DOA estimation method and system based on global dynamic convolutional neural network

Abstract

The invention discloses a DOA estimation method and system based on a global dynamic convolutional neural network, and the method comprises the steps: firstly obtaining a time domain signal received by a uniform linear array, and carrying out the preprocessing; secondly, constructing a DOA estimation model, and extracting deep features from the preprocessed signals by using a global dynamic convolutional neural network; and based on the extracted deep features, estimating the direction of arrival (DOA) of the signal source in a classification mode by using a classification head. And finally, performing DOA estimation model pre-training by using the two-information-source ideal covariance matrix data set to obtain an initial weight, performing fine adjustment on the initial weight by using the four-information-source sampling covariance matrix data set, and performing verification. According to the method, the DOA estimation precision is improved, and a relatively large signal source number can be well estimated without losing accuracy in a relatively severe environment.

Description

Technical Field

[0001] This invention pertains to signal direction of arrival (DOA) estimation based on deep learning, which is widely used in radar, sonar, target tracking, satellite communication and other technical fields. Specifically, it relates to a DOA estimation method and system based on a global dynamic convolutional neural network. Background Technology

[0002] DOA estimation of sound or signal sources is a core task in array signal processing, widely used in radar, sonar, wireless communication, speech enhancement, and robotic hearing. Traditional DOA estimation methods, such as Multiple Signal Classification (MUSIC) and Rotation Invariant Subspace (ESPRIT) algorithms, based on rigorous mathematical models and statistical assumptions, can achieve super-resolution estimation under ideal conditions, but suffer from insufficient robustness in real-world complex environments. Performance degrades significantly when coherent sources, low signal-to-noise ratios, or limited snapshot data are present, and computational complexity is high, making it difficult to meet real-time processing requirements. In recent years, deep learning has provided a new approach to DOA estimation, employing a data-driven paradigm to construct DOA estimation as an angle classification or regression problem, directly learning directional mapping relationships from array signal data.

[0003] However, existing deep learning methods have significant limitations. Traditional convolutional neural networks (CNNs) have limited receptive fields, making it difficult to capture long-range dependencies in the array covariance matrix; their fixed weights lack adaptability to signal characteristics, making them ill-suited for dynamic source environments; furthermore, CNNs struggle to effectively utilize prior knowledge such as array geometry. While attention-based methods can capture global dependencies, their computational complexity increases quadratically with the input size, resulting in high computational costs for high-dimensional covariance matrices; moreover, they often overemphasize global interactions, neglecting local details and lacking effective integration of array physical constraints.

[0004] GKPapageorgiou et al. proposed a DOA estimation method based on convolutional neural networks, but the model has three limitations: First, the large number of network parameters leads to excessive computational resources required for training; second, the training data relies too much on ideal conditions, using only the ideal covariance matrix instead of the more generalized sampling covariance matrix, which affects the model's generalization ability; finally, the model can only process a maximum of 3 information sources, and its processing capability is far lower than that of traditional methods.

[0005] SLZheng et al. proposed a DOA estimation method based on original in-phase and orthogonal signal data. However, this method has two shortcomings: in feature construction, simply concatenating the original in-phase and orthogonal signal data into a two-dimensional matrix may not be the optimal representation; in terms of network structure, a single residual network structure is difficult to capture both the local details and global context of the signal at the same time, and it is also difficult to effectively utilize prior knowledge such as array geometry.

[0006] Deep learning's advantages lie in its powerful nonlinear fitting and generalization capabilities, fast inference speed once training is complete, and adaptability to complex environments. Nevertheless, existing deep learning-based DOA estimation methods still face challenges: performance depends on the scale and quality of training data, and its generalization ability is insufficient in unknown scenarios; high-dimensional inputs lead to large computational and storage overheads, making deployment on edge devices difficult; and its ability to handle dynamic sources and fuse physical models still needs improvement. Therefore, there is an urgent need in this field for a new DOA estimation method that balances high accuracy, strong generalization, low complexity, and adaptability to dynamic environments. Summary of the Invention

[0007] Objective: To address the limitations of traditional convolutional kernels with their small receptive fields in feature extraction of array covariance matrices and the inability of traditional training methods to accurately estimate the DOA of large numbers of sources, this invention proposes a DOA estimation method based on a global dynamic convolutional neural network. This method integrates the feature information of the covariance matrix through matrix operations without sacrificing the resolution of the covariance matrix, obtaining a globally perceptive convolutional kernel for convolution operations. This ensures the utilization rate of feature information, and through pre-training and fine-tuning methods, it achieves DOA estimation for a wider range of sources.

