Channel prediction reliability improvement method based on proxy model

By using proxy models to monitor and evaluate the output results of the CSI prediction model in 5G communication system, the problems of CSI feedback overhead and AI model instability in large-scale MIMO technology are solved, and the reliability improvement of channel prediction and downlink are achieved. Improved stability.

CN118890119APending Publication Date: 2024-11-01SOUTHEAST UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411144333.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-20
Publication Date
2024-11-01

AI Technical Summary

Technical Problem

In 5G communication systems, large-scale MIMO technology has limited spectral efficiency and energy efficiency due to excessive CSI feedback overhead, and AI-based CSI prediction model has instability and performance fluctuations, affecting the stability of downlink transmission.

Method used

Using proxy models to monitor the output results of CSI prediction model, train channel prediction models and proxy models through data compression and reconstruction, deploy them on user equipment and base stations, evaluate the reliability of prediction models in real time, eliminate low-precision results, and improve the stability of CSI prediction. sex.

Benefits of technology

Without increasing the burden on the equipment, the reliability of channel prediction is significantly improved, the impact of performance fluctuations of CSI prediction model on communication stability is reduced, and the transmission stability of the downlink is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118890119A_ABST
    Figure CN118890119A_ABST
Patent Text Reader

Abstract

The invention discloses a channel prediction reliability improvement method based on a proxy model, and belongs to the technical field of communication. The method comprises the following steps: collecting and compressing channel data offline; training a channel prediction model by using the collected original channel data and the compressed data; training a proxy model using the output of the prediction model; and deploying the proxy model and the channel prediction model to the user equipment and the base station respectively and running on line. According to the method provided by the invention, in the real-time channel prediction process, on the premise of slightly increasing the equipment burden, the reliability of channel prediction is improved, and the influence of the performance fluctuation of the channel prediction model on the communication stability is reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a method for improving channel prediction reliability based on an agent. Background Art

[0002] As a key technology of the fifth generation (5G) communication system, Massive Multiple-Input Multiple-Output (Massive MIMO) technology greatly improves spectral efficiency and energy efficiency by equipping the base station (BS) with a large number of transmit antennas. In order to obtain the performance gain of massive MIMO, it is necessary to obtain accurate downlink channel state information (CSI) at the BS, and the accuracy of the obtained CSI is crucial to the performance gain of MIMO. In the frequency division duplexing (FDD) system, CSI is estimated at the user equipment (UE) and fed back to the BS through the UE's uplink transmission. However, due to the increase in the number of antennas, CSI feedback generates too much feedback overhead, which puts a huge burden on the precious bandwidth.

[0003] In order to reduce the CSI feedback overhead, CSI prediction technology predicts future channels by utilizing the time correlation between historical CSI and future CSI. This technology effectively reduces the number of CSI feedbacks, thereby reducing feedback overhead. At the same time, artificial intelligence (AI) technology has been used to implement CSI prediction, and its performance is significantly better than traditional autoregressive and parameter model algorithms. However, due to the inherent instability of AI, the performance results of the AI-based CSI prediction model fluctuate greatly, and some results have poor performance, which will lead to instability in downlink transmission. Therefore, it is urgent to develop a mechanism to improve the stability of AI-based CSI prediction. Summary of the invention

[0004] The present invention provides a method for improving the reliability of channel prediction based on a proxy model, which monitors the output results of a CSI prediction model through a proxy model and excludes outputs with poor performance, thereby improving the stability of CSI prediction based on AI.

[0005] The embodiment of the present invention provides a method for improving channel prediction reliability based on a proxy model, comprising the following steps:

[0006] Step 1: compress the channel data collected by the user equipment manufacturer, and transmit the original channel data collected by the user equipment manufacturer and the compressed channel data to the base station manufacturer;

[0007] Step 2: Reconstruct the compressed channel data received by the base station manufacturer, use the reconstructed channel data as input, and use the original channel data as output labels to train the channel prediction model on the base station side;

[0008] Step 3, after the channel prediction model training is completed, the output of the trained channel prediction model is transmitted to the user equipment manufacturer by the base station manufacturer, so that the user equipment manufacturer uses the original channel data as input and the output of the channel prediction model as a label to train the proxy model;

[0009] Step 4: After the proxy model training is completed, the obtained proxy model and channel prediction model are deployed on the user equipment and the base station respectively and run online.

