Methods and apparatus of region-based ai model update
The region-based AI model update framework addresses performance degradation by defining and switching AI models based on predefined regions, ensuring effective AI model updates across different areas.
Patent Information
- Application Number
- PCT/CN2024/082266
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-03-18
- Publication Date
- 2025-09-25
AI Technical Summary
Existing AI models in mobile wireless communication systems deteriorate in performance when devices move outside their suitable regions due to limited computational capacity, making it difficult to apply online training or pre-trained models effectively.
A region-based framework and procedure for updating AI models in wireless networks, defined by cell lists or geographical regions, allowing devices to switch or update models based on predefined regions, using UE or RAN nodes for identification and model activation/deactivation.
Enables efficient AI model updates across different regions, maintaining performance by leveraging predefined regions and network nodes for timely model switching, applicable to 5G and beyond.
Smart Images

Figure CN2024082266_25092025_PF_FP_ABST
Abstract
Description
METHODS AND APPARATUS OF REGION-BASED AI MODEL UPDATEFIELD
[0001] The present disclosure relates generally to communication systems, and more particularly, the method and apparatus of region-based AI model update.BACKGROUND
[0002] Artificial Intelligence (AI) and Machine Learning (ML) have permeated a wide spectrum of industries, ushering in substantial productivity enhancements. In the realm of mobile communications systems, these technologies are orchestrating transformative shifts. Mobile devices are progressively supplanting conventional algorithms with AI-ML models.
[0003] One key challenge in applying AI for mobile wireless communication is maintaining appropriate AI models when considering the mobility of the device. Due to the limitation of the AI model generation, the performance of the AI model will decrease when devices move outside the suitable region of the AI model.
[0004] One solutionis to fine-tune the parameter or apply the online training approach such that AI model can be modified when the environment is changed. However, these approaches require high computation capability and are difficult to apply to devices with limited computation capacity, especially for implementations that require AI to make decisions or predictions in milliseconds. In contrast, another approach assumes the AI model is pre-trained offline, and devices will update and / or switch AI model when the device exceeds the appropriate region of the model.
[0005] In this invention, apparatus and mechanisms are sought to provide the framework and the corresponding procedure to maintain the update of the AI model when the device moves across different regions.SUMMARY
[0006] The following presents a simplified summary of one or more aspects to provide a basic understanding of such aspects. This summary is not an extensive overview of all contemplated aspects, and is intended to neither identify key or critical elements of all aspects nor delineate the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to the more detailed description that is presented later.
[0007] Apparatus and methods are provided for region-based AI-ML model update in the wireless network. The apparatus can be a UE or any moving device. In one novel aspect, the AI model is updated based on the predefined regions and the region to which the UE belongs. In one embodiment, the region is defined by the cell list where one region contains one or multiple cell ID. The UE belongs to a region if its serving cell is within the cell list of the corresponding region. In different embodiments, the region is defined by one or more of the following information, including UE location / position, gNB coverage, serving cell id, available measurement at the UE, available AI mode input, and other information that can describe the concept of regions. Different regions can be disjoint or overlapping, and some places in the entire space are allowed that do not belong to any defined region.
[0008] The general framework and corresponding procedure are provided to update AI-ML model when UE moves across different regions. The overall region-based AI-ML model update procedure contains region definition, region identification, candidate AI model update, and AI model update. In one embodiment, the region definition procedure contains an AI-ML model / dataset center, which can be a UE server, network server, OAM, or OTT server, performs the dataset and labeled data collection, AI-ML model training. Based on the collected data and training results, the model / dataset center provides the description and / or definition of region and the mapping between the region and the corresponding AI model. The model / dataset center will deliver / coordinate the region definition and mapping to the component that performs the region identification procedure. In one embodiment, the region identification is performed by a RAN node, which can be a gNB, LCM, OAM, . . etc. The RAN node sends the region report configuration to the UE and UE performs the region report to the RAN node by following the configuration. The RAN node identifies the region to which the UE belongs and decides to send the model update command. In one embodiment, when UE receives the model update command, the UE selects the model from the pre-downloaded candidate model list based on the model update command. The UE deactivates the original AI model and activates the new AI model for the corresponding use case. In another embodiment, UE performs the model update without any pre-downloaded candidate model list. The UE sends the model update request to the model / dataset center to download the corresponding AI model. The UE deactivates the original model and activates the new AI model.
