A plant 5g coverage automatic optimization method and system based on internet of things

Abstract

This invention discloses an automatic optimization method and system for 5G coverage in factory areas based on the Internet of Things (IoT). To address the problems of existing automatic parameter adjustments lacking safety boundaries, easily affecting production and business continuity, and lacking automatic verification and rollback of optimization effects, this invention collects measurement data from IoT terminals in the factory area and obtains a snapshot library of initial wireless configuration parameters and stable parameters. It constructs a heterogeneous dynamic spatiotemporal graph sequence representing the relationship between terminals and 5G cells. Using a pre-trained heterogeneous dynamic spatiotemporal graph neural network, it outputs predicted values ​​and uncertainties of business performance indicators and calculates lower confidence boundaries to determine business continuity safety conditions. Under feasibility discrimination constraints and change magnitude constraints, it uses constrained Bayesian optimization to generate a set of candidate wireless configuration parameters. Based on the experimental cell group and the control cell group, it performs a comparative sequential statistical test to form verification results. Based on the verification results, it performs hierarchical rollback or writes to stable snapshots. This achieves the technical effects of automatically optimizing 5G coverage parameters while ensuring business continuity, rapidly verifying optimization effects, and automatically recovering in case of anomalies.

Description

Technical Field

[0001] This invention relates to the field of mobile communication network optimization, and in particular to an automatic optimization method and system for 5G coverage in factory areas based on the Internet of Things. Background Technology

[0002] With the integration of 5G mobile communication technology and the Industrial Internet, an increasing number of manufacturing plants are building dedicated 5G mobile communication networks to support data collection, equipment control, machine vision, and mobile operations from industrial IoT terminals. Factory environments typically feature numerous metal obstructions, strong reflections, frequent personnel and vehicle movement, and rapid changes in production line status, leading to dynamic variations in wireless propagation conditions over time. This can result in localized weak coverage, handover anomalies, and fluctuations in service performance. To improve coverage quality and service experience, existing technologies generally rely on network management servers to continuously collect measurement data and service performance indicators reported by IoT terminals and optimize wireless configuration parameters such as cell transmit power, beam configuration, and handover parameters.

[0003] Existing network optimization technologies have evolved from manual parameter tuning based on experience to automation and intelligence. Early methods primarily relied on on-site surveys and engineering experience for static parameter configuration. Later, self-organizing network optimization mechanisms based on rule bases and threshold triggers were introduced for scenarios such as coverage optimization and handover optimization. In recent years, with enhanced data acquisition and computing capabilities, some solutions have begun to utilize machine learning models to predict service performance and automatically generate candidate configuration schemes using parameter search strategies. These schemes are then gradually rolled out after verification in pilot cells, reducing manual intervention and improving optimization efficiency.

[0004] However, existing technologies still have the following shortcomings in factory production scenarios:

[0005] 1) Automatic parameter tuning lacks safety boundary constraints oriented towards business continuity and does not adequately consider the uncertainties of prediction and assessment, which may lead to unacceptable performance degradation of critical production operations during the parameter tuning process.

[0006] 2) The verification mechanism after the optimization plan takes effect is not perfect. Common practices rely on offline evaluation or post-event statistical analysis. There is a lack of online comparison verification and rapid judgment methods between test and control areas, making it difficult to detect and block negative impacts in a timely manner.

[0007] 3) The lack of automated rollback and stable configuration management mechanisms means that manual rollback is often required when optimization results are not as expected or when anomalies occur. Recovery takes a long time and is not conducive to ensuring the continuity of production and business operations.

[0008] Therefore, there is a need for an automatic optimization method and system for 5G coverage in factory areas that can address the shortcomings of existing technologies. Summary of the Invention

[0009] One objective of this invention is to propose an automatic optimization method for 5G coverage in factory areas based on the Internet of Things (IoT). Addressing the problems of existing technologies where automatic parameter adjustment lacks a safety boundary oriented towards business continuity, insufficient consideration of predictive uncertainty leading to disruptions in production and business continuity, and the lack of automatic verification and rollback mechanisms after optimization, the following technical solution is proposed: Collecting measurement data reported by IoT terminals in the factory area and obtaining the service level or service type; reading the initial radio configuration parameter set and stable parameter snapshot library of 5G cells; constructing a heterogeneous dynamic spatiotemporal graph sequence representing the relationship between IoT terminals and 5G cells based on the measurement data; calling a pre-trained heterogeneous dynamic spatiotemporal graph neural network model to output the predicted mean and uncertainty of service performance indicators; determining the confidence coefficient and calculating the lower confidence bound based on the service level; comparing the lower confidence bound with the service continuity threshold vector to form a safety judgment; generating a candidate radio configuration parameter set using constrained Bayesian optimization under feasibility discrimination constraints and change magnitude constraints; performing a comparative sequential statistical test based on the test cell group and the control cell group to form verification results; and implementing graded rollback or writing the candidate parameters into the stable parameter snapshot library based on the verification results. This invention has the technical effect of automatically optimizing 5G coverage parameters while ensuring business continuity, quickly verifying the optimization effect, and automatically recovering in case of anomalies.

[0010] This invention provides an automatic optimization method for 5G coverage in factory areas based on the Internet of Things, comprising:

[0011] S1. Collect IoT measurement data reported by IoT terminals within the factory area, including service level or service type, and obtain the initial wireless configuration parameter set and stable parameter snapshot library corresponding to the 5G cell; S2. Construct a heterogeneous dynamic spatiotemporal graph sequence representing the IoT terminal and the 5G cell and their correlation based on the IoT measurement data; S3. Call the pre-trained heterogeneous dynamic spatiotemporal graph neural network model to infer the heterogeneous dynamic spatiotemporal graph sequence, obtain the predicted value and uncertainty of the service performance index, determine the confidence coefficient according to the service level, calculate the lower confidence bound, and compare the lower confidence bound with the preset service continuity threshold vector to determine the service continuity security condition; S4. Use constrained Bayesian optimization iteration to generate the set of wireless configuration parameters to be evaluated, and apply feasibility discrimination constraints and change magnitude constraints. Call the pre-trained feasibility classifier to output the feasibility probability. The feasibility discrimination constraint includes that the feasibility probability is not less than the preset feasibility threshold, and the change magnitude constraint includes that the parameter change amount is not greater than the preset change magnitude threshold. For the set of parameters to be evaluated that meet the constraints, the set of parameters to be evaluated is evaluated. S5. Evaluate the lower confidence bound of the wireless configuration parameter set and determine whether it meets the service continuity security conditions. Determine the candidate wireless configuration parameter set based on the acquisition function. S6. Determine the test cell group based on the candidate wireless configuration parameter set and the control cell group based on the initial wireless configuration parameter set. Send the candidate wireless configuration parameter set to the test cell group while the control cell group maintains the initial wireless configuration parameter set. Collect the corresponding IoT measurement data during the verification period, perform a comparative sequential statistical test to obtain the sequential test result, and determine the compliance of the service performance indicators of the test cell group with the service continuity threshold vector to obtain the compliance judgment result, thus forming the verification result. S7. Determine the rollback level based on the verification result. When the sequential test result indicates failure or the compliance judgment result indicates non-compliance, select the target rollback parameter set from the stable parameter snapshot library according to the rollback level and send it to the test cell group. When the sequential test result indicates success and the compliance judgment result indicates compliance, write the candidate wireless configuration parameter set into the stable parameter snapshot library.

[0012] Optionally, S1 includes:

[0013] A preset time window is set in the network management server, and IoT measurement data reported by IoT terminals in the factory area is received within the preset time window. The IoT measurement data includes terminal identifier, terminal location information, timestamp, wireless measurement indicators, service performance indicators, and service level or service type. The IoT measurement data is time-aligned and data integrity is verified, and IoT measurement data that does not meet the preset integrity conditions is removed, forming IoT measurement data for subsequent processing. The initial wireless configuration parameter set corresponding to the 5G cell that has established a wireless connection with the IoT terminal is read. The initial wireless configuration parameter set includes transmit power parameters, beam parameters, and handover parameters. The stable parameter snapshot library is read. The stable parameter snapshot library stores at least one historically stable wireless configuration parameter set and the effective cell identifier and effective time information corresponding to the historically stable wireless configuration parameter set. The preset confidence coefficient table is read. The preset confidence coefficient table stores the correspondence between service level and confidence coefficient. The IoT measurement data, the initial wireless configuration parameter set, the stable parameter snapshot library, and the preset confidence coefficient table are output.

[0014] Furthermore, a service continuity threshold vector is read or configured in the network management server. The service continuity threshold vector includes a latency threshold, a packet loss rate threshold, and a throughput threshold, and the service continuity threshold vector is stored in association with the service level or service type.

[0015] Optionally, S2 includes:

[0016] The preset time window is divided into multiple consecutive time slices, and the set of IoT terminals and the set of 5G cells corresponding to the time slice are determined based on the IoT measurement data within each time slice.

[0017] IoT terminal nodes are established for the IoT terminal set, and 5G cell nodes are established for the 5G cell set;

[0018] Based on the connection relationship between the terminal and the 5G cell in the IoT measurement data, a connection edge is established between the corresponding IoT terminal node and the corresponding 5G cell node, and connection edge features are configured for the connection edge according to the wireless measurement indicators.

[0019] Based on the terminal location information in the IoT measurement data, neighboring edges are established between IoT terminal nodes that meet preset proximity conditions, and neighboring edge features are configured for the neighboring edges based on the terminal location information.

[0020] The cell node characteristics are configured for the 5G cell node based on the initial wireless configuration parameter set; the gating coefficients corresponding to each time slice are calculated based on the scene state characteristics in the IoT measurement data, and the gating coefficients are used to characterize the impact of scene state on changes in wireless propagation conditions.

[0021] The gating coefficient is used to update the connecting edge features of the connecting edge and the neighboring edge features of the neighboring edge to form a time-varying edge weight gating mechanism.

[0022] The graphs constructed from each time slice and updated by the edge weight gating mechanism are combined in chronological order to obtain the heterogeneous dynamic spatiotemporal graph sequence.

[0023] Furthermore, the heterogeneous dynamic spatiotemporal graph sequence further includes environmental nodes, which include at least one of building nodes, metal obstruction nodes, and environmental sensor nodes, and a propagation influence edge is established between the environmental nodes and the 5G cell nodes. The edge features of the propagation influence edge include at least one of obstruction indication, material attenuation coefficient, and temperature, humidity, or dust concentration measured by the environmental sensor.