[0008] Technical solution: In one aspect, this invention provides a DOA estimation method based on a global dynamic convolutional neural network, comprising the following steps:

[0009] S1: Obtain the time-domain signal received by the uniform linear array, perform preprocessing, generate one-hot encoded tags for the corresponding time-domain signal, and reshape the dimensions of the time-domain signal;

[0010] S2: Construct a DOA estimation model and use a global dynamic convolutional neural network to extract deep features from the preprocessed time-domain signal; based on the extracted deep features, use a classification head to estimate the direction of arrival (DOA) of the signal source through classification.

[0011] S3: First, use a two-source ideal covariance matrix dataset to pre-train the DOA estimation model to obtain the initial weights. Then, use a four-source sampled covariance matrix dataset to fine-tune the initial weights and verify them.

[0012] In another aspect, the present invention also provides a DOA estimation system based on a global dynamic convolutional neural network, comprising the following modules:

[0013] The signal data processing module is used to acquire the time-domain signal received by the uniform linear array and perform preprocessing.

[0014] The Direction of Arrival (DOA) estimation module is used to construct a DOA estimation model and estimate the DOA of the signal source based on the preprocessed time-domain signal.

[0015] The fine-tuning training module is used to first pre-train the DOA estimation model using a two-source ideal covariance matrix dataset to obtain initial weights, and then fine-tune the initial weights using a four-source sampled covariance matrix dataset and verify them.

[0016] The design principle of this invention is as follows:

[0017] 1. A uniform linear array receives signals from the signal sources and integrates them using a global dynamic convolutional kernel to fully utilize features, improving the utilization rate of feature information and the DOA estimation accuracy under additive white Gaussian noise. Furthermore, a fine-tuning strategy is proposed to effectively increase the number of signal sources to be estimated. The advantage of this strategy is that it does not require the same number of training epochs as the pre-training stage, but instead achieves better estimation results for a larger number of signal sources with fewer epochs, thus reducing runtime.

[0018] 2. In the method of the present invention, a uniform linear array is used to receive data, a pseudo-covariance matrix is ​​obtained through a global dynamic convolutional neural network, a classifier is used to obtain the probability values ​​of 121 angles from the generated pseudo-covariance matrix, and then the accurate DOA estimated angle is selected based on the probability values.

[0019] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0020] The uniform linear array receives signals from the signal source. By utilizing the robustness and generalization ability of the network model of this invention, the accuracy of DOA estimation is improved. It can estimate a large number of signal sources with good accuracy even in harsh environments. At the same time, it introduces relative position coding information and effectively utilizes prior knowledge such as array geometry.

[0021] Compared to other traditional deep learning DOA estimation algorithms, the global dynamic perception convolution kernel effectively integrates the feature information of the covariance matrix, taking into account both local details and global features, greatly improving feature utilization. Furthermore, the fine-tuning technique can obtain estimation results for a larger number of sources with less training overhead. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the uniform linear array structure involved in this invention;

[0023] Figure 2This is a schematic diagram of the deep learning model structure involved in this invention;

[0024] Figure 3 The results of the four-source prediction of this invention, IQResNet, and CNN;

[0025] Figure 4 The results of the five-source prediction of this invention, IQResNet, and CNN;

[0026] Figure 5 The results of the 6-source prediction of this invention, IQResNet, and CNN;

[0027] Figure 6 The results of the seven-source prediction of this invention, IQResNet, and CNN;

[0028] Figure 7 This is a performance comparison chart of the method of this invention with other algorithms under different signal-to-noise ratio conditions;

[0029] Figure 8 This is a performance comparison chart of the method of this invention with other algorithms under different snapshot numbers. Detailed Implementation

[0030] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0031] This embodiment is based on the following principle: In one aspect, the present invention provides a DOA estimation method based on a global dynamic convolutional neural network. The specific method of this embodiment is as follows:

[0032] I. Signal Model and Preprocessing

[0033] Signal model:

[0034] like Figure 1 The uniform linear array antenna structure shown is composed of It consists of several array elements, with the first sensor as the reference element, and the spacing between the array elements is... The wavelength is .

