AI / ML model training method and product
The AI/ML model training method controlled by user terminals solves the data privacy and security risks caused by network-side training and improves the generalization performance and data security of model training.
Patent Information
- Application Number
- CN202410316989.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-20
- Publication Date
- 2025-09-23
AI Technical Summary
In existing technologies, AI/ML model training for R18 air interfaces is performed entirely by the network side, resulting in a lack of control over the model training process by user terminals and posing data privacy and security risks.
The user terminal can control the model training process according to its product characteristics. By collecting and uploading data to the server configured on the network side, the authorized user controls the server to perform model training, and the network side provides hardware capabilities and data transmission support.
The generalization performance and data privacy security of model training have been improved. User terminal manufacturers can conduct model training based on product characteristics, reducing data privacy and security risks.
Smart Images
Figure CN120692531A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of communication technologies, and more specifically, to an AI / ML model training method and product. Background Art
[0002] In existing technologies, AI / ML model training over the R18 air interface is performed by the network. The network performs centralized model training based on collected data, providing operational-grade hardware capabilities and compilation efficiency. The network also intervenes in data transmission, reducing data privacy and security risks. However, because model training is entirely performed by the network, vendors' user terminals lack control over the model training process. Summary of the Invention
[0003] In order to solve the above problems, the present invention proposes an AI / ML model training method and product, and the user terminal can control the model training according to its product characteristics.
[0004] An embodiment of the present invention provides an AI / ML model training method, the method comprising:
[0005] Collect data uploaded by user terminals and upload the data to a server configured on the network side;
[0006] The user who configures the data usage permission controls the server to use the data for model training.
[0007] Preferably, before the authorized user controls the server to perform model training, the method further includes:
[0008] The server receives uploaded network-side configuration information, condition information, and scene auxiliary information.
[0009] As a preferred solution, the data used by the authorized user for model training includes all data within the scope permitted by the data usage authority.
[0010] Preferably, before the user terminal uploads data, the method further includes:
[0011] The user terminal acquires data and converts the data into a standardized format to obtain the data; or
[0012] The user terminal obtains the data according to a standardized data format.
[0013] Preferably, the collecting of data uploaded by the user terminal includes:
[0014] Collect data uploaded by user terminals through data collection methods;
[0015] The data collection method includes at least one of the following: Logged MDT, Immediate MDT, L1 measurement, and L3 measurement.
[0016] Preferably, the user terminal includes the uploaded data in an RRC container, and / or,
[0017] The user terminal uploads data through the CP channel or the user terminal uploads data through the UP channel.
[0018] As a preferred solution, before uploading the data to the server configured on the network side, the method further includes:
[0019] The configuration information, condition information and scenario auxiliary information of the RAN are added to the data.
[0020] Preferably, the method further comprises:
[0021] The server receives permission configuration data and configures data use permissions for different user terminals according to the permission configuration data.
[0022] Preferably, the network side includes a core network and / or Ethernet OAM and / or a server.
[0023] An embodiment of the present invention provides a computer program product, including a computer program / instruction, which implements the steps of any method described in the above embodiments when executed by a processor.
[0024] The present invention provides an AI / ML model training method and product. The method includes: collecting data uploaded by user terminals and uploading the data to a server configured on the network side; and configuring data usage permissions so that users can control the server to use the data for model training. User terminals can control model training based on their product characteristics. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 This is a flowchart of an AI / ML model training method provided by an embodiment of the present invention;
[0026] Figure 2 This is another flowchart of the AI / ML model training method provided by an embodiment of the present invention;
[0027] Figure 3 This is another flowchart of the AI / ML model training method provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.
[0029] The present invention provides an AI / ML model training method. Figure 1 , is a flow chart of an AI / ML model training method provided by an embodiment of the present invention, the method comprising the following steps:
[0030] Step S1, collecting data uploaded by the user terminal and uploading the data to a server configured on the network side;
[0031] Step S2, the authorized user who configures the data usage permission controls the server to use the data for model training.
