Method for reconstructing channel state information of large-scale multiple-input-multiple-output system

Abstract

The invention discloses a channel state information reconstruction method for a large-scale MIMO (Multiple Input Multiple Output) system, which comprises the following steps: S101, each user equipment of the large-scale MIMO communication system receives a local original downlink channel state information matrix and inputs the matrix to a feature encoder to generate a low-dimensional feature; and S102, inputting the low-dimensional features to a decoder head of user equipment, and generating an intermediate representation. And S103, transmitting the intermediate representation to a base station of the large-scale multiple-input multiple-output communication system, receiving the intermediate representation by the base station, and inputting the intermediate representation to the tail part of a decoder for processing, thereby completing information feedback. Compared with the prior art, by combining distributed calculation, federated learning, pipeline parallel optimization and a privacy protection mechanism, the problems of high communication overhead, poor multi-user applicability, insufficient real-time performance and limited privacy protection capability in a large-scale mMIMO system are solved.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication and signal processing technology, and more specifically, to a method for reconstructing channel state information of a large-scale multiple-input multiple-output system. Background Art

[0002] Massive multiple-input multiple-output (mMIMO) technology is one of the core technologies of modern wireless communication systems such as 5G and future 6G networks. Spatial diversity and multiplexing gain can be achieved through large-scale antenna arrays, which can significantly improve spectrum efficiency and communication capacity. However, to achieve these performance advantages, the base station (BTS) needs accurate channel state information to perform precoding and user scheduling.

[0003] In frequency division duplex (FDD) mode, downlink channel state information (CSI) must be uploaded from the user equipment (UE) to the base station through a feedback link. t ) and the number of subcarriers (N c ) increases, the data dimension of the CSI matrix is ​​extremely high, and the traditional feedback method introduces the following problems: 1. Huge communication overhead: The CSI feedback link occupies a large amount of bandwidth, limiting the actual performance of the system. 2. Heavy computational burden: The processing of high-dimensional CSI matrices places strict requirements on the computing resources of base stations and user devices. 3. Insufficient privacy protection: Directly uploading raw CSI data may lead to the leakage of user sensitive information. These problems have spawned many CSI feedback solutions based on compression and distributed processing, but the existing technologies are still insufficient in terms of efficiency, applicability and privacy protection.

[0004] The prior art discloses a codebook-based CSI feedback technology. This solution uses a predefined codebook shared by the base station and the user equipment. The user equipment maps the high-dimensional CSI into a codebook index and uploads the index to the base station to feedback the channel information. The disadvantage is that the codebook length increases linearly with the antenna scale, resulting in a large amount of feedback data in large-scale mMIMO scenarios. The codebook design requires complex optimization, which poses a challenge to the actual implementation of the system. CSI feedback based on compressed sensing (CS) assumes that the CSI matrix is ​​sparse, and realizes compression and reconstruction through sparse representation and reconstruction technology. The disadvantage is that the sparsity assumption does not hold in a complex channel environment, resulting in a decrease in reconstruction performance. Relying on iterative algorithms for reconstruction increases the computational delay. CSI feedback based on deep learning (such as CsiNet, CLNet) uses neural networks to compress and reconstruct the CSI matrix. It usually includes encoder and decoder modules, which are deployed on the UE and BTS ends respectively. The disadvantage is that most of them are designed for single users and are difficult to expand to multi-user scenarios. The existing distributed learning framework requires frequent parameter exchange or model synchronization, and the communication overhead is huge.

[0005] Although the existing technologies have made progress in different aspects, they still have the following key defects when facing actual large-scale mMIMO systems: 1. High communication overhead: Although the codebook and compressed sensing-based methods can reduce the amount of CSI data directly fed back, there is still a problem of a surge in the amount of feedback data when the antenna scale and the number of subcarriers increase. Deep learning and joint learning frameworks require frequent exchange of gradients or model parameters in distributed training, which further increases the communication overhead. 2. Poor user applicability: Most existing methods are designed for single-user scenarios and are difficult to expand in multi-user collaborative environments. Even if they are expanded to multi-user scenarios, the performance is often affected and there is a lack of effective multi-user adaptation mechanisms. 3. Insufficient real-time performance: Compressed sensing methods and some deep learning solutions rely on iterative calculations, resulting in high CSI reconstruction delays, which are difficult to meet the low latency requirements in dynamic channel environments. 4. Limited privacy protection capabilities: Directly uploading raw data or model parameters may leak the privacy information of user devices, especially when it comes to sensitive scenarios such as medical Internet of Things or industrial Internet of Things. This problem is particularly prominent. 5. Insufficient utilization of computing resources: Most solutions do not effectively allocate computing resources for base stations and user devices, resulting in low resource utilization, which limits the feasibility of actual deployment.

[0006] To this end, in combination with the above requirements and the defects of the prior art, the present application proposes a method for reconstructing channel state information of a large-scale multiple-input multiple-output system. Summary of the invention

[0007] The present invention provides a method for reconstructing channel state information of a large-scale multiple-input multiple-output system, aiming to optimize the performance of the mMIMO system in modern wireless communication networks. By combining distributed computing, federated learning, pipeline parallel optimization and privacy protection mechanisms, the high communication overhead, poor multi-user applicability, insufficient real-time performance and limited privacy protection capabilities in large-scale mMIMO systems are solved. Through modular design and a federated learning framework, efficient CSI feedback and reconstruction are achieved, while protecting user privacy and improving the applicability and reconstruction accuracy of the system.

