Multi-terminal multi-service co-network transmission access detection and link state information estimation method

Abstract

The application discloses a multi-terminal multi-service co-network transmission access detection and link state information estimation method to solve the problem of high correlation of terminal link characteristics and poor generalization in the prior art. The method introduces a graph neural network (GNN) to enhance an orthogonal approximate message passing (OAMP) algorithm, realizes fast convergence and accurate detection, significantly reduces the number of iterations and improves performance. The method comprises the following steps: a certain number of active terminals use a single handshake protocol to initiate parallel transmission to a center node, transform the link from a space-frequency domain to an angle-frequency domain, and preprocess the model; an advanced learning model based on a deep unfolding architecture is used for holographic perception of the terminal-level link state; and based on the link state panoramic information obtained in the early stage, accurate terminal activity detection is performed.

Description

Technical Field

[0001] This invention relates to the field of signal processing and data transmission technology in communication networks, specifically to a method for multi-terminal, multi-service shared network transmission access detection and link state information estimation. Background Technology

[0002] With the deepening of digitalization, the Internet of Things has become a significant trend in the information society. Under this trend, the number of various terminal devices has surged, and business types have become increasingly diversified. Multi-terminal, multi-service network transmission, as an efficient resource-sharing mode, can effectively support parallel communication of massive numbers of terminals on a unified network platform, but it still faces many challenges in practical applications. Multi-terminal, multi-service network transmission scenarios typically exhibit three main characteristics: First, the data traffic generated by massive numbers of devices exhibits significant asymmetry, with the amount of data generated by a device far exceeding its data reception; second, business activities are highly sudden and random, with terminal behavior being highly asynchronous in time; third, the transmitted data units are generally small and compact, placing extremely stringent requirements on end-to-end latency. These characteristics collectively pose significant challenges to network connection management, the flexibility of dynamic resource allocation, and the ability to handle high-concurrency requests. Traditional sequential authorization protocols that rely on scheduling are difficult to meet the practical requirements of high efficiency and low latency due to frequent handshakes and complex control signaling. Single-handshake access technologies and non-orthogonal resource-sharing technologies are widely used to reduce access latency and resource overhead.

[0003] In high-concurrency scenarios, system performance faces severe challenges. The significant correlation between inter-terminal link characteristics makes it difficult for traditional detection mechanisms to accurately distinguish individual terminal states. Simultaneously, the ultra-high-dimensional data processing demands resulting from massive terminal access lead to a sharp increase in system computational complexity. The combined effect of these two factors causes a systematic decline in detection accuracy and a significant deterioration in overall performance. Existing technical solutions have multiple limitations. While compressed sensing methods based on sparsity can utilize the distribution characteristics of terminal activity, their computational process is exceptionally complex, overly reliant on idealized model assumptions, and vulnerable in real-world environments. Deep learning technology offers a new path for performance improvement; however, purely data-driven methods suffer from interpretability deficiencies, and the model decision-making process lacks transparency. These inherent defects severely restrict the practical value of existing methods in high-concurrency scenarios. Therefore, a method for access detection and link state information estimation that balances efficiency, generalization, and scalability is urgently needed. Summary of the Invention

[0004] In view of this, the present invention provides a method for access detection and link state information estimation for multi-terminal and multi-service network transmission. This method introduces a graph neural network (GNN) to enhance the orthogonal approximate message passing (OAMP) algorithm, which significantly reduces the iteration requirements while maintaining low computational complexity and greatly improves the convergence speed to efficiently support high-concurrency scenarios. This method achieves access detection and link state information estimation that balances efficiency, generalization and scalability.

[0005] To achieve the above objectives, the technical solution of the present invention includes the following steps:

[0006] Step 1: A certain number of active terminals initiate parallel transmissions to the central node using a single handshake protocol.

