Multi-branch convolution DOA estimation complex value method based on variable kernel

Through the multi-branch convolution method based on variable cores, the accuracy and cost problems of traditional DOA algorithm are solved, and efficient DOA estimation is achieved, especially in low signal-to-noise ratio and few quick shots. Better performance under conditions.

CN118962578AActive Publication Date: 2024-11-15SUN YAT SEN UNIV

Patent Information

Application Number
CN202411023843.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-15
Estimated Expiration
2044-07-29

AI Technical Summary

Technical Problem

Traditional DOA algorithms need to increase array element and snapshot number when improving angle estimation accuracy, resulting in high cost and calculation costs, and poor angle estimation capabilities in low signal-to-noise ratio environments. Ordinary convolutional networks have large parameters and calculation amounts. The problem of insufficient receptive wildness in Yamato.

Method used

Using a multi-branch convolution method based on variable cores, DOA estimation is realized by calculating, expanding, residual super-resolution processing and fusion of real and imaginary information on complex values, using Inception Residual Block and complex convolution to improve estimation accuracy.

Benefits of technology

Achieve higher DOA estimation accuracy under the same signal-to-noise ratio conditions, reducing the requirements of array element and snapshot count, especially in the conditions of low signal-to-noise ratio and fewer snapshot count, and perform better than traditional methods, improving estimation performance .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118962578A_ABST
    Figure CN118962578A_ABST
Patent Text Reader

Abstract

The embodiment of the invention provides a DOA estimation complex value method based on variable kernel multi-branch convolution, and belongs to the technical field of signal processing. According to the scheme, a complex value signal is calculated to obtain real part and imaginary part information of the signal; expanding the signal according to the real part and imaginary part information to obtain first data; calculating the first data through an enlightenment residual super-resolution module to obtain second data; and performing fusion processing on the second data to obtain a DOA estimation result. Within the application range of the DOA method, the precision of the method is basically superior to that of a traditional algorithm under the condition of the same signal-to-noise ratio.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of signal processing, and in particular to a DOA estimation complex value method based on multi-branch convolution with a variable kernel. Background Art

[0002] In the existing DOA application field, traditional DOA algorithms need to increase the number of radar elements and snapshots to achieve higher angle estimation accuracy, which results in high usage and computational costs, and traditional technologies have poor angle estimation capabilities when the environmental signal-to-noise ratio is low. At the same time, using deep learning methods to solve DOA has become a new hot research direction in recent years, but currently, most of them use ordinary convolutional networks for deep learning, which have problems such as too many parameters, large computational complexity, and insufficient receptive field. Summary of the invention

[0003] The main purpose of the embodiments of the present application is to provide a complex-valued DOA estimation method based on multi-branch convolution with a variable kernel.

[0004] On the one hand, an embodiment of the present invention provides a DOA estimation complex value method based on multi-branch convolution with a variable kernel, the method comprising the following steps:

[0005] Calculate the complex-valued signal to obtain the real and imaginary part information of the signal;

[0006] Expanding the signal according to the real part and the imaginary part information to obtain first data;

[0007] Calculating the first data through a residual super-resolution module to obtain second data;

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

[0009] Furthermore, the complex-valued signal is calculated to obtain the real and imaginary information of the signal, and the formula used includes:

[0010]

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

[0012] Furthermore, the step of calculating the complex-valued signal to obtain the real part and imaginary part information of the signal comprises the following steps:

[0013] Establish a complex signal processing module;

[0014] According to the complex signal processing module, the length characteristic number of the input complex signal is expanded through the complex linear layer to obtain the real part and imaginary part information of the signal.

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

[0016] According to the real part and imaginary part information, a complex-valued convolution kernel is used to expand the characteristic channel of the signal to obtain first data.

[0017] Further, the step of expanding the characteristic channel of the signal using a complex-valued convolution kernel according to the real part and the imaginary part information to obtain the first data comprises the following steps:

[0018] An intermediate value is obtained by using a plurality of independent complex convolution kernels to act on the real part and the imaginary part information; the intermediate value includes an intermediate value real part and an intermediate value imaginary part;

[0019] Taking a modulus value of the intermediate value, completing the expansion of the characteristic channel of the signal, and obtaining first data;

[0020] The calculation formula for obtaining the intermediate value by using multiple independent complex convolution kernels to act on the real and imaginary information is:

[0021]

[0022]

[0023] in, is the real part of the intermediate value, is the imaginary part of the intermediate value; W1 is the complex-valued convolution kernel of size K×3×3; K is the number of the complex convolution kernels; is the real part and imaginary part information;

[0024] The calculation formula for obtaining the first data by taking the modulus value of the intermediate value and completing the expansion of the characteristic channel of the signal is:

[0025]

[0026] in, is the first data.

