Adaptive intelligent antenna system based on deep learning
By constructing a signal construction module, a topology modeling module, a weighted prediction module, and an interference suppression module, and utilizing graph neural networks and generative adversarial networks, the problem of unstable beam pointing in smart antenna systems under strong interference and multi-user scenarios was solved, achieving higher signal stability and resource utilization efficiency.
Patent Information
- Application Number
- CN202511525083.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-24
- Publication Date
- 2026-02-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In existing technologies, the array element weighting coefficients of smart antenna systems are unstable in strong interference and multi-user concurrent scenarios, leading to unstable beam pointing, increased bit error rate, and decreased throughput. Furthermore, traditional methods are not refined enough in terms of input representation, coupling representation, and stability constraint operations, resulting in insufficient robustness.
By constructing a signal construction module, a topology modeling module, a weighted prediction module, and an interference suppression module, and utilizing graph neural networks and generative adversarial networks, the array weighting parameters are gradually filtered and updated to achieve a clear expression of the array topology features and systematic compression of interference signals.
It improves the convergence and anti-interference ability of array weighting parameters, enhances the dynamic tracking capability of signals, reduces the bit error rate, and improves the stability and resource allocation efficiency of the system.
Smart Images

Figure CN121479290A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart antenna array technology, and more particularly to an adaptive smart antenna system based on deep learning. Background Technology
[0002] The field of intelligent antenna array technology mainly involves constructing multi-antenna array structures and combining signal processing algorithms and control strategies to achieve adaptive directivity adjustment, beamforming, interference suppression, and channel optimization of signals. The core principle is to utilize the spatial degrees of freedom of the array antenna and change the radiation pattern by adjusting the weighting coefficients of the array elements, thereby enhancing the target signal and attenuating and shielding the interference signal, improving spectrum utilization, transmission quality, and enhancing anti-interference capabilities.
[0003] An adaptive smart antenna system based on deep learning combines a deep learning model with an adaptive antenna control mechanism. By constructing a deep neural network, it models and predicts wireless channel states, interference patterns, and user movement trajectories, and then dynamically adjusts the radiation pattern and gain distribution of the antenna array. The goal is to achieve high-precision acquisition and tracking of target signals, improve signal reception quality, reduce bit error rate, enhance system throughput, and rationally allocate space and spectrum resources in multi-user access scenarios, thereby achieving efficient wireless communication. Through the nonlinear mapping capability of the deep learning model, it maintains stable performance under conditions of rapid time-varying, multipath fading, and multi-user interference, meeting the future requirements for high-capacity, low-latency, and high-reliability communication.
[0004] Existing technologies adjust radiation patterns using array element weighting coefficients and rely on deep networks to model and predict channels and interference. In actual operation, common problems include inconsistent input organization, unintuitive representation of coupling information, and a lack of clear and stable criteria and interpretable structures for weight updates. When the correspondence between the time and spatial dimensions is not clear, differential comparison is easily affected by instantaneous noise, causing weights to fluctuate between adjacent time slots, resulting in unstable beam pointing. In scenarios with strong interference and multi-user concurrency, the update process is not robust enough to abnormal samples, and errors propagate to subsequent time slots, leading to increased bit error rate and decreased throughput, as well as decreased short-term reliability and resource allocation efficiency. Traditional methods are not refined enough in terms of input representation, coupling expression, and stability constraints, which can easily lead to convergence instability, interference residues, and time-domain response lag. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an adaptive smart antenna system based on deep learning.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: an adaptive smart antenna system based on deep learning includes: Signal construction module: Receives amplitude, phase and time index through antenna array elements, multiplies amplitude and phase channel by channel to generate composite signal, arranges composite signals according to time index order and combines them in a matrix to establish a receiving array matrix; Topology modeling module: Based on the received array matrix, calculate the element difference and compare it with the spatial coordinates, screen out array element pairs that exceed the threshold and mark them, combine graph neural network to generate an adjacency matrix and iteratively merge them to form array topology features; Weighted prediction module: Based on the array topology features, expand the node features and determine the differences. Low differences are superimposed, while high differences are retained independently. A generative adversarial network is introduced to compare the generated vectors with the theoretical weights and select the parameter set with the lowest differences. The parameter set is repeatedly updated until it is stable, and array weighted parameters are generated. Interference suppression module: Based on the array weighting parameters, the parameters are multiplied with the corresponding elements of the receiving array matrix to generate a weighting matrix, the main signal vector is extracted and the remaining vectors are compared, and vectors with differences exceeding the threshold are determined to be interference terms and stripped to obtain the interference suppression signal; State evolution module: Based on the interference suppression signal, the signal is segmented in time sequence and frequency components are expanded to form a multi-dimensional matrix. Then, trend factors are extracted, the factor sequence is extended over time and combined with array weighting parameters to obtain channel evolution parameters.
