Memory-saving information source positioning method based on distributed deep learning

Abstract

The invention provides a memory-saving information source positioning method based on distributed deep learning. According to the method, the memory requirement of equipment calculation in a large-scale antenna scene is reduced. An overlapped sub-array selection strategy is adopted, a complete array is divided into a plurality of sub-arrays, and part of array elements are allowed to be overlapped between adjacent sub-arrays. The strategy effectively compensates for the loss of cross-correlation information between the subarrays, thereby improving the estimation precision of the direction of arrival. Each sub-array is paired with an independent sub-processor and is responsible for compressing received data and transmitting a result to a fusion center. And the fusion center effectively extracts direction-of-arrival estimation information from the complex data set by using a graph neural network. Direction of arrival estimation is regarded as a regression task, and high-precision meshless estimation is realized through post-processing by using Toeplitz prior. Simulation results show that the provided distributed deep learning method achieves estimation accuracy equivalent to that of a centralized deep learning method, and meanwhile, the distributed deep learning method has low equipment memory requirements.

Description

Technical Field

[0001] The present invention relates to a memory-saving information source localization method based on distributed deep learning, and belongs to the technical field of antenna array signal processing. Background Art

[0002] Source localization is an important research area in array signal processing and has been widely used in wireless communications, radar, mobile phone arrays and other fields. It uses an array composed of multiple sensors or antennas to receive the source signal and then find the location of the source. When the source is far away from the array or the array is not too large, when the wave is incident on the array, it can be regarded as a plane wave. Therefore, source localization is called arrival direction estimation.

[0003] With the advancement of technology, various applications such as mobile communications have increasingly stringent requirements for high-precision positioning. For example, the integration of communication and perception has attracted more and more attention from academia and the wireless industry. In order to achieve ultra-high precision positioning, researchers have proposed the use of ultra-large-scale antenna arrays consisting of hundreds or even thousands of antenna elements. However, the complexity of ultra-large-scale antenna arrays has become very obvious, which has brought a huge computational burden to a single processor. Therefore, in order to avoid centralized processing of received signals, distributed direction of arrival estimation algorithms have received more attention. However, these methods exhibit high computational complexity, resulting in a large amount of resource consumption during processing. Deep learning methods have low computational complexity due to their data-driven characteristics, which can make up for the shortcomings of distributed direction of arrival estimation algorithms. However, due to their high degree of model parameterization, they need to rely on a large amount of training data. As the size of the array increases, the dimension of the input data also increases. The massive training data composed of large-dimensional matrices will bring huge memory pressure to the device. The algorithm's demand for memory is a great challenge in ultra-large-scale antenna scenarios. Therefore, distributed deep learning is a feasible direction, which can not only significantly reduce the memory requirements of sub-processors, but also meet the needs of real-time estimation with its low complexity. Summary of the invention

[0004] In the present invention, an overlapping subarray selection strategy is adopted to divide the complete array into multiple subarrays, allowing some array elements to overlap between adjacent subarrays. This strategy effectively compensates for the loss of mutual correlation information between subarrays, thereby improving the accuracy of wave direction estimation. Each subarray is paired with an independent subprocessor, which is responsible for compressing the received data and transmitting the results to the fusion center. The fusion center uses graph neural networks to effectively extract angle estimation information from complex data sets. The angle estimation is regarded as a regression task, and the Toeplitz prior is used to achieve high-precision gridless angle estimation through post-processing. The wave direction estimation accuracy comparable to the centralized deep learning method is achieved with low complexity, and the device has low memory requirements.

[0005] The technical solution of the present invention is:

[0006] A memory-saving source localization method based on distributed deep learning, characterized in that it comprises the following steps:

[0007] S1. Construct a receiving signal model of a uniform linear array, adopt an overlapping subarray selection strategy, and divide the complete array into D subarrays;

[0008] S2. Build a distributed covariance reconstruction network D-CRN, which consists of a sub-processor and a fusion center. The sub-processor includes a pre-processing module and a residual compression neural network RCNN module. The fusion center includes a distributed graph neural network D-GraphSAGE module and a post-processing module.