[0024] Furthermore, the scene state features include at least one of the following: the start / stop status of production line equipment, the cycle time of robotic arm operation, the density and speed of automated guided vehicles or forklifts, the position of overhead cranes or hoists, and the stacking status of metal coils or large tooling; and a gating coefficient is calculated based on the scene state features, such that when the scene state features indicate enhanced occlusion or drastic changes in the reflection environment, the gating coefficient reduces the weight of the connecting edge features or adjacent edge features to suppress the prediction offset caused by sudden scene changes.

[0025] Optionally, S3 includes:

[0026] The pre-trained heterogeneous dynamic spatiotemporal graph neural network model is invoked to infer the heterogeneous dynamic spatiotemporal graph sequence, outputting the predicted mean vector and the predicted standard deviation vector of service performance indicators for each 5G cell. A confidence coefficient corresponding to the service level is determined from a pre-set confidence coefficient table. Based on the predicted mean vector, the predicted standard deviation vector, and the confidence coefficient, a lower confidence bound vector is calculated for each 5G cell. Each component of the lower confidence bound vector is the predicted mean of the corresponding service performance indicator minus the product of the confidence coefficient and the predicted standard deviation. Based on the comparison between the lower confidence bound vector and the service continuity threshold vector, the judgment result of the service continuity security condition is determined. When each component of the lower confidence bound vector is not less than the corresponding component of the service continuity threshold vector, the judgment result of the service continuity security condition is determined to be satisfied; otherwise, it is determined to be unsatisfied. The predicted mean vector, the predicted standard deviation vector, the lower confidence bound vector, and the judgment result of the service continuity security condition are output.

[0027] Furthermore, before calculating the lower confidence bound, the method further includes: performing uncertainty calibration on the prediction standard deviation vector of the business performance index, wherein the uncertainty calibration includes quantile calibration or conformal prediction calibration based on the prediction error within the historical time window, and calculating the lower confidence bound vector based on the calibrated prediction standard deviation vector.

[0028] Optionally, S4 includes:

[0029] The parameter search range is determined by the initial set of wireless configuration parameters, and a set of candidate wireless configuration parameters to be evaluated is generated iteratively within the parameter search range by constrained Bayesian optimization.

[0030] In each iteration, scene features are extracted based on the IoT measurement data, and the feasibility classifier is called to output the feasibility probability based on the scene features and the set of candidate wireless configuration parameters to be evaluated. The feasibility discrimination constraint is determined based on the comparison result between the feasibility probability and the preset feasibility threshold.

[0031] The parameter change index is calculated based on the parameter change amount of the candidate wireless configuration parameter set to be evaluated and the initial wireless configuration parameter set. The change magnitude constraint is then determined based on the comparison between the parameter change index and the preset change magnitude threshold. When both the feasibility discrimination constraint and the change magnitude constraint are satisfied, a corresponding heterogeneous dynamic spatiotemporal graph sequence is generated based on the candidate wireless configuration parameter set to be evaluated. The heterogeneous dynamic spatiotemporal graph neural network model is then called to output the corresponding service performance index prediction mean vector and the corresponding service performance index prediction standard deviation vector. The corresponding lower confidence bound vector is then calculated. The comparison between the corresponding lower confidence bound vector and the service continuity threshold vector determines whether the candidate wireless configuration parameter set to be evaluated meets the service continuity security condition. A target value is calculated for the candidate wireless configuration parameter set to be evaluated that meets the service continuity security condition based on the target service performance index improvement rules. Finally, a collection function value is calculated based on the target value, the feasibility probability, and the parameter change index.

[0032] The set of candidate wireless configuration parameters to be evaluated with the largest acquisition function value is taken as the iterative output, and when the preset iteration termination condition is met, the set of candidate wireless configuration parameters with the largest acquisition function value during the iteration process is determined as the candidate wireless configuration parameter set.

[0033] Furthermore, before executing S5, the process includes: constructing a wireless propagation simulation model based on the 3D map of the factory area, building material information, and the location information of the IoT terminal, and performing a coverage pre-simulation on the candidate wireless configuration parameter set to obtain a pre-simulation coverage index; when the pre-simulation coverage index does not meet the preset pre-simulation conditions, the distribution of the candidate wireless configuration parameter set to the test cell group is prohibited, and the process returns to execute S4 to update the candidate wireless configuration parameter set.

[0034] Optionally, S5 includes:

[0035] A test cell group is determined based on the candidate radio configuration parameter set, and a control cell group is determined based on the initial radio configuration parameter set. The test cell group and the control cell group meet preset comparability conditions, which include the same cell type and adjacent or overlapping coverage areas. The candidate radio configuration parameter set is sent to the test cell group, while the control cell group uses the initial radio configuration parameter set.

[0036] The verification period is divided into multiple consecutive verification sub-time windows, and IoT measurement data corresponding to the test cell group and the control cell group are collected in each verification sub-time window. Based on the IoT measurement data in each verification sub-time window, the statistical values ​​of the service performance indicators of the test cell group and the control cell group are calculated, and the difference between the two is used to construct a service performance indicator difference sequence.

[0037] A sequential statistical test is performed on the differential sequence of the business performance indicators. The sequential statistical test includes updating the test statistic at the end of each verification sub-time window and comparing the test statistic with a pass threshold and a fail threshold. When the test statistic reaches the pass threshold, the sequential test result is determined to be pass. When the test statistic reaches the fail threshold, the sequential test result is determined to be fail. When the verification period ends and the test statistic does not reach the pass threshold and the fail threshold, the sequential test result is determined according to a preset judgment rule.

[0038] The minimum value vector of service performance indicators for the test cell group during the verification period is calculated based on the IoT measurement data, and the minimum value vector of service performance indicators is compared with the service continuity threshold vector to obtain a compliance judgment result; the verification result is output, including the sequential test result and the compliance judgment result.

[0039] Optionally, S6 includes:

[0040] The rollback level is determined based on the sequential test results and compliance judgment results in the verification results. Specifically, when the compliance judgment result indicates non-compliance, the rollback level is determined to be the highest level. When the compliance judgment result indicates compliance and the sequential test result indicates failure, the rollback level is determined to be the second highest level. When the compliance judgment result indicates compliance and the sequential test result indicates success, the rollback is determined not to be triggered.

[0041] When a rollback is triggered, a set of historically stable radio configuration parameters matching the test cell group is retrieved from the stable parameter snapshot library according to the rollback level, and the retrieved set of historically stable radio configuration parameters is determined as the target rollback parameter set. The target rollback parameter set corresponding to the highest level is the set of historically stable radio configuration parameters that was most recently effective and covers all cells in the test cell group in the stable parameter snapshot library, and the target rollback parameter set corresponding to the second highest level is the set of historically stable radio configuration parameters in the stable parameter snapshot library that has the same parameter dimension as the candidate radio configuration parameter set and has the smallest parameter change amount compared to the candidate radio configuration parameter set.

[0042] The target rollback parameter set is sent to the test cell group to complete the rollback, and the rollback radio configuration parameter set is determined as the final radio configuration parameter set.

[0043] When no rollback is triggered, the candidate wireless configuration parameter set is written into the stable parameter snapshot library as a new stable parameter snapshot, and the candidate wireless configuration parameter set is determined as the final wireless configuration parameter set.

[0044] Furthermore, when a rollback is triggered, the method further includes: rolling back the candidate wireless configuration parameter set parameter by parameter according to preset parameter priorities, wherein the parameter priorities include rolling back switching parameters before rolling back beam parameters and rolling back beam parameters before rolling back transmit power parameters; collecting IoT measurement data after each parameter rollback and recalculating the minimum value vector of service performance indicators after each parameter rollback; and upgrading the rollback level to the highest level and using the target rollback parameter set corresponding to the highest level to complete the rollback.

[0045] Optionally, the pre-trained heterogeneous dynamic spatiotemporal graph neural network model and the pre-trained feasibility classifier are obtained through the following training method:

[0046] The system collects IoT measurement data within multiple historical time windows and simultaneously collects the wireless configuration parameter set corresponding to each historical time window. Based on the IoT measurement data, it determines the actual service performance index vector for each 5G cell within each historical time window and generates a feasibility label based on the comparison between the actual service performance index vector and the service continuity threshold vector. It extracts scene features based on the IoT measurement data, including scene state features. Based on the terminal distribution, terminal mobility information, and connection relationship between terminals and 5G cells represented by the IoT measurement data, it constructs a training heterogeneous dynamic spatiotemporal graph sequence. Nodes include IoT terminal nodes and 5G cell nodes, and edges include connecting edges and adjacent edges. The edge weights of the edges are updated over time using an edge weight gating mechanism based on the scene state features. The heterogeneous dynamic spatiotemporal graph neural network model is driven to output the predicted mean vector and the predicted standard deviation vector of business performance indicators for the training heterogeneous dynamic spatiotemporal graph sequence. A loss function is constructed based on the actual business performance indicator vector, the predicted mean vector, and the predicted standard deviation vector. The model parameters are iteratively updated to make the loss function converge, resulting in the trained heterogeneous dynamic spatiotemporal graph neural network model. The wireless configuration parameter set is combined with the scene features to form training sample features. The feasibility classifier is driven to output the feasibility probability based on the training sample features. A classification loss function is constructed based on the feasibility probability and the feasibility label. The classifier parameters are iteratively updated to make the classification loss function converge, resulting in the trained feasibility classifier.

[0047] On the other hand, the present invention also provides an automatic optimization system for 5G coverage in factory areas based on the Internet of Things, comprising:

[0048] The data acquisition module is used to collect IoT measurement data reported by IoT terminals in the factory area and obtain the service level or service type, and read the initial wireless configuration parameter set and stable parameter snapshot library of 5G cells.

[0049] The graph construction module is used to construct a heterogeneous dynamic spatiotemporal graph sequence representing the relationship between IoT terminals and 5G cells based on the IoT measurement data.

[0050] The reasoning and security boundary module is used to call the pre-trained heterogeneous dynamic spatiotemporal graph neural network model to output the predicted value of business performance indicators and their uncertainty, determine the confidence coefficient based on the business level to calculate the lower confidence boundary, and compare the lower confidence boundary with the business continuity threshold vector to determine the business continuity security conditions.

