Communication fault prediction method

Abstract

The invention provides a communication fault prediction method, and relates to the technical field of communication fault prediction, and the method comprises the steps: collecting the real-time operation data and historical fault data of each communication node in a target communication network; based on a preset time-space association threshold, screening adjacent nodes having data interaction association with the target communication node, and constructing a node time-space data set; performing feature extraction on the node spatio-temporal data set to obtain a fusion feature vector containing a time dimension operation trend feature and a space dimension correlation feature; and inputting the fusion feature vector into a pre-trained time-space diagram neural network prediction model, and outputting a communication fault prediction result of the target communication node. The technical problem of insufficient prediction accuracy of communication fault prediction in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of communication fault prediction, and in particular to a communication fault prediction method. BACKGROUND

[0002] Precise prediction of communication faults can convert passive repair of faults into proactive prevention, thereby ensuring stable operation of communication networks and reducing losses caused by link interruptions and equipment abnormalities.

[0003] However, existing fault prediction methods mostly use single-node data for prediction, without fully considering the spatial interaction and correlation between nodes in the communication network, and ignoring the trend characteristics of node operation data over time, resulting in low data information utilization, one-sided dimension, and susceptibility to accidental data fluctuations, which further leads to insufficient accuracy of fault prediction and difficulty in achieving effective early warning of communication faults.

[0004] Therefore, there is an urgent need for a communication fault prediction method that can integrate spatial and temporal correlation characteristics to improve the accuracy and foresight of fault prediction. SUMMARY

[0005] The present application provides a communication fault prediction method to solve the technical problem of insufficient prediction accuracy in the prior art.

[0006] The technical solution of the present application to solve the above technical problem is as follows: The present application provides a communication fault prediction method, comprising: Collecting real-time operation data and historical fault data of each communication node in the target communication network; Based on a preset spatiotemporal correlation threshold, filtering adjacent nodes that have data interaction and correlation with the target communication node, and constructing a node spatiotemporal data set; Performing feature extraction on the node spatiotemporal data set to obtain a fusion feature vector containing time dimension operation trend characteristics and space dimension correlation characteristics; Inputting the fusion feature vector into a pre-trained spatiotemporal graph neural network prediction model to output the communication fault prediction result of the target communication node.

[0007] The present application has the following beneficial effects: Compared with the prior art, firstly, the real-time running data and historical fault data of each communication node in the target communication network are collected, thereby providing quantifiable and strongly correlated basic data support for subsequent feature extraction and fault judgment. Secondly, based on a preset space-time correlation threshold, neighboring nodes having data interaction correlation with the target communication node are screened, a node space-time data set is constructed, effective neighboring nodes are screened out, and the prediction accuracy is prevented from being disturbed by irrelevant node data, and the node space-time data set is obtained, thereby providing a reliable data basis for subsequent space-time correlation feature analysis. Thirdly, feature extraction is performed on the node space-time data set, a fusion feature vector containing time dimension running trend features and space dimension correlation features is obtained, the node space-time data set is subjected to feature extraction in the time dimension and the space dimension, and a fusion feature vector reflecting the fault evolution law is formed, thereby providing high-quality input data for subsequent model prediction. Finally, the fusion feature vector is input into a pre-trained space-time graph neural network prediction model, and a communication fault prediction result of the target communication node is output, thereby providing a high-credibility basis for fault forward-looking operation and maintenance decision-making.

[0008] Through the technical solution, the fusion feature vector containing the time dimension running trend features and the space dimension correlation features is extracted, the problem of low utilization rate of single-dimension data is solved, the accidental data fluctuation interference is reduced through complementary verification of the space-time features, the implicit coupling law of the fault precursor is fully mined, and finally the communication fault prediction result is predicted and output through the space-time graph neural network prediction model. In this way, the accuracy of the communication fault prediction is improved, the fault precursor can be identified in advance, the traditional passive repair is upgraded to active prevention and control, the communication interruption risk and operation and maintenance cost caused by the fault are effectively reduced, and the stable operation of the target communication network is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 A flowchart of a communication fault prediction method provided by the application is shown in the figure. Figure 2 A flowchart of a space-time graph neural network prediction model pre-training step in a communication fault prediction method provided by the application is shown in the figure. DETAILED DESCRIPTION

[0010] The technical solutions in the embodiments of the application will be clearly and completely described in combination with the accompanying drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0011] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0012] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0013] Examples, such as Figure 1 As shown, this embodiment of the invention provides a communication fault prediction method, including: S10: Collect real-time operating data and historical fault data of each communication node in the target communication network.

[0014] There is a strong correlation between real-time operating data and historical fault data of communication network nodes and communication faults. Abnormal fluctuations in real-time operating data are often an immediate signal of a fault, while historical fault data can provide a regular reference for the correspondence between such anomalies and faults. Collecting these two types of data can accurately capture the logical chain of communication fault occurrence.

[0015] To address the aforementioned issues, this application collects real-time operational data and historical fault data of each communication node in the target communication network.

[0016] Specifically, step S10 in the method includes: Collect real-time operating data and historical fault data of each communication node in the target communication network. The real-time operating data includes at least the bandwidth utilization, data transmission latency, and data packet loss rate of the target communication node and its neighboring nodes.

[0017] In this embodiment, real-time operational data of each communication node in the target communication network is first collected. The real-time operational data includes at least the bandwidth utilization, data transmission latency, and packet loss rate of the target communication node and its neighboring nodes. The target communication node refers to the specific network device in the target communication network that requires fault prediction, such as a router, switch, or server. The neighboring nodes refer to other network devices that interact with the target communication node.

[0018] Bandwidth utilization rate refers to the ratio of the currently used bandwidth to the theoretical maximum bandwidth of the link. It can reflect the real-time load and congestion risk between the target communication node and its neighboring nodes. Generally, the higher the bandwidth utilization rate, the more fully the link is used. However, when the utilization rate is close to 100%, data packets need to be queued in the node buffer to wait for transmission, which can easily lead to problems such as increased data transmission delay and data packet loss. It is an important precursor to network congestion.

