Model life cycle management method and system, central node and distributed nodes
Through clustering federated learning and model life cycle management systems, the problems of federated learning models in data heterogeneity and complexity are solved, and the ability to efficiently train and adapt to new scenarios is realized, which is suitable for wireless communication systems.
Patent Information
- Application Number
- CN202410980473.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-26
AI Technical Summary
During the training process, the existing federated learning model has reduced convergence speed and training accuracy due to data heterogeneity, and the high model complexity leads to cost increase. It has failed to adapt to new scenarios or deployed on newly added nodes, and the communication system has There is a lack of effective signaling interaction mechanism and configuration information design.
Using clustering federated learning method, the distributed nodes are clustered through central nodes to form a training group, reduce the complexity of model training, improve training efficiency, and realize the identification, deployment, monitoring and update of models through the model life cycle management system, which is suitable for Provide new scenarios and newly added nodes.
It effectively reduces the complexity and cost of model training, improves training efficiency and generalization, adapts to new scenarios and newly added nodes, and implements data privacy security and signaling interaction mechanisms in the communication system.
Smart Images

Figure CN119025939A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of wireless communications, and in particular to a model lifecycle management method and system, a central node, and a distributed node. Background Art
[0002] Related technologies conduct standardization research on key technologies of air interface AI (Artificial Intelligence). Among them, federated learning, as a distributed framework, can effectively improve the training efficiency of AI models, protect the privacy and security of local data while reducing transmission overhead. Therefore, federated learning has important research value and broad application prospects in 6G (6th Generation Mobile Communication Technology) networks.
[0003] Data heterogeneity is widely present in wireless communication systems, which is manifested in that the data obtained by each UE (User Equipment) through data collection has significant differences in data distribution (usually non-independent and identically distributed characteristics). The main reasons are the differences in user behavior and the dynamic changes in the wireless channel environment. The existence of data heterogeneity has an adverse effect on the training of federated learning models (for example, in terms of convergence speed and training accuracy).
[0004] In the relevant technical solutions, the federated learning solutions to combat data heterogeneity mainly adopt solutions based on data enhancement or model enhancement. The data enhancement-based methods mainly use generative methods to improve data diversity to achieve the purpose of data enhancement, while the model enhancement-based methods mainly reduce the impact of data heterogeneity by designing complex model structures and model training methods. Summary of the invention
[0005] The inventors discovered through research that due to the high complexity of the model, the cost of model training in related technologies is significantly increased.
[0006] In view of at least one of the above technical problems, the present disclosure provides a model LCM (Life Cycle Management) method and system, a central node and a distributed node, and adopts a clustering federated learning method to reduce the complexity of model training, improve the efficiency of model training, and reduce the cost of model training.
[0007] According to one aspect of the present disclosure, a model lifecycle management method is provided, which is executed by a central node and includes:
[0008] Clustering all distributed nodes to form at least one training group, wherein each training group includes a plurality of distributed nodes;
[0009] In each training group, the cluster federated learning model is trained together with the multiple distributed nodes of the training group to obtain a trained cluster federated learning model.
[0010] In some embodiments of the present disclosure, clustering all distributed nodes to form at least one training group includes:
[0011] Determine data distribution similarity of the distributed nodes according to the distributed node data, wherein the distributed node data includes at least one of the distance of the loss value after pre-training of the local data and the cosine similarity of the gradient;
[0012] Based on the data distribution similarity, distributed node clustering is performed to form at least one training group.
[0013] In some embodiments of the present disclosure, performing distributed node clustering based on the data distribution similarity to form at least one training group includes:
[0014] For each distributed node, determining a benefit function value of the distributed node according to the data distribution similarity of the distributed node;
[0015] Determining a cost function value of the distributed node according to a transmission overhead of a local uploaded gradient of the distributed node;
[0016] Determining a utility function value of the distributed node according to the benefit function value and the cost function value of the distributed node;
[0017] According to the utility function value of each distributed node, distributed nodes are clustered to form at least one training group.
[0018] In some embodiments of the present disclosure, clustering distributed nodes according to the utility function value of each distributed node to form at least one training group includes:
[0019] Given an initial grouping, traverse each distributed node and iteratively execute switching rules until the clustering result no longer changes, thereby forming a final distributed node grouping result, wherein the switching rules include: for each current training group as a target training group in turn, execute a switching operation, wherein the switching operation includes: when a distributed node attempts to join a target training group, if the utility function value of the distributed node itself becomes larger, and the utility function value of any distributed node in the target training group does not decrease, then the distributed node is added to the target training group, otherwise, the switching operation is repeated for the next target training group until the distributed node joins a target training group.
[0020] In some embodiments of the present disclosure, in each training group, performing model training with the multiple distributed nodes of the training group to generate a cluster federated learning model includes:
[0021] Receive local model parameters updated and reported by distributed nodes;
[0022] Aggregate the local model parameters of all distributed nodes in the training group and update the global model parameters;
[0023] The global model parameters are sent to the distributed nodes, wherein the global model parameters are used by the distributed nodes to update the local model parameters of the distributed nodes.
[0024] In some embodiments of the present disclosure, in each training group, performing model training with the multiple distributed nodes of the training group to generate a cluster federated learning model further includes:
[0025] Repeat the steps of receiving local model parameters updated and reported by distributed nodes, aggregating the local model parameters of all distributed nodes in the training group, updating global model parameters, and sending the global model parameters to the distributed nodes until the clustering federated learning model converges.
[0026] In some embodiments of the present disclosure, the training group includes a group identifier, and the group identifier includes at least one of a cell or cell group identifier and a training group identifier.
[0027] In some embodiments of the present disclosure, the model lifecycle management method further includes:
[0028] Receiving user equipment capabilities reported by the distributed node, wherein the user equipment capabilities are used to indicate at least one of a model and a function supported by the user equipment;
[0029] At least one of a model and a function is identified.
[0030] In some embodiments of the present disclosure, the identifying at least one of the model and the function includes:
[0031] At least one of the model and the function is identified by at least one of an offline method and an online method.
[0032] In some embodiments of the present disclosure, the model lifecycle management method further includes:
[0033] A global model is transmitted or handed over to the distributed node, wherein the transmitted or handed over global model includes at least one of a first model and a second model, the first model is at least one clustered federated learning model of the cell or cell group where the distributed node is located obtained through model training, and the second model is at least one clustered federated learning model of the cell or cell group where the distributed node is not located obtained through model training.
[0034] In some embodiments of the present disclosure, the model lifecycle management method further includes:
[0035] Determining at least one of a model and a function matching the distributed node based on at least one of a scenario, a configuration, a site, and a user equipment capability;
[0036] Send model management configuration signaling for model reasoning related to model deployment to the distributed nodes, wherein the model management configuration signaling is used for the distributed nodes to select, activate, deactivate, switch, and rollback at least one of the models and functions.
[0037] In some embodiments of the present disclosure, the model lifecycle management method further includes:
[0038] Receive monitoring output results of at least one of the functions and models reported by the distributed nodes.
[0039] In some embodiments of the present disclosure, the model lifecycle management method further includes:
[0040] Determine whether to adjust the generated cluster federated learning model according to the monitoring output result and the change of at least one of the scenario, configuration, site and user equipment capability, wherein the adjustment includes at least one of adding members to the training group, deleting members from the training group, retraining and fine-tuning;
[0041] When it is determined that the generated cluster federated learning model needs to be adjusted, a model update configuration signaling is sent to the distributed nodes.
[0042] According to another aspect of the present disclosure, a model lifecycle management method is provided, which is executed by a distributed node and includes:
[0043] Reporting the distributed node data to the central node, wherein the distributed node data is used for the central node to cluster all the distributed nodes to form at least one training group, each training group including multiple distributed nodes;
[0044] In each training group, the central node corresponding to the training group jointly trains the cluster federated learning model to obtain a trained cluster federated learning model.
[0045] In some embodiments of the present disclosure, in each training group, the central node corresponding to the training group performs model training to generate a cluster federated learning model, including:
[0046] Update the local model and obtain local model parameters;
[0047] Uploading the local model parameters to the central node, wherein the local model parameters are used for the central node to aggregate the local model parameters of all distributed nodes in the training group and update the global model parameters;
[0048] The global model parameters sent by the central node are received, wherein the global model parameters are used by the distributed nodes to update the local model parameters of the distributed nodes.