[0051] The constrained Bayesian optimization module is used to iteratively generate a set of wireless configuration parameters to be evaluated under feasibility discrimination constraints and change magnitude constraints. It determines the set of candidate wireless configuration parameters based on the lower confidence bound and the acquisition function. The feasibility discrimination constraints are determined by the feasibility probability output by the pre-trained feasibility classifier and the preset feasibility threshold.

[0052] The comparative sequential verification module is used to establish a test cell group based on the candidate radio configuration parameter set and a control cell group based on the initial radio configuration parameter set, distribute the candidate radio configuration parameter set and collect verification data, perform comparative sequential statistical tests and generate verification results.

[0053] The graded rollback and snapshot management module is used to determine the rollback level based on the verification results and select the target rollback parameter set from the stable parameter snapshot library for rollback, or to write the candidate wireless configuration parameter set into the stable parameter snapshot library when the verification is passed.

[0054] The beneficial effects of this invention are:

[0055] 1. By using a heterogeneous dynamic spatiotemporal graph neural network to output predicted values ​​and uncertainties of business performance indicators, and by introducing a confidence coefficient to calculate the lower confidence boundary as a safety boundary based on the business level, the candidate parameter schemes are limited to the safe and feasible domain that meets the business continuity threshold vector, thereby reducing the risk of performance degradation of key production operations caused by automatic parameter tuning and improving business continuity assurance capabilities.

[0056] 2. By implementing constrained Bayesian optimization under feasibility discrimination constraints and change range constraints, and combining the feasibility probability output by the feasibility classifier to screen candidate solutions and limit the amount of parameter changes, we can improve the efficiency of parameter search and optimization benefits, and avoid coverage fluctuations and network instability caused by large parameter changes, thereby improving the controllability and robustness of automatic optimization.

[0057] 3. The optimization effect is quickly verified online by using a comparative sequential statistical test between the experimental and control groups. When the test fails or is non-compliant, a hierarchical rollback and parameter-by-parameter rollback are triggered based on the stable parameter snapshot library. This enables automatic verification of the optimization effect and automatic recovery from anomalies, shortening the time affected by faults and reducing the cost of manual intervention. Attached Figure Description

[0058] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0059] Figure 1 This is a flowchart of an automatic optimization method for 5G coverage in a factory area based on the Internet of Things proposed in this invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0061] refer to Figure 1 An automatic optimization method for 5G coverage in factory areas based on the Internet of Things includes:

[0062] S1. Collect IoT measurement data reported by IoT terminals within the factory area, including service level or service type, and obtain the initial wireless configuration parameter set and stable parameter snapshot library corresponding to the 5G cell; S2. Construct a heterogeneous dynamic spatiotemporal graph sequence representing the IoT terminal and the 5G cell and their correlation based on the IoT measurement data; S3. Call the pre-trained heterogeneous dynamic spatiotemporal graph neural network model to infer the heterogeneous dynamic spatiotemporal graph sequence, obtain the predicted value and uncertainty of the service performance index, determine the confidence coefficient according to the service level, calculate the lower confidence bound, and compare the lower confidence bound with the preset service continuity threshold vector to determine the service continuity security condition; S4. Use constrained Bayesian optimization iteration to generate the set of wireless configuration parameters to be evaluated, and apply feasibility discrimination constraints and change magnitude constraints. Call the pre-trained feasibility classifier to output the feasibility probability. The feasibility discrimination constraint includes that the feasibility probability is not less than the preset feasibility threshold, and the change magnitude constraint includes that the parameter change amount is not greater than the preset change magnitude threshold. For the set of parameters to be evaluated that meet the constraints, the set of parameters to be evaluated is evaluated. S5. Evaluate the lower confidence bound of the wireless configuration parameter set and determine whether it meets the service continuity security conditions. Determine the candidate wireless configuration parameter set based on the acquisition function. S6. Determine the test cell group based on the candidate wireless configuration parameter set and the control cell group based on the initial wireless configuration parameter set. Send the candidate wireless configuration parameter set to the test cell group while the control cell group maintains the initial wireless configuration parameter set. Collect the corresponding IoT measurement data during the verification period, perform a comparative sequential statistical test to obtain the sequential test result, and determine the compliance of the service performance indicators of the test cell group with the service continuity threshold vector to obtain the compliance judgment result, thus forming the verification result. S7. Determine the rollback level based on the verification result. When the sequential test result indicates failure or the compliance judgment result indicates non-compliance, select the target rollback parameter set from the stable parameter snapshot library according to the rollback level and send it to the test cell group. When the sequential test result indicates success and the compliance judgment result indicates compliance, write the candidate wireless configuration parameter set into the stable parameter snapshot library.

[0063] In this specific embodiment, S1 includes:

[0064] When the network management server starts up, it sets the preset time window to 300 seconds and defines the data collection range based on the start and end times of the window synchronized by the server. It receives IoT measurement data reported by IoT terminals in the factory area within the preset time window through the MQTT access channel. The IoT measurement data is stored as structured data rows at the record granularity, and each row contains terminal identifier, terminal location information, timestamp, wireless measurement indicators, service performance indicators, and service level and service type. The terminal location information is represented by three-dimensional coordinates and positioning accuracy fields in the factory area coordinate system. The timestamp is represented by UTC millisecond-level time. The wireless measurement indicators are fixed to include RSRP, RSRQ, and SINR. The service performance indicators are fixed to include end-to-end latency, packet loss rate, and throughput. The service level is represented by integer levels, with smaller level values ​​indicating more critical services.

[0065] The network management server performs time alignment and data integrity verification on the received IoT measurement data. Time alignment maps the timestamp to the corresponding aligned time point with an alignment granularity of 1 second, and for multiple records from the same terminal at the same aligned time point, the median of each field is taken to form a single aligned record. Data integrity verification filters and removes IoT measurement data that does not meet the integrity conditions one by one according to integrity criteria. The integrity conditions are fixed and include field integrity conditions, range consistency conditions, and sampling density conditions. Field integrity conditions require that the terminal identifier, terminal location information, timestamp, wireless measurement indicators, service performance indicators, and service level fields are all non-empty and resolvable values. Range consistency conditions require that the terminal location information falls within the factory boundary polygon and the positioning accuracy is no greater than 5 meters, and that the RSRP is in a certain position. And the RSRQ is in And the SINR is in And the packet loss rate is at Furthermore, the throughput is not less than 0 Mbps, and the sampling density condition requires that the number of aligned records formed by each terminal within the preset time window is not less than 240, and the time span of the aligned records covers the start and end boundaries of the preset time window.

[0066] The network management server aggregates the verified alignment records to form an IoT measurement dataset for subsequent processing and establishes an index with the cell identifier as the key. Then, through the network management northbound interface, it reads the currently effective configuration corresponding to the 5G cell that has established a wireless connection with the IoT terminal at the start time of the preset time window, forming an initial wireless configuration parameter set. The initial wireless configuration parameter set includes transmit power parameters, beam parameters, and handover parameters at the cell granularity. The transmit power parameters include the reference signal transmit power and the maximum transmit power. The beam parameters include the beam set identifier, the beamforming weight set, and the beam handover threshold. The handover parameters include event A3 offset, hysteresis, and trigger delay.

[0067] The network management server reads the stable parameter snapshot library from the relational database. The stable parameter snapshot library stores a set of historically stable wireless configuration parameters in the form of snapshot records and stores the corresponding set of effective cell identifiers and effective time information for each snapshot record. The stable parameter snapshot library only stores snapshot records with stable tags to ensure that the rollback target has a traceable source of stability.

[0068] The network management server reads a pre-set confidence coefficient table from the configuration database. The pre-set confidence coefficient table stores confidence coefficients that correspond one-to-one with the service level, and is fixedly configured so that the confidence coefficient is 2.33 when the service level is 1, 1.96 when the service level is 2, and 1.64 when the service level is 3.

[0069] The network management server reads the service continuity threshold vector from the threshold configuration table and stores it in association with service levels. The service continuity threshold vector is defined as follows:

[0070] ;

[0071] in Indicates the business level is The business continuity threshold vector and Indicates the business level. This represents the delay threshold component corresponding to the end-to-end delay and is stored as a negative maximum allowable delay value to ensure that the threshold comparison direction is consistent with subsequent security decisions. This represents the packet loss rate threshold component corresponding to the stated packet loss rate and is stored as a negative maximum allowable packet loss rate value. This represents the throughput threshold component corresponding to the stated throughput and is stored as the minimum allowed throughput value, with a fixed configuration when the service level is 1. Fixed configuration when business level is 2 Fixed configuration when business level is 3 The IoT measurement dataset, the initial wireless configuration parameter set, the stable parameter snapshot library, the pre-set confidence coefficient table, and the service continuity threshold vector associated with each service level are output together to subsequent steps for heterogeneous dynamic spatiotemporal graph construction and service continuity security condition determination.

[0072] In this specific embodiment, S2 includes:

[0073] The network management server receives IoT measurement datasets, initial wireless configuration parameter sets, and factory boundary and basic ledger data corresponding to a preset time window. It divides the preset time window of 300 s into 30 consecutive time slices, with each time slice having a fixed duration of 10 s. The aligned records belonging to that time slice within each time slice are aggregated according to terminal identifiers to obtain the IoT terminal set for that time slice, and aggregated according to cell identifiers to obtain the 5G cell set for that time slice.

[0074] For the IoT terminal set, IoT terminal nodes are established with the terminal identifier as the node unique key; for the 5G cell set, 5G cell nodes are established with the cell identifier as the node unique key.

[0075] Based on the "connection relationship between terminal and 5G cell" in the alignment record, a connection edge is established between the corresponding IoT terminal node and the corresponding 5G cell node, and the connection edge type is marked as a connection edge. The connection relationship is determined by the serving cell identifier field in the alignment record, and a connection edge is established for each alignment record. For multiple connection edges generated by the same terminal in the same time slice, the median of their wireless measurement indicators is taken as the connection edge feature from the terminal to the serving cell in that time slice. The connection edge feature is fixedly constructed as a three-dimensional vector, and the three-dimensional components are RSRP, RSRQ and SINR in sequence. Before being entered into the graph, linear normalization is performed according to the normalization parameter table built into the server to ensure that the numerical range of each component is consistent.

