Deep learning-based meshless DOA estimation method and device in cross-eye interference confrontation environment

Abstract

The invention discloses a meshless DOA estimation method based on deep learning in a cross-eye interference confrontation environment, and belongs to the technical field of radars, and the method comprises the steps: obtaining an echo pulse signal received by a radar array, the echo pulse signal containing cross-eye interference; separating a real component and an imaginary component of each echo pulse signal, performing zero-mean standardization on the real component and the imaginary component, and recombining the real component and the imaginary component into a four-dimensional tensor; and inputting the four-dimensional tensor into the angle estimation model to obtain an estimated value of the target azimuth angle. According to the method, the neural network model is introduced, continuous accurate estimation of the azimuth angle of the radar target in the cross-eye interference environment is achieved, the continuous azimuth angle estimation value can be directly output through the end-to-end supervised learning framework, quantization errors caused by traditional classification network discretization are avoided, and meshless high-precision angle measurement estimation is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular relates to a gridless DOA estimation method and device based on deep learning in a cross-eye interference countermeasure environment. Background Art

[0002] In the field of complex electromagnetic countermeasures, cross-eye jamming, a typical example of mainlobe deceptive jamming, creates a strongly coupled false beam within the radar mainlobe through the space-time joint modulation of multi-source coherent signals, rendering traditional direction-finding systems ineffective. This jamming mechanism leverages the principles of radar beamforming to precisely control the amplitude, phase, and spatial directionality of the jamming signal, thereby causing systematic deviations in single-pulse amplitude / phase comparison angle measurement methods.

[0003] At present, existing anti-interference technologies are mostly based on single-dimensional processing based on spatial filtering or polarization filtering. Although they can effectively suppress interference in sidelobe interference scenarios, they lack the feature decoupling capability for cross-eye interference that overlaps with the target signal in the mainlobe. Especially when the jammer dynamically adjusts the waveform parameters, the existing technology cannot perceive the spatiotemporal correlation characteristics of the interference signal in real time, resulting in a lag in the convergence of the adaptive beamforming weights and difficulty in maintaining stable angle measurement accuracy.

[0004] Although frequency domain filtering and blind source separation technologies have developed in recent years, they are not robust enough under low signal-to-noise ratio and high interference-to-noise ratio conditions, and they do not fully utilize the spatial-temporal joint domain feature information of radar echoes.

[0005] Therefore, how to deal with the tactical threat of mainlobe cross-eye interference is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] In order to solve the above problems existing in the prior art, the present invention provides a gridless DOA estimation method and device based on deep learning in a cross-eye interference countermeasure environment. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0007] In a first aspect, the present invention provides a gridless DOA estimation method based on deep learning in a cross-eye interference countermeasure environment, comprising:

[0008] Acquiring an echo pulse signal received by a radar array, wherein the echo pulse signal includes cross-eye interference;

[0009] Separating the real component and the imaginary component of each echo pulse signal, performing zero-mean normalization on the real component and the imaginary component, and then recombining them into a four-dimensional tensor;

[0010] The four-dimensional tensor is input into an angle estimation model to obtain an estimated value of the target azimuth angle.

[0011] In one embodiment of the present invention, the angle estimation model is trained according to the following steps:

[0012] Acquire a plurality of training samples, each of the training samples comprising a sample echo signal and a label of the sample echo signal, wherein the label is an actual target azimuth of the sample echo signal;

[0013] The labels of each sample echo signal are respectively combined with the scaling factor f scale Multiply them together to get the optimized label, the scaling factor f scale is a constant greater than 1;

[0014] Inputting multiple sample echo signals into the neural network to be trained to obtain the predicted target azimuth of each sample echo signal;

[0015] Calculating a loss value of a preset loss function based on the optimized label and the predicted target azimuth;

[0016] When the loss value converges, an angle estimation model is obtained.

[0017] In one embodiment of the present invention, the neural network to be trained is an LSTM network or a ViT network.