[0049] In some embodiments of the present disclosure, in each training group, the central node corresponding to the training group performs model training to generate a cluster federated learning model further comprising:
[0050] Repeat the steps of updating the local model, obtaining local model parameters, uploading the local model parameters to the central node, and receiving the global model parameters issued by the central node until the clustering federated learning model converges.
[0051] In some embodiments of the present disclosure, the model lifecycle management method further includes:
[0052] Reporting user equipment capabilities to the central node, wherein the user equipment capabilities are used to indicate at least one of the models and functions supported by the user equipment, so that the central node can identify at least one of the models and functions.
[0053] In some embodiments of the present disclosure, reporting the user equipment capability to the central node includes at least one of the following steps, wherein:
[0054] The model is reported in the form of a model identifier or a model identifier list, wherein the supported model types include at least one of a generative vector machine, a neural network, and a random forest;
[0055] The function is reported in the form of a function identifier or a function identifier list, wherein the supported function types include at least one of classification, regression, clustering, dimensionality reduction and association.
[0056] In some embodiments of the present disclosure, the model lifecycle management method further includes:
[0057] A global model transmitted or handed over by a receiving central node, wherein the transmitted or handed over global model includes at least one of a first model and a second model, the first model being at least one clustered federated learning model of a cell or cell group where a distributed node is located obtained through model training, and the second model being at least one clustered federated learning model of a cell or cell group where a distributed node is not located obtained through model training.
[0058] In some embodiments of the present disclosure, the model lifecycle management method further includes:
[0059] Based on at least one of the scenario, configuration, site, and user device capability, for the artificial intelligence-related feature, determining at least one of a function and a model that matches the artificial intelligence-related feature;
[0060] For at least one of the determined functions and models, perform at least one of selection, activation, deactivation, switching, and rollback;
[0061] Feedback model management configuration signaling related to model deployment for model reasoning to the central node.
[0062] In some embodiments of the present disclosure, the model lifecycle management method further includes:
[0063] At least one model is selected and activated according to the model management configuration signaling for performing model reasoning.
[0064] In some embodiments of the present disclosure, the model lifecycle management method further includes:
[0065] The cluster federated learning model is monitored to obtain a monitoring output result of the cluster federated learning model.
[0066] In some embodiments of the present disclosure, the model lifecycle management method further includes:
[0067] Determine whether to adjust the generated cluster federated learning model according to the monitoring output result and the change of at least one of the scenario, configuration, site and user equipment capability, wherein the adjustment includes at least one of adding members to the training group, deleting members from the training group, retraining and fine-tuning;
[0068] When it is determined that the generated cluster federated learning model needs to be adjusted, a model update configuration signaling is sent to the central node.
[0069] According to another aspect of the present disclosure, there is provided a central node, comprising:
[0070] A grouping module is configured to cluster all distributed nodes to form at least one training group, wherein each training group includes a plurality of distributed nodes;
[0071] The model training module is configured to train the cluster federated learning model together with the multiple distributed nodes in the training group in each training group to obtain a trained cluster federated learning model.
[0072] According to another aspect of the present disclosure, there is provided a central node, comprising:
[0073] a memory configured to store instructions;
[0074] The processor is configured to execute the instructions so that the central node executes the model lifecycle management method as described in any of the above embodiments.
[0075] According to another aspect of the present disclosure, there is provided a distributed node, comprising:
[0076] A data reporting module is configured to report distributed node data to a central node, wherein the distributed node data is used for the central node to cluster all distributed nodes to form at least one training group, each training group including a plurality of distributed nodes;
[0077] The distributed training module is configured to train the cluster federated learning model together with the central node corresponding to the training group in each training group to obtain a trained cluster federated learning model.
[0078] According to another aspect of the present disclosure, there is provided a distributed node, comprising:
[0079] a memory configured to store instructions;
[0080] The processor is configured to execute the instructions so that the distributed nodes execute the model lifecycle management method as described in any of the above embodiments.
[0081] According to another aspect of the present disclosure, a model lifecycle management system is provided, comprising a central node as in any of the above embodiments and a distributed node as in any of the above embodiments.
[0082] According to another aspect of the present disclosure, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and when the instructions are executed by a processor, the model lifecycle management method as described in any of the above embodiments is implemented.
[0083] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, wherein when the computer program is executed by a processor, the model lifecycle management method as described in any of the above embodiments is implemented.
[0084] The present invention adopts a clustering federated learning method to reduce the complexity of model training, improve the efficiency of model training, and reduce the cost of model training. BRIEF DESCRIPTION OF THE DRAWINGS
[0085] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0086] Figure 1 It is a schematic diagram of some embodiments of the model lifecycle management system disclosed in the present invention.
[0087] Figure 2 It is a schematic diagram of some embodiments of the model lifecycle management method disclosed in the present invention.
[0088] Figure 3 Schematic diagrams of other embodiments of the model lifecycle management method disclosed in the present invention.
[0089] Figure 4 Schematic diagrams of some further embodiments of the model lifecycle management method disclosed herein.
[0090] Figure 5 Schematic diagrams of some further embodiments of the model lifecycle management method disclosed herein.
[0091] Figure 6 Schematic diagram of some embodiments of the central node of the present disclosure.
[0092] Figure 7 Schematic diagram of the structure of other embodiments of the central node of the present disclosure.
[0093] Figure 8 Schematic diagram of some embodiments of the distributed nodes disclosed herein.
[0094] Fig. 9 Schematic diagram of the structure of some other embodiments of the distributed node disclosed in the present invention. DETAILED DESCRIPTION
[0095] The technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present disclosure and its application or use. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.
[0096] Unless specifically stated otherwise, the relative arrangement of components and steps, the numerical expressions and numerical values set forth in these embodiments do not limit the scope of the present disclosure.
[0097] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.
[0098] Technologies, methods, and apparatus known to ordinary technicians in the relevant field may not be discussed in detail, but where appropriate, such technologies, methods, and apparatus should be considered part of the authorization specification.
[0099] In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values.
[0100] It should be noted that like reference numerals and letters refer to similar items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.
[0101] Through research, the inventors found that there are still some technical problems in the related technology: 1. The federated learning model of the related technology cannot adapt to new scenarios or be deployed on newly added nodes; 2. Due to the high complexity of the related technology model, the cost of model training or model reasoning is significantly increased; the related technical solution does not consider the application of the proposed solution in the communication system, especially the corresponding signaling interaction mechanism and configuration information design in the wireless air interface.
[0102] In view of at least one of the above technical problems, the present disclosure provides a model lifecycle management method and system, a central node and a distributed node. The present disclosure is described below through specific embodiments.
[0103] Figure 1 Schematic diagram of some embodiments of the model lifecycle management system disclosed in the present invention. Figure 1 As shown, the model lifecycle management system disclosed in the present invention may include a central node 1 and multiple distributed nodes 2.
[0104] The central node 1 is configured to cluster all distributed nodes to form at least one training group, wherein each training group includes a plurality of distributed nodes.
[0105] Distributed node 2 is configured to train a local model in each training group.
[0106] In some embodiments of the present disclosure, the central node 1 may also be configured to aggregate and issue update model parameters.
[0107] In some embodiments of the present disclosure, the distributed node 2 may be implemented as a local client.
[0108] In some embodiments of the present disclosure, the distributed node 2 may be implemented as a UE (User Equipment).
[0109] In some embodiments of the present disclosure, the central node 1 may be implemented as a server.
[0110] In some embodiments of the present disclosure, the central node 1 may be implemented as a NW (network side device).
[0111] In some embodiments of the present disclosure, the central node 1 can be implemented as at least one of gNB (the next Generation Node B, next generation base station, 5G base station), LMF (Location Management Function, positioning management function network element), and OAM (Orchestration and Management, orchestration and management network element).
[0112] In some embodiments of the present disclosure, for model training: within any training group, the distributed node 2 trains the local model, and the central node 1 aggregates and sends down updated model parameters.
[0113] In some embodiments of the present disclosure, model reasoning is performed by distributed nodes 2 .
[0114] The above embodiments of the present disclosure provide a configuration system suitable for lifecycle management of a wireless intelligent air interface clustering federated learning model. The above embodiments of the present disclosure define a distributed network architecture based on federated learning.
