Communication management method and system based on Internet of Things

Abstract

The invention relates to the technical field of communication of the Internet of Things, and particularly discloses a communication management method and system based on the Internet of Things. The method comprises the following steps: acquiring bandwidth, throughput and time delay parameters of equipment, and generating a standardized communication data set through timestamp alignment and topology construction; constructing a dynamic graph structure by using the data set, and generating a 32-dimensional compression feature vector through space-time convolution and compression coding; based on the vector, a real-time communication coefficient value is output by fusing multiple parameters according to a dynamic weight coefficient; obtaining a historical communication coefficient database, cooperatively training a threshold model through a meta-learning framework, and outputting an adaptive determination threshold; comparing the real-time communication coefficient value with an adaptive threshold value, and executing network state simulation to generate a base station switching or channel resetting instruction; and after executing the instruction, collecting communication parameter change data to generate feedback information, and returning the feedback information to the historical database and the standardized data set. According to the invention, accurate communication state monitoring in a dynamic environment is realized, and the base station switching decision accuracy and the network stability are improved.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things (IoT) communication technology, and particularly relates to a communication management method and system based on IoT. Background Technology

[0002] Traditional IoT communication management systems typically employ fixed threshold monitoring mechanisms to determine abnormal device communication status. For example, an alarm is triggered when bandwidth utilization exceeds a preset threshold (e.g., 80%) or latency exceeds a static threshold (e.g., 100ms). However, this approach suffers from two major technical bottlenecks:

[0003] Poor adaptability to dynamic environments: IoT devices are often deployed in complex and ever-changing environments, where network topology fluctuations and channel interference cause communication parameters to oscillate violently. Fixed thresholds cannot distinguish between normal fluctuations and real faults, resulting in a false alarm rate (such as false alarms triggered by brief signal interference) or a missed alarm rate (such as gradual channel degradation not being identified) of up to 40%.

[0004] Cold start issue for new devices: When a new device connects to the network, the lack of historical data makes threshold initialization difficult. Traditional solutions use the average value of similar devices as a temporary threshold, but due to differences in device type, location, and usage scenarios, the distribution of communication coefficients can deviate by up to 35%, leading to decision-making errors.

[0005] Existing improvement plans attempt to mitigate the problem in the following ways:

[0006] (1) The threshold is dynamically adjusted by using the average of historical parameters to calculate the sliding window, but it can only adapt to slowly changing scenarios and the response delay to sudden anomalies exceeds 500ms.

[0007] (2) Using clustering algorithms to divide devices into groups and setting group thresholds still cannot solve the problem of individual differences in devices within a group (e.g., the standard deviation of the communication coefficient within a group reaches 0.25).

[0008] (3) A simple neural network is introduced to predict the threshold, but the model needs to be trained in a concentrated manner and depends on massive labeled data. New devices need to be retrained, and the deployment delay exceeds 10 minutes.

[0009] Fundamental flaws: Existing technologies lack joint mining of device network topology correlations and spatiotemporal evolution characteristics of parameters, and threshold models cannot quickly adapt in data-scarce scenarios. This leads to decreased communication stability of IoT cellular modules (such as NB-IoT / eMTC modules) in dynamic environments, high base station handover false trigger rate, and slow mean time to recovery. Summary of the Invention

[0010] This invention provides a communication management method and system based on the Internet of Things (IoT) to solve the problem of how to dynamically generate communication coefficients based on the bandwidth, throughput, and latency parameters of IoT devices, and to achieve accurate monitoring of cellular module communication status and optimization of base station handover decisions through a meta-learning framework for collaborative training of an adaptive threshold model.

[0011] To address the aforementioned technical problems, this invention provides a communication management method based on the Internet of Things (IoT), comprising:

[0012] Collect bandwidth, throughput, and latency parameters from IoT devices, perform timestamp alignment and topology construction, and generate a standardized communication dataset.

[0013] A dynamic graph structure is constructed using a standardized communication dataset, and a 32-dimensional compressed feature vector is generated through spatiotemporal convolution and compression coding operations.

[0014] Based on a 32-dimensional compressed feature vector, the bandwidth parameter, throughput parameter, and latency parameter are weighted and fused according to dynamic weight coefficients to output the real-time communication coefficient value.

[0015] Obtain a historical communication coefficient database, collaboratively train a threshold calculation model through a meta-learning framework, and output an adaptive judgment threshold to a real-time threshold library.

[0016] The expression for the collaborative training threshold calculation model of the meta-learning framework is:

[0017]

[0018] in, It is the meta-eigenvector; For a three-channel feature fusion unit; For manifold coordinates; Information entropy of environmental characteristics; For fault labeling, a graph neural network embedding layer; This indicates a tensor concatenation operation; Label historical faults;

[0019] The expression for the output adaptive judgment threshold to the real-time threshold library is:

[0020]

[0021] in, This represents the final release threshold for the device (dev). It is a dynamic smoothing factor; The cache threshold for device dev; For exponential moving average operators; Output the original threshold;

[0022] Compare the real-time communication coefficient value with the adaptive judgment threshold, perform network status simulation operation, and generate base station handover command or channel reset command;

[0023] Execute base station switching commands or channel reset commands, collect communication parameter change values ​​to generate execution feedback data, and feed it back to the historical communication coefficient database and standardized communication dataset.

[0024] Furthermore, generating a standardized communication dataset includes:

[0025] Obtain bandwidth parameters, throughput parameters, and latency parameters, perform timestamp alignment and missing value filling to obtain preprocessed communication parameters;

[0026] The network topology relationships of devices are extracted from the preprocessed communication parameters, and an adjacency matrix is ​​constructed to obtain network topology data.

[0027] Normalization is performed on the network topology data and preprocessed communication parameters to generate a standardized communication dataset.

[0028] Furthermore, generating the 32-dimensional compressed feature vector includes:

[0029] By combining historical fault records with standardized communication datasets, a dynamic graph structure with spatiotemporal correlation is constructed.

[0030] Spatiotemporal graph convolution is performed on the dynamic graph structure to extract cross-layer correlation features and obtain spatiotemporal feature vectors;

[0031] The spatiotemporal feature vectors are compressed and encoded to generate 32-dimensional compressed feature vectors.

[0032] Furthermore, the output real-time communication coefficient values ​​include:

[0033] Load the dynamic weighting coefficients for the current device type;

[0034] The bandwidth parameter, throughput parameter, and latency parameter are weighted and fused according to the dynamic weighting coefficients.

[0035] The real-time communication coefficient value is output based on the weighted fusion calculation results.

[0036] Furthermore, the output adaptive decision threshold includes:

[0037] The threshold calculation model is initialized using a meta-learning framework;

[0038] Cross-node collaborative training is performed on the threshold calculation model based on the historical communication coefficient database;

[0039] Output adaptive judgment threshold and update it to the real-time threshold library.

[0040] Furthermore, cross-node collaborative training includes:

[0041] Historical communication coefficient data are obtained synchronously from multiple network nodes;

[0042] Aggregate distributed parameter gradients while protecting data privacy;

[0043] Update the global parameters of the threshold calculation model.

[0044] Furthermore, generating base station handover commands or channel reset commands includes:

[0045] Anomaly determination is performed based on the real-time communication coefficient value and the adaptive determination threshold.

[0046] Simulate the communication status of abnormal devices after base station handover;

[0047] Generate base station handover commands or channel reset commands.

[0048] Furthermore, the communication state simulation includes:

[0049] Construct a digital twin network environment for the target base station;

[0050] Predict the real-time communication coefficient changes after base station handover;

[0051] Instruction generation is triggered when the predicted value is higher than the current anomaly threshold.