[0018] In one embodiment of the present invention, the ViT network includes a first Transformer encoder, a second Transformer encoder, a third Transformer encoder, a fourth Transformer encoder, a fifth Transformer encoder, and a sixth Transformer encoder connected in sequence.

[0019] In one embodiment of the present invention, the LSTM network includes a first LSTM layer, a second LSTM layer and a fully connected layer;

[0020] The first LSTM layer includes 800 first LSTM units connected in sequence, and the second LSTM layer includes 16 second LSTM layers connected in sequence, wherein the nth second LSTM unit in the second LSTM layer is connected to the 50*nth first LSTM unit in the first LSTM layer.

[0021] In one embodiment of the present invention, the preset loss function is a mean square error loss function.

[0022] In one embodiment of the present invention, the step of inputting the four-dimensional tensor into an angle estimation model to obtain an estimated value of the target azimuth angle includes:

[0023] After the four-dimensional tensor is input into the angle estimation model, the output of the angle estimation model is divided by the scaling factor fscale , obtaining an estimated value of the target azimuth angle; the dimension of the four-dimensional tensor is expressed as 1×2×N×K, where N represents the number of array elements and K represents the number of snapshots.

[0024] In a second aspect, the present invention further provides a gridless DOA estimation device based on deep learning in a cross-eye interference countermeasure environment, comprising:

[0025] An acquisition module is used to acquire an echo pulse signal received by a radar array, wherein the echo pulse signal contains cross-eye interference;

[0026] A separation module, configured to separate the real component and the imaginary component of each echo pulse signal, and to perform zero-mean normalization on the real component and the imaginary component respectively and then reconstruct them into a four-dimensional tensor;

[0027] The estimation module is used to input the four-dimensional tensor into the angle estimation model to obtain an estimated value of the target azimuth.

[0028] In a third aspect, the present invention further provides an electronic device comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0029] Memory for storing computer programs;

[0030] The processor is configured to implement the method steps described in the first aspect when executing the program stored in the memory.

[0031] In a fourth aspect, the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps described in the first aspect are implemented.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The present invention provides a gridless DOA estimation method and device based on deep learning in a cross-eye interference countermeasure environment. By introducing a neural network model, continuous and accurate estimation of the azimuth angle of radar targets in a cross-eye interference environment is achieved. The neural network model is a pre-trained LSTM network or ViT network. The LSTM network is used to capture the dynamic characteristics of a time-domain snapshot sequence, while the ViT network is used to extract spatial nonlinear correlations. This method breaks through the linear constraints of traditional spatial filtering methods and significantly improves the modeling capability of complex interference signals. This end-to-end supervised learning framework can directly output continuous azimuth angle estimates, avoiding the quantization errors caused by the discretization of traditional classification networks and achieving gridless high-precision angle measurement.

[0034] In addition, during the training process of the angle estimation model, a scaling factor is introduced to dynamically amplify the label of the sample echo signal, which effectively solves the gradient vanishing and early stopping problems in small-scale regression tasks, improves the convergence efficiency and stability of the model, and demonstrates excellent anti-interference ability and robustness under harsh interference conditions.

[0035] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of a gridless DOA estimation method based on deep learning in a cross-eye interference countermeasure environment provided by an embodiment of the present invention;

[0037] Figure 2 This is another flow chart of a gridless DOA estimation method based on deep learning in a cross-eye interference countermeasure environment provided by an embodiment of the present invention;

[0038] Figure 3 is a schematic diagram of generating a four-dimensional tensor provided by an embodiment of the present invention;

[0039] Figure 4 Schematic diagram of the structure of the LSTM network provided by an embodiment of the present invention;

[0040] Figure 5a This is a graph showing how the training loss and validation loss of the LSTM network change with the number of iterations.