[0007] As a further embodiment of the present invention, the receiving array matrix includes an amplitude matrix, a phase matrix, and a time index matrix; the array topology features include an element difference matrix, a set of coupled elements, and an adjacency matrix; the array weighting parameters include node feature vectors, a set of node differences, and a parameter set; the interference suppression signal includes a main signal vector, a residual signal, and an interference term; and the channel evolution parameters include a trend factor sequence, an extension sequence, and a parameter matrix.
[0008] As a further aspect of the present invention, the signal construction module includes: Amplitude-phase synthesis submodule: Based on the amplitude, phase and time index received by the antenna array elements, the amplitude and phase are multiplied one channel at a time to generate a composite signal, and the composite signals are arranged in the order of the time index and continuously spliced to form a vector sequence to generate a serialized composite signal; Matrix generation submodule: Based on the serialized composite signal, the channel vectors are combined and arranged in the order of array element numbers. The merged vectors are then placed in the row and column positions of the matrix to reorganize the matrix structure and establish the receiving array matrix.
[0009] As a further aspect of the present invention, the topology modeling module includes: Difference calculation submodule: Based on the receiving array matrix, calculate the difference in amplitude between two consecutive array elements and the difference in phase between two array elements, write the amplitude difference and phase difference into the matrix position and sort them according to the array element index to generate the array element difference matrix. Threshold filtering submodule: Based on the array element difference matrix, compare each element of the amplitude difference and phase difference in the matrix with the spatial coordinate distance, and mark the elements that exceed the threshold. Combine the marking results into a set and index and classify them to generate a coupled array element set. The adjacency synthesis submodule: Based on the coupled array element set, pair elements are matched one by one and association chains are established. The matrices of each chain are combined to generate an adjacency matrix. A graph neural network is used to perform multi-layer convolution on the adjacency matrix and extract the spatial features of nodes and neighboring nodes. Then, pooling compression is performed to generate a low-dimensional representation and establish array topology features.
[0010] As a further aspect of the present invention, the specific execution process of the graph neural network is as follows: based on the adjacency matrix, the product of the feature value of each node and the feature value of the adjacent nodes is calculated and summed to form the combined feature of the node. The combined feature is then passed to the next layer of convolution in sequence. The layer-by-layer calculation is repeated until the node is updated. Then, the updated node features are pooled and compressed into a low-dimensional matrix representation to establish the array topology feature.
[0011] As a further aspect of the present invention, the specific steps of matching each pair of elements and establishing an association chain are as follows: in the coupled element set, each element is read sequentially according to the element index order, the current element is paired with the remaining unpaired elements in the set, and the resulting pairing is recorded in a linked list structure. The linked list node contains the number index of the two elements and the position index in the coupled set. After the pairing is completed, the linked list nodes are connected sequentially, with the pointer of the previous node pointing to the next node, forming an association chain of multiple nodes connected sequentially.
[0012] As a further aspect of the present invention, the weighted prediction module includes: Feature expansion submodule: Based on the array topology features, expand the features of each node, record the feature values of each node into a vector in a specified order, and merge the node features to generate a node feature vector; Difference determination submodule: Based on the node feature vector, calculate the difference between the features of adjacent nodes element by element, judge the difference by setting a threshold, add the difference less than the threshold to the feature value of the previous node, and retain the difference greater than the threshold separately to generate a node difference set; Parameter update submodule: Based on the node difference set, a generative adversarial network is introduced to generate candidate parameters for the generated vector and the discriminator compares them with the true values. A threshold is set and the difference is judged. The parameter set with low difference is selected and weighted. The new parameters after adjustment are compared with the original parameters until the difference is no longer significant, and the array weighted parameters are obtained.
[0013] As a further aspect of the present invention, based on the node difference set, the generator generates candidate parameters item by item according to the set vector and outputs them as a parameter sequence. The discriminator receives the parameter sequence and compares it element by element with the theoretical weight sequence, outputs the judgment result and feeds it back to the generator. The generator adjusts the generated parameters according to the feedback. After multiple iterations, the parameter set with lower differences is selected to obtain the array weighted parameters.
