense, thereby improving the prediction accuracy of the local clutter power at the mask, and ultimately improving the CFAR detection accuracy under complex non-uniform backgrounds.

[0062] In addition, this application avoids using real manually labeled data for training. In actual use, large-scale unlabeled data can be used for training, which improves the generalization of the method to measured data compared to smart radar targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Schematic diagram of measured HFSWRRD spectrum data;

[0064] Figure 2 This is the detection flow chart of ULCPE-CFAR;

[0065] Figure 3 Schematic diagram of the U-Net network structure;

[0066] Figure 4 Schematic diagram of ROC curves for different algorithms for single-beam HFSWR measured data;

[0067] Figure 5 P fa =1×10 -4 When the P value of different algorithms for single-beam HFSWR measured data is d -SNR curve;

[0068] Figure 6 A comparison of target track detection between ULCPE-CFAR and TP-CFAR in multi-beam HFSWR measured data;

[0069] Figure 7 A comparison chart of target track detection between ULCPE-CFAR and supervised intelligent target detector in multi-beam HFSWR measured data. DETAILED DESCRIPTION

[0070] It should be noted that, in the absence of conflict, the various embodiments disclosed in this application can be combined with each other.

[0071] Specific embodiment 1: The radar unsupervised neural network CFAR detection method based on local clutter power estimation described in this embodiment includes:

[0072] Step 1: Self-supervised training process of ULCPE-CFAR. The HFSWR range-Doppler spectrum (RD spectrum) to be trained is processed, and the clutter data is used to mask the target area and enhance the data, generating pure clutter data with the target unit removed. The pure clutter data is randomly masked to construct self-supervised training samples. The training samples are input into the U-Net network to be trained. By calculating the MSE loss between the pure clutter data and the network output, the network is iteratively optimized and trained to enable the network to predict the local clutter power at the masked location. In each round of network training, the validation set is used to select and save the optimal model. The network is trained using the strategy of fine-tuning the optimal model to obtain a U-Net network that can estimate the local clutter power at the masked location after training.

[0073] Step 2: ULCPE-CFAR group inference process. Peak detection is performed on the HFSWR RD spectrum to be tested, and the units to be detected (peak points) are obtained and divided into multiple groups. The RD spectrum is masked at each group of units to be detected to generate group-masked samples. The masked samples are input into the trained U-Net network to obtain the network output results after local clutter power prediction. The error between the original RD spectrum and the local clutter power prediction results corresponding to each group of masks is calculated, and the errors at each unit to be detected are accumulated to estimate the local signal-to-noise ratio matrix. A threshold decision is made on the local signal-to-noise ratio estimation matrix to achieve adaptive CFAR detection at the units to be detected.

[0074] 1. Figure 2 The detection flow chart of the proposed ULCPE-CFAR is presented. The method includes the network training process and the network inference process.

[0075] The network training process of ULCPE-CFAR includes the following steps:

[0076] Step 1: Radar data collection and preliminary processing.

[0077] 1.1 Data acquisition and data domain processing:

[0078] A high-frequency ground wave radar (HFSWR) system transmits electromagnetic wave signals in a specific frequency band and receives echo signals reflected from the sea surface through an antenna array. These echo signals are processed to generate the original range-Doppler spectrum (RD spectrum) data matrix. (like Figure 1 (as shown), where N1 is the total number of range bins in the RD spectrum, and N2 is the total number of Doppler bins in the RD spectrum. This data is the result of square-law detection, and each value in the RD spectrum corresponds to a power value.

[0079] To facilitate subsequent analysis and processing, the power value of the original RD spectrum is converted to decibel (dB) units according to the logarithmic relationship. The specific conversion formula is as follows:

[0080] X dB =10·log 10 (X) (1)

[0081] Where X dB is the RD spectrum matrix converted to decibel values.

