Wireless base model training method and device for channel estimation and channel prediction, computer equipment and readable storage medium

Abstract

The invention relates to a training method and device of a wireless base model for channel estimation and channel prediction, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: acquiring sample channel state information of a base station side; performing feature extraction processing on the sample channel state information through a feature dimension conversion network and a task corresponding to the base model to obtain a target token; and encoding the target token through an expert model in the position encoding module and each expert joint encoding module and a gating network corresponding to the task to obtain a first output token, and processing the first output token and a preset token to be learned through a decoder to obtain reconstructed channel state information and a wireless base model. By adopting the method, the adaptive capacity of the base model on different CSI data sets can be improved, the conflict of different pre-training tasks is solved, and the reasoning performance and the processing precision of the base model in processing multiple tasks are improved.

Description

Technical Field

[0001] The present application relates to the field of wireless communication technology, and in particular to a training method, apparatus, computer equipment, computer-readable storage medium, and computer program product for a wireless base model for channel estimation and channel prediction. Background Art

[0002] Massive Multiple-Input Multiple-Output (mMIMO) and Orthogonal Frequency-Division Multiplexing (OFDM) are two fundamental technologies widely adopted in fifth-generation mobile communication systems (5G). In MIMO-OFDM systems, accurate channel state information (CSI) is crucial for tasks such as precoding, beamforming, power allocation, modulation selection, and transmit antenna selection. Channel estimation and channel prediction are the two basic methods for obtaining CSI. The former generally estimates CSI from received pilot signals, while the latter predicts unknown CSI in the time and frequency domains based on known partial CSI.

[0003] Related channel estimation and channel prediction employ parameterized model schemes, such as the least squares (LS) method and minimum mean squared error (MMSE). However, the performance of these parameterized model schemes is heavily dependent on the accuracy of the channel model. When the prior assumptions do not match the actual channel, performance is severely impaired. Deep learning technology, due to its strong nonlinear fitting capabilities, has been widely used in channel estimation and prediction. With the development of deep learning technology, foundation models have emerged, which have achieved great success in fields such as natural language processing. These models, which employ neural networks with large parameters, can demonstrate excellent few-shot learning capabilities in downstream tasks after self-supervised pre-training on large datasets. The development of foundation models for various tasks, such as channel estimation and prediction, is a pressing technical challenge. Summary of the Invention

[0004] Based on this, it is necessary to provide a training method, device, computer equipment, computer-readable storage medium and computer program product for a wireless base model for channel estimation and channel prediction that can improve the versatility of the base model in order to address the above technical problems.

[0005] In a first aspect, the present application provides a training method for a wireless base model for channel estimation and channel prediction, wherein the base model includes at least a feature dimension conversion network, an encoder, and a decoder, wherein the encoder includes a position encoding module and at least one expert joint encoding module; the method includes:

[0006] Acquire sample channel state information on the base station side, where the sample channel state information includes information of target dimensions, where the target dimensions include a time dimension, a space dimension, and a frequency dimension;

[0007] Performing feature extraction processing on the sample channel state information through the feature dimension conversion network and the tasks corresponding to the base model to obtain a target token;

[0008] Encoding the target token through the position encoding module, the expert models in each of the expert joint encoding modules, and the gating network corresponding to the task to obtain a first output token;

[0009] Based on the task and at least one expert joint decoding module contained in the decoder, the first output token and the preset token to be learned are processed to obtain reconstructed channel state information corresponding to the sample channel state information, and the base model is processed based on the reconstructed channel state information to obtain a trained base model.

[0010] In one embodiment, encoding the target token by the position encoding module, the expert models in each of the expert joint encoding modules, and the gating network corresponding to the task to obtain the first output token includes:

[0011] Position-encode the target token using a position encoding module to obtain a first position-encoded token;

[0012] The first position coding token is encoded by using multiple expert models in each of the expert joint coding modules connected end to end and a gating network corresponding to the task to obtain a first output token.

[0013] In one embodiment, the expert joint encoding module includes a pre-processing layer and an expert processing layer, the expert processing layer includes multiple expert models and multiple gating networks; encoding the first position coding token through the multiple expert models in each of the expert joint encoding modules connected end to end and the gating network corresponding to the task to obtain the first output token includes:

[0014] For the i-th expert joint encoding module, normalize and process the input token of the i-th expert joint encoding module through the pre-processing layer in the i-th expert joint encoding module to obtain a first token;

[0015] Based on each expert model in the i-th expert joint encoding module, encode the first token respectively to obtain the expert output result corresponding to each expert model; and based on the gating network corresponding to the task and the input token corresponding to the i-th expert joint encoding module, process the first token to obtain the weight corresponding to each expert model;

[0016] Based on the expert output results of each expert model and the weights corresponding to each expert model, obtaining the output token of the i-th expert joint encoding module;

[0017] Determine that the output token of the expert joint coding module located at the last position is the first output token; the input token of the expert joint coding module located at the initial position is the first position coding token, and the input token of the i-th expert joint coding module is the output token of the i-1-th expert joint coding module.

[0018] In one embodiment, the encoder further includes a position encoding module; the processing of the first output token and a preset token to be learned by the task and at least one expert joint decoding module included in the decoder to obtain reconstructed channel state information corresponding to the sample channel state information includes:

[0019] Based on the token splicing strategy corresponding to the task, the first output token and the preset token to be learned are spliced ​​to obtain a spliced ​​token;

[0020] Position-encoding the spliced ​​token by the position encoding module to obtain a second position encoding token;

[0021] The second position coding token is decoded by multiple expert models in each of the expert joint decoding modules connected end to end and the gating network corresponding to the task to obtain a second output token, and based on the second output token, the reconstructed channel state information corresponding to the sample channel state information is obtained.