[0032] When this embodiment is specifically implemented, see Figure 2 , is another flow chart of the AI / ML model training method provided by an embodiment of the present invention. Figure 2 In the process, the user terminal UE uploads data to the radio access network RAN, the radio access network RAN collects the data uploaded by different user terminals, and uploads the data to the server configured by the network NW, and the server stores the uploaded data.
[0033] It should be noted that the server in this embodiment is specifically a server used for model training.
[0034] It should be noted that the data collected in this embodiment specifically includes data used for model training.
[0035] It should be noted that in this embodiment, the wireless access network is used as the execution entity to perform data collection. In other embodiments, the core network or a third-party server may also perform data collection.
[0036] It should be noted that authorized users include user terminals, and may also be "operators" and / or "chip manufacturers" and / Or "any user who has configured data usage rights". When model training is required, the user who has pre-configured data usage rights sends a control instruction to the server configured on the network NW to control the server to use the data for model training.
[0037] When implementing this solution, the server will be configured on the network NW.
[0038] The server is configured in the network NW, which performs unified model training based on the collected data. This is conducive to integrating data sets collected by multiple user terminal manufacturers for model training, and can provide operational-level hardware capabilities and compilation efficiency for model training. In addition, the network NW intervenes in data transmission, reducing data privacy and security risks.
[0039] It should be noted that in this embodiment, the network side device uploads the configuration information, condition information and auxiliary information of the network device, which is a preferred implementation method. Other third-party servers that can obtain the above information can also upload the above information to the server.
[0040] The control process of model training is executed by the user terminal, which avoids the problem that when the server training is configured on the network side, the model training is completely performed by the network side, and the user terminal device lacks control over the model training process, resulting in low support for this solution from the user terminal and chip manufacturers. When the model of this application solution is trained, the user terminal can control the model training according to its product characteristics.
[0041] In another embodiment provided by the present invention, before the authorized user controls the server configured on the network side to perform model training, the network side may also participate in the model training process.
[0042] It should be noted that the network NW is a preferred embodiment of the network side and performs the functions of the network side in this application.
[0043] That is, the network side uploads its network side configuration information, condition information and auxiliary information of the scene to the server.
[0044] The network NW provides relevant configuration, conditions, and scenario-related auxiliary information for the model training process, improving the generalization of the trained model. While the network training process is controlled by the user terminal (UE), the network NW can also participate in the model training process, avoiding the drawbacks of training being solely performed by the user terminal (UE) or the network NW. While providing operational-level data privacy and security, each user terminal manufacturer can combine the unique hardware, firmware, and software features of their products for model training.
[0045] In another embodiment provided by the present invention, see Figure 3 , is another flowchart of the AI / ML model training method provided by an embodiment of the present invention.
[0046] When the server configured by the network NW performs model training, the data used includes all data within the scope permitted by the data usage authority.
[0047] When the network NW performs model training, all data within the scope of data usage permissions can be used. Not only can the data uploaded by the user terminal be used, but also the configuration information, condition information, auxiliary information of the scene uploaded by the network side, and other network-side information can be used.
[0048] The data used by authorized users includes data uploaded by all user terminals that have been assigned the data usage permission. That is, the data used by authorized users includes the first data Date1 uploaded by the first user terminal UE1, the second data Date2 uploaded by the second user terminal UE2, and the third data Date3 uploaded by the third user terminal UE3, all within the scope of the data usage permission.
[0049] Therefore, during data collection, the first user terminal UE1 uploads its first data Date1 to the radio access network RAN, the second user terminal UE2 uploads its second data Date2 to the radio access network RAN, and the third user terminal UE3 uploads its third data Date3 to the radio access network RAN. The radio access network RAN collects the first data Date1, the second data Date2 and the third data Date3, and uploads the collected data to the server configured by the network NW.
[0050] Data collected on the network-side server is available to all UE manufacturers that report data for use in model training. Each UE manufacturer can use data reported by its own UE, as well as data shared by multiple UE manufacturers within the scope permitted by the agreement. Data on this server is only used for model training and cannot be downloaded or stored outside of the server.
[0051] Users with configured data usage permissions use the data reported by them and the data reported by other user terminals that have configured the data usage permissions to train models, which can improve the breadth, universality, and completeness of the data used for training, thereby improving the generalization performance of the model.
