An AI-based eSIM code number full life cycle management system

Abstract

The application relates to the field of eSIM code number management, in particular to an eSIM code number full life cycle management system based on AI. The system comprises a collection module, a regional calculation module, a demand analysis module, a planning module and a strategy adjustment module. The collection module is used for collecting full-amount behavior data of eSIM code numbers, performing data preprocessing and outputting code number behavior data flow. The regional calculation module is used for performing regional effective activation calculation according to the code number behavior data flow and constructing activation amount curves of different regions. The demand analysis module is used for performing multi-point activation growth rate calculation on the activation amount curves of different regions, performing code number activation demand estimation and outputting activation demand values of different regions. The planning module is used for performing dynamic planning and intelligent distribution according to the activation demand values and outputting code number putting strategies. The application realizes accurate putting of eSIM code number resources, thereby improving the resource utilization efficiency of code numbers.

Description

Technical Field

[0001] This invention relates to the field of eSIM number management, and more particularly to an AI-based eSIM number lifecycle management system. Background Technology

[0002] In the actual operation of eSIM numbers, their lifecycle typically encompasses multiple stages, including activation, usage, suspension, resumption, and cancellation, each accompanied by the generation of a large amount of dynamic behavioral data. These behaviors are not only influenced by user habits but also closely related to factors such as regional network load, business strategy adjustments, and changes in market demand. As the scale of eSIM numbers continues to expand, traditional management methods relying on manual experience or simple rules are gradually revealing problems such as low efficiency, delayed response, and uneven resource allocation. In some areas, eSIM resources may be scarce, while in others, resources may be idle, leading to a decline in overall resource utilization. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes an AI-based eSIM number lifecycle management system to resolve at least one of the aforementioned technical issues.

[0004] To achieve the above objectives, this invention provides an AI-based eSIM number lifecycle management system, which includes a data acquisition module, a regional calculation module, a demand analysis module, a planning module, and a strategy adjustment module. The acquisition module is used to: acquire all behavioral data of the eSIM number, perform data preprocessing, and output the number behavior data stream; The region calculation module is used to: perform effective activation calculation of regions based on code behavior data stream, and construct activation curves for different regions; The demand analysis module is used to: calculate the multi-point activation growth rate of the activation curves in different regions, estimate the activation demand for code numbers, and output the activation demand values ​​for different regions. The planning module is used to: perform dynamic planning and intelligent allocation based on the activation demand value, and output a code number delivery strategy; The strategy adjustment module is used to: execute code number distribution operations in different regions based on the code number distribution strategy, and collect the actual activation volume of each region after a period of time; calculate the prediction accuracy of different regions based on the actual activation volume of each region, and then perform distribution strategy parameter adjustment processing.

[0005] In this invention, the acquisition module is used to: acquire full behavioral data of eSIM numbers, perform data preprocessing, and output a number behavior data stream, specifically including: Collect all behavioral data of the eSIM number; the all behavioral data includes activation, suspension, number retention, and account cancellation data; The full set of behavioral data is subjected to denoising, filtering, and outlier removal to obtain standardized data; The standardized data is aligned with statistical timestamps and divided into sliding time windows to output a code-behavior data stream.

[0006] In this invention, the region calculation module is used to: perform effective activation calculation of regions based on code behavior data stream, and construct activation curves for different regions, specifically including: Business areas are identified and classified based on code behavior data streams, and data streams from different areas are extracted. The activation behavior of the data stream in different regions is calculated one by one to obtain activation behavior data for multiple regions; Perform abnormal behavior detection on the data stream and mark invalid behaviors; Based on invalid behaviors, the activation behavior data is used to calculate the effective activation volume and count the effective activations. The effective activation amount is continuously time-series fitted to construct activation amount curves for different regions.

[0007] In this invention, the specific steps for calculating the activation frequency, terminal binding relationship, and communication usage based on data streams from different regions to obtain the user's usage trajectory are as follows: Set invalid service rules, including short-term batch activation, frequent terminal changes, and long-term number occupation without data traffic; Based on the invalid business rules, abnormal behavior detection is performed on the user's usage trajectory, and invalid behavior is marked. Invalid behaviors are risk-marked and classified for management, and behavioral information is collected, uploaded, and stored for management.

[0008] In this invention, the demand analysis module is used to: calculate the multi-point activation growth rate of the activation curves in different regions, estimate the activation demand for code numbers, and output the activation demand values ​​for different regions, specifically including: Multi-point activation growth rate calculation was performed on the activation curves of different regions to obtain the growth rate at multiple time points. Identify the inflection points of phased growth based on the growth rate; Based on the aforementioned stage-specific growth inflection point, a growth trend prediction is made to obtain the activation volume prediction trend. Based on the activation volume prediction trend, the activation demand for code numbers is estimated, and the activation demand value for different regions is output.

[0009] In this invention, the planning module is used to: perform dynamic planning and intelligent allocation based on the activation demand value, and output a code number delivery strategy, specifically including: Identify the regional code inventory pool; Based on the regional code inventory pool, the number of allocable codes, the number of occupied codes, and the number of recyclable codes in different regions are determined to obtain the real-time balance of each region. Based on the activation demand value and the real-time remaining capacity in each region, dynamic planning and intelligent allocation are performed to output the code number delivery strategy.

[0010] In this invention, the specific steps for dynamically planning and intelligently allocating based on the activation demand value and the real-time remaining capacity of each region, and outputting the code number delivery strategy, are as follows: Based on the activation demand value and the real-time remaining resources in each region, the supply and demand relationship is calculated to identify resource gap areas; Calculate the gap value of the resource gap region; Based on the gap values, a hierarchical sorting is performed to obtain an allocation priority sequence; Based on the real-time remaining capacity of each region, fine-grained segmentation is performed to obtain the availability and release cycle of different code numbers; Construct a multi-level resource pool based on the availability and release cycle; Based on the allocation priority sequence and multi-level resource pools, a hierarchical allocation plan is performed, and a code number delivery strategy is output. The hierarchical allocation plan is as follows: priority is given to using immediately allocable codes in the local area to meet the demand. When local resources are insufficient, short-term recyclable codes and cross-regional transfer codes are used in sequence to supplement the resources, forming a phased resource replenishment process. In this invention, the strategy adjustment module is used to: execute code number distribution and allocation operations for different regions based on the code number distribution strategy, and collect the actual activation volume of each region after a period of time; calculate the prediction accuracy of different regions based on the actual activation volume of each region, and then perform distribution strategy parameter adjustment processing, specifically including: The code number distribution strategy is used to perform code number distribution and allocation operations in different regions, and the actual activation volume in each region is collected after a period of time. Based on the actual activation volume in each region, the deviation of the activation demand value is calculated to obtain the predicted deviation value. Statistical tracking of prediction deviations is performed to generate time-series deviations. Prediction accuracy is calculated based on time series deviation values ​​to obtain prediction accuracy for different regions; Based on the prediction accuracy, the delivery strategy parameters are adjusted to output an intelligent delivery management strategy.