[0019] Data transmission latency refers to the time it takes for a data packet to travel from the target communication node to a neighboring node or from a neighboring node to the target communication node. It can reflect the smoothness of the network link and the processing efficiency of the communication node. If the data transmission latency suddenly increases significantly, such as more than twice the normal range, it usually indicates that there is an anomaly in the network path or communication node, such as route detour or memory overload.

[0020] Packet loss rate refers to the proportion of sent data packets that fail to reach the receiving end. It can reflect the reliability of data transmission and the stability of the link. In a normal network, the packet loss rate is usually less than 0.1%. If the packet loss rate increases significantly, such as exceeding 5%, it is often caused by serious fault risks, such as physical damage to device ports, link interruption, node buffer overflow, etc.

[0021] For example, bandwidth utilization and packet loss rate can be read from the Management Information Base (MIB) of the communication nodes, and traffic data between nodes can be collected through the NetFlow / IPFIX protocol. The timestamp difference of packet transmission can be recorded to calculate the data transmission latency. In this way, real-time operating data of each communication node in the target communication network can be obtained. Real-time operating data is directly related to the link health status of the communication nodes and reflects the early signs of failure, which is an important data prerequisite for accurate fault prediction.

[0022] Secondly, historical fault data of each communication node in the target communication network is collected. For example, historical fault data, including historical fault type, fault time information, fault location information, and fault association information, can be obtained from historical operation databases, fault event logs, and manual operation and maintenance records. Historical fault data can be compared with real-time operation data over time to uncover abnormal patterns of indicators before the fault occurred, providing historical reference for time trend analysis in feature extraction, thereby improving the accuracy of fault prediction.

[0023] In summary, compared to existing technologies, this application collects real-time operational data and historical fault data of each communication node in the target communication network. This provides quantifiable and strongly correlated foundational data support for subsequent feature extraction and fault diagnosis.

[0024] S20: Based on a preset spatiotemporal correlation threshold, filter out neighboring nodes that have data interaction correlation with the target communication node and construct a node spatiotemporal dataset.

[0025] The neighboring nodes connected to the target communication node have varying degrees of impact on the target communication node's failure risk. Generally speaking, neighboring nodes with high data interaction intensity with the target communication node are more likely to have their operational anomalies directly transmitted to the target communication node. Conversely, neighboring nodes with low data interaction intensity with the target communication node have less impact on the target communication node's failure risk even if they experience anomalies, and are therefore not selected as valid nodes. This differentiated screening can accurately focus on key nodes that have a substantial impact on the target node's failure risk, avoiding interference from irrelevant node data on prediction accuracy.

[0026] To address the aforementioned issues, this application uses a preset spatiotemporal correlation threshold to filter neighboring nodes that have data interaction relationships with the target communication node, and constructs a node spatiotemporal dataset.

[0027] Specifically, step S20 in the method includes: Calculate the average data interaction volume between the target communication node and each neighboring node within a preset statistical period, and determine the neighboring nodes whose average data interaction volume is greater than a preset spatiotemporal correlation threshold as valid neighboring nodes. The real-time operation data and historical fault data of the target communication node and its effective neighboring nodes are integrated and sorted by timestamp to construct a node spatiotemporal dataset.

[0028] In this embodiment, the average data interaction volume between the target communication node and each neighboring node within a preset statistical period is first calculated. Neighboring nodes whose average data interaction volume exceeds a preset spatiotemporal correlation threshold are then identified as valid neighboring nodes. The preset statistical period is a time window for calculating the average interaction volume. Determining the preset statistical period requires balancing correlation stability and real-time performance. Setting it too short may lead to misjudgments of correlation due to occasional peaks, while setting it too long may fail to reflect recent interaction changes. Preferably, this application recommends dynamically setting the period based on the fluctuation frequency of network data interaction; for example, it could be set to 1 hour for enterprise networks and 10 minutes for high-frequency trading networks.

[0029] The average data interaction volume is the ratio of the total amount of data transmitted bidirectionally between the target communication node and its neighboring nodes within a preset statistical period to the preset statistical period itself. For example, if the preset statistical period is 24 hours and the target communication node and its neighboring node A have transmitted a total of 12GB of data in the past 24 hours, then the average data interaction volume = 12 / 24 = 0.5GB / hour. The average data interaction volume reflects the intensity of data interaction between the target communication node and its neighboring nodes per unit time. The higher the average data interaction volume, the more frequent and closely related the data interaction between the target communication node and its neighboring nodes is within the preset statistical period. This means that the real-time operating status of the neighboring nodes is more likely to affect the transmission stability of the target communication node, and the failure risk of the target communication node may also be directly related to the interaction behavior of the neighboring nodes. Therefore, neighboring nodes with high interaction intensity are more valuable for predicting and analyzing the failure of the target communication node and should be given priority attention as effective neighboring nodes.

[0030] The preset spatiotemporal correlation threshold is a quantitative standard for judging whether the correlation between a target communication node and its neighboring nodes is significant. Setting the preset spatiotemporal correlation threshold too high may miss weakly correlated but critical nodes, while setting it too low may introduce redundant data. Therefore, it can be dynamically determined based on the specific network scenario and historical interaction data. For example, in a data center network, the core switch and access switch interact frequently, so the preset spatiotemporal correlation threshold can be set higher, such as 1GB / hour, while the edge device interacts less with the core switch, so the preset spatiotemporal correlation threshold can be set lower, such as 100MB / hour.

[0031] For example, the average data interaction volume is compared with a preset spatiotemporal association threshold. For instance, if the average data interaction volume between the target communication node and neighboring node A is 0.5 GB / hour and the preset spatiotemporal association threshold is 0.4 GB / hour, the average data interaction volume is greater than the preset spatiotemporal association threshold. Then, neighboring node A is determined to be a valid neighboring node, indicating that neighboring node A has a significant data association with the target communication node. Conversely, if the average data interaction volume is less than or equal to the preset spatiotemporal association threshold, it indicates that the association between the neighboring node and the target communication node is weak and the impact of the fault on the target communication node is negligible. In this case, the neighboring node is excluded.

