An internet of things card management platform

Abstract

The present application relates to the technical field of Internet of Things, and in particular to an Internet of Things card management platform, comprising: a data calculation module configured to generate a plurality of device network quality data according to real-time network parameters corresponding to a plurality of operator networks respectively; a map generation module configured to gather the plurality of device network quality data and generate a regional network quality map; a strategy management module configured to perform network state prediction and global optimization analysis based on the regional network quality map and generate a target switching strategy; and an output execution module configured to issue the target switching strategy to a target terminal, so that the target terminal performs a switching operation of an operator network based on the target switching strategy, thereby effectively solving the problem that the existing Internet of Things card management cannot guarantee the connection continuity and service quality during the cross-network movement of a device.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) technology, and more specifically to an IoT card management platform. Background Technology

[0002] IoT SIM card management platforms play a crucial role in enabling the connection of massive numbers of IoT devices. Through intelligent connection management, they enhance the stability and resource utilization efficiency of IoT systems.

[0003] Existing IoT SIM card management mainly relies on preset fixed rules or simple threshold judgments based on a single network parameter to achieve operator network switching. For example, network switching operations are triggered based on the signal strength threshold received by the device, or a manually configured priority list is used for network switching. However, because it is impossible to dynamically perceive and globally optimize based on the real-time differences in the status of multiple operator networks, the switching decision is lagging or inaccurate when the network environment changes, making it difficult to guarantee the continuity of connection and quality of service when the device moves across networks. Summary of the Invention

[0004] To address the technical problem that existing IoT SIM card management systems cannot guarantee connection continuity and quality of service during device movement across networks, this application provides an IoT SIM card management platform.

[0005] The IoT card management platform provided in this application adopts the following technical solution:

[0006] An IoT card management platform, comprising:

[0007] The data calculation module is used to generate network quality data for multiple devices based on the real-time network parameters corresponding to multiple operator networks.

[0008] The map generation module is used to aggregate network quality data from multiple devices and generate a regional network quality map.

[0009] The strategy management module is used to predict network status and perform global optimization analysis based on the regional network quality map, and generate target switching strategies.

[0010] The output execution module is used to send the target handover policy to the target terminal so that the target terminal can perform the handover operation of the operator network based on the target handover policy.

[0011] Furthermore, real-time network parameters are collected through a data acquisition terminal. The steps for generating network quality data for multiple devices based on the real-time network parameters corresponding to multiple operator networks include:

[0012] Multi-dimensional feature extraction is performed on real-time network parameters to obtain feature vectors;

[0013] The feature vectors are processed by a pre-set machine learning model to predict their quality, resulting in a predicted quality score.

[0014] Based on the device context information of the acquisition terminal corresponding to the real-time network parameters, the predicted quality score is calibrated to obtain device network quality data.

[0015] Furthermore, the steps for calibrating the predicted quality score and obtaining device network quality data based on the device context information of the acquisition terminal corresponding to the real-time network parameters include:

[0016] Context type recognition is performed based on device context information. After obtaining the context type label, features are extracted from the preset quantization model corresponding to the context type label to obtain the context feature vector.

[0017] The calibration weights are obtained by calculating the weights of the context feature vectors using a pre-defined calibration model.

[0018] Based on the calibration weights, the predicted quality scores are nonlinearly fused to obtain the device network quality data.

[0019] Furthermore, the steps for aggregating network quality data from multiple devices to generate a regional network quality map include:

[0020] Based on the acquisition device and acquisition timestamp of each device's network quality data, spatiotemporal clustering is performed on the network quality data of each device to obtain quality clustering regions;

[0021] Anomaly filtering is performed on the quality clustering regions to obtain quality data;

[0022] A regional network quality map is obtained by performing map rendering processing on the quality data using a preset weighted fusion algorithm.

[0023] Furthermore, based on the acquisition device and acquisition timestamp to which each device's network quality data belongs, the steps for performing spatiotemporal clustering processing on the network quality data of each device to obtain quality clustering regions include:

[0024] The spatial location of each acquisition device at the corresponding acquisition timestamp is obtained. The spatial location, acquisition timestamp and network quality data of each acquisition device are encoded to obtain the encoding features corresponding to each acquisition device.

[0025] Calculate the distance of each encoded feature in the feature space, and obtain the target acquisition devices to which the two encoded features belong (the distance difference is less than the preset distance). Take one of the target acquisition devices as the center point and the distance between the two target acquisition devices as the radius to determine the quality clustering region.

[0026] Furthermore, the steps for obtaining a regional network quality map by performing map rendering processing on the quality data using a preset weighted fusion algorithm include:

[0027] Quality indicators are extracted from the quality data to obtain multiple sets of quality indicators;

[0028] After optimizing the quality indicator set to obtain the quality fusion weight, the quality data is rasterized and rendered according to the quality fusion weight to obtain the regional network instruction map.

[0029] Furthermore, the steps for generating target switching strategies based on regional network quality maps to predict network status and perform global optimization analysis include:

[0030] Spatiotemporal sequence prediction is performed based on regional network quality maps to obtain the predicted network state.

[0031] A multi-objective optimization algorithm is used to perform global optimization analysis on the predicted network state. After obtaining the optimization target weights, the optimization target weights are substituted into the preset policy synthesis model for policy generation processing to obtain the target switching policy.

[0032] Furthermore, the steps of substituting the optimized target weights into the preset policy synthesis model for policy generation to obtain the target switching policy include:

[0033] Based on the optimization objective weights, network constraints are inferred to obtain the policy constraint set;

[0034] Based on each policy constraint in the policy constraint set, output the candidate policies that match each policy constraint respectively.

[0035] Based on the current network parameters when the target device accesses the operator network corresponding to each candidate strategy, the network connection quality of each candidate strategy is calculated, and the candidate strategy corresponding to the network connection quality with the first-ranked value is determined as the target handover strategy.