[0009] In one embodiment, the region identification is performed by a UE, and the AI model update is triggered by UE.
[0010] In one embodiment, the region-based AI-ML model update procedure does not contain candidate models and corresponding procedures.
[0011] The invention is not limited to 5G NR, it may also apply to other communication systems, such as 6G or further-generation communication systems
[0012] To the accomplishment of the foregoing and related ends, the one or more aspects comprise the features hereinafter fully described and particularly pointed out in the claims. The following description and the annexed drawings set forth in detail certain illustrative features of the one or more aspects. These features are indicative, however, of but a few of the various ways in which the principles of various aspects may be employed, and this description is intended to include all such aspects and their equivalents.BRIEF DESCRIPTION OF THE DRAWINGSDETAILED DESCRIPTION
[0013] Figure 1 illustrates an exemplary scenario where the UE updates the AI-ML model according to the pre-defined region. The region definition and the mapping between the region and AI model are provided by the model / dataset center. In one embodiment, the center can be UE server, NW server, OAM, or OTT server. In one embodiment, the definition of regions is described by a cell list that contains one or multiple cell IDs. For those UE with serving cell ID in the cell list is referred to the corresponding region. In different embodiments, the region is defined by one or more of the following information, including UE location / position, gNB coverage, serving cell id, available measurement at the UE, available AI mode input, and other information that can describe the concept of regions. In one embodiment, the region and AI model are identified by region ID and AI model ID, respectively.
[0014] The model / dataset center delivers the related information of region definition to the node that performs region identification. In one embodiment, the region identification is made by the RAN node, which can be LCM, OAM, CN, or gNB. The information of region definition can be delivered from the model / dataset center to the RAN node via a standardized or non-standardized interface. The RAN node receives the necessary information reported from UE and identifies the region to which the UE belongs. In one embodiment, the information includes existing measurement reports, assistance information, and / or other new pre-configured region reports. Based on the region to which the UE belongs, the RAN node triggers the model update procedure. In one embodiment, UE updates the AI-ML model from the candidate model list which is pre-downloaded. In another embodiment, UE updates the AI-ML model by the model which downloaded from the model / dataset center.
[0015] In one embodiment, the regions can be defined without overlapping. Figure 2 illustrates an exemplary region definition where region 1, 2, and 3 are disjoint. The region identification can be made based on the region to which the UE belongs.
[0016] In another embodiment, the regions can be defined with overlapping. Some places are allowed that do not belong to any pre-defined regions. Figure 3 illustrates an exemplary region definition where two different regions, e.g., region 1 and 3, are overlapping. Figure 3 also illustrates the case where some place, e.g. location of UE2, does not belong to any pre-defined region.
[0017] Figure 4 illustrates an exemplary general procedure to perform the region-based AI-ML model update. The procedure includes region definition, region identification, and model update.
[0018] The region definition refers to the method to describe the boundary of the regions. It can be defined based on different physical quantities and / or different abstract concepts. In one embodiment, the definition of regions is described by a cell list that contains one or multiple cell IDs. For those UE with serving cell ID in the cell list is referred to the corresponding region. In another embodiment, the region is defined by latitude and longitude. In another embodiment, the region is defined by a range within a reference point, which can be a gNB. In another embodiment, the region is defined by available AI model input on the UE.The available AI model input includes one or more following elements, UE side measurement, RSRP, CSI, AoA, TA, UE capability, UE speed, UE location.
[0019] The region identification contains the procedure to identify which region the UE belongs to. In one embodiment, based on the region definition if a UE’s information refers to one and only one region, the RAN node identifies the UE’s region as the corresponding one. If a UE’s information refers to more than one region (e.g., UE3 in Fig 2) . The RAN node can identify UE’s region as one of them. In one embodiment, the UE is identified by the nearest region. In another embodiment, UE can be identified as belonging to multiple regions. If a UE’s information refers to no pre-defined region. The RAN node can identify UE’s region as empty, previous identification, or the nearest region. The nearest region can be determined based on the concept of distance given in the region definition.
[0020] In one embodiment, the distance between one region and UE is the distance between UE and the nearest region boundary. In another embodiment, the distance between a region and a UE is the difference between the available measurements on the UE and the available measurements defined in the region.
