Base station, user equipment, network and method for machine learning related communication

Abstract

The invention relates to a base station, a terminal and a method for machine learning (AI / ML)-related communication (C) between a network device (BS, SV) and a user equipment (UE, UEs, UE1, UE2... UEk)-in which the user equipment (UE, UEs) is grouped (G) into at least one of a number of user groups (UG1, UG2, USG1, USG2) for handling (M) detected (D) of a machine learning (AI / ML) model drift-in which, upon detecting (D) of the machine learning (AI / ML) model drift, the machine learning (AI / ML) model drift is determined for the user groups (UG1, UG2, USG1, USG2) of the user groups (UG1, UG2, USG1, USG2). USG2) selects (S) a machine learning (AI / ML) model drift processing mode of operation (M).

Description

[0001] The present invention relates to a base station, a user equipment, a network and a method for machine learning (AI / ML) related communication between a mobile telecommunication network and a user equipment.

[0002] For example, details related to machine learning are discussed in the following literature: US2022 / 0006485A1, US2021 / 0132941 A1, US10,382,296B2, US10,878,342B2, US 9,961,574 B1, US2016 / 0342903 A1, US2020 / 0202171A1, US2021 / 0390455 A1, US2020 / 0151619 A1, US2020 / 0099713A1, EP3422262A1, US2021 / 0157704A1; "A survey on concept drift adaptation", J. Gama et al., ACM Computing Surveys, Vol. 46, No. 4, April 2014; "Combining Offline and Online Classifiers for Life-long Learning, Lydia Fischer et al., 2015 International Joint Conference on Neural Networks (IJCNN), October 1, 2015; "Learning from streaming data with concept drift and imbalance-an overview", T. Ryan Hoens et al., Prog ArtifIntell, April 2012; "Channel Charting-Locating Users Within the Radio Environment Using Channel State Information", C. STUDER et al., IEEE Access, September 29, 2020; "Efficient User Clustering Using a Low-Complexity Artificial Neural Network (ANN) for 5G NOMA Systems" Efficient User Clustering in NOMA System]”, S. PRABHAKUMARESAN et al., IEEE Access, September 29, 2020, all of which are incorporated into this application by reference.

[0003] The object of the present invention is to improve machine learning (AI / ML) related communication between a mobile telecommunication network (device) and (multiple) user equipment. This object is achieved by the teachings of the independent patent claims, respectively. Embodiments of the present invention allow for improved machine learning (AI / ML) related communication between a mobile telecommunication network (device) and (multiple) user equipment, such as more efficient communication and / or better resource allocation.

[0004] The dependent patent claims describe some advantageous embodiments of the invention.

[0005] Some embodiments of the present invention are described below, which can also present further preferred embodiments of the present invention in any combination.

[0006] Embodiments of the present invention will be described below with reference to the accompanying drawings, as shown below:

[0007] Figure 1 shows the traditional approach versus the so-called machine learning and / or artificial intelligence (AI / ML) approach.

[0008] Figure 2 It is the life cycle of AI / ML machine learning.

[0009] Figure 3 is a user equipment having an artificial intelligence and / or machine learning module for communicating with a network device, such as a cellular mobile telecommunication network,

[0010] Figure 4 is the overall performance of the wireless communication network,

[0011] Figure 5 It is to group user devices.

[0012] Figure 6 is the life cycle profile graph,

[0013] Figure 7 is the statistical relationship pattern between dataset categories and lifecycle time ranges,

[0014] Figure 8 is the UE index re-ranked based on multiple threshold levels,

[0015] Fig. 9 is to subdivide user devices into groups and / or subgroups,

[0016] Fig.10 is a top-down grouping based on mixed criteria,

[0017] Fig.11 It is a signaling flow chart of the user equipment grouping process.

[0018] Fig.12 is a flow chart based on groupings such as a life cycle profile diagram,

[0019] Fig.13 The number of UE groups is selected based on the similarity level threshold.

[0020] Fig.14 It is the AI / ML machine learning mode switching selection table.

[0021] Fig.15 It is an (advanced) procedure for AI / ML mode switching triggered by model drift.

[0022] Fig.16 Model drift detection and AI / ML mode switching selection based on event triggering and / or codebook,

[0023] Fig.17 It is a model drift detection based on periodic timer,

[0024] Fig.18 This is a flowchart of the AI / ML mode switching selection procedure.

[0025] Fig.19 Share trigger mode switching information with user device groups,

[0026] Fig. 20 It is user device grouping and mode switching.

[0027] Figure 1 generally illustrates an example of the difference between a traditional model-based approach that uses input data and a fixed algorithmic scheme to generate output data and (lower) a so-called artificial intelligence and / or machine learning (AI / ML) model approach that uses input data and training / test data to generate output data. Instead of a fixed algorithmic scheme, a data input process (algorithm) based on training / test data / local (UE) data may be used.