[0032] Secondly, real-time operational data of the target communication node and its effective neighboring nodes are collected at a fixed acquisition frequency. Historical fault data of the target communication node and its effective neighboring nodes within a preset historical time range are obtained. Then, the real-time operational data and historical fault data of the target communication node and its effective neighboring nodes are integrated and sorted by timestamp to construct a node spatiotemporal dataset. The acquisition frequency and preset historical time range can be dynamically set according to the actual application scenario and needs. For example, for communication networks with high real-time requirements, the acquisition frequency can be set to once per second and the historical time range to the past month. For communication networks with stability as a priority, the acquisition frequency can be set to once every 10 minutes and the historical time range to the past 6 months. Dynamic adjustment ensures that the dataset can capture both immediate state fluctuations and contain sufficient historical correlation information.

[0033] For example, if the target communication node is router R1, and the selected effective neighboring nodes are switches S1 and S2, with a collection frequency of once every 5 minutes and a preset historical time range of the past 3 months, then the real-time operating data of router R1, switch S1, and switch S2 and the historical fault records of the past 3 months are collected at a collection frequency of once every 5 minutes. Then, the real-time operating data and historical fault data are sorted and integrated according to the collection timestamp to form a node spatiotemporal dataset that simultaneously contains the time dynamic information and spatial correlation information of the target communication node and effective neighboring nodes. For example, {2024-09-05 10:00:00, router R1's bandwidth utilization is 98%, data transmission latency is 50ms, packet loss rate is 5%, port congestion is 15min, switch S1's bandwidth utilization is 90%, data transmission latency is 40ms, packet loss rate is 4%, switch S2's bandwidth utilization is 95%, data transmission latency is 55ms, packet loss rate is 6%; 2024-09-05 10:05:00, ...}.

[0034] In summary, compared to existing technologies, this application, based on a preset spatiotemporal correlation threshold, filters neighboring nodes that have data interaction relationships with the target communication node, constructing a node spatiotemporal dataset. This process identifies effective neighboring nodes, avoids interference from irrelevant node data with prediction accuracy, and obtains a node spatiotemporal dataset, providing a reliable data foundation for subsequent spatiotemporal correlation feature analysis.

[0035] S30: Extract features from the spatiotemporal dataset of the nodes to obtain a fused feature vector containing time-dimensional trend features and spatial-dimensional correlation features.

[0036] The propagation of faults in the target communication network exhibits significant spatiotemporal coupling characteristics: From a temporal perspective, faults are often not sudden but rather evolve gradually with abnormal indicators. For example, if a node's bandwidth utilization is too high, it will simultaneously show a trend of increased latency and packet loss rate. This continuous deterioration of the operating state will accumulate over time and eventually evolve into an explicit fault. From a spatial perspective, faults will spread and propagate along related nodes. For example, if a node experiences a high packet loss rate due to a port failure, other nodes that have close data interaction with it will experience a surge in bandwidth usage due to data packet retransmission, which in turn will lead to increased latency. This diffusion of fault impact from neighboring nodes to the target node reflects the spatial correlation of fault propagation.

[0037] To address the aforementioned issues, this application extracts features from the spatiotemporal dataset of the nodes to obtain a fused feature vector that includes time-dimensional trend features and spatial-dimensional correlation features.

[0038] Specifically, step S30 in the method includes: Using a preset time window as the sliding unit, the real-time running data of each communication node in the spatiotemporal dataset of the node is slidably calculated to obtain the standard deviation of the fluctuation of each communication node in different time windows, which is used as the running trend feature in the time dimension. Calculate the correlation coefficient between the real-time running data of the target communication node and each effective neighboring node, and combine the physical topological distance weight of each communication node to generate spatial dimension correlation features. The time-dimensional trend features and the spatial-dimensional correlation features are concatenated according to the feature dimensions, and then standardized to obtain the fused feature vector.

[0039] In this embodiment, a sliding calculation is first performed on the real-time operating data of each communication node in the node spatiotemporal dataset using a preset time window as the sliding unit. This yields the standard deviation of the fluctuation of each communication node within different time windows, which serves as the time-dimensional operating trend feature. The preset time window can be dynamically set according to the actual application scenario and historical fault data, such as 30 minutes or 1 hour.

[0040] For example, if the preset time window is 30 minutes and the sliding step size is set to 30 minutes (i.e., the windows do not overlap), multiple consecutive time windows are generated, such as 10:00-10:30, 10:30-11:00, 11:00-11:30, etc. A sliding calculation is performed on the real-time operating data (including bandwidth utilization, data transmission latency, and packet loss rate) of the target communication node and each effective neighboring node in the node spatiotemporal dataset. Within each time window, based on all real-time operating data within that time period, such as a collection frequency of once every 5 minutes, the 30-minute time window contains 6 sets of real-time operating data. The standard deviation of fluctuation for each set of real-time operating data is calculated. The standard deviation of fluctuation reflects the stability of the real-time operating data; the larger the standard deviation, the more drastic the fluctuation of the real-time operating data within that time window, and drastic fluctuations are often a typical precursor to a failure.

[0041] For example, within the 10:00-10:30 time window, the bandwidth utilization of the target communication node router R1 is 65%, 68%, 72%, 69%, 75%, and 70%, respectively. The calculated standard deviation of bandwidth utilization fluctuation is 5.2%. Using the same method, the standard deviation of data transmission latency fluctuation is calculated to be 8.5ms, and the standard deviation of packet loss rate fluctuation is 1.3%. Thus, the standard deviation of fluctuation within the 10:00-10:30 time window is obtained. Following the same logic and method, the standard deviations of fluctuation within different time windows are calculated. Finally, the standard deviations of fluctuation within all time windows are arranged in chronological order to obtain the time-dimensional operational trend characteristics. These time-dimensional operational trend characteristics can reflect the evolution of the fluctuation level of real-time operational data over time, providing a temporal basis for detecting early signs of faults.