[0036] Beneficial effects achieved:

[0037] This application provides an IoT SIM card management platform, comprising: a data calculation module for generating network quality data for multiple devices based on real-time network parameters corresponding to multiple operator networks; a map generation module for aggregating network quality data from multiple devices to generate a regional network quality map; a policy management module for performing network status prediction and global optimization analysis based on the regional network quality map to generate a target handover policy; and an output execution module for distributing the target handover policy to the target terminal so that the target terminal can execute the operator network handover operation based on the target handover policy.

[0038] In this application, a data calculation module generates network quality data for multiple devices based on real-time network parameters corresponding to multiple operator networks, thereby capturing the dynamic status changes of each operator network in real time. A map generation module aggregates the network quality data of multiple devices to generate a regional network quality map, providing an overall view of the regional network quality. A policy management module performs network status prediction and global optimization analysis based on the regional network quality map to generate a target handover policy, enabling the prediction of the future status of the operator network and the planning of the optimal handover path. An output execution module distributes the target handover policy to the target terminal to execute the handover operation, so that the target terminal can automatically switch to the best operator network based on real-time and predicted network conditions when moving across networks, thereby effectively avoiding connection interruptions caused by handover lag or improper handover, and maintaining a high-quality service level. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of a module of an IoT card management platform according to this application;

[0040] Figure 2 This is a flowchart illustrating the steps performed by the data calculation module in this application.

[0041] Figure 3 This is a flowchart illustrating the steps performed by the map generation module in this application.

[0042] Figure 4 This is a flowchart illustrating the steps performed by the strategy management module of this application.

[0043] Explanation of icon numbers:

[0044] 10. Data Calculation Module; 20. Map Generation Module; 30. Strategy Management Module; 40. Output Execution Module. Detailed Implementation

[0045] The following combination Figures 1-4 This application will be described in further detail.

[0046] In the description of this application, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0047] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0048] This application discloses an IoT card management platform.

[0049] Please refer to Figure 1 The IoT card management platform proposed in this embodiment includes:

[0050] The data calculation module generates network quality data for multiple devices based on real-time network parameters corresponding to multiple operator networks; the map generation module aggregates network quality data from multiple devices to generate a regional network quality map; the policy management module performs network status prediction and global optimization analysis based on the regional network quality map to generate target handover policies; and the output execution module distributes the target handover policies to the target terminals so that the target terminals can execute the operator network handover operation based on the target handover policies.

[0051] The IoT SIM card management platform proposed in this embodiment has a data calculation module that can calculate and generate network quality data for each device based on real-time network parameters of multiple operator networks, providing real-time, quantitative device network quality data for the management system on the IoT SIM card management platform. The map generation module can aggregate and merge the scattered device network quality data to generate a regional network quality map reflecting the overall operator network status, thereby elevating point data into analyzable global spatial intelligence. The policy management module can perform predictive analysis and global optimization of operator network status based on this regional network quality map, thereby generating the optimal target switching strategy, completing the key transformation from situational awareness to intelligent decision-making. The output execution module can distribute the generated target switching strategy to the target terminal and drive it to execute the switching operation, realizing closed-loop execution of the decision. These four modules, combined in this way, form a complete closed loop of data acquisition, situational awareness generation, intelligent decision-making, and precise execution. This enables terminal devices to automatically, promptly, and accurately switch to the optimal network during movement, based on in-depth analysis of the global real-time and predicted situational awareness of multiple operator networks, thus fundamentally ensuring connection continuity and service quality during cross-network movement at the system level.

[0052] It should be noted that the output execution module encapsulates the determined target handover policy into a structured instruction message and sends it to the target terminal through a secure, low-latency communication link, such as based on the operator signaling channel or an encrypted IP connection. This instruction message contains clear handover triggering conditions, the target operator network identifier, and necessary access parameters. Upon receiving the instruction message, the policy execution agent of the target terminal immediately parses and loads it, transforming it into an internal subnet instruction executable by the target device. It continuously monitors its own status and network environment. Once the preset handover trigger conditions in the target handover policy are met, such as reaching a specified time, entering a specific geographical area, or the current network quality falling below a threshold, the target terminal automatically invokes its network connection management interface. Following the target handover policy instructions, it executes the entire handover operation, from disconnecting from the current network and initiating attachment and authentication to the target operator's network, until the data connection is rebuilt on the target operator's network. This accurately and reliably translates the generated target handover policy into actual network handover actions on the target device, thereby physically closing the entire control loop from "perception-analysis-decision" to "execution." Ultimately, this ensures that every device can achieve automatic, accurate, and timely network migration in a dynamically changing network environment based on globally optimal decisions, guaranteeing the continuity of service connections and optimizing the service experience.

[0053] In one feasible implementation, refer to Figure 2 As shown, the specific execution steps of the data calculation module include steps S11 to S13, wherein real-time network parameters are collected through the acquisition terminal:

[0054] Step S11: Perform multi-dimensional feature extraction processing on the real-time network parameters to obtain feature vectors.

[0055] This embodiment combines, transforms, and mines deep information from multiple complementary dimensions of real-time network parameters. For example, ① it extracts statistical features such as mean, variance, and short-term fluctuation trends of real-time network parameters from the time domain dimension. Specifically, for the signal strength in continuously collected real-time network parameters, the average of all signal strengths is calculated within a preset time window to obtain the mean feature reflecting the average level of the signal within that time window. Secondly, based on the calculated mean, the square of the difference between each signal strength and the mean is calculated for each signal strength within the same time window, and the average of these squared values ​​is then calculated to obtain the variance, which is used to quantify the fluctuation amplitude and stability of the signal strength within that time window. Finally, to capture short-term fluctuation trends, a smaller sliding sub-window is usually used to directly calculate the difference sequence between adjacent signal strengths within the sliding sub-window. By analyzing the direction and amplitude of this difference sequence, the short-term fluctuation trend of the signal in the near future is characterized as whether it is rising, falling, or stable.

