A wireless channel prediction method based on a large model prior and a cloud edge collaborative liquid network

Abstract

The application provides a wireless channel prediction method based on a large model prior and a cloud edge collaborative liquid network, steps comprising: dividing an original channel state sequence into a long-period slow fading sequence and a short-period fast fading sequence; aligning the long-period slow fading sequence to a pre-trained large language model through a cross-modal projection layer to extract a macroscopic occlusion prior and output a scaling and translation vector; performing space-time feature extraction on the short-period fast fading sequence by a liquid neural network to obtain end-side basic feature maps; performing channel-by-channel affine transformation on the end-side basic feature maps by using the scaling and translation vector to generate deep correction feature maps; in a training stage, reversely feeding back gradients along a cross-scale feature linear modulation layer through a reconstruction loss function; and in a mapping stage, reconstructing a channel prediction matrix through an output mapping module. The application overcomes prediction lag caused by insufficient end-side computing power and channel mutation, realizes high-precision channel tracking under low communication overhead, and provides high-reliability wireless communication guarantee for embodied intelligent agents.

Description

Technical Field

[0001] This invention relates to the field of wireless communication and embodied intelligence, and in particular to a wireless channel prediction method based on large model priors and cloud-edge collaborative liquid networks for highly dynamic physical scenarios of embodied intelligent agents. Background Technology

[0002] With the deep integration of next-generation wireless communication systems (such as 6G) and Embodied AI, highly dynamic physical scenarios such as smart manufacturing and unmanned factories place extremely high demands on the reliability and low latency of wireless communication. In these scenarios, when embodied agents move at high speeds or perform complex operations, the wireless channel experiences severe Doppler shift and multipath effects due to frequent obstruction by large machinery within the factory and the displacement of the equipment itself. Therefore, accurate and real-time prediction of wireless channel state information (CSI) has become a crucial prerequisite for ensuring highly reliable communication between the cloud and embodied agents.

[0003] Traditional channel prediction methods based on statistics or conventional deep learning fix the network weights after training. These "static" or "solid" network architectures perform reasonably well in stable channels, but they often react slowly to the rapid fading and high-frequency phase fluctuations caused by sudden movements or momentary occlusions of embodied agents. They lack the ability to adapt to the continuous dynamic physical environment in real time, which not only leads to a sharp drop in prediction accuracy but also easily produces prediction lag, making it impossible to meet the fast and real-time channel tracking requirements in highly dynamic scenarios.

[0004] In recent years, prediction schemes based on Large-Scale Language Models (LLMs) have been gradually introduced into the field of wireless communication, aiming to capture the macroscopic evolution of channels by leveraging their powerful long-context reasoning capabilities and pre-trained prior knowledge. However, large language models have an extremely large number of parameters, resulting in extremely high computational complexity and inference latency, making them completely unsuitable for direct deployment on embodied edge devices with extremely limited computing power, power consumption, and memory. If they are deployed purely in the cloud, the round-trip data transmission latency between the cloud and edge, often reaching hundreds of milliseconds, severely violates the principle of extremely fast response required for embodied control, making it unable to cope with the microscopic rapid fading characteristics of millisecond-level evolution at the edge. Current solutions generally lack an effective decoupling and coordination mechanism across time scales and physical spaces.

[0005] In summary, how to leverage the macroscopic environmental perception priors of large language models in long-period slow fading, while also overcoming the bottleneck of edge computing power to achieve agile microscopic prediction of edge-side nonlinear fast fading signals, and constructing a cloud-edge collaborative prediction architecture with extremely low overhead, is a technical challenge that urgently needs to be solved in the current intersection of embodied intelligence and wireless communication.

[0006] Patent application number 202410790833.6 discloses a channel prediction method based on a pre-trained large language model. This method deploys a constructed channel prediction network model at the base station side, predicting the user's downlink channel state information (CSI) for the next L time steps based on the user's uplink channel state information (CSI) for the previous P time steps, thus achieving online channel prediction. The method includes dataset construction, network construction and training, and real-time deployment steps. This invention predicts the user's downlink channel state information for the next L time steps based on the user's uplink channel state information for the previous P time steps, reducing the overhead of channel estimation, improving the spectral efficiency of wireless communication systems, and achieving high-precision and strong generalization of time-division duplex (TDD) and frequency-division duplex (FDD) channel prediction. However, the aforementioned patent still suffers from the problem of weak computational power when deploying the large language model at the terminal side. Summary of the Invention

[0007] To address the technical challenges of traditional static networks struggling to handle high-frequency mutations in highly dynamic physical scenarios for embodied agents, and the computational bottleneck faced by large language models in edge deployment, this invention proposes a wireless channel prediction method based on large model priors and a cloud-edge collaborative liquid network. Designed for highly dynamic physical scenarios for embodied agents, this method effectively overcomes the prediction lag caused by weak edge device computing power and high-frequency channel mutations. It achieves high-precision, millisecond-level microchannel tracking with extremely low communication overhead, providing highly reliable and extremely low-latency wireless communication guarantees for embodied agents.

[0008] To achieve the above objectives, the technical solution of the present invention is implemented as follows, comprising the following steps:

[0009] Step 1: Utilize channel reciprocity to divide the original historical complex channel state information sequence into a long-period slow fading sequence and a short-period fast fading sequence;

[0010] Step 2: The long-period slow fading sequence is segmented and labeled through a cross-modal projection layer and aligned to the input embedding space of the pre-trained large language model. The macroscopic occlusion prior is extracted through the pre-trained large language model and output as scaling and translation vectors for embodied end-side modulation through a dimensionality reduction network.

[0011] Step 3: The liquid neural network extracts spatiotemporal features from the short-period fast-fading sequence based on continuous-time ordinary differential equations to obtain the basic feature map at the edge.