[0021] The model update procedure contains the download model and model selection. In one embodiment, the candidate model list is pre-downloaded before the model update procedure is triggered. Once UE receives the update model command, it will update the model by selecting from the candidate list. In another embodiment, when the model update procedure is triggered, UE downloads the new model from the model / dataset. Once the updated model is available, the UE deactivates the original AI model and activates the new AI model for the corresponding AI use cases.
[0022] Figure 5 illustrates an exemplary procedure to region-based AI model update for the case where candidate models are pre-downloaded. The model / dataset center collects the training dataset and labelled data for AI model training and region definition. The AI model region coordination is performed between the model / dataset center and RAN nodes. In one embodiment, the coordination contains one or more elements including region definition, region ID, corresponding AI use cases, AI model ID, the mapping between region and AI model, the description used to identify UE region, etc.
[0023] The UE sends a UE capability report to the RAN node. In one embodiment, the UE capability report contains one or more elements including supported AI use cases, such as AI BM, AI CSI compression, AI positioning, and AI mobility, etc., the number of candidate models that can be stored, UE storage memory, speed, location, position, available assistant information, types and quantities of available measurements, etc.
[0024] The RAN node sends the region report configuration to UE. The configuration provided the information and guide for UE to report UE-related information to RAN nodes. In one embodiment, the configuration contains one or more elements including report periodicity, trigger condition, report type, report content…, etc. Based on the configuration, UE sends the UE region info report to RAN nodes.
[0025] The RAN node performs the region identification when receives the UE region info report. Based on the region identification results, the RAN node sends the model update command to UE if the AI model needs to be updated. In one embodiment, the RAN node sends the update command when UE’s region is changed. In another embodiment, the RAN node sends the update command when the UE is moving out of the suitable region of the current AI model. In another embodiment, the RAN node sends the update command when UE’s serving cell is not included in the cell list of the current region.
[0026] The model update command contains one or multiple elements includes the AI use cases, UE region ID, AI model ID, and the indicator for UE to select model from the pre-downloaded candidate model list. UE performs the mode update where the model is selected from the pre-download candidate model. In one embodiment, UE is identified as a region that refers to one and only one candidate model, UE updates its AI model by the candidate model. In another embodiment, UE belongs to multiple regions that refer to more than one candidate model, UE select the model corresponding to the nearest region. In another embodiment, UE selects the model based on the indicator given in the model update command. Once the updated model is available, the UE deactivates the original AI model and activates the new AI model for the corresponding AI use cases. After updating the new AI model, UE sends the model update complete to the RAN nodes to inform the update procedure is completed.
[0027] The region-based AI model update in this invention provide a general framework and procedure to implement those AI use cases where the AI training and / or inference process relies on the characteristics of the specific region / area. In one embodiment, AI approaches are applied on the UE to predict RRM measurement. The available AI model input can be the previous measurement, e.g., RSRP, of the serving cell and the neighbor cells. The AI model output can be the prediction of measurement, which can be temporal / spatial / frequency domain prediction. In one embodiment, the AI model is pre-trained based on the measurement available on the specific region / area. As UE moves across different regions / areas, e.g. UE handover to different serving cell, the available measurements on the UE are different and can not be utilized as the input of the original AI model. The region-based AI model update provides the method to update the AI model and to support the region-based AI use cases.
[0028] For the region-based AI model update procedure illustrated in Figure 5, the candidate model is assumed to be pre-downloaded. In this invention, a method of updating candidate model list based on the region-based method is also provided. Figure 6 illustrates an exemplary general procedure to perform the region-based candidate model list update. The procedure includes region definition, region identification, and candidate model list update.
[0029] Figure 7 illustrates an exemplary procedure for region-based candidate model list update. The model / dataset center collects the training dataset and labelled data for AI model training and region definition for candidate model update. The candidate model region coordination is performed between the model / dataset center and RAN nodes. In one embodiment, the coordination contains one or more elements including region definition, region ID, corresponding AI use cases, AI model ID, the mapping between region and AI model, the description used to identify UE region, etc..