[0056] ② Analyze the distribution characteristics of real-time network parameters in different frequency bands from the frequency domain dimension. Specifically, firstly, the time-series signal in the collected real-time network parameters, i.e., the signal strength that changes with time, is converted into a frequency domain signal through a fast Fourier transform to obtain its spectrum. Then, based on the obtained spectrum, the spectrum is divided according to the preset frequency band boundaries, such as dividing it into low-frequency band, mid-frequency band, and high-frequency band, and the integral or summation of the square of the amplitude corresponding to all frequency components in each frequency band is calculated to quantify the signal energy carried by that frequency band. Finally, the distribution characteristics of real-time network parameters in different frequency bands are characterized by analyzing the proportion of energy values ​​of each frequency band to the total energy, or by observing the main frequency band positions where the spectral energy is concentrated.

[0057] ③ Generate combined features from the correlation dimension, such as "interaction term between the signal-to-noise ratio of real-time network parameters and the data packet loss rate of real-time network parameters". Specifically, first, at the same time point or within the same time window, obtain different but potentially correlated real-time network parameters, such as the signal-to-noise ratio and the data packet loss rate of real-time network parameters; then, multiply, divide, calculate the difference, or construct a more complex polynomial to process the two parameters, thereby generating a new feature value. This newly generated feature value is the interaction term.

[0058] ④ Constructing spatial and temporal correlation features reflecting the stability of the operator's network status by integrating real-time network parameters reported by the acquisition devices. Specifically: First, in the spatial dimension, based on the geographical location information of each acquisition device, acquisition devices within a preset proximity range are divided into a device group. Then, the variance or standard deviation of the signal strength of the same real-time network parameter reported by all acquisition devices in the device group at the same time is calculated to obtain the spatial correlation features characterizing the consistency and stability of the network status in the area. Second, in the temporal dimension, for the same acquisition device, the real-time network parameters reported at multiple consecutive sampling times are constructed into a time series. By calculating the autocorrelation coefficient of this time series at different time lags, the dependence and pattern of the network parameters themselves changing over time are quantified, thereby extracting the temporal correlation features. Finally, by aligning and combining the spatial correlation feature sequence extracted from the same group of devices with the temporal correlation features extracted from each stimulus device, for example, constructing a joint feature vector that simultaneously includes spatial regional stability indicators and temporal autocorrelation indicators, a spatial and temporal correlation feature that can comprehensively reflect the consistency of the operator's network status in spatial distribution and its persistence in temporal evolution is formed.

[0059] This series of processes transforms the raw, potentially independent parameters into a structured feature vector with significantly increased information density (i.e., statistical features, distribution features, combined features, and spatial and temporal correlation features). The primary effect of this is to provide the pre-set machine learning model in subsequent steps with inputs that are more representative and reveal the complex nonlinear relationships of network quality. This greatly improves the accuracy and robustness of the pre-set machine learning model in quality prediction, enabling the final generated device network quality data to more accurately reflect the true state of the network.

[0060] Step S12: The feature vector is processed by a preset machine learning model to obtain a predicted quality score.

[0061] The feature vector, which integrates statistical features, distribution features, combined features, and spatial and temporal correlation features, is provided as input data to a pre-trained machine learning model, such as a gradient boosting decision tree (GBDT) model. During the training phase, the pre-trained machine learning model has learned from a large number of historical samples how to map complex feature vectors to a quantifiable network quality score.

[0062] During current operation, the pre-built machine learning model automatically and hierarchically performs conditional judgments and information integration on each feature dimension of the input feature vector based on its numerous internally constructed decision rule trees. The feature vector containing all the above-mentioned types of features is input as a whole into the pre-trained machine learning model. When processing, each decision tree within the pre-built machine learning model automatically and non-linearly evaluates different features in the input vector (whether it is time-domain statistical features reflecting signal stability, distribution features revealing frequency-domain energy structure, interactive combination features characterizing the coupling relationship between parameters, or correlation features depicting the spatiotemporal evolution pattern of network state) according to the rules learned during the training phase. By dividing the samples into paths at the nodes of each tree according to the thresholds of different features, the complex information contained in all features is gradually integrated and refined. Finally, by integrating the output results of all decision trees and converting them into a continuous numerical value, this is used as a predicted quality score reflecting the overall quality of the network. Through an algorithm that can automatically capture and model the complex non-linear relationships between high-dimensional features, it replaces the traditional evaluation methods that rely on expert experience or simple linear formulas, thereby making more accurate, stable, and interpretable quantitative predictions of network quality.

[0063] Step S13: Based on the device context information of the acquisition terminal corresponding to the real-time network parameters, the predicted quality score is calibrated to obtain device network quality data.

[0064] This embodiment calibrates the predicted quality score based on the device context information of the acquisition terminal corresponding to the real-time network parameters. This overcomes the potential universality bias of the predicted quality score generated by the preset machine learning model in the previous step, which is based solely on network-side feature vectors. This is because the general preset machine learning model does not fully consider the unique state and operating scenario of the specific acquisition terminal. This step introduces device context information such as acquisition terminal type, real-time mobile speed, service priority of the currently running application, or remaining battery power as correction factors into the evaluation system. This allows for contextualized correction and personalized adaptation of the predicted quality score, making the obtained device network quality data no longer an abstract score detached from the specific user, but a high-fidelity quality representation that closely integrates objective network status and subjective device context.

[0065] Furthermore, step S13 also includes steps S131 to S133:

[0066] Step S131: Perform context type recognition processing based on device context information, obtain context type label, and then extract features from the preset quantization model corresponding to the context type label to obtain context feature vector.