[0008] The primary purpose of the present invention is to solve the above technical problems. The technical solutions of the present invention are as follows:

[0009] A first aspect of the present invention provides a method for feeding back channel state information of a large-scale multiple-input multiple-output system, the method comprising the following steps:

[0010] S101. Each user equipment of a large-scale multiple-input multiple-output communication system receives a local original downlink channel state information matrix, inputs it into a feature encoder, and generates a low-dimensional feature.

[0011] S102: Input the low-dimensional features into a decoder head of a user device to generate an intermediate representation.

[0012] S103, transmitting the intermediate representation to a base station of a large-scale multi-input multi-output communication system, and the base station receives the intermediate representation and inputs it into the tail of the decoder for processing to complete information feedback.

[0013] Furthermore, the large-scale multiple-input multiple-output communication system includes a user equipment end and a base station end, wherein the base station end includes a base station, the user equipment end includes N user equipments, and the base station is equipped with N t transmit antennas to communicate with multiple user equipments, each of which is equipped with a receive antenna and communicates with the base station via N c subcarriers to establish a connection; the user equipment is provided with a modular distributed architecture, including: a feature encoder F n,1 and decoder header F n,3 , the feature encoder F n,1 The local downlink channel state information matrix H nm For feature extraction, the decoder head F n,3 Cooperate with the feature encoder to complete part of the decoding task; the base station is provided with a decoder tail F 2 , used to centrally process low-dimensional features uploaded from multiple user devices.

[0014] Furthermore, the specific process of receiving the local original downlink channel state information matrix, generating low-dimensional features, and generating intermediate representations is as follows: at the feature encoder F located at the user equipment n,1 The high-dimensional downlink channel state information matrix is ​​compressed into a low-dimensional feature representation Z n,1 , the decoder header F of the user equipment n,3 The low-dimensional feature representation Z n,1 Further processing is performed to generate an intermediate representation Z 2 .

[0015] Furthermore, at the user equipment side, the downlink channel state information matrix received by the nth user equipment is expressed as follows:

[0016]

[0017] Through the feature encoder F n,1 , using a convolutional layer or a fully connected layer to transform the high-dimensional downlink channel state information matrix H n Mapped to a low-dimensional feature representation Z n,1 :

[0018] Z n,1 =F n,1 (H n )

[0019] In the decoder header F of the user equipment n,3 Responsible for further processing Zn,1 :

[0020] Z 2 =F n,3 (Z n,1 )

[0021] in is the intermediate representation, c 2 <<c 1 ;

[0022] The decoder tail F at the base station 2 The feature data Z from multiple user devices n,1 Centralized processing is performed, and the signal received by the nth user equipment is:

[0023]

[0024] Where n∈[1,N], Indicates H n The conjugate transpose of W n represents the precoding matrix of the nth user equipment, represents the input data stream of the nth user equipment; represents additive white Gaussian noise, whose components have a mean of zero and a variance of σ 2 The independent and identically distributed complex Gaussian noise represents the random disturbance in the signal reception process.

[0025] A second aspect of the present invention provides a method for reconstructing channel state information of a large-scale multi-input multi-output system, the method being used to reconstruct an intermediate representation obtained by a large-scale multi-input multi-output system channel state information feedback method into a downlink channel state information matrix, comprising the following steps:

[0026] S201. After receiving the intermediate representation, the base station of the massive MIMO communication system inputs it into the tail of the decoder to reconstruct the downlink channel state information matrix.

[0027] S202: Compare the reconstructed downlink channel state information matrix with the original downlink channel state information matrix, and calculate a normalized mean square error between the two.

[0028] S203: The user equipment calculates the gradients of the encoder and decoder heads, and the base station calculates the global gradient of the decoder tail.

[0029] S204, using the Adam optimizer to update the parameters of the encoder, decoder head and decoder tail, decompose the decoder tail into multiple sub-modules, and process each micro-batch in multi-GPU parallel computing.

[0030] S205. Use pipeline parallel output at the end of the decoder and repeat steps S201-S204 until the encoder and decoder converge and the downlink channel state information reconstruction is completed.

[0031] Further, after the base station receives the intermediate representation, the decoder tail F of the base station 2 Receive Z 2 And generate the final reconstructed downlink channel state information matrix:

[0032]

[0033] The m-th reconstructed downlink channel state information matrix of the n-th user equipment is:

[0034]

[0035] in, represents the reconstructed matrix, H n,m is the channel state information matrix of the m-th channel of the n-th user. The process of calculating the normalized mean square error in step S202 is:

[0036]

[0037] Minimize the normalized mean square error through collaborative optimization.

[0038] Furthermore, in step S203, the process of calculating the gradients of the encoder, decoder head and decoder tail is:

[0039]

[0040] in, represents the gradient of the encoder, represents the gradient of the decoder head, represents the global gradient of the decoder tail. In step S204, the parameters of the encoder, decoder head and decoder tail follow the following update rules at each iteration k:

[0041]

[0042] where η>0 is the learning rate.

[0043] Further, in step S205, the pipeline parallel output at the end of the decoder includes the following specific steps: 2 Divide into several sub-modules, each sub-module is assigned to multiple GPUs for parallel processing; the feature data Z uploaded by the user device is n,1 Divide into multiple micro-batches so that each computing node can process in parallel.

