Channel time-frequency joint extrapolation method based on multi-layer perceptron hybrid architecture

Abstract

The invention discloses a channel time-frequency joint extrapolation method based on a multi-layer perceptron hybrid architecture. Aiming at the problems of high acquisition cost of actual communication scene channel data, high time consumption of model online training and difficulty in covering the actual communication scene by simulation data, the method comprises the following steps of: making a pre-training model by using a large-scale simulation channel data set to extract channel change characteristics; and carrying out online rapid adaptation according to the trace data of the actual measurement scene. Aiming at the characteristics that the time delay spread of an indoor wireless channel is relatively small and the number of antennas is relatively small, a multi-layer perceptron hybrid architecture is applied to the time-frequency domain extrapolation of a single-carrier broadband channel. The model can quickly adapt to an actual communication scene through fine tuning of a small amount of data. The method has the advantages of being high in adaptation speed, small in data dependence, high in scene migration capacity, high in calculation parallelism degree and the like, and is suitable for edge deployment and rapid adaptation to actual scenes.

Description

Technical Field

[0001] This invention relates to the field of wireless channel prediction, and more specifically to a channel time-frequency joint extrapolation method based on a multilayer perceptron hybrid (MLP-Mixer) architecture. Background Technology

[0002] In recent years, with the rapid development of 6G and IoT technologies, systems have placed increasingly stringent demands on the accuracy and timeliness of Channel State Information (CSI) acquisition. Indoor environments, due to complex building structures, dense scatterers, and frequent human activity, exhibit rich multipath effects and complex spatiotemporal correlations. In this highly dynamic and non-stationary environment, real-time prediction and extrapolation of wireless channel characteristics are expected to become a key technology for ensuring high-quality communication links.

[0003] Traditional channel analysis methods are often based on the assumption of quasi-stationarity of the channel, making it difficult to capture the nonlinear changes in channel characteristics caused by people moving or terminals moving in indoor scenarios. A wireless channel is a high-dimensional stochastic process in time, frequency, and space. Existing technologies often use neural networks to analyze such multidimensional coupled characteristics. However, current research methods have several shortcomings: Convolutional Neural Networks (CNNs) are limited by their local receptive field, making it difficult to capture frequency domain correlations under long-term delays and temporal correlations over long time spans; Recurrent Neural Networks (RNNs) are limited by their computational logic, making it difficult to achieve highly parallel real-time inference. The latest self-attention mechanism (Transformer) has a computational complexity that increases quadratically with channel bandwidth and antenna dimension, greatly limiting its deployment on edge terminals with limited computing power.

[0004] However, for indoor environments where multipath effects are significant and spatiotemporally complex, deep learning-based solutions require corresponding modifications. Indoor scenarios exhibit distinct characteristics; different room layouts and decoration materials can lead to substantial variations in the statistical properties of the channel. While simulation data generated using standard channel models can provide physical priors for the model, these simulations still fall short of covering a wide range of real-world communication scenarios. In real-world environments, the accuracy of models trained using simulation results can severely degrade or even fail. Relying entirely on measured data for training from scratch presents a contradiction: large datasets result in high model accuracy but also in high data acquisition and training costs and slow model convergence; small datasets result in lower model accuracy and a high risk of getting trapped in local optima. Therefore, achieving rapid online adaptation and high-precision extrapolation of models for specific scenarios with minimal computational complexity using very small amounts of measured data remains a major challenge in current channel prediction and extrapolation. Summary of the Invention

[0005] Technical Problem: The purpose of this invention is to provide a channel time-frequency joint extrapolation method based on a multilayer perceptron hybrid architecture. This method proposes a way to predict future, high-bandwidth channel impulse responses from historical, low-bandwidth channel impulse responses, and provides a method for rapid adaptation to actual communication scenarios. This invention exhibits good performance in indoor channel time-frequency joint extrapolation, and requires only a small amount of measured data to quickly adapt to actual communication scenarios. It has advantages such as fast adaptation speed, low data dependency, strong scenario transfer capability, and high computational parallelism.

