An internet of things monitoring scene risk prediction system and method based on a space-time graph neural network

Abstract

The application provides an Internet of Things monitoring scene risk prediction system and method based on a space-time graph neural network, and relates to the technical field of Internet of Things edge computing. The system comprises: a multivariate time series data acquisition module for acquiring standardized multivariate time series data of a current period; an orthogonal decomposition module for performing orthogonal decomposition on the standardized multivariate time series data of the current period to output three types of orthogonal risk propagation modes, namely instantaneous risk, local propagation and global diffusion; a dynamic risk conduction graph generation module for generating a dynamic adjacency matrix that truly reflects a risk propagation path based on mode data; a mode high-dimensional feature generation module for extracting multi-modal space-time features in parallel in combination with the dynamic adjacency matrix and the mode data; and a prediction module for completing risk prediction based on the high-dimensional features. The system can simultaneously capture different scale risk propagation modes through multi-scale risk decoupling and exclusive operator design, thereby significantly improving prediction accuracy.

Description

Technical Field

[0001] This application relates to the field of IoT edge computing technology, specifically to a risk prediction system and method for IoT monitoring scenarios based on spatiotemporal graph neural networks. Background Technology

[0002] With the widespread application of IoT technology in key areas such as industrial control, smart cities, and energy, the scale of IoT systems is growing exponentially, and their security risks are becoming increasingly prominent. Traditional IoT security monitoring systems are mainly based on rule matching and anomaly detection of single devices, which cannot effectively capture the spatiotemporal propagation characteristics of risks between devices, resulting in delayed early warnings and high false negative rates.

[0003] In recent years, Spatio-Temporal Graph Neural Networks (STGNNs) have been widely used in IoT risk prediction tasks due to their ability to simultaneously model temporal dependencies and spatial correlations. However, existing STGNN technologies have the following drawbacks in practical applications: The contradiction between the static graph structure assumption and dynamic risk propagation: Most existing methods construct graph structures based on predefined static adjacency matrices (such as physical connections and geographical distances), which cannot adapt to the dynamic changes of IoT devices (offline, new additions, and mobile devices) and the real-time evolution of risk propagation paths. Although some studies have proposed dynamic graph learning methods based on attention mechanisms, these methods can only capture the instantaneous correlation between nodes and cannot reflect the intrinsic dynamic mechanism of risk propagation.

[0004] Insufficient modeling of multi-scale risk propagation characteristics: Risk propagation in IoT systems exhibits significant multi-scale characteristics: instantaneous risks (such as device failure) affect only a single node, locally propagating risks (such as network attacks) spread between adjacent devices, while globally propagating risks (such as worm viruses) can spread rapidly throughout the entire network. Existing STGNNs employ a unified spatiotemporal convolution operator, which cannot effectively capture risk propagation patterns at different scales simultaneously.

[0005] The ability to predict long-tail risks is severely inadequate: IoT risk data exhibits an extreme long-tail distribution, with over 99% of samples representing normal states, while samples of high-risk events (such as major safety incidents) are extremely scarce. Existing models are trained based on traditional loss functions such as mean squared error, tending to predict majority class samples, resulting in extremely low accuracy in predicting low-probability, high-risk events. Summary of the Invention

[0006] The purpose of this invention is to provide a risk prediction system and method for IoT monitoring scenarios based on spatiotemporal graph neural networks to at least solve the above-mentioned technical problems.

[0007] One aspect of the present invention provides a risk prediction system for IoT monitoring scenarios based on spatiotemporal graph neural networks, the IoT monitoring scenario risk prediction system based on spatiotemporal graph neural networks comprising: A multivariate time series data acquisition module, which is used to acquire standardized multivariate time series data for the current period; The orthogonal decomposition module is used to perform orthogonal decomposition on the standardized multivariate time series data of the current period to obtain the risk propagation mode data set of the current period, wherein the risk propagation mode data set includes instantaneous risk mode, local propagation mode and global diffusion mode; A dynamic risk transmission diagram generation module is used to generate a dynamic adjacency matrix that can truly reflect the risk transmission path in the current period based on the risk propagation modality data set. The modal high-dimensional feature generation module is used to perform parallel extraction of multimodal spatiotemporal features based on the dynamic adjacency matrix that can truly reflect the risk propagation path in the current period and the risk propagation modal data group in the current period, thereby obtaining the modal high-dimensional feature information of the current period. The prediction module is used to predict the risk of the current period based on the modal high-dimensional feature information of the current period, thereby obtaining the prediction result of the current period.

[0008] Optionally, the IoT monitoring scenario risk prediction system based on spatiotemporal graph neural networks further includes: The risk tracing and analysis module is used to obtain the risk tracing results for the current period based on the prediction results of the current period and the modal high-dimensional feature information of the current period.

[0009] Optionally, the orthogonal decomposition module includes: The instantaneous risk mode extraction module is used to extract modes from the standardized multivariate time series data of the current period based on the topology of the Internet of Things (IoT) devices, thereby obtaining the instantaneous risk mode tensor of the current period; wherein, the topology of the IoT devices includes multiple device nodes; A local propagation mode extraction module is used to generate a local propagation mode tensor for the current period based on the instantaneous risk mode tensor for the current period. A global diffusion mode extraction module is used to generate a global diffusion mode tensor for the current period based on the instantaneous risk mode tensor for the current period.

[0010] Optionally, the orthogonal decomposition module further includes: The reconstruction calibration module is used to perform orthogonality constraints and lossless reconstruction calibration on the instantaneous risk mode tensor, the local propagation mode tensor, the global diffusion mode tensor, and the standardized multivariate time series data of the current period during the training phase, so as to obtain the updated decomposition parameters.

[0011] Optionally, the dynamic risk transmission diagram generation module includes: The node-to-modal mutual information calculation module is used to generate an instantaneous modal mutual information matrix, a local modal mutual information matrix, and a global modal mutual information matrix based on the instantaneous risk modal tensor, the local propagation modal tensor, and the global diffusion modal tensor of the current period. A weighted fusion module for comprehensive risk correlation is used to generate the original comprehensive risk correlation matrix for the current period based on the instantaneous modal mutual information matrix, the local modal mutual information matrix, and the global modal mutual information matrix. A row-level sparsification module for the correlation matrix is ​​used to generate a sparse comprehensive risk correlation matrix for the current period based on the original comprehensive risk correlation matrix for the current period and the topology of IoT devices. An adjacency matrix construction module is used to generate a dynamic adjacency matrix for the current period based on the sparse comprehensive risk correlation matrix of the current period and the dynamic adjacency matrix of the previous period.

[0012] Optionally, the modal high-dimensional feature generation module includes: A transient risk mode temporal feature extraction module is used to extract the transient risk mode feature tensor of the current period based on the transient risk mode tensor of the current period; A local propagation mode spatiotemporal feature extraction module is used to extract the local propagation mode feature tensor of the current period based on the local propagation mode tensor of the current period; A global diffusion mode spatiotemporal feature extraction module is used to generate a global diffusion mode feature tensor for the current period based on the global diffusion mode tensor for the current period.

[0013] Optionally, the prediction module includes: A multimodal feature dimension unification module is used to unify the dimensions of the instantaneous risk modal feature tensor, the local propagation modal feature tensor, and the global diffusion modal feature tensor of the current period, thereby obtaining a unified instantaneous risk modal feature tensor, a unified local propagation modal feature tensor, and a unified global diffusion modal feature tensor. The feature fusion module is used to fuse the instantaneous risk modal feature tensor, the local propagation modal feature tensor, and the global diffusion modal feature tensor of the same dimension in a unified dimension, so as to obtain the fused comprehensive spatiotemporal feature tensor and the modal weight vector of each device. A fully connected layer computing module is used to generate the original risk prediction value tensor from the fused integrated spatiotemporal feature tensor. The post-processing and standardization module is used to generate the final risk prediction value tensor for the current period based on the original risk prediction value tensor.

[0014] Another aspect of the present invention provides a method for risk prediction in IoT monitoring scenarios based on spatiotemporal graph neural networks, the method comprising: Obtain standardized multivariate time series data for the current period; Orthogonal decomposition is performed on the standardized multivariate time series data of the current period to obtain the risk propagation modality data set of the current period, which includes instantaneous risk modality, local propagation modality, and global diffusion modality. Generate a dynamic adjacency matrix that accurately reflects the risk propagation path for the current period based on the risk propagation modality data set; Based on the dynamic adjacency matrix that can truly reflect the risk propagation path in the current period and the risk propagation modal data group in the current period, multimodal spatiotemporal features are extracted in parallel to obtain the modal high-dimensional feature information of the current period. Risk prediction for the current period is performed based on the modal high-dimensional feature information of the current period, thereby obtaining the prediction result for the current period.

[0015] Optionally, the IoT monitoring scenario risk prediction method based on spatiotemporal graph neural networks further includes: The risk tracing results for the current period are obtained based on the prediction results and the modal high-dimensional feature information of the current period.

[0016] The IoT monitoring scenario risk prediction system based on spatiotemporal graph neural network proposed in this application decomposes the original data into three orthogonal risk propagation modes and designs a special spatiotemporal convolution operator for each mode, which can effectively capture risk propagation patterns at different scales at the same time and improve prediction accuracy. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of an IoT monitoring scenario risk prediction system based on a spatiotemporal graph neural network according to an embodiment of this application; Figure 2 This is a flowchart of a risk prediction method for IoT monitoring scenarios based on a spatiotemporal graph neural network according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some, but not all, embodiments of this application. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0019] like Figure 1 The IoT monitoring scenario risk prediction system based on spatiotemporal graph neural networks shown includes: A multivariate time series data acquisition module, which is used to acquire standardized multivariate time series data for the current period; The orthogonal decomposition module is used to perform orthogonal decomposition on the standardized multivariate time series data of the current period to obtain the risk propagation mode data set of the current period, wherein the risk propagation mode data set includes instantaneous risk mode, local propagation mode and global diffusion mode; A dynamic risk transmission diagram generation module is used to generate a dynamic adjacency matrix that can truly reflect the risk transmission path in the current period based on the risk propagation modality data set. The modal high-dimensional feature generation module is used to perform parallel extraction of multimodal spatiotemporal features based on the dynamic adjacency matrix that can truly reflect the risk propagation path in the current period and the risk propagation modal data group in the current period, thereby obtaining the modal high-dimensional feature information of the current period. The prediction module is used to predict the risk of the current period based on the modal high-dimensional feature information of the current period, thereby obtaining the prediction result of the current period.

[0020] In this embodiment, the application adopts a periodic time series prediction architecture based on a sliding time window. All processing steps are repeatedly executed within a fixed time period ΔT, and each period corresponds to a complete processing flow of a sliding window.

[0021] In this embodiment, the system operation cycle is ΔT (configurable, typically from 1 minute to 1 hour, adjusted according to the real-time requirements of the monitoring scenario). The length of the historical sliding window is T (including the current time and T consecutive time steps). Predict the future step size as K (predict the risk value for the next K time steps). Time indexing rules: The nth running cycle corresponds to the current time tn, the historical data window processed is [tn-T+1, tn], and the predicted future window is [tn+1, tn+K]. State transfer mechanism: The output state of each cycle (dynamic adjacency matrix, hidden state of the loop layer) will be used as the input history state of the next cycle.

[0022] In this embodiment, the multivariate time series data for IoT device status monitoring refers to structured data formed by synchronously collecting multiple operating status parameters of multiple IoT devices at consecutive time points. Each data sample contains F monitoring feature values ​​of N devices at T consecutive time steps, forming a three-dimensional tensor structure.

[0023] In this embodiment, the three-dimensional tensor representation of the input data is as follows: The physical meaning and value range of each dimension are as follows: The first dimension, N, represents the total number of IoT devices. The value range depends on the scale of the monitoring scenario and is typically between 100 and 10,000. For example, in industrial control scenarios, N is usually between 500 and 2,000; in smart city scenarios, N is usually between 1,000 and 10,000. It should be noted that this example is merely illustrative and does not limit the scope of this embodiment.

[0024] The second dimension, T, is the length of the historical sliding window, which is the number of historical data time steps used in each prediction period. It is a hyperparameter with a value range of 12 to 24, and a recommended value is 16.

[0025] The third dimension, F, is the monitoring feature dimension for each device, which is the number of operating status parameters collected for each device. The value ranges from 5 to 50, depending on the device type and monitoring requirements.