[0012] Step 4: Using the obtained scaling and translation vectors, perform channel-wise affine transformation on the end-side basic feature map through cross-scale feature linear modulation to generate a depth-corrected feature map;

[0013] Step 5: Based on the deep correction feature map, reconstruct the channel state information prediction matrix for multiple consecutive time steps in the future through the output mapping module.

[0014] The original historical complex channel state information sequence is obtained based on the channel model of the wireless communication system;

[0015] The method for obtaining the original historical complex channel state information sequence is as follows: the edge cloud and the physical end-side devices receive uplink and downlink pilot signals respectively, and the original historical complex channel state information sequence is directly obtained by parsing using the least squares channel estimation technique; the future is predicted based on the original historical complex channel state information sequence. Frequency domain channel state information for each time step;

[0016] The method for establishing the channel model of a wireless communication system is as follows:

[0017] Configure a uniform planar array at the edge cloud, the uniform planar array containing Several antenna elements are regularly arranged in a two-dimensional xz plane to resolve the departure angle of the signal; the departure angle is determined by the azimuth angle. and pitch angle Composition; uniform planar array pairs from specific directions The response generated by the signal is guided by the array guide vector definition;

[0018] The channel is modeled based on the geometric random channel model. The superposition of discrete propagation paths, and for the first The path is configured with the following time-varying physical parameters: time-varying complex gain Doppler frequency shift ,Delay and leaving the corner Based on delay Frequency response vector and array steering vector Perform synthesis to generate in time Observed frequency domain channel state information matrix ,in Represents the field of complex numbers. Indicates the number of frequency domain subcarriers. Indicates the number of antenna elements.

[0019] The method for dividing the original historical complex channel state information sequence into a long-period slow-fading sequence and a short-period fast-fading sequence using channel reciprocity is as follows:

[0020] At the edge cloud, based on the reciprocity of time-division duplex channels, a low-pass filter is applied to the original historical complex channel state information sequence to filter out high-frequency random noise and microscopic fast fading components, extracting macroscopic complex gain attenuation characteristics, and employing a time step... Downsampling is performed to extract long-period slow-fading sequences. ,in For long-term historical observation steps, For the number of frequency domain subcarriers, This refers to the number of antenna elements;

[0021] At the embodiment end, a time step is used to process the original historical complex channel state information sequence. Perform intensive sampling to extract short-period fast-fading sequences. ,in This represents the number of historical observation steps over a short period; and ;

[0022] Long-period slow-fading sequences are uploaded to the edge cloud; short-period fast-fading sequences are kept on the device side for local processing.

[0023] The cross-modal projection layer includes a feature stitching layer, a flattening operation layer, and a one-dimensional convolutional layer connected in sequence; the pre-trained large language model and the low-rank adaptive fine-tuning module are set in parallel, and the output of the pre-trained large language model and the output of the low-rank adaptive fine-tuning module are added together and then input to the global average pooling layer; the dimensionality reduction network adopts a multilayer perceptron.

[0024] The method for extracting macroscopic occlusion priors using a pre-trained large language model is as follows:

[0025] Embedding matrix Input into the pre-trained large language model; freeze the original weight matrix of the pre-trained large language model. ;

[0026] Low-rank adaptive fine-tuning modules are connected in parallel to each attention module of the pre-trained large language model;

[0027] After processing by a network layer with a low-rank adaptive fine-tuning module, the output of the pre-trained large language model is added to the fine-tuning residual of the low-rank adaptive fine-tuning module. Then, a global average pooling layer is used to eliminate the time dimension, extracting macroscopic feature vectors representing the prior of the global environment. .

[0028] The method for segmenting and tagging long-period slow-fading sequences is as follows:

[0029] Feature splicing layer extracts long-period slow-fading sequences The real and imaginary parts are concatenated along the feature dimension;

[0030] The flattening operation layer flattens the frequency domain subcarrier dimension and the antenna space dimension, thus flattening the long-period slow fading sequence. Transform into a two-dimensional real number sequence ;

[0031] Two-dimensional real number sequence Long-term historical observation steps along the time dimension Divided into A length of Non-overlapping sequence blocks are used to obtain a sequence block set. ;

[0032] Using a one-dimensional convolutional layer as a tokenizer, each sequence block is mapped to a dimension of... Radio frequency terms, plus learnable position codes. This yields the embedding matrix of the pre-trained large language model input space. ,in This indicates the feature embedding dimension of the pre-trained large language model;

[0033] The method for outputting the scaling and translation vectors for end-side modulation via the dimensionality reduction network is as follows:

[0034] By using a multilayer perceptron containing activation functions, macroscopic feature vectors are transformed... Dimension reduction projection is a scaling vector Translation vector , represented as:

[0035] ;

[0036] in, For activation function, , For the projection matrix weights, , For bias vectors, The channel dimension of the feature map of the embedded end-side network;

[0037] After the calculation is complete, the edge cloud will scale the vector. Translation vector Distribute to the device's cache.

[0038] The liquid neural network extracts spatiotemporal features from short-period fast-fading sequences based on continuous-time ordinary differential equations, and obtains the edge-side basic feature map using the following method:

[0039] Short-period fast-fading sequences Input to a liquid neural network; the liquid neural network uses the explicit Euler method to discretize and solve continuous-time ordinary differential equations, and sets the time step of the discretization solution. Aligned with the sampling step size of the short-period fast fading sequence ,Right now ;

[0040] The sampling step size of the current wireless channel sampling is obtained from the device-side communication protocol stack. and the sampling step size As the forward propagation computation layer of the hyperparameter-input liquid neural network, the time step is made equal to the sampling step in the numerical integrator. This allows the time dimension difference of the physical layer to be embedded as a scalar multiplication factor in the neuron state update calculation;

[0041] Through continuous time-step evolutionary calculations, the basic feature map of the end side is extracted. ,in The number of characteristic channels of the end-side network. For the temporal hidden layer dimension.