[0010] Optionally, in one embodiment of the present invention, in step 2, the channel prediction model training adopts an end-to-end learning method to train and improve the model parameters so as to minimize the cost function, and the cost function is described as follows:

[0011]

[0012] Where N is the number of output results of the channel prediction model, N s is the number of subbands in the channel, ‖·‖2 is the Euclidean norm, is the original channel matrix at time m+n, It is the prediction model output matrix at time m+n.

[0013] Optionally, in one embodiment of the present invention, in step 3, the agent model training adopts an end-to-end learning method to train and improve the model parameters so as to minimize the cost function, and the cost function is described as follows:

[0014]

[0015] Where N is the number of output results of the channel prediction model, N s is the number of subbands in the channel, ‖·‖2 is the Euclidean norm, is the proxy model output matrix at time m+n, It is the prediction model output matrix at time m+n.

[0016] Optionally, in one embodiment of the present invention, in the online operation phase of step 4, the proxy model takes the channel matrix estimated at the past moment as input, compares the output matrix of the proxy model with the channel matrix estimated at the current moment, calculates the cosine similarity, and adds a bias as an evaluation result; when the evaluation result is higher than a preset threshold, the output result of the prediction model on the current base station side is considered to be reliable, and the user equipment will send a signal to the base station to indicate the use of the prediction model output at the current moment for downlink transmission; otherwise, the output result of the prediction model on the current base station side is considered to be unreliable, and the user equipment will feed back the channel matrix estimated at the current moment to the base station.

[0017] Optionally, in one embodiment of the present invention, the calculation method of the evaluation result of the online operation stage is:

[0018]

[0019] Among them, ∈ is the bias, N s is the number of subbands in the channel, ‖·‖2 is the Euclidean norm, is the proxy model output matrix at time m+n, is the original channel matrix at time m+n.

[0020] Optionally, in one embodiment of the present invention, in step 1, an autoencoder structure is used to compress channel data.

[0021] The agent-based channel prediction reliability improvement method of the embodiment of the present invention improves the reliability of channel prediction and reduces the impact of channel prediction model performance fluctuations on communication stability while slightly increasing the burden on equipment during the implementation of channel prediction.

[0022] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0024] Figure 1 A flowchart of a method for improving channel prediction reliability based on an agent according to an embodiment of the present invention;

[0025] Figure 2 It is an overall architecture diagram of the agent-based channel prediction reliability improvement method according to an embodiment of the present invention;

[0026] Figure 3 This is an overall architecture diagram before the reliability improvement method is deployed in an embodiment of the present invention;

[0027] Figure 4 A schematic diagram of a data collection process during offline training in an embodiment;

[0028] Figure 5 Schematic diagram of the training of the CSI prediction model and the proxy model during the offline training process in the embodiment;

[0029] Figure 6 This is an overall architecture diagram after the reliability improvement method is deployed in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0031] In the channel data prediction task, due to the powerful feature extraction capability of AI technology, the average performance of AI-based channel data prediction is significantly better than that of traditional channel data prediction methods. However, due to the inherent instability of AI, the performance fluctuation of its output is significantly greater than that of traditional methods, and the accuracy of a small number of results is too low, which seriously affects the stability of the downlink and hinders the deployment of AI-based channel data prediction. Therefore, developing a method to improve the output stability of AI-based channel data prediction is the key to promoting its deployment.

[0032] Figure 1 The present invention provides a flowchart of a method for improving channel prediction reliability based on an agent according to an embodiment of the present invention.

[0033] like Figure 1 As shown, the agent-based channel prediction reliability improvement method includes the following steps:

[0034] Step 1: compress the channel data collected by the user equipment manufacturer, and transmit the original channel data collected by the user equipment manufacturer and the compressed channel data to the base station manufacturer.

[0035] The user equipment manufacturer first collects channel data and compresses it using a pre-deployed autoencoder, and then transmits the collected original data and compressed channel data to the base station manufacturer.

[0036] In step 1, the channel data collection, compression and transmission process simulates the acquisition, compression and feedback process of the channel matrix when the real system is running. The data compression process must be consistent with the channel data compression feedback process in reality, in order to facilitate the use of the autoencoder structure to compress the channel data.

[0037] Step 2: Reconstruct the compressed channel data received by the base station manufacturer, use the reconstructed channel data as input, and the original channel data as output label to train the channel prediction model on the base station side to meet the channel prediction operation process on the base station side in actual deployment.