[0026] In this embodiment, the application is for a communication base station or communication tower scenario.

[0027] In this embodiment, the first dimension N represents the total number of communication base stations within a single monitoring area, typically ranging from 50 to 500, corresponding to the base station scale of a city or county. Each base station serves as an independent graph node, containing all monitoring parameters for the equipment room, tower-mounted equipment, and tower structure.

[0028] The second dimension, T, is the length of the historical sliding window, a hyperparameter. Given the rapid spread of risks from communication base stations, a value of 12 is recommended, corresponding to 1 hour of historical data (collected at a frequency of 5 minutes per instance).

[0029] The third dimension F is the monitoring feature dimension for each base station, covering the following: (I) Operating parameters of core equipment in base station equipment room (10 dimensions) These parameters reflect the operational status of the core communication equipment and supporting facilities inside the base station equipment room, and are the foundation for the normal operation of the base station. They mainly include: BBU (Baseband Unit) core board temperature: The real-time temperature of the core chip of the baseband processing unit, in degrees Celsius. The normal range is 0 to 60 degrees Celsius, and values ​​exceeding 70 degrees Celsius are abnormal.

[0030] BBUCPU (Central Processing Unit) utilization: The real-time percentage of CPU usage in the baseband processing unit. The normal range is 10% to 60%, and anything above 80% is abnormal.

[0031] BBU memory usage: The real-time percentage of memory occupied by the baseband processing unit. The normal range is 20% to 70%, and anything above 90% is abnormal.

[0032] RRU (Remote Radio Unit) Average Transmit Power: The average transmit power of the remote radio unit, in dBm. The normal range is 43 to 46, and a value below 40 is abnormal.

[0033] RRU receiver sensitivity: The receiver signal sensitivity of the remote radio unit, measured in dBm. The normal range is -110 to -90, and values ​​above -85 are abnormal.

[0034] DC bus voltage: The bus voltage of the base station DC power supply system, in volts. The normal range is -48±5. Values ​​below -40 or above -56 are abnormal.

[0035] Battery pack remaining capacity: The percentage of remaining capacity of the backup battery pack. The normal range is 20% to 100%. A value below 10% is abnormal.

[0036] Air conditioning return air temperature: The return air temperature of the computer room air conditioner, in degrees Celsius. The normal range is 20 to 28 degrees Celsius. A value exceeding 30 degrees Celsius is considered abnormal.

[0037] Data center humidity: The relative humidity inside the data center, expressed as a percentage. The normal range is 30 to 70%. Values ​​below 20 or above 80 are considered abnormal.

[0038] Smoke sensor status: Alarm status of the smoke sensor in the computer room, 0 indicates normal, 1 indicates alarm.

[0039] (II) Operating parameters of radio frequency equipment on the tower (6 dimensions) These parameters reflect the operating status of the radio frequency equipment and antenna system installed on the communication tower, directly affecting the coverage quality of the wireless network. They mainly include: Antenna azimuth deviation: The difference between the actual azimuth of the antenna and the designed azimuth, in degrees. The normal range is -5 to +5, and a value exceeding ±10 is abnormal.

[0040] Antenna downtilt deviation: The difference between the actual downtilt angle of the antenna and the designed downtilt angle, in degrees. The normal range is -2 to +2, and a value exceeding ±5 is abnormal.

[0041] Feeder VSWR: The voltage VSWR of the antenna feeder system is normally within the range of 1.0 to 1.5. A value exceeding 2.0 is abnormal.

[0042] Tower amplifier gain: The actual gain of the amplifier at the top of the tower, in dB. The normal range is 10 to 15, and a value below 8 is abnormal.

[0043] RRU chassis temperature: The internal temperature of the remote radio frequency unit chassis on the tower, in degrees Celsius. The normal range is -40 to 65 degrees Celsius, and values ​​exceeding 70 degrees Celsius are abnormal.

[0044] GPS (Global Positioning System) synchronization status: The working status of the base station GPS synchronization module. 0 indicates normal synchronization, and 1 indicates synchronization loss.

[0045] (III) Safety parameters of tower structure (7 dimensions) These parameters reflect the structural safety status of the communication tower itself and are crucial for preventing major safety accidents such as tower collapse. They mainly include: Tower tilt angle in the north-south direction: The tilt angle of the communication tower in the north-south direction, in degrees. The normal range is -0.5 to +0.5. Any value exceeding ±1.0 is abnormal.

[0046] East-west tilt angle of the tower: The tilt angle of the communication tower in the east-west direction, in degrees. The normal range is -0.5 to +0.5. An angle exceeding ±1.0 is abnormal.

[0047] Tower vibration acceleration: The amplitude of vibration acceleration at the top of the communication tower, in m / s², with a normal range of 0 to 0.1, and values ​​exceeding 0.5 are considered abnormal.

[0048] Tower foundation settlement: The uneven settlement of the communication tower foundation, measured in millimeters. The normal range is 0 to 5, and values ​​exceeding 20 are considered abnormal.

[0049] Bolt stress value: The stress value of the key connecting bolts of the tower body, in MPa. The normal range is 0 to 100, and values ​​exceeding 200 are abnormal.

[0050] Tower ice thickness: The thickness of ice on the tower and antenna, in millimeters. The normal range is 0 to 10, and values ​​exceeding 50 are abnormal.

[0051] Wind speed: Real-time wind speed at the location of the communication tower, in m / s. The normal range is 0 to 20, and values ​​exceeding 30 are considered abnormal.

[0052] (iv) Wireless Network Performance Parameters (6 Dimensions) These parameters reflect the quality of wireless network service provided by the base station, directly reflecting user perception and serving as an important indicator of network faults. They mainly include: Downlink average throughput: The average downlink throughput of users in the cell, in Mbps. The normal range is 10 to 100, and a value below 5 is abnormal.

[0053] Average uplink throughput: Average uplink throughput per user in the cell, in Mbps. The normal range is 2 to 20, and a value below 1 is abnormal.

[0054] Wireless connection success rate: The success rate of establishing a wireless channel in a cell, expressed as a percentage. The normal range is 99.0% to 100%, and values ​​below 95% are considered abnormal.

[0055] Call drop rate: The rate of dropped calls in a cell, expressed as a percentage. The normal range is 0 to 1.0, and a value exceeding 3.0 is considered abnormal.

[0056] Handover success rate: The success rate of handover between small intervals, in percentage. The normal range is 98.0% to 100%. A value below 90% is abnormal.

[0057] PRB utilization rate: The average utilization rate of physical resource blocks in a cell, expressed as a percentage. The normal range is 20% to 70%, and values ​​exceeding 90% are considered abnormal.

[0058] (v) Surrounding environmental parameters (3D) These parameters reflect the natural environmental conditions surrounding communication base stations and are an important basis for predicting base station failures caused by natural disasters. They mainly include: Ambient temperature: The ambient temperature at the location of the base station, in degrees Celsius, ranging from -40 to 50.

[0059] Rainfall: Hourly rainfall at the location of the base station, in millimeters, ranging from 0 to 100.

[0060] Lightning frequency: The number of lightning strikes per hour within a 1-kilometer radius of the base station location, ranging from 0 to 100.

[0061] III. Scenario-Specific Data Acquisition and Preprocessing Specifications (a) Acquisition frequency Different sampling frequencies are used for different types of parameters, taking into account the operational characteristics of communication base stations: Operating parameters of equipment in the computer room and performance parameters of the wireless network: The sampling frequency is 5 minutes / time, which is the standard sampling frequency of the operator's existing monitoring system.

[0062] Tower structure safety parameters: Under normal circumstances, the sampling frequency is once every hour. When the wind speed exceeds 15m / s or the rainfall exceeds 20mm / h, the sampling frequency is automatically increased to once every 5 minutes.

[0063] Surrounding environmental parameters: The sampling frequency is 1 hour / time.

[0064] Before inputting into the network, all parameters are uniformly aligned to a time step of 5 minutes per iteration using an interpolation method.

[0065] (II) Preprocessing Specifications In addition to the common missing value imputation, outlier truncation, and Z-score normalization, the following preprocessing steps are added to address the characteristics of communication base station data: Base station status filtering: Filter out base station data that is under maintenance or out of service to avoid interfering with model training.

[0066] Holiday feature coding: Add a one-dimensional holiday feature, where 0 represents a weekday and 1 represents a weekend or statutory holiday, to capture the periodic changes in network traffic during holidays.

[0067] Time feature encoding: Add two time features, namely hour features (0 to 23) and week features (0 to 6), to capture the daily and weekly periodicity of network traffic.

[0068] After the above preprocessing, the final input feature dimension of each base station is 35 dimensions (32-dimensional monitoring parameters + 1-dimensional holiday features + 2-dimensional time features).

[0069] In this embodiment, this application employs a base station-level integrated risk label, where each base station corresponds to a risk value at each time step, comprehensively reflecting the base station's equipment failure risk, structural security risk, and network service risk. The risk value ranges from 0 to 1, and the specific level classification is as follows: 0.00-0.20: Low risk: The base station is operating normally, all parameters are within the normal range, and the network service quality is good.

[0070] 0.21-0.50: Medium risk: The base station has minor anomalies, with some parameters exceeding the normal range, but network services are not affected. Closer monitoring is required.

[0071] 0.51-0.80: High risk: The base station has obvious abnormalities, multiple parameters are outside the normal range, network service quality is degraded, and it needs to be dealt with within 24 hours.

[0072] 0.81-1.00: Extremely high risk: The base station is about to experience a major failure such as station failure or tower tilting, or partial service interruption has already occurred, requiring immediate action.

[0073] In this embodiment, the standardized multivariate time series data for the nth period Where N is the total number of IoT devices, T is the length of the historical sliding window (the number of continuous time steps including the current moment, usually 12 to 24), and F is the monitoring feature dimension of each device.

[0074] Standardized eigenvalues ​​of the last time step in the (n-1)th period ,in For device node indexing. Indexed by feature dimensions.

[0075] In this embodiment, the present application first obtains the topology of the IoT devices. The initial topology is an undirected graph with communication base stations as nodes, which is used to provide a basic spatial reference for the first prediction cycle of the system's initial operation. During the first system operation, without the dynamic risk transmission graph generated in the previous cycle, the risk spatiotemporal prism decomposition and dynamic graph generation calculations for the first cycle must be completed based on the initial topology. From the second prediction cycle onwards, the system will use the dynamic risk transmission graph generated based on modal mutual information evolution proposed in this application for all calculations, and the initial topology will no longer participate in any subsequent processing.

[0076] In this embodiment, the initial topology is obtained using the following method: Node definition: Each independent communication base station is treated as a node in the initial topology graph, and the number of nodes is the total number of base stations N in the monitoring area.

[0077] Initial first-order neighbor set Get: Basic data extraction: Extract latitude and longitude coordinate data of all monitored base stations from the operator's base station resource management system, and extract the standard neighbor cell relationship table of all base stations (i.e., the preset measurement neighbor cell list of each base station during the handover process) from the core network configuration system.

[0078] Physical proximity neighbor filtering: Based on latitude and longitude, calculate the spherical distance between any two base stations, filter out all base stations whose distance to the target base station is less than 3 kilometers (the general threshold for macro base stations), and form a set of physical proximity neighbors.

[0079] Logical adjacency neighbor supplementation: Supplement all base stations in the target base station's neighbor relationship table that are not in the physical neighbor set, and finally form the initial first-order neighbor set of the base station, which is the union of physical neighbor and logical adjacency neighbors.

[0080] Edge definition: If two base stations are initial first-order neighbors of each other, then there is an undirected edge between these two nodes in the initial topology graph.

[0081] Initial adjacency matrix generation: Based on the definitions of nodes and edges above, an initial adjacency matrix is ​​generated. An element of 1 in the matrix indicates that an edge exists between two base stations, while an element of 0 indicates that no edge exists. All diagonal elements of the matrix are set to zero, and the matrix is ​​kept symmetric.

[0082] Initial graph Laplacian matrix calculation: The symmetric normalized graph Laplacian calculation method in the field of spatiotemporal graph neural networks is adopted. The initial symmetric normalized graph Laplacian matrix is ​​calculated based on the initial adjacency matrix and used for local propagation mode extraction in the first cycle.

[0083] In this embodiment, the orthogonal decomposition module may include: an instantaneous risk mode extraction module, a local propagation mode extraction module, and a global diffusion mode extraction module; these will be described in detail below.

