Large model-based AP access method applied to CF mMIMO system

Abstract

The invention discloses a large-model-based AP access method applied to a CF mMIMO system, and belongs to the technical field of wireless communication. The method specifically comprises the steps that firstly, a base model suitable for an AP selection scene is loaded on a server to obtain a large model # imgabs0 #, and then original data of a wireless network environment is collected and preprocessed; inputting a natural language obtained through preprocessing into a large model # imgabs1 # to generate an AP selection prediction result # imgabs2 #, then applying the AP selection prediction result # imgabs3 # to a real wireless network environment, and designing a feedback prompt pfeed; and according to the feedback prompt pfeed, optimizing an AP selection prediction result by using a large model # imgabs4 #, obtaining a final AP selection prediction result # imgabs5 #, transmitting the final AP selection prediction result # imgabs6 # to a corresponding server, converting the final AP selection prediction result # imgabs5 # into a form y [t + l] which can be identified by a computer, and notifying each access point to access all users meeting the AP access condition through a backhaul link to serve the users. According to the method, the multi-dimensional channel characteristics can be comprehensively considered, so that the predicted data is more accurate.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication technology and artificial intelligence (AI), and in particular relates to an AP access method based on a large model applied to a CF mMIMO system. Background Art

[0002] Cell-free massive multiple-input multiple-output (CF mMIMO) has the characteristics of large-scale connection, low latency, high rate and high reliability, and is a technology with great potential in 5G and 6G networks.

[0003] CF mMIMO is a user-centric distributed MIMO system. In the CF mMIMO system, all access points (APs) are connected through fronthaul links and cooperate with the central processing unit (CPU) to provide services to multiple users simultaneously on the same time-frequency resources through time division duplexing technology. Due to the capacity limitation and hardware damage of the fronthaul load, it is not realistic for the AP to provide services to all users at the same time during the uplink and downlink data transmission phases. Therefore, it is necessary to select a specific subset of users for service to maximize the resource utilization of the communication system.

[0004] The common AP selection algorithms in the industry are mainly divided into: algorithms based on channel feature sorting and algorithms based on artificial intelligence, such as Figure 1 As shown;

[0005] Among them, algorithms based on channel feature sorting include the Euclidean distance between AP and user, large-scale fading coefficient, and signal-to-noise ratio. Based on the channel features between AP and user, the system sorts all surrounding APs in descending order based on a certain channel feature with the user as the center, and selects the top-ranked AP subset for access. However, this algorithm only uses a single (or a small number of) channel features to sort APs, which will result in a lot of channel features being wasted and cannot achieve global optimization of data.

[0006] Therefore, the characteristics of artificial intelligence algorithms in pattern recognition can be utilized to provide further support for AP selection algorithms.

[0007] AI-based algorithms include K-means, convolutional neural networks, and reinforcement learning. Although the addition of AI algorithms can utilize more channel features to improve spectrum efficiency, in practice, these algorithms usually perform poorly in transfer learning, mainly because the geographical locations, AP deployments, channel conditions, and user distributions faced during model training and model application are very different, resulting in weak generalization capabilities of various algorithms. Summary of the invention

[0008] In order to solve the problems of the prior art in AP selection in the CF mMIMO system: low utilization of channel characteristics and weak generalization ability of the algorithm; the present invention proposes an AP access method based on a large model applied to the CF mMIMO system, which can not only improve the utilization of channel characteristics, enhance the generalization ability of the algorithm, optimize the AP selection process, but also obtain the predicted value for AP access in real time.

[0009] The AP access method based on a large model applied to the CF mMIMO system has the following specific steps:

[0010] Step 1: Select a base model suitable for the AP selection scenario based on the model parameters and its comprehensive score on the open source evaluation platform OpenCompass, and load it on the server to obtain a large model.

[0011] Step 2: Collect raw data of the wireless network environment from real scenarios or twin networks;

[0012] The original data includes user location information, user traffic demand information, AP location information, channel estimation information, user access information, user signal-to-noise ratio, access point resource block allocation information, and network congestion information.

