Channel prediction method and system based on multi-cycle feature convolutional network

Abstract

The invention provides a channel prediction method and system based on a multi-cycle feature convolutional network. According to the method, context sensing weighted input and a dynamic multi-model pool mechanism are introduced. A terminal device obtains environmental context features in real time, dynamically calculates the weight of each time step in a historical channel state information sequence based on the features, and generates a weighted optimized input sequence. Meanwhile, the network equipment pre-trains and maintains a plurality of multi-period feature convolutional network models specific to different environments, and the network can effectively capture the complex time-varying rules in the period and during the weeks in the channel, so that the accuracy of channel state information prediction is improved. And the terminal equipment immediately selects and activates the optimal sub-model in the model pool for prediction according to the real-time context features, and locally performs increment fine adjustment. According to the invention, through management of the data input end and rapid and adaptive switching of the prediction model end, the precision of channel prediction and the adaptive capacity to environmental sudden change are effectively improved.

Description

Technical Field

[0001] This application relates to a channel prediction method and system based on multi-periodic feature convolutional networks, belonging to the field of wireless communication technology. Background Technology

[0002] In 6th generation (6G) mobile communication systems and their future evolution, network devices need to acquire Channel State Information (CSI) between terminal devices and network devices, and further perform resource scheduling and beam management for uplink / downlink data transmission based on the CSI. Since there is an inherent time delay in CSI feedback, predicting future CSI becomes a necessary means to improve system performance.

[0003] Currently, channel state information prediction schemes based on artificial intelligence (AI) / machine learning (ML) are among the mainstream technologies. These schemes typically use historical and / or current channel state information as input to the channel state information prediction model, perform time-series modeling through deep learning networks, and ultimately output predictions of channel state information for future moments.

[0004] However, existing AI / ML-based channel state information prediction schemes have significant technical limitations: the prediction models struggle to balance environmental adaptability, input data quality, and real-time updates. Specifically, after local training or updates, the models cannot quickly adapt to drastic changes in the channel environment, resulting in poor generalization. Furthermore, existing models directly use raw historical channel state information sequences as input, failing to effectively identify and filter transient noise, brief obstructions, or low-correlation data within the sequences, thus impacting prediction accuracy.

[0005] Furthermore, existing prediction networks (such as RNNs, LSTMs, or Transformers) primarily focus on capturing short- and long-term dependencies in sequences, lacking the ability to uncover multi-scale periodic patterns in channel characteristics. In real-world communication scenarios, channels are affected by multipath effects, terminal motion, and environmental reflections, exhibiting complex intra- and inter-periodic variation patterns.

[0006] In summary, existing technologies struggle to consistently provide high-precision channel state information (CSO) predictions in highly dynamic and complex wireless communication scenarios. Therefore, there is an urgent need in this field for a novel CSO prediction method that leverages the advantages of multi-period feature convolutional networks to effectively capture complex time-varying patterns within and between periods in the channel, thereby improving prediction accuracy. Furthermore, this method addresses the aforementioned technical challenges by optimizing input data quality management and enabling real-time model environment switching. Summary of the Invention

[0007] The purpose of this application is to propose a channel prediction method and system based on a multi-period feature convolutional network, to address the shortcomings of existing pure time-series prediction models, which suffer from poor prediction accuracy and robustness due to a lack of quality management of input data and difficulty in quickly adapting to environmental changes. Furthermore, the multi-period feature convolutional network proposed in this application automatically mines multiple main and secondary periods from historical channel state information and performs multi-scale convolution extraction based on the reconstructed two-dimensional structure, achieving high-precision prediction of channel state information for future multiple frames.

[0008] This application is implemented using the following technical solution:

[0009] The first aspect of this invention provides a channel prediction method based on multi-period feature convolutional networks:

[0010] Acquire historical channel state information and current environmental context features in the communication scenario of the terminal device;

[0011] Based on the environmental context features at the current moment, a prediction model corresponding to the environmental context features is selected from a model pool that pre-stores multiple multi-period feature convolutional network channel prediction models.

[0012] The historical channel state information is weighted based on the environmental context characteristics at the current moment;

[0013] The weighted historical channel state information is preprocessed to generate a channel sequence with embedded features;

[0014] The prediction model is selected based on the channel sequence input with embedded features to predict channel data at future times.

[0015] Furthermore, methods for weighting historical channel state information based on the environmental context characteristics at the current moment include:

[0016] Current context features Input to a lightweight weighted network, the output of which is for each historical time step weight ;

[0017] Weighted historical channel state information ;

[0018] in, Z represents the samples in the historical channel state information, z is the historical time step index, used to represent the samples in the historical channel state information at different historical moments; Z represents the total number of samples in the fused historical channel state information.

[0019] Furthermore, methods for predicting channel data at future times using multi-periodic feature convolutional network channel prediction models include:

[0020] Multiple periods are extracted from the channel sequence with embedded features using Fast Fourier Transform:

[0021] ;

[0022] Where A is the frequency energy distribution vector obtained by averaging the amplitude spectrum of the channel time series along the characteristic dimension after performing a Fast Fourier Transform, which is used to characterize the energy intensity of channel variations at different frequency components. This represents a one-dimensional time series composed of channel sequences; and These represent the calculation of the Fast Fourier Transform and the amplitude value, respectively, selecting the first... Each amplitude value was obtained. Frequency , corresponding to Length of each cycle ;

[0023] Reconstruct the one-dimensional time series into multiple two-dimensional tensors based on each period:

[0024] ;

[0025] in, This represents the two-dimensional tensor obtained after reconstruction based on the i-th principal period. Used to extend the time series with zeros in the time dimension, so that it is consistent with... compatible, and This represents the number of rows and columns of the transformed two-dimensional tensor; in the two-dimensional tensor, one dimension represents the variation within a period, and the other dimension represents the variation during the period.

[0026] Two-dimensional convolution features are extracted from the two-dimensional tensors and fused with periodic features to capture changes within and during the period:

[0027] ;

[0028] in, Represents the two-dimensional tensor corresponding to the i-th period. The periodic feature representation tensor after two-dimensional convolution feature extraction;

[0029] The two-dimensional features are converted back to one-dimensional features, and the sequence length is truncated to the original length T using Trunc(∙):

[0030] ;

[0031] in, This represents the tensor of two-dimensional periodic features of the i-th period. The one-dimensional periodic feature sequence obtained after dimensional transformation;

[0032] Use the Softmax function to The amplitudes are normalized to obtain the aggregate weights. :

[0033] ;

[0034] The aggregation weight and The weighted summation of the one-dimensional features yields the output of the multi-period feature convolutional network.

