An artificial intelligence-based heterogeneous network space multi-level mobile internet of things cloud system

Abstract

The application relates to the technical field of mobile communication and Internet of Things, and discloses a heterogeneous network space multi-level mobile Internet of Things cloud system based on artificial intelligence, which comprises the following modules: a data acquisition module that acquires real-time state data of a mobile Internet of Things terminal and generates a standardized network-service joint situation vector; an endogenous interference sensing module that generates a network interference correlation graph based on the vector; a service-network dynamic decoupling module that inversely reconstructs terminal service characteristics according to the correlation graph and decouples a key network state sequence; a service intention abstraction module that constructs a heterogeneous network cooperative state target according to the sequence; a cooperative scheduling strategy generation module that generates a multi-dimensional cooperative scheduling instruction by using a cooperative decision model; a network resource self-growth module that preplays a state sequence evolution after the instruction and constructs a compensation network plane; and an instruction distribution and verification module that executes the instruction of the fusion compensation logic and updates the correlation graph. The application effectively solves the problems of mobile Internet of Things heterogeneous network interference and scheduling lag.

Description

Technical Field

[0001] This invention relates to the fields of mobile communication and Internet of Things (IoT) technology, and in particular to a multi-layered mobile IoT cloud system based on artificial intelligence in a heterogeneous network space. Background Technology

[0002] In mobile IoT scenarios, terminals typically need to access various heterogeneous networks such as cellular networks, Wi-Fi, and satellite networks to ensure service continuity. Existing network management systems often focus on resource optimization within a single network or use simple load balancing strategies for network selection, lacking in-depth perception and quantitative modeling of complex endogenous interference between heterogeneous networks. Existing solutions usually treat service requirements and network states as independent inputs, adopting a static or passive response mode of "service requirements first, then network resources are matched." This tightly coupled architecture cannot proactively decouple and predict service intentions by reverse analyzing the network state sequence before service requirements emerge, resulting in delayed resource scheduling and difficulty in meeting the low latency and high reliability requirements of highly dynamic mobile IoT services.

[0003] Furthermore, current collaborative scheduling strategies are directly deployed and executed after generation, lacking accurate prediction of network state evolution after execution, and unable to proactively construct compensatory resources to address resource bottlenecks discovered during the prediction process. The system can only detect and repair performance degradation after it occurs, resulting in slow response and a lack of a closed-loop mechanism to feed back actual execution deviations to update the internal interference model, hindering continuous evolution of system intelligence. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a multi-layered mobile IoT cloud system based on artificial intelligence in a heterogeneous network space. By constructing a dynamic network interference correlation model based on a spatiotemporal graph convolutional network and a service-network reverse decoupling mechanism, this invention can deeply understand and quantify the endogenous interference between heterogeneous networks and its dynamic correlation with services and mobility, and proactively decouple service intent and collaborative carrying requirements. Furthermore, by introducing a scheduling pre-playing and compensation network plane dynamic construction mechanism based on virtual simulation, it achieves pre-verification of collaborative scheduling strategies and proactive compensation for resource bottlenecks, and continuously updates the interference cognition model based on execution verification feedback. This effectively overcomes the problems of inefficient collaboration and rigid resource adaptation caused by isolated interference cognition, passive service response, and open-loop scheduling execution in existing technologies. It achieves a fundamental transformation from isolated optimization to correlated cognition, from passive response to intent-driven, and from open-loop execution to autonomous evolution, thereby ensuring the quality of service and resource utilization efficiency of mobile IoT services in a highly dynamic heterogeneous network environment.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention provides a multi-layered mobile Internet of Things cloud system based on artificial intelligence in a heterogeneous cyberspace, comprising the following steps:

[0008] Data acquisition module: used to acquire real-time status data of mobile IoT terminals, perform feature decoupling and standardization mapping on the real-time status data, and generate a standardized network-service joint situational vector;

[0009] Endogenous interference perception module: Based on the network-service joint situational vector, it identifies and quantifies the endogenous interference intensity between any two heterogeneous networks through a spatiotemporal graph convolutional network, analyzes the dynamic correlation between the endogenous interference intensity and terminal service characteristics and terminal mobility characteristics, and generates a network interference correlation graph.

[0010] Service-Network Dynamic Decoupling Module: Used to reverse reconstruct the terminal service characteristics based on the network interference correlation diagram, and decouple the key network state sequence;

[0011] Business Intent Abstraction Module: Used to extract the terminal service characteristics based on the key network state sequence, obtain the collaborative bearing requirements of the heterogeneous network space, and construct the required collaborative state target of the heterogeneous network based on the collaborative bearing requirements.

[0012] The collaborative scheduling strategy generation module is used to generate multi-dimensional collaborative scheduling instructions based on the collaborative state objective as the optimization guide, the network-service joint situation vector and the key network state sequence as state inputs, and the collaborative decision model.

[0013] Network resource self-growth module: used to pre-simulate the evolution trajectory of the key network state sequence after the instruction is executed, and dynamically construct a compensation network plane for possible resource bottlenecks;

[0014] Instruction distribution and verification module: used to execute multi-dimensional collaborative scheduling instructions that integrate the compensation network plane logic, monitor the real-time status data generated after execution, match and analyze the actual state evolution trajectory with the virtual pre-simulation trajectory, lock the interference transmission chain corresponding to the root cause of the deviation, and update the network interference correlation diagram.

[0015] Furthermore, real-time status data of mobile IoT terminals is acquired, and feature decoupling and standardization mapping are performed on the real-time status data to generate a standardized network-service joint situational vector, as follows:

[0016] By using sensors and network interfaces deployed on mobile IoT terminals, real-time status data of the terminals in heterogeneous network spaces can be collected.

[0017] The real-time status data is subjected to feature decoupling processing, and service feature subsets, network feature subsets, and mobility feature subsets are obtained based on data attributes;

[0018] The service feature subset, the network feature subset, and the mobility feature subset are standardized respectively, and the data values ​​in each subset are mapped to a unified numerical range to obtain the corresponding standardized service feature set, standardized network feature set, and standardized mobility feature set.

[0019] The standardized service feature set, the standardized network feature set, and the standardized mobile feature set are vectorized and concatenated to generate the standardized network-service joint situational vector.

[0020] Furthermore, based on the network-service joint situational awareness vector, the intrinsic interference intensity between any two heterogeneous networks is identified and quantified through a spatiotemporal graph convolutional network. The dynamic correlation between the intrinsic interference intensity and terminal service characteristics and terminal mobility characteristics is analyzed to generate a network interference correlation graph, including:

[0021] Each heterogeneous network is defined as a graph node, and the network-service joint situational vector is assigned to the corresponding graph node as a node feature to construct a heterogeneous network topology graph.

[0022] Based on the physical coverage overlap and authorized spectrum interval between networks, calculate the connection weight between any two graph nodes, establish edges for node pairs with connection weights greater than zero, and generate a weighted initial network graph.

[0023] The initial network graph and the node feature sequences on continuous time slices are input into the spatiotemporal graph convolutional network, and the spatiotemporal coupling feature tensor of each node pair is output.

[0024] The spatiotemporal coupling feature tensor is input into a fully connected layer, and regression is performed to obtain a scalar sequence of endogenous interference intensity between each pair of network nodes within the corresponding time window.

[0025] The service data rate time series and the terminal distribution density time series are extracted from the standardized service feature set and respectively, and used as key feature sequences.

[0026] Calculate the Pearson correlation coefficient between the scalar sequence of endogenous interference intensity and each key feature sequence, and generate a correlation coefficient matrix with network node pairs as rows and key features as columns;

[0027] The network interference correlation graph is constructed by using the mean of the scalar sequence of endogenous interference intensity as the edge weight between corresponding network node pairs and the correlation coefficient matrix as the correlation attribute of the corresponding edges.

[0028] Furthermore, the calculation of the Pearson correlation coefficient between the scalar sequence of endogenous interference intensity and each key feature sequence, generating a correlation coefficient matrix with network node pairs as rows and key features as columns, includes:

[0029] For each pair of network nodes, its scalar sequence of endogenous interference intensity is treated as a variable;

[0030] The service data rate time series and the terminal distribution density time series are respectively used as two other variables;

[0031] Based on the Pearson correlation coefficient formula, the correlation coefficient between the scalar sequence of endogenous interference intensity and the time series of service data rate, as well as the correlation coefficient between the scalar sequence of endogenous interference intensity and the time series of terminal distribution density, are calculated.

[0032] The two calculated correlation coefficients are arranged in order into a row vector, where each network node corresponds to one row vector;

[0033] Stack the row vectors corresponding to all network nodes to form the correlation coefficient matrix.

[0034] Furthermore, based on the network interference correlation graph, the terminal service characteristics are reverse-reconstructed to decouple the key network state sequence, as follows:

[0035] Based on the scalar sequence of endogenous interference intensity corresponding to each edge in the network interference correlation graph, the statistical characteristics of each sequence within a preset time window are calculated, and an interference feature matrix with heterogeneous networks as nodes is constructed.

[0036] Using the interference feature matrix as graph structure data and the standardized business feature set as input node features, a graph neural network is constructed.

[0037] The standardized business feature set is input into the graph neural network for forward propagation, and the reconstructed business feature vector of each heterogeneous network node is output.

[0038] Calculate the element-wise absolute difference between the standardized business feature set and the reconstructed business feature vector to obtain a sequence of business feature residual vectors;

[0039] Principal component analysis is performed on the time dimension of the business feature residual vector sequence to extract the principal component components. The principal component components are then arranged in chronological order to generate the key network state sequence.