[0013] Step 3: Preprocess the original data;

[0014] The processing of raw data includes data cleaning, conversion into natural language form and data segmentation.

[0015] The specific process is:

[0016] First, in the data collection scenario, a data sample is obtained every 24 / T hours, so T samples are generated every day. The sample collection timestamp is t, and the collected original data sample is recorded as x[t].

[0017] Then, for the current timestamp t, the information y[t] of whether the AP is connected is calculated based on the original data sample x[t], and the sample x[t] and the corresponding access information y[t] are converted into natural languages ​​X[t] and Y[t].

[0018] Finally, the natural language collected over a period of time is divided into original training data and optimization data in chronological order; the original data from time tw to time t is taken as the initial input, denoted as X ini [t]; take the optimization data from time t to time t+l-1 as the input for generating feedback and optimizing prediction, denoted as X fr [t].

[0019] Among them, t+l is the expected prediction time.

[0020] Step 4: Process the natural language obtained through preprocessing and input it into the large model Generate AP selection prediction results

[0021] First, the original training data is formally connected to obtain the input prompt p at time t input [t], the expression is:

[0022]

[0023] Indicates connection;

[0024] Then, set the problem prompt p in the uniform format required for the large model ques And the format prompts for large model output form ;

[0025] p ques Indicates a query for future AP access. form Used to normalize the output of large models.

[0026] Finally, given the input prompt p input [t]、Question promptp ques And format tips p form , get the information whether the initial AP is connected at time t Its expression is:

[0027]

[0028] Step 5: Select the prediction result of AP Applied in real wireless network environment, design feedback prompts p feed ;

[0029] Feedback Tips feed Includes overall performance feedback prompts Network congestion feedback prompt Network idle feedback prompt Formatting and completeness feedback tips The calculation formula is as follows:

[0030]

[0031] Among them, the overall performance feedback prompts It refers to the error between the actual wireless network environment and the initial AP selection prediction result.

[0032] Network congestion feedback prompt Refers to the network congestion caused by the AP selection scheme generated using the initial AP selection prediction results.

[0033] Network idle feedback prompt Refers to: the network idleness that will be caused by the AP selection scheme generated using the initial AP selection prediction result;

[0034] Formatting and completeness feedback tips It is a constraint on the output of the large model to ensure that the generated model can output complete prediction information at each timestamp.

[0035] Step 6: Follow the feedback prompts feed , using a large model Optimize the AP selection prediction results to obtain the final AP selection prediction results

[0036] The specific process is:

[0037] First, determine the prediction results The corresponding feedback p feed [t] Whether the preset stop condition is met (stop) feed ,l), if not satisfied, the feedback prompt p at time t feed [t]、Optimization tips refine Input large model Select prediction results for AP Optimize and get the prediction results for:

[0038]

[0039] Stop condition stop(s feed ,l) refers to the self-optimization process when the feedback meets the condition s feed Or stop after the number of iterations meets the length l.

[0040] Then, judge the prediction results Whether the corresponding feedback meets the preset stop condition stop(s feed ,l), if it is not satisfied, then in the next iteration, all the outputs and feedbacks generated by the previous iteration are input into the large model;

[0041] For the t+l0th iteration, the calculation process of the prediction result is expressed as follows:

[0042]

[0043] l0≤l,

[0044]

[0045] If the feedback of the t+l0th iteration meets the stopping condition, the large model The output has been optimized to the desired state, select the current optimization result The data fitting model used when , is used as output, and then output If the number of iterations meets a certain length l, then directly output

[0046] Step 7: Select the final AP prediction result It is transmitted to the corresponding server, converted into a computer-recognizable form y[t+l], and notified through the backhaul link that each access point can access all users who meet the AP access conditions and provide services for them.