[0082] 1.2RD spectrum region division:

[0083] Based on the HFSWR detection characteristics analysis, the ship target speed is usually significantly lower than the system's theoretical detection speed limit. This characteristic causes the targets in the HFSWR RD spectrum data to be concentrated in a specific Doppler range.

[0084] To more effectively analyze RD spectrum data, this application divides the RD spectrum data into two areas: target area and non-target area based on the theoretical location of first-order sea clutter. The positive and negative Doppler frequencies corresponding to the Bragg peaks in the first-order sea clutter can be calculated using the following formula:

[0085]

[0086] In the above formula, f represents the radar carrier frequency, g represents the acceleration of gravity, and c is the speed of light. According to the above formula, the Doppler unit numbers y1 and y2 corresponding to the Bragg peak theory can be obtained:

[0087]

[0088] Among them, f single The frequency resolution corresponding to a single Doppler unit. Considering the characteristic that the first-order sea clutter on one side occupies 2Δy Doppler units in the RD spectrum, the Doppler range of the first-order sea clutter of the HFSWR can be expressed as y∈[y1-Δy,Y1)∪(Y2,y2+Δy], where Y1=y1+Δy and Y2=y2-Δy.

[0089] The target area after the first-order sea clutter division is located in the area between the first-order sea clutter of HFSWR, that is, the Doppler unit satisfies the range of y∈[Y1,Y2]. The ship targets inside it are highly concentrated, and the number of targets accounts for most of the entire RD spectrum; the non-target area is located on both sides of the first-order sea clutter of HFSWR, and the Doppler unit satisfies the range of y∈[0,Y1)∪(Y2,N2]. The ship targets inside it are sparsely distributed, and the number accounts for a very low proportion (such as Figure 1 ).

[0090] 1.3 Dataset Construction:

[0091] The multiple RD spectrum data converted into decibel values are divided into a training sample set and a test sample set. The specific division method is as follows:

[0092] The training sample set is represented as Used for parameter optimization and feature learning of neural network models, where X i Represents a single RD spectrum data matrix, N train Indicates the number of RD spectra contained in the training set;

[0093] The test sample set is represented as Used to evaluate the target detection performance of the detector constructed by the neural network model, where N test Indicates the number of RD spectra contained in the test set.

[0094] Step 2: Preprocessing of training RD spectrum data.

[0095] 2.1 RD spectrum target area data replacement:

[0096] In the RD spectrum, ship targets within the target area can significantly affect the network's ability to estimate local clutter power during training. To avoid this effect, this application proposes a target area data replacement method, which uses non-target area data from the same RD spectrum to replace the target area data.

[0097] From the RD spectrum X i Select a non-target area data block of the same size as the target area from the non-target area. Replace the RD spectrum target area data with the selected clutter data block data. The ship target originally contained in the target area is replaced by clutter, and an RD spectrum similar to pure clutter is obtained. (like Figure 2 ). The mathematical process is expressed as follows:

[0098]

[0099] Among them, N d =(Y2-Y1) / 2. This data replacement method can effectively eliminate the interference of the target on the training network, improve the network's ability to learn the statistical characteristics of clutter, and thus achieve accurate local clutter power prediction.

[0100] 2.2 Data Enhancement:

[0101] In order to further improve the generalization ability of the model, the generated pure clutter RD spectrum Data enhancement is performed through horizontal translation and horizontal flipping operations to obtain the same data as the processed data. If the translation distance is Δy, then the data It can be expressed as:

[0102]

[0103] Step 3: random masking.

[0104] This application randomly masks the RD spectrum data to encourage the neural network to learn the knowledge of the clutter distribution in the RD spectrum. Figure 2 The specific implementation process is as follows:

[0105] 3.1 Mask center position sampling:

[0106] After preprocessing, the data Randomly select N in the range of N1×N2 mask coordinates As the center position of the random mask, its coordinate range is a j ∈[1,N1],b j ∈[1,N2].

[0107] This random sampling mechanism ensures that the network can learn the clutter characteristics of each region of the RD spectrum and avoids the learning blind spots caused by fixed masks.