[0040] Furthermore, based on the key network state sequence, the terminal service characteristics are extracted to obtain the collaborative bearing requirements of the heterogeneous network space. Based on the collaborative bearing requirements, the required collaborative state target of the heterogeneous network is constructed as follows:

[0041] Spectral analysis was performed on the key network state sequence to extract its power spectral density distribution in different frequency bands;

[0042] The frequency band components in the power spectral density distribution whose amplitude exceeds a preset threshold are identified as the dominant fluctuation modes;

[0043] Calculate the frequency and phase corresponding to the dominant fluctuation pattern, and perform time-frequency domain matching with the periodic business indicator sequence extracted from the standardized business feature set to determine the terminal business type and key performance indicators affected by the dominant fluctuation pattern.

[0044] Using historical demand baselines as a benchmark, dynamic demand ranges are constructed for each key performance indicator based on the strength of the dominant fluctuation pattern.

[0045] Transform the demand ranges of each key performance indicator into constraints and establish a set of constraints.

[0046] To address the transmission capability characteristics of different heterogeneous networks, the construction problem of the cooperative state objective is modeled as a constrained optimization problem.

[0047] The optimization problem is solved using the Lagrange multiplier method, and the optimal set of performance parameters required for each heterogeneous network to meet the collaborative carrying requirements is obtained, which serves as the collaborative state objective.

[0048] Furthermore, based on the aforementioned collaborative state objective as the optimization guide, and using the network-service joint situational vector and the key network state sequence as state inputs, a multi-dimensional collaborative scheduling instruction is generated using a collaborative decision-making model, as follows:

[0049] Align the network-service joint situational vector with the key network state sequence in the time dimension and concatenate them along the feature dimension to generate a comprehensive state input vector;

[0050] The comprehensive state input vector is input into a neural network model trained based on historical data, and the scheduling action parameters are output.

[0051] Using the joint performance boundary of the cooperative state objective as a constraint, numerical boundary checks and feasibility adjustments are performed on the scheduling action parameters to generate verification scheduling action parameters;

[0052] Based on the standard control interfaces of each network element in the heterogeneous network, the verification and scheduling action parameters are converted into specific spectrum reconfiguration instructions, transmit power update instructions, and access point switching instructions.

[0053] The spectrum reconfiguration command, transmit power update command, and access point switching command are combined to form the multi-dimensional collaborative scheduling command.

[0054] Furthermore, the evolution trajectory of the key network state sequence after the pre-executive command is executed, and a compensation network plane is dynamically constructed to address potential resource bottlenecks, as follows:

[0055] A virtual simulation environment is established based on the network interference correlation diagram and historical state data.

[0056] The multi-dimensional collaborative scheduling instructions are input into the virtual simulation environment to simulate the network operation process after the parameters of each network element are adjusted, and the estimated network performance index sequence during the simulation operation is output. The network performance index sequence is used as the virtual pre-simulation trajectory.

[0057] The time delay sequence, packet loss rate sequence, and bandwidth utilization rate sequence in the virtual pre-simulation trajectory are compared with the corresponding time delay upper limit, packet loss rate upper limit, and bandwidth utilization rate upper limit in the cooperative state target, respectively, on a time-by-time basis.

[0058] Identify the time period in which any performance index sequence value continuously exceeds its corresponding upper limit and the target network element with the highest load during that time period. Mark the time period and the target network element together as a resource bottleneck, and calculate the average difference between the performance index sequence value and its corresponding upper limit as the performance gap.

[0059] Based on the performance gap and the network type of the target network element, determine the type and specifications of the virtual resources that need to be supplemented;

[0060] In the virtual simulation environment, during the time period corresponding to the resource bottleneck, virtual network function nodes of appropriate specifications are deployed and configured for the affected service flows.

[0061] A logically isolated compensation network plane is formed by establishing connections between the virtual network functional nodes and the relevant physical networks. Further, a multi-dimensional collaborative scheduling instruction incorporating the logic of the compensation network plane is executed. Real-time status data generated after execution is monitored, and the actual state evolution trajectory is matched and analyzed with the virtual pre-simulation trajectory to pinpoint the interference propagation chain corresponding to the root cause of the deviation. Based on this, the network interference correlation graph is updated as follows:

[0062] The multi-dimensional collaborative scheduling command is executed, and the actual performance index data of each network element in the monitoring window are collected after the command takes effect through the network probe and terminal reporting mechanism.

[0063] The actual performance index data is time-aligned and the corresponding performance index sequence in the virtual pre-simulation trajectory generated by the network resource self-growth module is calculated;

[0064] Identify continuous time periods in which the difference exceeds a preset deviation threshold, and mark the performance index with the largest difference and its corresponding network element within that time period as significant performance deviation and the network element where the deviation occurred;

[0065] Starting from the network element where the deviation occurs, in the network interference association graph, based on the weight of the edges in the graph, iteratively backtrack to the upstream network node that is its neighbor and has a higher edge weight;

[0066] The edges traversed by the backtracking search are connected in sequence to form a path from an upstream network node to the network element where the deviation occurred. This path is identified as the interference propagation chain that caused the significant performance deviation, and the upstream network node is the root cause node of the deviation.

[0067] Based on the historical correlation strength of each edge in the interference propagation chain within the time period corresponding to the significant performance deviation, the weights of the corresponding edges in the network interference correlation graph are enhanced and updated.

[0068] (III) Beneficial Effects

[0069] The proposed heterogeneous network space multi-layered mobile IoT cloud system based on artificial intelligence has the following beneficial technical effects:

[0070] 1. This invention, by constructing an endogenous interference perception module and a service-network dynamic decoupling module, achieves for the first time the quantitative correlation of dynamic endogenous interference between heterogeneous networks and the inverse decoupling of service characteristics. Specifically, the system uses a spatiotemporal graph convolutional network to process the network-service joint situation vector, which not only quantifies the interference intensity between any two networks but also analyzes the dynamic correlation between the interference and features such as service rate and terminal density, forming a network interference correlation graph. Then, guided by this graph, a graph neural network is used to inversely reconstruct the service characteristics and extract the principal components from the residuals, decoupling the key network state sequences driving service changes. This mechanism overcomes the shortcomings of existing technologies that treat interference as isolated noise and cannot establish a dynamic correlation model between it and services / mobility.

[0071] 2. This invention achieves intelligent scheduling that moves from passive response to intent-driven, multi-dimensional collaboration through a business intent abstraction module and a collaborative scheduling strategy generation module. The system performs spectral analysis on the decoupled key network state sequences, identifies dominant fluctuation patterns, and matches them with business indicators to abstract collaborative carrying requirements and construct precise heterogeneous network collaborative state targets that meet these requirements. Guided by these targets, a collaborative decision model is used to generate multi-dimensional scheduling instructions that integrate spectrum, power, and access point switching. This method changes the rigid traditional "demand-matching" model, proactively anticipating business intents from network state evolution and directly driving the generation of scheduling actions with global collaborative targets. This makes resource allocation more forward-looking and overall optimal, effectively meeting the stringent and dynamically changing requirements of mobile IoT services for network performance.

[0072] 3. This invention constructs a complete autonomous closed loop of virtual pre-simulation, compensation construction, execution verification, and model update through a network resource self-growth module and an instruction distribution and verification module. Before execution, the system pre-simulates the effect of scheduling instructions in a virtual environment. For resource bottlenecks discovered during the pre-simulation, it dynamically constructs logically isolated compensation network planes and integrates the compensation logic with the main instructions. After execution, by comparing the actual trajectory with the pre-simulated trajectory, it identifies the interference propagation chain corresponding to the root cause of the deviation and updates the model parameters of the network interference correlation graph accordingly. This mechanism achieves a leap from open-loop execution to closed-loop autonomy. It not only significantly reduces the risk of resource bottlenecks and improves scheduling reliability through "pre-simulation-compensation," but also enables the system to learn from actual deviations and continuously optimize its inherent interference cognition model through "verification-update," overcoming the inherent defects of existing systems such as slow response and inability to autonomously evolve their intelligence level. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of a heterogeneous network space multi-layered mobile Internet of Things cloud system based on artificial intelligence, according to the present invention. Detailed Implementation

[0074] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0075] like Figure 1 The illustrated heterogeneous cyberspace multi-layered mobile IoT cloud system based on artificial intelligence includes the following steps:

[0076] Data acquisition module: used to acquire real-time status data of mobile IoT terminals, perform feature decoupling and standardization mapping on the real-time status data, and generate a standardized network-service joint situational vector;

[0077] In this embodiment, real-time status data of the mobile IoT terminal is acquired, and the real-time status data is subjected to feature decoupling and standardized mapping to generate a standardized network-service joint situational vector, as follows:

[0078] By using sensors and network interfaces deployed on mobile IoT terminals, real-time status data of the terminals in heterogeneous network spaces can be collected.

[0079] The real-time status data is subjected to feature decoupling processing, and service feature subsets, network feature subsets, and mobility feature subsets are obtained based on data attributes;

[0080] The service feature subset, the network feature subset, and the mobility feature subset are standardized respectively, and the data values ​​in each subset are mapped to a unified numerical range to obtain the corresponding standardized service feature set, standardized network feature set, and standardized mobility feature set.

[0081] The standardized service feature set, the standardized network feature set, and the standardized mobile feature set are vectorized and concatenated to generate the standardized network-service joint situational vector.

[0082] Preferably, the process of generating standardized network-service joint situational awareness vectors by the data acquisition module is described in detail. The core purpose of this process is to transform the diverse and heterogeneous raw monitoring data reported by mobile IoT terminals into digital feature vectors with standardized format, uniform scale, and ease of processing by subsequent artificial intelligence modules, thereby providing basic data support for the collaborative analysis of the entire system.