[0067] The device context information, such as whether the acquisition device is in a high-speed moving state, a low-speed moving state, or a stationary state, and whether the service carried by the acquisition device is time-sensitive, bandwidth-sensitive, or normal background, is parsed and matched by a preset classification logic, and classified into one of a limited number of preset discrete categories, such as outputting the context type label "high-speed moving - time-sensitive". Then, feature vectors are extracted from the preset quantization model corresponding to the obtained context type label. Specifically, the management system pre-stores preset quantization models specifically designed or trained for each defined context type label. For example, for the "high-speed movement - latency sensitive" type, the corresponding preset quantization model has a built-in series of specific transformation and combination rules for key context parameters such as movement speed, rate of change of direction, and service request interval. The preset quantization model first activates its internal specific processing logic and parameter set bound to that type label based on the input context type label. These logics and parameter sets are specifically optimized for the characteristics of this scenario during the preset quantization model construction phase. Next, the preset quantization model combines the device context information input or associated with the current context type label, such as movement speed value, curvature of the motion trajectory, and the specific latency limit declared by the currently running application, and calculates according to preset mathematical transformations, encoding rules, or through a lightweight sub-neural network. This condenses the information into a set of values ​​with clear physical or business meaning. For example, for mathematical transformations, the device context information corresponding to the context type label is directly calculated according to preset formulas. For example, the moving speed value is mapped to the [0,1] interval through linear normalization, or the trajectory curvature is converted into discrete levels through a piecewise function (such as fixed values ​​corresponding to low, medium, and high curvature); for the encoding rules, the context type label is converted into a numerical value according to a predefined mapping table, for example, the context type label "video stream" is represented as a vector [1, 0, 0] through one-hot encoding; and for more complex nonlinear relationships, a lightweight sub-neural network (such as a fully connected network containing only one hidden layer) is used for processing, and the numerical value or one-hot encoding obtained after the above processing is used as the input. The numerical or one-hot encoding of the input layer of the sub-neural network is first multiplied and summed with the corresponding weights in the weight matrix of the first layer, and then the bias term of the neuron in that layer is added to form the weighted input of each neuron. Next, the weighted input is fed into the activation function (such as ReLU or Sigmoid) specified for that layer for nonlinear transformation. The purpose of this transformation is to introduce nonlinear factors so that the sub-neural network can learn and represent more complex patterns. The transformed output is used as the activation value and becomes the input of the next layer (hidden layer or output layer), and the above steps of multiplying with the new weight matrix, summing, adding bias, and processing through the activation function are repeated.The signal propagates layer by layer, with each layer's weight matrix defining how it combines all features from the previous layer. The activation function applies a non-linear mapping after each combination. Finally, when the signal reaches the output layer, the activation values ​​of its neurons (the number determined by the desired output feature dimension) undergo possible final transformations (such as linear output or through the Softmax function) to form and output a set of numerical values ​​with specific representational meaning, such as a contextual feature vector containing scene fitness coefficients. This process transforms raw, heterogeneous device context information into structured features that are highly relevant to the device's current scene, have a uniform format, and high information density. This provides accurate and targeted input for subsequent calibration weight quantization calculations, ensuring that calibration closely matches the device's actual operating mode and significantly improving the scene adaptability and individual accuracy of the final device network quality data.

[0068] The preset classification logic is a set of rules that maps device context information (such as mobility status and service type) to context type labels (such as "high-speed mobility - latency sensitive type"). The preset quantization model is a calculation module or algorithm specifically designed for each context type label to convert the device context information classified into that category into a structured numerical feature vector.

[0069] Step S132: Calculate the weights of the context feature vectors using a preset calibration model to obtain the calibration weights.

[0070] It should be noted that the preset calibration model is a pre-trained machine learning model (e.g., a multilayer perceptron) designed to establish a mapping relationship between the context feature vector of the acquisition device and the calibration weights. During computation, the input layer of the preset calibration model receives the context feature vector. Each value in this context feature vector is multiplied by the weight parameter corresponding to each neuron in the first hidden layer. All products are summed with the bias parameter of that neuron to form the net input. This net input is passed through the activation function (e.g., ReLU) of this layer to generate a non-linear output, which serves as the activation value of that neuron. This activation value is passed as a signal to the next hidden layer, and the process of multiplying with a new weight matrix, summing, and transforming through the activation function is repeated. With each layer of this operation, the information is abstracted and integrated at a higher level. After several such nonlinear interlayer propagations, the signal reaches the output layer. Each neuron in the output layer (the number of which is consistent with the number of calibration weights required; for example, if an overall adjustment coefficient for the predicted quality score needs to be generated, then one neuron is used) performs a final linear or slightly nonlinear calculation, summing all its inputs with the corresponding weights and adding the final bias. The calculation result is directly output as a scalar value, which is the calibration weight. This enables the determination of the adjustment direction and intensity to be applied when correcting the predicted quality score on the network side based on the specific device context information. This achieves a leap from fixed rules to adaptive, scenario-based decision-making in the calibration process, ensuring that the final generated device network quality data can accurately reflect the actual connection quality expectation of the corresponding acquisition device in a specific context.

[0071] Step S133: Perform nonlinear fusion processing on the predicted quality score based on the calibration weight to obtain the device network quality data.

[0072] The predicted quality score and calibration weights are simultaneously input into a nonlinear fusion function (e.g., a parameterized variant of the sigmoid function). The calibration weights act as adjustment parameters within this nonlinear fusion function, controlling the shape and offset of the nonlinear fusion curve. Based on its mathematical definition, this nonlinear fusion function performs a nonlinear transformation mapping on the predicted quality score. For example, it substitutes the predicted quality score and calibration weights together into a nonlinear mapping function, such as a parameterized sigmoid function, with the form: ), where k and The value and sign of the calibration weight determine the result; when the calibration weight is positive, it increases k and... Decreasing the calibration weights causes the entire nonlinear fusion curve to become steeper and shift to the left in the input scoring region. This significantly enhances the predicted quality score, highlighting its high quality advantage. However, when the predicted quality score approaches its theoretical extreme, the saturation characteristic of the nonlinear mapping function prevents the output from growing indefinitely. Conversely, when the calibration weights are negative, k decreases and... The increase in weight causes the nonlinear fusion curve to flatten and shift to the right, effectively compressing or suppressing the predicted quality score to reflect the negative impact of the context feature vector. Through this nonlinear function controlled by calibration weights, the predicted quality score is remapped to a new, calibrated numerical range that more accurately reflects the expected experience in the current device context. The final output function value is the network quality data for that device. Through a nonlinear mathematical transformation driven by device context information, the predicted quality score is corrected and adapted to the actual quality expectation of a specific acquisition device in a specific scenario. This results in device network quality data that carries both objective network state information and deeply integrates individual device context awareness, ensuring that the data foundation upon which subsequent network switching decisions rely has high individual relevance and scenario accuracy.