[0047] The advantages of the present invention are:

[0048] A large-model-based AP access method applied to CF mMIMO systems generates a model based on the large model that can be continuously optimized and iterated based on its own feedback to generate accurate prediction data. Compared with the traditional algorithm that only uses a single channel feature to rank and select APs, this solution can comprehensively consider multi-dimensional channel features to make the predicted data more accurate. Compared with other AP selection algorithms based on artificial intelligence algorithms, this solution uses the same large model for feedback optimization, which means that the model accurately customizes exclusive prediction dialogues for each CPU, achieving the model's super generalization ability. At the same time, the data segmentation algorithm of this solution provides a guarantee for the real-time nature of the data, and the unique stop mechanism enables the algorithm to obtain accurate prediction information before the use time. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a schematic diagram of a common AP selection algorithm in the prior art;

[0050] Figure 2 It is a flow chart of a large model-based AP access method applied to a CF mMIMO system of the present invention;

[0051] Figure 3 This is a network architecture diagram of cloud-edge collaboration of the present invention;

[0052] Figure 4 It is an example of the description method of converting natural language in the present invention;

[0053] Figure 5 This is an example of the present invention inputting the preprocessed natural language into a large model to generate an AP selection prediction result; DETAILED DESCRIPTION

[0054] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0055] The AP access method based on a large model applied to the CF mMIMO system is as follows: Figure 2 As shown, the specific steps are as follows:

[0056] Step 1: Select a base model suitable for the AP selection scenario based on the model parameters and its comprehensive score on the open source evaluation platform OpenCompass, and load it on the server to obtain a large model.

[0057] Step 2: Collect raw data of the wireless network environment from real scenarios or twin networks;

[0058] The original data includes user location information, user traffic demand information, AP location information, channel estimation information, user access information, user signal-to-noise ratio, access point resource block allocation information, and network congestion information.

[0059] Step 3: Preprocess the original data;

[0060] The processing of raw data includes data cleaning, conversion into natural language form and data segmentation.

[0061] The specific process is:

[0062] First, in the data collection scenario, a data sample is obtained every 24 / T hours, so T samples are generated every day. The sample collection timestamp is t, and the collected original data sample is recorded as x[t].

[0063] Then, for the current timestamp t, the information y[t] of whether the AP is connected is calculated based on the original data sample x[t], and the sample x[t] and the corresponding access information y[t] are converted into natural languages ​​X[t] and Y[t].

[0064] Finally, the natural language collected over a period of time is divided into original training data and optimization data in chronological order; the original data from time tw to time t is taken as the initial input, denoted as X ini [t]; take the optimization data from time t to time t+l-1 as the input for generating feedback and optimizing prediction, denoted as X fr [t]. This segmentation method needs to be X fr [t]Leave data of a certain length to ensure the optimization effect.

[0065] Among them, t+l is the expected prediction time.

[0066] Step 4: Process the natural language obtained through preprocessing and input it into the large model Generate AP selection prediction results

[0067] First, the original training data is formally connected to obtain the input prompt p at time t input [t], the expression is:

[0068]

[0069] Indicates connection;

[0070] Then, set the problem prompt p in the uniform format required for the large model ques And the format prompts for large model output form ;

[0071] p ques Indicates a query for future AP access. form Used to normalize the output of large models.

[0072] Finally, given the input prompt p input [t]、Question promptp ques And format tips p form , get the information whether the initial AP is connected at time t Its expression is:

[0073]

[0074] Step 5: Select the prediction result of AP Applied in real wireless network environment, design feedback prompts p feed ;

[0075] Feedback Tips feed Includes overall performance feedback prompts Network congestion feedback prompt Network idle feedback prompt Formatting and completeness feedback tips The calculation formula is as follows:

[0076]

[0077] Among them, the overall performance feedback prompts It refers to the error (such as mean square error, etc.) between the actual wireless network environment and the initial AP selection prediction result.

[0078] Network congestion feedback prompt Refers to the network congestion caused by the AP selection scheme generated using the initial AP selection prediction results.

[0079] Network idle feedback prompt Refers to: the network idle state that will be caused by the AP selection scheme generated by using the initial AP selection prediction result; further support is provided for the scalability of large models. The network idle state under the error convergence condition can instruct the AP to adjust the on / off state to save energy.

[0080] Formatting and completeness feedback tips It is a constraint on the output of the large model to ensure that the generated model can output complete prediction information at each timestamp.