[0041] Figure 5b It is a graph showing the change of the training loss value and the verification loss value of the ViT network with the number of iterations;

[0042] Figure 6 This is a graph showing the variation of the angle measurement error of the LSTM network and the ViT network with SNR;

[0043] Figure 7 This is a graph showing the variation of the angle measurement error of the LSTM network and the ViT network with the JNR;

[0044] Figure 8 2 is a schematic structural diagram of a gridless DOA estimation device based on deep learning in a cross-eye interference countermeasure environment provided by an embodiment of the present invention;

[0045] Figure 9 It is a structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0047] Figure 1This is a flow chart of a gridless DOA estimation method based on deep learning in a cross-eye interference countermeasure environment provided by an embodiment of the present invention. Figure 2 This is another flow chart of a gridless DOA estimation method based on deep learning in a cross-eye interference countermeasure environment provided by an embodiment of the present invention. Figures 1-2 As shown, this embodiment provides a gridless DOA estimation method based on deep learning in a cross-eye interference countermeasure environment, including:

[0048] S1. Acquire an echo pulse signal received by a radar array, where the echo pulse signal contains cross-eye interference.

[0049] Exemplarily, this embodiment uses single-target and two-source cross-eye interference as the application scenario, and the above-mentioned radar array is a uniform linear array.

[0050] S2. Separate the real component and the imaginary component of each echo pulse signal, perform zero-mean normalization on the real component and the imaginary component, and then reconstruct them into a four-dimensional tensor.

[0051] Figure 3 Schematic diagram of generating a four-dimensional tensor according to an embodiment of the present invention. Figure 3 As shown in the figure, the echo pulse signal received by the radar array is orthogonally demodulated to separate the real part (I path) and the imaginary part (Q path). Then, the real and imaginary components are respectively normalized by zero mean and recombined into a four-dimensional tensor with the dimension of 1×2×N×K, where N is the number of array elements and K is the number of snapshots. The four-dimensional tensor contains the joint space-time domain features of the echo pulse signal.

[0052] S3. Input the four-dimensional tensor into the angle estimation model to obtain an estimated value of the target azimuth.

[0053] Optionally, the angle estimation model is trained according to the following steps:

[0054] S3a. Acquire multiple training samples, each training sample including a sample echo signal and a label of the sample echo signal, wherein the label is the actual target azimuth of the sample echo signal.

[0055] S3b, the labels of each sample echo signal are respectively compared with the proportional factor f scale Multiply to get the optimized label, the scaling factor f scale is a constant greater than 1.

[0056] It should be noted that in order to solve the problem of premature training stop or convergence to suboptimal solution due to the small value range of the label, this embodiment introduces a scaling factor f in the training process. scale , define the optimized label as: label scale =f scale×label, label represents the label of the sample echo signal, which can expand the prediction range of the angle estimation model.

[0057] Preferably, f scale =10.

[0058] S3c. Inputting the multiple sample echo signals into the neural network to be trained to obtain the predicted target azimuth of each sample echo signal.

[0059] In this step, the neural network to be trained can selectively use an LSTM (Long Short-Term Memory) network. Figure 4 : is a structural diagram of an LSTM network provided by an embodiment of the present invention. Specifically, the LSTM network includes a first LSTM layer, a second LSTM layer, and a fully connected layer. The first LSTM layer includes 800 first LSTM units connected in sequence, and the second LSTM layer includes 16 second LSTM layers connected in sequence. The nth second LSTM unit in the second LSTM layer is connected to the 50th*nth first LSTM unit in the first LSTM layer, where n = 1, 2, ..., 16. The dimension of the hidden layer in the first LSTM unit and the second LSTM unit is 128, and the batch size is 80.

[0060] In addition, the neural network to be trained can also be a ViT (Vision Transformer) network. The ViT network includes a first Transformer encoder, a second Transformer encoder, a third Transformer encoder, a fourth Transformer encoder, a fifth Transformer encoder, and a sixth Transformer encoder connected in sequence. The number of heads of the multi-head attention mechanism in each Transformer encoder is 8, and the patch size is 16×16.

[0061] S3d. Calculate the loss value of the preset loss function based on the optimized label and the predicted target azimuth.