[0014] As a further aspect of the present invention, the interference suppression module includes: Weighted generation submodule: Based on the array weighting parameters, each parameter is multiplied by the element at the corresponding position in the receiving array matrix, and the resulting product is sequentially filled into the matrix unit, arranged and combined according to the row and column numbers to establish the matrix structure and form a weighted matrix; Interference determination submodule: Based on the weighting matrix, extract the main signal vector and calculate the difference between each item and the remaining vector. Compare the difference with the threshold and determine the vector that exceeds the threshold. Separate the vectors one by one and record the remaining signal to obtain the interference suppression signal.
[0015] As a further aspect of the present invention, the state evolution module includes: Time-frequency submodule: Based on the interference suppression signal, the signal is segmented and numbered in time sequence, the frequency is quickly expanded and each frequency component is extracted, the components are filled into the matrix row and column units to form a multi-dimensional matrix, the matrix is decomposed row by row and the trend factors are extracted to generate a trend factor sequence. Factor fusion submodule: Based on the trend factor sequence, extend the time of the factor sequence and generate an extended sequence. Combine the extended sequence with the array weighted parameters at the corresponding positions to obtain a new parameter matrix. Summarize the matrix column by column and record it as a parameter sequence to obtain the channel evolution parameters.
[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: 1. In this invention, by combining the amplitude difference and phase difference of array elements with the spatial coordinate distance as a screening method, array element pairs exceeding the threshold are marked and transformed into adjacency relationships. By utilizing the multi-layer iterative merging topology, the array features present a clearer coupling hierarchy. 2. In this invention, by iteratively selecting a set with lower differences until a stable condition is reached, a weighting parameter with stronger convergence is obtained. During the formation of the weighting matrix, the main signal vector and the remaining vector are differentially divided item by item and compared with a threshold to achieve systematic compression of interference components in the signal. 3. In this invention, by using a multidimensional matrix structure to carry frequency components and extract trend factors during the signal time expansion stage, and combining the factor sequence with weighted parameters after extending it in the time dimension, a parameter set that can express the dynamic evolution of the channel is generated, which makes the result have stronger anti-interference and better dynamic tracking performance. Attached Figure Description
[0017] Figure 1 This is a system flowchart of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0019] Example 1 Please see Figure 1 This invention provides a technical solution: an adaptive smart antenna system based on deep learning, comprising: Signal construction module: Receives amplitude, phase and time index through antenna array elements, multiplies amplitude and phase channel by channel to generate composite signal, arranges composite signals according to time index order and combines them in a matrix to establish a receiving array matrix; Topology modeling module: Based on the receiving array matrix, calculate the difference between array elements and compare it with spatial coordinates, screen array element pairs that exceed the threshold and then mark them, combine graph neural network to generate adjacency matrix and iteratively merge to form array topology features; Weighted prediction module: Based on array topology features, expand node features and determine differences. Low differences are superimposed, while high differences are retained independently. Generative adversarial network is introduced to compare the generated vectors with theoretical weights and select the parameter set with lower differences. The parameter set is repeatedly updated until it is stable, and array weighted parameters are generated. Interference suppression module: Based on the array weighting parameters, the parameters are multiplied with the corresponding elements of the receiving array matrix to generate a weighting matrix, the main signal vector is extracted and the remaining vectors are compared, and vectors with differences exceeding the threshold are identified as interference terms and removed to obtain the interference suppression signal; State evolution module: Based on the interference suppression signal, the signal is segmented in time sequence and frequency components are expanded to form a multi-dimensional matrix. Then, trend factors are extracted, the factor sequence is extended over time and combined with array weighting parameters to obtain channel evolution parameters.
[0020] The receiving array matrix includes an amplitude matrix, a phase matrix, and a time index matrix. The array topology features include an element difference matrix, a set of coupled elements, and an adjacency matrix. The array weighting parameters include node feature vectors, a set of node differences, and a set of parameters. The interference suppression signals include the main signal vector, the remaining signal, and the interference terms. The channel evolution parameters include a trend factor sequence, an extension sequence, and a parameter matrix.