[0028] Figure 2 In general, the AI / ML life cycle can be divided into several stages, as shown in the figure. If the model performance cannot be maintained continuously due to drift, it may be necessary to monitor after the AI / ML (artificial intelligence and / or machine learning) model is deployed. For example, in simple terms, the AI / ML model has learned to detect vehicles or cars at noon in summer, and now it must detect cars on snowy and icy streets at midnight in foggy weather, and there will be (car) local data drift and AI / ML machine learning model drift. This requires providing updated feedback to readjust the model. As a reason for the decline in model performance, fundamentally, the data set statistics may change, and the model reasoning ability will also be affected by unseen data as input.

[0029] For example, such a scenario may use a data collection and preprocessing step, an (AI / ML) model training step, a model testing and validation step, a model deployment and inference step, and a monitoring step (feedback on (repeated) previous steps).

[0030] Figure 3 It shows that when applying AI / ML models to any use case or application, one of the challenging issues may be managing the lifecycle of the AI / ML model (e.g., how long is the current AI / ML model valid / usable given the changes in local data).

[0031] This may be mainly due to data / model drift occurring during the AI / ML model deployment / inference phase and may lead to degradation of AI / ML model performance.

[0032] For example, changes in the statistical characteristics of the data set of local data in the user equipment (UE, UEs, UE1, UE2..UEk) may also cause model drift;

[0033] For example, as a very simple example, a dataset of local data of user equipment (UE, UEs, UE1, UE2..UEk) now obtains speed / optical / etc data of cars driving on snowy and icy surfaces at night (instead of cars driving on dry streets in the summer in the past datasets) and the relationship between the input and output of the trained model.

[0034] Therefore, it may be very important to handle AI / ML data / model drift by tracking model performance (such as predictability, accuracy, etc.) as the AI / ML model needs to be updated when drift occurs.

[0035] When deploying wireless communication networks that support AI / ML models, it is important to consider how to handle drift in the AI / ML model lifecycle when the wireless device is operating, such as model training, inference, and updates.

[0036] Figure 3 A cellular mobile telecommunication network (e.g. 3G, 4G, 5G, 6G, 7G, 8G, Wi-Fi, etc.) is shown having a network device as a base station BS and a core network element CN as, for example, a backend server SV (e.g. for storing and / or updating and / or providing to the BS / UE a machine learning model drift processing operation mode for use in a user equipment UE or BS and / or, for example, in the user equipment UE or BS).

[0037] Figure 3A user equipment UE (e.g. UE1, UE2, UE3, UE4, UE5, ... UEk in a vehicle / car V or a mobile phone etc.) and a (cellular) telecommunication end user device MS (e.g. TCU or a mobile phone or a USB cellular stick etc.) with an artificial intelligence and / or machine learning module AI / MLmod are shown,

[0038] For communicating C with a network device (eg a base station BS) of a cellular mobile telecommunication network NW.

[0039] Figure 4 It shows that the overall performance of wireless communication networks that apply AI / ML models may be significantly affected by model drift, and the statistical characteristics will change as the model performance degrades over time.

[0040] In addition, when AI / ML-related signaling overhead occurs between the BS and the UE for lifecycle processing, the additional signaling traffic overhead may be very large.

[0041] Figure 4 Illustrative examples of AI / ML model performance degradation over time when drift occurs are shown. The exact model performance degradation pattern varies in different environments and settings. Therefore, studying how to compensate for model performance degradation due to drift may be important in addressing the challenges of managing AI / ML model drift in wireless communication networks.

[0042] Figure 4 The change of the model performance level (eg, prediction accuracy) over time is shown, where the area above the curve is an acceptable region, the area below the curve is an unacceptable region, and the dotted line represents the performance threshold.

[0043] Some key changes / impact areas in 3GPP may include

[0044] User equipment UE measurement report (e.g. AI / ML data set / feature statistical metrics, etc.), user equipment UE capability information (e.g. AI / ML training / model upper limit), RRC configuration (e.g. AI / ML user grouping), data / control signaling (e.g. DCI, UCI, MAC CE, etc., PHY38.21x (x:3,4,5), MAC 38.321, RRC 38.331, UECap 38.306).

[0045] Some embodiments of the present invention will be explained below.

[0046] To address the problem of AI / ML models’ performance decreasing over time when drift occurs,

[0047] User grouping based on AI / ML (artificial intelligence and / or machine learning) dataset categories may be used.

[0048] Embodiments of the present invention may involve user grouping G methods and / or (AI / ML model / data) drift detection D and / or mode M switching selection S,

[0049] In particular, forming user groups G for AI / ML drift processing and / or

[0050] AI / ML operation mode M selects S for AI / ML drift processing.