[0083] Specifically, the acquisition of real-time status data is accomplished collaboratively by the terminal's hardware sensing units and software protocol stack interfaces. The hardware sensing units include, but are not limited to, a performance counter for monitoring the terminal's application processor utilization and an inertial measurement unit for sensing the terminal's three-dimensional spatial acceleration and angular velocity. The software protocol stack interfaces involve application programming interfaces provided by the terminal's operating system and communication stack, used to acquire reference signal received power reported by the cellular network modem, channel busy rate reported by the wireless LAN driver, and data packet transmission intervals and sizes reported by upper-layer service applications. This real-time acquisition is executed by a data acquisition agent program deployed on the terminal side. This agent, triggered by a fixed sampling period or significant status change events, synchronously reads data from the aforementioned interfaces to form raw status data records containing timestamps.

[0084] More specifically, the collected real-time status data undergoes feature decoupling processing. Decoupling refers to automatically classifying mixed raw data records into different predefined categories based on the physical layer nature or logical layer to which the data belongs. In this embodiment, the data attribute separation operation is implemented through a pre-configured feature classification rule engine: this engine routes data to the corresponding processing channel based on the data field name, source interface identifier, and metadata tag.

[0085] For example, end-to-end latency, jitter, and packet loss rate fields sourced from "Service Application QoS Monitor" are assigned to the service characteristic channel; fields sourced from "L1 / L2 Layer Measurement Report" with names containing "RSRP," "SINR," and "Bandwidth" are assigned to the network characteristic channel; and fields sourced from "Location and Motion Service" with names containing "Latitude," "Longitude," and "Velocity" are assigned to the mobility characteristic channel. Through this operation, the raw data is structured into three independent datasets, corresponding to service characteristics, network characteristics, and mobility characteristics, respectively.

[0086] Next, the three datasets are individually standardized. Standardization is a data transformation process designed to eliminate analytical biases caused by differences in units and numerical ranges of different features. The process of mapping data values ​​within each subset to a unified numerical range is specifically performed using the min-max normalization algorithm.

[0087] For each numerical feature within each feature channel, the algorithm reads its maximum and minimum values ​​within a historical time window, and then uses a linear scaling formula to convert each reported real-time value to a range between zero and one. For non-numerical categorical features, one-hot encoding is used to convert them into binary vectors. After this processing, the values ​​of all features in the output dataset of each channel are constrained to the same numerical range, forming standardized business feature sets, standardized network feature sets, and standardized mobile feature sets, respectively.

[0088] Finally, vectorized concatenation is performed to generate the final standardized network-service joint situational awareness vector. Vectorized concatenation is a data aggregation operation, specifically referring to merging the three standardized feature sets mentioned above in memory according to a predefined, fixed feature dimension order. The order of arrangement is, for example, as follows: first arrange all feature values ​​in the standardized service feature set, then arrange all feature values ​​in the standardized network feature set, and finally arrange all feature values ​​in the standardized mobility feature set.

[0089] The resulting one-dimensional floating-point array after merging is the network-service joint situational awareness vector generated in this step. This vector, in a compact mathematical form, integrates the terminal's service demand intensity, network connection quality, and spatial mobility status at a specific moment, forming the input basis for subsequent modules to perform quantitative analysis and intelligent decision-making on the network-space joint situational awareness.

[0090] Endogenous interference perception module: Based on the network-service joint situational vector, it identifies and quantifies the endogenous interference intensity between any two heterogeneous networks through a spatiotemporal graph convolutional network, analyzes the dynamic correlation between the endogenous interference intensity and terminal service characteristics and terminal mobility characteristics, and generates a network interference correlation graph.

[0091] In this embodiment, based on the network-service joint situational awareness vector, the intrinsic interference intensity between any two heterogeneous networks is identified and quantified using a spatiotemporal graph convolutional network. The dynamic correlation between the intrinsic interference intensity and terminal service characteristics and terminal mobility characteristics is analyzed to generate a network interference correlation graph, including:

[0092] Each heterogeneous network is defined as a graph node, and the network-service joint situational vector is assigned to the corresponding graph node as a node feature to construct a heterogeneous network topology graph.

[0093] Based on the physical coverage overlap and authorized spectrum interval between networks, calculate the connection weight between any two graph nodes, establish edges for node pairs with connection weights greater than zero, and generate a weighted initial network graph.

[0094] The initial network graph and the node feature sequences on continuous time slices are input into the spatiotemporal graph convolutional network, and the spatiotemporal coupling feature tensor of each node pair is output.

[0095] The spatiotemporal coupling feature tensor is input into a fully connected layer, and regression is performed to obtain a scalar sequence of endogenous interference intensity between each pair of network nodes within the corresponding time window.

[0096] The service data rate time series and the terminal distribution density time series are extracted from the standardized service feature set and respectively, and used as key feature sequences.

[0097] Calculate the Pearson correlation coefficient between the scalar sequence of endogenous interference intensity and each key feature sequence, and generate a correlation coefficient matrix with network node pairs as rows and key features as columns;

[0098] The network interference correlation graph is constructed by using the mean of the scalar sequence of endogenous interference intensity as the edge weight between corresponding network node pairs and the correlation coefficient matrix as the correlation attribute of the corresponding edges.

[0099] Preferably, the specific implementation process for generating the network interference correlation graph is described in detail. The purpose of this process is to establish a quantitative model to dynamically characterize the mutual interference intensity between different heterogeneous networks due to resource sharing, and to reveal the statistical correlation between this interference and the characteristics of the services carried by the network and the terminal mobility mode, ultimately forming an interference relationship graph with rich attributes.

[0100] Specifically, the operation of defining each heterogeneous network managed in the cloud system as a graph node is performed. Here, each managed independent radio access network entity, such as a 4G cell, a 5G NR cell, or a set of Wi-Fi BSS access points, is abstracted as an independent graph node and assigned a unique identifier. Next, the operation of assigning the network-service joint situational vector to the corresponding graph node as a node feature is performed.

[0101] The specific implementation of this operation is as follows: Within each data collection cycle, the cloud system aggregates the network-service joint situational vectors reported by all mobile IoT terminals currently attached to the heterogeneous network; subsequently, it calculates the average value of these vectors at the terminal level to generate an aggregated feature vector representing the current overall average situational state of the network; finally, it binds and stores this aggregated feature vector with the identifier of the network node as the instantaneous feature of that node. The action of constructing the heterogeneous network topology graph involves initializing a graph data structure containing all the aforementioned network nodes and their corresponding instantaneous feature vectors. At this point, the graph only contains the set of nodes and node features, while the edge set is empty.

[0102] More specifically, the operation calculates the connection weight between any two graph nodes based on the physical coverage overlap and the licensed spectrum interval between the networks. The physical coverage overlap is obtained from the geographic information system database of the network deployment and is specifically represented as the ratio of the intersection area to the union area of ​​the polygon regions covered by the two networks. The licensed spectrum interval is obtained from the spectrum management information database and is specifically represented as the absolute difference in the center frequencies of the primary carriers of the two networks, in megahertz.

[0103] The specific logic for calculating the weights is as follows: Based on the well-known interference principles in this field, when the geographical coverage areas of two networks significantly overlap and the spectrum they use is adjacent to each other, the possibility of potential co-channel or adjacent-channel interference between them is higher. Therefore, a judgment is made based on a preset coverage overlap threshold and a spectrum interval threshold. If the coverage overlap of a network node pair is greater than the preset coverage overlap threshold and its spectrum interval is less than the preset spectrum interval threshold, then it is determined that there is a potential interference relationship between the node pair, and a non-zero initial connection weight is assigned to it to represent this relationship; otherwise, its initial connection weight is zero.

[0104] The specific value of the initial connection weights can be a fixed non-zero constant to simplify the initial topology construction; a more precise value will be learned from the data through subsequent spatiotemporal graph convolutional networks. Next, the operation of creating edges for node pairs with connection weights greater than zero is performed. This operation iterates through all network node pairs, checks their calculated initial connection weights, and if the weight is greater than zero, an undirected edge is added between the node pairs in the graph data structure, and the initial weight value is assigned to the edge as its initial weight attribute. Thus, a weighted initial network graph is generated, where edges represent potential interference between networks, and weights initially identify this relationship.

[0105] Subsequently, the weighted initial network graph and the node feature sequences from consecutive time slices are input into the spatiotemporal graph convolutional network. The node feature sequences from consecutive time slices are generated by repeatedly performing node feature aggregation operations over multiple consecutive acquisition cycles, producing a feature vector sequence of length T for each node.

[0106] The input action involves arranging the aggregated feature vectors of all nodes from the past T time slices into a three-dimensional tensor in chronological order, while simultaneously providing the edge weight matrix of the initial network graph as spatial adjacency relationships to a pre-trained spatiotemporal graph convolutional network model. This model handles both spatial adjacency relationships and temporal series changes through its internal structure. One of the core tasks of this spatiotemporal graph convolutional network is to learn and refine the edge weights between network node pairs using the historical sequence of node features, optimizing them from initial simple identifiers to values ​​that can accurately quantify the intensity of dynamic disturbances.

[0107] The operation of outputting the spatiotemporal coupled feature tensor of each node pair means that after processing, the spatiotemporal graph convolutional network model will generate a fixed-dimensional feature vector for each edge in the graph. This vector integrates the joint pattern information of the situational features of the two ends of the edge within T time windows, which propagate along the network topology and evolve over time. It is a deep representation of the learned disturbance dynamics.