[0038] In step 2, the channel prediction model is trained in an end-to-end learning manner to improve the model parameters so as to minimize the cost function. The cost function is described as follows:

[0039]

[0040] Where N is the number of output results of the channel prediction model, N s is the number of subbands in the channel, ‖·‖2 is the Euclidean norm, is the original channel matrix at time m+n, It is the prediction model output matrix at time m+n.

[0041] Step 3: After the channel prediction model training is completed, the base station manufacturer transmits the output of the trained channel prediction model to the user equipment manufacturer, so that the user equipment manufacturer uses the original channel data as input and the output of the channel prediction model as a label to train the proxy model.

[0042] After the channel prediction model is trained, the base station manufacturer transmits the output of the channel prediction model to the user equipment manufacturer. The user equipment manufacturer uses the original data as input and the output of the channel prediction model as a label to train the proxy model so that the user equipment side agent simulates the output of the base station side channel prediction model.

[0043] In step 3, the agent model training adopts an end-to-end learning method to train and improve the model parameters to minimize the cost function. The cost function is described as follows:

[0044]

[0045] Where N is the number of output results of the channel prediction model, N s is the number of subbands in the channel, ‖·‖2 is the Euclidean norm, is the proxy model output matrix at time m+n, It is the prediction model output matrix at time m+n.

[0046] Step 4: After the proxy model training is completed, the obtained proxy model and channel prediction model are deployed on the user equipment and the base station respectively and run online.

[0047] In the online operation phase in step 4, the proxy model takes the channel matrix estimated at the past moment as input; then, the output matrix of the proxy model is compared with the channel matrix estimated at the current moment, the cosine similarity is calculated, and the bias is added as the evaluation result. When the evaluation result is higher than the preset threshold, the prediction model output result on the current base station side is considered reliable, and the user equipment will send a signal to the base station to indicate the use of the prediction model output at the current moment for downlink transmission; otherwise, the prediction model output result on the current base station side is considered unreliable, and the user equipment will feed back the channel matrix estimated at the current moment to the base station.

[0048] In an embodiment of the present invention, the calculation method of the evaluation result of the online operation stage is:

[0049]

[0050] Among them, ∈ is the bias, N s is the number of subbands in the channel, ‖·‖2 is the Euclidean norm, is the proxy model output matrix at time m+n, is the original channel matrix at time m+n.

[0051] The proxy model-based channel prediction reliability improvement method of the present invention develops a performance monitoring framework and uses a lightweight proxy model deployed on the UE to simulate the output of the CSI prediction model to evaluate the performance of the CSI prediction model in real time. When a low-precision result is detected, the UE will feed back the current CSI to the base station to eliminate the low-precision result and improve the output stability of the AI-based CSI prediction. In addition, a set of implementation processes including offline training and online deployment is proposed, such as Figure 2 shown.

[0052] The invention specifically comprises the following steps:

[0053] (1) In the downlink of a MIMO system, the entire bandwidth is divided into N s subbands, the user end uses N r Root receiving antennas, the base station uses N t = 32 transmitting antennas, the downlink center frequency is 2.1GHz, and the bandwidth is 10MHz. The scenario is an urban microcell scenario with no line of sight. The user is randomly initialized in a 100-meter area around the base station, randomly generates a movement direction, and moves at a speed of 10 meters per second. According to the above settings, a data set is generated, which contains 4,000,000 samples, of which the training set includes 3,200,000 samples, and the validation set and test set each include 400,000 samples.

[0054] (2) Figure 3The AI-based CSI prediction framework is presented in the paper, in which the UE first feeds back the estimated CSI within K moments, and the BS stores the K reconstructed CSIs; then, the BS inputs the K stored reconstructed CSIs into the CSI prediction model to obtain N predicted CSIs. In the next N moments, the BS can use the predicted CSI for downlink transmission, thereby saving the CSI feedback overhead. Based on this framework, the present invention proposes a method for improving CSI prediction reliability based on a proxy model.

[0055] (3) Figure 4 and Figure 5 The offline training phase of the display proxy model and CSI prediction model is completed by the UE and BS manufacturers. First, the UE manufacturer obtains the CSI data set through a series of methods such as simulation and real-world collection, compresses it using a predetermined CSI compression encoder, and transmits the original and compressed data to the BS manufacturer. Next, the BS manufacturer reconstructs the CSI data and divides the data set into groups of K+N data. The first K compressed and reconstructed CSI prediction model inputs are used in each group, and the last N original CSI data are used as output labels.

