Wireless communication physical layer multitasking method based on pre-trained large language model

Abstract

The application discloses a wireless communication physical layer multitask processing method based on a pre-trained large language model, belongs to the technical field of wireless communication, and constructs a wireless communication system physical layer multitask large model based on a pre-trained large language model, is deployed at a base station side, is trained through user historical data, and after training is completed, online inference is carried out, can be used for completing multiple tasks such as channel estimation, channel prediction, distance estimation, road loss estimation and beam forming, and the model has low deployment difficulty, small training cost and maintained model precision.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of wireless communication, and relates to a wireless communication system physical layer task processing technology, in particular to a wireless communication system physical layer multi-task processing solution based on a pre-trained large language model. BACKGROUND

[0002] With the popularity of the Internet of Vehicles scene, the number and density of communication demands gradually increase. This poses a huge challenge to the rate and stability of the fifth generation communication system (5G). This challenge affects every aspect of communication design. At the network level, more efficient congestion control and traffic management strategies are needed to cope with dynamic loads; at the physical transmission level, since accurate channel state information (CSI) is the premise for the normal operation of subsequent precoding, power allocation, modulation selection, and transmit antenna selection, etc., with the introduction of massive multiple-input multiple-output (mMIMO) technology, the number of antennas has increased significantly, resulting in an increase in pilot resource overhead for CSI estimation, and the channel aging problem brought about by high speed also exacerbates this phenomenon, resulting in a decrease in system spectral efficiency; these challenges all need to be addressed by designing appropriate solutions.

[0003] Due to the different optimization objectives of these problems, traditional solutions all focus on single task optimization or single module optimization, for example, for pilot-based channel estimation tasks, many methods choose generative artificial intelligence (GAI) as a tool to achieve higher accuracy of channel estimation, because the powerful generation capability of GAI is in line with the task requirements of channel estimation; while in the channel prediction task, most methods use recurrent neural networks (RNN) due to their excellent performance in handling time series tasks; but designing an optimization method for each module separately in a system will make the design process very complex, and the lack of consideration of the error in information transmission between modules when designing separately will also lead to a decrease in system stability. At the same time, the relatively independent optimization between modules can only obtain a local optimal solution, and it is difficult to obtain a global optimal solution that combines all modules.

[0004] The multi-task learning method can jointly train multiple communication tasks and obtain universal representations between tasks to obtain better model performance, while also reducing the complexity of the training process. Therefore, some works consider two key tasks of wireless signal recognition, wireless signal classification and modulation mode recognition, and use a bottom-sharing multi-task learning network for joint training; there are also works that consider the joint training and inference of direct channel (DC) and cascaded channel (CC) estimation in the reconfigurable intelligent surface (RIS) system. By jointly learning two tasks, DC and CC can be obtained using one network, reducing the overhead of pilots in the actual system. However, these methods also have obvious shortcomings. The bottom-sharing multi-task learning method is prone to data imbalance and seesaw phenomenon, and due to the small size of the model, it is difficult to improve the types and range of learning tasks, that is, most existing methods only jointly train two strongly related communication tasks, and how to jointly train more communication tasks or communication tasks with no obvious correlation remains to be solved.

[0005] Large language models (LLMs) are emerging multi-task learners. Unlike traditional deep learning methods, large models have shown amazing reasoning and generalization performance in various fields such as natural language processing, medicine, law, finance, etc. The GPT series of models, as the earliest language large models, have shown good performance in various natural language processing tasks. In the field of computer vision, large models such as BeiT have achieved the best performance in tasks such as object detection, semantic segmentation, and visual question answering. Although there have been attempts to apply pre-trained LLMs to channel prediction tasks through fine-tuning, this approach is only suitable for single tasks and does not fully exploit the multi-task characteristics of LLMs. Therefore, applying pre-trained large language models to handle multiple tasks in the physical layer of wireless communication systems through fine-tuning still has some difficulties. Due to the diverse mapping relationships of physical layer tasks, there are many weakly related tasks, such as channel prediction and positioning enhancement, which increase the difficulty of joint training. Moreover, due to the huge domain gap between CSI and natural language, and the need to avoid the occurrence of mutual exclusion phenomenon caused by weak correlation between tasks, the characteristics of the physical layer of wireless communication systems further increase the difficulty of network design. SUMMARY

[0006] In order to overcome the deficiencies of the prior art, the present application proposes a multi-input multi-output-orthogonal frequency division multiplexing (MIMO-OFDM) wireless communication system physical layer multi-task processing solution based on a pre-trained large language model. The designed wireless communication system physical layer multi-task network model is trained on the physical layer multi-task hybrid data set, achieving high-precision physical layer multi-task effect and high-efficiency training process.