[0115] The above embodiments of the present disclosure use the base station, OAM or LMF as the central node and the UE as the distributed node. At the same time, the above embodiments of the present disclosure are applicable to various solutions such as distributed deployment, centralized deployment or hybrid deployment of computing resources on the RAN side in the future.
[0116] The model lifecycle management method disclosed herein and the structure and function of the central node 1 and the multiple distributed nodes 2 are further described below through specific embodiments.
[0117] Figure 2 Schematic diagram of some embodiments of the model lifecycle management method of the present disclosure. Preferably, this embodiment can be executed by the central node of the present disclosure or the model lifecycle management system of the present disclosure. The method includes at least one step from step 21 to step 22.
[0118] In some embodiments of the present disclosure, lifecycle management (LCM) is an important concept proposed based on the wireless AI framework. As a full-process management solution covering AI models, LCM includes data collection, model training, model reasoning, model management (based on functionality / model ID selection, activation / deactivation, switching and fallback operations, monitoring, updating), model transmission / handover, etc.
[0119] Step 21: The central node clusters all distributed nodes to form at least one training group, wherein each training group includes a plurality of distributed nodes.
[0120] In some embodiments of the present disclosure, step 21 may include: the central node executes a clustering algorithm to form K user groups, where K is a natural number greater than or equal to 1.
[0121] In some embodiments of the present disclosure, the clustering grouping algorithm (the basic method for performing UE clustering) may be an unsupervised learning algorithm or a semi-supervised learning algorithm.
[0122] In some embodiments of the present disclosure, the unsupervised learning algorithm is: not relying on additional labeled data information in the clustering task, such as: prototype clustering, density clustering, hierarchical clustering, etc.
[0123] In some embodiments of the present disclosure, the semi-supervised learning algorithm is: using some additional labeled data information in the clustering task, such as: constraint-based methods, methods based on the spatial structure of data sets, grid-based methods, etc.
[0124] In some embodiments of the present disclosure, step 21 may include at least one of steps 211 to 212 .
[0125] Step 211, determining data distribution similarity of distributed nodes based on each distributed node data, wherein the distributed node data includes at least one of the distance of the loss value after local data pre-training and the cosine similarity of the gradient.
[0126] In some embodiments of the present disclosure, the distance may be L1 distance (Manhattan distance) or L2 distance (Euclidean distance).
[0127] Step 212: Perform distributed node clustering based on the data distribution similarity to form at least one training group.
[0128] The above-mentioned embodiments of the present disclosure can perform distributed node clustering according to the data distribution similarity of the distributed nodes, thereby improving the performance of federated learning against data heterogeneity.
[0129] In some embodiments of the present disclosure, step 212 may include at least one of steps 2121 to 2124 .
[0130] Step 2121: for each distributed node, determine the benefit function value of the distributed node according to the data distribution similarity of the distributed node.
[0131] Step 2122: Determine the cost function value of the distributed node according to the transmission overhead of the local uploaded gradient of the distributed node.
[0132] Step 2123, determining the utility function value of the distributed node according to the benefit function value and the cost function value of the distributed node.
[0133] In some embodiments of the present disclosure, the UE clustering distributed algorithm uses a coalition formation game algorithm, which defines the utility function value of any user as the benefit function value minus the cost function value, where the benefit function value is the data distribution similarity, and the cost function value is the transmission overhead of the local user uploading the gradient.
[0134] Step 2124, clustering the distributed nodes according to the utility function value of each distributed node to form at least one training group.
[0135] In some embodiments of the present disclosure, step 2124 may include: given an initial grouping, and traversing each distributed node to iteratively execute switching rules until the clustering result no longer changes, forming a final distributed node grouping result, wherein the switching rules include: for each current training group as a target training group in turn, performing a switching operation, wherein the switching operation includes: when a distributed node attempts to join a target training group, if the utility function value of the distributed node itself becomes larger, and the utility function value of any distributed node in the target training group does not decrease, then the distributed node is added to the target training group, otherwise, the switching operation is repeated for the next target training group until the distributed node joins a target training group.
[0136] Step 22: In each training group, the central node trains the cluster federated learning model together with the multiple distributed nodes in the training group to obtain a trained cluster federated learning model.
[0137] In some embodiments of the present disclosure, the trained cluster federated learning model can be used to perform model lifecycle management operations such as model transmission or handover, model deployment, model monitoring, model update, and model reasoning.
[0138] In some embodiments of the present disclosure, after the training is completed, the central node obtains the global model of the trained cluster federated learning model, and the distributed nodes obtain the local model of the trained cluster federated learning model.
[0139] In some embodiments of the present disclosure, step 22 may include at least one of steps 221 to 224 .
[0140] Step 221: receiving local model parameters updated and reported by distributed nodes.
[0141] Step 222: aggregate the local model parameters of all distributed nodes in the training group and update the global model parameters.
[0142] Step 223: Send the global model parameters to the distributed nodes, wherein the global model parameters are used by the distributed nodes to update the local model parameters of the distributed nodes.
[0143] Step 224, repeatedly execute the steps of receiving the local model parameters updated and reported by the distributed nodes, aggregating the local model parameters of all distributed nodes in the training group, updating the global model parameters and sending the global model parameters to the distributed nodes (for example, steps 221 to 223) until the clustering federated learning model converges.
[0144] In some embodiments of the present disclosure, the training group includes a group identifier, and the group identifier includes at least one of a cell or cell group identifier and a training group identifier.
[0145] In some embodiments of the present disclosure, the cluster federated learning model training group has a group ID, which is unique within the cell / cell group. The format of the group ID is composed of two parts of bits connected in series, including at least one of the following bits: an identification bit of the cell / cell group, a bit used to distinguish the same cell / cell group; and a bit used to distinguish different cluster federated learning model training groups.
[0146] In some embodiments of the present disclosure, an M-bit binary bit string is used to identify a cell / cell group, for example: cellID / cell group ID; an N-bit binary bit string is used to distinguish different clustering federated learning models within the same cell / cell group, for example: multiple models trained using different gradient descent methods, different optimization objectives or different data sets, or can be identified based on UE location, etc.
[0147] The relevant technology has not yet defined a solution for data heterogeneity. The above-mentioned embodiment of the present disclosure can effectively reduce the data distribution differences within each user group based on the user clustering method, thereby overcoming the impact of data heterogeneity on the training and model reasoning performance of the federated learning model.
[0148] Figure 3 Schematic diagrams of other embodiments of the model lifecycle management method disclosed in the present invention. Preferably, this embodiment can be executed by the central node disclosed in the present invention or the model lifecycle management system disclosed in the present invention. In addition to including Figure 2 In addition to the method of the embodiment, the present invention may also include Figure 3 At least one of steps 31 to 38 of the embodiment.
[0149] Step 31: The central node receives the user equipment capabilities reported by the distributed nodes, wherein the user equipment capabilities are used to indicate at least one of the models and functions supported by the user equipment.
[0150] In step 32, the central node identifies at least one of a model and a function.
[0151] In some embodiments of the present disclosure, step 32 may include: identifying at least one of the model and the function in at least one of an offline manner and an online manner.
[0152] In some embodiments of the present disclosure, step 32 may include at least one of the following steps: identification of at least one of an offline model and a function; and identification of at least one of an online model and a function.
[0153] In step 33, the central node transmits or hands over the global model to the distributed node, wherein the transmitted or handed over global model includes at least one of a first model and a second model, the first model is at least one clustered federated learning model of the cell or cell group where the distributed node is located obtained through model training, and the second model is at least one clustered federated learning model of the cell or cell group where the distributed node is not located obtained through model training.
[0154] Step 34: The central node determines at least one of a model and a function that matches the distributed node based on at least one of a scenario, a configuration, a site, and a user equipment capability.
[0155] In step 35, the central node sends a model management configuration signaling related to model deployment for model reasoning to the distributed nodes, wherein the model management configuration signaling is used for the distributed nodes to select, activate, deactivate, switch, and roll back at least one of the models and functions.
[0156] Step 36: The central node receives the monitoring output results of the clustering federated learning model reported by the distributed nodes.
[0157] In some embodiments of the present disclosure, step 36 may include: the central node receives the monitoring output result of at least one of the functions and models reported by the distributed nodes.
[0158] Step 37, the central node determines whether to adjust the generated cluster federated learning model based on the monitoring output results and changes in at least one of the scenarios, configurations, sites, and user equipment capabilities, wherein the adjustment includes at least one of adding members to the training group, deleting members from the training group, retraining, and fine-tuning.
