Perceptual auxiliary channel prediction method and system based on pre-trained large language model

Abstract

The invention relates to the technical field of wireless communication, in particular to a perceptual auxiliary channel prediction method and system based on a pre-trained large language model. The invention provides perception auxiliary channel prediction based on a pre-trained large language model, on one hand, inherent environment correlation between a single-base-station perception channel and a double-base-station communication channel is utilized, a perception channel is taken into consideration, and correct modeling is performed on the communication channel and the perception channel so as to determine the inherent time-space relationship; and on the other hand, a perception auxiliary channel prediction network based on the pre-trained large language model is constructed, a perception channel is used as assistance, and future communication CSI data is predicted through joint learning of a time-space relationship in historical perception CSI data and historical communication CSI data. Compared with an existing method, the method has the advantages of higher precision and strong generalization channel prediction effect. The channel prediction method solves the problem that an existing channel prediction method is insufficient in precision and generalization.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and more specifically to: 1. a perception-assisted channel prediction method based on a pre-trained large language model; 2. a perception-assisted channel prediction system based on a pre-trained large language model. Background Art

[0002] Obtaining accurate channel state information (CSI) is increasingly crucial for beamforming and precoding in multi-antenna wireless communication systems. CSI is typically acquired through channel estimation, where the estimation frequency depends on the channel coherence time. In high-speed mobility scenarios, the channel coherence time decreases significantly, increasing the channel estimation overhead and reducing system spectral efficiency. Consequently, traditional CSI acquisition methods inevitably reduce the spectral efficiency of large-scale multi-antenna communication systems.

[0003] To reduce channel estimation overhead and improve system spectral efficiency, channel prediction has attracted widespread attention. This approach leverages the temporal correlation between past and future channel states to predict the CSI of future communications based on historical observations. Existing channel prediction methods, such as those based on network models or deep learning, exhibit limitations in accuracy and generalization. For example, Chinese invention patent No. 202410790833.6 discloses a channel prediction method based on a pre-trained large language model. This method is applied to a typical base station-single-user communication scenario. Based on the pre-trained large language model, a channel prediction network model is designed and constructed. This model is trained on a channel prediction dataset to achieve channel prediction for both time division duplex (TDD) and frequency division duplex (FDD). While this patent differs from the specific application scenario in this case, it shares the same challenges as existing methods: in actual use, accuracy and generalization fall short of ideal performance. An analysis of this patent reveals that it relies solely on historical CSI as input, failing to fully capture the complexity of real-world channel behavior, resulting in poor performance. Summary of the Invention

[0004] Based on this, it is necessary to provide a perception-assisted channel prediction method and system based on a pre-trained large language model to address the problems of insufficient accuracy and generalization of existing channel prediction methods.

[0005] The present invention is achieved by adopting the following technical solutions:

[0006] In a first aspect, the present invention discloses a perception-assisted channel prediction method based on a pre-trained large language model, which is used in a target scenario where a base station communicates with a mobile user while receiving an echo signal for perception.

[0007] The perception-assisted channel prediction method based on the pre-trained large language model includes the following steps:

[0008] Step 1: Model the communication channel and the perception channel to obtain the communication channel model h c,t (τ), the perceptual channel model H s,t (τ);

[0009] Step 2: Based on h c,t (τ), H s,t (τ) Construct a training data set;

[0010] The training data set includes M data samples; each data sample includes: the communication CSI data of P+Q time slots between the base station and the mobile communication user in a continuous period of time And the base station perceives the CSI data of the first P time slots received

[0011] Step 3: construct a perception-assisted channel prediction network and train it on the training data set until a trained perception-assisted channel prediction network is obtained;

[0012] Among them, the perception-assisted channel prediction network has dual inputs and a single output;

[0013] During training, the dual input includes: input one Enter 2H P ; Single output includes: predicted communication CSI data h for the next Q time slots Q ;H P for or Obtained by dimensionality reduction As a truth value, and used with h Q Comparison for network parameter adjustment; express Communication CSI data of the first P time slots; express Communication CSI data of the last Q time slots;

[0014] The perception-assisted channel prediction network includes: preprocessing layer, feature extraction and fusion layer, pre-trained large language layer, and output layer; the preprocessing layer is used to H P Preprocessing to obtain numerical output Feature extraction and fusion layers are used to Perform feature extraction and fusion to obtain embedded features The pre-trained large language layer is used to combine X based on the pre-trained large language model PE,n right Processing to obtain feature output X LLM,n; The output layer is used to LLM,n Further processing to obtain prediction results in,

[0015] Step 4: Use the trained perception-assisted channel prediction network to make predictions.

