Model training method and device and processor readable storage medium

Abstract

The invention relates to a model training method and device and a processor readable storage medium, and a center node carries out model training in cooperation with a first distributed node according to node data of the first distributed node, so that the first distributed node can complete a model training task under the cooperative help of other nodes. The effects of improving the network performance of the whole distributed network and optimizing the network and application service experience are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication, in particular to a model training method and device and processor readable storage medium. BACKGROUND

[0002] With the development of artificial intelligence technology in the field of mobile communication, computing power, data and model are presented as the three elements of artificial intelligence.

[0003] Taking a distributed mobile communication network scenario as an example, a large computing power node located at a center node may not be able to access a node with massive data located at an edge in time, and the node with massive data located at the edge does not have computing power to convert data into a model, resulting in a problem that a single node cannot complete a model training task under a distributed network. SUMMARY

[0004] Therefore, it is necessary to provide a model training method, device and processor readable storage medium to enable a single node to complete a model training task under a distributed network.

[0005] In a first aspect, an embodiment of the present application provides a model training method applied to an intelligent network element of a center node, and the method comprises:

[0006] cooperating with the first distributed node to perform model training according to node data of the first distributed node; the first distributed node is any one of distributed nodes in a distributed network.

[0007] In one of the embodiments, before cooperating with the first distributed node to perform model training according to node data of the first distributed node, the method comprises:

[0008] receiving node data sent by an intelligent network element of the first distributed node.

[0009] In one of the embodiments, before cooperating with the first distributed node to perform model training according to node data of the first distributed node, the method comprises:

[0010] sending a data reporting request to an intelligent network element of the first distributed node, the data reporting request being used to instruct the intelligent network element of the first distributed node to report node data;

[0011] receiving node data sent by an intelligent network element of the first distributed node.

[0012] In one of the embodiments, the node data comprises:

[0013] a local model and parameters trained by the first distributed node, and node information of the first distributed node; or

[0014] The first distributed node trains the local model and parameters, the node information of the first distributed node, the data that the first distributed node cannot train, and the collaborative training request.

[0015] In one of the embodiments, the model training of the first distributed node is performed collaboratively according to the node data, including:

[0016] The local model of the first distributed node is updated according to the node data.

[0017] In one of the embodiments, the node data includes the local model and parameters trained by the first distributed node, and the node information of the first distributed node; the local model of the first distributed node is updated according to the node data, including:

[0018] Based on the local model and parameters, the node information, and the global model of the distributed network, the local model update parameters of the first distributed node are determined;

[0019] The local model of the first distributed node is updated according to the local model update parameters.

[0020] In one of the embodiments, the first distributed node supports model update training; the local model of the first distributed node is updated according to the local model update parameters, including:

[0021] The local model update parameters are sent to the intelligent network element of the first distributed node, instructing the intelligent network element of the first distributed node to update and train the local model.

[0022] In one of the embodiments, the first distributed node does not support model update training; the local model of the first distributed node is updated according to the node data, including:

[0023] The local model is updated and trained according to the local model update parameters;

[0024] The model obtained by the update and training is sent to the intelligent network element of the first distributed node, and model training instruction information is sent, the model training instruction information being used to instruct that the subsequent model training and update of the first distributed node are performed at the center node.

[0025] In one of the embodiments, the method further includes:

[0026] The local models reported by each distributed node in the distributed network are obtained;

[0027] The local models reported by each distributed node are subjected to a weighted aggregation operation to obtain the global model of the distributed network.

[0028] In one of the embodiments, the method further includes:

[0029] determine whether the global model converges;

[0030] In the case that the global model does not converge, the process of training the model in cooperation with the first distributed node is cyclically performed until the latest obtained global model converges.

[0031] In one of the embodiments, the method of training the model in cooperation with the first distributed node according to the node data further comprises:

[0032] training the model according to the node data on the data that the first distributed node is unable to train.

[0033] In one of the embodiments, the node data further comprises the data that the first distributed node is unable to train and the cooperative training request; and the method of training the model according to the node data on the data that the first distributed node is unable to train comprises:

[0034] in response to the cooperative training request, cooperatively training the model on the data that is unable to train.

[0035] In one of the embodiments, the first distributed node supports model update training; and the method further comprises:

[0036] integrating the cooperatively trained model and the local model;

[0037] sending the integrated model and an update indication to the intelligent network element of the first distributed node, instructing the intelligent network element of the first distributed node to update the integrated model.

[0038] In one of the embodiments, the first distributed node does not support model update training; and the method further comprises:

[0039] integrating the cooperatively trained model and the local model, and updating the integrated model;

[0040] sending the integrated and updated model and model training indication information to the intelligent network element of the first distributed node, the model training indication information being used to instruct the first distributed node to perform subsequent model training and update at the center node.

[0041] In one of the embodiments, the node data further comprises the data that the first distributed node is unable to train and the cooperative training request; and the method of training the model according to the node data on the data that the first distributed node is unable to train comprises:

[0042] in response to the cooperative training request, determining a second distributed node according to the node information of the other distributed nodes; the second distributed node representing a distributed node capable of cooperatively training the model on the data that is unable to train.

[0043] In one of the embodiments, the method further comprises:

[0044] sending, to an intelligent network element of the first distributed node, identification information of the second distributed node.

[0045] In one of the embodiments, the method further includes:

[0046] sending, to an intelligent network element of the second distributed node, the untrainable data, the node information of the first distributed node, the local model and the parameters.

[0047] In one of the embodiments, the method further includes:

[0048] sending, to an intelligent network element of the first distributed node, a data sending notification; the data sending notification is used to instruct the intelligent network element of the first distributed node to send, to the second distributed node, the untrainable data, the node information of the first distributed node, the local model and the parameters according to the identification information.

[0049] In one of the embodiments, before the model training on the untrainable data of the first distributed node according to the node data, the method further includes:

[0050] determining, according to the computing power information of the first distributed node, whether the first distributed node can complete the model training.

[0051] in the case that the first distributed node cannot complete the model training, requesting the untrainable data of the first distributed node from the first distributed node.

[0052] In a second aspect, the embodiments of the present application provide a model training method, applied to an intelligent network element of a first distributed node, and the method includes:

[0053] sending, to an intelligent network element of a center node, node data; the node data is used for the intelligent network element of the center node to cooperatively perform model training of the first distributed node; the first distributed node is any one of distributed nodes in a distributed network.

[0054] In one of the embodiments, the sending of the node data to the intelligent network element of the center node includes:

[0055] receiving a data reporting request sent by the intelligent network element of the center node;

[0056] sending, to the intelligent network element of the center node, the node data according to the data reporting request.

[0057] In one of the embodiments, the node data includes:

[0058] the local model and the parameters trained by the first distributed node, and the node information of the first distributed node; or

[0059] The first distributed node trains a local model and parameters, node information of the first distributed node, data that the first distributed node cannot train, and a collaborative training request.

[0060] In one of the embodiments, the method further comprises:

[0061] Collecting model training related data in a distributed network;

[0062] Training a local model according to the related data.

[0063] In one of the embodiments, collecting the related data in the distributed network comprises:

[0064] Sending a data collection request to each network function in a core network and each network function in an access network in the first distributed node;

[0065] Receiving response data sent by each network function to obtain model training related data.

[0066] In one of the embodiments, collecting the model training related data in the distributed network comprises:

[0067] Sending a data collection request to each data plane in the first distributed node;

[0068] Receiving response data sent by each data plane to obtain model training related data.

[0069] In one of the embodiments, the method further comprises:

[0070] Receiving a local model update message sent by an intelligent network element of the center node.

[0071] In one of the embodiments, the local model update message comprises local model update parameters, and the method further comprises:

[0072] Updating and training the local model according to the local model update parameters.

[0073] In one of the embodiments, the local model update message comprises an updated local model and model training indication information, and the model training indication information is used to indicate that subsequent model training and updating of the first distributed node are both performed at the center node.

[0074] In one of the embodiments, the method further comprises:

[0075] Receiving an integrated model and update indication sent by an intelligent network element of the center node; the integrated model is obtained by integrating a collaborative model and a local model by the intelligent network element of the center node; the collaborative model is trained according to data that the first distributed node cannot train;

[0076] updating the integrated model.

[0077] In one of the embodiments, the method further comprises:

[0078] receiving the integrated model and the model training indication information from the intelligent network element of the center node, the model training indication information indicating that the subsequent model training and updating of the first distributed node are to be performed at the center node.

[0079] In one of the embodiments, the method further comprises:

[0080] receiving the identification information of the second distributed node from the intelligent network element of the center node, the second distributed node representing a node for training the collaborative model according to the data that the first distributed node is unable to train.

[0081] In one of the embodiments, the method further comprises:

[0082] in response to the data sending notification from the intelligent network element of the center node, sending the data that the first distributed node is unable to train, the node information of the first distributed node, the local model and the parameters to the intelligent network element of the second distributed node according to the identification information.

[0083] In one of the embodiments, the method further comprises:

[0084] receiving the integrated model and the updating indication from the intelligent network element of the second distributed node, the integrated model being obtained by the intelligent network element of the second distributed node by integrating the collaborative model and the local model;

[0085] updating the integrated model.

[0086] In one of the embodiments, the method further comprises:

[0087] receiving the integrated model and the model training indication information from the intelligent network element of the second distributed node, the model training indication information indicating that the subsequent model training and updating of the first distributed node are to be performed at the second distributed node.

[0088] In one of the embodiments, before receiving the local model updating parameters from the intelligent network element of the center node, the method further comprises:

[0089] determining the second distributed node based on the pre-maintained network node neighbor information.

[0090] sending, to an intelligent network element of the second distributed node, untrainable data of the first distributed node, node information of the first distributed node, a local model and parameters trained by the first distributed node, and a collaborative training request to instruct the second distributed node to perform model training on the untrainable data of the first distributed node.

[0091] In one of the embodiments, the method further includes:

[0092] If a rejection of the collaborative request sent by the intelligent network element of the second distributed node is received, a new second distributed node is determined based on pre-maintained network node neighbor information.

[0093] In a third aspect, an embodiment of the present application provides a model training method applied to an intelligent network element of a second distributed node, and the method includes:

[0094] receiving node data of the first distributed node; the first distributed node being any one of the distributed nodes in the distributed network;

[0095] collaboratively performing model training of the first distributed node according to the node data.

[0096] In one of the embodiments, the receiving of the node data of the first distributed node includes:

[0097] receiving node data sent by an intelligent network element of a center node; or

[0098] receiving node data sent by an intelligent network element of the first distributed node.

[0099] In one of the embodiments, the node data includes untrainable data of the first distributed node, node information of the first distributed node, and a local model and parameters trained by the first distributed node.

[0100] In one of the embodiments, the collaboratively performing of the model training of the first distributed node according to the node data of the first distributed node includes:

[0101] performing collaborative model training on the untrainable data.

[0102] In one of the embodiments, the first distributed node supports update training, and the method further includes:

[0103] sending, to the intelligent network element of the first distributed node, the integrated model and an update instruction to instruct the intelligent network element of the first distributed node to update the integrated model.

[0104] In one of the embodiments, the first distributed node does not support update training, and the method further includes:

[0105] Integrate the cooperative model and the local model, and update the integrated model;

[0106] Send the integrated and updated model and model training indication information to the intelligent network element of the first distributed node, the model training indication information being used to indicate that subsequent model training and updating of the first distributed node are both to be performed at the second distributed node.

[0107] In one of the embodiments, the method further includes:

[0108] In response to the cooperative training request sent by the intelligent network element of the first distributed node, send a rejection cooperative request to the intelligent network element of the first distributed node, the rejection cooperative request being used to instruct the first distributed node to determine a new second distributed node based on the pre-maintained network node neighbor information again.

[0109] In a fourth aspect, an embodiment of the present application provides a model training device, which includes:

[0110] The first training module is configured to perform model training in cooperation with the first distributed node according to node data of the first distributed node; the first distributed node is any one of the distributed nodes in the distributed network.

[0111] In a fifth aspect, an embodiment of the present application provides a model training device, which includes:

[0112] The node data sending module is configured to send node data to the intelligent network element of the center node; the node data is used for the intelligent network element of the center node to perform model training in cooperation with the first distributed node; the first distributed node is any one of the distributed nodes in the distributed network.

[0113] In a sixth aspect, an embodiment of the present application provides a model training device, which includes:

[0114] The data receiving module is configured to receive node data of the first distributed node; the first distributed node is any one of the distributed nodes in the distributed network.

[0115] The second training module is configured to perform model training in cooperation with the first distributed node according to the node data.

[0116] In a seventh aspect, an embodiment of the present application provides a processor readable storage medium, which stores a program, the program being used to make a processor execute the steps of the model training method provided in any one of the first aspect to the third aspect.