[0042] The method for generating the depth-corrected feature map is as follows:

[0043] End-side basic feature map Each channel The affine transformation is:

[0044] ;

[0045] in, Represents the basic feature map of the end side The Middle The time-series feature row vectors corresponding to each channel Represents depth-corrected feature maps The Middle Each channel corresponds to a depth-corrected temporal feature row vector. Indicates the first The scaling characteristic modulation scalar of each channel, Indicates the first The translation feature modulation scalar of each channel.

[0046] The output mapping module includes a layer normalization layer, a fully connected output projection layer, and a reshaping and complex number restoration module connected in sequence.

[0047] Depth-corrected feature map Perform layer normalization;

[0048] The normalized features are input into the fully connected output projection layer, reducing the feature dimension from the temporal hidden layer dimension. Mapped to ,in The number of time steps for future prediction. For the number of frequency domain subcarriers, The number of antenna elements, factor The real and imaginary parts are separated.

[0049] Perform a tensor reshaping operation on the output tensor of the fully connected output projection layer, converting it to a dimension of... Multidimensional real tensors;

[0050] The multidimensional real tensor is sliced ​​along the last feature dimension and split into real feature tensors. and the imaginary part feature tensor ;

[0051] For real part characteristic tensor and the imaginary part feature tensor Perform complex number restoration to obtain the future. Channel state information prediction matrix of the step .

[0052] The method for asynchronous joint optimization of large language models and liquid neural networks at both ends of the cloud edge by reconstructing the loss function and backpropagating the gradient along the cross-scale feature linear modulation layer is as follows:

[0053] The normalized mean square error is used as the loss function; the prediction error gradient is backpropagated locally on the embodiment side, and the weights of the liquid neural network and the parameters of the fully connected output projection layer are updated sequentially.

[0054] The embodied end calculates the loss function on the scaling vector according to the chain rule. Translation vector The gradient stream is generated and transmitted back to the edge cloud via a wireless uplink.

[0055] After receiving the gradient stream, the edge cloud updates the dimensionality reduction network and the low-rank adaptive fine-tuning module in sequence.

[0056] The loss function: ;

[0057] Constraints: ;

[0058] in, For the expectation of the joint distribution of historical observations and future actual channels, The square of the Frobenius norm of the matrix. These are trainable parameters in the cloud and trainable parameters on the device. For the future The actual channel matrix observations at each time step For the future Channel matrix prediction values ​​at each time step The number of time steps for future prediction. This is a cloud-edge collaborative mapping function. To predict the future Channel state over a time step.

[0059] The present invention has the following beneficial effects and advantages:

[0060] 1. This invention constructs a cloud-edge collaborative decoupled architecture that spans time scales and physical space, effectively overcoming the computational power barrier of deploying large models on the edge. Addressing the pain points of weak computational power and extremely high latency requirements of embodied intelligent agents, it decouples channel evolution characteristics into "macroscopic slow fading" and "microscopic fast fading." A large language model with massive parameters is deployed in the cloud to extract macroscopic physical laws, while a lightweight network is deployed on the edge to cope with high-frequency mutations. This fully utilizes the powerful long-context inference priors of the large language model while effectively avoiding the computational power bottleneck of running large models on the edge, ensuring low power consumption and millisecond-level operating efficiency of the edge devices.

[0061] 2. This invention introduces a continuous-time liquid neural network (LNN) and constructs a liquid neural network based on ordinary differential equations (ODE). Its neuron time constants and synaptic connections can evolve dynamically in real time according to environmental stimuli, adapting to rapidly changing fast fading signals. This significantly alleviates the problems of prediction lag and error surge in traditional models, and realizes low-latency and high-precision microchannel tracking in highly dynamic scenarios.

[0062] 3. This invention designs a cross-scale fusion mechanism based on Feature Linear Modulation (FiLM), which significantly reduces the bandwidth overhead of cloud-edge communication. During cloud-edge collaboration, macroscopic features are extracted from a large model in the cloud, then dimensionality-reduced and projected into lightweight affine commands, which are then sent to the edge via a physical downlink channel. The edge uses these affine commands to perform channel-level linear modulation on the microscopic feature map. This achieves high-precision fusion of macroscopic and microscopic spatiotemporal features while compressing the bandwidth consumption of cloud-edge communication to an extremely low level, highly meeting the high-concurrency constraints of actual wireless communication networks.

[0063] 4. This invention employs a low-rank adaptive (LoRA) joint fine-tuning mechanism to achieve efficient transfer of domain knowledge with low computational cost. While retaining the macroscopic prior knowledge of the pre-trained large model, it freezes the original weights of the large model and only connects a small-scale low-rank trainable matrix (LoRA) in parallel with the attention module. Combined with an end-to-end asynchronous backpropagation mechanism, joint optimization is performed. This allows the model to complete the cross-domain transfer of knowledge to the specific domain of wireless channel prediction with a small amount of parameter updates without losing its original generalization ability, effectively improving the engineering feasibility and commercialization value of this invention. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1This is a flowchart of the present invention.

[0066] Figure 2 This is a schematic diagram of the physical scenario for the channel model of wireless channel prediction according to the present invention.