[0044] Furthermore, the decoder tail F 2 It is divided into several sub-modules:

[0045] F 2 =f 2,1 °f 2,2 °…°f 2,K

[0046] where f 2,K represents the Kth submodule, each f 2,K is assigned to an independent GPU device, and the feature data Z uploaded by the user device is n,1 The division into multiple micro-batches is as follows:

[0047]

[0048] in, represents the bth micro-batch, B is the number of micro-batches, and each micro-batch Pass through the submodule f at the end of the decoder 2,k For each micro-batch, Follow these steps to process in the pipeline:

[0049]

[0050] in is the output of the first submodule, represents the output of the kth submodule, k>1, is the reconstructed downlink channel state information matrix of the bth micro-batch. The micro-batch results are merged to obtain the final reconstructed downlink channel state information matrix

[0051] The pipeline process loops through the following steps in each iteration:

[0052] According to steps S203 and S204, the model parameters are updated by the user equipment and the base station.

[0053] Re-pass the optimized features to the pipeline to complete the new reconstruction calculation

[0054] Continue iterating until the normalized mean square error of the model converges to the set threshold.

[0055] Time complexity of pipeline parallelism

[0056] The time complexity of pipeline parallelism can be expressed as:

[0057]

[0058] in, is the computation time of the kth submodule, T comm is the communication time of a micro-batch, and B / K represents the number of micro-batches utilized by the submodules in parallel.

[0059] The pipeline mechanism significantly improves the computational efficiency of the decoder tail through parallel processing.

[0060] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0061] The present invention provides a channel state information reconstruction method for a large-scale multi-input multi-output system. By compressing a large-scale CSI matrix into small-volume feature data, the feedback communication volume between a user device and a base station can be significantly reduced. Meanwhile, only the encrypted and crushed compressed feature data is transmitted, and the original CSI data is retained locally in the user device, thereby increasing the privacy protection of user data. By decomposing the decoder tail into multiple sub-modules and reconstructing the information matrix in a pipeline parallel output manner, multiple user devices are supported to participate in distributed training at the same time, thereby improving the scalability of multi-user scenarios, and shortening the training time and decoding delay, thereby meeting the demand for low-latency feedback in a dynamic channel environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 The present invention is a flow chart of a method for feeding back channel state information of a large-scale multiple-input multiple-output system.

[0063] Figure 2 The present invention is a flow chart of a method for reconstructing channel state information of a large-scale multiple-input multiple-output system.

[0064] Figure 3 A schematic diagram of a large-scale multi-input multi-output system in an embodiment of the present invention.

[0065] Figure 4 The figure is a parallel diagram of a pipeline including three working nodes in one embodiment of the present invention.

[0066] Figure 5 The figure is a flow chart of an algorithm in one embodiment of the present invention.

[0067] Figure 6 Schematic diagram of a deep convolutional neural network used as an autoencoder in one embodiment of the present invention.

[0068] Figure 7 This is a result diagram of an encoding dimension of 256 in an internal environment in an embodiment of the present invention.

[0069] Figure 8 This is a diagram showing the result of encoding dimension 512 in an indoor environment in an embodiment of the present invention.

[0070] Fig. 9This is a diagram showing the result of an encoding dimension of 256 in an outdoor environment in an embodiment of the present invention.

[0071] Fig.10 This is a result diagram of an encoding dimension of 512 in an outdoor environment in an embodiment of the present invention. DETAILED DESCRIPTION

[0072] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0073] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0074] Example 1

[0075] like Figure 1 As shown, the present invention provides a large-scale multi-input multi-output system channel state information feedback method, the method comprising the following steps:

[0076] S101. Each user equipment of a large-scale multiple-input multiple-output communication system receives a local original downlink channel state information matrix, inputs it into a feature encoder, and generates a low-dimensional feature.

[0077] S102: Input the low-dimensional features into a decoder head of a user device to generate an intermediate representation.

[0078] S103, transmitting the intermediate representation to a base station of a large-scale multi-input multi-output communication system, and the base station receives the intermediate representation and inputs it into the tail of the decoder for processing to complete information feedback.

[0079] like Figure 2 As shown, based on the above intermediate representation, the present invention also provides a method for reconstructing channel state information of a large-scale multi-input multi-output system, which is used to reconstruct the intermediate representation obtained by the large-scale multi-input multi-output system channel state information feedback method into a downlink channel state information matrix, including the following steps:

[0080] S201. After receiving the intermediate representation, the base station of the massive MIMO communication system inputs it into the tail of the decoder to reconstruct the downlink channel state information matrix.

[0081] S202: Compare the reconstructed downlink channel state information matrix with the original downlink channel state information matrix, and calculate a normalized mean square error between the two.

[0082] S203: The user equipment calculates the gradients of the encoder and decoder heads, and the base station calculates the global gradient of the decoder tail.

[0083] S204, using the Adam optimizer to update the parameters of the encoder, decoder head and decoder tail, decompose the decoder tail into multiple sub-modules, and process each micro-batch in multi-GPU parallel computing.

[0084] S205. Use pipeline parallel output at the end of the decoder and repeat steps S201-S204 until the encoder and decoder converge and the downlink channel state information reconstruction is completed.

