Method for model management and reliability enhancement of ai-driven mobility prediction in wireless network

Abstract

This invention discloses a model management and reliability enhancement method for AI-driven mobility prediction in wireless networks, belonging to the field of mobile communication network technology. The method includes: an offline phase, collecting historical data including UE measurement data; training and testing the AI ​​model using a subset of data based on UE mobility level, functional requirements, and OW / PW step size; identifying the model's predictive performance and overhead; establishing a mapping relationship between model identifiers and network / UE side parameters; and selecting the optimal unique sub-model based on uploaded mobility speed and functional requirements; and an online phase, performing instance-level predictability evaluation of the AI ​​model based on the local dataset before and after each model call, followed by model activation / deactivation determination. This invention improves the inference reliability and stability of the AI ​​model deployed in the access network, offering advantages such as convenient implementation, low additional signaling overhead, and low engineering deployment cost, making it more suitable for widespread application in practical mobile communication systems.

Description

Technical Field

[0001] This invention belongs to the field of mobile communication network technology, specifically referring to a model management and reliability enhancement method for AI (artificial intelligence) driven mobility prediction in wireless networks. Background Technology

[0002] In mobile communication networks, when a UE (User Equipment) moves from one cell to another, or when communication quality degrades due to external interference, and due to the needs of Radio Resource Management (RRM) such as inter-cell load balancing, the sector serving the UE will change to ensure stable communication quality. During this process, the 3rd Generation Partnership Project (3GPP) specifies several RRM events, such as those most commonly used in intra-system handover, to ensure successful and high-quality handover. Monitoring Radio Link Failure (RLF) and subsequent cell reselection serve as a "safety net" mechanism for stable communication quality during this process, while RLF is considered a serious fault that needs to be avoided.

[0003] Whether it's triggering RRM events, inter-cell handover, or monitoring RLF, the goal is to ensure continuous cell coverage in mobile communication services, and UE measurements are the foundation for achieving all of this. In recent years, the integration of AI and radio access networks has become an important research area. The superior nonlinear fitting capabilities of AI models can often capture and predict complex measurement sequence change patterns under time-varying channels, breaking away from traditional rule-based mobility management methods.

[0004] Since version 18, 3GPP Radio Access Network (RAN) Working Group 1 has studied multiple applications of AI at the physical layer. Among them, Channel State Information (CSI) prediction, beam prediction, and UE positioning enhancement are all implemented based on temporal or spatial prediction in mobility scenarios. More importantly, the 3GPP RAN Technical Specification Group's report TR 38.843 also established a macro framework called Life Cycle Management (LCM) for subsequent AI standardization research. In version 19, RAN3, the RAN Architecture Working Group, further enhanced the use cases of UE handover optimization and handover trajectory prediction based on version 18, and continued to rely on data collected in mobility scenarios such as UE historical measurement reports, historical trajectories, and historical event reports, and proposed several other use cases combined with specific service characteristics. RAN2, the Radio Resource Control Working Group, focuses on AI-based mobility prediction and enhancement in the RRC_CONNECTED (Radio Resource Control Connected State) mode, motivated by reducing UE measurement overhead and achieving measurement event prediction.

[0005] In summary, mobility prediction, as a core component, is present in various AI-related proposals across 3GPP working groups, and almost all scenarios require spatiotemporal prediction of UE measurement reports to achieve research objectives. However, most use cases and related academic research neglect practical management strategies for network-side AI models and datasets, lack reasonable training data segmentation methods, and fail to address the balancing of model cost and performance. Furthermore, regardless of the specific use case, AI model errors cannot be completely eliminated; some unreliable AI model outputs or decisions based on model AI outputs can worsen the quality of current communication services.

[0006] In document 1 (Khunteta S, Chavva AK R. Deep learning based link failuremitigation[C]. 2017 16th IEEE International Conference on Machine Learning and Applications (ICMLA), Cancun, Mexico, 2017: 806-811.), the historical reference signal received power (RSRP) sequence of the serving cell and the strongest neighbor cell is used as input to a Long Short-Term Memory (LSTM) network. Then, the state vectors of the serving cell and the strongest neighbor cell output by the LSTM are input into a classification model to predict whether the cell handover is successful. In document 2 (Boutiba K, Bagaa M, Ksentini A. Radio link failure prediction in 5G networks[C]. 2021 IEEE Global Communications Conference (GLOBECOM), Madrid, Spain, 2021: 1-6.), three historical measurement sequences—Reference Signal Receive Quality (RSRQ), Channel Quality Indicator (CQI), and Power Headroom (PHR)—are used as inputs to the LSTM. The LSTM also outputs the predicted values ​​of these three indicators simultaneously. Then, an SVM (Support Vector Machine) is used to determine whether an RLF has occurred based on the predicted values ​​of these three indicators.

[0007] The aforementioned works all utilize a common event prediction paradigm: first, historical measurement report sequences are input into a time-domain prediction model, often a deep learning-based regression model; then, the measurement sequence predictions output by the AI ​​model are input into a specific classifier to determine whether a specific event, such as A1-A5, HOF, or RLF, has occurred. These methods have the advantage of clear structure and strong versatility. However, the final prediction result is affected by the errors of both the time-series prediction model and the classification model. In other words, even if the error in the time-domain prediction is small enough, when there is a lack of sufficient causality between the selected features and the final event, the classifier will struggle to estimate whether the event has occurred using these features.

[0008] Some researchers have chosen to increase the number of input features to achieve end-to-end prediction. Document 3 (Islam MA, Siddique H, Zhang W, et al. A deep neural network-based communication failure prediction scheme in 5G RAN[J]. IEEE Transactions on Network and Service Management, 2023, 20(2): 1140-1152.) not only utilizes a large amount of time-series data from the base station side, including 21 dimensions such as background bit error, but also utilizes the coordinates of the base station, altitude and surrounding environmental information, and even weather features recorded by the weather station. After feature encoding, the feature dimension of the input LSTM model even reaches 728 dimensions. The authors designed an LSTM-based autoencoder that treats RLF prediction as an anomaly detection problem, and achieved end-to-end prediction of RLF. Document 4 (Klus R, Klus L, Solomitckii D, et al. Deep learning basedlocalization and HO optimization in 5G NR Networks[C]. 2020 International Conference on Localization and GNSS, Tampere, Finland, 2020: 1-6.) designs a multi-task model. First, it uses beam-level RSRP to output the UE's location through a fully connected network. Then, it uses historical beam-level RSRP sequences to jointly predict future beam selection, taking the current UE location and the selected beam ID (identifier) ​​as auxiliary features. Document 5 (Lee C, Cho H, Song S, et al. Prediction-based conditional handover for 5G mm-wave networks: a deep-learning approach[J].IEEE Vehicular Technology Magazine, 2020, 15(1): 54-62.) studies the prediction task of conditional handover in 5G millimeter-wave networks, using uplink and downlink throughput, measurement reports (RSSI, RSRP, RSRQ) of serving cell and neighboring cells, TA (Time Advance), and handover identifier (whether there is a handover within the same system within the input time window) as input features.