[0007] The present application is a technology for fine-tuning a pre-trained large language model to enable it to solve multiple physical layer tasks in a MIMO-OFDM wireless communication system. The physical layer tasks include but are not limited to channel estimation, channel prediction, beam management, positioning enhancement, etc. The present application considers a dual-mode communication system that simultaneously supports low-frequency and millimeter-wave wireless communication links. A pre-trained large language model is used to build a wireless communication system physical layer multi-task large model, which is deployed at the base station side. The model is trained using user historical data and performs online inference after training. It can be used to complete tasks such as channel estimation, channel prediction, distance estimation, path loss estimation, and beamforming. The input of the physical layer multi-task large model is the user's uplink low-frequency pilot signal at the first P time points, and the output is the target output for each task, which is the user's uplink low-frequency CSI at the first P time points, the user's downlink low-frequency CSI at the last L time points, the user's distance from the base station, and the millimeter-wave downlink beam vector. By integrating multiple tasks into one model, the inference efficiency of the model is improved without compromising the inference accuracy, making it easier to maintain and update the model and the system.

[0008] To achieve the above purpose, the present application mainly includes a wireless communication system physical layer multi-task network model based on a pre-trained large language model (GPT-2), which includes a preprocessing module, a multi-task adapter module, a pre-trained large model backbone network module, and a multi-task output module. It is deployed at the base station side and performs online inference of multiple physical layer tasks using user uplink low-frequency pilot signals. The GPT-2 model is an open-source large language model based on the Transformer architecture developed by OpenAI. To improve the multi-task learning efficiency of the model, most of the parameters of the GPT-2 model are frozen, and only the linear layers of the feedforward network are fine-tuned using the MoE-Lora designed by the present application. MoE-Lora refers to a multi-task fine-tuning method that combines hybrid expert models and Lora fine-tuning. The present application first applies this multi-task fine-tuning method to physical layer tasks in a wireless communication system and combines pre-trained large models with multi-task learning to achieve efficient multi-task fine-tuning of large models. In addition, the present application designs a multi-task expert weight distribution network to better dynamically combine the decisions of the expert network based on the characteristics of the physical layer tasks.

[0009] The wireless communication system physical layer multi-task processing solution based on a pre-trained large language model comprises the following steps:

[0010] 1) A user obtains a MIMO-OFDM channel prediction training data set through field channel measurement or simulation, each sample comprising pilot signals received by an uplink low-frequency link in a continuous time period (model input), CSI data of a downlink low-frequency link in a continuous time period (model input), the distance and path loss of a user and a base station, and the optimal beam sequence number of a millimeter wave downlink (model output), and the sampling intervals of the time sampling points are the same.

[0011] 2) A wireless communication system physical layer multi-task network model based on a pre-trained large language model (GPT-2) is constructed, comprising a preprocessing module, a multi-task adapter module, a pre-trained large model backbone network module and a multi-task output module. The network input is the pilot signals of the user uplink low-frequency link in the first P time (the pilot signals received by the uplink low-frequency link in the first P time sampling points in the training set), and the output is the target output of each task, i.e. the user uplink low-frequency CSI in the first P time, the user downlink low-frequency CSI in the last L time, the distance and path loss of the user from the base station, and the millimeter wave downlink beam vector. The constructed physical layer multi-task network is trained on the training data set in step 1).

[0012] 3) The trained wireless communication system physical layer multi-task network model is deployed to the base station side of the actual scene of the wireless communication system, and a plurality of physical layer tasks are completed online through reasoning according to the user uplink low-frequency pilot signals.

[0013] Through the above steps, the physical layer multi-task processing of the base station side of the wireless communication system is realized.

[0014] Compared with the prior art, the beneficial effects of the present application are:

[0015] The present application provides a wireless communication system physical layer multi-task processing solution based on a pre-trained large language model, which applies the pre-trained large language model to multiple physical layer tasks through network design and fine-tuning training, fully utilizes the powerful modeling and generalization capabilities of the large language model, and achieves good performance on each task. At the same time, compared with the previous way of designing a network for each module separately, the overall design complexity of the model is greatly reduced, and the model training process is simplified. The technology proposed in the present application has the following advantages:

[0016] (I) Task processing accuracy: The large model multi-task network extracts appropriate features for each task through the preprocessing module, which facilitates the model to capture task characteristics. The multi-task adapter module aligns the task features of each task with the semantic feature space of the pre-trained large model, and the constructed multi-task output head module selects the most suitable model architecture for each task. These specially designed network structures improve the accuracy of the model on each task.