[0035] Assumption The sources are respectively Narrowband incoherent signals are incident on a signal with On a uniform linear array (ULA) with n array elements, then the nth The sensor at the first The complex envelope received in the next snapshot is:

[0036] (1)

[0037] For the first The signals received by each sensor Represents the first One signal source, For the number of snapshots, , For the first The sensor receives the first signal relative to the reference element. Phase delay of the signal propagation.

[0038] The received data model can be represented as:

[0039] (2)

[0040] For the received signal matrix, The direction matrix, in It is a direction vector. It is a signal vector. It is a noise vector.

[0041] Preprocessing:

[0042] From the received signal Obtaining the ideal covariance matrix and sampling covariance matrix The real, imaginary, and phase parts of these two covariance matrices are extracted respectively, serving as the three channel information of the input data. These three parts are then stacked along the 0th dimension to generate a final dimension... The three-dimensional tensor adapts to the data dimension of the network input.

[0043] II. Deep Feature Extraction

[0044] As shown in the figure, a deep neural network containing the following modules was constructed:

[0045] 1. Embedded layer module

[0046] The embedding layer module employs a three-level convolutional structure, extracting spatial features from low to high order layer by layer. Specifically, it consists of three convolutional layers, with each kernel size being [missing information]. The step size is 1, the padding is 1, and the number of channels is 32, 64, and 128 respectively, using two-dimensional batch normalization and the SiLU activation function. The embedding layer module's role is to encode the raw, redundant, and noisy covariance matrix data into features, gradually transforming it into a highly refined feature representation most effective for the DOA estimation task, converting the physical signal into depth features, and realizing an end-to-end mapping from the array covariance matrix to depth space features. Its mathematical expression is as follows:

[0047] (3)

[0048] In the formula, For input data, For two-dimensional batch normalization layer, The kernel size is Convolutional layers, These are the high-level features after encoding.

[0049] 2. Global Dynamic Convolution Module

[0050] A feature calculation method based on dual-scale dynamic convolution, which specifically includes the following steps:

[0051] Step 1: Input Feature Preprocessing

[0052] Obtain the input feature tensor Its dimensions are ,in For batch size, For the number of channels, , The height and width of the feature map are defined; the input features are sequentially processed by depthwise separable convolution, layer normalization, and gated transformation. The gated transformation is achieved through... The convolutional layer is used to learn the feature importance weights; the depthwise separable convolution is used to subsequently implement the relative position bias.

[0053] Step 2: Generation of affinity matrix and global dynamic convolution kernel:

[0054] First, copy the preprocessed input feature data three times. The first copy is then processed by... The convolutional layer and dimension reshaping operation yield the Query matrix; the second part is passed through an adaptive pooling layer, The convolutional layers and dimension reshaping operations yield the Keys matrix; the third matrix is ​​reshaped to obtain the Values ​​matrix, the purpose of which is to adjust the dimensions to fit subsequent convolutional operations. Then, the Einstein summation convention is applied to the Query and Keys matrices to obtain the affinity matrix. , where the matrix It integrates global information about the array covariance matrix.

[0055] matrix pass Convolutional layers expand the number of channels; through a partitioning operation, two independent weight tensors are obtained, corresponding to a small convolutional kernel and a large convolutional kernel, respectively. and The small kernel's corresponding weight tensor is responsible for capturing local details, while the large kernel's corresponding weight tensor is responsible for extracting global context, thus forming a complementary dual-scale perception mechanism.

