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 application relates to the technical field of radar detection, in particular to a radar unsupervised neural network CFAR detection method based on local clutter power estimation. BACKGROUND

[0002] In recent years, high frequency surface wave radar (HFSWR) is widely used in ocean monitoring, but its detection background is interfered by ionospheric clutter, sea clutter, ground clutter and other interference factors, which leads to great challenges for target detection.

[0003] The constant false alarm rate (CFAR) detector estimates the local clutter power (LCPE) of the to-be-detected unit by using the reference unit in real time, dynamically adjusts the detection threshold, and is widely used in actual radar systems. In order to adapt to various complex non-uniform scenes, adaptive CFAR detectors are gradually proposed, which can realize more accurate local clutter power estimation in complex non-uniform environments, dynamically adjust the detection threshold, and further improve the detection performance. At present, adaptive CFAR detectors can be divided into two categories according to their local clutter power estimation strategy types. The first type improves the LCPE accuracy by eliminating the interference units in the reference window, such as ADCCA-CFAR, CML-CFAR, etc. The second type selects the optimal strategy based on the difference of the detection background to improve the accuracy of the local clutter power estimation, such as VI-CFAR, RWVI-CFAR, etc. However, the adaptive CFAR detector still has the problem of low accuracy of local clutter power estimation in complex non-uniform environments, which further leads to the problem of low accuracy of CFAR detection. SUMMARY

[0004] The purpose of the present application is to solve the problem of low accuracy of local clutter power estimation in complex non-uniform environments of existing adaptive CFAR detectors, which further leads to the problem of low accuracy of CFAR detection, and to provide a radar unsupervised neural network CFAR detection method based on local clutter power estimation.

[0005] The technical scheme adopted by the present application to solve the above technical problem is:

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

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

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

[0009] ​Step three: randomly divide all peak points into D non-overlapping groups;

[0010] Step four: mask each group to obtain an RD spectrum data mask hiding matrix, and input the RD spectrum data mask hiding 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 five: calculate the local signal-to-noise ratio estimation result of each group mask by using the difference between the RD spectrum data matrix and the prediction result ;

[0012] Step six: repeat steps four and five to obtain the local signal-to-noise ratio estimation result of the mask of the D groups, and accumulate the estimation results of the D groups to obtain the accumulation result, that is, the local signal-to-noise ratio estimation result of the peak point. Then, according to the comparison between the local signal-to-noise ratio estimation result of the peak point and the threshold set by the false alarm probability, the presence or absence of the target at each peak point is determined to realize CFAR detection;

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

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

[0015] Step B: distinguish the target area and the 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: data augmentation is performed on the pure clutter RD spectrum data matrix;

[0018] Step E: randomly mask the data augmented pure clutter RD spectrum data matrix to obtain an RD spectrum data mask hiding matrix ;

[0019] Step F: input the RD spectrum data mask hiding matrix , and output the data augmented pure clutter RD spectrum data matrix, train the neural network by fine-tuning the optimal model training strategy, and obtain the trained neural network.

[0020] ​Further, the step B distinguishes the target region and the non-target region in the RD spectrum data matrix by obtaining the positive and negative Doppler frequencies corresponding to the Bragg peak in the first-order sea clutter, which are expressed as:

[0021] ,

[0022] wherein, represents the radar carrier frequency, represents the gravity acceleration, is the speed of light.

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

[0024] ,

[0025] ,

[0026] ,

[0027] ,

[0028] ,

[0029] wherein, , represents a single RD spectrum data matrix, , , , and represents an intermediate variable, represents the number of Doppler units occupied by the single first-order sea clutter, represents the frequency resolution corresponding to a single Doppler unit, represents the total number of Doppler units of the RD spectrum.

[0030] Further, the mask masking in the step E is a random mask masking, specifically:

[0031] Step E1: obtaining the pure clutter RD spectrum data matrix after data enhancement , wherein is the total number of distance units of the RD spectrum, is the total number of Doppler units of the RD spectrum;

[0032] Step E2: randomly selecting coordinates as the center positions of the random mask, and the coordinate values range from ​;

[0033] Step E3: Generate a fundamental matrix with the same dimensions as the data-enhanced pure clutter RD spectrum data matrix. The fundamental matrix contains all elements that are 1.