[0052] Furthermore, the collection of communication parameter change values ​​to generate execution feedback data includes:

[0053] Record instruction execution latency and success rate;

[0054] Collect the changed values ​​of bandwidth parameters, throughput parameters, and latency parameters;

[0055] The changed values ​​are written into the historical communication coefficient database and the standardized communication dataset.

[0056] An Internet of Things (IoT)-based communication management system, applied to any one of the IoT-based communication management methods described above, comprising:

[0057] The data acquisition unit is configured to collect bandwidth parameters, throughput parameters, and latency parameters and generate a standardized communication dataset.

[0058] The feature fusion unit is configured to construct a dynamic graph structure and output a 32-dimensional compressed feature vector.

[0059] The coefficient calculation unit is configured to calculate the real-time communication coefficient value based on the dynamic weight coefficient.

[0060] The threshold generation unit is configured to train a model using a meta-learning framework and output an adaptive decision threshold.

[0061] The decision execution unit is configured to perform communication status simulation operations and generate base station switching or channel reset commands.

[0062] The feedback learning unit is configured to collect instruction execution data and feed it back to the historical communication coefficient database and the standardized communication dataset.

[0063] The key innovations of this invention include:

[0064] (1) Integrate multi-dimensional parameters such as bandwidth, throughput, and latency, and generate real-time communication coefficients through weighted calculation to dynamically reflect the communication status of the device.

[0065] (2) Construct a threshold calculation framework based on meta-learning, and use historical data to collaboratively train an adaptive threshold generator to solve the threshold drift problem in dynamic environments.

[0066] (3) Introduce digital twin network simulation technology to predict changes in communication status after base station switching or channel reset, and realize data-driven accurate decision-making.

[0067] (4) Spatiotemporal graph convolution is used to extract network topology association features, and low-dimensional feature vectors are generated by efficient compression encoding to balance feature expression ability and computational efficiency.

[0068] The following are its main beneficial effects:

[0069] (1) By dynamically fusing bandwidth, throughput and latency parameters to generate communication coefficients, and combining them with an adaptive threshold model trained by a meta-learning framework, the network can be accurately distinguished from normal fluctuations and real faults, significantly reducing false alarms and false negatives.

[0070] (2) The meta-feature encoding and cross-node collaborative training mechanism are adopted to enable the adaptive threshold model to converge quickly in the absence of historical data, ensuring that newly added devices can obtain accurate judgment thresholds in a timely manner.

[0071] (3) Based on the simulation and prediction of the state after the handover using the digital twin network, the optimal handover command or channel reset strategy is generated, which greatly shortens the fault recovery time and reduces the number of invalid handovers. Attached Figure Description

[0072] Figure 1 A flowchart illustrating a communication management method based on the Internet of Things (IoT) provided in this application embodiment;

[0073] Figure 2 This is a structural block diagram of an Internet of Things-based communication management system provided in an embodiment of this application. Detailed Implementation

[0074] Example 1: Refer to Figure 1This is a flowchart illustrating a communication management method based on the Internet of Things (IoT) provided in an embodiment of the present invention. The process may include at least steps S100-S600:

[0075] S100 collects bandwidth, throughput, and latency parameters from IoT devices, performs timestamp alignment and topology construction, and generates a standardized communication dataset.

[0076] S200: A dynamic graph structure is constructed using a standardized communication dataset. After spatiotemporal convolution and compression coding operations, a 32-dimensional compressed feature vector is generated.

[0077] The S300, based on a 32-dimensional compressed feature vector, outputs real-time communication coefficient values ​​by weighting and fusing bandwidth parameters, throughput parameters, and latency parameters according to dynamic weight coefficients.

[0078] S400: Obtain the historical communication coefficient database, collaboratively train the threshold calculation model through the meta-learning framework, and output the adaptive judgment threshold to the real-time threshold library.

[0079] S500 compares the real-time communication coefficient value with the adaptive judgment threshold, performs network status simulation operation, and generates base station switching command or channel reset command.

[0080] S600 executes base station switching commands or channel reset commands, collects communication parameter change values ​​to generate execution feedback data, and feeds it back to the historical communication coefficient database and standardized communication dataset.

[0081] Step S100 includes at least steps S110-S130:

[0082] S110. Obtain the bandwidth parameters, throughput parameters, and latency parameters of the IoT device, perform timestamp alignment and fill in missing values ​​to obtain preprocessed communication parameters.

[0083] Specifically, the bandwidth parameters of the target device are collected using the built-in probes of the cellular module. These bandwidth parameters represent the statistical value of the amount of data transmitted per unit time, and the sampling frequency is set to once per millisecond. Simultaneously, throughput parameters are obtained by parsing the transport layer protocol header fields. These throughput parameters include the ratio of successfully transmitted data packets to retransmitted data packets. Furthermore, hardware timestamp technology is used to record latency parameters, which include the round-trip time difference of data packets from the sending end to the receiving end.

[0084] Understandably, a timestamp alignment operation is performed on the bandwidth, throughput, and latency parameters. Specifically, the hardware clock signal at the acquisition time of each parameter is extracted. Using the generation time of the latency parameter as a reference point, the time axes of the bandwidth and throughput parameters are linearly interpolated to align, generating a time-synchronized three-dimensional parameter sequence. Further, missing values ​​in the three-dimensional parameter sequence are detected. When the duration of consecutive missing values ​​is less than a preset threshold, cubic spline interpolation is used to fill them. When the duration exceeds the preset threshold, a neighboring node parameter substitution mechanism is triggered. Specifically, the neighboring node parameter substitution mechanism selects the neighboring node with the highest signal strength based on the device network topology and copies its parameter values ​​for the corresponding time period to fill the missing values. The final output is a preprocessed communication parameter containing a complete time series. The data structure of the preprocessed communication parameter is a set of <bandwidth, throughput, latency> triples with timestamps.

[0085] The preprocessed communication parameters will serve as input data for network topology construction in S120. Their time alignment characteristics ensure the temporal consistency of subsequent adjacency matrix calculations, while the missing value imputation mechanism guarantees parameter continuity and avoids feature fusion failure due to data interruption.

[0086] S120. Extract the device network topology relationship from the preprocessed communication parameters, construct the adjacency matrix, and obtain the network topology data.

[0087] Specifically, the latency parameter sequence in the preprocessed communication parameters is analyzed, and the latency fluctuation correlation coefficient between any two devices is calculated. The specific calculation method for the correlation coefficient is as follows: latency change curves within the same time window are selected, and the signal coupling strength between devices is determined through Pearson correlation analysis. When the correlation coefficient exceeds a dynamic threshold, it is determined that there is a direct communication link between the two devices. The dynamic threshold is dynamically adjusted according to the current network load rate, and its adjustment rule is: the threshold decreases linearly when the load rate increases and increases exponentially when the load rate decreases.

[0088] Furthermore, a device adjacency matrix is ​​constructed based on the aforementioned communication link relationships. The element values ​​of the adjacency matrix are determined by the following rules: elements on the main diagonal are set to zero; device pairs with direct links are assigned their correlation coefficients; and device pairs without direct links are assigned negative infinity. Understandably, the adjacency matrix undergoes symmetry normalization processing. Specifically, each row element is divided by the sum of the non-negative elements in that row, and then the arithmetic mean is taken with the transpose matrix, ultimately generating a double-random matrix with a row sum of one. This double-random matrix, together with the device identifier mapping table, constitutes network topology data, and its physical meaning is the probability transition matrix for signal transmission between devices.

[0089] The network topology data will serve as the reference for normalization processing in S130. Its double random matrix characteristics ensure numerical stability during subsequent feature scaling, while the dynamic threshold mechanism adapts to the load fluctuation characteristics of the Internet of Things environment.

[0090] S130. Normalize the network topology data and preprocessed communication parameters to generate a standardized communication dataset.