[0085] According to the above technical features, the present invention compresses large-scale CSI matrices into small-volume feature data through modular design and distributed segmentation algorithm, significantly reducing the feedback communication volume between user equipment and base stations; through joint learning optimization algorithm, it supports multiple user equipment to participate in distributed training at the same time, improves the scalability of multi-user scenarios, and can use pipeline parallel mechanism to shorten training time and decoding delay, meeting the demand for low-latency feedback in dynamic channel environment; by retaining the original CSI data in the local user equipment, it avoids the leakage of unprocessed data, and only transmits encrypted and crushed compressed feature data, providing a safe and reliable feedback solution. In actual application scenarios, it is suitable for high-density user scenarios (such as stadiums, exhibition centers), industrial Internet of Things, medical Internet of Things and autonomous driving, solving the contradiction between communication overhead and privacy protection, and promoting the development of wireless communication technology to a wider range of application fields.

[0086] It should be noted that, compared with the prior art, the present invention has the following technical advantages:

[0087] 1. Significantly reduce communication overhead

[0088] Through modular distributed architecture and strict control of sub-model boundary feature dimensions, the amount of data transmission between user equipment (UE) and base station (BTS) is significantly reduced.

[0089] Compared with the traditional direct feedback method, the present invention can significantly reduce the feedback overhead in the frequency division duplex (FDD) mode while maintaining the reconstruction accuracy.

[0090] In multi-user scenarios, the communication efficiency of the system is further improved through data aggregation and distribution optimization strategies.

[0091] 2. Improve system training efficiency

[0092] Pipeline parallel technology is introduced at the tail end of the decoder on the base station side, which decomposes the tasks into micro-batches and distributes them to multiple GPUs or computing nodes for parallel processing, significantly reducing the wall clock time required for training.

[0093] Through the momentum and adaptive learning rate mechanism of the Adam optimizer, it quickly converges to the optimal model, reduces the number of training iterations, and improves the stability of training.

[0094] 3. Enhance data privacy protection

[0095] The user equipment holds the local CSI matrix and only uploads the encoded compressed features, avoiding the direct leakage of the original data.

[0096] It adopts a distributed update mechanism of boundary features and gradients to ensure data privacy and security. It is suitable for application scenarios with high requirements for privacy protection, such as medical Internet of Things and industrial Internet of Things.

[0097] 4. Improve system applicability

[0098] The present invention is suitable for high-density multi-user scenarios, supports multiple user devices to participate in training at the same time, and realizes collaborative optimization through a federated learning framework.

[0099] It can be widely used in scenarios such as smart cities, industrial manufacturing, and autonomous driving in 5G and future 6G communication networks, providing support for efficient communication.

[0100] 5. Improve reconstruction accuracy

[0101] Through innovative distributed design and efficient gradient calculation, the normalized mean square error is optimized and accurate reconstruction of the CSI matrix is ​​achieved.

[0102] Excellent reconstruction performance can be maintained even in large-scale antenna arrays and complex multi-user environments.

[0103] Example 2

[0104] Based on the above embodiment 1, combined with Figure 1-Figure 6 , this embodiment describes the specific contents of the present invention in detail.

[0105] like Figure 3 and Figure 4 As shown, the large-scale multiple-input multiple-output communication system includes a user equipment end and a base station end, wherein the base station end includes a base station, the user equipment end includes N user equipments, and the base station is equipped with N t transmit antennas to communicate with multiple user equipments, each of which is equipped with a receive antenna and communicates with the base station via N c subcarriers to establish a connection; the user equipment is provided with a modular distributed architecture, including: a feature encoder F n,1 and decoder header F n,3 , the feature encoder F n,1 The local downlink channel state information matrix H nmFor feature extraction, the decoder head F n,3 Cooperate with the feature encoder to complete part of the decoding task; the base station is provided with a decoder tail F 2 , used to centrally process low-dimensional features uploaded from multiple user devices.

[0106] The specific process of receiving the local original downlink channel state information matrix, generating low-dimensional features, and generating intermediate representations is as follows: at the feature encoder F located at the user equipment n,1 The high-dimensional downlink channel state information matrix is ​​compressed into a low-dimensional feature representation Z n,1 , the decoder header F of the user equipment n,3 The low-dimensional feature representation Z n,1 Further processing is performed to generate an intermediate representation Z 2 .

[0107] The following key variables are defined in this system:

[0108] At the user equipment side, the downlink channel state information matrix received by the nth user equipment is expressed as follows:

[0109]

[0110] Through the feature encoder F n,1 , using a convolutional layer or a fully connected layer to transform the high-dimensional downlink channel state information matrix H n Mapped to a low-dimensional feature representation Z n,1 :

[0111] Z n,1 =F n,1 (H n )

[0112] In the decoder header F of the user equipment n,3 Responsible for further processing Z n,1 :

[0113] Z 2 =F n,3 (Z n,1 )

[0114] in is the intermediate representation, c 2 <<c 1 ;

[0115] The decoder tail F at the base station 2 The feature data Z from multiple user devices n,1 Centralized processing is performed, and the signal received by the nth user equipment is:

[0116]

[0117] Where n∈[1,N], Indicates H n The conjugate transpose of W n represents the precoding matrix of the nth user equipment, represents the input data stream of the nth user equipment; represents additive white Gaussian noise, whose components have a mean of zero and a variance of σ 2 The independent and identically distributed complex Gaussian noise represents the random disturbance in the signal reception process.