[0009] S3. Construct a training set for D-CRN, which consists of input data and labels. Each subarray corresponds to a subprocessor, which first calculates the covariance matrix R of the subarray received signal through a preprocessing module. d , and use it as the input data of D-CRN; calculate the noise-free covariance matrix of the complete array and extract its first row element u as the label of the D-CRN training set;

[0010] S4. Use the training set to train D-CRN and use the RCNN module to train R d Compress to get vector c d ; c d As the input of the fusion center, the output is obtained through the D-GraphSAGE module Will The output θ, the desired direction of arrival estimate, is given as input to the post-processing module.

[0011] Furthermore, in step S1, it is necessary to first establish a receiving signal model of a uniform linear array, specifically,

[0012] S11. The array used by the receiving end is a uniform linear array with M array elements. The spacing between adjacent array elements is d=λ / 2, where λ is the wavelength of the signal. The reference array element is the center array element of the array, and its coordinates are set to (0,0).

[0013] S12, suppose there are K narrow-band far-field sources From θ={θ1,θ2,…,θ K} is incident on a uniform linear array consisting of M array elements. The complete array is evenly divided into D sub-arrays, each with M p The signal received by the dth sub-array at time t can be expressed as follows:

[0014]

[0015] where n d (t) represents Gaussian white noise, ∑(·) represents the cumulative summation symbol, A d (θ)=[a d (θ1),a d (θ2),…,a d (θ K )],a d (θ k ) represents the steering vector of the kth source, which is specifically expressed as follows:

[0016]

[0017] Where e represents the natural index, π is the ratio of a circle to its circumference.

[0018] Furthermore, in step S2, it is necessary to build a D-CRN consisting of a sub-processor and a fusion center, specifically,

[0019] S21, build d sub-processors, each of which consists of a preprocessing module and RCNN;

[0020] S211, the preprocessing module is responsible for calculating the covariance matrix R of the d-th subarray received signal d , specifically expressed as follows,

[0021]

[0022] Where E[·] represents the expected calculation, R sd represents the source covariance matrix received by the d-th subarray, σ 2 represents the noise power, I represents the unit matrix;

[0023] Split the real and imaginary parts into two dimensions, so R d The dimension is (M p ×M p ×2), and use it as the input of the RCNN module;

[0024] S212, RCNN module consists of a residual network and a fully connected layer, where the residual network is integrated by P ResBlocks; each convolutional layer in ResBlockp contains k p convolution kernel, the output of ResBlockp is expressed as,

[0025]

[0026] where Y p-1 Represents the output of ResBlockp-1, Represents the input Y of the network and ResBlockp p-1 The mapping relationship between represents a skip connection; the output Y of the residual network P Through the fully connected layer, the dimension is (1×M p )’s compressed vector;

[0027] S22, build a fusion center, which consists of a D-GraphSAGE module and a post-processing module;

[0028] S221, the output of the sub-processor c d As nodes, fusion centers are viewed as an undirected graph in Represents the node matrix; adjacency matrix Represents the connection relationship between nodes. If nodes i and j are connected by an edge, then ε i,j =1, otherwise 0; the update process of the GraphSAGE module is as follows,

[0029]

[0030] in represents the set of neighbor nodes, Representative Node The local features of all neighbor nodes at layer l, and Represents the feature information of the node at the l-1 layer and the updated l layer respectively, AGGREGATE represents the aggregation operation using the average method, CONCAT represents the feature concatenation operation, and W l is the weight matrix of the lth layer, g(·) is the activation function;

[0031] S222, the post-processing module uses the Toeplitz structure to restore u to a noise-free covariance matrix, and uses the classic algorithm Root-MUSIC to obtain an accurate direction of arrival estimate θ.

[0032] Further, in step S3, the subarray received signal is preprocessed to generate R d , as the input data of the D-CRN training set. u is the label of the D-CRN training set, and its generation process is as follows,

[0033]

[0034] Where R represents the covariance matrix of the complete array received signal, R s represents the covariance matrix of the source s(t), T(u) represents the Toeplitz matrix, and u is its first row element.