[0091] Specifically, based on the double random matrix in the network topology data, the centrality weight of each device is calculated. The specific calculation process for the centrality weight is as follows: power-law iterative decomposition is performed on the matrix eigenvalues, and the principal eigenvector is taken as the device weight distribution. Further, based on the centrality weight, weighted normalization is performed on the preprocessed communication parameters. Specific implementation methods include:

[0092] Bandwidth parameter normalization: Divide the original bandwidth value by the base bandwidth of the communication subnet where the device is located, and then multiply by the centrality weight to obtain the relative bandwidth coefficient;

[0093] Throughput parameter normalization: The original throughput ratio is processed by logarithmic compression, mapped to the [0,1] interval, and then convolved with the centrality weights.

[0094] Delay parameter normalization: Based on the lowest delay in the entire network, calculate the reciprocal of the relative delay difference, and then superimpose the square root transformation result of the centrality weight.

[0095] Understandably, the normalized bandwidth coefficient, throughput coefficient, and delay coefficient are recombined according to timestamps to generate a standardized communication dataset. The data structure of the standardized communication dataset is a time-aligned sequence of <relative bandwidth, convolutional throughput, weighted delay> triples, whose numerical range is uniformly constrained within the interval [0,1].

[0096] The standardized communication dataset will serve as the input source for the dynamic graph construction in S210. Its centrality weighting mechanism highlights the parameter characteristics of key nodes, while the unified numerical range ensures the convergence efficiency of subsequent spatiotemporal convolution operations. Through the execution feedback data feedback in S630, the standardized communication dataset continuously updates historical records, forming a closed-loop optimization system.

[0097] Technical connection instructions between modules:

[0098] S110→S120: The time delay parameter sequence in the preprocessed communication parameters directly drives the calculation of network topology relationship, and its time alignment characteristics ensure the accuracy of correlation coefficient analysis;

[0099] S120→S130: The double random matrix in the network topology data serves as a normalization benchmark, and its device weight distribution determines the parameter scaling ratio.

[0100] S130→S210: The standardized communication dataset is called by S210 to construct a dynamic graph structure, and its normalized values ​​accelerate the feature extraction process of spatiotemporal convolution.

[0101] S600→S130: Update the historical record through the execution feedback data returned by S630, and dynamically optimize the weight allocation strategy for subsequent standardized processing.

[0102] Step S200 includes at least steps S210-S230:

[0103] S210. Obtain a standardized communication dataset and construct a dynamic graph structure with spatiotemporal correlation by combining it with historical fault records.

[0104] Specifically, a sequence of triples with timestamps, representing <relative bandwidth, convolutional throughput, weighted latency>, is extracted from the standardized communication dataset. This triple sequence is generated through normalization in S130, and its numerical range is constrained to the [0,1] interval. Further, a historical fault record database is queried. This database stores communication anomaly events of various devices over the past twelve months, including three fields: event occurrence time, anomaly type, and recovery measures. Understandably, the triple sequence is spatiotemporally aligned and matched with the historical fault records. The specific implementation method is as follows:

[0105] Time-dimensional correlation: Based on the moment of the fault event, extract the triple sequence within three seconds before and after as keyframes;

[0106] Spatial dimension correlation: Based on the network topology of the devices, select all node data within a two-hop range around the faulty device;

[0107] Feature label injection: Fault type labels are embedded in the triples of keyframes to form spatiotemporal data blocks with semantic annotations.

[0108] Furthermore, a dynamic graph structure is constructed based on the spatiotemporal data blocks. The node attributes of the dynamic graph structure include the triplet values ​​at the current moment, and the edge attributes include the signal strength between nodes and the number of historical co-occurring faults. The update frequency of the dynamic graph structure is synchronized with the data acquisition cycle, and the graph state of the previous three frames is retained during each update to achieve temporal continuity. The final output dynamic graph structure includes three data entities: topological connectivity, node feature matrix, and edge weight matrix. These data entities are serialized and encapsulated using the storage format of a graph database.

[0109] The dynamic graph structure serves as the input carrier for the spatiotemporal convolution operation in S220. Its labeled keyframe data provides supervision signals for subsequent feature extraction, while the dynamic update mechanism ensures the graph structure's real-time response to changes in network state. The historical fault record database is continuously updated based on the communication parameter changes collected in S620, forming a closed-loop optimization system.

[0110] S220. Perform spatiotemporal graph convolution operation on the dynamic graph structure to extract cross-layer correlation features and obtain spatiotemporal feature vectors.

[0111] Specifically, spatiotemporal graph convolution is performed on the dynamic graph structure, and the operation process is divided into two stages: spatial domain convolution and temporal domain convolution.

[0112] Spatial domain convolution stage. A multi-order neighborhood aggregation strategy is used to process the node feature matrix. The specific implementation of this strategy is as follows:

[0113] First-order neighborhood aggregation: The triplet value of the central node is taken as a weighted average of the values ​​of its directly connected neighboring nodes, and the weights are determined by the edge weight matrix.

[0114] Second-order neighborhood aggregation: extends to nodes within two hops of the central node, and adjusts the aggregation coefficient based on the number of historical co-occurring faults;

[0115] Residual connection: The original node features and the aggregation result are concatenated by channels to preserve the original information of the underlying layer.

[0116] Temporal convolution stage. The graph state changes across three consecutive frames are analyzed by sliding along the time axis. This analysis process includes:

[0117] The temporal gradient features are obtained by performing difference calculations on the feature values ​​of the same node at different times;

[0118] The edge weight change rate of adjacent frames is exponentially smoothed to obtain topological evolution features;

[0119] The fault labeling information is used as an attention mask to enhance the intensity of the characteristic response during abnormal periods.

[0120] Understandably, the convolution results in the spatial and temporal domains are fused using a feature fusion method: after performing element-wise addition on the two feature matrices, a non-linear activation function is used to map them to a high-dimensional space. The final output spatiotemporal feature vector is a 256-dimensional floating-point array, which is arranged in order of device node number, forming a feature matrix with a dimension of [number of nodes × 256].

[0121] The spatiotemporal feature vector will serve as the input data for compressed encoding in S230. Its multi-order neighborhood aggregation mechanism captures the implicit correlations between devices, while temporal difference analysis enhances the sensitivity of the features to dynamic changes in the network. Through feature dimension compression in S230, the high-dimensional feature matrix is ​​refined into a more discriminative low-dimensional representation.

[0122] S230. Perform feature dimension compression encoding on the spatiotemporal feature vector to generate a 32-dimensional compressed feature vector.

[0123] Specifically, a dual attention mechanism is used to perform compressed encoding on the spatiotemporal feature vector, which includes two branches: feature importance scoring and node influence scoring.

[0124] Feature importance scoring branch:

[0125] Calculate the mutual information between the 256-dimensional feature channels to generate a feature correlation matrix;

[0126] Singular value decomposition was performed on the correlation matrix, and the first 32 principal component directions were selected.

[0127] The original features are projected along the principal component direction to obtain primary compressed features.

[0128] Node Influence Rating Branch:

[0129] Calculate the propagation influence weight of each node based on the centrality index of the device in the network topology;

[0130] The weights are combined with the spatiotemporal feature vectors to perform a Hadamard product operation, highlighting the features of key nodes.

[0131] Global average pooling is performed on the weighted feature matrix to generate secondary compressed features.

[0132] Furthermore, the primary compression features and secondary compression features are cross-fused, and the fusion process is as follows:

[0133] Perform element-wise multiplication on the two types of feature vectors;

[0134] Feature scale is constrained by L2 normalization;

[0135] Weak response dimensions with absolute values ​​less than 0.01 are removed by sparsification.