[0022] In one embodiment, the base model further includes an output module, and obtaining the reconstructed channel state information corresponding to the sample channel state information based on the second output token includes:

[0023] The second output token is reconstructed through the fully connected layer in the output module to obtain reconstructed channel state information corresponding to the sample channel state information.

[0024] In one embodiment, the time dimension includes the number of time sampling points, the space dimension includes the number of antennas of the base station, and the frequency dimension includes the number of subcarriers. The feature dimension conversion network and the task corresponding to the base model are used to perform feature extraction processing on the sample channel state information to obtain a target token, including:

[0025] performing block processing on the sample channel state information to obtain a plurality of sample channel state blocks;

[0026] Perform feature extraction processing on each sample channel state block through a feature dimension conversion network to obtain multiple unit tokens;

[0027] Based on the task, a plurality of unit tokens are screened to obtain a target token.

[0028] In one embodiment, the processing the base model based on the reconstructed channel state information to obtain a trained base model includes:

[0029] Performing loss prediction based on the sample channel state information and reconstructed channel state information corresponding to the sample channel state information to obtain a channel reconstruction loss; and determining a token routing ratio based on the sample channel state information and the target token;

[0030] Obtaining a task load balancing loss based on the tasks corresponding to the base model, the token routing ratio, and the average weight corresponding to the gating network;

[0031] Obtaining a model training loss based on the channel reconstruction loss, a preset weighting coefficient, and the task load balancing loss;

[0032] The model parameters of the base model are updated through the model training loss until the preset training completion conditions are met to obtain a trained base model.

[0033] In a second aspect, the present application further provides a training device for a wireless base model for channel estimation and channel prediction, wherein the base model includes at least a feature dimension conversion network, an encoder, and a decoder, wherein the encoder includes a position encoding module and at least one expert joint encoding module; the device includes:

[0034] A first acquisition module is configured to acquire sample channel state information from a base station, where the sample channel state information includes information of a target dimension, and the target dimension includes a time dimension, a space dimension, and a frequency dimension;

[0035] A first processing module is configured to perform feature extraction processing on the sample channel state information through the feature dimension conversion network and the tasks corresponding to the base model to obtain a target token;

[0036] A first encoding module is configured to encode the target token using the position encoding module, the expert models in the expert joint encoding modules, and the gating network corresponding to the task to obtain a first output token;

[0037] The second processing module is used to process the first output token and the preset token to be learned based on the task and at least one expert joint decoding module included in the decoder, obtain reconstructed channel state information corresponding to the sample channel state information, and process the base model based on the reconstructed channel state information to obtain a trained base model.

[0038] In a third aspect, the present application further provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps in this embodiment when executing the computer program.

[0039] In a fourth aspect, the present application further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps in this embodiment are implemented.

[0040] In a fifth aspect, the present application also provides a computer program product, including a computer program, which implements the steps in this embodiment when executed by a processor.

[0041] The above-mentioned training method, device, computer equipment, computer-readable storage medium and computer program product for the wireless base model for channel estimation and channel prediction, wherein the method includes: obtaining sample channel state information on the base station side, the sample channel state information includes information of the target dimension, and the target dimension includes time dimension, space dimension and frequency dimension; performing feature extraction processing on the sample channel state information through the feature dimension conversion network and the task corresponding to the base model to obtain the target token; encoding the target token through the position encoding module, the expert model in each expert joint encoding module and the gating network corresponding to the task to obtain the first output token; based on the task and at least one expert joint decoding module contained in the decoder, processing the first output token and the preset token to be learned to obtain reconstructed channel state information corresponding to the sample channel state information, and processing the base model based on the reconstructed channel state information to obtain the trained base model. By adopting this method, channel state information of different data distributions can be better processed, the adaptability of the base model on different CSI data sets is improved, and while considering the heterogeneity between different pre-training tasks, the conflicts between different pre-training tasks are resolved, thereby improving the reasoning performance and task processing accuracy of the trained base model in simultaneously processing multiple tasks, and being able to meet the real-time reasoning requirements of multiple tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments of the present application or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying any creative work.

[0043] Figure 1 A flowchart of a method for training a wireless base model for channel estimation and channel prediction in one embodiment is provided;

[0044] Figure 2 Schematic diagram of a process for obtaining a first output token in one embodiment;

[0045] Figure 3 is a flow chart of processing steps of an encoder in one embodiment;

[0046] Figure 4 1 is a flow chart of a method for training a wireless base model for channel estimation and channel prediction in another embodiment;

[0047] Figure 5 Schematic diagram of the structure of the pre-processing layer in one embodiment;

[0048] Figure 6 A schematic diagram of the structure of an expert processing layer in one embodiment;

[0049] Figure 7 A block diagram of a training device for a wireless base station model for channel estimation and channel prediction in one embodiment;

[0050] Figure 8 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0052] It should be noted that the terms "first", "second", etc. used in this application may be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "including" and "having" used in this application and any variations thereof are intended to cover non-exclusive inclusions. The term "plurality" used in this application refers to two or more. The term "and / or" used in this application refers to one of the solutions or any combination of multiple solutions.