[0117] The model training method, device and processor readable storage medium provided by the embodiments of the present application can be used for the center node to perform model training in cooperation with the first distributed node according to the node data of the first distributed node, so that the first distributed node can better complete the model training task with the cooperation of other nodes. Moreover, since the first distributed node represents any distributed node in the distributed network, it is equivalent to that the center node can perform model training in cooperation with any distributed node in the distributed network according to the node data of the distributed node. In this way, the model training task that cannot be completed by each distributed node can be completed under the cooperation of the center node, so that the model training of the entire distributed network can achieve global optimization, thereby achieving the effect of improving the network performance of the entire distributed network, optimizing the network and the application service experience. BRIEF DESCRIPTION OF DRAWINGS

[0118] Figure 1 A network function schematic diagram of the 6G network architecture provided in an embodiment;

[0119] Figure 2 An application scenario schematic diagram of the model training method provided in an embodiment;

[0120] Figure 3 A flow schematic diagram of the model training method provided in an embodiment;

[0121] Figure 4 A flow schematic diagram of the model training method provided in another embodiment;

[0122] Figure 5 A flow schematic diagram of the model training method provided in another embodiment;

[0123] Figure 6 A flow schematic diagram of the model training method provided in another embodiment;

[0124] Figure 7 An interaction schematic diagram of the model training method provided in an embodiment;

[0125] Figure 8 An interaction schematic diagram of the model training method provided in another embodiment;

[0126] Figure 9 An interaction schematic diagram of the model training method provided in another embodiment;

[0127] Figure 10 An interaction schematic diagram of the model training method provided in another embodiment;

[0128] Figure 11 An interaction schematic diagram of the model training method provided in another embodiment;

[0129] Figure 12 An interaction schematic diagram of the model training method provided in another embodiment;

[0130] Figure 13 An interaction schematic diagram of the model training method provided in another embodiment;

[0131] Figure 14 An interaction schematic diagram of the model training method provided in another embodiment;

[0132] Figure 15 An interaction schematic diagram of the model training method provided in another embodiment;

[0133] Figure 16 A structural block diagram of the model training apparatus provided in an embodiment. DETAILED DESCRIPTION

[0134] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0135] In the embodiments of the present application, the character " / " generally represents that the associated objects before and after it are in an "or" relationship. The term "multiple" means two or more, and other quantifiers are similar.

[0136] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0137] The technical background of the present application will be described first as follows.

[0138] With the continuous evolution of communication technology, 1G, 2G, 3G, 4G, up to the current 5G and the future 6G, each generation of communication standard and each specific system has its own network architecture. Among them, the network architecture of 5G and previous generations is originally designed to be centrally controlled, that is, the core network function is centrally deployed based on the requirement of centrally deploying the core network machine room. However, with the rapid development of the network, the dual driving force of business scenarios and technology development makes 6G face the multi-scenario and network demand of air, sky, land and sea, and the centralized network architecture cannot meet all scenarios.

[0139] Based on this, the 6G architecture will be a new network architecture that is a fusion of centralized control mobile communication networks and open Internet, and a coexistence of centralized and distributed networks. It is necessary to go beyond centralized control, which gradually evolves to a distributed architecture, expands more network functions (such as authentication and authorization) to the network edge, and establishes distributed distributed homogeneous micro cloud units with different function levels. Each micro cloud unit is self-contained and has all the functions of control and data forwarding. Figure 1 A schematic diagram of a 6G network architecture is shown. Multiple micro cloud units can form an autonomous micro network according to business requirements, and provide network services according to specific business scenarios, user scale, geographic environment, and other requirements. These micro cloud units are also referred to as distributed network nodes. These distributed network nodes are micro cloud units deployed at the network edge, or edge networks, representing a customized or fully functional core network, and these distributed network nodes will be managed and controlled by a central node. In this way, the 6G network establishes distributed network nodes with different function levels and complete functions. Network nodes form autonomous micro networks according to business requirements, and provide network services according to specific business scenarios, user scale, geographic environment, and other requirements.

[0140] In addition, in the mobile communication network 2030 architecture, the 6G network will build native intelligence with artificial intelligence (AI) business-oriented perception capabilities. In the development process of AI technology, computing power, data, and models are presented as the three elements of AI. With the increasing computing power of smartphones and Internet of Things devices, users are increasingly concerned about data privacy, which makes it difficult for a single node in a distributed network scenario to access a node with massive data at the edge in a timely manner, and a node with massive data at the edge does not have the computing power to convert data into a model, resulting in a single node that cannot complete the model training task because it does not fully meet the three elements.

[0141] Therefore, how a single node in a distributed network completes the model training task to achieve global optimization of the model training and improve network performance, optimize network and application service experience becomes a technical problem to be solved. To this end, the present application embodiment provides a model training method, device and processor readable storage medium, which can solve the above problems. Of course, the technical solutions provided in the present application embodiment are not limited to solving only the above problems, but also have other technical effects. For details, please refer to the following embodiment description.

[0142] Before the present application embodiment is described, the application scenario of the present application embodiment is described.

[0143] The model training method provided in the embodiments of the present application can be applied to, but is not limited to, 5G, 6G and future evolved network architectures and the like, for example, as shown in Figure 2 The application scenario of the 6G network structure is taken as an example. The application scenario includes a distributed network composed of a center node and a plurality of distributed nodes. In the distributed network, the center node and the plurality of distributed nodes can communicate with each other in pairs, and the center node has strong computing power and resources, and can have the ability and function of managing the distributed nodes in the distributed network.

[0144] In addition, in the distributed network provided in the embodiments of the present application, intelligent network elements are deployed on the center node and each distributed node, Figure 2 The intelligent network element is not shown in the center node), which can be used to execute the steps in the model training method in the embodiments of the present application.

[0145] In the present application, embodiments are provided in which the intelligent network element deployed on the center node in the distributed network, the intelligent network element deployed on the first distributed node, and the intelligent network element deployed on the second distributed node are the execution subjects. The first distributed node is any one of the distributed nodes in the distributed network, and the second distributed node represents any one of the distributed nodes other than the first distributed node.

[0146] The embodiment in which the intelligent network element of the center node is the execution subject will be described first.

[0147] As shown in Figure 3 In one embodiment, a model training method is provided, which can include the following steps:

[0148] S101, according to the node data of the first distributed node, the model training is performed in cooperation with the first distributed node; the first distributed node is any one of the distributed nodes in the distributed network.

[0149] In the embodiments of the present application, node data generally refers to data related to the first distributed node, which can include data related to the node itself of the first distributed node, data related to the model on the first distributed node, data on some devices deployed on the first distributed node, or data generated by devices that do not deploy on the first distributed node but interact with the first distributed node, and the like.

[0150] The node itself related data can be node information, which in one embodiment includes but is not limited to the following contents: load information, identification information, computing power information, location information.

[0151] The load information represents load-related information of an upper bearer of the first distributed node. The load of the bearer includes a load that has been borne and a load that is expected to be borne in the future, etc. The load-related information can include various parameters of the load, such as size, usage rate, working voltage, working current, working time, and size of required various processing resources, etc. The identification information is information for identifying and distinguishing the first distributed node from other distributed nodes, which can be set by using numbers, letters, or a combination of numbers and letters, and embodiments of the present application are not limited in this regard. The computing power information represents information of the processing capability of the first distributed node for information, which can include information computing capability, network carrying capability, data storage capability, etc., which can be estimated based on available processor resources, network resources, storage resources of storage devices, etc. The location information represents the location of the first distributed node, for example, a location in the distributed network or a geographic location coordinate in a world coordinate system. The location information can be determined by the central node or the first distributed node based on various positioning methods, and embodiments of the present application are not limited in this regard.

[0152] The model-related data of the first distributed node can include a model trained by the first distributed node in the node itself, and parameters of the model, wherein the parameters of the model can be model gradients, model test data, or the model itself. Of course, the model-related data can also have data that the first distributed node cannot train, for example, after the first distributed node completes the training of the model in the node itself, the model parameters and computing power resources in the node itself cannot support the training of the data, and this part of the data that cannot support the training can be some non-private data.

[0153] Based on the node data of the first distributed node, the intelligent network element in the central node cooperates with the first distributed node to perform model training, so that the first distributed node can better complete the model training task with the help of other nodes. There are three cases of cooperation between the central node and the first distributed node for model training:

[0154] The first case is that the first distributed node has the ability to complete the local model training task, in which case the central node can guide and update the model trained by the first distributed node to make the model training of the first distributed node reach the global optimum.

[0155] The second case is that the first distributed node does not have the ability to complete the local model training task, in which case the central node cooperates with the first distributed node to complete the model training of the data that cannot be trained in the node of the first distributed node, and cooperates with the first distributed node to update the trained model to make the model training of the first distributed node reach the global optimum.

[0156] The third case is that the first distributed node does not have the ability to complete the local model training task. In this case, after the first distributed node completes the training of the model that can be completed in the node, the center node determines a second distributed node from the distributed network to help the first distributed node complete the model training of the data that cannot be trained in the node, and cooperates to update the model of the first distributed node that is completed, so that the model training of the first distributed node reaches the global optimum.

[0157] The second distributed node can be any distributed node in the distributed network except the first distributed node. Of course, for the case that the first distributed node does not have the ability to complete the local model training task, in addition to the center node actively determining a second distributed node to help the first distributed node to train the model, the first distributed node can also actively determine a second distributed node from other distributed nodes to help itself to train the model.

[0158] In the embodiment of the application, the center node cooperates with the first distributed node to train the model according to the node data of the first distributed node, so that the first distributed node can better complete the model training task with the help of other nodes. And since the first distributed node represents any distributed node in the distributed network, it is equivalent to that the center node can cooperate with any distributed node to train the model according to its node data when the distributed node trains the model. Thus, the model training task that cannot be completed by each distributed node can be completed under the cooperation of the center node, so that the model training of the entire distributed network can reach the global optimum, thereby realizing the effect of improving the network performance, optimizing the network and application service experience of the entire distributed network.

[0159] Each distributed node in the distributed network can train the model. For example, taking the first distributed node as a representative of any distributed node, the intelligent network element of the first distributed node can first collect model training related data in the distributed network when performing the local model training task, and then train the model in the node according to the model training related data.

[0160] The model training related data can be data from the distributed network, including network function (Network Function, NF) data, UE data, node information, and data required for network operation and management, including configuration, performance, log, alarm information, etc. The NF data can include NF load, NF performance configuration, etc. The node information can include node resource information, computing power information, and identification information, etc. The UE data can include service data and sensing data, wherein the sensing data includes raw data, sensing measurement data, preprocessed data, and sensing results, etc.

[0161] Optionally, the intelligent network element of the first distributed node can collect the model training related data in the distributed network in the following manner: the intelligent network element of the first distributed node sends a data collection request to each network function in the core network and each network function in the access network in the first distributed node, and then receives the response data sent by each network function to obtain the model training related data.

[0162] Optionally, the intelligent network element of the first distributed node can collect the model training related data in the distributed network in the following manner: the intelligent network element of the first distributed node sends a data collection request to each data plane in the first distributed node, and then receives the response data sent by each data plane to obtain the model training related data.

[0163] The intelligent network element of the first distributed node performs model training according to the model training related data in the node. As mentioned in the foregoing embodiments, the intelligent network element of the first distributed node can have the ability to complete the local model training task, or can not have the ability to complete the local model training task. Therefore, the first distributed node can complete the entire local model training task, or can only complete a part of the local model training task. For example, the local model after completing the entire local model training task has completed 100% of the model training task, and the local model after only completing a part of the local model training task has completed 50% of the model training task.

[0164] When the intelligent network element of the first distributed node completes the entire local model training task, the center node needs to cooperate with the model guidance update. When the intelligent network element of the first distributed node only completes a part of the local model training task, the center node needs to cooperate with the training of the other part of the model training task that cannot be completed and the model guidance update. Therefore, the contents included in the node data can be different in the two cases, as described below:

[0165] In an embodiment, the node data includes: the trained local model and parameters of the first distributed node, and the node information of the first distributed node; or the trained local model and parameters of the first distributed node, the node information of the first distributed node, the untrainable data of the first distributed node, and the collaborative training request.

[0166] In the embodiments of the present application, the model trained by the first distributed node in the node itself is referred to as the local model of the first distributed node. Then:

[0167] In one case, the node data only includes the trained local model and parameters of the first distributed node, and the node information of the first distributed node, indicating that the first distributed node has completed all local model training tasks.

[0168] In another case, the node data includes the trained local model and parameters of the first distributed node, the node information of the first distributed node, and the untrainable data of the first distributed node, indicating that the first distributed node has only completed part of the local model training task, and the other part of the model training task has not been completed. The data corresponding to this part of the uncompleted model training task is the untrainable data of the first distributed node. It should be noted that if the untrainable data in the first distributed node is all private data, it cannot be reported to the center node, but needs to be negotiated with the node administrator to add computing power equipment to the first distributed node.

[0169] In addition, in the embodiments of the present application, the center node intelligent network element collaborates with the first distributed node to perform model training according to the node data of the first distributed node, indicating that the intelligent network element of the center node has obtained the node data of the first distributed node at this moment, so if the node data includes the untrainable data of the first distributed node, it can also include the collaborative training request sent by the first distributed node.

[0170] In the embodiments of the present application, the process of the center node obtaining the node data of the first distributed node can be completed at any time point before the intelligent network element of the center node collaborates with the first distributed node to perform model training according to the node data of the first distributed node, which is not limited in the embodiments of the present application.