[0056] Finally, based on the input features Tensor Size For two independent weight tensors respectively and Generate two learnable relative position bias tensors. and The system initializes the input feature map using a truncated normal distribution. An index generation function generates coordinate indices for each position within two independent weight tensors, including height and width indices. A one-dimensional index in the relative position bias matrix is ​​obtained by multiplying the height index by the width of the relative position bias matrix and then adding the width index. Finally, the input feature map's height is used as the basis for the initialization. and width Create repeating pattern tensors in the height and width directions, generate a two-dimensional index for each position in the relative position offset tensor through repeated interpolation operations, and then add it to the one-dimensional index to obtain the complete relative position offset tensor. and Then the relative position bias tensor and Corresponding element-wise with two independent weight tensors and The summation ultimately generates a location-aware tensor that incorporates location information. and ;

[0057] Step 3: Implementation of global dynamic convolution:

[0058] Position-aware tensor and After performing a linear transformation, a global dynamic vector is obtained through a Softmax layer. Then, a dimension reshaping operation is performed on this vector to obtain the final global dynamic convolution kernel. and Then, a convolution operation similar to that in a convolutional neural network is performed with the Values ​​matrix to achieve global dynamic convolution. Finally, different global dynamic convolution kernels are... and The outputs of the global dynamic convolution module are concatenated along the channel dimension. This enables the effective fusion of local detailed features and global contextual information.

[0059] Multi-task DOA estimation

[0060] The classification head output layer consists of an adaptive average pooling layer and a convolutional layer. It outputs probability values ​​in 121 directions using a sigmoid function, and finally selects the probability with the highest probability through spectral peak search. Each value is used as the DOA output, and its mathematical expression is:

[0061] (4)

[0062] In the formula, Estimate the angle for DOA. For peak search function, For neural networks, This is the Sigmoid function.

[0063] IV. Model Training Steps

[0064] This step employs a phased training strategy, namely transfer learning. First, pre-training is performed on an idealized dataset, followed by fine-tuning on a sampled dataset that more closely resembles real-world scenarios, ensuring the model possesses both strong generalization ability and high prediction accuracy. The specific steps are as follows:

[0065] 1. Pre-training stage: Based on a dataset of ideal covariance matrices from two sources.

[0066] (1) Construction of pre-training dataset

[0067] Based on the uniform linear array antenna structure, the signal-to-noise ratio parameter is... arrive by Discretize the step size; simultaneously, the incident angle parameter of each signal source is... to Within the azimuth range, with The resolution is discretized. Then, from the discretized angle values, all unique pairwise combinations are exhaustively enumerated to simulate different dual-source spatial distribution scenarios. For each scenario defined by a specific signal-to-noise ratio and a signal angle pair, an independent training sample is generated by calculating its corresponding ideal covariance matrix. By iterating through all parameter combinations, the dual-source ideal covariance matrix dataset is integrated.

[0068] (2) Pre-training process

[0069] Using the aforementioned two-source ideal covariance matrix dataset, the model is initially trained on all learnable parameters. The training objective of this stage is to enable the model to initially grasp the basic rules for extracting effective features from the covariance matrix and mapping them to the desired output from clean, low-complexity data, thus laying a solid initial weight foundation for subsequent learning.

[0070] 2. Fine-tuning stage: Based on the four-source sampling covariance matrix dataset

[0071] (1) Construction of the fine-tuning dataset

[0072] The construction of the fine-tuning dataset follows the same steps as the pre-training dataset construction in terms of the discretization rules for the signal-to-noise ratio and signal angle parameters. The key difference lies in expanding the number of signal sources in the scene to four, i.e., performing all four-tuple combinations on the discretized angle values ​​to simulate a four-source environment. For each set of parameters, the sampling covariance matrix is ​​calculated using a finite number of snapshots of the received signal sequence as an estimate of the scene. Finally, the four-source sampling covariance matrix fine-tuning dataset is integrated.

[0073] (2) Fine-tuning process

[0074] Based on the initial model weights obtained in the pre-training phase, the model is trained a second time using the four-source sampling covariance matrix dataset. This second phase employs a significantly reduced learning rate strategy to fine-tune the model parameters in a progressive manner. This training process aims to adapt the model to non-ideal factors in real-world electromagnetic environments, including inherent system noise, multipath propagation interference, and statistical estimation errors caused by a limited number of snapshots. This transforms the model's basic feature mapping ability learned under ideal data distributions into robust inference capabilities that maintain high-precision estimations in complex observation data, ultimately comprehensively improving the model's generalization performance and practical value in real-world scenarios.