[0009] Utilizing more features, especially geographic and high-level features, undoubtedly has a positive effect on enhancing the performance of AI models. However, the selection of which auxiliary information to use has always been a controversial topic. For example, the physical layer working group RAN1 has debated the privacy and legality of using UE location-related information to assist beam prediction. On the other hand, using high-level metrics significantly increases the complexity of the model itself and the signaling overhead of cross-layer data collection. Furthermore, challenges such as data heterogeneity between different operators and the asynchronous measurement cycles of different metrics can lead the model into a trap of "significant theoretical gains but difficult engineering implementation."

[0010] LCM, as a fundamental framework proposed by 3GPP, fully considers the application of AI models in practical communication scenarios and plays a positive role in avoiding models falling into the "Ivory Tower Trap". In fact, some works before its proposal already reflected some of the connotations of LCM. Document 6 (He H, Li X, Feng Z, et al. An adaptive handovertrigger strategy for 5G C / U plane split heterogeneous Network[C]. 2017 IEEE14th International Conference on Mobile Ad Hoc and Sensor Systems (MASS), Orlando, FL, USA, 2017: 476-480.) uses the Fast Dynamic Time Warping (FastDTW) algorithm and the density-based clustering (DBSCAN) algorithm to cluster a large amount of UE trajectory data into several subsets, and trains a Hidden Markov Model (HMM) for each subset to predict RSSI. For the received input data, the similarity with the existing subsets is first calculated, and then the corresponding HMM is assigned to perform inference. Document 7 (Bai Y, Zhang J, Sun C, et al. AI-based beam management in 3GPP: Optimizing data collection time window for temporal beam prediction[J]. IEEE Open Journal of Vehicular Technology, 2024,5: 48-55.) quantitatively analyzes the relationship between the UE's motion speed and the length of the input sequence during the data collection phase when studying beam management based on a time-domain prediction model, and proposes that the input length of the model should depend on the UE's speed.Under the same research objective, document 8 (Sun C, Zhao L, Cui T, et al. AI model selection and monitoring for beam management in 5G-Advanced[J]. IEEE Open Journal of the Communications Society, 2024, 5: 38-50.) distinguishes the propagation environment of the UE based on the downlink base station beam measurement report, trains different AI models for different propagation environments, and introduces an additional performance evaluation model to monitor the effect of the beam prediction model, and determines the activation and deactivation of the AI ​​model based on the results of the evaluation model. Summary of the Invention

[0011] To address AI-enhanced mobility prediction tasks and improve the inference reliability and stability of AI models deployed in the access network, this invention provides a model management and reliability enhancement method for AI-driven mobility prediction in a wireless network. By utilizing certain aspects of model lifecycle management, the network side possesses a precise management strategy that can allocate the optimal and unique AI model to different UEs. Furthermore, by quantifying the uncertainty of the AI ​​model, the reliability of each AI model call is evaluated, and corresponding model activation / deactivation mechanisms are deployed accordingly.

[0012] The present invention provides a model management and reliability enhancement method for AI-driven mobility prediction in wireless networks, comprising:

[0013] Step 1, during the offline phase, perform the following steps:

[0014] Step 11: Collect historical data to obtain a full dataset; the historical data includes UE measurement data, UE mobility level, event trigger records, and cell handover records;

[0015] Step 12: The model manager first segments the full dataset according to the UE mobility level, and then configures matching sub-datasets based on the user-side functional requirements and OW / PW step size. The sub-models are then trained and monitored, and their predictive performance and overhead are identified. These sub-models are AI models used to perform temporal prediction of UE measurement data for corresponding functional requirements. User-side functional requirements include radio resource management measurement event prediction, beam management tasks, and RLF prediction tasks. All trained sub-models form a global sub-model set.

[0016] Step 13: The model manager establishes the mapping relationship between model identifier and network / UE side parameters, receives the mobile speed and functional requirements uploaded by the network side or UE side, and selects the optimal unique sub-model from the global sub-model set;

[0017] Step 2: During the online phase, perform the following steps:

[0018] Step 21: Deploy the AI ​​model selected by the model manager to the network side or UE side, collect the most recent historical data for the AI ​​model on the network side or UE side to build a local dataset, evaluate the predictability of the input sequence of the AI ​​model based on the local dataset, and calculate the prediction error of the AI ​​model.

[0019] Step 22: If the prediction error of the AI ​​model does not meet the requirements, set the AI ​​model to deactivated mode and use traditional measurement methods to complete the business requirements; otherwise, activate the AI ​​model, input the historical data most recent to the current time into the AI ​​model, and input the output predicted time series into the event discriminator to determine whether the event has occurred.

[0020] In step 13, the model manager selects models according to the model management strategy, including: for each sub-model Configure comprehensive feature tuples This tuple contains the prediction performance and cost metrics obtained by the sub-model during the monitoring phase, as well as the duration of the sub-model's observation and prediction windows; (Comprehensive feature tuple) It is a sub-model A unique identifier; the model manager determines the user's UE mobility level. and functional requirements Select a set of matching candidate sub-models from the global sub-model set. Model management strategies utilize pre-defined objective functions. , combined with , Derived policy constraints and weighting factors From the candidate sub-model set Select the unique optimal sub-model from the selection process. Objective function Used to quantify the degree of matching between the overall performance of the sub-model and the needs of network-side users.

[0021] In step 21, firstly, four indicators are calculated for the input sequence of each sample in the local dataset: spectral predictability (SP), maximum Lyapunov exponent (LLE), two-state entropy (H2reg), and augmented Dickey-Fuller ADF statistic, to obtain a predictability quantization vector; then, a high-error sample threshold is set. , will be greater than The sample is labeled as negative, otherwise it is labeled as positive. A machine learning classifier is trained. The input of the classifier is the predictability quantization vector of the sample, and the output is positive or negative. The prediction error of the AI ​​model is calculated by counting the number of positive and negative samples.

[0022] In step 2, before and after each AI model call, the predictability assessment in step 21 is performed to calculate the prediction error of the AI ​​model, and then step 22 is executed to trigger the activation or deactivation control of the AI ​​model.

