FDA radar target positioning method based on Transform neural network

Abstract

The invention relates to the technical field of array signal processing, in particular to a frequency diverse array (FDA) radar target positioning method based on a Transform neural network. In order to solve the problem that the far-field target positioning effect is not ideal, the beam characteristics of a frequency control array radar are deeply analyzed, a Transform network model is combined, the target positioning problem is converted into the nonlinear mapping problem from an echo signal covariance matrix to a target azimuth angle and distance information, and the FDA radar target positioning method based on the Transform neural network is provided. The azimuth angle and distance decoupling is efficiently completed, the azimuth angle and distance information is directly obtained, and the target positioning precision is improved at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of array signal processing, and in particular to an FDA radar target positioning method based on a Transformer neural network. Background Art

[0002] Radar is an electronic detection device. Radar technology has developed rapidly and is widely used in many fields, including military, aviation, maritime, meteorological observation, traffic monitoring, earth science research, and disaster prevention. For example, in the military field, radar is used for reconnaissance, surveillance, and target positioning and tracking; in the civilian field, it can help aircraft and ships avoid collisions, and is also used to detect storms and other weather phenomena in weather forecasts. The research on radar target positioning technology is in full swing. The radar echo signal is processed in different ways and means to obtain the specific location of the target and achieve accurate positioning of the target.

[0003] In the early stages of radar technology, the radar system's interference suppression capability was insufficient. When electromagnetic interference existed in the working environment, the radar's performance would drop significantly. With the advancement of radar technology, especially the development of phased array radar technology, the performance has been significantly improved. Phased array radar achieves rapid scanning and shape change of antenna beams by adjusting the signal phase and amplitude of each unit of the array antenna. With its flexible beam adjustment, powerful interference blocking and high resolution, phased array radar provides technical advantages for expanding radar coverage, improving measurement accuracy and detecting low-altitude visible targets including stealth targets. It has been widely used in target positioning for a long time. However, the application of phased array in radar target positioning has certain limitations: its beam pattern corresponds to different distances at the same azimuth angle, and it is difficult to directly obtain the distance information of the target from the change of beam direction, resulting in distance ambiguity, and it is not possible to achieve the distance measurement and angle measurement of the target at the same time. Specifically, radar target positioning is to estimate the azimuth and distance information of the target's spatial position by processing the radar echo signal.

[0004] In the traditional phased array architecture, it is quite difficult to solve the above problems. FDA radar came into being. The small frequency offset between array elements can produce a beam pattern with azimuth distance dependence that is different from the phased array. Such a coupled beam brings opportunities for target positioning, so FDA radar is widely used in target positioning. FDA radar effectively avoids the large bandwidth and nonlinear problems associated with broadband signals by making each array element emit a single-frequency continuous wave of different frequencies. In addition, only a single linear array is needed to simultaneously achieve the azimuth and distance positioning of the target, effectively simplifying the complexity of the array structure. However, it is precisely because the azimuth distance of the FDA beam pattern is coupled that the azimuth and distance information of the target cannot be directly extracted from the radar echo data by traditional methods. Therefore, exploring how to decouple the azimuth distance from the noisy echo data, extract the target azimuth and distance information, and thus complete the positioning of the target has become a key issue that needs to be solved in the field of array signal processing.

[0005] At present, with the rapid progress of artificial intelligence, especially the widespread application of deep learning methods, outstanding results have been achieved in many fields. The integration and intersection of disciplines have now become the main development direction of scientific research in the future. Deep learning can learn the important information of input data to form a function mapping. Compared with traditional methods, it can learn the effective information of radar echo signals and then directly map them to target-related parameters. It does not require the consistency of input and output data, has excellent feature extraction and generalization capabilities, and can be generalized to various unknown scenarios. Deep learning methods provide an innovative solution to overcome the core problems in radar signal processing.

[0006] In summary, how to combine the Transformer network model with the FDA radar to achieve high-precision target positioning has become an important topic with great application value. Summary of the invention

[0007] In order to solve the above problems, the present invention provides an FDA radar target positioning method based on Transformer neural network. The specific technical solution is as follows:

[0008] A FDA radar target positioning method based on Transformer neural network includes the following steps:

[0009] Step S1: Construct a data set. Build a single-base FDA radar transceiver array signal model, set various parameters of the signal model, obtain the FDA radar echo signal matrix, calculate the FDA radar echo matrix covariance, obtain the covariance matrix, separate the real from the virtual and normalize the covariance matrix, and obtain the network input data;

[0010] Step S2: Build a Transformer network model. The model structure mainly consists of data input, Transformer module, feedforward network and data output part. The data input part includes data segmentation, position encoding and linear projection;

[0011] Step S3: training model and outputting results. The built Transformer network model is trained using the training set of input data. After the model training is completed, the model is tested using the test set of input data. Finally, the azimuth and distance of the far-field target are output to complete the estimation of the target position.