[0062] In this step, the preset loss function is the mean square error loss function. Since the optimized label samples used in training are the labels of the echo signal and the scaling factor f scale The product of , then reflected in the mean square error loss function will magnify the loss value times, thereby appropriately amplifying the neural network's prediction error for each sample echo signal during training. In other words, while keeping the complexity of the data pre-processing and post-processing steps essentially unchanged, this approach appropriately increases the amplitude of gradient updates, preventing the neural network from prematurely converging due to excessively small loss values. This effectively increases the sensitivity of the preset loss function to prediction error, thereby promoting more comprehensive model learning and optimization.

[0063] S3e. When the loss value converges, the angle estimation model is obtained.

[0064] Furthermore, in step S3, the step of inputting the four-dimensional tensor into the angle estimation model to obtain an estimated value of the target azimuth angle includes:

[0065] After the four-dimensional tensor is input into the angle estimation model, the output of the angle estimation model is divided by the scale factor f scale , and obtain the estimated value of the target azimuth angle; the dimension of the four-dimensional tensor is expressed as 1×2×N×K, where N represents the number of array elements and K represents the number of snapshots.

[0066] Below, the gridless DOA estimation method based on deep learning in the cross-eye interference countermeasure environment provided by the present invention is further illustrated through simulation experiments.

[0067] First, we constructed a dataset for supervised regression learning. This simulation used a scenario involving a single target and two-source cross-eye interference. A single echo pulse signal received by the radar array was used as a sample echo signal. Data was generated through the synergistic effect of multiple random perturbation factors, specifically covering the following parameters:

[0068] The signal-to-noise ratio is randomly and evenly distributed within -10 to 10dB, the interference-to-noise ratio is randomly and evenly distributed within 0 to 30dB, and the phase difference between the interference and the signal Randomly and uniformly distributed within 0~π, the amplitude ratio of the two jammers in the jamming loop is a n The phase difference between the two jammers is randomly and uniformly distributed between 0 and 2. n Randomly and uniformly distributed within the range of 0 to 2π. The above parameter value range meets the actual radar working requirements. The Monte Carlo algorithm is used to perform probability distribution sampling on the parameter space, thereby maximizing coverage efficiency within a limited sample size and ensuring the diversity and distribution balance of training samples.

[0069] In order to construct a complete data set, the actual target azimuth is further set as the label of each training sample. The actual target azimuth is between -3.5° and 3.5°, divided into 71 types at intervals of 0.1°. 10,000 samples are generated for each label and randomly divided into training set, test set and validation set according to the ratios of 0.70, 0.15 and 0.15.

[0070] In this embodiment, the radar is a uniform linear array with N=32 elements, the element spacing d=λ / 2, λ represents the signal wavelength, the jammer array rotation angle is the steering angle of the baseline platform where the jammer is located relative to the radar boresight, and the jammer baseline length is the length of the baseline platform where the jammer is located.

[0071] The remaining parameters are shown in Table 1:

[0072] Table 1 Radar and jammer simulation parameter settings

[0073]

[0074]

[0075] Figure 5a This is a graph showing how the training loss and validation loss of the LSTM network change with the number of iterations. Figure 5a It can be seen that the LSTM network converges quickly in the early stages of training (iterations 0 to 20), with the training loss value rapidly decreasing from an initial value of 25 to around 1, and the validation loss value synchronously decreasing from an initial value slightly lower than the training loss value to 0.9, indicating that the LSTM network quickly learns the signal characteristics in the interference environment in the early stages and has preliminary learning capabilities. In the middle and late stages of training (iterations 20 to 114), the training loss value stabilizes within the range of 0.29±0.01, while the validation loss value fluctuates between 0.29 and 0.35, with no significant downward trend, indicating that the LSTM network has reached convergence at this stage. This trend demonstrates the good learning ability of the LSTM model in the DOA estimation task.

[0076] Figure 5b This is a graph showing how the training loss and validation loss of the ViT network change with the number of iterations. Figure 5b As shown, the ViT network converges rapidly in the early stages of training (iterations 0 to 20), with the training loss rapidly decreasing from an initial value of 16 to around 0.5. The validation loss also decreases simultaneously from an initial value slightly lower than the training loss to 0.5, demonstrating that the ViT network effectively captures data features early on and demonstrates initial learning capabilities. In the mid-to-late stages (iterations 20 to 114), the training loss of the ViT network stabilizes within a range of 0.16±0.01, while the validation loss fluctuates between 0.18 and 0.20, showing no significant downward trend. This indicates that the ViT network achieves a high degree of fit to the training data and has converged at this stage. This trend demonstrates the ViT model's stable learning capabilities in the DOA estimation task.