[0118] After the base station receives the intermediate representation, the base station's decoder tail F 2 Receive Z 2 And generate the final reconstructed downlink channel state information matrix:

[0119]

[0120] The m-th reconstructed downlink channel state information matrix of the n-th user equipment is:

[0121]

[0122] in, represents the reconstructed matrix, H n,m is the channel state information matrix of the m-th channel of the n-th user. The process of calculating the normalized mean square error in step S202 is:

[0123]

[0124] Through collaborative optimization, the normalized mean square error is minimized. In step S203, the process of calculating the gradients of the encoder, decoder head and decoder tail is:

[0125]

[0126] in, represents the gradient of the encoder, represents the gradient of the decoder head, represents the global gradient of the decoder tail. In step S204, the parameters of the encoder, decoder head and decoder tail follow the following update rules at each iteration k:

[0127]

[0128] where η>0 is the learning rate.

[0129] like Figure 4 As shown, the pipeline parallel output at the end of the decoder includes the following specific steps:

[0130] Data aggregation: low-dimensional features Z uploaded by user devices n,1 The base station integrates the low-dimensional feature data from multiple user devices into an intermediate representation Z 2 The initial part of

[0131] Modular decomposition of the decoder tail, decoder tail F 2 Decomposed into several submodules f 2,k , each submodule completes a specific decoding task:

[0132] F 2 =f 2,1 °f 2,2 °…°f 2,k

[0133] f 2,k represents the kth submodule, K is the total number of submodules, and each submodule is assigned to an independent GPU (such as GPU1, GPU2, GPU3, etc. in the figure).

[0134] It should be noted that this partitioning allows the decoding process to be performed in a pipelined manner, reducing the computational burden of each GPU and improving the overall throughput.

[0135] Micro-batch division: feature data Z uploaded by user devices 2 Divided into B micro-batches:

[0136]

[0137] in, represents the b-th micro-batch, and B is the number of micro-batches.

[0138] Pipeline processing flow: For each micro-batch The calculation process of pipeline processing is as follows:

[0139] 1) The first submodule processes:

[0140]

[0141] in It is the output of the first submodule.

[0142] 2) Processing of the kth submodule (k>1):

[0143]

[0144] The output of the k-1th submodule Passed to k+1 submodules.

[0145] 3) The last submodule outputs:

[0146]

[0147] in is the reconstructed CSI matrix of the b-th mini-batch.

[0148] Distribution of reconstruction results: reconstruction results of all micro-batches are integrated to generate the completed reconstructed CSI matrix:

[0149]

[0150] Reconstruction results Distributed back to each user device for final use.

[0151] Repeat the iteration until convergence: Through the feedback regression mechanism, the encoder F of the model n,1 、Decoder Head F n,3 、F at the end 2 The parameters are continuously optimized.

[0152] Gradient calculation and update rules:

[0153]

[0154] The optimization process continues until the normalized mean square error (NMSE) of the model converges to a preset threshold.

[0155] In this embodiment, the CSI feedback model is divided into the following three modules through a distributed architecture:

[0156] Feature encoder (located on the user device side, F n,1 ): Compress the high-dimensional CSI matrix into a low-dimensional feature representation Z n,1 , only the compressed feature data is uploaded to the base station.

[0157] Decoder tail (located at the base station, F 2 ): Feature data Z from multiple user devices n,1 Perform centralized processing to generate an intermediate representation Z 2 .

[0158] Decoder head (located at the user device end, F n,3 ):The user equipment receives the intermediate representation Z 2 , complete the final CSI matrix reconstruction.

[0159] In this embodiment, in step S205, the pipeline parallel output at the end of the decoder includes the following specific steps: 2 It is divided into several sub-modules, each of which is assigned to multiple GPUs for parallel processing; the feature data Z uploaded by the user device is n,1Divide into multiple micro-batches so that each computing node can process in parallel.

[0160] The decoder tail F 2 It is divided into several sub-modules:

[0161] F 2 =f 2,1 °f 2,2 °…°f 2,K

[0162] where f 2,K represents the Kth submodule, each f 2,K is assigned to an independent GPU device, and the feature data Z uploaded by the user device is n,1 The division into multiple micro-batches is as follows:

[0163]

[0164] in, represents the bth micro-batch, B is the number of micro-batches, and each micro-batch Pass through the submodule f at the end of the decoder 2,k For each micro-batch, Follow these steps to process in the pipeline:

[0165]

[0166] in is the output of the first submodule, represents the output of the kth submodule, k>1, is the reconstructed downlink channel state information matrix of the bth micro-batch. The micro-batch results are merged to obtain the final reconstructed downlink channel state information matrix

[0167] The pipeline process loops through the following steps in each iteration:

[0168] According to steps S203 and S204, the model parameters are updated by the user equipment and the base station.

[0169] Re-pass the optimized features to the pipeline to complete the new reconstruction calculation

[0170] Continue iterating until the normalized mean square error of the model converges to the set threshold.

[0171] Time complexity of pipeline parallelism

[0172] The time complexity of pipeline parallelism can be expressed as:

[0173]

[0174] in, is the computation time of the kth submodule, T comm is the communication time of a micro-batch, and B / K represents the number of micro-batches utilized by the submodules in parallel.

[0175] The pipeline mechanism significantly improves the computational efficiency of the decoder tail through parallel processing.