[0053] In an exemplary embodiment, Figure 1 As shown, a training method for a wireless base model for channel estimation and channel prediction is provided. The method is explained by taking the application of the method to a terminal device as an example. The method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. The above-mentioned terminal can be, but is not limited to, various personal computers, laptops, smart phones, and tablet computers. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services. The base model includes at least a feature dimension conversion network, an encoder, and a decoder. The encoder includes a position encoding module and at least one expert joint encoding module. In this embodiment, the training method for a wireless base model for channel estimation and channel prediction includes the following steps:

[0054] Step 102: Acquire sample channel state information on the base station side.

[0055] The sample channel state information includes information of the target dimension, which includes the time dimension, the space dimension and the frequency dimension. The time dimension includes the number of time sampling points, the space dimension includes the number of antennas of the base station, and the frequency dimension includes the number of subcarriers.

[0056] Optionally, the method in this embodiment can be applied to a MIMO-OFDM system, wherein the base station side is equipped with a planar array multi-antenna and the user side is equipped with a single antenna. The large base model obtained after training in this embodiment can be a model deployed on the base station side, which can simultaneously process the channel estimation and channel prediction tasks of the "time-frequency-space" three-dimensional CSI. Among them, channel estimation refers to the CSI estimation value of the known partial position and the restoration of the complete CSI. The channel prediction task includes time domain channel prediction and frequency domain channel prediction, which refers to predicting unknown CSI through partial CSI in the time and frequency directions respectively.

[0057] Specifically, the terminal can obtain information from the base station of the MIMO-OFDM system in multiple dimensions, namely, the base station's CSI data (t, n, k), which includes information in the time dimension t, the spatial dimension n, and the frequency dimension k. In one example, the terminal needs to obtain the number of sampling points, the number of antennas in the base station, and the number of subcarriers. The information obtained by the terminal is the sample channel state information, namely the 3D CSI sample.

[0058] Step 104: Perform feature extraction processing on the sample channel state information through the feature dimension conversion network and the tasks corresponding to the base model to obtain the target token.

[0059] The feature dimension conversion network includes an embedding module and a masking module. The task corresponding to the base model can be the task that the base model is used to process, such as a channel estimation task or a channel prediction task. The task can also include a pre-training task, which can include one or more of a random mask reconstruction task, a time domain mask reconstruction task, a spatial domain mask reconstruction task, and an interpolation denoising task. The target token is the token corresponding to the task obtained after processing the sample channel state information.

[0060] Specifically, the terminal performs feature extraction processing on the sample channel state information through the embedding module and the mask module in the feature dimension conversion network. For example, the sample channel state information can be subjected to feature conversion processing through the embedding module to obtain multiple tokens, and the mask operation is performed on the obtained multiple tokens based on the mask module and the pre-training task to obtain the target token.

[0061] Step 106: Encode the target token through the position encoding module, the expert models in the expert joint encoding modules, and the gating network corresponding to the task to obtain the first output token.

[0062] The encoder includes a position encoding module and at least one expert joint encoding module; each expert joint encoding module includes a gating network and multiple expert models, and the gating network can output the model weights of each expert model, etc.

[0063] Specifically, the terminal can input the target token into the encoder, and perform position encoding processing on the target token through the position encoding module in the encoder to obtain a first position encoding token, and input the first position encoding token into at least one expert joint encoding module, and perform encoding processing through the gating network and multiple expert models in the expert joint encoding module to obtain the first output token output by the encoder.

[0064] Step 108: Based on the task and at least one expert joint decoding module included in the decoder, the first output token and the preset token to be learned are processed to obtain reconstructed channel state information corresponding to the sample channel state information, and the base model is processed based on the reconstructed channel state information to obtain a trained base model.

[0065] The decoder includes a position encoding module and at least one expert joint decoding module. The preset token to be learned can be a learnable token, for example, a token that can be learned during the model training process.

[0066] Specifically, the terminal can splice the first output token and the preset token to be learned based on the pre-training task to obtain a spliced ​​token. For example, it can be a one-to-one splicing of dimensions, etc. In this way, the terminal can input the spliced ​​token into the position coding module in the decoder, and perform position coding on the spliced ​​token through the position coding module to obtain a second position coding token. In this way, the terminal can decode the second position coding token through multiple end-to-end expert joint decoding modules to obtain a second output token. And the second output token is reconstructed through the fully connected layer in the output module to obtain reconstructed channel state information corresponding to the sample channel state information. In this way, the trained base large model can be obtained based on the reconstructed channel state information.

[0067] In the above-mentioned wireless base model training method for channel estimation and channel prediction, sample channel state information is obtained from the base station. The sample channel state information includes information on target dimensions, including time, space, and frequency. Feature extraction is performed on the sample channel state information using a feature dimension conversion network and the tasks corresponding to the base model to obtain a target token. The target token is encoded using a position encoding module, expert models in each expert joint encoding module, and a gating network corresponding to the task to obtain a first output token. Based on the task and at least one expert joint decoding module included in the decoder, the first output token and a preset token to be learned are processed to obtain reconstructed channel state information corresponding to the sample channel state information. The base model is then processed based on the reconstructed channel state information to obtain a trained base model. This method can better process channel state information with different data distributions, improving the base model's adaptability to different CSI datasets. While considering the heterogeneity of different pre-training tasks, conflicts between different pre-training tasks are resolved. This improves the inference performance and task processing accuracy of the trained base model when simultaneously processing multiple tasks, enabling it to meet the real-time inference requirements of multiple tasks.

[0068] In an exemplary embodiment, Figure 2 As shown in FIG, the specific execution process of "encoding the target token through the position encoding module, the expert models in the expert joint encoding modules, and the gating network corresponding to the task to obtain the first output token" includes:

[0069] Step 202: Position-encode the target token through a position encoding module to obtain a first position-encoded token.