[0075] In another aspect, the present invention also provides a DOA estimation system based on a global dynamic convolutional neural network, comprising the following modules:

[0076] The signal data processing module is used to acquire the time-domain signal received by the uniform linear array and perform preprocessing.

[0077] The Direction of Arrival (DOA) estimation module is used to construct a DOA estimation model and estimate the DOA of the signal source based on the preprocessed time-domain signal.

[0078] The fine-tuning training module is used to first pre-train the DOA estimation model using a two-source ideal covariance matrix dataset to obtain initial weights, and then fine-tune the initial weights using a four-source sampled covariance matrix dataset and verify them.

[0079] The direction of arrival (DOA) estimation module includes a global dynamic convolutional neural network unit and a classification unit.

[0080] The global dynamic convolutional neural network unit is used to construct an affinity matrix and extract deep features from the preprocessed time-domain signal using the global dynamic convolutional neural network.

[0081] The classification unit is used to estimate the direction of arrival (DOA) of the signal source by using a classification head based on the extracted deep features.

[0082] V. Model Training Configuration and Parameter Settings

[0083] 1. Detailed configuration of network structure parameters

[0084] (1) Embedded layer module parameters

[0085] Table 1 shows the specific parameter configurations for the embedding layer in the neural network structure of this method.

[0086] Table 1

[0087]

[0088] (2) Parameters of the global dynamic convolution module

[0089] Table 2 shows the specific parameter configurations for global dynamic convolution in the neural network structure of this method.

[0090] Table 2

[0091]

[0092] (3) Classification Header Module Parameters

[0093] Table 3 shows the specific parameter configuration of the classification head in the neural network structure of this method.

[0094] Table 3

[0095]

[0096] 3. Pre-training phase configuration

[0097] Table 4 shows the specific parameter configurations for the pre-training of this method.

[0098] Table 4

[0099]

[0100] 4. Fine-tuning stage configuration

[0101] Table 5 shows the specific parameter configurations for fine-tuning this method.

[0102] Table 5

[0103]

[0104] 4. Hardware and software environment configuration

[0105] Table 6 provides the specific experimental environment configuration.

[0106] Table 6

[0107]

[0108] VI. Performance Analysis and Experimental Analysis

[0109] The model methods compared to this invention (DGCNet) are a residual network based on original in-phase and orthogonal signal components (IQResNet) and a deep convolutional neural network (CNN). Table 7 shows the specific model index parameters of the model of this invention and the comparative models.

[0110] 1. Analysis of Model Parameter Quantity and Complexity

[0111] The formula for calculating the total number of parameters in the global dynamic convolutional neural network is as follows:

[0112] (5)

[0113] In the formula, For the number of parameters in the embedding layer, For residual connection parameters, The number of global dynamic convolution parameters. The number of parameters is the number of classification head parameters. Calculations show that the total number of parameters in the model of this invention is 1.15M, which is lower than that of the comparison model.

[0114] The computational complexity of the model is measured by the number of floating-point operations (FLOPs) during the forward propagation process. The total model complexity is:

[0115] (6)

[0116] In the formula, The number of floating-point operations performed in the embedding layer. The number of floating-point operations performed to join the residuals. This represents the number of floating-point operations performed in the global dynamic convolution. The number of floating-point operations is for the classification header. Calculations show that the model of this invention performs [a certain number of] operations in [a certain way]. The total FLOPs under the input dimension is 280.06M.

[0117] Table 7

[0118]

[0119] 2. Experimental Analysis

[0120] To verify the effectiveness of the above method, multiple simulation experiments were conducted in this embodiment, and the experimental performance was analyzed, as follows:

[0121] (1) Experimental performance evaluation indicators

[0122] Signal-to-noise ratio (SNR) is defined as:

[0123] (7)

[0124] in, For signal power, This represents noise power.

[0125] The performance estimation criterion is the Joint Root Mean Square Error (RMSE), defined as follows:

[0126] (8)

[0127] in, For the first The Monte Carlo Trial The precise estimate of DOA for each source. Indicates the number of information sources. This indicates the number of Monte Carlo trials.

