A neural network training and inference method in a cell-free massive MIMO system based on federated learning

Abstract

The application discloses a neural network training and inference method in a cell-free massive MIMO system based on federated learning, and belongs to the cell-free massive MIMO field; specifically, first, a communication scenario composed of a cloud CPU and multiple edge sub-networks is built; then, the users of a single sub-network respectively complete the data feature extraction of each other and transmit the data to the edge CPU for integration. Next, each edge CPU and the cloud CPU cooperatively complete federated learning FL training: the edge CPU takes the integrated data as the input of the latter half model, performs training and updating, and regularly returns the fault gradient to the user to help the user update the feature extractor. The cloud CPU receives the latter half model of each sub-network to perform weighted averaging to obtain a global latter half model, and the global latter half model is updated and distributed. The trained latter half model is deployed on each AP, waits for a new real-time request from the user, and outputs the final inference result to complete the service. The application provides an AI-enabled service with low latency and high efficiency for the CF mMIMO system.

Description

Technical Field

[0001] This invention relates to the field of cellular-free massive MIMO, and in particular to a method for training and inference neural networks in cellular-free massive MIMO systems based on federated learning. Background Technology

[0002] Future communication networks will leverage cutting-edge technologies to achieve unprecedented breakthroughs in data transmission rates, large-scale network coverage, high reliability, and low latency, thereby providing more efficient and higher-quality wireless communication services than ever before. This development trend places stringent requirements on networks, including high spectral efficiency, low latency, high reliability, and high energy efficiency.

[0003] Cell-free massive multiple-input multiple-output (CF mMIMO) systems have become a key technology for overcoming the aforementioned bottlenecks. In CF mMIMO systems, a large number of distributed access points (APs) are connected to a central processing unit (CPU), providing services to a small number of users using the same time-frequency resources. Compared to mMIMO systems in the same frequency band and traditional small cell systems, CF mMIMO systems demonstrate enormous potential for improving network performance in multiple dimensions.

[0004] However, the design philosophy of future 6G networks is undergoing a fundamental shift, with its core no longer limited to providing powerful communication performance. Under the trend of computing-communication-network convergence, CF mMIMO systems should also integrate with various cutting-edge technologies such as AI and edge computing, possessing intelligent coordination and collaboration capabilities to achieve intrinsic intelligence, thereby improving network performance or providing personalized services to users.

[0005] In existing research, Multi-access Edge Computing (MEC) has been introduced to the network edge of CFmMIMO systems to provide additional computing power, thereby improving the performance of latency-sensitive applications. This MEC-based CF mMIMO network has significant potential for intrinsic intelligence. Furthermore, combining it with various AI technologies can efficiently improve network coverage quality or provide different AI-enabled services, thus enhancing network functionality.

[0006] For example, centralized deep reinforcement learning (DRL) or distributed DRL can be used to assist CF mMIMO systems in achieving efficient AP selection and improving system capacity. Federated learning (FL) can be used to achieve decentralized precoding and reduce communication overhead. The future development of networks is essentially a systemic evolution of the deep integration of communication, computing, and artificial intelligence. The ultimate goal of new CF mMIMO networks is to build a network architecture that integrates communication and perception, intrinsically coordinates computing power, and is ubiquitously distributed with intelligence.

[0007] Most existing CF mMIMO network designs rely on traditional centralized AI implementations. However, this centralized approach to data collection for training and deploying AI models conflicts with the architecture of multi-AP collaborative communication. Especially at the network edge, challenges such as data privacy, resource constraints, and the need for low-latency services hinder the realization of AI-driven resource allocation, AP scheduling, and other intrinsic intelligence, as well as user-facing AI-enabled services. CF mMIMO faces barriers in achieving integrated computing and improving the diversity of network functions.