[0006] Technical solution: In order to achieve the above-mentioned objectives, the present invention provides a channel time-frequency joint extrapolation method based on a multilayer perceptron hybrid architecture. First, a large-scale channel dataset is obtained by simulating a public channel model to ensure that the data can cover common indoor communication scenarios in order to improve the generalization ability of the pre-trained model as much as possible.

[0007] Next, the dataset is preprocessed. Open-source simulation frameworks often guarantee that the average power of the channel data in a single frame is 1. In this case, there is no need to normalize the data. The multidimensional tensor of the single-carrier broadband channel data from the time series dimension, transmit and receive antenna dimension, delay tap dimension and complex dimension is transformed into a two-dimensional tensor of the time series dimension and other feature dimensions, which is conducive to the MLP-Mixer to stir the features separately.

[0008] Then, a channel time-frequency joint extrapolation model based on the MLP-Mixer architecture is constructed. The processed two-dimensional channel tensor is encoded by time position to enhance the temporal order information of the data. After encoding, it enters the alternating stacked time mixing layer and feature mixing layer to mix the time dimension and other dimensions of the two-dimensional tensor respectively.

[0009] Pre-training is conducted using channel measurement or simulation data. After training convergence, a pre-trained model with strong generalization ability is successfully obtained. Channel data of actual communication scenarios is acquired using general software radio peripherals. The acquired channel data needs to be preprocessed by sliding time averaging to minimize the impact of large-scale fading and retain the characteristics of small-scale fading. The pre-trained model can be corrected by randomly sampling measured channel data or a portion of the channel data at the initial observation time, so as to achieve rapid adaptation of the model to the scenario.

[0010] The method includes the following steps:

[0011] Step 1: Use open channel model simulation to obtain a large-scale channel dataset. That is, conduct a large number of simulations based on the standard channel model, covering multiple channel models, delay spread of common indoor environments and normal indoor walking speed. Sample the channel impulse response to different bandwidths to create past, low bandwidth-future, and high bandwidth single-carrier broadband channel data pairs.

[0012] Step 2, data reshaping, is to reshape the original channel data into a two-dimensional tensor, including the time dimension and other dimensions, where the other dimensions mix information from the transmitting antenna, receiving antenna, time delay domain, and complex numbers;

[0013] Step 3: Simulation pre-training, which involves inputting the training dataset into the channel time-frequency extrapolation model based on the multilayer perceptron hybrid architecture for training, and obtaining the pre-trained model. This involves embedding the channel data into time, and then alternately mixing the time dimension of the channel data with other dimensions. After thorough mixing, average pooling is performed on the original time dimension, and the prediction result is obtained through the decoding head.

[0014] Step 4: Loading pre-trained weights, which involves loading the pre-trained model weights into the channel time-frequency joint extrapolation model to be adapted;

[0015] Step 5: Online rapid adaptation, which involves acquiring a small amount of channel observation data in the actual communication scenario, preprocessing it, and then fine-tuning the model to be adapted online to quickly adapt the model to the actual communication scenario.

[0016] The dataset is generated using a standard channel model, covering multiple channel models for indoor communication scenarios, including common indoor channel delay spread and mobile speed. It is sampled across different bandwidths to obtain single-carrier broadband channel impulse responses of a certain specification. The input data is the historical low-bandwidth MIMO channel impulse response, and the output data is the future high-bandwidth MIMO channel impulse response. Measured data is collected, processed to obtain the channel impulse response under the corresponding bandwidth, and a sliding time average is performed to ensure that the influence of large-scale fading is removed as much as possible while maintaining the characteristics of small-scale fading.