[0159] Step 38: When the central node determines to adjust the generated cluster federated learning model, the central node sends a model update configuration signaling to the distributed nodes.
[0160] The present invention provides a wireless intelligent air interface model lifecycle management configuration method based on clustering federated learning, which adopts a clustering federated learning method to reduce the complexity of model training, improve the efficiency of model training, and reduce the cost of model training.
[0161] The above-mentioned embodiment scheme of the present disclosure defines the signaling process required for the lifecycle management of the cluster federated learning model and the configuration information that must be included in the signaling, and is a technical reserve for implementing federated learning model management on the RAN side.
[0162] The above-mentioned embodiments of the present disclosure define a configuration method for implementing the lifecycle management of the federated learning model on the RAN side. Through the model management configuration signaling, a model management solution based on functions and model IDs is supported, including at least one of the functions of selecting, activating, deactivating, switching, and falling back for the clustered federated learning model in the model reasoning phase, thereby realizing RAN's support for the basic functions of federated learning model reasoning.
[0163] The above-mentioned embodiment of the present disclosure proposes a configuration method suitable for the lifecycle management of the wireless intelligent air interface clustering federated learning model, which can implement at least one of the model training, reasoning, management, monitoring and deployment based on clustering federated learning on the RAN side, protect data privacy and security, and reduce the impact of data heterogeneity on the performance of federated learning.
[0164] Figure 4 Schematic diagram of some other embodiments of the model lifecycle management method of the present disclosure. Preferably, this embodiment can be executed by the distributed node of the present disclosure or the model lifecycle management system of the present disclosure. The model lifecycle management method of the present disclosure can include at least one step from step 41 to step 42.
[0165] Step 41, the distributed nodes report the distributed node data to the central node, wherein the distributed node data is used for the central node to cluster all the distributed nodes to form at least one training group, each training group including multiple distributed nodes.
[0166] In step 42, the distributed nodes in each training group train the cluster federated learning model together with the central node corresponding to the training group to obtain a trained cluster federated learning model.
[0167] In some embodiments of the present disclosure, the trained cluster federated learning model can be used to perform model lifecycle management operations such as model transmission or handover, model deployment, model monitoring, model update, and model reasoning.
[0168] In some embodiments of the present disclosure, after the training is completed, the central node obtains the global model of the trained cluster federated learning model, and the distributed nodes obtain the local model of the trained cluster federated learning model.
[0169] In some embodiments of the present disclosure, step 42 may include at least one of steps 421 to 424 .
[0170] Step 421: The distributed node updates the local model and obtains local model parameters.
[0171] In step 422, the distributed nodes upload the local model parameters to the central node, wherein the local model parameters are used for the central node to aggregate the local model parameters of all distributed nodes in the training group and update the global model parameters.
[0172] Step 423: The distributed node receives the global model parameters sent by the central node, wherein the global model parameters are used by the distributed node to update the local model parameters of the distributed node.
[0173] In step 424, the distributed nodes repeatedly execute the steps of updating the local model, obtaining the local model parameters, uploading the local model parameters to the central node, and receiving the global model parameters issued by the central node (i.e., steps 421 to 423) until the clustering federated learning model converges.
[0174] In some embodiments of the present disclosure, Figure 4 As shown, the model lifecycle management method disclosed in the present invention may further include at least one step from step 43 to step 49.
[0175] Step 43, the distributed node reports the user equipment capability to the central node, wherein the user equipment capability is used to indicate at least one of the models and functions supported by the user equipment, so that the central node can identify at least one of the models and functions.
[0176] In some embodiments of the present disclosure, step 43 may include at least one of steps 431 to 432 .
[0177] Step 431, reporting in the form of a model identifier and at least one of a model identifier list, wherein the supported model types include at least one of a generation vector machine, a neural network, and a random forest.
[0178] Step 433: reporting in the form of a function identifier and at least one item in a function identifier list, wherein the supported function types include at least one of classification, regression, clustering, dimensionality reduction and association.
[0179] In step 44, the distributed node receives the global model transmitted or handed over by the central node, wherein the transmitted or handed over global model includes at least one of a first model and a second model, the first model is at least one clustered federated learning model of the cell or cell group where the distributed node is located obtained through model training, and the second model is at least one clustered federated learning model of the cell or cell group where the distributed node is not located obtained through model training.
[0180] Step 45, the distributed node determines at least one of the functions and models that match the artificial intelligence-related features based on the scenario, configuration, site, and at least one of the user equipment capabilities; the distributed node selects, activates, deactivates, switches, and rolls back at least one of the determined functions and models; the distributed node feeds back model management configuration signaling related to model deployment for model reasoning to the central node.
[0181] Step 46: The distributed node selects and activates at least one model according to the model management configuration signaling to perform model reasoning.
[0182] Step 47: The distributed nodes monitor the cluster federated learning model to obtain a monitoring output result of the cluster federated learning model.
[0183] In some embodiments of the present disclosure, step 47 may include: the distributed node monitors at least one of the function and the model to obtain a monitoring output result of at least one of the function and the model.
[0184] Step 48, the distributed node determines whether to adjust the generated cluster federated learning model based on the monitoring output results and changes in at least one of the scenarios, configurations, sites, and user equipment capabilities, wherein the adjustment includes at least one of adding members to the training group, deleting members from the training group, retraining, and fine-tuning.
[0185] Step 49: When the distributed node determines to adjust the generated cluster federated learning model, the distributed node sends a model update configuration signaling to the central node.
[0186] The above embodiments of the present disclosure can be adapted to new scenarios and can be deployed on newly added nodes.
[0187] The above embodiments of the present disclosure may be applied in a communication system. In particular, the above embodiments of the present disclosure provide a corresponding signaling interaction mechanism and configuration information design in a wireless air interface.
[0188] The above-mentioned embodiments of the present disclosure can be applied to various use cases such as distributed positioning, distributed beam management, distributed wireless resource management, etc. by defining the basic configuration and functional process of federated learning model training, thereby protecting the security and privacy of user-related data, reducing the transmission overhead related to the air interface and model training, and increasing the generalization and robustness of the model.
[0189] Figure 5 Schematic diagram of some other embodiments of the model lifecycle management method of the present disclosure. Preferably, this embodiment can be executed by the distributed node of the present disclosure or the model lifecycle management system of the present disclosure. The model lifecycle management method of the present disclosure can include at least one step from step 51 to step 54.
[0190] Step 51: Identify at least one of a model and a function based on cluster federated learning.
[0191] In some embodiments of the present disclosure, step 51 may include at least one of steps 511 to 512 .
[0192] Step 511: The central node identifies at least one of a model and a function.
[0193] In some embodiments of the present disclosure, step 511 (identification scheme of at least one of a model and a function) may include at least one of step A and step B.
[0194] Step A, offline mode (offline): NW and UE sides do not identify at least one of the model and function through air interface signaling: each model or function is assigned a model ID or function ID for model identification, which can be used in air interface interaction after model identification.
[0195] Step B, online: the NW and UE sides identify at least one of the model and the function through air interface signaling.
[0196] In some embodiments of the present disclosure, step B may include at least one of step B1 and step B2.
[0197] Step B1, the UE side initiates the identification of at least one of the models and functions, and reports the UE capability to indicate at least one of the models and functions supported by the UE, and the NW side assists in completing the remaining steps of the identification of at least one of the models and functions. During the identification of at least one of the models and functions, one or more IDs of at least one of the models and functions may be allocated, and the IDs may be used in the air interface interaction after the model identification.
[0198] In step B2, the NW side initiates the identification of at least one of the model and function, and the UE side responds and reports the UE capability to indicate the model / function supported by the UE. During the identification of at least one of the model and function, one or more IDs of at least one of the model and function may be allocated, and the ID may be used in the air interface interaction after the model identification.
[0199] Step 512: The distributed node reports UE capabilities, which are used to indicate at least one of the models and functions supported by the UE.
[0200] In some embodiments of the present disclosure, step 512 may include at least one of steps 5121 to 5122 .
[0201] Step 5121, model: reported in the form of model ID and at least one item in the model ID list. The supported model types include one or more, for example: generation vector machine (SVM), neural network, random forest, etc.
[0202] Step 5122, function: reported in the form of function ID and at least one item in the function ID list. The supported function types include one or more, for example: classification, regression, clustering, dimensionality reduction, association, etc.