[0067] Figure 3 This is a system architecture diagram of the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] This invention proposes a wireless channel prediction method based on large model priors and cloud-edge collaborative liquid networks, specifically involving ultra-low latency prediction of high dynamic channel state information (CSI) in 6G industrial embodied intelligence scenarios. The process of this invention is as follows: Figure 1 As shown, the invention includes the following steps: establishing a multi-timescale channel data representation, extracting slow fading sequences and fast fading sequences using channel reciprocity; designing a cloud-based macroscopic perception module based on a large language model to patch and tokenize the slow fading radio frequency sequences, aligning them to the semantic space of the large language model, extracting global occlusion priors, and generating lightweight modulation parameters; designing an embodied end-side microscopic prediction module based on a liquid neural network to process fast fading sequences through continuous-time ordinary differential equations, overcoming the high latency and generalization bottlenecks of traditional time-series models; designing a cross-scale characteristic linear modulation (FiLM) layer in the end-side network to perform instantaneous affine transformations on end-side features using parameters distributed from the cloud; and designing an output mapping and end-to-end joint training module to directly output the future channel prediction matrix. This invention is the first to combine the cross-modal radio frequency understanding capability of a large language model with the ultra-fast reflection mechanism of an embodied end-side liquid network, providing a high-precision, low-overhead underlying prediction solution for 6G industrial intelligent devices through signaling-level asynchronous modulation.

[0070] This invention employs an asynchronous decoupled architecture based on MEC (Mobile Edge Computing) edge-cloud cognitive brain and embodied edge-side fluid cerebellum to perform time-series modeling and prediction of highly dynamic industrial channel state information, including the following steps:

[0071] Step 1: Utilize channel reciprocity to divide the original historical complex channel state information sequence into a long-period slow fading sequence and a short-period fast fading sequence.

[0072] (1) Establish the channel model of the wireless communication system.

[0073] like Figure 2 As shown, the channel statistical characteristics, propagation path, antenna array response, and noise impact are clearly defined, and a channel model of the wireless communication system is first established. At the edge cloud, a system containing... A uniform planar array (UPA) of antenna elements, which is regularly arranged in a two-dimensional plane (e.g., the xz plane), containing... Units, of which and Let X and Y represent the number of antenna elements in the horizontal (x-axis) and vertical (z-axis) directions of the uniform planar array, respectively, and the antenna spacing be X and Y. and Uniform planar arrays (UPAs) enable base stations to resolve signal departure angles, which are typically determined by the azimuth angle. and pitch angle Commonly described, uniform planar arrays from a specific direction The signal response is determined by the array steering vector. To depict.

[0074] For a uniform planar array UPA in the xz plane, its steering vector can be expressed as:

[0075] ,

[0076] in, Two-dimensional index for antennas ( The linearized representation of, and These represent the index numbers of the antenna element on the x-axis and z-axis, respectively. Indicates the carrier wavelength.

[0077] Regarding time Based on the widely used Geometric Random Channel (GSCM) model, this model simulates real-world, highly dynamic multipath wireless propagation environments. It realistically recreates the fading characteristics of the channel by characterizing the geometric and physical features of multiple discrete propagation paths. In highly dynamic environments, the channel is... The superposition of discrete propagation paths, the first The path has time-varying complex gain Doppler frequency shift ,Delay and leaving the corner These physical parameters ultimately determine the time... Observed frequency domain CSI matrix , represented as:

[0078] ,

[0079] in, For the number of frequency domain subcarriers, , The item reflects the Doppler effect. It is the time step. It is the frequency response vector. It is the conjugate transpose of the guiding vector of the spatial response. They represent the first The azimuth and elevation angles of the path.

[0080] Embodied intelligent agents at sampling time steps Displacement within the plant and obstruction by large machinery inside the plant cause changes in the propagation between the plant and the edge cloud.

[0081] (2) The original historical complex channel state information (CSI) sequence is divided into a long-period slow fading sequence with a large time step and a short-period fast fading sequence with a small time step by utilizing channel reciprocity.

[0082] To separate macroscopic environmental changes from microscopic instantaneous mutations and adapt to the cloud-edge collaborative two-layer network architecture proposed in this invention, this embodiment performs physical scale decoupling and dual sampling diversion on the above-mentioned original sequence. The original sequence is obtained as follows: the cloud and the physical end-side devices receive uplink and downlink pilot signals respectively, and use the least squares (LS) channel estimation technique to directly parse the historical complex channel state information sequence. Furthermore, due to the multipath effect in the real physical wireless propagation environment, and the inevitable presence of additive white Gaussian noise (AWGN) and external electromagnetic interference in the receiver's RF hardware, the original sequence observed by the channel estimation objectively and naturally contains high-frequency fast fading components and random noise terms.

[0083] In the edge cloud, based on the reciprocity of time-division duplex channels, the macroscopic complex gain attenuation characteristics of the path affected by large dynamic obstacles are extracted. A low-pass filter is applied to the original historical complex channel state information sequence for smoothing, accurately filtering out high-frequency random noise and microscopic fast fading components, thereby extracting the macroscopic complex gain attenuation characteristics of the path affected by large dynamic obstacles. A large time step is then applied to the smoothed historical complex channel state information sequence after the above low-pass filtering process. Downsampling was performed to extract a three-dimensional complex long-period slow fading sequence that reflects large-scale occlusion and slow movement patterns. The sequence will be uploaded to the MEC edge cloud for processing to match the context-aware window of the large language model on the edge cloud. This represents the number of historical observation steps over a long period.

[0084] On the embodied end side, regarding the Doppler frequency shift The resulting high-frequency phase fluctuations are addressed by applying a minimal time step to the original historical complex channel state information sequence. Dense sampling was performed to extract a three-dimensional short-period fast fading sequence that reflects Doppler frequency shift and multipath fast fading characteristics. The sequence is preserved and processed locally at the embodied agent's edge to match the millisecond-level micro-evolutionary response of the edge-side liquid neural network, wherein, This represents the number of historical observation steps in a short period, and To strictly adapt to the physical channel evolution characteristics of highly dynamic industrial scenarios, the extremely small time step described in this embodiment... The value range is limited to 1 millisecond to 5 milliseconds, and the large time step... The value range is limited to 50 milliseconds to 200 milliseconds.