[0035] Furthermore, in step S4, D-CRN is trained using the MSE loss function as shown below:

[0036]

[0037] in represents the actual output of the kth source in the ith training set, u (k,i) Represents the label of the k-th source in the i-th training set.

[0038] The beneficial effects of the present invention are: the present invention can significantly reduce the memory usage of the device when processing large data sets. By introducing the concept of overlapping subarrays, the network's deficiencies in information utilization are made up, and GraphSAGE is used as the fusion center to improve the accuracy of the direction of arrival estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flowchart of a memory-saving source localization method based on distributed deep learning according to an embodiment of the present invention;

[0040] Figure 2 is a receiving signal model of a uniform linear array and a subarray selection strategy in an embodiment;

[0041] Figure 3 2 is a schematic diagram of the covariance matrix of the array received signal under the subarray selection strategy in the embodiment;

[0042] Figure 4 is an architecture diagram of a D-CRN in an embodiment;

[0043] Figure 5 : is the architecture diagram of RCNN in the embodiment;

[0044] Figure 6 : is a comparison of the direction of arrival estimation performance of the method proposed in the embodiment with that of the existing method; wherein (a) is a comparison of the direction of arrival estimation performance of the method proposed in the embodiment with that of the existing method under different signal-to-noise ratios, and (b) is a comparison of the direction of arrival estimation performance of the method proposed in the embodiment with that of the existing method under different snapshots;

[0045] Figure 7 This is a comparison of the minimum memory required by the distributed deep learning method proposed in the embodiment and the traditional centralized deep learning method. DETAILED DESCRIPTION

[0046] The preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0047] Example

[0048] like Figure 1-5 As shown, the present invention proposes a memory-saving source localization method based on distributed deep learning, comprising the following steps:

[0049] S1. Construct a receiving signal model of a uniform linear array, adopt an overlapping subarray selection strategy, and divide the complete array into D subarrays;

[0050] S2. Build a distributed covariance reconstruction network D-CRN, which consists of a sub-processor and a fusion center. The sub-processor includes a pre-processing module and a residual compression neural network RCNN module. The fusion center includes a distributed graph neural network D-GraphSAGE module and a post-processing module.

[0051] S3. Construct a training set for D-CRN, which consists of input data and labels. Each subarray corresponds to a subprocessor, which first calculates the covariance matrix R of the subarray received signal through a preprocessing module. d , and use it as the input data of D-CRN; calculate the noise-free covariance matrix of the complete array and extract its first row element u as the label of the D-CRN training set;

[0052] S4. Use the training set to train D-CRN and use the RCNN module to train R d Compress to get vector c d ; c d As the input of the fusion center, the output is obtained through the D-GraphSAGE module Will The output θ, the desired direction of arrival estimate, is given as input to the post-processing module.

[0053] In this embodiment, in step S1, it is necessary to first establish a receiving signal model of a uniform linear array, specifically,

[0054] S11. The array used by the receiving end is a uniform linear array with M array elements. The spacing between adjacent array elements is d=λ / 2, where λ is the wavelength of the signal. The reference array element is the center array element of the array, and its coordinates are set to (0,0).

[0055] S12, suppose there are K narrow-band far-field sources From θ={θ1,θ2,…,θ K} is incident on a uniform linear array consisting of M array elements; the complete array is divided into D sub-arrays, each with M p array elements; the signal received by the dth sub-array at time t is expressed as follows,

[0056]

[0057] where n d (t) represents Gaussian white noise, ∑(·) represents the cumulative summation symbol, A d (θ)=[a d (θ1),ad (θ2),…,a d (θ K )],a d (θ k ) represents the steering vector of the kth source, which is specifically expressed as follows:

[0058]

[0059] Where e represents the natural index, π is the ratio of a circle to its circumference.