[0070] Step 204 : encoding the first position coding token through the multiple expert models in the end-to-end expert joint coding modules and the gating network corresponding to the task to obtain a first output token.

[0071] Among them, each expert joint coding module is connected end to end, that is, the output data of the first expert joint coding module is the input data of the second expert joint coding module, the output data of the position coding module is the input data of the first expert joint coding module, and the output data of the expert joint coding module at the end position in the encoder is the output data of the encoder.

[0072] Specifically, the terminal can input the target token into the position encoding module in the encoder, which then performs position encoding on the concatenated token to obtain a first position encoding token. The terminal can then input the first position encoding token into the first expert joint encoding module, which then encodes the first position encoding token through multiple end-to-end expert joint encoding modules to obtain a first output token. The terminal can then input the first output token into the decoder.

[0073] In this embodiment, by encoding an expert processing layer including a gating network and multiple expert models, different expert activation characteristics can be learned for different tasks, and the performance of the base model can be improved while considering data heterogeneity.

[0074] In an exemplary embodiment, the expert joint encoding module includes a pre-processing layer and an expert processing layer, and the expert processing layer includes multiple expert models and multiple gating networks.

[0075] Accordingly, if Figure 3 As shown in FIG, the specific execution process of “encoding the first position encoding token by using multiple expert models in the end-to-end expert joint encoding modules and the gating network corresponding to the task” includes:

[0076] In step 302 , for the i-th expert joint encoding module, the input token of the i-th expert joint encoding module is normalized and processed by the attention mechanism through the pre-processing layer in the i-th expert joint encoding module to obtain a first token.

[0077] Among them, the pre-processing layer includes the first normalization layer, the multi-head attention layer, the second normalization layer, and the expert processing layer can be a moe feedforward layer.

[0078] Specifically, the input token of the i-th expert joint encoding module is normalized by the first normalization layer in the pre-processing layer to obtain a first normalized result. The first normalized result is input to the multi-head attention layer to obtain an attention result. The attention result and the input token are superimposed to obtain a superimposed result. The superimposed result is input to the second normalization layer to obtain a second normalized result. In other words, the terminal can determine that the second normalized result is the first token.

[0079] In step 304, the first token is encoded based on each expert model in the i-th expert joint encoding module to obtain the expert output corresponding to each expert model. Furthermore, the first token is processed based on the gating network corresponding to the task and the input token corresponding to the i-th expert joint encoding module to obtain the weight corresponding to each expert model.

[0080] Specifically, the terminal can input the second normalization result (first token) into the expert processing layer; the expert processing layer includes multiple expert models, and each expert model encodes the first token to obtain the expert output results of each expert model for the first token; at the same time, the terminal can determine the weights of each expert model based on the pre-training task. For example, the first token can be processed by the fully connected layer in the gating network to obtain N weights, and then normalized by the Softmax layer. Finally, the top K weights are retained, and the remaining weights are reset to zero to obtain the output of the model weights of each expert model by the gating network. Optionally, the terminal can also determine the activated expert model from the multiple expert models based on the type of pre-training task, and encode the first token through each activated expert model to obtain the expert output results of the first token corresponding to each expert model. Wherein, K and N are both positive integers.

[0081] Step 306: Based on the expert output results of each expert model and the weight corresponding to each expert model, the output token of the i-th expert joint encoding module is obtained.

[0082] Specifically, the terminal may perform a weighted calculation based on the expert output results of each expert model and the model weights corresponding to each expert model to obtain a weighted result, and determine the weighted result as the output token of the i-th expert joint encoding module. Alternatively, the terminal may perform a weighted calculation based on the expert output results of each activated expert model and the model weights corresponding to each activated expert model to obtain a weighted result, and determine the weighted result as the output token of the i-th expert joint encoding module.

[0083] In step 308, the output token of the expert joint encoding module at the last position is determined to be the first output token. The input token of the expert joint encoding module at the initial position is the first position encoding token, and the input token of the i-th expert joint encoding module is the output token of the i-1-th expert joint encoding module.

[0084] Specifically, the terminal may determine the output token of the expert joint encoding module located at the end of the encoder as the first output token of the encoder. That is, if there are n expert joint encoding modules in the encoder, the terminal may determine the output token of the nth expert joint encoding module as the first output token of the encoder.

[0085] Among them, the input token of the first expert joint coding module located in the encoder is the first position coding token, that is, the terminal can input the first position coding token into the first expert joint coding module; the input token of the i-th expert joint coding module is the output token of the i-th expert joint coding module, that is, the input token of the i-th expert joint coding module is the output token of the previous expert joint coding module of this module.

[0086] In this embodiment, by encoding an expert processing layer including a gating network and multiple expert models, different expert activation characteristics can be learned for different tasks, and the performance of the base model can be improved while considering data heterogeneity.

[0087] In an exemplary embodiment, the encoder further includes a position encoding module. The specific execution process of the step of "processing the first output token and the preset token to be learned by the task and at least one expert joint decoding module included in the decoder to obtain reconstructed channel state information corresponding to the sample channel state information" includes:

[0088] Based on the task-specific token concatenation strategy, the first output token and the preset token to be learned are concatenated to obtain a concatenated token. The concatenated token is positionally encoded using a position encoding module to obtain a second position-encoded token. The second position-encoded token is decoded using multiple expert models in end-to-end expert joint decoding modules and a task-specific gating network to obtain a second output token. Based on the second output token, reconstructed channel state information corresponding to the sample channel state information is obtained.