[0027] Further, the step of calculating the first data by using the inspiration residual super-resolution module to obtain the second data comprises the following steps:

[0028] Using a plurality of the inspiration residual blocks to establish an inspiration residual super-resolution module;

[0029] The first data is calculated through the inspiration residual super-resolution module to obtain second data.

[0030] Further, the using of a plurality of the inspiration residual blocks to establish an inspiration residual super-resolution module comprises the following steps:

[0031] Setting a plurality of inspiration residual blocks; each inspiration residual block contains a plurality of paths;

[0032] Setting the convolution kernel size of the inspiration residual block;

[0033] Set the channel ratio of the path output;

[0034] Introducing a residual connection inside each of the inspiration residual blocks;

[0035] Adding a batch normalization layer in each of the inspiration residual blocks;

[0036] A plurality of the inspiration residual blocks processed by the above steps are stacked to complete the establishment of the inspiration residual super-resolution module.

[0037] Furthermore, the first data is calculated by the inspiration residual super-resolution module to obtain the second data, and the formula used includes:

[0038]

[0039] Among them, F represents the residual function in the network, W L represents the weight parameter of the Lth inspiration residual block;

[0040] is the second data; the second data With the first data Have the same dimensions.

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

[0042] Acquire the second data on dimension two and dimension three;

[0043] Convolving the second data using a convolution kernel to obtain convolution data;

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

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

[0046]

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

[0048] Among them, out() is the output of the fusion module, is the expression data of the neural network, N is the number of discrete points of the output; CrossEntropyLoss is the cross entropy loss function.

[0049] On the other hand, an embodiment of the present invention further provides a complex-valued DOA estimation device based on multi-branch convolution with a variable kernel, the device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the aforementioned method when executing the computer program.

[0050] On the other hand, an embodiment of 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 described above is implemented.

[0051] The embodiments of the present application include at least the following beneficial effects: The present application provides a complex-valued method for DOA estimation based on multi-branch convolution with a variable kernel. The scheme of the present invention calculates a complex-valued signal to obtain the real and imaginary information of the signal; expands the signal according to the real and imaginary information to obtain first data; calculates the first data through the inspiration residual super-resolution module to obtain second data; and fuses the second data to obtain a DOA estimation result. Within the applicable scope of the DOA method, the accuracy of the present method is substantially better than that of the traditional algorithm under the same signal-to-noise ratio conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of a DOA estimation complex value method based on multi-branch convolution with a variable kernel provided in an embodiment of the present invention.

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

[0054] Figure 3 It is a schematic diagram of the design of the Inception module provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in conjunction with the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. When the following description refers to the 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 the embodiments of the present application. They are only examples of devices and methods consistent with some aspects of the embodiments of the present application as detailed in the attached claims.

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

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

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

[0059] Before describing the embodiments of the present application in detail, some nouns and terms involved in the embodiments of the present application are first described. The nouns and terms involved in the embodiments of the present application are subject to the following explanations.

[0060] 1) DOA (Direction of Arrival) is the direction cosine. The DOA estimation method is usually 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, a neural network layer used to process complex inputs and complex weights.

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

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

[0064] 5) Inception Residual Block is a deep neural network module that combines the Inception module and 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 are further described below in conjunction with the accompanying drawings.

[0066] On the one hand, the embodiment of the present invention provides a DOA estimation complex value method based on multi-branch convolution with a variable kernel, specifically, referring to Figure 1 , the method comprises the following steps:

[0067] S100, calculating the complex-valued signal to obtain real and imaginary part information of the signal;

[0068] S200, expanding the signal according to the real part and the imaginary part information to obtain first data;

[0069] S300, calculating the first data through a residual super-resolution module to obtain second data;

[0070] S400: Perform fusion processing on the second data to obtain a DOA estimation result.