[0012] Preferably, the step S1 specifically includes the following steps:

[0013] In step S11, the selected FDA is a uniform linear array containing 11 array elements, each array element is arranged along a straight line, and the distance between every two adjacent array elements is λ / 2.

[0014] Step S12, consider that there are 24 far-field narrowband incoherent target signal sources with an initial carrier frequency of 10GHZ and a frequency offset of 5kHz incident on the array. Assume that the signal-to-noise ratio is a constant 10dB, the number of snapshots is a constant 100, and the noise is independent and obeys a zero-mean Gaussian distribution of white noise. The target azimuth angle range is [5°, 40°], the change interval is 5°, and there are 8 angles in total; the target distance range is [4km, 12km], the change interval is 4km, and there are 3 distances in total. The combination of 8 angles and 3 distances results in 24 target positions.

[0015] R=E]Y(t)Y H (t)] (1)

[0016] R′=[R1′,R2′,R3′,…,R′ N ] (2)

[0017] Among them, Y(t)=[y1(t),y2(t),···,y N (t)] T is the received signal vector, H represents the Hermitian operator, E[·] represents the mathematical expectation, R is the echo signal covariance matrix, and R′ represents the data obtained after the covariance matrix R is preprocessed by performing operations such as virtual-real separation and normalization.

[0018] Step S13, substitute the above 24 targets into equations (1) and (2) in turn for calculation to obtain 24 sets of data. It should be noted that the real and imaginary parts of the signal are separated in the process of constructing the data, so the number of experimental samples is twice the size of the beam space. Repeat the same method 100 times to obtain 2424 sets of data as the input sample data of the network. Select the training set and the validation set at a ratio of 100:1, and select the results of the first 100 calculations from the 24 targets, that is, 2400 groups, a total of 580,800 sample data, as the input data for network model training. Then select the results of the last calculation from the 24 targets, that is, 24 groups, a total of 5808 sample data, as the input data for verifying the performance of the network model. The data used in network training are all equipped with the corresponding target true azimuth and distance values ​​as training labels. The validation data is predicted by the network model to obtain the target azimuth and distance estimates. The actual azimuth and distance of the 24 targets are the test labels of the network;

[0019] Preferably, the Transformer network model structure and algorithm flow of step S2 are as follows:

[0020] The Transformer network model structure is mainly composed of data input, Transformer module, feedforward network and data output part. The data input part includes data segmentation, position encoding and linear projection.

[0021] Step S21, input data enters the model, the network model receives the prepared data set, and preprocesses the original data set to meet the shape requirements of the network model for the input data. The size of each input data is 22×22×3. Then the input data is cut into blocks of a fixed size of 1×1×3, and each small block is used as a patch, so the number of patches in each input data is 484. After cutting, the input data sequence is flattened, that is, the obtained 484 patches are sent to the fully connected layer to be mapped into the input data dimension that matches the input dimension of the Transformer module, and 484 embedded vectors are output. The dimension of each embedded vector is 1×1×3=3 after flattening, so the dimension of the output data after linear transformation mapping by the fully connected layer is 484×3. Because Transformer uses a fully connected method to model the global relationship of each block when processing input data, it cannot directly identify the spatial position information in the data sequence, that is, the model cannot "remember" the relative position information of each block in the input data. To solve this problem, the network model performs position encoding on each embedded vector, so that the Transformer can "remember" the spatial relative position information of each embedded vector when modeling the input data features, thereby deepening the understanding of the spatial relative position information of the input data. Finally, the sequence with completed position encoding is sent to the Transformer module;