[0011] The beneficial effects of this invention are as follows: By uniformly collecting and preprocessing all eSIM number behavior data, structured integration of multi-source data is achieved, effectively eliminating data format differences and time deviations between different systems, and improving data consistency and integrity. By performing effective regional activation calculations based on number behavior data streams and constructing activation curves for different regions, the actual business development status of each region can be intuitively reflected. This helps identify high-activity and low-activity regions, providing a quantitative basis for resource allocation and improving the refinement of regional dimension analysis. By calculating the growth rate of the activation curve at multiple points and further estimating number activation demand, the growth trend changes in different regions can be dynamically captured, avoiding biases caused by single-point-of-time analysis. Potential demand growth areas can be identified in advance, enabling forward-looking prediction of future activation demand, thereby improving the accuracy and timeliness of demand judgment and providing a scientific basis for subsequent planning.

[0012] Dynamic planning and intelligent allocation based on activation demand values ​​enable optimal allocation of code resources across different regions, avoiding over-concentration or under-allocation. Quantifying demand-driven decision-making makes resource allocation more rational, improving overall resource utilization efficiency and enhancing the system's responsiveness to market changes, achieving a dynamic balance between supply and demand. Feedback analysis of code allocation execution results, along with calculation of prediction accuracy based on actual activation volumes, forms an optimization mechanism. This allows for continuous correction of allocation strategy parameters, gradually improving prediction and allocation accuracy and reducing long-term error accumulation. Furthermore, periodic adjustments enable the system to adapt to changes in regional demand, improving overall strategy execution effectiveness and stability. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the structure of an AI-based eSIM number lifecycle management system according to the present invention; Figure 2 A schematic diagram of the effective activation curves for each region; Figure 3 A schematic diagram showing the distribution of activation volume by region; Figure 4 This is a statistical diagram illustrating the effective activation and invalidation of markers. Detailed Implementation

[0014] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0015] This application provides an AI-based eSIM number lifecycle management system. The executing entities of the AI-based eSIM number lifecycle management system include, but are not limited to, mechanical equipment, data processing platforms, cloud server nodes, network upload devices, etc., which can be considered general computing nodes of this application. The data processing platform includes, but is not limited to, at least one of an audio / image management system, an information management system, and a cloud data management system.

[0016] In the embodiments of the present invention, see Figures 1 to 4 The diagram below illustrates the steps of an AI-based eSIM number lifecycle management system according to the present invention. In this example, the steps of the AI-based eSIM number lifecycle management system include: The acquisition module is used to: acquire all behavioral data of the eSIM number, perform data preprocessing, and output the number behavior data stream; The region calculation module is used to: perform effective activation calculation of regions based on code behavior data stream, and construct activation curves for different regions; The demand analysis module is used to: calculate the multi-point activation growth rate of the activation curves in different regions, estimate the activation demand for code numbers, and output the activation demand values ​​for different regions. The planning module is used to: perform dynamic planning and intelligent allocation based on the activation demand value, and output a code number delivery strategy; The strategy adjustment module is used to: execute code number distribution operations in different regions based on the code number distribution strategy, and collect the actual activation volume of each region after a period of time; calculate the prediction accuracy of different regions based on the actual activation volume of each region, and then perform distribution strategy parameter adjustment processing.

[0017] In this embodiment, eSIM number lifecycle behavior data is collected and preprocessed. Data sources include platform logs, core network status data, billing system records, and device-side logs, providing activation, suspension, number retention, account cancellation, and traffic information, respectively. A unified data structure is defined, with core fields including entity ID as a unique identifier, event_type representing the behavior type, event_time as a millisecond-level timestamp, device_id as the terminal identifier, and traffic_volume as the traffic value. Data collection adopts a combined stream and batch approach, with real-time data accessed through a message queue, latency controlled within 2 seconds, and offline data synchronized once daily. The preprocessing process includes deduplication, noise reduction, anomaly removal, and standardization. The deduplication rule is to retain only one record for the same subject and the same behavior within 1 second. Anomaly removal is based on behavioral logic constraints, such as deleting a canceled account that is then reactivated without a new account opening record. Time is uniformly aligned with a granularity of 1 second, and a sliding window is used to construct the behavior sequence, with a window length of 3600 seconds and a step size of 300 seconds. Statistical features are extracted for each window, including the number of activations, the number of downtimes, and the average traffic.

[0018] Effective activation calculations are performed for different regions based on behavioral data streams. Region division is based on base station affiliation and user registration information, using a unified regional coding system. Effective activation is defined as an activation behavior that is not marked as abnormal and meets the minimum duration condition, where the duration threshold is set to 60 seconds. Each entity ID is sorted by time, and the first activation behavior is selected as the activation record, while abnormal behaviors are excluded. Subsequently, activation behaviors are aggregated and statistically analyzed by region using a sliding time window with a window length of 3600 seconds and a step size of 300 seconds. The number of effective activations is counted within each window. To ensure data stability, a smoothing process is introduced, using a moving average method with a length of 5 to smooth the activation count.

[0019] Growth rate analysis was performed on the activation curves of each region to estimate future demand. First, the curves were discretized with a time granularity of 300 seconds. The growth rate was defined as the difference in activation volume between adjacent time points divided by the time interval. A three-point difference method was used to calculate the smoothed growth rate, weighted with 0.6 and 0.4 weights. Then, the growth trend was identified, and the range of variation was determined by calculating the mean and standard deviation of the growth rate. Demand estimation was based on trend extrapolation, dividing the next 24 hours into eight 3-hour periods. Linear or quadratic fitting was used to predict each period, and the model with the smallest mean square error was selected as the result. The prediction results were summed to obtain the total demand value, and a safety factor of 1.2 was introduced to cope with sudden increases. All parameters were derived from historical statistical analysis; for example, the time division and safety factor were determined based on historical fluctuation ranges.

[0020] The system matches activation demand with inventory resources in each region and generates an optimal deployment strategy using dynamic programming. Input parameters include regional demand values, current inventory levels, and resource pool structure. First, an objective function is established to minimize unmet demand and resource waste. Constraints include that the allocation to each region cannot exceed its demand value, and the total allocation across the entire network cannot exceed the total inventory. Regional priority weights are introduced, calculated based on demand urgency and historical growth rate, ranging from 0 to 1. Dynamic programming is implemented by allocating resources region-by-region, prioritizing the needs of high-weight regions. The allocation strategy has three levels: first, using available resources within the current region; second, using short-term recoverable resources; and finally, cross-regional transfers. Inventory status is updated in real-time throughout the allocation process. The output is a code-based deployment strategy, including the allocation quantity for each region, resource source, and deployment time schedule.

[0021] The code allocation process was executed according to the deployment strategy, releasing codes to various regions and initiating actual use. The deployment rhythm was set to once per hour, adjusted to once every 30 minutes if necessary. A 24-hour observation period was then set to collect actual activation data for each region, sourced from activation records in the behavioral data stream. The actual activation volume was compared with the predicted demand to calculate the prediction deviation, and further, the prediction accuracy was calculated, defined as 1 minus the average absolute percentage error. A 7-day sliding window was used to smooth the accuracy rate, avoiding the impact of daily fluctuations. Deployment strategy parameters were adjusted based on the accuracy results: when the accuracy rate was below 0.7, the deployment rhythm was shortened and the single deployment volume was reduced; when the accuracy rate was above 0.9, the deployment volume was appropriately increased and the interval extended. Simultaneously, the regional allocation ratio and scheduling priority were adjusted.