[0030] The UE sends a UE capability report to the RAN node. In one embodiment, the UE capability report contains one or more elements including supported AI use cases, such as AI BM, AI CSI compression, AI positioning, and AI mobility, etc., the number of candidate models that can be stored, UE storage memory, speed, location, position, available assistant information, types and quantities of available measurements, etc.
[0031] The RAN node sends the region report configuration to UE. The configuration provided the information and guide for UE to send UE-related information to RAN nodes to identify UE’s region. In one embodiment, the configuration contains one or more elements including report periodicity, trigger condition, report type, report content…, etc. Based on the configuration, UE sends the UE region info report to RAN nodes.
[0032] The RAN node performs the region identification when receives the UE region info report. Based on the region identification results, the RAN node sends the candidate model list update command to UE if the candidate model list needs to be updated. In one embodiment, the RAN node sends the update command when UE’s region is changed. In another embodiment, the RAN node sends the update command when the UE is moving out of the suitable region of the current candidate model list. In another embodiment, the RAN node sends the update command when UE’s serving cell is not included in the cell list of the current region.
[0033] The candidate model update command contains one or multiple elements including UE region identification, list of candidate model ID and the indicator of update method, e.g., delta update or direct replace. UE sends the update candidate model list request to the model / dataset center. In one embodiment, the candidate model list request contains one or multiple elements including the information of the UE’s serving cell ID, the UE region, UE’s current candidate model list, and the UE capability, e.g., the number of candidate models that UE can store, UE memory storage, etc. The candidate model update can be a direct replacement or delta update. In one embodiment, UE updates the entire list by downloading new candidate models from the model / dataset center. In another embodiment, UE downloads the model which is the delta part of the original candidate model list and new candidate model list, and updates the list with the new model. After the candidate model list is updated, UE sends the candidate model update complete to the RAN nodes to inform the update procedure is completed.
[0034] In one embodiment, some similar steps in both AI model update and candidate model list update can share the same procedure, and the necessary information for two procedures can be carried in the same configuration or signaling, e.g., data collection, AI model training, region definition, region coordination, UE capability report, region report configuration, UE region info report, and region identification.
[0035] In one embodiment, the candidate model list is not necessary, the region-based AI model update procedure does not require UE to pre-download AI model.
[0036] Figure 8 illustrates an exemplary procedure for region-based AI model update for the case where no pre-download model is available on UE for model update. In one embodiment, the procedure before receiving model update command in the case with the pre-download candidate model list can be reused. After receiving the model update command, UE sends the model download request to the model / dataset center. In one embodiment, the model download request contains one or more elements including the UE’s serving cell ID, the UE region and the UE capability, e.g., the available AI model input, the available UE measurement, UE memory storage, etc. UE then downloads the AI model from the model / dataset center. The UE deactivates the original AI model and activates the new AI model for corresponding AI use cases. After the UE updates the AI model, it sends a model update complete message to the RAN node to inform the update procedure is completed
[0037] Figure 9 illustrates an exemplary procedure for region-based AI model update without network awareness.
[0038] The region definition is performed at the model / dataset center as the same flow for the model update described in Figure 5.
[0039] The UE sends init AI model update request to the model / dataset center to initialize the update procedure. The request contains the UE capability, the number of candidate model stored on UE, the UE’s storage memory, the available measurement on UE, the AI use cases, the available AI model input, etc.
[0040] The model / dataset center delivers the region description to UE, where the description contains one or more elements including region definition, cell list corresponding to the region, the information to identify UE’s region, and the mapping between region ID and AI model ID, etc.
[0041] The UE performs the region identification based on available information which contains one or more elements including serving cell ID, speed, location, and type and number of available measurements, available AI model input, and other assistance information, etc.
[0042] Based on the region identification results, the UE decides to update AI model if necessary. In one embodiment, the model update is triggered when UE’s region is changed. In another embodiment, the model update is triggered when the UE is moving out of the suitable region of the current AI model. In another embodiment, the mode update is triggered when UE’s serving cell is not included in the cell list of the current region.
[0043] In one embodiment, the model update is performed where the model is selected from the candidate model list based on the region identification.
[0044] In another embodiment, the UE sends the model update request to download the corresponding AI model for the case where no candidate model list is available.
[0045] Figure 10 illustrates an exemplary procedure for region-based candidate model list update without network awareness.