[0051] Figure 5 Some embodiments of the present invention show that:

[0052] In order to solve the problem of performance degradation based on AI / ML model drift, it is proposed to group user equipment UE, UE1, UE2, UE3, UE4, UE5, ... UEk of a wireless access network RAN ​​(such as a communication C network such as cellular and / or Wi-Fi), where, for example, an AI / ML (artificial intelligence / machine learning) model is deployed for a specific use case or application.

[0053] Based on, for example, the data set category DC of the AI / ML model and the anti-drift time duration (for example, the duration change of local data in the user equipment UE (related to the AI / ML model; also called the life cycle time range LT, LTi) does not exceed the threshold value ti, R, Ri), the distributed user equipment UE connected to one (or more) base stations BS or communicating with it C is classified into at least one user equipment UE user equipment group UG1, UG2, etc.

[0054] The UE grouping criteria here are, for example / in particular, based on similarity metrics of predetermined parameters (e.g. data set categories DC, DCi; lifecycle time range categories LT, LTi); and can continue with zero or one or more subdivisions based on (multiple) parameters such as service type / priority, traffic pattern, QoS, etc.

[0055] However, here for example the baseline user equipment UE grouping G is based on the dataset categories DC, DCi and the anti-drift time durations LT, LTi (of the AI / ML model).

[0056] The definition of similarity level may be implementation specific, and reference metrics may include, for example, using Euclidean distance or correlation-based metrics. Figure 5 An overall abstract block diagram of the UE grouping concept for ML drift handling is shown.

[0057] Figure 5 The lower part shows that after the step of grouping the user equipment UE G into groups UG1, UG2,

[0058] The user equipments UE1, UE2, UE3 are now assigned to the group UG1, communicating (sending / transmitting; individually and / or as a group) with the network NW (in particular the RAN, in particular one (or more) base stations BS),

[0059] And the user equipments UE4, .. , UEk are now assigned to the group UG2, communicating (sending / transmitting; individually and / or as a group) with the network NW (especially the RAN, especially one (or more) base stations BS).

[0060] Figure 6 Some embodiments of the present invention show that:

[0061] An example of the newly proposed lifecycle profile graph, which may consist of a list L of parameters (eg DC, LT, DD) as index (i) values ​​(eg DCi, LTi, DDi).

[0062] Each parameter (eg DC, LT, DD; DCi, LTi, DDi) may be marked by each user equipment UE, UE1, UE2, UE3, UE4, UE5, ... UEk or a means for mapping.

[0063] The lifecycle profile graph (e.g., list L) may be continuously maintained (e.g., in a server SV) for updating based on, for example, historical data of AI / ML trained models applied to different use cases / applications and / or real-time communication feedback information (e.g., information incoming via a radio access network RAN).

[0064] The initial lifecycle profile map may be retrieved from a database (eg in a server SV of the network NW) and may be stored at the database and extracted / retrieved when, for example, a new UE grouping event G is set.

[0065] When retrieving and updating the lifecycle profile map, multiple BSs and mobile edge networks may share it with each other so that related user equipment UE groups G may be formed and regrouped in different communication scenarios and / or under, for example, several base stations BS.

[0066] For a number (i) of user equipments (UE, UEs, UE1, UE2 .. UEk) and / or a plurality of groups UG1, UG2, a lifecycle profile map or list L may for example comprise a table with a dataset category DC, a LT lifecycle and a data distribution index DD,

[0067] For example, for user equipment UE1 or (i=1), the lifecycle profile map or list L may store / include a data set category DC1, a lifecycle LT1 and a data distribution index DD1,

[0068] And / or for the group UG1 (i=1), the lifecycle profile map or list L may store / include the dataset category DC1, the lifecycle LT1 and the data distribution index DD1.

[0069] Figure 7 Some embodiments of the present invention show that:

[0070] In the group G for the user equipment UE,

[0071] The statistical relationship pattern between the monitorable data set categories DC, DCi and the life cycle time ranges LT, LTi (LT is the user device number i) in the life cycle profile diagram.

[0072] For example, multiple threshold levels ti (regarding the degree of difference between the current data and the initial / model data) can be defined to classify each user equipment UE into different subgroups UG1, UG2 based on, for example, the estimated life cycle LT time range thresholds t1, t2, t3, t4, tR (of time or stable duration t) and in combination with, for example, data set category characteristics.

[0073] For example, the threshold t1 here is the shortest life cycle time range value,

[0074] This indicates that the AI / ML model lifecycle LT may be very short, i.e., for example, the change of local data about or in the user equipment UEi compared to the initial data forming the initial AI / ML model exceeds a threshold t1.