[0017] (II) Model generalization performance: The backbone network of the large model multi-task network is a pre-trained GPT-2 model, most of whose parameters are frozen, and only MoE-Lora is used to fine-tune the feedforward network. The model adapts to different tasks by dynamically combining Lora weights through the threshold routing network, achieving dynamic adaptation of the network to the input and thus obtaining strong generalization ability. The proposed large model multi-task network can maintain good model accuracy when the channel data distribution changes.

[0018] (III) Model training overhead: To meet the demand of multiple physical layer tasks, the large model multi-task network only needs to be trained once and can be deployed to the base station for operation. Compared with the traditional method of selecting a suitable network for each module and training and deploying them separately, the training process of this method is simpler and the deployment difficulty is greatly reduced. At the same time, since the large model multi-task network is based on the fine-tuning of the pre-trained large language model, the training cost is comparable to that of ordinary small models (such as CNN, RNN, etc.), and there is no huge training overhead. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is the network architecture diagram based on the pre-trained large model when the present application is implemented.

[0020] Figure 2 is the multi-task adapter module diagram constructed when the present application is implemented. DETAILED DESCRIPTION

[0021] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below in combination with the drawings and specific embodiments.

[0022] The application is applied to a typical base station-single user communication scene, wherein the base station side is equipped with a linear array multi-antenna, and the user side is equipped with a single antenna. The signal is in the form of an orthogonal frequency-division multiplexing (OFDM) signal. The tasks of model processing include: channel estimation, channel prediction, sub-6G assisted millimeter wave beamforming, user distance estimation, and path loss estimation. During the model running process, the user side sends uplink low-frequency pilot signals to the base station, and the base station obtains the current time uplink low-frequency CSI, future time downlink low-frequency CSI, millimeter wave beamforming vector, user distance information, and path loss information through the proposed physical layer multi-task model, thereby efficiently completing the downlink data transmission task.

[0023] The application includes three steps of data set construction, network construction and training, and real-time deployment. The specific working steps are as follows:

[0024] S10: Construct a MIMO-OFDM multi-task hybrid data set;

[0025] The data sets of multiple wireless communication system physical layer tasks are combined for subsequent model multi-task joint training on the obtained multi-task hybrid data set. Each sample in the multi-task hybrid data set includes pilot signals received by the uplink low-frequency link in the first P time sampling points in a continuous time, CSI data of the downlink low-frequency link in the last L time sampling points, and the distance and path loss of the user and the base station, and the optimal beam sequence number of the millimeter wave downlink, and the sampling intervals of each time sampling point are the same.

[0026] S20: Construct a physical layer multi-task network model and train it on the data set in S10. The physical layer multi-task network includes a preprocessing module, a multi-task adapter module, a pre-trained large model backbone network module, and a multi-task output module. The input of the network is the user uplink low-frequency pilot signal at the first P time, and the output is the target output corresponding to each task, i.e., the user uplink low-frequency CSI at the first P time, the user downlink low-frequency CSI at the last L time, the millimeter wave downlink beam vector, and the distance and path loss of the user to the base station.

[0027] S30: Deploy the trained network to an actual scene or apply it to a test set for testing. Input the user's historical P time uplink low-frequency pilot signal data, and output the output data corresponding to each task.

[0028] In step S10: According to the actual demand, the system related parameters and the data set construction need to be determined. This step includes the following processes S11-S12:

[0029] S11: Firstly, determine the dataset parameters, i.e. the number of antennas of the low-frequency band and the millimeter wave band, the center frequency point, the number of subcarriers and the interval, the CSI time domain sampling interval and the sampling point number, and the configuration of the millimeter wave band communication codebook. Assume that the base station side is equipped with N t = N h × 1 planar array antennas, the millimeter wave band is equipped with planar array antennas, the user side is equipped with a single antenna, the time sampling interval is ΔT, the uplink low-frequency pilot sampling point number is P, and the downlink low-frequency CSI sampling point number is L. The uplink and downlink frequency points of the low-frequency band are f u and f d respectively, and the bandwidths are the same, both of which adopt OFDM symbols, both of which contain K subcarriers and the subcarrier interval is Δf, the uplink and downlink frequency points of the millimeter wave band are and respectively, and the bandwidths are the same, both of which adopt OFDM symbols, both of which contain subcarriers and the subcarrier interval is The low-frequency communication adopts digital precoding, and the millimeter wave communication adopts analog precoding, wherein the uplink pilot adopts a comb pilot distribution, and the pilot interval is m subcarriers, wherein the millimeter wave communication codebook uses a super-resolution discrete Fourier codebook C = {c1, c2, … c B}, wherein represents the b-th steering vector, and B represents the number of codebooks.