[0084] In this embodiment, the instantaneous risk mode extraction module is used to extract modes from the standardized multivariate time series data of the current period according to the topology of the Internet of Things (IoT) devices, thereby obtaining the instantaneous risk mode tensor of the current period; wherein, the topology of the IoT devices includes multiple device nodes.

[0085] In this embodiment, modal extraction is performed on the standardized multivariate time series data of the current period based on the topology of the IoT devices to obtain the instantaneous risk modal tensor of the current period, including: Traverse all time steps within the history window (12 represents the historical sliding window length), all base station nodes and all 35 feature dimensions For each position Perform the following calculations in sequence: Calculate the time abrupt change component, if =1 (the first time step of the current window), then the standardized eigenvalue of the last time step of the previous period is used. (The standardized eigenvalue of the last time step of the (n-1)th cycle) is used as the value of the previous time step; otherwise, the eigenvalue of the previous time step within the current window is used. The absolute difference between the feature values ​​at the current time and the previous time is calculated to obtain the time mutation component at that location. This component accurately reflects the degree of sudden change of the corresponding feature in the time dimension. For example, instantaneous anomalies such as a sudden drop in the transmission power of the base station RRU or a sudden rise in the temperature of the equipment room will show a significantly high value in this component.

[0086] Calculate spatially isolated components for each base station. Calculate the current time step of all its initial first-order neighbor nodes. and characteristics The arithmetic mean of the following: ,in For base stations The initial number of neighbors is calculated, and then the absolute difference between the current base station characteristic value and the mean is calculated to obtain the spatial isolation component of the location. This component can effectively distinguish the anomaly of a single base station from the common fluctuation of all base stations in the area. For example, when the GPS synchronization status of a base station is suddenly lost, its spatial isolation component will increase significantly, while when thunderstorms cause the reception sensitivity of all base stations in the area to decrease generally, this component will remain at a low value.

[0087] Perform feature-level weighted fusion using learnable feature-level weight coefficients. and (respectively the first) The weights of the temporal abrupt component and the spatially isolated component of each feature are initially set to 0.5. During training, they will be automatically optimized according to the risk contribution of different features. For example, for structural safety parameters such as the tower tilt angle, the weight of the spatially isolated component will automatically increase, while for network performance parameters such as PRB (Physical Resource Block) utilization, the weight of the temporal abrupt component will automatically increase. The instantaneous risk component at that location is obtained by weighted summation of the temporal abrupt component and the spatially isolated component. The calculation formula is as follows: ; in The feature value at the current time. The eigenvalue is the characteristic value of the previous time step.

[0088] After the complete position-by-position calculation described above, the instantaneous risk mode tensor of the nth period is finally obtained. Each element of this tensor contains only a single burst anomaly signal of the corresponding base station at the corresponding time step, completely filtering out risk information from local propagation and global diffusion. For example, when the remaining capacity of a base station's battery pack suddenly drops from 80% to 10% at a certain time step, the instantaneous risk component at that location will show a significantly high value, while the normal capacity fluctuations of surrounding base stations will not have any impact on it. This tensor will be directly input into the instantaneous modal GRU (Gated Recurrent Unit) branch of the subsequent multimodal spatiotemporal convolutional layer to extract the temporal evolution risk features of a single base station.

[0089] A local propagation mode extraction module is used to generate a local propagation mode tensor for the current period based on the instantaneous risk mode tensor for the current period.

[0090] In this embodiment, generating the local propagation mode tensor for the current period based on the instantaneous risk mode tensor of the current period includes: Pre-compute the initial symmetric normalized graph Laplacian matrix, which forms the basis for extracting local spatial correlations: First, construct the pair angle matrix based on the initial adjacency matrix. The diagonal element D( , ) equals base station The initial number of neighbors, with all off-diagonal elements being 0; then using the formula: The symmetric normalized graph Laplacian matrix is ​​calculated, where The matrix is ​​an N×N identity matrix. Symmetric normalization is used to ensure the numerical stability of subsequent graph convolution operations and to avoid uneven feature distribution caused by differences in node degree.

[0091] Then iterate through all time steps. All base station nodes and all 35 feature dimensions For each position Perform the following calculations in sequence: The first step is to calculate the time moving average, taking the current time and the previous time. ( WThe local time window size is the sliding window length used to extract local time trends. It's a hyperparameter, and given the characteristics of the local risk propagation speed at communication base stations (typically 10-20 minutes), a value of 3 is taken, corresponding to a 15-minute time span. The arithmetic mean of the feature values ​​at each time point is calculated. If the time index is less than 1 (i.e., ... If <1), then the eigenvalue of the corresponding position in the previous cycle is used. (The standardized eigenvalues ​​corresponding to the time step of the (n-1)th period are used for padding. The formula for calculating the time moving average is: ; This operation can effectively filter out high-frequency instantaneous change signals and retain the local time trends of medium and low frequencies. For example, interference signals within the same base station cluster will gradually increase and affect adjacent base stations within 15 minutes. This slow-changing trend will be fully preserved, while the instantaneous power fluctuations of a single base station will be smoothly filtered out.

[0092] Execution space Laplace smoothing, for each time step and characteristics The spatial feature vector after time moving average and Multiply, where This is a learnable spatial smoothness coefficient (initially set to 0.5, and automatically optimized during training based on the spatial propagation characteristics of different features: for parameters with strong spatial propagation characteristics such as network interference and signal coverage), It will automatically rise to 0.7-0.9; for parameters with weak spatial propagation characteristics, such as computer room air conditioning temperature and battery capacity, It will automatically decrease to 0.2-0.4). The formula for calculating spatial Laplace smoothing is: ; This operation can weightedly fuse the features of a single base station with the features of its neighboring base stations to extract spatial correlations that exist only in local base station clusters. For example, when the downlink throughput of a certain base station decreases due to interference, the throughput of its neighboring base stations will also show a similar downward trend. This local spatial correlation will be amplified and preserved by this operation, while the independent anomalies of a single base station or the common fluctuations of all base stations in the network will be suppressed.

[0093] After the complete position-by-position calculation described above, the local propagation mode tensor of the nth period is finally obtained. Each element of this tensor contains only risk signals that gradually spread between adjacent base stations, completely filtering out instantaneous single-point anomalies and large-scale risk information that spreads globally. Typical local propagation risks will exhibit clear spatiotemporal patterns in this tensor: for example, co-channel interference within a base station cluster will start from the central base station and gradually spread to 3-5 surrounding base stations within 3-5 time steps, manifested as a synchronous increase in the local propagation components of parameters such as PRB utilization and downlink throughput of these base stations; another example is that thunderstorms in a local area will cause a synchronous decrease in the receiving sensitivity of all base stations in that area, and such regional common fluctuations will also be fully extracted.

[0094] A global diffusion mode extraction module is used to generate a global diffusion mode tensor for the current period based on the instantaneous risk mode tensor for the current period.

[0095] In this embodiment, the global diffusion mode tensor of the current period is generated based on the instantaneous risk mode tensor of the current period. This is specifically used to extract the global risk signal that spreads synchronously throughout the entire monitoring area (city / county). The global diffusion risk has network-wide consistency, low-frequency trend and strong global spatial correlation. It is the dominant component with the highest energy proportion among all base station status changes, while instantaneous single-point anomalies and local area fluctuations are only secondary components.

[0096] By extracting the main global trend components from the data through singular value decomposition, it is possible to accurately separate risk signals that evolve synchronously only across the entire network, forming a strict orthogonal distinction from the high-frequency single-point characteristics of instantaneous risks and the medium- and low-frequency regional characteristics of local risks.

[0097] Perform tensor dimension transformation to reconstruct the standardized multivariate time series data of the communication base station in the nth period into a two-dimensional spatiotemporal matrix for singular value decomposition: Each base station is treated as a row in a matrix. The 35 features of that base station at 12 time steps are concatenated in chronological order to form a one-dimensional vector of length 420, ultimately yielding a two-dimensional spatiotemporal matrix. This dimensionality transformation method can completely preserve all historical state information of each base station, while transforming global risk into low-rank principal components of the matrix. This is because the risk changes synchronized across the entire network will cause the feature vectors of all base stations to exhibit a high degree of linear correlation, thus appearing as the first few principal components with concentrated energy in singular value decomposition.

[0098] For two-dimensional spacetime matrix Performing singular value decomposition yields three matrices: the left singular matrix. Each column of the singular value matrix corresponds to a global spatial pattern. Its diagonal elements are singular values ​​sorted in descending order, and each singular value corresponds to the energy magnitude of a principal component; the right singular matrix Each column corresponds to a global time pattern. The singular value decomposition satisfies the equation... It can decompose the original data into the sum of multiple mutually orthogonal principal components.

[0099] Principal component selection is performed, and the energy is accumulated in descending order of singular values ​​before calculation. The cumulative energy percentage of each singular value: ,in Let be the rank of the matrix, which is the maximum number of non-zero singular values.

[0100] Choose the smallest integer This results in a cumulative energy percentage ≥0.95, meaning retaining the top performers that explain more than 95% of the data variance. One principal component.

[0101] This principal component selection method ensures the complete retention of all global risk information while filtering out local fluctuations and transient noise with an energy share of less than 5%. In communication base station scenarios, the value of K is typically between 3 and 8, corresponding to a few dominant global factors such as core network status, overall network power supply status, and large-scale weather conditions.

[0102] Perform global matrix reconstruction, before use Reconstructing the two-dimensional global diffusion matrix from principal components: ; in, The first part of the left singular matrix represents the front part of the left singular matrix. List, The first part of the singular value matrix represents the first part of the singular value matrix. Before departure List, The first part of the right singular matrix represents the first part of the right singular matrix. List.

[0103] The reconstructed matrix contains only the global trend component from the original data; fluctuations in all local base station clusters and transient changes in individual base stations are completely filtered out. Finally, tensor dimension restoration is performed to reconstruct the two-dimensional global diffusion matrix. It is re-deformed into a three-dimensional tensor with the same dimensions as the original input, i.e. ,in For base station indexing, For time step index, Indexed by feature dimensions.

[0104] After the above complete calculation, the global diffusion mode tensor of the nth period is finally obtained. Each element of this tensor contains only global risk signals that are synchronously spreading throughout the entire monitoring area, completely filtering out instantaneous single-point anomalies and locally spreading regional risk information. Typical global spread risks will present a clear network-wide synchronous pattern in this tensor: for example, a core network failure causing a sudden drop in the wireless connection rate and handover success rate of all base stations in the entire city will result in a significant high value in the global spread mode; another example is a large-scale power outage causing a synchronous decrease in the remaining battery capacity of base stations in multiple districts and counties, which will be fully extracted as a cross-regional power supply risk; and typhoon weather causing a synchronous increase in parameters such as tower vibration acceleration, wind speed, and rainfall of all base stations in the entire monitoring area will also be retained as a global trend. Non-global risks such as power failures of individual base stations and co-channel interference of local base station clusters will not appear in this mode because their energy proportion is low and they are filtered out during principal component reconstruction.

[0105] In this embodiment, the orthogonal decomposition module may further include a reconstruction calibration module, which is used during the training phase.

[0106] The reconstruction calibration module is used to perform orthogonality constraints and lossless reconstruction calibration on the instantaneous risk mode tensor, the local propagation mode tensor, the global diffusion mode tensor, and the standardized multivariate time series data of the current period during the training phase, so as to obtain the updated decomposition parameters.

[0107] In this embodiment, during the training phase, orthogonality constraints and lossless reconstruction calibration are applied to the instantaneous risk mode tensor, the local propagation mode tensor, the global diffusion mode tensor, and the standardized multivariate time series data of the current period, thereby obtaining the updated decomposition parameters, including: Perform a tensor flattening operation to transform the three three-dimensional risk mode tensors and the instantaneous risk mode tensor of the nth period. The local propagation mode tensor of the nth period The global diffusion mode tensor of the nth period Convert each matrix into a two-dimensional matrix to calculate the correlation between the matrices and the reconstruction error: By concatenating the base station dimension and time dimension of each modality tensor in sequence, we obtain a shape of... The flattened matrix, i.e. This flattening approach treats the 35-dimensional feature vector of each base station at each time step as an independent sample, ensuring that subsequent calculations can measure the overall correlation of the three modalities across all features of all samples.