[0203] In some embodiments of the present disclosure, UE capability reporting is associated with at least one of a specific configuration and a condition under the UE capability of the distributed node feature / feature group.
[0204] In some embodiments of the present disclosure, UE capability reporting is associated with auxiliary conditions (scenario, site, data set) directly decided and confirmed by the UE and the NW.
[0205] Step 52, model training process based on clustering federated learning.
[0206] The above-mentioned embodiments of the present disclosure define a clustering federated learning model training process, which may include a user clustering algorithm, local update, model air interface transmission, and global gradient aggregation.
[0207] In some embodiments of the present disclosure, step 52 may include at least one of steps 521 to 525 .
[0208] Step 521: The central node executes a clustering algorithm to form K user groups, where K is a natural number greater than or equal to 1.
[0209] In some embodiments of the present disclosure, step 521 may include: Figure 2 Step 21 of the embodiment.
[0210] In some embodiments of the present disclosure, step 521 may include: initialization.
[0211] In some embodiments of the present disclosure, step 521 may include: all distributed nodes initialize model parameters, and the central node executes a user grouping algorithm to form K user groups.
[0212] In some embodiments of the present disclosure, a distributed algorithm based on data distribution similarity is used when performing UE clustering, and the measurement method of data distribution similarity is at least one of the following parameters: L1 distance (Manhattan distance) or L2 distance (Euclidean distance) of the user's local loss value; cosine similarity of the user's local gradient.
[0213] The above embodiments of the present disclosure define that the basic method for performing UE clustering includes at least one of an unsupervised learning algorithm and a semi-supervised learning algorithm.
[0214] Step 522: In each user group, the distributed nodes perform local model updates to obtain local model gradients associated with the local data set.
[0215] In some embodiments of the present disclosure, each distributed node uses the stochastic gradient descent algorithm (SGD) to perform local model update. The specific rule is expressed as formula (1):
[0216]
[0217] In formula (1), w i,t is the distributed node model parameter; S i,t It is a data batch obtained by random sampling based on the local data set of distributed nodes; is the gradient of the loss function at the data point; i represents the index value of the distributed node, U irepresents the i-th distributed node; t represents the t-th iteration of the model training process. Model training requires several rounds of iterations to converge the global model parameter v; x, y represents a data point, where x is the data and y is the label value (true value) corresponding to x, which is used to verify how much the prediction result of the model w on x differs from y, which can also measure the size of the loss function.
[0218] Step 523: The distributed node uploads the local model gradient to the central node corresponding to the user group to which it belongs.
[0219] Step 524: The central node performs gradient aggregation, updates global model parameters, and sends them to distributed nodes.
[0220] In some embodiments of the present disclosure, the gradient aggregation rule can be expressed as formula (2).
[0221]
[0222] In formula (2), β i,t Indicates whether the central node successfully receives the distributed node U i The gradient of v k,t is the global model parameter of the central node; η is the learning rate; k represents the index value of the user group, C k represents the kth user group, U i ∈C k represents all distributed nodes belonging to the kth user group; i is the index value of the distributed node; t represents the tth iteration of the model training process, the local model update in step 402 performs one iteration, and the global model update in step 404 performs one iteration accordingly; D i Represents the number of data points of the i-th distributed node.
[0223] Step 525: Repeat steps 522 to 524 until the clustering federated learning model converges.
[0224] In some embodiments of the present disclosure, the criterion for judging model convergence is at least one of the following: the total number of iteration rounds reaches a preset value; the loss function value or training accuracy of two adjacent iteration rounds is less than a preset threshold.
[0225] In some embodiments of the present disclosure, steps 522 to 525 may include: model training: within any training group (user group), any distributed node updates local model parameters and transmits the model parameters to the central node through the air interface. The central node updates the global model parameters, and through several rounds of iterative updates until the global model parameters converge, K clustering federated learning models are formed.
[0226] In some embodiments of the present disclosure, after the training is completed, the central node obtains the global model of the trained cluster federated learning model, and the distributed nodes obtain the local model of the trained cluster federated learning model.
[0227] In some embodiments of the present disclosure, the trained cluster federated learning model can be used to perform model lifecycle management operations such as model transmission or handover, model deployment, model monitoring, model update, and model reasoning.
[0228] Step 53, model deployment and model reasoning process based on clustering federated learning.
[0229] In some embodiments of the present disclosure, step 53 may include: model reasoning: for any reasoning task, the central node performs at least one operation of selection, activation, deactivation, switching, and fallback on at least one clustered federated learning model, transmits or transfers the global model to the distributed node for performing model reasoning, and obtains model monitoring results.
[0230] In some embodiments of the present disclosure, step 53 may include at least one of steps 531 to 534 .
[0231] Step 531: The central node transmits or transfers the global model to the distributed nodes.
[0232] In some embodiments of the present disclosure, the global model of transmission or handover includes at least one of a first model and a second model, the first model is at least one clustered federated learning model of the cell or cell group where the distributed node is located obtained through model training, and the second model is at least one clustered federated learning model of the cell or cell group where the distributed node is not located obtained through model training.
[0233] In some embodiments of the present disclosure, the global model transmission or handover is in at least one of the following forms: sent in a private format; sent in a public format, and the UE knows the model structure; sent in a public format, and the UE does not know the model structure.
[0234] In some embodiments of the present disclosure, a terminal-specific binary executable file format is provided.
[0235] Step 532: The distributed nodes deploy the AI model based on the model management configuration signaling.
[0236] In some embodiments of the present disclosure, model management configuration signaling includes at least one of selection, activation, deactivation, switching, and fallback of a cluster federated learning model during the model reasoning phase.
[0237] In some embodiments of the present disclosure, the model management configuration signaling includes at least one of model management configuration signaling based on model function and model management configuration signaling based on model ID.
[0238] In some embodiments of the present disclosure, the signaling interaction of the model deployment process includes at least at least one of step 5321 and step 5322.
[0239] Step 5321, UE side triggering: the distributed node determines at least one of the functions and models that match a certain artificial intelligence (AI)-related feature based on at least one of the scenario, configuration, site and user equipment capabilities; the distributed node selects, activates, deactivates, switches, and falls back on at least one of the determined functions and models; the distributed node feeds back model management configuration signaling related to model deployment for model reasoning to the central node.
[0240] Step 5322, NW side trigger: the central node determines at least one of the models and functions that matches the distributed node based on at least one of the scenarios, configurations, sites and user equipment capabilities; the central node sends a model management configuration signaling for model reasoning related to model deployment to the distributed node, wherein the model management configuration signaling is used for the distributed node to select, activate, deactivate, switch, and fallback at least one of the models and functions.
[0241] In some embodiments of the present disclosure, the model management configuration signaling may include at least one of a system message, an RRC signaling, a MAC CE signaling, a DCI signaling, and a NAS message.
[0242] Step 533: The distributed node selects and activates one or more models according to the model management configuration signaling to perform model reasoning.
[0243] In some embodiments of the present disclosure, the model inference scheme includes at least one of the following: an online model inference scheme, for example, for application scenarios with high latency requirements, data arrives in the form of data streams; an offline model inference scheme, for example, for application scenarios with low latency requirements, data is stored in advance on distributed nodes in the form of data blocks.
[0244] In some embodiments of the present disclosure, in an online model inference scheme, for data sample points arriving in real time, the output at the sample points is calculated based on the deployed model as the model inference result.
[0245] The above-mentioned embodiments of the present disclosure define that the model reasoning scheme includes at least one of an online model reasoning scheme and an offline model reasoning scheme.
[0246] Step 534: The distributed nodes monitor the cluster federated learning model to obtain a monitoring output result of the cluster federated learning model.
[0247] In some embodiments of the present disclosure, step 534 may include: the distributed node monitors at least one of the functions and the model, and obtains the monitoring output result of at least one of the functions and the model. In some embodiments of the present disclosure, the KPI used for performance monitoring of at least one of the functions and the model includes at least one of the following: system performance indicators: for example, throughput, block error rate, latency, etc.; accuracy indicators: for example, accuracy of model output, MSE / NMSE / cosine similarity between model input and output, etc.; overhead and complexity of model monitoring: for example, model size (number of parameters), computational complexity (FLOPs), etc.; generalization performance.
[0248] Step 54, model update process based on clustering federated learning.