[0034] Step E4: In the fundamental matrix In, with each coordinate Generate from center A zero-mask of size;

[0035] Step E5: Repeat step E4 until all coordinates have been processed, resulting in a random mask matrix. Random mask matrix Represented as:

[0036] ,

[0037] in, Preset mask size;

[0038] Step E6: Apply the random mask matrix The RD spectrum data matrix of the data augmented pure clutter is multiplied by a dot to generate an RD spectrum data masking matrix.

[0039] Furthermore, the masking in step three is group masking, specifically:

[0040] Step 31: Generate a fundamental matrix with the same dimensions as the data-enhanced pure clutter RD spectrum data matrix. The fundamental matrix contains all elements that are 1.

[0041] Step 32: In the basic matrix In the middle, based on the position of each peak point Generate from center A zero-mask of size;

[0042] Step 33: Repeat step 32 until the positions of all peak points have been processed, and the mask matrix is ​​obtained. , Represented as:

[0043] ,

[0044] Steps 3 and 4: Repeat steps 3 and 3 to obtain the set of mask matrices for D groups. ;

[0045] Step 35: Set the mask matrix The RD spectrum data matrix of the data augmented pure clutter is multiplied by a dot to generate an RD spectrum data masking matrix.

[0046] Further, the local signal-to-noise ratio estimation results of each group of masks are represented as:

[0047] ,

[0048] ,

[0049] ,

[0050] wherein, , is a preset size, usually smaller than the mask size , , , represents an intermediate variable, represents the position label matrix of all peak points of the jth group, represents the prediction result of the jth group the difference between each mask and the RD spectrum data matrix , represents the RD spectrum matrix to be detected, represents the prediction result of the jth group, represents the jth group of mask matrices.

[0051] Further, the local signal-to-noise ratio estimation results of the peak points are represented as:

[0052] ,

[0053] Further, the loss function of the neural network is:

[0054] ,

[0055] wherein, represents the number of RD spectra to be trained, represents the pure clutter RD spectrum data matrix after data enhancement, represents the prediction result.

[0056] Further, the training of the neural network is performed by fine-tuning the optimal model training strategy, specifically: first, through a fixed number of basic training stages, the current optimal model saved by using the artificial verification set is used as the basic model, 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 and iteratively trained, and through multiple rounds of parameter optimization, the final network model with the optimal detection performance is obtained;

[0057] The artificial verification set is constructed by adjusting the signal-to-noise ratio to inject the target area of the verification set RD spectrum with the real target in the HFSWR RD spectrum;

[0058] The optimal model is obtained by the following steps:

[0059] Through the non-target region false alarm rate Under the condition of keeping constant, select the simulation target detection probability The highest model as the performance optimal model, that is, the trained neural network:

[0060] ,

[0061] ,

[0062] Wherein, Indicates the number of detected simulation targets, Indicates the total number of Doppler units in the non-target area, Indicates the number of false alarm points in the non-target area.

[0063] Further, the data enhancement includes horizontal translation and horizontal flip.

[0064] The beneficial effects of the present application are:

[0065] The present application accurately estimates the local clutter power at the unit to be detected by using a neural network, and the technical solution of the present application prevents the loss of background information after masking caused by too dense CUT (peak point), thereby improving the prediction accuracy of the local clutter power at the mask, and ultimately improving the CFAR detection accuracy in a complex non-uniform background.

[0066] Moreover, the present application avoids using real artificial labeled data for training, and can use large-scale unlabeled data for training in actual use, thereby improving the generalization of the method to measured data compared with an intelligent radar target. DETAILED DESCRIPTION

[0067] Figure 1 It is a schematic diagram of measured HFSWR RD spectrum data;

[0068] Figure 2 It is a detection flowchart of ULCPE-CFAR;

[0069] Figure 3 It is a schematic diagram of U-Net network structure;

[0070] Figure 4 It is a ROC curve schematic diagram of detection results of different algorithms for single-beam HFSWR measured data;

[0071] Figure 5 When , the PR curve of detection results of different algorithms for single-beam HFSWR measured data is shown in the following figure: ​

[0072] Figure 6 Figure 4 is a comparison chart of target track detection of ULCPE-CFAR and TP-CFAR in the measured data of multi-beam HFSWR.

