A 6g wireless communication predicted channel modeling method based on large language model fine tuning

Abstract

The application discloses a 6G wireless communication prediction channel modeling method based on large language model fine tuning, relates to the technical field of channel prediction, and comprises the following steps: processing channel measurement data, matching corresponding text data, dividing a training set and a test set, and constructing a channel prediction data set; designing a channel encoder and a dual-domain fusion module to extract channel features, and designing a text-driven encoder to extract text features; fine tuning a large language model by using the extracted channel and text features, enhancing multi-modal perception and transfer learning capability; designing a fine tuning module, fine tuning the output of the large language model, and projecting to predicted future channel state information; designing an angle consistency loss function, training a prediction algorithm based on large language model fine tuning in combination with prediction loss, and obtaining a trained network architecture; and iteratively predicting the space-time domain channel state by using the trained network, and outputting a channel prediction result. The system has high-precision prediction performance, and has outstanding practical value and popularization prospect.

Description

Technical Field

[0001] This invention relates to the field of channel prediction technology, and in particular to a 6G wireless communication prediction channel modeling method based on large language model fine-tuning. Background Technology

[0002] With the rapid development of sixth-generation (6G) mobile communication and vehicular networks, wireless communication systems face severe challenges from complex time-varying channel environments. Facing the development focus of 6G technology—"global coverage, full spectrum, full application, full sensing, full digitalization, and strong security"—future communication systems will exhibit characteristics such as increased antenna numbers, expanded user base, and more complex service types, leading to increased pilot overhead and system complexity. Channel State Information (CSI) directly affects the performance of 6G technologies such as beamforming, adaptive modulation and coding, and dynamic spectrum allocation. Therefore, low-overhead, low-complexity, and high-accuracy CSI acquisition methods are indispensable. Traditional channel estimation methods utilize pilot signals within the communication system to obtain CSI. However, with the trend towards higher frequencies, greater mobility, and larger antenna arrays, channel estimation faces challenges of limited accuracy, high latency, and high computational complexity.

[0003] On the other hand, the close coordination between the physical environment and channel characteristics, coupled with the enhanced ability to perceive environmental information through multimodal data, makes CSI prediction possible. Predictive channel modeling utilizes historical CSI to predict future channel states. Traditional methods mainly rely on physical models, which suffer from insufficient modeling accuracy and excessive computational complexity. Neural network methods, limited by the number of parameters, exhibit certain shortcomings in prediction accuracy and generalization ability.

[0004] Large Language Models (LLMs) offer a new paradigm for deep learning by fine-tuning a limited number of parameters to adapt to various downstream tasks. In wireless communication, the complex real-world propagation environment poses a major challenge to CSI prediction. LLMs' multimodal sensing capabilities enable them to simultaneously process data from channels, images, and text, providing multi-perspective environmental context information. Furthermore, LLMs exhibit excellent generalization capabilities, maintaining robust predictive performance even in data-constrained scenarios, which is particularly relevant for practical deployment. The excellent transfer learning capabilities of LLMs lay the foundation for fine-tuning. They facilitate knowledge transfer from the textual domain to different channel environments, aiding in rapid model adaptation. LLMs support multi-task learning, enabling various tasks such as channel prediction and beamforming prediction within a unified framework, thus ensuring system flexibility and efficiency. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a 6G wireless communication prediction channel modeling method based on large language model fine-tuning, so as to solve the technical problems of difficulty in obtaining channel state information and insufficient prediction accuracy, achieve efficient and accurate channel state information prediction, and have strong generalization performance.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A 6G wireless communication prediction channel modeling method based on large language model fine-tuning proposed in this invention includes:

[0008] Step S1: Normalize the channel measurement data, match and normalize the text data corresponding to the surrounding environment of the channel measurement data, and divide the normalized channel measurement data and text data into training set and test set. The training set and test set are used to construct a channel prediction dataset.

[0009] Step S2: Extract channel features from the normalized channel measurement data using a channel encoder and a dual-domain fusion module, and extract text features from the text data using a text-driven encoder.

[0010] Step S3: Fine-tune the large language model using the extracted channel features and text features;

[0011] Step S4: Use the fine-tuning module to adjust the output of the fine-tuned large language model, and project the adjusted output of the large language model onto the predicted future channel state information;

[0012] Step S5: Design the angle consistency loss function. The angle consistency loss function is combined with the mean square error prediction loss function to train the prediction algorithm based on the fine-tuning of the large language model on the training set, so as to obtain the trained and fine-tuned network architecture of the large language model.

[0013] Step S6: Using the trained large language model network architecture, iteratively predict the spatiotemporal channel state on the test set and output the channel prediction results.