[0035] ;

[0036] The above algorithm constitutes a temporal two-dimensional variational module. The multi-period feature convolutional network is composed of multiple temporal two-dimensional variational modules stacked together and uses residual connections. The output of the layer is represented as:

[0037] ;

[0038] Will By using a variable dimension projection layer, the internal feature dimensions of the model are projected... Actual channel dimension projected onto the original channel state information Subsequently, the output layer extracts from the projected output sequence. Data from each time step is used as a prediction of channel state information for future times. .

[0039] Furthermore, methods for pre-storing model pools containing multiple multi-period feature convolutional network channel prediction models include:

[0040] Model pooling for training multiple multi-cycle feature convolutional network models Each model is trained for a specific environmental context.

[0041] Channel training dataset They are divided into channel training samples and channel verification samples;

[0042] Based on the prediction results of the channel training samples and the channel identification verification samples, the loss function is calculated;

[0043] Based on the gradient backpropagation algorithm and the loss function, the parameters of the channel prediction model are adjusted to obtain the trained channel prediction model.

[0044] Furthermore, the loss function is:

[0045] ;

[0046] in, Represents the loss function. Indicates the first The aforementioned channel identification verification samples Indicates the first The prediction results of the channel training samples. The number of channel samples. This represents the square of the L2 norm.

[0047] Furthermore, the pooling used for training the model training dataset The following conditions must be met:

[0048] Each channel state information sample in the sample is appended with an environmental context label. ;

[0049] This includes one or more types of information such as the terminal device's mobile status, signal quality, and environmental geometry.

[0050] according to Will Divided into multiple environment-specific training subsets The subset is then used to train the corresponding channel prediction model.

[0051] Furthermore, it also includes: using newly acquired channel state information samples to perform local incremental fine-tuning of the selected multi-periodic feature convolutional network channel prediction model, obtaining the model update increment. ;

[0052] use Update the corresponding multi-period feature convolutional channel prediction model in the model pool.

[0053] The second invention provides a channel prediction system based on a multi-period feature convolutional network, including a network device and a terminal device;

[0054] The network device is equipped with a model pool, which stores multiple multi-period feature convolutional network channel prediction models.

[0055] The terminal device is used to acquire historical channel state information and current environmental context features in the communication scenario.

[0056] Based on the environmental context features at the current moment, a multi-period feature convolutional network channel prediction model corresponding to the environmental context features is selected from the model pool;

[0057] The historical channel state information is weighted based on the environmental context characteristics at the current moment;

[0058] The weighted historical channel state information is preprocessed to generate a channel sequence with embedded features;

[0059] The prediction model is selected based on the channel sequence input with embedded features to predict channel data at future times.

[0060] Furthermore, the multi-periodic feature convolutional network channel prediction model includes:

[0061] The periodicity mining module is used to extract multiple main periods from the channel sequence via Fast Fourier Transform:

[0062] ;

[0063] Where A is the frequency energy distribution vector obtained by averaging the amplitude spectrum of the channel time series along the characteristic dimension after performing a Fast Fourier Transform, which is used to characterize the energy intensity of channel variations at different frequency components. This represents a one-dimensional time series composed of channel sequences; and These represent the calculation of the Fast Fourier Transform and the amplitude value, respectively, selecting the first... Each amplitude value was obtained. Frequency , corresponding to Length of each cycle ;

[0064] The periodic reconstruction module is used to reshape a one-dimensional time series into multiple two-dimensional tensors based on each period:

[0065] ;

[0066] in, This represents the two-dimensional tensor obtained after reconstruction based on the i-th principal period. Used to extend the time series with zeros in the time dimension, so that it is consistent with... compatible, and The number of rows and columns of the transformed two-dimensional tensor are represented; in the two-dimensional tensor, one dimension represents the change within a period, and the other dimension represents the change during the period.

[0067] A multi-periodic feature convolution module is used to extract periodic features from the two-dimensional tensor through two-dimensional convolution and fuse them. A parameter-efficient Inception module is used as the two-dimensional convolution module to capture changes within and during the period.

[0068] ;

[0069] in, Represents the two-dimensional tensor corresponding to the i-th period. The periodic feature representation tensor after two-dimensional convolution feature extraction;

[0070] The feature aggregation module is used to convert the two-dimensional feature representation back to a one-dimensional feature representation, and uses Trunc(∙) to truncate the sequence length to the original length T:

[0071] ;

[0072] in, This represents the tensor of two-dimensional periodic features of the i-th period. The one-dimensional periodic feature sequence obtained after dimensional transformation;

[0073] Use the Softmax function to apply this The amplitudes are normalized to obtain the aggregate weights. :

[0074] ;

[0075] The aggregation weight and The weighted summation of the one-dimensional features yields the output of the multi-period feature convolutional network.

[0076] ;

[0077] The aforementioned periodic mining module, periodic reconstruction module, multi-period feature convolution module, and feature aggregation module together constitute a temporal two-dimensional variational module. This temporal two-dimensional variational module is composed of multiple stacked temporal two-dimensional variational modules connected using residual connections. The layer output is:

[0078] ;

[0079] The output layer first will By using a variable dimension projection layer, the internal feature dimensions of the model are projected... Actual channel dimension projected onto the original channel state information The output layer extracts the output from the projected output sequence. Data from each time step is used as a prediction of channel state information for future times. .

[0080] Furthermore, the network device delivers the channel prediction model pool to the terminal device via RRC signaling;

[0081] The terminal device uses newly acquired channel state information samples to perform local incremental fine-tuning of the selected multi-periodic feature convolutional network channel prediction model, obtaining the model update increment. ;

[0082] Terminal devices will transmit via uplink The report is sent to the network device, which then uses the aforementioned... Refresh the corresponding multi-period feature convolutional channel prediction model in the model pool.