[0108] 3.2 Mask matrix generation:

[0109] Generating and preprocessing data All-1 fundamental matrix of the same dimension I is masked at each mask center position, i.e., the coordinates (a j ,b j ) is used as the center to generate a q×q all-zero mask (q is a preset size, which is similar to the typical target size in the HFSWR RD spectrum). The random mask matrix is obtained

[0110]

[0111] 3.3 Mask data shielding:

[0112] The random mask matrix M i With preprocessed data Multiply the points to generate the RD spectrum data after masking

[0113] Step 4: U-Net network self-supervised training.

[0114] This application proposes a two-stage optimized U-Net training strategy based on fine-tuning the optimal model, and achieves unsupervised high-precision local clutter power prediction through a self-supervised learning mechanism. The specific implementation is as follows:

[0115] 4.1 Network prediction process:

[0116] Define the U-Net network (structure as Figure 3 ) is the mathematical model of f θ(·), where θ represents the set of parameters that can be optimized in the network. The masked data (generated in step 3) Input network, output prediction result Y i pred :

[0117]

[0118] Predicted value Y i pred Contains the network's estimation of the local clutter power in the obscured area, whose accuracy is optimized through a subsequent loss function.

[0119] 4.2 Loss Function:

[0120] The mean square error (MSE) is used as the training loss function to measure the prediction result Y i pred and the true value The difference between the data (i.e., the data that is not masked after preprocessing) is expressed as follows:

[0121]

[0122] By minimizing the loss function, the network can gradually optimize the network parameters θ, thereby improving the prediction accuracy of the local clutter power in the shielded area.

[0123] 4.3 Fine-tune the optimal model training strategy:

[0124] This method adopts a two-stage optimization strategy of fine-tuning the optimal model to enable the network to accurately grasp the clutter distribution characteristics in complex environments. The specific implementation process is as follows:

[0125] 1. Basic Training Phase: 500 rounds of iterative training are performed. Each round uses the backpropagation algorithm to calculate the gradient of the loss function with respect to the network parameters, and the Adam optimizer is used to iteratively update the network parameters. After each round, the optimal model is selected based on a manually constructed validation set. This process enables the network to learn basic knowledge about the clutter distribution in target and non-target areas.

[0126] First, a simulation verification set is constructed. To simulate the target characteristics of the real verification set, real target unit data is extracted from the real radar RD spectrum as the injection target template. N random T The real target template is superimposed on the center of the simulated target by adjusting the local signal-to-noise ratio (SNR) to obtain the simulated RD spectrum, and then the artificial verification set is constructed.

[0127] Then the optimal model is screened using the simulation verification set. Under the condition of keeping constant, select the probability of detecting the simulated target The highest model is taken as the best performing model. and The mathematical representation of is as follows:

[0128]

[0129] in, Indicates the number of simulated targets detected, Indicates the total number of Doppler cells in the non-target area, Indicates the number of false alarm points in the non-target area.

[0130] 2. Fine-tuning training phase: Load the optimal model parameters from the basic training phase and fix the parameters of all batch normalization layers in the U-Net network to prevent changes in the parameter distribution. Continue fine-tuning the network for 1000 rounds of iterative training to optimize the remaining trainable parameters, ultimately achieving a U-Net capable of performing highly accurate local clutter power prediction. This process further optimizes the network's adaptability to complex environments.

[0131] The network inference process (target detection process) of ULCPE-CFAR includes the following steps:

[0132] Step 1: Peak detection.

[0133] In order to reduce the computational complexity of the algorithm for local clutter power estimation for each unit in the RD spectrum, this application first uses peak detection to screen the units to be detected based on the peak characteristics of the target in the HFSWR RD spectrum. The specific implementation steps are as follows:

[0134] The test set to be tested RD spectrum after decibel value processing X i Perform local extreme value detection and filter all peak points. Single RD spectrum X i within The peak point coordinates constitute a set According to the peak point position, the binary peak point position label matrix corresponding to each RD spectrum can be obtained

[0135]

[0136] Through the above steps, the peak point screened out is the unit under test (CUT) that needs to be tested in the RD spectrum.