[0016] This perception-assisted channel prediction method based on a pre-trained large language model implements the method or process according to an embodiment of the present disclosure.

[0017] In a second aspect, the present invention discloses a perception-assisted channel prediction system based on a pre-trained large language model, which uses the perception-assisted channel prediction method based on a pre-trained large language model disclosed in the first aspect.

[0018] The perception-assisted channel prediction system based on the pre-trained large language model includes: a channel modeling module, a dataset construction module, a model training module, and a model prediction module.

[0019] The channel modeling module is used to model the communication channel and the perception channel to obtain the communication channel model h c,t (τ), the perceptual channel model H s,t (τ). The dataset building module is used to build c,t (τ), H s,t (τ) constructs a training dataset. The model training module is used to construct a perception-assisted channel prediction network and train it on the training dataset until a trained perception-assisted channel prediction network is obtained. The model prediction module is used to perform prediction using the trained perception-assisted channel prediction network.

[0020] This perception-assisted channel prediction system based on a pre-trained large language model implements the method or process according to an embodiment of the present disclosure.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] This paper provides perception-assisted channel prediction based on a pre-trained large language model. This method, on the one hand, leverages the inherent environmental correlation between the single-base station perception channel and the dual-base station communication channel, taking the perception channel into account and accurately modeling the communication and perception channels to determine their inherent spatiotemporal relationships. Furthermore, it constructs a perception-assisted channel prediction network based on the pre-trained large language model, using the perception channel as an aid to predict future communication CSI data by jointly learning the spatiotemporal relationships between historical perception CSI data and historical communication CSI data. Compared to existing methods, this method achieves higher accuracy and stronger generalization in channel prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0024] Figure 1 Flowchart of the perception-assisted channel prediction method based on a pre-trained large language model provided in Example 1 of the present invention;

[0025] Figure 2 A simulation environment diagram provided in Example 1 of the present invention;

[0026] Figure 3 This is a simulation result diagram provided by Example 1 of the present invention;

[0027] Figure 4 for Figure 1 The structure diagram of the perception-assisted channel prediction network;

[0028] Figure 5 This is a comparison chart of simulation verification performance provided by Example 1 of the present invention. DETAILED DESCRIPTION

[0029] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0030] It should be noted that when a component is referred to as being "mounted on" another component, it may be directly on the other component or there may be a central component. When a component is considered to be "set on" another component, it may be directly set on the other component or there may be a central component. When a component is considered to be "fixed to" another component, it may be directly fixed to the other component or there may be a central component.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used herein in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0032] First of all, it should be explained that the present invention is applicable to: target scenarios in which a base station communicates with a mobile user while receiving an echo signal for perception; wherein the base station is equipped with dual functions of communication and perception - the base station side is equipped with a duplex planar array antenna for communicating with the mobile user and simultaneously perceiving the surrounding environment through radar; that is, the base station side has both communication and perception functions; the mobile user side is equipped with a single antenna; and the signal form is an orthogonal frequency division multiplexing (OFDM) signal.

[0033] Based on the theory of Integrated Communication and Sensing (ISAC), the communication channel from the base station to the mobile user (a dual-station scenario) and the perception channel from the base station to the environment and back to the base station (a single-station scenario) are both determined by the scatterers in the environment. Since the physical propagation environment is shared, the scatterers in the target scene are classified into three types: 1. Shared scatterers; 2. Communication scatterers; and 3. Perception scatterers. Shared scatterers affect both the communication channel and the perception channel; communication scatterers affect only the communication channel; and perception scatterers affect only the perception channel.

