CSI (Channel State Information) compression and feedback model adaptive scheduling method and device based on digital twinning

Abstract

The invention relates to the technical field of CSI compression and feedback, in particular to a CSI compression and feedback model adaptive scheduling method, device and equipment based on digital twinning and a computer storage medium. According to the adaptive scheduling method for the CSI compression and feedback model, a full-life-cycle management strategy based on data acquisition-model library construction-user request-network response and scheduling is given, centralized management and scheduling of all models on a network side are realized, unified maintenance is facilitated, and the pressure of a terminal is relieved; meanwhile, the digital twin network provided by the invention can deploy models suitable for users in a maximum range, deploy the most suitable models for the users according to real-time states and personalized requirements, and ensure that the maximum performance gain is obtained on the basis of economy and high efficiency.

Description

Technical Field

[0001] The present invention relates to the field of CSI compression and feedback technology, and in particular to a CSI compression and feedback model adaptation scheduling method, device, equipment and computer storage medium based on digital twins. Background Art

[0002] In massive MIMO communication systems, channel state information (CSI) compression and feedback methods based on AI / ML models have great application potential and have been proven to offer significant performance gains compared to traditional codebook-based compression schemes. However, different CSI / AI models can exhibit performance differences due to inconsistent feedback overhead and the size of model parameters, resulting in varying CSI recovery accuracy. Furthermore, user status is constantly changing, and rapidly evolving physical environments and communication requirements place higher demands on the efficiency and accuracy of real-time model scheduling. These issues raise the question of adapting models to user needs. A single AI model trained in a fixed pattern cannot adapt to all users. Therefore, to best meet the personalized needs of diverse users, multiple CSI compression and feedback models need to be pre-trained and deployed. Due to limited computing and storage capabilities on terminals, addressing these issues requires establishing a unified response and decision-making mechanism on the network side for all CSI compression and feedback models, enabling efficient model scheduling and management that best meets user expectations. Summary of the Invention

[0003] Therefore, the technical problem to be solved by the present invention is to overcome the problem in the prior art that a single CSI compression and feedback model cannot adapt to all users.

[0004] To solve the above technical problems, the present invention provides a CSI compression and feedback model adaptive scheduling method, comprising:

[0005] Collecting historical user CSI data to construct a first training set, and generating multiple different CSI compression and feedback function models based on the first training set;

[0006] Sending the multiple different CSI compression and feedback function models to user samples for testing, obtaining status information, demand information, and corresponding usage feedback information of the user samples, and constructing a second training set;

[0007] Learning a mapping relationship between user status information and demand information and an optimal matching CSI compression and feedback function model based on the second training set, and training and generating a basic model;

[0008] Build the target digital twin network based on multiple CSI compression and feedback function models and basic models;

[0009] The target user status information and demand information are input into the target digital twin network, and the target optimal CSI compression and feedback function model is scheduled to perform CSI compression and feedback.

[0010] Preferably, the method of generating a plurality of different CSI compression and feedback function models based on the first training set training includes:

[0011] Based on different model structures, a plurality of different CSI compression and feedback function models are generated according to the first training set training, or

[0012] Based on the same model structure, a plurality of different CSI compression and feedback function models are generated according to the first training set training under different quantization feedback bit numbers, or

[0013] Based on different model structures, and for each different model structure, a plurality of different CSI compression and feedback function models are generated according to the first training set training under different quantization feedback bit numbers.

[0014] Preferably, the status information includes one or more of reference signal received power, signal-to-noise ratio and interference ratio, channel quality indicator, large-scale fading, and user location.

[0015] Preferably, the demand information includes one or more of throughput demand information and file transfer protocol service request information.

[0016] Preferably, the feedback information used is a weighted sum of a user test feedback information evaluation index and a feedback overhead of the model, wherein the user test feedback information evaluation index is the square of the cosine similarity between the ideal CSI information and the model inference output information or the throughput within the user test time period.

[0017] Preferably, learning the mapping relationship between user status information and demand information and the optimal matching CSI compression and feedback function model based on the second training set, and training and generating the basic model includes:

[0018] Taking user status information and demand information as input and the optimal matching CSI compression and feedback function model as output, a neural collaborative filtering network model is constructed;

[0019] According to the second training set, the neural collaborative filtering network model is iteratively trained using a gradient descent algorithm and a back propagation algorithm until the loss function converges to generate a basic model.