[0081] Step 6: Follow the feedback prompts feed , using a large model Optimize the AP selection prediction results to obtain the final AP selection prediction results

[0082] The specific process is:

[0083] First, determine the prediction results The corresponding feedback p feed [t] Whether the preset stop condition is met (stop) feed ,l), if not satisfied, the feedback prompt p at time t feed [t]、Optimization tips refine Input large model Select prediction results for AP Optimize and get the prediction results for:

[0084]

[0085] Stop condition stop(s feed ,l) refers to the self-optimization process when the feedback meets the condition s feed Or stop after the number of iterations meets the length l.

[0086] Then, judge the prediction results Whether the corresponding feedback meets the preset stop condition stop(s feed ,l), if it is not satisfied, then in the next iteration, all the outputs and feedbacks generated by the previous iteration are input into the large model;

[0087] In order to achieve self-optimization of the large model and enable it to learn from past feedback, it is necessary to input all outputs and feedback generated by previous iterations into the large model during the iteration. For the t+l0th iteration, the calculation process of the prediction result is expressed as follows:

[0088]

[0089] l0≤l,

[0090]

[0091] If the feedback of the t+l0th iteration meets the stopping condition, the large model The output has been optimized to the desired state, select the current optimization result The data fitting model used when , is used as output, and then output If the number of iterations meets a certain length l, then directly output

[0092] Step 7: Select the final AP prediction result It is transmitted to the corresponding server, converted into a computer-recognizable form y[t+l], and notified through the backhaul link that each access point can access all users who meet the AP access conditions and provide services for them.

[0093] The AP selection technology proposed in the present invention can be deployed on various network architectures such as cellular networks, wireless local area networks, Mesh networks, and cloud-edge collaborative networks. Considering the specifications of the model and the future development trend of wireless networks, this embodiment takes the cloud-edge network architecture as an example to illustrate the proposed AP access technology based on the large model. The cloud-edge collaborative model combines the powerful computing power of cloud computing and the low latency characteristics of edge computing, laying the foundation for the deployment of large models.

[0094] The network architecture of cloud-edge collaboration is as follows Figure 3 As shown: It is divided into four layers: cloud processing layer, edge processing layer, access point layer and terminal layer.

[0095] The cloud processing layer deploys high-performance cloud servers, which can quickly expand or reduce resources such as CPU, memory, etc. according to demand without manually replacing hardware. Cloud servers can be accessed through the Internet from anywhere in the world, providing great flexibility and convenience. In a public cloud environment, multiple edge servers share the same physical hardware resources but are isolated from each other to ensure data security.

[0096] The edge processing layer is responsible for filtering and preliminarily processing various types of data uploaded by the access point in the data uplink to reduce the amount of data transmitted to the cloud and improve network efficiency; in the data downlink, it is responsible for receiving instructions from the cloud server and then regulating the actions of each access point. The edge processing layer can quickly process and analyze real-time data to support applications that are sensitive to latency.

[0097] The APs in the access point layer are small in size and numerous in number. They are connected to the CPU in the edge processor layer via a backhaul link and are coordinated and controlled by the CPU to achieve collaborative processing across geographical distribution and provide services to users (User Equipment, UE) at all terminal layers. This enables the architecture to support large-scale device connections and is suitable for scenarios such as the Internet of Things.

[0098] The terminal layer includes all user UEs, and each UE is served by multiple APs. In this cloud-edge collaborative network architecture, services are optimized with UE as the center, that is, the allocation and processing of network resources are aimed at providing uniform and high-quality services to each user.

[0099] The specific process steps are as follows:

[0100] S1: Building a large model

[0101] The large model applicable to the AP selection scenario mentioned in this embodiment is deployed on the server, and its base model is obtained by the following two steps.

[0102] S1.1: Select a base model. There are many open source base models, including but not limited to ChatGLM3-6B from Tsinghua University, Qwen-14B from Alibaba, InternLM-20B from SenseTime, etc. When selecting a base model, you should consider the number of model parameters and its comprehensive score on the open source evaluation platform OpenCompass.