[0034] That is, the communication channel consists of signals associated with the shared scatterer and the communication scatterer; the sensing channel consists of signals associated with the shared scatterer and the sensing scatterer.

[0035] Therefore, the present invention is actually a channel prediction technology based on a pre-trained large language model (for example, using GPT-2) applied to a multi-antenna orthogonal frequency division multiplexing (OFDM) communication and perception integration (ISAC) system.

[0036] Example 1

[0037] See Figure 1 , is a flowchart of a perception-assisted channel prediction method based on a pre-trained large language model provided in this embodiment 1, which includes:

[0038] Step 1: Model the communication channel and the perception channel to obtain the communication channel model h c,t (τ), the perceptual channel model H s,t (τ).

[0039] Step 1 is actually based on the ISAC theoretical foundation: the single-station sensing channel and the dual-station communication channel share the same physical propagation environment and exhibit overlapping multipath components involving common scatterers, so that the sensing data can be used to reveal useful information about the communication channel behavior.

[0040] It should be noted, however, that while ISAC proposes combining sensing to improve channel acquisition accuracy, this remains a challenging task due to the difficulty in accurately modeling the fundamental relationship between sensing and the communication channel. This stems from the dynamic nature of wireless environments, characterized by rich multipath propagation effects such as numerous scatterers and reflections.

[0041] So in step one:

[0042] 1. Model the communication channel in the delay domain. The specific mathematical expression of the model is as follows:

[0043]

[0044] Where h c,t (τ) represents the overall impulse response model of the communication channel in time slot t in the delay domain; subscript c represents the channel; subscript t represents the time slot; τ represents the delay domain; δ(τ-τ i ) represents the delay characteristic of the communication path; τ i represents the communication path delay; represents a collection of shared scatterers; represents the set of communication scatterers; α i (t) represents the communication path gain coefficient; f d,i represents the Doppler shift; represents the antenna’s steering vector; θ i Indicates the pitch angle; Indicates the azimuth.

[0045] 2. Model the perception channel in the delay domain. The specific mathematical expression of the model is as follows:

[0046]

[0047] Where H s,t (τ) represents the overall impulse response model of the perception channel in the delay domain at time slot t; subscript s represents perception; subscript t represents time slot; τ represents delay domain; represents the delay characteristics of the sensing path; represents the perceived path delay; represents a collection of shared scatterers; represents the set of perceptual scatterers; β i (t) represents the perceived path gain coefficient; f d,i represents the Doppler shift; represents the antenna’s steering vector; θ i Indicates the pitch angle; Indicates azimuth; the subscript T indicates transposition.

[0048] It should be noted that by comparing the expressions of the communication channel and the perception channel after modeling, we can know that their specific parameters are the same. ——Indicates The paths associated with the shared scatterers in , resulting in overlapping signal paths with the same height, also illustrate that the perceived channel after the above modeling actually contains useful information about the behavior of the communication channel.

[0049] Furthermore, the above modeling was verified by simulation; the simulation environment is as follows: Figure 2 As shown in Figure 1: Consider a typical street environment, the middle square represents a moving vehicle - that is, a mobile user, the base station communicates with one of the vehicles, and the surrounding cubes represent common buildings. The simulation results are shown in Figure 1. Figure 3 As shown, it can be seen that the main signal paths of the communication channel and the sensing channel have a high overlap in azimuth and elevation angle distribution - this corresponds to a shared scatterer.

[0050] Therefore, step one accurately models the perception channel and the communication channel, reflecting the basic relationship between them.

[0051] Step 2: Based on h c,t (τ), H s,t (τ) Construct a training data set.

[0052] The training data set includes M data samples; each data sample includes: communication CSI data of P+Q time slots between the base station and the mobile communication user in a continuous period of time and perception CSI data of the first P time slots perceived and received by the base station.

[0053] It should be emphasized that the intervals between adjacent time slots are the same.

[0054] Step 2 is actually to determine the relevant parameters according to the requirements and obtain CSI data through field channel measurement or simulation.

