Zero-overhead environment adaptive intelligent channel feedback method
Through the introduction of scene feature information by the autoencoder and hypernetwork structure, the channel feedback neural network model is fine-tuned, which solves the data collection and computing power requirements caused by environmental changes in the existing technology, and realizes zero-overhead environmental adaptive channel feedback, improving the accuracy and efficiency of channel feedback.
Patent Information
- Application Number
- CN202510331823.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-20
- Publication Date
- 2025-07-04
AI Technical Summary
The existing intelligent CSI feedback method cannot quickly adapt to the propagation environment changes, resulting in additional data collection costs and computing resource requirements, and cannot achieve efficient environmental adaptation.
The autoencoder structure and hypernetwork or equivalent structure are used to introduce characterization scene feature information, pre-train channel feedback neural network model, fine-tune parameters to adapt to the new environment, and realize zero-overhead environmental adaptive channel feedback.
Without adding additional overhead and computing power costs, the scenario adaptability and accuracy of the channel feedback network are improved, and data collection needs are reduced.
Smart Images
Figure CN120263242A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of communication technologies, and particularly to a zero-overhead environment-adaptive intelligent channel feedback method. Background Art
[0002] Massive Multiple-Input Multiple-Output (Massive MIMO) technology is an important technology for the fifth-generation (5G) communication system. By configuring a large number of transmit antennas at the Base Station (BS), it significantly improves the utilization efficiency of the spectrum and energy of the communication system. To achieve the performance improvement brought by Massive MIMO, the base station needs to accurately obtain the downlink channel state information (Channel State Information, CSI), and the accuracy of CSI is crucial for enhancing the performance of MIMO. In a Frequency-Division Duplexing (FDD) system, CSI is estimated by the User Equipment (UE) and fed back to the base station through the uplink of the UE. In recent years, Artificial Intelligence (AI) technology has been applied to realize CSI feedback and significant gains have been obtained. However, when the propagation environment changes, the distribution of CSI also changes, which means that the previously trained feedback neural network based on the AI method may not achieve stable results in the new environment.
[0003] Some current existing works usually introduce online learning schemes into the intelligent CSI feedback task, including methods such as transfer learning and domain adaptation. However, the online learning scheme requires collecting new CSI samples for real-time training, which will result in additional data collection costs and requires powerful computing resources and cannot quickly adapt to the new environment. In recent years, digital twin technology has provided new opportunities for changing the channel acquisition process. It infers various information about the wireless channel according to different precise 3D scene feature information and ray tracing technology, such as propagation path parameters, channel covariance information, and link quality. Generally, the information characterizing the scene features can fully reflect the propagation environment and has an obvious correlation with the CSI sample distribution, so it has a certain auxiliary effect on CSI feedback reconstruction. In order to make full use of the communication characteristics in different scenarios, improve the generalization of the channel feedback network, and without bringing additional overhead and computing costs, it is necessary to study how to design a zero-overhead environment-adaptive intelligent CSI feedback method on the basis of considering introducing the information characterizing the scene features. Summary of the Invention
[0004] The present invention provides a zero-overhead environment-adaptive intelligent channel feedback method, which uses an autoencoder structure to compress and reconstruct channel data, and at the same time uses a hypernetwork or an equivalent structure to introduce information representing scene features, enabling the feedback neural network model to learn the feature information in the propagation scene, and effectively improving the scene adaptability of the model without increasing additional feedback overhead.
[0005] An embodiment of the present invention provides a zero-overhead environment-adaptive intelligent channel feedback method, including the following steps:
[0006] Step 1, collect channel data from different scenes and information data representing scene features corresponding to the scenes;
[0007] Step 2, construct an autoencoder neural network, and use the collected channel data of the mixed scenes to pre-train a channel feedback neural network model;
[0008] Step 3, use a hypernetwork or an equivalent structure to introduce information representing scene features, and fine-tune the parameters of the pre-trained feedback neural network model;
[0009] Step 4, after the fine-tuning of the pre-trained feedback neural network model is completed, introduce the feature information of the new scene to obtain a channel feedback network with high adaptability to the new scene.
[0010] Optionally, in an embodiment of the present invention, in Step 1, the channel data is the channel matrix H. In order to enable the channel feedback network to learn the relationship between different scene features and the distribution of the channel matrix H, each channel matrix data corresponds to the feature data of the specific scene from which it is collected.
[0011] Optionally, in an embodiment of the present invention, in Step 2, the input of the channel feedback neural network model is the channel matrix H, and the output is the reconstructed channel matrix. This network consists of a fully connected layer, a batch normalization layer, and a convolutional layer.