[0020] Preferably, inputting the target user status information and demand information into the target digital twin network and scheduling the target optimal CSI compression and feedback function model to perform CSI compression and feedback includes:

[0021] Obtain the target user status information and demand information at preset time intervals, and input the target digital twin network to determine the target optimal CSI compression and feedback function model. If all models of the current target user are not the target optimal CSI compression and feedback function model, start the scheduling process, or

[0022] When the target user detects that all its models are not the optimal models in the current state, the target user status information and demand information are input into the target digital twin network to determine the target optimal CSI compression and feedback function model, and start the scheduling process.

[0023] The present invention also provides a CSI compression and feedback model adaptation scheduling device, comprising:

[0024] A functional model training module is configured to collect historical user CSI data to construct a first training set, and to generate a plurality of different CSI compression and feedback functional models based on the first training set;

[0025] A training set construction module is used to send the multiple different CSI compression and feedback function models to user samples for testing, obtain status information, demand information and corresponding usage feedback information of the user samples, and construct a second training set;

[0026] A basic model training module, configured to learn a mapping relationship between user status information and demand information and an optimal matching CSI compression and feedback function model based on the second training set, and train and generate a basic model;

[0027] A digital twin network construction module, used to build a target digital twin network based on multiple CSI compression and feedback function models and basic models;

[0028] The target adaptation scheduling module is used to input the target user status information and demand information into the target digital twin network, and schedule the target optimal CSI compression and feedback function model to perform CSI compression and feedback.

[0029] The present invention also provides a CSI compression and feedback model adaptation scheduling device, including:

[0030] Memory for storing computer programs;

[0031] A processor is used to implement the above-mentioned steps of the CSI compression and feedback model adaptation scheduling method when executing the computer program.

[0032] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned CSI compression and feedback model adaptation scheduling method are implemented.

[0033] The above technical solution of the present invention has the following advantages over the prior art:

[0034] The CSI compression and feedback model adaptation scheduling method described in the present invention provides a management strategy based on the entire life cycle of data acquisition, model library construction, user request, network response and scheduling. It implements centralized management and scheduling of all models on the network side, facilitates unified maintenance, and reduces pressure on terminals. At the same time, the digital twin network provided by the present invention can deploy user-applicable models to the greatest extent possible and allocate the most appropriate model to users based on real-time status and personalized needs, ensuring maximum performance gains on an economical and efficient basis. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0036] Figure 1 This is a flow chart of an implementation method of a CSI compression and feedback model adaptation scheduling method provided by the present invention;

[0037] Figure 2 This is a schematic diagram of the basic model and functional model architecture and interaction relationship of a scheduling method based on a digital twin CSIu, compression, and feedback model provided by an embodiment of the present invention;

[0038] Figure 3 This is a flowchart of the interaction between a base station and a terminal based on a digital twin CSI compression and feedback model scheduling method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The core of the present invention is to provide a CSI compression and feedback model adaptation scheduling method, device, equipment and computer storage medium, which can effectively allocate the most suitable model to users according to real-time status and personalized needs.

[0040] In order to enable those skilled in the art to better understand the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0041] Please refer to Figure 1 , Figure 1 This is a flowchart of an implementation method for CSI compression and feedback model adaptation scheduling provided by the present invention; the specific operation steps are as follows:

[0042] S101: Collect historical user CSI data to construct a first training set, and generate multiple different CSI compression and feedback function models based on the first training set;

[0043] S102: Send the multiple different CSI compression and feedback function models to user samples for testing, obtain status information, demand information, and corresponding usage feedback information of the user samples, and construct a second training set;

[0044] S103: Learning a mapping relationship between user status information and demand information and an optimal matching CSI compression and feedback function model based on the second training set, and training to generate a basic model;

[0045] S104: Build a target digital twin network based on multiple CSI compression and feedback function models and basic models;

[0046] S105: Input the target user status information and demand information into the target digital twin network, and schedule the target optimal CSI compression and feedback function model to perform CSI compression and feedback.

[0047] Based on the above embodiment, this embodiment describes step S101 in detail:

[0048] In one embodiment, a large number of user samples with different statuses in a cell are collected, their historical CSI data are obtained, and a first training set of the functional model is generated.