[0055] First, the above-mentioned "target scenario where a base station communicates with a mobile user and simultaneously receives and senses echo signals" is described in detail:

[0056] 1. The duplex planar array antenna equipped on the base station side is N=N v ×N h .

[0057] Where N represents the total number of antennas; N v 、N h The antenna spacing is half the signal wavelength.

[0058] That is to say, there are N transmitting and receiving antenna pairs in the target scene.

[0059] 2. The mobile user side is equipped with a single antenna.

[0060] 3. The time sampling interval of CSI is the same.

[0061] 4. The bandwidth of the uplink and downlink is the same, and both use OFDM signals; each transmitting and receiving antenna pair contains K subcarriers, and the subcarrier spacing is the same.

[0062] Then, first h c,t (τ), H s,t (τ) is converted (using discrete Fourier transform or other methods) into the frequency domain to obtain the communication CSI data of the kth subcarrier at time slot t Perceiving CSI data k∈[1,K].

[0063] Then based on Construct the communication CSI data at time slot t Perceiving CSI data

[0064] in,

[0065] Then based on A training data set is constructed, and each data sample includes: communication CSI data of P+Q time slots and perception CSI data of the first P time slots.

[0066] Therefore, the communication CSI data of P+Q time slots can be expressed as: or in, is the communication CSI data of the first P time slots; is the communication CSI data of the next Q time slots and is taken as the true value.

[0067] The perceived CSI data of the first P time slots can be expressed as: or

[0068] In addition, it should be noted that:

[0069] For a communication channel, there are Nt1 (generally 4) transmitting antennas and 1 receiving antenna; therefore, the communication CSI data in a single time slot is 2-dimensional Nt1*K.

[0070] Then, for a single data sample in the training dataset, is 3-dimensional Nt1*K*(P+Q), is 3-dimensional Nt1*K*P, It is 3-dimensional Nt1*K*Q.

[0071] For the perception channel, the number of transmitting antennas is Nt2 (generally 4) and the number of receiving antennas is Nr2 (generally 4); therefore, the perception CSI data in a single time slot is 3-dimensional Nt2*Nr2*K.

[0072] Then, for a single data sample in the training dataset, It is 4-dimensional Nt2*Nr2*K*P.

[0073] Step three: construct a perception-assisted channel prediction network and train it in the training data set until a trained perception-assisted channel prediction network is obtained.

[0074] See Figure 4 , showing the structure diagram of the perception-assisted channel prediction network, which includes: preprocessing layer, feature extraction and fusion layer, pre-trained large language layer, and output layer.

[0075] In general, the perception-assisted channel prediction network processes the CSI of each transmit / receive antenna pair in parallel to improve computational efficiency.

[0076] Specifically, when the network is trained, its input is The output is the predicted communication CSI data of the next Q time slots With h Q Compare the results to adjust the network parameters. Specifically, calculate the normalized mean square error as the loss function based on the two, and adjust the network parameters through backpropagation.

[0077] It should be noted that, considering The dimension is larger. If it is directly input into the network, it will significantly increase the computational complexity. You can also first Reduce the dimension to 3 dimensions and use it as input. The diagonal elements on Nt2*Nr2, thus obtaining the perceived CSI data of Nr2*K*P (Can also be written ).

[0078] That is to say, the network has dual input and single output; during training, the dual input includes: input one Enter 2H P ; Single output includes: predicted communication CSI data h for the next Q time slots Q Among them, h Q You can use You can also use Obtained by dimensionality reduction

[0079] The following combination Figure 4 Describe each layer:

[0080] 1. The preprocessing layer is used to H P Preprocessing to obtain numerical output

[0081] in, express The corresponding frequency domain value; express The corresponding delay domain value; Indicates H P The corresponding frequency domain value; Indicates H P The corresponding delay domain value; the subscript n represents the nth transmitting and receiving antenna pair, n∈[1,N].

[0082] like Figure 4 As shown, the preprocessing layer includes: a perception data preprocessing unit and a communication data preprocessing unit.