[0249] In some embodiments of the present disclosure, step 54 may include: model update: based on the model monitoring results, the distributed nodes select at least one of the following adjustments to the cluster federated learning model: adding members, deleting members from the training group, retraining, and fine-tuning.
[0250] In some embodiments of the present disclosure, the distributed nodes adjust the generated cluster federated learning model based on the model update configuration signaling, wherein the adjustment includes at least one of adding members to the training group, deleting members from the training group, retraining, and fine-tuning.
[0251] In some embodiments of the present disclosure, the model update configuration signaling interaction includes at least one of step 541 and step 542 .
[0252] Step 541, UE side trigger: The distributed node independently decides to add members to the training group of the cluster federated learning model training, delete members from the training group, retrain and fine-tune the training group based on the monitoring output results of at least one of the functions and models, and the changes in at least one of the scenarios, configurations, sites and user equipment capabilities, and feeds back the model update configuration signaling to the central node.
[0253] Step 542, NW side trigger: The central node decides to add members to the training group of the cluster federated learning model training, delete members from the training group, retrain and fine-tune, etc. based on the monitoring output results of at least one of the functions and models, and the changes in at least one of the scenarios, configurations, sites and user equipment capabilities, and sends model update configuration signaling to the distributed nodes.
[0254] The above-mentioned embodiment of the present disclosure proposes a configuration method suitable for lifecycle management of a wireless intelligent air interface clustering federated learning model.
[0255] The above-mentioned embodiments of the present disclosure propose a method for configuring the lifecycle management of a federated learning model that is resistant to data heterogeneity, and design a training scheme for a federated learning model that is resistant to data heterogeneity, such as adopting a clustering federated learning method, thereby reducing the complexity of model training, improving the efficiency of model training, and reducing the cost of model training.
[0256] The above-mentioned embodiments of the present disclosure design a model lifecycle management solution based on clustering federated learning.
[0257] The above-mentioned embodiments of the present disclosure design a signaling interaction mechanism and configuration information corresponding to a lifecycle management solution based on cluster federated learning.
[0258] In order to improve the efficiency of federated learning model management, the above-mentioned embodiments of the present disclosure can support function-based and model ID-based model lifecycle management solutions, including at least one operation of selection, activation, deactivation, switching, fallback, retraining and fine-tuning of clustered federated learning models.
[0259] The above-mentioned embodiment of the present disclosure proposes a configuration method for lifecycle management of federated learning models in 5G-Advanced and 6G wireless intelligent networks, which can realize federated learning model training, model reasoning, model monitoring, model management, and model deployment framework under wireless air interface, help to increase the generalization of the model, combat the impact of data heterogeneity, and be applicable to a variety of RAN-side AI use cases.
[0260] The above embodiments of the present disclosure can be applied to 6G, 5G-Advanced, wireless AI, high frequency / millimeter wave / terahertz, industrial Internet and other scenarios.
[0261] Figure 6 Schematic diagram of some embodiments of the central node of the present disclosure. Figure 6 As shown, the central node of the present disclosure may include a grouping module 61 and a model training module 62.
[0262] The grouping module 61 is configured to cluster all distributed nodes to form at least one training group, wherein each training group includes a plurality of distributed nodes.
[0263] In some embodiments of the present disclosure, the grouping module 61 can be configured to determine the data distribution similarity of the distributed nodes based on the data of each distributed node, wherein the distributed node data includes at least one of the distance of the loss value after pre-training of the local data and the cosine similarity of the gradient; and perform distributed node clustering based on the data distribution similarity to form at least one training group.
[0264] In some embodiments of the present disclosure, when the grouping module 61 performs distributed node clustering based on the data distribution similarity to form at least one training group, it can be configured to determine, for each distributed node, a benefit function value of the distributed node according to the data distribution similarity of the distributed node; determine a cost function value of the distributed node according to the transmission overhead of the local uploaded gradient of the distributed node; determine the utility function value of the distributed node according to the benefit function value and the cost function value of the distributed node; and perform distributed node clustering according to the utility function value of each distributed node to form at least one training group.
[0265] In some embodiments of the present disclosure, when the grouping module 61 clusters distributed nodes according to the utility function value of each distributed node to form at least one training group, it can be configured to give an initial grouping and iteratively execute switching rules by traversing each distributed node until the clustering result no longer changes, thereby forming a final distributed node grouping result, wherein the switching rules include: performing a switching operation on each current training group as a target training group in turn, wherein the switching operation includes: when a distributed node attempts to join a target training group, if the utility function value of the distributed node itself becomes larger and the utility function value of any distributed node in the target training group does not decrease, then the distributed node is added to the target training group; otherwise, the switching operation is repeated for the next target training group until the distributed node joins a target training group.
[0266] The model training module 62 is configured to train the cluster federated learning model together with the multiple distributed nodes in the training group in each training group to obtain a trained cluster federated learning model.
[0267] In some embodiments of the present disclosure, the trained cluster federated learning model can be used to perform model lifecycle management operations such as model transmission or handover, model deployment, model monitoring, model update, and model reasoning.
[0268] In some embodiments of the present disclosure, after the training is completed, the central node obtains the global model of the trained cluster federated learning model, and the distributed nodes obtain the local model of the trained cluster federated learning model.
[0269] In some embodiments of the present disclosure, the model training module 62 can be configured to receive local model parameters updated and reported by distributed nodes; aggregate the local model parameters of all distributed nodes in the training group to update global model parameters; and send the global model parameters to the distributed nodes, wherein the global model parameters are used by the distributed nodes to update the local model parameters of the distributed nodes.
[0270] In some embodiments of the present disclosure, the model training module 62 can also be configured to repeatedly perform the operations of receiving local model parameters updated and reported by distributed nodes, aggregating the local model parameters of all distributed nodes in the training group, updating global model parameters, and sending the global model parameters to the distributed nodes until the clustering federated learning model converges.
[0271] In some embodiments of the present disclosure, the training group includes a group identifier, and the group identifier includes at least one of a cell or cell group identifier and a training group identifier.
[0272] In some embodiments of the present disclosure, the central node of the present disclosure may also be configured to receive user equipment capabilities reported by the distributed nodes, wherein the user equipment capabilities are used to indicate at least one of the models and functions supported by the user equipment; and to identify at least one of the models and functions.
[0273] In some embodiments of the present disclosure, when the central node identifies at least one of a model and a function, it may be configured to identify at least one of the model and the function using at least one of an offline method and an online method.
[0274] In some embodiments of the present disclosure, the central node of the present disclosure may also be configured to transmit or hand over a global model to the distributed node, wherein the transmitted or handed over global model includes at least one of a first model and a second model, the first model being at least one clustered federated learning model of the cell or cell group where the distributed node is located obtained through model training, and the second model being at least one clustered federated learning model of the cell or cell group where the distributed node is not located obtained through model training.
[0275] In some embodiments of the present disclosure, the central node of the present disclosure may also be configured to determine at least one of the models and functions that matches the distributed node based on at least one of the scenarios, configurations, sites, and user equipment capabilities; and send model management configuration signaling for model reasoning related to model deployment to the distributed node, wherein the model management configuration signaling is used for the distributed node to select, activate, deactivate, switch, and fallback at least one of the models and functions.
[0276] In some embodiments of the present disclosure, the central node of the present disclosure may also be configured to receive monitoring output results of the clustering federated learning model reported by the distributed nodes.
[0277] In some embodiments of the present disclosure, the central node of the present disclosure may also be configured to determine whether to adjust the generated cluster federated learning model based on the monitoring output results and changes in at least one of the scenarios, configurations, sites, and user equipment capabilities, wherein the adjustment includes at least one of adding members to the training group, deleting members from the training group, retraining, and fine-tuning; and in the case where it is determined that the generated cluster federated learning model is to be adjusted, a model update configuration signaling is sent to the distributed nodes.
[0278] In some embodiments of the present disclosure, the central node of the present disclosure may be configured to execute any of the above embodiments of the present disclosure (for example Figures 2 to 3 At least one embodiment, Figure 5 The model lifecycle management method described in at least one of step 512, steps A and B, step 521, step 524, step 531 and step 542 in the embodiment.
[0279] Figure 7 Schematic diagram of the structure of some other embodiments of the central node of the present disclosure. Figure 7 As shown, the central node includes a memory 71 and a processor 72 .