[0077] Furthermore, the generalization and robustness of the LSTM network and ViT network used in the deep learning-based gridless DOA estimation method under the cross-eye interference countermeasure environment are analyzed respectively.

[0078] Figure 6 The following graph shows the variation of the angular measurement error of the LSTM network and the ViT network with the SNR (signal-to-noise ratio). To construct the required test set, the SNR was first discretized. Nine independent test sets were set in the range [-20dB, 20dB] with a step size of 5dB, covering scenarios ranging from extremely low SNR to common detection scenarios. Next, angle samples were generated: for each SNR condition, dense test samples were generated in the azimuth range [-3.5°, 3.5°] with a 0.1° interval. Each angle point contained 5000 independent samples, bringing the total number of test sets to 9 × 71 × 5000 sets.

[0079] The simulation results are as follows Figure 6 As shown in the results, the ViT network outperforms the LSTM network in terms of accuracy and generalization. The angle measurement error of the ViT network is significantly lower than that of the LSTM network in all SNR ranges (including extreme environments within and outside the training set), with an average reduction of 22% to 27%. The ViT network directly models the global relationship between array elements through the self-attention mechanism, avoiding the information attenuation problem of the recursive structure of the LSTM network, thereby maintaining high accuracy in the cross-eye interference environment. In addition, the generalization error fluctuation of the ViT network is comparable to that of the LSTM network, but the absolute error is lower, indicating that its model architecture is less dependent on the data distribution. Therefore, in practical application scenarios where the SNR distribution range is wide or contains extreme values, the ViT network is more practical; if the system has strict stability requirements and the SNR range is strictly limited, the LSTM network can be given priority.

[0080] Figure 7 The following plots the angular measurement error of the LSTM network and the ViT network as a function of the JNR (Interference-to-Noise Ratio). To construct the required test set, the JNR discretization design was first performed: 11 independent test sets were set up in the range [-10dB, 40dB] with a step size of 5dB, covering scenarios ranging from extremely low signal-to-noise ratios to common detection scenarios. Next, angle sample generation was performed: for each JNR condition, dense test samples were generated in the azimuth range [-3.5°, 3.5°] at 0.1° intervals, with each angle point containing 5000 independent samples. The total sample size of the entire test set reached 11 × 71 × 5000 sets.

[0081] The simulation results can be found in Figure 7In the robustness analysis, the LSTM network's error ranged from 0.0521° to 0.0602°, with a standard deviation of 0.0031. Overall fluctuations were small, but the error values ​​were high. For example, the error peaked at 0.0598° at 5 dB and dropped to a minimum of 0.0521° at 20 dB, indicating limitations in its adaptability to moderate JNR environments. In contrast, the ViT network's error significantly decreased to 0.0387° to 0.0424°, with a standard deviation of only 0.0016, reaching its lowest value in the 20 dB to 25 dB range. This result demonstrates the ViT network's ability to model complex signal features through a global attention mechanism, enabling it to more stably suppress the impact of JNR fluctuations on angle measurement accuracy.

[0082] In the low JNR (-10dB to 0dB) generalization range, the LSTM network's error stabilized at 0.0560 to 0.0575, consistent with its performance in the training range, but without showing an optimization trend. The ViT network's error decreased from 0.0418° at -10dB to 0.0411° at 0dB, a decrease of 1.7%, and was lower than some of the results in its training range, indicating its excellent generalization ability in low JNR noise environments.