[0022] Step S22, the data then enters a small Transformer encoding module. The input and output dimensions of this module remain consistent, so multiple such small modules can be stacked in the network model; L = 3, the core of this module is self-attention and multi-head attention mechanism; the convolution operation captures the connection relationship between the input sequence data elements within the receptive field by locally encoding information, and uses network stacking or fully connected layers to capture long-distance feature information. By calculating the relationship between each element in the sequence and all other elements in the sequence, the intrinsic correlation information between the elements in the sequence is captured, and at the same time, each component of the sequence is updated by integrating the overall information of the input sequence, so that the model can focus on different positions of the input sequence, thereby helping the model to better complete the prediction. A feedforward network is added at the end of each self-attention block. The feedforward network includes two transformation layers, one of which uses a nonlinear activation function;

[0023] Step S23, after that, the data enters the next stage, the multi-head attention mechanism is turned on, and the matrix Q=[q1,…,q M] to establish multiple dependencies from the input sequence data in parallel, where the attention weight of each sub-head corresponds to a specific type of dependency. By integrating the outputs generated by different subspaces, the final output result of the multi-head self-attention mechanism is obtained, thereby realizing the establishment of diversified dependencies in the sequence data;

[0024] Step S24, finally, the Transformer module outputs the sequence, which then passes through the MLP Head part. The dimension of the output layer of the MLP Head will match the number of categories, and the probability value of each category is output through the Softmax activation function. Subsequently, the node is multiplied by the probability value and summed up, and finally the azimuth and distance values ​​of the far-field target are directly output to complete the target position estimation and realize target positioning.

[0025] Preferably, the training of the network model in step S3 specifically includes the following steps:

[0026] Step S31, input the sample data into the built network model to extract features and train learning, the batch size batch_size is set to 64, the number of network iterations is 100, the learning rate is 0.0001, MSE is selected as the loss function, and the optimizer is Adam. In the forward propagation stage, the model will calculate the output value according to the input data according to the FDA-Transformer algorithm process;

[0027] Step S32, in the error back propagation stage, by comparing the difference between the output value and the true value, the error is calculated and back propagated layer by layer to update the network parameters, thereby optimizing the model. Through repeated iterations, the model gradually learns and improves the accuracy of target positioning. Once the training error approaches the established threshold or reaches the preset number of iterations, the training is considered complete.

[0028] The beneficial effects of the present invention are as follows: the frequency-controlled array (FDA) radar beam is favored in target positioning because of its azimuth-distance dependence. When the FDA radar accurately locates far-field targets, two aspects of the problem are mainly considered: one is how to efficiently complete the coupling of azimuth distance; the other is how to accurately extract the azimuth and distance information of the target. The existing methods are difficult to extract the azimuth and distance information of the target from the FDA radar echo data. In view of the above two problems, the present invention mainly utilizes the beam characteristics of the FDA radar, combines the Transformer network model, studies and explores the target positioning method, transforms the target positioning problem into a nonlinear mapping problem from the echo signal covariance matrix to the target azimuth and distance information, and proposes a FDA radar target positioning method based on the Transformer neural network. The experimental results show that the method proposed by the present invention can extract features from the echo data, efficiently complete the decoupling of the azimuth distance, directly obtain the azimuth and distance information, and realize the accurate positioning of the far-field target. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0030] Figure 1 It is a uniform linear FDA structure;

[0031] Figure 2 It is the FDA signal receiving and sending model;

[0032] Figure 3 It is the Transformer network model;

[0033] Figure 4 Comparison of target azimuth and distance estimation results;

[0034] Figure 5 The influence of different signal-to-noise ratios on the accuracy of the algorithm in estimating the target azimuth angle;

[0035] Figure 6 The influence of different signal-to-noise ratios on the accuracy of the algorithm in estimating target distance;

[0036] Figure 7 The influence of different snapshot numbers on the accuracy of the algorithm in estimating the target azimuth;

[0037] Figure 8 The impact of different snapshot numbers on the accuracy of the algorithm in estimating target distance; DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0040] It should also be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0041] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0042] like Figure 1 , Figure 2 As shown, a specific embodiment of the present invention provides an FDA radar target positioning method based on a Transformer neural network, comprising the following steps:

[0043] Step S1: Build a single-base FDA radar transceiver array signal model, set various parameters of the signal model, obtain the FDA radar echo signal matrix, calculate the FDA radar echo matrix covariance, obtain the covariance matrix, separate the real and virtual covariance matrix and normalize it to obtain the network input data.

[0044] The specific steps include:

[0045] In step S11, the selected FDA is a uniform linear array containing 11 array elements, each array element is arranged along a straight line, and the distance between every two adjacent array elements is λ / 2.