[0136] The resulting 32-dimensional compressed feature vector contains both channel-level and node-level compressed information, and each dimension of the feature vector corresponds to a cross-layer association mode. The 32-dimensional compressed feature vector is encapsulated in a binary encoding format for use by the S310 for real-time communication coefficient calculation.

[0137] The 32-dimensional compressed feature vector will serve as the input data for S300. Its dual attention mechanism balances the relationship between feature selection and topological structure, while sparsity processing improves the compactness of feature representation. The effectiveness of the compressed feature vector is continuously verified through the real-time communication coefficient values ​​output by S330, forming a technical closed loop.

[0138] Technical connection instructions between modules:

[0139] S210→S220: The node feature matrix and edge weight matrix in the dynamic graph structure directly drive the spatial domain convolution operation, and its labeled keyframe data guides the attention allocation of the temporal domain convolution.

[0140] S220→S230: The 256-dimensional array of the spatiotemporal feature vector is used as the input of the dual attention mechanism, and its high-dimensional space mapping result is compressed and encoded for use;

[0141] S230→S310: The 32-dimensional compressed feature vector is called by S310 to load dynamic weight coefficients, and its sparse feature dimension accelerates the weighted fusion calculation.

[0142] S600→S210: Update historical fault records by collecting communication parameter change values ​​through S620, and optimize the labeling accuracy of subsequent dynamic graph construction.

[0143] Step S300 includes at least steps S310-S330:

[0144] S310. Obtain the 32-dimensional compressed feature vector and load the dynamic weight coefficients for the current device type.

[0145] Specifically, the 32-dimensional compressed feature vector is read from the distributed cache. This 32-dimensional compressed feature vector is generated through feature dimension compression encoding in S230, and its data structure is a floating-point array arranged in order of device node number. Further, the weight configuration library is queried based on the type identifier in the device registry. This weight configuration library stores dynamic weight coefficient combinations corresponding to different device types, including three dimensions: bandwidth weight coefficient, throughput weight coefficient, and latency weight coefficient. Understandably, a real-time calibration operation is performed on the dynamic weight coefficients. The specific implementation method is as follows:

[0146] Environmental factor compensation: Collect the current environmental noise intensity and signal interference index, and adjust the original weight values ​​through a preset compensation coefficient table;

[0147] Load balancing adaptation: Based on the load status of adjacent nodes in the network topology, the weight ratio is redistributed according to the inverse relationship of load.

[0148] Timeliness verification: Check the last update timestamp of the weight coefficients, and trigger a remote weight server to synchronize the update when the validity period threshold is exceeded.

[0149] The loading process of the dynamic weight coefficients adopts a double-buffering mechanism. While maintaining the stable operation of the current coefficients, the background preloads the calibrated coefficients for the next calculation cycle. The final output weight combination contains three sets of floating-point values ​​accurate to four decimal places, corresponding to the weighting ratios of bandwidth, throughput, and latency parameters, respectively.

[0150] The dynamic weighting coefficients will serve as the benchmark parameters for weighted fusion calculations in S320. Their environmental compensation mechanism ensures the adaptability of the coefficients to real-time network conditions, while the double-buffered loading method avoids sudden weight changes during the calculation process. Based on the communication parameter change values ​​collected by S620, the weight configuration library continuously optimizes the preset weight values ​​for various devices.

[0151] S320. Perform weighted fusion calculation on bandwidth parameters, throughput parameters, and latency parameters based on dynamic weighting coefficients.

[0152] Specifically, the bandwidth feature sub-vector, throughput feature sub-vector, and latency feature sub-vector contained in the 32-dimensional compressed feature vector are analyzed, and the dimension allocation ratio of the sub-vectors is 12:10:10. Further, each sub-vector is fused with its corresponding dynamic weight coefficient. This fusion process is divided into two stages: feature activation and cross-domain aggregation.

[0153] (1) Feature activation stage:

[0154] A linear transformation is performed on the bandwidth feature vector to map the 12-dimensional features to a 3-dimensional space. The result is obtained by multiplying the vector by the bandwidth weight coefficient.

[0155] Logarithmically scale the throughput feature subvector to compensate for long-tail distribution bias, and then perform a dot product operation with the throughput weight coefficient.

[0156] The time delay feature vector is input into the time decay function, and a timeliness correction factor is generated by combining the time delay weight coefficient.

[0157] (2) Cross-domain aggregation stage:

[0158] Spatially align the bandwidth score vector with the throughput dot product result, and generate a joint feature matrix through the Kronecker product.

[0159] The timeliness correction factor is used as a matrix element-level multiplier to adjust the numerical distribution of the joint feature matrix;

[0160] Perform singular value decomposition on the adjusted matrix and take the maximum singular value as the base fusion score.

[0161] The weighted fusion calculation employs a pipelined architecture, with the processing of the three sub-vectors executed in parallel and then synchronously merged in the aggregation phase. During the calculation process, the risk of numerical overflow at each stage is monitored in real time, and an automatic switch to a safe calculation mode is initiated when an outlier is detected. The final intermediate fusion result includes two sets of data: a base score and the contribution percentage of each sub-item.

[0162] The weighted fusion result will serve as the basis for calculating the coefficient values ​​output in S330. Its phased processing mechanism balances computational efficiency and accuracy requirements, while the safe computation mode ensures operational stability under extreme network conditions. Through the abnormal state determination in S510, the backpropagation gradient of the intermediate fusion result can be used to optimize the weight coefficient allocation strategy.

[0163] S330: Outputs real-time communication coefficient values ​​based on weighted fusion results.

[0164] Specifically, the weighted fusion result is standardized and quantified, and the process includes four key steps:

[0165] Dynamic range adjustment: Based on the statistical distribution in the historical communication coefficient database, the base score is linearly transformed to the [0,100] interval;

[0166] Contribution weighting: The transformation results are calibrated a second time according to the contribution ratio of each sub-item to highlight the influence of the dominant parameters;

[0167] Time smoothing filter: Employs an exponentially weighted moving average algorithm, combining the current calculation result with historical values ​​from the previous six periods;

[0168] Discretization output: The final value is rounded to one decimal place to generate a standard format real-time communication coefficient value.

[0169] The output interface for the real-time communication coefficient values ​​adopts a publish-subscribe model and supports both HTTP REST and MQTT protocols. Each coefficient value comes with a complete metadata description, including fields such as generation timestamp, data source device identifier, and computing node information. When network congestion is detected, a coefficient value compression transmission mode is automatically activated, transmitting only the difference data relative to the last update.

[0170] The real-time communication coefficient value will serve as input data for anomaly detection in S510. Its dynamic range adjustment ensures comparability between different device types, while time smoothing filtering suppresses misjudgments caused by instantaneous fluctuations. Through the adaptive threshold update mechanism in S430, the historical distribution characteristics of the coefficient value are continuously analyzed to optimize the threshold calculation model.

[0171] Technical connection instructions between modules:

[0172] S310→S320: The dynamic weighting coefficients serve as benchmark parameters to drive the weighted fusion calculation, and their double buffering mechanism ensures the smoothness of coefficient switching.

[0173] S320→S330: The weighted fusion result is processed through multiple stages to generate a standardized output, and its contribution analysis provides feedback for weight optimization.

[0174] S330→S510: The real-time communication coefficient value is used as the core criterion input to the anomaly detection module, and its time smoothing characteristics improve the robustness of the state determination.

[0175] S600→S310: The weight configuration library is optimized in reverse by using the execution feedback data collected by S620 to form a closed-loop learning system.

[0176] Step S400 includes at least steps S410-S430:

[0177] S410. Obtain the historical communication coefficient database and initialize the threshold calculation model using the meta-learning framework.