[0083] In the generalization range of high JNR (30dB-40dB), the LSTM network's error increased from 0.0602° at 30dB to 0.0607° at 40dB, a fluctuation of only 0.8%, but the overall error was still higher than that of the ViT network. The ViT network's error increased from 0.0423° at 30dB to 0.0456° at 40dB, a 7.8% increase, but remained lower than the corresponding value of the LSTM network. Comparing the generalization range with the training range, the standard deviation of the ViT network's error fluctuation was 32% lower than that of the LSTM network. Further analysis revealed that the ViT network's increased error at high JNR may be due to a lack of extremely high JNR samples in the training data, but its global feature extraction capability still effectively suppressed performance degradation.

[0084] Figure 8 Schematic diagram of the structure of a gridless DOA estimation device based on deep learning in a cross-eye interference countermeasure environment provided by an embodiment of the present invention. Figure 8 As shown, an embodiment of the present invention further provides a gridless DOA estimation device based on deep learning in a cross-eye interference countermeasure environment, comprising:

[0085] An acquisition module 810 is configured to acquire an echo pulse signal received by a radar array, wherein the echo pulse signal includes cross-eye interference;

[0086] A separation module 820 is used to separate the real component and the imaginary component of each echo pulse signal, perform zero-mean normalization on the real component and the imaginary component, and then reconstruct them into a four-dimensional tensor;

[0087] The estimation module 830 is used to input the four-dimensional tensor into the angle estimation model to obtain an estimated value of the target azimuth.

[0088] It can be seen from the above embodiments that the beneficial effects of the present invention are:

[0089] The present invention provides a gridless DOA estimation method and device based on deep learning in a cross-eye interference countermeasure environment. By introducing a neural network model, continuous and accurate estimation of the azimuth angle of radar targets in a cross-eye interference environment is achieved. The neural network model is a pre-trained LSTM network or ViT network. The LSTM network is used to capture the dynamic characteristics of a time-domain snapshot sequence, while the ViT network is used to extract spatial nonlinear correlations. This method breaks through the linear constraints of traditional spatial filtering methods and significantly improves the modeling capability of complex interference signals. This end-to-end supervised learning framework can directly output continuous azimuth angle estimates, avoiding the quantization errors caused by the discretization of traditional classification networks and achieving gridless high-precision angle measurement.

[0090] In addition, during the training process of the angle estimation model, a scaling factor is introduced to dynamically amplify the label of the sample echo signal, which effectively solves the gradient vanishing and early stopping problems in small-scale regression tasks, improves the convergence efficiency and stability of the model, and demonstrates excellent anti-interference ability and robustness under harsh interference conditions.

[0091] The embodiment of the present invention further provides an electronic device, such as Figure 9 As shown, it includes a processor 901, a communication interface 902, a memory 903 and a communication bus 904, wherein the processor 901, the communication interface 902, and the memory 903 communicate with each other through the communication bus 904.

[0092] Memory 903, used for storing computer programs;

[0093] The processor 901 is configured to execute the program stored in the memory 903, and implement the following steps:

[0094] Acquiring an echo pulse signal received by a radar array, wherein the echo pulse signal includes cross-eye interference;

[0095] Separating the real component and the imaginary component of each echo pulse signal, performing zero-mean normalization on the real component and the imaginary component, and then recombining them into a four-dimensional tensor;

[0096] The four-dimensional tensor is input into an angle estimation model to obtain an estimated value of the target azimuth angle.

[0097] The communication bus mentioned in the electronic device mentioned above may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. This communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, only one thick line is used in the figure, but this does not mean that there is only one bus or only one type of bus.

[0098] The communication interface is used for communication between the above electronic device and other devices.

[0099] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage. Alternatively, the memory may be at least one storage device located away from the processor.

[0100] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0101] The method provided in the embodiments of the present invention can be applied to electronic devices. Specifically, the electronic devices can be desktop computers, portable computers, smart mobile terminals, servers, etc. This is not limited here; any electronic device that can implement the present invention falls within the scope of protection of the present invention.

[0102] As for the device / electronic device / storage medium embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0103] It should be noted that the device, electronic device and storage medium of the embodiments of the present invention are respectively the device, electronic device and storage medium for applying the above-mentioned gridless DOA estimation method based on deep learning in the cross-eye interference countermeasure environment. All embodiments of the above-mentioned gridless DOA estimation method based on deep learning in the cross-eye interference countermeasure environment are applicable to the device, electronic device and storage medium, and can achieve the same or similar beneficial effects.