[0014] As a further optimization scheme of the 6G wireless communication prediction channel modeling method based on large language model fine-tuning described in this invention, step S2 includes:

[0015] S201. Through one-dimensional discrete Fourier transform, the channel state information is transformed from the time delay domain to the frequency domain:

[0016]

[0017] in, and These represent the channel state information in the time delay domain and the frequency domain, respectively. Represents the Discrete Fourier Transform;

[0018] S202, Design a channel encoder to extract channel state information in the time delay domain. Channel state information in the frequency domain The deep features are as follows:

[0019] S2021. Extract the real and imaginary parts of the channel state information in the time delay domain and frequency domain respectively, and concatenate them. The specific formula is as follows:

[0020]

[0021]

[0022] in, and These represent the channel state information in the time delay domain and frequency domain, respectively, after concatenation. This indicates a serial operation. and These represent taking the real and imaginary parts of a complex matrix, respectively.

[0023] S2022. Using convolution calculations and activation functions, the concatenated time-delay and frequency-domain features are projected into a high-dimensional space to extract the corresponding high-dimensional features. The specific formula is as follows:

[0024]

[0025]

[0026] in, and These represent the extracted high-dimensional channel features in the time delay domain and frequency domain, respectively. This represents convolution calculation. This represents the LeakyReLU activation function;

[0027] S203. Design a dual-domain fusion module. The dual-domain fusion module fuses the extracted high-dimensional features in the time delay domain and frequency domain to obtain the fused channel features. The fused channel features refer to the channel features after fusing the high-dimensional features in the time delay domain and the high-dimensional features in the frequency domain. The specific content is as follows:

[0028] S2031, the dual-domain fusion module includes a parallel adaptive convolution module and a dual-domain conditional injection module;

[0029] S2032. The high-dimensional channel features extracted from the time delay domain and frequency domain are concatenated, and a preliminary fusion is performed using a convolution operation. The specific formula is as follows:

[0030]

[0031] in, This represents the channel characteristics after initial fusion;

[0032] S2033, to A parallel adaptive convolution module is used to extract adaptive coefficients and adaptive features. Convolutions with different window sizes are then used to extract high-dimensional features at multiple scales. Adaptive coefficients are used to weight these multi-scale features. The specific formula is as follows:

[0033]

[0034]

[0035]

[0036] in, , These represent three different adaptive coefficients. This represents the extracted adaptive features in the time-delay domain; Indicates a separation operation; Indicates the convolution window is Convolution calculation, The values ​​are 1, 3, and 5; This indicates that the corresponding window size is The time-delay domain features obtained after convolution calculation at different scales This represents the weighted adaptive feature in the time-delay domain;

[0037] S2034, to A parallel adaptive convolution module is used to extract adaptive coefficients and adaptive features. High-dimensional features at multiple scales are then extracted from the adaptive features. The adaptive coefficients are used to weight these multi-scale features. The specific formula is as follows:

[0038]

[0039]

[0040]

[0041] in, This represents the extracted adaptive features in the frequency domain. This indicates that the corresponding window size is The frequency domain features at different scales obtained after convolution calculation This represents the weighted frequency domain adaptive feature;

[0042] S2035. Using the dual-domain conditional injection module, adaptive features in the time delay domain and frequency domain are injected into the preliminary fusion features; first, ... and The affine transformation weights and biases are calculated using convolution after concatenation. The specific formula is as follows:

[0043]

[0044] in, and These are the coefficients and bias of the affine transformation, respectively; then, layer normalization is used to adjust the distribution of the initial fused features, with the specific formula as follows:

[0045]

[0046] in, This represents the fusion characteristics after normalization. The representation layer undergoes normalization; after convolution calculation on the normalized fused features, an affine transformation is performed, followed by another convolution operation, and then a residual connection is established with the unnormalized fused features. The specific formula is as follows:

[0047]

[0048]

[0049] in, Represents affine transformation, This represents the element-wise product of a matrix. This represents the fusion features after dual-domain conditional injection;

[0050] A feedforward neural network consisting of convolutional layers and activation functions is used to model the fused features after conditional injection. The intrinsic relationship is expressed by the following formula:

[0051]

[0052] in, This represents the channel characteristics after fine-tuning. This represents the GELU activation function;

[0053] S204. Design a text-driven encoder to extract text features, embed the text data into a vector space, and adjust the distribution of the text vectors to obtain the specific content of the text features:

[0054] S2041. Using a pre-trained Sentence Transformer model, the text data is projected into a feature space vector. The specific formula is as follows:

[0055]

[0056] in, This represents the embedded text vector. This represents the Sentence Transformer model. Represents text data;

[0057] S2042. By redistributing the text vectors, using the L2 norm and linear transformation, the distribution is adjusted. The specific formula is as follows:

[0058]

[0059] in, This represents the text features obtained after adjusting the distribution. and Represents the trainable weights and biases of a linear transformation. This represents the L2 norm.