[0128] (2) Experimental results diagram

[0129] Figure 1 A schematic diagram of the uniform linear array structure involved in this invention is shown. Figure 2 A schematic diagram of the deep learning model structure involved in this invention is shown. Figure 3 , Figure 4 , Figure 5 and Figure 6 The number of sources is given under additive white Gaussian noise conditions. The estimation results for incoherent signals 4, 5, 6, and 7 are shown in the figure. Figure 3 The prediction results for the four sources corresponding to DGCNet, IQResNet, and CNN; Figure 4 The prediction results for the five sources corresponding to DGCNet, IQResNet, and CNN; Figure 5 The prediction results for the six sources corresponding to DGCNet, IQResNet, and CNN; Figure 6 The prediction results for 7 sources correspond to DGCNet, IQResNet, and CNN. When the number of sources... At that time, the signal-to-noise ratio is With a snapshot count of 100, the signal-to-noise ratio and snapshot count are both 100 when the number of information sources is 5, 6, and 7. And 400. It can be seen that, under the pre-training and fine-tuning method, DGCNet maintains good performance, and the model can accurately estimate the cases with 4, 5, and 6 information sources. In this case, the model can still accurately estimate all sources, but a spurious peak appears in the DOA spectrum. Considering that the number of sources already far exceeds the number of sources contained in the dataset, a slight decrease in model performance is acceptable. Although IQResNet and CNN can also accurately estimate with 4 and 5 sources, when the number of sources is further increased, the predicted amplitude of both methods is not as good as DGCNet.

[0130] Figure 7 In signal-to-noise ratio Quick shot number The two incident angles are respectively Monte Carlo trials The RMSE plot of the compared models in the scenario. From Figure 4 As can be seen, the RMSE of all models gradually decreases with increasing signal-to-noise ratio, but the performance of the comparison models IQResNet and CNN is inferior to that of DGCNet. Furthermore, due to the limitation of grid resolution, DGCNet... It has already reached its performance limit, indicating that DGCNet can accurately estimate DOA even under low signal-to-noise ratio conditions, while the other two algorithms need to further increase the signal-to-noise ratio to reach their performance limit.

[0131] Figure 8 In the number of snapshots Signal-to-noise ratio The two incident angles are respectively Monte Carlo trials The RMSE plot of the compared models in the scenario. From Figure 5 As can be seen, the RMSE of all models gradually decreases with the increase of the number of snapshots, and DGCNet consistently outperforms IQResNet and CNN, while maintaining good estimation accuracy even with a small number of snapshots.

Claims

1. A DOA estimation method based on a global dynamic convolutional neural network, characterized in that, Includes the following steps: S1: Obtain the time-domain signal received by the uniform linear array and perform preprocessing; S2: Construct a DOA estimation model and use a global dynamic convolutional neural network to extract deep features from the preprocessed time-domain signal; Based on the extracted deep features, the direction of arrival (DOA) of the signal source is estimated by using a classification head and a classification method. S3: First, use a two-source ideal covariance matrix dataset to pre-train the DOA estimation model to obtain the initial weights. Then, use a four-source sampled covariance matrix dataset to fine-tune the initial weights and verify them.

2. The DOA estimation method based on a global dynamic convolutional neural network according to claim 1, characterized in that, The specific implementation process of preprocessing in step S1 is as follows: generating one-hot encoded labels for the corresponding time-domain signals and reshaping the dimensions of the time-domain signals.

3. The DOA estimation method based on a global dynamic convolutional neural network according to claim 1, characterized in that, The specific implementation process of the global dynamic convolutional neural network is as follows: S3.1: Embedding Layer Module. The embedding layer module adopts a three-level convolutional structure to extract spatial features layer by layer, realizing an end-to-end mapping from the array covariance matrix to the depth spatial features, and obtaining the feature tensor. ; S3.2: Construct a global dynamic convolution module for feature tensors. Feature calculation is performed to obtain the output of the global dynamic convolution module. .