[0137] Step 2: Group masking.

[0138] To prevent the loss of background information after masking due to the over-dense CUT (peak point), which in turn affects the network's accuracy in predicting the local clutter power at the masked location, this application designs a group masking strategy. The specific implementation steps are as follows:

[0139] 2.1 Peak point grouping:

[0140] The RD spectrum X i The CUT is randomly divided into D non-overlapping subgroups. The RD spectrum X is obtained. i Corresponding sets of multiple CUT position label matrices The CUT label matrix P of the jth group ij It can be expressed as:

[0141]

[0142] In the above formula, P ij represents the CUT label matrix divided into the jth group, represents the CUT coordinate set divided into the j-th subgroup, Indicates the number of CUTs divided into the j-th subgroup.

[0143] 2.2 Group mask generation:

[0144] Taking group j as an example, generate a All-1 fundamental matrix of the same dimension In I, the CUT positions Generate a full 0 mask of size q×q for each center and get the corresponding mask matrix Its mathematical expression is:

[0145]

[0146] Then we get a set of D CUT mask matrices

[0147] 2.3 Group mask data shielding:

[0148] The RD spectrum X i Respectively with each set of mask matrices Multiply the inner samples to get the data set after each group of masks

[0149]

[0150] Step 3: U-Net network inference (local clutter power estimation).

[0151] Data sets after masking of each group All internal samples are passed through the trained U-Net network f θ (·) is processed to obtain the network output result set Y for local clutter power estimation at the mask i pred :

[0152]

[0153] Step 4: local signal-to-noise ratio estimation.

[0154] This application designs a method to accurately predict the local signal-to-noise ratio (SNR) at the CUT based on the local clutter power estimation results and original power at the CUT. The specific implementation steps are as follows:

[0155] 4.1 Calculation of local signal-to-noise ratio matrix of each group:

[0156] Taking the jth group as an example, first calculate the prediction results at each group mask and the true value X i The difference between E ij :

[0157]

[0158] Where I represents a matrix of size N1×N2 with all 1s. The difference here is E ij Indicates the real power X of the unit to be detected (at the mask) i and local clutter predicted power The difference between.

[0159] Then the CUT position that is divided into the group As the center, the estimated error matrix E of the nearby p×p area is ij The accumulated result is the local SNR estimation result of the peak point. In this way, the local SNR estimation matrix of the jth group can be obtained:

[0160]

[0161] Where k = (p-1) / 2. Then we get the set of D peak point local SNR estimation matrices:

[0162] 4.2 Total local signal-to-noise ratio matrix calculation:

[0163] Accumulate the local SNR estimation matrices corresponding to the peak points of group D to obtain the RD spectrum X i The total local SNR estimation matrix

[0164]

[0165] Step 5: Target detection.

[0166] For the total local signal-to-noise ratio estimation matrix For each unit in the , the higher the local signal-to-noise ratio, the more likely the unit is to be the target. Compare with the threshold T: If the unit to be detected is It indicates that there is a target at (x, y) in the RD spectrum; and if the unit to be detected is This indicates whether the target exists at (x, y) in the RD spectrum. This process can be expressed by the following formula:

[0167]

[0168] Where H0 indicates that there is no target at (x, y), and H1 indicates that there is a target at (x, y). In this process, the false alarm probability is controlled by adjusting the detection threshold T to achieve adaptive CFAR detection.

[0169] Through the above steps, the present application can achieve high detection performance and strong generalization performance of the algorithm by utilizing the strong local clutter power prediction ability of the neural network in complex non-uniform environments without the need for manual labeling.

[0170] Finally, the effectiveness of the algorithm is verified by measuring the RD spectrum data of single-beam and multi-beam HFSWR.