[0007] Step 2: Transform the link from the spatial-frequency domain to the angle-frequency domain and preprocess the model;

[0008] Step 3: Employ an advanced learning model based on a deep unfolding architecture to achieve holographic perception of the terminal-level link status;

[0009] Step 4: Based on the comprehensive link status information obtained in the previous stage, perform precise terminal activity detection.

[0010] Further, in step 1, each terminal sends symbols carrying information through a specific communication mechanism, and each terminal is configured with a unique non-orthogonal spreading sequence as an identification feature code; the feature code is multiplied by a binary indicator representing the active state of the terminal to generate a terminal-specific signal with terminal-specific characteristics; the terminal-specific signal is distributed and carried on multiple resource block units through a resource mapping mechanism.

[0011] Furthermore, step 1 addresses scenarios with massive terminal access, where massive terminals refer to a minimum of 32 terminals and the number of ports on the central node is [missing information]. The total number of terminals within the service area is But at any given moment, only One terminal has an uplink access request. It employs a multi-channel mechanism with high parallel transmission capabilities to support simultaneous access by a massive number of devices. The multi-terminal, multi-service transmission system includes... Each parallel transmission channel; simultaneously, each potential access terminal is equipped with a unique and non-orthogonal preamble sequence. Where M is the length of the leader sequence, Let k represent the complex number field, where k is the kth terminal and its value range is 1-K; This is a binary activity indicator, set to 1 when the terminal is active and 0 when it is inactive. For the first The terminal to the central node Sub-links on each transmission channel Let be the additive white Gaussian noise on the p-th transmission channel, where p ranges from 1 to p, and the noise power is . Then the central node is... The received signal on each transmission channel is The received signals from all transmission channels are spliced ​​together to obtain a splicing matrix. Meanwhile, let the concatenation matrix of the preceding sequence The fusion matrix of link status and terminal activity status splicing matrix Additive white Gaussian noise splicing matrix ,get: .

[0012] Further, step 2, the specific process is as follows:

[0013] Step 2.1: Transform the original feature space from the initial representation domain to the new transformation domain;

[0014] Step 2.2: Convert the complex signal model into a real signal model.

[0015] Further, in step 2.1, the original feature space is transformed from the initial representation domain to a new transform domain, specifically by defining... It is a dimension of The discrete Fourier transform matrix, columns of the dictionary matrix , , Let be a dictionary matrix, the first... The initial representation domain link of each terminal is The link in the new transform domain corresponding to the k-th terminal is:

[0016]

[0017] Further After vectorization and transpose, we get:

[0018]

[0019]

[0020] in Represents the Kronecker product operation; Definition , , yes The conjugate of , we get: .

[0021] Further, in step 2.2, the complex signal model is converted into a real signal model, specifically as follows:

[0022] definition , , ,as well as ,in Indicates taking the real part, To indicate taking the imaginary part, we have: .

[0023] Further, step 3, the specific process is as follows:

[0024] First, we consider the leading sequence matrix under the real number model. Perform singular value decomposition to obtain the truncated singular value matrix. And right singular matrix , and Let the probability density function and cumulative probability distribution function of the standard Gaussian distribution be represented respectively. The data acquisition channel index set using high-precision sampling units is... The system is configured with the following number of high-precision sampling units: , express The first of the real part List;

[0025] Perform initialization: , , , , , ; use this as the initial value The next iteration; To estimate the initial value of the noise variance, and These are intermediate variables in the algorithm. for The non-zero probability of the element in the k-th row and n-th column, i.e., sparsity. No. Estimate the link variance corresponding to the column. Hidden state variables The initial value of ; c is the independent variable, which takes any real number greater than zero; Representing vectors The 2-norm, Representation matrix The F-norm;

[0026] In the In this iteration, a linear estimation is performed using the linear minimum mean square error (LMMSE) estimator.