[0022] In a specific embodiment, region A is selected for illustration, and the sampling time window Δt is set to 300 seconds, resulting in 288 time points per day. The collected fields include entity ID, behavior type, timestamp, terminal identifier, and traffic value, with time uniformly converted to milliseconds. Data preprocessing includes deduplication and anomaly removal, where the deduplication rule is to retain only one record for the same entity ID and behavior type within ±1 second. Valid activation is defined as a duration greater than or equal to 60 seconds that has not been marked as an anomaly. Assuming the total activation count within a certain time window is 120 and the abnormal activation count is 20, the valid activation count is calculated as follows: Effective activations = Total activations - Abnormal activations = 120 - 20 = 100.

[0023] The activation sequences obtained from time series statistics are 80, 100, 130, and 150. After smoothing the sequences using a moving average with a window size W=3, the smoothed value at the second point is: Smoothing value = (80 + 100 + 130) ÷ 3 = 310 ÷ 3 ≈ 103.3.

[0024] The final result is the region activation curve Y.

[0025] The growth rate is calculated using the finite difference method. The formula for the first-order growth rate is: r1 = (Yt - Yt-1) ÷ Δt.

[0026] For example: r 12 = (100-80)÷300=20÷300=0.067; r 13 = (130-100) ÷ 300 = 30 ÷ 300 = 0.1.

[0027] The differential growth rate is: r2 = (Yt+1) - (Yt-1) ÷ 2Δt = 150 - 80 ÷ 600 = 70 ÷ 600 ≈ 0.117.

[0028] The overall growth rate is calculated using the weighted formula: r = 0.6 × r1 + 0.4 × r2 = 0.6 × 0.1 + 0.4 × 0.117 = 0.06 + 0.0468 = 0.1068.

[0029] Demand forecasting uses a linear model: Yt = Y0 + r × t, where t is time. Let Y0 = 150, t = 288, then: Y288=150+0.1068×288≈150+30.76≈180.76.

[0030] Introducing a safety factor α = 1.2, the demand value D is: D=Y288×α=180.76×1.2≈216.91≈217.

[0031] If the real-time margin S in region A is 180, then the gap value G is: G = DS = 217 - 180 = 37.

[0032] The priority scoring formula is defined as: Score = 0.5 × Gap Ratio + 0.3 × Supply-Demand Ratio + 0.2 × Growth Rate. The gap ratio = G ÷ D = 37 ÷ 217 ≈ 0.17, and the supply-demand ratio = S ÷ D = 180 ÷ 217 ≈ 0.83. Therefore: Score≈0.5×0.17+0.3×0.83+0.2×0.1068≈0.085+0.249+0.021≈0.355.

[0033] Resource allocation is performed based on the score, for example, 40 code numbers are allocated.

[0034] After the deployment, the observation period was set to 24 hours. The actual activation count was Aactual=200, and the predicted count was Apredict=217. The deviation value E was calculated as follows: E=Actual-Apredict=200-217=-17.

[0035] The relative error RE is: RE = absolute value E ÷ Apredict = 17 ÷ 217 ≈ 0.078.

[0036] The prediction accuracy Acc is defined as: Acc = 1 - RE = 1 - 0.078 = 0.922.

[0037] Set a threshold; a high accuracy rate is defined as an Acc rate greater than 0.9. Adjust the delivery strategy parameters accordingly. The formula for adjusting the amount of waste is: New waste = Original waste × 1 + β, where β = 0.1.

[0038] If the original quantity was 40, then the new quantity would be: 40 × 1.1 = 44.

[0039] The delivery schedule has been adjusted from 60 minutes to 90 minutes, with an adjustment coefficient γ=1.5.

[0040] If the accuracy rate is below 0.7, then adjust in the opposite direction, for example, multiply the delivery volume by 0.9 and shorten the delivery time to 30 minutes.

[0041] In this embodiment, the acquisition module is used to: acquire full behavioral data of the eSIM number, perform data preprocessing, and output a number behavior data stream, specifically including: Collect all behavioral data of the eSIM number; the all behavioral data includes activation, suspension, number retention, and account cancellation data; The full set of behavioral data is subjected to denoising, filtering, and outlier removal to obtain standardized data; The standardized data is aligned with statistical timestamps and divided into sliding time windows to output a code-behavior data stream.

[0042] In this embodiment, multi-source data related to the entire lifecycle of the eSIM number is uniformly collected and modeled. The eSIM number is defined as a unique identifier for a configuration file allocated in the operator platform, namely profile_id. This identifier is mapped to the user number msisdn through the subscription relationship table in the core network user data management system. The mapping data comes from the real-time synchronized data table of the user data management system. All behavioral data covers four core events: activation, suspension, number retention, and account cancellation. Data sources are divided into four systems: SM-DP+ platform logs provide activation behavior data, the core network system provides suspension and number retention status change data, the billing system provides account cancellation data, and device-side logs provide supplementary behavioral information. The collection method adopts a stream-batch integrated architecture, where platform logs and device logs are accessed in real time through message queues, with data latency controlled within 2 seconds. Core network and billing system data are supplemented using a daily batch synchronization method. During data collection, a unified data structure is defined, including: entity ID as the primary key (preferably profile_id); event_type representing the behavior type (values ​​1 to 4 corresponding to activation, shutdown, account retention, and account cancellation, respectively); event_time as a unified timestamp in milliseconds; source_system representing the data source; and confidence_score as the data confidence level (values ​​range from 0 to 1). The confidence level is calculated by weighting data source reliability and time consistency. Data source reliability is set as follows: platform logs 0.95, core network 0.9, billing system 0.92, and device logs 0.8, with weights of 0.7 and 0.3 respectively. To avoid duplicate data, a deduplication rule is defined: when the entity ID plus event_type plus event_time is repeated within ±1 second, only the earliest record is retained. All times are uniformly converted to millisecond timestamps, and the data latency tolerance window is set to 5 minutes.

[0043] The process involves denoising by deleting data that failed to activate and had no subsequent successful activation records. Specifically, records where activation failed and no subsequent successful activation occurred are removed. For the same entity ID exhibiting the same behavior repeatedly within one second, only the earliest record is retained. State transition constraints are also implemented; for example, if an activation occurs after account closure and there is no record of account reopening, it is considered an illegal state transition and deleted. Next, confidence-based filtering is performed with a threshold of 0.6. Data below this threshold is directly removed. This threshold, determined based on historical statistical experiments, ensures that the vast majority of valid data is retained. Finally, time-series filtering is applied using median filtering with a window size of 5 consecutive events. If a behavior occurs less than 20% of the time within the window and is inconsistent with preceding and following behaviors, it is considered transient noise and removed. For example, in a sequence of behaviors like activation, shutdown, activation, activation, activation, the shutdown event in the middle will be deleted. Next, outlier detection is performed, categorized into time and frequency dimensions. For the time dimension, the interquartile range (IQR) method is used to calculate the IQR of the data. Data exceeding the upper bound or falling below the lower bound is considered anomaly; for example, activation-to-cancellation time less than 10 seconds or activation delay exceeding 300 seconds are considered anomalous data. For the frequency dimension, the frequency of behavior per unit time is defined, i.e., the number of behaviors per hour. When this value is greater than 50, it is considered abnormal traffic or a system error. Finally, data standardization is performed, including unifying time to millisecond-level timestamps, converting behavior types to four-dimensional vector representations, and normalizing the download latency of continuous variables, mapping 0 milliseconds to 5000 milliseconds to the 0-1 interval.