[0021] The signal building module includes: Amplitude-phase synthesis submodule: Based on the amplitude, phase and time index received by the antenna array elements, the amplitude and phase are multiplied one channel at a time to generate a composite signal, and the composite signals are arranged in the order of the time index and continuously spliced to form a vector sequence to generate a serialized composite signal; Matrix generation submodule: Based on the serialized composite signal, it combines channel vectors and arranges them in the order of array element numbers. The merged vectors are then placed in the row and column positions of the matrix to reorganize the matrix structure and establish the receiving array matrix. Amplitude and phase synthesis submodule: Based on the amplitude, phase and time index received by the antenna array elements, it adopts complex number operation method, sets the amplitude range to a decimal between 0 and 1, the phase range to an angle value of 0 to 360 degrees, and the time index to increase sequentially at 1-millisecond intervals. It multiplies the amplitude and phase of each channel one by one, and records the result as a composite signal. The composite signal is arranged in the order of the time index. After the arrangement is completed, the output length is set to 1024 elements. The composite signal is continuously spliced to form a one-dimensional vector sequence of length 1024, generating a serialized composite signal. Matrix generation submodule: Based on serialized composite signals, the number of channels is set to 16, and the length of one channel vector is 64 elements. The one-dimensional vector sequence is divided into 16 channel vectors. The index of the channel vectors is arranged in the order of the array element numbers, which increase from 1 to 16. The divided channel vectors are placed in the row and column positions of the matrix in sequence. The matrix size is fixed at 16 rows and 64 columns. After the vectors are filled, a matrix structure is formed, and the receiving array matrix is generated.
[0022] The topology modeling module includes: Difference Calculation Submodule: Based on the receiving array matrix, calculate the difference in amplitude between two consecutive array elements and the difference in phase between two array elements, write the amplitude difference and phase difference into the matrix position and sort them according to the array element index to generate the array element difference matrix. Threshold filtering submodule: Based on the array element difference matrix, compare each element of the amplitude difference and phase difference in the matrix with the spatial coordinate distance, and mark the elements that exceed the threshold. Combine the marking results into a set and index and classify them to generate a coupled array element set. The adjacency synthesis submodule: Based on the coupled array element set, it matches the array element pairs one by one and establishes the association chain. It combines the matrices of each chain and generates the adjacency matrix. It uses a graph neural network to perform multi-layer convolution on the adjacency matrix and extracts the spatial features of nodes and neighboring nodes. Then, it performs pooling compression to generate a low-dimensional representation and establishes the array topology features. The difference calculation submodule is based on the receiving array matrix. The number of array elements is set to 16, the amplitude range of adjacent array elements is 0 to 1, and the phase range is 0 to 360 degrees. During the operation, the amplitude of two adjacent array elements is read and the difference is calculated. The difference is calculated by subtracting the amplitude from the amplitude 1 and keeping three decimal places. At the same time, the phase difference between two adjacent array elements is calculated by subtracting the phase from the phase 1 and taking the modulus when it exceeds 360 degrees. The obtained amplitude difference and phase difference are written into a 16-row 16-column matrix. The matrix is sorted according to the array element index 1 to 16. After sorting, the array element difference matrix is generated. Threshold Filtering Submodule: Based on the element difference matrix, the matrix elements are filtered. The threshold values for amplitude difference are set to 0.15, phase difference to 10 degrees, and spatial coordinate distance to 0.5 meters. Each amplitude difference element in the matrix is compared with the threshold value of the phase difference element. If the difference is greater than the set threshold, it is marked as 1; if it is less than or equal to the threshold value, it is marked as 0. At the same time, the spatial coordinates of the corresponding element are read. The spatial coordinates are represented by three-dimensional Cartesian coordinates and rounded to two decimal places. The Euclidean distance between the coordinates is compared to see if it exceeds 0.5 meters. If it does, a mark is added. The marked elements are combined into a set and classified according to the element index to generate a coupled element set. The adjacency synthesis submodule, based on a coupled array element set, uses a graph neural network to matrix-process the array element pairs. A coupled array element set is limited to a maximum of 32 pairs. Each pair is paired sequentially, and an association chain is established. The association chain uses a linked list structure to store node information, including the two array element numbers and their index positions in the set. The linked list nodes are connected in the input order to form a chain structure. When matrix-processing the chain structure, the array element numbers are mapped to rows and columns of the adjacency matrix. If a coupling relationship exists, the matrix position is marked as 1; otherwise, it is marked as 0, forming a 16x16 adjacency matrix. Graph convolution is applied to the adjacency matrix using three layers of convolution, with the feature dimensions of each convolutional node set to 64, 32, and 16. After convolution, pooling is performed using average pooling, compressing the matrix into a low-dimensional representation matrix of 16x16, generating array topological features.