[0076] Based on the terminal location information in the alignment record, neighboring edges are established between IoT terminal nodes that meet preset proximity conditions, and the neighboring edge type is marked as a neighboring edge. The preset proximity conditions are fixed as follows: the three-dimensional Euclidean distance between two terminals within the same time slice is not greater than 20 m, and the overlap ratio of the alignment time points of the two terminals is not less than 0.8. The neighboring edge features are fixedly constructed as a four-dimensional vector, and the four-dimensional components are the relative displacements of the two terminals in sequence. 3D Euclidean distance ,in The result is obtained by subtracting the mean values ​​of the position sequences of the two terminals within that time slice. And perform the same linear normalization on the four-dimensional vector as on the features of the connecting edges;

[0077] Based on the initial wireless configuration parameter set, cell node characteristics are configured for each 5G cell node, and these characteristics are fixedly constructed into a nine-dimensional vector. The nine components are, in order, the reference signal transmit power, maximum transmit power, beam set identifier, main beam azimuth angle, main beam downtilt angle, beam switching threshold, event A3 offset, hysteresis, and trigger delay. The beam set identifier, as a discrete quantity, is mapped to an integer according to a preset mapping table and then normalized. The interval and the angle measurement is in degrees and according to or Range normalization;

[0078] The network management server simultaneously constructs an environmental node set based on the 3D map of the factory area and the environmental ledger, and incorporates the environmental nodes into each time slice image. The environmental nodes are fixed to include building nodes, metal obstruction nodes, and environmental sensor nodes, and all of them use the object identifier in the ledger as the node unique key. Among them, the building nodes store the building outline and material category, and obtain the material attenuation coefficient by looking up the material category in a table. The metal obstruction nodes store the obstruction bounding box and obstruction height, and fix their material attenuation coefficient to 1.0. The environmental sensor nodes store the temperature, humidity, and dust concentration readings aligned with the time slice.

[0079] The network management server establishes a propagation influence edge between the environmental node and the 5G cell node and marks the propagation influence edge type as a propagation influence edge. The establishment rule of the propagation influence edge is fixed as the horizontal distance from the geometric center of the environmental node to the cell antenna coordinates is no more than 80 m. The propagation influence edge features are fixedly constructed as a five-dimensional vector, and the five-dimensional components are, in order, occlusion indication, material attenuation coefficient, temperature, humidity and dust concentration. The occlusion indication is calculated by whether the building outline or metal occlusion box and the line connecting the "cell antenna to the average position of the terminal" intersect spatially, and the value is 1 if they intersect and 0 if they do not. The temperature, humidity and dust concentration are obtained by taking the average of the aligned readings of the environmental sensor node in the time slice and filling the missing readings with the most recent valid reading of the sensor.

[0080] The network management server reads and aligns scene state features from the factory production management system and equipment control system, and calculates the gating coefficient corresponding to each time slice based on these scene state features. The scene state features are fixed and include the start / stop status of production line equipment, the cycle time of robotic arms, the density and speed of automated guided vehicles or forklifts, the position of overhead cranes or hoists, and the stacking status of metal coils or large tooling. The gating coefficient is calculated at the time slice level and its value range is limited to [specific values ​​to be filled in]. Furthermore, when the scene state features represent occlusion enhancement or drastic changes in the reflection environment, the gating coefficient takes a smaller value and is used to suppress the prediction offset caused by sudden scene changes;

[0081] The network management server uses the gating coefficients to perform an edge weight gating mechanism update on the features of connected edges and neighboring edges, while keeping the features affected by propagation from participating in the gating update. The edge weight gating mechanism uses the following update formula:

[0082] ;

[0083] in Indicates the time slice index is The edge index is The updated feature vector of the edge and the edge index is The edges are either connecting edges or adjacent edges. Indicates the time slice index is The gating coefficient is calculated from the scene state features. Indicates the time slice index is The edge index is The updated feature vectors are 3D vectors for connecting edges and 4D vectors for neighboring edges. Represents the time slice number and is an integer ranging from 1 to 30. Indicates time slice The edge number is obtained in the order in which the edges are constructed and is used to uniquely locate an edge.

[0084] The network management server combines the heterogeneous graphs constructed in each time slice and updated by the edge weight gating mechanism into a sequence of heterogeneous dynamic spatiotemporal graphs in ascending order of time slice number.

[0085] In this specific embodiment, S3 includes:

[0086] The network management server loads a pre-trained heterogeneous dynamic spatiotemporal graph neural network model from the model repository and switches to inference mode. This model takes a heterogeneous dynamic spatiotemporal graph sequence as input and 5G cell nodes as output. The model structure is fixed, consisting of a spatial heterogeneous encoder, a time-series encoder, and an indicator output head, with the following fixed parameter values: The spatial heterogeneous encoder includes two relation-aware graph attention layers and establishes independent attention weight matrices for three edge types: connection edges, neighbor edges, and propagation influence edges. The number of attention heads per layer is fixed at 4, and the output dimension of each head is fixed at 16, resulting in a single-layer hidden dimension of 64. The time-series encoder uses a gated recurrent unit to encode the hidden representation of cell nodes sorted by time slice, with the gated recurrent unit's hidden dimension fixed at 64. The hidden state of the last time slice is used as the spatiotemporal fusion representation of the cell. The indicator output head uses two multilayer perceptrons to output the mean vector and standard deviation vector of the predicted service performance indicators, respectively. The number of layers in the two multilayer perceptrons is fixed at 2, and the intermediate layer dimension is fixed at 32. The output path of the prediction standard deviation is activated at the end using softplus and then... The numerical stability term is used to ensure that the prediction standard deviation is positive;

[0087] The network management server encodes the heterogeneous dynamic spatiotemporal graph sequence into a model input tensor in time slice order and performs forward inference once to obtain the predicted mean vector and the predicted standard deviation vector of the service performance index for each 5G cell. In this embodiment, the service performance index is uniformly adopted as a three-dimensional index vector with the same direction as the service continuity threshold vector, and the three-dimensional components represent negative end-to-end latency, negative packet loss rate and throughput in sequence, thereby ensuring the consistency of subsequent judgment direction of "not less than the threshold".

[0088] The network management server determines the service level of each 5G cell based on the connection relationship between terminals and cells in the IoT measurement data collected in step S1. The service level is the minimum value of the service levels of all terminals served by the cell within the preset time window to characterize the criticality of the service carried by the cell. The server then reads the confidence coefficient from a preset confidence coefficient table based on the service level and records it as follows. Simultaneously, the uncertainty calibration coefficient vector corresponding to this business level is read from the uncertainty calibration table and denoted as... The uncertainty calibration table is continuously maintained by the network management server during operation, using the most recent 7 consecutive days of historical time windows as the calibration set. For each service level, the component ratio sequence of "the absolute value of the deviation between the actual service performance index vector and the predicted mean vector divided by the predicted standard deviation vector" is calculated and taken. Quantiles form a three-dimensional calibration coefficient vector. This enables conformal prediction calibration of the predicted standard deviation vector and ensures that the lower confidence boundary coverage is consistent with the true risk.

[0089] The network management server calculates the lower confidence bound vector for each 5G cell and compares it with the service continuity threshold vector corresponding to that service level to determine the service continuity security conditions. The formula for calculating the lower confidence bound vector is as follows:

[0090] ;

[0091] in The community sign is The lower confidence bound vector of the 5G cell is a three-dimensional vector. This represents the predicted mean vector of the service performance indicators of the cell, and is a three-dimensional vector. This represents the predicted standard deviation vector of the service performance indicators of the cell, and is a three-dimensional vector. Indicates the business level is The confidence coefficient is obtained and read from a preset confidence coefficient table. Indicates the business level is The uncertainty calibration coefficient vector is a three-dimensional vector and is read from the uncertainty calibration table. This represents element-wise multiplication. This indicates the service level corresponding to the cell and is determined by the minimum service level value of the serving terminal in that cell. Indicates a 5G cell identifier;

[0092] The network management server will use the lower confidence bound vector With business continuity threshold vector Component-by-component comparison and when Each component is not less than The corresponding component determines whether the cell meets the service continuity security conditions; otherwise, it determines that the cell does not meet the conditions. The mean vector of service performance index prediction, the standard deviation vector of service performance index prediction, the lower confidence boundary vector, and the service continuity security condition determination result of all cells are output.

[0093] In this specific embodiment, S4 includes:

[0094] The network management server receives the initial wireless configuration parameter set and the heterogeneous dynamic spatiotemporal graph sequence, and receives the service continuity threshold vector and confidence coefficient configuration required for service continuity security condition determination. The server represents the "wireless configuration parameter set to be evaluated" as a parameter configuration set simultaneously provided for all 12 5G cells in the factory area. The parameter configuration of each cell is fixed and consists of reference signal transmit power, main beam downtilt angle, event A3 offset, hysteresis, and trigger delay, and is consistent with the parameter dimensions of the initial wireless configuration parameter set. The server determines the parameter search range based on the initial wireless configuration parameter set and fixes the discretization step size to 0.5 dB for the reference signal transmit power and 0.5 dB for the main beam downtilt angle. Event A3 bias step size is 0.5 dB, hysteresis step size is 0.5 dB, and trigger delay step size is 20 ms. Simultaneously, the change amplitude constraint is fixed at no more than 1 dB for the change in reference signal transmit power and no more than [missing value] for each cell. The change in bias of event A3 is no greater than 1 dB, the change in hysteresis is no greater than 1 dB, and the change in trigger delay is no greater than 40 ms. When generating parameters, components that exceed the range are truncated to ensure that all sets of wireless configuration parameters to be evaluated meet the change range constraints.

[0095] The server uses constrained Bayesian optimization to iteratively generate a set of wireless configuration parameters to be evaluated within the parameter search range. The surrogate model for the constrained Bayesian optimization is fixedly Gaussian process regression to fit the target value output by the "target service performance indicator improvement rule". The kernel function of the Gaussian process regression is fixedly Matérnn. The kernel input is a concatenated vector obtained by linearly normalizing all cell parameter vectors according to their components. After each iteration, the kernel function hyperparameters are updated by maximizing the log marginal likelihood. The initial sampling is fixed by generating 20 sets of parameters that satisfy the change magnitude constraint using the Sobol sequence as the initial observation points, and their target values ​​are obtained by calling the inference process of step S3.