[0178] Specifically, the historical communication coefficient database is loaded from a distributed storage cluster. This database is continuously updated using execution feedback data from the S630, and its storage structure is a time-series data warehouse partitioned by device type. Further, the communication coefficient sequences in the database are parsed to extract three key features:

[0179] Statistical distribution characteristics: including mean, variance, and skewness coefficient within the sliding window;

[0180] Abnormal pattern characteristics: Mark the coefficient fluctuation trajectory within 20 seconds before and after historical failure events;

[0181] Environmental correlation characteristics: Correlate the network topology status and channel load indicators of the same period in storage.

[0182] Understandably, a meta-learning framework is used to construct the initialization structure of the threshold calculation model, and the framework includes the following components:

[0183] Meta-feature extractor: Encodes the three types of key features into a 128-dimensional meta-feature vector;

[0184] The basic learner consists of a three-layer fully connected network, with the input dimension matching the meta-feature vector;

[0185] Quick Adapter: Includes a pluggable parameter fine-tuning module that supports incremental learning.

[0186] The initialization process employs a hierarchical pre-training strategy: first, a meta-feature extractor is trained based on global data, and then the base learner is fine-tuned according to device type. Model parameters are transmitted encrypted using secure multi-party computation technology to ensure data privacy during cross-node collaboration. The final generated initial model contains general feature extraction capabilities and device-specific discrimination boundaries, which are then used by the S420 for collaborative training and optimization.

[0187] The threshold calculation model will serve as the baseline architecture for distributed training in S420. Its meta-feature encoding mechanism unifies the data representation across different devices, while the hierarchical pre-training strategy accelerates the subsequent convergence process. Through threshold update feedback in S430, the model's initialization parameters are continuously optimized to adapt to changes in the network environment.

[0188] In another implementation, the process of obtaining a historical communication coefficient database and initializing the threshold calculation model using a meta-learning framework includes the following steps:

[0189] Specifically, historical communication coefficient database is loaded from the distributed storage cluster. The database is continuously updated through execution feedback data from S630, and its storage structure is as follows:

[0190]

[0191] in, For the device index number ( ), Indicates the total number of devices; For equipment of Time-series vector of communication coefficients (generated by S330); This is a three-dimensional environmental feature vector (elements represent network topology load rates in order). Channel interference index Noise intensity ); For historical fault labels (0 indicates normal state, 1 indicates abnormal state).

[0192] Furthermore, an improved Riemannian manifold element learning algorithm (RML) is used to construct the model initialization framework: ① Manifold space projection:

[0193]

[0194] in, For manifold coordinates; For time window indexing, This represents the total number of windows. For window Internal communication coefficient Covariance matrix; The mean is the Fréchet mean. This is the time decay weight.

[0195] The formula outputs the manifold coordinates. This will serve as the core input to Formula ②, enabling the transformation from the original communication coefficients to temporal geometric features. This formula maps the dynamic communication coefficients to a symmetric positive definite manifold space, capturing the geometric evolution characteristics of the device state.

[0196] ② Meta-feature encoding:

[0197]

[0198] in, It is the meta-eigenvector; For a three-channel feature fusion processor (fully connected neural network); Information entropy of environmental characteristics; For fault labeling, a graph neural network embedding layer; This indicates a tensor splicing operation.

[0199] Output As equipment The meta-feature vector is used to initialize the basic parameters of the threshold model. Through collaborative training with S420, It will become the core carrier for cross-device knowledge transfer.

[0200] Generated meta-feature vectors The model training (Formulas ③ and ④) will be directly used for S420, and its encoding process integrates three key types of information: historical equipment status, environmental characteristics, and fault labels.

[0201] S420. Cross-node collaborative training of the threshold calculation model based on the historical communication coefficient database.

[0202] Specifically, the initial model is deployed to an edge computing node cluster, with each node partitioning the training data into geographical regions. The collaborative training process is executed in three phases:

[0203] (1) Local training phase. Historical communication coefficient data are synchronously acquired from multiple network nodes:

[0204] The node loads historical data matching the device type from local storage and divides the training set and validation set in a 7:3 ratio.

[0205] Perform time series augmentation operations on the training set to generate noisy replica data to improve robustness;

[0206] A dynamic learning rate strategy is used to update model parameters, and an early stopping mechanism is used to evaluate the validation set after each round of training.

[0207] (2) Gradient exchange stage: Aggregate distributed parameter gradients while protecting data privacy.

[0208] Nodes transmit parameter gradients through a secure aggregation protocol, which employs homomorphic encryption and differential privacy technology.

[0209] The master node performs outlier detection and normalization on the gradient, and generates the global gradient update direction;

[0210] The model's expressive power in sparse regions is protected by weighted fusion gradients based on the amount of node data.

[0211] (3) Model aggregation stage. Update the global parameters of the threshold calculation model:

[0212] The federated averaging algorithm is applied to the parameters of the base learner to update the globally shared low-level feature representation;

[0213] Preserve the local-specific parameters of the fast adapter to adapt to differences in regional network environments;

[0214] Verify the generalization performance of the aggregated model on the cross-regional test set and trigger the hierarchical retraining mechanism.

[0215] The collaborative training employs an asynchronous parallel architecture, allowing nodes to train locally offline and synchronize with the latest global model upon going online. Model bias is monitored in real time during training, and regional weight coefficients are automatically adjusted when data distribution drift is detected. The resulting collaborative model possesses environmental adaptability and cross-device generalization, providing the inference basis for threshold generation on the S430.

[0216] The collaborative training mechanism serves as the core component for continuous model optimization. Its secure aggregation protocol ensures data privacy, while its asynchronous architecture adapts to the heterogeneity of edge computing scenarios. Using newly collected data from the S620, the training process forms a closed-loop iterative optimization.

[0217] In another implementation, the process of cross-node collaborative training of the threshold calculation model based on the historical communication coefficient database includes the following steps:

[0218] Specifically, in Threshold calculation model for deploying an edge node cluster , This is the mathematical representation of the threshold calculation model. Where: θ is the model parameter vector, optimized through S420 co-training; A 128-dimensional eigenvector generated for S410; function mapping f: →[0,1], output device status confidence;

[0219] Furthermore, an improved Möbius federated aggregation protocol is adopted:

[0220] ③ Hyperbolic parameter aggregation:

[0221]

[0222] in, For the first Global model parameters after rounds of federated training; The coordinates of the origin in hyperbolic space; For node indexing, For training rounds; Weights for distribution perception; Represents a node No. The local model parameters are trained in rounds. This formula achieves parameter aggregation in hyperbolic space, overcoming the heterogeneity limitations of Euclidean space.

[0223] ④ Privacy gradient updates:

[0224]

[0225] in, The model parameter increment in the t-th training round; For gradient clipping operators;

[0226] For nodes Model gradient ( For the gradient operator in the model parameter space; The loss function; (Pseudo-labels generated by node j) It is Gaussian noise.

[0227] The training employs an alternating optimization strategy: first, federated aggregation updates the global parameters. Then fine-tune using local data. This allows formulas ③ and ④ to form a privacy-precision balanced closed loop. Among them, the gradient... The calculations directly depend on the S410 generated. And noise The injection ensures the safety of the threshold output by formula ⑤.

[0228] S430: Output adaptive judgment threshold and update it to the real-time threshold library.

[0229] Specifically, the trained threshold calculation model is deployed to an online inference engine, which receives real-time communication coefficient values ​​generated by the S330 as input. The threshold generation process includes the following steps:

[0230] (1) Dynamic reasoning stage:

[0231] Extract sliding window features from the input coefficient sequence and perform similarity matching with historical fault patterns;

[0232] Load the corresponding model branch based on the device type and perform real-time forward propagation calculations;

[0233] Output the basic threshold estimate and confidence score, which reflects the model's grasp of the device's state.