[0030] S12: Use Γ = {T1, T2, … T N} to represent the task set to be processed by the model and the corresponding task sub-dataset, wherein N represents the total number of tasks, and in this example N = 5. For each task T i , there is a sub-dataset D i = {TI i , TO i}, wherein TI i represents the input data of the i-th task, and TO i represents the output data of the i-th task. Use to represent the CSI of the k-th subcarrier and the i-th time point of the downlink, use to represent the pilot signal of the k-th subcarrier and the i-th time point of the uplink, and use to represent the CSI of the k-th subcarrier and the i-th time point of the uplink. Then the downlink CSI data of all frequency points at the i-th time point is represented as

[0031]

[0032] where i is an integer in the range of [P+1, P+L]. Similarly, the uplink CSI data at the i-th time instant at all frequency points, and the uplink pilot data at the i-th time instant at the pilot frequency point can be represented as

[0033]

[0034] where i is an integer in the range of [1, P]. Let denote the optimal mmWave beamforming vector of the user at the i-th time instant. Let i ∈C 1 denote the distance between the user and the base station at the i-th time instant. Let i ∈C 1 denote the path loss at the location where the user is at the i-th time instant. Thus for the channel estimation task,

[0035]

[0036] for the channel prediction task,

[0037]

[0038] for the Sub-6G assisted mmWave beamforming task,

[0039]

[0040] for the distance estimation task,

[0041]

[0042] TO4=l P ∈C 1 ,

[0043] for the path loss estimation task,

[0044]

[0045] TO5=pl P ∈C 1

[0046] At step S20: construct the physical layer multi-task network, input TI={TI1, TI2, … TI N}, output TO={TO1, TO2, … TO N}. This step mainly includes the following processes S21-S26:

[0047] Step S21: Since the channel characteristics required by each task are different, uniform preprocessing is not conducive to fully utilizing the characteristics of each task, so each task is designed with a corresponding preprocessing function to preprocess the input data. Let the preprocessed data of the i-th task be i.e.

[0048]

[0049] The prediction processing operation of the channel estimation and channel prediction tasks is to perform Tokenization operation on the CSI at each time, i.e. to perform flattening operation on the spatial and frequency domain features of the CSI as follows:

[0050]

[0051] Where flatten(X, n) means flattening the n-th dimension and all subsequent dimensions of tensor X into one dimension. The beamforming task, distance estimation task and path loss estimation task need channel angular domain features, so the domain transformation operation is performed on the CSI data to convert the spatial domain CSI to the angle domain CSI, i.e.

[0052]

[0053] Where is an N t dimensional DFT matrix.

[0054] Step S22: This step performs feature alignment operation on the preprocessed data, so that the feature space of each task is aligned with the semantic feature space of the pre-trained large model. In order to fully consider the unique characteristics of each task, a multi-task adapter module is defined for each task, denoted as Adapter in,i , whose specific network architecture is shown in Figure 2 . The multi-task adapter module includes a linear mapping layer and a residual connection network. The input and output are described in detail below.

[0055] The input of the multi-task adapter module is the preprocessed data of the i-th task is simply aligned through the linear mapping layer, i.e. mapped to Where L is the sequence length input to the pre-trained LLM, and 768 is the hidden layer dimension of the selected pre-trained GPT-2. That is

[0056]

[0057] Where Linear(·) contains at least two fully connected operations, i.e. linearly mapping the first and second dimensions to the specified dimensions, respectively. Then the feature semantics are aligned through the residual connection network to obtain the task representation of the i-th task That is,

[0058]

[0059] where Res(·) contains two 1-dimensional convolution kernels and an activation function Relu, the convolution kernel size is 3, and the step is 1. The above steps can be simplified as

[0060]

[0061] Step S23: This step inputs the aligned features into the backbone network, i.e., the pre-trained large language model, for processing, corresponding to the pre-trained large language model part in Figure 1 . The backbone network adopts the architecture of GPT-2, and its main body is stacked by multiple layers of decoders of transformer, where each layer of decoder includes a multi-head attention layer, a summation and normalization layer, and a feedforward layer. In actual deployment, the number of layers N L of GPT-2 can be flexibly adjusted, and the MoE-Lora fine-tuning method is used to fine-tune the linear layer weights in the feedforward layer. The final output of the large language model is represented as That is,

[0062]

[0063] where LLM(·) represents the forward inference process of the pre-trained large language model.