[0049] The accuracy performance differences of the models can come from differences in model structure and cost feedback. The specific model structure is not specified in this invention. Common CsiNet structures such as convolutional neural networks, MLP, and Transformer can be used to enrich the model library:

[0050] In one embodiment, based on different model structures, a plurality of different CSI compression and feedback function models are generated according to the first training set training. Common model structures may be based on convolutional neural networks, MLP, Transformer, etc.

[0051] In one embodiment, based on the same model structure, multiple different CSI compression and feedback function models are generated by training according to the first training set at different quantized feedback bit numbers, and the feedback bit numbers from low to high correspond to the trained models having performance gains from low to high.

[0052] In one embodiment, based on different model structures, and for each different model structure, a plurality of different CSI compression and feedback function models are generated according to the first training set training under different quantization feedback bit numbers. If models with N model structures are to be deployed in the function model library, and each model is trained according to M types of feedback overheads, then the function model library will have a total of N*M models with different performance gains.

[0053] Based on the above embodiment, this embodiment describes step S102 in detail:

[0054] In one embodiment, all trained functional models are sent to the test user samples, and each user uses all functional models to simultaneously perform the CSI feedback process within the test time range. The length of the test time is reasonably set according to the time interval recommended by the model.

[0055] Regarding the selected initial state information, the present invention provides the following suggested implementation methods:

[0056] In one embodiment, the reference signal received power (RSRP), signal-to-noise ratio and interference ratio (SINR), and channel quality indicator (CQI) are used as the channel state information S of the user. i

[0057] In one embodiment, large-scale fading is used as the channel state information S of the user. i

[0058] In one embodiment, the user's location is used as the user's status information S i

[0059] In one embodiment, the status information includes one or more of reference signal received power, signal-to-noise ratio and interference ratio, channel quality indicator, large-scale fading, user location, etc.

[0060] Regarding the selected user demand information, the present invention provides the following suggested implementation methods:

[0061] In one embodiment, throughput requirements are used as the user's personalized requirement information feature.

[0062] In one embodiment, the file transfer protocol service request information (the current packet length P of the user under the file transfer protocol service model) is used. i , service packet arrival rate A i ) as the user's personalized demand information feature R i ={P i ,A i}.

[0063] In one embodiment, the demand information includes one or more of throughput demand information and file transfer protocol service request information.

[0064] The user's status information and demand information are combined to form the feature vector X representing the i-th user's status and demand i ={S i ,R i}, as the input of the basic model.

[0065] Evaluation metrics selected for user testing feedback:

[0066] In one embodiment, the SGCS between the ideal CSI information and the model inference output information is used as the measurement feedback information evaluation index T of the model. ij :

[0067]

[0068] where w k The eigenvector representing the ideal CSI within the kth resource unit, w′ k It represents the feature vector of the model prediction output CSI within the k-th resource unit; GCS represents cosine similarity, and SGCS is the squared GCS.

[0069] In one embodiment, the throughput during the user test period is used as the measurement information feedback evaluation index T of the model. ij Specifically, the user uses the currently tested model to test the CSI compression and feedback process within a specified time period, and the total throughput of the user during this period is counted as the feedback evaluation indicator.

[0070] For the settings of training labels:

[0071] In one embodiment, a comprehensive evaluation index method is adopted to evaluate the user test feedback information evaluation index T ij And the feedback cost F of the model ij The trade-off relationship between (where i represents the i-th user and j represents the j-th model) is established by setting certain weights for the two quantities to synthesize the current user's evaluation score for the model, namely:

[0072] S ij =W1*T ij +W2*F ij

[0073] The model with the highest score is taken as the current optimal model, which is the label.

[0074] Based on the above embodiment, this embodiment describes step S103 in detail:

[0075] The implementation of the basic model can be technically described as follows: during the training process, it is necessary to comprehensively learn the comprehensive characteristics formed by user needs and user channel status according to a set of reasonable evaluation strategies, construct a potential mapping relationship between their matching degree with different models, and thus output the optimal model selection strategy. The evaluation metrics used to measure whether the selected model is optimal need to comprehensively consider multiple aspects such as user performance gain and overhead burden. In essence, it is to find the optimal trade-off between meeting the user's basic and personalized needs as much as possible and the actual overhead burden. The set evaluation strategy is applied to the training process of the basic model in the form of training labels. When the model training reaches convergence, it is considered that it has learned the optimal matching strategy.