[0017] The data reshaping method specifically involves reshaping the processed single-carrier broadband channel data according to the order of time series, antenna index, delay tap, and complex dimension. The two-dimensional tensor is fed into the model; where T is the number of time steps, A is the number of antenna pairs, and D is the number of time delay taps.

[0018] The MLP-Mixer architecture's channel time-frequency extrapolation model first sends historical channels into a time embedding layer for position encoding, strengthening the chronological order of channel observation data in the time dimension and making full use of the channel's time evolution characteristics.

[0019] The alternating mixing of the channel data's time dimension and other dimensions involves feeding the time-position-encoded two-dimensional tensor into an alternately stacked time mixing layer and a feature mixing layer to mix the time dimension and other dimensions of the two-dimensional tensor respectively; wherein the time mixing layer is used to extract the temporal evolution law of the channel, and the feature mixing layer is used to extract the spatial and frequency domain correlation features of the channel.

[0020] The channel data, after being thoroughly mixed by alternating stacked time mixing layers and feature mixing layers, is averaged in the time dimension to effectively reduce the number of parameters and avoid getting trapped in local optima when training with a small amount of data.

[0021] When conducting rapid online model adaptation, it is only necessary to move the receiver at walking speed in the actual communication scenario to collect a small amount of channel data, randomly sample a small amount of data or sample the observation data of the initial period to effectively correct the model, so that the model can quickly adapt to the actual communication scenario.

[0022] Beneficial Effects: This invention discloses a method for joint time-frequency extrapolation and online rapid adaptation of channels based on MLP-Mixer. It can effectively predict future channel impulse responses with high bandwidth by using historical channel impulse responses with low bandwidth. This method uses time-position coding to enhance the model's analytical capabilities in the time dimension, while fully mixing the time and feature dimensions to deeply analyze the highly coupled Doppler and multipath effects in wireless channels. Average pooling is applied to the time dimension to suppress model overfitting and ensure the model's transferability. Fine-tuning with a small amount of data allows the model to quickly adapt to real-world communication scenarios. This invention has advantages such as fast adaptation speed, low data dependency, strong scenario transferability, and high computational parallelism, making it suitable for edge deployment and rapid adaptation to real-world scenarios. Attached Figure Description

[0023] Figure 1 This is a structural diagram of the channel time-frequency joint extrapolation model based on the MLP-Mixer architecture of this invention;

[0024] Figure 2 This is a comparison chart of the online rapid adaptation convergence speed of the present invention;

[0025] Figure 3 This invention provides an online, rapid adaptation and prediction accuracy map. Detailed Implementation

[0026] This invention provides a channel time-frequency joint extrapolation method based on a multilayer perceptron hybrid architecture. It utilizes a large-scale simulated channel dataset to create a pre-trained model to extract channel variation features, and then performs online rapid adaptation based on minimal data from actual test scenarios. The specific steps are as follows:

[0027] First, a large-scale channel dataset was obtained through simulation using the TR38.901 channel model. This ensured that the simulation data covered common indoor communication scenarios to maximize the generalization ability of the pre-trained model. Typically, the dataset includes various scenarios such as line-of-sight and non-line-of-sight, and abundant and scarce scatterers. The mean square delay spread of the channel data basically covers common indoor communication scenarios, allowing the pre-trained model to easily adapt to communication scenarios with spacious or confined spaces. The terminal movement speed vector in the channel data has a random direction and a magnitude equal to walking speed. Randomly selecting these parameters ensures a sufficiently rich dataset and prevents severe overfitting during training.

[0028] Next, the dataset undergoes data preprocessing. In this invention, the power mean of the data is first ensured to be 1, or normalization is performed. To adapt to the channel time-frequency joint extrapolation model based on the MLP-Mixer architecture proposed in this invention, the data needs to be dimensionally transformed. The multidimensional tensor of single-carrier broadband channel data, consisting of time series, transmit / receive antenna, delay tap, and complex dimensions, is transformed into a two-dimensional tensor with time series dimensions and other feature dimensions. Convert to ,in For time steps, , These represent the number of transmit and receive antennas, The number of delay taps, The number of transmit and receive antenna pairs is increased to decouple the Doppler and multipath effects of the wireless channel, while also facilitating the separate mixing characteristics of the MLP-Mixer.