[0073] In one feasible implementation, refer to Figure 3 As shown, the specific execution steps of the map generation module include steps S21 to S23:

[0074] Step S21: Based on the acquisition device and acquisition timestamp to which the network quality data of each device belongs, perform spatiotemporal clustering processing on the network quality data of each device to obtain quality clustering regions.

[0075] Based on the acquisition device and acquisition timestamp of each device's network quality data, spatiotemporal clustering is performed on the network quality data of each device to obtain quality clustering regions. In this way, continuous spatiotemporal blocks with high similarity in network quality characteristics are identified and divided from massive, scattered device network quality data that are mixed with temporal and spatial attributes. This solves the problem that individual device network quality data is too isolated and difficult to be directly used to represent the overall network status of a region. It realizes the aggregation and merging of discrete device network quality data through their inherent spatiotemporal correlation, thereby elevating point information into quality clustering regions with clear spatiotemporal boundaries and representative quality characteristics. This provides a basic geographic spatiotemporal framework for subsequent regional network status analysis and optimization decisions.

[0076] Furthermore, step S21 also includes steps S211~S212:

[0077] Step S211: Obtain the spatial location of each acquisition device at the corresponding acquisition timestamp, and encode the spatial location, acquisition timestamp, and network quality data of each acquisition device to obtain the encoding features corresponding to each acquisition device.

[0078] First, the spatial location of each data acquisition device at its corresponding data acquisition timestamp is obtained. This is achieved by using the device's built-in positioning module (such as GPS) or by leveraging the operator's base station positioning technology to record and report the geographic coordinates (such as latitude and longitude) of each device in real time when generating each real-time network parameter. Next, the spatial location, data acquisition timestamp, and network quality data of each device are encoded to obtain corresponding coding features. Specifically, these three types of heterogeneous information are fused and vectorized. For example, the latitude and longitude of the spatial location are converted into planar coordinate values ​​through map projection, and the data acquisition timestamp is decomposed and encoded into packets. By incorporating periodic information such as hours and days of the week, the device network quality data is treated as a scalar value. All these values ​​are then concatenated and combined in a preset order to form a coded feature. This transforms the originally independent and multi-dimensional spatial, temporal, and device network quality data into a unified, complete, and machine-readable numerical feature representation. This provides a directly calculable and processable basic data unit for the next step of spatiotemporal clustering algorithms based on feature spatial distance. It enables the clustering process to simultaneously consider the geographical proximity, temporal similarity, and quality consistency of devices, laying a crucial data foundation for accurately dividing quality clustering regions.

[0079] Step S212: Calculate the distance of each coding feature in the feature space, and obtain the target acquisition devices to which the two coding features belong (the distance difference is less than the preset distance). Take one of the target acquisition devices as the center point and the distance between the two target acquisition devices as the radius to determine the quality clustering region.

[0080] The Euclidean distance formula is used to calculate the distance between each encoded feature in the feature space. Each encoded feature is regarded as a point in a high-dimensional space. The square root of the sum of the squared differences between each pair of points in all feature dimensions is calculated to obtain a quantified distance value, which is used to characterize the similarity between the corresponding two acquisition devices in multiple dimensions such as space, time and device network quality data.

[0081] After calculating the pairwise distances between all encoded features, all acquisition device pairs are traversed, and those with distance values ​​less than a pre-defined distance for defining proximity are selected. Each of these selected acquisition devices is marked as a target acquisition device. For each pair of acquisition devices, a quality clustering region is determined with one of the target acquisition devices as the center point and the distance between the two target acquisition devices as the radius. This is done by randomly selecting one target acquisition device from the pair as the initial center point, using the calculated distance between the two target acquisition devices in the feature space as a mapping reference for the geographic radius, and delineating a circular area in the real geographic space. This circular area is the quality clustering region. This process is recursively or iteratively performed on all target acquisition device pairs and their adjacent devices that meet the distance condition, thereby merging overlapping areas and finally forming a complete and continuous quality clustering region. This achieves the identification of the geographic distribution of network quality by using a distance metric that integrates spatiotemporal and quality multidimensional information, grouping acquisition devices with similar physical locations, similar sampling times, and consistent network quality data into the same continuous geographic region. This provides a basis for spatial unit division for subsequent regional network state analysis and optimization.

[0082] Step S22: Filter out abnormal data in the quality clustering region to obtain quality data.

[0083] For each quality cluster region, the network quality data of all acquisition devices within the cluster region are first treated as a dataset. Then, statistical distribution-based methods, such as box plots or Z-scores, are used to identify and remove outliers. Taking the Z-score method as an example, the management system calculates the mean and standard deviation of the dataset, and then calculates the Z-score for each data point, which is the difference between the data point and the mean divided by the standard deviation. Data points with absolute Z-scores exceeding a preset value are identified as statistically significant outliers and removed from the dataset. After this process, the remaining data points in the dataset are the filtered quality data that can represent the general network quality status of the corresponding quality cluster region. This effectively eliminates the impact of outlier noise data introduced by temporary equipment failures, random measurement errors, or instantaneous strong local interference on the overall regional quality assessment. This ensures that the quality data used to construct the regional network quality map is stable and regionally representative, which helps improve the accuracy, reliability, and practical value of the generated regional network quality map.

[0084] Step S23: The quality data is processed by a preset weighted fusion algorithm to obtain a regional network quality map.

[0085] A regional network quality map is obtained by performing map rendering processing on the quality data through a preset weighted fusion algorithm. This transforms the quality data, which has been clustered and purified in the previous steps but still exists in a discrete form in various geographical areas, into a continuous, intuitive regional network quality map that can be used for macro-level decision-making. This regional network quality map not only clearly reveals the distribution differences of network quality, strong and weak areas and boundary transitions throughout the geographical area, but also ensures that the quality status reflected by the regional network quality map is robust and representative through weighting.

[0086] Among them, the preset weighted fusion algorithm is a predefined and encapsulated computing model. Its core is to extract quality indicators and optimize weights for quality data from different sources or locations based on geospatial relationships and data attributes, and generate quality fusion weights for rendering the quality data into a regional network instruction map.

