Search

CN-121984553-A - Millimeter wave MIMO system beam selection method based on deep learning

CN121984553ACN 121984553 ACN121984553 ACN 121984553ACN-121984553-A

Abstract

The invention belongs to the field of wireless communication, and particularly relates to a millimeter wave MIMO system beam selection method based on deep learning. The method comprises the steps of synchronously obtaining an environment image and a radio frequency pilot signal, extracting physical environment characteristics by using a convolutional neural network, converting the physical environment characteristics into a spatial sparsity constraint matrix, mapping the radio frequency signal from a spatial domain to a beam domain to construct a characteristic map, carrying out weighting and clipping under physical guidance on the beam domain characteristics by using the sparsity constraint matrix by adopting a cross-modal attention fusion mechanism, and outputting an optimal beam pair index through a full connection layer. The invention realizes the deep coupling of vision and radio frequency information, avoids the physical blocking direction, reduces the pilot frequency overhead and time delay of beam scanning, and enhances the robustness and the access efficiency of millimeter wave links.

Inventors

  • CHEN XUEHUA
  • DENG HANYUAN
  • CAI YAJUN
  • YE YILANG
  • ZHUO ZHANGYU
  • XIE ZHIMAO

Assignees

  • 国脉科技股份有限公司

Dates

Publication Date
20260505
Application Date
20260402

Claims (10)