[0108] Next, the spatiotemporal coupling feature tensor is input into a fully connected layer. This fully connected layer is part of the output of the spatiotemporal graph convolutional network model. The input operation involves feeding the spatiotemporal coupling feature vector corresponding to each edge obtained in the previous step into this fully connected layer. The operation of regressing to obtain the scalar sequence of endogenous interference intensity between each pair of network nodes within the corresponding time window means that the fully connected layer maps the high-dimensional feature vector into a scalar value, which represents the interference intensity estimated for the corresponding edge (network node pair) in the current time slice. By processing T time slices sequentially, a sequence of interference intensity values ​​of length T is finally generated for each edge, i.e., the scalar sequence of endogenous interference intensity. Each value in this sequence is a dimensionless relative intensity index; the larger the value, the stronger the interference.

[0109] Simultaneously, operations are performed in parallel to extract the service data rate time series from the standardized service feature set and the terminal distribution density time series from the standardized mobile feature set, as key feature sequences. Specifically, on the cloud system side, for the entire managed area, in each collection cycle, the "data rate" field value is extracted from the standardized service feature set reported by all terminals, and the average value of all terminals is calculated to form a global service data rate time series.

[0110] Similarly, "location grid identifiers" are extracted from the standardized mobility feature set reported by all terminals. The terminal distribution density of each network is estimated by dividing the number of terminals located within each network's coverage area in each time slice by the coverage area. The average value across all networks is then taken to form a global terminal distribution density time series. These two global series serve as key features characterizing macro-level service load and terminal mobility.

[0111] Next, the operation of calculating the Pearson correlation coefficient between the scalar sequence of endogenous interference intensity and each key feature sequence is performed. For each edge in the graph, its scalar sequence of endogenous interference intensity of length T is compared with the global service data rate time series of the same length T to calculate the Pearson correlation coefficient. Then, the Pearson correlation coefficient is calculated between the time series of global terminal distribution density and the data to obtain the correlation coefficient. The operation of generating a correlation coefficient matrix with network node pairs as rows and key features as columns refers to creating a two-dimensional matrix where each row corresponds to an edge, and each row contains two elements: the first column stores the key features corresponding to that edge. The second column stores the edge corresponding to the given edge. .

[0112] Finally, the final operation of constructing the network interference correlation graph is performed, using the mean of the scalar sequence of endogenous interference intensity as the edge weights between corresponding network node pairs and the correlation coefficient matrix as the association attribute of the corresponding edges. The construction action includes two steps: First, traversing all edges in the graph, calculating the arithmetic mean of the scalar sequence of endogenous interference intensity corresponding to each edge, and replacing the original initial weight of the edge with this mean. At this time, the edge weights have been updated to the precisely quantized average interference intensity estimate learned by the spatiotemporal graph convolutional network.

[0113] Secondly, each row of the correlation coefficient matrix (i.e., the two correlation coefficients corresponding to each pair of network nodes) is appended as an attribute vector to the corresponding edge. The resulting network interference correlation graph has nodes representing heterogeneous networks, edge weights representing the endogenous interference intensity quantified by model learning, and edge attributes revealing the statistical correlation between this interference intensity and global service changes and terminal mobility changes. This completes the construction of an interference relationship graph with causal explanatory potential from raw data.

[0114] In this embodiment, calculating the Pearson correlation coefficient between the scalar sequence of endogenous interference intensity and each key feature sequence, and generating a correlation coefficient matrix with network node pairs as rows and key features as columns, includes:

[0115] For each pair of network nodes, its scalar sequence of endogenous interference intensity is treated as a variable;

[0116] The service data rate time series and the terminal distribution density time series are respectively used as two other variables;

[0117] Based on the Pearson correlation coefficient formula, the correlation coefficient between the scalar sequence of endogenous interference intensity and the time series of service data rate, as well as the correlation coefficient between the scalar sequence of endogenous interference intensity and the time series of terminal distribution density, are calculated.

[0118] The two calculated correlation coefficients are arranged in order into a row vector, where each network node corresponds to one row vector;

[0119] Stack the row vectors corresponding to all network nodes to form the correlation coefficient matrix.

[0120] Preferably, the specific implementation method for calculating the Pearson correlation coefficient between the scalar sequence of endogenous interference intensity and each key feature sequence, and generating the correlation coefficient matrix, is described in detail. The purpose of this step is to attach a quantified correlation attribute to each edge in the network interference correlation graph (i.e., each pair of networks with interference relationship), which reveals the degree of statistical linear correlation between the dynamic interference intensity represented by the edge and the dynamic service load and terminal distribution of the system as a whole.

[0121] Specifically, for each pair of network nodes, the operation of treating its endogenous disturbance intensity scalar sequence as a variable is performed. Here, a pair of network nodes specifically refers to any two heterogeneous network nodes in the initial network graph whose connection weights are greater than zero, thus establishing an edge. Its endogenous disturbance intensity scalar sequence is a sequence of disturbance intensity values ​​of length T, output after processing by the spatiotemporal graph convolutional network model and regression by a fully connected layer, where T is the number of continuous time slices processed by the model. Treating this sequence as a variable means that it will be considered as a random variable X to be examined in the subsequent Pearson correlation analysis.

[0122] More specifically, the operation involves treating the service data rate time series and the terminal distribution density time series as two additional variables. The service data rate time series referred to here is a sequence of length T formed by extracting and calculating the global average from the standardized service feature set. The terminal distribution density time series referred to here is a sequence of length T formed by extracting and calculating the global average from the standardized mobility feature set.

[0123] Before performing the correlation calculation in this step, a crucial prerequisite must be ensured: the three sequences mentioned above—namely, the scalar sequence of endogenous interference intensity for each network node pair, the global service data rate time series, and the global terminal distribution density time series—must be strictly aligned in the time dimension. This alignment operation is guaranteed by the data preprocessing unit in the cloud system, ensuring that they represent the exact same time window, the same sampling point, and the same sequence length T, so that the interference intensity value at each moment can match the corresponding global service load and terminal distribution status values.

[0124] Next, operations are performed based on the Pearson correlation coefficient formula to calculate the correlation coefficient between the scalar sequence of endogenous interference intensity and the time series of service data rate, as well as the correlation coefficient between the scalar sequence of endogenous interference intensity and the time series of terminal distribution density.

[0125] For the network node pair currently being processed, the calculation is performed independently in two steps: First, the scalar sequence of the endogenous interference intensity of the node pair and the time series of the global service data rate are input together into the Pearson correlation coefficient calculation function to obtain the correlation coefficient. Secondly, the scalar sequence of the endogenous interference intensity of the node pair and the time series of the global terminal distribution density are input together into the Pearson correlation coefficient calculation function to obtain the correlation coefficient. Each correlation coefficient is a scalar value between -1 and 1. Its absolute value represents the strength of the linear correlation, and the positive or negative sign represents the direction of the correlation.

[0126] Then, the operation of arranging the two calculated correlation coefficients into a row vector in order is performed. Arranging in order means, according to a predefined fixed format, arranging the network node pairs according to the calculated correlation coefficients with the service data rate time series. The correlation coefficient with the terminal distribution density time series is placed in the first column. This is placed in the second column. This creates a one-dimensional array containing two elements, i.e., a row vector. Each network node pair generates such a unique row vector after completing the above calculation.

[0127] Finally, the operation of stacking the row vectors corresponding to all network node pairs to form the correlation coefficient matrix is ​​performed. Specifically, the stacking operation involves concatenating the row vectors generated by all node pairs sequentially along the row direction according to a predetermined order (e.g., lexicographical order) of their globally unique identifiers. This results in a two-dimensional matrix with two columns, the number of rows equal to the total number of network node pairs with edges. The i-th row of the matrix corresponds to the i-th network node pair, and its first column stores the information of that node pair. The second column stores the elements of this node pair. This matrix is ​​the correlation coefficient matrix required in the claims, and it will serve as the direct data source for assigning "association attributes" to each edge when constructing the network interference correlation graph.

[0128] Service-Network Dynamic Decoupling Module: Used to reverse reconstruct the terminal service characteristics based on the network interference correlation diagram, and decouple the key network state sequence;

[0129] In this embodiment, based on the network interference correlation graph, the terminal service characteristics are reversely reconstructed to decouple the key network state sequence, as follows:

[0130] Based on the scalar sequence of endogenous interference intensity corresponding to each edge in the network interference correlation graph, the statistical characteristics of each sequence within a preset time window are calculated, and an interference feature matrix with heterogeneous networks as nodes is constructed.

[0131] Using the interference feature matrix as graph structure data and the standardized business feature set as input node features, a graph neural network is constructed.

[0132] The standardized business feature set is input into the graph neural network for forward propagation, and the reconstructed business feature vector of each heterogeneous network node is output.

[0133] Calculate the element-wise absolute difference between the standardized business feature set and the reconstructed business feature vector to obtain a sequence of business feature residual vectors;

[0134] Principal component analysis is performed on the time dimension of the business feature residual vector sequence to extract the principal component components. The principal component components are then arranged in chronological order to generate the key network state sequence.

[0135] Preferably, the operation of the service-network dynamic decoupling module is described in detail. The technical challenge of this module lies in inferring, from directly observable inter-network interference relationships and the performance of aggregated services on affected terminals, the dynamic sequence of the internal states of each network itself, which cannot be directly observed. The technical concept is to model the observed service characteristics as consisting of two parts—one part originating from the network's own basic transmission capabilities and state, and the other part originating from dynamic interference between networks. By utilizing a known inter-network interference relationship graph to filter out the impact of interference, the network's own state can be estimated.

[0136] Specifically, the operation of constructing a network state inference graph neural network based on the network interference correlation graph is performed. First, the graph structure is defined: a weighted undirected graph is constructed with each heterogeneous network as a node and the weights of the edges in the network interference correlation graph (i.e., the mean of endogenous interference intensity) as edge attributes. Next, the node input features are defined: for each network node, its input feature is a scalar, which is the aggregated value of a specific key performance indicator extracted from the standardized service feature set of all terminals belonging to the network at the current time (e.g., the arithmetic mean of the service latency of all terminals). This scalar is denoted as the observed service indicator.