[0103] S2.2: Install the environment in the cloud and load the base model, denoted as

[0104] S2: Get the original data

[0105] The original data includes access layer data and access point layer data, and the terminal layer data includes user location information, user traffic demand information, etc. The access point layer data includes AP location information, channel estimation information, user access information, user signal-to-noise ratio, access point resource block allocation information, network congestion information, etc. All the above information will be collectively referred to as wireless network environment information in the following text.

[0106] S3: Processing Raw Data

[0107] Data cleaning and conversion to natural language form are completed at the edge processing layer, and data segmentation is completed at the cloud processing layer.

[0108] S3.1: Cleaning data. Let the sample collection timestamp be t, and the collected wireless network environment data sample be x[t]. For the convenience of description, x[t] in the following text takes the SNR of user access as an example.

[0109] S3.2: Calculate the information y[t] for determining AP access at this time according to x[t]. The present invention is explained by taking the minimum SNR value of each access point access user as an example, that is, at this time y[t] represents the SNR threshold.

[0110] S3.3: Convert x[t], y[t] into natural language X[t], Y[t] and upload it to the cloud processing layer. The conversion into natural language form is to use natural language to describe the key information in the data. It is particularly important to note that this description is generally in a fixed format to facilitate the understanding of the large model.

[0111] like Figure 4 As shown, the sample x[t] is {“EDGE_ID”:“EDGE-BJ-001”, “TIME”:“2024-01-0100:00”, “USER_ACCESS”:[[3,7],[3.5,5.2],…]}

[0112] Converted into natural language X[t]: At 00:00 on January 1, 2024, the signal-to-noise ratios of users accessing different access points from the edge processor EDGE-BJ-001 are 3, 7; 3.5, 5.2,... respectively.

[0113] The information y[t] for determining AP access is: {“EDGE_ID”:“EDGE-BJ-001”, “TIME”:“2024-01-0100:00”, “SNR_THRESHOLD”:[3,3.5,…]}

[0114] Converted into natural language Y[t]: At 00:00 on January 1, 2024, the signal-to-noise ratio thresholds of users accessing different access points from the edge processor EDGE-BJ-001 are 3dB, 3.5dB,... respectively.

[0115] In the figure, the collected data sample x[t] is a computer-readable dictionary that records the edge processor number (EDGE_ID), data collection time (TIME), and user access SNR (USER_ACCESS), where the user access SNR is given in the form of a two-dimensional array. The first dimension of the array represents different APs, and the second dimension of the array represents the SNR value of each AP access user. The natural language form X[t] of x[t] is the key-value pair content in the above dictionary described in Chinese.

[0116] Similarly, the SNR threshold y[t] is a computer-readable dictionary that records the edge processor number (EDGE_ID), data collection time (TIME), and predicted SNR threshold (SNR_THRESHOLD), where the predicted SNR threshold is given in a list, and each element in the list represents the predicted threshold of an AP. The natural language form Y[t] of y[t] is the key-value pair content in the above dictionary described in Chinese.

[0117] S3.4: The cloud processing layer collects data from the edge processing layer and segments it.

[0118] S4: Generate initial prediction results for AP selection

[0119] The present invention uses the initial prediction results generated by the large model in the cloud processing layer, and obtains accurate prediction values ​​by training the large model through feedback and optimization iteration. AP access prediction is expressed in the form of questions and answers, such as Figure 5 As shown, the problem prompt p is given ques And the format prompts for large model output form .

[0120] S5: Generate Feedback

[0121] Since large models cannot obtain accurate predictions immediately at the start-up stage, it is necessary to design feedback prompts corresponding to the current prediction results provided by the edge processing layer and the cloud processing layer to generate more accurate outputs.

[0122] Design feedback tips feed Including but not limited to overall performance feedback tips provided by the edge processing layer Network congestion feedback and network idle feedback and format and completeness feedback provided by the cloud processing layer

[0123] Among them, the overall performance: the mean square error between the predicted signal-to-noise ratio threshold and the true signal-to-noise ratio threshold is 6.34, indicating that the overall performance error should be reduced;

[0124] Network congestion: After the user accesses based on the predicted SNR threshold, network congestion of 0.2, 0, ... appears at each access point. These congestion conditions should be eliminated or reduced.