[0044] The time stamp alignment process is performed to unify the time base and set the alignment granularity to 1 second. The alignment method combines nearest neighbor merging and forward padding. When the time difference between two events is less than 2 seconds, they are merged into the same time point. For missing time points, forward padding is used to fill in the gaps. The maximum padding window is set to 30 seconds; if the time difference exceeds this range, no padding is performed. After time alignment, a sliding window is constructed for feature extraction. The window length is set to 3600 seconds, and the step size is set to 300 seconds, meaning a data window based on the past hour is generated every 5 minutes. Multiple statistical features are calculated within each window, including activation count, shutdown count, number retention count, account cancellation flag (0 or 1), average latency calculated based on normalized latency, and behavior distribution entropy used to measure behavior complexity, calculated based on the proportion of each behavior within the window. Each time window generates a fixed-dimensional feature vector, ranging from 6 to 10 dimensions, and a behavior sequence is constructed in chronological order, with a maximum sequence length of 288, corresponding to one window every 5 minutes within 24 hours. The final data is streamed through a message queue. The topic name is defined as esim_behavior_stream, the key is the entity ID, and the value is the time window feature vector. The overall system latency is controlled within 10 seconds.

[0045] In this embodiment, the region calculation module is used to: perform effective activation calculation of regions based on the code behavior data stream, and construct activation curves for different regions, specifically including: Business areas are identified and classified based on code behavior data streams, and data streams from different areas are extracted. The activation behavior of the data stream in different regions is calculated one by one to obtain activation behavior data for multiple regions; Perform abnormal behavior detection on the data stream and mark invalid behaviors; Based on invalid behaviors, the activation behavior data is used to calculate the effective activation volume and count the effective activations. The effective activation amount is continuously time-series fitted to construct activation amount curves for different regions.

[0046] In this embodiment, service areas are identified and divided based on location and network attribute information in the code number behavior data stream. The area definitions originate from operator network planning data, including base station coverage area codes and administrative area codes. Base station areas are obtained through cell identifier mapping, while administrative areas are obtained through number location or user registration information. For each behavior data entry, location feature fields are extracted, including cell identifier, location area code, and network access identifier. Area attribution is determined according to priority, prioritizing real-time access base station information, followed by historical location information. To avoid area jitter caused by frequent handovers, area stability constraints are introduced: if the number of area changes exceeds 3 within a 300-second time window, the area with the highest frequency is used as the area attribution for that window. The area classification granularity is set to a three-level structure, including region, city, and base station cluster, with a unified integer coding rule. After completing the area division, code number activation behavior is statistically analyzed for each area's data stream. Activation behavior is defined as the first activation event, i.e., the first activation record in the code number's lifecycle. The identification method involves sorting each entity ID by time, filtering for the earliest activation event, and ensuring that no prior activation records exist. A sliding time window is used in the statistical process, with a window length of 3600 seconds and a step size of 300 seconds, consistent with the preceding data structure. The number of activation actions is counted within each window to obtain a regional-level activation action time series. To ensure statistical accuracy, a deduplication constraint is introduced, meaning that the same entity ID is counted only once throughout its entire lifecycle. Further activation action features are extracted, including activation time distribution, activation density, and the percentage of activations within a window. Activation density is defined as the number of activations per unit time, and the activation percentage is the ratio of the number of activations to the total number of actions.

[0047] Anomaly detection is performed on behaviors in the regional data stream to identify invalid or abnormal operations. Abnormal behavior is defined as actions that do not conform to normal business logic or statistical distribution, including high-frequency activation, abnormal timing jumps, and abnormal latency. The detection method combines rules and statistical models. First, rule thresholds are set; for example, if the same entity ID is activated more than 3 times within 1 hour, it is marked as abnormal, and an interval of less than 10 seconds between activation and deactivation is also considered abnormal. Second, detection is based on statistical distribution. The mean and standard deviation of features such as behavior frequency, time interval, and latency are calculated, and the three-standard-deviation principle is adopted. When data deviates from the mean by more than three times the standard deviation, it is judged as abnormal. Furthermore, a local anomaly factor method is introduced to perform density analysis on multi-dimensional features and identify low-density points as abnormal samples. All behaviors judged as abnormal are uniformly marked as invalid behaviors, with an invalid_flag value of 1, and the rest are set to 0.

[0048] A valid activation is defined as an activation behavior that is not marked as invalid, i.e., an activation record with an invalid_flag of 0 and that meets the lifecycle logical constraints. During the calculation process, the data stream for each region is filtered, removing all entity ID activation records corresponding to invalid behaviors, and the number of valid activations within the window is recounted. To ensure statistical consistency, a valid activation must meet two conditions: it must persist for at least 60 seconds after activation and must not be deactivated within a short period. The statistical method still uses a sliding window mechanism with a window length of 3600 seconds and a step size of 300 seconds, calculating the number of valid activations within each window. Simultaneously, an efficiency index is calculated, i.e., the ratio of valid activations to the original activations, to measure data quality.

[0049] The time series data undergoes preprocessing, including missing value imputation and outlier smoothing. Missing values ​​are imputed using linear interpolation, and outliers are replaced using local means. Subsequently, a time series fitting method is used for modeling. This embodiment employs a combination of weighted moving average and polynomial fitting, with the moving average window set to 5 time steps and the polynomial order set to 3 to capture trend changes. To enhance fitting stability, a regularization parameter is introduced to prevent overfitting. During the fitting process, time is used as the independent variable and the effective activation level as the dependent variable, outputting an activation curve in the form of a continuous function. Independent modeling is performed for each region, resulting in multiple regional activation curves that reflect the business development trends of different regions.

[0050] In a specific embodiment, a time window of Δt = 300 seconds is set to process the code number behavior data stream. Each data entry contains an entity ID, a timestamp, a base station identifier, and a behavior type. Area identification uses a base station mapping rule to map the base station identifier to an area code; for example, base station 101 maps to area A, and base station 202 maps to area B. For the behavior data collected within a certain time window, the data stream for area A contains 10 records, of which 5 are activation behaviors. An activation behavior is defined as the first activation record of a certain entity ID. The 5 activation records in area A are sorted by time. If two of them are duplicate activations of the same entity ID, only the first record is retained. Therefore, the number of activation behaviors is calculated as follows:

[0051] ;

[0052] Similarly, region B was calculated to have 4 activations. The final set of activation behavior data for each region was obtained; this data is entirely derived from the statistical results of the original behavior flow, without any implicit inferences.