[0083] Compared with the prior art, the beneficial effects achieved by this application are as follows:

[0084] 1. This application employs a multi-periodic feature convolutional network as the core prediction model, which can transform a one-dimensional channel state information sequence into a multi-periodic two-dimensional feature tensor. This structure effectively extracts multi-scale periodic features, such as periodic fading, receiver oscillations, and terminal maneuverability changes, using Fast Fourier Transform, and combines periodic reconstruction with the multi-periodic convolutional structure to jointly model the two-dimensional features. This enables the model to simultaneously capture and effectively model both short-period (e.g., changes caused by Doppler shift) and long-period (e.g., changes caused by environmental or ship hull obstruction) changes in the channel, overcoming the limitation of existing technologies (such as traditional RNNs and LSTMs) in effectively handling complex multi-scale temporal changes in a single network.

[0085] 2. This application deploys a dynamic multi-model pool to pre-train multiple multi-periodic feature convolutional network channel prediction sub-models optimized for specific environments. The terminal device switches and activates the optimal sub-model in real time based on real-time context features. This switching mechanism avoids the long communication and computational overhead required for traditional model updates, ensuring that the highly specialized and optimal channel prediction sub-model remains active, thereby greatly improving the model's sustained robustness under complex propagation conditions.

[0086] 3. This application introduces a context-aware weighted input mechanism, which dynamically calculates the weights of historical samples based on contextual features through a lightweight network. This mechanism can intelligently filter out historical data that is irrelevant to the current environment or is noisy, ensuring that the input to the multi-period feature convolutional network is a high-quality, highly relevant optimized sequence. This effectively enhances the model's robustness in practical deployments and further improves the accuracy of predictions.

[0087] 4. This application adopts a strategy of instant switching and incremental fine-tuning. The terminal device performs small-batch, low-power incremental fine-tuning locally and reports only minimal model update increments to the network device. This avoids the huge uplink communication overhead caused by the periodic uploading of complete model parameters in federated learning, and reduces the computational power consumption on the terminal side. Attached Figure Description

[0088] Figure 1 This is a schematic diagram of the communication system.

[0089] Figure 2 A flowchart of a channel prediction method provided in an embodiment of this application;

[0090] Figure 3 A schematic flowchart of a channel prediction method based on a multi-period feature convolutional network provided in this application embodiment;

[0091] Figure 4 This application provides a schematic diagram of a maritime unmanned aerial vehicle (UAV) ship communication scenario.

[0092] Figure 5 A schematic diagram of a multi-period feature convolutional network prediction model structure provided in an embodiment of this application;

[0093] Figure 6 This is a graph showing the predicted NMSE results for different drone and ship speeds provided in the embodiments of this application;

[0094] Figure 7 An NMSE result diagram of different models is provided for an embodiment of this application;

[0095] Figure 8 This is a schematic diagram of the structure of a communication device provided in this embodiment.

[0096] Figure 9 This is a schematic diagram of another communication device provided in an embodiment of this application. Detailed Implementation

[0097] The terms "first" and "second," etc., used in the specification and drawings of this application are used to distinguish different objects or different treatments of the same object, rather than to describe a specific order of objects. Furthermore, the terms "comprising" and "having," and any variations thereof, mentioned in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices. In the embodiments of this application, "multiple" includes two or more. In the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Additionally, the network architecture and business scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. Those skilled in the art will understand that with the evolution of network architecture and the emergence of new business scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0098] The technical solutions of this application can be applied to various communication systems, such as 3GPP communication systems, fourth-generation (4G) mobile communication technologies, such as long-term evolution (LTE) systems, 5G mobile communication systems, 5G new radio (NR) communication systems, new radio vehicle-to-everything (NRV2X) systems, and also to LTE and 5G hybrid networking systems, or nonterrestrial network (NTN) systems, device-to-device (D2D) communication systems, machine-to-machine (M2M) communication systems, Internet of Things (IoT) systems, and other next-generation communication systems, such as sixth-generation (6G) communication systems and other communication systems evolved after 5G. They can also be non-3GPP communication systems, and this application does not limit them.

[0099] Figure 1 This is a schematic diagram of a communication system 100 provided in an embodiment of this application.

[0100] Communication system 100 may include network devices, such as Figure 1 The network device 110 is shown. The communication system 100 may also include terminal devices, such as... Figure 1The terminal device 120 shown. The network device 110 and the terminal device 120 can communicate via a wireless link.

[0101] Figure 1 An exemplary network device 110 and a terminal device 120 are shown. In one possible implementation, the communication system 100 may also include multiple network devices and / or multiple terminal devices.

[0102] The network devices in this application embodiment can be network-side devices such as access network devices and core network devices. Access network devices are sometimes also called access nodes. Access network devices have wireless transceiver capabilities and are used to communicate with terminal devices. Access network devices include, but are not limited to, base stations, evolved NodeBs (eNodeBs), transmit / receive points (TRPs) in the aforementioned communication systems, NR nodes (gNBs) in 5G mobile communication systems, next-generation eNodeBs (ng-eNBs) in 5G mobile communication systems, access network devices or modules of access network devices in open RAN (ORAN) systems, satellites in non-terrestrial network (NTN) communication systems, base stations in future mobile communication systems, or access nodes in wireless fidelity (Wi-Fi) systems. Access network devices can also be modules or units capable of implementing some of the functions of a base station. Access network equipment can be macro base stations, micro base stations, indoor stations, relay nodes, donor nodes, or wireless controllers in cloud radio access network (CRAN) scenarios. Access network equipment can also be servers, wearable devices, or vehicle-mounted devices. Multiple access network devices in a communication system can be base stations of the same type or different types. Base stations can communicate directly with terminal devices or through relay stations. Terminal devices can communicate with multiple base stations using different access technologies. This application does not limit the specific technologies or device forms used in the access network equipment.

[0103] In this embodiment of the application, the device used to implement the function of the network device can be the network device itself, or it can be a device that enables the network device to implement the function, such as a processor, circuit, chip or chip system. The device can be installed in the network device or connected to the network device for use.