[0125] Network idleness: After the user is connected based on the predicted signal-to-noise ratio threshold, the network idleness of each access point is 0%, 15%, etc., and the network resource utilization should be improved.

[0126] Format and completeness: The predicted SNR thresholds should be in the same format as the actual SNR thresholds, and a complete forecast should be provided for each hour.

[0127] S6: Optimize forecasts

[0128] The present invention uses the feedback in S5 at the cloud processing layer Optimize its latest output. Optimization hints are used to help large models determine the optimization direction.

[0129] Through the above feedback information, the signal-to-noise ratio prediction threshold of all access points of the edge processor EDGE-BJ-001 is predicted again every hour between 00:00 and 23:00 on January 2, 2024. A more accurate method should be used in the prediction to reduce the overall performance error, and the predicted signal-to-noise ratio threshold should be closer to the actual signal-to-noise ratio. In addition, the predicted signal-to-noise ratio threshold should be consistent with the actual signal-to-noise ratio threshold format and provide a complete prediction for each hour.

[0130] S7: AP access

[0131] S7.1: Information generated by the large model to determine AP access It will be transferred from the cloud processing layer to the edge processing layer and then to the corresponding CPU.

[0132] S7.2: The edge processing layer CPU receives prediction data in natural language It is then converted into a computer-recognizable form y[t+l], and each access point is notified through the backhaul link to access all users who meet the AP access conditions and provide services for them.

[0133] S7.3: Each access point uploads the wireless network environment data to the CPU, and the communication system repeatedly executes S2→S7 to achieve the operation of the system proposed by the present invention.

[0134] The AP selection technology proposed in the present invention not only uses multi-dimensional channel features, but also continuously learns global information in repeated inquiries, so that the algorithm can obtain globally optimal prediction results and has strong generalization capabilities; the details are as follows:

[0135] 1. Channel characteristic utilization:

[0136] The AP access algorithm based on the large model can generate prediction information by using a variety of channel features as model inputs. These features include but are not limited to signal strength, signal-to-noise ratio, channel state information, user location information, etc. In the feedback phase, these channel features can also be used as a metric to optimize the output of the model. In theory, all channel feature data uploaded from the access point layer and the terminal layer can be used as model input and feedback, so that the large model can make better use of the channel feature data.

[0137] 2. Model generalization ability:

[0138] The core of this solution is to use a large model deployed in the cloud processing layer to perform unified prediction tasks. This large model not only has powerful data processing capabilities, but also can provide accurate prediction results through its own continuous feedback and optimization. In actual operation, the model receives data input from multiple CPUs in the edge processing layer, each of which is connected to a different set of APs and has unique network environment and user behavior characteristics.

[0139] To adapt the model to this diversity, each CPU will communicate with the big model at the cloud processing layer based on the specific needs and historical performance of the APs it manages. This dialogue includes the input data that is continuously received and the feedback based on the predicted results. In this way, the big model can continuously obtain the latest information from the edge processing layer and optimize its own prediction algorithm.

[0140] This dynamic interaction and iteration process enables the large model to generate customized predictions that are adapted to different CPUs and the APs they manage. Each prediction is tailored to the characteristics of a specific AP cluster and the current network conditions, ensuring the accuracy and practicality of the prediction.

[0141] The design of this process significantly improves the generalization ability of the model. Generalization ability refers to the ability of the model to maintain high prediction accuracy when facing unseen data or environments. Through continuous feedback and optimization, the large model can capture a wider range of data distribution and potential pattern changes, thereby providing reliable predictions in different network environments.

[0142] 3. System delay:

[0143] This solution performs special segmentation on the data so that the model can make predictions in advance, and the unique stopping mechanism enables the overall algorithm to obtain accurate prediction information before the use time. Therefore, compared with AP selection solutions using other types of algorithms, this solution can significantly reduce system latency. In network communications, reduced latency means faster response time and higher network efficiency, which is especially important for application scenarios that require real-time or near real-time services. Through this efficient prediction method, the system can adapt to changes in network conditions more quickly and provide better access point selection for user devices, thereby improving overall network performance and user experience.