[0096] In each iteration, the server extracts scene features consistent with step S2 based on IoT measurement data and concatenates them with the set of wireless configuration parameters to be evaluated at the cell level to form a classification input. It then calls a pre-trained feasibility classifier to output the feasibility probability. The feasibility classifier is fixed as a three-layer multilayer perceptron, with the input layer dimension fixed at the sum of "scene feature dimension 10" and "single cell parameter dimension 5" (15). The first hidden layer has 64 neurons with ReLU activation function, the second hidden layer has 32 neurons with ReLU activation function, and the output layer has 1 neuron with Sigmoid activation function, outputting a range of... The server obtains the feasibility probability for each cell and takes the minimum value of the feasibility probabilities of all cells as the feasibility probability of the wireless configuration parameter set to be evaluated, and fixes the feasibility judgment constraint to the feasibility probability as not less than 0.8.

[0097] When the set of wireless configuration parameters to be evaluated simultaneously meets the feasibility judgment constraint and the change magnitude constraint, the server copies the heterogeneous dynamic spatiotemporal graph sequence to the evaluation heterogeneous dynamic spatiotemporal graph sequence corresponding to the set of wireless configuration parameters to be evaluated, and only replaces the 5G cell node features of each time slice with the parameter features normalized according to the set of wireless configuration parameters to be evaluated to maintain the semantic consistency of nodes and edges. Then, the same heterogeneous dynamic spatiotemporal graph neural network model in step S3 is called to infer the evaluation heterogeneous dynamic spatiotemporal graph sequence and calculate the lower confidence bound vector of each cell and compare it with the service continuity threshold vector of the corresponding service level to determine the service continuity safety condition. The server defines the "target service performance index improvement rule" as maximizing the lower confidence bound improvement of the throughput component of all cells and maximizing the lower confidence bound margin of the negative end-to-end delay component and the negative packet loss rate component under the premise of meeting the service continuity safety condition, and outputs the rule as a single target value for proxy model fitting.

[0098] In each iteration, the server uses a fixed set of 2048 candidate points to perform a discrete search on the acquisition function and selects the set of wireless configuration parameters to be evaluated with the largest acquisition function value to proceed to the next verification step. The acquisition function is defined as follows:

[0099] ;

[0100] in This represents the collected function value and is used to select the iterative output from the candidate point set. This represents the parameter vector, which is the result of concatenating and normalizing the set of wireless configuration parameters to be evaluated by cell and parameter components. This represents the target value calculated from the lower confidence bound vector based on the target service performance indicator improvement rules, with a larger value indicating a greater predicted improvement. This represents the feasibility probability obtained by taking the minimum value from the output of the feasibility classifier for each cell. This indicates the parameter change index and its value range is... The value is obtained by dividing the absolute change of each component of the set of wireless configuration parameters to be evaluated relative to the initial set of wireless configuration parameters by the corresponding maximum allowable change, and then taking the maximum value among all components.

[0101] When a set of wireless configuration parameters to be evaluated does not meet the business continuity and security requirements, the server will change its target value. Set the value to 0 to prevent the collection function from selecting this set.

[0102] The server is set to a preset iteration termination condition of 40 maximum iterations or a maximum acquisition function value increase of no more than 5 consecutive iterations. And at the termination, the set of wireless configuration parameters with the largest collected function value during the iteration process is determined as the candidate set of wireless configuration parameters;

[0103] Before outputting the candidate wireless configuration parameter set, the server performs a coverage pre-simulation to implement the pre-distribution blocking mechanism. The coverage pre-simulation is fixedly completed by a wireless propagation simulation model, which takes a 3D map of the factory area, building material information, and IoT terminal location information as inputs, and uses the cell reference signal transmission power and main beam downtilt angle from the candidate wireless configuration parameter set as transmission-side parameters. The simulation calculation is fixedly performed using 3D ray tracing, and the path loss of the direct path and the first reflection path from each cell is calculated for each terminal location. The reflection and blocking losses are determined by the material attenuation coefficient, and the interference power is obtained by linearly superimposing the received power of the non-serving cell. Combined with the noise floor of -100 dBm, the pre-simulated RSRP and pre-simulated SINR of the terminal location are calculated. Based on this, the server calculates the pre-simulated coverage index and defines the pre-simulated coverage index as "the proportion of terminal locations with a pre-simulated RSRP of not less than -95 dBm and a pre-simulated SINR of not less than 5 dB", and the preset pre-simulation condition is fixedly set to the proportion of not less than 0.95. Furthermore, the proportion relative to the initial wireless configuration parameter set will not decrease. If the candidate wireless configuration parameter set does not meet the preset pre-simulation conditions, the server will prohibit the use of the candidate wireless configuration parameter set for subsequent distribution and remove it from the candidate set before returning to continue executing the constraint Bayes optimization to update the candidate wireless configuration parameter set until a candidate wireless configuration parameter set that meets the preset pre-simulation conditions is obtained.

[0104] In this specific embodiment, S5 includes:

[0105] The network management server receives the candidate radio configuration parameter set and determines the test cell group and control cell group accordingly. The server first calculates the parameter difference cell by cell and parameter by parameter between the candidate radio configuration parameter set and the initial radio configuration parameter set, and marks the cell with at least one parameter difference as the candidate test cell. Then, based on the initial coverage area polygon obtained in step S4 coverage pre-simulation, a control cell is matched for each candidate test cell to meet the preset comparability conditions. The preset comparability conditions are fixed and include consistent cell types and adjacent or overlapping coverage areas. The consistent cell types are determined by the equality of the "cell type" field in the network management ledger. The adjacent coverage areas are determined by the minimum distance between the boundaries of the two coverage area polygons being no greater than 0 m and no greater than 10 m. The overlapping coverage areas are determined by the ratio of the overlapping area of ​​the two coverage area polygons to the coverage area of ​​the candidate test cell being no less than 0.2.

[0106] The server uses the matching rule of "the control cell is not occupied by any candidate test cell and has the smallest horizontal distance from the antenna of the candidate test cell" to search for control cells in the set of cells not marked as candidate test cells and form a one-to-one pairing relationship. The candidate test cells that have successfully completed the pairing are formed into test cell groups and their paired control cells are formed into control cell groups.

[0107] The server sends a set of candidate radio configuration parameters to the test cell group through the northbound interface of the network management system, while keeping the initial radio configuration parameter set of the control cell group unchanged. The server sends a parameter change task with a transaction identifier and starts the verification period after the network management system returns a confirmation receipt that "all test cell parameters are effective". The verification period is fixed at 600 seconds and divided into 12 consecutive verification sub-time windows, and the duration of each verification sub-time window is fixed at 50 seconds.

[0108] Within each verification sub-time window, the server collects IoT measurement data corresponding to the test cell group and IoT measurement data corresponding to the control cell group according to the same data reception and time alignment rules as in step S1. Here, "corresponding" is limited to the alignment record of the serving cell identifier belonging to the test cell group or the control cell group within the verification sub-time window. At the end of each verification sub-time window, the service performance index statistics are calculated at the cell granularity and further aggregated into group statistics. The service performance index statistics at the cell granularity are fixedly taken as the median end-to-end latency, median packet loss rate, and median throughput of all terminals in the cell within the verification sub-time window. Based on this, a three-dimensional service performance index statistics vector in the same direction as in step S3 is constructed. The group statistics are fixedly weighted by the number of effective terminals in the cell to calculate the weighted average of the three-dimensional service performance index statistics vectors of each cell in the group to obtain the three-dimensional service performance index statistics vectors of the test cell group and the control cell group within the verification sub-time window.

[0109] The server uses the throughput component in the three-dimensional business performance indicator statistical vector as the indicator to be tested in the sequential statistical test and constructs a difference sequence of business performance indicators. The first element of the difference sequence is... The difference value is denoted as And by the first The throughput statistics of the test cell group are obtained by subtracting the throughput statistics of the control cell group from the throughput statistics of the test cell group within each verification sub-time window. The server performs a controlled sequential statistical test on the differential sequence and updates the test statistic at the end of each verification sub-time window, comparing it with the pass and fail thresholds. The sequential statistical test is always a sequential probability ratio test for mean improvement, and the test statistic is defined as:

[0110] ;

[0111] in Indicates before completion Test statistic after each validation sub-time window An integer representing the number of completed verification sub-time windows, ranging from 1 to 12. This indicates the index of the verification sub-time window, with a value range of 1 to... integers, This represents the minimum acceptable throughput increase and is fixed at a certain value. This represents the variance of the difference value noise, calculated by the server based on the historical time window difference value sequence of the same test cell and control cell paired within the most recent consecutive 24 hours before the verification starts, and remains fixed during this verification period. Indicates the first Throughput difference within each verification sub-time window;

[0112] The server will set the pass threshold to a fixed 2.89 and the fail threshold to a fixed -2.25. When the threshold is reached, the sequential test result is determined to be passed, and the subsequent verification sub-time window acquisition is immediately terminated. When the failure threshold is reached, the sequential test result is determined to be a failure, and the subsequent verification sub-time window data collection is immediately terminated. This continues until all 12 verification sub-time windows are completed. When the pass threshold or fail threshold is not met, the server determines the sequential test result based on whether the mean of the difference sequence is not less than 0 according to the preset judgment rules.

[0113] The server calculates the compliance judgment result while generating the sequential verification result. The compliance judgment is completed by comparing the minimum value vector of service performance indicators of the test cell group during the verification period with the service continuity threshold vector of the corresponding service level. The service level of the test cell group is fixed as the minimum value of the service level of the terminals covered by the aligned records of all test cell groups during the verification period, and is recorded as . The minimum value vector of service performance indicators for the test cell group during the verification period is fixed by taking the minimum value of the three-dimensional service performance indicator statistical vector of the test cell group for each verification sub-time window according to three components, while keeping the end-to-end latency component and packet loss rate component in the negative direction and the throughput component in the positive direction. Therefore, when each component of the minimum value vector of the service performance indicators is not less than the output of step S1 and is consistent with the service level... Associated business continuity threshold vector When the corresponding component is determined, the compliance judgment result is determined to be compliant; otherwise, it is determined to be non-compliant.

[0114] The server writes the sequential test results and compliance judgment results together into the verification result record, and outputs them along with the test cell group identifier set, the control cell group identifier set, the start and end times of the verification period, and the differential sequence for rollback level determination and stable snapshot management.