[0108] The orthogonality constraint loss is calculated to penalize information overlap among the three modes: the squared Frobenius norm of the pairwise product of each of the three flattened mode matrices is calculated separately, and the three results are summed to obtain the orthogonality constraint loss. The calculation formula is: ; in Let Frobenius norm be the square root of the sum of the squares of all elements of a matrix. If two modes are perfectly orthogonal, their product matrix will be zero, and the square of the corresponding Frobenius norm will also be zero. In communication base station scenarios, this constraint ensures that transient power failures of a single base station are not incorrectly included in local propagation modes, and co-channel interference from a cluster of local base stations is not incorrectly included in global propagation modes, thus avoiding confusion of risk types at the source.

[0109] The lossless reconstruction constraint loss is calculated to penalize information loss during the decomposition process: the Frobenius norm squared is obtained by calculating the difference between the element-wise sum of the three risk mode tensors and the standardized multivariate time series data of the communication base station in the nth period. The calculation formula is: ; This constraint ensures that the linear combination of the three modes can completely reconstruct the original data, guaranteeing that no important risk information is lost during the decomposition process. In the context of communication base stations, this means that whether it is a momentary anomaly of a single base station, a propagation risk in a local area, or a global risk across the entire network, it can be completely assigned to the corresponding mode without any risk signal being missed.

[0110] Perform a loss fusion operation, multiplying the orthogonality constraint loss and the lossless reconstruction constraint loss by their respective balance coefficients, and adding them to the model's total loss function. That is, the total loss function includes... The values ​​of the two balance coefficients have been specifically optimized for the communication base station scenario: the orthogonality constraint is a soft constraint with a low weight to avoid excessive constraints that could lead to decomposition distortion; the lossless reconstruction constraint is a hard constraint with a high weight to ensure information integrity.

[0111] Parameter optimization is performed using the backpropagation algorithm: the gradients of the total loss function over all decomposed relevant learnable parameters are used to update the values ​​of these parameters, including the weights of the temporal abrupt component corresponding to each feature. Spatial isolated component weights And the spatial smoothness coefficient of the local propagation mode. .

[0112] During training, these parameters are automatically optimized based on the characteristics of the communication base station data: for structural safety parameters such as tower tilt angle and bolt stress value, the spatial isolated component weights are adjusted. It will automatically increase to 0.7-0.9 because such anomalies usually occur independently of a single base station; for network performance parameters such as PRB utilization and downlink throughput, the weight of the time-varying component... It will automatically rise to 0.6-0.8 because anomalies in these parameters usually manifest as sudden traffic fluctuations; for parameters such as network interference and signal coverage, the spatial smoothing coefficient... It will automatically rise to 0.7-0.9 because this type of anomaly has strong local spatial propagation characteristics.

[0113] After the above complete calculations, the orthogonality constraint loss value is finally obtained. and lossless reconstruction constraint value These two values, along with the prediction loss and adversarial loss, constitute the model's total loss, guiding the parameter updates of the entire network. Secondly, there are the updated decomposition-related learnable parameters, which are continuously optimized in subsequent training iterations until the three modalities achieve optimal orthogonality and reconstruction results. Once the model training is complete, these parameters are fixed, and the fixed parameters are directly used for risk mode decomposition during the inference phase, eliminating the need for constraint calculations in this step.

[0114] In this embodiment, the dynamic risk transmission graph generation module may include: a node pair modal mutual information calculation module, a comprehensive risk correlation degree weighted fusion module, a correlation degree matrix row-level sparsification module, and an adjacency matrix construction module; these will be described in detail below.

[0115] The node-to-modal mutual information calculation module is used to generate an instantaneous modal mutual information matrix, a local modal mutual information matrix, and a global modal mutual information matrix based on the instantaneous risk modal tensor, the local propagation modal tensor, and the global diffusion modal tensor of the current period.

[0116] In this embodiment, generating the instantaneous modal mutual information matrix, the local modal mutual information matrix, and the global modal mutual information matrix based on the instantaneous risk mode tensor, the local propagation mode tensor, and the global diffusion mode tensor of the current period includes: Perform modal feature sequence aggregation for each base station and each mode The arithmetic mean of 35 feature dimensions is used to obtain the one-dimensional time series of the base station under the corresponding mode. The calculation formula is: ; in This serves as the index for time steps within the history window.

[0117] This aggregation operation compresses the multi-feature risk status of each base station into a single variable time series, which can comprehensively reflect the overall risk level of the base station under that mode and avoid interference from random fluctuations of individual features in mutual information calculation. In the context of communication base stations, this means that if a base station simultaneously experiences an anomaly of reduced RRU transmit power and reduced downlink throughput, the aggregated time series will show a unified trend of increasing risk, more accurately reflecting the risk propagation relationship between base stations.

[0118] Perform global equal-frequency binning discretization, collect time series values ​​of all modes from all base stations, and form a bin of length [length missing]. The global value set; sort the set by value in ascending order and divide it into... One equal frequency interval (number of equal frequency sub-blocks) The number of intervals for discretizing the continuous time series, a hyperparameter, is recommended to be 15 for communication base station scenarios, satisfying the following conditions. (using empirical formulas to avoid overfitting), each interval contains the exact same number of samples; record the boundary values ​​of each interval. ,in , Using global equal-frequency binning instead of individual binning for each node ensures that the time series of all base stations use a unified discretization standard, guaranteeing the comparability of mutual information calculations between different base stations; at the same time, equal-frequency binning is not sensitive to outliers and can effectively handle occasional extreme outliers in communication base station data.

[0119] Perform time series discretization mapping: for each modal time series of each base station, map each continuous value to the corresponding discrete interval number to obtain the discretized time series. If Discretization, which transforms continuous risk values ​​into discrete risk levels, is a necessary prerequisite for calculating mutual information.

[0120] Calculate the marginal probability distribution for any two base stations. and And arbitrary mode m, statistical discretized time series In each interval Number of times Calculate the marginal probability distribution: ; Similarly, calculate the base station Marginal probability distribution ,in The edge probability distribution reflects the probability that a single base station is at different risk levels in this mode.

[0121] Next, the joint probability distribution is calculated, and the discretized time series of the two base stations are statistically analyzed in interval pairs. The number of times the above appears together That is, simultaneously satisfying and Time step Calculate the joint probability distribution based on the number of elements: ; The joint probability distribution reflects the probability that two base stations are simultaneously in different risk level combinations under this mode.

[0122] Then, modal mutual information is calculated, and node pairs are computed according to the definitions of entropy and joint entropy. Mutual information in mode m:

[0123] Among them when When = 0, the summation term is set to 0 to avoid the logarithm becoming meaningless.

[0124] A higher mutual information value indicates a stronger correlation between the time series of two base stations in that mode, and a higher probability of a risk propagation relationship. In the context of communication base stations, this characteristic manifests as follows: if co-channel interference from base station A propagates to the adjacent base station B, their local propagation mode time series will exhibit a high mutual information value; while if the two base stations are far apart and have no direct signal propagation relationship, their mutual information value will be close to 0.

[0125] Finally, iterate through all base station pairs and repeat the above steps of edge probability distribution calculation, joint probability distribution calculation, and mutual information calculation to obtain the mutual information matrices corresponding to the three modes: Instantaneous modal mutual information matrix Local modal mutual information matrix and global modal mutual information matrix , where matrix elements Indicates base station and base stations The mutual information value under mode m. These three matrices reflect the correlation of risk propagation among base stations at three different scales: the instantaneous modal mutual information matrix reflects the accidental correlation between sudden anomalies of individual base stations, the local modal mutual information matrix reflects the strong correlation of gradual risk propagation within a local area, and the global modal mutual information matrix reflects the strong correlation of synchronous risk evolution across the entire network.

[0126] The comprehensive risk correlation weighted fusion module is used to generate the original comprehensive risk correlation matrix for the current period based on the instantaneous modal mutual information matrix, the local modal mutual information matrix, and the global modal mutual information matrix.

[0127] In this embodiment, generating the original comprehensive risk correlation matrix for the current period based on the instantaneous modal mutual information matrix, the local modal mutual information matrix, and the global modal mutual information matrix includes: Row normalization of the mutual information matrix is ​​performed on the mutual information matrix of each modality, mapping the element values ​​of each row to the [0,1] interval to eliminate the differences in the numerical range of mutual information between different modalities. Due to the different risk characteristics of the three modalities, the absolute value range of their mutual information differs significantly: global modal mutual information usually reflects the strong correlation of network-wide synchronization and is generally high; instantaneous modal mutual information usually reflects the independent anomalies of a single base station and is generally low. If weighted fusion is performed directly without normalization, the modality with the larger numerical range will dominate the overall correlation, causing the contributions of other modalities to be masked. The formula for calculating row normalization is: ; Where the denominator is added This is to avoid division by zero when all elements in a row are 0. After normalization, the sum of the row vector elements corresponding to each base station is 1, and each element represents the relative correlation strength between that base station and the corresponding base station in that mode.

[0128] Subsequently, a weighted fusion calculation is performed to determine the overall correlation degree for each base station. The weighted sum of the normalized mutual information values ​​of the three modalities yields the original comprehensive risk correlation degree: ; During training in a communication base station scenario, the modality fusion weights will automatically converge to values ​​that conform to the actual risk distribution: typically It will converge to 0.6-0.8 because most of the risks to communication base stations (such as same-channel interference, regional power outages, and local weather effects) are local propagation risks. It will converge to 0.15-0.3, corresponding to global risks such as core network failure and large-scale typhoons; It will converge to 0.05-0.1 because transient anomalies usually occur independently of a single base station and rarely propagate between base stations. This adaptive weight adjustment ensures that the overall correlation accurately reflects the actual propagation intensity of different types of risks.

[0129] Next, the diagonal elements are set to zero, meaning all diagonal elements of the comprehensive correlation matrix are set to zero. This is because risk propagation involves the interaction between different base stations. There is no risk propagation relationship between base stations themselves. In mutual information calculation, the mutual information value between base stations is the maximum value. If it is not set to zero, it will lead to the highest self-loop weight of each base station in the subsequent graph structure, completely masking the true relationship with other base stations.

[0130] Then, the nonnegativity constraint is applied, setting all elements less than 0 to zero, i.e.: like but Theoretically, the mutual information value is always greater than or equal to 0. However, due to numerical errors or data noise in the calculation process, extremely small negative numbers may appear. The non-negativity constraint can ensure that the correlation value conforms to the physical meaning, that is, the risk propagation intensity cannot be negative.

[0131] After the above calculations, the original comprehensive risk correlation matrix for the nth period is finally obtained. Each element Indicates base station to base station The overall risk propagation intensity; the higher the value, the greater the risk from the base station. propagated to base station The higher the probability, the faster the speed. In the context of communication base stations, this matrix can accurately reflect the real risk propagation relationship: for example, there will be a high degree of correlation between adjacent base stations with direct signal coverage overlap, while there will be a medium degree of correlation between base stations belonging to the same core network node but geographically far apart, and the correlation between base stations without any physical or network connection is close to 0.

[0132] A row-level sparsification module for the correlation matrix is ​​used to generate a sparse comprehensive risk correlation matrix for the current period based on the original comprehensive risk correlation matrix for the current period and the topology of IoT devices.

[0133] In this embodiment, generating the sparse comprehensive risk correlation matrix for the current period based on the original comprehensive risk correlation matrix for the current period and the topology of IoT devices includes: First, iterate through each base station node. Extract its corresponding row vector in the original comprehensive risk correlation matrix. Each element of this row vector represents a base station. The risk propagation intensity to all other base stations. In communication base station scenarios, this row vector typically exhibits obvious local clustering characteristics: with base stations The element values ​​of adjacent base stations with a geographical distance of less than 2 kilometers are significantly higher, while the element values ​​of base stations with a distance of more than 5 kilometers are generally close to 0.

[0134] Then, all elements of the row vector are sorted in descending order, and the base station index corresponding to each element is recorded, resulting in a sorted list of element values ​​and an index list. After sorting, the first K elements correspond to the base stations. The K neighbors with the highest probability of risk propagation are typically those associated with the base station. There are adjacent base stations that have direct fiber optic connections, belong to the same base station cluster, or have signal coverage overlap of more than 30%; while the element values ​​after rank K are usually less than 0.01, and there is almost no real risk propagation relationship between the corresponding base stations, so they can be safely deleted.