[0012] Optionally, in an embodiment of the present invention, in Step 2, the channel feedback neural network model uses the mean square error between the input channel matrix H and the reconstructed channel matrix as the cost function for training to minimize the cost function. The cost function is described as follows:
[0013]
[0014] where f AE is the channel feedback autoencoder, and |·|2 is the Euclidean norm.
[0015] Optionally, in an embodiment of the present invention, in step 3, the network for introducing information characterizing scene features consists of a neural network, whose input is the information characterizing scene features and output is the parameter Θ' for fine-tuning the pre-trained feedback neural network model. AE .
[0016] Optionally, in an embodiment of the present invention, in step 3, the network is trained using the mean square error between the input channel matrix H and the reconstructed channel matrix as the cost function to minimize the cost function, and the cost function is described as follows:
[0017]
[0018] where Θ' AE is the parameter for fine-tuning the pre-trained feedback autoencoder model by the hypernetwork or equivalent structure, f' AE is the fine-tuned channel feedback autoencoder, and |·|2 is the Euclidean norm.
[0019] Optionally, in an embodiment of the present invention, in step 4, after the fine-tuning of the feedback neural network model is completed, the network will introduce the feature information from new scenes. The feature information of these new scenes can help the network readjust its internal parameters under new environmental conditions to ensure the high adaptability of the model to new scenes.
[0020] Optionally, the information characterizing scene features includes at least one of physical layout, obstacle distribution, or environmental conditions.
[0021] Optionally, the channel feedback neural network is deployed at the user side and the base station side. The user side executes the encoder and the base station side executes the decoder to achieve the compression and reconstruction of channel state information.
[0022] Optionally, the method is applicable to a frequency division duplexing (FDD) massive multiple input multiple output (Massive MIMO) system and does not require online collection of channel data for new scenes.
[0023] The zero-overhead environment-adaptive intelligent channel feedback method of the embodiment of the present invention considers problems such as the additional data collection cost and the large demand for computing power resources in the current online learning intelligent CSI feedback scheme, effectively utilizes the hypernetwork or equivalent structure to learn the important information about the propagation environment contained in the information characterizing different scene features, and customizes the adjustment of the parameters corresponding to different scene characteristics for the pre-trained encoder, realizing an intelligent channel state information feedback scheme that is environment-adaptive without additional overhead.
[0024] The additional aspects and advantages of the present invention will be given in part in the following description, become apparent in part from the following description, or be understood through the practice of the present invention. Brief Description of the Drawings
[0025] The above and / or additional aspects and advantages of the present invention will become apparent and be readily understood from the following description of embodiments in conjunction with the accompanying drawings, where:
[0026] Figure 1 FIG. is a flowchart of a zero-overhead environment-adaptive intelligent channel feedback method provided according to an embodiment of the present invention;
[0027] Figure 2 FIG. is a framework diagram of a zero-overhead environment-adaptive intelligent channel feedback method according to an embodiment of the present invention;
[0028] Figure 3 FIG. is a neural network architecture diagram of a zero-overhead environment-adaptive intelligent channel feedback method according to an embodiment of the present invention. Detailed Description of the Embodiments
[0029] Embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present invention and should not be construed as limiting the present invention.
[0030] In the task of improving the generalization of the intelligent CSI feedback scenario, since online learning schemes such as transfer learning and domain adaptation methods need to collect new CSI samples for real-time training, it will result in additional data collection costs and require powerful computing resources, and cannot quickly adapt to new environments. Therefore, it is necessary to design an intelligent CSI feedback method that can use information characterizing the communication scenario features, make full use of the communication characteristics in different scenarios, improve the generalization of the CSI feedback network, and do not bring additional overhead and computing costs.
[0031] Figure 1 FIG. is a flowchart of a zero-overhead environment-adaptive intelligent channel feedback method provided according to an embodiment of the present invention.
[0032] As Figure 1 shown, the zero-overhead environment-adaptive intelligent channel feedback method includes the following steps:
[0033] Step 1, collect channel data from different scenarios and information data characterizing the scenario features of the corresponding scenarios.
[0034] In an embodiment of the present invention, the channel data is the channel matrix H, which is the CSI collected in different scenarios; while the data representing the scenario features can represent the features of these different scenarios, such as the physical layout, obstacle distribution, environmental conditions, etc. of the corresponding scenario. In order for the intelligent channel feedback network to learn the relationship between different scenario features and the channel matrix distribution, each channel matrix data corresponds to the feature data of the specific scenario from which it is collected.