[0089] The preset token to be learned can be a special token that can be learned during the model training process and can interact with other tokens. The token splicing strategy corresponding to the task can be the token splicing strategy corresponding to the pre-trained task, which is used to indicate the splicing method between the first output token and the preset token to be learned. The decoder includes a splicing module, a position encoding module, and multiple expert joint decoding modules.

[0090] Specifically, the terminal can determine the token splicing strategy corresponding to the pre-training task according to the type of pre-training task, and based on the token splicing method indicated by the token splicing strategy, splice the first output token output by the encoder with the preset token to be learned to obtain a spliced ​​token. For example, it can be a one-to-one splicing of dimensions, etc. In this way, the terminal can input the spliced ​​token into the position coding module in the decoder, and perform position coding on the spliced ​​token through the position coding module to obtain a second position coding token. In this way, the terminal can input the second position coding token into the first expert joint decoding module, that is, the expert joint decoding module at the initial position, that is, the terminal can use multiple expert joint decoding modules connected end to end to decode the second position coding token to obtain a second output token. And the second output token is reconstructed through the fully connected layer in the output module to obtain the reconstructed channel state information corresponding to the sample channel state information.

[0091] Optionally, the second position coding token is decoded by multiple end-to-end expert joint decoding modules to obtain the second output token. The specific process may be similar to the encoding process in the encoder, which will not be repeated here.

[0092] In this embodiment, the splicing of learning tokens and output tokens can improve the model's ability to focus on important information during training, improve the base model's backtracking and reasoning capabilities during training and application, and further improve model performance.

[0093] In an exemplary embodiment, the base model further includes an output module, and the specific implementation process of the step of "obtaining reconstructed channel state information corresponding to the sample channel state information based on the second output token" includes:

[0094] The second output token is reconstructed through the fully connected layer in the output module to obtain reconstructed channel state information corresponding to the sample channel state information.

[0095] Among them, the output module can include a fully connected layer.

[0096] Specifically, the terminal can reconstruct the second output token output by the decoder through the fully connected layer in the output module. For example, the fully connected layer includes a network layer composed of fully connected neurons. The terminal outputs the second output token to the fully connected layer. The features of the second output token can be linearly transformed through the weight matrix and bias term of the fully connected layer, and then a nonlinear mapping is introduced through the activation function. The second output token is transformed by the nonlinear mapping through the activation function, and finally a reconstruction result that matches the dimension of the original token (that is, the sample channel state information) is output, that is, the reconstructed channel state information is obtained.

[0097] In this embodiment, the base model can learn how to recover the original information from the compressed or abstract token representation, thereby optimizing the feature extraction capability of the model.

[0098] In an exemplary embodiment, the time dimension includes the number of time sampling points, the spatial dimension includes the number of antennas of the base station, and the frequency dimension includes the number of subcarriers. The step of "performing feature extraction processing on the sample channel state information through the feature dimension conversion network and the tasks corresponding to the base model to obtain the target token" includes:

[0099] The sample channel state information is divided into blocks to obtain multiple sample channel state blocks. Feature extraction is performed on each sample channel state block using a feature dimension conversion network to obtain multiple unit tokens. Based on the task, the multiple unit tokens are screened to obtain the target token.

[0100] Specifically, the terminal may first perform block processing on the sample channel state information to obtain a plurality of sample channel state information blocks, and perform embedding processing on each sample channel state information block through the embedding module in the feature dimension conversion network to obtain a unit token corresponding to each sample channel state information block, and the dimension of the unit token may be a one-dimensional token. For example, the sample channel state information may be three-dimensional data, and the corresponding sample channel state information blocks may be three-dimensional CSI blocks. The embedding module in the feature dimension conversion network may perform embedding processing on each three-dimensional CSI block, that is, perform feature extraction processing, and obtain a one-dimensional unit token. In this way, the terminal may perform a masking operation on the multiple one-dimensional unit tokens corresponding to the multiple three-dimensional CSI blocks based on the pre-training task through the mask module in the feature dimension conversion network to obtain a target token, that is, obtain a target token that matches the pre-training task.

[0101] Optionally, the terminal can perform different mask operations based on different pre-training tasks, that is, selectively retain some unit tokens among all unit tokens or input all tokens into the subsequent network, that is, determine some unit tokens as target tokens, or determine all unit tokens as target tokens, etc.

[0102] Optionally, each token has three-dimensional position information (y), which identifies the (time, space, and frequency) information of the token in the original 3D CSI. Random mask reconstruction randomly retains a certain proportion of tokens in the mask module. Time-domain mask reconstruction retains tokens whose time dimension coordinates are less than a certain value. Frequency-domain mask reconstruction retains tokens whose frequency dimension coordinates are less than a certain value. For the three mask reconstruction pre-training tasks mentioned above, some learnable tokens are spliced ​​into the decoder module, corresponding to the tokens masked in the mask module, so that the number of tokens input to the decoder network is the same as the number of tokens corresponding to the original 3D CSI.

[0103] For interpolation denoising pre-training, the terminal downsamples the original 3D CSI at regular intervals in time, space, and frequency, then performs linear interpolation to obtain CSI data of the same size as the original 3D CSI. This step mimics the process of estimating the overall CSI by interpolating the CSI at the pilot signals during channel estimation. The interpolated CSI data is processed by the network and does not pass through the mask module or mask splicing operations.

[0104] In this embodiment, different mask operations are performed on different pre-training tasks to obtain different target tokens, and accuracy optimization is performed for different tasks. This can not only improve the performance of a single task, but also provide a basis for the modular design and dynamic adaptation of complex systems, further improving the model performance of the base large model.