[0135] Next, a sparse comprehensive risk correlation matrix is ​​constructed, and a zero matrix with the same dimension as the original matrix is ​​initialized. For each base station Only the element values ​​corresponding to the top K highest correlation neighbors after sorting are copied from the original matrix to the sparse matrix, that is, if the base station It is a base station If a given element is one of its first K neighbors, then: ,otherwise ; This operation ensures that only the edges with the greatest risk of propagation are retained in the sparse matrix, while fully preserving the original correlation values ​​of these edges.

[0136] Perform optional undirected graph symmetry verification. This step is suitable for situations where risk propagation in communication base station scenarios is an undirected process. For example, risks such as co-channel interference and signal coverage impact are usually bidirectional: the signal from base station A interferes with base station B, and the signal from base station B also interferes with base station A. If symmetry verification is enabled, the sparse correlation matrix is ​​converted into a symmetric matrix, calculated using the following formula: ; This operation eliminates one-way spurious associations caused by data noise, making the graph structure more consistent with the physical characteristics of risk propagation in communication base stations. For one-way propagation risks such as power failures and transmission link interruptions, this step is unnecessary; the directed graph structure can be retained.

[0137] After the above complete calculations, the sparse comprehensive risk correlation matrix for the nth period is finally obtained. This matrix has two core characteristics: First, each row has only K non-zero elements, reducing the computational complexity from O(N²) to O(NK). When N=500 and K=8, the computational cost is only 1.6% of the original matrix, enabling real-time computation on ordinary edge computing devices. Second, it retains only the truly existing risk propagation edges, filtering out over 98% of noise correlations. In communication base station scenarios, this matrix clearly presents the structural characteristics of base station clusters: base stations within the same cluster are tightly connected, while connections between different base station clusters are sparse, highly consistent with actual network topology and risk propagation patterns.

[0138] An adjacency matrix construction module is used to generate a dynamic adjacency matrix for the current period based on the sparse comprehensive risk correlation matrix of the current period and the dynamic adjacency matrix of the previous period.

[0139] In this embodiment, generating the dynamic adjacency matrix for the current period based on the sparse comprehensive risk correlation matrix of the current period and the dynamic adjacency matrix of the previous period includes: First, perform the alignment operation for non-zero elements, creating two temporary matrices with the same dimensions as the original matrix. and The initial values ​​are all zero. Iterate through all base station pairs. ,like >0 or Then the element value at the corresponding position is copied to the two temporary matrices respectively. and This step ensures that both matrices have non-zero elements in exactly the same positions, avoiding smoothing calculation errors caused by inconsistent non-zero positions. In the context of communication base stations, this means that if a risk propagation edge existed in the previous cycle but temporarily disappeared due to noise in the current cycle, or if it newly appeared in the current cycle but did not exist in the previous cycle, it will be retained in the temporary matrix for smoothing calculation and will not be directly discarded.

[0140] Then, an element-wise exponential moving average calculation is performed, taking the positions of all non-zero elements in the two temporary matrices according to the formula: Calculate the smoothed temporary correlation degree.

[0141] smoothness coefficient (The hyperparameter controlling the update speed of the historical graph structure, with a value range of [0.7, 0.9], and a recommended value of 0.8 for communication base station scenarios, is specifically optimized to filter random noise within a 5-minute data collection period while promptly capturing real risk propagation changes within 1-2 hours.) This determines the weight of the historical graph structure: when When the value is 0.8, the graph structure of the current period contributes 20% of the weight, and the historical graph structure contributes 80% of the weight.

[0142] This weighting method gives the graph structure changes inertia. The actual risk propagation relationships will gradually be established or disappear within 2-3 cycles, while the noise fluctuations of a single cycle will be averaged out. For example, when a cluster of base stations begins to experience co-channel interference, the correlation between base stations will gradually increase from 0 to 0.6 within 3 cycles, rather than suddenly jumping to 0.6 due to data noise in a single cycle; when the interference is eliminated, the correlation will also gradually decrease to 0 within 3 cycles, rather than suddenly disappearing.

[0143] Next, sparsity preservation operations are performed on the temporary matrix. All values ​​less than the sparsity threshold (Threshold used to filter weakly related edges, hyperparameter, recommended value is) The elements of ) are set to zero.

[0144] This is because the exponential moving average calculation turns elements that were originally zero into extremely small non-zero values. These values ​​are spurious weak correlations generated by data noise and have no real significance for risk propagation. Without filtering, the sparsity of the matrix will gradually be lost with periodic iterations, the computational load will increase exponentially, and a large amount of noise will be introduced that interferes with subsequent graph convolution calculations. In communication base station scenarios, values ​​smaller than... The probability of risk transmission corresponding to the correlation is less than 0.01%, which can be safely filtered out.

[0145] Perform the final row normalization operation on the temporary matrix. A second row normalization process is performed to ensure that the sum of elements in each row is 1, satisfying the numerical stability requirements of subsequent graph convolution operations. The normalization formula is: After normalization, the sum of the weights of all outgoing edges for each base station is 1, and each weight value represents the probability that the risk propagates from that base station to its corresponding neighboring base station.

[0146] After the above complete calculations, the final dynamic adjacency matrix for the nth period is finally obtained. This matrix accurately reflects the propagation path and intensity of risks among communication base stations within the current time window, while also exhibiting good temporal continuity and stability. In a communication base station scenario, this matrix can accurately present the dynamic evolution of risk propagation: when thunderstorms occur in a local area, the connectivity between base stations within that area gradually strengthens; when a core network failure occurs, the connectivity between all base stations and core network nodes simultaneously strengthens; after the failure is repaired, these connections gradually weaken and return to normal.

[0147] In this embodiment, the modal high-dimensional feature generation module may include: an instantaneous risk modal temporal feature extraction module, a local propagation modal spatiotemporal feature extraction module, and a global diffusion modal spatiotemporal feature extraction module; these will be described in detail below.

[0148] The instantaneous risk mode temporal feature extraction module is used to extract the instantaneous risk mode feature tensor of the current period based on the instantaneous risk mode tensor of the current period.

[0149] In this embodiment, extracting the instantaneous risk mode feature tensor of the current period based on the instantaneous risk mode tensor of the current period includes: Perform an input dimension transformation to convert the instantaneous risk mode tensor of the nth period from a device-priority dimension to a time-priority dimension, resulting in... ( After the dimensional transformation, the first dimension of the tensor corresponds to the time step, the second dimension corresponds to the base station node, and the third dimension corresponds to the feature dimension. This arrangement makes it easier for GRU to process the feature data of all base stations step by step, realize batch parallel computing, and greatly improve the inference speed.

[0150] In communication base station scenarios, this batch processing method can simultaneously calculate the instantaneous risk characteristics of hundreds of base stations, meeting the real-time monitoring needs of large-scale networks. Subsequently, the initial hidden state of the GRU is set, and the final hidden state of the instantaneous modal GRU in the (n-1)th cycle is... ( This serves as the initial hidden state for GRU this week. The key to this step is the cross-cycle hidden state propagation, which enables the continuation of risk state information from the previous cycle to the current cycle, achieving continuous modeling of instantaneous risks lasting longer than one hour. For example, when the remaining capacity of a base station's battery pack suddenly drops to 10% in the last time step of cycle n-1, this abnormal state is saved in the hidden state and propagated to cycle n, allowing the model to continuously track the evolution of the fault without losing information due to cycle switching. Next, a time-step GRU calculation is performed, for each time step... The hidden state at the current time step is calculated using the same GRU parameters, and the calculation formula is as follows: ; ; ; ; in To update the door, control how much historical hidden state information is retained; To reset the door, control how much historical hidden state information is forgotten; The sigmoid activation function maps the output to the [0,1] interval; For element-wise multiplication; , , The weight matrix is ​​a learnable matrix with dimensions respectively. ,in, is the dimension of the instantaneous modality hidden layer (number of neurons in the GRU hidden layer, a hyperparameter with a value range of 32 to 128). In communication base station scenarios, GRU's gating mechanism can automatically distinguish between instantaneous fluctuations and persistent faults: for instantaneous spikes in base station CPU utilization, the reset gate will quickly forget this historical information; while for a continuous increase in the temperature of the BBU core board, the update gate will retain this risky state for a long time until the fault is repaired.

[0151] Perform dropout regularization on all hidden states output by the GRU. Apply dropout operation, with a random dropout ratio of 1:1. (Dropout probability, a hyperparameter ranging from 0.1 to 0.3, used to prevent overfitting in neurons.) This operation effectively prevents model overfitting and avoids the model memorizing individual noisy samples from the training set. In communication base station scenarios, since base stations are in normal operation most of the time, high-risk samples in the training set are relatively few, and dropout regularization can significantly improve the model's generalization ability to unknown faults.

[0152] Perform a dimension restoration operation to transform the time-priority feature tensor output by the GRU back to the device-priority dimension, obtaining the final instantaneous modal feature tensor. After dimensionality restoration, the first dimension of the tensor corresponds to the base station node, the second dimension corresponds to the time step, and the third dimension corresponds to the hidden layer features.

[0153] After the above calculations, the instantaneous risk mode feature tensor of the nth period is finally obtained. ,in Indicates base station At time step The 64-dimensional instantaneous risk feature vector. This feature vector comprehensively reflects the base station's... The instantaneous risk evolution trend over the past hour can accurately capture the temporal characteristics of sudden risks such as power fluctuations, board failures, and sensor anomalies in individual base stations. For example, when the RRU transmit power of a base station suddenly drops in the third time step, the corresponding feature vector will show a significant abrupt change in the third time step and remain at a high value in subsequent time steps until the power returns to normal. This tensor will be directly input into the dimension unification sublayer of the subsequent adaptive feature fusion layer, and fused with the local propagation mode features and the global diffusion mode features. Simultaneously, the final hidden state of the GRU in this cycle... It will be saved to the system state buffer as the initial hidden state of the GRU in the (n+1)th cycle.

[0154] A local propagation mode spatiotemporal feature extraction module is used to extract the local propagation mode feature tensor of the current period based on the local propagation mode tensor of the current period.

[0155] In this embodiment, extracting the local propagation mode feature tensor of the current period based on the local propagation mode tensor of the current period includes: Pre-compute the Chebyshev polynomial basis, which is the foundation for ChebConv to extract local space features. This is based on the final dynamic adjacency matrix of the nth period. Calculate the Laplacian matrix of the symmetric normalized graph ,in for The diagonal matrix, diagonal elements Equal to base station The number of neighbors in the dynamic adjacency matrix; Calculate the scaled Laplacian matrix ,in for The scaling operation maps the eigenvalues ​​of the Laplacian matrix to the interval [-1,1], ensuring the numerical stability of Chebyshev polynomial calculations. Before calculation Chebyshev polynomial basis , ,in Higher-order polynomials are obtained through recursive formulas Calculated.

[0156] In communication base station scenarios, the polynomial basis is recalculated each cycle based on the latest dynamic adjacency matrix to ensure real-time capture of dynamic changes in risk propagation paths. Subsequently, time-step Chebyshev graph convolution is performed for each time step. Local spatial features are extracted using pre-computed Chebyshev polynomial bases. The calculation formula is as follows:

[0157] in The weight matrix is ​​a learnable matrix. This is a spatial feature tensor. This operation weights and fuses the features of each base station with the features of its neighbors within a 3-hop range, enabling accurate capture of risk propagation patterns within local base station clusters. In communication base station scenarios, this step can effectively extract the spatial diffusion features of co-channel interference: when the interference signal of a base station increases, the features of its neighboring base stations will change synchronously, and ChebConv will encode this spatial correlation into a high-dimensional feature vector.

[0158] Next, TCN (Temporal Convolutional Network) is performed for temporal feature extraction. First, the spatial feature tensor is converted into a device-priority dimension. Where N base stations are used as the batch dimension and 12 time steps are used as the time dimension. As a channel dimension, this arrangement conforms to the input format requirements of TCN, enabling parallel computation across all base stations. The transformed tensor is then input into... In a causal TCN composed of residual blocks, each residual block contains two dilated causal convolutional layers, with the dilation coefficients arranged according to... Exponential growth ( (For layer indexing, starting from 0), therefore the maximum temporal receptive field of the 3rd layer TCN is This allows for coverage of the entire 12-timestep historical window. Each dilated causal convolutional layer is followed by a ReLU activation function and a dropout operation, with a random dropout ratio of [missing value]. The system consists of neurons; residual connections are added within each residual block to directly add the input to the output, mitigating the vanishing gradient problem; finally, layer normalization is applied to the output of each residual block to improve training stability. The TCN calculation formula is: ; in This indicates a causal temporal convolutional network operation where the output dimension remains consistent with the input dimension. In communication base station scenarios, TCN can accurately capture the temporal evolution of local risks. For example, the decrease in base station receiver sensitivity caused by thunderstorms in a local area will exhibit a temporal pattern of gradually increasing, then remaining stable, and then gradually decreasing. This pattern will be effectively encoded into high-dimensional features by TCN.