[0052] In another embodiment provided by the present invention, the user terminal needs to standardize the data before uploading the data, that is, the user terminal obtains the data and converts it into a standardized data transmission format to obtain the data.
[0053] Alternatively, the user terminal obtains data in a standardized data format as the data, that is, the user terminal obtains data in a standardized data format and does not need to convert the data into a standardized data transmission format after obtaining the data.
[0054] The format and content of data collected from user terminals (UEs) to the network (NW) need to be standardized. User terminals upload data in a standardized data transmission format. The network (NW) conducts unified model training based on the collected data. This allows data to be obtained from as many user terminals (UEs) as possible, and ensures that the data collected by the network (NW) reaches the server and is correctly identified and used by the server, thereby improving model training efficiency.
[0055] In another embodiment of the present invention, when the wireless access network collects the data uploaded by the user terminal, the data uploaded by the user terminal is collected using a preset data collection method.
[0056] See also Figure 3 The radio access network (RAN) needs to collect data uploaded by different user terminals. It reuses several data collection methods currently confirmed by the RAN2. RAN2 is the working group within 3GPP responsible for developing technical standards for radio interface architecture and protocols. During the 5G standardization process, it developed numerous technical specifications and documents related to these architectures and protocols.
[0057] The data collection method includes at least one of the following: Logged MDT, Immediate MDT, L1 measurement, and L3 measurement.
[0058] Minimization Drive Test (MDT) is an automated drive test technology introduced by 3G UMTS / 4G LTE / 5G NR that collects, reports, and pre-processes measurement data by configuring ordinary user / commercial terminals on the network.
[0059] As long as the user terminal has GPS enabled and supports MDT, it can automatically report MDT data containing user location information to the base station. MDT, similar to MR (Measurement Report), contains fields such as RSRP and RSRQ, as well as GPS latitude and longitude information, which can be used for big data analysis.
[0060] Logged MDT records MDT-related measurements in the idle state. Periodic measurements are performed to determine the terminal's received power strength, including basic coverage RSRP / RSRQ, as well as RF fingerprint and GNSS location information. The user terminal associates the available detailed location information with the measurement records / reports for the next measurement cycle. Each location information is valid only once. The user terminal collects measurement data in the idle and inactive states. After the measurement, the UE temporarily stores the measurement results in the UE memory in the form of a log file and reports them the next time the UE enters the connected state.
[0061] It should be noted that, in this embodiment, the radio access network collects the data uploaded by the user terminal by multiplexing the data collection method of RAN2. In other embodiments, the radio access network may also collect the data uploaded by the user terminal by other methods.
[0062] Reuse several data collection methods currently confirmed by RAN2, allowing UE and UE to jointly intervene and participate in the model training process.
[0063] In another embodiment of the present invention, the user terminal includes the uploaded data in an RRC container, and / or,
[0064] The user terminal uploads data through the CP channel or the user terminal uploads data through the UP channel.
[0065] Specifically, the data of the user terminal is included in the RRC container, and the user terminal uploads the communication data through the data flow and uploads the data through the CP channel.
[0066] That is, the user terminal uploads data to the radio access network RAN through the CP channel, and the radio access network RAN uploads data to the network NW.
[0067] Alternatively, the user terminal uploads data through the UP channel.
[0068] Specifically, the data of the user terminal is uploaded through the UP channel.
[0069] That is, the user terminal uploads data to the radio access network RAN through the UP channel, and the radio access network RAN uploads data to the network NW.
[0070] The CP channel operates in CP mode, also known as the control plane, signaling radio bearer plane, or control plane bearer. This also places user data within the control plane data (specifically, the NAS layer) and sends it out. The advantage is that when the data volume is small, the transmission speed is fast.
[0071] The UP channel is carried out in UP mode, also known as the user plane, data plane, or data radio bearer. User data and control data are separated, and this path carries only user data. This has the advantage of fast transmission when there is a lot of data.
[0072] In another embodiment of the present invention, before the radio access network uploads the data to a server configured on the network side, the radio access network adds RAN configuration information, condition information, and scenario auxiliary information to the data.