[0076] In one embodiment, learning a mapping relationship between user status information and demand information and an optimal matching CSI compression and feedback function model based on the second training set, and training and generating a basic model includes:

[0077] Taking user status information and demand information as input and the optimal matching CSI compression and feedback function model as output, a neural collaborative filtering network model (NCF) is constructed;

[0078] According to the second training set, the neural collaborative filtering network model is iteratively trained using a gradient descent algorithm and a back propagation algorithm until the loss function converges to generate a basic model.

[0079] In one embodiment, the loss function may be a mean square error (MSE) loss function, a mean absolute error (MAE) loss function, a Huber loss function, or the like.

[0080] In one embodiment, for the process of determining the hyperparameters of the network model, the number of learning rounds can be set to T. The setting of the learning rounds needs to weigh the impact of the model training speed, training cost, and model training accuracy; the learning rate is set to 1; and the weight initialization method selects random weight initialization.

[0081] In one embodiment, the base station calculates a training loss value based on the output result of each grouping model and the label value information, for example, using a mean square error loss function to calculate the training loss value:

[0082]

[0083] Among them, I represents the number of samples belonging to model training, y i Represents the output of the model. Represents the label value of data i. In the embodiment, DTN uses the gradient descent algorithm (SGD) and an optimizer (such as Adam) in each iteration to continuously narrow the gap between the model output and the label. After the iteration, the model parameters are updated. Taking the SGD algorithm as an example:

[0084]

[0085] in, represents the model parameters to be updated in round t, Represents the model parameters after the tth round of update, Represents the gradient of the training loss value calculated in the tth round, Represents the learning rate for round t. When the loss function stops decreasing after a certain number of iterations, the model converges and the NCF model is considered to have learned the mapping relationship between user status and needs and the optimal matching model.

[0086] After step S101, step S102 and step S103, the offline training phase of the DTN in the present invention is described. After the generated DTN is deployed on the network side, further, refer to Figure 2 and Figure 3 ,In the subsequent steps S104 and S105, the specific logical framework of the ,deployed DTN, and the detailed interaction process of the ,basic model and the functional model in the online application ,phase are described in detail.

[0087] Based on the above embodiment, this embodiment describes step S104 in detail:

[0088] In one embodiment:

[0089] The present invention constructs a specific digital twin network (DTN) corresponding to the physical layer air interface for the CSI feedback task. Specifically, the bottom layer is constructed as the wireless side air interface, which includes the physical layer uplink communication link between the terminal and the base station; the middle layer is divided into two parts: the functional model and the basic model, which are connected to two different data interfaces respectively. The top layer is constructed as the request and response interface between the user and the digital twin network. The detailed description of the function of each part is given below.

[0090] ● Functional model and first data interface:

[0091] A large number of CSI compression feedback models with different performances need to be trained and stored in the functional model library to perform inference compression tasks.

[0092] The first data interface is connected to the functional model library. When building the model library offline, the digital twin network can obtain the actual CSI data in the physical wireless side network cell through this interface for training.

[0093] ●Basic model and second data interface:

[0094] In this invention, the basic model simulates and constructs user demand characteristics in the real physical world through a large amount of historical data. By observing only a small amount of real-time user information, it can quickly make matches and selections, thereby helping users select the most suitable model from the current functional model library and execute the scheduling policy. Specifically, its input should include two parts: one is a specific quantity that reflects the user's current communication needs, such as throughput requirements and FTP service request information; the other is information describing the user's current channel status, such as RSRP, SINR, CQI, etc. The output should be the selection action of the optimal model.

[0095] The digital twin network deploys a second data interface on the uplink of the wireless side, and obtains user needs and status information through feedback as input to the basic model, which is then used for model training and online real-time reasoning process.

[0096] ●Third data interface:

[0097] During the online application process after DTN is deployed, users submit model request information to DTN through the third data interface and feedback its status information at the same time. This information will be used for DTN's response and decision-making.

[0098] Based on the above embodiment, this embodiment describes step S105 in detail:

[0099] In one embodiment, the target user status information and demand information are obtained at preset time intervals and input into the target digital twin network to determine the target optimal CSI compression and feedback function model. If all models of the current target user are not the target optimal CSI compression and feedback function model, the scheduling process is started, that is, the DTN network performs a model scheduling recommendation process for the user at a fixed time interval, that is, the user does not need to detect the adaptability of the model he currently has. After the model recommendation cycle arrives, if the DTN detects that the model the user has at this time is the current optimal model, it will not trigger a new model recommendation process of the DTN; if the DTN detects that the model the user has at this time is not the current optimal model, the DTN executes the model scheduling process to recommend a new function model to the user.