[0171] In order to discuss the detection performance of the algorithm, the performance evaluation is first carried out using the single-beam HFSWR measured data. Figure 4 The ROC curves of the proposed ULCPE-CFAR and the traditional CFAR detector used in the HFSWR field for single-beam HFSWR measured data are shown. As can be seen from the figure, the detection performance of ULCPE-CFAR is significantly better than that of other comparison algorithms at various false alarm probabilities. fa =1×10 -4 When ULCPE-CFAR is compared with TP-CFAR, the probability of finding P d An increase of approximately 30.89%. Figure 5 Shows when P fa =1×10 -4 When each algorithm detects the P of the single-beam HFSWR measured data, d -SNR curve. Where SNR represents the local signal-to-noise ratio of the target. As can be seen from the figure, the detection performance of ULCPE-CFAR is better than that of other comparison algorithms at each SNR. When SNR = 15dB, the detection probability of ULCPE-CFAR is P higher than that of TP-CFAR. dAn increase of approximately 50.3%.

[0172] The algorithm performance is then evaluated using multi-beam HFSWR measured RD spectrum data. Figure 6 Shows P fa =1×10 -4 The target tracks detected by ULCPE-CFAR and TP-CFAR for the multi-beam HFSWR measured data. As can be seen from the figure, the number of tracks detected by ULCPE-CFAR is significantly more than that of TP-CFAR. At the same time, ULCPE-CFAR has better detection performance than TP-CFAR in the close-range multi-target background. For further quantitative analysis, Table 1 shows the statistical results of the target point detection probability and the track detection probability after tracking of the above detection results. When P fa =1×10 -4 When compared to TP-CFAR, the target point detection probability of ULCPE-CFAR increased by approximately 11.9%, and the target point detection probability of ULCPE-CFAR increased by 16.7%. These test results show that both single-beam and multi-beam field measurements show that ULCPE-CFAR significantly outperforms traditional methods in target detection. This verifies that the proposed method can achieve adaptive detection and high detection performance in complex and non-uniform field measurement environments, demonstrating excellent performance in actual radar target detection tasks.

[0173] Table 1 Comparison of detection probability of multi-beam HFSWR measured data

[0174]

[0175] In addition, to discuss the generalization of the algorithm, Figure 7 Shows P fa =1×10 -4 When P is used, the target tracks detected by ULCPE-CFAR and the supervised intelligent radar target detector using the U-Net network for the multi-beam HFSWR measured data are the same. The supervised intelligent radar target detector is trained only with labeled single-beam data. As can be seen from the figure, the number of tracks detected by ULCPE-CFAR is significantly greater than that of the supervised intelligent detector. At the same time, ULCPE-CFAR has better detection performance for long-range targets. According to the quantitative detection results in Table 1, when P fa =1×10 -4When compared to supervised intelligent detectors, the target point detection probability of ULCPE-CFAR is improved by about 24.5%, and the target point detection probability of ULCPE-CFAR is improved by 25.8% compared to supervised intelligent detectors. The above detection results show that compared to supervised intelligent object detectors, ULCPE-CFAR can utilize large-scale data for training by eliminating the need for manual annotation of large-scale datasets. This advantage makes it more suitable for actual radar target detection systems.

[0176] The proposed ULCPE-CFAR detector effectively addresses the limitations of traditional CFAR algorithms and supervised intelligent detectors in practical radar target detection applications. It not only maintains high detection performance in complex and non-uniform environments, but also has strong generalization for time-varying data, and therefore has important engineering value.

[0177] The present application utilizes a neural network to accurately estimate the local clutter power at the unit to be detected, thereby realizing CFAR detection in a complex non-uniform background. During the training process, the technical solution of the present application effectively reduces the detection performance loss of the algorithm under unknown non-uniform distribution through a training strategy of fine-tuning the optimal model. During the reasoning process, the technical solution of the present application prevents the loss of background information after masking due to the excessive density of the units to be detected, thereby improving the prediction accuracy of the local clutter power at the mask. Compared with the traditional CFAR detection method, the present application improves the detection accuracy in complex non-uniform measured data.