[0159] After the above calculations, the feature tensor of the local propagation mode in the nth period is finally obtained. ,in Indicates base station At time step The 128-dimensional local propagation risk feature vector is derived. This feature vector integrates spatial and temporal information, comprehensively reflecting the base station's... By analyzing the local risk evolution trends of the base station cluster and its neighbors over the past hour, the spatiotemporal characteristics of typical local propagation risks such as co-channel interference, regional power outages, and localized weather impacts can be accurately captured. For example, when co-channel interference occurs within a base station cluster, the local propagation feature vectors of all base stations within the cluster will increase synchronously within 2-3 time steps after the interference occurs, and remain at a high value during the duration of the interference. This tensor will be directly input into the dimension unification sublayer of the subsequent adaptive feature fusion layer, where it will be fused with the instantaneous risk modal features and the global diffusion modal features.

[0160] A global diffusion mode spatiotemporal feature extraction module is used to generate a global diffusion mode feature tensor for the current period based on the global diffusion mode tensor for the current period.

[0161] In this embodiment, generating the global diffusion mode feature tensor for the current period based on the global diffusion mode tensor for the current period includes: Perform time-step multi-head graph attention calculation, for each time step A multi-head GAT (Graph Attention Network) is used to extract long-range spatial features. For each attention head... First, the original 35-dimensional features are mapped to 32-dimensional intermediate features using a learnable projection matrix. Then, the base station is calculated. For base stations Attention coefficient: ; ; in The projection matrix is ​​learnable. For learnable attention vectors, For base stations The set of neighbors (including itself) in the final dynamic adjacency matrix of the nth period.

[0162] Attention coefficient This indicates that under the h-th attention head, the base station Status of base station The degree of impact. In communication base station scenarios, GAT automatically assigns higher attention weights to key nodes such as core network nodes and aggregation base stations, because changes in the state of these nodes will affect all base stations in the entire area; while the attention weights of ordinary edge base stations are relatively low.

[0163] The outputs of all four attention heads are then concatenated to obtain the spatial features of the current time step: ; in This indicates a splicing operation. This is a spatial feature tensor. The concatenation operation can fuse attention information from multiple different perspectives, improving the expressive power of spatial features. Next, GAT dropout regularization is performed on all attention coefficients. Apply dropout operation, with a random dropout ratio of 1:1. Attention-based connections. This operation prevents the model from over-relying on connections from a few base stations, improving the model's robustness to changes in network topology. In communication base station scenarios, this effectively avoids distortion in global feature extraction caused by temporary outages of individual base stations.

[0164] To perform long-term temporal feature extraction using LSTM (Long Short-Term Memory) networks, the spatial feature tensor is first converted into a time-first dimension. The algorithm uses 12 time steps as the sequence dimension, N base stations as the batch dimension, and 256-dimensional features as the channel dimension. This arrangement enables parallel computation of the LSTM across all base stations, significantly improving inference speed. The final hidden state and cell state of the global modality LSTM in the (n-1)th cycle are used as the initial states of the LSTM in the current cycle. The hidden state and cell state of the current cycle are calculated step-by-step using the following formula: ; ; ; ; ; ; in This is an input gate that controls the input of new information; The forget gate controls the forgetting of old information; It is an output gate that controls the output of cell state; For element-wise multiplication; , , , The weight matrix is ​​learnable. Cross-cycle state propagation enables continuous modeling of global risks lasting longer than one hour. For example, base station failures caused by typhoons may last for several days; LSTM can continuously track the evolution of this risk through state propagation. Subsequently, LSTM dropout regularization is performed, applying dropout to the hidden states of each LSTM layer, with a random dropout ratio of [percentage missing]. The neurons are used to prevent the model from overfitting. Finally, a dimension restoration operation is performed to convert the time-priority feature tensor output by the LSTM back to the device-priority dimension, resulting in the final global diffusion modality feature tensor. .

[0165] Simultaneously, the average of the attention weights across all time steps is calculated to obtain the GAT attention weight matrix for the nth period. This matrix reflects the global impact between different base stations and can be used for subsequent risk tracing analysis. After the above complete calculation, two outputs are finally obtained: First, the global diffusion mode characteristic tensor of the nth period. ,in Indicates base station At time step The 256-dimensional global diffusion risk feature vector integrates long-distance spatial information and long-term temporal information, which can accurately capture the spatiotemporal characteristics of typical global diffusion risks such as core network failures, large-scale power outages, and typhoon disasters. Second, the GAT attention weight matrix in the nth period. This matrix clearly presents the propagation path of global risk. For example, when a core network failure occurs, the rows and columns corresponding to core network nodes will exhibit significantly high attention values. The global diffusion mode feature tensor will be directly input into the dimension unification sublayer of the subsequent adaptive feature fusion layer, and fused with the instantaneous risk mode features and local propagation mode features; the GAT attention weight matrix will be saved for subsequent interpretable risk attribution analysis. Simultaneously, the final hidden state of the LSTM in this cycle... and cell state It will be saved to the system state buffer as the initial state of the LSTM in the (n+1)th cycle.

[0166] In this embodiment, the prediction module may include: a multimodal feature dimension unification module, a feature fusion module, a fully connected layer calculation module, and a post-processing and normalization module; these will be described in detail below.

[0167] A multimodal feature dimension unification module is used to unify the dimensions of the instantaneous risk modal feature tensor, the local propagation modal feature tensor, and the global diffusion modal feature tensor of the current period, thereby obtaining a unified instantaneous risk modal feature tensor, a unified local propagation modal feature tensor, and a unified global diffusion modal feature tensor.

[0168] In this embodiment, the instantaneous risk mode feature tensor, the local propagation mode feature tensor, and the global diffusion mode feature tensor of the current period are unified in dimension to obtain a unified instantaneous risk mode feature tensor, a unified local propagation mode feature tensor, and a unified global diffusion mode feature tensor, including: Define three learnable linear transformation weight matrices and bias vectors: Instantaneous mode transformation weights bias ; Local mode transformation weights bias ; Global mode transformation weights bias ; A linear transformation is performed on the final time step features of each mode, and the calculation formula is as follows: ; ; ; Applying the ReLU activation function to the transformed features introduces nonlinear expressive power: ; ; ; Finally, the instantaneous modal features with unified dimensions are obtained. Local modal features after unification of dimensions Global modal features after unifying dimensions ;in, This is the unified dimension after multimodal feature fusion, a hyperparameter with a value range of 128 to 512. It is usually taken as the maximum or average value of the three modal hidden layer dimensions.

[0169] The feature fusion module is used to fuse the instantaneous risk modal feature tensor, the local propagation modal feature tensor, and the global diffusion modal feature tensor of the same dimension in a unified dimension, so as to obtain the fused comprehensive spatiotemporal feature tensor and the modal weight vector of each device.

[0170] In this embodiment, the instantaneous risk mode feature tensor, the local propagation mode feature tensor, and the global diffusion mode feature tensor of a unified dimension are fused to obtain the fused comprehensive spatiotemporal feature tensor and the mode weight vector of each device, including: Define the learnable parameters of a modal attention network: First layer weight bias ;in, The intermediate hidden layer dimension is calculated for modal attention; the hyperparameter has a value range of 32 to 128. Second layer weights bias ; Calculate the attention score for each modality's features separately:

[0171] in Indicates device Attention score in modality m.

[0172] The attention scores for the three modalities are normalized using the Softmax function to obtain the weights for each modality: ; ; ; satisfy + + =1.

[0173] Based on the weights, the features of the three modalities are fused element-wise to obtain the comprehensive spatiotemporal features: Finally, the fused spatiotemporal feature tensor is obtained. And the modal weight vector for each device, used for subsequent interpretability risk attribution analysis.

[0174] A fully connected layer computing module is used to generate the original risk prediction value tensor from the fused integrated spatiotemporal feature tensor.

[0175] In this embodiment, generating the original risk prediction value tensor from the fused integrated spatiotemporal feature tensor includes: Define the learnable parameters of a fully connected network for risk prediction: First layer weight bias ; Second layer weights bias ;in, The intermediate hidden layer dimension of the fully connected layer for risk prediction is a hyperparameter with a value range of 64 to 256. First-layer fully connected computation and activation: ; ; Dropout regularization: Apply dropout operation to the output of the first fully connected layer, randomly discarding a certain percentage of values. (Dropout probability of fully connected layers, hyperparameter, with a value range of 0.1 to 0.3) neurons to prevent overfitting.

[0176] The second layer of fully connected computation outputs the future. The initial risk forecast for the step:

[0177] in Indicates device In the future time steps (i.e.) The original risk prediction value at time ( ) is used to obtain the original risk prediction value tensor. .

[0178] The post-processing and standardization module is used to generate the final risk prediction value tensor for the current period based on the original risk prediction value tensor.

[0179] In this embodiment, generating the final risk prediction value tensor for the current period based on the original risk prediction value tensor includes: Outlier truncation: Remove outliers from the original risk predictions that are less than [a certain value]. The element is truncated as , will be greater than The element is truncated as .

[0180] In this embodiment, To minimize the risk value of the training set, The maximum risk value in the training set is the minimum and maximum value of the true risk label of all training samples obtained during the offline training phase, which are fixed values. The outlier truncation threshold is a hyperparameter with a value range of 0.01 to 0.05. Min-Max Standardization: This method linearly maps the truncated risk prediction value to the [0,1] interval. The calculation formula is as follows: ; Boundary constraints: After standardization, elements less than 0 are set to 0, and elements greater than 1 are set to 1, ensuring that the final risk value is strictly within the interval [0,1]. This yields the final risk prediction tensor for the nth period. ,in Indicates device In the future Standardized risk levels for each time step.

[0181] In this embodiment, the IoT monitoring scenario risk prediction system based on spatiotemporal graph neural networks may further include: The risk tracing and analysis module is used to obtain the risk tracing results for the current period based on the prediction results of the current period and the modal high-dimensional feature information of the current period.

[0182] In this embodiment, obtaining the risk tracing result for the current period based on the prediction result and the modal high-dimensional feature information of the current period includes: Calculate the contribution of each risk mode to the final prediction result (using modal contribution analysis) to determine the dominant propagation type of the current predicted risk; Specifically, extract the modal weights corresponding to alarm base station i. (Instantaneous modal weights) (Local Modal Weights) (Global Modality Weights) Compare the three weights; the mode with the highest weight is the dominant risk type for that base station. like Maximum: Determined as an instantaneous single-point risk, typical faults include base station power failure, board burnout, sensor damage, and other faults that occur independently of a single base station. like Maximum: Determined as a local propagation risk, typical faults include co-channel interference, regional power outages, and the impact of localized thunderstorms, which are faults that only propagate within the base station cluster; like Maximum: Determined as a global spread risk, typical faults include core network failures, large-scale power outages, typhoon disasters, and other faults that affect the entire monitoring area.

[0183] It also outputs the specific contribution percentages of the three modalities, quantifying the impact of different types of risks on the prediction results.

[0184] By combining the predicted risk value of each device with its out-degree in the attention weight matrix, the risk importance score of each node is calculated to identify key risk nodes.

[0185] In this embodiment, the application employs the graph attention backtracking method, which uses the output GAT attention weights to backtrack the upstream nodes of risk propagation and locate the original risk source.

[0186] For example, alarm base stations Starting from the base station, backtrack along the directed edges of the dynamic adjacency matrix to extract all points pointing to the base station. The GAT attention weights corresponding to the edges (Indicates base station) For base stations (The degree of global impact). All upstream base stations are sorted from largest to smallest according to their attention weight, and the top 3 base stations with the highest weights are the most likely candidates for risk sources.

[0187] In a communication base station scenario, if the risk source is a core network node, its attention weight to all alarm base stations will be significantly higher than that of other nodes; if the risk source is a certain edge base station, only its neighboring base stations will have a higher attention weight.