[0083] The perception data preprocessing unit is used to: P First convert to real number, then normalize it, and get H P First convert to the time delay domain, convert to real number, and then normalize it to get

[0084] The communication data preprocessing unit is used to: First convert to real number, then normalize it, and get Will First convert it into the time delay domain, then convert it into real numbers, and then normalize it to get

[0085] For ease of understanding, the following is a further explanation of the preprocessing layer:

[0086] In the form of communication CSI data of a single time slot and a single subcarrier, H P 、 Rewrite, then H P Expressed as Expressed as

[0087] because are all frequency domain complex numbers, which can be directly used as complex values ​​X c,n , complex value three X s,n .

[0088] At the same time They are converted into the time delay domain (using methods such as inverse discrete Fourier transform) and expressed as: So, As complex values ​​of two X c,n,τ , complex value four X s,n,τ .

[0089] Then, X c,n 、X c,n,τ 、X s,n 、X s,n,τ It can be viewed as a complex-valued matrix and expressed as:

[0090]

[0091] Considering that the subsequent layers of the network can only process real number data, X c,n 、X c,n,τ 、X s,n 、X s,n,τ Convert them to real numbers and normalize them to get:

[0092] So, It can be regarded as a real matrix and expressed as:

[0093]

[0094] 2. Feature extraction and fusion layers are used to Perform feature extraction and fusion to obtain embedded features

[0095] like Figure 4 As shown, the feature extraction and fusion layer includes: a perception data feature extraction unit, a communication data feature extraction unit, a feature fusion unit, and a fully connected layer.

[0096] The perception data feature extraction unit is used to: use a channel attention mechanism with N1 series connections Perform feature extraction and use another N1 series channel attention mechanism to Perform feature extraction and add the two feature extraction results to obtain the perception feature X CL,s .

[0097] The communication data feature extraction unit is used to: use a channel attention mechanism with N2 series connections Perform feature extraction and use another N2 series channel attention mechanism to Perform feature extraction and add the two feature extraction results to obtain the communication feature X CL,c .

[0098] It should be noted that the channel attention mechanism consists of a series of convolutional long-term memory (ConvLSTM) units. Therefore, for the channel attention mechanism, the input data is first extracted in the spatial dimension through convolution operations in the network. Simultaneously, the temporal dynamic feature map is captured through the LSTM gating mechanism. The feature map is compressed through the pooling layer and then passed through the LSTM unit for spatiotemporal feature modeling, thereby extracting deeper and more comprehensive features.

[0099] Among them, N1 and N2 can be flexibly adjusted according to actual conditions.

[0100] The feature fusion part is used to: use the cross attention mechanism to CL,s 、X CL,c Fusion is performed to obtain the fusion feature X CA,n .

[0101] in, It should be noted that the cross attention mechanism is based on X CL,c As the main feature, dynamically from X CL,s Extract useful context information from X. That is, CL,s In the form of keys and values, CL,c Provides additional contextual support.

[0102] The fully connected layer is used to: CA,n The dimension of is aligned with the feature dimension of the pre-trained large language layer, and we get in, F represents the feature dimension of the pre-trained large language layer.

[0103] Of course, the process of feature extraction and fusion layer can be formulated as:

[0104]

[0105] X CA,n =CA(X CL,c +X CL,s );

[0106]

[0107] Where, Represents N1 series channel attention mechanism; represents N2 channel attention mechanisms in series; CA(.) represents the cross attention mechanism; linear(.) represents the fully connected layer.

[0108] 3. Pre-trained large language layer is used to combine X based on pre-trained large language modelPE,n right Processing to obtain feature output X LLM,n .

[0109] like Figure 4 As shown, the pre-trained large language layer includes: position encoding layer and pre-trained large language model.

[0110] The position encoding layer is used to: encode the position of the tensor X PE,n and Add together to get the encoded feature X EB,n .

[0111] Pre-trained large language model is used for: EB,n Perform feature extraction and feature mapping to obtain X LLM,n .