[0073] Figure 7 Figure 5 is a comparison chart of target track detection of ULCPE-CFAR and supervised intelligent target detector in the measured data of multi-beam HFSWR. DETAILED DESCRIPTION

[0074] It should be particularly pointed out that the various embodiments disclosed in the present application can be combined with each other without conflict.

[0075] Embodiment one: the radar unsupervised neural network CFAR detection method based on local clutter power estimation in the embodiment includes:

[0076] Step one, self-supervised training process of ULCPE-CFAR. The HFSWR range-doppler spectrum (RD spectrum) to be trained is processed, the target area is shielded by using the clutter data, and the pure clutter data is generated by data enhancement to remove the target unit; the pure clutter data is randomly masked to construct a self-supervised training sample; the training sample is input into the U-Net network to be trained, the MSE loss between the pure clutter data and the network output is calculated, the network is iteratively optimized and trained, and the network is prompted to have the ability to predict the local clutter power at the mask; in each round of network training, the optimal model is selected and saved by using the validation set; the network is trained using the strategy of fine-tuning the optimal model, and the U-Net network after training can realize the estimation of the local clutter power at the mask.

[0077] Step two, grouping reasoning process of ULCPE-CFAR. The peak value of the HFSWR RD spectrum to be tested is detected, and the detection unit (peak point) to be detected is divided into multiple groups; the RD spectrum is masked at each group of detection units to generate a sample after grouping masking; the masked sample is input into the trained U-Net network to obtain the network output result after local clutter power prediction; the error between the original RD spectrum and the local clutter power prediction result corresponding to each mask is calculated, and the error at each detection unit is accumulated to estimate the local signal-to-noise ratio matrix; the threshold decision is made on the local signal-to-noise ratio estimation matrix to realize the adaptive CFAR detection at the detection unit.

[0078] 1. Figure 2 The detection flowchart of the applied ULCPE-CFAR is shown. The method includes network training process and network reasoning process.

[0079] The network training process of ULCPE-CFAR includes the following steps:

[0080] Step 1: Radar data acquisition and preliminary processing.

[0081] 1.1 Data Acquisition and Data Field Processing:

[0082] A high-frequency ground wave radar (HFSWR) system transmits electromagnetic wave signals in a specific frequency band and receives the echo signals reflected from the sea surface via an antenna array. These echo signals are then processed to generate the raw range-Doppler spectrum (RD spectrum) data matrix. (like Figure 1 As shown), where The total number of distance units in the RD spectrum. This represents the total number of Doppler elements in the RD spectrum. This data is the result after square-law detection, and each value in the RD spectrum corresponds to a power value.

[0083] To facilitate subsequent analysis and processing, the power values ​​of the original RD spectrum were converted to decibels (dB) using a logarithmic relationship. The specific conversion formula is as follows:

[0084] (1)

[0085] In the formula This is the RD spectrum matrix converted to decibel values.

[0086] 1.2 RD spectral region division:

[0087] Based on the detection characteristics analysis of HFSWR, the speed of ship targets is usually significantly lower than the theoretical detection speed limit of the system. This characteristic causes the targets in the HFSWR RD spectrum data to be concentrated within a specific Doppler interval.

[0088] To more effectively analyze RD spectrum data, this application divides the RD spectrum data into two regions—target region and non-target region—based on the theoretical location of first-order sea clutter. The positive and negative Doppler frequencies corresponding to the Bragg peaks within the first-order sea clutter can be calculated using the following formula:

[0089] (2)

[0090] In the above formula, Indicates the radar carrier frequency. Represents gravitational acceleration. The speed of light. Based on the above formula, the number of Doppler units corresponding to the Bragg peak theory can be obtained. , :

[0091] (3)

[0092] in, Frequency resolution of a single Doppler bin. Combining the characteristics of unilateral first-order Bragg scatter in RD spectrum with one Doppler bin, the Doppler range of HFSWR first-order Bragg scatter can be expressed as , where , , .