[0087] Furthermore, step S23 may also include steps S231 to S232:

[0088] Step S231: Extract quality indicators from the quality data to obtain multiple sets of quality indicators.

[0089] From the quality data of each quality cluster region, various statistical measures or derived indicators that can characterize the network quality characteristics of the corresponding quality cluster region from different perspectives are calculated and extracted. For example, the mean of the quality data of the quality cluster region is calculated to reflect the overall level, the variance is calculated to reflect the degree of fluctuation, and specific quantiles (such as 90th percentile delay) are calculated to reflect extreme cases. More complex statistical measures such as skewness, kurtosis, and coefficient of variation can also be calculated. Indicators reflecting the same type of characteristics are grouped into a subset, thus forming multiple quality indicator sets with clear physical or business orientations, such as "central trend indicator set", "dispersion indicator set", "tail performance indicator set" and "spatiotemporal stability indicator set". This realizes the expansion of a single-dimensional quality value into a multi-dimensional and refined quality feature profile, providing input information with clear semantics for weight optimization in subsequent steps. This allows map rendering to no longer rely solely on a general quality data, but to adaptively select and integrate the most relevant quality dimensions according to different optimization objectives (such as emphasizing stability or peak performance), thereby improving the information content, decision support capability, and scene adaptability of the generated regional network quality map.

[0090] Step S232: Optimize the weights of the quality index set to obtain the quality fusion weights, and then perform raster rendering of the quality data based on the quality fusion weights to obtain the regional network instruction map.

[0091] Multiple quality index sets are input into a multi-objective optimization model. This model performs a trade-off calculation based on preset optimization objectives, such as simultaneously pursuing accurate reflection of average quality, significant warning of quality fluctuations, and prominent display of extremely degraded areas when rendering maps. The objective of "accurately reflecting average quality" is quantified into a sub-objective function. Its mathematical expression is to minimize the mean square error between the predicted map values ​​and the actual observed regional average quality index, that is... The objective of "significantly warning of quality fluctuations" is quantified into another sub-objective function. Its expression is to minimize the degree to which map rendering values ​​ignore the regional quality variance index (corresponding to the set of dispersion indexes), that is... The aim is to optimize the visual representation of areas with high variance on the map; the objective of "highlighting extremely degraded areas" is quantified into a sub-objective function. Its expression is to minimize the map's representation error of regional tail performance indicators (such as the 95th percentile delay, corresponding to the tail performance indicator set), i.e. To ensure that extremely degraded regions are not masked by smoothing, these three sub-objective functions together constitute the mathematical description of the multi-objective optimization problem.

[0092] A multi-objective optimization algorithm is used to solve this multi-objective optimization problem. The specific process is as follows: The multi-objective optimization algorithm searches within a solution space composed of all weight coefficients. For example, the Non-Dominated Sorting Genetic Algorithm (NSGA-II) initializes a population composed of random weight coefficient combinations. In each iteration, the algorithm calculates the sub-objective function values ​​corresponding to each weight combination (i.e., a candidate solution). The specific numerical values ​​of the objective function are used to evaluate its performance, and non-dominated sorting and crowding calculations are performed based on these objective function values ​​to compare the quality of the solutions. Through selection, crossover, and mutation operations, a new population evolves. This process is repeated continuously, driving the population to evolve towards simultaneously optimizing all sub-objectives, and ultimately approaching the Pareto optimal frontier, that is, finding a set of solutions in which any sub-objective (such as reducing...) is optimal. Further improvements to any of these sub-objectives will inevitably lead to at least one other sub-objective (such as...). or As the Pareto front deteriorates, an additional decision criterion is used to select the final solution that best suits the current decision preference. The specific coefficient values ​​of each quality index set contained in this selected solution are the quality fusion weights obtained by solving a multi-objective optimization problem and weighing multiple conflicting objectives.

[0093] After obtaining the quality fusion weights, the target geographic area is divided into a fine, regular grid. For each grid, its geographical location is first used to find its own or neighboring quality clusters. Then, the quality data of these quality clusters is obtained. Next, the optimized quality fusion weights are used to weight and sum these quality data to calculate a comprehensive value representing the quality of the grid network. Finally, the comprehensive quality values ​​of all grids are mapped to corresponding colors and filled according to preset color levels, thereby generating a complete regional network quality map with continuously gradient colors. Through weight optimization, the map rendering process can adaptively highlight the quality dimensions that are most important for current network management and optimization decisions (such as focusing on average quality in daily operation and maintenance, and focusing on volatility during fault diagnosis). Through weighted rasterization, a regional network quality map that directly supports geospatial analysis is generated, thereby greatly improving the practicality and flexibility of the regional network quality map as a decision-making tool.

[0094] In one feasible implementation, refer to Figure 4 As shown, the specific execution steps of the policy management module include steps S31 to S32:

[0095] Step S31: Perform spatiotemporal sequence prediction based on the regional network quality map to obtain the predicted network state.

[0096] The system takes the sequence of regional network quality maps across multiple historical time slices as input. Each historical time slice map is a two-dimensional raster data recording network quality values ​​at different geographical locations. These sequences are processed by a spatiotemporal sequence prediction model, such as a spatiotemporal graph neural network or a three-dimensional convolutional neural network. This spatiotemporal sequence prediction model can simultaneously capture the spatial correlation and temporal evolution of network quality. For example, it can learn how signal attenuation in a certain area usually affects its surrounding areas, or the periodic network congestion patterns that occur in commercial areas during specific times of the day. The spatiotemporal sequence prediction model encodes, memorizes, and infers the input sequences through its internal hierarchical structure, and finally outputs a predicted network quality raster map covering the same geographical area in one or more future time slices. The future network quality distribution represented by this map is the predicted network state. This expands the decision-making basis from a static perception of the current network state to a dynamic inference of the network situation in the future. This makes the subsequently generated network switching strategy forward-looking, guiding target devices to switch to networks that are about to maintain high quality or are improving in advance, thereby effectively avoiding connection interruptions or service quality degradation caused by passively responding to network changes.