[0105] In an exemplary embodiment, the step of “processing the base model based on the reconstructed channel state information to obtain a trained base model” includes:

[0106] Based on the sample channel state information and the reconstructed channel state information corresponding to the sample channel state information, a loss prediction is performed to obtain the channel reconstruction loss. Furthermore, based on the sample channel state information and the target token, a token routing ratio is determined. Based on the tasks corresponding to the base model, the token routing ratio, and the average weight corresponding to the gating network, a task load balancing loss is obtained. Based on the channel reconstruction loss, a preset weighting coefficient, and the task load balancing loss, a model training loss is obtained. The model parameters of the base model are updated using the model training loss until the preset training completion conditions are met, resulting in a trained base model.

[0107] The preset training completion condition may be that the number of training times reaches the target number, or that the loss function obtained during the training process has satisfied the convergence condition. The convergence condition may be that the loss function has remained unchanged, or that the loss function has approached the target value. This embodiment does not limit the specific values ​​of the number of training times and the target value, and those skilled in the art may determine these values ​​based on the needs of the actual application scenario. The preset weighting coefficients include a first coefficient for channel reconstruction loss and a second coefficient for task load balancing loss.

[0108] Specifically, the channel reconstruction loss L1 can be the loss of channel estimation or channel prediction; the task load balancing loss L2 is the load balancing loss for the moe architecture. The terminal can determine the channel reconstruction loss through the mean square error (MSE). For example, the channel reconstruction loss L1 can be calculated using the following formula:

[0109]

[0110] Among them, H is the original CSI, that is, the sample channel state information, It is to reconstruct the channel state information.

[0111] In this way, the terminal can obtain the task load balancing loss L2 through the pre-training task corresponding to the base model, the token routing ratio and the average weight corresponding to the gating network. For example, the terminal may be executing the i-th pre-training task, and there are T target tokens in the sample channel state data Ω of a training batch, where T is a positive integer; for example, the token routing ratio can be D n , n represents the number of expert models, the token routing ratio represents the ratio of all target tokens routed to expert model n, and the average activation weight of the gating network for each expert model n is P n , then the terminal can calculate the load balancing loss L2 using the following formula:

[0112]

[0113] in, is an indicator function, which is 1 when the condition in the brackets is met, otherwise it is 0; Indicates the model weight of the nth expert model corresponding to the target token x for the i-th pre-training task output by the gated network G; in this way, the terminal can perform weighted calculation by the channel reconstruction loss, the task load balancing loss, the first coefficient of the channel reconstruction loss, and the second coefficient of the task load balancing loss to obtain the model training loss corresponding to the base model. After the terminal obtains the model training loss, it can determine whether the preset training completion condition is currently met. If the preset training completion condition is not met, the terminal can update the model parameters of the base model based on the model training loss to obtain an updated base model, and re-execute the method in the above embodiment with the updated base model until the preset training completion condition is met to obtain a trained base model.

[0114] In this embodiment, by comprehensively considering the loss function in terms of task load balancing and the loss function in terms of channel information, it is possible to balance multi-objective optimization, avoid single-objective deviation, improve training efficiency, and improve the performance of the model obtained after training.

[0115] The following describes in detail the specific implementation process of the above-mentioned wireless base model training method for channel estimation and channel prediction in conjunction with a specific embodiment:

[0116] The training method of a wireless base model for channel estimation and channel prediction provided in this embodiment is a training method for a wireless base model for channel estimation and channel prediction, which mainly involves an improved sparse MoE network architecture. By replacing the FNN layer in the transformer block with a specially designed MoE layer, the network can better process CSI data with different distributions, resolve conflicts between different pre-training tasks, and have better inference acceleration performance. For example, it is applied to a MIMO-OFDM system, in which the base station side is equipped with a planar array multi-antenna and the user side is equipped with a single antenna. The proposed large base model is deployed on the base station side and can simultaneously process the channel estimation and channel prediction tasks of the "time-frequency-space" three-dimensional CSI. Among them, channel estimation refers to the CSI estimation value of a known partial position and the restoration of the complete CSI. The channel prediction task includes time domain channel prediction and frequency domain channel prediction, which refers to predicting unknown CSI through partial CSI in the time and frequency directions respectively.

[0117] like Figure 4 As shown, the network architecture of this embodiment may include an embedding module, a mask module, an encoder module, a decoder module, and an output module. The input data of the embedding module may be sample channel state data, i.e., 3D CSI samples. The encoder module may include encoder position encoding (position encoding module) and a TTS-MoE architecture. The TTS-MoE architecture may include multiple module blocks, each of which may be an expert joint encoding module. The decoder module may include mask splicing (splicing module), decoder position encoding (position encoding module), and a TTS-MoE architecture. The TTS-MoE architecture may include multiple module blocks, each of which may be an expert joint decoding module. That is, the decoder includes multiple expert joint decoding modules connected end to end. The reconstructed CSI obtained may be reconstructed channel state data.

[0118] The embedding module takes 3D CSI samples as input, processes them into a series of tokens. Specifically, the 3D CSI samples are first divided into blocks to obtain a series of 3D CSI blocks. These blocks are then embedded into a series of 1D tokens using a 3D convolutional network, resulting in multiple unit tokens.

[0119] Mask module: Performs different mask operations according to different pre-training tasks, that is, selectively retains some tokens or inputs all tokens into the subsequent network, and determines the target token based on the pre-training task among multiple unit tokens.

[0120] Encoder module: The input tokens fed into the encoder module are first positionally encoded and then processed through a series of transformer blocks, where each transformer block adopts the proposed TTS-MoE architecture to obtain the first output token output by the encoder module.