[0060] As a further optimization scheme for the 6G wireless communication prediction channel modeling method based on large language model fine-tuning described in this invention, step S3 involves fine-tuning the large language model using the extracted channel and text features to enhance multimodal perception and transfer learning capabilities, including the following steps:

[0061] S301, Channel characteristics and text features The concatenation, as input to the large language model, utilizes the multimodal perception and transfer learning of the large language model to obtain the output features after information exchange, and separates the channel features and text features after information exchange from them. The specific formula is as follows:

[0062]

[0063]

[0064] in, This represents the output of a large language model. Representing a large language model, and These represent the channel features and text features separated from the output of the large language model, respectively.

[0065] As a further optimization scheme for the 6G wireless communication prediction channel modeling method based on large language model fine-tuning described in this invention, step S4 includes:

[0066] S401. Using a linear layer, the output channel features are divided into fine-grained features and coarse-grained features. Then, a feedforward neural network is applied to the fine-grained features, while keeping the coarse-grained features unchanged. The feedforward neural network consists of two linear layers and a LeakyReLU activation function. Finally, the fine-grained and coarse-grained features are combined through residual connections, as shown in the following formula:

[0067]

[0068]

[0069] in, This represents the combined channel characteristics. and These are the trainable parameters of the corresponding linear layer. This represents a feedforward neural network. and These represent fine-grained and coarse-grained features, respectively.

[0070] S402. Apply a linear layer and rearrangement operation to the combined features to project back the predicted future channel state information. The specific formula is as follows:

[0071]

[0072]

[0073] in, This represents the real and imaginary parts of the predicted channel state information after projection. This represents the final predicted complex channel state information. and These are the trainable parameters of the corresponding linear layer. It is the imaginary unit. This indicates a rearrangement operation.

[0074] As a further optimization scheme of the 6G wireless communication prediction channel modeling method based on large language model fine-tuning described in this invention, step S5 includes:

[0075] S501. Simultaneously input the channel measurement data and text data from the training set into the 6G wireless communication prediction channel modeling algorithm based on large language model fine-tuning to obtain the predicted channel state information. The specific formula is as follows:

[0076]

[0077] in, Represents the parameters of the network. This represents a 6G wireless communication predictive channel modeling algorithm based on fine-tuning of a large language model.

[0078] S502. Design the angle consistency loss function, and jointly calculate the difference between the predicted channel state information and the actual channel state information using the prediction loss calculation. The expression for the loss function is:

[0079]

[0080] in, Represents the overall loss function. This represents the prediction loss function. This represents the angle consistency loss function. The balancing factor is used; the prediction loss is calculated using the mean square error to determine the difference between the predicted channel state information and the actual channel state information. The specific formula is as follows:

[0081]

[0082] in, Represents the actual channel state information, This indicates the calculation of the mean square error; the angle consistency loss uses Fourier transform to convert the predicted channel state information and the true channel state information to the angle domain, and then calculates the mean square error of the amplitude spectrum of the two in the angle domain. The specific formula is as follows:

[0083]

[0084]

[0085]

[0086] in, and These represent the amplitude-frequency response of the actual channel state information and the predicted channel state information in the angle domain, respectively. Indicates the modulus of the complex number;

[0087] S503. Calculate the gradient of the trainable parameters of the 6G wireless communication prediction channel modeling method based on large language model fine-tuning using backpropagation of the loss function, and update the parameters. The specific formula is as follows:

[0088]

[0089] in, This indicates the updated network parameters. Indicates the learning rate. The gradient represents the network parameters;

[0090] S504. Repeat steps S501-S503 until the loss function converges. At this point, the training ends, and the trained 6G wireless communication prediction channel modeling method based on large language model fine-tuning is obtained.

[0091] As a further optimization scheme of the 6G wireless communication prediction channel modeling method based on large language model fine-tuning described in this invention, step S6 includes:

[0092] The test set from the channel prediction dataset is input into the trained 6G wireless communication prediction channel modeling algorithm based on large language model fine-tuning. Continuous future channel state information is obtained through autoregressive iterative prediction. .

[0093] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predictive channel modeling of 6G wireless communication based on fine-tuning of a large language model.