[0042] Secondly, the correlation coefficients of real-time operational data between the target communication node and each effective neighboring node are calculated, and combined with the physical topological distance weights of each communication node, spatial dimension correlation features are generated. Specifically, communication failures often exhibit significant propagation in networks. Their propagation probability and impact mainly depend on the tightness of data interaction between nodes and their physical proximity. Generally speaking, the more frequent the data interaction and the closer the physical distance, the easier it is for the failure to propagate quickly through the link, and the greater the impact on the target node. Therefore, by combining the correlation coefficients of real-time operational data between the target communication node and each effective neighboring node (which reflects the synchronicity of changes in real-time operational data; higher synchronicity indicates tighter coupling between nodes) with the physical topological distance weights of each communication node (which reflects the spatial proximity between nodes; the closer the distance, the higher the probability of failure propagation), spatial dimension correlation features are obtained. This can accurately quantify the failure propagation risk intensity between the target communication node and each effective neighboring node, providing a reliable basis for subsequent models to capture cross-node failure correlation signals.

[0043] Finally, the time-dimensional trend features and spatial-dimensional correlation features are concatenated according to the feature dimensions and standardized to obtain a fused feature vector. Specifically, based on the sliding order of the preset time windows, the standard deviations of bandwidth utilization, data transmission latency, and packet loss rate fluctuations under each time window are sequentially concatenated to form a one-dimensional time feature vector; based on the topology numbering order of effective neighboring nodes, the spatial correlation contribution values ​​of each effective neighboring node are sequentially concatenated to form a one-dimensional spatial feature vector. The one-dimensional time feature vector and the one-dimensional spatial feature vector are concatenated end to end in the order of the one-dimensional time feature vector first and the one-dimensional spatial feature vector last to obtain an initial fused feature vector. The initial fused feature vector is then standardized to obtain the final fused feature vector.

[0044] Specifically, the step of "calculating the correlation coefficient of real-time operating data between the target communication node and each effective neighboring node, and generating spatial dimension correlation features by combining the physical topological distance weights of each communication node" includes: Obtain real-time operational data of the target communication node and each effective neighboring node; For each piece of real-time operational data, the correlation coefficient between the target communication node and a single effective neighboring node within a preset statistical period is calculated to obtain three basic correlation coefficients. The average of the three basic correlation coefficients is taken as the comprehensive correlation coefficient between the target communication node and its corresponding effective neighboring nodes. Based on the node physical topology map of the target communication network, the physical topology coordinates of the target communication node and each effective neighboring node are extracted, and the physical topology distance between them is calculated using the Euclidean distance formula. The physical topology distance is normalized to obtain a distance normalization value, and the physical topology distance weight of the corresponding effective neighboring node is calculated by a preset weight conversion formula. Multiply the comprehensive correlation coefficient of a single effective neighbor node by the corresponding physical topological distance weight to obtain the spatial correlation contribution value of the corresponding effective neighbor node to the target communication node. Integrate the spatial association contribution values ​​of all valid neighboring nodes, sort them by the topological number of the valid neighboring nodes, and generate spatial dimension association features.

[0045] In this embodiment of the application, real-time operating data of the target communication node and each effective neighboring node, including bandwidth utilization, data transmission latency, and data packet loss rate, are first obtained.

[0046] Secondly, for bandwidth utilization, data transmission latency, and packet loss rate, correlation coefficients (such as Pearson correlation coefficients) are calculated between the target communication node and a single effective neighboring node within a preset statistical period, resulting in three basic correlation coefficients. The Pearson correlation coefficient ranges from -1 to 1. The closer the Pearson correlation coefficient is to 1, the more synchronized the changes in that real-time data between the target communication node and the single effective neighboring node are within the preset statistical period. For example, within a preset statistical period (e.g., 1 hour), the basic correlation coefficient for bandwidth utilization between the target communication node (e.g., router R1) and the effective neighboring node (e.g., switch S1) is 0.9, the basic correlation coefficient for data transmission latency is 0.85, and the basic correlation coefficient for packet loss rate is 0.8. These three factors together constitute the three basic coefficients.

[0047] Secondly, the average of the three basic correlation coefficients is taken as the comprehensive correlation coefficient between the target communication node and its corresponding effective neighboring nodes. This comprehensive correlation coefficient integrates the synchronization of bandwidth utilization, data transmission latency, and packet loss rate, avoiding the randomness of a single indicator and comprehensively reflecting the multi-dimensional correlation strength between the target communication node and its corresponding effective neighboring nodes. For example, the comprehensive correlation coefficient between the target communication node (e.g., router R1) and its effective neighboring node (e.g., switch S1) is (0.9 + 0.85 + 0.8) / 3 = 0.85, which serves as the overall correlation strength between the two.

[0048] Furthermore, based on the physical topology map of the target communication network, the physical topology coordinates of the target communication node and each effective neighboring node are extracted, and the physical topology distance between them is calculated using the Euclidean distance formula. For example, the physical topology map of the target communication network can be generated based on technical documents (such as network architecture blueprints and equipment deployment plans) formed during the early planning and design phase of the target communication network, or through professional network topology management tools (such as Visio). Then, the physical topology coordinates of the target communication node and each effective neighboring node are extracted from the physical topology map of the target communication network. For example, the physical topology coordinates of the target communication node (such as router R1) are (10, 20), and the physical topology coordinates of the effective neighboring node (such as switch S1) are (15, 25). Then, the physical topology distance between them is calculated using the Euclidean distance formula √[(x2-x1)²+(y2-y1)²]. For example, the physical topology distance = √[(15-10)²+(25-20)²] = 7.07.