[0085] Long-period slow-decay sequence With short-period fast-fading sequences The historical observation sequence constituted The channel prediction task aims to establish a cross-spatial decoupled cloud-edge collaborative mapping function. , To predict the future, including all trainable parameters from both the cloud and the edge. The channel state at a minimum time step is:

[0086] ;

[0087] in, For the future The true channel matrix observations at each time step are obtained during the model training phase by performing physical layer channel estimation using the pilot signals that will actually arrive in the future, and are used as training labels for calculating the loss. The corresponding predicted value is obtained from the cloud-edge collaborative mapping function. Historical observation sequence The forward computation reasoning yielded the following results: The number of time steps for future prediction is limited to a range of 1 to 10. The above channel state and cloud-edge collaborative mapping function... A nonlinear mapping from historical physical multi-scale features to the future full channel matrix was constructed.

[0088] Step 2: The long-period slow fading sequence is segmented and labeled through a cross-modal projection layer and aligned to the input embedding space of the pre-trained large language model. The macroscopic occlusion prior is extracted through the pre-trained large language model and output as scaling and translation vectors for end-side modulation through a dimensionality reduction network.

[0089] like Figure 3As shown in the upper part, in order to break down the modal barrier between high-dimensional complex radio frequency sequences and the discrete real text space of large language models, the MEC edge cloud receives three-dimensional complex long-period slow-fading sequences. Subsequently, cross-modal feature extraction and macroscopic prior perception are performed. The cloud-based macroscopic perception module based on the large language model consists of a cross-modal projection layer, a pre-trained large language model, and a dimensionality reduction network connected in sequence. The cross-modal projection layer is composed of feature concatenation, flattening operations, and a one-dimensional convolutional network (Conv1D), which aligns the high-dimensional complex radio frequency signals to the semantic space of the large language model. The pre-trained large language model is used to extract macroscopic occlusion priors. The dimensionality reduction network is composed of a multilayer perceptron (MLP), which compresses the high-dimensional features output by the large language model into lightweight affine modulation parameters.

[0090] First, extract the three-dimensional complex long-period slow decay sequence. The real and imaginary parts are concatenated along the feature dimension, and then the frequency domain subcarriers and antenna space dimension are flattened (the original feature dimension is flattened). and (by merging them into a one-dimensional structure through tensor reshape operations), it is transformed into a two-dimensional sequence of real numbers. Next, the two-dimensional real number sequence Long-term historical observation steps along the time dimension Divided into A length of Non-overlapping sequence blocks are denoted as the sequence block set. .in, The value can be 16 or 32.

[0091] Using a one-dimensional convolutional layer as a tokenizer, each sequence block is mapped to a dimension of 1. Radio frequency terms, plus learnable position codes. This yields the embedding matrix for aligning the input space of the large language model. , represented as:

[0092] ,

[0093] in, This represents a one-dimensional convolution operation. This represents the feature embedding dimension of a large language model.

[0094] Embedding matrix In this embodiment, to reduce computational complexity and preserve pre-training priors, the original weight matrix of the large language model is frozen when input into the pre-trained large language model. A low-rank adaptive module is connected in parallel to each attention module, and a low-rank dimensionality reduction matrix is ​​connected in parallel next to the original frozen fully connected layer. and the increasing dimension matrix The bypass network is constructed to specifically learn the channel characteristics specific to the current industrial plant or physical environment. During the model initialization phase, a low-rank dimensionality reduction matrix... Random initialization using a Gaussian distribution, increasing the dimensionality of the matrix. Initialize the matrix to all zeros to ensure that the bypass output is zero in the initial fine-tuning phase and does not interfere with the original output of the pre-trained model; subsequently, during the cloud-edge joint training phase, the low-rank dimensionality reduction matrix is ​​performed based on the reconstruction loss function using the backpropagation algorithm. and the increasing dimension matrix The weight parameters are obtained by iteratively updating the input features. The forward computation process of the low-rank adaptive module is represented as follows:

[0095] ,

[0096] in, The output characteristics of the low-rank adaptive module, and For a low-rank trainable matrix, the rank is... , This is the scaling factor.

[0097] After processing by a network layer with low-rank adaptive fine-tuning, the output of the large language model LLM is compared with the fine-tuned residual of the low-rank adaptive module LoRA. After addition, the time dimension is eliminated by a global average pooling layer, extracting a macroscopic physical environment feature vector representing the prior knowledge of the global environment. .

[0098] Subsequently, the macroscopic physical environment feature vector is processed using a multilayer perceptron containing activation functions. Dimensionality reduction projection is a lightweight affine instruction, i.e., scaling vector. With translation vector , represented as:

[0099] ,

[0100] in, This indicates a modified linear unit activation function. , All are projection matrix weights. and For the corresponding bias vector, This represents the channel dimension of the edge network feature map. The weights and bias vectors mentioned above are obtained as follows: In the initial stage of model training, initial values ​​are assigned using the He initialization method, and then during the cloud-edge collaborative training process, they are obtained through iterative optimization using the backpropagation algorithm based on the loss function between the predicted and true values.

[0101] After the calculation is completed, the MEC edge cloud transmits the scaling vector via the wireless link downlink channel. Translation vector The data is rapidly distributed to the on-device cache of the embodied intelligent agent.

[0102] For complex radio frequency signals, this invention has created a unique pre-pipeline of "complex real and imaginary parts separate splicing + flattening + 1D convolution non-overlapping block division (Conv1D) + position encoding", which forcibly compresses and aligns high-dimensional communication baseband data to the hidden layer dimension of the large language model.

[0103] Step 3: The liquid neural network extracts spatiotemporal features from the short-period fast fading sequence based on continuous-time ordinary differential equations to obtain the basic feature map of the end side.

[0104] like Figure 3 The lower part shows the highly nonlinear short-period fast fading sequence received by the end-side antenna of the embodied intelligent agent. Due to the limited computing power of edge devices and the need to meet extremely low latency requirements, this embodiment uses a liquid neural network (LNN) as the edge backbone network to dynamically and adaptively capture high-frequency time-varying features and output edge micro representations.