[0071] The embodiment of the present invention discloses that step S100 calculates the complex-valued signal to obtain the real part and imaginary part information of the signal, and the formula used includes:

[0072]

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

[0074] The embodiment of the present invention discloses that step S100 calculates the complex-valued signal to obtain the real part and imaginary part information of the signal, including the following steps:

[0075] S110, establishing a complex signal processing module;

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

[0077] As an optional implementation, the present invention S110 first establishes a complex signal processing module, which is intended to process the input complex signal. The design purpose of the complex signal processing module is to be able to effectively process data in the complex domain and perform the operations and conversions required for subsequent steps. This module usually includes mathematical operations, transformation functions, and appropriate data structures to ensure the correct processing and analysis of complex signals.

[0078] The present invention S120 utilizes the established complex signal processing module to process the input complex signal through the complex linear layer to achieve the expansion of the length characteristic number. This process is intended to convert or expand the input complex signal into its real and imaginary information for subsequent analysis and application. The role of the complex 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] The embodiment of the present invention discloses that step S200 expands the signal according to the real part and the imaginary part information to obtain the first data, including the following steps:

[0080] S210. Expand the characteristic channel of the signal using a complex-valued convolution kernel according to the real part and the imaginary part information to obtain first data.

[0081] The embodiment of the present invention discloses that step S210 uses a complex-valued convolution kernel to expand a characteristic channel of a signal according to real part and imaginary part information to obtain first data, including the following steps:

[0082] S211, obtaining an intermediate value by using a plurality of independent complex convolution kernels to act on real part and imaginary part information; the intermediate value includes an intermediate value real part and an intermediate value imaginary part;

[0083] S212, taking a modulus value of the intermediate value, completing the expansion of the characteristic channel of the signal, and obtaining first data;

[0084] By using multiple independent complex convolution kernels to act on the real and imaginary information, the calculation formula for the intermediate value is:

[0085]

[0086]

[0087] in, is the real part of the intermediate value, is the imaginary part of the intermediate value; W1 is a complex-valued convolution kernel of size K×3×3; K is the number of complex convolution kernels; is the real and imaginary part information;

[0088] The modulus value of the intermediate value is taken to complete the expansion of the characteristic channel of the signal, and the calculation formula for the first data is obtained as follows:

[0089]

[0090] in, is the first data.

[0091] The embodiment of the present invention discloses that step S300 calculates the first data through the inspiration residual super-resolution module to obtain the second data, including the following steps:

[0092] S310, using a plurality of inspiration residual blocks to establish an inspiration residual super-resolution module;

[0093] S320, calculating the first data through the inspiration residual super-resolution module to obtain the second data.

[0094] The embodiment of the present invention discloses that step S310 uses a plurality of inspiration residual blocks to establish an inspiration residual super-resolution module, including the following steps:

[0095] S311, setting a plurality of inspiration residual blocks; each inspiration residual block contains a plurality of paths;

[0096] S312, setting the convolution kernel size of the inspiration residual block;

[0097] S313, setting the ratio of the number of channels output by the path;

[0098] S314, introducing a residual connection inside each inspiration residual block;

[0099] S315, adding a batch normalization layer in each revelation residual block;

[0100] S316. Stack several inspiration residual blocks processed by the above steps to complete the establishment of an inspiration residual super-resolution module.

[0101] As an optional implementation, the present invention S311 first determines how many inspiration residual blocks are used to construct a super-resolution module. Each inspiration residual block refers to a specific neural network structure, which contains multiple paths or branches for processing different aspects or features of input data. These paths can achieve different feature extraction and transformation through different convolution kernels, activation functions or other hierarchical structures.

[0102] In the present invention, S312, setting the size of the convolution kernel within each revelation residual block is a critical step. The size of the convolution kernel determines the granularity and scope of the spatial feature extraction performed in a specific revelation residual block. Generally, the size of the convolution kernel can be selected based on the requirements of the problem and the size of the input data to ensure that the dimension is effectively reduced or the features are enhanced while retaining important information.

[0103] Different paths or branches in each revelation residual block S313 of the present invention usually generate feature maps of different numbers and types. In this step, setting the ratio of the number of channels output by each path can adjust the information flow and importance between different feature maps so as to more effectively integrate these features to improve the performance of the model.

[0104] The S314 residual connection of the present invention refers to adding direct jump connections between different layers of the neural network, with the purpose of reducing the loss of information when it is transmitted in the network. This connection method helps to alleviate the gradient vanishing problem and can effectively help the network learn deeper feature representations. Introducing such a connection in each inspiration residual block can effectively enhance the performance and training speed of the model.