[0112] The pre-trained large language layer takes advantage of the large language model's superior ability to model complex relationships in time, frequency, and space domains. It should be noted that in this embodiment 1, the pre-trained large language model is recommended to use GPT-2, which is composed of N L The decoder of each layer of Transformer is stacked; each layer of decoder includes: a multi-head attention layer, a first-level addition and normalization layer, a feedforward layer, and a second-level addition and normalization layer. L The pre-trained large language model can be flexibly adjusted according to the actual situation. Of course, other networks such as Deepseek, Qwen, and ChatGPT can also be used, but they must meet the performance requirements.

[0113] Of course, the pre-training large language layer processing process can be formulated as:

[0114]

[0115] X LLM,n =LLM(X EB,n );

[0116] Where, LLM(.) means pre-trained large language model;

[0117]

[0118] 4. The output layer is used to LLM,n Further processing to obtain prediction results

[0119] like Figure 4 As shown, the output layer first passes through the fully connected layer to convert X LLM,n Map to low-dimensional space, then rearrange the dimensions, and then perform denormalization to obtain the transformation result Xout,n , and finally based on X out,n Build

[0120] Characterize the communication CSI prediction value of the nth transmit and receive antenna pair; then:

[0121] Of course, the output layer processing can be formulated as:

[0122] X out,n =De-Norm(Rearrange(FC(X LLM,n )));

[0123]

[0124] In the formula, FC(.) represents the fully connected layer; Rearrange(.) represents dimension rearrangement; De-Norm(.) represents denormalization processing;

[0125] X out,n [1,:,:] represents X out,n The first slice of the first dimension is The real part, that is, X out,n All elements whose first dimension index is 1 in X out,n [2,:,:] represents X out,n The second slice of the first dimension is The real part, that is, the corresponding X out,n All elements whose first dimension index is 2 in .

[0126] Therefore, the expression corresponding to the loss function is:

[0127]

[0128] Where, represents the normalized mean square error; represents the square of the Frobenius norm of the matrix;

[0129] express The true value of the communication CSI of the nth transmitting and receiving antenna pair in ;

[0130] It should be noted that the subscripts n and t represent the set index, so The meaning is different: Refers to the communication CSI of all K subcarriers and all Q time slots of the nth transmit / receive antenna pair; Refers to the communication CSI of all K subcarriers of all N antennas at time slot t.

[0131] Step 4: Use the trained perception-assisted channel prediction network to make predictions.

[0132] Generally, the trained channel prediction network model is deployed to the base station side in the actual scenario.

[0133] Similar to the training phase, in step 4: the communication CSI data and perception CSI data of the previous P time slots in history are input to the trained perception-assisted channel prediction network. Then, after processing by the trained perception-assisted channel prediction network, the predicted communication CSI data of the next Q time slots is output.

[0134] Simulation Verification

[0135] In order to illustrate the superiority of the method of Example 1, the performance of the method of Example 1 (referred to as Method 1) is compared with that of the other four methods. Figure 5 .

[0136] Among them, the other four methods include: non-perceptual auxiliary channel prediction method based on pre-trained large model (referred to as method two, that is, removing perceptual channel data on the basis of method one), channel prediction method based on long short-term memory neural network (referred to as method three), channel prediction method based on convolutional neural network (referred to as method four), and converter-based channel prediction method (referred to as method five).

[0137] Depend on Figure 5 For mobile users at different speeds, the normalized mean square error of method one is the smallest, indicating that its prediction accuracy is the highest; and when the speed of method one is in the range of 10km / h to 90km / h, the change range of the normalized mean square error is also the smallest, indicating that it also has good generalization performance.

[0138] Example 2

[0139] This embodiment 2 provides a perception-assisted channel prediction system based on a pre-trained large language model, which uses the perception-assisted channel prediction method based on a pre-trained large language model provided in embodiment 1.

[0140] The perception-assisted channel prediction system based on the pre-trained large language model includes: a channel modeling module, a dataset construction module, a model training module, and a model prediction module.

[0141] The channel modeling module is used to model the communication channel and the perception channel to obtain the communication channel model h c,t (τ), the perceptual channel model H s,t(τ). The dataset building module is used to build c,t (τ), H s,t (τ) constructs a training dataset. The model training module is used to construct a perception-assisted channel prediction network and train it on the training dataset until a trained perception-assisted channel prediction network is obtained. The model prediction module is used to perform prediction using the trained perception-assisted channel prediction network.