[0176] To support multi-user collaborative training, the present invention adopts a federated learning framework so that user devices do not need to share original data, but participate in training through local gradient calculation and update: each user device calculates the gradient of the encoder and decoder head based on local data, and the base station calculates the global gradient of the decoder tail. Efficient compression and accurate reconstruction are achieved by minimizing the normalized mean square error (NMSE). The Adam optimizer is used to dynamically adjust the learning rate, which significantly improves the model convergence speed and training stability.

[0177] In order to optimize the computing performance of the base station, the present invention introduces pipeline parallel technology at the end of the decoder: the end of the decoder is divided into multiple sub-modules F 2 =f 2,4 °f 2,3 °f 2,2 °f 2,1 , deployed on multiple computing nodes respectively. The uploaded feature data Z n,1 Divide into multiple micro-batches to enable each computing node to process in parallel and improve resource utilization. The parallel output of the pipeline is dynamically aggregated to ensure that the output Z at the end of the decoder 2 Maintain integrity.

[0178] In terms of training strategy, the present invention adopts forward propagation and back propagation strategies to solve the problems of unstable gradient update and low training efficiency in the traditional single propagation strategy. n,1 The input CSI matrix H n,m Compressed into low-dimensional features Z n,1 The decoder tail F 2 For feature Z from all user devices n,1 Further processing is performed to generate an intermediate representation Z 2 The decoder header F n,3 The intermediate representation Z returned by the base station 2 Convert to the reconstructed CSI matrix Therefore, in the case of N user devices, the total number of forward propagations per iteration is 2N+1.

[0179] During the back-propagation process, the encoder F n,1 According to the gradient calculation of the loss function, from the decoder tail F 2The gradient is back-propagated to the encoder; the decoder tail F 2 Calculate the global gradient of the decoder tail based on the loss function of all user devices The decoder header F n,3 Compute the gradient of the decoder head according to its reconstruction loss Therefore, in the scenario of N user devices, the total number of back propagation per iteration is 2N+1.

[0180] In order to reduce the communication overhead between the base station and the user equipment, the feature dimension c at the boundary of the submodel 1 and c 2 Strictly controlled:

[0181] c 1 <<min{d 1 ,d 2}

[0182] c 2 <<min{d 1 ,d 3}

[0183] Among them, d 1 Represents the feature encoder F n,1 The parameter dimension, d 2 Indicates the decoder header F 2 The parameter dimension, d 3 Denotes the decoder tail F n,3 The parameter dimension of .

[0184] Example 3

[0185] Based on the above-mentioned embodiment 1 and embodiment 2, this embodiment combines Figure 5-Figure 10 The specific implementation process of the present invention is described in detail.

[0186] like Figure 5 As shown, the specific steps are as follows:

[0187] Step 1: User Equipment (UE) side coding-forward propagation process includes each UE receiving the local original CSI matrix H n,m , H n,m Input feature encoder F n,1 , generate low-dimensional features Z n,1 .

[0188] Step 2: The centralized processing at the base station (BTS) side includes the user equipment sending feature Z n,1 Upload to the base station, the base station receives all the features uploaded by users Z n,1 , input decoder tail F 2 . Centrally process the feature data and generate the intermediate representation Z 2 .

[0189] Step 3: User equipment side decoding - forward propagation process includes the base station converting the intermediate result Z 2 Distributed to each user device. The user device receives Z 2 , input decoder head F n,3 , reconstruct the CSI matrix

[0190] Step 4: Calculate the loss function: Compare the reconstructed CSI matrix With the real CSI matrix H n,m , calculate the normalized mean square error (NMSE).

[0191] Step 5: Gradient calculation on the user device side - The back propagation process involves calculating the encoder F based on the gradients returned by the decoder tail. n,1 Gradient According to the reconstruction loss, the decoder head F is calculated n,3 Gradient

[0192] Step 6: Gradient calculation at the base station side - The back propagation process involves the base station calculating the decoder tail F according to the loss function of all user devices. 2 The global gradient

[0193] Step 7: The parameter update process involves updating the encoder F using the Adam optimizer n,1 and decoder header F n,3 The parameters of the decoder tail F are updated using the Adam optimizer 2 Parameters.

[0194] Step 8: The pipeline parallelization and communication optimization process includes the base station transferring the decoder tail F 2 Decomposed into multiple submodules F 2 =f 2,4 °f 2,3 °f 2,2 °f 2,1 , process each micro-batch in multi-GPU parallel computing to improve computing efficiency. Strictly control the feature dimension c 1 and c 2 Significantly reduce the amount of data transmission between UE and BTS.

[0195] Step 9: Repeat the above process until the model converges and completes efficient CSI compression and accurate reconstruction.

[0196] According to the above technical features, this embodiment is based on Figure 6The convolutional neural network (CNN) of the autoencoder shown in FIG. reconstructs the downlink channel state information (CSI) matrix. The network structure of the convolutional neural network consists of an encoder, a decoder tail, and a decoder head. In order to ensure that the size of each convolutional layer output is the same as the size of the CSI matrix N t ×N t To match, an appropriate padding strategy is used.

[0197] To ensure the fairness of the performance comparison, we use the same dataset as CsiNet (Channel State Information Compression and Reconstruction Model Based on Deep Learning). The dataset contains a training set and a test set, where there are 100,000 original samples and 30,000 validation samples in the training set; there are 20,000 samples in the test set. The size of each data sample is 32×32. The mMIMO system is configured with 5 user devices. Training iterations: 20,000. Mini-Batch size: 800. Learning rate: 5×105. The momentum factor of the Adam optimizer is set to (0.9, 0.9).