[0137] Next, a graph neural network model is defined: a message-passing-based graph neural network architecture, such as a graph convolutional network, is employed. The training objective of this network is to learn a function that enables each node to predict a "de-interferenced" reference value for a business indicator based on its own and its neighbors' observed business indicators and the interference weights of the connecting edges. This model requires training using historical data, which should include business indicators observed during periods of relatively stable network conditions or periods of low known interference as approximate ground truth values ​​to supervise the model's learning of how to eliminate the influence of interference.

[0138] More specifically, the process involves performing forward propagation using the graph neural network to obtain reference values ​​for de-interference service metrics for each network node. After model training is complete, the observed service metrics scalars for each network node at the current time, along with the topology and edge weights of the network interference correlation graph, are input into the graph neural network. Through forward propagation, the network outputs a scalar value for each node, called the de-interference service metric reference value. This value represents the value that the model infers based on the current overall network interference structure, suggesting that the network might achieve its service metrics at the current time if it were not interfered with by other networks in the graph.

[0139] Then, the operation of calculating the residual between the observed service indicators and the de-interference reference values ​​is performed to form a network state residual sequence. For each network node, the difference between its current observed service indicator and the de-interference service indicator reference value output by the model is calculated. This difference, i.e., the residual, reflects the gap between the actual service performance of the network node and its expected performance after eliminating known interference. This gap is attributed to the dynamic changes in the network's internal state (such as its own load bursts, equipment hardware and software anomalies, other unmodeled interference sources, etc.). For each network node, this residual is continuously calculated over a period of time to obtain a one-dimensional residual time series for each node, i.e., the network state residual sequence. This sequence directly characterizes the net impact of the network's own state fluctuations on services.

[0140] Finally, a joint analysis of the state residual sequences of all network nodes is performed to extract key network state sequences. The time-aligned network state residual sequences from all heterogeneous network nodes are combined into a matrix, where each row represents a network and each column represents a time point. Principal component analysis is then performed on this matrix in the spatial dimension (i.e., across network nodes).

[0141] The first K principal components are extracted, each a time series representing the state fluctuations of a specific pattern that coexists or is widely correlated across the entire network. Arranging these K principal component time series sequentially generates the key network state sequence. Each component of this sequence characterizes a potential key factor, either global or regional, that drives the synchronous fluctuations of the states of multiple networks.

[0142] Business Intent Abstraction Module: Used to extract the terminal service characteristics based on the key network state sequence, obtain the collaborative bearing requirements of the heterogeneous network space, and construct the required collaborative state target of the heterogeneous network based on the collaborative bearing requirements.

[0143] In this embodiment, based on the key network state sequence, the terminal service characteristics are extracted to obtain the collaborative bearing requirements of the heterogeneous network space. Based on the collaborative bearing requirements, the required collaborative state target of the heterogeneous network is constructed as follows:

[0144] Spectral analysis was performed on the key network state sequence to extract its power spectral density distribution in different frequency bands;

[0145] The frequency band components in the power spectral density distribution whose amplitude exceeds a preset threshold are identified as the dominant fluctuation modes;

[0146] Calculate the frequency and phase corresponding to the dominant fluctuation pattern, and perform time-frequency domain matching with the periodic business indicator sequence extracted from the standardized business feature set to determine the terminal business type and key performance indicators affected by the dominant fluctuation pattern.

[0147] Using historical demand baselines as a benchmark, dynamic demand ranges are constructed for each key performance indicator based on the strength of the dominant fluctuation pattern.

[0148] Transform the demand ranges of each key performance indicator into constraints and establish a set of constraints.

[0149] To address the transmission capability characteristics of different heterogeneous networks, the construction problem of the cooperative state objective is modeled as a constrained optimization problem.

[0150] The optimization problem is solved using the Lagrange multiplier method, and the optimal set of performance parameters required for each heterogeneous network to meet the collaborative carrying requirements is obtained, which serves as the collaborative state objective.

[0151] Preferably, the specific implementation method of the business intent abstraction module for deriving collaborative bearing requirements and constructing collaborative state targets based on key network state sequences is described in detail. The technical problem solved by this module is: how to interpret and quantify key network state sequences that characterize the fluctuations of the underlying network into deterministic requirements for future business performance, and to transform these requirements into specific and measurable performance targets that various heterogeneous networks need to achieve collaboratively.

[0152] Specifically, the analysis examines the correlation between key network state sequences and historical service performance to deduce the operations required for collaborative support. A key network state sequence is a set of abstract time series. First, an analysis database is established, containing a matrix of key network state sequences for historical time periods, as well as historical sequences of system-level key service performance indicators (such as average network latency, 95th percentile latency, and service request rejection rate) aggregated from standardized service feature sets during the same period.

[0153] Then, for each key network state sequence component (i.e., each fluctuation pattern), the Granger causality or strong lag correlation with the historical sequence of each key business performance indicator is calculated. This analysis identifies key network state components that have a significant predictive or driving effect on specific business performance indicators and labels them as demand drivers. For each labeled demand driver, a regression model is established between its numerical value (or its first-order difference, envelope, etc.) and the degree of deterioration of the corresponding business performance indicator.

[0154] This model describes "the extent to which service performance is expected to deteriorate when the network state factor fluctuates." Based on this model and the current value and short-term forecast trend of the demand-driven factors, the target values ​​that relevant key performance indicators need to be guaranteed within a future planning cycle to prevent service performance deterioration can be derived, such as "the average latency of the entire network needs to be guaranteed to be below 20ms." All the performance indicators that need to be guaranteed and their target values ​​together constitute the collaborative carrying requirements.

[0155] More specifically, the process involves decomposing and mapping collaborative bearer requirements to various heterogeneous networks to form initial collaborative state goals. Collaborative bearer requirements are system-level goals, which need to be decomposed across different heterogeneous networks. First, based on network topology, historical load data, and service flow path predictions, the contribution ratio or responsibility weight that each heterogeneous network should bear in achieving each system-level performance goal within the future planning period is estimated. For example, based on service flow distribution, core network element A contributes 40% to end-to-end latency, while access networks B and C each contribute 30%. Then, based on the contribution weights and system-level goal values, localized performance goals are calculated for each network. For example, to achieve an average latency of <20ms, network A needs to control the average latency of its processed service flows to <8ms, while networks B and C need to control it to <6ms respectively. This set of localized performance goals calculated for all networks constitutes the initial collaborative state goals.

[0156] Finally, the feasibility of the initial coordinated state objectives is verified and optimized. Since the above decomposition is based on estimation, and network capabilities are subject to hard constraints (such as maximum bandwidth and minimum theoretical latency), the initial objectives may be infeasible or conflicting. Therefore, a coordinated optimization model needs to be established. The decision variables of this model are the adjustable internal parameters of each network that affect its local performance objectives, such as the proportion of bandwidth reserved for specific services and scheduling priority weights.

[0157] The objective function is to minimize the sum of squared weighted deviations between all network local objectives and the initial assigned objectives, so that the final result is as close as possible to the ideal decomposition. Constraints include: adjustments to the local parameters of each network must be within their physical and capacity limitations; the performance achieved by each network after adjustment, after being weighted by contribution, must meet the system-level collaborative carrying requirements. This optimization problem is typically a quadratic programming problem with linear or nonlinear constraints. Using a mature numerical optimization algorithm library, a set of feasible and coordinated local performance objective values ​​for each network can be obtained, which is the final determined collaborative state objective. This objective is a set of clear, specific, measurable, and feasible-verified network-level KPIs, which will directly guide the generation of the scheduling strategy for the next module.

[0158] The collaborative scheduling strategy generation module is used to generate multi-dimensional collaborative scheduling instructions based on the collaborative state objective as the optimization guide, the network-service joint situation vector and the key network state sequence as state inputs, and the collaborative decision model.

[0159] In this embodiment, based on the cooperative state objective as the optimization guide, and using the network-service joint situational vector and the key network state sequence as state inputs, a multi-dimensional cooperative scheduling instruction is generated using a cooperative decision-making model, as follows:

[0160] Align the network-service joint situational vector with the key network state sequence in the time dimension and concatenate them along the feature dimension to generate a comprehensive state input vector;

[0161] The comprehensive state input vector is input into a neural network model trained based on historical data, and the scheduling action parameters are output.

[0162] Using the joint performance boundary of the cooperative state objective as a constraint, numerical boundary checks and feasibility adjustments are performed on the scheduling action parameters to generate verification scheduling action parameters;

[0163] Based on the standard control interfaces of each network element in the heterogeneous network, the verification and scheduling action parameters are converted into specific spectrum reconfiguration instructions, transmit power update instructions, and access point switching instructions.

[0164] The spectrum reconfiguration command, transmit power update command, and access point switching command are combined to form the multi-dimensional collaborative scheduling command.

[0165] Preferably, the specific implementation method of the cooperative scheduling strategy generation module for generating multi-dimensional cooperative scheduling instructions based on cooperative state objectives is described in detail. The technical task of this module is to transform the abstract cooperative state objectives (i.e., the performance indicators that each network needs to achieve) into a series of specific, executable network control instructions that can work together to achieve the objectives.

[0166] Specifically, the operation involves fusing the network-service joint situational awareness vector with the key network state sequence to form a decision state representation. Decision-making requires a unified state input. The latest acquired network-service joint situational awareness vector is concatenated with the key network state sequence aligned to the same timestamp to form a comprehensive state feature vector. The alignment operation is as follows: using the acquisition time of the joint situational awareness vector as a reference, data from the corresponding time and its preceding adjacent time are extracted from the key network state sequence; if the sequence times do not match perfectly, linear interpolation is used to complete the data, ensuring that the data from both sources refer to the same observation window in time.