[0049] Furthermore, several physical topological distances are normalized to obtain distance normalization values. Then, the physical topological distance weights of the corresponding effective neighboring nodes are calculated using a preset weight transformation formula. For example, Z-score normalization can be used for normalization to obtain distance normalization values. The closer the physical topological distance, the smaller the distance normalization value. Thus, the distance normalization value can intuitively reflect the relative distance relationship. For example, physical topological distance weight = 1 - distance normalization value. The closer the physical topological distance, the smaller the distance normalization value, and the higher the physical topological distance weight. This is because faults in neighboring nodes are more easily propagated to the target communication node through physical links, while the impact of distant nodes is weaker. Therefore, assigning higher weights to nearby nodes allows spatial correlation characteristics to more accurately reflect the fault propagation potential between nodes.

[0050] Furthermore, the comprehensive correlation coefficient of a single effective neighbor node is multiplied by its corresponding physical topology distance weight to obtain the spatial correlation contribution value of the corresponding effective neighbor node to the target communication node. For example, if the comprehensive correlation coefficient between the target communication node (e.g., router R1) and the effective neighbor node (e.g., switch S1) is 0.85, and the corresponding physical topology distance weight is 0.7, then the spatial correlation contribution value of the effective neighbor node (e.g., switch S1) to the target communication node (e.g., router R1) is 0.85 × 0.7 = 0.595. Following the same method, the spatial correlation contribution values ​​of all effective neighbor nodes are calculated. The spatial correlation contribution value combines the comprehensive correlation coefficient and the corresponding physical topology distance weight, reflecting the actual correlation impact of a single effective neighbor node on the failure risk of the target communication node, thus balancing the data-level correlation tightness and the spatial failure propagation potential.

[0051] Finally, the spatial correlation contribution values ​​of all effective neighboring nodes are integrated and sorted according to their unique topology numbers in the node physical topology map to generate spatial dimension correlation features. For example, if the spatial correlation contribution value of effective neighboring node switch S1 is 0.595 and the spatial correlation contribution value of effective neighboring node switch S2 is 0.35, after sorting by topology number, the spatial dimension correlation features generated are: 0.595 and 0.35. The spatial dimension correlation features can comprehensively reflect the actual correlation impact of all effective neighboring nodes on the failure risk of the target communication node.

[0052] Specifically, the step of "concatenating the time-dimensional trend features and spatial-dimensional correlation features according to feature dimensions, and obtaining the fused feature vector after standardization" includes: Based on the sliding order of the preset time windows, the standard deviations of the fluctuations of bandwidth utilization, data transmission latency, and data packet loss rate under each time window are sequentially concatenated to form a one-dimensional time feature vector; Based on the topological numbering order of the effective neighboring nodes, the spatial correlation contribution values ​​of each effective neighboring node are sequentially concatenated to form a one-dimensional spatial feature vector; The one-dimensional time feature vector and the one-dimensional space feature vector are concatenated end to end in the order of one-dimensional time feature vector first and one-dimensional space feature vector last to obtain the initial fused feature vector. The initial fused feature vector is standardized to obtain the fused feature vector.

[0053] In this embodiment, the standard deviations of bandwidth utilization, data transmission delay, and packet loss rate under each time window are first concatenated based on the sliding order of the preset time windows to form a one-dimensional time feature vector. For example, according to the sliding order of the preset time windows, the standard deviations of bandwidth utilization, data transmission delay, and packet loss rate under each time window (10:00-10:30, 10:30-11:00, 11:00-11:30, ...) are concatenated. For example, if the standard deviation of bandwidth utilization is 5.2%, the standard deviation of data transmission delay is 8.5ms, and the standard deviation of packet loss rate is 1.3% within the 10:00-10:30 time window, then the one-dimensional time feature vector is [5.2%, 8.5ms, 1.3%, ...].

[0054] Secondly, based on the topological numbering order of the effective neighboring nodes, the spatial correlation contribution values ​​of each effective neighboring node are sequentially concatenated to form a one-dimensional spatial feature vector, such as [0.595, 0.35, ...].

[0055] Next, following the order of one-dimensional time feature vector first and one-dimensional space feature vector last, the one-dimensional time feature vector and the one-dimensional space feature vector are concatenated end to end to obtain the initial fused feature vector, such as [5.2%, 8.5ms, 1.3%, ..., 0.595, 0.35, ...].

[0056] Finally, the initial fused feature vector is standardized to obtain the fused feature vector. For example, Z-score standardization can be used to standardize the initial fused feature vector to eliminate the influence of different units and dimensions, resulting in fused feature vectors of the same order of magnitude.

[0057] Specifically, the phrase "for each piece of real-time operational data, calculate the correlation coefficient between the target communication node and a single effective neighboring node within a preset statistical period to obtain three basic correlation coefficients" includes: Regarding bandwidth utilization, the bandwidth utilization data of the target communication node and a single effective neighboring node are extracted for each timestamp within a preset statistical period to construct a time series of bandwidth utilization for both parties. A linear correlation analysis was performed on the two bandwidth utilization time series, and the obtained correlation coefficient value was used as the first basic correlation coefficient. To address data transmission latency, the transmission latency data between the target communication node and a single effective neighboring node is collected for each timestamp within a preset statistical period, and a transmission latency time series between the two parties is constructed. A linear correlation analysis was performed on the two transmission delay time series, and the obtained correlation coefficient value was used as the second basic correlation coefficient. To address the packet loss rate, the packet loss rate records for each timestamp between the target communication node and a single effective neighboring node within a preset statistical period are obtained, and a time series of packet loss rates for both parties is constructed. Linear correlation analysis was performed on the two packet loss rate time series, and the obtained correlation coefficient value was used as the third basic correlation coefficient.

[0058] In this embodiment, firstly, regarding bandwidth utilization, bandwidth utilization data corresponding to each timestamp of the target communication node and a single effective neighboring node within a preset statistical period is extracted to construct a bandwidth utilization time series for both parties. Then, a linear correlation analysis is performed on the two bandwidth utilization time series, and the resulting correlation coefficient is used as the first basic correlation coefficient. The preset statistical period can be dynamically set according to actual application scenarios and needs, such as 1 hour. The linear correlation analysis can be performed using the Pearson coefficient to calculate the correlation coefficient.