[0060] In this embodiment, in step S2, it is necessary to build a D-CRN consisting of a sub-processor and a fusion center, specifically,

[0061] S21, build d sub-processors, each of which consists of a preprocessing module and RCNN;

[0062] S211, the preprocessing module is responsible for calculating the covariance matrix R of the d-th subarray received signal d , specifically expressed as follows,

[0063]

[0064] Where E[·] represents the expected calculation, R sd represents the source covariance matrix received by the d-th subarray, σ 2 represents the noise power, I represents the unit matrix;

[0065] Split the real and imaginary parts into two dimensions, so R d The dimension is (M p ×M p ×2), and use it as the input of the RCNN module;

[0066] S212, RCNN module consists of a residual network and a fully connected layer, where the residual network is integrated by P ResBlocks; each convolutional layer in ResBlockp contains k p convolution kernel, the output of ResBlockp is expressed as,

[0067]

[0068] where Y p-1 Represents the output of ResBlockp-1, Represents the input Y of the network and ResBlockp p-1 The mapping relationship between represents a skip connection; the output Y of the residual network P Through the fully connected layer, the dimension is (1×M p )’s compressed vector;

[0069] S22, build a fusion center, which consists of a D-GraphSAGE module and a post-processing module;

[0070] S221, the output of the sub-processor c d As nodes, fusion centers are viewed as an undirected graph in Represents the node matrix; adjacency matrix Represents the connection relationship between nodes. If nodes i and j are connected by an edge, then ε i,j =1, otherwise 0; the update process of the GraphSAGE module is as follows,

[0071]

[0072] in represents the set of neighbor nodes, Representative Node The local features of all neighbor nodes at layer l, and Represents the feature information of the node at the l-1 layer and the updated l layer respectively, AGGREGATE represents the aggregation operation using the average method, CONCAT represents the feature concatenation operation, and W l is the weight matrix of the lth layer, g(·) is the activation function;

[0073] S222, the post-processing module uses the Toeplitz structure to restore u to a noise-free covariance matrix, and uses the classic algorithm Root-MUSIC to obtain an accurate direction of arrival estimate θ.

[0074] In this embodiment, in step S3, the subarray received signal is preprocessed to generate R d , as the input data of the D-CRN training set. u is the label of the D-CRN training set, and its generation process is as follows,

[0075]

[0076] Where R represents the covariance matrix of the complete array received signal, R s represents the covariance matrix of the source s(t), T(u) represents the Toeplitz matrix, and u is its first row element.

[0077] In this embodiment, in step S4, the D-CRN is trained using the MSE loss function as shown below:

[0078]

[0079] in represents the actual output of the kth source in the ith training set, u (k,i) Represents the label of the k-th source in the i-th training set.

[0080] This embodiment is verified by simulation experiments as follows:

[0081] Simulation Example 1: Assume that there are two far-field signal sources incident on a 40-element uniform linear array, and the number of received snapshots is set to 200. The complete array is divided into 3 sub-arrays, each containing 20 elements, and the number of overlapping elements of adjacent sub-arrays is 10. The angle of the far-field signal source is [-15.6, 2.3], and the signal-to-noise ratio range is set from -10dB to 14dB, with an interval of 2dB. The performance comparison of the embodiment method and the existing method under the uniform linear array is shown in the figure. Figure 6 As shown, the memory required by the embodiment method and the existing centralized deep learning method is compared as shown in Figure 7 As shown, it can be seen from the figure that the embodiment method has better performance than the existing method.

[0082] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A memory-saving source localization method based on distributed deep learning, characterized in that: The following steps are included: S1. Construct a receiving signal model of a uniform linear array, adopt an overlapping subarray selection strategy, and divide the complete array into D subarrays; S2. Build a distributed covariance reconstruction network D-CRN, which consists of a sub-processor and a fusion center. The sub-processor includes a pre-processing module and a residual compression neural network RCNN module. The fusion center includes a distributed graph neural network D-GraphSAGE module and a post-processing module. S3. Construct a training set for D-CRN, which consists of input data and labels. Each subarray corresponds to a subprocessor, which first calculates the covariance matrix R of the subarray received signal through a preprocessing module. d , and use it as the input data of D-CRN; calculate the noise-free covariance matrix of the complete array and extract its first row element u as the label of the D-CRN training set; S4. Use the training set to train D-CRN and use the RCNN module to train R d Compress to get vector c d ; c d As the input of the fusion center, the output is obtained through the D-GraphSAGE module Will The output θ, the desired direction of arrival estimate, is given as input to the post-processing module.