[0023] The specific execution process of the graph neural network is as follows: based on the adjacency matrix, the product of the feature value of each node and the feature value of the adjacent nodes is calculated and summed to form the combined feature of the node. The combined feature is then passed to the next layer of convolution in turn. The layer-by-layer calculation is repeated until the node is updated. Then, the updated node features are pooled and compressed into a low-dimensional matrix representation to establish the array topological feature. Graph neural networks, according to the formula: ;
[0024] in: For the first The node at the th The layer outputs a feature vector. For the first The node at the th Layer input feature vector, For the first Layer weight matrix, For activation function, For the first A set of neighboring nodes, For the first The node and the first Each node normalization constant, Spatial distance weighting coefficient, This is the amplitude difference weighting coefficient. The phase difference weighting coefficient, The weighting coefficient is the difference in time index. For the first The node and the first The spatial Euclidean distance between nodes For the first The node and the first The difference in amplitude at each node For the first The node and the first Phase difference between nodes For the first The node and the first The time index difference of each node For node indexing, Indexing neighboring nodes, For convolutional layer index; Execution process: Given the adjacency matrix and initial node features, index by layer. Traverse the graph convolutional layers and index each node. Read the neighbor set And retrieve the neighbor index one by one. Calculated based on three-dimensional coordinates Using meters as the unit and rounding to two decimal places, calculate based on the amplitude sequence. Take the interval from zero to one and retain three decimal places, then calculate based on the phase sequence. The unit is degrees, and the angle is wrapped around to 0 to 360 degrees. Calculation is based on the time index. Using sampling points as the unit and limiting them to non-negative integers, set the weighting coefficients. , , , And fix it, calculate the weighted combination and with Normalization is performed, and matrix dimension matching is used to... Acting on And sum them up, the summed result is... The output is And write the node index into the output tensor of this layer. After the update is complete, proceed to the next level of index. Repeat the same steps until the last layer is output.
[0025] The specific steps for matching element pairs and establishing association chains are as follows: In the coupled element set, each element is read sequentially according to its index. The current element is paired with the remaining unpaired elements in the set. The resulting pairings are recorded in a linked list structure. Each linked list node contains the index of the two elements and their position index in the coupled set. After pairing, the linked list nodes are connected sequentially, with the pointer of the previous node pointing to the pointer of the next node, forming an association chain of multiple nodes connected sequentially.
[0026] The weighted prediction module includes: Feature expansion submodule: Based on the array topology features, expand the features of each node, record the feature values of each node into a vector in a specified order, and merge the node features to generate a node feature vector; The difference determination submodule calculates the difference between the features of adjacent nodes element by element based on the node feature vector. It judges the difference by setting a threshold, and adds the difference less than the threshold to the feature value of the previous node. The difference greater than the threshold is kept separately to generate a set of node differences. The parameter update submodule: Based on the node difference set, a generative adversarial network is introduced to generate candidate parameters for the generated vector and the discriminator compares them with the ground truth. A threshold is set and the difference is judged. The parameter set with low difference is selected and weighted. The new parameters after adjustment are compared with the original parameters until the difference is no longer significant, and the array weighted parameters are obtained. Feature expansion submodule: Based on array topology features, organize node features, set the total number of nodes to 32, and the node features include three types of data: amplitude, phase and position index. The amplitude range is a decimal from 0 to 1, the phase range is 0 to 360 degrees, and the position index is an integer from 1 to 32. During the expansion process, the node features are read in the order of number from 1 to 32. The amplitude, phase and position index of each node are recorded in a vector unit of length 3. Then, the 32 units are concatenated in the order of number to form a vector of length 96, and the vectors are merged to generate the node feature vector. The difference determination submodule calculates the difference element by element based on the node feature vector. It sets the threshold for amplitude difference between adjacent nodes to 0.05, the threshold for phase difference to 5 degrees, and the threshold for position index difference to 2. It compares each pair of adjacent nodes in turn. Elements with amplitude difference less than 0.05 are numerically superimposed with the amplitude of the previous node and retained to one decimal place. Elements with phase difference less than 5 degrees are numerically superimposed with the phase of the previous node and retained to an integer. Elements with position index difference less than 2 are calculated and summed with the index of the previous node and retained to an integer. If the difference is greater than the preset threshold, the element is directly retained without superposition. The result set after determination is output to generate the node difference set. The parameter update submodule uses a generative adversarial network (GAN) to generate and filter candidate parameters based on the node difference set. The generator's input vector dimension is set to 64, the output vector dimension to 32, the batch size to 16, the learning rate to 0.001, and the number of iterations to 100. The generator transforms the input node difference set into a sequence of candidate parameters. The discriminator's input vector dimension is set to 32, and the output class is binary (true / false). The discriminator compares the candidate parameter sequence with the theoretical weight sequence, with a comparison threshold of 0.1. If the difference is less than 0.1, the parameter is considered valid; otherwise, it is considered invalid. The selected valid parameter set is then weighted and adjusted, with a fixed weighting factor of 0.5. The adjusted new parameters are compared with the original parameters one by one. When the difference is less than 0.05 for 10 consecutive iterations, the parameter is considered stable, and the array-weighted parameters are output.