[0023] The advantages and positive effects of this invention are as follows:

[0024] (1) The method of this invention realizes a closed-loop workflow for the entire lifecycle of AI-enhanced mobility prediction. Unlike existing solutions that mainly focus on partial links such as "data collection - model training - deployment and application", this invention constructs a complete and implementable LCM workflow, covering dataset segmentation, sub-model training, model monitoring, model identification, model selection, model delivery, as well as instance-level predictability assessment and model activation / deactivation control in the online stage. It connects the entire link of AI model from offline construction to online operation, effectively enhances the stability and reliability of AI model deployment and optimization in RAN environment, and enables AI model management to support mobility enhancement tasks in a systematic and closed-loop manner, thereby achieving the enhancement of communication QoS.

[0025] (2) The present invention proposes a unique and optimal sub-model selection mechanism based on Model Management Strategy (MMS). This invention integrates UE mobility level, functional requirements, and OW / PW (observation window / prediction window) structural attributes with model monitoring results into a unified model identification system. Furthermore, it constructs a deterministic model selection mechanism through constraint filtering and weighted cost optimization, enabling the network side to index the unique and optimal target model from the candidate sub-model set based solely on simple state parameters reported by the UE side or available to the network side. This mechanism simultaneously considers prediction accuracy, computational overhead, and timeliness requirements, achieving precise alignment between model resources and service requirements (i.e., functional requirements). Furthermore, since this MMS does not require the introduction of sensitive user service content, personal behavior data, or other highly privacy-sensitive features during model selection, there are no additional UE privacy leakage risks. Simultaneously, its decision-making basis is primarily based on basic state information obtainable from existing systems, eliminating the need for complex additional information collection and interaction processes. Therefore, it offers advantages such as ease of implementation, low additional signaling overhead, and low engineering deployment costs, making it more suitable for widespread application in actual mobile communication systems.

[0026] (3) The method of this invention realizes an instance-level predictability quantification and error control mechanism for online deployment of AI models. Addressing the problem that AI model prediction errors cannot be completely eliminated, making it difficult to guarantee online operational reliability solely based on offline accuracy, this invention introduces an instance-level predictability evaluation scheme for online deployment scenarios. By comprehensively analyzing the spectral characteristics (SP), phase space divergence (LLE), local morphological complexity (H2reg), and stationarity (ADF) of the input measurement sequence, the predictability of the current sample is quantified, and a correlation judgment mechanism is established between the predictability of the input measurement sequence and the reliability of AI model inference. Based on this, combined with a model activation / deactivation strategy, when the reliability of the AI ​​model inference corresponding to the current input sequence is determined to be within an acceptable range, the AI ​​model is activated to meet the functional requirements of the UE; when the reliability of the AI ​​model inference is determined to be low, the AI ​​model is deactivated to prevent it from entering the subsequent decision-making link and falls back to the traditional process defined by the protocol to meet the functional requirements of the UE. Therefore, this invention can effectively identify and avoid unreliable inference during the online operation of AI models, and to a certain extent decouple AI performance from communication QoS assurance, significantly improving the inference reliability and stability of AI models deployed in the access network. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the model management and reliability enhancement method for AI-driven mobility prediction in wireless networks according to the present invention.

[0028] Figure 2 This is a schematic diagram of the event prediction process according to an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram of the model selection process according to an embodiment of the present invention;

[0030] Figure 4 This is a schematic diagram of the model recognition process according to an embodiment of the present invention;

[0031] Figure 5 This is a schematic diagram of the model selection process in an embodiment of the present invention;

[0032] Figure 6 This is a schematic diagram of the process for evaluating the reliability of model inference in an embodiment of the present invention;

[0033] Figure 7 This is a flowchart illustrating the inference reliability enhancement based on model deactivation / activation in an embodiment of the present invention. Detailed Implementation

[0034] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0035] The model management and reliability enhancement method for AI-driven mobility prediction in wireless networks implemented in this invention, compared with the traditional AI-enhanced mobility prediction workflow, namely the strategy of data collection-model training, testing-model deployment and optimization, considers more details and further ahead, and is more feasible and instructive. It can solve several pain points of traditional AI-based mobility prediction solutions: (a) How to assign AI models to UEs with specific needs? In this process, how to assign a unique "ID" to different AI models, that is, to achieve AI model recognition, this ID needs to be correctly recognized by both the UE and the network side. (b) How to establish a mapping between UE / network side parameters and AI model IDs is also a problem. Only by establishing such a mapping relationship can we ensure that the AI ​​model that meets the functional requirements of the network / UE side is correctly delivered, that is, to achieve AI model delivery. Furthermore, how to quantify UE / network side parameters and AI model IDs, the relative quantity between the two determines the training overhead in the training phase and the signaling overhead in the deployment phase, which must be a direction that needs to be balanced. According to the "no free lunch" theorem, it is unrealistic to expect that training a general model to outperform sub-models specifically trained for each separable subset of the dataset is feasible. Conversely, training an excessive number of sub-models violates Occam's razor, wasting excessive training overhead and further increasing signaling overhead during deployment. (c) Ensuring the correct delivery of AI models to the network or UE side is crucial for maximizing AI-enhanced mobile communication QoS. To this end, by integrating model monitoring results with the mapping logic between model ID and network / UE side parameters, and establishing clear model management guidelines, precise alignment between AI-side model identification, overall performance, and network / UE side parameters and requirements granularity can be achieved.

[0036] By addressing these pain points, the method of this invention realizes the logical link from offline to deployment of AI models, enabling the network / UE side to select the unique and optimal AI model for delivery based on given model management criteria, thus achieving model selection.

[0037] Deploying AI models in a network environment presents numerous challenges, most notably that traditional AI-based mobility prediction schemes rely heavily on the accuracy of the AI ​​model for final communication QoS, particularly end-to-end optimization schemes. In fact, even with strong overall performance on test sets, a certain number of unreliable calls occur in a per-call online deployment environment. These unreliable AI model outputs degrade communication QoS, but this damage can be masked to some extent by the "average" evaluation method. Therefore, this invention deploys a predictability quantizer on the online AI model to control its activation / deactivation, thereby mitigating the communication QoS degradation caused by unreliable AI model outputs, improving overall system reliability, and decoupling the dependency between AI-side performance and the final communication QoS optimization to some extent.