[0121] Decoder module: The first output token from the encoder module is concatenated with the token to be learned to produce a concatenated token. The concatenation method depends on the pre-training task. The concatenated token is first positionally encoded and then processed through a transformer block based on the TTS-MoE architecture to produce the second output token.

[0122] Output module: The output module obtains the reconstructed CSI through the fully connected layer for the second output token output by the decoder.

[0123] like Figure 5 As shown, it can be the pre-processing layer in the expert joint encoding module / expert joint decoding module, such as Figure 6 As shown, it can be a structural diagram of the expert processing layer in the expert joint encoding module / expert joint decoding module, that is, the MoE feedforward layer. The following is an example of the expert joint encoding module:

[0124] For example, the expert joint encoding module can include N expert networks, or N expert models, with K of them activated at a time. These K activated expert models can be the expert models used in the current training batch. The expert joint encoding module can include four gating networks, corresponding to four pre-training tasks: random mask reconstruction, time-domain mask reconstruction, frequency-domain mask reconstruction, and interpolation denoising. When processing each corresponding task, the relevant gate is activated, and N expert weights are output based on the input tokens. The expert outputs are weighted to obtain the output of the MoE feedforward layer.

[0125] For example, the pre-training task currently being performed by the base model may be the i-th task, and the token input to the MoE feed-forward layer is , the corresponding activation network is G i , and its weight distribution to the nth expert is , the nth expert network is E n . Assume that the execution When there are tasks, the output of the MoE feedforward layer is y i , then the following formula is satisfied:

[0126] ;

[0127] Among them, each expert network is a FFN architecture, namely:

[0128] ;

[0129] The gating network first processes the input x through the fully connected layer to obtain N weights, then normalizes it through the Softmax layer, and finally retains the top K largest weights and resets the rest to zero. The output of the gating network is

[0130]

[0131] Among them, FC represents the fully connected layer, and FC(x) represents the processing of the input x through the fully connected layer.

[0132] This embodiment is a base model for channel prediction and channel estimation tasks, and a token-aware and task-specific sparse mixture of experts (TTS-MoE) model is designed. The Transformer blocks of the model encoder and decoder contain a mixture of experts (MoE) layer. In this way, the weights and activation selection of the experts in the mixture of experts layer are determined by the gating network, and the output of the gating network is determined by the input token. This can fully consider data heterogeneity, that is, the expert activation and weight distribution are related to different input data; each pre-training task corresponds to a dedicated gating network, so that the network learns specialized expert activation characteristics for different tasks; after training, the base model only needs to activate some weights simultaneously during task reasoning, which greatly reduces the network reasoning latency.

[0133] That is to say, the training method of the wireless base model for channel estimation and channel prediction in this embodiment can improve the performance of the base model in solving channel prediction and channel estimation tasks simultaneously while considering the heterogeneity and conflicts between pre-training tasks; it can also consider the differences in CSI data distribution to improve the adaptability of the base model on different CSI data sets, thereby improving the estimation prediction accuracy and generalization performance.

[0134] It should be understood that, although the various steps in the flowcharts involved in the above embodiments are displayed in sequence according to the instructions of the arrows, these steps are not necessarily performed in sequence in the order indicated by the arrows. Unless clearly stated herein, the execution of these steps is not strictly limited in order, and these steps can be performed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of the steps or stages in other steps or other steps. It is understandable that the various steps in different embodiments can be freely combined as needed, and the various non-contradictory schemes formed by the combination all fall within the scope of protection of this application.

[0135] Based on the same inventive concept, an embodiment of the present application also provides a training device for a wireless base model for channel estimation and channel prediction for implementing the above-mentioned training method for a wireless base model for channel estimation and channel prediction. The implementation solution provided by the device is similar to the implementation solution described in the above-mentioned method. Therefore, the specific limitations of the embodiments of one or more training devices for wireless base models for channel estimation and channel prediction provided below can be found in the above-mentioned limitations on the training method for wireless base models for channel estimation and channel prediction, and will not be repeated here.

[0136] In an exemplary embodiment, Figure 7 As shown, a training device 700 for a wireless base model for channel estimation and channel prediction is provided. The base model includes at least a feature dimension conversion network, an encoder, and a decoder. The encoder includes a position encoding module and at least one expert joint encoding module. The device includes:

[0137] A first acquisition module 702 is configured to acquire sample channel state information from a base station, where the sample channel state information includes information of target dimensions, including time, space, and frequency dimensions.

[0138] The first processing module 704 is used to perform feature extraction processing on the sample channel state information through the feature dimension conversion network and the tasks corresponding to the base model to obtain the target token;

[0139] A first encoding module 706 is configured to encode the target token using the position encoding module, the expert models in the expert joint encoding modules, and the gating network corresponding to the task to obtain a first output token;

[0140] The second processing module 708 is used to process the first output token and the preset token to be learned based on the task and at least one expert joint decoding module included in the decoder, obtain reconstructed channel state information corresponding to the sample channel state information, and process the base model based on the reconstructed channel state information to obtain a trained base model.

[0141] Each module in the aforementioned wireless base station model training device for channel estimation and channel prediction can be implemented in whole or in part via software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0142] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Figure 8 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store sample data and reconstruction data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a training method for a wireless base model for channel estimation and channel prediction is implemented.

[0143] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the embodiments of the present application when executing the computer program.

[0144] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the embodiments of the present application are implemented.

[0145] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the embodiments of the present application when executed by a processor.

[0146] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0147] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), quantum computing-based data processing logic devices, artificial intelligence (AI) processors, and the like.