[0104] The terminal device in this application embodiment can be a wireless terminal device capable of receiving network device scheduling and instruction information. The wireless terminal device can be a device providing voice and / or data connectivity to a user, a handheld device with wireless connectivity, or other processing devices connected to a wireless modem. For example, the terminal device can communicate with one or more core networks or the Internet via a radio access network (RAN). The terminal device can also be referred to as a terminal, user equipment (UE), mobile station, mobile terminal, etc. Terminal devices can be widely used in various scenarios, such as device-to-device (D2D), vehicle-to-everything (V2X) communication, machine-type communication (MTC), Internet of Things (IoT), ultra-reliable low-latency communication (URLLC), virtual reality, augmented reality, industrial control, autonomous driving, telemedicine, smart grids, smart furniture, smart offices, smart wearables, smart transportation, smart cities, or satellite communication, etc. The terminal device can be a mobile phone, tablet computer, computer with wireless transceiver function, wearable device, vehicle, aircraft (such as drone, helicopter, airplane), hot air balloon, ship, robot, robotic arm, or smart home device, etc. The embodiments of this application do not limit the form of the terminal device.

[0105] In this embodiment of the application, the device used to implement the function of the terminal device can be the terminal device itself, or it can be a device that supports the terminal device in implementing the function, such as a processor, circuit, chip or chip system. The device can be installed in the terminal device or connected to the terminal device for use.

[0106] Access network equipment and / or terminal equipment can be fixed or mobile. They can be deployed on land, including indoors or outdoors, handheld or vehicle-mounted; on water; or in the air on aircraft, balloons, and satellites. This application does not limit the application scenarios of the access network equipment and terminal equipment. They can be deployed in the same or different scenarios; for example, both can be deployed on land simultaneously; or the access network equipment can be deployed on land while the terminal equipment is deployed on water, etc., and so on.

[0107] The technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of this application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.

[0108] Example 1

[0109] Please see Figure 2 The diagram illustrates a flowchart of a channel prediction method provided in an embodiment of this application, which can be applied to... Figure 1 In the communication system shown, such as Figure 2 As shown, the method may include the following steps:

[0110] S101, the network device performs model pool specialization training and delivery. During the offline phase, the network device utilizes a large-scale historical channel state information set. The Each channel state information sample in the sample is appended with an environmental context label. The aforementioned It includes at least one or more pieces of information such as terminal device mobility status, signal quality, and environmental geometry; the network device, based on the aforementioned... Will Divided into multiple environment-specific training subsets Network devices utilize each subset For the corresponding multi-period feature convolutional network sub-model Perform independent optimization training to enable To achieve optimal prediction accuracy within its specific environmental context, a dynamic multi-model pool is ultimately constructed. Subsequently, the network device delivers the model pool (or its core parameters) to the terminal device via RRC signaling or pre-configuration.

[0111] For example, network devices may divide channel state information samples labeled as high-speed mobile and non-line-of-sight (NLOS) into subsets. And use this subset to independently train a multi-period feature convolutional network sub-model. In this model pool, each sub-model has the ability to achieve optimal prediction accuracy within a specific environmental context, avoiding the performance trade-offs that occur with a single general-purpose model under different complex environments.

[0112] S102, Real-time Context and Historical Channel State Information (CSI) Acquisition. The UE continuously acquires the historical channel state information sequence. As input. Simultaneously, the terminal device collects the environmental context features at the current moment in real time. The It includes at least key information such as the terminal's instantaneous speed and wireless signal strength.

[0113] For example, the environmental context features The data collection is not based on a single metric; the terminal device measures and integrates data from multiple dimensions in real time. These features include at least: instantaneous velocity (used to calculate Doppler shift, reflecting the rate of change in channel time correlation), RSRP (Reference Signal Received Power, used to help determine whether the current environment is line-of-sight / non-line-of-sight), and delay spread (reflecting multipath richness). The real-time acquisition of these multi-dimensional features provides accurate decision-making basis for subsequent steps such as context-aware weighting and instant model switching.

[0114] S103, Data weighted input generation and preprocessing. The terminal will use real-time context features. The input is fed into a locally pre-trained lightweight weighted network. The weighted network immediately outputs a set of weights. The weight Reflects historical channel state information samples Correlation with the current channel environment. Terminal equipment utilizes... Historical channel state information sequence Perform a weighting operation to obtain the weighted historical channel state information. This step effectively filters out historical data with low correlation or that is affected by transient interference. Subsequently, the terminal device processes the weighted historical channel state information. Preprocessing is performed, including decomposing the real and imaginary parts in complex form, normalization, and then converting them into a unified-dimensional embedded feature sequence through linear mapping. As input to a multi-period feature convolutional network.

[0115] For example, when a terminal device performs channel state information prediction in a high-speed mobile state, it collects environmental context features in real time. This serves as the basis for judging data quality. If a lightweight weighted network identifies a certain historical sample... It is a recently acquired sample and its state is highly consistent (e.g., the sample's velocity and signal strength are both similar to...). If the network matches the current environmental state it reflects, then a higher weight is assigned to it. Conversely, if the network identifies a historical sample... Abnormal amplitude spikes appear due to transient external disturbances, and their characteristics differ from those of stable ones. If a sample does not match, it is assigned a very low weight. Similarly, outdated samples collected from low-speed terminal devices are also assigned low weights. This weighting process effectively suppresses the contamination and misleading effects of noise spikes and low-correlation data on the predicted sequence. Ultimately, through this weighting process, the channel prediction network receives an optimized sequence with a high signal-to-noise ratio, improving the model's robustness and accuracy in complex channels.

[0116] S104, the optimal multi-period feature convolutional channel prediction model is activated instantly. The terminal device uses real-time acquired contextual features. By using locally stored decision rules, a specialized sub-model that best matches the current environment is selected and activated in real time from a dynamic multi-model pool. .

[0117] For example, if When specific high-speed movement and non-line-of-sight (NLOS) conditions are met, the terminal device immediately activates the features optimized for that scenario. The activation process is completed locally on the terminal side, enabling millisecond-level response to environmental changes.

[0118] S105, the terminal device performs channel state information prediction. The terminal device will then use the preprocessed embedded feature sequence... Input to activation In the model, channel state information prediction is performed, and the predicted channel state information for future time periods is output. .

[0119] S106, Local Incremental Fine-tuning by the Terminal Device. After completing the channel state information prediction, the terminal device uses the newly acquired channel state information samples to... Perform small-batch, low-power local incremental fine-tuning to obtain the model update increment. .

[0120] S107, the terminal device performs incremental update reporting. The terminal device sends a model update incremental message to the network device, the message carrying... The unique identifier and the incremental parameter .