[0234] (2) Environmental calibration phase:

[0235] Obtain the current network topology status and channel load indicators, and adjust the basic thresholds using the environmental compensation matrix;

[0236] Perform a consistency check on the thresholds of adjacent devices to eliminate spatially abnormal points;

[0237] A time-smoothing filter is applied to fuse the current calculation results with the historical thresholds from the previous six periods.

[0238] (3) Threshold release stage:

[0239] The calibrated thresholds are written into a real-time threshold library by device identifier, and the library adopts a memory-disk dual storage architecture.

[0240] Generate a threshold change log to record the adjustment range, triggering factors, and affected devices;

[0241] Update notifications are pushed to the anomaly detection module of S510 via a publish-subscribe model.

[0242] The adaptive threshold adopts a progressive update strategy: it takes effect immediately when the confidence score is higher than 0.9, and triggers a manual review process when it is lower than 0.6. The threshold library version management adopts a multi-timeline snapshot mechanism to support rapid rollback operations in abnormal situations. The final output threshold data includes four required fields: device identifier, threshold value, effective time, and confidence score.

[0243] The adaptive threshold will serve as the core basis for anomaly detection in S510. Its dynamic inference mechanism enables personalized threshold setting, while environmental calibration ensures global network consistency. Through the switching command feedback from S530, the threshold generation strategy forms a closed-loop verification system.

[0244] In another implementation, the process of outputting an adaptive decision threshold and updating it to the real-time threshold library includes:

[0245] Specifically, based on a fully trained model Generate device-level dynamic thresholds. Among them, To train the optimal threshold calculation model after convergence:

[0246] ⑤ Topology-adaptive threshold:

[0247]

[0248] in, For equipment The original threshold output; For equipment The average communication coefficient of neighboring nodes; The Sigmoid confidence score of the model output; This is a quantification term for dynamic topological changes; This is the starting time of the current analysis; Barcode entropy; For continuous homology mapping; It is a time series.

[0249] ⑥ Threshold smoothing output:

[0250]

[0251] in, This represents the final release threshold for the device (dev). The cache threshold for device dev; It is a dynamic smoothing factor; This is the exponential moving average operator.

[0252] The final threshold is appended to the tuple when written to the real-time library. (Device ID, effective time, confidence level, data version) are pushed to the anomaly detection module of S510 via an incremental synchronization protocol. The threshold sequence processed by formula ⑥ can significantly suppress network oscillations in S520 state simulation.

[0253] Technical connection instructions between modules:

[0254] S410→S420: The initialization parameters of the threshold calculation model serve as the starting point for collaborative training, and its meta-feature encoder ensures cross-node feature consistency.

[0255] S420→S430: The model parameters generated by the collaborative training drive the online inference engine, whose federated learning architecture adapts to the needs of distributed deployment;

[0256] S430→S510: The adaptive judgment threshold is used as a dynamic benchmark input for anomaly detection, and its environmental calibration result affects the state judgment boundary.

[0257] S600→S410: Update the historical database using the execution feedback data returned from S630 to optimize the quality of subsequent model initialization.

[0258] Step 500 includes at least steps S510-S530:

[0259] S510: Obtain the real-time communication coefficient value and adaptive judgment threshold, and perform abnormal state judgment.

[0260] Specifically, the real-time communication coefficient value is read from the distributed message queue. This coefficient value is generated through weighted fusion calculation in S330, and its data structure includes three mandatory fields: device identifier, timestamp, and coefficient value. Simultaneously, the adaptive judgment threshold for the corresponding device type is obtained by querying the real-time threshold library. This threshold is generated through model inference in S430 and includes three attributes: basic threshold, confidence score, and effective time. Further, an abnormal state judgment operation is performed, specifically implemented as follows:

[0261] Time series comparison analysis: Extract the communication coefficient sequence of the target device for the most recent five periods and calculate its cumulative deviation value relative to the threshold;

[0262] Spatial consistency verification: Obtain the communication coefficients of adjacent devices in the same area, establish a local spatial distribution model, and detect whether the target device is a statistical outlier;

[0263] Pattern matching detection: The current coefficient fluctuation trajectory is dynamically time-normalized and matched with the abnormal patterns in the historical fault database to calculate the similarity score.

[0264] Understandably, a three-level judgment strategy is adopted for comprehensive decision-making: when the time sequence deviation exceeds a preset threshold and the spatial outlier probability is greater than 90%, a primary anomaly labeling is triggered; when the pattern similarity score exceeds a dynamic threshold, it is upgraded to a confirmed anomaly state; when all three indicators exceed the limits simultaneously, it is immediately labeled as an emergency anomaly. The judgment result includes three types of metadata: anomaly level, confidence level, and associated fault type, which are available for use by the S520's state simulation module.

[0265] The abnormal state determination result will serve as the trigger condition for network state simulation. Its temporal-spatial joint analysis reduces the false judgment rate, while the multi-level determination strategy realizes hierarchical early warning of abnormalities. The abnormal patterns in the historical fault database are continuously updated and optimized based on the parameter change values ​​collected after the switchover by the S620.

[0266] S520: Simulate the communication status after base station handover for abnormal devices.

[0267] Specifically, a digital twin network environment is constructed for the abnormal device, and the environment includes the following elements:

[0268] Constructing a digital twin network environment for the target base station:

[0269] Base station topology model: Reconstruct the coverage relationship between the target base station and candidate base stations based on network planning data;

[0270] Channel state mirroring: replicates real-time parameters such as the current channel quality indicator and reference signal received power;

[0271] Device Behavior Simulator: Loads the historical communication patterns of abnormal devices as a benchmark for behavior prediction.

[0272] Further, a base station handover simulation operation is performed, the specific process of which is as follows:

[0273] Predict the real-time communication coefficient change after base station handover.

[0274] Candidate base station evaluation: Calculate the signal strength attenuation coefficient, load balancing index, and historical handover success rate of each candidate base station;

[0275] Virtual handover execution: Simulate signaling processes in a digital twin environment and record virtual handover latency and link reconstruction time;

[0276] Effect prediction analysis: Based on the simulation results, predict the change curve of the communication coefficient after the switch and calculate the steady-state improvement.

[0277] The simulation process employs the Monte Carlo method with multiple iterations, introducing random noise in each iteration to simulate wireless channel fluctuations. Statistical analysis of the output results generates three key indicators: predicted handover success probability, expected coefficient improvement range, and optimal handover time window. When the simulation results indicate multiple base stations are available, a reinforcement learning-based intelligent decision-making algorithm is triggered to select the optimal target.

[0278] When the predicted value exceeds the current anomaly threshold, an instruction is generated. The base station handover simulation results serve as the basis for instruction generation. Its digital twin environment enables lossless testing of the real network, while the Monte Carlo method improves the reliability of the prediction results. The parameters of the simulation model are continuously calibrated and optimized using the actual handover delay data recorded by S610.

[0279] S530 generates base station handover instructions or channel reset instructions based on the prediction results.

[0280] Specifically, the predicted indicators output by the simulation analysis are parsed, and control instructions are generated according to the following rules:

[0281] Base station handover command generation: When the expected coefficient increases beyond the threshold and the success probability is greater than 95%, a handover command containing the target base station ID, handover time window, and frequency parameters is packaged.

[0282] Channel reset command generation: When base station handover is not feasible but channel quality fluctuates drastically, a reset command containing a new modulation and coding scheme and power adjustment parameters is generated;

[0283] Hybrid strategy generation: It can simultaneously generate switching and reset combination commands for multi-module devices, and set execution priority and timing relationship.