[0142] Since this system uses the perception-assisted channel prediction method based on the pre-trained large language model in Example 1, it also has the same effect and will not be repeated here.

[0143] Example 3

[0144] This embodiment 3 discloses a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps of the perception-assisted channel prediction method based on a pre-trained large language model disclosed in embodiment 1 are implemented.

[0145] Computer devices may include: mobile terminals and fixed terminals. Examples of the former include mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), and in-vehicle terminals (such as in-vehicle navigation terminals); examples of the latter include digital TVs and desktop computers.

[0146] This embodiment 3 also discloses a readable storage medium, which stores computer program instructions. When the computer program instructions are read and executed by a processor, the steps of the perception-assisted channel prediction method based on the pre-trained large language model disclosed in embodiment 1 are executed.

[0147] Among them, the readable storage medium may include, but is not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0148] This embodiment 3 further discloses a computer program product, including a computer program. When the computer program is executed by a processor, the steps of the perception-assisted channel prediction method based on a pre-trained large language model disclosed in embodiment 1 are implemented.

[0149] It should be noted that the computer program for executing the above-mentioned program can be written in one or more programming languages or a combination thereof. Among them, the programming language includes object-oriented programming languages such as Java, Smalltalk, C++, and also includes conventional procedural programming languages such as "C" language or similar programming languages. The above-mentioned computer program can be executed completely on the user's computer, or partially on the user's computer, or partially on the user's computer and partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN).

[0150] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A perception-assisted channel prediction method based on a pre-trained large language model, characterized in that: It is used in target scenarios where a base station communicates with a mobile user while receiving echo signals for perception; It includes the following steps: Step 1: Model the communication channel and the perception channel to obtain the communication channel model h c,t (τ), the perceptual channel model H s,t (τ); Step 2: Based on h c,t (τ), H s,t (τ) Construct a training data set; The training data set includes M data samples; each data sample includes: the communication CSI data of P+Q time slots between the base station and the mobile communication user in a continuous period of time And the base station perceives the CSI data of the first P time slots received Step 3: construct a perception-assisted channel prediction network and train it on the training data set until a trained perception-assisted channel prediction network is obtained; Among them, the perception-assisted channel prediction network has dual inputs and a single output; During training, the dual input includes: input one Enter 2H P ; Single output includes: predicted communication CSI data h for the next Q time slots Q ;H P for or Obtained by dimensionality reduction As a truth value, and used with h Q Comparison for network parameter adjustment; express Communication CSI data of the first P time slots; express Communication CSI data of the middle and last Q time slots; The perception-assisted channel prediction network includes: preprocessing layer, feature extraction and fusion layer, pre-trained large language layer, and output layer; the preprocessing layer is used to H P Preprocessing to obtain numerical output Feature extraction and fusion layers are used to Perform feature extraction and fusion to obtain embedded features The pre-trained large language layer is used to combine X based on the pre-trained large language model PE,n right Processing to obtain feature output X LLM,n ; The output layer is used to LLM,n Further processing to obtain prediction results in, Step 4: Use the trained perception-assisted channel prediction network to make predictions.

2. The perception-assisted channel prediction method based on a pre-trained large language model according to claim 1, characterized in that The scatterers in the target scene include: shared scatterers, communication scatterers, and perception scatterers; There are N transmit / receive antenna pairs in the target scenario; each transmit / receive antenna pair contains K subcarriers, and the subcarrier spacing is the same.

3. The perception-assisted channel prediction method based on a pre-trained large language model according to claim 2, characterized in that In step 1, h c,t The mathematical expression of (τ) is: Where, subscript c represents channel; subscript t represents time slot; τ represents delay domain; δ(τ-τ i ) represents the delay characteristic of the communication path; τ i represents the communication path delay; represents a collection of shared scatterers; represents the set of communication scatterers; α i (t) represents the communication path gain coefficient; f d,i represents the Doppler shift; represents the antenna’s steering vector; θ i Indicates the pitch angle; Indicates azimuth; H s,t The mathematical expression of (τ) is: Where, subscript s represents perception; subscript t represents time slot; τ represents delay domain; represents the delay characteristics of the sensing path; represents the perceived path delay; represents a collection of shared scatterers; represents the set of perceptual scatterers; β i (t) represents the perceived path gain coefficient; fd,i represents the Doppler shift; represents the antenna’s steering vector; θ i Indicates the pitch angle; Indicates azimuth; the subscript T indicates transposition.

