A DOA estimation complex-valued method based on variable kernel multi-branch convolution

Abstract

This application provides a complex-valued DOA estimation method based on multi-branch convolution with a variable kernel, belonging to the technical field of signal processing. The method calculates the real and imaginary parts of the complex-valued signal; based on these parts, it expands the signal to obtain first data; it then calculates the first data using a heuristic residual super-resolution module to obtain second data; finally, it fuses the second data to obtain the DOA estimation result. Within the applicable range of DOA methods, this method consistently outperforms traditional algorithms under the same signal-to-noise ratio conditions.

Description

Technical Field

[0001] This application relates to the field of signal processing, and in particular to a method for DOA estimation of complex values ​​based on multi-branch convolution with variable kernel. Background Technology

[0002] In existing DOA applications, traditional DOA algorithms require increasing the number of radar elements and snapshots to achieve higher angle estimation accuracy, resulting in high usage and computational costs. Furthermore, traditional techniques have poor angle estimation capabilities in low-signal-to-noise environments. Meanwhile, using deep learning methods to solve DOA problems has become a popular research direction in recent years. However, currently, most methods employ ordinary convolutional networks, which suffer from problems such as excessive parameters, high computational cost, and insufficient receptive field. Summary of the Invention

[0003] The main objective of this application is to provide a method for DOA estimation complex values ​​based on multi-branch convolution with variable kernels.

[0004] On one hand, embodiments of the present invention provide a method for DOA estimation of complex values ​​based on multi-branch convolution with variable kernels, the method comprising the following steps:

[0005] The real and imaginary parts of the signal are calculated by performing calculations on the complex-valued signal.

[0006] Based on the real and imaginary part information, the signal is extended to obtain the first data;

[0007] The second data is obtained by calculating the first data using the heuristic residual super-resolution module;

[0008] The second data is fused to obtain the DOA estimation result.

[0009] Furthermore, the formulas used to calculate the real and imaginary parts of the complex-valued signal include:

[0010]

[0011] in, W is the normalized matrix of the complex-valued signal with size 1×N×N, and M is the complex weight parameter matrix with size N×M, where M represents the number of discrete points in the probability distribution of the final output of the neural network. represents the real and imaginary parts of the signal; j is the imaginary number.

[0012] Further, the calculation of the complex-valued signal to obtain the real and imaginary parts of the signal includes the following steps:

[0013] Establish a complex signal processing module;

[0014] According to the complex signal processing module, the input complex signal is extended by length feature number through a complex-valued linear layer to obtain the real and imaginary parts of the signal.

[0015] Further, the step of expanding the signal based on the real and imaginary part information to obtain the first data includes the following steps:

[0016] Based on the real and imaginary part information, the feature channels of the signal are expanded using a complex-valued convolution kernel to obtain the first data.

[0017] Further, the step of expanding the feature channels of the signal using a complex-valued convolution kernel based on the real and imaginary part information to obtain the first data includes the following steps:

[0018] An intermediate value is obtained by applying multiple independent complex convolution kernels to the real and imaginary parts of the information; the intermediate value includes the real part and the imaginary part of the intermediate value.

[0019] The modulo value of the intermediate value is taken to expand the characteristic channels of the signal and obtain the first data;

[0020] The formula for calculating the intermediate value by using multiple independent complex convolution kernels to process the real and imaginary parts of the information is as follows:

[0021]

[0022]

[0023] in, Let be the real part of the intermediate value. W1 is the imaginary part of the intermediate value; W1 is the complex convolution kernel with a size of K×3×3; K is the number of complex convolution kernels; The real and imaginary parts are the information;

[0024] The process of taking the modulo value of the intermediate value to expand the feature channels of the signal yields the following calculation formula for the first data:

[0025]

[0026] in, This refers to the first data.

[0027] Further, the step of calculating the second data from the first data using the heuristic residual super-resolution module includes the following steps:

[0028] A heuristic residual super-resolution module is constructed using several of the aforementioned heuristic residual blocks;

[0029] The second data is obtained by calculating the first data using the aforementioned residual super-resolution module.