[0280] The memory 71 is used to store instructions. The processor 72 is coupled to the memory 71. The processor 72 is configured to execute and implement any of the above embodiments (for example, Figures 2 to 3 At least one embodiment, Figure 5 The model lifecycle management method described in at least one of step 512, steps A and B, step 521, step 524, step 531 and step 542 in the embodiment.
[0281] like Figure 7 As shown, the central node also includes a communication interface 73 for information exchange with other devices. At the same time, the central node also includes a bus 74, through which the processor 72, the communication interface 73, and the memory 71 communicate with each other.
[0282] The memory 71 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory. The memory 71 may also be a memory array. The memory 71 may also be divided into blocks, and the blocks may be combined into virtual volumes according to certain rules.
[0283] In addition, the processor 72 may be a central processing unit (CPU), or may be an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present disclosure.
[0284] Figure 8 Schematic diagram of some embodiments of the distributed nodes disclosed in the present invention. Figure 8 As shown, the distributed node disclosed herein may include a data reporting module 81 and a distributed training module 82 .
[0285] The data reporting module 81 is configured to report the distributed node data to the central node, wherein the distributed node data is used for the central node to cluster all distributed nodes to form at least one training group, each training group including multiple distributed nodes.
[0286] The distributed training module 82 is configured to train the cluster federated learning model together with the central node corresponding to the training group in each training group to obtain a trained cluster federated learning model.
[0287] In some embodiments of the present disclosure, the trained cluster federated learning model can be used to perform model lifecycle management operations such as model transmission or handover, model deployment, model monitoring, model update, and model reasoning.
[0288] In some embodiments of the present disclosure, the distributed training module 82 can be configured to update the local model and obtain local model parameters; upload the local model parameters to the central node, wherein the local model parameters are used for the central node to aggregate the local model parameters of all distributed nodes in the training group and update the global model parameters; receive the global model parameters sent by the central node, wherein the global model parameters are used for the distributed nodes to update the local model parameters of the distributed nodes.
[0289] In some embodiments of the present disclosure, the distributed training module 82 can also be configured to repeatedly perform the operations of updating the local model, obtaining local model parameters, uploading the local model parameters to the central node, and receiving the global model parameters issued by the central node until the clustering federated learning model converges.
[0290] In some embodiments of the present disclosure, the distributed nodes of the present disclosure may also be configured to report user equipment capabilities to a central node, wherein the user equipment capabilities are used to indicate at least one of the models and functions supported by the user equipment, so that the central node can identify at least one of the models and functions.
[0291] In some embodiments of the present disclosure, when the distributed nodes of the present disclosure report the user equipment capabilities to the central node, they can be configured to report in the form of a model identifier or a model identifier list, wherein the supported model types include at least one of a generation vector machine, a neural network, and a random forest; and report in the form of a function identifier or a function identifier list, wherein the supported function types include at least one of classification, regression, clustering, dimensionality reduction, and association.
[0292] In some embodiments of the present disclosure, the distributed node of the present disclosure may also be configured to receive a global model transmitted or handed over by a central node, wherein the transmitted or handed over global model includes at least one of a first model and a second model, the first model being at least one clustered federated learning model of the cell or cell group where the distributed node is located obtained through model training, and the second model being at least one clustered federated learning model of the cell or cell group where the distributed node is not located obtained through model training.
[0293] In some embodiments of the present disclosure, the distributed nodes of the present disclosure may also be configured to determine, for features related to artificial intelligence, at least one of the functions and models that match the features related to artificial intelligence based on at least one of the scenarios, configurations, sites, and user equipment capabilities; perform at least one of selection, activation, deactivation, switching, and fallback on at least one of the determined functions and models; and feed back model management configuration signaling related to model deployment for model reasoning to the central node.
[0294] In some embodiments of the present disclosure, the distributed node of the present disclosure may also be configured to select and activate at least one model according to the model management configuration signaling for performing model reasoning.
[0295] In some embodiments of the present disclosure, the distributed nodes of the present disclosure may also be configured to monitor the cluster federated learning model to obtain a monitoring output result of the cluster federated learning model.
[0296] In some embodiments of the present disclosure, the distributed nodes of the present disclosure may also be configured to determine whether to adjust the generated cluster federated learning model based on the monitoring output results and changes in at least one of the scenarios, configurations, sites, and user equipment capabilities, wherein the adjustment includes at least one of adding members to the training group, deleting members from the training group, retraining, and fine-tuning; and in the case where it is determined that the generated cluster federated learning model is to be adjusted, a model update configuration signaling is sent to the central node.
[0297] In some embodiments of the present disclosure, the central node of the present disclosure may be configured to execute any of the above embodiments of the present disclosure (for example Figure 4 Example, Figure 5The model lifecycle management method described in at least one of step 511, steps A and B, steps 522-523, steps 532-534 and step 541 in the embodiment.
[0298] Fig. 9 Schematic diagram of the structure of some other embodiments of the distributed node disclosed in the present invention. Fig. 9 Embodiment Distributed Nodes and Figure 7 The structures of the central nodes of the embodiments are the same. Fig. 9 Embodiment Distributed Nodes and Figure 7 The difference between the central nodes of the embodiments is that: Fig. 9 In the embodiment, the processor 72 is configured to execute any of the above embodiments (for example, Figure 4 Example, Figure 5 The model lifecycle management method described in at least one of step 511, steps A and B, steps 522-523, steps 532-534 and step 541 in the embodiment.
[0299] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, wherein when the computer program is executed by a processor, the computer program implements any of the above embodiments (for example Figure 2-Figure 5 The model lifecycle management method described in any embodiment).
[0300] According to another aspect of the present disclosure, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and when the instructions are executed by a processor, any of the above embodiments (for example Figure 2-Figure 5 The model lifecycle management method described in any embodiment).
[0301] In some embodiments of the present disclosure, the computer-readable storage medium may be a non-transitory computer-readable storage medium.
[0302] Those skilled in the art will appreciate that the embodiments of the present disclosure may be provided as methods, devices, or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present disclosure may take the form of a computer program product implemented on one or more computer-usable non-transient storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0303] The present disclosure is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present disclosure. It should be understood that each process and / or block in the flowchart and / or block diagram and the combination of processes and / or blocks in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0304] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0305] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0306] The central node, grouping module, model training module, distributed nodes, data reporting module and distributed training module described above can be implemented as a general-purpose processor, a programmable logic controller (PLC), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component or any appropriate combination thereof for performing the functions described in this application.
[0307] So far, the present disclosure has been described in detail. In order to avoid obscuring the concept of the present disclosure, some details known in the art are not described. Based on the above description, those skilled in the art can fully understand how to implement the technical solution disclosed here.
[0308] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or may be accomplished by instructing the relevant hardware through a program, and the program may be stored in a non-transitory computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.
[0309] The description of the present disclosure is given for the purpose of illustration and description, and is not intended to be exhaustive or to limit the present disclosure to the disclosed form. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present disclosure, and to enable those of ordinary skill in the art to understand the present disclosure and thereby design various embodiments with various modifications suitable for specific uses.
Claims
1. A model lifecycle management method, executed by a central node, comprising: Clustering all distributed nodes to form at least one training group, wherein each training group includes a plurality of distributed nodes; In each training group, the cluster federated learning model is trained together with the multiple distributed nodes of the training group to obtain a trained cluster federated learning model.
2. The model lifecycle management method according to claim 1, wherein: The clustering of all distributed nodes to form at least one training group includes: Determine data distribution similarity of the distributed nodes according to the distributed node data, wherein the distributed node data includes at least one of the distance of the loss value after pre-training of the local data and the cosine similarity of the gradient; Based on the data distribution similarity, distributed node clustering is performed to form at least one training group.
3. The model lifecycle management method according to claim 2, wherein: The performing distributed node clustering based on the data distribution similarity to form at least one training group comprises: For each distributed node, determining a benefit function value of the distributed node according to the data distribution similarity of the distributed node; Determining a cost function value of the distributed node according to a transmission overhead of a local uploaded gradient of the distributed node; Determining a utility function value of the distributed node according to the benefit function value and the cost function value of the distributed node; According to the utility function value of each distributed node, distributed nodes are clustered to form at least one training group.