[0105] Liquid neural networks are based on continuous-time ordinary differential equations. For the first... A hidden neuron, its membrane potential state With continuous time The underlying dynamic equation is defined as:

[0106] ,

[0107] in, Representative and the Index of other presynaptic neurons connected to each neuron Representing the A presynaptic neuron in The membrane potential state at time t. Based on the time constant, Inject features into external inputs. It refers to synaptic connection weights and nonlinear synaptic activation functions. The attenuation rate is dynamically adjusted based on the current network status.

[0108] Short-period fast-fading sequences The data is fed into a liquid neural network, and the explicit Euler method is used to solve continuous-time ordinary differential equations with a time step of [missing information]. Discretization solution, where the time step of the integration within the network is... The physical sampling step, i.e., the minimum time step, is numerically aligned with the short-period fast-fading sequence. The specific control logic is as follows: the actual time interval constant of the current wireless channel sampling is directly obtained from the end-side communication protocol stack. This parameter is explicitly injected as a hyperparameter into the forward propagation layer of the liquid neural network, and the integration step size parameter is set in the underlying numerical integrator. This allows the time dimension difference of the physical layer to be directly embedded as a scalar multiplication factor into the neuron state update calculation. The explicit Euler update formula for discretized iterative solution is expressed as:

[0109] ,

[0110] in, Representing the A presynaptic neuron in The membrane potential state at time t. This indicates that the m-th presynaptic neuron is in The membrane potential state at time t.

[0111] Through continuous time-step evolutionary calculations, the microscopic fundamental feature map of the end side is extracted. ,in, The number of feature channels in the edge network. For the temporal hidden layer dimension.

[0112] This invention breaks with the conventional setting of the integral step size of the ordinary differential equation of liquid neural network as a purely mathematical hyperparameter, and forces its value to be locked as the short-period sampling step size of the physical antenna. This solves the problem of time axis misalignment in conventional network prediction of physical waveforms and achieves resonance between mathematical evolution and real electromagnetic wave multipath evolution.

[0113] Step 4: Using the obtained scaling and translation vectors, perform channel-wise affine transformation on the end-side basic feature map through cross-scale feature linear modulation to generate a depth-corrected feature map.

[0114] To achieve a perfect combination of cloud computing power and edge sensitivity, and to avoid the extremely high communication overhead caused by transmitting large-size feature maps, this embodiment designs a cross-scale linear characteristic modulation (FiLM) layer on the edge. A cross-scale linear characteristic modulation (FiLM) collaborative layer is designed in the edge network and embedded into the liquid neural network state update computation graph described in step three, receiving the scaling vector sent from the cloud in step two. Translation vector Microscopic basic feature diagram of the opposite side Perform instantaneous channel-wise affine transformation to generate a deep-modified feature map that is resilient to sudden environmental changes. .

[0115] End-side networks acquire edge-side micro-basic feature maps Subsequently, a cross-scale feature linear modulation layer is embedded before the edge feature mapping network. By embedding the cross-scale feature linear modulation into the liquid neural network state update computation graph, the cross-scale feature linear modulation layer directly reads the latest scaling vectors issued from the edge cloud from the device's local cache. With translation vector The FiLM layer performs element-wise operations on the microscopic feature maps at the edge. Each channel Execution along the temporal hidden layer dimension The instantaneous affine transformation of a broadcast is represented as:

[0116] ;

[0117] in, Represents the microscopic basic features of the end side. The first in the matrix The time-series feature row vectors corresponding to each channel; Represents depth-corrected feature maps The first in the matrix Each channel corresponds to a depth-corrected temporal feature row vector; Indicates the first Lightweight scaling feature modulation scalar for each channel; Indicates the first Lightweight translation feature modulation scalar for each channel.

[0118] After scalar multiplication and addition, the edge network instantly reshapes the feature distribution based on macroscopic priors in the cloud, and outputs the corrected depth-corrected feature map. ,in, , This represents the transpose of a matrix.

[0119] This invention physically cuts off the FiLM (Feature Linear Modulation) layer that originally ran inside a single machine. The large model in the cloud extracts macroscopic priors and reduces them to extremely lightweight affine parameters, which are then sent to the liquid neural network on the edge via the downlink, where feature cascade modulation is completed. The transmission of two extremely small vectors replaces the sending of a huge feature map, fundamentally solving the high latency bottleneck of cloud-edge collaboration.

[0120] Step 5: Based on the deep correction feature map, reconstruct the channel state information prediction matrix for multiple consecutive time steps in the future through the output mapping module.

[0121] (1) During the training phase, the gradient is backpropagated along the cross-scale feature linear modulation layer by reconstructing the loss function, thereby realizing the asynchronous joint optimization of the models at both ends of the cloud edge.

[0122] Training is performed by minimizing the loss function between the real channel and the predicted channel, using the normalized mean square error (NMSE) as the objective function, which is defined as follows:

[0123] ,

[0124] ,

[0125] in, It is the expectation of the joint distribution of historical observations and future actual channels. The square of the Frobenius norm of the matrix. It includes trainable parameters in the cloud and trainable parameters on the device. For the future The actual channel matrix observations at each time step For the corresponding predicted value, The number of time steps for future predictions.