[0075] On the other hand, the threshold tRi here is the longest lifecycle time range value, which indicates that the AI / ML model lifecycle may be quite long.

[0076] Considering the density of the lifecycle time horizon estimation of the user equipment UE, multiple threshold levels may be derived, for example, by density-based classification ("multiple" means, for example, depending on more than one parameter, such as DC and LT).

[0077] like Figure 8 As shown in the figure, the UE index is reordered G (eg, first grouping G or regrouping again) based on, for example, multiple threshold levels,

[0078] For example, it depends on whether it is between two lifetime LT levels t0, t1, t2; or between t2 and t3; or between t3 and t4, etc.

[0079] exist Figure 8In the example, after user equipments UE1, UE2, UEk are grouped into groups UG1 and UG2, for example, user group UG1 has the shortest life cycle LT time range value from t0 to t1. That is, all user equipments UE1, UE2, UE3 in the user group UG1 have life cycle LT time range values ​​higher than threshold t0 but lower than threshold t1.

[0080] Hence, the group UG1 eg gets (eg as machine learning model drift processing operation mode M) assigned the highest priority (thus eg more frequent and / or higher priority and / or with more channels / time slots), eg for communication with the base station BS.

[0081] For example, the user group UGRi has the lowest priority (for communication with the base station BS, etc.) because the UGRi member UEk of this group has the longest lifecycle time range LT value, which indicates that the AI / ML model lifecycle of this group (UE members) can be very / quite long or stable.

[0082] After the UE groups UG1, UGRi, UG1, UG2 (e.g., relative to each other) are prioritized as described above, resource allocation and / or traffic signaling for (multiple) priority user equipment UE groups can be scheduled according to the corresponding priorities (e.g., a certain user group can obtain more or less air interface resources on one or more base stations BS for data communication C related to, for example, machine learning data updates).

[0083] Fig. 9 For some embodiments of the present invention, group G is shown.

[0084] Or two stages in a two-stage grouping (depending on the respective data set category of the user equipment, such as DC1 or DC2 or DCs), the user equipment UE1, UE2..UEk is divided into groups UG1, UG2,.., and then (alternatively in one step) the user equipment UE1, UE2..UEk is subdivided into subgroups USG1, USG2,.., etc.

[0085] Fig. 9 The process diagram of the user equipment UE1, UE2, .. UEk in FIG. 1 shows how UE group classification may be performed, for example using a lifecycle profile map index (ie depending on the lifecycle time range LT, LTi and the data set category DC, DCi).

[0086] Any further subdivision of the multi-level UE grouping may be performed using any predetermined parameters, such as based on similarity as a criterion.After classifying the UE groups, resource allocation and / or traffic signaling of the priority UE group(s) may be preferentially scheduled.

[0087] exist Fig. 9During the UE grouping process, for example

[0088] Assume that the UE is a group of K UEs, such as {UE1, UE2, ..., UE K}.

[0089] DC (dataset category) is a set of {DC1, DC2, ..., DC S}, where S is the size of the dataset category.

[0090] LT (Life Cycle Time Range) is a set of {LT DCi,1 , L T DCi,2 ,…,LT DCi,Ri}, where each element in the group represents a UE group in the i-th dataset category, and R i is the size of the time range threshold.

[0091] Fig.10 Some embodiments of the present invention are shown in block diagrams.

[0092] Regarding the multi-level grouping of the user equipment UE, for example, a top-down grouping method may be adopted based on a hybrid standard.

[0093] For multi-level user grouping, multiple parameter indexes can be used to divide users into multiple levels. Candidate parameters for user grouping criteria may include, for example / especially, index values ​​of lifecycle profile graphs, QoS (including latency constraints), service types / priorities, traffic patterns, etc. Fig.10 In the solution, first-level user equipments (UEs) are grouped into first-level user equipment subgroups USG1 and USG2 based on index values ​​of a lifecycle profile map (eg, on LT, DC).

[0094] The number of levels and the size of the parameter index may be implementation specific and configurable based on different applications.

[0095] Fig.11 A signaling flow diagram of a user equipment UE grouping process G is shown for some embodiments of the present invention.

[0096] Fig.11 The signaling between one or here several user equipments UE and the base station BS for performing a procedure for grouping the user equipments UE is shown.

[0097] After RRC establishment signaling and signaling regarding AI / ML configuration / measurement information exchange,

[0098] The base station BS (or another network element) configures the user equipment UE grouping based on, for example, predetermined indication index(es) for the user equipment UE.

[0099] The base station BS (or another network element) assigns a user equipment UE group ID to each user equipment group UG1, UG2.

[0100] Through DCI (PDCCH) and / or PDSCH channels, the base station informs all user equipments UE1, UE2, UEk (eg point-to-point or here through a broadcast channel) of their user equipment UE group ID (the ID of the group to which the user equipment is now assigned).