[0008] However, FL's inherent distributed training architecture can help CF mMIMO solve the aforementioned problems. By empowering AP's storage and computing capabilities and combining it with FL's distributed training philosophy, cross-network AI model collaborative training and deployment can be achieved, thereby enabling intrinsic or AI-enabled services. Furthermore, the CF mMIMO communication architecture provides efficient network communication performance for AI model training. Combining these two technologies is expected to promote network intelligence in CF mMIMO systems and accelerate the integration of computing and communication. Summary of the Invention

[0009] In resource-constrained network edge scenarios, centralized AI struggles to assist CF mMIMO systems in achieving intrinsic intelligence or AI-enabled services. This invention proposes a neural network training and inference method based on federated learning in non-cellular massive MIMO systems. It leverages the architecture and performance of CF mMIMO to support FL, thereby promoting the intelligence of the CF mMIMO system. Based on a distributed multi-AP collaborative communication architecture, it enables the training of multiple CF mMIMO collaborative AI models and distributes and deploys global models globally to APs at the edge, providing low-latency and high-efficiency AI-enabled services.

[0010] The specific steps are as follows:

[0011] Step 1: Build a communication scenario consisting of a cloud CPU and multiple edge CF mMIMO subnetworks;

[0012] Edge CF mMIMO subnetwork has In a single edge CF mMIMO subnetwork, there are multiple users holding data, multiple APs, and one edge CPU. A single AP can simultaneously receive signals from all users, and each user is connected to all APs simultaneously.

[0013] Step 2: System initialization, multiple edge CF mMIMO subnetworks establish communication connections with the cloud CPU;

[0014] No. Users in a CF mMIMO subnetwork Holding the original training data and feature extractor , No. Each edge CPU holds the second half of the model. The cloud CPU holds the global second half model. .

[0015] The AI ​​model is a complete neural network, consisting of a feature extractor and a second half of the model; each user trains their own corresponding feature extractor, while the edge CPU trains the second half of the model.

[0016] Specifically, the user calculates the features of the original data locally using a feature extractor, and transmits the processed feature data to the edge CPU through multiple corresponding APs. The edge CPU then uses this data as input to the second half of the model for training and updating, and periodically returns the torsional gradient to the user side to help the user side update the feature extractor.

[0017] Cloud CPU reception The second half of each edge CF mMIMO subnetwork is weighted and averaged to obtain the final global second half model, which is then deployed for updates.

[0018] Step 3, regarding the first Each user completes their own data feature extraction in a CF mMIMO sub-network, which is then transmitted to the edge CPU for integration. .

[0019] The specific steps include:

[0020] Step 3.1, regarding the first Each CF mMIMO subnetwork allows individual users to obtain their own raw data features using their respective feature extractors.

[0021] user Using a feature extractor Obtain the characteristics of the raw data for:

[0022]

[0023] in for The forward propagation function, Original training data eigenvectors, For the first The user set of a CFmMIMO subnetwork;

[0024] Step 3.2: Each user transmits its original data characteristics to each AP in the sub-network. Each AP receives the transmitted data from all users and transmits the signal to the edge CPU through the fronthaul link.

[0025] Step 3.3: The edge CPU receives and reconstructs the feature data of each user, and integrates them to obtain:

[0026] .

[0027] Step 4: Each edge CPU uses the integrated data to collaborate with the cloud CPU to complete federated learning (FL) training and obtain updated second-half model parameters.

[0028] The specific steps are as follows:

[0029] Step 4.1, regarding the first In the global training round, each edge CPU samples from the integrated dataset to obtain the training set;

[0030] No. One edge CPU, from the integrated dataset The data sampled from the middle to participate in this round of training, namely:

[0031] , ,

[0032] The initial value is 0.

[0033] Step 4.2: The edge CPUs use the training set to train the parameters of their respective deployed second-half models, obtaining the updated parameters. The parameters of the second half of the model are transmitted to the cloud CPU respectively;

[0034] No. The parameters of the second half model updated by the edge CPU are as follows:

[0035] For learning rate, , For the first In the first round of global training The parameters of the second half of the edge CPU model.

[0036] Step 4.3: The cloud CPU receives the second-half model parameters transmitted from each edge CPU and performs a weighted average to obtain the first... Global second-half model parameters updated in each round:

[0037]

[0038]

[0039]

[0040] For the first Global second-half model parameters in cloud CPU during round-wide training;

[0041] Step 4.4, the cloud CPU will... The global second half model updated in each round is delegated to the edge CPUs, and the edge CPUs update their respective second half models. Rear wheel model parameters:

[0042]

[0043] Step 4.5: Determine whether the number of rounds of global training has reached the set threshold. If yes, end; otherwise, resample the training set and return to step 4.2.