[0171] For the way in which the center node obtains the node data of the first distributed node, the embodiments of the present application provide two implementable ways, which are as follows:

[0172] The first way is that the first distributed node actively sends the node data to the center node. In an embodiment, the model training method of the present application further includes: the intelligent network element of the center node receives the node data sent by the intelligent network element of the first distributed node.

[0173] In this mode, the intelligent network element of the first distributed node actively sends node data to the intelligent network element of the central node. The intelligent network element of the first distributed node can actively send its node data to the central node in a case where the first distributed node needs the central node to cooperate with it to perform model training, so that the central node can learn the requirement for cooperating to train the model and the related information for cooperating with the first distributed node to perform model training.

[0174] In another mode, the central node actively finds the first distributed node to report node data of the first distributed node. In an embodiment, the model training method further includes: the intelligent network element of the central node sends a data reporting request to the intelligent network element of the first distributed node, and receives node data sent by the intelligent network element of the first distributed node; the data reporting request is used for the intelligent network element of the central node to instruct the intelligent network element of the first distributed node to report node data.

[0175] In this mode, the intelligent network element of the central node can periodically send a data reporting request to the intelligent network element of the first distributed node according to a preconfigured frequency, to instruct the intelligent network element of the first distributed node to report node data of the first distributed node. Alternatively, the intelligent network element of the central node can evaluate and calculate the computing power of each distributed node in the distributed network, and determine that the first distributed node does not have the capability to complete model training, and then the intelligent network element of the central node sends a data reporting request to the intelligent network element of the first distributed node to obtain node data of the intelligent network element of the first distributed node, to facilitate the central node to cooperate with the first distributed node to perform model training.

[0176] In the above two modes of obtaining node data of the first distributed node, no matter which mode is used, if the node data obtained by the central node includes a local model and parameters trained by the first distributed node, and node information of the first distributed node, it means that the central node needs to cooperate with the first distributed node to update the local model of the first distributed node. If the node data obtained by the central node includes a local model and parameters trained by the first distributed node, node information of the first distributed node, and data that cannot be trained by the first distributed node and a request for cooperative training, it means that the central node needs to cooperate with the first distributed node to train a model based on the data that cannot be trained by the first distributed node, and then guide the update of the trained model. For the specific cooperation process, refer to the detailed description of the subsequent embodiments. It is only emphasized that the mode of obtaining node data of the first distributed node by the central node does not limit the specific content of the obtained node data.

[0177] In the embodiments of the present application, the center node obtains the node data of the first distributed node in two ways: the first distributed node actively reports or the center node requests reporting. The flexibility of the way the center node obtains the node data of the first distributed node is improved, and because the purpose of the center node obtaining the node data of the first distributed node is to coordinate the first distributed node to perform model training, the center node can request the first distributed node to report when the first distributed node does not actively report, which can ensure that the first distributed node completes the model training task and improves the timeliness and reliability of the model training task completed by the first distributed node.

[0178] The detailed process of the center node coordinating the first distributed node to perform model training will be described below through specific embodiments.

[0179] In one embodiment, the model training of the first distributed node in S101 according to the node data of the first distributed node includes that the intelligent network element of the center node updates the local model of the first distributed node according to the node data.

[0180] That is, the coordination of the center node to the first distributed node can be only to coordinate the first distributed node to perform the local model update process of the first distributed node. In this case, the center node can obtain the local model and parameters trained by the first distributed node and the node information of the first distributed node in the node data. Then, as shown in Figure 4 The process of the intelligent network element of the center node updating the local model of the first distributed node according to the node data includes the following steps:

[0181] S201, based on the local model and parameters, the node information, and the global model of the distributed network, determining the local model update parameters of the first distributed node.

[0182] The local model and parameters refer to the local model and parameters trained by the first distributed node included in the node data. Because the center node is to coordinate the first distributed node to perform model update, the local model trained by the first distributed node here refers to the complete local model after training, that is, the local model after 100% training.

[0183] The global model of the distributed network refers to the large model corresponding to all distributed nodes in the entire distributed network, and the global model is established and maintained by the center node.

[0184] In one embodiment, the process of the center node establishing the global model includes that the intelligent network element of the center node obtains the local model reported by each distributed node in the distributed network, performs a weighted aggregation operation according to the local model reported by each distributed node, and obtains the global model.

[0185] Each distributed node in the distributed network can have a local model, and the intelligent network element of the center node obtains the local models, and based on the weights pre-configured for each distributed node, performs a weighted aggregation operation on the local models of all distributed nodes to generate a large model as a global model of the distributed network. The global model in the embodiment of the application is obtained based on the weighted aggregation of the local models of all distributed nodes in the distributed network, so that the obtained global model can comprehensively and accurately reflect the real situation of the model of the entire distributed network, thereby enabling the subsequent determination of the local model update parameters of the distributed nodes based on the global model to be more accurate.

[0186] Based on the local model and parameters trained by the first distributed node, the node information of the first distributed node, and the global model of the distributed network, the intelligent network element of the center node determines the local model update parameters of the first distributed node. The local model update parameters of each distributed node can be calculated based on the resource situation, location information, computing power, and load situation of the distributed node, and the global model is run and verified, and the update weight and update gradient of the local model of each distributed node are calculated based on the result of the running verification and historical experience data, as the local model update parameters.

[0187] S202, updating the local model of the first distributed node according to the local model update parameters.

[0188] After the intelligent network element of the center node calculates the local model update parameters of each distributed node, the local model update parameters of the first distributed node can be used to update the local model of the first distributed node in cooperation with the first distributed node.

[0189] In the embodiment of the application, the intelligent network element of the center node determines the local model update parameters of the first distributed node based on the local model and parameters, node information, and global model of the distributed network, and updates the local model of the first distributed node according to the local model update parameters. Since the center node considers multiple factors in the distributed network when determining the local model update parameters of the first distributed node, the local model update parameters of the first distributed node are determined based on the comprehensive data of the entire distributed network, so that the local model update parameters of the first distributed node are more accurate.

[0190] When the intelligent network element of the center node updates the local model of the first distributed node according to the local model update parameters, the intelligent network element of the center node can perform different collaborative updates according to whether the first distributed node has the ability to complete the entire model update or not.

[0191] In one embodiment, S202 comprises: the intelligent network element of the center node sending the local model update parameter to the intelligent network element of the first distributed node, instructing the intelligent network element of the first distributed node to update train the local model.

[0192] This embodiment is directed to the case where the first distributed node supports model update training, i.e., the first distributed node has the ability to complete model update.

[0193] In this case, the intelligent network element of the center node only needs to send the local model update parameter of the first distributed node directly to the intelligent network element of the first distributed node, and the intelligent network element of the first distributed node will automatically update train the complete local model it has trained after receiving the local model update parameter. In this way, the center node sends the local model update parameter to the first distributed node for its own local model update in the case where the first distributed node has the ability to perform model update training, which does not affect the local model update of the first distributed node and also saves the resources of the center node.

[0194] In another embodiment, S202 comprises: the intelligent network element of the center node updating training the local model according to the local model update parameter, sending the model obtained by updating training and model training instruction information to the intelligent network element of the first distributed node, the model training instruction information being used to instruct the first distributed node to perform subsequent model training and update at the center node.

[0195] This embodiment is directed to the case where the first distributed node does not support model update training, i.e., the first distributed node does not have the ability to complete model update.

[0196] In this case, the intelligent network element of the center node updates trains the local model of the first distributed node obtained in advance according to the determined local model update parameter of the first distributed node, and after the update training is completed, the intelligent network element of the center node sends the model obtained by updating training to the intelligent network element of the first distributed node, and also sends a model training instruction information to the intelligent network element of the first distributed node, which instructs the first distributed node to perform subsequent model training and update after model training at the center node. In this embodiment, the intelligent network element of the center node directly helps the first distributed node to complete the update training of the local model when the first distributed node does not have the ability to perform model update training, so that the local model of the first distributed node can be updated, and the intelligent network element of the center node also sends the model obtained by updating training to the first distributed node, ensuring the timeliness of the first distributed node obtaining the local model after updating training.

[0197] In addition, in the embodiment of the present application, the two cases of the first distributed node supporting model update training and not supporting model update training can be that the first distributed node actively informs the center node, or that the center node calculates according to the node information of the first distributed node and the computing power required for local model update training, and the embodiment of the present application does not limit this. According to the two cases of the first distributed node supporting model update training and not supporting model update training, the center node adopts different strategies to complete the update training of the first distributed node, so that the mode of cooperating with the first distributed node for model update training is more flexible.

[0198] The local model trained by the first distributed node and sent to the center node is the model corresponding to 100% model training task, so the center node cooperates with the first distributed node for model guidance update.

[0199] However, if the local model trained by the first distributed node and sent to the center node is only a part of the training, for example, a local model obtained by completing 50% model training task, in this case, the center node cooperates with the first distributed node to complete the model corresponding to the other 50% model training task which cannot be trained by the first distributed node, and after the model corresponding to the other 50% model training task is trained, the center node continues to cooperate with the first distributed node for model guidance update.

[0200] The process will be described below through specific embodiments.

[0201] In one embodiment, the model training of the center node in cooperation with the first distributed node according to the node data in S101 further includes that the intelligent network element of the center node trains the model for the data which cannot be trained by the first distributed node according to the node data.

[0202] In the embodiment, the precondition for the center node to train the model for the data which cannot be trained by the first distributed node is that the center node needs to obtain the data which cannot be trained by the first distributed node.

[0203] In the foregoing embodiments, it is mentioned that the node data of the first distributed node obtained by the center node includes the data which cannot be trained by the first distributed node and the request for cooperative training, or does not include the data which cannot be trained by the first distributed node and the request for cooperative training.

[0204] For the case that the node data obtained by the center node does not include the data which cannot be trained by the first distributed node and the request for cooperative training, the center node can actively request the first distributed node to request the data.

[0205] In an embodiment, the model training method provided by the application further includes: the intelligent network element of the center node determining whether the first distributed node can complete model training according to the computing power information of the first distributed node, and requesting the data that the first distributed node cannot train from the first distributed node in the case that the first distributed node cannot complete model training.

[0206] The intelligent network element of the center node pre-evaluates the node information, computing power and resource information, etc. of the first distributed node to determine whether the first distributed node can complete its local model training. If it is determined that the first distributed node can complete model training, no operation is needed. If it is determined that the first distributed node cannot complete local model training, the center node actively requests the data that the first distributed node cannot train from the first distributed node. Specifically, the active request of the first distributed node can be that the intelligent network element of the center node sends a request for data that cannot be trained to the intelligent network element of the first distributed node, and receives the data that cannot be trained sent by the intelligent network element of the first distributed node.

[0207] In the embodiment, when the center node pre-perceives that the first distributed node has insufficient capability to perform local model training, the center node actively requests the first distributed node to report the data that cannot be trained, so that the center node can cooperate with the first distributed node to complete model training and timely help the first distributed node complete its local model training task.

[0208] The center node cooperating with the first distributed node to perform model training on the data that cannot be trained by the first distributed node can be divided into two cases: the center node itself cooperating with the first distributed node to perform model training on the data that cannot be trained by the first distributed node, and the center node determining a second distributed node to cooperate with the first distributed node to perform model training on the data that cannot be trained by the first distributed node. In the case that the center node determines a second distributed node to help the first distributed node perform model training, the computing resources of the center node can be greatly saved.

[0209] The process of the center node itself cooperating with the first distributed node to perform model training on the data that cannot be trained by the first distributed node is described below.

[0210] In an embodiment, the intelligent network element of the center node performing model training on the data that cannot be trained by the first distributed node according to node data includes: performing cooperative model training on the data that cannot be trained in response to a cooperative training request.

[0211] In this embodiment, the node data includes the untrainable data of the first distributed node and the collaborative training request, and the intelligent network element of the center node performs model training on the untrainable data of the first distributed node in response to the collaborative training request, and the obtained model can be referred to as a collaborative model. The collaborative model is only the model corresponding to the part of data that the first distributed node cannot train, and belongs to a part of the complete model that the first distributed node should train. Therefore, after the intelligent network element of the center node performs model training on the untrainable data of the first distributed node, the collaborative model obtained this time needs to be integrated with another part of the local model that the first distributed node has trained before, and the complete model obtained by integration is the model corresponding to the 100% model training task that the first distributed node should complete.

[0212] After the model integration is completed, the complete model obtained by integration also needs to be updated.

[0213] Based on this, if the first distributed node supports model update training, in an embodiment, the method further includes: integrating the trained collaborative model and the local model; and sending the integrated model and an update instruction to the intelligent network element of the first distributed node, instructing the intelligent network element of the first distributed node to update the integrated model.

[0214] That is, if the first distributed node has the ability to complete subsequent model update, the intelligent network element of the center node integrates the trained collaborative model and the part of the local model that the first distributed node has trained, and then sends the integrated model and an update instruction to the intelligent network element of the first distributed node. The local model update parameter of the first distributed node can be carried in the update instruction, and the purpose is to instruct the intelligent network element of the first distributed node to receive the integrated model and perform update training on the integrated model according to the update instruction and the local model update parameter of the first distributed node.

[0215] The local model update parameter of the first distributed node carried in the update instruction can be determined by the intelligent network element of the center node based on the global model, the node information of the first distributed node, and the integrated model. The specific determination process is the same as the foregoing embodiment, and will not be described here.