[0030] Furthermore, the step of establishing a heuristic residual super-resolution module using a plurality of the heuristic residual blocks includes the following steps:

[0031] A number of revelation residual blocks are set up; each of the revelation residual blocks contains multiple paths;

[0032] Set the kernel size of the heuristic residual block;

[0033] Set the proportion of the number of channels in the path output;

[0034] Introduce residual connections within each of the aforementioned heuristic residual blocks;

[0035] Add a batch normalization layer to each of the aforementioned revelation residual blocks;

[0036] By stacking several of the revelation residual blocks processed in the above steps, the revelation residual super-resolution module is completed.

[0037] Furthermore, the second data is obtained by calculating the first data using the heuristic residual super-resolution module, and the formula used includes:

[0038]

[0039] Where F represents the residual function in the network, W L This represents the weight parameters of the Lth heuristic residual block;

[0040] This is the second data; the second data With the first data They have the same size.

[0041] Further, the process of fusing the second data to obtain the DOA estimation result includes the following steps:

[0042] Retrieve the second data in both dimension two and dimension three;

[0043] The second data is convolved using a convolution kernel to obtain convolutional data.

[0044] The convolutional data are fused to obtain the DOA estimation result.

[0045] Furthermore, the formula used in the DOA estimation complex value method based on variable kernel multi-branch convolution includes:

[0046]

[0047] p(n) = Softmax(out(n))

[0048] Where out() is the output of the fusion module. N represents the neural network's representation data, where N is the number of discrete points in the output; CrossEntropyLoss is the cross-entropy loss function.

[0049] On the other hand, embodiments of the present invention also provide a DOA estimation complex value device based on multi-branch convolution with variable kernel. The device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0050] On the other hand, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described above.

[0051] The embodiments of this application include at least the following beneficial effects: This application provides a complex-valued DOA estimation method based on multi-branch convolution with a variable kernel. The present invention calculates the real and imaginary parts of the complex-valued signal; based on the real and imaginary parts, the signal is extended to obtain first data; the first data is calculated using a heuristic residual super-resolution module to obtain second data; and the second data is fused to obtain the DOA estimation result. Within the applicable range of DOA methods, the accuracy of this method is generally superior to traditional algorithms under the same signal-to-noise ratio conditions. Attached Figure Description

[0052] Figure 1 This is a flowchart of the complex value estimation method for DOA based on multi-branch convolution with variable kernel provided in this embodiment of the invention.

[0053] Figure 2 This is a schematic diagram of the DOA estimation architecture of IRMAEN provided in an embodiment of the present invention.

[0054] Figure 3 This is a schematic diagram of the Inception module design provided in an embodiment of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0056] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0057] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0059] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0060] 1) DOA (Direction of Arrival) is the direction cosine. DOA estimation methods are commonly used in the field of signal processing. In array signal processing, it is used to determine the direction of the signal source relative to the receiving array.

[0061] 2) Complex Linear Layer is a type of neural network layer used to process complex inputs and complex weights.

[0062] 3) Complex Convolutional Layer is a special layer in convolutional neural networks used to process convolution operations between complex inputs and complex kernels.

[0063] 4) Residual Layer is a structure that is typically used to build deep residual networks.

[0064] 5) The Inception Residual Block is a deep neural network module that combines the Inception module and the Residual Connection. The Inception module is a multi-branch structure proposed by Google, and the Residual Connection is a connection method proposed by ResNet.

[0065] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0066] On one hand, embodiments of the present invention provide a method for DOA estimation complex values ​​based on multi-branch convolution with variable kernels. Specifically, refer to... Figure 1 The method includes the following steps:

[0067] S100. Calculate the complex-valued signal to obtain the real and imaginary parts of the signal;

[0068] S200: Based on the real and imaginary part information, the signal is extended to obtain the first data;

[0069] S300: The first data is calculated using the heuristic residual super-resolution module to obtain the second data;

[0070] S400. The second data is fused to obtain the DOA estimation result.

[0071] This invention discloses step S100, which calculates the complex-valued signal to obtain the real and imaginary parts of the signal. The formula used includes:

[0072]

[0073] in, W is the normalized matrix of the complex-valued signal with size 1×N×N, and M is the complex weight parameter matrix with size N×M, where M represents the number of discrete points in the probability distribution of the final output of the neural network. represents the real and imaginary parts of the signal; j is the imaginary number.

[0074] This invention discloses step S100, which involves calculating the real and imaginary parts of a complex signal to obtain the signal's information. This includes the following steps:

[0075] S110. Establish a complex signal processing module;

[0076] S120. According to the complex signal processing module, the input complex signal is extended by the length feature number through the complex-valued linear layer to obtain the real and imaginary part information of the signal.