[0053] Anomaly detection is performed on the regional data stream, with anomaly rule parameters set as T_1 = 60 seconds and T_2 = 3 activations per window. If the activation duration is less than T_1, it is considered an anomaly; if the same entity ID is activated more than T_2 within a time window, it is considered a frequency anomaly. Taking region A as an example, among the three activation behaviors, one has a duration of 30 seconds, which is less than the threshold T_1, and another has 4 activations, which is greater than the threshold T_2. Therefore, the number of abnormal behaviors is 2. The formula for calculating the effective activation quantity is: ;

[0054] In region B, if the activation behavior is 4 and the abnormal behavior is 1, then: ;

[0055] All abnormal behaviors are represented by the flag variable `invalid_flag`, with a value of 1 indicating invalidity and 0 indicating validity. The final output is the statistical results of the effective activation volume for each region.

[0056] The effective activation values ​​for each region are fitted using a time series. Let the effective activation values ​​for region A over four consecutive time windows be Y={1,2,3,5}, corresponding to the time series t={1,2,3,4}. A linear fitting model is used: ;

[0057] Parameters a and b are calculated using the least squares method, with the following formula:

[0058] Calculate each item:

[0059] Substitute into the calculation:

[0060] The final activation curve for region A is as follows: .

[0061] In this embodiment, the specific steps for calculating the activation frequency, terminal binding relationship and communication usage based on the data stream of different regions to obtain the user's usage trajectory are as follows: Set invalid service rules, including short-term batch activation, frequent terminal changes, and long-term number occupation without data traffic; Based on the invalid business rules, abnormal behavior detection is performed on the user's usage trajectory, and invalid behavior is marked. Invalid behaviors are risk-marked and classified for management, and behavioral information is collected, uploaded, and stored for management.

[0062] In this embodiment, invalid service rules are set, including short-term batch activation, frequent terminal changes, and long-term number occupation without data traffic. Based on the invalid business rules, abnormal behavior detection is performed on the user's usage trajectory, and invalid behavior is marked. Invalid behaviors are risk-marked and classified for management, and behavioral information is collected, uploaded, and stored for management.

[0063] User usage trajectories are defined as time-ordered sequences of behaviors, including multi-dimensional information such as activation, device changes, and data usage. The detection process employs a combination of sliding window and status statistics. A time-series window is constructed for each entity ID. For example, short-term batch activation detection uses a 300-second sliding window to calculate activation counts, with a window step size of 60 seconds; frequent device changes use a 24-hour cumulative statistical window; and long-term account occupancy with no data usage uses a 7-day observation window. For each trajectory data point, key features are extracted, including activation frequency, number of device changes, cumulative data usage, and behavior interval time, and compared item by item with rule thresholds. When any rule is triggered, the behavior is marked as invalid, and the triggering rule number is recorded. To improve detection stability, a minimum support count constraint is introduced, requiring the same rule to be triggered at least twice within a cycle before an anomaly is confirmed, to avoid the influence of occasional noise. All marking results are written to the `invalid_flag` data field, with a value of 1 indicating invalid behavior and 0 indicating normal behavior. The `invalid_type` field is also recorded to identify the anomaly type.

[0064] A risk scoring model is defined, mapping different types of abnormal behavior to risk scores ranging from 0 to 100. Short-term batch activations have a risk weight of 0.5, frequent terminal changes have a weight of 0.3, and long-term account occupancy without traffic has a weight of 0.2. The risk score for a single behavior is calculated as the sum of the products of each rule's trigger result and its corresponding weight, amplified by trigger frequency; for example, each additional trigger of the same rule increases the risk score by 5 points. Risk levels are categorized based on these scores, with thresholds of 0-30 for low risk, 30-70 for medium risk, and greater than 70 for high risk. For each entity ID, risk scores and trigger counts are aggregated within the statistical period to form a risk profile, including the total number of abnormal behaviors, main abnormal types, and risk levels. Statistical information includes indicators such as time distribution, regional distribution, and the proportion of behavior types; all indicators are directly calculated from labeled data. Finally, the risk data and behavioral statistics are written to a centralized storage system, managed by date and risk level partitions, and updated in real-time via a message queue.

[0065] In this embodiment, the demand analysis module is used to: calculate the multi-point activation growth rate of the activation curves in different regions, estimate the activation demand for code numbers, and output the activation demand values ​​for different regions, specifically including: Multi-point activation growth rate calculation was performed on the activation curves of different regions to obtain the growth rate at multiple time points. Identify the inflection points of phased growth based on the growth rate; Based on the aforementioned stage-specific growth inflection point, a growth trend prediction is made to obtain the activation volume prediction trend. Based on the activation volume prediction trend, the activation demand for code numbers is estimated, and the activation demand value for different regions is output.

[0066] In this embodiment, the activation curves constructed for each region are discretized to form time-series data, with a time granularity of 300 seconds, corresponding to a sampling point every 5 minutes. The activation quantity is defined as the number of effective activations within this time window. First, the curves are smoothed using a moving average window of length 5 to eliminate short-term fluctuations. Then, the growth rate is calculated, defined as the ratio of the difference in activation quantity between adjacent time points to the time interval, i.e., the change per unit time, with the time interval fixed at 300 seconds. To improve stability, a multi-point differencing method is introduced, simultaneously calculating first-order and three-point differencing at each time point. The first-order differencing reflects instantaneous changes, while the three-point differencing calculates the average rate of change using one point before and after the first point, reducing noise impact. The final growth rate is a weighted average of the two methods, with weights set to 0.6 and 0.4. All calculations are based on explicit time-series data with no hidden variables. The output is a growth rate sequence for each region at each time point, identifying inflection points in the activation quantity change process based on the growth rate sequence. An inflection point is defined as the location where the growth rate changes significantly, including the point where growth turns into decline or slow growth turns into rapid growth. The identification method combines threshold judgment with rate of change analysis. First, the change in growth rate is calculated, i.e., the difference in growth rates between adjacent time points, and their mean and standard deviation are calculated. A threshold is set as the mean plus or minus two standard deviations; when the change exceeds this range, it is considered a candidate inflection point. Second, a persistence constraint is introduced, meaning the trend must remain consistent over three consecutive time points to avoid misjudging single-point fluctuations. To further improve accuracy, candidate inflection points are validated a second time by calculating the average difference in growth rate within a 1800-second window before and after the point; when the difference exceeds 20%, the point is confirmed as a valid inflection point. All parameters are derived from historical curve statistical analysis; for example, the standard deviation range and the 20% difference threshold are determined based on the sample data distribution.

[0067] The activation curve is divided into multiple stages based on inflection points, with a relatively stable growth trend assumed within each stage. Trend fitting is performed for each stage, comparing linear and quadratic polynomial fitting methods. The optimal model is selected based on the fitting error, with mean squared error as the error metric. During the fitting process, time is used as the independent variable and activation amount as the dependent variable, with all data derived from historical observations. Extrapolation predictions are then performed for future time periods, with a prediction period set at 24 hours, or 288 time points. To ensure prediction continuity, smoothing is applied at stage boundaries using a weighted transition method to ensure continuity between adjacent stage predictions. Furthermore, a trend correction factor is introduced, calculated based on the average actual growth rate over the most recent 6 hours, to dynamically adjust the prediction slope.