[0046] Step S12, consider that there are 24 far-field narrowband incoherent target signal sources with an initial carrier frequency of 10GHZ and a frequency offset of 5kHz incident on the array. Assume that the signal-to-noise ratio is a constant 10dB, the number of snapshots is a constant 100, and the noise is independent and obeys a zero-mean Gaussian distribution of white noise. The target azimuth angle range is [5°, 40°], the change interval is 5°, and there are 8 angles in total; the target distance range is [4km, 12km], the change interval is 4km, and there are 3 distances in total. The combination of 8 angles and 3 distances results in 24 target positions.

[0047] R=E[Y(t)Y H (t)] (1)

[0048] R′=[R1′,R2′,R3′,…,R′ N ] (2)

[0049] Among them, Y(t)=[y1(t),y2(t),···,y N (t)] T is the received signal vector, H represents the Hermitian operator, E[·] represents the mathematical expectation, R is the echo signal covariance matrix, and R′ represents the data obtained after the covariance matrix R is preprocessed by performing operations such as virtual-real separation and normalization.

[0050] Step S13, substitute the above 24 targets into equations (1) and (2) in turn for calculation to obtain 24 sets of data. It should be noted that the real and imaginary parts of the signal are separated in the process of constructing the data, so the number of experimental samples is twice the size of the beam space. Repeat the same method 100 times to obtain 2424 sets of data as the input sample data of the network. Select the training set and the validation set at a ratio of 100:1, and select the results of the first 100 calculations from the 24 targets, that is, 2400 groups, a total of 580,800 sample data, as the input data for network model training. Then select the results of the last calculation from the 24 targets, that is, 24 groups, a total of 5808 sample data, as the input data for verifying the performance of the network model. The data used in network training are all equipped with the corresponding target true azimuth and distance values ​​as training labels. The validation data is predicted by the network model to obtain the target azimuth and distance estimates. The actual azimuth and distance of the 24 targets are the test labels of the network;

[0051] Step S2: Figure 3 The figure shows the completed Transformer network model. The Transformer network model structure is mainly composed of data input, Transformer module, feedforward network and data output. The data input part includes data segmentation, position encoding and linear projection.

[0052] The specific steps include:

[0053] Step S21, input data enters the model, the network model receives the prepared data set, and preprocesses the original data set to meet the shape requirements of the network model for the input data. The size of each input data is 22×22×3. Then the input data is cut into blocks of a fixed size of 1×1×3, and each small block is used as a patch, so the number of patches in each input data is 484. After cutting, the input data sequence is flattened, that is, the obtained 484 patches are sent to the fully connected layer to be mapped into the input data dimension that matches the input dimension of the Transformer module, and 484 embedded vectors are output. The dimension of each embedded vector is 1×1×3=3 after flattening, so the dimension of the output data after linear transformation mapping by the fully connected layer is 484×3. Because Transformer uses a fully connected method to model the global relationship of each block when processing input data, it cannot directly identify the spatial position information in the data sequence, that is, the model cannot "remember" the relative position information of each block in the input data. To solve this problem, the network model performs position encoding on each embedded vector, so that the Transformer can "remember" the spatial relative position information of each embedded vector when modeling the input data features, thereby deepening the understanding of the spatial relative position information of the input data. Finally, the sequence with completed position encoding is sent to the Transformer module;

[0054] Step S22, the data then enters a small Transformer encoding module. The input and output dimensions of this module remain consistent, so multiple such small modules can be stacked in the network model; L = 3, the core of this module is self-attention and multi-head attention mechanism; the convolution operation captures the connection relationship between the input sequence data elements within the receptive field by locally encoding information, and uses network stacking or fully connected layers to capture long-distance feature information. By calculating the relationship between each element in the sequence and all other elements in the sequence, the intrinsic correlation information between the elements in the sequence is captured, and at the same time, each component of the sequence is updated by integrating the overall information of the input sequence, so that the model can focus on different positions of the input sequence, thereby helping the model to better complete the prediction. A feedforward network is added at the end of each self-attention block. The feedforward network includes two transformation layers, one of which uses a nonlinear activation function;

[0055] Step S23, after that, the data enters the next stage, the multi-head attention mechanism is turned on, and the matrix Q=[q1,…,q M] to establish multiple dependencies from the input sequence data in parallel, where the attention weight of each sub-head corresponds to a specific type of dependency. By integrating the outputs generated by different subspaces, the final output result of the multi-head self-attention mechanism is obtained, thereby realizing the establishment of diversified dependencies in the sequence data;

[0056] Step S24, finally, the Transformer module outputs the sequence, which then passes through the MLP Head part. The dimension of the output layer of the MLP Head will match the number of categories, and the probability value of each category is output through the Softmax activation function. Subsequently, the node is multiplied by the probability value and summed up, and finally the azimuth and distance values ​​of the far-field target are directly output to complete the target position estimation and realize target positioning.