[0121] S108, the network device performs model pool refresh and maintenance. The network device receives data from multiple terminal devices. Then, these incremental parameters are processed by weighted averaging or other aggregation algorithms to obtain the aggregated increment. and utilize Update the corresponding model pool Model parameters.

[0122] The wireless communication method provided in this embodiment addresses the shortcomings of traditional model update methods (such as federated learning) by introducing context-aware weighted input and a dynamic multi-model pooling mechanism, which suffer from high update latency and poor model environmental adaptability. Specifically, this method uses a context-aware weighted input mechanism to dynamically filter and adjust the weights of historical channel state information collected locally by the terminal device. This ensures that the data input into the multi-periodic feature convolutional network is an optimized, highly correlated sequence, significantly enhancing the anti-interference capability and instantaneous prediction accuracy of the prediction model. Simultaneously, the dynamic multi-model pooling allows the terminal device to instantly select and activate the most matching specialized sub-model based on real-time environmental context features, avoiding the long network aggregation wait required for traditional model updates and significantly improving the model's environmental adaptability and prediction robustness. Furthermore, by locally incrementally fine-tuning and only reporting incremental parameters, compared to reporting complete model parameters or performing complex similarity discrimination coding, this scheme further reduces uplink communication overhead, improves system response efficiency, and ensures the long-term validity of the parameters of each specialized multi-periodic feature convolutional channel prediction model in the model pool.

[0123] Example 2

[0124] like Figure 3 As shown in the figure, this embodiment introduces a channel prediction method based on a multi-periodic feature convolutional network in the scenario of UAV-to-ship communication at sea, including the following steps:

[0125] S201, acquire historical channel status information and channel data in the scenario of communication between UAVs and ships at sea;

[0126] Figure 4 This is a schematic diagram illustrating a maritime unmanned aerial vehicle (UAV)-to-ship communication scenario applied in an embodiment of the present invention. The UAV is equipped with... Root antenna, service Each single-antenna user, the drone on the first root antenna to the first individual users at any time The channel impulse response can be expressed as:

[0127]

[0128] in It is the number of single-bounce (SB) clusters. It is the first The number of rays within a cluster. It is the number of multi-bounce (MB) clusters. It is the first The number of rays within a cluster. It's the Rice factor. The first on the drone. root antenna to the first The latency of the Loss path between users can be expressed as:

[0129]

[0130] in, It's the speed of light. It is the first drone root antenna to the first The distance between users.

[0131] No. The first single cluster ray, in the The first transmitter and the first The SB delay between the receivers can be expressed as:

[0132]

[0133] No. The first single cluster ray, in the The first transmitter and the first The MB delay between the receivers can be expressed as:

[0134]

[0135] Channel gain in LOS case It can be represented as:

[0136]

[0137] in, , , , These represent the azimuth angle of departure (AAoD), the elevation angle of departure (EAoD), the azimuth angle of arrival (AAoA), and the elevation angle of arrival (EAoA) of the Loss component, respectively. This is the initial phase. Angle vector. and The function consists of the Angle of Arrival (AoA) and the Angle of Departure (AoD) in the Loss of Sorrow (LoS) case. , , , Antenna patterns representing Tx and Rx, time-varying Doppler frequency shift. express. and This can represent the rotation matrices of Tx and RX arrays. Additionally, the position vectors... It is the first The first Tx antenna and the first 3D position of each Rx antenna.

[0138] In the SB case, channel gain It can be represented as:

[0139] For the SB component and It is the first The first Tx and the first The time-varying Doppler frequency caused by the movement of Rx antenna elements. Indicates the first In the single cluster, the first The time-varying angle of arrival / departure vector between each scatterer and Tx / Rx.

[0140] In this invention, the initial altitude of the maritime UAV is 100 meters, and it moves along a preset three-dimensional track. The UAV carries 128 antennas to serve single-antenna ship users, and the communication carrier frequency is 5.5 GHz. Historical channel state information data is generated using the aforementioned channel model of the maritime UAV's communication with ships, and a historical channel state information sample pool is constructed.

[0141] S202, preprocess the channel state information;

[0142] Furthermore, in one embodiment of the present invention, the preprocessing of the channel measurement data includes:

[0143] S202.1, the complex-form channel data is split into real and imaginary parts:

[0144] ;

[0145] In the formula, Indicates the real part, Indicates the imaginary part.

[0146] S202.2, Perform mean-variance normalization on the split data:

[0147] ;

[0148] in, and These are the mean and standard deviation of all channel state information in the training set, respectively.

[0149] S202.3 converts channel data into a fixed-dimensional feature sequence through linear mapping, which is then used as input to a multi-period feature convolutional network model.

[0150] S202.4 uses channel state information from several historical frames as model input and channel state information from subsequent frames as prediction labels to form a training sample matrix:

[0151] ;

[0152] in It is a historical frame. It is a predicted frame.

[0153] S203, Construct a multi-periodic feature convolutional network channel prediction model, such as Figure 5 As shown, the multi-period feature convolutional network of the present invention includes the following modules:

[0154] S203.1, the periodic mining module extracts multiple main periods from the channel sequence using Fast Fourier Transform:

[0155]

[0156] in and These represent the calculation of the Fast Fourier Transform and the amplitude value, respectively, selecting the first... Each amplitude value was obtained. The most important frequency These selected frequencies also correspond to Length of each cycle .

[0157] S203.2, the periodic reconstruction module reshapes the one-dimensional time series into multiple two-dimensional tensors based on each period:

[0158]

[0159] in Used to extend the time series with zeros in the time dimension, so that it is consistent with... compatible, and This represents the number of rows and columns of the transformed two-dimensional tensor. In this two-dimensional tensor, one dimension represents the variation within a period, and the other dimension represents the variation during the period.

[0160] S203.3, the multi-periodic feature convolution module performs two-dimensional convolution feature extraction on the two-dimensional tensor and fuses the periodic features. It uses a parameter-efficient Inception module as the two-dimensional convolution module to simultaneously capture changes within and during the period:

[0161]

[0162] S203.4, the feature aggregation module includes:

[0163] S203.4.1, the two-dimensional feature representation is converted back to a one-dimensional feature representation, and the sequence length is truncated to the original length T using Trunc(∙):

[0164]

[0165] S203.4.2, using the Softmax function to apply this... The amplitudes are normalized to obtain the aggregate weights. :

[0166] ;

[0167] S203.4.3, the aggregate weights and The weighted summation of the one-dimensional features yields the output of the multi-period feature convolutional network.