  1. 1. The millimeter wave multiple-input multiple-output system beam selection method based on deep learning is characterized by comprising the following specific steps of: step 1, acquiring environment sensing data and radio frequency original signals, acquiring visual environment images in a communication coverage area by using an auxiliary sensing module arranged at a base station end, and synchronously transmitting a preset number of pilot signals through a baseband processing unit to acquire channel state information in an initial state; Step 2, performing visual semantic feature extraction, inputting an acquired visual environment image into a preset lightweight convolutional neural network model for semantic segmentation, identifying and extracting obstacle mask features and reflecting surface physical features existing in the image, and further converting the environment physical information into a space sparsity constraint matrix; Step 3, performing beam domain feature reconstruction, performing transformation processing on the acquired initial channel state information through the baseband processing unit, mapping the channel state information from a space domain to a beam domain, and constructing a beam domain feature map for representing the radio frequency energy distribution feature; Step 4, performing cross-modal attention fusion processing, namely inputting the spatial sparsity constraint matrix and the beam domain feature map into a cross-modal attention fusion network, and weighting and clipping the beam domain feature map by using the spatial sparsity constraint matrix as a physical guide mask so as to inhibit the beam weight in the blocking direction and enhance the beam feature in the visual path direction; And 5, executing beam decision output, inputting the fused cross-mode feature vector into a full-connection layer for classification prediction, and outputting an index number of an optimal beam pair in the current communication environment.
  2. 2. The method for selecting a millimeter wave mimo system beam based on deep learning according to claim 1, wherein in step 1, the auxiliary sensing module comprises a low resolution vision sensor disposed on top of the base station, the vision sensor comprises an optical camera and a laser detection and measurement system; The visual sensor keeps synchronous with a baseband clock of the multi-input multi-output system through a preset synchronous trigger interface, and the baseband processing unit sends a trigger pulse instruction to the visual sensor at the moment when each pilot frequency time slot starts through the synchronous trigger interface, so that each captured frame of visual environment image is aligned with the corresponding pilot frequency signal in the time dimension; The acquired visual environment image is subjected to dimension reduction processing through a preset data compression algorithm, redundant color information and high-frequency noise in the visual environment image are removed in the dimension reduction processing, a brightness component and a contrast component reflecting the boundary of a physical barrier are reserved, and a delay error of the synchronous triggering interface is set in a preset nanosecond range.
  3. 3. The method for selecting a millimeter wave mimo system beam based on deep learning according to claim 1, wherein in step 2, the lightweight convolutional neural network model is composed of a plurality of cascaded feature extraction units, and each feature extraction unit comprises a convolutional layer, a linear rectification activation layer and a pooling layer; The convolution layer performs sliding convolution operation in an image space by using a convolution kernel operator with a preset size so as to extract edge, texture and geometric structure characteristics in the image; the semantic segmentation processing identifies physical barriers including building outlines, pedestrian targets and vehicle targets by marking each pixel unit in the image in a category, and generates corresponding barrier masks; the obstacle mask exists in a matrix form, wherein a region with a value approaching a first preset value represents a spatial direction in which the physical blocking probability is greater than a preset probability threshold value, and a region with a value approaching a second preset value represents an open region.
  4. 4. The method for selecting the millimeter wave multiple-input multiple-output system beam based on deep learning according to claim 1, wherein in the step 2, the generating process of the space sparsity constraint matrix comprises the steps of establishing a three-dimensional space mapping model, mapping an obstacle mask obtained by semantic segmentation processing under an angular domain coordinate system, and setting a preset attenuation factor at a corresponding matrix coordinate position according to an azimuth angle range and a pitch angle range occupied by an obstacle in a physical space; Setting a preset enhancement factor at a corresponding spatial angle position for the identified physical characteristics of the reflecting surface; The physical characteristics of the reflecting surface comprise a metal vertical surface, a glass curtain wall and a flat pavement, and corresponding reflection gain weights are distributed in the space sparsity constraint matrix according to a preset electromagnetic wave reflection coefficient table; the dimension of the spatial sparsity constraint matrix is kept logically consistent with the dimension of the beam domain feature map, and is used for indicating signal blocking probabilities or reflection gains in different spatial directions at a physical level.
  5. 5. The method for selecting a millimeter wave mimo system beam based on deep learning according to claim 1, wherein in step 3, the initial channel state information acquisition process is implemented by transmitting orthogonal pilot sequences, and the number of pilot signals is smaller than the number of pilots required for full beam scanning; The transformation processing comprises denoising the initial channel state information, extracting main characteristic components in a channel matrix by adopting a filtering algorithm based on singular value decomposition, and inhibiting background interference noise in a multipath environment; the process of mapping to the beam domain adopts a discrete Fourier transform matrix as a transform basis, and converts an antenna array space sampling signal of a multi-input multi-output system into angular domain components representing different beam orientations; The beam domain feature map represents the radio frequency energy distribution topology in the form of a two-dimensional matrix, wherein an abscissa represents a beam index corresponding to a horizontal azimuth angle, an ordinate represents a beam index corresponding to a vertical elevation angle, and a numerical value in the matrix represents the intensity of received power or gain in the direction of the corresponding beam.
  6. 6. The method for selecting a millimeter wave mimo system beam based on deep learning according to claim 1, wherein in step 4, the cross-modal attention fusion network adopts a feature fusion mechanism based on a transformer architecture, and the transformer architecture comprises a position coding module, a multi-head self-attention module and a feedforward neural network module; The position coding module codes coordinate information in the beam domain feature map in a combination mode of sine function values and cosine function values, and performs addition fusion with input feature vectors, so that a network perceives the relative position relation of the beam in a space angle domain; in the processing process, the beam domain feature map is used as a query vector, and the space sparsity constraint matrix is converted into a key vector and a value vector; The multi-head self-attention module obtains attention weight distribution through calculation of dot product similarity between the query vector and the key vector and scaling and normalization.
  7. 7. The method for selecting a millimeter wave mimo system beam based on deep learning as claimed in claim 6, wherein said clipping and weighting process in step 4 includes performing element-level multiplication of said generated spatial sparsity constraint matrix with said beam domain feature map; When the visual semantic feature extraction result judges that a certain space area is blocked by a physical barrier, the numerical value of the space sparsity constraint matrix at the corresponding coordinate is close to 0, and the feature expression of blocked wave beams in the direction is restrained; When the visual semantic feature extraction result judges that a certain direction is a visual distance visible path or a reflection path exists, setting a numerical value of the space sparsity constraint matrix at a corresponding coordinate as a preset high gain constant, and enhancing the contribution degree of the directional beam features in the deep neural network; the clipping acts on the attention weight distribution as a physical mask through the spatial sparsity constraint matrix, forcing the model to concentrate points of interest on a physically visible signal propagation path.
  8. 8. The method for selecting the millimeter wave mimo system beam based on deep learning according to claim 1, wherein in step 5, the fully connected layer is composed of a multi-layer cascaded neuron structure, each layer of neurons are linearly combined through a preset connection weight matrix, and nonlinear feature transformation is realized through activation function processing; the number of neurons of the final output layer is equal to the total number of candidate beam pairs in the mimo system; the output layer adopts probability normalization function processing to convert the output vector into probability distribution selected by each candidate beam pair, and selects the beam index with the maximum probability value as the final prediction result; the prediction result is sent to a beam forming module of the base station, and the beam forming module calls a pre-stored complex weight vector according to the prediction result, and generates a high-gain narrow beam pointing to target user equipment by adjusting the phase and amplitude of each unit in the antenna array.
  9. 9. The method for selecting a millimeter wave multiple-input multiple-output system beam based on deep learning as set forth in claim 1, further comprising an environment change detection mechanism and dynamic adjustment logic, wherein said baseband processing unit continuously compares said visual environment images in a plurality of consecutive periods, and when it is recognized that a physical environment is changed and the pixel change rate in the images exceeds a preset change threshold, the system automatically increases the weight coefficient of said spatial sparsity constraint matrix in said cross-modal attention fusion network to improve the avoidance capability of newly-appearing obstacles; when the visual environment is in a static stable state, the system increases the weight of the radio frequency characteristics; In a multi-user scene, the baseband processing unit divides a local window in the panoramic image according to the estimated spatial position of each user device, generates a corresponding local spatial sparsity constraint matrix for each user, and guides the beam weight to avoid the angular direction causing the mutual interference among multiple users by introducing an interference suppression factor into the local spatial sparsity constraint matrix.
  10. 10. The millimeter wave multiple-input multiple-output system beam selection method based on deep learning of claim 1, further comprising a model training stage, wherein in the training stage, the lightweight convolutional neural network model and the cross-modal attention fusion network are subjected to joint parameter optimization by utilizing a pre-collected multi-modal data set; the multi-mode data set comprises a matched visual image, radio frequency channel data and an optimal beam tag obtained through full beam scanning; The parameter optimization process adopts a cross entropy loss function as a target optimization function, calculates a gradient value through a back propagation algorithm, and carries out iterative updating on connection weights and bias items in a network by utilizing a self-adaptive moment estimation optimizer until the classification accuracy of the model on a verification set reaches a preset convergence threshold; The model training stage further includes introducing a severe weather training parameter set, and enabling the semantic segmentation process to identify building edge contours in low-contrast images by adding noise textures and blurring effects to the training data.