[0057] Step S3: train the model and output the results. The training learning parameters are set, the batch size batch_size is set to 64, the number of network iterations is 100, the learning rate is 0.0001, MSE is selected as the loss function, and the optimizer is Adam. The built Transformer network model is trained with the training set of the input data. After the model training is completed, the model is tested with the test set of the input data. Finally, the far-field target azimuth and distance are output to complete the target position estimation. The target azimuth and distance estimation results are obtained, as well as the impact of different signal-to-noise ratios and snapshot numbers on the accuracy of the algorithm in estimating the target azimuth, such as Figure 4-Figure 8 shown.

[0058] Figure 4The experimental results of the method of the present invention for estimating the target azimuth and distance are compared. The experiment verifies whether the proposed algorithm is accurate in estimating the azimuth and distance of the set 24 far-field targets through simulation, and compares the estimation results with the FDA-CNN algorithm, FDA-MUSIC algorithm, and FDA-BP algorithm. The number of snapshots is fixed to 100, the SNR is fixed to 10dB, the number of training set samples is 580800, the number of test samples is 5808, and the size of a single estimation result is 48. In order to more intuitively express the positioning effect of the FDA-Transformer algorithm on the target, the values ​​predicted by the algorithm and the FDA-CNN algorithm are presented in a two-dimensional grid diagram. The black circle represents the position estimated by the FDA-Transformer algorithm, the brown asterisk represents the real target position, and the black asterisk represents the position estimated by the FDA-CNN algorithm. Not only can the accuracy of the algorithm be judged from the comparison of the numerical values, but also the accuracy of the algorithm positioning can be judged according to the distance of the predicted target position from the real target position. The distance from the real target position indicates that the accuracy is low, and vice versa. It is not difficult to see from the figure that the position of each black circle overlaps with the position of the brown star, and, except for the estimated values ​​at (10°, 8000m), (15°, 8000m), (30°, 8000m), and (40°, 8000m), which are slightly deviated, the other positions are almost completely coincident with the true target position. On the other hand, the estimated values ​​of the positions of the black circles at (15°, 8000m), (15°, 12000m), (20°, 8000m), (20°, 12000m), (25°, 8000m), (25°, 12000m), (30°, 4000m), (30°, 12000m), and (40°, 8000m) do not overlap with the position of the brown star, and are seriously deviated. The overlap between the other positions and the true target position is not high, and the positioning effect of the two algorithms is obvious. In summary, the simulation experiment results show that the FDA-Transformer algorithm has high positioning accuracy and its positioning effect is significantly better than that of the FDA-CNN algorithm.

[0059] Figure 5 The influence of different signal-to-noise ratios on the accuracy of the algorithm in estimating the target azimuth angle is shown in Figure 2. Figure 6 The influence of different signal-to-noise ratios on the accuracy of the algorithm in estimating the target distance. Figure 7 The influence of different snapshot numbers on the accuracy of the algorithm in estimating the target azimuth angle is shown in Figure 2. Figure 8 The effect of different snapshot numbers on the accuracy of the algorithm in estimating target distance.

[0060] Depend on Figure 5-Figure 6The performance of the algorithm in estimating the target distance angle under different signal-to-noise ratios can be analyzed. The horizontal axis is the signal-to-noise ratio, ranging from -10dB to 15dB, and the vertical axis is the RMSE value of the azimuth and distance, ranging from 0° to 1° and 0m to 270m respectively; It can be clearly seen from the figure that when the SNR changes in the range of [-10dB, 15dB], under the same signal-to-noise ratio, the corresponding azimuth and distance RMSE values ​​predicted by the FDA-Transformer algorithm are smaller than those of the FDA-CNN algorithm and the FDA-BP algorithm; With the increase of the signal-to-noise ratio, whether it is azimuth or distance estimation, the RMSE corresponding to the FDA-CNN algorithm and the FDA-BP algorithm is decreasing, but the RMSE corresponding to the FDA-Transformer algorithm does not change significantly, and the RMSE value tends to 0; Compared with the FDA-CNN algorithm and the FDA-BP algorithm, the FDA-Transformer algorithm is not sensitive to changes in the signal-to-noise ratio. Experimental results show that within the range of signal-to-noise ratio variation, compared with the FDA-CNN algorithm and the FDA-BP algorithm, the FDA-Transformer algorithm has better and more stable estimation performance and exhibits higher estimation accuracy.