[0068] The activation demand for future eSIM numbers is quantitatively estimated based on the predicted activation trend. Activation demand is defined as the total number of new activations and peak capacity required to support the system within the prediction period. First, the prediction curve is integrated to obtain the cumulative activation volume for the next 24 hours, serving as the base demand value. Second, peak points in the prediction curve are identified, and the maximum activation volume per unit time is calculated to assess the system's instantaneous carrying capacity requirement. To improve estimation accuracy, a safety redundancy coefficient of 1.2 is introduced to handle sudden growth; this parameter is determined based on historical maximum deviation statistics. The final demand value is calculated by multiplying the cumulative activation volume by the redundancy coefficient, while recording the peak demand as a capacity reference. The output includes the total demand, peak demand, and time distribution information for each region. All calculations are directly derived from the prediction curve without implicit model parameters. This result can be used for resource allocation and network scheduling, enabling intelligent management of the entire eSIM number lifecycle.

[0069] In this embodiment, the planning module is used to: perform dynamic planning and intelligent allocation based on the activation demand value, and output a code number delivery strategy, specifically including: Identify the regional code inventory pool; Based on the regional code inventory pool, the number of allocable codes, the number of occupied codes, and the number of recyclable codes in different regions are determined to obtain the real-time balance of each region. Based on the activation demand value and the real-time remaining capacity in each region, dynamic planning and intelligent allocation are performed to output the code number delivery strategy.

[0070] In this embodiment, a regional-level eSIM number inventory pool model is constructed to describe the manageable eSIM number resource set in different regions. The number inventory pool is defined as the set of numbers generated in the system but not yet fully consumed. Its basic data comes from the number management system and configuration distribution platform, including the number segment allocation table and inventory status table. The inventory pool is divided by region, with the region identifier consistent with the aforementioned region division rules, derived from the base station affiliation and administrative region code mapping table. For each number, its status field is extracted, including unallocated, allocated but inactive, activated, and deactivated, and categorized based on its lifecycle status. Status mapping rules are set during the identification process; for example, unallocated numbers are directly added to the inventory pool, while deactivated numbers are re-added when recycling conditions are met. To ensure data consistency, the inventory data uses a minute-level refresh mechanism, synchronizing status data from the core database every 60 seconds. The inventory pool structure is defined to include the region identifier, total number of numbers, number of each status, and update timestamp. A data verification mechanism is further introduced; when the total inventory does not match the sum of the distributions of each status, a verification process is triggered to ensure data accuracy.

[0071] Detailed calculations are performed on the regional inventory pool to obtain the quantity and real-time availability of various resources. First, the number of allocable codes is defined as the number of codes currently unallocated and ready for immediate distribution; this data is directly derived from the unallocated field in the inventory status table. The number of occupied codes is defined as the total number of allocated but not yet deactivated codes, including activated codes and deactivated codes bound to users. The number of reclaimable codes is defined as deactivated or long-term inactive codes that meet the reclamation criteria, which are set to be codes that have been deactivated for more than 72 hours or have had no traffic usage for more than 7 days; these thresholds are determined based on historical reclamation strategy statistics. Real-time availability is defined as the sum of the allocable and reclaimable quantities minus the reserved buffer. The buffer is used to cope with sudden demand and is set to 20% of the average activation volume of the current region over the past 24 hours. All quantity calculations are based on statistics from explicitly defined fields in the inventory pool and do not involve implicit inferences. To ensure real-time performance, all indicators are updated every minute and timestamps are recorded for subsequent scheduling synchronization.

[0072] This method combines predicted regional activation demand with real-time inventory reserves to dynamically allocate and deploy code resources. An optimization objective function is constructed to minimize resource shortage risk and inventory waste, specifically defined as the sum of the absolute values ​​of the differences between demand and allocated quantities for each region. Input parameters include the activation demand value, real-time inventory reserves, and maximum total available resources for each region, derived from the total network inventory. A dynamic programming approach is used to solve the problem, decomposing the total resource allocation problem into multiple regional sub-problems and progressively calculating the optimal allocation scheme. Constraints include ensuring that the allocated quantity for each region does not exceed its maximum carrying capacity, defined as the historical peak activation quantity multiplied by an expansion coefficient of 1.2, while guaranteeing that the total allocated quantity for the entire network does not exceed the total inventory. To improve allocation flexibility, priority weights are introduced, calculated based on the region's historical growth rate and current demand urgency, ranging from 0 to 1. The final allocated quantity is adjusted proportionally according to the weights. The output is a code deployment strategy, including the allocation quantity for each region, the deployment time window, and priority indicators.

[0073] The objective function expression is optimized as follows: ;

[0074] Where i: region index number, used to identify the i-th region; N: Total number of regions, representing the number of regions participating in allocation optimization; : Activation demand value (predicted demand) for the i-th region; : The actual number of code numbers allocated to the i-th region (decision variable); : The absolute deviation between the demand and allocation in the i-th region, used to measure the matching error; The total system error is calculated by summing the deviations from all regions. The goal is to find the optimal solution that minimizes the total deviation among all feasible allocation schemes. In a specific embodiment, the system is divided into three regions, with the activation requirements for each region being... The real-time inventory balances are respectively The maximum load-bearing capacities are respectively The total available inventory across the entire network is S=180, and the priority weights for each region are as follows: .

[0075] Based on the above parameters, construct the optimization objective function:

[0076] in This represents the actual allocation quantity for the i-th region, while satisfying the total inventory constraint. Regional carrying capacity constraints and nonnegativity constraints Solve under the given conditions.

[0077] First, determine the available basic quantity by combining real-time margin and demand. ,get Further prioritization weights are introduced to adjust the requirements, resulting in weighted requirements. The total weight requirement is 171. Based on this, the entire network inventory of 180 is allocated proportionally to obtain the initial solution. .

[0078] After load-bearing capacity verification, none of the regions exceeded their upper limits, therefore the allocation results remained unchanged. Substituting these values ​​into the objective function, the total system deviation was calculated as follows: This indicates that the allocation scheme achieves a better balance between inventory constraints and demand constraints.

[0079] The final code number delivery strategy is as follows: 114 codes are allocated to region 1 and delivered immediately; 51 codes are allocated to region 2 and delivered after a 5-minute delay; and 15 codes are allocated to region 3 and delivered after a 10-minute delay. At the same time, the priority execution order is determined according to the weight, thereby achieving the optimal allocation of multi-region code number resources based on dynamic programming.

[0080] In this embodiment, the specific steps for dynamically planning and intelligently allocating based on the activation demand value and the real-time remaining capacity of each region, and outputting the code number delivery strategy are as follows: Based on the activation demand value and the real-time remaining resources in each region, the supply and demand relationship is calculated to identify resource gap areas; Calculate the gap value of the resource gap region; Based on the gap values, a hierarchical sorting is performed to obtain an allocation priority sequence; Based on the real-time remaining capacity of each region, fine-grained segmentation is performed to obtain the availability and release cycle of different code numbers; Construct a multi-level resource pool based on the availability and release cycle; Based on the allocation priority sequence and multi-level resource pools, a layer-by-layer allocation plan is performed, and a code number delivery strategy is output.