[0198] To achieve the parallelization of the pipeline of two micro-batches, the decoder tail F 2 It is decomposed into two sub-models, and the split point is located at the connection of two CRBlocks (Convolutional Residual Blocks). Each sub-model is assigned to an independent GPU accelerator for calculation. The specific experimental settings are as follows:

[0199] Baseline method:

[0200] In the “no pipeline parallelism” baseline approach, the entire decoder tail is deployed on a single GPU accelerator.

[0201] Experimental setup:

[0202] Encoding dimension c 1 and c 2 They are set to 256 and 512 respectively. Using indoor and outdoor CSI data, the wall clock time of the “pipeline parallel” and “no pipeline parallel” methods is compared.

[0203] Experimental results (see Table 1):

[0204] like Figure 7 As shown in Figure 2, when the Mini-Batch size of each user device is 400 and the encoding dimension is 256, the wall clock time of indoor data in training 20,000 iterations is reduced by 4.11%, as shown in Figure 2. Fig. 9As shown in Figure 2, the wall clock time of outdoor data is reduced by 3.74% after 20,000 training iterations. When the Mini-Batch size is increased to 800 and the encoding dimension is still 256, the wall clock time of indoor data is reduced by 21.85%, and that of outdoor data is reduced by 21.98%.

[0205] like Figure 8 and Fig.10 As shown, similar performance improvement is also observed when the encoding dimension is 512.

[0206] Table 1 CSILOCAL algorithm results

[0207]

[0208]

[0209] To demonstrate the advantage of our proposed CSILocal algorithm in terms of communication efficiency, we compare its convergence performance with two baseline algorithms, the Federated Averaging Algorithm (FedAvg) and the Federated Averaging with Personalization (FedAvgPer). Figure 7-10 The following results can be observed:

[0210] like Figure 7 As shown in the figure, when the coding dimension is 256 in an indoor environment, CSILocal shows the lowest communication overhead, and only requires very few parameter exchanges to achieve the same test loss as FedAvg and FedAvgPer (such as 10-4). FedAvg and FedAvgPer require significantly more parameter exchanges and have greater communication overhead. In the figure, the x-axis represents the number of exchanged parameters, and the y-axis represents the test loss. The CSILocal curve shifts significantly to the left, indicating higher communication efficiency.

[0211] like Figure 8 As shown in the figure, when the coding dimension is 512 in an indoor environment, the increase in coding dimension increases the overall communication overhead compared to the case of coding dimension 256. CSILocal still maintains the lowest communication overhead and can achieve low test loss (such as 10-4) with fewer parameter exchanges. The curves of FedAvg and FedAvgPer shift to the right, indicating a greater communication overhead requirement. The number of exchanged parameters of CSILocal on the X-axis is significantly lower than that of other algorithms, verifying its superiority under high-dimensional feature data.

[0212] like Fig. 9As shown in the figure, when the coding dimension is 256 in the outdoor environment, CSILocal still shows the best communication efficiency in the outdoor environment. When the test loss (such as 10-3) is reached, the number of exchange parameters required by CSILocal is much lower than FedAvg and FedAvgPer. Compared with the indoor scene, the communication overhead in the outdoor scene is lower overall, but the efficiency advantage of CSILocal is still very significant. Compared with the indoor environment, the test loss curve is more concentrated, but CSILocal still has leading communication efficiency.

[0213] like Fig.10 As shown in the figure, when the coding dimension is 512 in an outdoor environment, the increase in coding dimension leads to an increase in communication overhead, but CSILocal still maintains the lowest parameter exchange requirement. FedAvg and FedAvgPer require significantly higher numbers of exchanged parameters to achieve the same test loss. In outdoor environments, CSILocal shows excellent scalability and can effectively reduce communication costs even at higher coding dimensions. CSILocal's curve position continues to lead other algorithms, with lower test losses and significantly reduced communication overhead.

[0214] In the embodiments provided in the present application, it should be understood that the above-mentioned embodiments of the present invention can also be stored in a computer-readable storage medium if they are implemented in the form of software function modules and sold or used as independent products. Based on this understanding, the technical solution of the embodiments of the present invention can essentially or partly contribute to the prior art in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media capable of storing program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0215] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. The icons describing the structural positional relationships in the accompanying drawings are only used for exemplary purposes and are not to be construed as limiting the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and is not possible to list all the embodiments here. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the claims of the present invention.

Claims

1. A method for feedback of channel state information in a large-scale multiple-input multiple-output system, characterized in that: The following steps are involved: S101, each user equipment of the massive multiple-input multiple-output communication system receives a local original downlink channel state information matrix, inputs it into a feature encoder, and generates a low-dimensional feature; S102, inputting the low-dimensional features into a decoder head of a user device to generate an intermediate representation; S103, transmitting the intermediate representation to a base station of a large-scale multi-input multi-output communication system, and the base station receives the intermediate representation and inputs it into the tail of the decoder for processing to complete information feedback.