[0097] Step S32: Perform global optimization analysis on the predicted network state using a multi-objective optimization algorithm to obtain the optimized objective weights. Then, substitute the optimized objective weights into the preset strategy synthesis model for strategy generation processing to obtain the target switching strategy.

[0098] The fundamental purpose of using a multi-objective optimization algorithm to perform global optimization analysis on the predicted network state is to systematically coordinate and balance multiple conflicting decision objectives when formulating switching strategies. This is because the predicted network state reveals the complex situation of the future network in terms of performance, coverage, and other dimensions, while decision objectives such as lowest cost, lowest latency, and highest reliability often cannot be satisfied simultaneously. Therefore, a mechanism is needed to determine the relative importance of each objective in a specific global scenario. In this embodiment, the multi-objective optimization algorithm takes the future network conditions revealed by the predicted network state and preset business objectives (such as cost, latency, and reliability) as input, and finds a balance point that best meets the current global management needs on the Pareto optimal frontier of multiple objectives through mathematical optimization. It then outputs a set of quantified optimization objective weights, which accurately represent the priority of each decision objective in the current and foreseeable network environment. Specifically, the process of obtaining the optimization objective weights is as follows: The multi-objective optimization algorithm first initializes a population consisting of multiple randomly generated weight combinations, each weight combination representing a potential objective priority allocation scheme. In each iteration, the multi-objective optimization algorithm evaluates the comprehensive performance of each weight combination. This is done by applying the weight combination to the pre-defined network state. The network state is measured by calculating quantified values ​​of how well the overall network will achieve various objectives (such as total cost, average latency, and worst-case reliability) in the future under a given weight. Subsequently, a multi-objective optimization algorithm, based on these calculated quantified values, uses non-dominated sorting to stratify all individuals in the population to identify Pareto non-dominated solutions that are no worse than other solutions on all objectives and are better than at least one objective. The algorithm also calculates the crowding distance of each solution within the same non-dominated layer to measure its diversity. Then, selection is based on the individual's non-dominated layer (preferring higher layers) and crowding distance (preferring larger distances within the same layer). The selection process involves retaining superior individuals and performing crossover and mutation operations on them to generate new weight combinations, thus forming the next generation of the population. This iterative process of evaluation, sorting, selection, and recombination continues, driving the entire population to evolve continuously in the solution space towards the direction of simultaneously optimizing all objectives, i.e., the Pareto optimal front. When the preset number of iterations or convergence conditions are reached, a final solution is selected from the solution set located at the Pareto front in the final population, based on an additional decision criterion (such as selecting the minimum distance to the ideal point or a preference specified by the decision-maker). The specific weight value corresponding to this selected solution is the optimization objective weight.

[0099] Subsequently, the optimized target weights are substituted into a pre-defined policy synthesis model for policy generation. Its core function is to transform abstract, quantified target priorities into specific, executable network handover instruction rules. The pre-defined policy synthesis model is a pre-defined or pre-trained decision logic framework that receives the optimized target weights and related predicted network state and constraint information as input. Internally, it contains mapping rules from the target to specific network selection actions (e.g., rule-based decision trees or reinforcement learning-based policy networks). By processing these inputs, the prediction policy model synthesizes and generates a complete policy that clearly guides the target device on which operator's network to switch to under what conditions—the target handover policy. This entire process achieves a closed loop from multi-dimensional prediction of future networks to global optimization trade-offs, and finally to the generation of an executable policy, ensuring the forward-looking nature, global optimality, and operability of the handover decision.

[0100] Furthermore, step S32 may also include steps S321 to S323:

[0101] Step S321: Infer network constraints based on the optimization target weights to obtain the policy constraint set.

[0102] The optimization target weights are taken as input and processed through a predefined inference logic model, such as a mapper based on a rule engine or a lightweight neural network. This inference logic model stores the correlation between different optimization target weights and potential network limiting factors. For example, a high latency weight is considered to be sensitive to network transmission speed, thus inferring that "end-to-end latency must not exceed X milliseconds" as a limiting factor; a high cost weight is considered to be sensitive to tariffs, thus inferring that "single handover cost must not exceed Y yuan" as a limiting factor; and a high reliability weight is considered to be sensitive to connection stability, thus inferring that "received signal strength must not be lower than Z dBm" as a limiting factor. The inference logic model activates corresponding association rules based on the magnitude of the input optimization target weights. For each activated constraint, it generates policy constraints with specific thresholds or range descriptions. All these policy constraints together constitute a policy constraint set, which transforms the abstract optimization target weights representing the priority of business objectives into a series of operable network performance and resource boundary conditions. This provides a clear range of feasible solutions for subsequent candidate policy generation, ensuring that the final generated target switching policy is not only globally optimized in terms of objectives, but also strictly constrained by actual network conditions and achievable at the specific execution level.

[0103] Step S322: Based on the policy constraints in the policy constraint set, output the candidate policies that match each policy constraint.

[0104] The management system internally pre-sets or dynamically generates a series of basic network handover action options, such as "switch to operator A's 5G network," "switch to operator B's 4G network," and "maintain current network connection." Then, using a constraint-based problem-solving logic, it iterates through these basic action options, treating each policy constraint in the step policy constraint set, such as "latency ≤ 50ms," "signal strength ≥ -85dBm," and "single handover cost ≤ 0.3 yuan," as a set of filtering conditions that must be met simultaneously. It then examines each basic action option to see if it satisfies all policy constraints under the predicted network state. Those basic action options that pass the test of all policy constraints are selected. Each feasible solution is marked as a viable option and is formatted as "Policy Alpha: In location X region, at future time T, switch to operator A's 5G network." This viable solution is then output as a candidate policy that matches the policy constraints. All these candidate policies that pass the test together constitute the candidate policy set. This transforms the abstract policy constraints into concrete and feasible network handover action plans, ensuring that the subsequent policy optimization stage is carried out within a feasible solution space that is strictly limited by actual network conditions. This fundamentally avoids generating theoretically optimal but practically unexecutable or invalid candidate policies that violate key network constraints.