[0059] For example, the bandwidth utilization data of the target communication node (such as router R1) and the effective neighboring node (such as switch S1) at each timestamp within a preset statistical period (such as 1 hour) is extracted. If the collection frequency is once every 5 minutes, there are a total of 12 pairs of bandwidth utilization data. Then, the data is aligned according to the collection timestamp to obtain the bandwidth utilization corresponding to each timestamp, such as router R1: [60%, 65%, 70%, ...], switch S1: [55%, 60%, 65%, ...], which serve as the time series of bandwidth utilization for both parties. Then, the correlation coefficient value is calculated using the Pearson coefficient. For example, first calculate the mean of two bandwidth utilization time series: the mean (μ1) of the target communication node (e.g., router R1) = (60% + 65% + ...) / 12 ≈ 70.08%, and the mean (μ2) of the effective neighboring node (e.g., switch S1) = (55% + 60% + ...) / 12 ≈ 65.08%; then calculate the covariance: multiply (target communication node bandwidth utilization - μ1) and (effective neighboring node bandwidth utilization - μ2) for each timestamp, sum them, and then divide by the sample size n, such as Σ[(xᵢ-μ1)(yᵢ-μ2)] / n = [(60%-70.08%)(55%-65.08%) + (65%- 70.08%)(60%-65.08%)+…] / 12≈71.375%; Next, calculate the standard deviation of the two bandwidth utilization time series: the standard deviation (σ1) of the target communication node (such as router R1) = √[Σ(xᵢ-μ1)² / 12]≈8.32%, and the standard deviation (σ2) of the effective neighboring node (such as switch S1) = √[Σ(yᵢ-μ2)² / 12]≈9.53%; Finally, calculate the correlation coefficient value = covariance / (σ1×σ2)≈71.375%% / (8.32%×9.53%)≈0.9, which serves as the first basic correlation coefficient, indicating that the bandwidth utilization trends of the two are relatively synchronized.

[0060] Secondly, following the same logic and calculation method, for data transmission delay, the transmission delay data corresponding to each timestamp of the target communication node and a single effective neighboring node within a preset statistical period is collected to construct the transmission delay time series of both parties. A linear correlation analysis is then performed on the two transmission delay time series, and the obtained correlation coefficient value is used as the second basic correlation coefficient. For example, if the data transmission delay of the target communication node (e.g., router R1) and the effective neighboring node (e.g., switch S1) at each timestamp is: Router R1: [20ms, 25ms, 30ms, ...], Switch S1: [18ms, 23ms, 28ms, ...], this is used as the data transmission delay time series of both parties. Then, the correlation coefficient value is calculated using the Pearson coefficient, such as 0.85, which is used as the second basic correlation coefficient, indicating that the data transmission delay changes of both parties have the same trend.

[0061] Finally, following the same logic and calculation method, for the packet loss rate, the packet loss rate records corresponding to each timestamp of the target communication node and a single effective neighboring node within a preset statistical period are obtained, constructing a packet loss rate time series for both parties. A linear correlation analysis is then performed on the two packet loss rate time series, and the obtained correlation coefficient value is used as the third basic correlation coefficient. For example, if the packet loss rates of the target communication node (e.g., router R1) and the effective neighboring node (e.g., switch S1) at each timestamp are: Router R1: [0.1%, 0.2%, 0.3%, ...], Switch S1: [0.15%, 0.25%, 0.35%, ...], this is used as the packet loss rate time series for both parties. Then, the correlation coefficient value is calculated using the Pearson coefficient, such as 0.8, which is used as the third basic correlation coefficient, indicating that the packet loss rate changes of both parties are synchronized.

[0062] Thus, the three basic correlation coefficients can quantify the association status between the target communication node and a single effective neighboring node from three dimensions: capacity load, transmission speed, and reliability. If an effective neighboring node is highly correlated with the target communication node in all three basic correlation coefficients, it indicates that it has a strong correlation with the failure risk of the target communication node.

[0063] In summary, compared to existing technologies, this application extracts features from the aforementioned spatiotemporal dataset of nodes to obtain a fused feature vector that includes temporal trend features and spatial correlation features. Thus, feature extraction is performed on the spatiotemporal dataset of nodes in both temporal and spatial dimensions, forming a fused feature vector that reflects the evolution of faults, providing high-quality input data for subsequent model predictions.

[0064] S40: Input the fused feature vector into the pre-trained spatiotemporal graph neural network prediction model, and output the communication failure prediction result of the target communication node.

[0065] The aforementioned steps, through in-depth mining and multi-dimensional feature extraction and fusion of the node spatiotemporal dataset, yield a fused feature vector that takes into account both the temporal dimension of operational trends and the spatial dimension of correlation, which can be used to predict communication failures.

[0066] Furthermore, the Spatiotemporal Graph Neural Network (STGNN) is a cross-model that integrates Graph Neural Network (GNN) and temporal modeling techniques. It is suitable for processing complex data that simultaneously contains spatial correlations and temporal dynamics. Therefore, it is naturally adapted to the characteristics of spatiotemporal fusion features such as fused feature vectors, which can maximize the value of features and improve the prediction accuracy of spatiotemporally coupled faults in communication networks.

[0067] To address the aforementioned issues, this application inputs the fused feature vector into a pre-trained spatiotemporal graph neural network prediction model and outputs the communication failure prediction result of the target communication node.

[0068] Specifically, such as Figure 2 As shown, step S40 in the method includes: Collect historical spatiotemporal data of multiple communication network nodes, label the fault occurrence corresponding to each historical spatiotemporal data node, and construct a sample fault dataset. Feature extraction and fusion processing are performed on the historical node spatiotemporal data in the sample fault dataset to obtain a sample fusion feature vector set; Using the sample fusion feature vector set as input and the fault labeling information of the sample fault dataset as supervision signal, a spatiotemporal graph neural network prediction model is constructed.