[0027] The specific execution process of the Generative Adversarial Network is as follows: Based on the set of node differences, the generator generates candidate parameters item by item according to the set vector and outputs them as a parameter sequence. The discriminator receives the parameter sequence and compares it element by element with the theoretical weight sequence, outputs the judgment result and feeds it back to the generator. The generator adjusts the generated parameters according to the feedback. After multiple iterations, the set of parameters with lower differences is selected to obtain the array weighted parameters. Generate adversarial networks according to the formula: ;
[0028] in: To modify the adversarial objective function scalar, The generator parameter set vector. Let the discriminator parameter set vector be... The input vector is the real sample. The noise input vector, For the generator mapping operator Mapped to a candidate parameter vector. The discriminator mapping operator maps the input vector to a score between zero and one. For the true sample distribution, For noise distribution, These are the weighting coefficients for the squared difference term. These are the weighting coefficients for the inner product terms. For the theoretical reference parameter sequence vector, This is an auxiliary vector obtained from node difference statistics. For expectation operator, For the natural logarithm operator, It is a 2-norm. For vector dot product operators, For vector element index, Represents the candidate parameter vector of the th One element, Indicates the theoretical reference parameter number One element, Represents the auxiliary vector. One element, and These represent the scores given by the discriminator to the real samples and the candidate samples, respectively. Execution process: In the weighted prediction stage, batch data is set and obtained from the distribution... Extract real samples From the distribution Extracting noise Generator In parameters General Mapping yields candidate parameter vectors And denote each element as By the discriminator In parameters The following calculations are performed separately for each pair. and The rating is to get and Batch expectations for both items, reading the theoretical reference sequence. And perform a L2 norm squared difference term with the candidate vectors. Where the element-wise difference is Read auxiliary vectors And perform an inner product with the candidate vector. Set weights and And linear synthesis After one forward computation, for and Based on respectively The gradient is updated alternately, and the sampling and update steps are repeated until the parameters are obtained and output within a predetermined number of iterations. The given array weighted parameter vector.
[0029] The interference suppression module includes: Weighted generation submodule: Based on the array weighting parameters, each parameter is multiplied by the corresponding element in the receiving array matrix, and the resulting product is filled into the matrix unit in sequence, arranged and combined according to the row and column numbers to establish the matrix structure and form a weighted matrix; Interference determination submodule: Based on the weighting matrix, extract the main signal vector and calculate the difference between each item and the remaining vector. Compare the difference with the threshold and determine the vector that exceeds the threshold. Separate the vectors one by one and record the remaining signal to obtain the interference suppression signal. Weighted Generation Submodule: Based on the array weighting parameters, it calculates the parameters and the receiving array matrix. The array size is set to 16 rows and 64 columns. The array weighting parameters range from 0.1 to 1.0 floating-point numbers. The elements of the receiving array matrix range from -1 to 1 decimal values. During the operation, the element at each matrix position is multiplied by the parameter at the corresponding position. The product result is retained to three decimal places. The product value is filled into the matrix cells in the order of row number from 1 to 16. The elements in each row are arranged and combined in the order of column number from 1 to 64. After filling, a new matrix structure is formed, generating the weighted matrix. Interference Detection Submodule: Based on the weighted matrix, the main signal vector and the residual vector are compared item by item. The main signal vector is set as the 16 elements in the first column of the weighted matrix, and the residual vectors are the vectors in columns 2 to 64. The difference between the main signal vector and the residual vector is calculated column by column. The threshold for amplitude difference is set to 0.2, and the threshold for phase difference is set to 15 degrees. The difference of each column vector is calculated element by element and compared with the corresponding threshold. If the difference is greater than the set threshold, the vector is marked as an interference vector. If the difference is less than or equal to the threshold, it is retained in the residual set. The marked interference vectors are stripped and removed from the matrix in turn. The vectors that are not stripped are recorded as the signal set, and the interference suppression signal is generated.