[0064] Step S24: This step inputs the features output by the large model into the multi-task adapter again to further realize feature alignment, and obtains the features The specific operation is the same as that in step S22, and can be represented as

[0065]

[0066] where Adapter out,i is the output side of the multi-task adapter corresponding to the i-th task.

[0067] Step S25: This step mainly further processes the output of the adapter to obtain the output result TO′ i of each task. The specific operation is to use a multi-task output head module to map the two dimensions of the i-th task feature output by Adapter out,i to the dimensions of the i-th task output result, respectively. Since the output result dimensions of beamforming, distance estimation, and loss estimation are relatively low, an MLP architecture is used as the multi-task output head module, and since the output dimensions of channel estimation and channel prediction are relatively high, a CNN architecture is used as the multi-task output head module. The specific operation can be represented as

[0068]

[0069] where Output i denotes the output head of the i-th task.

[0070] The output head network is the network after the adapter, which can map the features of each task to the corresponding target output of the task. For example, in the distance estimation task, the output head (MLP) maps the L*D-dimensional feature vector to a 1*1 distance estimation value. [L=K / m, D=768].

[0071] Step S26: Steps S21-S25 describe each module of the network in detail, the input is {T i , i=1, 2, …, 5}, and the output is {T i , i=1, 2, …, 5}. The total loss Loss of network training is the weighted sum of the loss of each task, that is,

[0072]

[0073] where f i l denotes the loss function of task i, and NMSE is used as the loss function for the remaining tasks except for the beam management task which uses the cross-entropy loss function. ω i denotes the proportion of the loss of each task, and the default value is

[0074] In step S30, the network model trained in step S20 is deployed to the actual scene for testing, which includes the following processes S31-S32:

[0075] Step S31: First, the user's historical P-time uplink low-frequency pilot signal data is converted into the tensor format of the network input. Then the converted data is input into the trained network model for testing to obtain the network output data corresponding to each physical layer task.

[0076] Step S32: The tensor data output by the network is converted into a matrix format, and then subsequent precoding operations are performed according to the obtained various channel information to realize online inference.

[0077] The present application can be used to complete channel estimation, channel prediction, distance estimation, path loss estimation, beamforming and other tasks, has low model training cost, can improve the processing accuracy of each task, and has strong generalization ability.

[0078] It is to be understood that the embodiments which have been described are merely illustrative of the principles of this application and that numerous and various modifications can be made by those skilled in the art without departing from the scope of the application as defined in the appended claims.

Claims

1. A method for wireless communication physical layer multitasking based on a pre-trained large language model, characterized in that, A multi-task large model for the physical layer of the wireless communication system is constructed based on the pre-trained large language model and deployed on the base station side. The model is trained using historical user data and then performs online inference after training, thereby completing multiple tasks at the physical layer of the wireless communication system. The steps include: 1) Obtain a multi-task mixed dataset of a wireless communication system as a model training dataset; Each data set sample includes the pilot signal received by the uplink low-frequency link at the first P time sampling points within a continuous period, the channel state information (CSI) data of the downlink low-frequency link at the last L time sampling points, as well as the distance and path loss between the user and the base station, and the optimal beam number of the millimeter wave downlink; 2) Construct a physical layer multi-task network model for a wireless communication system based on a pre-trained large language model, including a pre-processing module, a multi-task adapter module, a pre-trained large model backbone network module, and a multi-task output module; The network model inputs the user's uplink low-frequency pilot signal for the first P time periods, and outputs the target output corresponding to each physical layer task, namely, the user's uplink low-frequency CSI for the first P time periods, the user's downlink low-frequency CSI for the last L time periods, the user's distance from the base station and path loss, and the millimeter-wave downlink beam vector. Using the training data set obtained in step 1), the constructed wireless communication system physical layer multi-task network model is trained to obtain a trained wireless communication system physical layer multi-task network model; 3) Deploy the trained wireless communication system physical layer multi-task network model to the base station side of the wireless communication system, and complete multiple physical layer tasks based on the user's uplink low-frequency pilot signal through online inference; Through the above steps, the physical layer multi-task processing at the base station side of the wireless communication system can be realized.