[0168] ;

[0169] S203.5, together with S203.1 to S203.4, constitutes a "temporal two-dimensional variational module". The entire prediction network can be composed of multiple such modules stacked together, using residual connections, i.e., the... The output of the layer can be represented as:

[0170] ;

[0171] S203.6: The output layer first will By using a variable dimension projection layer (i.e., a fully connected layer), the internal feature dimensions of the model are projected... Actual channel dimension projected onto the original channel state information The output layer then extracts the output from the projected output sequence. The data from each time step is used as the final channel state information prediction value. .

[0172] S204, The channel prediction model is trained based on the channel training dataset to obtain the trained channel prediction model, including:

[0173] S204.1, the channel training dataset is divided into channel training samples and channel identification verification samples. In this embodiment, the channel state information samples obtained in step 1 are divided proportionally into: 80% training set, 10% verification set, and 10% test set. The channel state information samples are divided into multiple sets, each set consisting of P+L frames. In each set, the previous P frames are designated as the past P frames, and the subsequent L frames are designated as the next L frames.

[0174] S204.2, Based on the prediction results of the channel training samples and the channel identification verification samples, calculate the regression loss function;

[0175] S204.3, Based on the gradient backpropagation algorithm and the loss function, the parameters of the channel prediction model are adjusted to obtain the trained channel prediction model. In this embodiment, Adam is used as the optimizer with a learning rate of 0.0001. The loss function is determined by the following formula:

[0176] ;

[0177] in Represents the loss function. Indicates the first The aforementioned channel identification verification samples Indicates the first The prediction results of the channel training samples. The number of channel samples. Represents the square of the L2 norm;

[0178] S205, Input the preprocessed channel data from step 2 into the trained channel prediction model for prediction, including:

[0179] S205.1, the channel state information sequence is subjected to FFT transformation by the periodic mining module to extract the main period and secondary period corresponding to multiple energy peaks;

[0180] S205.2, the periodic reconstruction module maps the channel state information sequence into multiple two-dimensional tensors of different sizes based on different periods, which are used to represent the correlation within the period;

[0181] S205.3 uses a multi-period feature convolution module to perform convolution operations on the reconstructed matrix using a fixed-size two-dimensional convolution kernel to extract periodic features;

[0182] S205.4 fuses the multi-scale features obtained by periodic convolution according to the channel dimension and enhances the network stability through residual connections;

[0183] S205.5 maps the fused features to predicted channel gain values ​​for several future frames through the output layer, and outputs a predicted sequence of channel state information.

[0184] Figure 6 The NMSE performance of the proposed channel prediction method and existing time-series channel prediction schemes (RNN, LSTM, and Transformer) was compared with that of the other schemes as the number of prediction frames increased. It can be observed that the NMSE performance of all schemes decreases with the increase of the number of prediction frames. As the number of frames increases, the temporal correlation of channel state information weakens, affecting the performance of the prediction schemes. However, the NMSE performance of the proposed scheme consistently outperforms the other schemes. Figure 7 The NMSE performance of the proposed solution was compared with the change in prediction frames at different drone and ship speeds. As the speed increases, the Doppler frequency shift increases, and the predicted NMSE performance decreases.

[0185] The channel prediction method based on multi-periodic feature convolutional networks provided in this embodiment can effectively address the drastic time-varying and non-stationary characteristics caused by the high-speed maneuvering of UAVs and ocean waves in maritime communication links. By constructing a multi-periodic feature convolutional network model, and performing periodic mining, periodic reconstruction, multi-periodic convolutional feature extraction, and residual feature fusion on the channel state information sequence, the channel prediction method based on multi-periodic feature convolutional networks provided in this embodiment can capture the hidden temporal structure of multiple scales and multiple periods from historical channel state information, thereby accurately modeling the complex changing trends of maritime communication channels and achieving high-precision prediction of future channels. Compared with traditional channel prediction methods based on statistical models or single-scale neural networks, the channel prediction method based on multi-periodic feature convolutional networks provided in this embodiment can make fuller use of the implicit periodic structure and cross-period correlation in the channel state information sequence, significantly improving the stability and robustness of the prediction.

[0186] In the embodiments described above, the terminal device may execute some or all of the steps in each embodiment; the network device may execute some or all of the steps in each embodiment. These steps or operations are merely examples, and the embodiments of this application may also perform other operations or variations thereof. Furthermore, the steps may be executed in different orders as presented in the embodiments, and it is not necessary to execute all the operations in the embodiments of this application. Moreover, the sequence number of each step does not imply the order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0187] Please see Figure 8 , Figure 8 This is a schematic diagram of a communication device provided in an embodiment of this application. Figure 8As shown, the communication device 400 may include a communication module 410. The communication module 410 can implement corresponding communication functions, which can be internal communication functions of the communication device 400 or communication functions between the communication device 400 and other devices. Optionally, the communication module 410 may also be referred to as a communication interface or transceiver module. Optionally, the communication device 400 further includes a processing module 420. The processing module 420 can implement corresponding processing functions.

[0188] Optionally, the communication device 400 further includes a storage module, which can be used to store instructions and / or data; the processing module 420 can read the instructions and / or data in the storage module so that the communication device 400 can implement the aforementioned method embodiments.

[0189] In one possible design, the communication device 400 may correspond to the terminal device in the above method embodiments, or to a component (such as a circuit, chip, or chip system) configured in the terminal device. The communication device 400 can be used to execute the steps or processes performed by the terminal device in any of the above method embodiments.

[0190] Figure 9 This is another schematic block diagram of the communication device 500 provided in the embodiments of this application. The communication device 500 may be a chip, chip system, or processor, etc., in a terminal device or network device that implements the above-described methods. The communication device 500 can be used to implement the methods described in the above-described method embodiments; for details, please refer to the descriptions in the above-described method embodiments.