4. The model lifecycle management method according to claim 3, wherein: The step of clustering distributed nodes according to the utility function value of each distributed node to form at least one training group comprises: Given an initial grouping, traverse each distributed node and iteratively execute switching rules until the clustering result no longer changes, thereby forming a final distributed node grouping result, wherein the switching rules include: for each current training group as a target training group in turn, execute a switching operation, wherein the switching operation includes: when a distributed node attempts to join a target training group, if the utility function value of the distributed node itself becomes larger, and the utility function value of any distributed node in the target training group does not decrease, then the distributed node is added to the target training group, otherwise, the switching operation is repeated for the next target training group until the distributed node joins a target training group.
5. The model lifecycle management method according to any one of claims 1 to 4, wherein: The step of performing model training with the multiple distributed nodes of the training group in each training group to generate a cluster federated learning model includes: Receive local model parameters updated and reported by distributed nodes; Aggregate the local model parameters of all distributed nodes in the training group and update the global model parameters; The global model parameters are sent to the distributed nodes, wherein the global model parameters are used by the distributed nodes to update the local model parameters of the distributed nodes.
6. The model lifecycle management method according to claim 5, wherein: The step of performing model training with the plurality of distributed nodes in the training group in each training group to generate a cluster federated learning model further comprises: Repeat the steps of receiving local model parameters updated and reported by distributed nodes, aggregating the local model parameters of all distributed nodes in the training group, updating global model parameters, and sending the global model parameters to the distributed nodes until the clustering federated learning model converges.
7. The model lifecycle management method according to any one of claims 1 to 4, wherein: The training group includes a group identifier, and the group identifier includes at least one of a cell or cell group identifier and a training group identifier.
8. The model lifecycle management method according to any one of claims 1 to 4, further comprising: Receiving user equipment capabilities reported by the distributed node, wherein the user equipment capabilities are used to indicate at least one of a model and a function supported by the user equipment; At least one of a model and a function is identified.
9. The model lifecycle management method according to claim 8, wherein: The identifying of at least one of a model and a function comprises: At least one of the model and the function is identified by at least one of an offline method and an online method.
10. The model lifecycle management method according to any one of claims 1 to 4, further comprising: A global model is transmitted or handed over to the distributed node, wherein the transmitted or handed over global model includes at least one of a first model and a second model, the first model is at least one clustered federated learning model of the cell or cell group where the distributed node is located obtained through model training, and the second model is at least one clustered federated learning model of the cell or cell group where the distributed node is not located obtained through model training.
11. The model lifecycle management method according to any one of claims 1 to 4, further comprising: Determining at least one of a model and a function matching the distributed node based on at least one of a scenario, a configuration, a site, and a user equipment capability; Model management configuration signaling for model reasoning related to model deployment is sent to the distributed node, wherein the model management configuration signaling is used for the distributed node to select, activate, deactivate, switch, and rollback at least one of the model and function.
12. The model lifecycle management method according to any one of claims 1 to 4, further comprising: Receive the monitoring output result of the clustering federated learning model reported by the distributed node.
13. The model lifecycle management method according to claim 12, further comprising: Determine whether to adjust the generated cluster federated learning model according to the monitoring output result and the change of at least one of the scenario, configuration, site and user equipment capability, wherein the adjustment includes at least one of adding members to the training group, deleting members from the training group, retraining and fine-tuning; When it is determined that the generated cluster federated learning model needs to be adjusted, a model update configuration signaling is sent to the distributed nodes.
14. A model lifecycle management method, executed by distributed nodes, comprising: Reporting the distributed node data to the central node, wherein the distributed node data is used for the central node to cluster all the distributed nodes to form at least one training group, each training group including a plurality of distributed nodes; In each training group, the central node corresponding to the training group jointly trains the cluster federated learning model to obtain a trained cluster federated learning model.
15. The model lifecycle management method according to claim 14, wherein: In each training group, the central node corresponding to the training group performs model training to generate a cluster federated learning model, including: Update the local model and obtain local model parameters; Uploading the local model parameters to the central node, wherein the local model parameters are used for the central node to aggregate the local model parameters of all distributed nodes in the training group and update the global model parameters; The global model parameters sent by the central node are received, wherein the global model parameters are used by the distributed nodes to update the local model parameters of the distributed nodes.
16. The model lifecycle management method according to claim 15, wherein: In each training group, the central node corresponding to the training group performs model training to generate a cluster federated learning model, and further includes: Repeat the steps of updating the local model, obtaining local model parameters, uploading the local model parameters to the central node, and receiving the global model parameters issued by the central node until the clustering federated learning model converges.
17. The model lifecycle management method according to any one of claims 14 to 16, further comprising: Reporting user equipment capabilities to the central node, wherein the user equipment capabilities are used to indicate at least one of the models and functions supported by the user equipment, so that the central node can identify at least one of the models and functions.
18. The model lifecycle management method according to claim 17, wherein: Reporting the user equipment capability to the central node includes at least one of the following steps, wherein: The model is reported in the form of a model identifier or a model identifier list, wherein the supported model types include at least one of a generative vector machine, a neural network, and a random forest; The function is reported in the form of a function identifier or a function identifier list, wherein the supported function types include at least one of classification, regression, clustering, dimensionality reduction and association.
19. The model lifecycle management method according to any one of claims 14 to 16, further comprising: A global model transmitted or handed over by a receiving central node, wherein the transmitted or handed over global model includes at least one of a first model and a second model, the first model being at least one clustered federated learning model of a cell or cell group where a distributed node is located obtained through model training, and the second model being at least one clustered federated learning model of a cell or cell group where a distributed node is not located obtained through model training.
20. The model lifecycle management method according to any one of claims 14 to 16, further comprising: Based on at least one of the scenario, configuration, site, and user device capability, for the artificial intelligence-related feature, determining at least one of a function and a model that matches the artificial intelligence-related feature; For at least one of the determined functions and models, perform at least one of selection, activation, deactivation, switching, and rollback; Feedback model management configuration signaling related to model deployment for model reasoning to the central node.
21. The model lifecycle management method according to claim 20, further comprising: At least one model is selected and activated according to the model management configuration signaling for performing model reasoning.
22. The model lifecycle management method according to any one of claims 1 to 4, further comprising: The cluster federated learning model is monitored to obtain a monitoring output result of the cluster federated learning model.
23. The model lifecycle management method according to claim 22, further comprising: Determine whether to adjust the generated cluster federated learning model according to the monitoring output result and the change of at least one of the scenario, configuration, site and user equipment capability, wherein the adjustment includes at least one of adding members to the training group, deleting members from the training group, retraining and fine-tuning; When it is determined that the generated cluster federated learning model needs to be adjusted, a model update configuration signaling is sent to the central node.
24. A central node, comprising: A grouping module is configured to cluster all distributed nodes to form at least one training group, wherein each training group includes a plurality of distributed nodes; The model training module is configured to train the cluster federated learning model together with the multiple distributed nodes in the training group in each training group to obtain a trained cluster federated learning model.
25. A central node, comprising: a memory configured to store instructions; The processor is configured to execute the instructions so that the central node executes the model lifecycle management method as described in any one of claims 1-13.
26. A distributed node, comprising: A data reporting module is configured to report distributed node data to a central node, wherein the distributed node data is used for the central node to cluster all distributed nodes to form at least one training group, each training group including a plurality of distributed nodes; The distributed training module is configured to train the cluster federated learning model together with the central node corresponding to the training group in each training group to obtain a trained cluster federated learning model.
27. A distributed node, comprising: a memory configured to store instructions; The processor is configured to execute the instructions so that the distributed nodes perform the model lifecycle management method as described in any one of claims 14-23.
28. A model lifecycle management system, comprising a central node as claimed in claim 24 or 25 and a distributed node as claimed in claim 26 or 27.
29. A computer-readable storage medium, wherein: The computer-readable storage medium stores computer instructions, and when the instructions are executed by a processor, the model lifecycle management method according to any one of claims 1 to 23 is implemented.
30. A computer program product comprising a computer program, wherein: When the computer program is executed by a processor, the model lifecycle management method according to any one of claims 1 to 23 is implemented.
Citation Information
Patent Citations
Data isomerism federal learning method based on multi-level intelligent connection and storage medium
CN117875454A
Method, device and equipment for managing model
CN117997770A
Method and system for configuring neural networks of a set of nodes in a communication network
FR3126575A1
Multi-layer federated learning system and methodology based on distributed clustering
TWI807961B