[0073] Specifically, the radio access network RAN and the core network CN or OAM of the network NW are both control nodes for data collection, and can both obtain the data information content transmitted by the user terminal. In addition, the network NW can still accurately control the consumption of network resources during the data collection process.
[0074] When the data uploaded by the user terminal reaches the radio access network (RAN), the RAN processes the data reported by the user terminal. At the same time, the RAN can also attach auxiliary information related to the RAN configuration, conditions, and scenarios to the corresponding data set, and finally transmit the user terminal data to the core network (CN) or the data modeling server within the OAM.
[0075] In another embodiment of the present invention, the data usage rights are configured by the server according to the received rights configuration data, specifically:
[0076] The server receives the permission configuration data sent by the operator; and configures data use permissions for different user terminals according to the permission configuration data.
[0077] The server is controlled by the operator, who grants data usage rights to the user terminals of the manufacturers that participate in reporting data, that is, sends permission configuration data to the server;
[0078] The server receives the permission configuration data and configures data usage permissions for different user terminals according to the permission configuration data. The permission configuration data includes data usage permissions for different user terminals, that is, whether it supports user terminals of manufacturers with the same permissions to use their data when conducting model training.
[0079] Each manufacturer's user terminals can use the data reported by their own manufacturer's user terminals and the data shared by user terminals of multiple manufacturers within the scope permitted by the agreement. This can improve the breadth, universality and completeness of the data used for training, thereby improving the generalization performance of the model.
[0080] In another embodiment of the present invention, the network side includes a core network and / or Ethernet OAM and / or a server.
[0081] That is, the network side is specifically configured in the core network CN or Ethernet OAM or server, and can also be configured in two or three of them at the same time.
[0082] The "server" here refers to the "server" that can be controlled by the operator, rather than a third-party server that is not the operator.
[0083] This application solution integrates a server into the core network (CN) and / or Ethernet OAM on the network side, and performs model training by granting data access permissions to user terminals. This reuses several data collection methods currently confirmed by RAN2, allowing user terminals and the network to jointly participate in the model training process. While providing operational-level data privacy and security, user terminals from various manufacturers can utilize the unique hardware, firmware, and software features of their products for model training.
[0084] An embodiment of the present invention provides a computer program product, including a computer program / instruction, which implements the steps of any method described in the above embodiments when executed by a processor.
[0085] It should be noted that the functions to be performed and the technical effects achieved by the computer program product provided in this embodiment correspond to the AI / ML model training method provided in any of the above embodiments, and will not be elaborated here.
[0086] It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. An AI / ML model training method, characterized in that: The method comprises: Collect data uploaded by user terminals and upload the data to a server configured on the network side; The user who configures the data usage permission controls the server to use the data for model training.
2. The AI / ML model training method according to claim 1, characterized in that: Before the authorized user controls the server to perform model training, the method further includes: The server receives uploaded network-side configuration information, condition information, and scene auxiliary information.
3. The AI / ML model training method according to claim 1, characterized in that The data used by the authorized user for model training includes all data within the scope permitted by the data usage authority.
4. The AI / ML model training method according to claim 1, characterized in that Before the user terminal uploads data, the method further includes: The user terminal acquires data and converts the data into a standardized format to obtain the data; or The user terminal obtains the data according to a standardized data format.
5. The AI / ML model training method according to claim 1, characterized in that The collecting of data uploaded by the user terminal includes: Collect data uploaded by user terminals through data collection methods; The data collection method includes at least one of the following: Logged MDT, Immediate MDT, L1 measurement, and L3 measurement.
6. The AI / ML model training method according to claim 1, characterized in that The user terminal includes the uploaded data in an RRC container, and / or The user terminal uploads data through the CP channel or the user terminal uploads data through the UP channel.
7. The AI / ML model training method according to claim 1, characterized in that: Before uploading the data to the server configured on the network side, the method further includes: The configuration information, condition information and scenario auxiliary information of the RAN are added to the data.
8. The AI / ML model training method according to claim 1, characterized in that: The method further comprises: The server receives permission configuration data and configures data use permissions for different user terminals according to the permission configuration data.
9. The AI / ML model training method according to claim 1, characterized in that: The network side includes a core network and / or Ethernet OAM and / or a server.
10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.