[0148] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0149] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A training method for a wireless base station model for channel estimation and channel prediction, characterized in that: The base model comprises at least a feature dimension conversion network, an encoder and a decoder, wherein the encoder comprises a position encoding module and at least one expert joint encoding module; the method comprises: Acquire sample channel state information on the base station side, where the sample channel state information includes information of target dimensions, where the target dimensions include a time dimension, a space dimension, and a frequency dimension; Performing feature extraction processing on the sample channel state information through the feature dimension conversion network and the tasks corresponding to the base model to obtain a target token; Encoding the target token through the position encoding module, the expert models in each of the expert joint encoding modules, and the gating network corresponding to the task to obtain a first output token; Based on the task and at least one expert joint decoding module contained in the decoder, the first output token and the preset token to be learned are processed to obtain reconstructed channel state information corresponding to the sample channel state information, and the base model is processed based on the reconstructed channel state information to obtain a trained base model.

2. The method according to claim 1, characterized in that The encoding process of the target token by the position encoding module, the expert models in each of the expert joint encoding modules, and the gating network corresponding to the task to obtain a first output token includes: Position-encode the target token using a position encoding module to obtain a first position-encoded token; The first position coding token is encoded by using multiple expert models in each of the expert joint coding modules connected end to end and a gating network corresponding to the task to obtain a first output token.

3. The method according to claim 2, characterized in that The expert joint encoding module includes a pre-processing layer and an expert processing layer, wherein the expert processing layer includes multiple expert models and multiple gating networks; the first position coding token is encoded by the multiple expert models in each of the expert joint encoding modules connected end to end and the gating network corresponding to the task to obtain a first output token, including: For the i-th expert joint encoding module, normalize and process the input token of the i-th expert joint encoding module through the pre-processing layer in the i-th expert joint encoding module to obtain a first token; Based on each expert model in the i-th expert joint encoding module, encode the first token respectively to obtain the expert output result corresponding to each expert model; and based on the gating network corresponding to the task and the input token corresponding to the i-th expert joint encoding module, process the first token to obtain the weight corresponding to each expert model; Based on the expert output results of each expert model and the weights corresponding to each expert model, obtaining the output token of the i-th expert joint encoding module; Determine that the output token of the expert joint coding module located at the last position is the first output token; the input token of the expert joint coding module located at the initial position is the first position coding token, and the input token of the i-th expert joint coding module is the output token of the i-1-th expert joint coding module.

4. The method according to claim 1, wherein The encoder further includes a position encoding module; the processing of the first output token and a preset token to be learned by the task and at least one expert joint decoding module included in the decoder to obtain reconstructed channel state information corresponding to the sample channel state information includes: Based on the token splicing strategy corresponding to the task, the first output token and the preset token to be learned are spliced ​​to obtain a spliced ​​token; Position-encoding the spliced ​​token by the position encoding module to obtain a second position encoding token; The second position coding token is decoded by multiple expert models in each of the expert joint decoding modules connected end to end and the gating network corresponding to the task to obtain a second output token, and based on the second output token, the reconstructed channel state information corresponding to the sample channel state information is obtained.

5. The method according to claim 4, characterized in that The base model further includes an output module, and obtaining, based on the second output token, reconstructed channel state information corresponding to the sample channel state information includes: The second output token is reconstructed through the fully connected layer in the output module to obtain reconstructed channel state information corresponding to the sample channel state information.

6. The method according to claim 1, wherein The time dimension includes the number of time sampling points, the space dimension includes the number of antennas of the base station, and the frequency dimension includes the number of subcarriers. The feature dimension conversion network and the task corresponding to the base model are used to perform feature extraction processing on the sample channel state information to obtain a target token, including: performing block processing on the sample channel state information to obtain a plurality of sample channel state blocks; Perform feature extraction processing on each sample channel state block through a feature dimension conversion network to obtain multiple unit tokens; Based on the task, a plurality of unit tokens are screened to obtain a target token.

7. The method according to claim 1, characterized in that The processing of the base model based on the reconstructed channel state information to obtain a trained base model includes: Performing loss prediction based on the sample channel state information and reconstructed channel state information corresponding to the sample channel state information to obtain a channel reconstruction loss; and determining a token routing ratio based on the sample channel state information and the target token; Obtaining a task load balancing loss based on the tasks corresponding to the base model, the token routing ratio, and the average weight corresponding to the gating network; Obtaining a model training loss based on the channel reconstruction loss, a preset weighting coefficient, and the task load balancing loss; The model parameters of the base model are updated through the model training loss until the preset training completion conditions are met to obtain a trained base model.

8. A training device for a wireless base station model for channel estimation and channel prediction, characterized in that: The base model comprises at least a feature dimension conversion network, an encoder and a decoder, wherein the encoder comprises a position encoding module and at least one expert joint encoding module; the apparatus comprises: A first acquisition module is configured to acquire sample channel state information from a base station, where the sample channel state information includes information of a target dimension, and the target dimension includes a time dimension, a space dimension, and a frequency dimension; A first processing module is configured to perform feature extraction processing on the sample channel state information through the feature dimension conversion network and the tasks corresponding to the base model to obtain a target token; A first encoding module is configured to encode the target token using the position encoding module, the expert models in the expert joint encoding modules, and the gating network corresponding to the task to obtain a first output token; The second processing module is used to process the first output token and the preset token to be learned based on the task and at least one expert joint decoding module included in the decoder, obtain reconstructed channel state information corresponding to the sample channel state information, and process the base model based on the reconstructed channel state information to obtain a trained base model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

11. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images