[0093] The target area divided by the first-order Bragg scatter is located in the region between the first-order Bragg scatter of HFSWR, that is, the Doppler bin satisfies the range of , and the ship targets inside show a high degree of concentration, and the number of targets accounts for most of the RD spectrum; the non-target area is located in the region on both sides of the first-order Bragg scatter of HFSWR, and the Doppler bin satisfies the range of , and the ship targets inside are sparse, and the number of targets accounts for a very low proportion (such as Figure 1 ).

[0094] 1.3 Construction of data set:

[0095] A plurality of RD spectrum data processed into decibel values is divided into a training sample set and a test sample set, and the specific division method is as follows:

[0096] The training sample set is represented as , which is used for parameter optimization and feature learning of the neural network model, where represents a single RD spectrum data matrix, represents the number of RD spectrums contained in the training set;

[0097] The test sample set is represented as , which is used to evaluate the target detection performance of the detector constructed by the neural network model, where represents the number of RD spectrums contained in the test set.

[0098] Step 2, training RD spectrum data preprocessing.

[0099] 2.1 RD spectrum target area data replacement:

[0100] In the RD spectrum, the ship targets in the target area will significantly affect the network's ability to estimate the local scatter power during training. In order to avoid this influence, the present application proposes a target area data replacement method, that is, using the non-target area data in the same RD spectrum to replace the target area data.

[0101] A block of non-target area data with the same size as the target area is selected from the non-target area of the RD spectrum . The target area data of the RD spectrum is replaced with the selected scatter data block. The ship targets originally contained in the target area are replaced by the scatter, and an RD spectrum with approximately pure scatter is obtained (such as Figure 2). The mathematical process is shown as follows:

[0102] (4)

[0103] wherein, By this data replacement method, the interference of targets on the training network can be effectively eliminated, the learning ability of the network for clutter statistical characteristics can be improved, and accurate local clutter power prediction can be realized.

[0104] 2.2 Data augmentation:

[0105] To further improve the generalization ability of the model, the generated pure clutter RD spectrum is augmented by horizontal translation and horizontal flip to obtain the processed data . If the translation distance is , the data can be represented as:

[0106] (5)

[0107] Step 3, random mask masking.

[0108] The present application masks the RD spectrum data randomly to prompt the neural network to learn the clutter distribution knowledge in the RD spectrum. As shown in Figure 2 , the specific implementation process is as follows:

[0109] 3.1 Mask center position sampling:

[0110] Randomly select coordinates in the range of of the preprocessed data as the center positions of the random mask, and the coordinate value range is .

[0111] 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 area caused by fixed mask.

[0112] 3.2 Mask matrix generation:

[0113] Generate a full-1 basic matrix with the same dimension as the preprocessed data . Mask at each mask center position, that is, generate a full-0 mask with size (the preset size is approximately the same as the typical target size in the HFSWR RD spectrum) with the coordinate as the center. The random mask matrix is obtained:

[0114] (6)

[0115] 3.3 Mask data masking:

[0116] The random mask matrix is multiplied with the pre-processed data to generate the mask-masking RD-spectrum data .

[0117] (7)

[0118] Step 4, U-Net network self-supervised training.

[0119] The present application proposes a two-stage optimization U-Net training strategy based on fine-tuning the optimal model, which realizes unsupervised high-precision local clutter power prediction through a self-supervised learning mechanism. The specific implementation is as follows:

[0120] 4.1 Network prediction process:

[0121] The mathematical model of the U-Net network (structure as Figure 3 ) is defined as , where represents the set of parameters that can be optimized in the network. The data masked by the mask (generated in step 3) is input into the network, and the prediction result is output:

[0122] (8)

[0123] The prediction value contains the estimation result of the local clutter power in the masked area by the network, and its accuracy is optimized through the subsequent loss function.

[0124] 4.2 Loss function:

[0125] The mean squared error (MSE) is used as the training loss function to measure the difference between the prediction result and the true value (i.e., the pre-processed data that has not been masked), which is represented as follows:

[0126] (9)

[0127] By minimizing this loss function, the network can gradually optimize the network parameters , thereby improving the prediction accuracy of the local clutter power in the masked area.