[0178] In addition, this application is based on self-supervised learning of pure clutter data, avoiding the use of real manually labeled data for training. In actual use, large-scale unlabeled data can be used for training. Compared with the intelligent radar target detection algorithm, the method improves the generalization of the method to the measured data.

[0179] It should be noted that the specific embodiments are merely explanations and illustrations of the technical solutions of the present invention and cannot be used to limit the scope of protection. Any minor changes made based on the claims and description of the present invention shall still fall within the scope of protection of the present invention.

Claims

1. Radar unsupervised neural network CFAR detection method based on local clutter power estimation, characterized by The following steps are involved: Step 1: Obtain the sea surface reflected echo signal to be identified, and obtain the RD spectrum data matrix X of the sea surface reflected echo signal; Step 2: Perform local extreme value detection on the RD spectrum data matrix X of the echo signal reflected from the sea surface to determine all peak points; Step 3: Randomly divide all peak points into D non-overlapping groups; Step 4: Mask each group to obtain the RD spectrum data mask matrix, and input the RD spectrum data mask matrix into the trained neural network to obtain the prediction result of the output RD spectrum data matrix That is, the prediction result of the local clutter power at the mask; Step 5: Use RD spectrum data matrix X and prediction results The difference between them is used to calculate the local signal-to-noise ratio estimation result S at each group of masks ij ; Step 6: Repeat steps 4 and 5 to obtain D groups of local SNR estimation results at the mask, and accumulate the estimation results of the D groups to obtain the accumulated result, i.e. the local SNR estimation result of the peak point. Then, the local SNR estimation result of the peak point is compared with the threshold set by the false alarm probability to determine whether there is a target at each peak point, thereby achieving CFAR detection; The trained neural network is obtained by the following steps: Step A: Acquire the sea surface reflected echo signal and obtain the RD spectrum data matrix of the sea surface reflected echo signal; Step B: distinguishing the target area and non-target area in the RD spectrum data matrix; Step C: Replace the target area in the RD spectrum data matrix with a non-target area of the same size to obtain a pure clutter RD spectrum data matrix; Step D: perform data enhancement on the pure clutter RD spectrum data matrix; Step E: Perform random masking on the pure clutter RD spectrum data matrix after data enhancement to obtain the RD spectrum data mask matrix Step F: Mask the RD spectrum data to the matrix As input, the pure clutter RD spectrum data matrix after data enhancement is taken as output, and the neural network is trained by fine-tuning the optimal model training strategy to obtain a trained neural network.

2. The radar unsupervised neural network CFAR detection method based on local clutter power estimation according to claim 1 is characterized in that In step B, the target area and the non-target area in the RD spectrum data matrix are distinguished by obtaining the positive and negative Doppler frequencies corresponding to the Bragg peaks in the first-order sea clutter. The positive and negative Doppler frequencies corresponding to the Bragg peaks in the first-order sea clutter are expressed as: Where f is the radar carrier frequency, g is the acceleration due to gravity, and c is the speed of light.

3. The radar unsupervised neural network CFAR detection method based on local clutter power estimation according to claim 2 is characterized in that The pure clutter RD spectrum data matrix is expressed as: Y1=y1+Δy / 2 Y2=y2-Δy / 2 Among them, N d =(Y2-Y1) / 2, X i Represents a single RD spectrum data matrix, N d , Y1, y1, Y2 and y2 represent intermediate variables, Δy represents the number of Doppler units occupied by the unilateral first-order sea clutter, f single It represents the frequency resolution corresponding to a single Doppler unit, and N2 represents the total number of Doppler units in the RD spectrum.