[0035] Step 2: Construct an autoencoder neural network and pre-train a channel feedback neural network model using the collected channel data of the mixed scenarios.
[0036] In an embodiment of the present invention, the input of the channel feedback neural network model is the channel matrix H, and the output is the reconstructed channel matrix. This network adopts an autoencoder structure and consists of fully connected layers, batch normalization layers, and convolutional layers. The channel feedback neural network model uses the mean square error between the input channel matrix H and the reconstructed channel matrix as the cost function for training to minimize the cost function, and the cost function is described as follows:
[0037]
[0038] where f AE is the channel feedback autoencoder, and |·|2 is the Euclidean norm.
[0039] Step 3: Use a hypernetwork or an equivalent structure to introduce the information representing the scenario features and fine-tune the parameters of the pre-trained feedback neural network model.
[0040] In an embodiment of the present invention, the network for introducing the information representing the scenario features consists of a neural network, whose input is the information representing the scenario features, and the output is the parameter Θ' for fine-tuning the pre-trained feedback neural network model. AE . The neural network learns the communication environment knowledge existing in the scenario feature information, thereby obtaining the relationship between the communication scenario and the channel matrix distribution, and outputs it in the form of parameters to the pre-trained feedback neural network model for updating the model parameters. The network uses the mean square error between the input channel matrix H and the reconstructed channel matrix as the cost function for training to minimize the cost function, and the cost function is described as follows:
[0041]
[0042] where Θ' AE is the parameter of the hypernetwork or equivalent structure for fine-tuning the pre-trained feedback autoencoder model, f' AE is the fine-tuned channel feedback autoencoder, and |·|2 is the Euclidean norm.
[0043] Step 4: After the fine-tuning of the pre-trained feedback neural network model is completed, introduce the feature information of the new scenario to obtain a channel feedback network with high adaptability to the new scenario.
[0044] In an embodiment of the present invention, after the fine-tuning is completed, the network will introduce the feature information from the new scenario. This feature information of the new scenario can help the network readjust its internal parameters under the new environmental conditions to ensure the high adaptability of the model to the new scenario. Through this process, the feedback neural network can more accurately infer the channel state information in the new scenario, thereby improving the accuracy and efficiency of channel feedback.
[0045] The following will further elaborate on the details of the present invention in combination with specific embodiments and drawings.
[0046] The zero-overhead environment-adaptive intelligent channel feedback method uses an autoencoder structure to achieve the compression and reconstruction of channel data, and at the same time uses a hypernetwork or an equivalent structure to introduce information representing scene features, so that the feedback neural network model can learn the feature information in the propagation scene, effectively improving the scene adaptability of the model without increasing additional feedback overhead. The specific steps are as follows:
[0047] (1) In an FDD large-scale MIMO system, consider a typical single-cell downlink transmission link. Among them, a ULA is configured at the base station side, and the number of its antennas is N t = 8, and a single antenna is configured at the user side. The system adopts an OFDM modulation scheme, the number of subcarriers is N' c = 256, the center frequency is 5.8 GHz, and the bandwidth is 20 MHz. The scenario is a random scenario, and the users are randomly initialized in the scenario space. According to the above settings, a data set is generated. The data set contains 200 different scenarios, and each scenario contains 2000 CSI samples. A total of 160 scenarios are used in the training set, and 2000 CSI are selected from each scenario to form a mixed training set with a total of 320,000 data; a total of 20 scenarios are used in the validation set, and 2000 CSI are selected from each scenario to synthesize a mixed validation set with a total of 40,000 data; a total of 20 scenarios are used in the test set, and the data of each scenario is used as the test set alone.
[0048] (2) The framework diagram of the zero-overhead environment-adaptive intelligent channel feedback method as Figure 2 shown includes the following steps.
[0049] First, collect CSI samples of different scenarios, and use these samples to pre-train a channel feedback network of mixed samples. The pre-trained network model uses the mean square error between the input channel matrix H and the reconstructed channel matrix as the cost function for training to minimize the cost function. The cost function is described as follows:
[0050]
[0051] Among them, f AE is the channel feedback autoencoder, and |·|2 is the Euclidean norm.
[0052] Then, in the second step, the channel feedback network will be fine-tuned offline using the data containing scene feature information. Collect and use the scene feature information corresponding to the hybrid scene CSI samples, and use a hypernetwork or an equivalent structure to learn the implicit relationship between the communication scene feature information and the CSI samples, and generate the updated parameters of the feedback network. The fine-tuned channel feedback network still uses the mean square error between the input channel matrix H and the reconstructed channel matrix as the cost function for training to minimize the cost function, and the cost function is described as follows:
[0053]
[0054] Among them, Θ' AE is the parameter for the hypernetwork or equivalent structure to fine-tune the pre-trained feedback autoencoder model, and f' AE is the fine-tuned channel feedback autoencoder, and |·|2 is the Euclidean norm.