[0030] The state evolution module includes: Time-frequency submodule: Based on the interference suppression signal, the signal is segmented and numbered in time sequence, the frequency is quickly expanded and each frequency component is extracted, the components are filled into the matrix row and column units to form a multi-dimensional matrix, the matrix is decomposed row by row and the trend factors are extracted to generate a trend factor sequence. Factor fusion submodule: Based on the trend factor sequence, extend the time of the factor sequence and generate an extended sequence. Combine the extended sequence with the array weighted parameters at the corresponding positions to obtain a new parameter matrix. Summarize the matrix column by column and record it as a parameter sequence to obtain the channel evolution parameters. Time-Frequency Submodule: Based on the interference suppression signal, the signal is frequency-expanded. The signal sampling rate is set to 1000Hz, and the total signal length is 1024 sampling points. The interference suppression signal is divided into 4 segments of 256 sampling points each, and each segment is numbered from 1 to 4. Each segment is processed by Fast Fourier Transform, and the frequency component resolution is set to 3.9Hz. Frequency components in the range of 0 to 500Hz are extracted and sequentially filled into a two-dimensional matrix of size 4 rows and 128 columns. The row number of the matrix corresponds to the segment number, and the column number corresponds to the frequency component number. Then, the matrix is decomposed row by row, with a decomposition step size of 16 components. Each row and 128 columns is decomposed into 8 small vectors. The length of each small vector is fixed at 16 elements. The mean and standard deviation of each small vector are calculated as trend factor values, and they are sequentially combined into a sequence to generate a trend factor sequence. Factor Fusion Submodule: Based on the trend factor sequence, the length of the trend factor sequence is set to 32 elements, and the time extension factor is set to 2. The original 32 elements are copied and appended sequentially to obtain an extended sequence of length 64. The extended sequence is combined with the array weighted parameters at the corresponding positions. The array weighted parameters have a fixed dimension of 16 rows and 4 columns. The extended sequence is divided into 4 groups of 16 elements each, corresponding to the 4 columns of the matrix. During the combination process, the 16 elements of each group are superimposed in row order to the corresponding positions of the parameter matrix to generate a new matrix with a size of 16 rows and 4 columns. Then, the matrix is summarized by column, and the arithmetic mean of the elements in each column is set as the column summary result to obtain a parameter sequence of length 4, which is recorded as the signal evolution result and generates the channel evolution parameters.
[0031] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A deep learning based adaptive smart antenna system, characterized by, The system comprises: a signal construction module: receiving amplitude, phase and time index through antenna array elements, multiplying amplitude and phase to generate complex signals channel by channel, arranging complex signals in time index order and matrix combination, establishing a receiving array matrix; a topology modeling module: based on the receiving array matrix, calculating array element difference and comparing with spatial coordinates, screening and marking super-threshold array element pairs, combining graph neural network to generate adjacency matrix and iterative merging, forming array topology features; a weighted prediction module: based on the array topology features, expanding node features and determining differences, performing superposition on low differences, independently retaining high differences, introducing generative adversarial network to compare generated vectors and theoretical weights and select parameter set with lower difference, repeatedly updating parameter set until stable, generating array weighting parameters; an interference suppression module: based on the array weighting parameters, multiplying parameters and corresponding elements of the receiving array matrix to generate a weighted matrix, extracting a main signal vector and comparing the remaining vectors, determining that the vectors with differences exceeding the threshold are interference terms and stripping them to obtain an interference suppression signal; a state evolution module: based on the interference suppression signal, cutting the signal in time order and expanding the frequency components, forming a multi-dimensional matrix, extracting trend factors, time extending the factor sequence and combining with the array weighting parameters to obtain channel evolution parameters.
2. The deep learning based adaptive smart antenna system of claim 1, wherein, The receiving array matrix includes amplitude matrix, phase matrix and time index matrix, the array topology features include array element difference matrix, coupled array element set and adjacency matrix, the array weighting parameters include node feature vector, node difference set and parameter set, the interference suppression signal includes main signal vector, residual signal and interference term, and the channel evolution parameters include trend factor sequence, extended sequence and parameter matrix.
3. The deep learning based adaptive smart antenna system of claim 1, wherein, The signal construction module comprises: an amplitude-phase synthesis submodule: based on the amplitude, phase and time index received by the antenna array elements, multiplying the amplitude and phase to generate complex signals channel by channel, and arranging the complex signals in time index order, continuously splicing to form a vector sequence, generating a serialized complex signal; a matrix generation submodule: based on the serialized complex signal, combining channel vectors and merging and arranging them in array element number order, placing the merged vectors in matrix row and column positions in turn, reorganizing the matrix structure, and establishing a receiving array matrix.