[0216] In this embodiment, when the first distributed node has the update training capability, the center node directly instructs the first distributed node to complete the model update training itself, so that the model training with the first distributed node is not affected, and the resources of the center node are also saved.

[0217] If the first distributed node does not support model update training, in an embodiment, the method further comprises: integrating the collaborative model and the local model, and updating the integrated model; and sending the integrated and updated model and model training indication information to the intelligent network element of the first distributed node, the model training indication information being used to indicate that subsequent model training and update of the first distributed node need to be performed at the center node.

[0218] If the first distributed node does not have the ability to complete subsequent model update, the intelligent network element of the center node integrates the collaborative model and the part of the local model that has been trained by the first distributed node, and then continues to update the integrated model, and sends the integrated and updated model to the intelligent network element of the first distributed node, so that the first distributed node can learn the latest state of the model in time. In addition, because the first distributed node neither has the ability to complete model training nor has the ability to perform model update, the intelligent network element of the center node also sends model training indication information to the intelligent network element of the first distributed node, the model training indication information being used to indicate that subsequent model training and update of the first distributed node need to be performed at the center node, which means that the center node will assist the first distributed node to complete each model training and update, thereby ensuring that the first distributed node can complete the model training task.

[0219] The process in which the center node determines the second distributed node to assist the first distributed node to perform model training on the data that the first distributed node cannot train is described below.

[0220] In an embodiment, the intelligent network element of the center node performs model training on the data that the first distributed node cannot train according to node data, comprising: determining a second distributed node according to node information of other distributed nodes in response to a collaborative training request, the second distributed node representing a distributed node that can perform collaborative model training on the data that cannot be trained.

[0221] If the node data obtained by the intelligent network element of the center node includes the data that the first distributed node cannot train and the collaborative training request, the intelligent network element of the center node determines a second distributed node according to node information of other distributed nodes in response to the collaborative training request.

[0222] For example, the second distributed node can be determined according to position information and computing power and resource information of each distributed node. The second distributed node determined in this way first has sufficient computing power and resources to help the first distributed node to complete model training on the data that cannot be trained, and second, needs to be close to the first distributed node in position to save transmission resources and improve transmission efficiency, so that the second distributed node can quickly assist the first distributed node to perform model training.

[0223] The central node can synchronize the relevant information to the second distributed node and the first distributed node after determining the second distributed node, facilitating subsequent interaction between the first distributed node and the second distributed node. Therefore, in an embodiment, the method further comprises: sending the untrainable data, the node information of the first distributed node, the local model and the parameters to the intelligent network element of the second distributed node.

[0224] The intelligent network element of the central node determines the second distributed node in order to help the first distributed node to collaboratively train the model, so it is necessary to send the obtained node information of the first distributed node, the untrainable data of the first distributed node, and the part of the local model and the parameters that have been trained by the first distributed node to the second distributed node. After receiving the information, the intelligent network element of the second distributed node performs model training based on the untrainable data of the distributed node to obtain a collaborative model, and then integrates the collaborative model and the part of the local model that has been trained by the first distributed node to obtain a completed model after integration.

[0225] After the integration is completed, if the first distributed node supports model update training, the intelligent network element of the second distributed node sends the model obtained by the integration to the intelligent network element of the first distributed node, and at the same time sends an update instruction to the intelligent network element of the first distributed node, instructing the first distributed node to perform update training on the model after integration. The update training of the first distributed node on the model after integration can be based on update parameters, such as updating the gradient, updating the weight, etc. These update parameters can be determined by the intelligent network element of the central node based on the model after integration, the global model and the node information of the first distributed node, and then sent directly to the first distributed node. Alternatively, the intelligent network element of the central node can determine the update parameters based on the model after integration, the global model and the node information of the first distributed node, and then send them to the intelligent network element of the second distributed node, and then the intelligent network element of the second distributed node sends the update parameters to the first distributed node in the update instruction.

[0226] If the first distributed node does not support model update training, the intelligent network element of the second distributed node directly updates and trains the integrated model, and after the update is completed, sends the integrated and updated model to the intelligent network element of the first distributed node, so that the first distributed node can learn the latest state of the model in time. Similarly, the update parameters used by the intelligent network element of the second distributed node to update the integrated model can also be the update parameters determined by the intelligent network element of the center node based on the integrated model, the global model and the node information of the first distributed node and sent to the intelligent network element of the second distributed node. In addition, because the first distributed node has neither the ability to train the model nor the ability to update the model, the intelligent network element of the second distributed node will also send model training instruction information to the intelligent network element of the first distributed node to inform the first distributed node that subsequent model training and model update of the first distributed node need to be performed in the second distributed node, so that subsequent training tasks of the complete model of the first distributed node can be performed in time.

[0227] In another embodiment, the method further comprises: the intelligent network element of the center node sending identification information of the second distributed node to the intelligent network element of the first distributed node.

[0228] The intelligent network element of the center node determines the second distributed node and sends identification information of the second distributed node to the intelligent network element of the first distributed node, and the identification information is used for the first distributed node to learn which node the second distributed node is, to facilitate subsequent information interaction.

[0229] In another embodiment, the method further comprises: the intelligent network element of the center node sending data sending notification to the intelligent network element of the first distributed node; the data sending notification is used to instruct the intelligent network element of the first distributed node to send the untrainable data, the node information of the first distributed node, the local model and the parameters to the second distributed node according to the identification information.

[0230] In actual application, after the center node determines the second distributed node, it can also send a data sending notification to the first distributed node to instruct the first distributed node to send the untrainable data, the node information of the first distributed node, the local model and the parameters to the second distributed node.

[0231] The purpose of the embodiment is to enable the second distributed node to learn the relevant information required for model training in cooperation with the first distributed node. Based on this, the process in which the intelligent network element of the center node sends the data that cannot be trained, the node information of the first distributed node, the local model and the parameters to the intelligent network element of the second distributed node in the foregoing embodiments can be executed alternatively in actual application. Specifically, it can be based on the data sending notification in the embodiment. If the center node sends these relevant information to the second distributed node, the center node does not need to send the data sending notification to the first distributed node, and the first distributed node does not need to send these relevant information to the second distributed node. However, if the center node does not send these relevant information to the second distributed node, the center node needs to send the data sending notification to the first distributed node, so that the first distributed node sends these relevant information to the second distributed node.

[0232] In the embodiment, the intelligent network element of the center node sends the relevant information required for model training in cooperation with the first distributed node to the intelligent network element of the second distributed node directly, or sends the data sending notification to the intelligent network element of the first distributed node to instruct the intelligent network element of the first distributed node to send the relevant information required for model training in cooperation with the first distributed node to the intelligent network element of the second distributed node. The two modes are executed alternatively, which improves the flexibility of the way in which the second distributed node learns the relevant information required for model training in cooperation with the first distributed node.

[0233] The above embodiments are all the processes of the center node cooperating with the first distributed node for model training. In actual application, the first distributed node needs other nodes to cooperate for model training. In addition to finding the center node in the distributed network for cooperation, it can also directly find other distributed nodes for cooperation. That is, the first distributed node can directly determine a second distributed node from other distributed nodes in the distributed network to cooperate with itself for model training without going through the center node.

[0234] The process in which the first distributed node determines a second distributed node from other distributed nodes is described below.

[0235] In one embodiment, the intelligent network element of the first distributed node determines the second distributed node based on the pre-maintained network node neighbor information, sends the data that cannot be trained by itself, the node information of itself, the local model and the parameters trained by itself, and the cooperation training request to the intelligent network element of the second distributed node to instruct the second distributed node to perform model training on the data that cannot be trained by itself.

[0236] The intelligent network element of the first distributed node can have node information of other distributed nodes in the distributed network, such as location information, identification information, computing power information, and load information, and the like, based on pre-maintained network node neighbor information. The intelligent network element of the first distributed node selects the second distributed node based on the ability to cooperate with itself to complete model training, and the distance from itself is relatively close, which is convenient to save transmission resources and improve transmission efficiency. Of course, it can also be to preferentially select other distributed nodes that have cooperated before, and the embodiments of the present application do not limit this.

[0237] After the intelligent network element of the first distributed node determines the second distributed node, the intelligent network element of the first distributed node sends relevant information for the second distributed node to cooperate with itself to perform model training, that is, data that cannot be trained by itself, node information of itself, a local model and parameters that have been trained by itself, and a cooperative training request.

[0238] After the intelligent network element of the second distributed node receives the relevant information, the intelligent network element of the second distributed node performs cooperative model training on the data that cannot be trained by the first distributed node, integrates the cooperative model obtained by training and the part of the local model that has been trained by the first distributed node, and then sends the integrated model and parameters to the intelligent network element of the first distributed node. After the intelligent network element of the first distributed node receives the integrated model, the intelligent network element of the first distributed node sends the integrated model and parameters and the node information of itself to the intelligent network element of the center node for model updating guidance. The intelligent network element of the center node analyzes the updating parameters of the integrated model based on the global model according to the integrated model reported by the first distributed node.

[0239] Then, for the case that the first distributed node supports updating training, the intelligent network element of the center node can send the updating parameters to the intelligent network element of the first distributed node, so that the intelligent network element of the first distributed node updates and trains the integrated model according to the updating parameters.

[0240] But for the case that the first distributed node does not support updating training, in one way, the intelligent network element of the center node can still send the updating parameters to the intelligent network element of the first distributed node, and the intelligent network element of the first distributed node sends the updating parameters to the intelligent network element of the second distributed node to request the second distributed node to cooperate with itself to update and train the integrated model. In another way, the intelligent network element of the center node can directly send the updating parameters to the second distributed node to instruct the second distributed node to cooperate with the first distributed node to update and train the integrated model.

[0241] Of course, for the case that the first distributed node does not support the update training, the intelligent network element of the second distributed node can send the integrated and updated model and the model training indication information to the intelligent network element of the first distributed node after the intelligent network element of the second distributed node cooperates with the first distributed node to update and train the integrated model, wherein the model training indication information is used to indicate that the subsequent model training and update of the first distributed node need to be performed in the second distributed node.

[0242] In addition, after the intelligent network element of the first distributed node determines the second distributed node and sends the above related information to the intelligent network element of the second distributed node, the second distributed node can refuse the first distributed node, so in an embodiment, the intelligent network element of the second distributed node sends a refusal cooperation request to the intelligent network element of the first distributed node in response to the cooperation training request sent by the intelligent network element of the first distributed node, and the refusal cooperation request is used to instruct the first distributed node to determine a new second distributed node based on the pre-maintained network node neighbor information again.

[0243] It should be noted that since the second distributed node and the first distributed node are both distributed nodes in the distributed network, the functions and responsibilities of the two are flat, so the second distributed node can refuse the request of the first distributed node if it has more important things to complete after receiving the cooperation training request of the first distributed node. The central node is a node in the distributed network that has the function of managing all distributed nodes, so in the foregoing embodiment, when the central node instructs the second distributed node to cooperate with the first distributed node to perform model training after determining the second distributed node, the second distributed node will not refuse. It should be further noted that the second distributed node determined by the first distributed node from the distributed network can be the same distributed node as the second distributed node determined by the central node from the distributed network in the foregoing embodiment, or can be a different distributed node, and the embodiments of the present application do not limit this.

[0244] In the embodiments of the present application, the first distributed node determines the second distributed node that can cooperate with itself to perform model training from the maintained network node neighbor information, so that when the first distributed node cannot complete the model training task, it can find any distributed node in the distributed network to cooperate to perform the model training task, so that the model training of the first distributed node reaches the global optimum.

[0245] The first distributed node represents any distributed node in the distributed network, so the combination of the aforementioned first distributed node and the search for a center node and the search for other any distributed node cooperates to perform a model training task, which is equivalent to the respective distributed nodes in the embodiments of the present application transmitting the respective node-in-model parameters and the data that cannot be supported by the computing power resources to cause the center node / other cooperative distributed node (second distributed node) to update or train and update after completing the training of the node-in-model, so that the distributed nodes in the distributed network cooperate in data and knowledge sharing, computing power and resource sharing, etc., so that each distributed node cannot complete the training task, and the global training model is optimized, and the network performance of the entire distributed network is improved, and the network resources are reasonably utilized.

[0246] Generally, in the distributed network, each distributed node performs local model training tasks, which are trained multiple times in a loop until the model is optimized, and the model training task is considered to be completed. In the embodiments of the present application, the condition for the local model of each distributed node to be optimal is to judge based on whether the global model in the center node meets the convergence condition.

[0247] In an embodiment, the method further comprises: the intelligent network element of the center node detecting whether the global model converges; and in the case where the global model does not converge, cyclically performing the process of cooperating with the first distributed node to perform model training until the latest obtained global model converges.

[0248] Specifically, the first distributed node represents any distributed node in the distributed network, so it is equivalent that each distributed node in the distributed network needs to continue to send the latest state model to the intelligent network element of the center node after obtaining the latest state model in each round of training task. The intelligent network element of the center node updates the global model based on the latest state model of the distributed nodes, and detects whether the updated latest global model converges. If the latest global model converges, the intelligent network element of the center node sends an instruction to stop training to each distributed node, indicating that the local model of each distributed node also converges. If the latest global model does not converge, the intelligent network element of the center node re-executes any one of the cooperative processes provided in the embodiments of the present application for cooperating with the first distributed node and other distributed nodes to perform model training until the latest state model updated by each distributed node converges.