[0105] The present invention S315 adds a batch normalization layer inside each revelation residual block, which helps to standardize the distribution of input data and accelerate the convergence process by reducing internal covariate shift.

[0106] The present invention S316 stacks the inspiration residual blocks processed and configured in the above steps in a designed order and quantity. Such a stacking process is to build a complete inspiration residual super-resolution module, which can achieve efficient super-resolution reconstruction tasks by learning complex nonlinear mappings. The stacking process of each inspiration residual block ensures that the model can effectively extract and reconstruct high-quality details and structures from the input image.

[0107] The embodiment of the present invention discloses that step S300 calculates the first data through the inspiration residual super-resolution module to obtain the second data, and the formula used includes:

[0108]

[0109] Among them, F represents the residual function in the network, W L represents the weight parameter of the Lth inspiration residual block;

[0110] is the second data; the second data With the first data Have the same dimensions.

[0111] The embodiment of the present invention discloses that step S400 performs fusion processing on the second data to obtain a DOA estimation result, including the following steps:

[0112] S410, obtaining second data on dimension two and dimension three;

[0113] S420, convolving the second data using a convolution kernel to obtain convolution data;

[0114] S430: Fusing the convolution data to obtain a DOA estimation result.

[0115] The DOA estimation complex value method based on multi-branch convolution with a variable kernel disclosed in an embodiment of the present invention uses the following formula:

[0116]

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

[0118] Among them, out() is the output of the fusion module, is the expression data of the neural network, N is the number of discrete points of the output; CrossEntropyLoss is the cross entropy loss function.

[0119] The present invention takes into account that DOA (Direction of Arrival) estimation refers to determining the direction of a signal source relative to a receiver by analyzing received signals in signal processing. This direction usually refers to the angle or direction vector of the signal source relative to the receiver, that is, 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 level, the application of DOA estimation has gradually expanded to the civilian field. In the process of scientific and technological development, with the continuous progress in sensor technology, signal processing algorithms, etc., the accuracy and performance of DOA estimation have been greatly improved, and the demand for spatial information in the civilian field has been increasing. For example, communication systems, drone navigation, intelligent transportation and other fields all need to locate and track signal sources, which has also prompted the application and development of DOA estimation technology in the civilian field.

[0120] Deep learning is mainly based on the construction and training process of neural networks. Neural networks initially simulated the connection relationship between neurons in the human brain. Through a multi-level network of neurons, they can extract features and abstractly represent input data, and finally predict and classify output data. In the early 2010s, with the emergence of deep learning models such as convolutional neural networks (CNN) and recurrent neural networks (RNN), deep learning has made breakthrough progress in image recognition, natural language processing and other fields. Compared with traditional methods, deep learning methods have the advantages of strong feature extraction capabilities, strong ability to handle complex relationships, strong generalization capabilities, end-to-end learning, strong adaptability, and a wide range of applications. They can overcome the shortcomings of traditional methods in the field of DOA and become an important tool for solving complex problems and processing large-scale data.

[0121] The embodiment of the present invention only requires a smaller number of array elements and snapshots to achieve higher estimation accuracy. Within the applicable scope of the DOA method, the accuracy of the method is substantially better than that of the traditional algorithm under the same signal-to-noise ratio, and is more obvious under the conditions of low signal-to-noise ratio and small number of snapshots.

[0122] In the direction of arrival (DOA) estimation, the received signal is usually represented in the form of complex numbers, including amplitude and phase information. In this paper, a neural network method called inception residual multisnapshot angle estimation network (IRMAEN) is proposed. Figure 2 This is the DOA estimation architecture of IRMAEN according to an embodiment of the present invention.

[0123] The main feature of this method is to receive an N×K signal matrix X consisting of K snapshots received by the array through The formed numerical matrix is ​​input into the complex neural network as a whole, instead of treating the real and imaginary parts as two channels, and is also used through the Inception module to further improve the performance of the network. The advantage of the IRMAEN network is that it improves the accuracy of multi-snapshot DOA estimation, and its accuracy and efficiency are better than traditional methods 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 the performance of DOA estimation, especially in challenging noisy environments and with a small number of snapshots.