[0044] Step 5: Each edge CPU calculates the gradient of the fault based on the currently updated second-half model parameters and its own integrated data, and transmits it to each user to update the feature extractor for each user.

[0045] No. The corresponding user in each edge CPU Data Calculate the gradient of the fault :

[0046]

[0047] .

[0048] user Feature extractor The updated formula is:

[0049]

[0050] The equals sign represents assignment.

[0051] Step 6: Globally aggregate the updated feature extractors of all users in the same CF mMIMO subnetwork to obtain the updated feature extractor parameters;

[0052] The global aggregation formula is:

[0053]

[0054] Step 7: Return to Step 3, use the feature extractor with updated parameters to extract and integrate the data features again, continue federated learning (FL) training, update the parameters of the second half of the edge CPU model until the required number of FL training iterations or the parameters of the second half of the model are qualified.

[0055] Step 8: Deploy the updated model to each AP, wait for users to issue new real-time requests, and output the final inference results to complete the service.

[0056] The specific steps are as follows:

[0057] Step 8.1, in the... In a CF mMIMO subnetwork, AP Deploy and maintain the model ;

[0058] Step 8.2, User Initiate an inference request and submit feature data. The data is transmitted to the corresponding edge CPU, where the CPU reconstructs the inference feature data.

[0059] Step 8.3: The edge CPU schedules the AP for collaborative inference for the current user based on information such as the user's simultaneous request for inference and channel parameters. The following constraints must be met:

[0060]

[0061]

[0062] in This refers to the set of users who initiate inference requests in the same edge CF mMIMO network at the current moment.

[0063] For the first A set of APs in a CF mMIMO network. Indicates user With AP Channel gain between The threshold value for filtering APs using the above formula is a scalar. Use the first A set of APs for each user;

[0064] Step 8.4, the edge CPU will connect the user Data that needs to be reasoned Transmitted to the corresponding set The selected AP uses the locally deployed model for inference:

[0065]

[0066] in For the forward propagation function of the second half of the model, For the model The calculated logits values, after normalization, can be used to calculate the labels. .

[0067] Step 8.5: Multiple APs synchronously transmit the final results, perform joint inference through over-the-air computation, and provide the user with the final inference result, thus completing the service.

[0068]

[0069]

[0070] ).

[0071] For downlink AP With users The transmit power coefficient between The normalized downlink signal-to-noise ratio, For users in the downlink Received noise.

[0072] Furthermore, the model deployed by AP can be fine-tuned based on the current input data:

[0073] .

[0074] The advantages of this invention are:

[0075] 1. This invention proposes a neural network training and inference method based on federated learning in a non-cellular large-scale MIMO system. Users in each edge CF mMIMO sub-network convert the raw data into features and transmit them to the corresponding edge CPU. Each edge CPU integrates the data and performs FL training collaboratively. This effectively avoids the limitation that multiple APs in the sub-network cannot independently transmit or reconstruct single-user data, thereby realizing federated learning.

[0076] 2. This invention proposes a neural network training and inference method based on federated learning in a non-cellular large-scale MIMO system. It coordinates user participation in the edge CF mMIMO sub-network to achieve efficient federated learning training with lower communication overhead and higher accuracy of the trained model compared to traditional federated learning methods.

[0077] 3. This invention proposes a neural network training and inference method based on federated learning in a non-cellular massive MIMO system. It empowers access points (APs) as intelligent edge nodes, fully leveraging the collaborative characteristics of CF mMIMO communication. Through multi-AP joint inference, AI intelligent services are achieved. This inference mode not only reduces service latency but also supports personalized AI services, propelling CFmMIMO subnetworks to break through the conceptual bottlenecks of traditional communication systems and achieve intelligent operation. Attached Figure Description

[0078] Figure 1 This is a flowchart illustrating a neural network training and inference method in a cellular-free massive MIMO system based on federated learning, according to the present invention.