[0249] In the embodiments of the present application, whether the model training of each distributed node in the distributed network converges is determined based on whether the global model of the center node converges. Since the global model is constructed based on the information and model of all distributed nodes in the distributed network, it can more comprehensively reflect the model situation of the entire distributed network. Therefore, taking whether the global model converges as the basis can more accurately determine whether the model of each distributed node is trained to the best state, so that the model training of each distributed node can reach the global optimum, thereby improving the network performance of the entire distributed network and optimizing the network and application service experience.

[0250] Based on the same technical concept, the embodiments of the model training method with the intelligent network element of the first distributed node as the execution subject are also provided. The principles, manners and technical effects involved in the implementation process of each step in the embodiments of the model training method with the intelligent network element of the first distributed node as the execution subject are the same as those in the foregoing embodiments with the intelligent network element of the center node as the execution subject, and the same parts and beneficial effects in the embodiments of the present application as the foregoing method embodiments will not be described in detail hereinafter.

[0251] The following are embodiments related to the model training method with the intelligent network element of the first distributed node as the execution subject.

[0252] As shown in Figure 5 The embodiments of the present application provide a model training method applied to an intelligent network element of a first distributed node, which comprises the following steps:

[0253] S301, sending node data to the intelligent network element of the center node; the node data is used for the intelligent network element of the center node to cooperatively perform model training of the first distributed node; the first distributed node is any one of the distributed nodes in the distributed network.

[0254] In one embodiment, the step of sending the node data to the intelligent network element of the center node comprises the following steps:

[0255] receiving a data reporting request sent by the intelligent network element of the center node;

[0256] According to the data reporting request, the node data is sent to the intelligent network element of the center node.

[0257] In one embodiment, the node data comprises:

[0258] the local model and parameters of the first distributed node after training, and the node information of the first distributed node; or

[0259] The first distributed node trains the local model and parameters, node information of the first distributed node, data that the first distributed node cannot train, and a collaborative training request.

[0260] In one embodiment, the method further comprises:

[0261] Collecting model training related data in the distributed network;

[0262] Training a local model according to the related data.

[0263] In one embodiment, collecting the related data in the distributed network comprises:

[0264] Sending a data collection request to each network function in the core network and each network function in the access network in the first distributed node;

[0265] Receiving response data sent by each network function to obtain model training related data.

[0266] In one embodiment, collecting the model training related data in the distributed network comprises:

[0267] Sending a data collection request to each data plane in the first distributed node;

[0268] Receiving response data sent by each data plane to obtain model training related data.

[0269] In one embodiment, the method further comprises:

[0270] Receiving a local model update message sent by the intelligent network element of the center node.

[0271] In one embodiment, the local model update message includes local model update parameters, and the method further comprises:

[0272] Updating and training the local model according to the local model update parameters.

[0273] In one embodiment, the local model update message includes an updated local model and model training indication information, and the model training indication information is used to indicate that subsequent model training and updating of the first distributed node need to be performed at the center node.

[0274] In one embodiment, the method further comprises:

[0275] Receiving an integrated model and update indication sent by the intelligent network element of the center node; the integrated model is obtained by integrating a collaborative model and a local model by the intelligent network element of the center node; the collaborative model is trained according to the data that the first distributed node cannot train;

[0276] Updating the integrated model.

[0277] In an embodiment, the method further comprises:

[0278] receiving, from the intelligent network element of the center node, a model integrated from the local model and the collaborative model and updated model training indication information, the model training indication information indicating that subsequent model training and updating of the first distributed node are to be performed at the center node.

[0279] In an embodiment, the method further comprises:

[0280] receiving, from the intelligent network element of the center node, identification information of a second distributed node, the second distributed node representing a node that trains the collaborative model according to the data that the first distributed node cannot train.

[0281] In an embodiment, the method further comprises:

[0282] in response to a data sending notification sent by the intelligent network element of the center node, sending, to the intelligent network element of the second distributed node, the data that cannot be trained, node information of the first distributed node, the local model, and parameters to the intelligent network element of the second distributed node according to the identification information.

[0283] In an embodiment, the method further comprises:

[0284] receiving, from the intelligent network element of the second distributed node, an integrated model and an update indication, the integrated model being obtained by integrating the local model and the collaborative model by the intelligent network element of the second distributed node.

[0285] updating the integrated model.

[0286] In an embodiment, the method further comprises:

[0287] receiving, from the intelligent network element of the second distributed node, a model integrated from the local model and the collaborative model and updated model training indication information, the model training indication information indicating that subsequent model training and updating of the first distributed node are to be performed at the second distributed node.

[0288] In an embodiment, before receiving the local model update parameters sent by the intelligent network element of the center node, the method further comprises:

[0289] determining the second distributed node based on pre-maintained network node neighbor information.

[0290] sending, to the intelligent network element of the second distributed node, the data that cannot be trained of the first distributed node, node information of the first distributed node, the local model and parameters trained by the first distributed node, and a collaborative training request to instruct the second distributed node to perform model training on the data that cannot be trained of the first distributed node.

[0291] In an embodiment, the method further comprises:

[0292] If a rejection of the coordination request sent by the intelligent network element of the second distributed node is received, a new second distributed node is determined based on the pre-maintained network node neighbor information.

[0293] Likewise, based on the same technical concept, the embodiments of the present application also provide an embodiment of a model training method with the intelligent network element of the second distributed node as the execution subject, wherein the principles, manners and technical effects that the implementation processes of the steps in the embodiment of the model training method with the intelligent network element of the second distributed node as the execution subject are involved are also the same as those of the foregoing embodiment with the intelligent network element of the center node as the execution subject, and the beneficial effects that can be achieved are also the same, and thus the same parts and beneficial effects in the embodiments of the present application as the foregoing method embodiments will not be described in detail hereinafter.

[0294] The following are embodiments related to the model training method with the intelligent network element of the second distributed node as the execution subject.

[0295] As shown in Figure 6 The embodiments of the present application provide a model training method, applied to an intelligent network element of a second distributed node, which comprises:

[0296] S401, receiving node data of a first distributed node; the first distributed node is any one of the distributed nodes in the distributed network.

[0297] S402, performing model training of the first distributed node in coordination according to the node data.

[0298] In an embodiment, the receiving of the node data of the first distributed node comprises:

[0299] receiving node data sent by an intelligent network element of a center node; or,

[0300] receiving node data sent by an intelligent network element of the first distributed node.

[0301] In an embodiment, the node data comprises untrainable data of the first distributed node, node information of the first distributed node, a local model and parameters of the first distributed node that have been trained.

[0302] In an embodiment, the performing of the model training of the first distributed node in coordination according to the node data of the first distributed node comprises:

[0303] performing coordinated model training on the untrainable data.

[0304] In an embodiment, the first distributed node supports update training, and the method further comprises:

[0305] sending the integrated model and the update indication to the intelligent network element of the first distributed node, instructing the intelligent network element of the first distributed node to update the integrated model.

[0306] In one embodiment, the first distributed node does not support updating training, and the method further includes:

[0307] integrating the collaborative model and the local model, and updating the integrated model;

[0308] sending the integrated and updated model and model training indication information to the intelligent network element of the first distributed node, the model training indication information being used to instruct that subsequent model training and updating of the first distributed node are to be performed at the second distributed node.

[0309] In one embodiment, the method further includes:

[0310] In response to the collaborative training request sent by the intelligent network element of the first distributed node, sending a rejection collaborative request to the intelligent network element of the first distributed node, the rejection collaborative request being used to instruct the first distributed node to determine a new second distributed node based on pre-maintained network node neighbor information again.

[0311] On the basis of the above embodiments, an embodiment of a model training method is further provided, which relates to the process of collecting data by the intelligent network elements of each distributed node in a distributed network and performing model training and then sending to a center node.

[0312] Embodiment 1.1: The intelligent network element of the distributed node collects model related data from each NF in the node, and performs model related data analysis and model training based on the computing power and resources of the node. After the training is completed, data is reported to the center node, and the center node guides and updates the training result. As shown in Figure 7 the embodiment includes the following steps:

[0313] S111, the intelligent network element of the distributed node sends a data collection request to the NF(s) in the distributed node.

[0314] The data collection request is used to request model related data. The model related data includes NF data, UE data, and node information, etc.

[0315] S112, the NF(s) reply to the intelligent network element of the distributed node with a data collection response.

[0316] S113, the intelligent network element of the distributed node performs data analysis and model training according to the collected model related data.

[0317] S114, the intelligent network element of the distributed node reports the trained model and model parameters, node information, etc. to the intelligent network element of the center node.

[0318] If the intelligent network element of the distributed node fails to train due to node computing power or resource information, the unanalyzed / trained non-private data (i.e. the data that cannot be trained in the foregoing embodiment) is reported together. The reporting here can be divided into active reporting by the intelligent network element of the distributed node and requested reporting by the intelligent network element of the center node.

[0319] If the untrained data of the intelligent network element of the distributed node is all private data, the node administrator needs to be consulted for increasing computing power equipment and the like.

[0320] S115, after receiving the reported data from the intelligent network element of the distributed node, the intelligent network element of the center node guides the model and parameters.

[0321] The specific principles and implementation processes of the steps in the foregoing embodiments can be referred to the descriptions of the foregoing embodiments, which will not be described here.

[0322] Embodiment 1.2: The intelligent network element of the distributed node collects model-related data from the data plane in the node, analyzes the model-related data and trains the model based on the computing power and resources of the node. After the training is completed, the intelligent network element reports to the center node, and the center node guides and updates the training result. As shown in FIG. 1B, this embodiment includes the following steps: Figure 8

[0323] S121, the intelligent network element of the distributed node sends a data collection request to the data plane in the distributed node, and the data collection request is used to request model-related data. The model-related data includes NF data, UE data, node information, etc.

[0324] S122, the data plane replies to the intelligent network element of the distributed node with a data collection response.

[0325] S123, the intelligent network element of the distributed node analyzes the data and trains the model based on the collected model-related data.

[0326] S124, the intelligent network element of the distributed node reports the trained model and model parameters, node information, etc. to the intelligent network element of the center node.

[0327] If the intelligent network element of the distributed node fails to train due to node computing power or resource information, the unanalyzed / trained non-private data (i.e. the data that cannot be trained in the foregoing embodiment) is reported together. The reporting here can be divided into active reporting by the intelligent network element of the distributed node and requested reporting by the intelligent network element of the center node.

[0328] ​If the untrained data of the intelligent network element of the distributed node is all privacy data, the node administrator needs to be consulted to increase the computing power equipment and the like.

[0329] S125, after the intelligent network element of the center node receives the reported data of the intelligent network element of the distributed node, the model and the parameter are guided.

[0330] The specific principles and implementation processes of the steps in the above embodiments can be referred to the descriptions of the foregoing embodiments, which will not be described here.

[0331] On the basis of the foregoing embodiments, an embodiment of a model training method is further provided, which relates to the process of the intelligent network element of the center node guiding the model and the parameter of each distributed node.

[0332] When the distributed node analyzes and trains the model locally, the computing power and resources of the distributed node may be limited and unable to complete the data analysis and model training. At this time, the center node needs to analyze the specific conditions of each distributed node and determine whether to cooperate with the distributed node to reasonably allocate resources and cooperatively complete the model training of the distributed node.

[0333] (1) When the computing power and resources of the distributed node are sufficient to complete the data analysis and model training in the node, the center node only needs to update and guide the parameters of the model according to the global state.

[0334] Embodiment 2.1: The distributed node actively reports the node data to the center node, so that the center node only needs to update and guide the parameters of the model of the distributed node according to the global state. As shown in the following table, this embodiment includes the following steps: Figure 9

[0335] S211, the intelligent network element of the distributed node reports the locally trained model and the parameter, the node information to the intelligent network element of the center node.

[0336] S212, the intelligent network element of the center node updates the model parameters based on the global model.

[0337] Specifically, the intelligent network element of the center node performs a weighted aggregation operation on the local model reported by each distributed node based on the weight to obtain a global model, calculates the update weight and the update gradient of the local model of each distributed node based on the global model and the global data, resource conditions and the like, and guides and updates the model of the distributed node.

[0338] S213, the intelligent network element of the center node sends the updated model parameters to the intelligent network element of the distributed node.

[0339] ​S214, the intelligent network element of the distributed node adjusts and trains the local model according to the updated model parameters.

[0340] The steps S211-S214 are repeated subsequently until the global model converges to the optimum.

[0341] The specific principles and implementation processes of the steps in the above embodiments can be referred to the descriptions of the foregoing embodiments, which will not be repeated here.

[0342] Embodiment 2.2: The central node requests the distributed node to report node data, so as to update the model parameters of the distributed node according to the global network state and guide the model. As shown in Figure 10 , this embodiment includes the following steps:

[0343] S221, the intelligent network element of the central node requests the intelligent network element of the distributed node to report node information and model and parameters.

[0344] S222, the intelligent network element of the distributed node reports the locally trained model and parameters, node information to the intelligent network element of the central node.