2. The method of claim 1, wherein the pre-trained large language model is a GPT model. The wireless communication system is a base station-single user communication scenario, where the base station side is equipped with multiple linear array antennas and the user side is equipped with a single antenna; The signal is in the form of orthogonal frequency division multiplexing signal; the wireless communication physical layer tasks processed by the wireless communication system physical layer multi-task network model include: channel estimation, channel prediction, Sub-6G assisted millimeter wave beamforming, user distance estimation, and path loss estimation. 3.The method of claim 1, wherein the pre-trained large language model is a GPT model. In step 1), the sampling intervals of the time sampling points are the same; constructing a wireless communication system physical layer multi-task hybrid data set specifically involves merging data sets of multiple wireless communication system physical layer tasks; including: S11: Determine the training data set parameters, including the number of antennas for the low-frequency band and millimeter-wave band, the center frequency, the number and spacing of subcarriers, the CSI time-domain sampling interval and number of sampling points, and the configuration of the millimeter-wave band communication codebook; S12: defining a set of wireless communication system physical layer tasks Γ = {T1, T2,..., T i ,…,T N} of the model processing and sub-data sets corresponding to the tasks; Each task T in the physical layer task set Γ i , each corresponding to a sub-dataset D i , expressed as: D i ={TI i ,TO i }, where TI i Represents the input data of the i-th task, TO i Represents the output data of the i-th task.

4. The method of claim 3, wherein the pre-trained large language model is a GPT model. Step 2) constructing a wireless communication system physical layer multi-task network model based on a pre-trained large language model, the input of the model is TI={TI1, TI2, … TI N}, and the output is TO={TO1, TO2, … TO N}; the model construction includes the following steps: S21: Design corresponding preprocessing functions to preprocess the input data of different tasks; S22: Performing feature alignment on the preprocessed data so that the feature space of each task is aligned with the semantic feature space of the physical layer multi-task network model built based on the pre-trained large model; Define a multi-task adapter module for each task; the multi-task adapter includes a linear mapping layer and a residual connection network; input the preprocessed data of the i-th task into the multi-task adapter, and first perform a simple dimension alignment through the linear mapping layer; The aligned features are input into a backbone network, i.e., a pre-trained large language model, for processing in S23; The backbone network is stacked by multiple layers of transformer decoders, each of which includes a multi-head attention layer, a summation and normalization layer, and a feedforward layer; The number of layers of the backbone network is adjusted, and the linear layer weights in the feedforward layer are fine-tuned by using a fine-tuning method; After the forward inference process of the large language model, the final output representation of the large language model is obtained; S24: The features output by the large language model are input into a multi-task adapter again to further realize feature alignment and output further aligned features; S25: The output features of the multi-task adapter are further processed by using a multi-task output head module to map the two dimensions of the adapter output features to the dimensions of the corresponding task output data, respectively, to obtain the output results of each task; S26: The loss weighted sum of each task is used as the overall loss function for network model training, and the model is trained to obtain a trained wireless communication system physical layer multi-task network model based on a pre-trained large language model. The mapping through the linear mapping layer includes at least two full connection operations, i.e., linear mapping of the first dimension and the second dimension to specified dimensions, respectively; the residual connection network includes two 1-dimensional convolution kernels and an activation function Relu.

5. The method of claim 4, wherein the pre-trained large language model is a GPT model. The input of the multi-task adapter is First, the dimensions are simply aligned, i.e., mapped to where L is the sequence length input to the pre-trained LLM, and 768 is the hidden layer dimension of the pre-trained GPT-2 selected; and then the feature semantics are aligned through a residual connection network to obtain is expressed as: Adapter in,i The multi-task adapter module corresponding to the i-th task; In S23, the backbone network uses a GPT-2 large language model; the fine-tuning method uses a MoE-Lora fine-tuning method.

6. The method of claim 4, wherein the pre-trained large language model is a GPT model. In S24, the features output by the large model are input into a multi-task adapter again to further realize feature alignment, which is represented as:

7. The method of claim 4, wherein the pre-trained large language model is a GPT model. When using the trained wireless communication physical layer multi-task processing model for offline prediction, first, the user's historical P-time uplink low-frequency pilot signal data is converted into the format of network input by tensorization operation. wherein, is a final output feature after feature alignment; Adapter out,i is a multi-task adapter on the output side corresponding to the i-th task. 8.The method of claim 1, wherein the method further comprises: ​

Images