[0128] 4.3 Fine-tuning optimal model training strategy:

[0129] The method adopts a two-stage optimization strategy of fine-tuning the optimal model, enabling the network to accurately grasp the clutter distribution characteristics in complex environments. The specific implementation process is as follows:

[0130] 1. Basic training stage: 500 iterations of training are performed, and the gradient of the loss function with respect to the network parameters is calculated using the back propagation algorithm in each round. The network parameters are updated iteratively through the Adam optimizer. After each round, the current optimal model is selected based on the artificially constructed validation set. This process enables the network to learn the basic knowledge of clutter distribution in the target area and non-target area.

[0131] First, the simulation validation set is constructed. To simulate the target characteristics of the real validation set, real target cell data is extracted from the real radar RD spectrum as the injected target template. Randomly select points in the target area of the RD spectrum as the center of the simulation target. The real target template is superimposed on the simulation target center by adjusting the local signal-to-noise ratio (SNR), and the simulation RD spectrum is obtained, and then the artificial validation set is constructed.

[0132] Then, the optimal model is selected using the simulation validation set. Under the condition that the false alarm rate of the non-target area remains constant, the model with the highest simulation target detection probability is selected as the optimal model. and The mathematical representation is as follows:

[0133] (10)

[0134] where represents the number of detected simulation targets, represents the total number of Doppler cells in the non-target area, represents the number of false alarm points in the non-target area.

[0135] 2. Fine-tuning training stage: load the optimal model parameters of the basic training stage, and fix the parameters of all BN layers in the U-Net network to prevent parameter distribution changes. Continue to fine-tune the network for 1000 iterations of training, optimize the remaining trainable parameters, and finally obtain a U-Net network with high-accuracy local clutter power prediction capability. This process further optimizes the network's ability to adapt to complex environments.

[0136] The network inference process (target detection process) of ULCPE-CFAR includes the following steps:

[0137] Step 1, peak detection.

[0138] In order to reduce the computational complexity of the algorithm for local clutter power estimation of each unit in the RD spectrum, the application first uses peak detection to screen the units to be detected according to the characteristics of the target peak value in the HFSWR RD spectrum. The specific implementation steps are as follows:

[0139] The test set to be detected The RD spectrum after the decibel value processing Perform local extreme value detection to screen all peak points. A single RD spectrum 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 :

[0140] (11)

[0141] Through the above steps, the screened peak points are the units to be detected (CUT) in the RD spectrum.

[0142] Step 2, group mask masking.

[0143] In order to prevent the loss of background information after mask masking due to the excessive density of CUT (peak point), which in turn affects the accuracy of the network's local clutter power prediction at the mask, the application designs a group mask masking strategy. The specific implementation steps are as follows:

[0144] 2.1 Peak point grouping:

[0145] Randomly divide the CUT of the RD spectrum into D non-overlapping subgroups. Get the corresponding multi-group CUT position label matrix set of the RD spectrum , wherein the CUT label matrix of the jth group can be expressed as:

[0146] (12)

[0147] In the above formula, denotes the CUT label matrix divided into the jth group, denotes the CUT coordinate set divided into the jth subgroup, denotes the number of CUTs divided into the jth subgroup.

[0148] 2.2 Group mask generation:

[0149] Take the jth group as an example, generate a full 1 basic matrix with the same dimension as the preprocessed data In I, the CUT positions that are divided into the group A full 0 mask of size is generated for each center, resulting in a corresponding mask matrix The mathematical expression is:

[0150] (13)

[0151] Further, a set of D group CUT mask matrices is obtained.

[0152] 2.3 Group mask data masking:

[0153] The RD spectrum is respectively multiplied by the sample points in each group mask matrix set , resulting in a set of data after each group mask masking :

[0154] (14)

[0155] Step 3, U-Net network inference (local clutter power estimation).

[0156] The samples in each group of masked data set are processed through the trained U-Net network , resulting in a set of network output results for local clutter power estimation at the mask :

[0157] (15)

[0158] Step 4, local signal-to-noise ratio estimation.

[0159] According to the local clutter power estimation result at the CUT and the original power, the application designs a method for accurately predicting the local signal-to-noise ratio (SNR) at the unit. The specific implementation steps are as follows:

[0160] 4.1 Each group of local signal-to-noise ratio matrix calculation:

[0161] Take the jth group as an example. First, calculate the difference between the prediction result at each group mask and the true value :

[0162] (16)

[0163] wherein represents a full 1 matrix of size . The difference ​Indicates the actual power of the unit under test (at the mask). With local clutter prediction power The differences between them.