[0345] S223, the intelligent network element of the central node updates the model parameters based on the global model.

[0346] Based on the global model and the global network data, resource situation, etc., the update weight and update gradient of the local model of each distributed node are calculated, and the model of the distributed node is guided and updated.

[0347] S224, the intelligent network element of the central node sends the updated model parameters to the intelligent network element of the distributed node.

[0348] S225, the intelligent network element of the distributed node adjusts and trains the local model according to the updated model parameters.

[0349] The steps S222-S225 are repeated subsequently until the global model converges to the optimum.

[0350] (2) When the distributed node analyzes data and trains the model locally, the computing power and resources of the distributed node are limited and cannot complete data analysis and model training, and a cooperative training request can be sent to the central node, and the central node cooperatively trains the model.

[0351] Embodiment 2.3: The central node guides and cooperatively trains the model of the distributed node based on the cooperative request of the distributed node. As shown in Figure 11 , this embodiment includes the following steps:

[0352] S231, the intelligent network element of the distributed node reports the locally trained model and parameters, node information, untrainable data, and a collaborative model training request to the intelligent network element of the center node.

[0353] The untrainable data includes non-private data that cannot be trained due to limited computing power and resources.

[0354] S232, the intelligent network element of the center node trains the untrainable data reported by the distributed node to obtain a collaborative model, integrates the local model and the collaborative model, and updates the parameters of the integrated model of the distributed node according to the global model.

[0355] S233, the intelligent network element of the center node sends the integrated model and the updated model parameters to the intelligent network element of the distributed node.

[0356] If the computing power and resources of the distributed node can support the update training of the model, the intelligent network element of the center node sends the trained model and parameters to the intelligent network element of the distributed node. If the computing power and resources of the distributed node cannot support the update training of the model, the intelligent network element of the center node needs to inform the distributed node of the training result and inform the distributed node that subsequent model training and update will be performed at the center node.

[0357] S234, the intelligent network element of the distributed node adjusts and trains the integrated model according to the updated model parameters.

[0358] The steps S231-S234 are repeated until the model converges to the optimal.

[0359] The specific principles and implementation processes of each step in the above embodiments can be referred to the description of the foregoing embodiments, which will not be repeated here.

[0360] Embodiment 2.4: When the distributed node analyzes data and trains a model locally, the computing power and resources of the distributed node are limited and cannot complete data analysis and model training, a collaborative training request can be sent to the center node, and the center node selects collaborative distributed nodes for the distributed node with insufficient capability to perform collaborative model training. As shown in the following figure, this embodiment includes the following steps: Figure 12

[0361] S241, the intelligent network element of the distributed node 1 reports the locally trained model and parameters, node information, untrainable data, and a collaborative model training request to the intelligent network element of the center node.

[0362] The untrainable data includes non-private data that cannot be trained due to limited computing power and resources.

[0363] ​S242, the intelligent network element of the center node updates the model parameters of the distributed node 1 according to the global model, and selects the distributed node 2 as the collaborative distributed node for collaborative training of the distributed node 1 according to the data reported by the distributed node 1 and the resource conditions of other distributed nodes.

[0364] S243, the intelligent network element of the center node sends the updated model parameters to the intelligent network element of the distributed node 1, and carries the node ID of the distributed node 2.

[0365] S244a: the intelligent network element of the center node sends the node information of the distributed node 1, the locally trained model of the distributed node 1 and the untrainable data to the intelligent network element of the distributed node 2.

[0366] S244b: the intelligent network element of the distributed node 1 sends its own node information, the locally trained model of the distributed node 1 and the untrainable data to the intelligent network element of the distributed node 2.

[0367] Among them, steps S244a and S244b are optional, for example, if the intelligent network element of the center node sends a data sending notification to the intelligent network element of the distributed node 1, S244b is executed, and if the intelligent network element of the center node does not send a data sending notification to the intelligent network element of the distributed node 1, S244a is executed.

[0368] S245, the intelligent network element of the distributed node 2 analyzes and collaboratively trains the untrainable data of the distributed node 1, and integrates the collaborative model and the locally trained model of the distributed node 1.

[0369] S246, the intelligent network element of the distributed node 2 sends the integrated model and parameters to the distributed node 1.

[0370] S247, the intelligent network element of the distributed node 1 adjusts and trains the integrated model according to the updated model parameters of the center node.

[0371] If the computing power and resources of the distributed node 1 can support the update training of the model, the distributed node 2 sends the integrated model and parameters to the distributed node 1, so that the distributed node 1 itself completes the update training of the integrated model. If the computing power and resources of the distributed node 1 cannot support the update training of the model, the distributed node 2 updates the integrated model using the updated model parameters of the center node, sends the updated model to the distributed node 1, and notifies the distributed node 1 that the subsequent model training and update will be performed on the distributed node 2.

[0372] It should be noted that, Figure 12It is not shown that the computing power and resources of the distributed node 1 cannot support the update training of the model. Only the computing power and resources of the distributed node 1 can support the update training of the model.

[0373] In addition, it should be noted that whether the distributed node 1 updates and trains the integrated model or the distributed node 2 updates and trains the integrated model, the update model parameters used are the update model parameters determined by the center node in combination with the global model for the integrated model.

[0374] The steps S242-S247 are repeated subsequently until the model converges to the optimal.

[0375] The specific principles and implementation processes of each step in the above embodiments can be referred to the description of the foregoing embodiments, which will not be repeated here.

[0376] Embodiment 2.5 When the distributed node analyzes data and trains the model locally, the computing power and resources of the distributed node are limited and cannot complete data analysis and model training, the center node actively cooperatively trains the model. As shown in Figure 13 The embodiment includes the following steps:

[0377] S251, the intelligent network element of the distributed node reports the locally trained model and parameters, node information to the intelligent network element of the center node.

[0378] S252, the intelligent network element of the center node judges that the distributed node cannot perform model training, requests data from the distributed node and receives the non-private data that cannot be trained from the distributed node.

[0379] Specifically, the intelligent network element of the center node judges that the distributed node resources are limited and cannot complete data analysis and model training by itself according to the computing power and resource information reported by the distributed node, requests data from the distributed node and receives non-private data that cannot be trained from the distributed node.

[0380] Note: If the center node perceives that the computing power and resource information of the distributed node cannot complete the training of all local models before the distributed node reports the information, it is not necessary to perform S251 directly, and directly performs step S252.

[0381] S253, the intelligent network element of the center node trains the model for the data that cannot be trained reported by the distributed node to obtain a cooperative model, integrates the local model and the cooperative model, and updates the parameters of the integrated model according to the global model.

[0382] S254, the intelligent network element of the center node sends the integrated model and the updated model parameters to the intelligent network element of the distributed node.

[0383] If the computing power and resources of the distributed node can support the update training of the model, the center node sends the trained model and parameters to the distributed node. If the computing power and resources of the distributed node cannot support the update training of the model, the center node needs to inform the distributed node of the training result, and inform the distributed node that subsequent model training and update are performed at the center node.

[0384] S255, the intelligent network element of the distributed node adjusts and trains the integrated model according to the updated model parameters.

[0385] The subsequent steps S251-S255 are repeated until the model converges to the optimal.

[0386] The specific principles and implementation processes of each step in the above embodiments can be referred to the description of the foregoing embodiments, which will not be repeated here.

[0387] Embodiment 2.6: When the distributed node analyzes data and trains the model locally, the computing power and resources of the distributed node are limited and cannot complete data analysis and model training, the center node selects a collaborative distributed node for the distributed node with insufficient capability to perform collaborative model training. As shown in the following figure, this embodiment includes the following steps: Figure 14

[0388] S261, the intelligent network element of the distributed node 1 reports the locally trained model and parameters, node information to the intelligent network element of the center node.

[0389] S262, the intelligent network element of the center node judges that the distributed node 1 cannot perform model training, requests data from the distributed node 1 and receives the data from the distributed node 1 that cannot be trained.

[0390] The intelligent network element of the center node judges that the distributed node resource is limited and cannot complete data analysis and model training by itself according to the computing power and resource information reported by the distributed node 1, requests non-private data from the distributed node 1 and receives the data from the distributed node 1.

[0391] Note: If the center node perceives that the computing power and resource information of the distributed node cannot complete the training of all local models before the distributed node reports the information, it directly executes step S262 without executing S261.

[0392] S263, the intelligent network element of the center node updates the model parameters of the distributed node 1 according to the global situation, and selects a distributed node 2 as a collaborative distributed node for collaborative training for the distributed node 1 according to the data responded by the distributed node 1 and the resource situation of other distributed nodes.

[0393] ​S264, the intelligent network element of the center node sends the updated model parameters to the intelligent network element of the distributed node 1, and carries the node ID of the distributed node 2.

[0394] S265a, the intelligent network element of the center node sends the node information of the distributed node 1, the locally trained model and the untrainable data to the intelligent network element of the distributed node 2.

[0395] S265b, the intelligent network element of the distributed node 1 sends its own node information, locally trained model and untrainable data to the intelligent network element of the distributed node 2.

[0396] Note: Steps S265a and S265b are optional. For example, if the intelligent network element of the center node sends a data sending notification to the intelligent network element of the distributed node 1, S265b is executed, and if the intelligent network element of the center node does not send a data sending notification to the intelligent network element of the distributed node 1, S265a is executed.

[0397] S266, the intelligent network element of the distributed node 2 analyzes the untrainable data of the intelligent network element of the distributed node 1 and cooperatively trains the model, and integrates the cooperative model and the locally trained model of the distributed node 1.

[0398] S267, the intelligent network element of the distributed node 2 sends the integrated model and parameters to the intelligent network element of the distributed node 1.

[0399] S268, the intelligent network element of the distributed node adjusts and trains the integrated model according to the updated model parameters of the center node.

[0400] If the computing power and resources of the distributed node 1 can support the update training of the model, the distributed node 2 sends the integrated model and parameters to the distributed node 1, so that the distributed node 1 itself completes the update training of the integrated model. If the computing power and resources of the distributed node 1 cannot support the update training of the model, the distributed node 2 updates the integrated model using the updated model parameters of the center node, sends the integrated and updated model to the distributed node 1, and notifies the distributed node 1 that subsequent model training and update will be performed on the distributed node 2.

[0401] It should be noted that, Figure 14 The case where the computing power and resources of the distributed node 1 cannot support the update training of the model is not shown in the figure, and only the case where the computing power and resources of the distributed node 1 can support the update training of the model is shown. In addition, it should be noted that whether the distributed node 1 or the distributed node 2 updates and trains the integrated model, the updated model parameters used are the updated model parameters determined by the center node in combination with the global model for the integrated model.

[0402] The steps S262-S268 are repeated until the model converges to the optimum.

[0403] The specific principles and implementation processes of each step in the above embodiments can refer to the descriptions of the foregoing embodiments, which will not be repeated here.

[0404] Embodiment 2.7: When the distributed node can maintain the information of the neighbor distributed node, the distributed node can select the collaborative distributed node for data analysis and model training by itself. As shown in the following figure, this embodiment includes the following steps: Figure 15

[0405] S271: The intelligent network element of the distributed node 1 selects the collaborative distributed node 2 by itself based on the neighbor information maintained by itself.

[0406] S272: The intelligent network element of the distributed node 1 sends its node information, the local model and parameters trained by itself, the non-private data that cannot be trained, and the collaborative training request to the intelligent network element of the distributed node 2.

[0407] Note: If the distributed node 2 refuses the collaborative training request of the distributed node 1, the distributed node 1 needs to select a new distributed node 2 as the collaborative distributed node.

[0408] S273: The intelligent network element of the distributed node 2 analyzes the data reported by the intelligent network element of the distributed node 1 and collaboratively trains the model, and then integrates the collaborative model with the local model trained by the distributed node 1.

[0409] S274: The intelligent network element of the distributed node 2 sends the integrated model and parameters to the intelligent network element of the distributed node 1.

[0410] S275: The intelligent network element of the distributed node 1 reports the integrated model and parameters, node information to the intelligent network element of the center node.

[0411] S276: The intelligent network element of the center node updates the integrated model parameters reported by the distributed node 1 according to the global model.

[0412] S277: The intelligent network element of the center node sends the updated model parameters to the intelligent network element of the distributed node 1.

[0413] ​If the computing power and resources of the distributed node 1 can support the update training of the model, the intelligent network element of the center node sends the updated model parameters to the distributed node 1. If the computing power and resources of the distributed node 1 cannot support the update training of the model, the intelligent network element of the center node sends the updated model parameters to the distributed node 1, and the distributed node 1 sends the updated model parameters to the distributed node 2; or the intelligent network element of the center node directly sends the updated model parameters to the distributed node 2, and the intelligent network element of the distributed node 2 needs to inform the distributed node 1 of the model result of the update, and the subsequent model training and update are performed on the distributed node 2.

[0414] It should be noted that, Figure 14 In the above embodiment, only the case that the computing power and resources of the distributed node 1 can support the update training of the model is shown, and the case that the computing power and resources of the distributed node 1 cannot support the update training of the model is not shown.

[0415] S278, the intelligent network element of the distributed node 1 adjusts and trains the integrated model according to the model parameters updated by the center node.

[0416] The subsequent steps S272-S278 are repeated until the model converges to the optimal.