[0029] Let the low-bandwidth channel impulse response sequence acquired over multiple time steps be... ,in For historical time steps, It is the feature dimension under low bandwidth. This represents the set of real numbers, including information such as multipath delay taps, antenna pairs, and real and imaginary parts. The goal is to learn a nonlinear mapping function. To predict future time High-bandwidth channel impulse response tensor This joint extrapolation process can be described as follows:

[0030] ,

[0031] in, This represents the predicted high-bandwidth channel impulse response. Represents the set of learnable parameters of a neural network. This is a historical low-bandwidth channel impulse response sequence. A joint time-frequency extrapolation model of the channel based on the MLP-Mixer architecture is constructed, with the specific structure as follows: Figure 1As shown, the processed two-dimensional channel tensor undergoes temporal position encoding to enhance the temporal order information of the data, which is beneficial for model analysis of channel evolution over time. After encoding, it enters alternating stacked time-stirring layers and feature-stirring layers, stirring the time dimension and other dimensions of the two-dimensional tensor respectively, for a total of 8 stackings. The hidden layer dimension of the time-stirring layer is equal to the number of input feature dimensions, and the hidden layer dimension of the feature-stirring layer is twice the number of input feature dimensions. Layer normalization is performed before each stirring, and the input of each stirring layer is connected to the input of the next stirring layer in a skip manner, which is beneficial for the stacking of fully connected layers. The time-stirring layer is used to extract the temporal evolution law of the channel, and the feature-stirring layer is used to extract the spatial and frequency domain correlation features of the channel. After sufficient stirring, global average pooling is performed on the time dimension to effectively reduce the number of parameters and prevent the model from overfitting or getting trapped in local optima.

[0032] After training convergence, a pre-trained model with strong generalization ability is successfully obtained. Channel data from a real-world communication scenario is acquired using a general-purpose software radio peripheral; the receiver can move at a walking speed. The acquired channel data needs to undergo sliding-time averaging preprocessing to minimize the impact of large-scale fading while preserving the characteristics of small-scale fading. Randomly sampling a certain amount of measured channel data, or sampling a portion of the channel data at the initial observation time, can effectively correct the pre-trained model, enabling rapid adaptation of the model to the scenario.

[0033] This invention presents a pre-trained model and performs rapid adaptation in a typical indoor communication environment to evaluate the method of MLP-Mixer-based channel time-frequency joint extrapolation and online rapid adaptation.

[0034] A channel time-frequency joint extrapolation model based on the MLP-Mixer architecture was designed. The input data consisted of the channel impulse response performance at 16 consecutive historical time steps under a 20MHz bandwidth, and the regression objective was the channel impulse response performance at one future time step under an 80MHz bandwidth. The specific neural network structure configuration is shown in Table 1, mainly including an input preprocessing layer, a feature mapping layer, a temporal position encoding layer, a mixer skeleton, a temporal pooling layer, an output decoding head, and a physical reconstruction layer. A pre-trained model was obtained based on the above network structure. The L1 loss function was used to leverage the sparsity of the channel to focus on correcting the main multipath propagation, which is more conducive to model convergence. Due to the large number of model parameters, the training batch size was chosen to be 2048, while the maximum gradient threshold was limited to 1 to prevent gradient explosion. The Adam operator was used for model training, and cosine annealing was used for the learning rate.