Claims

1. A large model-based AP access method applied to a CF mMIMO system, characterized in that: The specific steps are as follows: Step 1: Select a base model suitable for the AP selection scenario based on the model parameters and its comprehensive score on the open source evaluation platform OpenCompass, and load it on the server to obtain a large model. Step 2: Collect raw data of the wireless network environment from real scenarios or twin networks; Step 3: Preprocess the original data; The processing of raw data includes data cleaning, conversion into natural language form and data segmentation; Step 4: Process the natural language obtained through preprocessing and input it into the large model Generate AP selection prediction results First, the original training data is formally connected to obtain the input prompt p at time t input [t], the expression is: Indicates connection; Then, set the problem prompt p in the uniform format required for the large model ques And the format prompts for large model output form ; p ques Indicates a query for future AP access; p form Used to standardize the output of large models; Finally, given the input prompt p input [t]、Question promptp ques And format tips p form , get the information whether the initial AP is connected at time t Its expression is: Step 5: Select the prediction result of AP Applied in real wireless network environment, design feedback prompts p feed ; Feedback Tips feed Includes overall performance feedback prompts Network congestion feedback prompt Network idle feedback prompt Formatting and completeness feedback tips The calculation formula is as follows: Step 6: Follow the feedback prompts feed , using a large model Optimize the AP selection prediction results to obtain the final AP selection prediction results The specific process is: First, determine the prediction results The corresponding feedback p feed [t] Whether the preset stop condition stop(sfeed,l) is met. If not, the feedback prompt p at time t is displayed. feed [t]、Optimization tips refine Input large model Select prediction results for AP Optimize and get the prediction results for: Stop condition stop(s feed ,l) refers to the self-optimization process when the feedback meets the condition s feed Or stop after the number of iterations meets the length l; Then, judge the prediction results Whether the corresponding feedback meets the preset stop condition stop(s feed ,l), if it is not satisfied, then in the next iteration, all the outputs and feedbacks generated by the previous iteration are input into the large model; For the t+l0th iteration, the calculation process of the prediction result is expressed as follows: If the feedback of the t+l0th iteration meets the stopping condition, the large model The output has been optimized to the desired state, select the current optimization result The data fitting model used when , is used as output, and then output If the number of iterations meets a certain length l, then directly output Step 7: Select the final AP prediction result It is transmitted to the corresponding server, converted into a computer-recognizable form y[t+l], and notified through the backhaul link that each access point can access all users who meet the AP access conditions and provide services for them.

2. The large model-based AP access method for a CF mMIMO system according to claim 1, characterized in that: In the step 2, the original data includes user location information, user traffic demand information, AP location information, channel estimation information, user access information, user signal-to-noise ratio, access point resource block allocation information and network congestion information.

3. The large model-based AP access method applied to a CF mMIMO system as claimed in claim 1, characterized in that: The specific process of step three is: First, in the data collection scenario, a data sample is obtained every 24 / T hours, so T samples are generated every day. The sample collection timestamp is t, and the collected original data sample is recorded as x[t]; Then, for the current timestamp t, the information y[t] of whether the AP is connected is calculated based on the original data sample x[t], and the sample x[t] and the corresponding access information y[t] are converted into natural languages ​​X[t] and Y[t]; Finally, the natural language collected over a period of time is divided into original training data and optimization data in chronological order; the original data from time tw to time t is taken as the initial input, denoted as X ini [t]; take the optimization data from time t to time t+l-1 as the input for generating feedback and optimizing prediction, denoted as X fr [t]; Among them, t+l is the expected prediction time.

4. The large model-based AP access method for a CF mMIMO system according to claim 1, characterized in that: In step 5, the overall performance feedback prompts It refers to the error between the actual wireless network environment and the initial AP selection prediction result; Network congestion feedback prompt Refers to: Network congestion caused by the AP selection scheme generated using the initial AP selection prediction results; Network idle feedback prompt Refers to: the network idleness that will be caused by the AP selection scheme generated using the initial AP selection prediction result; Formatting and completeness feedback tips It is a constraint on the output of the large model to ensure that the generated model can output complete prediction information at each timestamp.

Images