[0115] In this specific embodiment, S6 includes:

[0116] The network management server receives the verification results and completes the rollback level determination and parameter distribution control within the same transaction context. The server records the sequential verification results in the verification results as follows: And the compliance assessment result will be recorded as ,in The value is limited to pass, fail, or expired, and the expired determination is converted from the preset determination rule in step S5 to pass or fail for subsequent unified processing. The value is limited to either compliant or non-compliant;

[0117] Server based on and Determine the rollback level, specifically when To prevent non-compliance, the rollback level will be set to the highest level and a rollback will be triggered. For compliance and If the rollback fails, the rollback level will be set to the second highest level and a rollback will be triggered. For compliance and To ensure that a rollback is not triggered upon successful completion and to proceed with the stable snapshot write process;

[0118] When a rollback is triggered, the server accesses the stable parameter snapshot library and retrieves a set of historically stable radio configuration parameters that match the test cell group to form the target rollback parameter set. The stable parameter snapshot library stores radio configuration parameter sets in units of snapshot records, and each snapshot record contains the effective cell identifier set, effective time information, and a stability flag. A true stability flag indicates that the snapshot has passed online or offline consistency verification in historical operation and is allowed for rollback. The server's search rule for the highest level is fixed as follows: select the most recently effective snapshot record in reverse chronological order of effective time, whose effective cell identifier set covers all cells in the test cell group, and use the radio configuration parameter set corresponding to this snapshot record as the target rollback parameter set for the highest level. The server's search rule for the second-highest level is fixed as follows: filter snapshot records with a true stability flag that have parameter dimensions consistent with the candidate radio configuration parameter set and whose effective cell identifier set intersects with the test cell group, calculate the parameter change distance between each candidate snapshot record and the candidate radio configuration parameter set, and select the one with the smallest distance as the target rollback parameter set for the second-highest level. The parameter change distance is calculated using the following formula:

[0119] ;

[0120] in Indicates that the snapshot index is The normalized change of the set of wireless configuration parameters corresponding to the snapshot record relative to the set of candidate wireless configuration parameters is distance and The smaller the value, the closer the two are. The parameter component set, in this embodiment, is fixed to include five types of parameter components: reference signal transmit power, main beam downtilt angle, event A3 offset, hysteresis, and trigger delay, and is expanded into a component set for all cells. This represents the parameter component index and is used to uniquely indicate a certain type of parameter component in a specific cell. This indicates the set of candidate wireless configuration parameters in the parameter components. The value on, Indicates that the snapshot index is The target rollback parameter set in the parameter components The value on, Represents parameter components The maximum allowable change amount, and its value is consistent with the change amplitude constraint in step S4, and is fixed as 1 dB corresponding to the reference signal transmit power and the main beam downtilt angle corresponding to... Event A3 bias corresponds to 1 dB, hysteresis corresponds to 1 dB, and trigger delay corresponds to 40 ms. This represents a snapshot index and a specific snapshot record in the corresponding stable parameter snapshot library;

[0121] After determining the target rollback parameter set, the server performs parameter-by-parameter rollback to reduce the instantaneous impact of rollback on on-network services. The parameter priority of parameter-by-parameter rollback is fixed as follows: rollback of switching parameters takes precedence over rollback of beam parameters, and rollback of beam parameters takes precedence over rollback of transmit power parameters. In this embodiment, the switching parameters correspond to event A3 offset, hysteresis, and trigger delay; the beam parameters correspond to the main beam downtilt angle; and the transmit power parameters correspond to the reference signal transmit power.

[0122] The server uses the candidate wireless configuration parameter set as the current network parameter set and the target rollback parameter set as the rollback target set. First, it only issues handover parameters to the test cell group, reverting the candidate values ​​to the target values, creating a unique transaction identifier for this issuance, and waiting for network management confirmation that all test cell group handover parameters have taken effect. Then, the server initiates a rollback evaluation time window, fixed at 120 seconds. Within the rollback evaluation time window, it collects IoT measurement data corresponding to the test cell group according to the data reception and time alignment rules in step S1, and calculates the minimum value vector of service performance indicators according to the statistical rules in step S5, comparing it with the service continuity threshold vector. A comparison is made to determine whether compliance has been restored, including the business level. Within the fixed retrieval and rollback evaluation time window, the minimum service level of the service terminal in the test cell group and the service performance indicators that are consistent with the direction of step S3 are negative end-to-end latency, negative packet loss rate and throughput.

[0123] When compliance is restored, the server stops subsequent rollbacks and determines the current set of parameters on the network as the final set of wireless configuration parameters. The rollback event, the target rollback parameter set identifier, and the final set of wireless configuration parameters are written to the audit log for traceability. When compliance is not restored, the server continues to issue a subset of beam parameter rollbacks to the test cell group and evaluates it according to the same process. Then, it issues a subset of transmit power parameter rollbacks to the test cell group and evaluates it.

[0124] If compliance is not restored after all parameter rollbacks are completed, the server will upgrade the rollback level to the highest level and re-retrieve the target rollback parameter set corresponding to the highest level according to the highest level search rules. Then, the server will send the target rollback parameter set corresponding to the highest level to the test cell group in one go and wait for the network management to confirm the effect. The parameter set after the effect will be determined as the final wireless configuration parameter set.

[0125] When no rollback is triggered, the server writes the candidate wireless configuration parameter set into the stable parameter snapshot library as a stable snapshot. In the snapshot record, the effective cell identifier set is all cells in the test cell group, the effective time information is the current writing timestamp, the stable flag is true, and the parameter dimension information and parameter value hash digest used for snapshot consistency verification are retained to ensure the determinism of subsequent retrieval matching. Then, the candidate wireless configuration parameter set is determined as the final wireless configuration parameter set and output to the subsequent operation and maintenance interface for display and archiving.

[0126] In this specific embodiment, the pre-trained heterogeneous dynamic spatiotemporal graph neural network model and the pre-trained feasibility classifier are generated by the model training server in the offline training environment and written into the model repository for use in steps S3 and S4. The model training server uses 30 consecutive days of historical running data as the training data source and divides each day into continuous historical time windows according to the same rules as in step S1, with each historical time window having a fixed length of 300 seconds. Within each historical time window, IoT measurement data and the set of wireless configuration parameters in effect at that time are collected synchronously, and time alignment and data integrity verification are performed respectively to obtain the IoT measurement dataset and the set of wireless configuration parameters for training.

[0127] The model training server calculates the actual service performance index vector of each 5G cell within the historical time window according to the IoT measurement dataset used for training, and maintains the same index direction and component order as in step S3. The actual service performance index vector is fixed as a three-dimensional vector, and the three-dimensional components are, in order, the negative median end-to-end latency, the negative median packet loss rate, and the median throughput. The negative signs of end-to-end latency and packet loss rate are used to ensure the unified comparison direction of "the larger the value, the better the performance".

[0128] The model training server determines the service level of each 5G cell within the historical time window based on the IoT measurement dataset used for training, and takes the minimum service level of the service terminal in the cell as the service level of the cell. Then, it reads the service continuity threshold vector associated with the service level and compares the actual service performance index vector with the service continuity threshold vector component by component to generate a feasibility label. When each component of the actual service performance index vector is not less than the corresponding component of the corresponding service continuity threshold vector, the feasibility label is set to 1; otherwise, it is set to 0.

[0129] The model training server extracts scene features based on the IoT measurement dataset used for training and fixes them, including the start and stop status of production line equipment, the working cycle of robotic arms, the density of automated guided vehicles or forklifts, the speed of automated guided vehicles or forklifts, the position of overhead cranes or hoists, and the stacking status of metal coils or large tooling. The discrete state quantities are mapped to numerical codes through a preset dictionary and the continuous quantities are linearly normalized according to a preset range, thereby forming a scene feature vector with a dimension of 10, which is used for subsequent training of feasible classifiers.

[0130] Within each historical time window, the model training server generates a sequence of heterogeneous dynamic spatiotemporal graphs for training according to the same node and edge construction rules as in step S2, and maintains a time slice number of 30 and a time slice duration of 10 s. The sequence of heterogeneous dynamic spatiotemporal graphs for training is fixed to include IoT terminal nodes, 5G cell nodes and environmental nodes, and the edge types are fixed to include connection edges, neighboring edges and propagation influence edges. Based on the scene features, the gating coefficient corresponding to each time slice is calculated, and the edge weight gating mechanism is updated on the connection edge features and neighboring edge features to ensure that the graph semantics of the training input and the output of step S2 are consistent.

[0131] The structure and parameter values ​​of the heterogeneous dynamic spatiotemporal graph neural network model are completely consistent with those in step S3 and are fixed. It consists of two layers of relational perception graph attention layers, a gated recurrent unit time encoder, and two output heads of two-layer multilayer perceptrons. It outputs the mean vector and standard deviation vector of the predicted service performance indicators for each 5G cell. The model training server uses Gaussian negative log-likelihood as the training loss and sums the loss function of the heterogeneous dynamic spatiotemporal graph neural network model by summing the loss function of all historical time windows, all 5G cells, and the three-dimensional indicator components.

[0132] ;

[0133] in This represents the training loss of a heterogeneous dynamic spatiotemporal graph neural network model. Indicates the historical time window index and This represents the total number of historical time windows used for training. Indicates 5G cell index and This represents the total number of 5G cells in the factory area, which is fixed in this embodiment. Represents the component index of business performance metrics and Indicates the negative end-to-end delay component and Indicates the negative packet loss rate component and Represents throughput components, Indicates the cell index is The 5G cell indexed in the historical time window is The first of the actual business performance index vectors One portion, This indicates that the heterogeneous dynamic spatiotemporal graph neural network model indexes the cells as follows: The 5G cell indexed in the historical time window is The first business performance indicator prediction mean vector One portion, This indicates that the heterogeneous dynamic spatiotemporal graph neural network model indexes the cells as follows: The 5G cell indexed in the historical time window is The first standard deviation vector of the business performance index prediction Each component is guaranteed by the softplus activation of the output head and the numerical stability term. ;

[0134] The model training server uses the Adam optimizer to iteratively train the heterogeneous dynamic spatiotemporal graph neural network model with a fixed learning rate. Furthermore, the batch size is fixed at 8 historical time windows, and all historical time windows are divided into training set proportions according to their chronological order. Validation set ratio Percentage of test set When the validation set Training is terminated when the loss does not decrease after 5 consecutive training rounds, and the model parameters with the lowest loss on the validation set are solidified into the pre-trained heterogeneous dynamic spatiotemporal graph neural network model.