[0077] As an optional implementation, in S110 of the present invention, a complex signal processing module is first established, which is designed to process the input complex signal. The complex signal processing module is designed to efficiently process data in the complex domain and perform the operations and transformations required for subsequent steps. This module typically includes mathematical operations, transformation functions, and appropriate data structures to ensure the correct processing and analysis of the complex signal.

[0078] In S120 of this invention, a complex signal processing module is established to process the input complex signal through a complex-valued linear layer, thereby expanding the length feature number. This process aims to convert or expand the input complex signal into its real and imaginary parts for subsequent analysis and application. The role of the complex-valued linear layer is to perform specific mathematical operations to ensure that the real and imaginary parts of the signal can be accurately extracted and recorded to meet the needs of subsequent algorithms or systems.

[0079] This invention discloses step S200, which involves expanding the signal based on the real and imaginary parts to obtain first data, including the following steps:

[0080] S210. Based on the real and imaginary part information, the feature channels of the signal are expanded using a complex-valued convolution kernel to obtain the first data.

[0081] This invention discloses step S210, which involves expanding the feature channels of a signal using a complex-valued convolution kernel based on the real and imaginary part information to obtain first data. This includes the following steps:

[0082] S211. By using multiple independent complex convolution kernels to process the real and imaginary information, an intermediate value is obtained; the intermediate value includes the real part and the imaginary part of the intermediate value.

[0083] S212. Take the modulus of the intermediate value to expand the characteristic channel of the signal and obtain the first data;

[0084] By using multiple independent complex convolution kernels to process the real and imaginary parts, the formula for calculating the intermediate value is as follows:

[0085]

[0086]

[0087] in, The real part of the intermediate value. The imaginary part of the intermediate value; W1 is a complex convolution kernel of size K×3×3; K is the number of complex convolution kernels; Information consisting of the real and imaginary parts;

[0088] Taking the modulo value of the intermediate value expands the characteristic channels of the signal, resulting in the following formula for calculating the first data:

[0089]

[0090] in, This is the first data point.

[0091] This invention discloses step S300, which involves calculating the first data using a heuristic residual super-resolution module to obtain the second data, including the following steps:

[0092] S310. Use several heuristic residual blocks to establish a heuristic residual super-resolution module;

[0093] S320. The first data is calculated using the heuristic residual super-resolution module to obtain the second data.

[0094] This invention discloses step S310, which involves establishing a heuristic residual super-resolution module using a plurality of heuristic residual blocks, including the following steps:

[0095] S311. Set up several revelation residual blocks; each revelation residual block contains multiple paths;

[0096] S312, Set the kernel size of the heuristic residual block;

[0097] S313, Set the channel ratio for path output;

[0098] S314. Introduce residual connections within each revelation residual block;

[0099] S315. Add a batch normalization layer to each revelation residual block;

[0100] S316. Stack several revelation residual blocks processed in the above steps to complete the construction of the revelation residual super-resolution module.

[0101] As an optional implementation, in step S311 of this invention, the number of heuristic residual blocks to use in constructing the super-resolution module is first determined. Each heuristic residual block refers to a specific neural network structure that contains multiple paths or branches for processing different aspects or features of the input data. These paths can achieve different feature extraction and transformations through different convolutional kernels, activation functions, or other hierarchical structures.

[0102] In S312 of this invention, setting the size of the convolution kernel within each heuristic residual block is a crucial step. The size of the convolution kernel determines the granularity and scope of spatial feature extraction performed within a specific heuristic residual block. Typically, the size of the convolution kernel can be selected based on the needs of the problem and the size of the input data to ensure that dimensionality is effectively reduced or features are enhanced while preserving important information.

[0103] In S313 of this invention, different paths or branches in each heuristic residual block typically generate different numbers and types of feature maps. In this step, setting the channel ratio of each path's output can adjust the information flow and importance between different feature maps, allowing for more effective integration of these features to improve model performance.

[0104] The S314 residual connection of this invention refers to adding direct skip connections between different layers of a neural network, with the aim of reducing information loss during transmission within the network. This connection method helps alleviate the vanishing gradient problem and effectively helps the network learn deeper feature representations. Introducing such connections in each heuristic residual block can effectively enhance model performance and training speed.