[0056] After completing the CSI prediction model training, the BS manufacturer transmits the output of the CSI prediction model to the UE manufacturer. The UE manufacturer uses the original CSI data as the input of the lightweight proxy model and the output of the CSI prediction model as the output label. The input and output dimensions are consistent with the CSI prediction model, and the proxy model is trained. The complexity of the proxy model is much smaller than that of the CSI prediction model. The training uses the Adam optimization algorithm to minimize the cost function. The learning rate of the Adam algorithm is first set to 0.0001. After the loss function converges, it is adjusted to 0.00001 and continues training until the loss function converges. 512 samples are used in each iteration.

[0057] (4) Figure 6 The online deployment phase of the proxy model and CSI prediction model is demonstrated. After completing offline training, the manufacturer deploys the proxy model and CSI prediction model to the UE and BS respectively. When running online, the proxy model takes the K CSIs estimated at the past moment as input to obtain the simulated N predicted CSIs; then, the simulated current predicted CSI is compared with the CSI estimated at the current moment, the cosine similarity is calculated, and a bias is added as the evaluation result. When the evaluation result is higher than a preset threshold, the current prediction result is considered reliable, and the user equipment will send a signal to the base station to indicate the use of the current prediction result; otherwise, the current prediction result is considered unreliable, and the user equipment will feedback the CSI estimated at the current moment to the base station.

[0058] In summary, the proxy-based channel prediction reliability improvement method proposed in the present invention can monitor the performance of the CSI prediction model on the BS side by deploying a proxy model without significantly increasing the burden on the UE, so as to improve the reliability of the CSI prediction.

[0059] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction 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, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0060] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "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 clearly and specifically defined.

[0061] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or N executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present invention belong.

Claims

1. A method for improving channel prediction reliability based on a proxy model, characterized in that: The following steps are involved: Step 1: compress the channel data collected by the user equipment manufacturer, and transmit the original channel data collected by the user equipment manufacturer and the compressed channel data to the base station manufacturer; Step 2: Reconstruct the compressed channel data received by the base station manufacturer, use the reconstructed channel data as input, and use the original channel data as output labels to train the channel prediction model on the base station side; Step 3, after the channel prediction model training is completed, the output of the trained channel prediction model is transmitted to the user equipment manufacturer by the base station manufacturer, so that the user equipment manufacturer uses the original channel data as input and the output of the channel prediction model as a label to train the proxy model; Step 4: After the proxy model training is completed, the obtained proxy model and channel prediction model are deployed on the user equipment and the base station respectively and run online.

2. The method according to claim 1, characterized in that In step 2, the channel prediction model is trained in an end-to-end learning manner to improve the model parameters so as to minimize the cost function. The cost function is described as follows: Where N is the number of output results of the channel prediction model, N s is the number of subbands in the channel, ‖·‖2 is the Euclidean norm, is the original channel matrix at time m+n, It is the prediction model output matrix at time m+n.

3. The method according to claim 1, characterized in that In step 3, the agent model training adopts an end-to-end learning method to train and improve the model parameters to minimize the cost function. The cost function is described as follows: Where N is the number of output results of the channel prediction model, N s is the number of subbands in the channel, ‖·‖2 is the Euclidean norm, is the proxy model output matrix at time m+n, It is the prediction model output matrix at time m+n.

4. The method according to claim 1, characterized in that: In the online operation phase of step 4, the proxy model takes the channel matrix estimated at the past moment as input, compares the output matrix of the proxy model with the channel matrix estimated at the current moment, calculates the cosine similarity, and adds a bias as the evaluation result; when the evaluation result is higher than a preset threshold, the output result of the prediction model on the current base station side is considered to be reliable, and the user equipment will send a signal to the base station to indicate the use of the prediction model output at the current moment for downlink transmission; otherwise, the output result of the prediction model on the current base station side is considered to be unreliable, and the user equipment will feed back the channel matrix estimated at the current moment to the base station.

5. The method according to claim 4, characterized in that The calculation method of the evaluation results of the online operation stage is: Among them, ∈ is the bias, N s is the number of subbands in the channel, ‖·‖2 is the Euclidean norm, is the proxy model output matrix at time m+n, is the original channel matrix at time m+n.

6. The method according to claim 1, characterized in that In step 1, an autoencoder structure is used to compress channel data.

Citation Information

Cited By

  • Channel prediction method and device, computer equipment, storage medium and computer program product

    CN120785447A