[0135] In terms of feasibility classifier training, the model training server extracts five parameter components from the training wireless configuration parameter set at the cell granularity, consistent with step S4, and concatenates them with the scene feature vectors of the corresponding historical time windows to form a training sample feature vector with a dimension of 15. The feasibility label is then used as a supervision signal to train a three-layer multi-layer perceptron structure feasibility classifier. Its input layer dimension is fixed at 15, the number of neurons in the first hidden layer is fixed at 64, the number of neurons in the second hidden layer is fixed at 32, the number of neurons in the output layer is fixed at 1, and the output layer activation function is fixed at Sigmoid, outputting a feasibility probability. The training loss is fixed at binary cross-entropy, the optimizer is fixed at Adam, and the learning rate is fixed at a certain value. Furthermore, the batch size is fixed at 256, and the same training set, validation set, and test set partitioning and early stopping rules as the heterogeneous dynamic spatiotemporal graph neural network model are used to obtain a pre-trained feasibility classifier. After the two models are trained, the model training server calculates the component-wise ratio sequence of "the absolute value of the deviation between the actual business performance index vector and the predicted mean vector divided by the predicted standard deviation vector" based on the validation set, and takes the result for each business level. The quantiles generate an uncertainty calibration coefficient vector and write it into the uncertainty calibration table for use in step S3.

[0136] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0137] This invention directly addresses the aforementioned technical problem through a closed-loop technical path of "predictive evaluation plus security constraints plus online verification plus automatic rollback": First, a heterogeneous dynamic spatiotemporal graph sequence of terminals and cells is constructed based on IoT terminal measurement data. A heterogeneous dynamic spatiotemporal graph neural network is used to predict service performance indicators while simultaneously providing uncertainty. Then, confidence coefficients are determined based on service levels, and the lower confidence bound of service performance indicators is calculated. This lower confidence bound is compared with a service continuity threshold vector, thus forming a safety boundary for service continuity through conservative estimation. This ensures that subsequent parameter searches are only performed on candidate solutions that meet safety conditions, reducing the probability of automatic parameter tuning affecting production service continuity. Furthermore, a comparative sequential statistical test between the experimental cell group and the control cell group is used to rapidly verify the actual effectiveness of the candidate parameters online. If the test fails or the threshold is not met, a rollback is triggered, ensuring timely restoration of stable configuration in abnormal situations. This achieves the technical effects of "controllable optimization, automatic verification, and recoverable failure."

[0138] Compared to general automatic parameter optimization schemes, this invention makes several improvements to the algorithm structure to further enhance the aforementioned technical effects, specifically addressing factory scenarios and business continuity requirements: First, it employs heterogeneous dynamic spatiotemporal graph modeling and introduces a scenario state gating mechanism, enabling the model to characterize the impact of terminal location, connection relationships, and production line status changes on propagation conditions, reducing prediction bias caused by sudden scenario changes; Second, it constructs a lower confidence boundary safety boundary using prediction uncertainty as input, and can calibrate the uncertainty to make safety judgments more consistent with real risks; Third, it introduces feasible probability constraints from a feasibility classifier and parameter change magnitude constraints in constrained Bayesian optimization, balancing search efficiency and change controllability, and avoiding business fluctuations caused by large-scale parameter tuning; Fourth, it proposes comparative sequential verification, hierarchical rollback, and stable snapshot library management, enabling rapid pass or rapid termination during the verification phase, and reverting to historical stable parameter sets according to levels in case of anomalies, thereby improving the guarantee capability of production and business continuity from a mechanism perspective.

Claims

1. An automatic optimization method for plant site 5G coverage based on Internet of Things, characterized in that, include: S1. Collect IoT measurement data reported by IoT terminals in the factory area, including service level or service type, and obtain the initial wireless configuration parameter set and stable parameter snapshot library corresponding to the 5G cell. S2. Construct a heterogeneous dynamic spatiotemporal graph sequence representing IoT terminals and 5G cells and their relationships based on IoT measurement data; S3. Call the pre-trained heterogeneous dynamic spatiotemporal graph neural network model to infer the heterogeneous dynamic spatiotemporal graph sequence, obtain the predicted values ​​and uncertainties of business performance indicators, determine the confidence coefficient according to the business level, calculate the lower confidence bound, and compare the lower confidence bound with the preset business continuity threshold vector to determine the business continuity security conditions. S4. Use constrained Bayesian optimization iteration to generate a set of wireless configuration parameters to be evaluated, and apply feasibility discrimination constraints and change magnitude constraints. Call the pre-trained feasibility classifier to output the feasibility probability. The feasibility discrimination constraints include that the feasibility probability is not less than a preset feasibility threshold, and the change magnitude constraints include that the parameter change amount is not greater than a preset change magnitude threshold. Calculate the lower confidence bound for the set of wireless configuration parameters to be evaluated that meet the constraints and determine whether the service continuity security conditions are met. Determine the candidate set of wireless configuration parameters based on the acquisition function. S5. Determine the test cell group based on the candidate radio configuration parameter set and the control cell group based on the initial radio configuration parameter set. Send the candidate radio configuration parameter set to the test cell group and keep the initial radio configuration parameter set in the control cell group. Collect the corresponding IoT measurement data during the verification period, perform a comparative sequential statistical test to obtain the sequential test result, and determine the compliance of the service performance indicators of the test cell group with the service continuity threshold vector to obtain the compliance judgment result, so as to form the verification result. S6. Determine the rollback level based on the verification results. When the sequential test result indicates failure or the compliance judgment result indicates non-compliance, select the target rollback parameter set from the stable parameter snapshot library according to the rollback level and send it to the test cell group. When the sequential test result indicates success and the compliance judgment result indicates compliance, write the candidate radio configuration parameter set into the stable parameter snapshot library.

2. The method of claim 1, wherein the method further comprises: S1 includes: Set a preset time window in the network management server, and receive IoT measurement data reported by IoT terminals in the factory area within the preset time window; The IoT measurement data includes terminal identifier, terminal location information, timestamp, wireless measurement indicators, service performance indicators, and service level or service type. The IoT measurement data undergoes time alignment and data integrity verification, discarding IoT measurement data that does not meet preset integrity conditions, and forming IoT measurement data for subsequent processing. The initial wireless configuration parameter set corresponding to the 5G cell establishing a wireless connection with the IoT terminal is read; the initial wireless configuration parameter set includes transmit power parameters, beam parameters, and handover parameters. A stable parameter snapshot library is read; the stable parameter snapshot library stores at least one historically stable wireless configuration parameter set and the effective cell identifier and effective time information corresponding to the historically stable wireless configuration parameter set. A preset confidence coefficient table is read; the preset confidence coefficient table stores the correspondence between service level and confidence coefficient. Output the IoT measurement data, the initial wireless configuration parameter set, the stable parameter snapshot library, and the pre-set confidence coefficient table. 3.The factory 5G coverage automatic optimization method based on the Internet of Things according to claim 1, wherein S2 include: The preset time window is divided into multiple consecutive time slices, and the set of IoT terminals and the set of 5G cells corresponding to the time slice are determined based on the IoT measurement data within each time slice. IoT terminal nodes are established for the IoT terminal set, and 5G cell nodes are established for the 5G cell set; Based on the connection relationship between the terminal and the 5G cell in the IoT measurement data, a connection edge is established between the corresponding IoT terminal node and the corresponding 5G cell node, and connection edge features are configured for the connection edge according to the wireless measurement indicators. Based on the terminal location information in the IoT measurement data, neighboring edges are established between IoT terminal nodes that meet preset proximity conditions, and neighboring edge features are configured for the neighboring edges based on the terminal location information. The cell node characteristics are configured for the 5G cell node based on the initial wireless configuration parameter set; the gating coefficients corresponding to each time slice are calculated based on the scene state characteristics in the IoT measurement data, and the gating coefficients are used to characterize the impact of scene state on changes in wireless propagation conditions. The gating coefficient is used to update the connecting edge features of the connecting edge and the neighboring edge features of the neighboring edge to form a time-varying edge weight gating mechanism. The graphs constructed from each time slice and updated by the edge weight gating mechanism are combined in chronological order to obtain the heterogeneous dynamic spatiotemporal graph sequence.

4. The method for automatic optimization of 5G coverage in a factory area based on the Internet of Things according to claim 1, characterized in that, S3 includes: The pre-trained heterogeneous dynamic spatiotemporal graph neural network model is invoked to infer the heterogeneous dynamic spatiotemporal graph sequence, outputting the predicted mean vector and the predicted standard deviation vector of service performance indicators for each 5G cell. A confidence coefficient corresponding to the service level is determined from a pre-set confidence coefficient table. Based on the predicted mean vector, the predicted standard deviation vector, and the confidence coefficient, a lower confidence bound vector is calculated for each 5G cell. Each component of the lower confidence bound vector is the predicted mean of the corresponding service performance indicator minus the product of the confidence coefficient and the predicted standard deviation. The judgment result of the service continuity security condition is determined based on the comparison result between the lower confidence bound vector and the service continuity threshold vector. When each component of the lower confidence bound vector is not less than the corresponding component of the service continuity threshold vector, the judgment result of the service continuity security condition is determined to be satisfied; otherwise, it is determined to be unsatisfied. The predicted mean vector, the predicted standard deviation vector, the lower confidence bound vector, and the judgment result of the service continuity security condition are output.

5. The method for automatic optimization of 5G coverage in a factory area based on the Internet of Things according to claim 1, characterized in that, S4 include: The parameter search range is determined by the initial set of wireless configuration parameters, and a set of candidate wireless configuration parameters to be evaluated is generated iteratively within the parameter search range by constrained Bayesian optimization. In each iteration, scene features are extracted based on the IoT measurement data, and the feasibility classifier is called to output the feasibility probability based on the scene features and the set of candidate wireless configuration parameters to be evaluated. The feasibility discrimination constraint is determined based on the comparison result between the feasibility probability and the preset feasibility threshold. The parameter change index is calculated based on the parameter change amount of the candidate wireless configuration parameter set to be evaluated and the initial wireless configuration parameter set. The change magnitude constraint is determined based on the comparison result of the parameter change index and the preset change magnitude threshold. When both the feasibility discrimination constraint and the change magnitude constraint are satisfied, a corresponding heterogeneous dynamic spatiotemporal graph sequence is generated based on the candidate wireless configuration parameter set to be evaluated. The heterogeneous dynamic spatiotemporal graph neural network model is called to output the corresponding service performance index prediction mean vector and the corresponding service performance index prediction standard deviation vector. Then, the corresponding lower confidence bound vector is calculated. The service continuity threshold vector is compared with the corresponding lower confidence bound vector to determine whether the candidate wireless configuration parameter set to be evaluated meets the service continuity security conditions. Based on the target service performance index improvement rules, a target value is calculated for the set of candidate wireless configuration parameters to be evaluated that meet the service continuity and security conditions, and a collection function value is calculated based on the target value, the feasibility probability, and the parameter change index. The set of candidate wireless configuration parameters with the largest acquisition function value is used as the iterative output, and when the preset iteration termination condition is met, the set of candidate wireless configuration parameters with the largest acquisition function value during the iteration process is determined as the candidate wireless configuration parameter set.