[0105] In S315 of this invention, a batch normalization layer is added inside each heuristic residual block, which helps to normalize the distribution of input data and accelerates the convergence process by reducing the internal covariate offset.

[0106] In step S316 of this invention, the heuristic residual blocks, processed and configured in the above steps, are stacked according to the designed order and number. This stacking process is to construct a complete heuristic residual super-resolution module, which can achieve efficient super-resolution reconstruction tasks by learning complex nonlinear mappings. The stacking process of each heuristic residual block ensures that the model can effectively extract and reconstruct high-quality details and structures from the input image.

[0107] This invention discloses step S300, which uses a heuristic residual super-resolution module to calculate the first data to obtain the second data. The formula used includes:

[0108]

[0109] Where F represents the residual function in the network, W L This represents the weight parameters of the Lth heuristic residual block;

[0110] This is the second data; the second data With the first data They have the same size.

[0111] This invention discloses step S400, which involves fusing the second data to obtain the DOA estimation result, including the following steps:

[0112] S410, Obtain the second data in both dimension two and dimension three;

[0113] S420. Convolve the second data using a convolution kernel to obtain convolutional data;

[0114] S430. Fuse the convolutional data to obtain the DOA estimation result.

[0115] The complex value estimation method for DOA based on multi-branch convolution with variable kernel disclosed in this invention includes the following formulas:

[0116]

[0117] p(n) = Softmax(out(n))

[0118] Where out() is the output of the fusion module. N represents the neural network's representation data, where N is the number of discrete points in the output; CrossEntropyLoss is the cross-entropy loss function.

[0119] This invention considers that DOA (Direction of Arrival) estimation, in signal processing, refers to determining the direction of a signal source relative to a receiver by analyzing the received signal. This direction typically refers to the angle or direction vector of the signal source relative to the receiver. With the development of science and technology and the improvement of technical levels, the application of DOA estimation has gradually expanded to the civilian field. In the process of technological development, with the continuous progress of sensor technology and signal processing algorithms, the accuracy and performance of DOA estimation have been greatly improved. The increasing demand for spatial information in the civilian field, such as in communication systems, UAV navigation, and intelligent transportation, all require the location and tracking of signal sources, which has also promoted the application and development of DOA estimation technology in the civilian sector.

[0120] Deep learning is primarily based on the construction and training process of neural networks. Neural networks initially simulate the connections between neurons in the human brain, forming a network of multi-layered neurons to extract features and abstractly represent input data, ultimately enabling the prediction and classification of output data. In the early 2010s, with the emergence of deep learning models such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), deep learning achieved groundbreaking progress in fields such as image recognition and natural language processing. Compared to traditional methods, deep learning methods possess advantages such as strong feature extraction capabilities, strong ability to handle complex relationships, strong generalization ability, end-to-end learning, strong adaptability, and wide applicability. They overcome the limitations of traditional methods in the domain of Data of Applications (DOA) and have become an important new tool for solving complex problems and processing large-scale data.

[0121] The embodiments of the present invention can achieve high estimation accuracy with only a small number of array elements and snapshots. Within the applicable range of the DOA method, the accuracy of this method is generally better than that of traditional algorithms under the same signal-to-noise ratio, and the performance is even more obvious under the conditions of low signal-to-noise ratio and few snapshots.

[0122] In direction of arrival (DOA) estimation, the received signal is typically represented in complex form, including amplitude and phase information. This invention proposes a neural network method called the Inception Residual Multisnapshot Angle Estimation Network (IRMAEN), referencing... Figure 2 This is the DOA estimation architecture of IRMAEN in this embodiment of the invention.

[0123] The main feature of this method is that it transforms the N×K signal matrix X, composed of K snapshots received by the array, into... The resulting numerical matrix is ​​input as a whole into the complex neural network, instead of treating the real and imaginary parts as two separate channels. Furthermore, an Inception module is used to further improve the network's performance. The advantage of the IRMAEN network is that it improves the accuracy of multi-shot DOA estimation, outperforming traditional methods in both accuracy and efficiency at the same signal-to-noise ratio. Using this strategy, the network can more effectively utilize the structural and operational characteristics of complex signals, more accurately capture signal features, and significantly improve DOA estimation performance, especially in challenging noisy environments and with a limited number of snapshots.