[0069] In this embodiment, historical spatiotemporal data of multiple communication network nodes are first collected, and the fault occurrence corresponding to each historical node spatiotemporal data is labeled to construct a sample fault dataset. For example, historical spatiotemporal data of multiple communication network nodes can be collected across different scenarios, such as data centers, campuses, and wide area networks. This includes real-time operating data (at least bandwidth utilization, data transmission latency, and packet loss rate) and historical fault records of various devices such as routers and switches at different historical time periods. Simultaneously, each set of data is labeled as faulty or fault-free, and the specific fault type and fault occurrence timestamp are refined to ensure that the data corresponds to the time sequence of the faults. Thus, a sample fault dataset is obtained.

[0070] Secondly, feature extraction and fusion processing are performed on the historical node spatiotemporal data in the sample fault dataset to obtain a sample fusion feature vector set. For example, following the same logic and method as the previous steps, feature extraction and fusion processing are performed on the historical node spatiotemporal data in the sample fault dataset: using a preset time window as the sliding unit, the historical node spatiotemporal data of each communication node is calculated to obtain the standard deviation of fluctuation of each communication node within different time windows, serving as the time-dimensional operational trend feature; the correlation coefficient between the target communication node and each effective neighboring node is calculated, and combined with the physical topological distance weight of each communication node, spatial-dimensional association features are generated; the time-dimensional operational trend features and spatial-dimensional association features are concatenated according to feature dimensions, and after standardization, the sample fusion feature vector set is obtained. This ensures that the structure and dimensions of the sample fusion feature vector and the fusion feature vector are completely matched, allowing the learned patterns to be directly transferred to practical applications.

[0071] Finally, using the sample fusion feature vector set as input and the fault labeling information of the sample fault dataset as supervision signal, a spatiotemporal graph neural network prediction model is constructed. For example, the spatiotemporal graph neural network prediction model can be constructed through the following technical path: 1. Data preparation: Divide the sample fusion feature vector set and the sample fault dataset into training, validation, and test sets in a 7:1.5:1.5 ratio. 2. Model construction: The model mainly consists of a graph convolution module, a temporal modeling module, a feature fusion layer, and an output layer. The graph convolution module (e.g., GCN, GAT) takes spatial dimension correlation features as input, captures the topological association between the target node and its effective neighboring nodes through an adjacency matrix, and outputs node-level spatial features. The temporal modeling module (e.g., LSTM, Transformer) receives temporal dimension running trend features, extracts the dynamic evolution law of indicators over time through gating or self-attention mechanisms, and outputs temporal features. The feature fusion layer integrates spatial and temporal features using concatenation and fully connected methods, and enhances nonlinear expressive power through the ReLU activation function. The output layer, set according to the prediction target, uses the Softmax activation function to output the probability distribution of multiple fault types. 3. Model Training: Using the fused feature vectors of the samples in the training set as input and the fault labeling information of the corresponding sample fault data as supervision signals, the cross-entropy loss function or MSE loss function is used to calculate the error between the predicted value and the true label during training. The optimizer is Adam, and the initial learning rate is set to 0.001. The learning rate is dynamically adjusted through a cosine annealing strategy (decreasing by 10% every 5 rounds) to balance the convergence speed and accuracy. The batch size is set to 32 to avoid the training instability caused by the large amount of input data in a single session. The accuracy of the validation set is monitored in real time. When the accuracy of the validation set is stable above 90% for 10 consecutive rounds and the accuracy of the test set reaches the target simultaneously, the training is stopped and the model parameters are saved to obtain the spatiotemporal graph neural network prediction model.

[0072] Furthermore, the spatiotemporal graph neural network prediction model, which integrates feature vectors as input, outputs the probability distribution of multiple types of faults of the target communication node as the communication fault prediction result.

[0073] In summary, compared to existing technologies, this application inputs the fused feature vector into a pre-trained spatiotemporal graph neural network prediction model, outputting a communication fault prediction result for the target communication node. Thus, based on the spatiotemporal graph neural network prediction model, accurate communication fault prediction results are output, providing a highly reliable basis for proactive fault-based operational and maintenance decisions.

[0074] In summary, the embodiments of this application have at least the following technical effects: Compared to existing technologies, this application first collects real-time operational data and historical fault data of each communication node in the target communication network. This provides quantifiable and strongly correlated basic data support for subsequent feature extraction and fault diagnosis.

[0075] Secondly, this application uses a preset spatiotemporal correlation threshold to filter neighboring nodes that have data interaction relationships with the target communication node, thus constructing a node spatiotemporal dataset. This process filters out effective neighboring nodes, avoids interference from irrelevant node data with prediction accuracy, and obtains a node spatiotemporal dataset, providing a reliable data foundation for subsequent spatiotemporal correlation feature analysis.

[0076] Furthermore, this application extracts features from the aforementioned spatiotemporal dataset of nodes to obtain a fused feature vector containing both temporal trend features and spatial correlation features. Thus, feature extraction is performed on the spatiotemporal dataset of nodes in both temporal and spatial dimensions, forming a fused feature vector that reflects the evolution of faults, providing high-quality input data for subsequent model predictions.

[0077] Finally, this application inputs the fused feature vector into the pre-trained spatiotemporal graph neural network prediction model, outputting the communication fault prediction result of the target communication node. Thus, based on the spatiotemporal graph neural network prediction model, accurate communication fault prediction results are output, providing a highly reliable basis for proactive fault-based operation and maintenance decisions.

[0078] Through the above technical solution, this application solves the problem of low utilization of single-dimensional data by extracting a fused feature vector of temporal trend features and spatial correlation features. Furthermore, it reduces the interference of accidental data fluctuations through complementary verification of spatiotemporal features, comprehensively uncovering the implicit coupling patterns of fault precursors. Finally, it uses a spatiotemporal graph neural network prediction model to predict and output communication fault prediction results. This improves the accuracy of communication fault prediction, enables early identification of fault precursors, upgrades traditional passive repair to proactive prevention, effectively reduces the risk of communication interruption and maintenance costs caused by faults, and ensures the stable operation of the target communication network.