[0027]

[0028]

[0029]

[0030]

[0031] in As an intermediate variable, Truncating the singular value matrix, It is the conjugate transpose of V; and For intermediate output, Take the (k,k)th element of the matrix; Represents the identity matrix;

[0032] Then, orthogonalization is used to reduce the correlation between the input and output:

[0033] LMMSE estimator for The error estimate of the noisy observation is

[0034]

[0035] LMMSE estimator for Noisy observations

[0036]

[0037] Next, we perform LMMSE posterior estimation.

[0038]

[0039]

[0040]

[0041] in For posterior sparsity; The posterior mean; For posterior variance;

[0042] The first in the GNN module The terminal's first The feature vector corresponding to each node is defined as follows: , No. The state vectors of each node are initialized to... , ~ This represents a fully connected network with 8 distinct and independent parameters; the eigenvectors of the edges are defined as follows: All elements in it are constants; This represents an index that is different from n, where ;

[0043] To ensure that all terminals share the same side messages, for the GNN's... In the next iteration, during the propagation phase, the messages on each edge are updated. GNN is called in each iteration. Next iteration:

[0044]

[0045] This represents the intermediate variable of the GNN in iteration t; the superscript T here indicates transpose.

[0046] During the aggregation phase, for the first The message at a given node is obtained by summing the messages on all its adjacent edges. ;

[0047] Aggregated node messages Further input into the GRU network of the gate control recurrent unit Obtain hidden state variables And update, ;

[0048] Then, based on the hidden state variables, the node's state variables are updated to complete the first step. Next iteration: ;

[0049] When GNN completes After iteration, the posterior mean optimized by GNN is obtained. ,variance and sparsity :

[0050]

[0051]

[0052]

[0053] Then, orthogonalization was performed.

[0054]

[0055] With damping coefficient To perform damping updates, The value of is a real number in the range [0, 1].

[0056]

[0057] Update sparsity and link variance to complete the first step. Round iteration;

[0058] Link variance is ;

[0059] sparsity is

[0060] go through After several iterations, the link estimation results, sparsity matrix, and posterior variance vector are obtained:

[0061] The link estimation results are as follows: ;

[0062] The sparsity matrix is: ;

[0063] The posterior variance vector is: .

[0064] Further, in step 4, terminal activity is detected using the following method:

[0065] First, the output variables of the link estimation network are averaged to obtain the mean values ​​of the link estimation results, the sparsity matrix, and the posterior variance vector:

[0066]

[0067]

[0068]

[0069] in Used to measure the effectiveness of link estimation Represents the SoftMax activation function. As trainable parameters, treating activity detection as a binary classification task, the output of the fully connected network is: ;

[0070] Selected First element Representing the The active probability of each terminal is used to determine the support set for detection based on the threshold. The support set is the set of active terminals. .

[0071] Beneficial effects:

[0072] This invention utilizes a GNN-enhanced OAMP algorithm to significantly reduce iteration requirements while maintaining low computational complexity, greatly improving convergence speed to efficiently support high-concurrency scenarios. It achieves access detection and link state information estimation that balances efficiency, generalization, and scalability. The constructed network structure possesses inherent permutation invariance, autonomously adapting to dynamic parameters such as link quality fluctuations and changes in the number of active terminals, enabling seamless migration across scenarios. By deeply mining the prior sparsity characteristics of terminal links in a multi-dimensional domain space, a composite feature extraction mechanism is constructed, achieving a significant leap in detection and estimation accuracy compared to traditional methods. Combined with a single handshake mechanism and intelligent resource mapping strategy, it effectively reduces control signaling overhead, overcomes physical interface limitations, and provides systematic support for high-concurrency access. Attached Figure Description

[0073] Figure 1 Network block diagram for estimating terminal link state information;

[0074] Figure 2 The normalized mean square error (NMSE) is used as the evaluation metric for different signal-to-noise ratios. The link state information estimation performance of this example and the comparison scheme is shown in the figure.