[0284] The instruction generation process incorporates a dual verification mechanism: first, digital signatures ensure instruction integrity, and then a time-sensitive hash lock is added to prevent replay attacks. The generated instruction message adopts a layered encoding structure, including:

[0285] Control header: Command type, validity period, priority identifier;

[0286] Parameter body: Specific configuration parameters and execution constraints;

[0287] Verification tail: A digital fingerprint generated based on the device key.

[0288] The final output command is synchronously sent to the target device and the base station controller via a multipath transmission protocol, with a transmission confirmation and execution status feedback mechanism set up. If the command is not confirmed for execution within a preset time, a compensation strategy is automatically triggered to generate an alternative solution.

[0289] The control commands will be executed by S610, and their dual verification mechanism ensures command security, while multi-path transmission improves delivery reliability. Through the execution feedback data returned by S630, the command generation strategy forms a closed-loop optimization system.

[0290] Technical connection instructions between modules:

[0291] S510→S520: The abnormal state determination result triggers digital twin simulation, and its multi-level determination criteria affect the initial parameter settings of the simulation;

[0292] S520→S530: The simulation prediction index directly drives the instruction generation logic, and its Monte Carlo analysis results determine the instruction type and parameters;

[0293] S530→S610: The control instructions are input to the execution unit, and their hierarchical coding structure is adapted to the protocol stack requirements of different devices;

[0294] S600→S510: Optimize the anomaly detection threshold using actual switching data collected by S620 to improve the accuracy of subsequent detection.

[0295] Step S600 includes at least steps S610-S630:

[0296] S610: Execute base station handover command or channel reset command, and record execution delay and success rate.

[0297] Specifically, the base station handover command or channel reset command is received through the IoT device management platform. This command is generated by the S530 and verified by a digital signature. The execution process consists of three stages:

[0298] (1) Command parsing stage: Extract the operation type identifier from the command header. When the identifier is base station handover, parse the target base station ID, frequency parameters, and time window constraints; when the identifier is channel reset, parse the modulation and coding scheme, power adjustment parameters, and effective time point. Verify the validity of the command hash lock to ensure that the command has not been tampered with and is within its validity period.

[0299] (2) Resource scheduling phase: A resource reservation request is initiated to the base station controller. The request includes the device identifier, quality of service requirements, and priority flag. For base station handover operations, the following sub-steps are executed synchronously:

[0300] Query the load status and signal coverage map of the target base station;

[0301] Negotiate and switch execution times to avoid channel congestion;

[0302] Pre-allocate wireless resource blocks and establish a transport context.

[0303] (3) Instruction execution phase:

[0304] For base station handover commands, the handover signaling process is triggered within the time window agreed upon in the protocol, and the total time from sending the handover command to receiving the handover completion confirmation is recorded;

[0305] For channel reset commands, adjust physical layer parameters according to the parameter configuration table, and record the configuration delay from sending the reset command to receiving confirmation of parameter effectiveness;

[0306] The system monitors the execution results of commands. When a successful response is received from the base station or device, the command is marked as successful. If no response is received within a timeout period or a failure code is received, the command is marked as failed.

[0307] The execution record includes four mandatory fields: operation type, start timestamp, end timestamp, and result status, and is transmitted to the data analysis node via an encrypted channel. Execution latency is calculated to the millisecond level, and a network time protocol is used to synchronize the clocks of each node. The accuracy of the execution results is cross-validated by real-time parameter changes collected by the S620.

[0308] S620 collects changes in bandwidth, throughput, and latency parameters, and generates execution feedback data.

[0309] Specifically, during the monitoring period following command execution, changes in key parameters are collected using the following methods:

[0310] Bandwidth parameter acquisition:

[0311] Deploy probes at the application layer to count the amount of payload data successfully transmitted per unit time.

[0312] Record the uplink and downlink directions separately, and calculate the average bidirectional bandwidth.

[0313] By comparing the bandwidth change rate before and after instruction execution, a relative improvement index is generated.

[0314] Throughput parameter collection:

[0315] Analyze the transport layer protocol header and calculate the retransmission rate and the percentage of valid data packets;

[0316] The throughput efficiency index is calculated using the formula: (1 - retransmission rate) × percentage of valid data packets.

[0317] A time series model is established, and the influence of inherent network fluctuations is removed to extract the trend of parameter changes.

[0318] Delay parameter acquisition:

[0319] End-to-end round-trip delay is measured using hardware timestamps;

[0320] Perform percentile statistics on the time delay sequence and record the P50, P90, and P99 percentile values;

[0321] Perform outlier filtering to eliminate measurement bias caused by sudden interference.

[0322] Furthermore, the collected raw parameters are spatiotemporally aligned and matched with the execution records to generate structured feedback data, including:

[0323] Execution context: referenced instruction ID, device identifier, operation type;

[0324] Parameter change matrix: bandwidth change rate, throughput efficiency increment, and latency percentile change;

[0325] Environmental status snapshot: network topology, channel quality, and interference index at the time of data collection.

[0326] After passing data integrity verification, the feedback data is marked as valid samples and stored in a temporary buffer. When an abnormal parameter acquisition is detected, a supplementary measurement mechanism is automatically triggered to complete the data. Through the S630's reflux operation, the feedback data becomes a key input to the optimization model.

[0327] S630, The execution feedback data is fed back to the historical communication coefficient database and the standardized communication dataset.

[0328] Specifically, the feedback data is processed as follows before being written to persistent storage:

[0329] Data cleaning stage:

[0330] To verify the continuity of the time series, neighbor interpolation was used to fill in the missing time periods;

[0331] Detect and correct systematic biases caused by sensor drift;

[0332] Remove abnormal records that exceed the theoretical limits of the physical layer.

[0333] Feature encoding stage:

[0334] Convert the parameter change matrix to the same standardized format as S130;

[0335] The execution context information is feature-hash encoded to generate a device-operation joint identifier;

[0336] Extract key factors from the environmental state snapshot and construct dynamic graph node attributes compatible with S210.

[0337] Distributed storage phase:

[0338] The standardized communication data is written into the partition table of the standardized communication dataset and stored in segments according to device type and time range;

[0339] After associating the performance metrics with the original instructions, update the case library in the historical communication coefficient database;

[0340] Cluster analysis is performed on abnormal patterns in the feedback data to generate new fault feature templates, which are then stored in the knowledge graph.

[0341] The reflow operation employs an eventual consistency model, allowing asynchronous data replication across geographically distributed nodes. Data updates are managed through a version control mechanism, supporting historical status queries by time point. Storage load is monitored in real-time during the reflow process, automatically downgrading to incremental update mode when system overload is detected. The updated data drives model retraining in S410 and normalization parameter adjustment in S130, forming a closed-loop learning system.

[0342] Technical connection instructions between modules:

[0343] S610→S620: The timestamp and result status in the execution record serve as the time reference and validity criterion for parameter collection;

[0344] S620→S630: The environmental state snapshot of the feedback data dynamically updates the graph structure construction rules of S210;

[0345] S630→S410 / S130: The standardized data after reflow is directly used in model training and normalization parameter calculation;

[0346] S500→S610: The instruction signature information generated in S530 is used in the S610 stage to verify the integrity of the instruction.

[0347] Example 2: Figure 2 A structural block diagram of an Internet of Things (IoT)-based communication management system according to an embodiment of the present invention is shown. Figure 2 As shown, the structure may include:

[0348] The data acquisition unit 10 is configured to collect bandwidth parameters (data transmission volume per unit time), throughput parameters (percentage of effective data packets), and latency parameters (end-to-end round-trip latency) of IoT devices via the built-in probes of the cellular module. It performs timestamp alignment and missing value imputation operations to generate preprocessed communication parameters. Furthermore, it extracts the device network topology from the preprocessed communication parameters and constructs an adjacency matrix to characterize the signal coupling strength between devices. Finally, it performs normalization processing on the network topology data and preprocessed communication parameters to output a standardized communication dataset. This unit is connected to the feature fusion unit via a distributed data bus to ensure that the data transmission latency is less than 50 milliseconds.