[0061] Depend on Figure 7-Figure 8 The performance of the algorithm in estimating the target distance angle under different snapshot numbers can be analyzed. The horizontal axis is the snapshot number, ranging from 100 to 600, and the vertical axis is the RMSE value of the azimuth and distance, ranging from 0° to 1° and 0m to 64m respectively; it can be clearly seen from the figure that under the same snapshot number, the azimuth and distance RMSE values ​​corresponding to the FDA-Transformer algorithm are smaller than those corresponding to the FDA-CNN algorithm and the FDA-BP algorithm; with the increase of the snapshot number, the RMSE values ​​of the FDA-CNN algorithm and the FDA-BP algorithm are decreasing, whether it is azimuth or distance estimation, but the azimuth and distance RMSE values ​​corresponding to the FDA-Transformer algorithm do not change significantly, and the RMSE value tends to 0; compared with the FDA-CNN algorithm and the FDA-BP algorithm, the FDA-Transformer algorithm is not sensitive to the change of the snapshot number. The experimental results show that within the snapshot number change range, the estimation performance of the FDA-Transformer algorithm is better and more stable than that of the FDA-CNN algorithm and the FDA-BP algorithm.

[0062] Those of ordinary skill in the art will appreciate that the units of each example described in conjunction with the embodiments disclosed in the present invention can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition of each example has been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0063] In the embodiments provided in the present application, it should be understood that the division of units is merely a logical function division, and there may be other division methods in actual implementation, for example, multiple units may be combined into one unit, one unit may be split into multiple units, or some features may be ignored.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.

[0065] In summary, the present invention has been disclosed above with preferred embodiments.

Claims

1. A FDA radar target positioning method based on Transformer neural network, characterized in that: The following steps are involved: Step S1: Construct a data set. Build a single-base FDA radar transceiver array signal model, set various parameters of the signal model, obtain the FDA radar echo signal matrix, calculate the FDA radar echo matrix covariance, obtain the covariance matrix, separate the real from the virtual and normalize the covariance matrix, and obtain the network input data; Step S2: Build a Transformer network model. The model structure mainly consists of data input, Transformer module, feedforward network and data output part. The data input part includes data segmentation, position encoding and linear projection; Step S3: training model and outputting results. The built Transformer network model is trained using the training set of input data. After the model training is completed, the model is tested using the test set of input data. Finally, the azimuth and distance of the far-field target are output to complete the estimation of the target position.

2. The FDA radar target positioning method based on Transformer neural network according to claim 1 is characterized in that: The step S1 specifically includes the following steps: In step S11, the selected FDA is a uniform linear array containing 11 array elements, each array element is arranged along a straight line, and the distance between every two adjacent array elements is λ / 2. Step S12, consider that there are 24 far-field narrowband incoherent target signal sources with an initial carrier frequency of 10GHZ and a frequency offset of 5kHz incident on the array. Assume that the signal-to-noise ratio is a constant 10dB, the number of snapshots is a constant 100, and the noise is independent and obeys a zero-mean Gaussian distribution of white noise. The target azimuth angle range is [5°, 40°], the change interval is 5°, and there are 8 angles in total; the target distance range is [4km, 12km], the change interval is 4km, and there are 3 distances in total. The combination of 8 angles and 3 distances results in 24 target positions. R=E[Y(t)Y H (t)] (1) R′=[R1′,R2′,R3′,…,R′ N ] (2) Among them, Y(t)=[y1(t),y2(t),···,y N (t)] T is the received signal vector, H represents the Hermitian operator, E[·] represents the mathematical expectation, R is the echo signal covariance matrix, and R′ represents the data obtained after the covariance matrix R is preprocessed by performing operations such as virtual-real separation and normalization. Step S13, the above 24 targets are substituted into equations (1) and (2) in turn for calculation to obtain 24 sets of data. It should be noted that the real part and the imaginary part of the signal are separated in the process of constructing the data, so the number of experimental samples is twice the size of the beam space. Repeat the same method 100 times, and obtain 2424 sets of data as the input sample data of the network. The training set and the validation set are selected at a ratio of 100:1, and the results of the first 100 calculations are selected from the 24 targets, that is, 2400 groups, a total of 580,800 sample data, as the input data for network model training. Then the results of the last calculation from the 24 targets, that is, 24 groups, a total of 5808 sample data, are used as input data to verify the performance of the network model. The data used in network training are all equipped with corresponding target real azimuth and distance values ​​as training labels. The validation data is predicted by the network model to obtain the target azimuth and distance estimation values. The actual azimuth and distance of the 24 targets are the test labels of the network.