[0046] The region definition is performed at the model / dataset center as the same flow for the model update described in Figure 7.
[0047] The UE sends init AI model update request to the model / dataset center to initialize the update procedure. The request contains the UE capability, the number of candidate model stored on UE, the UE’s storage memory, the available measurement on UE, the AI use cases, the available AI model input, etc.
[0048] The model / dataset center delivers the region description to UE, where the description contains one or more elements including region definition, cell list corresponding to the region, the information to identify UE’s region, and the mapping between region ID and AI model ID, etc.
[0049] The UE performs the region identification based on available information which contains one or more elements including UE’s serving cell ID, UE speed, location, and type and number of available measurements, available AI model input, and other assistance information, etc.
[0050] Based on the region identification results, the UE decides to update the candidate model list if necessary. In one embodiment, the candidate model update is triggered when UE’s region is changed. In another embodiment, the candidate model update is triggered when the UE is moving out of the suitable region of the current candidate model list. In another embodiment, the candidate mode update is triggered when UE’s serving cell is not included in the cell list of the current region.
[0051] The candidate model update can be a direct replacement or delta update. In one embodiment, UE updates the entire list by downloading new candidate models from the model / dataset center. In another embodiment, UE downloads the model which is the delta part of the original candidate model list and new candidate model list, and updates the list with the new model.
Claims
1.A method for UE to perform the region-based AI-ML model update for wireless communication system with network awareness that comprises the steps of:report the UE region info to the RAN node for region identification;receive the model update command from the RAN node based on the region identification result; andupdate AI model based on the model update command.2.The method of claim 1, wherein the region is defined by one or multiple elements including the cell list, latitude and longitude location, available measurement of the UE, serving cell of the UE, AI-ML model training result, and available input of AI-ML model.3.The method of claim 1, further comprising UE report the UE capability to the RAN node.4.The method of claim 3, wherein the UE capability contains one or more elements including supported AI use cases, the number of candidate models that can be stored, UE storage memory, speed, location, position, available assistant information, types and quantities of available measurements, etc.5.The method of claim 1, further comprising the UE received the region report configuration from the RAN node.6.The method of claim 5, wherein the report configuration contains one or more elements including report periodicity, trigger condition, report type, report content, etc.7.The method of claim 1, further comprising the UE sends the UE region info report to the RAN node based on the region report configuration.8.The method of claim 1, further comprising the UE received the model update command from the RAN node.9.The method of claim 1, wherein the UE performs model update by selecting AI model from the pre-downloaded candidate model based on the information in model update command.10.The method of claim 1, wherein UE update model with no pre-downloaded candidate model.11.The method of claim 10, wherein UE sends the model update request to the model / dataset center.12.The method of claim 10, wherein UE downloads the update model from the model / dataset center.13.The method of claim 1, further comprising the UE sends the model update complete to the RAN node to inform the completion of model update.14.A method for UE to perform the region-based candidate model list update for wireless communication systems with network awareness that comprises the steps of:report the UE region info to the RAN node for region identification;receive the candidate model update command from the RAN node based on the region identification result; andupdate candidate model, where UE send the update candidate model list request and download the candidate model from the model / dataset center.15.The method of claim 14, further comprising the UE received the region report configuration from the RAN node.16.The method of claim 15, wherein the report configuration contains one or more elements including report periodicity, trigger condition, report type, report content, etc.17.The method of claim 14, further comprising the UE sends the UE region info report to the RAN node based on the region report configuration.18.The method of claim 14, further comprising the UE received the candidate model update command from the RAN node.19.The method of claim 18, wherein the candidate model update command contains one or multiple elements including the AI use cases, UE region ID, list of candidate model ID and the indicator of update method, e.g., delta update or direct replace.20.The method of claim 14, wherein the UE sends the update candidate model list request to the model / dataset center.21.The method of claim 14, wherein the UE downloads the candidate model from the model / dataset center and updates the candidate model list.22.The method of claim 21, wherein the candidate model list update is a direct replacement.23.The method of claim 21, wherein the candidate model list update is delta update.24.A method for the model / dataset center, which can be a UE server, network server, OAM, and OTT server, to perform the region-based AI-ML model update for the wireless communication with network awareness that comprise the