[0126] The loss function backpropagates the gradient along the cross-scale feature linear modulation layer. Specifically, during model training, the prediction error is calculated based on the aforementioned target loss function. The error gradient is first backpropagated locally at the edge, sequentially updating the weights of the liquid neural network and the parameters of the fully connected output projection layer at the edge. Simultaneously, since the channel-wise affine transformation in the cross-scale feature linear modulation layer (FiLM) is mathematically a completely continuous and differentiable linear operation, the edge device directly calculates the loss function for the scaled vector sent from the cloud according to the chain rule. Translation vector The gradient flow is transmitted across the cloud-edge network interface and back to the edge cloud via a wireless uplink. Upon receiving this gradient flow, the edge cloud uses it as a starting point to continue the reverse transmission through a chain rule, updating the cloud-based dimensionality reduction network. Ultimately, it directly updates the weight parameter matrix of the low-rank adaptive module in the large language model on the cloud, i.e., the low-rank trainable matrix. and This enables joint closed-loop optimization of the cloud-edge two-layer network that transcends spatial decoupling.

[0127] (2) In the mapping stage, based on the deep correction feature map, the channel state information prediction matrix for multiple consecutive time steps in the future is reconstructed by the output mapping module.

[0128] The output mapping module is responsible for converting the modulated depth feature map The reconstruction is a specific prediction matrix for future channel state information; the output mapping module includes a layer normalization layer, a fully connected output projection layer, and a reshaping and complex number restoration layer connected in sequence.

[0129] First, the depth-corrected feature map The application layer normalizes the data, then inputs it to the edge fully connected output projection layer to reduce the feature dimension from... Mapping to the flattened real dimension required for prediction ,in, The number of time steps for future prediction. The number of frequency domain subcarriers, factor The real and imaginary parts, representing separation, are represented as:

[0130] ,

[0131] in, To output a tensor, For fully connected output projection layer weights, The bias vector for the fully connected output projection layer. This is a layer normalization operation.

[0132] Output tensor The structure is reshaped and complex number restoration is performed. Specifically, the flattened real number field output tensor is first processed. Through tensor reshaping operations, it is transformed into a dimension of The multidimensional real tensor is then sliced ​​along the last feature dimension (dimension size 2) into real feature tensors. and the imaginary part feature tensor Finally, according to the definition of complex numbers, through... (in By performing complex number restoration (using the imaginary unit), we obtain the final future. Step-wise complex channel state information prediction matrix .

[0133] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A wireless channel prediction method based on large model priors and cloud-edge collaborative liquid networks, characterized by the following steps: include: Step 1: Utilize channel reciprocity to divide the original historical complex channel state information sequence into a long-period slow fading sequence and a short-period fast fading sequence; Step 2: The long-period slow fading sequence is segmented and labeled through a cross-modal projection layer and aligned to the input embedding space of the pre-trained large language model. The macroscopic occlusion prior is extracted through the pre-trained large language model and output as scaling and translation vectors for embodied end-side modulation through a dimensionality reduction network. Step 3: The liquid neural network extracts spatiotemporal features from the short-period fast-fading sequence based on continuous-time ordinary differential equations to obtain the basic feature map at the edge. Step 4: Using the obtained scaling and translation vectors, perform channel-wise affine transformation on the end-side basic feature map through cross-scale feature linear modulation to generate a depth-corrected feature map; Step 5: Based on the deep correction feature map, reconstruct the channel state information prediction matrix for multiple consecutive time steps in the future through the output mapping module.

2. The wireless channel prediction method based on large model prior and cloud-edge collaborative liquid network according to claim 1, characterized in that, The original historical complex channel state information sequence is obtained based on the channel model of the wireless communication system; The method for obtaining the original historical complex channel state information sequence is as follows: the edge cloud and the physical end-side devices receive uplink and downlink pilot signals respectively, and the original historical complex channel state information sequence is directly obtained by parsing using the least squares channel estimation technique; the future is predicted based on the original historical complex channel state information sequence. Frequency domain channel state information for each time step; The method for establishing the channel model of a wireless communication system is as follows: Configure a uniform planar array at the edge cloud, the uniform planar array containing Several antenna elements are regularly arranged in a two-dimensional xz plane to resolve the departure angle of the signal; the departure angle is determined by the azimuth angle. and pitch angle Composition; uniform planar array pairs from specific directions The response generated by the signal is guided by the array guide vector definition; The channel is modeled based on the geometric random channel model. The superposition of discrete propagation paths, and for the first The path is configured with the following time-varying physical parameters: time-varying complex gain Doppler frequency shift ,Delay and leaving the corner Based on delay Frequency response vector and array steering vector Perform synthesis to generate in time Observed frequency domain channel state information matrix ,in Represents the field of complex numbers. Indicates the number of frequency domain subcarriers. Indicates the number of antenna elements.

3. The wireless channel prediction method based on large model prior and cloud-edge collaborative liquid network according to claim 2, characterized in that, The method for dividing the original historical complex channel state information sequence into a long-period slow-fading sequence and a short-period fast-fading sequence using channel reciprocity is as follows: At the edge cloud, based on the reciprocity of time-division duplex channels, a low-pass filter is applied to the original historical complex channel state information sequence to filter out high-frequency random noise and microscopic fast fading components, extracting macroscopic complex gain attenuation characteristics, and employing a time step... Downsampling is performed to extract long-period slow-fading sequences. ,in For long-term historical observation steps, For the number of frequency domain subcarriers, This refers to the number of antenna elements; At the embodiment end, a time step is used to process the original historical complex channel state information sequence. Perform intensive sampling to extract short-period fast-fading sequences. ,in This represents the number of historical observation steps over a short period; and ; Long-period slow-fading sequences are uploaded to the edge cloud; short-period fast-fading sequences are kept on the device side for local processing.

4. The wireless channel prediction method based on large model prior and cloud-edge collaborative liquid network according to claim 3, characterized in that, The cross-modal projection layer includes a feature stitching layer, a flattening operation layer, and a one-dimensional convolutional layer connected in sequence; the pre-trained large language model and the low-rank adaptive fine-tuning module are set in parallel, and the output of the pre-trained large language model and the output of the low-rank adaptive fine-tuning module are added together and then input to the global average pooling layer. The dimensionality reduction network employs a multilayer perceptron.