[0079] Figure 2 A scenario diagram of a cellular-free massive MIMO system enabled by federated learning, built for this invention.

[0080] Figure 3 This is a schematic diagram illustrating the user transmission features of the present invention;

[0081] Figure 4 This is a schematic diagram of the training of the latter half of the AI ​​model of this invention;

[0082] Figure 5 This is a schematic diagram of multi-AP joint inference of the present invention;

[0083] Figure 6 This is a flowchart illustrating how users of this invention each complete their own data feature extraction and transmission to the edge CPU for integration.

[0084] Figure 7 This is a flowchart of the method for training a collaborative AI model using multiple CF mMIMO subnetworks according to the present invention;

[0085] Figure 8 This is a flowchart of the multi-AP joint inference method of the present invention. Detailed Implementation

[0086] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are also within the scope of protection of this application.

[0087] This invention provides a method for neural network training and inference in a cellular-free massive MIMO system based on federated learning, comprising two methods: collaborative AI model training of multiple CF mMIMO subnetworks and joint inference of multiple APs; such as Figure 1As shown, the specific steps are as follows:

[0088] Step 1: Build a communication scenario consisting of a cloud CPU and multiple edge CF mMIMO subnetworks;

[0089] Edge CF mMIMO subnetwork has In a single edge CF mMIMO subnetwork, there are multiple users holding data, multiple APs, and one edge CPU. A single AP can simultaneously receive signals from all users, and each user is connected to all APs simultaneously.

[0090] Step 2: System initialization, multiple edge CF mMIMO subnetworks establish communication connections with the cloud CPU;

[0091] No. Users in a CF mMIMO subnetwork Holding the original training data and feature extractor , No. Each edge CPU holds the second half of the model. The cloud CPU holds the global second half model. .

[0092] During the training phase, this invention, based on the concept of model segmentation, divides the AI ​​model (neural network) trained by traditional machine learning into two parts: a first half and a second half. The first half of the model can perform feature extraction from the data, while the second half performs downstream tasks. In this framework, each user trains their corresponding first half of the model, i.e., the feature extractor, while the edge CPU is responsible for training the second half of the model.

[0093] Specifically, the user calculates the features of the original data locally using a feature extractor, and transmits the processed feature data to the edge CPU through multiple corresponding APs. The edge CPU then uses this data as input to the second half of the model for training and updating, and periodically returns the torsional gradient to the user side to help the user side update the feature extractor.

[0094] Cloud CPU reception The second half of each edge CF mMIMO subnetwork is weighted and averaged to obtain the final global second half model, which is then deployed for updates.

[0095] like Figure 2The diagram illustrates a scenario of a cellular-free, large-scale MIMO edge intelligence system enabled by federated learning. Overall, the system consists of multiple edge-based CF mMIMO sub-networks and a cloud CPU. Users in each sub-network transmit processed feature data to the edge CPU as a training dataset. The edge CPU and cloud CPU then collaborate on FL training based on their respective data, achieving efficient AI model training and providing the foundation for inference services, i.e., AI-enabled services. Within the overall system, AP nodes are empowered with computing and storage capabilities and deploy the FL-trained models. Under the unique multi-AP collaborative communication framework of CF mMIMO, this edge intelligence system can achieve multi-AP joint inference, providing users with inference services—AI services based on changes in the deployed model.

[0096] like Figure 3 The diagram illustrates user-transmitted features. In a single-edge CF mMIMO system network, users need to transmit processed feature data during the initialization and intermediate update phases. Specifically, users extract features from their local data using a locally maintained feature extractor and transmit them to the corresponding edge CPU based on the communication architecture of the CF mMIMO sub-network. The edge CPU then uses communication technology to recover the feature data transmitted by each user and finally integrates all the data as the dataset for subsequent FL training.

[0097] like Figure 4 The diagram illustrates the training process for the latter half of the AI ​​model. Multiple edge CF mMIMO subnetworks train their respective models on their CPUs. The cloud CPU, acting as the central server, collects the model parameters from each edge CPU in the current round, performs a weighted average, updates the global model, and distributes it to the edge CPUs. This process iterates until the model converges, achieving collaborative FL training.