[0094] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0095] This invention first preprocesses the acquired measured channel data of the scenario to be predicted to generate corresponding text data, thus obtaining a training channel dataset. The training channel dataset is then input into a pre-defined channel encoder, a dual-domain fusion module, and a text-driven encoder to extract corresponding multimodal features. These features are then input into a large language model for information exchange, and the channel prediction result is obtained through a fine-tuning module. This application utilizes a fine-tuned large language model to achieve efficient and accurate channel state information prediction, and it still exhibits strong generalization ability in different environments. Attached Figure Description

[0096] Figure 1 This is a schematic diagram of a process for predictive channel modeling of 6G wireless communication based on fine-tuning of a large language model, provided in an embodiment of this application.

[0097] Figure 2 This is a schematic diagram of the structure of a 6G wireless communication predictive channel modeling algorithm based on large language model fine-tuning provided in an embodiment of this application;

[0098] Figure 3 This is a schematic diagram of a dual-domain fusion module structure provided in an embodiment of this application; wherein, (a) is a parallel adaptive convolution module, and (b) is a dual-domain conditional injection module;

[0099] Figure 4 This is a schematic diagram of a text-driven encoder structure provided in an embodiment of this application;

[0100] Figure 5 These are the average simulation results for scenario one; where (a) is the average MAE result and (b) is the average RMSE result.

[0101] Figure 6 These are the average simulation results for scenario two; where (a) is the average MAE result and (b) is the average RMSE result.

[0102] Figure 7 The results are the average simulation results for scenario three; (a) is the average MAE result and (b) is the average RMSE result. Detailed Implementation

[0103] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0104] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0105] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0106] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0107] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0108] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0109] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0110] Figure 1 The illustration shows a flowchart of a 6G wireless communication prediction channel modeling method based on large language model fine-tuning provided in an embodiment of this application.

[0111] like Figure 1 As shown, this 6G wireless communication prediction channel modeling method based on large language model fine-tuning includes the following steps:

[0112] S1. Normalize the channel measurement data, match the text data corresponding to the surrounding environment of the normalized channel measurement data, and divide the normalized channel measurement data and text data into training set and test set. The training set and test set are used to construct a channel prediction dataset, including the following:

[0113] S101. Using text to describe the channel measurement data, the matched channel prediction dataset is divided into a training set and a test set in an 8:2 ratio. It should be noted that there are multiple methods to obtain the measured channel data for the scenario to be predicted, and the specific acquisition method is not limited. For example, the channel measurement environment is a suburban scene near a building. The receiver is placed at a fixed location within the building, while the transmitter's location changes continuously with the scene's location. The measurement frequencies are 7 GHz, 10 GHz, and 13 GHz. Each set of data records the measured channel matrix at different path points with the building as the origin.

[0114] S2. Channel features are extracted from normalized channel measurement data using a channel encoder and a dual-domain fusion module, and text features are extracted from text data using a text-driven encoder. The structure of the 6G wireless communication prediction channel modeling method based on large language model fine-tuning is as follows: Figure 2 As shown, it includes the following:

[0115] S201. Transform the channel state information from the time delay domain to the frequency domain using one-dimensional discrete Fourier transform:

[0116]

[0117] in, and These represent the channel state information in the time delay domain and the frequency domain, respectively. This represents the Discrete Fourier Transform.

[0118] S202, Design a channel alignment encoder to extract time-delay domain channel state information. and frequency domain channel state information The deep features are as follows:

[0119] S2021. Extract the real and imaginary parts of the channel state information in the time delay domain and frequency domain respectively, and concatenate them. The specific formula is as follows:

[0120]

[0121]

[0122] in, and These represent the channel state information in the time delay domain and frequency domain, respectively, after concatenation. This indicates a serial operation. and These represent taking the real and imaginary parts of a complex matrix, respectively.

[0123] S2022. Using convolution calculations and activation functions, the concatenated time-delay and frequency-domain features are projected into a high-dimensional space to extract the corresponding high-dimensional features. The specific formula is as follows:

[0124]

[0125]

[0126] in, and These represent the extracted high-dimensional channel features in the time delay domain and frequency domain, respectively. This represents convolution calculation. This represents the LeakyReLU activation function. In one embodiment, the kernel size is 3×3, and the hidden state dimensions are 64 and 768, respectively.

[0127] S203. Design a dual-domain fusion module to fuse the extracted high-dimensional features in the time delay domain and frequency domain to obtain the fused channel features. The specific details are as follows:

[0128] S2031, Dual-Domain Fusion Module Structure as follows Figure 3 As shown, Figure 3 This is a schematic diagram of the dual-domain fusion module structure provided in an embodiment of this application; Figure 3 (a) in the diagram represents the parallel adaptive convolution module. Figure 3 (b) in the diagram represents the dual-domain conditional injection module. It includes a parallel adaptive convolution module and a dual-domain conditional injection module. In one embodiment, a single-layer dual-domain fusion module is used, specifically:

[0129] S2032. The high-dimensional channel features extracted from the time delay domain and frequency domain are concatenated, and a preliminary fusion is performed using a convolution operation. The specific formula is as follows:

[0130]

[0131] in, This indicates the channel characteristics after initial fusion.