[0038] The model management and reliability enhancement method for AI-driven mobility prediction in wireless networks of this invention mainly includes functional modules such as a time-domain prediction model, an event discriminator, a model manager, a predictability evaluation module, and a model activation / deactivation mechanism module. For example... Figure 1As shown, the method of the present invention uses a time-domain prediction model and an event discriminator as functional execution units for specific mobility tasks, a model manager as a bridging unit between offline training and online deployment, and a predictability assessment and model activation / deactivation mechanism as a reliability assurance unit for the online operation phase, thereby jointly constituting a set of LCM workflows for AI-enhanced mobility prediction tasks. First, in the offline phase, the model manager segments the original dataset to form subsets that match the UE's mobility level, functional requirements, and OW / PW configuration. It then trains, monitors, identifies, and labels the corresponding sub-models, establishing a mapping between "model identifier—network / UE side parameters." Second, in the online phase, the network side, guided by model management criteria and based on the UE's current mobility level, service requirements, and deployment constraints, selects the unique and optimal sub-model from the candidate sub-model set and delivers it. Subsequently, the selected measurement time-domain prediction model outputs future measurement sequences, and the event discriminator classifies A3 events, RLF, or beam management-related events according to protocol rules to support the network side's advance decision-making. Furthermore, before and after each model call, the predictability assessment module evaluates the predictability and potential error risks of the current input instance and triggers model activation or deactivation control accordingly: if the call is deemed reliable, the AI ​​model output is allowed to participate in network decision-making; if the call is deemed unreliable, the process reverts to the original protocol flow to execute the corresponding function. Therefore, this invention achieves full-process collaboration from data collection, sub-model construction, model selection to online reliability control, which not only ensures that the AI ​​model can be correctly delivered, uniquely selected and stably invoked, but also effectively suppresses the negative impact of unreliable AI output on communication QoS, ultimately achieving a comprehensive enhancement of the stability, reliability and intelligence level of the mobile communication system.

[0039] First, the method of this invention collects data. The full dataset collected in this embodiment can be provided by a simulation platform or by field measurements. The features included are only basic UE measurement data, UE mobility level, event triggering records, and cell handover records. Among them, the UE measurement data may include: (1) core physical layer measurement parameters, such as Reference Signal Received Power (RSRP), Reference Signal Received Quality (RSRQ), Signal-to-Noise Ratio (SINR), Received Signal Strength Indicator (RSSI), Physical Cell Identifier (PCI), etc.; (2) positioning parameters, such as Global Cell ID, Positioning Reference Signal Identifier, etc.; (3) the frequency and cell type of the measurement object, etc. The specific data included in the UE measurement data can be determined according to the actual network situation. The UE mobility level is set according to different levels based on the user's movement speed. This embodiment of the invention sets three levels.

[0040] Secondly, the full dataset is segmented using the model manager to form sub-datasets that match the UE mobility level, functional requirements, and OW / PW configuration. The corresponding sub-models are then trained, monitored, identified, and labeled, establishing a mapping relationship between "model identifier—network / UE side parameters." Based on the mobility speed and service requirements uploaded by the network or UE side, the optimal unique sub-model is selected from the global sub-model set and delivered to the network or UE side for deployment.

[0041] Then, on the network side or UE side, the most recent historical data of the AI ​​model is collected to build a local dataset. Based on the local dataset, the predictability of the input sequence of the AI ​​model is evaluated, the prediction error of the AI ​​model is calculated, and the model activation / deactivation mechanism is judged. When the prediction error requirement is met, the selected AI model is activated to output the future measurement sequence. The event discriminator judges A3 events, RLF or beam management related events according to the protocol rules to support the network side to make decisions in advance.

[0042] (1) First, the implementation of the time-domain prediction model is explained. The time-domain prediction model, also known as the AI ​​model, is based on a deep learning model to predict UE measurement data in the time domain. The specific input and output content is closely related to the current communication optimization goal, for example:

[0043] For measurement event prediction in radio resource management, the input / output of the time-domain prediction model is usually the L3 RSRP (Reference Received Power) of the serving cell / neighboring cell in the past / future, where L3 represents the radio resource control layer. For beam management tasks, the input of the time-domain prediction model is usually the historical best beam ID or L1 beam-level RSRP, where L1 represents the physical layer. The current candidate beam can also be used as a covariate to help predict the best future beam. For RLF prediction tasks, the input and output are usually based on UE measurements and can include the SINR (Signal-to-Interference-plus-Noise Ratio) and RSRP of the serving cell.

[0044] Temporal prediction is a prerequisite and a necessary step in mobility prediction, and its predictive performance is directly related to the improvement of global communication QoS quality. Furthermore, temporal prediction can provide advance knowledge of future channel quality, which not only reduces measurement overhead to some extent but also provides a time window for network / UE to make decisions in advance.

[0045] (2) Next, the implementation of the event discriminator is explained. The event discriminator is a rule-based discriminator whose specific rules are aligned with the 3GPP standard. After receiving the future measurement sequence output by the measurement time-domain prediction model, it determines whether a specific event has occurred, providing a basis for subsequent decision-making.

[0046] For example, predicting A3 events requires deploying a measurement prediction model that can predict the L3 RSRP of the serving cell and neighboring cells in the future. Then, an A3 event discriminator is deployed. Based on the corresponding Time to Trigger (TTT) window length, it receives a portion of the real historical L3 RSRP of the serving cell and neighboring cells, along with the L3 RSRP output by the AI. Finally, it outputs whether the A3 event has occurred. The process is as follows: Figure 2 As shown.

[0047] (3) Explain the implementation of the model manager. The model manager contains a series of components for sub-dataset splitting of the collected full dataset, model monitoring, model identification, model selection and model delivery.

[0048] The core functionality of the model manager includes: for a large dataset, finding a feasible basis for alignment and segmentation, and performing offline training. After balancing model performance and training overhead, a finite and reasonably sized set of sub-models is obtained. These trained sub-models are then monitored offline, primarily evaluated in terms of temporal prediction accuracy, computational cost, and timeliness. Furthermore, each trained sub-model is uniquely labeled based on its characteristics, and a mapping is established between it and network / UE-side parameters. The goal is to ensure that for networks with different functional requirements or UEs with varying mobility levels, the appropriate sub-model with matching functionality and overall performance can be correctly identified and assigned when an AI model is needed. To achieve this, a model management framework needs to be designed, combining model monitoring results with the mapping logic between model identifiers and network / UE-side parameters, to truly address the question of "what kind of sub-model does the network / UE need under what circumstances?" At this point, if... Figure 3 As shown, once the current state of the network / UE side is determined, the required AI functions can be determined. Under the given model management criteria, based on the sub-model-network / UE side mapping logic, a unique and optimal sub-model can be selected for delivery and invocation.

[0049] (3.1) The model manager implements dataset splitting, sub-model training and model recognition.

[0050] For the offline training phase of the mobility prediction task, the dataset is first divided into subsets based on the UE's mobility level. The UE's mobility level is usually reflected by the UE's motion type or speed range. In this embodiment of the invention, the UE mobility level is divided into three levels: low, medium, and high, which can be represented by the numbers 1, 2, and 3. Level 1 usually corresponds to users walking, level 2 usually corresponds to users driving non-motorized vehicles, and level 3 usually corresponds to users driving motorized vehicles.