2. The memory-saving source localization method based on distributed deep learning according to claim 1, characterized in that: In step S1, a receiving signal model of a uniform linear array is constructed, specifically, S11. The array used by the receiving end is a uniform linear array with M array elements. The spacing between adjacent array elements is d=λ / 2, where λ is the wavelength of the signal. The reference array element is the center array element of the array, and its coordinates are set to (0,0). S12, suppose there are K narrow-band far-field sources From θ={θ1,θ2,…,θ K } is incident on a uniform linear array consisting of M array elements; the complete array is divided into D sub-arrays, each with M p array elements; the signal received by the dth sub-array at time t is expressed as follows, Where n d (t) represents Gaussian white noise, ∑(·) represents the cumulative summation symbol, A d (θ)=[a d (θ1),a d (θ2),…,a d (θ K )],a d (θ k ) represents the steering vector of the kth source, which is specifically expressed as follows: Where e represents the natural index, π is the ratio of a circle to its circumference.

3. The memory-saving source localization method based on distributed deep learning according to claim 2, characterized in that: In step S2, a D-CRN consisting of a sub-processor and a fusion center is constructed, specifically, S21, build d sub-processors, each of which consists of a preprocessing module and RCNN; S211, the preprocessing module is responsible for calculating the covariance matrix R of the d-th subarray received signal d , specifically expressed as follows, Where E[·] represents the expected calculation, R sd represents the source covariance matrix received by the d-th subarray, σ 2 represents the noise power, I represents the unit matrix; Split the real and imaginary parts into two dimensions, so R d The dimension is (M p ×M p ×2), and use it as the input of the RCNN module; S212, RCNN module consists of a residual network and a fully connected layer, where the residual network is integrated by P ResBlocks; each convolutional layer in ResBlockp contains k p convolution kernel, the output of ResBlockp is expressed as, where Y p-1 Represents the output of ResBlockp-1, Represents the input Y of the network and ResBlockp p-1 The mapping relationship between represents a skip connection; the output Y of the residual network P Through the fully connected layer, the dimension is (1×M p )’s compressed vector; S22, build a fusion center, which consists of a D-GraphSAGE module and a post-processing module; S221, the output of the sub-processor c d As nodes, fusion centers are viewed as an undirected graph in Represents the node matrix; adjacency matrix Represents the connection relationship between nodes. If nodes i and j are connected by an edge, then ε i,j =1, otherwise 0; the update process of the GraphSAGE module is as follows, in represents the set of neighbor nodes, Representative Node The local features of all neighbor nodes at layer l, and Represents the feature information of the node at the l-1 layer and the updated l layer respectively, AGGREGATE represents the aggregation operation using the average method, CONCAT represents the feature concatenation operation, and W l is the weight matrix of the lth layer, g(·) is the activation function; S222, the post-processing module uses the Toeplitz structure to restore u to a noise-free covariance matrix, and uses the classic algorithm Root-MUSIC to obtain an accurate direction of arrival estimate θ.

4. The memory-saving source localization method based on distributed deep learning according to claim 3, characterized in that: In step S3, the subarray received signal is preprocessed to generate R d , as the input data of the D-CRN training set; u is the label of the D-CRN training set, and its generation process is as follows, Where R represents the covariance matrix of the complete array received signal, R s represents the covariance matrix of the source s(t), T(u) represents the Toeplitz matrix, and u is its first row element.

5. The memory-saving source localization method based on distributed deep learning according to claim 4, characterized in that: In step S4, D-CRN is trained using the MSE loss function as shown below: in represents the actual output of the kth source in the ith training set, u (k,i) Represents the label of the k-th source in the i-th training set.

Images