[0035] This embodiment presents several pre-trained models based on other common neural network architectures for comparison, as detailed in Table 2. To quantitatively evaluate the accuracy of the predictions, Normalized Mean Squared Error (NMSE) is used as the validation metric, and its formula is defined as follows:

[0036] ,

[0037] in This indicates the number of test samples, while and Representing the first The target high-bandwidth channel impulse response matrix and the predicted high-bandwidth channel impulse response matrix of each sample are shown. It is evident that although CNN has fewer parameters, its model accuracy is too low to be acceptable. This also indicates that in indoor channel scenarios with fewer antennas and smaller latency spread, the extraction of local features from channel data is less effective under certain bandwidth conditions, while global features are more advantageous. LSTM has a large number of parameters and high model accuracy, but the computational parallelism of LSTM input to RNN architecture is very poor, making it unsuitable for edge deployment. Furthermore, LSTM convergence is very slow, pre-training costs are high, and it does not meet the requirements for rapid online adaptation. TCN performs poorly with a small number of parameters; continuously increasing the receptive field and number of channels improves performance, but this degenerates the model into an approximately fully connected architecture. Additionally, the large number of parameters results in very slow convergence. The MLP-Mixer designed in this invention has a moderate number of parameters, high computational parallelism, and fast convergence speed, which is beneficial for rapid online adaptation.

[0038] Finally, the pre-trained weights are loaded and fine-tuned using different amounts of data. For example... Figure 2 As shown, the convergence speeds of de novo training and online fast adaptation were compared. It was found that the method proposed in this invention achieved good results after the first round of training, while de novo training required several rounds. Furthermore, online fast adaptation, aided by the prior knowledge learned by the pre-trained model, can achieve higher accuracy. Figure 3 As shown, the model accuracy of training from scratch and online fast adaptation under different data volumes was compared. It is evident that the accuracy of online fast adaptation is better than that of training from scratch, especially when the data volume is small. This demonstrates the superiority of the channel time-frequency joint extrapolation and online fast adaptation method based on MLP-Mixer designed in this invention.

[0039] All aspects of this invention not described in detail are well-known to those skilled in the art. The preferred embodiments of this invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of this invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of this invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

[0040] Table 1. Structure information of the channel time-frequency joint extrapolation model of the present invention

[0041]

[0042] Table 2 Performance Comparison of the MLP-Mixer Pre-trained Model of the Present Invention

[0043]

Claims

1. A channel time-frequency joint extrapolation method based on a multilayer perceptron hybrid architecture (MLP-Mixer), characterized in that, First, a large-scale channel dataset is obtained by simulating open channel models to ensure that the data can cover common indoor communication scenarios in order to improve the generalization ability of the pre-trained model as much as possible. Next, the dataset is preprocessed. Open-source simulation frameworks often guarantee that the average power of the channel data in a single frame is 1. In this case, there is no need to normalize the data. The multidimensional tensor of the single-carrier broadband channel data from the time series dimension, transmit and receive antenna dimension, delay tap dimension and complex dimension is transformed into a two-dimensional tensor of the time series dimension and other feature dimensions, which is conducive to the MLP-Mixer to stir the features separately. Then, a channel time-frequency joint extrapolation model based on the MLP-Mixer architecture is constructed. The processed two-dimensional channel tensor is encoded by time position to enhance the temporal order information of the data. After encoding, it enters the alternating stacked time mixing layer and feature mixing layer to mix the time dimension and other dimensions of the two-dimensional tensor respectively. Pre-training is conducted using channel measurement or simulation data. After training convergence, a pre-trained model with strong generalization ability is successfully obtained. Channel data of actual communication scenarios is acquired using general software radio peripherals. The acquired channel data needs to be preprocessed by sliding time averaging to minimize the impact of large-scale fading and retain the characteristics of small-scale fading. The pre-trained model can be corrected by randomly sampling measured channel data or a portion of the channel data at the initial observation time, so as to achieve rapid adaptation of the model to the scenario.