[0051] Next, a corresponding set of sub-models is trained for each subset of data. For a given set of sub-models, each sub-model is primarily determined by three factors: UE requirements and input / output step size (Observation Window (OW) / Prediction Window (PW)). The functional requirements of the UE are most closely related to the features of the AI ​​model's input and output. For example, for requirements such as A3 event prediction and handover prediction, only cell-level RSRP within a certain OW and PW window is needed as input / output features to drive model training and inference; while for beam management tasks, beam-level RSRP in the spatial / temporal domains is required as input / output features. Therefore, selecting the feature combination of model input and output needs to match the UE's functional requirements, i.e., establishing a mapping relationship between the model's input / output features and the UE's requirements, allowing for a one-to-one correspondence. After determining the UE mobility level and functional requirements, the sub-models in each set of sub-models at this point only differ in OW and PW lengths.

[0052] like Figure 4 As shown, the model identification process can be achieved through three segmentation processes. First, the entire dataset is divided according to the UE mobility level. Then, the model types are divided according to the UE functional requirements. Finally, based on the different lengths of the model input / output data OW / PW, sub-models under the same function are trained to obtain a set of sub-models under each functional requirement of each UE mobility level.

[0053] (3.2) The model manager implements model monitoring, model selection, and model management criteria.

[0054] In this invention, model monitoring mainly refers to the testing of AI models on the offline side. The relevant test KPIs mainly focus on the prediction performance and overhead of each sub-model. The prediction performance indicators are related to the task of the current sub-model. For example, for regression tasks, prediction performance indicators can be MAE, MSE, MAPE, etc., and for classification tasks, prediction performance indicators can be accuracy, recall, F1 score, etc. The model overhead mainly focuses on training and inference overhead. The former can be reflected by indicators such as the number of parameters and the maximum memory usage under a specified batch size, while the latter can be reflected by the floating-point computation amount of each inference.

[0055] For each sub-model, before offline training and testing, its model identification is determined based on UE mobility level, UE functional requirements, and OW and PW lengths. After offline training and testing, model monitoring will add a new identifier to it, namely the prediction performance and overhead of the sub-model.

[0056] Model selection, as the core implementation step of this invention, aims to establish a deterministic optimization mechanism. This mechanism allows the network side to directly identify the unique sub-model that best suits the current scenario and offers the optimal overall performance simply by obtaining the user equipment's mobility level and functional requirements. To achieve this goal, this invention introduces a Model Management Strategy (MMS) as a key decision node. This strategy constructs a logical adaptation link between the network side's state awareness and requirement parameters and the sub-model's attribute identifiers. This serves as a bridge between upper-layer business requirements and lower-layer model resources, guiding the precise execution of the model selection process.

[0057] The implementation process of MMS essentially constructs a constraint optimization mechanism based on a multi-dimensional cost function, aiming to transform the qualitative requirements of the network side into quantitative optimization decisions for a set of sub-models. Specifically, this strategy abstracts the multi-dimensional performance indicators of each sub-model obtained during the aforementioned model monitoring phase—mainly including the theoretical effective lead time representing timeliness, the prediction error indicator representing reliability, and the computational overhead indicator representing resource consumption—as input variables of the objective function. Simultaneously, it maps the UE's current mobility level and functional requirements to weight coefficients and constraint boundaries in the objective function. Through this mapping mechanism, MMS can logically and dynamically adjust its preferences for different performance dimensions, such as automatically increasing the reliability weight as the UE's mobility level increases, thereby establishing an adaptation link between model identification attributes and the dynamic state of the network side at a mathematical level.

[0058] For a given UE mobility level Functional requirements From the global sub-model set Matching sub-models are selected from the samples to form a candidate sub-model set. ,satisfy:

[0059] (1)

[0060] in, For sub-model The structural attribute identifier is determined by the sub-model. and The duration composition, input and output feature types of the sub-model are determined by the UE functional requirements. The only certainty. The sub-models represented by OW, PW, and UE. There is a certain correspondence, and the detailed correspondence will be discussed below.

[0061] To achieve quantitative representation and accurate selection of sub-models, for each sub-model... Configure comprehensive feature tuples The structural properties of the tuple fusion sub-model and the quantitative indicators for model monitoring are as follows:

[0062] (2)

[0063] In the formula, For sub-model The model monitoring results, in this embodiment, include prediction error, which characterizes the prediction accuracy. And the quantized value of floating-point operations representing computational overhead. ; For the model observation window duration, This is the duration of the model prediction window.

[0064] MMS simultaneously receives parameters from the UE side. , and the comprehensive feature tuples of each sub-model in the candidate sub-model set Through a pre-defined strategy-specific objective function , combined with , Derived policy constraints and weighting factors From the candidate sub-model set The unique, definitive, and optimal sub-model is selected from the samples. This process can be described as follows:

[0065] (3)

[0066] Wherein, objective function Used to quantify the degree of matching between the overall performance of the sub-model and the needs of network-side users. The model selection is dynamically adjusted based on differences in UE mobility levels and functional requirements to achieve speed adaptation and requirement matching.

[0067] In the quantitative representation stage of model selection, each sub-model Using comprehensive feature tuples that include structural attributes and monitoring indicators To uniquely identify the user, and for the actual solution of formula (3), in order to achieve logical adaptation between the network-side user requirements and the sub-model identifier, MMS needs to... The four-dimensional tuple mapping represents three key decision dimensions: reliability, cost, and timeliness. Among these, Although not directly considered an independent term in the cost function, it implicitly acts as a core parameter determining the model's ability to capture temporal features. and Within. Specifically, increase By incorporating historical measurement information over a longer span, the long-term dependence characteristics of the channel can be captured, thereby significantly reducing prediction errors. However, the resulting increase in the dimensionality of the input sequence inevitably leads to higher training and inference overhead. The synchronous rise.

[0068] On the other hand, targeting Utility assessment needs to be based on the theoretical maximum lead time. Determined, defined as:

[0069] (4)

[0070] in, For business functional requirements Specific protocol time boundaries, such as the trigger time delay (TTT) in A3 events or T310 (Radio Link Failure Timer) and N310 (Out-of-Sync Counter) in RLF, represent the maximum effective advance time for event prediction. Since the triggering determination of measurement events strictly depends on the measurement sequence within this window, research indicates that AI models struggle to predict events triggered beyond this window length, and the event prediction advance time is primarily distributed around this time window. Therefore, this invention constructs a truncation function based on protocol parameter configuration, which truncates the single sub-model side parameters... Transformation into the theoretical maximum lead time related to both sides At this point, the model identifier in formula (3) is a four-dimensional tuple. Abstracted into a three-dimensional decision vector for policy computation This achieves the mapping from sub-model identifier attributes to the network-side decision space.