[0124] In IRMAEN, the present invention aims to take the characteristics of network data into consideration and extract features from horizontal, vertical and local aspects. To this end, the present invention adopts convolutional layer designs with different sizes and output channel ratios in the Inception module, and simplifies the difficulty of selecting hyperparameters to a certain extent. Figure 3Shows the design of the Inception module in IRMAEN.

[0125] Among them, n is the size parameter of the convolution kernel, p is the number of input channels, q is the number of output channels, and [ω1,ω2,ω3,ω4] is the ratio of the number of channels output by each branch. The 3n×n convolution branch can effectively extract the features between different snapshot signals, the n×3n convolution branch extracts the features of a single snapshot signal, and adding 3n×3n convolution branches and parallel pooling paths can further extract the overall features of the signal. According to the distribution characteristics of the output data, as the network deepens, the data close to the top-layer will show significant sparsity and locality. Therefore, it is necessary to use convolution kernels of different sizes at different stages and adjust the ratio of the number of output channels of different paths. By combining different receptive fields, IRMAEN has more outstanding feature extraction capabilities.

[0126] In order to reduce the difficulty of selecting hyperparameters and reduce manual intervention, the present invention sets the number of output channels and branch ratios of the Inception module in each module of IRMAEN, and removes some 1×1 convolution kernels, because the present invention will stack the same modules to expand the depth of the network instead of increasing the complexity in the Inception module.

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

[0128] 1) Complex signal processing module

[0129] 2) Inception residual super-resolution module

[0130] 3) Fusion module

[0131] (1) Complex signal processing module: First, the present invention processes the input data through a complex-valued linear layer. This step uses the mathematical framework of complex numbers to extract the real and imaginary information of the signal, effectively solving the problem of loss of real and imaginary information in general methods. Specifically, the present invention calculates through the following mathematical expression.

[0132]

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

[0134] Secondly, through complex value convolution processing, the present invention Specifically, the present invention is calculated by the following mathematical expression.

[0135]

[0136]

[0137]

[0138] Among them, W1 is a complex-valued convolution kernel with a size of K×3×3, and a K×N×M size is obtained through complex-valued convolution. In this process, K independent complex convolution kernels are used to act on the input K DOA estimation modes are obtained as the physical meaning of dimension three. Through complex convolution, the characteristic dimension of data can be increased, and intuitive insights are provided for understanding the internal mechanism of the network. Take the modulo value to get the final real number output for subsequent processing.

[0139] (2) Inception Residual Super-Resolution Module: In this module, the input real signal is processed by stacking 16 Inception Residual Blocks. To improve the resolution of angle estimation, the structure is as follows Figure 1As shown in the blue part in . For parameter selection, the present invention sets the convolution kernel size parameter n in each Inception Residual Block to 1, and specifies the ratio of the number of channels output by each path to be 2:3:2:1. This is because in the early stage of network feature extraction, the sparsity of the data is not significant. The present invention hopes to use smaller convolution kernels and increase the proportion of 3×3 convolution kernels in order to extract features in detail. The different paths in the Inception module enable the network to have a rich receptive field and extract data features more comprehensively. Here, the present invention also introduces residual connections, which can deepen the network without deteriorating performance. Adding BatchNorm accelerates convergence while alleviating the possible vanishing gradient problem of deep networks. To ensure that the size of the feature map does not change, zero padding technology is used. Specifically, for each revelation residual block, the present invention calculates through the following mathematical expression.

[0140]

[0141] Among them, F represents the residual function in the network, W L Represents the weight parameter of the lth inspiration residual block.

[0142] Final Output With input With the same size of K×N×M, we get K different estimation modes and N DOA estimation results. Figure 2 In the output, we can observe Compared to input Becoming sparse and concentrated, the resolution of angle estimation is effectively improved.

[0143] (3) Fusion module: In this module, the present invention hopes to fuse the input signal in dimension two and dimension three, that is, to fuse the different modes mentioned above and the results of each estimation, and finally obtain a probability density output of 1×M. The structure of the Fusion Module is as follows Figure 2As shown in the red part of , the number of input and output channels of each layer is marked in brackets. The framework adopted is similar to the traditional GoogLeNet [reference], but the Max Pooling layer connected in series with the Inception module is removed, and a convolution kernel of size 5×5 is used to reduce the length and width of the data and increase the feature channels of the data. It is worth noting that compared with general neural network DOA estimation, most of them use convolution kernels of fixed size without considering the distribution changes of data in the network. In view of the sparsity of data and its longitudinal distribution characteristics, the present invention increases the convolution kernel size parameter n in the Inception module to 3 at this stage, uses a larger convolution kernel to effectively perform comprehensive fusion and feature extraction of data, and the present invention also sets the ratio of the number of channels output by each path in the Inception module to 4:1:4:1, reducing the ratio of 9×9 convolution kernels, which can reduce unnecessary computational overhead to a certain extent. Finally, a linear layer is used to map the output of IRMAEN. Among them, the present invention sets the number of discrete points of the output to 256, which can strike a balance between the accuracy and computational efficiency of angle estimation.