[0349] The feature fusion unit 20 is configured to receive the standardized communication dataset, construct a spatiotemporally correlated dynamic graph structure by combining it with a historical fault record database, wherein node attributes include real-time communication parameters of the devices, and edge attributes include signal strength attenuation coefficients; perform spatiotemporal graph convolution operations on the dynamic graph structure, extract cross-layer correlation features using a multi-order neighborhood aggregation strategy, and generate a 256-dimensional spatiotemporal feature vector; further, perform feature dimension compression encoding through an autoencoder network to output a 32-dimensional compressed feature vector. This unit integrates a lightweight graph neural network accelerator, achieving a processing speed of 100,000 devices / second.

[0350] The coefficient calculation unit 30 is configured to load a dynamic weight coefficient matrix for the current device type, which is updated hourly using a quantum evolution algorithm. Based on the 32-dimensional compressed feature vector, it analyzes bandwidth feature sub-vectors, throughput feature sub-vectors, and latency feature sub-vectors. It then performs a weighted fusion calculation on the three types of sub-vectors according to the dynamic weight coefficients to generate real-time communication coefficient values. The unit supports a dynamic weight hot-switching mechanism, completing coefficient reconfiguration within 50 milliseconds when the device type changes.

[0351] The threshold generation unit 40 is configured to obtain a historical communication coefficient database from a distributed storage cluster, initialize the basic architecture of the threshold calculation model using a meta-learning framework, perform cross-node collaborative training on the edge node cluster, and achieve secure exchange of parameter gradients through the Möbius federated aggregation protocol; output an adaptive judgment threshold and write it into a real-time threshold library, attaching an effective timestamp and confidence label. This unit integrates a differential privacy module with a privacy protection budget ε ≤ 0.1.

[0352] The decision execution unit 50 is configured to compare real-time communication coefficient values ​​with an adaptive judgment threshold. When an abnormal state is detected, it triggers a digital twin simulation engine; constructs a virtual network environment for the target base station; predicts the trajectory of communication coefficient changes after base station handover or channel reset; and generates control instructions containing the target base station ID, frequency parameters, and execution time window. The unit's decision delay is controlled within 300 milliseconds, and the prediction accuracy reaches 92%.

[0353] The feedback learning unit 60 is configured to monitor the execution process of base station handover or channel reset commands, record execution latency and success rate indicators; collect changes in bandwidth, throughput, and latency parameters after the command takes effect; and feed the execution feedback data back to the historical communication coefficient database and standardized communication dataset via an incremental synchronization protocol. The unit employs a dual-channel verification mechanism, ensuring a data return integrity rate of ≥99.9%.

[0354] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A communication management method based on the Internet of Things, characterized in that, include: Collect bandwidth, throughput, and latency parameters from IoT devices, perform timestamp alignment and topology construction, and generate a standardized communication dataset. A dynamic graph structure is constructed using a standardized communication dataset, and a 32-dimensional compressed feature vector is generated through spatiotemporal convolution and compression coding operations. Based on a 32-dimensional compressed feature vector, the bandwidth parameter, throughput parameter, and latency parameter are weighted and fused according to dynamic weight coefficients to output the real-time communication coefficient value. Obtain a historical communication coefficient database, collaboratively train a threshold calculation model through a meta-learning framework, and output an adaptive judgment threshold to a real-time threshold library. The expression for the collaborative training threshold calculation model of the meta-learning framework is: in, It is the meta-eigenvector; For a three-channel feature fusion unit; For manifold coordinates; Information entropy of environmental characteristics; For fault labeling, a graph neural network embedding layer; This indicates a tensor concatenation operation; Label historical faults; The expression for the output adaptive judgment threshold to the real-time threshold library is: in, This represents the final release threshold for the device (dev). It is a dynamic smoothing factor; The cache threshold for device dev; For exponential moving average operators; Output the original threshold; Compare the real-time communication coefficient value with the adaptive judgment threshold, perform network status simulation operation, and generate base station handover command or channel reset command; Execute base station switching commands or channel reset commands, collect communication parameter change values ​​to generate execution feedback data, and feed it back to the historical communication coefficient database and standardized communication dataset.

2. The method according to claim 1, characterized in that, The generation of standardized communication datasets includes: Obtain bandwidth parameters, throughput parameters, and latency parameters, perform timestamp alignment and missing value filling to obtain preprocessed communication parameters; The network topology relationships of devices are extracted from the preprocessed communication parameters, and an adjacency matrix is ​​constructed to obtain network topology data. Normalization is performed on the network topology data and preprocessed communication parameters to generate a standardized communication dataset.

3. The method according to claim 1, characterized in that, Generating a 32-dimensional compressed feature vector includes: By combining historical fault records with standardized communication datasets, a dynamic graph structure with spatiotemporal correlation is constructed. Spatiotemporal graph convolution is performed on the dynamic graph structure to extract cross-layer correlation features and obtain spatiotemporal feature vectors; The spatiotemporal feature vectors are compressed and encoded to generate 32-dimensional compressed feature vectors.

4. The method according to claim 1, characterized in that, The output real-time communication coefficient values ​​include: Load the dynamic weighting coefficients for the current device type; The bandwidth parameter, throughput parameter, and latency parameter are weighted and fused according to the dynamic weighting coefficients. The real-time communication coefficient value is output based on the weighted fusion calculation results.

5. The method according to claim 1, characterized in that, The output adaptive decision thresholds include: The threshold calculation model is initialized using a meta-learning framework; Cross-node collaborative training is performed on the threshold calculation model based on the historical communication coefficient database; Output adaptive judgment threshold and update it to the real-time threshold library.

6. The method according to claim 5, characterized in that, Cross-node collaborative training includes: Historical communication coefficient data are obtained synchronously from multiple network nodes; Aggregate distributed parameter gradients while protecting data privacy; Update the global parameters of the threshold calculation model.

7. The method according to claim 1, characterized in that, Generating base station handover commands or channel reset commands includes: Anomaly determination is performed based on the real-time communication coefficient value and the adaptive determination threshold. Simulate the communication status of abnormal devices after base station handover; Generate base station handover commands or channel reset commands.

8. The method according to claim 7, characterized in that, Communication status simulation includes: Construct a digital twin network environment for the target base station; Predict the real-time communication coefficient changes after base station handover; Instruction generation is triggered when the predicted value is higher than the current anomaly threshold.

9. The method according to claim 1, characterized in that, The data collected includes changes in communication parameters to generate execution feedback data, including: Record instruction execution latency and success rate; Collect the changed values ​​of bandwidth parameters, throughput parameters, and latency parameters; The changed values ​​are written into the historical communication coefficient database and the standardized communication dataset.

10. A communication management system based on the Internet of Things (IoT), applied to the communication management method based on the IoT as described in any one of claims 1-9, characterized in that, include: The data acquisition unit is configured to collect bandwidth parameters, throughput parameters, and latency parameters and generate a standardized communication dataset. The feature fusion unit is configured to construct a dynamic graph structure and output a 32-dimensional compressed feature vector. The coefficient calculation unit is configured to calculate the real-time communication coefficient value based on the dynamic weight coefficient. The threshold generation unit is configured to train a model using a meta-learning framework and output an adaptive decision threshold. The decision execution unit is configured to perform communication status simulation operations and generate base station switching or channel reset commands. The feedback learning unit is configured to collect instruction execution data and feed it back to the historical communication coefficient database and the standardized communication dataset.

Images