[0071] Furthermore, the parameters received by MMS This is concretized as a set of constraints and weight vectors driven by the network-side state, i.e.:

[0072] (5)

[0073] in, It is a velocity adaptive weight vector, constraint set This constitutes a hard constraint boundary vector based on business requirements, and its mathematical expression is:

[0074] (6)

[0075] in:

[0076] This represents the task tolerance error threshold, which can be a pre-specified parameter, such as 3dB, or a parameter related to task requirements, such as A3 Offset, or a pre-specified parameter. It represents the minimum accuracy requirement required to ensure the accuracy of event prediction.

[0077] This refers to the deployment computing power budget, which is determined by the current hardware resource utilization and maximum allocable computing power of network edge nodes or user terminals.

[0078] This represents the minimum action execution delay, which is a UE task requirement parameter. Taking A3 event prediction as an example, at this time... It covers the sum of air interface signaling interaction time and hardware processing latency required from the model outputting prediction results to the issuance of the handover command;

[0079] This indicates the protocol time boundary, which is a parameter related to the UE task requirements, namely the event evaluation window length defined by the protocol.

[0080] Based on the above definition, due to the candidate sub-model set Based on a finite number The MMS first utilizes a set of discrete points composed of pre-defined combinations. Each constraint component in the algorithm filters the set of sub-models, eliminating sub-models that do not meet the constraints, in order to define the discrete feasible region. For any sub-model that enters the feasible region Its attribute indicators must meet the following constraints:

[0081] (7)

[0082] This constraint ensures that the selected sub-model not only has reliability at the model algorithm level, but also feasibility for network-side deployment and protocol compatibility.

[0083] After determining the feasible region, MMS constructs a weighted cost objective function. Achieve deterministic optimization of the optimal sub-model. Function Using velocity adaptive weight vector The objective function for dynamically adjusting preferences for different performance dimensions, with its overall cost, is:

[0084] (8)

[0085] in, and All are normalization operators for the corresponding physical quantities, with weights All three are functions relating to UE mobility level and functional requirements, with the weights varying according to the UE's mobility level. Dynamic scheduling follows the following monotonicity constraints:

[0086] (9)

[0087] According to formula (9), for a high-speed moving UE, fast channel fading leads to weakened time correlation and increased prediction uncertainty. MMS addresses this by increasing the reliability weight. To select a more accurate conservative sub-model, which typically corresponds to a shorter PW and a longer OW; while for UEs with low-speed movement and gradual channel changes, the strategy focuses on increasing the timeliness weight. In order to obtain a greater lead time for decision-making, As a supplement, for network-side deployments with ample inference budgets, there is a greater tendency to trade higher overhead for reliability; while for edge deployments such as UE-side deployments, a more conservative approach is taken towards the aforementioned operations.

[0088] Ultimately, MMS in the discrete feasible region The algorithm solves the global cost minimization problem and outputs a unique and optimal sub-model. :

[0089] (10)

[0090] like Figure 5 As shown, during the model selection execution phase, the MMS first narrows down the candidate sub-model set containing specific input-output structures based on the speed and service requirement parameters reported by the UE. Then, it applies hard constraints determined by protocol parameters or hardware conditions, such as maximum allowable latency, deployment computing power budget, or task-specific error thresholds, to filter the set and eliminate sub-models that do not meet the basic feasible domain. For the selected effective sub-models, the MMS calculates their comprehensive cost score by substituting weight coefficients. By solving the constraint-weighted cost minimization problem, it uniquely indexes the sub-model with globally optimal comprehensive performance from the set. This process ensures that the network side does not need to process massive amounts of underlying monitoring data; it can obtain a deterministic and unique optimal model configuration based solely on basic state parameters, achieving closed-loop control of the model management strategy between offline monitoring and online selection.

[0091] (4) Implement a predictability assessment module and a model activation / deactivation mechanism.

[0092] In the online phase after AI model deployment, predictability is evaluated and corresponding invocation decisions are generated before and after each AI model is invoked. The main content includes three parts: instance-level time series predictability calculation method, instance-level predictability to instance-level prediction error mapping algorithm, and model deactivation / activation strategy.

[0093] (4.1) Implementing instance-level predictability assessment. To assess the predictability of instance-level measurement sequences, primarily the input sequences of AI models, this invention selects four metrics to reflect their predictability:

[0094] 1) Spectral Predictability (SP). SP measures the concentration of frequency components across a time series to assess whether the time series exhibits a clear periodicity or quasi-periodicity. A SP closer to 1 indicates that energy is concentrated on a few frequency components, resulting in strong rhythmicity and predictability. A SP closer to 0 indicates a uniform energy distribution (close to white noise), lacking a dominant rhythm and exhibiting low predictability.

[0095] For the input sequence of SP The linear detrending method is used to eliminate the long-term trend in the sequence. The sequence after detrending is: ,right After performing an FFT (Fast Fourier Transform), the power spectral density (PSD) is calculated:

[0096] (11)

[0097] in, The square of the modulus of the complex number. This represents the normalization factor, which corresponds to the length of the input sequence.

[0098] right Normalization is performed, where the first... Normalized power of each frequency component for:

[0099] (12)

[0100] in, for Frequency division Count them. Spectral entropy for:

[0101] (13)

[0102] Input sequence The calculation is as follows:

[0103] (14)

[0104] 2) Largest Lyapunov Exponent (LLE). LLE directly assesses the sensitivity of a measured sequence to small perturbations by quantifying the exponential divergence rate of adjacent trajectories in phase space. When LLE is greater than 0, it indicates that the measured sequence trajectory is highly sensitive to subtle changes; even if similar segments exist in history, its subsequent evolution will diverge rapidly, exhibiting strong randomness or chaotic characteristics and low predictability. If LLE is less than 0, it indicates that the measured sequence trajectory has convergence or stability in phase space; its temporal evolution logic is less affected by initial condition perturbations, and its predictability is high.

[0105] Input sequence for LLE First, its state space is reconstructed through time-delay embedding:

[0106] (15)

[0107] in, The length of the input sequence for the LLE test. , The embedding dimension has a value of 3. For time delay, , Indicates the measurement period; For state index, the value range is: .

[0108] For each refactoring state Find the nearest neighbor state (Excluding points with nearby times to avoid interference from short-term correlations), their Euclidean distance for:

[0109] (16)

[0110] Over time Then, the distance between adjacent states :

[0111] (17)

[0112] For the divergence rate of a state pair :

[0113] (18)

[0114] From the The average of the divergence rates of the independent state pairs is obtained as follows:

[0115] (19)

[0116] 3) Two-Regimes Entropy (H2reg). H2reg measures the complexity of a time series' local trend structure by evaluating the degree of orderliness in the sequence during the switching between "growth" and "decline" states from a symbolic dynamics perspective, reflecting the regularity of the sequence's rise and fall modes. Unlike SP global frequency analysis, H2reg focuses on the arrangement and combination of local patterns within the time domain. The closer H2reg is to 0, the more dominant the measured sequence is, such as continuous rise or specific regular oscillations, containing more specific modes and having higher predictability; the closer H2reg is to 1, the more random the measured sequence switches between rises and falls, lacking specific structure, with high disorder and low predictability.