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

[0145] The DOA estimation of IRMAEN can be regarded as a multi-classification problem about the number of output discrete points, so the present invention selects the cross entropy loss function to evaluate the training of the network.

[0146]

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

[0148] Among them, out() is the output of Fusion Module, is the truth expression of the neural network, and N is the number of discrete points output, 256. By minimizing the cross entropy, an approximation of the probability distribution of the target angle can be obtained, thereby estimating the DOA.

[0149] In the training part of the network, the present invention randomly and uniformly generates training data with signal-to-noise ratios (SNR) ranging from -1 to 13 dB, snapshot numbers (K) ranging from 4 to 32, and true angles (θ) ranging from -30° to 30°. The 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. The present 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] The key points of the present invention are:

[0151] 1. Different from traditional algorithms such as MUSIC, Capon, and DML, the present invention adopts a deep learning end-to-end learning method to learn from the original input data to the final output results. The intermediate process does not require human intervention or manual adjustment of parameters. It may be more vivid to compare deep learning to the process of learning to play a musical instrument. Traditional machine learning methods are like learning the skills of playing various musical instruments, which require mastering various fingerings, rhythms, and music theories. Deep learning is like learning to perceive musical inspiration and express emotions. There is no need to understand the specific instrument playing skills in advance. Through imitation and training, you gradually learn to play beautiful music. By observing a large number of musical works and performance videos, the deep learning model automatically learns the performance skills and styles, and can eventually create unique musical works by itself, showing expressiveness and creativity that humans cannot achieve.

[0152] 2. In addition, unlike most convolutional neural networks, the present invention introduces and modifies the Inception module. This design can increase the receptive field while maintaining computational efficiency, thereby improving the network's representation ability. Ordinary convolutional networks have problems such as excessive parameter volume, large computational complexity, and insufficient receptive field. The Inception module uses convolution kernels of different sizes and pooling operations, and connects their outputs together to reduce the number of parameters, reduce the computational complexity, and increase the receptive field, effectively solving these problems and improving the network's representation ability and computational efficiency. The present invention uses a more vivid metaphor to describe: ordinary convolutional networks are like a child who can only see a local scene. Although some details around can be observed, the entire scene cannot be fully understood. The Inception module is like a multi-headed monster with different perspectives at the same time. It can look at problems from different perspectives at the same time, and combine these perspectives to form a more comprehensive understanding of the entire scene. In this way, even in the face of complex scenes, more accurate judgments and decisions can be made.

[0153] 3. Reduced 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 data.

[0155] Compared with the current existing technology, the method of the present invention has lower requirements on the number of array elements and snapshots of 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 environmental conditions of low signal-to-noise ratio and low snapshot number, better meeting the needs of practical applications.

[0156] On the other hand, an embodiment of the present invention provides a DOA estimation complex value device based on multi-branch convolution with a variable kernel, the device includes a memory and a processor, the memory stores a computer program, and the processor implements the above method when executing the computer program.

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

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

[0159] On the other hand, an embodiment of the present invention 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 above method is implemented.

[0160] It will be appreciated by those skilled in the art that all or some of the steps and systems in the disclosed method above may be implemented as software, firmware, hardware and appropriate combinations thereof. Some physical components or all physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor or a microprocessor, or may be implemented as hardware, or may be implemented as an integrated circuit, such as an application specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or a non-transitory medium) and a communication medium (or a temporary medium). As known to those skilled in the art, the term computer storage medium 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 include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, disk storage or other magnetic storage devices, or any other medium that may be used to store desired information and may be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically embodies computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0161] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention should be within the scope of the rights of the present invention.

Claims

1. A complex-valued DOA estimation method based on multi-branch convolution with variable kernel, characterized in that: The method comprises the following steps: Calculate the complex-valued signal to obtain the real and imaginary part information of the signal; Expanding the signal according to the real part and the imaginary part information to obtain first data; Calculating the first data through a residual super-resolution module to obtain second data; The second data is fused to obtain a DOA estimation result.