[0164] Subsequently, the CUT position was assigned to that group. Centered on, the surrounding area The estimation error matrix of the region The summation is performed. The resulting summation represents the local SNR estimate at the peak point. In this way, the local SNR estimate matrix for the j-th group can be obtained. :

[0165] (17)

[0166] in, This leads to the set of local SNR estimation matrices for the peak points of group D. .

[0167] 4.2 Calculation of the total local signal-to-noise ratio matrix:

[0168] The local SNR estimation matrices corresponding to the peak points of group D are summed to obtain the RD spectrum. Total local SNR estimation matrix :

[0169] (18)

[0170] Step 5, target detection.

[0171] For the total local signal-to-noise ratio estimation matrix For each cell in the model, a higher local signal-to-noise ratio (SNR) indicates that the cell is more likely to be a target. Therefore, this application requests the local SNR estimation results. With threshold Comparison: If the unit to be detected is located at... This indicates that in the RD spectrum The target exists at the location; while if the unit to be detected is located at... This indicates that in the RD spectrum The question is whether the target exists at a certain location. This process can be represented by the following formula:

[0172] (19)

[0173] in, express The target does not exist. express The target situation exists. During this process, the detection threshold is adjusted. Controlling the false alarm probability enables adaptive CFAR detection.

[0174] Through the above steps, this application can achieve high detection performance and strong generalization performance of the algorithm by utilizing the strong local clutter power prediction capability of neural networks in complex and non-uniform environments without the need for manual labeling.

[0175] Finally, the effectiveness of the algorithm was verified using HFSWR measured RD spectrum data for both single-beam and multi-beam applications.

[0176] To discuss the detection performance of the algorithm, we first used the measured data of single-beam HFSWR to evaluate its performance. Figure 4 The ROC curves of the proposed ULCPE-CFAR and a traditional CFAR detector used in HFSWR are presented for single-beam HFSWR measured data. As can be seen from the figures, the detection performance of ULCPE-CFAR is significantly better than the other comparative algorithms at all false alarm probabilities. At that time, the detection probability of ULCPE-CFAR compared to TP-CFAR It increased by approximately 30.89%. Figure 5 Showing when At that time, each algorithm detected the measured data of single-beam HFSWR. Curve. Among them... This represents the local signal-to-noise ratio of the target. As can be seen from the figure, ULCPE-CFAR exhibits [signal-to-noise ratio] in various [various aspects]. The detection performance of these algorithms is superior to other comparison algorithms. At that time, the detection probability of ULCPE-CFAR compared to TP-CFAR It increased by approximately 50.3%.

[0177] The algorithm's performance was then evaluated using measured RD spectrum data from multi-beam HFSWR. Figure 6 Showing The figure shows the target tracks detected by ULCPE-CFAR and TP-CFAR using multi-beam HFSWR measured data. As can be seen from the figure, ULCPE-CFAR detects significantly more tracks than TP-CFAR, and ULCPE-CFAR exhibits superior detection performance in close-range multi-target backgrounds. For further quantitative analysis, Table 1 presents the statistical results of the target point detection probability and the track detection probability after tracking based on the above detection results. When the target detection probability of ULCPE-CFAR is about 8.2% higher than that of TP-CFAR, and the track detection probability of ULCPE-CFAR is about 12.47% higher than that of TP-CFAR. The above detection results show that both single-beam and multi-beam experimental results show that ULCPE-CFAR has better target detection performance than traditional methods. It is verified that the proposed method can realize adaptive detection with high detection performance in complex non-uniform environments, and has excellent performance in actual radar target detection tasks.