[0117] For the input sequence of H2reg , To determine the length of the input sequence for the H2reg test, the continuous numerical sequence is first transformed into a binary symbolic sequence using a first-order difference symbolization method. , for the The first of the binary symbol sequences Points for:

[0118] (20)

[0119] Here, 1 represents a growth state, and 0 represents a decline or a stable state.

[0120] To capture the dynamic structure of the sequence, a structure of length is constructed. If there is a symbol block, then the total number of possible symbol patterns is Species. Count each symbol pattern. probability of occurrence Then the two-state entropy for:

[0121] ;(twenty one)

[0122] Normalize it to obtain as follows:

[0123] ;(twenty two)

[0124] 4) Augmented Dickey-Fuller (ADF) statistic. The ADF statistic is used to assess the stationarity of the statistical properties of a time series in the time domain through the ADF test. Its core function is to determine the stationarity of the series by detecting the presence of a unit root. The statistical properties of a stationary series do not drift significantly over time, providing a stable learning benchmark for AI models and exhibiting high predictability. Conversely, non-stationary series are often accompanied by random walks or trend abrupt changes, causing the statistical distribution to drift over time. AI models struggle to capture stable time-series dependencies, resulting in low predictability. The more negative the ADF statistic, the stronger the measurement of series stationarity. An ADF statistic ≥ -5 indicates significant non-stationarity of a single series.

[0125] For the input sequence An autoregressive model containing an intercept term and a time trend term is constructed for hypothesis testing:

[0126] ;(twenty three)

[0127] in, Representing time series The first difference, Indicates time, For drift term, The coefficient for the time trend term. It is a one-period lagging sequence The coefficient. The length of the input sequence for the ADF test is given by the ADF statistic, which is the logarithm of the coefficient. Performing significance testing Statistics.

[0128] For each sample in the training, validation, and test sets, four predictability quantization metrics are calculated, resulting in a predictability quantization vector for each sample. , The ADF statistic reflects the predictability of instance-level measurement sequences. On the network or user side where the AI ​​model is deployed, the most recent historical data is collected for the AI ​​model to obtain a local dataset. This local dataset is divided into a training set, a validation set, and a test set. The samples in the dataset are the input sequences of the AI ​​model, and the sample labels are the corresponding input sequences of the AI ​​model.

[0129] After obtaining the instance-level predictability quantization vector, it is necessary to further map the predictability to the prediction error of the AI ​​model. This invention achieves this based on a machine learning (ML) classifier:

[0130] Based on the local dataset and the AI ​​model's output data on that dataset, the error for each sample is calculated. For example... Figure 6 As shown, given a high error sample threshold Compared with equation (7), Focusing on the potential errors that may arise from a single AI model inference execution, It focuses more on describing the average error of a sub-model; while and Similarly, It can also be set to task-related or pre-specified. This invention will be greater than... Samples are labeled as negative if they are not, and as positive otherwise. An ML classifier is trained using the local dataset; the classifier can be SVM or RF. The input to the classifier is the predictability quantization vector of the sample, and the output is the positive / negative class. The prediction error of the AI ​​model is calculated by counting the number of positive and negative samples.

[0131] (4.2) Model Activation / Deactivation. Model activation and deactivation are based on the classification results output by the ML classifier. Specifically, when the proportion of positive class output by the classifier is high, it is considered that the measurement sequence received by the AI ​​model is highly predictable. After calling the AI ​​model, the predicted value output by its inference is reliable. In this case, the network side / UE side will implement the original functional requirements of the UE based on the AI ​​model, such as RLF prediction. Conversely, it is considered that the output value obtained by inference after calling the AI ​​model is unreliable. In this case, the network side will perform a "deactivation" operation on the AI ​​model, and will not call the AI ​​model but will implement the corresponding functions based on the original protocol process.

[0132] Taking RLF prediction as an example, this process is in Figure 7 The paper states that, based on the output of the ML classifier, if the proportion of high-error samples exceeds the set threshold, the AI ​​model is deactivated and traditional measurement methods are used for RLF prediction; otherwise, the AI ​​model is activated, the output time series is predicted based on the input time series, and the event discriminator is used to detect whether RLF has occurred.

[0133] As can be seen from the above description, the method of the present invention can achieve the following technical effects:

[0134] (1) A closed-loop model lifecycle management mechanism connecting offline construction and online operation has been constructed. This invention does not use AI models as isolated algorithm modules, but rather constructs a closed-loop workflow covering the entire lifecycle of the mobility enhancement task, including dataset segmentation, sub-model training, model monitoring, model identification, model selection, model delivery, and online activation / deactivation control. This opens up the complete technical link from offline construction to online deployment and operation of AI models, laying the foundation for the manageability, deployability, and controllability of models.

[0135] (2) Achieving a unique and optimal selection of candidate sub-models based on simple state parameters. The method of this invention integrates UE mobility level, functional requirements, OW / PW structural attributes, and model monitoring results into the model management strategy. Through constraint filtering and weighted cost optimization, it achieves deterministic selection of the target model from the candidate sub-model set. This process relies only on simple state parameters reported by the UE or directly obtained by the network side, without introducing additional sensitive information or complex additional interactions. Therefore, it has the advantages of selection accuracy, ease of implementation, low signaling overhead, and low privacy risk.

[0136] (3) Establishing an instance-level predictability quantification mechanism for online deployment scenarios. In view of the problems that AI model prediction errors are difficult to completely eliminate and offline average accuracy is difficult to fully represent the reliability of online calls, the method of this invention extracts features such as the spectral characteristics, phase space divergence, local morphological complexity and stationarity of the input measurement sequence to perform instance-level quantification of the predictability of the current input sample, and establishes a correlation judgment mechanism between predictability and the reliability of AI model inference, so that the system can perceive the current call risk during the actual operation of the model.

[0137] (4) Error suppression and process rollback for unreliable AI inference are achieved through activation / deactivation control. After completing the instance-level predictability assessment, the method of this invention further combines the model activation / deactivation mechanism to dynamically control whether the AI ​​model participates in subsequent service processing: when it is determined that the reliability of the model inference corresponding to the current input is within an acceptable range, the AI ​​model is activated to meet the UE functional requirements; when it is determined that its inference reliability is low, the AI ​​model is deactivated so that it does not enter the subsequent decision-making link and rolls back to the traditional process defined by the protocol to complete the corresponding function. Through this mechanism, the adverse effects of unreliable AI inference output on communication QoS can be effectively avoided.