2. The channel time-frequency joint extrapolation method based on a multilayer perceptron hybrid architecture according to claim 1, characterized in that, The method includes the following steps: Step 1: Use open channel model simulation to obtain a large-scale channel dataset. That is, conduct a large number of simulations based on the standard channel model, covering multiple channel models, delay spread of common indoor environments and normal indoor walking speed. Sample the channel impulse response to different bandwidths to create past, low bandwidth-future, and high bandwidth single-carrier broadband channel data pairs. Step 2, data reshaping, is to reshape the original channel data into a two-dimensional tensor, including the time dimension and other dimensions, where the other dimensions mix information from the transmitting antenna, receiving antenna, time delay domain, and complex numbers; Step 3: Simulation pre-training, which involves inputting the training dataset into the channel time-frequency extrapolation model based on the multilayer perceptron hybrid architecture for training, and obtaining the pre-trained model. This involves embedding the channel data into time, and then alternately mixing the time dimension of the channel data with other dimensions. After thorough mixing, average pooling is performed on the original time dimension, and the prediction result is obtained through the decoding head. Step 4: Loading pre-trained weights, which involves loading the pre-trained model weights into the channel time-frequency joint extrapolation model to be adapted; Step 5: Online rapid adaptation, which involves acquiring a small amount of channel observation data in the actual communication scenario, preprocessing it, and then fine-tuning the model to be adapted online to quickly adapt the model to the actual communication scenario.

3. The channel time-frequency joint extrapolation method based on a multilayer perceptron hybrid architecture according to claim 2, characterized in that, The dataset is generated using a standard channel model, covering multiple channel models for indoor communication scenarios, including common indoor channel delay spread and mobile speed. It is sampled across different bandwidths to obtain single-carrier broadband channel impulse responses of a certain specification. The input data is the historical low-bandwidth MIMO channel impulse response, and the output data is the future high-bandwidth MIMO channel impulse response. Measured data is collected, processed to obtain the channel impulse response under the corresponding bandwidth, and a sliding time average is performed to ensure that the influence of large-scale fading is removed as much as possible while maintaining the characteristics of small-scale fading.

4. The channel time-frequency joint extrapolation method based on a multilayer perceptron hybrid architecture according to claim 3, characterized in that, The data reshaping method specifically involves reshaping the processed single-carrier broadband channel data according to the order of time series, antenna index, delay tap, and complex dimension. The two-dimensional tensor is fed into the model; where, For time steps, For the number of antenna pairs, This represents the number of delay taps.

5. The channel time-frequency joint extrapolation method based on a multilayer perceptron hybrid architecture according to claim 4, characterized in that, The MLP-Mixer architecture's channel time-frequency extrapolation model first sends historical channels into a time embedding layer for position encoding, strengthening the chronological order of channel observation data in the time dimension and making full use of the channel's time evolution characteristics.

6. The channel time-frequency joint extrapolation method based on a multilayer perceptron hybrid architecture according to claim 5, characterized in that, The alternating mixing of the channel data's time dimension and other dimensions involves feeding the time-position-encoded two-dimensional tensor into an alternately stacked time mixing layer and a feature mixing layer to mix the time dimension and other dimensions of the two-dimensional tensor respectively; wherein the time mixing layer is used to extract the temporal evolution law of the channel, and the feature mixing layer is used to extract the spatial and frequency domain correlation features of the channel.

7. The channel time-frequency joint extrapolation method based on a multilayer perceptron hybrid architecture according to claim 6, characterized in that, The channel data, after being thoroughly mixed by alternating stacked time mixing layers and feature mixing layers, is averaged in the time dimension to effectively reduce the number of parameters and avoid getting trapped in local optima when training with a small amount of data.

8. The channel time-frequency joint extrapolation method based on a multilayer perceptron hybrid architecture according to claim 7, characterized in that, When conducting rapid online model adaptation, it is only necessary to move the receiver at walking speed in the actual communication scenario to collect a small amount of channel data, randomly sample a small amount of data or sample the observation data of the initial period to effectively correct the model, so that the model can quickly adapt to the actual communication scenario.

Images