2. The method according to claim 1, characterized in that The complex-valued signal is calculated to obtain the real and imaginary information of the signal, and the formula used includes: in, is a normalized matrix of complex-valued signals of size 1×N×N, W is a complex weight parameter matrix of size N×M, M represents the number of discrete points of the probability distribution of the final output of the neural network, is the real and imaginary information of the signal; j is an imaginary number.

3. The method according to claim 1, characterized in that The step of calculating the complex-valued signal to obtain the real part and the imaginary part information of the signal comprises the following steps: Establish a complex signal processing module; According to the complex signal processing module, the length characteristic number of the input complex signal is expanded through the complex linear layer to obtain the real part and imaginary part information of the signal.

4. The method according to claim 1, characterized in that: The step of expanding the signal according to the real part and the imaginary part information to obtain the first data comprises the following steps: According to the real part and imaginary part information, a complex-valued convolution kernel is used to expand the characteristic channel of the signal to obtain first data.

5. The method according to claim 4, characterized in that The step of expanding the characteristic channel of the signal using a complex-valued convolution kernel according to the real part and the imaginary part information to obtain the first data comprises the following steps: An intermediate value is obtained by using a plurality of independent complex convolution kernels to act on the real part and the imaginary part information; the intermediate value includes an intermediate value real part and an intermediate value imaginary part; Taking a modulus value of the intermediate value, completing the expansion of the characteristic channel of the signal, and obtaining first data; The calculation formula for obtaining the intermediate value by using multiple independent complex convolution kernels to act on the real and imaginary information is: in, is the real part of the intermediate value, is the imaginary part of the intermediate value; W1 is the complex-valued convolution kernel of size K×3×3; K is the number of the complex convolution kernels; is the real part and imaginary part information; The calculation formula for obtaining the first data by taking the modulus value of the intermediate value and completing the expansion of the characteristic channel of the signal is: in, is the first data.

6. The method according to claim 1, characterized in that The step of calculating the first data by using the inspiration residual super-resolution module to obtain the second data comprises the following steps: Using several inspiration residual blocks to build an inspiration residual super-resolution module; The first data is calculated through the inspiration residual super-resolution module to obtain second data.

7. The method according to claim 6, characterized in that The method of using a plurality of inspiration residual blocks to establish an inspiration residual super-resolution module comprises the following steps: Setting a plurality of inspiration residual blocks; each inspiration residual block contains a plurality of paths; Setting the convolution kernel size of the inspiration residual block; Set the channel ratio of the path output; Introducing a residual connection inside each of the inspiration residual blocks; Adding a batch normalization layer in each of the inspiration residual blocks; A plurality of the inspiration residual blocks processed by the above steps are stacked to complete the establishment of the inspiration residual super-resolution module.

8. The method according to claim 1, characterized in that The first data is calculated by the inspiration residual super-resolution module to obtain the second data, and the formula used includes: Among them, F represents the residual function in the network, W L represents the weight parameter of the Lth inspiration residual block; is the second data; the second data With the first data Have the same dimensions.

9. The method according to claim 1, characterized in that: The step of fusing the second data to obtain a DOA estimation result comprises the following steps: Acquire the second data on dimension two and dimension three; Convolving the second data using a convolution kernel to obtain convolution data; The convolution data are fused to obtain the DOA estimation result.

10. The method according to claim 1, characterized in that The DOA estimation complex value method based on multi-branch convolution with variable kernel uses the following formula: p(n)=Softmax(out(n)); Among them, out() is the output of the fusion module, is the expression data of the neural network, N is the number of discrete points of the output; CrossEntropyLoss is the cross entropy loss function.

Citation Information

Patent Citations

  • Two-dimensional DOA estimation method based on convolutional attention network

    CN116840774A

  • Robust direction of arrival estimation method based on ADMM-Net

    CN117331021A

  • DOA model training method and estimation method based on neural network

    CN117951495A

  • Deep Learning-Based Multiband Signal Fusion for 3-D SAR Super-Resolution

    US20240134037A1

Cited By

  • Planar array DOA estimation method of deep learning auxiliary subspace

    CN120847712A

  • A deep learning assisted subspace plane array DOA estimation method

    CN120847712B