[0132] S2033, to A parallel adaptive convolution module is used to extract adaptive coefficients and adaptive features. Convolutions with different window sizes are then used to extract high-dimensional features at multiple scales. Adaptive coefficients are used to weight these multi-scale features. The specific formula is as follows:

[0133]

[0134]

[0135]

[0136] in, , These represent three different adaptive coefficients. This represents the extracted adaptive features in the time-delay domain. This indicates a separation operation. Indicates the convolution window is Convolution calculation, The values ​​are 1, 3, and 5. This indicates that the corresponding window size is The time-delay domain features obtained after convolution calculation at different scales This represents the weighted time-delay domain adaptive feature.

[0137] S2034, to A parallel adaptive convolution module is used to extract adaptive coefficients and adaptive features. High-dimensional features at multiple scales are then extracted from the adaptive features. The adaptive coefficients are used to weight these multi-scale features. The specific formula is as follows:

[0138]

[0139]

[0140]

[0141] in, This represents the extracted adaptive features in the frequency domain. This indicates that the corresponding window size is The frequency domain features at different scales obtained after convolution calculation This represents the weighted frequency domain adaptive feature.

[0142] S2035. Using the dual-domain conditional injection module, adaptive features in the time delay domain and frequency domain are injected into the preliminary fusion features. First, ... and The affine transformation weights and biases are calculated using convolution after concatenation. The specific formula is as follows:

[0143]

[0144] in, and These represent the coefficients and bias of the affine transformation, respectively. Then, a layer normalization operation is used to adjust the distribution of the initial fused features; the specific formula is as follows:

[0145]

[0146] in, This represents the fusion characteristics after normalization. This represents the layer normalization operation. After applying convolution to the normalized fused features, an affine transformation is performed, followed by another convolution operation. Then, a residual connection is established between this convolution and the unnormalized fused features. The specific formula is as follows:

[0147]

[0148]

[0149] in, Represents affine transformation, This represents the element-wise product of a matrix. This represents the fused features after dual-domain conditional injection. A feedforward neural network consisting of convolutional layers and activation functions is used to model the features after conditional injection. The intrinsic relationship is expressed by the following formula:

[0150]

[0151] in, This represents the channel characteristics after fine-tuning. This represents the GELU activation function.

[0152] S204. Design a text-driven encoder to extract text features, embed text data into a vector space, and adjust the distribution of text vectors to obtain text features. The structure of the text-driven encoder is as follows: Figure 4 As shown, the specific content is as follows:

[0153] S2041. Using a pre-trained Sentence Transformer model, the text data is projected into a feature space vector. The specific formula is as follows:

[0154]

[0155] in, This represents the embedded text vector. This represents the Sentence Transformer model. This represents text data. In one embodiment, the size of the text vector is... .

[0156] S2042. By redistributing the text vectors, using the L2 norm and linear transformation, the distribution is adjusted. The specific formula is as follows:

[0157]

[0158] in, This represents the text features obtained after adjusting the distribution. and Represents the trainable weights and biases of a linear transformation. This represents the L2 norm.

[0159] S3. Fine-tune the large language model using the extracted channel and text features to enhance multimodal perception and transfer learning capabilities, including the following:

[0160] S301, Channel characteristics and text features The concatenation, as input to the large language model, utilizes the multimodal perception and transfer learning of the large language model to obtain the output features after information exchange, and separates the channel features and text features after information exchange from them. The specific formula is as follows:

[0161]

[0162]

[0163] in, This represents the output of a large language model. Represents a large language model. and These represent the channel features and text features separated from the output of the large language model, respectively. In one embodiment, a GPT-2 model with a hidden layer dimension of 768 is used as the skeleton of the large language model, and the GPT-2 model has 6 layers.

[0164] S4. The fine-tuning module is used to adjust the output of the fine-tuned large language model, and the adjusted output of the large language model is projected onto the predicted future channel state information, including the following:

[0165] S401. Using a linear layer, the output channel features are divided into fine-grained and coarse-grained features. Next, a feedforward neural network is applied to the fine-grained features while keeping the coarse-grained features unchanged. The feedforward neural network consists of two linear layers and a LeakyReLU activation function. Then, the fine-grained and coarse-grained features are combined through residual connections, as shown in the following formula:

[0166]

[0167]

[0168] in, This represents the combined channel characteristics. and These are the trainable parameters of the corresponding linear layer. This represents a feedforward neural network. and These represent fine-grained and coarse-grained features, respectively. In one embodiment, a 4-layer feedforward neural network is used.