[0101] In the signaling flow diagram, the criteria for UE grouping is predetermined, for example, based on a configured parameter index. Once the UE grouping is determined, a specific ID (such as a group RNTI) can be used to identify the UE groups USG1, USG2, etc. For each UE group, the DCI in the PDCCH and the PDSCH can then be transmitted.

[0102] Fig.12 For some embodiments of the present invention, a flow chart is shown showing UE grouping G based on, for example, a lifecycle profile diagram.

[0103] For AI / ML model deployment in wireless networks, for example, a new “Life Cycle Profile Diagram” is introduced to identify the characteristics of key parameters related to the AI / ML model life cycle LT, LTi applied by each user equipment UEi.

[0104] For example, the “Lifecycle Profile Graph” indicates the statistical patterns of the relationships between key parameters (including dataset category DC, data distribution DD, and drift resistance time horizon LT) applied to each specific use case of the AI / ML model.

[0105] It can also continuously update and store historical data based on AI / ML trained models for maintenance for each use case and application applied to AI / ML operations in the Radio Access Network RAN.

[0106] In the flow chart of the user equipment UE group G, for example, the network device (e.g., base station BS) performs several key steps of identification, classification and prioritization of UE groups and activation of AI / ML model adaptation due to drift. The life cycle profile diagram can be maintained during the base station-user equipment (BS-UE) communication phase of the UE group operation for update.

[0107] The target user equipment UE set may also be regrouped based on the change of the lifecycle profile map, and the mode switching may be performed based on the mode selection table.

[0108] For some embodiments of the present invention, based on the user equipment UE grouping G, for example based on the lifecycle profile map, the user equipment UE feedback report may be used to update the lifecycle profile map and / or regroup.

[0109] Feedback information to include could include

[0110] - Lifecycle profile graph information, such as {dataset category index, lifecycle time range index, data distribution index}

[0111] - Information indicating that the user equipment UE and / or the user equipment group has a drift indication

[0112] -AI / ML mode selection information

[0113] For example, feedback signaling may be considered for the following report types and / or channels (already supported by the standard specification):

[0114] -Periodic feedback report: PUCCH

[0115] - Semi-persistent feedback reporting: PUCCH or PUSCH with DCI triggering

[0116] -Aperiodic feedback report: PUSCH

[0117] Fig.13 Some embodiments of the present invention show that:

[0118] Selection of the number of user equipment UE groups based on similarity level thresholds.

[0119] The number of UE groups may be determined based on a predefined threshold of a similarity level, which may be measured by, for example, a Euclidean distance or a correlation-related metric.

[0120] The similarity level SL of the lifecycle time ranges in each UE group UG1, UG2, UGN may be used as a criterion for grouping G of user equipments UE.

[0121] The schematic diagram shows that as the similarity level threshold increases, the total number of UE groups also increases. In other words, by setting the similarity level threshold to be met to be higher, the difference in life cycle time ranges between user equipments UE within a user equipment group becomes smaller.

[0122] Fig.14 An AI / ML mode switching selection table is shown for some embodiments of the present invention.

[0123] For AI / ML model drift detection D, model performance monitoring can be activated to check whether any drift occurs after the AI / ML model is deployed in the radio access network RAN ​​for detection.

[0124] For example, based on a predefined threshold, the model drift indicator of drift detection D may include

[0125] - Changes in local (e.g., in user devices) data distribution

[0126] -Changes in model accuracy levels

[0127] For mode switch selection, the "mode switch selection table" may, for example, contain the mode selections to be used, such as:

[0128] - Baseline mode (non-AI / ML), e.g., this mode selects non-AI / ML based default actions based on a pre-programmed algorithm,

[0129] - Trained model mode, e.g., a mode that replaces a running activated AI / ML model with another candidate pre-trained model (e.g., offline or batch learning) for activation, or a mode where, for example, ensemble learning can be used with multiple pre-trained models,

[0130] - Partial retraining mode, e.g., this mode performs additional model training (e.g., online / incremental learning) using new incoming data to compensate the activated AI / ML model through model updates; if available, transfer / meta-learning can be used in this mode,

[0131] -Full retraining mode, for example, this mode discards the AI / ML model currently in use and performs AI / ML model training using a complete set of new data.

[0132] Fig.15 A high-level procedural flow chart of model drift triggering and AI / ML mode switching is shown for some embodiments of the present invention.

[0133] Model drift detection can be based on

[0134] - a periodic approach using a configurable timer based on the lifetime time range of a group of user equipments UE; and / or

[0135] - Aperiodic (or event-triggered) approaches, where a drift indicator with a predefined threshold can be based on a model performance metric.