[0167] More specifically, this involves executing operations based on the cooperative state objective and the current state, calculating scheduling action parameters through an optimized decision model. The goal of this step is to find a set of action parameters that allows network performance to approach the target value. First, the performance indicators in the cooperative state objective (such as upper limits for latency and lower limits for throughput) are explicitly quantified as constraints in the optimization problem. Second, adjustable network action parameters (such as the transmit power adjustment of each base station, the proportion of bandwidth reserved for different service types, and the preference weights for candidate access points) are defined as decision variables. These variables have physical upper and lower limits. Then, an optimization objective function is defined to improve overall network energy efficiency or fairness, such as minimizing total transmit power or maximizing the throughput of the worst-performing user.

[0168] Finally, a constrained mathematical optimization problem is constructed, with action parameters as decision variables and the objective function as the function defined above. Constraints include: performance constraints derived from the cooperative state objective; physical boundary constraints on the decision variables; and linear or nonlinear constraints derived from the network interference correlation graph or physical model that describe the coupling relationships between action parameters (e.g., the total allocated bandwidth cannot exceed the total system bandwidth). A mature mathematical programming solver (such as an interior-point solver or a sequential quadratic programming solver) is used to solve this problem, directly outputting a set of verification scheduling action parameters. This method directly encodes "goal orientation" into the constraints of the optimization model, ensuring that the output parameters mathematically meet the objective requirements and possess physical feasibility.

[0169] Then, the operation of converting the verification scheduling action parameters into network element commands is performed. The scheduling action parameters are the result of mathematical solutions. Each action parameter value needs to be filled into a pre-defined command template according to the standardized management interface protocol of each heterogeneous network element (e.g., a configuration management interface conforming to relevant 3GPP, IEEE, or IETF standards), generating structured configuration data recognizable by the device. For example, the parameter "Base station A transmit power adjustment amount -3dB" is converted into a power update configuration data block conforming to the base station's network management protocol (such as the NETCONF / YANG model); the parameter "Access point preference weight of service flow F tilted towards network B" is converted into a policy rule to be issued to the terminal or access controller. This conversion process is completed automatically based on a pre-defined parameter-command mapping rule base corresponding to the network element model and interface.

[0170] Finally, the process of encapsulating and generating multi-dimensional collaborative scheduling instructions is performed. Encapsulation refers to combining multiple independent control instructions for different network elements and types into a transactional instruction package according to their logical execution order and time synchronization requirements (such as certain spectrum configurations needing to be completed before service switching). This instruction package is the final generated multi-dimensional collaborative scheduling instruction, which contains a set of specific operation commands that work collaboratively in time and space, and will be sent to the instruction distribution and verification module for execution.

[0171] Network resource self-growth module: used to pre-simulate the evolution trajectory of the key network state sequence after the instruction is executed, and dynamically construct a compensation network plane for possible resource bottlenecks;

[0172] In this embodiment, the evolution trajectory of the key network state sequence after the pre-execution of the pre-instruction is observed, and a compensation network plane is dynamically constructed to address potential resource bottlenecks, as follows:

[0173] A virtual simulation environment is established based on the network interference correlation diagram and historical state data.

[0174] The multi-dimensional collaborative scheduling instructions are input into the virtual simulation environment to simulate the network operation process after the parameters of each network element are adjusted, and the estimated network performance index sequence during the simulation operation is output. The network performance index sequence is used as the virtual pre-simulation trajectory.

[0175] The time delay sequence, packet loss rate sequence, and bandwidth utilization rate sequence in the virtual pre-simulation trajectory are compared with the corresponding time delay upper limit, packet loss rate upper limit, and bandwidth utilization rate upper limit in the cooperative state target, respectively, on a time-by-time basis.

[0176] Identify the time period in which any performance index sequence value continuously exceeds its corresponding upper limit and the target network element with the highest load during that time period. Mark the time period and the target network element together as a resource bottleneck, and calculate the average difference between the performance index sequence value and its corresponding upper limit as the performance gap.

[0177] Based on the performance gap and the network type of the target network element, determine the type and specifications of the virtual resources that need to be supplemented;

[0178] In the virtual simulation environment, during the time period corresponding to the resource bottleneck, virtual network function nodes of appropriate specifications are deployed and configured for the affected service flows.

[0179] The virtual network functional nodes are connected to the relevant physical networks to form a logically isolated compensation network plane. Preferably, the specific implementation method of pre-simulating instruction execution trajectories and dynamically constructing the compensation network plane in the network resource self-growth module is described in detail. The technical problem solved by this module is: before actually executing scheduling instructions that may have a significant impact, to pre-evaluate their effects and intelligently predict and compensate for potential resource shortages, thereby improving the robustness of scheduling decisions.

[0180] Specifically, the operation involves establishing a virtual simulation environment based on the network interference correlation graph and historical state data. The virtual simulation environment is a software-defined network simulation system. The steps for establishing this environment include: first, importing the network interference correlation graph as the basic topology model of the simulation environment, where nodes represent heterogeneous network entities and edges represent the interference correlation relationships and strengths between networks.

[0181] Secondly, performance parameter models of each network node under typical service loads are extracted from the historical state database. These include latency-throughput relationship curves and packet loss rate models under specific loads. Finally, the aforementioned topology model and performance parameter models are instantiated and configured in a simulation platform to form a digital twin environment capable of simulating the impact of inter-network interference, service flow transmission processes, and changes in network element performance. This environment can receive control commands and simulate the dynamic evolution of network states.

[0182] More specifically, the process involves inputting the multi-dimensional collaborative scheduling instructions into the virtual simulation environment to simulate the network operation process after adjusting the parameters of each network element, and outputting a sequence of estimated network performance indicators during the simulation operation. This sequence of network performance indicators is then used as the virtual pre-simulation trajectory. The input instruction action refers to parsing the multi-dimensional collaborative scheduling instructions generated by the collaborative scheduling strategy generation module into configurable parameters of the corresponding network element model in the virtual simulation environment, and applying these parameters at the start point of the simulation timeline.

[0183] The simulation process refers to running a future time period in a simulation environment according to a preset service traffic model and terminal mobility model. During this process, the performance status of each node and link in the network is dynamically calculated based on the network interference correlation diagram and network element performance model. The action of outputting the predicted network performance index sequence refers to collecting performance indicators of the entire system or key nodes at fixed time intervals during the simulation, such as end-to-end latency, packet loss rate, and link bandwidth utilization, and recording the value of each indicator over time as a time series. This set of sequences reflecting the predicted performance at various future moments constitutes the virtual pre-simulation trajectory, which depicts the expected evolution path of the network state after the execution of scheduling instructions.

[0184] Next, the operation of comparing the time delay sequence, packet loss rate sequence, and bandwidth utilization rate sequence in the virtual pre-simulation trajectory with the corresponding time delay upper limit, packet loss rate upper limit, and bandwidth utilization rate upper limit in the cooperative state target is performed time-by-time.

[0185] The collaborative state objective defines quantified target requirements for key performance indicators, such as a latency limit of 50 milliseconds. Time-by-time comparison refers to comparing the estimated latency value recorded at each moment in the virtual simulation trajectory with this latency limit to determine if it exceeds the limit. Similarly, the same time-by-time comparison operation is performed on the estimated packet loss rate and its upper limit, and the estimated bandwidth utilization rate and its upper limit. This process generates a series of Boolean values, indicating whether each performance indicator meets the target at each moment.

[0186] Then, the process involves identifying the time period during which any performance indicator sequence value continuously exceeds its corresponding upper limit, and the target network element with the highest load during that time period. This time period and the target network element are jointly marked as a resource bottleneck, and the average difference between the performance indicator sequence value and its corresponding upper limit is calculated as the performance gap. Identifying resource bottlenecks is an analytical process. First, the Boolean value sequence is examined to find the time period with the longest duration in which the estimated performance value continuously exceeds the upper limit.

[0187] Next, during this period, the load of all network elements in the virtual simulation environment is analyzed to identify the network element with the highest load rate, which is then marked as the target network element. This time period and the target network element together constitute a resource bottleneck, indicating that this network element is the main reason for performance failure during this period. The performance gap is a quantitative indicator. The calculation involves subtracting the corresponding upper limit value from all performance indicator values ​​exceeding the upper limit during the bottleneck time period to obtain the gap at each moment. Then, the arithmetic mean of these gap values ​​is calculated. This average value represents the average severity of performance failure.

[0188] Next, the process involves determining the type and specifications of virtual resources that need to be supplemented based on the performance gap and the network type of the target network element. This is a resource mapping decision process. The performance gap indicates the amount of resources that need to be supplemented.

[0189] The network type of the target network element determines the form of resources. For example, for a wireless access network element, it may be necessary to supplement it with virtual wireless access points or virtual cells; for a core network forwarding node, it may be necessary to supplement it with virtual routers or virtual bandwidth enhancement units. The action of determining specifications refers to determining the specific capacity parameters of the virtual network functions to be deployed by querying a predefined resource specification mapping table based on the size of the performance gap. For example, it may be necessary to deploy a virtual forwarding node with a processing capacity of X Gbps.

[0190] Next, within the virtual simulation environment, during the time period corresponding to the resource bottleneck, the operation of deploying and configuring virtual network function nodes of appropriate specifications for the affected service flows is performed. Deployment and configuration are construction actions within the virtual environment. First, in the simulation environment software, a virtual network function node model is instantiated according to the type and specifications determined in the previous step. Then, this virtual node model is configured on the simulation timeline to be active only during the time period corresponding to the resource bottleneck. Finally, the routing strategy of the affected service flows in the simulation environment is modified so that the data flows during the bottleneck period are guided to this newly deployed virtual network function node for processing, thereby alleviating the load on the target network element.