[0124] In IRMAEN, this invention aims to take into account the characteristics of network data, extracting features from horizontal, vertical, and local perspectives. To this end, this invention employs convolutional layer designs with different sizes and output channel ratios in the Inception module, and simplifies the selection of hyperparameters to some extent. Figure 3The design of the Inception module in IRMAEN is shown.

[0125] Where n is the kernel size, p is the number of input channels, q is the number of output channels, and [ω1, ω2, ω3, ω4] are the proportions of the output channels for each branch. A 3n×n convolutional branch effectively extracts features between different snapshots, while an n×3n convolutional branch extracts features from a single snapshot. Furthermore, adding a 3n×3n convolutional branch along with parallel pooling paths further extracts the overall features of the signal. Based on the distribution characteristics of the output data, as the network deepens, the data near the top layer exhibits significant sparsity and locality. Therefore, it is necessary to use convolutional kernels of different sizes at different stages and adjust the proportions of the output channels for different paths. By combining different receptive fields, IRMAEN exhibits superior feature extraction capabilities.

[0126] To reduce the difficulty of selecting hyperparameters and minimize manual intervention, this invention sets the number of output channels and the branch ratio of the Inception module in each module of IRMAEN, and removes some 1×1 convolution kernels. This is because this invention stacks identical modules to expand the depth of the network rather than increasing the complexity of the Inception module.

[0127] As an optional implementation method, refer to Figure 2 The structure of the IRMAEN in this embodiment of the invention is mainly divided into three parts:

[0128] 1) Complex signal processing module

[0129] 2) Inception residual superresolution module

[0130] 3) Fusion module

[0131] (1) Complex signal processing module: First, through a complex-valued linear layer, this invention processes the input data. The process involved processing and expansion. This step utilizes the mathematical framework of complex numbers to extract the real and imaginary parts of the signal, effectively solving the problem of information loss in general methods. Specifically, this invention is calculated using the following mathematical expression.

[0132]

[0133] in, Let W be a normalized signal matrix of size 1×N×N, and W be a complex weight parameter matrix of size N×M, where M represents the number of discrete points in the probability distribution of the final output of the neural network, and j is an imaginary number. Through the processing of a complex-valued linear layer, this invention extends... With the length feature number M, an interpretable output was obtained. Specifically, considering The size is 1×N×M, and its physical meaning is the N DOA estimation results under a certain mode. This design not only emphasizes the network's multiple predictions of the direction of arrival, but also provides a more specific and practical interpretation of the output results.

[0134] Secondly, through complex-valued convolution processing, this invention... The number of feature channels is expanded. Specifically, this invention calculates this using the following mathematical expression.

[0135]

[0136]

[0137]

[0138] Wherein, W1 is a complex convolution kernel of size K×3×3, and through complex convolution, a kernel of size K×N×M is obtained. In this process, the input is processed by using K independent complex convolution kernels. K DOA estimation patterns were obtained, which serve as the physical meaning of dimension three. Complex convolution increases the feature dimension of the data and provides intuitive insights into the internal mechanisms of the network. Finally, this invention... The modulo value is used to obtain the final real number output. For subsequent processing.

[0139] (2) Inception Residual Super-Resolution Module: In this module, the input real-valued signal is processed by stacking 16 Inception Residual Blocks. To improve the resolution of angle estimation, the structure is as follows Figure 1The blue portion is shown in the diagram. Regarding parameter selection, this invention sets the kernel size parameter *n* in each Inception Residual Block to 1 and specifies the channel ratio for each path's output as 2:3:2:1. This is because data sparsity is not significant in the early stages of network feature extraction; this invention aims to use smaller kernels and increase the proportion of 3×3 kernels to achieve more detailed feature extraction. Different paths in the Inception module provide the network with a rich receptive field, enabling more comprehensive feature extraction. This invention also introduces residual connections, which deepen the network without degrading performance. Adding BatchNorm accelerates convergence while mitigating the gradient vanishing problem that may exist in deep networks. Zero-padding is used to ensure that the feature map size remains unchanged. Specifically, for each Inception Residual Block, this invention calculates using the following mathematical expression.

[0140]

[0141] Where F represents the residual function in the network, W L This represents the weight parameters of the l-th heuristic residual block.

[0142] Final output With input Having the same K×N×M dimensions, we obtained K different estimation models and N DOA estimation results. Figure 2 The output can be observed in the middle. Compared to input As the area becomes sparser and more concentrated, the resolution of angle estimation is effectively improved.