4. The deep learning based adaptive smart antenna system of claim 1, wherein, The topology modeling module comprises: a difference calculation submodule: based on the receiving array matrix, calculating the difference of the amplitudes of two consecutive array elements and the difference of the phases of the two array elements, writing the amplitude difference and phase difference into the matrix positions and sorting them by array element index, generating an array element difference matrix; a threshold screening submodule: based on the array element difference matrix, comparing each element of the amplitude difference and phase difference in the matrix with the spatial coordinate distance, marking the elements exceeding the threshold at the same time, combining the marking results into a set and indexing and classifying them, generating a coupled array element set; The adjacent synthesis submodule: based on the coupling array element set, pairing the element pairs one by one and establishing the associated chain, combining the matrices of each chain and generating the adjacency matrix, using the graph neural network to perform multi-layer convolution on the adjacency matrix and extract the spatial features of the nodes and adjacent nodes, then performing pooling compression to generate a low-dimensional representation form, and establishing the array topology feature.
5. The deep learning based adaptive smart antenna system of claim 4, wherein, The specific execution process of the graph neural network is to calculate the product and sum of the eigenvalues of each node and the eigenvalues of the adjacent nodes based on the adjacency matrix, form the combined features of the nodes, and sequentially pass the combined features to the next layer of convolution, repeatedly calculate layer by layer until the nodes are updated, and then perform pooling operation on each updated node feature and compress it into a low-dimensional matrix representation to establish the array topology feature.
6. The deep learning based adaptive smart antenna system of claim 4, wherein, The specific steps of pairing the element pairs one by one and establishing the associated chain are as follows: in the coupling array element set, read each array element in order according to the array element index, pair the current array element with the remaining unpaired array elements in the set, record the formed pairs in the linked list structure, the linked list node contains the index of the two array elements and the position index in the coupling set, and after pairing is completed, the linked list nodes are connected one by one, the pointer of the previous node points to the next node, forming an associated chain connected by multiple nodes in order.
7. The deep learning based adaptive smart antenna system of claim 1, wherein, The weighted prediction module comprises: The feature expansion submodule: based on the array topology feature, expand each node feature, and record each node feature value in a vector according to a specified order, and combine the node features to generate a node feature vector; The difference determination submodule: based on the node feature vector, calculate the difference value of adjacent node features element by element, determine the difference value by setting a threshold, add the difference value less than the threshold to the previous node feature value, and separately retain the difference value greater than the threshold to generate a node difference set; The parameter update submodule: based on the node difference set, introduce a generative adversarial network to generate candidate parameters for the generated vector and compare the true value by a discriminator, set a threshold and determine the difference, select a parameter set with lower difference and perform weighted adjustment, compare the new parameters after adjustment with the original parameters until the difference is no longer significant, and obtain the array weighting parameters.
8. The deep learning based adaptive smart antenna system of claim 7, wherein, The specific execution process of the generative adversarial network is to generate candidate parameters one by one from the generator based on the node difference set and output them as a parameter sequence, the discriminator receives the parameter sequence and compares it with the theoretical weight sequence element by element, outputs the determination result and feeds it back to the generator, the generator adjusts the generated parameters according to the feedback, and after multiple iterations, a parameter set with lower difference is selected to obtain the array weighting parameters.
9. The deep learning based adaptive smart antenna system of claim 1, wherein, The interference suppression module comprises: The weighted generation submodule: based on the array weighting parameters, multiply each parameter with the corresponding element in the received array matrix, fill the resulting product into the matrix cells in order, arrange and combine them according to the row and column numbers, establish the matrix structure and form the weighted matrix; The interference determination submodule: based on the weighted matrix, extract the main signal vector and calculate the difference value with the remaining vectors one by one, compare the difference value with the threshold value and determine the vectors exceeding the threshold value, peel off the vectors one by one and record the remaining signals to obtain the interference suppression signal.
10. The deep learning based adaptive smart antenna system of claim 1, wherein, The state evolution module comprises: The time-frequency sub-module: based on the interference suppression signal, the signal is divided into segments in time sequence and numbered, the frequency is quickly developed and each frequency component is extracted, the components are filled into matrix row and column units to form a multi-dimensional matrix, the matrix is operated row by row and the trend factor is extracted, and a trend factor sequence is generated; The factor fusion sub-module: based on the trend factor sequence, the time of the factor sequence is extended to generate an extended sequence, the extended sequence is combined with the array weighting parameters at the corresponding positions to obtain a new parameter matrix, the matrix is summarized column by column and recorded as a parameter sequence, and a channel evolution parameter is obtained.