[0417] The specific principles and implementation processes of each step in the above embodiments can be referred to the descriptions of the foregoing embodiments, which will not be described here.

[0418] It can be understood that in the embodiments 1.1-1.2 and 2.1-2.7, the distributed node and the distributed node 1 represent any distributed node in the distributed network, and the distributed node 2 represents any distributed node other than the distributed node 1.

[0419] Based on the above embodiments, it can be known that in the embodiments of the present application, the intelligent network element is introduced into each network node, which is used to collect data in the network and perform data analysis and model training. The intelligent network element of the distributed node can collect data from the NF in the node and the data plane of the node. In addition to collecting and analyzing the data and model of the node, the intelligent network element of the center node can also perform aggregation operation on the models of each distributed node based on the global information to obtain a global model, and update the models and parameters in each distributed node based on the global model to guide the model training of each distributed node. The models and parameters in each distributed node can be actively reported by the intelligent network element of the distributed node, or requested and reported by the intelligent network element of the center node to the intelligent network element of the distributed node.

[0420] When the distributed node itself has limited computing power and resources and cannot complete the intra-node data analysis and model training by itself, the center node can be requested to train cooperatively, and the center node itself or other distributed nodes can be selected as cooperative nodes to train cooperatively with the requesting distributed node. When the distributed node itself has limited computing power and resources and cannot complete the intra-node data analysis and model training by itself, the center node can be actively discovered, and the center node itself or other distributed nodes can be selected as cooperative nodes to train cooperatively with the requesting distributed node. When the distributed node itself has limited computing power and resources and cannot complete the intra-node data analysis and model training by itself, and the distributed node itself can maintain information of neighbor distributed nodes, the distributed node can select a cooperative distributed node to perform data analysis and model training.

[0421] In this way, through the cooperation of data and knowledge, the sharing of computing power and resources, etc. between distributed nodes, the training tasks that cannot be completed by a single distributed node are solved, the global training model is optimized, the network performance is improved, and the network resources are reasonably utilized.

[0422] The implementation principles, processes, and technical effects of each step in the above embodiments are the same as those in the above embodiments, and will not be repeated. In addition, each step in each embodiment of the present application is not necessarily selected, and other steps can be included between the steps. The steps in each embodiment of the present application are not limited to being executed in the order described. Some steps in each embodiment of the present application can be executed in parallel.

[0423] It should be understood that, although each step in the flowchart involved in each embodiment as described above is displayed in sequence according to the arrow, these steps are not necessarily executed in sequence according to the arrow. Unless otherwise stated herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or steps or stages in other steps.

[0424] Based on the same inventive concept, the present application also provides a model training device for implementing all the model training methods involved above.

[0425] As Figure 16As shown, the embodiment of the present application provides a model training device 160, which comprises: a first training module 1601, configured to perform model training in cooperation with a first distributed node according to node data of the first distributed node; the first distributed node is any one of the distributed nodes in the distributed network.

[0426] In one embodiment, the device comprises:

[0427] The first node data receiving module is configured to receive node data sent by an intelligent network element of the first distributed node.

[0428] In one embodiment, the device comprises:

[0429] The request reporting module is configured to send a data reporting request to the intelligent network element of the first distributed node, the data reporting request being used to instruct the intelligent network element of the first distributed node to report node data.

[0430] The second node data receiving module is configured to receive node data sent by the intelligent network element of the first distributed node.

[0431] In one embodiment, the node data comprises: a local model and parameters trained by the first distributed node, node information of the first distributed node; or the local model and parameters trained by the first distributed node, the node information of the first distributed node, data that cannot be trained by the first distributed node, and a cooperative training request.

[0432] In one embodiment, the first training module 1601 comprises:

[0433] The model updating unit is configured to perform local model updating of the first distributed node according to the node data.

[0434] In one embodiment, the node data comprises the local model and parameters trained by the first distributed node, and the node information of the first distributed node; and the model updating unit comprises:

[0435] The updating parameter determining sub-unit is configured to determine a local model updating parameter of the first distributed node based on the local model and parameters, the node information, and a global model of the distributed network.

[0436] The model updating sub-unit is configured to perform local model updating of the first distributed node according to the local model updating parameter.

[0437] In one embodiment, the first distributed node supports model updating training; and the model updating sub-unit is further configured to send the local model updating parameter to the intelligent network element of the first distributed node, so as to instruct the intelligent network element of the first distributed node to perform updating training on the local model.

[0438] In an embodiment, the first distributed node does not support model update training; the model update subunit is further configured to update the training local model according to the local model update parameter; and the model update subunit is further configured to send the model obtained by the update training and model training indication information to the intelligent network element of the first distributed node, where the model training indication information is used to indicate that subsequent model training and update of the first distributed node need to be performed at the center node.

[0439] In an embodiment, the apparatus further includes:

[0440] The local model obtaining module is configured to obtain the local model reported by each distributed node in the distributed network.

[0441] The global model obtaining module is configured to perform a weighted aggregation operation according to the local model reported by each distributed node to obtain a global model of the distributed network.

[0442] In an embodiment, the apparatus further includes:

[0443] The convergence detection module is configured to detect whether the global model converges.

[0444] The cycle coordination triggering module is configured to cyclically perform the process of coordinating the first distributed node to perform model training until the latest obtained global model converges, in a case where the global model does not converge.

[0445] In an embodiment, the first training module 1601 further includes:

[0446] The coordinated training unit is configured to perform model training on data that the first distributed node cannot train, according to the node data.

[0447] In an embodiment, the node data further includes data that the first distributed node cannot train and a coordinated training request; and the coordinated training unit includes:

[0448] The first coordinated training subunit is configured to perform coordinated model training on the data that cannot be trained, in response to the coordinated training request.

[0449] In an embodiment, the first distributed node supports model update training; and the apparatus further includes:

[0450] The integration module is configured to integrate the trained coordinated model and the local model.

[0451] The first indication module is configured to send the integrated model and an update indication to the intelligent network element of the first distributed node, to instruct the intelligent network element of the first distributed node to update the integrated model.

[0452] In an embodiment, the first distributed node does not support model update training; and the apparatus further includes:

[0453] a first integrated updating module configured to integrate the collaborative model and the local model and update the integrated model;

[0454] a second indication module configured to send the integrated and updated model and model training indication information to the intelligent network element of the first distributed node, the model training indication information being used to indicate that subsequent model training and updating of the first distributed node need to be performed at the center node.

[0455] In an embodiment, the node data further includes untrainable data of the first distributed node and a collaborative training request; and the collaborative training unit further includes:

[0456] a second node determination subunit configured to determine, in response to the collaborative training request, a second distributed node according to node information of other distributed nodes, the second distributed node representing a distributed node capable of collaborative model training on the untrainable data.

[0457] In an embodiment, the apparatus further includes:

[0458] an identification sending subunit configured to send identification information of the second distributed node to the intelligent network element of the first distributed node.

[0459] In an embodiment, the apparatus further includes:

[0460] a data sending subunit configured to send the untrainable data, node information of the first distributed node, the local model and parameters to the intelligent network element of the second distributed node.

[0461] In an embodiment, the apparatus further includes:

[0462] a notification unit configured to send a data sending notification to the intelligent network element of the first distributed node, the data sending notification being used to instruct the intelligent network element of the first distributed node to send the untrainable data, node information of the first distributed node, the local model and parameters to the second distributed node according to the identification information.

[0463] In an embodiment, the apparatus further includes:

[0464] a computing power judgment module configured to judge whether the first distributed node is capable of completing model training according to computing power information of the first distributed node.

[0465] a data sending module configured to request the untrainable data of the first distributed node from the first distributed node in a case where the first distributed node is incapable of completing model training.

[0466] Embodiments of the present application further provide a model training apparatus, which includes:

[0467] The second training module is configured to send node data to the intelligent network element of the center node, and the node data is used for the intelligent network element of the center node to cooperatively train a model of the first distributed node.

[0468] In an embodiment, the second training module comprises:

[0469] The request receiving unit is configured to receive a data reporting request sent by the intelligent network element of the center node.

[0470] The node data sending unit is configured to send node data to the intelligent network element of the center node according to the data reporting request.

[0471] In an embodiment, the node data comprises a trained local model and parameters of the first distributed node, node information of the first distributed node, or a trained local model and parameters of the first distributed node, node information of the first distributed node, untrainable data of the first distributed node, and a cooperative training request.

[0472] In an embodiment, the apparatus further comprises:

[0473] The data collection module is configured to collect model training related data in the distributed network.

[0474] The local model training module is configured to train a local model according to the related data.

[0475] In an embodiment, the data collection module comprises:

[0476] The first collection request sending unit is configured to send a data collection request to each network function in the core network and each network function in the access network in the first distributed node.

[0477] The first model data receiving unit is configured to receive response data sent by each network function to obtain model training related data.

[0478] In an embodiment, the data collection module comprises:

[0479] The second collection request sending unit is configured to send a data collection request to each data plane in the first distributed node.

[0480] The second model data receiving unit is further configured to receive response data sent by each data plane to obtain model training related data.

[0481] In an embodiment, the apparatus further comprises:

[0482] The update message receiving module is configured to receive a local model update message sent by the intelligent network element of the center node.

[0483] In an embodiment, the local model update message comprises local model update parameters, and the apparatus further comprises:

[0484] an update training module configured to update train the local model according to the local model update parameters.

[0485] In an embodiment, the local model update message comprises the updated trained local model and model training indication information, and the model training indication information is used to indicate that subsequent model training and updating of the first distributed node are both required to be performed at the central node.

[0486] In an embodiment, the apparatus further comprises:

[0487] a first indication receiving module configured to receive an integrated model and update indication sent by the intelligent network element of the central node, wherein the integrated model is obtained by integrating the collaborative model and the local model by the intelligent network element of the central node, and the collaborative model is trained according to the untrainable data of the first distributed node;

[0488] a first update module configured to update the integrated model.

[0489] In an embodiment, the apparatus further comprises:

[0490] a second indication receiving module configured to receive a model obtained by integrating and updating the collaborative model and the local model and model training indication information sent by the intelligent network element of the central node, and the model training indication information is used to indicate that subsequent model training and updating of the first distributed node are both required to be performed at the central node.

[0491] In an embodiment, the apparatus further comprises:

[0492] an identification information receiving module configured to receive identification information of a second distributed node sent by the intelligent network element of the central node, wherein the second distributed node represents a node for training the collaborative model according to the untrainable data of the first distributed node.

[0493] In an embodiment, the apparatus further comprises:

[0494] a training data sending module configured to send the untrainable data, node information of the first distributed node, the local model and parameters to the intelligent network element of the second distributed node according to the identification information in response to a data sending notification sent by the intelligent network element of the central node.

[0495] In an embodiment, the apparatus further comprises:

[0496] a third indication receiving module, configured to receive an integrated model and an update indication sent by the intelligent network element of the second distributed node; the integrated model is obtained by integrating the collaborative model and the local model by the intelligent network element of the second distributed node;

[0497] a second updating module, configured to update the integrated model.

[0498] In an embodiment, the apparatus further includes:

[0499] a fourth indication receiving module, configured to receive a model integrated and updated by integrating the collaborative model and the local model and model training indication information sent by the intelligent network element of the second distributed node, the model training indication information being used to indicate that subsequent model training and updating of the first distributed node need to be performed at the second distributed node.

[0500] In an embodiment, the apparatus further includes:

[0501] a second distributed node determining module, configured to determine the second distributed node based on pre-maintained network node neighbor information.

[0502] an information sending module, configured to send, to the intelligent network element of the second distributed node, data that cannot be trained by the first distributed node, node information of the first distributed node, a local model and parameters trained by the first distributed node, and a collaborative training request, to instruct the second distributed node to perform model training on the data that cannot be trained by the first distributed node.

[0503] In an embodiment, the second distributed node determining module is further configured to, if receiving a rejection of the collaborative request sent by the intelligent network element of the second distributed node, determine a new second distributed node based on the pre-maintained network node neighbor information again.

[0504] Embodiments of the present application further provide a model training apparatus, which includes:

[0505] a data receiving module, configured to receive node data of the first distributed node; the first distributed node is any one of distributed nodes in a distributed network;

[0506] a second training module, configured to collaboratively perform model training of the first distributed node according to the node data.

[0507] In an embodiment, the data receiving module is further configured to receive the node data sent by the intelligent network element of the center node, or receive the node data sent by the intelligent network element of the first distributed node.

[0508] In an embodiment, the node data includes data that cannot be trained by the first distributed node, node information of the first distributed node, a local model and parameters trained by the first distributed node.

[0509] In an embodiment, the second training module comprises:

[0510] The cooperative model training unit is configured to perform cooperative model training on the untrainable data.

[0511] In an embodiment, the first distributed node supports update training, and the apparatus further comprises:

[0512] The first indication sending module is configured to send the integrated model and an update indication to the intelligent network element of the first distributed node, instructing the intelligent network element of the first distributed node to update the integrated model.

[0513] In an embodiment, the first distributed node does not support update training, and the apparatus further comprises:

[0514] The second integration and update module is configured to integrate the cooperative model and the local model, and update the integrated model.