6. The method for automatic optimization of 5G coverage in a factory area based on the Internet of Things according to claim 1, characterized in that, S5 include: A test cell group is determined based on the candidate radio configuration parameter set, and a control cell group is determined based on the initial radio configuration parameter set. The test cell group and the control cell group meet preset comparability conditions, which include the same cell type and adjacent or overlapping coverage areas. The candidate radio configuration parameter set is sent to the test cell group, while the control cell group uses the initial radio configuration parameter set. The verification period is divided into multiple consecutive verification sub-time windows, and IoT measurement data corresponding to the test cell group and the control cell group are collected in each verification sub-time window. Based on the IoT measurement data in each verification sub-time window, the statistical values ​​of the service performance indicators of the test cell group and the control cell group are calculated, and the difference between the two is used to construct a service performance indicator difference sequence. A sequential statistical test is performed on the differential sequence of the business performance indicators. The sequential statistical test includes updating the test statistic at the end of each verification sub-time window and comparing the test statistic with a pass threshold and a fail threshold. When the test statistic reaches the pass threshold, the sequential test result is determined to be pass. When the test statistic reaches the fail threshold, the sequential test result is determined to be fail. When the verification period ends and the test statistic does not reach the pass threshold and the fail threshold, the sequential test result is determined according to a preset judgment rule. The minimum value vector of service performance indicators for the test cell group during the verification period is calculated based on the IoT measurement data, and the minimum value vector of service performance indicators is compared with the service continuity threshold vector to obtain a compliance judgment result. Output the verification results, including the sequential verification results and the compliance determination results.

7. The method for automatic optimization of 5G coverage in a factory area based on the Internet of Things according to claim 1, characterized in that, S6 include: The rollback level is determined based on the sequential test results and compliance judgment results in the verification results. Specifically, when the compliance judgment result indicates non-compliance, the rollback level is determined to be the highest level. When the compliance judgment result indicates compliance and the sequential test result indicates failure, the rollback level is determined to be the second highest level. When the compliance judgment result indicates compliance and the sequential test result indicates success, the rollback is determined not to be triggered. When a rollback is triggered, a set of historically stable radio configuration parameters matching the test cell group is retrieved from the stable parameter snapshot library according to the rollback level, and the retrieved set of historically stable radio configuration parameters is determined as the target rollback parameter set. The target rollback parameter set corresponding to the highest level is the set of historically stable radio configuration parameters that was most recently effective and covers all cells in the test cell group in the stable parameter snapshot library, and the target rollback parameter set corresponding to the second highest level is the set of historically stable radio configuration parameters in the stable parameter snapshot library that has the same parameter dimension as the candidate radio configuration parameter set and has the smallest parameter change amount compared to the candidate radio configuration parameter set. The target rollback parameter set is sent to the test cell group to complete the rollback, and the rollback radio configuration parameter set is determined as the final radio configuration parameter set. When no rollback is triggered, the candidate wireless configuration parameter set is written into the stable parameter snapshot library as a new stable parameter snapshot, and the candidate wireless configuration parameter set is determined as the final wireless configuration parameter set.

8. The method for automatic optimization of 5G coverage in factory areas based on the Internet of Things according to claim 1, characterized in that, The pre-trained heterogeneous dynamic spatiotemporal graph neural network model and the pre-trained feasibility classifier are obtained through the following training method: Collect IoT measurement data within multiple historical time windows, and simultaneously collect the wireless configuration parameter set corresponding to the historical time windows; determine the actual service performance index vector of each 5G cell within each historical time window based on the IoT measurement data, and generate a feasibility label based on the comparison result of the actual service performance index vector and the service continuity threshold vector; extract scene features based on the IoT measurement data, the scene features including scene state features; Based on the terminal distribution, terminal mobility information, and connection relationships between terminals and 5G cells represented by the IoT measurement data, a training heterogeneous dynamic spatiotemporal graph sequence is constructed. Nodes include IoT terminal nodes and 5G cell nodes, and edges include connecting edges and neighboring edges. The edge weights of the edges are updated over time using an edge weight gating mechanism based on the scene state characteristics. The heterogeneous dynamic spatiotemporal graph neural network model is driven to output the predicted mean vector and predicted standard deviation vector of service performance indicators to the training heterogeneous dynamic spatiotemporal graph sequence. A loss function is constructed based on the actual service performance indicator vector, the predicted mean vector, and the predicted standard deviation vector. The model parameters are iteratively updated to make the loss function converge, resulting in the trained heterogeneous dynamic spatiotemporal graph neural network model. The wireless configuration parameter set is combined with the scene features to form training sample features. The feasibility classifier is driven to output feasibility probabilities based on the training sample features. A classification loss function is constructed based on the feasibility probabilities and the feasibility labels. The classifier parameters are iteratively updated to make the classification loss function converge, resulting in the trained feasibility classifier.

9. The method for automatic optimization of 5G coverage in a factory area based on the Internet of Things according to claim 2, characterized in that, S1 further includes: reading or configuring a service continuity threshold vector in a network management server, wherein the service continuity threshold vector includes at least a latency threshold, a packet loss rate threshold, and a throughput threshold, and the service continuity threshold vector is stored in association with the service level or service type.

10. The method for automatic optimization of 5G coverage in a factory area based on the Internet of Things according to claim 3, characterized in that, The heterogeneous dynamic spatiotemporal graph sequence further includes environmental nodes, which include at least one of building nodes, metal obstruction nodes, and environmental sensor nodes. A propagation influence edge is established between the environmental nodes and the 5G cell nodes. The edge features of the propagation influence edge include at least one of obstruction indication, material attenuation coefficient, and temperature, humidity, or dust concentration measured by the environmental sensor.

11. The method for automatic optimization of 5G coverage in a factory area based on the Internet of Things according to claim 3, characterized in that, The scene state features include at least one of the following: the start / stop status of production line equipment, the cycle time of robotic arm operation, the density and speed of automated guided vehicles or forklifts, the position of overhead cranes or hoists, and the stacking status of metal coils or large tooling. A gating coefficient is calculated based on the scene state features, such that when the scene state features indicate enhanced occlusion or drastic changes in the reflection environment, the gating coefficient reduces the weight of the connecting edge features or adjacent edge features to suppress the prediction offset caused by abrupt scene changes.

12. The method for automatic optimization of 5G coverage in a factory area based on the Internet of Things according to claim 4, characterized in that, Before calculating the lower confidence bound, the method further includes: performing uncertainty calibration on the prediction standard deviation vector of the business performance index, wherein the uncertainty calibration includes quantile calibration or conformal prediction calibration based on the prediction error within the historical time window, and calculating the lower confidence bound vector based on the calibrated prediction standard deviation vector.

13. The method for automatic optimization of 5G coverage in a factory area based on the Internet of Things according to claim 7, characterized in that, When a rollback is triggered, the method further includes: rolling back the candidate wireless configuration parameter set parameter by parameter according to a preset parameter priority, wherein the parameter priority includes rolling back the switching parameter before rolling back the beam parameter and rolling back the beam parameter before rolling back the transmit power parameter; collecting IoT measurement data after each parameter rollback and recalculating the minimum value vector of the service performance index after each rollback; and upgrading the rollback level to the highest level and using the target rollback parameter set corresponding to the highest level to complete the rollback.

14. The method for automatic optimization of 5G coverage in a factory area based on the Internet of Things according to claim 1, characterized in that, Before executing S5, the process includes: constructing a wireless propagation simulation model based on the three-dimensional map of the factory area, building material information, and the location information of the IoT terminal, and performing a coverage pre-simulation on the candidate wireless configuration parameter set to obtain a pre-simulation coverage index; when the pre-simulation coverage index does not meet the preset pre-simulation conditions, the distribution of the candidate wireless configuration parameter set to the test cell group is prohibited, and the process returns to execute S4 to update the candidate wireless configuration parameter set.

15. An IoT-based automatic 5G coverage optimization system for factory areas, used to execute any one of the IoT-based automatic 5G coverage optimization methods for factory areas according to claims 1 to 14, comprising: The data acquisition module is used to collect IoT measurement data reported by IoT terminals in the factory area and obtain the service level or service type, and read the initial wireless configuration parameter set and stable parameter snapshot library of 5G cells. The graph construction module is used to construct a heterogeneous dynamic spatiotemporal graph sequence representing the relationship between IoT terminals and 5G cells based on the IoT measurement data. The reasoning and security boundary module is used to call the pre-trained heterogeneous dynamic spatiotemporal graph neural network model to output the predicted value of business performance indicators and their uncertainty, determine the confidence coefficient based on the business level to calculate the lower confidence boundary, and compare the lower confidence boundary with the business continuity threshold vector to determine the business continuity security conditions. The constrained Bayesian optimization module is used to iteratively generate a set of wireless configuration parameters to be evaluated under feasibility discrimination constraints and change magnitude constraints. It determines the set of candidate wireless configuration parameters based on the lower confidence bound and the acquisition function. The feasibility discrimination constraints are determined by the feasibility probability output by the pre-trained feasibility classifier and the preset feasibility threshold. The comparative sequential verification module is used to establish a test cell group based on the candidate radio configuration parameter set and a control cell group based on the initial radio configuration parameter set, distribute the candidate radio configuration parameter set and collect verification data, perform comparative sequential statistical tests and generate verification results. The graded rollback and snapshot management module is used to determine the rollback level based on the verification results and select the target rollback parameter set from the stable parameter snapshot library for rollback, or to write the candidate wireless configuration parameter set into the stable parameter snapshot library when the verification is passed.

Images