[0169] S402. Apply a linear layer and rearrangement operation to the combined features to project back the predicted future channel state information. The specific formula is as follows:

[0170]

[0171]

[0172] in, This represents the real and imaginary parts of the predicted channel state information after projection. This represents the final predicted complex channel state information. and These are the trainable parameters of the corresponding linear layer. It is the imaginary unit. This indicates a rearrangement operation.

[0173] S5. Design an angle consistency loss function. Combine the angle consistency loss function with the mean squared error prediction loss function to train a prediction algorithm based on fine-tuning of the large language model on the training set, resulting in the trained, fine-tuned large language model network architecture, including the following:

[0174] S501. Simultaneously input the channel data and text data from the training set into the 6G wireless communication prediction channel modeling algorithm based on large language model fine-tuning to obtain the predicted channel state information. The specific formula is as follows:

[0175]

[0176] in, Represents the parameters of the network. This represents a 6G wireless communication predictive channel modeling algorithm based on fine-tuning of a large language model.

[0177] S502. Design the angle consistency loss function, and jointly calculate the difference between the predicted channel state information and the actual channel state information using the prediction loss calculation. The expression for the loss function is:

[0178]

[0179] in, Represents the overall loss function. This represents the prediction loss function. This represents the angle consistency loss function. This is a balance factor. In one embodiment, The value is 0.7. The prediction loss is calculated using the mean square error to determine the difference between the predicted channel state information and the actual channel state information. The specific formula is as follows:

[0180]

[0181] in, Represents the actual channel state information, This indicates the calculation of the mean square error. The angle consistency loss uses Fourier transform to convert the predicted channel state information and the true channel state information to the angle domain, and then calculates the mean square error of their amplitude spectra in the angle domain. The specific formula is:

[0182]

[0183]

[0184]

[0185] in, and These represent the amplitude-frequency response of the real channel state information and the predicted channel state information in the angle domain, respectively. This indicates the modulus of the complex number.

[0186] S503. Calculate the gradient of the trainable parameters of the 6G wireless communication prediction channel modeling method based on large language model fine-tuning using backpropagation of the loss function, and update the parameters. The specific formula is as follows:

[0187]

[0188] in, This indicates the updated network parameters. Indicates the learning rate. This represents the gradient of the network parameters. In one embodiment, The value is 0.0003.

[0189] S504. Repeat steps S501-S503 until the loss function converges. At this point, the training ends, and the trained 6G wireless communication prediction channel modeling method based on large language model fine-tuning is obtained.

[0190] S6. Using the trained large language model network architecture, iteratively predict the spatiotemporal channel state on the test set, and output the channel prediction results, including the following:

[0191] S601. Input the test set from the channel prediction dataset into the trained 6G wireless communication prediction channel modeling algorithm based on large language model fine-tuning, and obtain continuous future channel state information through autoregressive iterative prediction. .

[0192] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0193] In one embodiment, the proposed method is used to perform channel prediction accuracy analysis against the most commonly used space-time predictive channel modeling networks. The comparison metrics are the MAE and RMSE of the delay power spectrum. The performance of different predictive channel modeling methods under different obstruction environments is also compared. Using multi-band, multi-scenario channel measurement data, the performance of different prediction models is evaluated. The average simulation results at different times in three frequency bands under three different scenarios are shown below. Figure 5 As shown in 6 and 7, Figure 5 This is the average simulation result for scenario one; Figure 5 (a) in the figure is the average MAE result. Figure 5 (b) in the figure shows the average RMSE results; Figure 6 This is the average simulation result diagram for scenario two; Figure 6 (a) in the figure is the average MAE result. Figure 6 (b) in the figure shows the average RMSE results; Figure 7 This is the average simulation result diagram for scenario three; Figure 7 (a) in the figure is the average MAE result. Figure 7 (b) in the figure shows the average RMSE result. By comparing the test results in different scenarios, it can be seen that the model of this application maintains high prediction accuracy in different scenarios, surpassing existing space-time prediction channel modeling methods.

[0194] Secondly, embodiments of this application provide a predictive channel modeling device based on a deep learning algorithm, comprising: a preprocessing module for preprocessing the acquired measured channel data of the scenario to be predicted, generating text data, and constructing a channel dataset; a training module for training a channel prediction model using the channel dataset through gradient descent of a loss function; and a prediction module for inputting the historical channel data into the trained channel prediction model to obtain a channel prediction result.

[0195] Thirdly, embodiments of this application provide a terminal device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned 6G wireless communication prediction channel modeling method based on large language model fine-tuning.

[0196] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned 6G wireless communication prediction channel modeling method based on large language model fine-tuning.