[0143] (3) Fusion Module: In this module, the present invention aims to fuse the input signal in dimensions two and three, that is, to fuse the different modes mentioned above and the results of each estimation, ultimately obtaining a 1×M probability density output. The structure of the Fusion Module is as follows: Figure 2The red portion in the diagram shows the number of input and output channels for each layer, indicated in parentheses. The framework is similar to the traditional GoogLeNet [reference], but the Max Pooling layer concatenated with the Inception module is removed. A 5×5 convolutional kernel is used to reduce the length and width of the data and increase the number of feature channels. It's worth noting that, compared to typical neural network DOA estimations, which mostly use fixed-size convolutional kernels without considering the distribution variations of the data within the network, this invention, considering the sparsity and vertical distribution characteristics of the data, increases the convolutional kernel size parameter n in the Inception module to 3 at this stage. Using a larger convolutional kernel effectively performs comprehensive data fusion and feature extraction. Furthermore, this invention sets the channel ratio of each path output in the Inception module to 4:1:4:1, reducing the proportion of 9×9 convolutional kernels, which can reduce unnecessary computational overhead to some extent. Finally, a linear layer is used to map the output of IRMAEN. This invention sets the number of discrete points in the output to 256, achieving a balance between the accuracy of angle estimation and computational efficiency.

[0144] As an optional implementation, the loss function and training process of this invention are as follows:

[0145] IRMAEN's DOA estimation can be viewed as a multi-class classification problem concerning the number of discrete output points. Therefore, this invention uses the cross-entropy loss function to evaluate the training of the network.

[0146]

[0147] p(n) = Softmax(out(n))

[0148] Where out() is the output of the Fusion Module, This is the truth expression of the neural network, where N is the number of discrete output points (256). By minimizing the cross-entropy, we can obtain an approximate value of the target angle probability distribution, and thus estimate the DOA.

[0149] In the network training part, this invention randomly and uniformly generates training data with a signal-to-noise ratio (SNR) ranging from -1 to 13 dB, a snapshot count (K) ranging from 4 to 32, and a true angle (θ) ranging from -30° to 30°. Weights are randomly initialized, and an Adam optimizer with exponential decay rates β1 and β2 of 0.9 and 0.999, respectively, and a batch size of 128 is used. This invention trains the IRMAEN model for 500 epochs with an initial learning rate of 1e-4, and the learning rate is reduced by 60% every 50 epochs.

[0150] Key points of this invention:

[0151] 1. Unlike traditional algorithms such as MUSIC, Capon, and DML, this invention employs an end-to-end deep learning approach, learning from the raw input data to the final output without requiring manual intervention or parameter adjustments. A more illustrative analogy for deep learning might be learning to play a musical instrument. Traditional machine learning methods are like learning the techniques of various instruments, requiring mastery of fingering, rhythm, and music theory. Deep learning, on the other hand, is like learning to perceive musical inspiration and express emotions; it doesn't require prior knowledge of specific instrument playing techniques. Through imitation and training, it gradually learns to play beautiful music. By observing a large number of musical works and performance videos, deep learning models automatically learn playing techniques and styles, ultimately enabling them to create unique musical works, showcasing expressiveness and creativity unattainable by humans.

[0152] 2. Furthermore, unlike most convolutional neural networks, this invention introduces and modifies the Inception module. This design allows the network to increase its receptive field while maintaining computational efficiency, thus improving its representational capabilities. Ordinary convolutional networks suffer from problems such as excessive parameters, high computational cost, and insufficient receptive field. The Inception module, by using convolutional kernels of different sizes and pooling operations, and concatenating their outputs, effectively solves these problems by reducing the number of parameters, lowering computational cost, and increasing the receptive field, thereby improving the network's representational capabilities and computational efficiency. To illustrate this, an ordinary convolutional network is like a child who can only see a partial scene; while it can observe some details, it cannot fully understand the entire scene. The Inception module, on the other hand, is like a multi-headed monster with different perspectives simultaneously. It can view the problem from different angles and synthesize these perspectives to form a more comprehensive understanding of the entire scene. This allows for more accurate judgments and decisions even in complex scenarios.

[0153] 3. It reduces the difficulty of selecting hyperparameters for the Inception multi-branch convolution module.