[0178] Table 1 Comparison of detection probability of multi-beam HFSWR measured data

[0179]

[0180] In addition, in order to discuss the generalization of the algorithm, Figure 7 It is shown that When the target detection probability of ULCPE-CFAR is about 8.2% higher than that of TP-CFAR, and the track detection probability of ULCPE-CFAR is about 12.47% higher than that of TP-CFAR. The above detection results show that both single-beam and multi-beam experimental results show that ULCPE-CFAR has better target detection performance than traditional methods. It is verified that the proposed method can realize adaptive detection with high detection performance in complex non-uniform environments, and has excellent performance in actual radar target detection tasks. When the target detection probability of ULCPE-CFAR is about 8.2% higher than that of TP-CFAR, and the track detection probability of ULCPE-CFAR is about 12.47% higher than that of TP-CFAR. The above detection results show that both single-beam and multi-beam experimental results show that ULCPE-CFAR has better target detection performance than traditional methods. It is verified that the proposed method can realize adaptive detection with high detection performance in complex non-uniform environments, and has excellent performance in actual radar target detection tasks.

[0181] The proposed ULCPE-CFAR detector effectively solves the limitations of traditional CFAR algorithms and supervised intelligent detectors in actual radar target detection applications. It not only maintains high detection performance in complex non-uniform environments, but also has strong generalization for time-varying data, so it has important engineering value.

[0182] The application accurately estimates the local clutter power at the to-be-detected unit by using a neural network, and realizes CFAR detection in a complex non-uniform background. In the training process, the technical solution of the application effectively reduces the detection performance loss of the algorithm under unknown non-uniform distribution by using the training strategy of optimal model fine-tuning. In the inference process, the technical solution of the application prevents the loss of background information after the mask shielding caused by the over-dense to-be-detected unit, and further improves the prediction accuracy of the local clutter power at the mask. Compared with the traditional CFAR detection method, the application improves the detection accuracy in complex non-uniform measured data.

[0183] Moreover, the self-supervised learning based on pure clutter data avoids using real human-labeled data for training, and large-scale unlabeled data can be used for training in actual use, thereby improving the generalization of the method to measured data compared with intelligent radar target detection algorithms.

[0184] It should be noted that the specific embodiments are only an explanation and description of the technical solutions of the application, and cannot limit the protection scope. Any partial change made according to the claims and description of the application shall still fall within the protection scope of the application.

Claims

1. Radar unsupervised neural network CFAR detection method based on local clutter power estimation, characterized in that The method comprises the following steps: Step one: obtaining the sea surface reflection echo signal to be identified, and obtaining the RD spectrum data matrix of the sea surface reflection echo signal ; Step two: RD spectrum data matrix of echo signal reflected by sea surface Local extremum detection is performed to determine all peak points; Step three: randomly divide all peak points into D non-overlapping groups; Step four: mask each group to obtain an 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 five: Compute local SNR estimates at each group mask using the difference between the RD spectrum data matrix and the predicted results ;​ Step six: repeat step four and step five to obtain the local signal-to-noise ratio estimation results of the masks of the D groups, and accumulate the estimation results of the D groups to obtain the local signal-to-noise ratio estimation results of the peak points, and then compare the local signal-to-noise ratio estimation results of the peak points with the threshold set by the false alarm probability to determine whether there is a target at each peak point, thereby realizing CFAR detection; The trained neural network is obtained by the following steps: Step A: obtaining a sea surface reflection echo signal and obtaining an RD spectrum data matrix of the sea surface reflection echo signal; Step B: distinguishing target regions and non-target regions in the RD spectrum data matrix; Step C: replacing the target regions in the RD spectrum data matrix with non-target regions of the same size to obtain a pure clutter RD spectrum data matrix; Step D: data enhancement on the pure clutter RD spectrum data matrix; Step E: Random mask masking is performed on the data-enhanced pure clutter RD spectrum data matrix to obtain an RD spectrum data mask masking matrix ; Step F: RD spectrum data mask masking matrix As input, the data enhanced pure clutter RD spectrum data matrix is output, the neural network is trained by fine-tuning the optimal model training strategy, and the trained neural network is obtained.

2. The radar unsupervised neural network CFAR detection method based on local clutter power estimation of claim 1, wherein The target regions and non-target regions in the RD spectrum data matrix are distinguished by obtaining the positive and negative Doppler frequencies corresponding to the first-order Bragg peak in sea clutter, and the positive and negative Doppler frequencies corresponding to the first-order Bragg peak in sea clutter are expressed as: ， wherein denotes the radar carrier frequency, denotes the gravitational acceleration, is the speed of light.