[0081] The hierarchical allocation plan is as follows: priority is given to using immediately allocable codes in the local area to meet the demand. When local resources are insufficient, short-term recyclable codes and cross-regional transfer codes are used in sequence to supplement the resources, forming a phased resource replenishment process. In this embodiment, a supply-demand comparison model is established for each region. The input parameters include the predicted activation demand value and the real-time reserve. The activation demand value is defined as the number of activations expected to be completed in the region within the next 24 hours, derived from previous time series forecasts. The real-time reserve is defined as the sum of the current number of allocable codes and the number of reclaimable codes minus the buffer amount, which is set to 20% of the average activation amount over the past 24 hours. The supply-demand relationship is calculated through the difference, i.e., the supply-demand difference equals the real-time reserve minus the activation demand value. When the supply-demand difference is less than 0, the region is determined to be a resource shortage region. To avoid misjudgments caused by short-term fluctuations, a stability constraint is introduced, i.e., a shortage status is confirmed only when three consecutive time windows show negative differences. The time window length is set to 300 seconds. The supply-demand ratio index is further calculated, i.e., the real-time reserve divided by the demand value, to measure the degree of tension. When this ratio is less than 0.8, it is marked as a high-risk shortage region. All parameters are derived from historical operational statistics; for example, the 0.8 threshold is determined based on the distribution of historical resource shortage periods.

[0082] The gap value is defined as the absolute value of the difference between the demand value and the real-time margin. That is, the gap value equals the activated demand value minus the real-time margin; a positive value indicates the size of the gap. To improve accuracy, a time-weighted mechanism is introduced, calculating demand in segments for different future time periods. For example, the 24-hour forecast period is divided into 8 sub-intervals, each 3 hours long, and the gap value for each interval is calculated. The gaps are then weighted and summed according to time weights, with weights set based on time urgency; the closer to the current time, the higher the weight. For example, the weight for the first 3 hours is 0.3, the weight for the middle stage is 0.2, and the weight for the later stage is 0.1. The final gap value is the sum of the gaps in each interval multiplied by their respective weights. Furthermore, a volatility correction factor is introduced, calculated based on the standard deviation of demand changes over the most recent 6 hours. When volatility is high, the gap value is appropriately amplified, with the amplification factor set to 1 plus the ratio of the standard deviation to the mean. All calculations are based on explicit data fields and statistical indicators, with no implicit inferences. The output is the quantified gap value for each gap region.

[0083] The gap values ​​are normalized, mapping all regional gap values ​​to the 0-1 range for easier comparison. A comprehensive scoring mechanism is then introduced, with the score composed of three parts: gap value, supply-demand ratio, and historical growth rate, with weights set to 0.5, 0.3, and 0.2 respectively. The historical growth rate, derived from previous growth analysis results, reflects regional development trends. After the comprehensive score is calculated, all regions are sorted from highest to lowest score to obtain a priority sequence. To avoid frequent ranking fluctuations, a ranking stabilization mechanism is introduced, maintaining the original order when the change between two consecutive ranking results is less than 5%. Further priority grading rules are established, dividing regions into high-priority, medium-priority, and low-priority categories, corresponding to score ranges greater than 0.7, 0.4 to 0.7, and less than 0.4 respectively.

[0084] For each code number in the inventory, extract the status field, including unassigned, assigned but inactive, activated, and deactivated, and calculate its usage period based on historical behavior data. Availability is defined as whether the code number is immediately available for allocation; if the status is unassigned, availability is 1; otherwise, it is 0. The release period is defined as the time required for a code number to transition from its current status to an available status. For example, deactivated codes require a 72-hour recovery period, while activated codes require a predicted release time based on user behavior. This prediction is calculated based on historical average lifecycle data, with the average value set at 30 days. Calculate the release period for all codes and categorize them into short-term and long-term resources based on period length: short-term is defined as less than 24 hours, and long-term as greater than 24 hours.

[0085] Fine-grained resources are divided into multi-level resource pools based on availability and release cycle. The resource pools are structured in three levels: Level 1 resource pools are immediately available resources (unassigned codes with an availability of 1); Level 2 resource pools are short-term reclaimable resources (codes with a release cycle of less than 24 hours); and Level 3 resource pools are cross-regional transfer resources (available resources in other regions). Each resource pool is established on a regional basis, recording the resource quantity, average release time, and update time. To ensure dynamic updates of the resource pools, a refresh cycle of 60 seconds is set, consistent with the inventory synchronization mechanism. A resource pool priority rule is further introduced, with Level 1 resource pools having the highest priority, followed by Level 2 and Level 3. All resource pool data originates from the splitting results in step four, without introducing any implicit calculations.

[0086] Resource allocation is performed for high-priority areas, prioritizing the use of available codes from the area's primary resource pool to meet demand. When primary resources are insufficient, short-lived recyclable codes from the secondary resource pool are used to supplement them. The supplementary amount is calculated based on the gap value and release cycle, prioritizing resources with the shortest release cycle. If local resources are still insufficient, a cross-regional allocation mechanism is initiated to allocate resources from the tertiary resource pools of other areas. The allocation amount is limited to 80% of the remaining resources in the target area to avoid impacting its own needs. The entire allocation process is executed iteratively region by region, updating the resource pool status at each step. The final output code distribution strategy includes the allocation quantity for each area, the resource source level, and the distribution time window.

[0087] In this embodiment, the strategy adjustment module is used to: execute code number distribution and allocation operations for different regions based on the code number distribution strategy, and collect the actual activation volume of each region after a period of time; calculate the prediction accuracy of different regions based on the actual activation volume of each region, and then perform distribution strategy parameter adjustment processing, specifically including: The code number distribution strategy is used to perform code number distribution and allocation operations in different regions, and the actual activation volume in each region is collected after a period of time. Based on the actual activation volume in each region, the deviation of the activation demand value is calculated to obtain the predicted deviation value. Statistical tracking of prediction deviations is performed to generate time-series deviations. Prediction accuracy is calculated based on time series deviation values ​​to obtain prediction accuracy for different regions; Based on the prediction accuracy, the code number delivery strategy parameters are adjusted to output an intelligent delivery management strategy. The adjustment of delivery strategy parameters includes adjustments to delivery pace, allocation ratio, and scheduling priority.

[0088] In this embodiment, the deployment strategy includes parameters such as the number of codes allocated to a region, the resource source level, and the deployment time window. The deployment time window is defined as the time period during which codes are released to the market, with a granularity of 300 seconds. During execution, a specified number of codes are transferred from the inventory pool to an allocable state through the code management system interface and bound with regional tags. The deployment rhythm is defined as the number of deployments per unit time, defaulted to once per hour, and can be adjusted to once every 30 minutes based on regional priority. After deployment is completed, the actual data collection phase begins, with a collection period set to 24 hours to observe the actual activation status after deployment. The actual activation volume is defined as the number of codes activated within this period, and the data comes from the activation event statistics in the behavioral data stream. To ensure data accuracy, a delay compensation mechanism is introduced, i.e., an additional 10 minutes are added after the collection is completed to capture delayed reported data.

[0089] The prediction deviation is calculated by comparing the activation demand value obtained in the prediction phase with the actual activation volume. The prediction deviation is defined as the actual activation volume minus the predicted demand value; a positive result indicates underprediction, while a negative result indicates overprediction. To facilitate cross-regional comparisons, a relative deviation index is introduced, which is the deviation value divided by the predicted demand value, yielding a proportional form of deviation. The calculation is performed region by region, aligned with a time window of 24 hours, consistent with the data collection period. To improve analysis accuracy, the deviation is divided into multiple sub-time periods for calculation; for example, eight sub-intervals are divided every three hours, and local deviation values ​​are calculated separately for each. All calculations are based on explicit data fields, meaning the predicted value is obtained by directly differencing the actual value, without any implicit models.