[0154] 4. The complex number processing module used can make full use of the characteristics of the received complex number data.

[0155] Compared with existing technologies, the method of this invention has lower requirements for the number of array elements and snapshots in the radar system, can achieve more accurate DOA estimation at the same signal-to-noise ratio, and has better estimation performance than traditional methods under low signal-to-noise ratio and low snapshot number conditions, thus better meeting the needs of practical applications.

[0156] On the other hand, embodiments of the present invention provide a DOA estimation complex value device based on multi-branch convolution with variable kernel. The device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned method.

[0157] The processor and memory can be connected via a bus or other means. Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0158] The non-transient software program and instructions required to implement the DOA estimation complex value method based on variable kernel multi-branch convolution in the above embodiments are stored in memory. When executed by the processor, the DOA estimation complex value method based on variable kernel multi-branch convolution in the above embodiments is executed.

[0159] On the other hand, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0160] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0161] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for estimating complex values ​​of DOA based on multi-branch convolution with variable kernel, characterized in that, The method includes the following steps: The real and imaginary parts of the signal are calculated by performing calculations on the complex-valued signal. Based on the real and imaginary part information, the signal is extended to obtain first data, including: Based on the real and imaginary part information, the feature channels of the signal are expanded using a complex-valued convolution kernel to obtain first data, including: An intermediate value is obtained by applying multiple independent complex convolution kernels to the real and imaginary parts of the information; the intermediate value includes the real part and the imaginary part of the intermediate value. The modulo value of the intermediate value is taken to expand the characteristic channels of the signal and obtain the first data; The formula for calculating the intermediate value by using multiple independent complex convolution kernels to process the real and imaginary parts of the information is as follows: in, Let be the real part of the intermediate value. W1 is the imaginary part of the intermediate value; W1 is the complex convolution kernel with a size of K×3×3; K is the number of complex convolution kernels; The real and imaginary parts are the information; The process of taking the modulo value of the intermediate value to expand the feature channels of the signal yields the following calculation formula for the first data: in, This refers to the first data; The second data is obtained by calculating the first data using the heuristic residual super-resolution module, including: A heuristic residual super-resolution module is constructed using several heuristic residual blocks, including: A number of revelation residual blocks are set up; each of the revelation residual blocks contains multiple paths; Set the kernel size of the heuristic residual block; Set the proportion of the number of channels in the path output; Introduce residual connections within each of the aforementioned revelation residual blocks; Add a batch normalization layer to each of the aforementioned heuristic residual blocks; By stacking several of the revelation residual blocks processed by the above steps, the revelation residual super-resolution module is established. The second data is obtained by calculating the first data using the aforementioned residual super-resolution module. The second data is fused to obtain the DOA estimation result.

2. The method according to claim 1, characterized in that, The formulas used to calculate the real and imaginary parts of the complex-valued signal include: in, W is the normalized matrix of the complex-valued signal with size 1×N×N, and M is the complex weight parameter matrix with size N×M, where M represents the number of discrete points in the probability distribution of the final output of the neural network. represents the real and imaginary parts of the signal; j is the imaginary number; N is the number of discrete points in the output.

3. The method according to claim 1, characterized in that, The calculation of the complex-valued signal to obtain the real and imaginary parts of the signal includes the following steps: Establish a complex signal processing module; According to the complex signal processing module, the input complex signal is extended by length feature number through a complex-valued linear layer to obtain the real and imaginary parts of the signal.

4. The method according to claim 1, characterized in that, The second data is obtained by calculating the first data using the heuristic residual super-resolution module, and the formula used includes: Where F represents the residual function in the network, W L This represents the weight parameters of the Lth heuristic residual block; This is the second data; the second data With the first data They have the same size.

5. The method according to claim 1, characterized in that, The process of fusing the second data to obtain the DOA estimation result includes the following steps: Retrieve the second data in both dimension two and dimension three; The second data is convolved using a convolution kernel to obtain convolutional data. The convolutional data are fused to obtain the DOA estimation result.

6. The method according to claim 1, characterized in that, The DOA estimation method based on variable kernel multi-branch convolution uses the following formulas: p(n) = Softmax(out(n)); Where out() is the output of the fusion module. N represents the neural network's representation data, where N is the number of discrete points in the output; CrossEntropyLoss is the cross-entropy loss function.

Images