[0075] Figure 3 The average error detection rate (AER) is used as the evaluation metric for different signal-to-noise ratios. The comparison chart shows the terminal access status recognition performance of this example and the comparison scheme. Detailed Implementation

[0076] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0077] This embodiment considers a scenario in a multi-terminal, multi-service transmission system where a multi-port central node serves a massive number of terminals. It is worth noting that the following embodiments are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0078] Figure 1 The diagram illustrates a network block diagram for terminal link state information estimation. The multi-terminal, multi-service shared network transmission access detection and link state information estimation method provided in this embodiment specifically includes the following steps:

[0079] Step 1, considering the scenario of massive terminal access, in this embodiment of the invention, massive terminals refer to terminals with a number of not less than 32. Assume the number of ports of the central node is... The total number of terminals within the service area is But at any given moment, only Several terminals have uplink access requests. A multi-channel mechanism with high parallel transmission capabilities is adopted to support simultaneous access by a massive number of devices. The multi-terminal, multi-service transmission system includes... Each parallel transmission channel is used. Simultaneously, each potential access terminal is equipped with a unique and non-orthogonal preamble sequence. Where M is the length of the leader sequence, Let represent the complex field, where k is the k-th terminal, and its value range is 1-K. Let This is a binary activity indicator, set to 1 when the terminal is active and 0 when it is inactive. For the first The terminal to the central node Sub-links on each transmission channel Let be the additive white Gaussian noise on the p-th transmission channel, where p ranges from 1 to p, and the noise power is . Then the central node is... The received signal on each transmission channel is

[0080]

[0081] The received signals from all transmission channels are spliced ​​together to obtain a splicing matrix. Meanwhile, let the concatenation matrix of the preceding sequence The fusion matrix of link status and terminal activity status splicing matrix Additive white Gaussian noise splicing matrix ,get:

[0082]

[0083] Step 2: Feature Domain Transformation; The original feature space is transformed to a new transform domain representation, and the model is preprocessed, including the following steps:

[0084] Step 2.1, transform the original feature space from the initial representation domain to the new transformation domain: Define It is a dimension of (The P here is the same as the previous P) is the discrete Fourier transform matrix, and the columns of the dictionary matrix. , , Let be a dictionary matrix, the first... The initial representation domain link of each terminal is The link in the new transform domain corresponding to the k-th terminal is:

[0085]

[0086] Further After vectorization and transpose, we get:

[0087]

[0088]

[0089] in This represents the Kronecker product operation. Definition , , yes The conjugate of , we get:

[0090]

[0091] Step 2.2, convert the complex signal model into a real signal model: Define , , ,as well as ,

[0092] in Indicates taking the real part, To indicate taking the imaginary part, we have:

[0093]

[0094] Step 3: Link State Estimation; The OAMP algorithm is further expanded using GNN. See the attached diagram in the manual for the process.

[0095] First, we consider the leading sequence matrix under the real number model. Performing Singular Value Decomposition (SVD) yields a truncated singular value matrix. And right singular matrix ,make and Let the probability density function and cumulative probability distribution function of the standard Gaussian distribution be represented respectively. The data acquisition channel index set using high-precision sampling units is... The system is configured with the following number of high-precision sampling units: , express The first of the real part Columns. Then initialization can be performed: , , , , , Using this as the initial value... The next iteration. To estimate the initial value of the noise variance, and These are intermediate variables in the algorithm. for The non-zero probability of the element in the k-th row and n-th column, i.e., sparsity. No. Estimate the link variance corresponding to the column. Hidden state variables The initial value of ; c is the independent variable, which takes any real number greater than zero; Representing vectors The 2-norm, Representation matrix The F-norm.

[0096] In the In this iteration, a linear estimation is performed using the Linear Minimum Mean Square Error (LMMSE) estimator.

[0097]

[0098]

[0099]

[0100]

[0101] in As an intermediate variable, Truncating the singular value matrix, It is the conjugate transpose of V; and For intermediate output, Take the (k,k)th element of the matrix; Represents the identity matrix;

[0102] LMMSE estimator for The error estimate of the noisy observation is

[0103]

[0104] LMMSE estimator for Noisy observations

[0105]

[0106] Next, we perform LMMSE posterior estimation.