[0191] like Figure 9 As shown, the communication device 500 may include one or more processors 510, which may also be referred to as processing units or processing modules, and can implement certain control functions. The processor 510 may be a general-purpose processor or a dedicated processor, such as a baseband processor or a central processing unit. The baseband processor can be used to process communication protocols and communication data, while the central processing unit can be used to control the communication device 500 (e.g., a base station, baseband chip, user, user chip), execute software programs, and process data from the software programs.

[0192] In an alternative design, the processor 510 may also store instructions and / or data that can be executed by the processor 510 to cause the communication device 500 to perform the methods described in the above method embodiments.

[0193] In another alternative design, the communication device 500 may include a communication interface 520 for implementing receiving and transmitting functions. For example, the communication interface 520 may be a transceiver circuit, interface, interface circuit, or transceiver. The transceiver circuit, interface, interface circuit, or transceiver for implementing receiving and transmitting functions may be separate or integrated. The aforementioned transceiver circuit, interface, interface circuit, or transceiver may be used for reading and writing code / data, or it may be used for transmitting or relaying signals.

[0194] Optionally, the communication device 500 may include one or more memories 530, which may store instructions that can be executed on the processor 510, causing the communication device 500 to perform the methods described in the above method embodiments. Optionally, the memories 530 may also store data. Optionally, the processor 510 may also store instructions and / or data. The processor 510 and the memories 530 may be provided separately or integrated together.

[0195] It should be understood that, in one possible design, the steps in the method embodiments provided in this application can be implemented by integrated logic circuits in the processor's hardware or by instructions in software form. The steps of the methods disclosed in the embodiments of this application can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are not provided here.

[0196] In one implementation, the communication device 500 may correspond to the terminal device in the above method embodiments and may be used to execute the various steps and / or processes executed by the terminal device in the above method embodiments. The processor 510 may be used to execute instructions stored in the memory 530, and when the processor 510 executes the instructions stored in the memory, the processor 510 is used to execute the various steps and / or processes of the above method embodiments corresponding to the terminal device.

[0197] In another implementation, the communication device 500 may correspond to the network device in the above method embodiments and may be used to execute the various steps and / or processes executed by the network device in the above method embodiments. The processor 510 may be used to execute instructions stored in the memory 530, and when the processor 510 executes the instructions stored in the memory, the processor 510 is used to execute the various steps and / or processes of the above method embodiments corresponding to the network device.

[0198] It should be understood that the aforementioned processing device can be one or more chips. For example, the processing device can be a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), a system-on-chip (SoC), a central processor unit (CPU), a network processor (NP), a digital signal processor (DSP), a microcontroller unit (MCU), a programmable logic device (PLD), or other integrated chips.

[0199] It is understood that the memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory used in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0200] According to the method provided in the embodiments of this application, this application also provides a chip system, which includes one or more processors for calling and executing instructions stored in memory, thereby causing the method described in the embodiments of this application to be executed. The chip system may be composed of chips or may include chips and other discrete devices.

[0201] The chip system may include input circuits or interfaces for transmitting information or data, and output circuits or interfaces for receiving information or data.

[0202] According to the method provided in the embodiments of this application, this application also provides a communication system, which includes the aforementioned network device and terminal device.

[0203] According to the method provided in the embodiments of this application, this application also provides a computer program product, which includes: computer program code, which, when run on a computer, causes the computer to execute the various steps or processes executed by the network device or terminal device in any of the foregoing method embodiments.

[0204] According to the method provided in the embodiments of this application, this application also provides a computer-readable storage medium storing program code, which, when run on a computer, causes the computer to execute the various steps or processes executed by the network device or terminal device in any of the foregoing method embodiments.

[0205] The computer-readable storage medium may be the aforementioned volatile memory or non-volatile memory, or it may include both volatile memory and non-volatile memory.

[0206] In the embodiments of this application, the terms and English abbreviations are exemplary examples given for ease of description and should not be construed as limiting the application in any way. This application does not preclude the possibility of defining other terms that can achieve the same or similar functions in existing or future agreements.

[0207] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When these computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated.

[0208] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0209] It should be understood that in the various embodiments of this application, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0210] In summary, the above description is merely a preferred embodiment of the technical solution of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A channel prediction method based on multi-periodic feature convolutional networks, characterized in that: Acquire historical channel state information and current environmental context features in the communication scenario of the terminal device; Based on the environmental context features at the current moment, a prediction model corresponding to the environmental context features is selected from a model pool that pre-stores multiple multi-period feature convolutional network channel prediction models. The historical channel state information is weighted based on the environmental context characteristics at the current moment; The weighted historical channel state information is preprocessed to generate a channel sequence with embedded features; The prediction model is selected based on the channel sequence input with embedded features to predict the channel state information at future times.

2. The channel prediction method based on multi-periodic feature convolutional networks according to claim 1, characterized in that, Methods for weighting historical channel state information based on the current environmental context features include: Current context features Input to a lightweight weighted network, the output of which is for each historical time step weight ; Weighted historical channel state information ; in, Z represents the samples in the historical channel state information, z is the historical time step index, used to represent the samples in the historical channel state information at different historical moments; Z represents the total number of samples in the fused historical channel state information.

3. The channel prediction method based on multi-periodic feature convolutional networks according to claim 2, characterized in that, Methods for predicting channel data at future times using multi-periodic feature convolutional network channel prediction models include: Multiple periods are extracted from the channel sequence with embedded features using Fast Fourier Transform: ； Where A is the frequency energy distribution vector obtained by averaging the amplitude spectrum of the channel time series along the characteristic dimension after performing a Fast Fourier Transform, which is used to characterize the energy intensity of channel variations at different frequency components. This represents a one-dimensional time series composed of channel sequences; and These represent the calculation of the Fast Fourier Transform and the amplitude value, respectively, selecting the first... Each amplitude value is obtained. frequency , corresponding to Length of each cycle ; Reconstruct the one-dimensional time series into multiple two-dimensional tensors based on each period: ； in, This represents the two-dimensional tensor obtained after reconstruction based on the i-th principal period. Used to extend the time series with zeros in the time dimension, so that it is consistent with... compatible, and This represents the number of rows and columns of the transformed two-dimensional tensor; in the two-dimensional tensor, one dimension represents the variation within a period, and the other dimension represents the variation during the period. Two-dimensional convolution features are extracted from the two-dimensional tensors and fused with periodic features to capture changes within and during the period: ； in, Represents the two-dimensional tensor corresponding to the i-th period. The periodic feature representation tensor after two-dimensional convolution feature extraction; The two-dimensional features are converted back to one-dimensional features, and the sequence length is truncated to the original length T using Trunc(∙): ； in, This represents the tensor of two-dimensional periodic features of the i-th period. The one-dimensional periodic feature sequence obtained after dimensional transformation; Use the Softmax function to The amplitudes are normalized to obtain the aggregate weights. : ； The aggregation weight and The weighted summation of the one-dimensional features yields the output of the multi-period feature convolutional network: ； The above algorithm constitutes a temporal two-dimensional variational module. The multi-period feature convolutional network is composed of multiple temporal two-dimensional variational modules stacked together and uses residual connections. The output of the layer is represented as: ； Will By using a variable dimension projection layer, the internal feature dimensions of the model are projected... Actual channel dimension projected onto the original channel state information Subsequently, the output layer extracts from the projected output sequence. Data from each time step is used as a prediction of channel state information for future times. .