[0079] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0080] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0081] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0082] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0083] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0084] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0085] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting communication failures, characterized in that, The method includes: Collect real-time operational data and historical fault data of each communication node in the target communication network; Based on a preset spatiotemporal correlation threshold, neighboring nodes that have data interaction correlation with the target communication node are selected to construct a node spatiotemporal dataset; Feature extraction is performed on the spatiotemporal dataset of the nodes to obtain a fused feature vector containing time-dimensional trend features and spatial-dimensional correlation features; The fused feature vector is input into the pre-trained spatiotemporal graph neural network prediction model, which outputs the communication failure prediction result of the target communication node.

2. The communication fault prediction method according to claim 1, characterized in that, Collect real-time operating data and historical fault data of each communication node in the target communication network. The real-time operating data includes at least the bandwidth utilization, data transmission latency, and data packet loss rate of the target communication node and its neighboring nodes.

3. The communication fault prediction method according to claim 1, characterized in that, Based on a preset spatiotemporal correlation threshold, neighboring nodes that have data interaction relationships with the target communication node are selected to construct a node spatiotemporal dataset, including: Calculate the average data interaction volume between the target communication node and each neighboring node within a preset statistical period, and determine the neighboring nodes whose average data interaction volume is greater than a preset spatiotemporal correlation threshold as valid neighboring nodes. The real-time operation data and historical fault data of the target communication node and its effective neighboring nodes are integrated and sorted by timestamp to construct a node spatiotemporal dataset.

4. The communication fault prediction method according to claim 1, characterized in that, Feature extraction is performed on the spatiotemporal dataset of the nodes to obtain a fused feature vector containing temporal trend features and spatial correlation features, including: Using a preset time window as the sliding unit, the real-time running data of each communication node in the spatiotemporal dataset of the node is slidably calculated to obtain the standard deviation of the fluctuation of each communication node in different time windows, which is used as the running trend feature in the time dimension. Calculate the correlation coefficient between the real-time running data of the target communication node and each effective neighboring node, and combine the physical topological distance weight of each communication node to generate spatial dimension correlation features. The time-dimensional trend features and the spatial-dimensional correlation features are concatenated according to the feature dimensions, and then standardized to obtain the fused feature vector.

5. The communication fault prediction method according to claim 4, characterized in that, Calculate the correlation coefficient between the real-time operating data of the target communication node and each effective neighboring node, and combine it with the physical topological distance weight of each communication node to generate spatial dimension correlation features, including: Obtain real-time operational data of the target communication node and each effective neighboring node; For each piece of real-time operational data, the correlation coefficient between the target communication node and a single effective neighboring node within a preset statistical period is calculated to obtain three basic correlation coefficients. The average of the three basic correlation coefficients is taken as the comprehensive correlation coefficient between the target communication node and its corresponding effective neighboring nodes. Based on the node physical topology map of the target communication network, the physical topology coordinates of the target communication node and each effective neighboring node are extracted, and the physical topology distance between them is calculated using the Euclidean distance formula. The physical topology distance is normalized to obtain a distance normalization value, and the physical topology distance weight of the corresponding effective neighboring node is calculated by a preset weight conversion formula. Multiply the comprehensive correlation coefficient of a single effective neighbor node by the corresponding physical topological distance weight to obtain the spatial correlation contribution value of the corresponding effective neighbor node to the target communication node. Integrate the spatial association contribution values ​​of all valid neighboring nodes, sort them by the topological number of the valid neighboring nodes, and generate spatial dimension association features.

6. The communication fault prediction method according to claim 4, characterized in that, The time-dimensional trend features and spatial-dimensional correlation features are concatenated along their respective feature dimensions, and after standardization, the fused feature vector is obtained, which includes: Based on the sliding order of the preset time windows, the standard deviations of the fluctuations of bandwidth utilization, data transmission latency, and data packet loss rate under each time window are sequentially concatenated to form a one-dimensional time feature vector; Based on the topological numbering order of the effective neighboring nodes, the spatial correlation contribution values ​​of each effective neighboring node are sequentially concatenated to form a one-dimensional spatial feature vector; The one-dimensional time feature vector and the one-dimensional space feature vector are concatenated end to end in the order of one-dimensional time feature vector first and one-dimensional space feature vector last to obtain the initial fused feature vector. The initial fused feature vector is standardized to obtain the fused feature vector.

7. The communication fault prediction method according to claim 5, characterized in that, For each piece of real-time operational data, the correlation coefficient between the target communication node and a single effective neighboring node within a preset statistical period is calculated, resulting in three basic correlation coefficients, including: Regarding bandwidth utilization, the bandwidth utilization data of the target communication node and a single effective neighboring node are extracted for each timestamp within a preset statistical period to construct a time series of bandwidth utilization for both parties. A linear correlation analysis was performed on the two bandwidth utilization time series, and the obtained correlation coefficient value was used as the first basic correlation coefficient. To address data transmission latency, the transmission latency data between the target communication node and a single effective neighboring node is collected for each timestamp within a preset statistical period, and a transmission latency time series between the two parties is constructed. A linear correlation analysis was performed on the two transmission delay time series, and the obtained correlation coefficient value was used as the second basic correlation coefficient. To address the packet loss rate, the packet loss rate records for each timestamp between the target communication node and a single effective neighboring node within a preset statistical period are obtained, and a time series of packet loss rates for both parties is constructed. Linear correlation analysis was performed on the two packet loss rate time series, and the obtained correlation coefficient value was used as the third basic correlation coefficient.

8. The communication fault prediction method according to claim 1, characterized in that, The pre-training steps of the spatiotemporal graph neural network prediction model include: Collect historical spatiotemporal data of multiple communication network nodes, label the fault occurrence corresponding to each historical spatiotemporal data node, and construct a sample fault dataset. Feature extraction and fusion processing are performed on the historical node spatiotemporal data in the sample fault dataset to obtain a sample fusion feature vector set; Using the sample fusion feature vector set as input and the fault labeling information of the sample fault dataset as supervision signal, a spatiotemporal graph neural network prediction model is constructed.

Images