[0098] like Figure 5 The diagram illustrates multi-AP joint inference. The system deploys the trained model to various APs in the edge CF mMIMO network. In this system, all APs possess storage and computing capabilities, enabling them to maintain the model and perform inference based on communication. When a user requests a service, the corresponding set of APs receives inference data transmitted from the edge CPU, performs inference based on the model, and then transmits the inference result to the corresponding user through joint communication with other APs. This result integrates knowledge from multiple AP models, resulting in higher accuracy. Based on this, the edge CF mMIMO network can leverage this system to achieve efficient AI-enabled services and expand network functionality.

[0099] Step 3, regarding the first Each user completes their own data feature extraction in a CF mMIMO sub-network, which is then transmitted to the edge CPU for integration via the CF mMIMO communication framework. .

[0100] like Figure 6 As shown, the specific steps include:

[0101] Step 3.1, regarding the first Each CF mMIMO subnetwork allows individual users to obtain their own raw data features using their respective pre-trained feature extractors.

[0102] user Using a feature extractor Obtain the characteristics of the raw data for:

[0103]

[0104] in for The forward propagation function, Original training data eigenvectors, For the first The user set of a CFmMIMO subnetwork;

[0105] Step 3.2: Each user transmits its original data characteristics to each AP in the sub-network. Each AP receives the transmitted data from all users and transmits the signal to the edge CPU through the fronthaul link to obtain the signal. .

[0106] From a physical layer perspective, the signals received by the CPU can be represented as:

[0107]

[0108]

[0109] in, It is the first A set of APs in a CF mMIMO subnetwork. In order to cooperate with AP The corresponding normalized uplink signal-to-noise ratio (SNR) This represents the power control coefficient. For users The standardized signal to be sent. For the corresponding AP Uplink noise. Indicates user With AP Channel gain between This is the estimated channel gain. These are large-scale attenuation modeling parameters, including geometric path loss, shading effects, etc. These are small-scale fading parameters;

[0110] Step 3.3: From the physical layer perspective, edge CPUs... The signal was restored in the middle. From the application layer perspective, the edge CPU receives and reconstructs the feature data of each user, integrates it, and uses it as input data for subsequent FL training.

[0111] .

[0112] Step 4: Each edge CPU uses the integrated data to collaborate with the cloud CPU to complete federated learning (FL) training and obtain updated second-half model parameters.

[0113] like Figure 7 As shown, the specific steps are as follows:

[0114] Step 4.1, regarding the first In the global training round, each edge CPU samples from the integrated dataset to obtain the training set;

[0115] No. One edge CPU, from the integrated dataset The data sampled from the middle to participate in this round of training, namely:

[0116] , ,

[0117] The initial value is 0.

[0118] Step 4.2: The edge CPUs use the training set to train the parameters of their respective deployed second-half models, obtaining the updated parameters. The parameters of the second half of the model are transmitted to the cloud CPU respectively;

[0119] No. The parameters of the second half model updated by the edge CPU are as follows:

[0120] For learning rate, , For the first In the first round of global training The parameters of the second half of the edge CPU model.

[0121] Step 4.3: The cloud CPU receives the second-half model parameters transmitted from each edge CPU and performs a weighted average to obtain the first... Global second-half model parameters updated in each round:

[0122]

[0123]

[0124]

[0125] For the first Global second-half model parameters in cloud CPU during round-wide training;

[0126] Step 4.4, the cloud CPU will... The global second half model updated in each round is delegated to the edge CPUs, and the edge CPUs update their respective second half models. Rear wheel model parameters:

[0127]

[0128] Step 4.5: Determine whether the number of rounds of global training has reached the set threshold. If yes, end; otherwise, resample the training set and return to step 4.2.

[0129] Step 5: Each edge CPU calculates the gradient of the fault based on the currently updated second-half model parameters and its own integrated data, and transmits it to each user to update the feature extractor for each user.

[0130] No. The corresponding user in each edge CPU Data Calculate the gradient of the fault :

[0131]

[0132] .

[0133] user Feature extractor The updated formula is:

[0134]

[0135] The equals sign represents assignment.