[0104] By using the terminal device provided by the embodiment of the present invention, proper nouns and / or fixed phrases can be displayed for user selection, thereby reducing user input time and improving user experience.

[0105] In the description of the present invention, reference to the terms "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification.

[0106] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A gridless DOA estimation method based on deep learning in a cross-eye interference countermeasure environment, characterized by: include: Acquiring an echo pulse signal received by a radar array, wherein the echo pulse signal includes cross-eye interference; Separating the real component and the imaginary component of each echo pulse signal, performing zero-mean normalization on the real component and the imaginary component, and then recombining them into a four-dimensional tensor; The four-dimensional tensor is input into an angle estimation model to obtain an estimated value of the target azimuth angle.

2. The gridless DOA estimation method based on deep learning in the cross-eye interference countermeasure environment according to claim 1 is characterized in that The angle estimation model is trained according to the following steps: Acquire a plurality of training samples, each of the training samples comprising a sample echo signal and a label of the sample echo signal, wherein the label is an actual target azimuth of the sample echo signal; The labels of each sample echo signal are respectively combined with the scaling factor f scale Multiply them together to get the optimized label, the scaling factor f scale is a constant greater than 1; Inputting multiple sample echo signals into the neural network to be trained to obtain the predicted target azimuth of each sample echo signal; Calculating a loss value of a preset loss function based on the optimized label and the predicted target azimuth; When the loss value converges, an angle estimation model is obtained.

3. The gridless DOA estimation method based on deep learning in the cross-eye interference countermeasure environment according to claim 2 is characterized in that: The neural network to be trained is an LSTM network or a ViT network.

4. The gridless DOA estimation method based on deep learning in the cross-eye interference countermeasure environment according to claim 3 is characterized in that The ViT network includes a first Transformer encoder, a second Transformer encoder, a third Transformer encoder, a fourth Transformer encoder, a fifth Transformer encoder, and a sixth Transformer encoder, which are connected in sequence.

5. The gridless DOA estimation method based on deep learning in the cross-eye interference countermeasure environment according to claim 3 is characterized in that: The LSTM network includes a first LSTM layer, a second LSTM layer and a fully connected layer; The first LSTM layer includes 800 first LSTM units connected in sequence, and the second LSTM layer includes 16 second LSTM layers connected in sequence, wherein the nth second LSTM unit in the second LSTM layer is connected to the 50*nth first LSTM unit in the first LSTM layer.

6. The gridless DOA estimation method based on deep learning in the cross-eye interference countermeasure environment according to claim 2 is characterized in that: The preset loss function is the mean square error loss function.

7. The gridless DOA estimation method based on deep learning in the cross-eye interference countermeasure environment according to claim 1 is characterized in that The step of inputting the four-dimensional tensor into an angle estimation model to obtain an estimated value of the target azimuth angle includes: After the four-dimensional tensor is input into the angle estimation model, the output of the angle estimation model is divided by the scaling factor f scale , obtaining an estimated value of the target azimuth angle; the dimension of the four-dimensional tensor is expressed as 1×2×N×K, where N represents the number of array elements and K represents the number of snapshots.

8. A gridless DOA estimation device based on deep learning in a cross-eye interference countermeasure environment, characterized in that: include: An acquisition module is used to acquire an echo pulse signal received by a radar array, wherein the echo pulse signal contains cross-eye interference; A separation module, configured to separate the real component and the imaginary component of each echo pulse signal, and to perform zero-mean normalization on the real component and the imaginary component respectively and then reconstruct them into a four-dimensional tensor; The estimation module is used to input the four-dimensional tensor into the angle estimation model to obtain an estimated value of the target azimuth.

9. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; Memory for storing computer programs; A processor, configured to implement the method steps described in any one of claims 1 to 7 when executing a program stored in a memory.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps of any one of claims 1 to 7 are implemented.

Images