5. The wireless channel prediction method based on large model prior and cloud-edge collaborative liquid network according to claim 4, characterized in that, The method for extracting macroscopic occlusion priors using a pre-trained large language model is as follows: Embedding matrix Input into the pre-trained large language model; freeze the original weight matrix of the pre-trained large language model. ; Low-rank adaptive fine-tuning modules are connected in parallel to each attention module of the pre-trained large language model; After processing by a network layer with a low-rank adaptive fine-tuning module, the output of the pre-trained large language model is added to the fine-tuning residual of the low-rank adaptive fine-tuning module. Then, a global average pooling layer is used to eliminate the time dimension, extracting macroscopic feature vectors representing the prior of the global environment. .

6. The wireless channel prediction method based on large model prior and cloud-edge collaborative liquid network according to claim 5, characterized in that, The method for segmenting and tagging long-period slow-fading sequences is as follows: Feature splicing layer extracts long-period slow-fading sequences The real and imaginary parts are concatenated along the feature dimension; The flattening operation layer flattens the frequency domain subcarrier dimension and the antenna space dimension, thus flattening the long-period slow fading sequence. Transform into a two-dimensional real number sequence ; Two-dimensional real number sequence Long-term historical observation steps along the time dimension Divided into A length of Non-overlapping sequence blocks are used to obtain a sequence block set. ; Using a one-dimensional convolutional layer as a tokenizer, each sequence block is mapped to a dimension of... Radio frequency terms, plus learnable position codes. This yields the embedding matrix of the pre-trained large language model input space. ,in This indicates the feature embedding dimension of the pre-trained large language model; The method for outputting the scaling and translation vectors for end-side modulation via the dimensionality reduction network is as follows: By using a multilayer perceptron containing activation functions, macroscopic feature vectors are transformed... Dimension reduction projection is a scaling vector Translation vector , represented as: ； in, For activation function, , For the projection matrix weights, , For bias vectors, The channel dimension of the feature map of the embedded end-side network; After the calculation is complete, the edge cloud will scale the vector. Translation vector Distribute to the device's cache.

7. The wireless channel prediction method based on large model prior and cloud-edge collaborative liquid network according to claim 3 or 6, characterized in that, The liquid neural network extracts spatiotemporal features from short-period fast-fading sequences based on continuous-time ordinary differential equations, and obtains the edge-side basic feature map using the following method: Short-period fast-fading sequences Input to a liquid neural network; the liquid neural network uses the explicit Euler method to discretize and solve continuous-time ordinary differential equations, and sets the time step of the discretization solution. Aligned with the sampling step size of the short-period fast fading sequence ,Right now ; The sampling step size of the current wireless channel sampling is obtained from the device-side communication protocol stack. and the sampling step size As the forward propagation computation layer of the hyperparameter-input liquid neural network, the time step is made equal to the sampling step in the numerical integrator. This allows the time dimension difference of the physical layer to be embedded as a scalar multiplication factor in the neuron state update calculation; Through continuous time-step evolutionary calculations, the basic feature map of the end side is extracted. ,in The number of characteristic channels of the end-side network. For the temporal hidden layer dimension.

8. The wireless channel prediction method based on large model prior and cloud-edge collaborative liquid network according to claim 7, characterized in that, The method for generating the depth-corrected feature map is as follows: End-side basic feature map Each channel The affine transformation is: ； in, Represents the basic feature map of the end side The Middle The time-series feature row vectors corresponding to each channel Represents depth-corrected feature maps The Middle Each channel corresponds to a depth-corrected temporal feature row vector. Indicates the first The scaling characteristic modulation scalar of each channel, Indicates the first The translation feature modulation scalar of each channel.

9. The wireless channel prediction method based on large model prior and cloud-edge collaborative liquid network according to claim 8, characterized in that, The output mapping module includes a layer normalization layer, a fully connected output projection layer, and a reshaping and complex number restoration module connected in sequence. Layer normalization layer depth-corrected feature map Perform layer normalization; The normalized features are input into the fully connected output projection layer, reducing the feature dimension from the temporal hidden layer dimension. Mapped to ,in The number of time steps for future prediction. For the number of frequency domain subcarriers, The number of antenna elements, factor The real and imaginary parts are separated. Perform a tensor reshaping operation on the output tensor of the fully connected output projection layer, converting it to a dimension of... Multidimensional real tensors; The multidimensional real tensor is sliced ​​along the last feature dimension and split into real feature tensors. and the imaginary part feature tensor ; For real part characteristic tensor and the imaginary part feature tensor Perform complex number restoration to obtain the future. Channel state information prediction matrix of the step .

10. The wireless channel prediction method based on large model prior and cloud-edge collaborative liquid network according to claim 9, characterized in that, The method for asynchronous joint optimization of large language models and liquid neural networks at both ends of the cloud edge by reconstructing the loss function and backpropagating the gradient along the cross-scale feature linear modulation layer is as follows: The normalized mean square error is used as the loss function; the prediction error gradient is backpropagated locally on the embodiment side, and the weights of the liquid neural network and the parameters of the fully connected output projection layer are updated sequentially. The embodied end calculates the loss function on the scaling vector according to the chain rule. Translation vector The gradient stream is generated and transmitted back to the edge cloud via a wireless uplink. After receiving the gradient stream, the edge cloud updates the dimensionality reduction network and the low-rank adaptive fine-tuning module in sequence. The loss function: ; Constraints: ; in, For the expectation of the joint distribution of historical observations and future actual channels, The square of the Frobenius norm of the matrix. These are trainable parameters in the cloud and trainable parameters on the device. For the future The actual channel matrix observations at each time step For the future Channel matrix prediction values ​​at each time step The number of time steps for future prediction. This is a cloud-edge collaborative mapping function. To predict the future Channel state over a time step.

Images