[0136] Step 6: Globally aggregate the updated feature extractors of all users in the same CF mMIMO subnetwork to obtain the updated feature extractor parameters;

[0137] After the local update is complete, the user Global updates are achieved by relying on the entire CF mMIMO network to update the global model.

[0138]

[0139] Step 7: Return to Step 3, use the feature extractor with updated parameters to extract and integrate the data features again, continue federated learning (FL) training, update the parameters of the second half of the edge CPU model until the required number of FL training iterations or the parameters of the second half of the model are qualified.

[0140] Step 8: Deploy the updated model to each AP, wait for users to issue new real-time requests, and output the final inference results to complete the service.

[0141] APs within the same CF mMIMO system network collaborate on model inference, assisting the CF mMIMO network in enabling AI-enabled services. For example... Figure 8 As shown, the specific steps are as follows:

[0142] Step 8.1: The system deploys the trained model to each AP. In a CF mMIMO subnetwork, AP Deploy and maintain the model ,in ;

[0143] Step 8.2: At different time periods, users Initiate an inference request and submit feature data. The data is transmitted to the corresponding edge CPU, where the CPU reconstructs the inference feature data.

[0144] Step 8.3: The edge CPU schedules the AP for collaborative inference for the current user based on information such as the user's simultaneous request for inference and channel parameters. The following constraints must be met:

[0145]

[0146]

[0147] in This refers to the set of users who initiate inference requests in the same edge CF mMIMO network at the current moment.

[0148] For the first A set of APs in a CF mMIMO network. Indicates user With AP Channel gain between The threshold value for filtering APs using the above formula is a scalar. Use the first A set of APs for each user;

[0149] Step 8.4, the edge CPU will connect the user Data that needs to be reasoned Transmitted to the corresponding set The selected AP uses the locally deployed model for inference:

[0150]

[0151] in For the forward propagation function of the second half of the model, For the model The calculated logits values, after normalization, can be used to calculate the labels. .

[0152] Step 8.5: Multiple APs synchronously transmit the final results, perform joint inference through over-the-air computation, and provide the user with the final inference result, thus completing the service.

[0153]

[0154]

[0155] ).

[0156] For downlink AP With users The transmit power coefficient between The normalized downlink signal-to-noise ratio, For users in the downlink Received noise.

[0157] Furthermore, the model deployed by AP can be fine-tuned based on the current input data:

[0158] .

[0159] This invention is not limited to the above-described embodiments. Any modifications, improvements, or substitutions that can be conceived by those skilled in the art without departing from the essential content of this invention fall within the protection scope of this invention.

Claims

1. A method for neural network training and inference in a cellular-free massive MIMO system based on federated learning, characterized in that, The specific steps are as follows: Step 1: Build a system using cloud CPU and... A communication scenario consisting of an edge CF mMIMO subnetwork; Each edge CF mMIMO subnetwork corresponds to an edge CPU; Step 2: System initialization, multiple edge CF mMIMO subnetworks establish communication connections with the cloud CPU; No. Users in a CF mMIMO subnetwork Holding the original training data and feature extractor The corresponding number Each edge CPU holds the second half of the model. The cloud CPU holds the global second half model. ; Step 3, regarding the first Each user completes their own data feature extraction in a CF mMIMO sub-network, which is then transmitted to the edge CPU for integration. ; Step 4: Each edge CPU uses the integrated data to collaborate with the cloud CPU to complete federated learning (FL) training and obtain updated second-half model parameters. The specific steps are as follows: Step 4.1, regarding the first In the global training round, each edge CPU samples from the integrated dataset to obtain the training set; Step 4.2: The edge CPUs use the training set to train the parameters of their respective deployed second-half models, obtaining the updated parameters. The parameters of the second half of the model are transmitted to the cloud CPU respectively; No. The parameters of the second half model updated by the edge CPU are as follows: For learning rate, , For the first In the first round of global training The parameters of the second half of the edge CPU model; Step 4.3: The cloud CPU receives the second-half model parameters transmitted from each edge CPU and performs a weighted average to obtain the first... Global second-half model parameters updated in each round: For the first Global second-half model parameters in cloud CPU during round-wide training; Step 4.4, the cloud CPU will... The global second half model updated in each round is delegated to the edge CPUs, and the edge CPUs update their respective second half models. Rear wheel model parameters: Step 4.5: Determine whether the number of rounds of global training has reached the set threshold. If yes, end; otherwise, resample the training set and return to step 4.