[0107]

[0108]

[0109]

[0110] in For posterior sparsity; The posterior mean; For posterior variance;

[0111] The first in the GNN module The terminal's first The feature vector corresponding to each node is defined as follows: , No. The state vectors of each node are initialized to... , ~ This represents a fully connected network with 8 distinct and independent parameters; the feature vectors of the edges are defined as follows: All elements in it are constants. This represents an index that is different from n, where .

[0112] To reduce computational complexity, and considering that GNNs learn generalized link features in this scenario, all terminals share the same side messages. For the GNN's... In the next iteration, during the propagation phase, the messages on each edge are updated. GNN is called in each iteration. iteration

[0113]

[0114] This represents the intermediate variable of the GNN in iteration t; the superscript T here indicates transpose.

[0115] During the aggregation phase, for the first The message at a given node is obtained by summing the messages on all its adjacent edges. .

[0116] Aggregated node messages Further fed into the gated recurrent unit (GRU) network Obtain hidden state variables And update,

[0117] Then, based on the hidden state variables, the node's state variables are updated to complete the first step. The next iteration.

[0118] ;

[0119] When GNN completes After iteration, the posterior mean optimized by GNN is obtained. ,variance sparsity :

[0120]

[0121]

[0122]

[0123] Then, orthogonalization was performed.

[0124]

[0125] With damping coefficient To perform damping updates, The value of is a real number in the range [0, 1].

[0126]

[0127] Update sparsity and link variance to complete the first step. Round iteration;

[0128] Link variance is ;

[0129] sparsity is

[0130] go through After several iterations, the link estimation results, sparsity matrix, and posterior variance vector are obtained:

[0131] The link estimation results are as follows: ;

[0132] The sparsity matrix is: ;

[0133] The posterior variance vector is: .

[0134] Step 4: Activity Detection; First, average the output variables of the link estimation network to obtain the mean values ​​of the link estimation results, sparsity matrix, and posterior variance vector:

[0135]

[0136]

[0137]

[0138] in Used to measure the effectiveness of link estimation Represents the SoftMax activation function. As trainable parameters, treating activity detection as a binary classification task, the output of the fully connected network is: ;

[0139] Selected First element Representing the The active probability of each terminal is used to determine the support set for detection based on the threshold. The support set is the set of active terminals. .

[0140] Step 3 has the results of estimating the link state information.

[0141] In the actual simulation, the total number of terminals was set to 400, the number of active terminals to 100, and the number of central node interfaces to 16. The OAMP-MMV algorithm, SOMP algorithm, EP algorithm, and LAMP algorithm were selected as comparative algorithms to verify the advantages of this invention in terminal access status identification and link state information estimation in multi-terminal concurrent access scenarios.

[0142] (1) OAMP-MMV Algorithm. The OAMP-MMV algorithm is designed for edge nodes in multi-terminal concurrent access mode, and it has excellent performance. Please refer to the following: "J. Ma and L. Ping, 'Orthogonal AMP,' in IEEE Access, vol. 5, pp. 2020-2033, 2017, doi: 10.1109 / ACCESS.2017.2653119." Comparing this invention with the OAMP-MMV algorithm illustrates the advantages of this invention in utilizing graph neural networks.

[0143] (2) SOMP Algorithm. Please refer to the following: "J.-F. Determe, J. Louveaux, L. Jacques and F. Horlin, 'On the Noise Robustness of Simultaneous Orthogonal Matching Pursuit,' in IEEE Transactions on Signal Processing, vol. 65, no. 4, pp. 864-875, 15 Feb. 1, 2017, doi: 10.1109 / TSP.2016.2626244." Compare this invention with the SOMP algorithm to illustrate the advantages of this invention utilizing graph neural networks.