2. The method for feedback of channel state information of a large-scale multiple-input multiple-output system according to claim 1, characterized in that: The large-scale multiple-input multiple-output communication system includes a user equipment end and a base station end, wherein the base station end includes a base station, the user equipment end includes N user equipments, and the base station is equipped with N t transmit antennas to communicate with multiple user equipments, each of which is equipped with a receive antenna and communicates with the base station via N c subcarriers to establish a connection; the user equipment is provided with a modular distributed architecture, Includes: Feature encoder F n,1 and decoder header F n,3 , the feature encoder F n,1 The local downlink channel state information matrix H nm For feature extraction, the decoder head F n,3 Cooperate with the feature encoder to complete part of the decoding task; the base station is provided with a decoder tail F2 for centrally processing low-dimensional features uploaded from multiple user devices.

3. The method for feedback of channel state information of a large-scale multiple-input multiple-output system according to claim 2, characterized in that: The specific process of receiving the local original downlink channel state information matrix, generating low-dimensional features, and generating intermediate representations is as follows: at the feature encoder F located at the user equipment n,1 The high-dimensional downlink channel state information matrix is ​​compressed into a low-dimensional feature representation Z n,1 , the decoder header F of the user equipment n,3 The low-dimensional feature representation Z n,1 Further processing is performed to generate an intermediate representation Z2.

4. The method for feedback of channel state information of a large-scale multiple-input multiple-output system according to claim 3, characterized in that: At the user equipment side, the downlink channel state information matrix received by the nth user equipment is expressed as follows: Through the feature encoder F n,1 , using a convolutional layer or a fully connected layer to transform the high-dimensional downlink channel state information matrix H n Mapped to a low-dimensional feature representation Z n,1 : Z n,1 =F n,1 (H n ) In the decoder header F of the user equipment n,3 Responsible for further processing Z n,1 : in is the intermediate representation, c2< <c1; The decoder tail F2 at the base station processes the feature data Z from multiple user equipments. n,1 Centralized processing is performed, and the signal received by the nth user equipment is: Where n∈[1,N], Indicates H n The conjugate transpose of W n represents the precoding matrix of the nth user equipment, represents the input data stream of the nth user equipment; represents additive white Gaussian noise, whose components have a mean of zero and a variance of σ 2 The independent and identically distributed complex Gaussian noise represents the random disturbance in the signal reception process.

5. A method for reconstructing channel state information of a large-scale multiple-input multiple-output system, the method for reconstructing the intermediate representation obtained by the method for feedback of channel state information of a large-scale multiple-input multiple-output system according to any one of claims 1 to 4 into a downlink channel state information matrix, characterized in that: The following steps are involved: S201, after receiving the intermediate representation, the base station of the massive MIMO communication system inputs it into the tail of the decoder to reconstruct the downlink channel state information matrix; S202, comparing the reconstructed downlink channel state information matrix and the original downlink channel state information matrix, and calculating a normalized mean square error between the two; S203, the user equipment calculates the gradients of the encoder and decoder heads, and the base station calculates the global gradient of the decoder tail; S204, using the Adam optimizer to update the parameters of the encoder, decoder head and decoder tail, decomposing the decoder tail into multiple sub-modules, and processing each micro-batch in multi-GPU parallel computing; S205. Use pipeline parallel output at the end of the decoder and repeat steps S201-S204 until the encoder and decoder converge and the downlink channel state information reconstruction is completed.

6. A method for reconstructing channel state information of a large-scale multiple-input multiple-output system according to claim 5, characterized in that: After the base station receives the intermediate representation, the decoder tail F2 of the base station receives Z2 and generates the final reconstructed downlink channel state information matrix: The m-th reconstructed downlink channel state information matrix of the n-th user equipment is: in, represents the reconstructed matrix, H n,m is the channel state information matrix of the m-th channel of the n-th user. The process of calculating the normalized mean square error in step S202 is: Minimize the normalized mean square error through collaborative optimization.

7. A method for reconstructing channel state information of a large-scale multiple-input multiple-output system according to claim 6, characterized in that: In step S203, the process of calculating the gradients of the encoder, decoder head and decoder tail is: in, represents the gradient of the encoder, represents the gradient of the decoder head, represents the global gradient at the end of the decoder.

8. The method for reconstructing channel state information of a large-scale multiple-input multiple-output system according to claim 7, characterized in that: In step S204, the parameters of the encoder, decoder header and decoder tail follow the following update rules at each iteration k: where η>0 is the learning rate.

9. A method for reconstructing channel state information of a large-scale multiple-input multiple-output system according to claim 8, characterized in that: In step S205, the pipeline parallel output at the decoder tail includes the following specific steps: dividing the decoder tail F2 into several submodules, each submodule is assigned to multiple GPUs for parallel processing; n,1 Divide into multiple micro-batches so that each computing node can process in parallel.

10. A method for reconstructing channel state information of a large-scale multiple-input multiple-output system according to claim 9, characterized in that: The decoder tail F2 is divided into several sub-modules: F2=f 2,1 °f 2,2 °…°f 2,K where f 2,K represents the Kth submodule, each f 2,K is assigned to an independent GPU device, and the feature data Z uploaded by the user device is n,1 The division into multiple micro-batches is as follows: in, represents the bth micro-batch, B is the number of micro-batches, and each micro-batch Pass through the submodule f at the end of the decoder 2,k For each micro-batch, Follow these steps to process in the pipeline: in is the output of the first submodule, represents the output of the kth submodule, k>1, is the reconstructed downlink channel state information matrix of the bth micro-batch. The micro-batch results are merged to obtain the final reconstructed downlink channel state information matrix

Images