2. Step 5: Each edge CPU calculates the gradient of the fault based on the currently updated second-half model parameters and its own integrated data, and transmits it to each user to update the feature extractor for each user. No. The corresponding user in each edge CPU Data Calculate the gradient of the fault : . user Feature extractor The updated formula is: Step 6: Globally aggregate the updated feature extractors of all users in the same CF mMIMO subnetwork to obtain the updated feature extractor parameters; The global aggregation formula is: Step 7: Return to Step 3, use the feature extractor with updated parameters to extract and integrate the data features again, continue federated learning (FL) training, update the parameters of the second half of the edge CPU model until the required number of FL training iterations or the parameters of the second half of the model are qualified. Step 8: Deploy the updated model to each AP, wait for users to issue new real-time requests, and output the final inference results to complete the service.

2. The method as described in claim 1, characterized in that, In step one, in a single edge CF mMIMO subnetwork, there are multiple users holding data, multiple APs, and one edge CPU; a single AP can receive signals from all users simultaneously, and each user is connected to all APs at the same time.

3. The method as described in claim 1, characterized in that, In step two, the feature extractor and the second half of the model form a complete neural network, which is the AI ​​model. In the same edge CF mMIMO subnetwork, each user calculates the features of the raw data locally using the feature extractor, and transmits the processed feature data to the edge CPU through multiple corresponding APs. The edge CPU uses this data as input to the second half of the model for training and updating, and periodically returns the fault gradient to the user side to help the user side update the feature extractor. Cloud CPU reception The parameters of the second half model of each edge CF mMIMO subnetwork are weighted and averaged to obtain the final global second half model, which is then distributed and updated.

4. The method as described in claim 1, characterized in that, Step three specifically includes the following steps: Step 3.1, regarding the first Each CF mMIMO subnetwork allows individual users to obtain their own raw data features using their respective feature extractors. user Using a feature extractor Obtain the characteristics of the raw data for: in for The forward propagation function, Original training data eigenvectors, For the first The user set of a CFmMIMO subnetwork; Step 3.2: Each user transmits its original data characteristics to each AP in the sub-network. Each AP receives the transmitted data from all users and transmits the signal to the edge CPU through the fronthaul link. Step 3.3: The edge CPU receives and reconstructs the feature data of each user, and integrates them to obtain: 。 5. The method as described in claim 1, characterized in that, In step 4.1, the first Each edge CPU from the integrated dataset The data sampled from the middle to participate in this round of training, namely: , , The initial value is 0.

6. The method as described in claim 1, characterized in that, The specific steps of step eight are as follows: Step 8.1, in the... In a CF mMIMO subnetwork, AP Deploy and maintain the model ; Step 8.2, User Initiate an inference request and submit feature data. The data is transmitted to the corresponding edge CPU, where the CPU reconstructs the inference feature data. Step 8.3: The edge CPU schedules the AP for collaborative inference for the current user based on information such as the user's simultaneous request for inference and channel parameters. The following constraints must be met: in The set of users who initiate inference requests in the same edge CF mMIMO network at the current moment; For the first A set of APs in a CF mMIMO network. Indicates user With AP Channel gain between This represents the threshold for filtering APs using the above formula; Use the AP set of the n'th user; Step 8.4, the edge CPU will connect the user Data that needs to be reasoned Transmitted to the corresponding set The selected AP uses the locally deployed model for inference: in For the forward propagation function of the second half of the model, For the model The calculated logits values, after normalization, can be used to calculate the labels. ; Step 8.5: Multiple APs synchronously transmit the final results, perform joint inference through over-the-air computation, and provide the user with the final inference result, thus completing the service. ) The normalized downlink signal-to-noise ratio, For downlink AP With users The transmit power coefficient between For users in the downlink Received noise.

7. The method as described in claim 6, characterized in that, The model deployed by AP can be fine-tuned based on the current input data: 。

Images