[0197] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A 6G wireless communication predictive channel modeling method based on large language model fine-tuning, characterized in that, include: Step S1: Normalize the channel measurement data, match and normalize the text data corresponding to the surrounding environment of the channel measurement data, and divide the normalized channel measurement data and text data into training set and test set. The training set and test set are used to construct a channel prediction dataset. Step S2: Extract channel features from the normalized channel measurement data using a channel encoder and a dual-domain fusion module, and extract text features from the text data using a text-driven encoder. Step S3: Fine-tune the large language model using the extracted channel features and text features; Step S4: Use the fine-tuning module to adjust the output of the fine-tuned large language model, and project the adjusted output of the large language model onto the predicted future channel state information; Step S5: Design the angle consistency loss function. The angle consistency loss function is combined with the mean square error prediction loss function to train the prediction algorithm based on the fine-tuning of the large language model on the training set, so as to obtain the trained and fine-tuned network architecture of the large language model. Step S6: Using the trained large language model network architecture, iteratively predict the spatiotemporal channel state on the test set and output the channel prediction results.

2. The 6G wireless communication predictive channel modeling method based on large language model fine-tuning according to claim 1, characterized in that, Step S2 includes: S201. Through one-dimensional discrete Fourier transform, the channel state information is transformed from the time delay domain to the frequency domain: ； in, and These represent the channel state information in the time delay domain and the frequency domain, respectively. Represents the Discrete Fourier Transform; S202, Design a channel encoder to extract channel state information in the time delay domain. Channel state information in the frequency domain The deep features are as follows: S2021. Extract the real and imaginary parts of the channel state information in the time delay domain and frequency domain respectively, and concatenate them. The specific formula is as follows: ； ； in, and These represent the channel state information in the time delay domain and frequency domain, respectively, after concatenation. This indicates a serial operation. and These represent taking the real and imaginary parts of a complex matrix, respectively. S2022. Using convolution calculations and activation functions, the concatenated time-delay and frequency-domain features are projected into a high-dimensional space to extract the corresponding high-dimensional features. The specific formula is as follows: ； ； in, and These represent the extracted high-dimensional channel features in the time delay domain and frequency domain, respectively. This represents convolution calculation. This represents the LeakyReLU activation function; S203. Design a dual-domain fusion module. The dual-domain fusion module fuses the extracted high-dimensional features in the time delay domain and frequency domain to obtain the fused channel features. The fused channel features refer to the channel features after fusing the high-dimensional features in the time delay domain and the high-dimensional features in the frequency domain. The specific content is as follows: S2031, the dual-domain fusion module includes a parallel adaptive convolution module and a dual-domain conditional injection module; S2032. The high-dimensional channel features extracted from the time delay domain and frequency domain are concatenated, and a preliminary fusion is performed using a convolution operation. The specific formula is as follows: ； in, This represents the channel characteristics after initial fusion; S2033, to A parallel adaptive convolution module is used to extract adaptive coefficients and adaptive features. Convolutions with different window sizes are then used to extract high-dimensional features at multiple scales. Adaptive coefficients are used to weight these multi-scale features. The specific formula is as follows: ； ； ； in, , These represent three different adaptive coefficients. This represents the extracted adaptive features in the time-delay domain; Indicates a separation operation; Indicates the convolution window is Convolution calculation, The values ​​are 1, 3, and 5; This indicates that the corresponding window size is The time-delay domain features obtained after convolution calculation at different scales This represents the weighted adaptive feature in the time-delay domain; S2034, to A parallel adaptive convolution module is used to extract adaptive coefficients and adaptive features. High-dimensional features at multiple scales are then extracted from the adaptive features. The adaptive coefficients are used to weight these multi-scale features. The specific formula is as follows: ； ； ； in, This represents the extracted adaptive features in the frequency domain. This indicates that the corresponding window size is The frequency domain features at different scales obtained after convolution calculation This represents the weighted frequency domain adaptive feature; S2035. Using the dual-domain conditional injection module, adaptive features in the time delay domain and frequency domain are injected into the preliminary fusion features; first, ... and The affine transformation weights and biases are calculated using convolution after concatenation. The specific formula is as follows: ； in, and These are the coefficients and bias of the affine transformation, respectively; then, layer normalization is used to adjust the distribution of the initial fused features, with the specific formula as follows: ； in, This represents the fusion characteristics after normalization. The representation layer undergoes normalization; after convolution calculation on the normalized fused features, an affine transformation is performed, followed by another convolution operation, and then a residual connection is established with the unnormalized fused features. The specific formula is as follows: ； ； in, Represents affine transformation, This represents the element-wise product of a matrix. This represents the fusion features after dual-domain conditional injection; A feedforward neural network consisting of convolutional layers and activation functions is used to model the fused features after conditional injection. The intrinsic relationship is expressed by the following formula: ； in, This represents the channel characteristics after fine-tuning. This represents the GELU activation function; S204. Design a text-driven encoder to extract text features, embed the text data into a vector space, and adjust the distribution of the text vectors to obtain the specific content of the text features: S2041. Using a pre-trained Sentence Transformer model, the text data is projected into a feature space vector. The specific formula is as follows: ； in, This represents the embedded text vector. This represents the Sentence Transformer model. Represents text data; S2042. By redistributing the text vectors, using the L2 norm and linear transformation, the distribution is adjusted. The specific formula is as follows: ； in, This represents the text features obtained after adjusting the distribution. and Represents the trainable weights and biases of a linear transformation. This represents the L2 norm.