Description

Millimeter wave MIMO system beam selection method based on deep learning Technical Field The invention belongs to the field of wireless communication, and particularly relates to a millimeter wave MIMO system beam selection method based on deep learning. Background With the evolution of the fifth generation and mobile communication technology, millimeter wave large-scale Multiple Input Multiple Output (MIMO) systems have become a key support for constructing high-speed wireless networks by virtue of their abundant spectrum resources and extremely high transmission rates. The technology compensates extremely high path loss of a high-frequency band by utilizing a narrow beam forming means, and improves signal gain and coverage area through precise directional connection. In a complex communication deployment environment, a beam management and selection mechanism is used as a core link for maintaining the stability of a communication link, and has decisive significance for guaranteeing the throughput of a system and the access quality of a user. The beam selection method based on deep learning utilizes the strong feature extraction and nonlinear mapping capability of the neural network, and aims to mine the optimal beam pairing rules from massive communication features. The direction is used for realizing real-time response to a dynamically-changing channel environment while reducing the traditional pilot scanning overhead by constructing an intelligent decision model. In practical deployment, how to integrate environment priori information into a deep learning architecture, and improve the accuracy and robustness of beam prediction has become an important subject of current research. The existing beam selection scheme mainly relies on channel estimation in a pure radio frequency domain, and when a millimeter wave signal is blocked by signals generated by physical barriers such as pedestrians or buildings, serious environment perception blind areas and information hysteresis exist. The conventional model cannot identify the real-time blocking state of the physical space, so that the system still frequently tries invalid beam connection when encountering the path loss mutation, and extremely high pilot redundancy and access delay are caused. Meanwhile, the prior art lacks the deep deconstructing capability for the multi-mode environment features, is difficult to organically integrate visual space semantics with the radio frequency beam domain features, and cannot dynamically correct the beam weights through a physical constraint mechanism, so that the beam alignment precision in a complex dynamic scene is difficult to meet the requirement of high-reliability communication. Disclosure of Invention The invention aims to provide a millimeter wave MIMO system beam selection method based on deep learning, which solves the problems in the background technology. In order to achieve the purpose, the technical scheme adopted by the invention is that the millimeter wave multiple-input multiple-output system beam selection method based on deep learning comprises the following specific steps: step 1, acquiring environment sensing data and radio frequency original signals, acquiring visual environment images in a communication coverage area by using an auxiliary sensing module arranged at a base station end, and synchronously transmitting a preset number of pilot signals through a baseband processing unit to acquire channel state information in an initial state; Step 2, performing visual semantic feature extraction, inputting an acquired visual environment image into a preset lightweight convolutional neural network model for semantic segmentation, identifying and extracting obstacle mask features and reflecting surface physical features existing in the image, and further converting the environment physical information into a space sparsity constraint matrix; step 3, performing beam domain feature reconstruction, performing transformation processing on the obtained initial channel state information through the baseband processing unit, mapping the channel state information from a space domain to a beam domain, and constructing a beam domain feature map for representing the radio frequency energy distribution feature; Step 4, performing cross-modal attention fusion processing, namely inputting the spatial sparsity constraint matrix and the beam domain feature map into a cross-modal attention fusion network, and weighting and clipping the beam domain feature map by using the spatial sparsity constraint matrix as a physical guide mask so as to inhibit the beam weight in the blocking direction and enhance the beam feature in the visual path direction; And 5, executing beam decision output, inputting the fused cross-mode feature vector into a full-connection layer for classification prediction, and outputting an index number of an optimal beam pair in the current communication environment. Preferably, in the step 1, the auxiliar