[0144] (3) EP Algorithm. Please refer to the following: "J. Céspedes, PM Olmos, M. Sánchez-Fernández and F. Perez-Cruz, 'Expectation Propagation Detection for High-Order High-Dimensional MIMO Systems,' in IEEE Transactions on Communications, vol. 62, no.8, pp. 2840-2849, Aug. 2014, doi: 10.1109 / TCOMM.2014.2332349." A comparison between this invention and the EP algorithm illustrates the advantages of this invention utilizing graph neural networks.

[0145] (4) LAMP Algorithm. Please refer to the following: "M. Borgerding, P. Schniter and S. Rangan, 'AMP-Inspired Deep Networks for Sparse Linear Inverse Problems,' in IEEE Transactions on Signal Processing, vol. 65, no. 16, pp. 4293-4308, 15 Aug.15, 2017, doi: 10.1109 / TSP.2017.2708040." Compare this invention with the LAMP algorithm to illustrate the advantages of this invention utilizing graph neural networks.

[0146] Normalized Mean Squared Error (NMSE) is used to measure the accuracy of link-state information estimation for each method. NMSE is defined as the sum of the Frobenius norms of the differences between the estimated and actual link-state information, divided by the sum of the Frobenius norms of the actual link-state information. Average Error Detection Rate (AER) is used to measure the accuracy of terminal access state identification for each method. AER is defined as the sum of the absolute values ​​of the differences between the estimated and actual terminal access states, divided by the total number of terminals. Figure 2 The graph shows a comparison of the link state information estimation performance of this example and the comparison scheme under different signal-to-noise ratios. Figure 3 The diagram shows a comparison of the terminal access status recognition performance of this example and the comparison scheme under different signal-to-noise ratios.

[0147] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for multi-terminal, multi-service shared network transmission access detection and link state information estimation, characterized in that, Includes the following steps: Step 1: A certain number of active terminals initiate parallel transmission to the central node using a single handshake protocol; Step 1 is for a scenario with massive terminal access, where massive terminals refer to a number of terminals of no less than 32, and the number of ports of the central node is [missing information]. The total number of terminals within the service area is But at any given moment, only One terminal has an uplink access request. It employs a multi-channel mechanism with high parallel transmission capabilities to support simultaneous access by a massive number of devices. The multi-terminal, multi-service transmission system includes... Each parallel transmission channel; simultaneously, each potential access terminal is equipped with a unique and non-orthogonal preamble sequence. Where M is the length of the leader sequence, Let k represent the complex number field, where k is the kth terminal and its value range is 1-K; This is a binary activity indicator, set to 1 when the terminal is active and 0 when it is inactive. For the first The terminal to the central node Sub-links on each transmission channel Let be the additive white Gaussian noise on the p-th transmission channel, where p ranges from 1 to p, and the noise power is . Then the central node is... The received signal on each transmission channel is The received signals from all transmission channels are spliced ​​together to obtain a splicing matrix. Meanwhile, let the concatenation matrix of the preceding sequence The fusion matrix of link status and terminal activity status splicing matrix Additive white Gaussian noise splicing matrix ,get: ; Step 2: Transform the link from the spatial-frequency domain to the angle-frequency domain and preprocess the model; the specific process of Step 2 is as follows: Step 2.1 transforms the original feature space from the initial representation domain to a new transform domain, specifically by defining... It is a dimension of The discrete Fourier transform matrix, columns of the dictionary matrix , , Let be a dictionary matrix, the first... The initial representation domain link of each terminal is The link in the new transform domain corresponding to the k-th terminal is: Further After vectorization and transpose, we get: in Represents the Kronecker product operation; Definition , , yes The conjugate of , we get: ; Step 2.2, convert the complex signal model into a real signal model, specifically as follows: definition , , ,as well as ,in Indicates taking the real part, To indicate taking the imaginary part, we have: ; Step 3: Employ an advanced learning model based on a deep unfolding architecture to achieve holographic perception of the terminal-level link status; Step 4: Based on the comprehensive link status information obtained in the previous stage, perform precise terminal activity detection.