[0188] A directed risk propagation graph is constructed based on the current dynamic adjacency matrix and attention weights. The top-3 main propagation paths from key risk nodes to other high-risk nodes are identified through depth-first search.

[0189] In this embodiment, the application uses the SHAP (SHapley Additive exPlanations) value method to calculate the contribution of each input feature to the predicted risk value of the alarm base station, and to locate the key monitoring parameters that lead to increased risk.

[0190] For example, using a trained risk prediction model and DeepExplainer (a SHAP interpreter for deep neural networks), alarm base stations can be calculated. The SHAP value of the 35 input features relative to their highest predicted risk value. A positive SHAP value indicates that the feature increases the risk value, while a negative SHAP value indicates that the feature decreases the risk value. The larger the absolute value, the higher the contribution.

[0191] Sort all features by absolute SHAP value from largest to smallest, and output the top 5 key features along with their specific values ​​and normal ranges. For example, if a base station is predicted to be high-risk, and SHAP analysis shows that the RRU transmit power has the highest and positive SHAP value, while its current value is 38dBm (normal range 43-46dBm), then it can be clearly determined that the risk of this base station is mainly caused by the abnormal decrease in RRU transmit power.

[0192] In this embodiment, for ease of understanding, the specific structure of the IoT monitoring scenario risk prediction system based on spatiotemporal graph neural network of this application is described as follows: Input preprocessing layer (multivariate time series data acquisition module) Internal structure: Linear interpolation operator for missing values; 3σ outlier truncation operator; Global Z-score normalization operator; Time feature encoding operator (holidays + hours + days of the week); II. Risk Spatiotemporal Decomposition Layer (Learnable Layer, Orthogonal Decomposition Module) 2.1 Instantaneous Risk Mode Extraction Branch Two parallel parameterless operators plus a learnable weighted fusion layer; First-order time difference operator (no parameters); Spatial isolated point detection operator (no parameters); Feature-level weighted fusion layer (learnable, 35 sets of weight coefficients) , ); 2.2 Local Propagation Mode Extraction Branch Internal structure: a parameter-free time operator + a learnable spatial smoothing layer; A time moving average operator of length 3 (without parameters); Thulaplatz smoothing layer (learnable, 1 spatial smoothing coefficient) ); 2.3 Global Diffusion Mode Extraction Branch Internal structure: pure matrix operations, no neural network layers. Tensor dimension transformation operator; Singular Value Decomposition (SVD) operator; Principal component selection operator (cumulative energy 95% threshold); Matrix reconstruction and dimension recovery operators; 2.4 Orthogonality Constraints and Non-destructive Reconstruction Calibration Branches (Training Phase Only) Internal structure: pure loss calculation logic, no neural network layer; Tensor flattening operator; Frobenius norm computation operator; Loss-weighted fusion operator; III. Dynamic Risk Transmission Diagram Generation Layer (Parameterless Calculation Layer, Dynamic Risk Transmission Diagram Generation Module) 3.1 Node-to-Modal Mutual Information Calculation Module Internal structure: Pure statistical calculation logic, without neural network layers; Feature sequence arithmetic mean aggregation operator; Global equal-frequency bin discretization operator (15 bins); Marginal probability and joint probability statistical calculation; Mutual information computation operator; 3.2 Comprehensive Risk Correlation Weighted Fusion Module Internal structure: pure matrix operations, containing 3 learnable weight parameters; Matrix row normalization operator; Weighted summation operator (learnable weights) , , ); Diagonal zeroing and nonnegativity constraint operators; 3.3 Row-level sparsity module for the correlation matrix Internal structure: pure matrix operations, no neural network layers; Row vector descending sort operator; Operator for retaining the first K elements (K=8); Optional undirected graph symmetry operators; 3.4 Graph Structure Exponential Moving Average Smoothing Update Module Internal structure: pure matrix operations, no neural network layers; Non-zero element alignment operator; Element-wise exponential moving average operator (smoothing coefficient) =0.8); Weakly related edge filtering operator (threshold) ); The final row normalization operator; IV. Multimodal Spatiotemporal Convolutional Layer (Learnable Layer, Core Feature Extraction Layer, Modal High-Dimensional Feature Generation Module) 4.1 Instantaneous Modal GRU Branch Internal structure: 1 layer GRU + 1 layer dropout; Input dimension transformation operator; Single-layer GRU layer (input dimension 35, hidden layer dimension 64); Dropout layer (dropout probability 0.2); Output dimension recovery operator; 4.2 Local Modal ChebConv+TCN Branch Internal structure: 1 layer ChebConv + 3 layers TCN residual blocks; Chebyshev polynomial basis precomputation operator; Single-layer Chebyshev plot convolutional layer (input dimension 35, output dimension 128, polynomial order 3). 3 serial TCN residual blocks (each residual block has the same structure): Layer 1: Dilated Causal Convolutional Layer (kernel size 5, dilation coefficient) , (for layer indexes) ReLU activation function; Dropout layer (dropout probability 0.2); Layer 2: Dilated Causal Convolutional Layer (kernel size 5, dilation coefficient) ); ReLU activation function; Dropout layer (dropout probability 0.2); Residual connectivity; Layer normalization layer; 4.3 Global Modal GAT+LSTM Branch Internal structure: 1 layer multi-head GAT + 2 layers LSTM + 2 layers dropout Single-layer multi-head graph attention layer (4 attention heads, each head outputs 32 dimensions, total output dimensions 128); GAT attention coefficient dropout layer (dropout probability 0.2); Input dimension transformation operator; Two stacked LSTM layers (input dimension 128, hidden layer dimension 256). LSTM hidden state dropout layer (dropout probability 0.2); Output dimension recovery operator; Attention weight averaging operator; V. Feature Fusion Layer (Learnable Layer) 5.1 Multimodal Feature Dimension Unified Sublayer Internal structure: 3 parallel linear layers + 3 ReLU activation functions; Instantaneous modal linear layer (input 64, output 256+ReLU); Local modal linear layer (input 128, output 256+ReLU); Global modal linear layer (input 256, output 256) + ReLU; 5.2 Modal Attention Fusion Sublayer Internal structure: Shared 2-layer linear attention network + Softmax normalization; Attention layer 1: Linear layer (input 256, output 64) + ReLU; Second attention layer: Linear layer (input 64, output 1); Softmax normalization operator; Element-wise weighted fusion operator; VI. Risk Prediction Output Layer (Learnable Layer + Parameterless Post-processing Layer); 6.1 Two Fully Connected Prediction Sublayers Internal structure: 2 layers of serial fully connected layers + 1 layer of dropout; The first fully connected layer (input 256, output 128) + ReLU; Dropout layer (dropout probability 0.2); The second fully connected layer (input 128, output K, where K is the prediction step size, 3 is recommended); 6.2 Risk Value Post-processing Sub-layer Internal structure: pure computational logic, no neural network layer; Outlier truncation operator (threshold 0.02); Minimum-maximum normalized operator; Boundary constraint operator (restricted to the interval [0,1]); VII. Cross-cycle state transfer mechanism (auxiliary mechanism); Internal structure: system memory buffer, no neural network layer; Transient modal GRU final hidden state buffer; The final hidden state buffer of the global modality LSTM; Global modal LSTM final cell state buffer; Used to implement cross-period modeling of continuous time series data; 8. Explainability Risk Source Analysis Module (Optional Post-processing Module) 8.1 Risk Type Tracing Module Internal structure: Pure logical judgment, no neural network layer; Modal weight comparison operator; Risk type classification operator; 8.2 Risk Source Location Module Internal structure: Pure graph analysis logic, no neural network layer; Attention-weighted backtracking operator; Upstream node sorting operator; 8.3 Key Feature Tracing Module Internal structure: SHAP-based interpreter; DeepExplainer initialization operator; SHAP value calculation operator; Feature contribution ranking operator.

[0193] This application has the following advantages: Existing technologies typically input raw data directly into graph neural networks, leading to overlapping features of instantaneous single-point faults, locally propagated interference, and globally spreading disasters, making it impossible for the model to distinguish between different types of risks. This solution, through orthogonality constraints and lossless reconstruction calibration, precisely decouples the raw data into three physically distinct and completely orthogonal risk modes, fundamentally solving the feature confusion problem. Each mode employs a dedicated feature extraction operator, enabling the use of the most suitable algorithm to process risks at different scales, significantly improving the ability to identify differentiated faults such as co-channel interference, core network faults, and tower tilting.

[0194] Most existing technologies use the physical connection topology of base stations as a fixed graph structure, which cannot reflect the true risk propagation patterns (for example, the propagation path of signal interference is not consistent with the fiber optic connection path). This solution adaptively learns a dynamic graph structure by calculating the mutual information of three modes. This automatically discovers risk propagation relationships that exist in the physical topology but are not physically present, while filtering out spurious connections that exist in the physical topology but have no risk propagation significance. The dynamic graph is updated every 5 minutes, enabling real-time adaptation to changes such as base station decommissioning, additions, and network topology adjustments, without requiring model retraining.

[0195] The accuracy of the system prediction in this application has been greatly improved, especially the ability to identify long-tail risks has been significantly enhanced.

[0196] Addressing the challenges of limited high-risk samples and diverse fault types in communication base stations, this solution employs multimodal decoupling, attention fusion, and weighted loss functions to improve prediction accuracy for various typical faults by 15%-25% compared to general GNN models. Specifically, it achieves a 92% accuracy rate for predicting instantaneous power failures, 88% for predicting localized co-channel interference, and 85% for predicting large-scale base station failures caused by typhoons. Alarms can be issued 15-60 minutes in advance, providing maintenance personnel with ample time to respond.

[0197] Extreme inference speed was achieved through three key optimizations: First, the three modal branches are computed in full parallelism with no data dependency; second, row-level sparsification of the correlation matrix reduces the computational complexity of graph convolution from O(N²) to O(NK) (K=8), with a computational cost of only 1.6% of the original dense graph; third, lightweight operators such as GRU and TCN are used to replace complex models. The total number of parameters in the entire network is approximately 300,000. On a typical edge computing server (a single entry-level GPU), the single inference time for 500 base stations is less than 100 milliseconds, fully meeting the real-time monitoring requirement of once every 5 minutes.

[0198] This scheme designs a cross-cycle state transfer mechanism, which transfers the final state of the instantaneous modal GRU and the global modal LSTM from one cycle to the next, breaking through the time limit of a single sliding window (1 hour). It can continuously model the slow evolution risks that last for several days or even weeks, such as the slow decline of battery capacity, the gradual increase of tower tilt, and the continuous change of typhoon path, thus avoiding the information loss problem of the traditional sliding window method.

[0199] By employing techniques such as orthogonality constraints, dropout regularization, and graph-structured exponential moving average smoothing, overfitting of the model is effectively prevented. Once trained on base station data from one city, the model can be directly transferred to other cities with only minor adjustments to achieve the same prediction accuracy. Furthermore, the model can adaptively identify the risk characteristics of different types of base stations (macro base stations, micro base stations, and indoor distributed base stations) without requiring separate model training for each type.

[0200] This application also provides a risk prediction method for IoT monitoring scenarios based on spatiotemporal graph neural networks, such as... Figure 2 As shown, the IoT monitoring scenario risk prediction method based on spatiotemporal graph neural network includes the following steps S201 to S205: Step S201: Obtain standardized multivariate time series data for the current period; Step S202: Perform orthogonal decomposition on the standardized multivariate time series data of the current period to obtain the risk propagation mode data set of the current period, wherein the risk propagation mode data set includes instantaneous risk mode, local propagation mode and global diffusion mode; Step S203: Generate a dynamic adjacency matrix that truly reflects the risk propagation path for the current period based on the risk propagation modality data set; Step S204: Based on the dynamic adjacency matrix that can truly reflect the risk propagation path in the current period and the risk propagation modal data group in the current period, perform parallel extraction of multimodal spatiotemporal features to obtain the modal high-dimensional feature information of the current period. Step S205: Based on the modal high-dimensional feature information of the current period, perform risk prediction for the current period to obtain the prediction result for the current period.

[0201] This embodiment orthogonally decomposes multivariate time series data into three types of risk propagation modes with clear physical meaning and complete orthogonality: instantaneous risk mode, local propagation mode, and global diffusion mode. Based on modal mutual information, it adaptively learns a dynamic risk transmission graph. Then, it designs a dedicated spatiotemporal feature extraction operator for each type of mode to perform parallel feature extraction. Finally, it uses a modal attention mechanism to adaptively fuse multimodal features to complete risk prediction and improve prediction accuracy.