[0100] In one embodiment, when the target user detects that all of its models are not the optimal models in the current state, the target user state information and demand information are input into the target digital twin network to determine the target optimal CSI compression and feedback function model, and start the scheduling process, that is: the time interval for model scheduling recommendation is not fixed, and the user needs to continuously detect the degree of adaptability of the model it currently has. If the user detects that the current model is not the optimal model in the current state, a model recommendation request will be sent to the DTN network. The DTN network will execute the model scheduling process only after receiving the user's request.

[0101] In one embodiment, the base station finds the corresponding model from the functional model library according to the indication output by the DTN basic model. Specifically, for example, based on the difference in the functional models described in step 102, if the decision action output by the basic model is to select the jth feedback bit number under the i-th model structure, the base station selects the ijth model from the functional model library and sends the model to the user. The user uses the received functional model to perform the CSI compression and feedback process.

[0102] This paper proposes a CSI feedback model scheduling method based on digital twins. DTN constructs interest mapping relationships for different CSI compression models under different user states and needs. A unified interest model is trained for users served by the same cell base station. Based on the user's request information, the optimal matching model for the user within the recommendation cycle is determined. This method minimizes the user's personalized needs while reducing the model's computational overhead and feedback resource overhead. DTN also deploys all models uniformly on the base station side for scheduling, reducing the computing pressure on the terminal and improving the management efficiency of the model throughout its lifecycle.

[0103] The specific design scheme and interaction relationship of the basic model and functional model in the digital twin network of the present invention utilizes the AI model to manage the AI model, constructing a mapping relationship between the user's current QoS requirements and the degree of matching with the CSI compression model, while meeting the user's personalized needs and accuracy as much as possible, while reducing feedback overhead.

[0104] The present invention is based on unified model deployment and model scheduling strategy on the base station side. The terminal only needs to make model requests to the DTN according to the prescribed recommended process, without having to store all models locally. While ensuring efficient management, it saves terminal storage space and computing resources.

[0105] An embodiment of the present invention further provides a CSI compression and feedback model adaptation scheduling device; the specific device may include:

[0106] A functional model training module is configured to collect historical user CSI data to construct a first training set, and to generate a plurality of different CSI compression and feedback functional models based on the first training set;

[0107] A training set construction module is used to send the multiple different CSI compression and feedback function models to user samples for testing, obtain status information, demand information and corresponding usage feedback information of the user samples, and construct a second training set;

[0108] A basic model training module, configured to learn a mapping relationship between user status information and demand information and an optimal matching CSI compression and feedback function model based on the second training set, and train and generate a basic model;

[0109] A digital twin network construction module, used to build a target digital twin network based on multiple CSI compression and feedback function models and basic models;

[0110] The target adaptation scheduling module is used to input the target user status information and demand information into the target digital twin network, and schedule the target optimal CSI compression and feedback function model to perform CSI compression and feedback.

[0111] The CSI compression and feedback model adaptation scheduling device of this embodiment is used to implement the aforementioned CSI compression and feedback model adaptation scheduling method. Therefore, the specific implementation methods of the CSI compression and feedback model adaptation scheduling device can be seen in the embodiment part of the CSI compression and feedback model adaptation scheduling method mentioned above. For example, the functional model training module, the training set construction module, the basic model training module, the digital twin network construction module, and the target adaptation scheduling module are respectively used to implement steps S101, S102, S103, S104 and S105 in the above-mentioned CSI compression and feedback model adaptation scheduling method. Therefore, its specific implementation methods can refer to the description of the corresponding embodiments of each part and will not be repeated here.

[0112] A specific embodiment of the present invention further provides a CSI compression and feedback model adaptation scheduling device, comprising: a memory for storing a computer program; and a processor for implementing the steps of the above-mentioned CSI compression and feedback model adaptation scheduling method when executing the computer program.

[0113] A specific embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned CSI compression and feedback model adaptive scheduling method are implemented.