[0136] The mode switching selection may, for example, take into account the latency constraints of the application data QoS and the historical data recorded by the mode switching selection based on the lifecycle profile graph.

[0137] Fig.16 According to some embodiments of the present invention, a flow chart is used to illustrate model drift detection based on event triggering and AI / ML mode M switching selection (S).

[0138] This flowchart illustrates the AI / ML mode M switching process with data distribution changes as the trigger factor.

[0139] For data distribution, for example, using GMM (Gaussian mixture model), a certain number of Gaussian distributions (GD i , where i=1,…,C).

[0140] A set of Gaussian distributions can then be quantized into codewords representing quantized values ​​of the data distribution.

[0141] Each local data distribution is then mapped to the codeword with the closest metric. The set of all codewords is then defined as the codebook. The local data distribution may change with the channel at any time, causing AI / ML model drift.

[0142] If the codebook index is different from the previously selected index, the AI / ML mode switching process is triggered.

[0143] Fig.17 A flow chart of model drift detection D based on a periodic timer is shown for some embodiments of the present invention.

[0144] Configurable timer (T c ) is set based on the life cycle time range of the user equipment UE groups UG1 and UG2.

[0145] If the user equipment UE groups UG1, UG2 are set to {UG1, UG2, . . . , UG R}, and the UE group index is prioritized based on the life cycle time range LT, LTi, then UG1 has the shortest time range and UG R Has the longest time frame.

[0146] The periodic timer is also configured based on the ordered UE groups UG1, UG2, ... Therefore, Tc is also set to {T c (UG1), T c (UG2),…,T c (UG R )}. These timer thresholds are, for example, T c (UG1) <T c (UG2) <T c (UG R ) in the order of .

[0147] The threshold parameter type of the event triggering method of drift detection D may be, for example:

[0148] -AI / ML model accuracy metrics

[0149] - Codebook index difference (when data distribution is quantized into codewords)

[0150] - Channel quality indicators (e.g. RSRP, RSRQ, SINR)

[0151] The above mentioned types of threshold parameters can also be used as a combination of the above mentioned parameters depending on the use case of the application.

[0152] For the timer of the periodic method of drift detection, a timer (T c ) to periodically measure model drift detection.

[0153] The timer value configured by RRC can be based on the life cycle time range parameter of the data set category. For example, the shorter the time range, the smaller the timer value, and vice versa, because the timer needs to be smaller than the time range. When the user equipment UE groups UG1, UG2, USG1, USG2 are (re)formed and sent to the user equipment UE via multicast, the timer can also be set to start as a periodic method. Once the timer expires, the user equipment UE can measure the model drift detection D, and the AI / ML model can operate to make a mode switching decision M. The timer can then be reset to the initial value to restart, and the user equipment UE continues to run the (selected) AI / ML model operation.

[0154] Fig.18 A flow chart of the AI / ML mode switching selection procedure S is shown.

[0155] Based on considerations such as latency-related constraints for application data QoS, the baseline mode is preferred to minimize the performance impact on ongoing communication flows. If other modes are allowed, the trained model is first used to check whether to replace the current model. The last option is to use a partial or full retraining mode, but depending on the length of the retraining time, the model performance quality may be time-limited.

[0156] Fig.19 According to some embodiments of the present invention, a flowchart of sharing triggered AI / ML mode switching information with member UEs, UE1, UE2, ..UEk of user equipment UE groups UG1, UG2, USG1, USG2 is shown.

[0157] After performing AI / ML mode switching on at least one user equipment UE in the user equipment UE groups UG1, UG2, USG1, and USG2,

[0158] This information is also shared with other user equipment UE members in UE groups UG1, UG2, USG1, USG2.

[0159] In this way, other user equipment UE members that may have similar drift D for triggering may adopt the priority mode switching selection M based on the shared information.

[0160] Then, the user equipment UE with triggered (D) drift can improve the performance impact by making a decision of mode switching selection (M) faster.

[0161] Fig. 20 A flow chart of user equipment UE grouping and mode switching M is shown for some embodiments of the present invention.

[0162] Fig. 20 shows the communication between a network element (eg a base station BS) and a user equipment UE for packet and mode switching,

[0163] And the boxes represent actions in network elements (such as base station BS and user equipment UE).