[0202] This application provides one possible implementation, which may also include obtaining the risk tracing result for the current period based on the prediction result of the current period and the modal high-dimensional feature information of the current period. For a specific embodiment using a communication base station monitoring scenario, please refer to the preceding description; further details are omitted here.

[0203] This embodiment achieves the following beneficial technical effects compared to existing technologies by employing key technologies such as multi-scale risk decoupling, dynamic graph learning, dedicated operator design, and cross-cycle state transfer: 1) Precise decoupling of multi-scale risks to solve feature confusion at the source: Existing technologies generally directly input raw data into graph neural networks, resulting in overlapping features of instantaneous single-point faults, localized propagation interference, and globally spreading disasters, making it impossible for the model to distinguish between different types of risks. This solution, through orthogonality constraints and lossless reconstruction calibration, precisely decouples the raw data into three physically meaningful and completely orthogonal risk modes, solving the feature confusion problem at the source. Each mode uses a dedicated feature extraction operator, enabling the use of the most suitable algorithm to process risks at different scales, significantly improving the ability to identify differentiated faults such as co-frequency interference, core network faults, and tower tilting.

[0204] 2) Adaptive Dynamic Graph Learning, Realistically Reflecting Risk Propagation Paths: Most existing technologies use the physical connection topology of base stations as a fixed graph structure, which cannot reflect the true risk propagation patterns (e.g., the propagation path of signal interference is not consistent with the fiber optic connection path). This solution adaptively learns a dynamic graph structure by calculating the mutual information of three modes. It can automatically discover risk propagation relationships that do not exist in the physical topology but do exist in reality, while filtering out false connections that exist in the physical topology but have no risk propagation significance. The dynamic graph is updated every 5 minutes, which can adapt in real time to changes such as base station decommissioning, addition, and network topology adjustments, without the need to retrain the model.

[0205] 3) Significantly improved prediction accuracy, especially enhanced ability to identify long-tail risks: Addressing the characteristics of limited high-risk samples and diverse fault types in communication base stations, this solution improves prediction accuracy for various typical faults by 15%-25% compared to general GNN models through multimodal decoupling, attention fusion, and weighted loss functions. Specifically, it achieves 92% accuracy in predicting instantaneous power failures, 88% accuracy in predicting localized co-channel interference, and 85% accuracy in predicting large-area base station failures caused by typhoons, issuing alarms 15-60 minutes in advance, providing maintenance personnel with ample time to handle the situation.

[0206] 4) Extreme inference speed to meet real-time monitoring requirements: Extreme inference speed is achieved through three key optimizations: First, the three modal branches are computed in full parallelism with no data dependency; second, row-level sparsification of the correlation matrix reduces the computational complexity of graph convolution from O(N²) to O(NK) (K=8), with a computational cost of only 1.6% of the original dense graph; third, lightweight operators such as GRU and TCN are used to replace complex models. The total number of parameters in the entire network is approximately 300,000. On a typical edge computing server (a single entry-level GPU), the single inference time for 500 base stations is less than 100 milliseconds, fully meeting the real-time monitoring requirement of once every 5 minutes.

[0207] 5) Cross-cycle state transfer, breaking through the time limit of sliding window: This scheme designs a cross-cycle state transfer mechanism to transfer the final state of instantaneous modal GRU and global modal LSTM from one cycle to the next, breaking through the time limit of a single sliding window (1 hour). It can continuously model the slow evolution risks that last for several days or even weeks, such as the slow decline of battery capacity, the gradual increase of tower tilt, and the continuous change of typhoon path, avoiding the information loss problem of traditional sliding window methods.

[0208] 6) Strong generalization ability and adaptability to multiple device types: Through techniques such as orthogonality constraints, dropout regularization, and graph structure exponential moving average smoothing, overfitting is effectively prevented. After training the model on base station data in one city, it can be directly transferred to other cities with only minor adjustments to achieve the same prediction accuracy. Furthermore, the model can adaptively identify the risk characteristics of different types of base stations (macro base stations, micro base stations, and indoor distributed base stations) without requiring separate model training for each type.

[0209] 7) Explainable risk tracing to support precise operation and maintenance decisions: This solution not only outputs risk prediction values, but also automatically locates risk types (instantaneous / local / global), identifies key risk source nodes, and clarifies key monitoring parameters that lead to increased risks through three interpretable technologies: modal contribution analysis, graph attention backtracking, and SHAP value method. This enables operation and maintenance personnel to take precise measures directly based on the prediction results, significantly reducing operation and maintenance costs and fault recovery time.

[0210] Based on the same inventive concept, embodiments of this application also provide an electronic device, including a processor and a memory, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the IoT monitoring scenario risk prediction method based on spatiotemporal graph neural network of any of the above embodiments.

[0211] In an exemplary embodiment, an electronic device is provided, such as Figure 3 As shown, Figure 3 The illustrated electronic device 300 includes a processor 301 and a memory 303. The processor 301 and the memory 303 are connected, for example, via a bus 302. Optionally, the electronic device 300 may also include a transceiver 304. It should be noted that in practical applications, the transceiver 304 is not limited to one type, and the structure of this electronic device 300 does not constitute a limitation on the embodiments of this application.

[0212] Processor 301 may be a CPU, GPU (Graphics Processing Unit), DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 301 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of DSP and microprocessor, etc.

[0213] Bus 302 may include a pathway for transmitting information between the aforementioned components. Bus 302 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 302 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0214] The memory 303 may be a ROM (Read Only Memory) or other type of static storage device capable of storing static information and instructions, RAM (Random Access Memory) or other type of dynamic storage device capable of storing information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0215] The memory 303 stores computer program code that executes the scheme of this application, and its execution is controlled by the processor 301. The processor 301 executes the computer program code stored in the memory 303 to implement the content shown in the foregoing method embodiments.

[0216] Among them, electronic devices include, but are not limited to: mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), and in-vehicle terminals (such as in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0217] Based on the same inventive concept, this application also provides a storage medium storing a computer program, wherein the computer program is configured to execute the IoT monitoring scenario risk prediction method based on spatiotemporal graph neural network of any of the above embodiments at runtime.

[0218] Based on the same inventive concept, this application also provides a computer program product, including a computer program configured to execute the IoT monitoring scenario risk prediction method based on spatiotemporal graph neural network of any of the above embodiments at runtime.

[0219] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A risk prediction system for IoT monitoring scenarios based on spatiotemporal graph neural networks, characterized in that, The IoT monitoring scenario risk prediction system based on spatiotemporal graph neural network includes: A multivariate time series data acquisition module, which is used to acquire standardized multivariate time series data for the current period; The orthogonal decomposition module is used to perform orthogonal decomposition on the standardized multivariate time series data of the current period to obtain the risk propagation mode data set of the current period, wherein the risk propagation mode data set includes instantaneous risk mode, local propagation mode and global diffusion mode; A dynamic risk transmission diagram generation module is used to generate a dynamic adjacency matrix that can truly reflect the risk transmission path in the current period based on the risk propagation modality data set. The modal high-dimensional feature generation module is used to perform parallel extraction of multimodal spatiotemporal features based on the dynamic adjacency matrix that can truly reflect the risk propagation path in the current period and the risk propagation modal data group in the current period, thereby obtaining the modal high-dimensional feature information of the current period. The prediction module is used to predict the risk of the current period based on the modal high-dimensional feature information of the current period, so as to obtain the prediction result of the current period. The orthogonal decomposition module includes: The instantaneous risk mode extraction module is used to extract modes from the standardized multivariate time series data of the current period based on the topology of the Internet of Things (IoT) devices, thereby obtaining the instantaneous risk mode tensor of the current period; wherein, the topology of the IoT devices includes multiple device nodes; A local propagation mode extraction module is used to generate a local propagation mode tensor for the current period based on the instantaneous risk mode tensor for the current period. A global diffusion mode extraction module is used to generate a global diffusion mode tensor for the current period based on the instantaneous risk mode tensor for the current period. The dynamic risk transmission diagram generation module includes: The node-to-modal mutual information calculation module is used to generate an instantaneous modal mutual information matrix, a local modal mutual information matrix, and a global modal mutual information matrix based on the instantaneous risk modal tensor, the local propagation modal tensor, and the global diffusion modal tensor of the current period. A weighted fusion module for comprehensive risk correlation is used to generate the original comprehensive risk correlation matrix for the current period based on the instantaneous modal mutual information matrix, the local modal mutual information matrix, and the global modal mutual information matrix. A row-level sparsification module for the correlation matrix is ​​used to generate a sparse comprehensive risk correlation matrix for the current period based on the original comprehensive risk correlation matrix for the current period and the topology of IoT devices. An adjacency matrix construction module is used to generate a dynamic adjacency matrix for the current period based on the sparse comprehensive risk correlation matrix of the current period and the dynamic adjacency matrix of the previous period.

2. The IoT monitoring scenario risk prediction system based on spatiotemporal graph neural network as described in claim 1, characterized in that, The IoT monitoring scenario risk prediction system based on spatiotemporal graph neural network further includes: The risk tracing and analysis module is used to obtain the risk tracing results for the current period based on the prediction results of the current period and the modal high-dimensional feature information of the current period.

3. The IoT monitoring scenario risk prediction system based on spatiotemporal graph neural network as described in claim 1, characterized in that, The orthogonal decomposition module further includes: The reconstruction calibration module is used to perform orthogonality constraints and lossless reconstruction calibration on the instantaneous risk mode tensor, the local propagation mode tensor, the global diffusion mode tensor, and the standardized multivariate time series data of the current period during the training phase, so as to obtain the updated decomposition parameters.

4. The IoT monitoring scenario risk prediction system based on spatiotemporal graph neural network as described in claim 1, characterized in that, The modal high-dimensional feature generation module includes: A transient risk mode temporal feature extraction module is used to extract the transient risk mode feature tensor of the current period based on the transient risk mode tensor of the current period; A local propagation mode spatiotemporal feature extraction module is used to extract the local propagation mode feature tensor of the current period based on the local propagation mode tensor of the current period; A global diffusion mode spatiotemporal feature extraction module is used to generate a global diffusion mode feature tensor for the current period based on the global diffusion mode tensor for the current period.

5. The IoT monitoring scenario risk prediction system based on spatiotemporal graph neural network as described in claim 4, characterized in that, The prediction module includes: A multimodal feature dimension unification module is used to unify the dimensions of the instantaneous risk modal feature tensor, the local propagation modal feature tensor, and the global diffusion modal feature tensor of the current period, thereby obtaining a unified instantaneous risk modal feature tensor, a unified local propagation modal feature tensor, and a unified global diffusion modal feature tensor. The feature fusion module is used to fuse the instantaneous risk modal feature tensor, the local propagation modal feature tensor, and the global diffusion modal feature tensor of the same dimension in a unified dimension, so as to obtain the fused comprehensive spatiotemporal feature tensor and the modal weight vector of each device. A fully connected layer computing module is used to generate the original risk prediction value tensor from the fused integrated spatiotemporal feature tensor. The post-processing and standardization module is used to generate the final risk prediction value tensor for the current period based on the original risk prediction value tensor.

6. A method for risk prediction in IoT monitoring scenarios based on spatiotemporal graph neural networks, implemented using the IoT monitoring scenario risk prediction system based on spatiotemporal graph neural networks as described in claim 1, characterized in that, The IoT monitoring scenario risk prediction method based on spatiotemporal graph neural network includes: Obtain standardized multivariate time series data for the current period; Orthogonal decomposition is performed on the standardized multivariate time series data of the current period to obtain the risk propagation modality data set of the current period, which includes instantaneous risk modality, local propagation modality, and global diffusion modality. Generate a dynamic adjacency matrix that accurately reflects the risk propagation path for the current period based on the risk propagation modality data set; Based on the dynamic adjacency matrix that can truly reflect the risk propagation path in the current period and the risk propagation modal data group in the current period, multimodal spatiotemporal features are extracted in parallel to obtain the modal high-dimensional feature information of the current period. Risk prediction for the current period is performed based on the modal high-dimensional feature information of the current period, thereby obtaining the prediction result for the current period.

7. The method for risk prediction in IoT monitoring scenarios based on spatiotemporal graph neural networks as described in claim 6, characterized in that, Also includes: The risk tracing results for the current period are obtained based on the prediction results and the modal high-dimensional feature information of the current period.

Images