3. The FDA radar target positioning method based on Transformer neural network according to claim 1 is characterized in that: The Transformer network model structure and algorithm flow of step S2 are as follows: The Transformer network model structure is mainly composed of data input, Transformer module, feedforward network and data output part. The data input part includes data segmentation, position encoding and linear projection. Step S21, input data enters the model, the network model receives the prepared data set, and preprocesses the original data set to meet the shape requirements of the network model for the input data. The size of each input data is 22×22×3. Then the input data is cut into blocks of a fixed size of 1×1×3, and each small block is used as a patch, so the number of patches in each input data is 484. After cutting, the input data sequence is flattened, that is, the obtained 484 patches are sent to the fully connected layer to be mapped into the input data dimension that matches the input dimension of the Transformer module, and 484 embedded vectors are output. The dimension of each embedded vector is 1×1×3=3 after flattening, so the dimension of the output data after linear transformation mapping by the fully connected layer is 484×3. Because Transformer uses a fully connected method to model the global relationship of each block when processing input data, it cannot directly identify the spatial position information in the data sequence, that is, the model cannot "remember" the relative position information of each block in the input data. To solve this problem, the network model performs position encoding on each embedded vector, so that the Transformer can "remember" the spatial relative position information of each embedded vector when modeling the input data features, thereby deepening the understanding of the spatial relative position information of the input data. Finally, the sequence with completed position encoding is sent to the Transformer module; Step S22, the data then enters a small Transformer encoding module. The input and output dimensions of this module remain consistent, so multiple such small modules can be stacked in the network model; L = 3, the core of this module is self-attention and multi-head attention mechanism; the convolution operation captures the connection relationship between the input sequence data elements within the receptive field by locally encoding information, and uses network stacking or fully connected layers to capture long-distance feature information. By calculating the relationship between each element in the sequence and all other elements in the sequence, the intrinsic correlation information between the elements in the sequence is captured, and at the same time, each component of the sequence is updated by integrating the overall information of the input sequence, so that the model can focus on different positions of the input sequence, thereby helping the model to better complete the prediction. A feedforward network is added at the end of each self-attention block. The feedforward network includes two transformation layers, one of which uses a nonlinear activation function; Step S23, after that, the data enters the next stage, the multi-head attention mechanism is turned on, and the matrix Q=[q1,…,q M ] to establish multiple dependencies from the input sequence data in parallel, where the attention weight of each sub-head corresponds to a specific type of dependency. By integrating the outputs generated by different subspaces, the final output result of the multi-head self-attention mechanism is obtained, thereby realizing the establishment of diversified dependencies in the sequence data; Step S24, finally, the Transformer module outputs the sequence, which then passes through the MLP Head part. The dimension of the output layer of the MLP Head will match the number of categories, and the probability value of each category is output through the Softmax activation function. Subsequently, the node is multiplied by the probability value and summed up, and finally the azimuth and distance values ​​of the far-field target are directly output to complete the target position estimation and realize target positioning.

4. The FDA radar target positioning method based on Transformer neural network according to claim 1 is characterized in that: The training of the network model in step S3 specifically includes the following steps: Step S31, input the sample data into the built network model to extract features and train learning, the batch size batch_size is set to 64, the number of network iterations is 100, the learning rate is 0.0001, MSE is selected as the loss function, and the optimizer is Adam. In the forward propagation stage, the model will calculate the output value according to the input data according to the FDA-Transformer algorithm process; Step S32, in the error back propagation stage, by comparing the difference between the output value and the true value, the error is calculated and back propagated layer by layer to update the network parameters, thereby optimizing the model. Through repeated iterations, the model gradually learns and improves the accuracy of target positioning. Once the training error approaches the established threshold or reaches the preset number of iterations, the training is considered complete.

Images