[0191] Finally, the operation of establishing connections between the virtual network functional nodes and the relevant physical networks to form logically isolated compensation network planes is performed. This is a crucial step in logicalizing the virtual compensation measures. Establishing connections refers to configuring connection channels between the virtual network functional node and one or more underlying physical network models at the logical level of the virtual simulation environment, enabling it to receive and send data.

[0192] Forming a logically isolated compensation network plane refers to logically organizing these newly deployed virtual network function nodes and their connections into an independent overlay network layer specifically designed to provide compensatory capacity within a specific time period. This overlay layer is logically isolated from the original physical network, but shares the service load, thereby dynamically expanding network resources.

[0193] Instruction distribution and verification module: used to execute multi-dimensional collaborative scheduling instructions that integrate the compensation network plane logic, monitor the real-time status data generated after execution, match and analyze the actual state evolution trajectory with the virtual pre-simulation trajectory, lock the interference transmission chain corresponding to the root cause of the deviation, and update the network interference correlation diagram.

[0194] In this embodiment, a multi-dimensional collaborative scheduling instruction incorporating the compensation network plane logic is executed. Real-time state data generated after execution is monitored, and the actual state evolution trajectory is matched and analyzed with the virtual pre-simulation trajectory to pinpoint the interference propagation chain corresponding to the root cause of the deviation. Based on this, the network interference correlation graph is updated as follows:

[0195] The multi-dimensional collaborative scheduling command is executed, and the actual performance index data of each network element in the monitoring window are collected after the command takes effect through the network probe and terminal reporting mechanism.

[0196] The actual performance index data is time-aligned and the corresponding performance index sequence in the virtual pre-simulation trajectory generated by the network resource self-growth module is calculated;

[0197] Identify continuous time periods in which the difference exceeds a preset deviation threshold, and mark the performance index with the largest difference and its corresponding network element within that time period as significant performance deviation and the network element where the deviation occurred;

[0198] Starting from the network element where the deviation occurs, in the network interference association graph, based on the weight of the edges in the graph, iteratively backtrack to the upstream network node that is its neighbor and has a higher edge weight;

[0199] The edges traversed by the backtracking search are connected in sequence to form a path from an upstream network node to the network element where the deviation occurred. This path is identified as the interference propagation chain that caused the significant performance deviation, and the upstream network node is the root cause node of the deviation.

[0200] Based on the historical correlation strength of each edge in the interference propagation chain within the time period corresponding to the significant performance deviation, the weights of the corresponding edges in the network interference correlation graph are enhanced and updated.

[0201] Preferably, the specific implementation method of the instruction distribution and verification module for executing scheduling instructions, verifying effects, and updating the network model is described in detail. This module constitutes the closed-loop control link of the system. Its purpose is to execute scheduling decisions, evaluate the accuracy of decisions by comparing actual and expected network behavior, and use actual data to correct the internal model to improve the reliability of future decisions.

[0202] Specifically, the process involves executing the multi-dimensional collaborative scheduling instructions and collecting actual performance data of each network element within the monitoring window after the instructions take effect, through network probes and terminal reporting mechanisms. Executing the instructions means securely and reliably sending the encapsulated multi-dimensional collaborative scheduling instructions to the corresponding physical and virtual network elements via the cloud system's control channel, and confirming their receipt and effectiveness. Simultaneously, within the preset monitoring window, active probes deployed at key network nodes are activated, and terminal devices are configured to report their perceived network performance data in enhanced mode. The data collection process continuously gathers raw performance data reported by probes and terminals, such as throughput, latency, packet loss rate, and signal strength, and aggregates and preprocesses this data to form actual performance data of each network element changing over time within the monitoring window.

[0203] More specifically, the process involves performing time-series alignment and difference calculation between the actual performance index data and the corresponding performance index sequences in the virtual pre-simulation trajectory generated by the network resource self-growth module. Time-series alignment refers to calibrating the actually collected sequences and the virtual pre-simulation sequences on the time axis, using the absolute time of the scheduling command's effective date as a reference point, ensuring that the comparison is of data within the same effective duration. Difference calculation involves subtracting the estimated performance value from the virtual pre-simulation value at each aligned time point to obtain the performance deviation value for that index at that moment. This operation is performed on all indices to generate a performance deviation trajectory with the same structure as the virtual pre-simulation trajectory but with deviated values.

[0204] Next, the process involves identifying consecutive time periods where the difference exceeds a preset deviation threshold, and marking the performance indicator with the largest difference within those time periods, along with its corresponding network element, as significant performance deviations and the network element where the deviation occurred. The preset deviation threshold is an acceptable error range set for different performance indicators. The identification process is as follows: traversing the performance deviation trajectory, for each performance indicator, identifying all time intervals in which its deviation value continuously exceeds its own threshold. For each such interval, identifying the moment when the deviation value reaches its maximum, and recording the performance indicator with the largest deviation at that moment, along with the physical or virtual network element corresponding to the source of that indicator's data. This network element is marked as the network element where the deviation occurred, and its corresponding indicator and deviation value constitute a significant performance deviation event.

[0205] Then, starting from the network element where the deviation occurred, an iterative backtracking search is performed in the network interference correlation graph, based on the weights of the edges in the graph, towards its adjacent upstream network nodes with higher edge weights. This step aims to assume that the performance deviation may be caused by the propagation of inter-network interference and to attempt to locate the possible source. First, the network element where the deviation occurred is set as the current search node. In the network interference correlation graph, all adjacent nodes that have connecting edges with the current search node are identified. The weights of the edges between these adjacent nodes and the current search node are compared.

[0206] The "upstream" direction is defined as a virtual direction pointing from adjacent nodes to the current search node, which aligns with the potential propagation direction of interference effects. The search rule is as follows: from all adjacent nodes, select the one with the largest edge weight between it and the current search node, and whose weight has recently shown an increasing or highly fluctuating trend, as the next hop backtracking node. If no suitable adjacent node is found, or the number of backtracking hops reaches a preset limit, the search terminates. This iterative process forms a node path.

[0207] Next, the system connects the edges traversed by the backtracking search in sequence to form a path from an upstream network node to the network element where the deviation occurred. This path is identified as the interference propagation chain causing the significant performance deviation, and the upstream network node is the root cause node of the deviation. The last node at the end of the search is marked as a candidate root cause node. The nodes and edges traversed from this root cause node to the network element where the deviation occurred are listed in sequence to form a candidate interference propagation chain. The system records this chain as a hypothesis, namely that the performance deviation may be caused by the superposition of interference propagation along this chain.

[0208] Finally, based on the historical correlation strength of each edge in the interference propagation chain within the time period corresponding to the significant performance deviation, the weights of the corresponding edges in the network interference correlation graph are enhanced and updated. This step is based on a dynamic adjustment model using new evidence. Historical correlation strength is specifically defined here as the dynamic similarity of the service load or key status indicators of the network nodes at both ends of each edge in the propagation chain within the time period corresponding to the current deviation event (e.g., obtained by calculating the correlation coefficient of the indicator sequences of the two nodes within that time period).

[0209] The update rule is as follows: for each edge in the propagation chain, its existing weight in the interference correlation graph is increased by an update amount. The magnitude of this update amount is positively correlated with two factors: the severity of the significant performance deviation identified this time (the normalized deviation value); and the historical correlation strength value of the edge within the deviation period. Simultaneously, the system can periodically slightly decay the weights of edges not associated with recent deviation events to maintain model adaptability and avoid weight saturation. Through this incremental update mechanism, the network interference correlation graph can continuously evolve as the system operates, more accurately reflecting the true interference relationships between networks.

[0210] 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.

Claims

1. A heterogeneous network space multi-layered mobile Internet of Things cloud system based on artificial intelligence, characterized in that, include: Data acquisition module: used to acquire real-time status data of mobile IoT terminals, perform feature decoupling and standardization mapping on the real-time status data, and generate a standardized network-service joint situational vector; Endogenous interference perception module: Based on the network-service joint situational vector, it identifies and quantifies the endogenous interference intensity between any two heterogeneous networks through a spatiotemporal graph convolutional network, analyzes the dynamic correlation between the endogenous interference intensity and terminal service characteristics and terminal mobility characteristics, and generates a network interference correlation graph. Service-Network Dynamic Decoupling Module: Used to reverse reconstruct the terminal service characteristics based on the network interference correlation diagram, and decouple the key network state sequence; Business Intent Abstraction Module: Used to extract the terminal service characteristics based on the key network state sequence, obtain the collaborative bearing requirements of the heterogeneous network space, and construct the required collaborative state target of the heterogeneous network based on the collaborative bearing requirements. The collaborative scheduling strategy generation module is used to generate multi-dimensional collaborative scheduling instructions based on the collaborative state objective as the optimization guide, the network-service joint situation vector and the key network state sequence as state inputs, and the collaborative decision model. Network resource self-growth module: used to pre-simulate the evolution trajectory of the key network state sequence after the instruction is executed, and dynamically construct a compensation network plane for possible resource bottlenecks; Instruction distribution and verification module: used to execute multi-dimensional collaborative scheduling instructions that integrate the compensation network plane logic, monitor the real-time status data generated after execution, match and analyze the actual state evolution trajectory with the virtual pre-simulation trajectory, lock the interference transmission chain corresponding to the root cause of the deviation, and update the network interference correlation diagram.