[0055] In the test phase, when CSI needs to be fed back in a new scene, only the information reflecting the new scene features needs to be provided to the hypernetwork or equivalent structure to obtain the feedback network parameters adapted to the new scene. The fine-tuned channel feedback network can obtain a channel feedback network with high adaptability to the new scene by applying these network parameters. Therefore, in actual use, this method does not require additional collection of CSI samples belonging to the new scene, nor does it require additional training of a new model belonging to the new scene, thus effectively reducing the data collection overhead.
[0056] (3) As Figure 3 is the neural network architecture diagram of a zero-overhead environment-adaptive intelligent channel feedback method according to an embodiment of the present invention.
[0057] This method designs a channel feedback neural network model, deploys an encoder at the user side and a decoder at the base station side to obtain the reconstructed CSI samples. Use a hypernetwork or an equivalent structure to learn the information representing the scene features, obtain the scene-customized network parameters, and provide them to the feedback network model to obtain a channel feedback network with high adaptability to the new scene.
[0058] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc., mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.
[0059] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "N" is at least two, such as two, three, etc., unless otherwise specifically defined.
[0060] Any process or method description shown in a flowchart or described in other ways herein may be understood to represent a module, segment, or portion of code including one or N executable instructions for implementing a customized logic function or process, and the scope of the preferred embodiments of the present invention includes additional implementations, where the functions may be executed in a substantially simultaneous manner or in a reverse order according to the functions involved, rather than in the order shown or discussed, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.
Claims
1. A zero-overhead environment-adaptive intelligent channel feedback method, characterized in that, It includes the following steps: Step 1: Collect channel data from different scenarios and information data representing scenario features corresponding to the scenarios; Step 2: Construct an autoencoder neural network and pre-train the channel feedback neural network model using the collected channel data of the mixed scenarios; Step 3: Use a hypernetwork or an equivalent structure to introduce the information representing scenario features and fine-tune the parameters of the pre-trained feedback neural network model; Step 4: After the fine-tuning of the pre-trained feedback neural network model is completed, introduce the feature information of the new scenario to obtain a channel feedback network adapted to the new scenario.
2. The method according to claim 1, wherein In Step 1, the channel data is the channel matrix H, and each channel matrix data corresponds to the feature data of the specific scenario from which it is collected.
3. The method according to claim 1, characterized in that, In step 2, the input of the channel feedback neural network model is the channel matrix H, and the output is the reconstructed channel matrix The network structure includes a fully connected layer, a batch normalization layer, and a convolutional layer.
4. The method according to claim 1, characterized in that, In Step 2, the channel feedback neural network model is trained using the mean square error between the input channel matrix H and the reconstructed channel matrix as the cost function to minimize the cost function, and the cost function is described as follows: where f AE is the channel feedback autoencoder, and |·|2 is the Euclidean norm.
5. The method according to claim 1, characterized in that, In step 3, the super network or equivalent structure consists of a neural network, whose input is information representing scene features and output is the parameter Θ' for fine-tuning the pre-trained feedback neural network model AE .
6. The method according to claim 1, characterized in that In Step 3, the hypernetwork or the equivalent structure is trained using the mean square error between the input channel matrix H and the reconstructed channel matrix as the cost function to minimize the cost function, and the cost function is described as follows: where, Θ' AE is the hypernetwork or equivalent structure for fine-tuning the parameters of the pre-trained feedback autoencoder model, f' AE is the fine-tuned channel feedback autoencoder, and |·|2 is the Euclidean norm.
7. The method according to claim 1, wherein In Step 4, the feature information of the new scenario is directly input into the hypernetwork or the equivalent structure to generate the channel feedback network parameters adapted to the scenario without additional training data.
8. The method according to claim 1, characterized in that The information representing scenario features includes at least one of physical layout, obstacle distribution, or environmental conditions.
9. The method according to claim 1, wherein The channel feedback neural network is deployed at the user side and the base station side. The user side executes the encoder, and the base station side executes the decoder to achieve compression and reconstruction of channel state information.
10. The method according to claim 1, characterized in that, The method is applicable to a frequency division duplexing (FDD) massive multiple input multiple output (Massive MIMO) system and does not require online collection of channel data of new scenarios.