[0515] The second indication sending module is configured to send the integrated and updated model and model training indication information to the intelligent network element of the first distributed node, the model training indication information being used to instruct the first distributed node to perform subsequent model training and update in the second distributed node.

[0516] In an embodiment, the apparatus further comprises:

[0517] The rejection cooperation module is configured to, in response to the cooperative training request sent by the intelligent network element of the first distributed node, send a rejection cooperation request to the intelligent network element of the first distributed node, the rejection cooperation request being used to instruct the first distributed node to determine a new second distributed node based on the pre-maintained network node neighbor information.

[0518] It should be noted that the implementation principles, processes, and technical effects of the various model training apparatuses provided in the embodiments of the present application are similar to those of the foregoing model training methods, and the same parts and beneficial effects of the model training apparatus embodiments as the model training method embodiments will not be described in detail.

[0519] It should be noted that the division of the modules, units, and sub-units in the embodiments of the present application is illustrative, and is merely a logical function division. In actual implementation, another division manner can be used. In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0520] The integrated unit described above, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a processor-readable storage medium. Based on such an understanding, the technical solutions of the present application essentially or the part of the prior art that contributes to the present application or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in the various embodiments of the present application.

[0521] It should be noted that the above-described apparatus provided by the embodiments of the present application can implement all the method steps achieved by the method embodiments and achieve the same technical effects. Therefore, the same parts and beneficial effects of the method embodiments will not be described in detail.

[0522] In one embodiment, a processor-readable storage medium storing a program for causing a processor to execute any one of the model training method embodiments is also provided. The processor-readable storage medium can be any available medium or data storage device that can be accessed by a processor, including but not limited to a magnetic storage (e.g., floppy diskette, hard disk, magnetic tape, MO, etc.), an optical storage (e.g., CD, DVD, BD, HVD, etc.), and a semiconductor memory (e.g., ROM, EPROM, EEPROM, NAND FLASH, SSD, etc.).

[0523] In one embodiment, a computer program product is also provided, which, when executed by a processor, can implement any one of the model training methods. The computer program product includes one or more computer instructions. When loaded and executed on a computer, these computer instructions can implement part or all of the above methods according to the processes or functions described in the embodiments of the present application.

[0524] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage, etc.) containing computer-usable program code.

[0525] Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A model training method, characterized in that, The method applied to a central node comprises: According to the node data of the first distributed node, the method cooperates with the first distributed node to perform model training.

2. The method of claim 1, wherein, Before the model training according to the node data of the first distributed node and cooperating with the first distributed node, the method comprises: Receiving the node data sent by the intelligent network element of the first distributed node.

3. The method of claim 1, wherein, Before the model training according to the node data of the first distributed node and cooperating with the first distributed node, the method comprises: Sending a data reporting request to the intelligent network element of the first distributed node, the data reporting request being used to instruct the intelligent network element of the first distributed node to report the node data; Receiving the node data sent by the intelligent network element of the first distributed node.

4. The method according to any one of claims 1 to 3, characterized in that, The node data comprises: The local model and parameters trained by the first distributed node, the node information of the first distributed node, or The local model and parameters trained by the first distributed node, the node information of the first distributed node, the data that cannot be trained by the first distributed node and the cooperative training request.

5. The method according to any one of claims 1 to 3, characterized in that, The model training according to the node data and cooperating with the first distributed node comprises: According to the node data, the local model of the first distributed node is updated.

6. The method of claim 5, wherein, The node data comprises the local model and parameters trained by the first distributed node and the node information of the first distributed node; According to the node data, the local model of the first distributed node is updated, comprising: Based on the local model and parameters, the node information and the global model of the distributed network, the local model update parameters of the first distributed node are determined; According to the local model update parameters, the local model of the first distributed node is updated.

7. The method of claim 6, wherein, The first distributed node supports model update training; according to the local model update parameters, the local model of the first distributed node is updated, comprising: The local model update parameters are sent to the intelligent network element of the first distributed node, instructing the intelligent network element of the first distributed node to update and train the local model.

8. The method of claim 6, wherein, The first distributed node does not support model update training; according to the node data, the local model of the first distributed node is updated, comprising: According to the local model update parameters, the local model of the first distributed node is updated; The model and model training instruction information obtained by the update training are sent to the intelligent network element of the first distributed node, and the model training instruction information is used to instruct the subsequent model training and update of the first distributed node to be performed in the central node.

9. The method of claim 6, wherein, The method further comprises: Obtaining the local model reported by each distributed node in the distributed network; According to the local model reported by each distributed node, a weighted aggregation operation is performed to obtain the global model of the distributed network.

10. The method of claim 9, wherein, The method further comprises: Detecting whether the global model converges or not; In the case that the global model does not converge, the process of model training cooperated with the first distributed node is cyclically performed until the latest obtained global model converges.

11. The method of claim 5, wherein, The model training cooperated with the first distributed node according to the node data further comprises: Model training is performed on the data that the first distributed node cannot train according to the node data.

12. The method of claim 11, wherein, The node data further comprises the data that the first distributed node cannot train and a cooperative training request; the model training performed on the data that the first distributed node cannot train according to the node data comprises: In response to the cooperative training request, cooperative model training is performed on the data that cannot be trained.

13. The method of claim 12, wherein, The first distributed node supports model update training; the method further comprises: Integrating the trained cooperative model and the local model; Sending the integrated model and an update instruction to the intelligent network element of the first distributed node, instructing the intelligent network element of the first distributed node to update the integrated model.

14. The method of claim 12, wherein, The first distributed node does not support model update training; the method further comprises: Integrating the cooperative model and the local model, and updating the integrated model; Sending the integrated and updated model and model training indication information to the intelligent network element of the first distributed node, the model training indication information being used to instruct the first distributed node to perform subsequent model training and update at the center node.

15. The method of claim 11, wherein, The node data further comprises the data that the first distributed node cannot train and a cooperative training request; the model training performed on the data that the first distributed node cannot train according to the node data comprises: In response to the cooperative training request, a second distributed node is determined according to the node information of other distributed nodes; the second distributed node represents a distributed node capable of cooperative model training on the data that cannot be trained.

16. The method of claim 15, wherein, The method further comprises: Sending the identification information of the second distributed node to the intelligent network element of the first distributed node.

17. The method of claim 16, wherein, The method further comprises: Sending the data that cannot be trained, the node information of the first distributed node, the local model and parameters to the intelligent network element of the second distributed node.

18. The method of claim 16, wherein, The method further comprises: Sending a data sending notification to the intelligent network element of the first distributed node; the data sending notification is used to instruct the intelligent network element of the first distributed node to send the data that cannot be trained, the node information of the first distributed node, the local model and parameters to the second distributed node according to the identification information.

19. The method of claim 11, wherein, Before the model training performed on the data that the first distributed node cannot train according to the node data, the method further comprises: Determining whether the first distributed node can complete model training according to the computing power information of the first distributed node; In the case that the first distributed node cannot complete model training, requesting the data that the first distributed node cannot train from the first distributed node.

20. A model training method comprising: An intelligent network element applied to a first distributed node, the method comprising: sending node data to an intelligent network element of a center node; the node data is used for the intelligent network element of the center node to cooperatively conduct model training of the first distributed node; the first distributed node is any one of distributed nodes in a distributed network.

21. The method of claim 20, wherein, The sending node data to the intelligent network element of the center node comprises: receiving a data reporting request sent by the intelligent network element of the center node; sending the node data to the intelligent network element of the center node according to the data reporting request.

22. The method of claim 20 or 21, wherein, The node data comprises: the local model and parameters trained by the first distributed node, node information of the first distributed node; or the local model and parameters trained by the first distributed node, node information of the first distributed node, untrainable data of the first distributed node and a cooperative training request.

23. The method of claim 20 or 21, wherein, The method further comprises: collecting model training related data in the distributed network; conducting local model training according to the related data.

24. The method of claim 23, wherein, The collecting related data in the distributed network comprises: sending a data collection request to each network function (NF) in a core network and each network function in an access network in the first distributed node; receiving response data sent by each network function to obtain the model training related data.

25. The method of claim 23, wherein, The collecting model training related data in the distributed network comprises: sending a data collection request to each data plane in the first distributed node; receiving response data sent by each data plane to obtain the model training related data.

26. The method of claim 20 or 21, wherein, The method further comprises: receiving a local model update message sent by the intelligent network element of the center node.

27. The method of claim 26, wherein, The local model update message comprises local model update parameters, and the method further comprises: updating the local model according to the local model update parameters.

28. The method of claim 26, wherein, The local model update message comprises an updated local model and model training indication information, and the model training indication information is used to indicate that subsequent model training and updating of the first distributed node need to be conducted at the center node.

29. The method of claim 26, wherein, The method further comprises: receiving an integrated model and update indication sent by the intelligent network element of the center node; the integrated model is obtained by integrating a cooperative model and a local model by the intelligent network element of the center node; the cooperative model is trained according to untrainable data of the first distributed node; updating the integrated model.

30. The method of claim 26, wherein, The method further comprises: receiving a model and model training indication information sent by the intelligent network element of the center node, the model being an integrated and updated model of a cooperative model and a local model, and the model training indication information being used to indicate that subsequent model training and updating of the first distributed node need to be conducted at the center node.

31. The method of claim 26, wherein, The method further comprises: receiving identification information of a second distributed node sent by the intelligent network element of the center node; the second distributed node represents a node for training a cooperative model according to untrainable data of the first distributed node.

32. The method of claim 31, wherein, The method further comprises: In response to the data sending notification sent by the intelligent network element of the center node, the untrainable data, the node information of the first distributed node, the local model and the parameters are sent to the intelligent network element of the second distributed node according to the identification information.

33. The method of claim 32, wherein, The method further comprises: receiving the integrated model and the update indication sent by the intelligent network element of the second distributed node; the integrated model is obtained by integrating the collaborative model and the local model by the intelligent network element of the second distributed node; updating the integrated model.

34. The method of claim 32, wherein, The method further comprises: receiving the model after the integration and the update of the collaborative model and the local model and the model training indication information sent by the intelligent network element of the second distributed node, and the model training indication information is used to indicate that the subsequent model training and update of the first distributed node need to be performed at the second distributed node.

35. The method of claim 26, wherein, Before the receiving of the local model update parameters sent by the intelligent network element of the center node, the method further comprises: determining the second distributed node based on the pre-maintained network node neighbor information; sending the untrainable data of the first distributed node, the node information of the first distributed node, the trained local model and the parameters of the first distributed node, and the collaborative training request to the intelligent network element of the second distributed node to indicate the second distributed node to perform model training on the untrainable data of the first distributed node.

36. The method of claim 35, wherein, The method further comprises: if the collaborative request is rejected by the intelligent network element of the second distributed node, a new second distributed node is determined based on the pre-maintained network node neighbor information.

37. A model training method, comprising: The method applied to the intelligent network element of the second distributed node comprises: receiving the node data of the first distributed node; the first distributed node is any one of the distributed nodes in the distributed network; collaboratively performing model training of the first distributed node according to the node data.

38. The method of claim 37, wherein, The receiving of the node data of the first distributed node comprises: receiving the node data sent by the intelligent network element of the center node; or receiving the node data sent by the intelligent network element of the first distributed node.

39. The method of claim 37 or 38, wherein, The node data comprises the untrainable data of the first distributed node, the node information of the first distributed node, the trained local model and the parameters of the first distributed node.

40. The method of claim 39, wherein, Collaboratively performing model training of the first distributed node according to the node data of the first distributed node comprises: performing collaborative model training on the untrainable data.

41. The method of claim 40, wherein, The first distributed node supports update training, and the method further comprises: sending the integrated model and the update indication to the intelligent network element of the first distributed node to indicate the intelligent network element of the first distributed node to update the integrated model.

42. The method of claim 40, wherein, The first distributed node does not support update training, and the method further comprises: integrating the collaborative model and the local model, and updating the integrated model; The integrated and updated model and model training instruction information are sent to the intelligent network element of the first distributed node, and the model training instruction information is used to indicate that subsequent model training and updating of the first distributed node are both performed at the second distributed node.

43. The method of claim 37 or 38, wherein, The method further includes: In response to the cooperative training request sent by the intelligent network element of the first distributed node, a rejection cooperative request is sent to the intelligent network element of the first distributed node, and the rejection cooperative request is used to instruct the first distributed node to determine a new second distributed node based on the pre-maintained network node neighbor information.

44. A model training apparatus, comprising: The apparatus includes: A first training module is configured to perform model training in cooperation with the first distributed node according to node data of the first distributed node, and the first distributed node is any one of the distributed nodes in the distributed network.

45. A model training apparatus, comprising: The apparatus includes: A node data sending module is configured to send node data to an intelligent network element of a center node, and the node data is used for the intelligent network element of the center node to perform model training in cooperation with the first distributed node, and the first distributed node is any one of the distributed nodes in the distributed network.

46. A model training apparatus, comprising: The apparatus includes: A data receiving module is configured to receive node data of a first distributed node, and the first distributed node is any one of the distributed nodes in the distributed network. A second training module is configured to perform model training in cooperation with the first distributed node according to the node data.

47. A processor-readable storage medium, comprising: The processor readable storage medium stores a program, and the program is used to make the processor execute the method in any one of claims 1 to 43.

Images