2. The method for multi-terminal, multi-service shared network transmission access detection and link state information estimation as described in claim 1, characterized in that, In step 1, each terminal sends symbols carrying information through a specific communication mechanism. Each terminal is configured with a unique non-orthogonal spreading sequence as an identification feature code. The identification feature code is multiplied by a binary indicator representing the active state of the terminal to generate a terminal-specific signal with terminal-specific characteristics. The terminal-specific signal is distributed and carried on multiple resource block units through a resource mapping mechanism.

3. The method for multi-terminal, multi-service shared network transmission access detection and link state information estimation as described in claim 1, characterized in that, Step 3, the specific process is as follows: First, we consider the leading sequence matrix under the real number model. Perform singular value decomposition to obtain the truncated singular value matrix. And right singular matrix , and Let the probability density function and cumulative probability distribution function of the standard Gaussian distribution be represented respectively. The data acquisition channel index set using high-precision sampling units is... The system is configured with the following number of high-precision sampling units: , express The first of the real part List; Perform initialization: , , , , , ; use this as the initial value The next iteration; To estimate the initial value of the noise variance, and These are intermediate variables in the algorithm. for The non-zero probability of the element in the k-th row and n-th column, i.e., sparsity. No. Estimate the link variance corresponding to the column. Hidden state variables The initial value of ; c is the independent variable, which takes any real number greater than zero; Representing vectors The 2-norm, Representation matrix The F-norm; In the In this iteration, a linear estimation is performed using the linear minimum mean square error (LMMSE) estimator. in As an intermediate variable, Truncating the singular value matrix, It is the conjugate transpose of V; and For intermediate output, Take the (k,k)th element of the matrix; Represents the identity matrix; Then, orthogonalization is used to reduce the correlation between the input and output: LMMSE estimator for The error estimate of the noisy observation is LMMSE estimator for Noisy observations Next, we perform LMMSE posterior estimation. in For posterior sparsity; The posterior mean; For posterior variance; The first in the GNN module The terminal's first The feature vector corresponding to each node is defined as follows: , No. The state vectors of each node are initialized to... , ~ This represents a fully connected network with 8 distinct and independent parameters; the eigenvectors of the edges are defined as follows: All elements in it are constants; This represents an index that is different from n, where ; To ensure that all terminals share the same side messages, for the GNN's... In the next iteration, during the propagation phase, the messages on each edge are updated. GNN is called in each iteration. Next iteration: in This represents the intermediate variable of the GNN in iteration t; the superscript T here indicates transpose. During the aggregation phase, for the first The message at a given node is obtained by summing the messages on all its adjacent edges. ; Aggregated node messages Further input into the GRU network of the gate control recurrent unit Obtain hidden state variables And update, ; Then, based on the hidden state variables, the node's state variables are updated to complete the first step. Next iteration: ; When GNN completes After iteration, the posterior mean optimized by GNN is obtained. ,variance sparsity : Then, orthogonalization was performed. With damping coefficient To perform damping updates, The value of is a real number in the range [0, 1]. Update sparsity and link variance to complete the first step. Round iteration; Link variance is ; sparsity is go through After several iterations, the link estimation results, sparsity matrix, and posterior variance vector are obtained: The link estimation results are as follows: ; The sparsity matrix is: ; The posterior variance vector is: .

4. The method for multi-terminal, multi-service shared network transmission access detection and link state information estimation as described in claim 3, characterized in that, Step 4 employs the following method to detect terminal activity: First, the output variables of the link estimation network are averaged to obtain the mean values ​​of the link estimation results, the sparsity matrix, and the posterior variance vector: in Used to measure the effectiveness of link estimation Represents the SoftMax activation function. As trainable parameters, treating activity detection as a binary classification task, the output of the fully connected network is: ; Selected First element Representing the The active probability of each terminal is used to determine the support set for detection based on the threshold. The support set is the set of active terminals. .

Images