[0114] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0115] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0116] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0117] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0118] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A CSI compression and feedback model adaptation scheduling method based on digital twin, characterized in that: include: Collecting historical user CSI data to construct a first training set, and generating multiple different CSI compression and feedback function models based on the first training set; Sending the multiple different CSI compression and feedback function models to user samples for testing, obtaining status information, demand information, and corresponding usage feedback information of the user samples, and constructing a second training set; Learning a mapping relationship between user status information and demand information and an optimal matching CSI compression and feedback function model based on the second training set, and training and generating a basic model; Build the target digital twin network based on multiple CSI compression and feedback function models and basic models; The target user status information and demand information are input into the target digital twin network, and the target optimal CSI compression and feedback function model is scheduled to perform CSI compression and feedback.

2. The CSI compression and feedback model adaptation scheduling method based on digital twin according to claim 1 is characterized in that: The method for generating a plurality of different CSI compression and feedback function models based on the first training set includes: Based on different model structures, a plurality of different CSI compression and feedback function models are generated according to the first training set training, or Based on the same model structure, a plurality of different CSI compression and feedback function models are generated according to the first training set training under different quantization feedback bit numbers, or Based on different model structures, and for each different model structure, a plurality of different CSI compression and feedback function models are generated according to the first training set training under different quantization feedback bit numbers.

3. The CSI compression and feedback model adaptation scheduling method based on digital twin according to claim 1 is characterized in that: The state information includes one or more of reference signal received power, signal-to-noise ratio and interference ratio, channel quality indicator, large-scale fading, and user location.

4. The CSI compression and feedback model adaptation scheduling method based on digital twin according to claim 1 is characterized in that: The demand information includes one or more of throughput demand information and file transfer protocol service request information.

5. The CSI compression and feedback model adaptation scheduling method based on digital twin according to claim 1 is characterized in that: The feedback information used is a weighted sum of a user test feedback information evaluation indicator and a feedback overhead of the model, wherein the user test feedback information evaluation indicator is the square of the cosine similarity between the ideal CSI information and the model inference output information or the throughput within the user test time period.

6. The CSI compression and feedback model adaptation scheduling method based on digital twin according to claim 1, characterized in that: The learning of the mapping relationship between the user status information and the demand information and the optimal matching CSI compression and feedback function model based on the second training set, and the training and generation of the basic model includes: Taking user status information and demand information as input and the optimal matching CSI compression and feedback function model as output, a neural collaborative filtering network model is constructed; According to the second training set, the neural collaborative filtering network model is iteratively trained using a gradient descent algorithm and a back propagation algorithm until the loss function converges to generate a basic model.

7. The CSI compression and feedback model adaptive scheduling method based on digital twin according to claim 1, characterized in that: Inputting the target user status information and demand information into the target digital twin network and scheduling the target optimal CSI compression and feedback function model to perform CSI compression and feedback includes: Obtain the target user status information and demand information at preset time intervals, and input the target digital twin network to determine the target optimal CSI compression and feedback function model. If all models of the current target user are not the target optimal CSI compression and feedback function model, start the scheduling process, or When the target user detects that all its models are not the optimal models in the current state, the target user status information and demand information are input into the target digital twin network to determine the target optimal CSI compression and feedback function model, and start the scheduling process.

8. A CSI compression and feedback model adaptation scheduling device based on digital twin, characterized in that: include: A functional model training module is configured to collect historical user CSI data to construct a first training set, and to generate a plurality of different CSI compression and feedback functional models based on the first training set; A training set construction module is used to send the multiple different CSI compression and feedback function models to user samples for testing, obtain status information, demand information and corresponding usage feedback information of the user samples, and construct a second training set; A basic model training module, configured to learn a mapping relationship between user status information and demand information and an optimal matching CSI compression and feedback function model based on the second training set, and train and generate a basic model; A digital twin network construction module, used to build a target digital twin network based on multiple CSI compression and feedback function models and basic models; The target adaptation scheduling module is used to input the target user status information and demand information into the target digital twin network, and schedule the target optimal CSI compression and feedback function model to perform CSI compression and feedback.

9. A CSI compression and feedback model adaptation scheduling device based on digital twin, characterized in that: include: memory for storing computer programs; A processor, configured to implement the steps of a CSI compression and feedback model adaptation scheduling method based on digital twins as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, it implements the steps of a CSI compression and feedback model adaptation scheduling method based on digital twins as described in any one of claims 1 to 7.

Images