[0164] Reference numerals

[0165] NW Network

[0166] RAN Telecom Network / Access

[0167] BS Base Station

[0168] SV Server

[0169] CN core network element

[0170] C communication, such as BS-UE

[0171] G UE Grouping

[0172] UG1 Group of user equipment (UE1, UE2..UE3)

[0173] UG2 User Equipment (UE4..UEk) Group

[0174] USG1 User Equipment UE subgroup

[0175] USG2 User Equipment UE Subgroup

[0176] D Detect machine learning (AI / ML) model drift,

[0177] S selects machine learning model drift processing (M) for at least one user group

[0178] M Model Drift Handling Operation Mode

[0179] DC is a dataset category in a set of dataset categories {DC1, DC2, …, DCS}, LT, LTDCi is a life cycle time range category

[0180] ti LT or threshold DD data distribution index of time t, t0, t1, t2, t3, t4, tR

[0181] DCI A list of user equipment group IDL parameters DCi, LTi, DDi for a group of user equipment UEs

[0182] V Vehicle, car

Claims

1. A network device (BS, SV), in particular a cellular mobile telecommunication network element (BS, SV) and / or a cellular mobile telecommunication network base station (BS), - configured to group (G) each of a number (k) of user equipments (UE, UEs, UE1, UE2 .. UEk) configured to communicate (C) with a network device (BS, SV) into at least one of a number of user groups (UG1, UG2, USG1, USG2), - configured to detect (D) a machine learning (AI / ML) model drift associated with at least one of the user equipments (UE, UEs, UE1, UE2 .. UEk), -Configured to, when a machine learning (AI / ML) model drift (D) is detected, for at least one user group (UG1, UG2, USG1, USG2), select (S) a machine learning model drift processing operation mode (M) for user equipment (UE1, UE2, UE3; UE4, UEk) in the user group (UG1, UG2, USG1, USG2).

2. Network device according to the preceding claim, in, The several user groups (UG1, UG2, USG1, USG2) are assigned different network resource allocation and / or traffic signaling priorities when communicating (C) with the network devices (BS, SV).

3. The network device according to any one of the preceding claims, Especially for user equipments (UE1, UE2, UE3; UE4, UEk) with different specific use cases or applications, - configured to group (G) the user equipments (UE1, UE2, UE3; UE4, UEk) into at least one or exactly one user group (UG1, UG2, USG1, USG2) based on similarity criterion parameters, these parameters in particular comprising one or two or more of the following: - a dataset category (DC) from a set of dataset categories {DC1, DC2, ..., DCS}, In particular, cQI or voice-related or non-voice-related communication data categories, - A set of life cycle time range (LT) categories {LT DCi,1 , LT DCi,2 ,…,LT DCi,Ri }, specifically, a set of times (t) during which the drift of the machine learning (AI / ML) model does not exceed a threshold (t1, t2, tR), - Data Distribution Index (DD), in particular indicating the local data distribution measured in the user equipment (UE, UEs, UE1, UE2..UEk) and / or the closest codeword in the assigned codebook.

4. The network device according to any one of the preceding claims, is configured to decide on the priority of network resource allocation and / or traffic signaling, in particular the priority of retraining or model update signaling, The priorities are different for the user groups (UG1, UG2, USG1, USG2) and / or the subgroups (USG1, USG2) and are the same for the user equipments (UE, UEs, UE1, UE2..UEk) in a group (UG1, UG2, USG1, USG2) and / or a subgroup (USG1, USG2), respectively. In particular, user equipment (UE, UEs, UE1, UE2..UEk) members in the group and / or subgroup with the shortest life cycle time range value obtain the highest priority, while user equipment (UE, UEs, UE1, UE2..UEk) members in the group and / or subgroup with the longest life cycle time range value obtain the lowest priority.

5. The network device according to any one of the preceding claims, - configured to group the user equipments (UE1, UE2, UE3; UE4, UEk) into subgroups and / or sub-subgroups of the user groups (UG1, UG2, USG1, USG2) based on similarity criterion parameters, these parameters in particular comprising one or more of the following: - service type requirements or measurements for the communication (C) between the mobile telecommunication network device (BS, SV) and the user equipment (UE, UEs) in the subgroup or sub-subgroup, - service priority requirements for the communication (C) between the mobile telecommunication network device (BS, SV) and the user equipment (UE, UEs) in the subgroup or sub-subgroup, - traffic pattern requirements or measurements of the communication (C) between the mobile telecommunication network device (BS, SV) and the user equipment (UE, UEs) in the subgroup or sub-subgroup, - QoS requirements or measurements of the communication (C) between the mobile telecommunication network device (BS, SV) and the user equipment (UE, UEs) in the subgroup.

6. The network device according to any one of the preceding claims, -Configured to assign a user equipment group ID (DCI) to each user group (UG1, UG2, USG1, USG2) of (k) user equipments (UE, UEs, UE1, UE2..UEk) and / or to send the user equipment group ID (DCI) of the group to all user equipments (UE, UEs, UE1, UE2..UEk) in the group.

7. The network device according to any one of the preceding claims, - configured to determine the number of user equipment groups (DCI) based on a predefined threshold of similarity levels of lifecycle time ranges (LT) within the user groups (UG1, UG2, USG1, USG2).