4. The DOA estimation method based on a global dynamic convolutional neural network according to claim 3, characterized in that, The global dynamic convolution module is implemented as follows: Input feature preprocessing: processing the input feature tensor The process involves sequentially performing depthwise separable convolution, layer normalization, and gating transformation; the gating transformation is achieved through... Convolutional layers are used to learn the weights of feature importance. Copy the preprocessed input feature data three times. The first copy is then processed... The convolutional layer and dimension reshaping operation yield the Query matrix; the second part is passed through an adaptive pooling layer, The convolutional layer and dimension reshaping operation yield the Keys matrix; the third matrix is ​​reshaped to obtain the Values ​​matrix. Subsequently, the Einstein summation convention is applied to the Query and Keys matrices to obtain the affinity matrix. ; matrix pass Convolutional layers expand the number of channels; through a partitioning operation, two independent weight tensors are obtained. and Based on input features For two independent weight tensors respectively and Generate two learnable relative position bias tensors. and Then, these two learnable relative position bias tensors are transformed into relative bias tensors of the two independent weight tensors using an index generation function. and ; the relative bias tensor and Corresponding element-wise with two independent weight tensors and Add them together to generate a location-aware tensor that fuses location information. and ; Position-aware tensor and After performing a linear transformation, a global dynamic vector is obtained through a Softmax layer. Then, a dimension reshaping operation is performed on this vector to obtain the final global dynamic convolution kernel. and Then, a convolution operation is performed with the Values ​​matrix to achieve a global dynamic convolution operation; finally, the Values ​​matrix is ​​convolved with... and The outputs of the convolutions are concatenated along the channel dimension to obtain the output of the global dynamic convolution module. .

5. The DOA estimation method based on a global dynamic convolutional neural network according to claim 4, characterized in that, The classification header is implemented as follows: The classification head output layer consists of an adaptive average pooling layer and a convolutional layer. It outputs probability values ​​using a sigmoid function and finally selects the values ​​with the highest probabilities through spectral peak search. Each value is output as a DOA.

6. The DOA estimation method based on a global dynamic convolutional neural network according to claim 5, characterized in that, The specific construction of the two-source ideal covariance matrix dataset is as follows: Within a uniform linear array antenna structure, the signal-to-noise ratio parameter changes from... arrive by Discretize the step size; simultaneously, the incident angle parameter of each signal source is... to Within the azimuth range, with Discretize the resolution values; then, exhaustively enumerate all unique pairwise combinations from the discretized angle values ​​to simulate different dual-source spatial distribution scenarios; for each scenario defined by the signal-to-noise ratio and a signal angle pair, generate an independent training sample by calculating its corresponding ideal covariance matrix; and integrate the dual-source ideal covariance matrix dataset by traversing all parameter combinations.

7. The DOA estimation method based on a global dynamic convolutional neural network according to claim 6, characterized in that, The four-source sampling covariance matrix dataset is specifically constructed as follows: The construction of the fine-tuning stage dataset follows the same steps as the aforementioned pre-training dataset construction in terms of the discretization rules for the signal-to-noise ratio and signal angle parameters. The difference lies in that the number of signal sources in the scene is expanded to four, i.e., all four-tuple combinations are performed on the discretized angle values ​​to simulate a four-source environment. The sampled covariance matrix of the received signal sequence from a finite number of snapshots is calculated as an estimate of the scene, and finally, the four-source sampled covariance matrix fine-tuning dataset is integrated.

8. A DOA estimation system based on a global dynamic convolutional neural network, used to implement the DOA estimation method according to any one of claims 1 to 7, characterized in that, Includes the following modules: The signal data processing module is used to acquire the time-domain signal received by the uniform linear array and perform preprocessing. The Direction of Arrival (DOA) estimation module is used to construct a DOA estimation model and estimate the DOA of the signal source based on the preprocessed time-domain signal. The fine-tuning training module is used to first pre-train the DOA estimation model using a two-source ideal covariance matrix dataset to obtain initial weights, and then fine-tune the initial weights using a four-source sampled covariance matrix dataset and verify them.

9. The DOA estimation system based on a global dynamic convolutional neural network according to claim 8, characterized in that, The Direction of Attraction (DOA) estimation module includes a global dynamic convolutional neural network unit and a classification unit; The global dynamic convolutional neural network unit is used to construct the affinity matrix and extract deep features from the preprocessed time-domain signal using the global dynamic convolutional neural network. The classification unit is used to estimate the direction of arrival (DOA) of the signal source by using a classification head based on the extracted deep features.

Images