3. The 6G wireless communication prediction channel modeling method based on large language model fine-tuning according to claim 1, characterized in that, In step S3, the extracted channel and text features are used to fine-tune the large language model to enhance its multimodal perception and transfer learning capabilities. This includes the following steps: S301, Channel characteristics and text features The concatenation, as input to the large language model, utilizes the multimodal perception and transfer learning of the large language model to obtain the output features after information exchange, and separates the channel features and text features after information exchange from them. The specific formula is as follows: ； ； in, This represents the output of a large language model. Representing a large language model, and These represent the channel features and text features separated from the output of the large language model, respectively.

4. The 6G wireless communication prediction channel modeling method based on large language model fine-tuning according to claim 3, characterized in that, Step S4 includes: S401. Using a linear layer, the output channel features are divided into fine-grained features and coarse-grained features. Then, a feedforward neural network is applied to the fine-grained features, while keeping the coarse-grained features unchanged. The feedforward neural network consists of two linear layers and a LeakyReLU activation function. Finally, the fine-grained and coarse-grained features are combined through residual connections, as shown in the following formula: ； ； in, This represents the combined channel characteristics. and These are the trainable parameters of the corresponding linear layer. This represents a feedforward neural network. and These represent fine-grained and coarse-grained features, respectively. S402. Apply a linear layer and rearrangement operation to the combined features to project back the predicted future channel state information. The specific formula is as follows: ； ； in, This represents the real and imaginary parts of the predicted channel state information after projection. This represents the final predicted complex channel state information. and These are the trainable parameters of the corresponding linear layer. It is the imaginary unit. This indicates a rearrangement operation.

5. The 6G wireless communication prediction channel modeling method based on large language model fine-tuning according to claim 4, characterized in that, Step S5 includes: S501. Simultaneously input the channel measurement data and text data from the training set into the 6G wireless communication prediction channel modeling algorithm based on large language model fine-tuning to obtain the predicted channel state information. The specific formula is as follows: ； in, Represents the parameters of the network. This represents a 6G wireless communication predictive channel modeling algorithm based on fine-tuning of a large language model. S502. Design the angle consistency loss function, and jointly calculate the difference between the predicted channel state information and the actual channel state information using the prediction loss calculation. The expression for the loss function is: ； in, Represents the overall loss function. This represents the prediction loss function. This represents the angle consistency loss function. The balancing factor is used; the prediction loss is calculated using the mean square error to determine the difference between the predicted channel state information and the actual channel state information. The specific formula is as follows: ； in, Represents the actual channel state information, This indicates the calculation of the mean square error; the angle consistency loss uses Fourier transform to convert the predicted channel state information and the true channel state information to the angle domain, and then calculates the mean square error of the amplitude spectrum of the two in the angle domain. The specific formula is as follows: ； ； ； in, and These represent the amplitude-frequency response of the actual channel state information and the predicted channel state information in the angle domain, respectively. Indicates the modulus of the complex number; S503. Calculate the gradient of the trainable parameters of the 6G wireless communication prediction channel modeling method based on large language model fine-tuning using backpropagation of the loss function, and update the parameters. The specific formula is as follows: ； in, This indicates the updated network parameters. Indicates the learning rate. The gradient represents the network parameters; S504. Repeat steps S501-S503 until the loss function converges. At this point, the training ends, and the trained 6G wireless communication prediction channel modeling method based on large language model fine-tuning is obtained.

6. The 6G wireless communication prediction channel modeling method based on large language model fine-tuning according to claim 5, characterized in that, Step S6 includes: The test set from the channel prediction dataset is input into the trained 6G wireless communication prediction channel modeling algorithm based on large language model fine-tuning. Continuous future channel state information is obtained through autoregressive iterative prediction. .

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the 6G wireless communication prediction channel modeling method based on large language model fine-tuning as described in any one of claims 1 to 6.

Images