8. The network device according to any one of the preceding claims, - is configured to store and / or update a life cycle profile map and / or parameter list (DC i , LT i , D.D. i ), these parameters respectively indicate at least one of the following items of the user equipment (UE, UEs): - a dataset category (DC) from a set of dataset categories {DC1, DC2, ..., DCS}, In particular, cQI or voice-related and non-voice-related communication data categories, - A set of life cycle time range (LT) categories {LT DCi,1 , LT DCi,2 ,…,LT DCi,Ri } in the life cycle time range (LT) category, especially the drift of the machine learning model does not exceed the threshold (R i ) a set of time, - A data distribution index (DD), in particular representing the local data distribution measured in the user equipment (UE, UEs, UE1, UE2..UEk) by the closest codeword in the assigned codebook.

9. The network device according to any one of the preceding claims, -Configured to provide lifecycle profiles and / or parameter lists to other network devices (DC i , LT i , D.D. i ), these parameters respectively indicate at least one of the following items of the user equipment (UE, UEs, UE1, UE2..UEk): - a dataset category (DC) from a set of dataset categories {DC1, DC2, ..., DCS}, In particular, cQI or voice-related or non-voice-related communication data categories, - A set of life cycle time range (LT) categories {LT DCi,1 , LT DCi,2 ,…,LT DCi,Ri } in the life cycle time range (LT) category, especially the drift of the machine learning model (AI / ML) does not exceed the threshold (R i ) a set of time, - A data distribution index (DD), in particular representing the local data distribution measured in the user equipment (UE, UEs, UE1, UE2..UEk) by the closest codeword in the assigned codebook.

10. The network device according to any one of the preceding claims, - is configured to select a machine learning mode switch upon detection of (D) machine learning (AI / ML) model drift, Machine learning modes include, in particular, one of the following: -Default machine learning model, - replace the default machine learning model with another one, - Partially retrain the machine learning model, updating the last machine learning model with new incoming local data of the user equipment (UE, UEs, UE1, UE2..UEk), -Completely retrain the machine learning model, build a new machine learning model using the new incoming local data of the user equipment (UE, UEs, UE1, UE2..UEk).

11. The network device according to any one of the preceding claims, Configured to detect (D) machine learning model drift based on periodic event-triggered monitoring of machine learning model performance for each user group (UG1, UG2, USG1, USG2).

12. The network device according to any one of the preceding claims, The method is configured to detect (D) machine learning model drift based on monitoring of non-periodic event triggers for each user group (UG1, UG2, USG1, USG2), in particular at least one of the following triggers: -Machine learning model accuracy metrics - Codebook index difference - Channel quality indication.

13. The network device according to any one of the preceding claims, Configured to detect (D) machine learning model drift based on received measurements of local data distribution in user equipment (UE, UEs, UE1, UE2..UEk), preferably using a closest codeword in an assigned codebook and / or a determination of a codebook index identifying the assigned codeword.

14. A user equipment (UE, UEs, UE1, UE2..UEk), - configured to send a feedback report to the network device (BS), in particular for a lifecycle profile map update and / or a regrouping of user equipments (UE, UEs, UE1, UE2 .. UEk), the feedback report in particular indicating at least one of the following: - Lifecycle profile graph information, especially the dataset category index (DC), lifecycle time range index (LT), data distribution index (DD), - information indicating that a user equipment (UE, UEs, UE1, UE2 .. UEk) or a group thereof has a detected indication of machine learning (AI / ML) drift, -Mode selection (M) information.

15. The user equipment (UE, UEs, UE1, UE2 .. UEk), in particular according to the preceding claim, being configured to - measure a local data distribution in the user equipment (UE, UEs, UE1, UE2 .. UEk), - assign the closest codeword in the codebook, - determining a codebook index identifying the assigned codeword, - determine if a codebook index change was triggered, - Sending data representing the result of the determination to the network device (BS), in particular for the machine learning mode (M) switching selection (S).

16. A telecommunications network (NW), having at least one network device (BS) according to any one of the preceding claims, and There is at least one user equipment (UE, UEs, UE1, UE2..UEk) according to any of the preceding claims.

17. A method for machine learning (AI / ML) related communication (C) between a network device (BS, SV), in particular according to any one of the preceding claims, and a user equipment (UE, UEs, UE1, UE2..UEk), in particular according to any one of the preceding claims, - in, The user equipments (UE, UEs) are grouped (G) into at least one of several user groups (UG1, UG2, USG1, USG2), To handle (M) detected (D) machine learning (AI / ML) model drift, - wherein, when machine learning (AI / ML) model drift (D) is detected, Select (S) a machine learning (AI / ML) model drift handling operation mode (M) for user equipment (UE, UEs) in user group (UG1, UG2, USG1, USG2).

Images