[0138] (5) Achieving coordination and unity between AI capability enhancement and communication QoS assurance. The key to the method of this invention is not simply to improve the offline prediction accuracy of the model, but to enable the AI ​​model to have the engineering characteristics of being deployable, schedulable, controllable and fallback in access network mobility enhancement scenarios through the overall design of "offline management - online selection - online evaluation - error control - process rollback". This ensures the system's operational stability and communication QoS while leveraging the AI ​​enhancement role, thereby enhancing the application value of the AI ​​model in actual networks.

[0139] Except for the technical features described in the specification, all other technologies are known to those skilled in the art. Descriptions of well-known components and technologies are omitted in this invention to avoid redundancy and unnecessary limitation. The embodiments described above do not represent all embodiments consistent with this application. Various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this invention are still within the protection scope of this invention.

Claims

1. A method for model management and reliability enhancement of AI-driven mobility prediction in wireless networks, characterized in that, include: Step 1, during the offline phase, perform the following steps: Step 11: Collect historical data to obtain a full dataset; the historical data includes UE measurement data, UE mobility level, event trigger records, and cell handover records; Step 12: The model manager first segments the full dataset according to the UE mobility level, and then configures matching sub-datasets based on the user-side functional requirements and OW / PW step size. The sub-models are trained and monitored, and the prediction performance and overhead of the sub-models are identified. The sub-models are AI models used to implement the temporal domain prediction of UE measurement data for the corresponding functional requirements. The types of user-side functional requirements include measurement event prediction for radio resource management, beam management tasks, and RLF prediction tasks. All trained sub-models form the global sub-model set; Step 13: The model manager establishes the mapping relationship between model identifier and network / UE side parameters, receives the mobile speed and functional requirements uploaded by the network side or UE side, and selects the optimal unique sub-model from the global sub-model set; Step 2: During the online phase, perform the following steps: Step 21: Deploy the AI ​​model selected by the model manager to the network side or UE side, collect the most recent historical data for the AI ​​model on the network side or UE side to build a local dataset, evaluate the predictability of the input sequence of the AI ​​model based on the local dataset, and calculate the prediction error of the AI ​​model. Step 22: If the prediction error of the AI ​​model does not meet the requirements, set the AI ​​model to deactivated mode and use traditional measurement methods to complete the functional requirements; otherwise, activate the AI ​​model, input the historical data most recent to the current time into the AI ​​model, and input the output predicted time series into the event discriminator to determine whether the event has occurred. In this context, UE stands for User Terminal, OW stands for Observation Window, PW stands for Prediction Window, RLF stands for Radio Link Failure, and AI stands for Artificial Intelligence.

2. The method of claim 1, wherein, In step 12, the UE mobility level is divided according to the user's movement type or movement speed range; for users with different UE mobility levels, the model types are divided according to functional requirements, and each model is trained to obtain different sub-models based on the OW / PW step size of the model input / output.

3. The method of claim 1, wherein, In step 12, the model manager monitors each sub-model after offline training, which means testing the sub-models to obtain their predictive performance and cost.

4. The method according to claim 1, characterized in that, In step 13, the model manager selects models according to the model management strategy, including: For each sub-model Configure comprehensive feature tuples This tuple contains the prediction performance and cost metrics obtained by the sub-model during the monitoring phase, as well as the duration of the sub-model's observation and prediction windows; (Comprehensive feature tuple) It is a sub-model A unique identifier; The model manager determines the user's UE mobility level. and functional requirements Select a set of matching candidate sub-models from the global sub-model set. ; Model management strategies utilize pre-defined objective functions. , combined with , Derived policy constraints and weighting factors From the candidate sub-model set Select the unique optimal sub-model from the selection process. Objective function Used to quantify the degree of matching between the overall performance of the sub-model and the needs of network-side users.

5. The method according to claim 4, characterized in that, In step 13, for each sub-model Configuration of comprehensive feature tuples ,in For sub-model The structural attribute identifier, including the model's observation window duration. and prediction window duration ; For sub-model The monitored prediction performance and cost metrics include prediction error, which characterizes prediction accuracy. and the quantization value of floating-point operations representing computational overhead ; Model management strategy from candidate sub-model set Find the unique optimal sub-model in The process includes: (1) Sub-model Prediction window duration Converted into theoretical maximum lead time express, ,in For functional requirements The protocol time boundary is the maximum effective lead time for event prediction; For sub-model The prediction window duration; the comprehensive feature tuple Mapped to a three-dimensional decision vector for policy computation ; (2) Integrating policy constraints with weighting factors Convert to set ;in, It is a set of constraints on the functional requirements of the network side or the UE side. , This is the task tolerance error threshold. For the deployment computing power budget on the network side or UE side, To minimize the execution delay of the action, This is the time boundary of the protocol; Let be the set of weight vectors, represented as Weight All of these are about UE mobility levels. and functional requirements The function satisfies the monotonicity constraint: ; Further, the objective function Represented as: ； in, and Sub-models Prediction error Quantization value of floating-point operations , and These are the normalization operators for prediction error and floating-point operation quantization values, respectively; Model management strategy first utilize Constraints on sets Perform filtering to eliminate sub-models that do not meet the constraints and determine the feasible region. Then, for the feasible region The sub-models within the model calculate the objective function value, and the optimal sub-model is found from them. .

6. The method according to claim 1, characterized in that, In step 21, firstly, four indicators are calculated for the input sequence of each sample in the local dataset: spectral predictability (SP), maximum Lyapunov exponent (LLE), two-state entropy (H2reg), and augmented Dickey-Fuller ADF statistic, to obtain a predictability quantization vector; then, a high-error sample threshold is set. , will be greater than The sample is labeled as negative, otherwise it is labeled as positive. A machine learning classifier is trained. The input of the classifier is the predictability quantization vector of the sample, and the output is positive or negative. The prediction error of the AI ​​model is calculated by counting the number of positive and negative samples.

7. The method according to claim 1, characterized in that, In step 2, before and after each AI model call, the predictability assessment in step 21 is performed to calculate the prediction error of the AI ​​model, and then step 22 is executed to trigger the activation or deactivation control of the AI ​​model.

8. The method according to claim 1, characterized in that, The AI ​​model described is a time-domain prediction model for UE measurement data based on a deep learning model. The model's inputs and outputs differ depending on the specific functional requirements, including: For measurement event prediction in radio resource management, the model input is the historical L3 RSRP of the serving cell / near cell, and the output is the future L3 RSRP of the serving cell / near cell. L3 represents the radio resource control layer, and RSRP is the reference signal received power. For beam management tasks, the model's input is the historical best beam ID or L1 beam-level RSRP, where LI represents the physical layer and ID represents the identifier; the output is the future best beam. For the RLF prediction task, the model input includes the historical SINR and RSRP of the available serving cells, and the output includes the future SINR and RSRP of the available serving cells. SINR represents the signal-to-interference-plus-noise ratio.

Images