4. The perception-assisted channel prediction method based on a pre-trained large language model according to claim 3, characterized in that Step 2 includes: h c,t (τ), H s,t (τ) is converted into the frequency domain to obtain the communication CSI data of the kth subcarrier at time slot t Perceiving CSI data k∈[1,K]; based on Construct the communication CSI data at time slot t Perceiving CSI data in, based on Construct a training dataset.

5. The perception-assisted channel prediction method based on a pre-trained large language model according to claim 1, characterized in that The preprocessing layer includes: a perception data preprocessing unit and a communication data preprocessing unit; In step three: The perception data preprocessing unit is used to: P First convert to real number, then normalize it, and get H P First convert to the time delay domain, convert to real number, and then normalize it to get The communication data preprocessing unit is used to: First convert to real number, then normalize it, and get Will First convert it into the time delay domain, then convert it into real numbers, and then normalize it to get 6. The perception-assisted channel prediction method based on a pre-trained large language model according to claim 1, characterized in that The feature extraction and fusion layer includes: perception data feature extraction unit, communication data feature extraction unit, feature fusion unit, and fully connected layer; In step three: The perception data feature extraction unit is used to: use a channel attention mechanism with N1 series connections Perform feature extraction and use another N1 series channel attention mechanism to Perform feature extraction and add the two feature extraction results to obtain the perception feature X CL,s ; The communication data feature extraction unit is used to: use a channel attention mechanism with N2 series connections Perform feature extraction and use another N2 series channel attention mechanism to Perform feature extraction and add the two feature extraction results to obtain the communication feature X CL,c ; The feature fusion part is used to: use the cross attention mechanism to CL,s 、X CL,c Fusion is performed to obtain the fusion feature X CA,n ; The fully connected layer is used to: CA,n The dimension of is aligned with the feature dimension of the pre-trained large language layer, and we get 7. The perception-assisted channel prediction method based on a pre-trained large language model according to claim 1, characterized in that The pre-trained large language layer includes: position encoding layer and pre-trained large language model; In step three: The position encoding layer is used to: encode the position of the tensor X PE,n and Add together to get the encoded feature X EB,n ; Pre-trained large language model is used for: EB,n Perform feature extraction and feature mapping to obtain X LLM,n .

8. The perception-assisted channel prediction method based on a pre-trained large language model according to claim 7, characterized in that In step three, the output layer first passes through the fully connected layer to convert X LLM,n Map to low-dimensional space, then rearrange the dimensions, and then perform denormalization to obtain the transformation result X out,n , and finally based on X out,n Build 9. The perception-assisted channel prediction method based on a pre-trained large language model according to claim 1, characterized in that In step 4, the communication CSI data and perception CSI data of the previous P time slots in history are input into the trained perception-assisted channel prediction network, which is then processed by the trained perception-assisted channel prediction network to output the predicted communication CSI data of the next Q time slots.

10. A perception-assisted channel prediction system based on a pre-trained large language model, characterized in that: It uses the perception-assisted channel prediction method based on a pre-trained large language model as described in any one of claims 1-9; The perception-assisted channel prediction system based on the pre-trained large language model includes: The channel modeling module is used to model the communication channel and the perception channel to obtain the communication channel model h c,t (τ), the perceptual channel model H s,t (τ); Dataset building module, which is used to build c,t (τ), H s,t (τ) Construct a training data set; A model training module is used to construct a perception-assisted channel prediction network and train it in a training dataset until a trained perception-assisted channel prediction network is obtained; as well as The model prediction module is used to make predictions using the trained perception-assisted channel prediction network.

Images