4. The radar unsupervised neural network CFAR detection method based on local clutter power estimation according to claim 1 is characterized in that The masking in step E is a random masking, specifically: Step E1: Obtain the pure clutter RD spectrum data matrix after data enhancement Where N1 is the total number of range cells of the RD spectrum, and N2 is the total number of Doppler cells of the RD spectrum; Step E2: Randomly select N in the range of N1×N2 mask coordinates As the center position of the random mask, its coordinate range is a j ∈[1,N1],b j ∈[1,N2]; Step E3: Generate a basic matrix with the same dimension as the pure clutter RD spectrum data matrix after data enhancement All elements in the basic matrix are 1; Step E4: In the basic matrix I, for each coordinate (a j ,b j ) Generate a q×q all-0 mask for the center; Step E5: Repeat step E4 until all coordinates are processed and a random mask matrix is obtained. Random mask matrix M i Expressed as: Where q is the preset mask size; Step E6: Random mask matrix M i Perform dot multiplication with the pure clutter RD spectrum data matrix after data enhancement to generate the RD spectrum data mask matrix.

5. The radar unsupervised neural network CFAR detection method based on local clutter power estimation according to claim 4 is characterized in that The masking in step 3 is group masking, specifically: Step 31: Generate a basic matrix with the same dimension as the pure clutter RD spectrum data matrix after data enhancement All elements in the basic matrix are 1; Step 32: In the basic matrix I, the position of each peak point Generate a q×q all-0 mask for the center; Step 33: Repeat step 32 until all peak points are processed and the mask matrix is obtained. M ij Expressed as: Step 34: Repeat steps 32 and 33 to obtain a set of mask matrices for D groups. Step 35: Set the mask matrix Perform dot multiplication with the pure clutter RD spectrum data matrix after data enhancement to generate the RD spectrum data mask matrix.

6. The radar unsupervised neural network CFAR detection method based on local clutter power estimation according to claim 5 is characterized in that The local signal-to-noise ratio estimation results S at each group of masks ij Expressed as: Where k = (p-1) / 2, p is the preset size, usually smaller than the mask size q, k, a, b represent intermediate variables, P ij Represents the position label matrix of all peak points in group j, E ij Represents the prediction results of group j Each mask and RD spectrum data matrix X i The difference between X i represents the RD spectrum matrix to be detected, represents the j-th group prediction result, M ij represents the j-th group mask matrix.

7. The radar unsupervised neural network CFAR detection method based on local clutter power estimation according to claim 6 is characterized in that The local signal-to-noise ratio estimation result of the peak point is expressed as:

8. The radar unsupervised neural network CFAR detection method based on local clutter power estimation according to claim 7 is characterized in that The loss function of the neural network is: Among them, N train represents the number of RD spectra to be trained, Represents the pure clutter RD spectrum data matrix after data enhancement, Y i pred (m,n) represents the prediction result.

9. The radar unsupervised neural network CFAR detection method based on local clutter power estimation according to claim 1 is characterized in that The training neural network is trained by fine-tuning the optimal model training strategy, specifically: first, a fixed number of basic training phases are performed, and the current optimal model is saved as the basic model using the manual verification set. Then, the optimal model obtained in the basic training phase is loaded and the batch normalization layer parameters are fixed. The unfrozen layers are fine-tuned and iteratively trained. The final network model with the best detection performance is obtained through multiple rounds of parameter optimization. The artificial validation set is constructed by injecting the real target in the RD spectrum of the HFSWR into the target area of the validation set RD spectrum by adjusting the signal-to-noise ratio; The optimal model is obtained by the following steps: False alarm rate of passing non-target area Under the condition of keeping constant, select the probability of detecting the simulated target The highest model is taken as the best performing model, that is, the trained neural network: in, Indicates the number of simulated targets detected, Indicates the total number of Doppler cells in the non-target area, Indicates the number of false alarm points in the non-target area.

10. The radar unsupervised neural network CFAR detection method based on local clutter power estimation according to claim 1 is characterized in that The data enhancement includes horizontal translation and horizontal flipping.

Images