[0090] The calculated deviation values ​​are stored chronologically to form a deviation time series, with the time granularity set to once per day. Statistical analysis is then performed on the deviation series to calculate indicators such as the mean, standard deviation, and maximum deviation, used to characterize the deviation distribution. To enhance stability, a sliding window statistical method is introduced, with a window length set to 7 days, to smooth the deviation values ​​within the window and eliminate abnormal daily fluctuations. Further, the deviation trend, i.e., the rate of change of deviation between adjacent time points, is calculated to identify the direction of change in predictive performance.

[0091] The mean absolute percentage error (MASE) is used as the basic indicator, calculated by taking the average of the absolute value of the deviation at each time point and the actual value. For ease of understanding, accuracy is defined as 1 minus this error value, with the result ranging from 0 to 1; a higher value indicates more accurate predictions. The calculation is performed separately for each region, using a 7-day sliding window for averaging to obtain a stable accuracy indicator. Further accuracy grading standards are established: greater than 0.9 is high accuracy, 0.7 to 0.9 is medium accuracy, and less than 0.7 is low accuracy. For low-accuracy regions, targeted adjustments are made. When accuracy is below 0.7, the delivery frequency is shortened, for example, from once per hour to once every 30 minutes, to improve response speed; simultaneously, the amount delivered per delivery is reduced to mitigate the risk from prediction errors. For medium-accuracy regions, the current frequency is maintained with minor adjustments to the allocation ratio, set at ±10%. For high-accuracy regions, the delivery volume can be appropriately increased and the delivery interval extended to improve resource utilization efficiency. Scheduling priority is determined by both accuracy and the urgency of the demand; regions with low accuracy and high demand receive increased priority.

[0092] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0093] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein are implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. An AI-based eSIM number lifecycle management system, characterized in that, It includes a data acquisition module, a regional calculation module, a demand analysis module, a planning module, and a strategy adjustment module. The acquisition module collects all behavioral data of the eSIM number and outputs a number behavior data stream; The region calculation module performs effective activation calculation of the region based on the code behavior data stream and constructs activation curves for different regions. The demand analysis module calculates the multi-point activation growth rate of the activation curves in different regions, estimates the activation demand for code numbers, and outputs the activation demand values ​​for different regions. The planning module performs dynamic planning and intelligent allocation based on the activation demand value, and outputs a code number delivery strategy. The strategy adjustment module performs code number distribution operations in different regions based on the code number distribution strategy, and collects the actual activation volume of each region after a period of time; it calculates the prediction accuracy of different regions based on the actual activation volume of each region, and then performs distribution strategy parameter adjustment processing. Specifically, the step of dynamically planning and intelligently allocating based on the activation demand value to output a code number delivery strategy includes: Identify the regional code inventory pool; Based on the regional code inventory pool, the number of allocable codes, the number of occupied codes, and the number of recyclable codes in different regions are determined to obtain the real-time balance of each region. Based on the activation demand value and the real-time remaining capacity of each region, dynamic planning and intelligent allocation are performed to output the code number delivery strategy. The specific steps for dynamically planning and intelligently allocating numbers based on the activation demand value and the real-time remaining capacity in each region, and outputting the number delivery strategy, are as follows: Based on the activation demand value and the real-time remaining resources in each region, the supply and demand relationship is calculated to identify resource gap areas; Calculate the gap value of the resource gap region; Based on the gap values, a hierarchical sorting is performed to obtain an allocation priority sequence; Based on the real-time remaining capacity of each region, fine-grained segmentation is performed to obtain the availability and release cycle of different code numbers; Construct a multi-level resource pool based on the aforementioned availability and release cycle; Based on the allocation priority sequence and multi-level resource pools, a layer-by-layer allocation plan is performed, and a code number delivery strategy is output.

2. The AI-based eSIM number lifecycle management system according to claim 1, characterized in that, The process of collecting full behavioral data of the eSIM number and outputting a number behavior data stream specifically includes: Collect all behavioral data of the eSIM number; the all behavioral data includes activation, suspension, number retention, and account cancellation data; The full set of behavioral data is subjected to denoising, filtering, and outlier removal to obtain standardized data; The standardized data is aligned with statistical timestamps and divided into sliding time windows to output a code-behavior data stream.

3. The AI-based eSIM number lifecycle management system according to claim 2, characterized in that, The step of calculating effective activation of regions based on code behavior data stream and constructing activation curves for different regions specifically includes: Business areas are identified and classified based on code behavior data streams, and data streams from different areas are extracted. The code activation behavior of data streams in different regions is calculated one by one to obtain activation behavior data for multiple regions; Perform abnormal behavior detection on the data stream and mark invalid behaviors; Based on invalid behaviors, the activation behavior data is used to calculate the effective activation volume and count the effective activations. The effective activation amount is continuously time-series fitted to construct activation amount curves for different regions.

4. The AI-based eSIM number lifecycle management system according to claim 3, characterized in that, The specific steps for detecting abnormal behavior and marking invalid behavior in the data stream are as follows: Set invalid service rules, including short-term batch activation, frequent terminal changes, and long-term number occupation without data traffic; The data stream is subjected to abnormal behavior detection based on the invalid business rules, and invalid behavior is marked. Invalid behaviors are flagged and classified for risk, and behavioral information is collected, uploaded, and stored for management.

5. The AI-based eSIM number lifecycle management system according to claim 4, characterized in that, The calculation of the multi-point activation growth rate of the activation curves in different regions, the estimation of code activation demand, and the output of activation demand values ​​for different regions specifically include: Multi-point activation growth rate calculation was performed on the activation curves of different regions to obtain the growth rate at multiple time points. Identify the inflection points of phased growth based on the growth rate; Based on the aforementioned stage-specific growth inflection point, a growth trend prediction is made to obtain the activation volume prediction trend. Based on the activation volume prediction trend, the activation demand for code numbers is estimated, and the activation demand value for different regions is output.

6. The AI-based eSIM number lifecycle management system according to claim 5, characterized in that, The method involves executing code distribution and allocation operations for different regions based on the code distribution strategy, and collecting the actual activation volume of each region after a certain period. Based on the actual activation volume of each region, the prediction accuracy of different regions is calculated, and then the distribution strategy parameters are adjusted. Specifically, this includes: Based on the code number distribution strategy, code number distribution operations are performed in different regions, and the actual activation volume of each region is collected after a period of time. Based on the actual activation volume in each region, the deviation of the activation demand value is calculated to obtain the predicted deviation value. Statistical tracking of prediction deviation values ​​is performed to generate time-series deviation values; Prediction accuracy is calculated based on time series deviation values ​​to obtain prediction accuracy for different regions; Based on the prediction accuracy, the code number delivery strategy parameters are adjusted to output an intelligent delivery management strategy.

7. The AI-based eSIM number lifecycle management system according to claim 6, characterized in that, The adjustment of the delivery strategy parameters includes adjustments to the delivery pace, allocation ratio, and scheduling priority.

Images