3. The radar unsupervised neural network CFAR detection method based on local clutter power estimation of claim 2, wherein The pure clutter RD spectrum data matrix is expressed as: ， ， ， ， ， wherein, , denotes a single RD spectrum data matrix, , , , and denotes an intermediate variable, denotes the number of Doppler bins occupied by a single-sided 1st order sea clutter, denotes the frequency resolution corresponding to a single Doppler bin, denotes the total number of Doppler bins of the RD spectrum.

4. The radar unsupervised neural network CFAR detection method based on local clutter power estimation of claim 1, wherein The mask masking in step E is random mask masking, specifically: Step E1 : Obtain the data enhanced pure clutter RD spectrum data matrix , wherein is the total number of distance units of the RD spectrum, is the total number of Doppler units of the RD spectrum; Step E2: In the range of 0 to 1 are randomly selected coordinates of the center position of the random mask, the value range of the coordinates is ; Step E3: Generate a base matrix with the same dimension as the data-augmented pure-radar data matrix, all elements of which are 1. , all elements of which are 1; Step E4: In the base matrix generate a full 0 mask of size centered at each coordinate ; Step E5: Repeat Step E4 until all coordinates are processed, resulting in a random mask matrix , random mask matrix is represented as: ， wherein, is a preset mask size; Step E6: Multiply the random mask matrix with the data-augmented pure-clutter RD-spectrum data matrix to generate the RD-spectrum data mask masking matrix.

5. The radar unsupervised neural network CFAR detection method based on local clutter power estimation of claim 4, wherein The mask masking in step three is group mask masking, specifically: Step three one: generate a basic matrix with the same dimension as the data enhanced pure clutter RD spectrum data matrix, all elements of the basic matrix are 1 ; Step three two: In the base matrix generate a full 0 mask of size centered at the position of each peak point ; Step three two: repeat step three one until the position of all peak points is processed, and the mask matrix is obtained , is represented as: ， Step three four: repeat step three two and step three three to get the set of mask matrix of D groups ; Step three five: generate a set of mask matrices Point multiply the data enhanced pure clutter RD spectrum data matrix to generate the RD spectrum data mask shielding matrix.

6. The radar unsupervised neural network CFAR detection method based on local clutter power estimation of claim 5, wherein local signal-to-noise ratio estimation results at the respective groups of masks is represented as: ， ， ， wherein, , is a preset size, usually smaller than the mask size , , , denotes an intermediate variable, denotes a position label matrix of all peak points of the jth group, denotes the prediction result of the jth group the difference between each mask and the RD spectrum data matrix , denotes the RD spectrum matrix to be detected, denotes the prediction result of the jth group, denotes the mask matrix of the jth group.

7. The radar unsupervised neural network CFAR detection method based on local clutter power estimation of claim 6, wherein 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 of claim 7, wherein The loss function of the neural network is: ， wherein, denotes the number of RD spectra to be trained, denotes the data augmented pure clutter RD spectra data matrix, denotes the prediction result.

9. The radar unsupervised neural network CFAR detection method based on local clutter power estimation of claim 1, wherein The trained neural network is trained by fine-tuning the optimal model training strategy, specifically: first, through a fixed number of basic training stages, save the current optimal model as a basic model by using an artificial verification set, then load the optimal model obtained in the basic training stage and fix the batch normalization layer parameters, and fine-tune the iteration training of the unfrozen layers, and obtain the final network model with the best detection performance through multiple rounds of parameter optimization; The artificial verification set is constructed by adjusting the signal-to-noise ratio of the real target in the RD spectrum of the HFSWR to inject the target region of the verification set RD spectrum; The optimal model is obtained by the following steps: By non-target region false alarm rate Selecting the simulation target detection probability under the condition of keeping constant The highest model as the performance optimal model, i.e. the trained neural network: ， ， wherein, represents the number of detected simulation targets, represents the total number of Doppler units in the non-target region, represents the number of false alarm points in the non-target region.

10. The radar unsupervised neural network CFAR detection method based on local clutter power estimation of claim 1, wherein The data enhancement includes horizontal translation and horizontal flip.

Images