4. The channel prediction method based on multi-periodic feature convolutional networks according to claim 1, characterized in that, Methods for pre-storing model pools containing multiple multi-period feature convolutional network channel prediction models include: Model pooling for training multiple multi-cycle feature convolutional network models Each model is trained for a specific environmental context. Channel training dataset They are divided into channel training samples and channel verification samples; Based on the prediction results of the channel training samples and the channel identification verification samples, the loss function is calculated; Based on the gradient backpropagation algorithm and the loss function, the parameters of the channel prediction model are adjusted to obtain the trained channel prediction model.

5. The channel prediction method based on multi-period feature convolutional networks according to claim 4, characterized in that, The loss function is: ； in, Represents the loss function. Indicates the first The aforementioned channel identification verification samples, Indicates the first The prediction results of the channel training samples. The number of channel samples. This represents the square of the L2 norm.

6. The channel prediction method based on multi-period feature convolutional networks according to claim 4, characterized in that, Used for training the model pool training dataset The following conditions must be met: Each channel state information sample in the sample is appended with an environmental context label. ; This includes one or more types of information such as the terminal device's mobile status, signal quality, and environmental geometry. according to Will Divided into multiple environment-specific training subsets The subset is then used to train the corresponding channel prediction model.

7. The channel prediction method based on multi-period feature convolutional networks according to claim 4, characterized in that, Also includes: Using newly acquired channel state information samples, the selected multi-periodic feature convolutional network channel prediction model is locally incrementally fine-tuned to obtain the model update increment. ; use Update the corresponding multi-period feature convolutional channel prediction model in the model pool.

8. A channel prediction system based on a multi-periodic feature convolutional network, characterized in that, This includes network equipment and terminal equipment; The network device is equipped with a model pool, which stores multiple multi-period feature convolutional network channel prediction models. The terminal device is used to acquire historical channel state information and current environmental context features in the communication scenario. Based on the environmental context features at the current moment, a multi-period feature convolutional network channel prediction model corresponding to the environmental context features is selected from the model pool; The historical channel state information is weighted based on the environmental context characteristics at the current moment; The weighted historical channel state information is preprocessed to generate a channel sequence with embedded features; The prediction model is selected based on the channel sequence input with embedded features to predict channel data at future times.

9. The channel prediction system based on a multi-period feature convolutional network according to claim 8, characterized in that, The multi-periodic feature convolutional network channel prediction model includes: The periodicity mining module is used to extract multiple main periods from the channel sequence via Fast Fourier Transform: ； Where A is the frequency energy distribution vector obtained by averaging the amplitude spectrum of the channel time series along the characteristic dimension after performing a Fast Fourier Transform, which is used to characterize the energy intensity of channel variations at different frequency components. This represents a one-dimensional time series composed of channel sequences; and These represent the calculation of the Fast Fourier Transform and the amplitude value, respectively, selecting the first... Each amplitude value is obtained. frequency , corresponding to Length of each cycle ; The periodic reconstruction module is used to reshape a one-dimensional time series into multiple two-dimensional tensors based on each period: ； in, This represents the two-dimensional tensor obtained after reconstruction based on the i-th principal period. Used to extend the time series with zeros in the time dimension, so that it is consistent with... compatible, and The number of rows and columns of the transformed two-dimensional tensor are represented; in the two-dimensional tensor, one dimension represents the change within a period, and the other dimension represents the change during the period. A multi-periodic feature convolution module is used to extract two-dimensional convolutional features from the two-dimensional tensor and fuse periodic features. A parameter-efficient Inception module is used as the two-dimensional convolution module to capture changes within and during the period. ； in, Represents the two-dimensional tensor corresponding to the i-th period. The periodic feature representation tensor after two-dimensional convolution feature extraction; The feature aggregation module is used to convert the two-dimensional feature representation back to a one-dimensional feature representation, and uses Trunc(∙) to truncate the sequence length to the original length T: ； in, This represents the tensor of two-dimensional periodic features of the i-th period. The one-dimensional periodic feature sequence obtained after dimensional transformation; Use the Softmax function to apply this The amplitudes are normalized to obtain the aggregate weights. : ； The aggregation weight and The weighted summation of the one-dimensional features yields the output of the multi-period feature convolutional network: ； The aforementioned periodic mining module, periodic reconstruction module, multi-period feature convolution module, and feature aggregation module together constitute a temporal two-dimensional variational module. This temporal two-dimensional variational module is composed of multiple stacked temporal two-dimensional variational modules connected using residual connections. The layer output is: ； The output layer first will By using a variable dimension projection layer, the internal feature dimensions of the model are projected... Actual channel dimension projected onto the original channel state information The output layer extracts the output from the projected output sequence. Data from each time step is used as a prediction of channel state information for future times. .

10. The channel prediction system based on a multi-period feature convolutional network according to claim 8, characterized in that, Network devices deliver the channel prediction model pool to terminal devices via RRC signaling; The terminal device uses newly acquired channel state information samples to perform local incremental fine-tuning of the selected multi-periodic feature convolutional network channel prediction model, obtaining the model update increment. ; Terminal devices will transmit via uplink The report is sent to the network device, which then uses the aforementioned... Refresh the corresponding multi-period feature convolutional channel prediction model in the model pool.

Images