step of:collect dataset and labeled data for AI-ML model training; anddetermine the region definition and the mapping between region and AI model.25.The method of claim 24, wherein the region is defined by one or multiple elements including the cell list, latitude and longitude location, available measurement of the UE, serving cell of the UE, AI-ML model training result, and available input of AI-ML model.26.The method of claim 25, wherein the regions are defined such that different regions do not overlap.27.The method of claim 25, wherein the regions are defined such that different regions can overlap.28.The method of claim 25, wherein the regions are defined to allow some location do not belong to any regions.29.The method of claim 25, wherein a cell list is assigned to each predefined region, when UE’s serving cell is within the cell list, UE is identified to the corresponding region, A UE can be identified as one, multiple predefined regions, or no predefined region.30.The method of claim 24, further comprising the model region coordination between model / dataset center and the RAN node for RAN node to perform region identification.31.The method of claim 30, wherein the coordination contains one or more elements including region definition, region ID, corresponding AI use cases, AI model ID, the mapping between region and AI model, the description used to identify UE region, etc..32.A method for the model / dataset center, which can be a UE server, network server, OAM, and OTT server, to perform the region-based candidate model list update for the wireless communication with network awareness that comprises the step of:collect dataset and labeled data for AI-ML model training; anddetermine the region definition and the mapping between region and candidate model list.33.A method for RAN node, which can be a LCM, OAM, CN, or gNB, to perform the region-based AI-ML model update for wireless communication system with network awareness that comprises the step of:region definition coordination with the dataset / model center; andregion identification that identifies the UE region based on the UE region info report.34.The method of claim 33, wherein the coordination contains one or more elements including region definition, region ID, corresponding AI use cases, AI model ID, the mapping between region and AI model, the description used to identify UE region, etc..35.The method of claim 33, further comprising the RAN send the region report configuration to the UE.36.The method of claim 35, wherein the report configuration contains one or more elements including report periodicity, trigger condition, report type, report content, etc.37.The method of claim 33, wherein the RAN node performs region identification based on the information of model region coordination, the region definition, the information at the RAN node, and the UE region info reported by UE.38.The method of claim 37, wherein the regions are defined by cell lists, When the serving cell of UE belongs to the cell list of a certain region, the RAN node identifies UE region as the region, A UE can be identified to one, multiple predefined regions, or no predefined region.39.The method of claim 33, further comprising the RAN node sends the model update command to the UE.40.The method of claim 39, wherein the model update command contains one or multiple elements includes the AI use cases, UE region ID, AI model ID, and the indicator for UE to select model from the pre-downloaded candidate model list.41.A method for RAN node, which can be a LCM, OAM, CN, or gNB, to perform the region-based candidate model list update for wireless communication system with network awareness that comprises the step of:region definition coordination with the dataset / model center; andregion identification that identifies the UE region based on the UE region info report.42.A method for UE to perform the region-based AI model update for wireless communication system without network awareness that comprise the step of:region definition coordination with the dataset / model center;region identification that identifies the UE region; andupdate AI model based on the region identification, where the model is selected from the pre-downloaded candidate model list or downloaded from the model / dataset center.43.The method of claim 42, wherein the coordination contains one or more elements including region definition, region ID, corresponding AI use cases, AI model ID, the mapping between region and AI model, the description used to identify UE region, etc..44.The method of claim 42, wherein the regions are defined by cell lists, When the serving cell of UE belongs to the cell list of a certain region, the UE region is identified as the region, A UE can be identified to one, multiple predefined regions, or no predefined region.45.The method of claim 42, wherein the UE perform model update by selecting AI model from the pre-downloaded candidate model based on the information in model update command.46.The method of claim 42, wherein UE update model with no pre-downloaded candidate model, UE send the model update request to the model / dataset center and downloads the model from the model / dataset center.47.A method for UE to perform the region-based candidate model list update for wireless communication systems without network awareness that comprises the step of:region definition coordination with the dataset / model center;region identification that identifies the UE region; andupdate candidate model, where UE send the update candidate model list request and download the candidate model from the model / dataset center.
Citation Information
Patent Citations
Model request method, model request processing method and related device
WO2023066288A1
Management and distribution of artificial intelligence model
WO2023131141A1
Method and apparatus for determining artificial intelligence (AI) model
WO2024012331A1