2. The heterogeneous network space multi-layered mobile Internet of Things cloud system based on artificial intelligence according to claim 1, characterized in that, Real-time status data of mobile IoT terminals is acquired, and feature decoupling and standardization mapping are performed on the real-time status data to generate a standardized network-service joint situational vector, as follows: By using sensors and network interfaces deployed on mobile IoT terminals, real-time status data of the terminals in heterogeneous network spaces can be collected. The real-time status data is subjected to feature decoupling processing, and service feature subsets, network feature subsets, and mobility feature subsets are obtained based on data attributes; The service feature subset, the network feature subset, and the mobility feature subset are standardized respectively, and the data values ​​in each subset are mapped to a unified numerical range to obtain the corresponding standardized service feature set, standardized network feature set, and standardized mobility feature set. The standardized service feature set, the standardized network feature set, and the standardized mobile feature set are vectorized and concatenated to generate the standardized network-service joint situational vector.

3. The heterogeneous network space multi-layered mobile Internet of Things cloud system based on artificial intelligence according to claim 1, characterized in that, Based on the network-service joint situational vector, the intrinsic interference intensity between any two heterogeneous networks is identified and quantified using a spatiotemporal graph convolutional network. The dynamic correlation between the intrinsic interference intensity and terminal service characteristics and terminal mobility characteristics is analyzed to generate a network interference correlation graph, including: Each heterogeneous network is defined as a graph node, and the network-service joint situational vector is assigned to the corresponding graph node as a node feature to construct a heterogeneous network topology graph. Based on the physical coverage overlap and authorized spectrum interval between networks, calculate the connection weight between any two graph nodes, establish edges for node pairs with connection weights greater than zero, and generate a weighted initial network graph. The initial network graph and the node feature sequences on continuous time slices are input into the spatiotemporal graph convolutional network, and the spatiotemporal coupling feature tensor of each node pair is output. The spatiotemporal coupling feature tensor is input into a fully connected layer, and regression is performed to obtain a scalar sequence of endogenous interference intensity between each pair of network nodes within the corresponding time window. The service data rate time series and the terminal distribution density time series are extracted from the standardized service feature set and respectively, and used as key feature sequences. Calculate the Pearson correlation coefficient between the scalar sequence of endogenous interference intensity and each key feature sequence, and generate a correlation coefficient matrix with network node pairs as rows and key features as columns; The network interference correlation graph is constructed by using the mean of the scalar sequence of endogenous interference intensity as the edge weight between corresponding network node pairs and the correlation coefficient matrix as the correlation attribute of the corresponding edges.

4. The heterogeneous network space multi-layered mobile Internet of Things cloud system based on artificial intelligence according to claim 3, characterized in that, The calculation of the Pearson correlation coefficient between the scalar sequence of endogenous interference intensity and each key feature sequence, generating a correlation coefficient matrix with network node pairs as rows and key features as columns, includes: For each pair of network nodes, its scalar sequence of endogenous interference intensity is treated as a variable; The service data rate time series and the terminal distribution density time series are respectively used as two other variables; Based on the Pearson correlation coefficient formula, the correlation coefficient between the scalar sequence of endogenous interference intensity and the time series of service data rate, as well as the correlation coefficient between the scalar sequence of endogenous interference intensity and the time series of terminal distribution density, are calculated. The two calculated correlation coefficients are arranged in order into a row vector, where each network node corresponds to one row vector; Stack the row vectors corresponding to all network nodes to form the correlation coefficient matrix.

5. The heterogeneous network space multi-layered mobile Internet of Things cloud system based on artificial intelligence according to claim 1, characterized in that, Based on the network interference correlation diagram, the terminal service characteristics are reverse-reconstructed to decouple the key network state sequence, as follows: Based on the scalar sequence of endogenous interference intensity corresponding to each edge in the network interference correlation graph, the statistical characteristics of each sequence within a preset time window are calculated, and an interference feature matrix with heterogeneous networks as nodes is constructed. Using the interference feature matrix as graph structure data and the standardized business feature set as input node features, a graph neural network is constructed. The standardized business feature set is input into the graph neural network for forward propagation, and the reconstructed business feature vector of each heterogeneous network node is output. Calculate the element-wise absolute difference between the standardized business feature set and the reconstructed business feature vector to obtain a sequence of business feature residual vectors; Principal component analysis is performed on the time dimension of the business feature residual vector sequence to extract the principal component components. The principal component components are then arranged in chronological order to generate the key network state sequence.

6. The heterogeneous network space multi-layered mobile Internet of Things cloud system based on artificial intelligence according to claim 1, characterized in that, Based on the key network state sequence, the terminal service characteristics are extracted to obtain the collaborative bearing requirements of the heterogeneous network space. Based on the collaborative bearing requirements, the required collaborative state targets of the heterogeneous network are constructed as follows: Spectral analysis was performed on the key network state sequence to extract its power spectral density distribution in different frequency bands; The frequency band components in the power spectral density distribution whose amplitude exceeds a preset threshold are identified as the dominant fluctuation modes; Calculate the frequency and phase corresponding to the dominant fluctuation pattern, and perform time-frequency domain matching with the periodic business indicator sequence extracted from the standardized business feature set to determine the terminal business type and key performance indicators affected by the dominant fluctuation pattern. Using historical demand baselines as a benchmark, dynamic demand ranges are constructed for each key performance indicator based on the strength of the dominant fluctuation pattern. Transform the demand ranges of each key performance indicator into constraints and establish a set of constraints. To address the transmission capability characteristics of different heterogeneous networks, the construction problem of the cooperative state objective is modeled as a constrained optimization problem. The optimization problem is solved using the Lagrange multiplier method, and the optimal set of performance parameters required for each heterogeneous network to meet the collaborative carrying requirements is obtained, which serves as the collaborative state objective.

7. The heterogeneous network space multi-layered mobile Internet of Things cloud system based on artificial intelligence according to claim 1, characterized in that, Based on the aforementioned collaborative state objective as the optimization guide, and using the network-service joint situational vector and the key network state sequence as state inputs, a multi-dimensional collaborative scheduling instruction is generated using a collaborative decision-making model, as follows: Align the network-service joint situational vector with the key network state sequence in the time dimension and concatenate them along the feature dimension to generate a comprehensive state input vector; The comprehensive state input vector is input into a neural network model trained based on historical data, and the scheduling action parameters are output. Using the joint performance boundary of the cooperative state objective as a constraint, numerical boundary checks and feasibility adjustments are performed on the scheduling action parameters to generate verification scheduling action parameters; Based on the standard control interfaces of each network element in the heterogeneous network, the verification and scheduling action parameters are converted into specific spectrum reconfiguration instructions, transmit power update instructions, and access point switching instructions. The spectrum reconfiguration command, transmit power update command, and access point switching command are combined to form the multi-dimensional collaborative scheduling command.

8. A heterogeneous network space multi-layered mobile Internet of Things cloud system based on artificial intelligence according to claim 1, characterized in that, The evolution trajectory of the key network state sequence after the pre-executive command is executed, and a compensation network plane is dynamically constructed to address potential resource bottlenecks, as follows: A virtual simulation environment is established based on the network interference correlation diagram and historical state data. The multi-dimensional collaborative scheduling instructions are input into the virtual simulation environment to simulate the network operation process after the parameters of each network element are adjusted, and the estimated network performance index sequence during the simulation operation is output. The network performance index sequence is used as the virtual pre-simulation trajectory. The time delay sequence, packet loss rate sequence, and bandwidth utilization rate sequence in the virtual pre-simulation trajectory are compared with the corresponding time delay upper limit, packet loss rate upper limit, and bandwidth utilization rate upper limit in the cooperative state target, respectively, on a time-by-time basis. Identify the time period in which any performance index sequence value continuously exceeds its corresponding upper limit and the target network element with the highest load during that time period. Mark the time period and the target network element together as a resource bottleneck, and calculate the average difference between the performance index sequence value and its corresponding upper limit as the performance gap. Based on the performance gap and the network type of the target network element, determine the type and specifications of the virtual resources that need to be supplemented; In the virtual simulation environment, during the time period corresponding to the resource bottleneck, virtual network function nodes of appropriate specifications are deployed and configured for the affected service flows. The virtual network functional nodes are connected to the relevant physical networks to form a logically isolated compensation network plane.

9. A heterogeneous network space multi-layered mobile Internet of Things cloud system based on artificial intelligence according to claim 1, characterized in that, Execute multi-dimensional collaborative scheduling instructions that incorporate the compensation network plane logic, monitor the real-time status data generated after execution, and match and analyze the actual state evolution trajectory with the virtual pre-simulation trajectory to pinpoint the interference propagation chain corresponding to the root cause of the deviation. Based on this, update the network interference correlation graph as follows: The multi-dimensional collaborative scheduling command is executed, and the actual performance index data of each network element in the monitoring window are collected after the command takes effect through the network probe and terminal reporting mechanism. The actual performance index data is time-aligned and the corresponding performance index sequence in the virtual pre-simulation trajectory generated by the network resource self-growth module is calculated; Identify continuous time periods in which the difference exceeds a preset deviation threshold, and mark the performance index with the largest difference and its corresponding network element within that time period as significant performance deviation and the network element where the deviation occurred; Starting from the network element where the deviation occurs, in the network interference association graph, based on the weight of the edges in the graph, iteratively backtrack to the upstream network node that is its neighbor and has a higher edge weight; The edges traversed by the backtracking search are connected in sequence to form a path from an upstream network node to the network element where the deviation occurred. This path is identified as the interference propagation chain that caused the significant performance deviation, and the upstream network node is the root cause node of the deviation. Based on the historical correlation strength of each edge in the interference propagation chain within the time period corresponding to the significant performance deviation, the weights of the corresponding edges in the network interference correlation graph are enhanced and updated.

Images