[0105] Step S323: Calculate the network connection quality of each candidate strategy based on the current network parameters when the target device accesses the operator network corresponding to each candidate strategy, and determine the candidate strategy corresponding to the network connection quality with the first-ranked value among the network connection quality as the target handover strategy.

[0106] The management system drives the target device to initiate a trial access or measurement request to the operator network specified by each candidate strategy, thereby obtaining the target device's current network parameters under the corresponding operator network in real time, such as signal strength, network latency, and packet loss rate. Subsequently, through a scoring function (e.g., a lightweight weighted calculation formula) similar to the quality prediction processing logic in step S12 but used for real-time evaluation, these real-time obtained current network parameters are calculated in real time. This scoring function receives the real-time obtained current network parameters corresponding to the operator network specified by the candidate strategy as input, and dynamically determines the weight coefficients of each parameter in the scoring function based on the optimization target weight. The management system presets a reasonable theoretical or empirical value range for each type of current network parameter to be evaluated. This range defines the boundary between the "worst" and "best" values ​​that such current network parameters may appear in the evaluation scenario. After obtaining a current network parameter, it is processed through a predefined linear transformation formula, such as... The calculations are performed to map the current network parameters from their range of values ​​with specific physical units to a unified dimensionless scaling range, such as between 0 and 1. This allows for rapid normalization of each current network parameter to eliminate the influence of dimensions. Subsequently, these normalized parameter values ​​are weighted and summed with their corresponding optimization weight coefficients to calculate a scalar score between 0 and 1. This score quantitatively reflects the network connectivity quality that the candidate strategy can provide at the current instant.

[0107] After calculating the network connection quality of all candidate strategies, the management system sorts these network connection qualities from high to low and automatically selects the candidate strategy corresponding to the highest-ranked network connection quality as the target switching strategy. This ensures that the final target switching strategy is not only based on historical predictions and global optimization, but also the best-performing solution that has withstood the test of the actual network environment at the current moment. This minimizes the risk of suboptimal decisions caused by prediction bias or sudden changes in network status, and greatly improves the instant success rate and connection quality assurance of each network switching action.

[0108] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An IoT card management platform, characterized in that, include: The data calculation module is used to generate network quality data for multiple devices based on the real-time network parameters corresponding to multiple operator networks. The map generation module is used to aggregate network quality data from multiple devices and generate a regional network quality map. The strategy management module is used to perform network status prediction and global optimization analysis based on the regional network quality map, and generate target switching strategies. The output execution module is used to send the target handover policy to the target terminal so that the target terminal can perform the handover operation of the operator network based on the target handover policy. The real-time network parameters are collected by a data acquisition terminal. The step of generating network quality data for multiple devices based on the real-time network parameters corresponding to multiple operator networks includes: The real-time network parameters are subjected to multi-dimensional feature extraction processing to obtain feature vectors; The feature vector is subjected to quality prediction processing by a preset machine learning model to obtain a predicted quality score; Based on the device context information of the acquisition terminal corresponding to the real-time network parameters, the predicted quality score is calibrated to obtain the device network quality data.

2. The IoT card management platform according to claim 1, characterized in that, The step of calibrating the predicted quality score based on the device context information of the acquisition terminal corresponding to the real-time network parameters to obtain the device network quality data includes: Based on the device context information, context type recognition processing is performed to obtain a context type label. Then, features are extracted from a preset quantization model corresponding to the context type label to obtain a context feature vector. The calibration weights are obtained by calculating the weights of the context feature vectors using a preset calibration model. The predicted quality score is subjected to nonlinear fusion processing based on the calibration weights to obtain the device network quality data.

3. The IoT card management platform according to claim 1, characterized in that, The step of aggregating network quality data from multiple devices to generate a regional network quality map includes: Based on the acquisition device and acquisition timestamp to which each device's network quality data belongs, spatiotemporal clustering is performed on each device's network quality data to obtain quality clustering regions; The quality clustering regions are filtered for outlier data to obtain quality data. The quality data is processed by a preset weighted fusion algorithm to obtain the regional network quality map.

4. The IoT card management platform according to claim 3, characterized in that, The step of performing spatiotemporal clustering processing on the network quality data of each device according to the acquisition device and acquisition timestamp to obtain the quality clustering region includes: The spatial location of each acquisition device at the corresponding acquisition timestamp is obtained, and the spatial location, acquisition timestamp and network quality data of each acquisition device are encoded to obtain the encoding features corresponding to each acquisition device. Calculate the distance between each of the encoded features in the feature space, and obtain the target acquisition devices to which two encoded features belong (the distance difference is less than a preset distance). Using one of the target acquisition devices as the center point and the distance between the two target acquisition devices as the radius, determine the quality clustering region.

5. The IoT card management platform according to claim 3, characterized in that, The step of performing map rendering processing on the quality data using a preset weighted fusion algorithm to obtain the regional network quality map includes: The quality data is processed to extract quality indicators, resulting in multiple sets of quality indicators. After optimizing the quality index set to obtain the quality fusion weight, the quality data is rasterized and rendered according to the quality fusion weight to obtain the regional network instruction map.

6. The IoT card management platform according to claim 1, characterized in that, The steps of performing network state prediction and global optimization analysis based on the regional network quality map to generate a target switching strategy include: Based on the aforementioned regional network quality map, spatiotemporal sequence prediction is performed to obtain the predicted network state. The predicted network state is globally optimized and analyzed using a multi-objective optimization algorithm. After obtaining the optimized objective weights, the optimized objective weights are substituted into a preset strategy synthesis model for strategy generation processing to obtain the target switching strategy.

7. The IoT card management platform according to claim 6, characterized in that, The step of substituting the optimized target weights into a preset strategy synthesis model for strategy generation to obtain the target switching strategy includes: Based on the optimization objective weights, network constraints are inferred to obtain a set of policy constraints. Based on each policy constraint in the policy constraint set, output candidate policies that match each policy constraint respectively; Based on the current network parameters when the target device accesses the operator network corresponding to each candidate strategy, the network connection quality of each candidate strategy is calculated, and the candidate strategy corresponding to the network connection quality with the first value among the network connection quality is determined as the target handover strategy.

Images