A knowledge graph-based intelligent operation and maintenance alarm system

By using a knowledge graph-based intelligent operation and maintenance alarm system, which leverages dynamic knowledge graphs and bidirectional propagation search technology, the problem of incomplete alarm convergence in IT operation and maintenance systems has been solved, achieving accurate alarm convergence and improving fault location efficiency.

CN123074908APending Publication Date: 2026-11-10GUIZHOU SUNAC ZHIYUN INFORMATION TECHNOLOGY CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611556108.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-09
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

When processing alarms, existing IT operations and maintenance systems have difficulty capturing both forward and backward propagation of alarms in dynamic dependencies, resulting in incomplete convergence results and an inability to accurately determine the alarm convergence range, leading to the loss of a large number of redundant alarms and critical alarms.

Method used

A knowledge graph-based intelligent operation and maintenance alarm system is adopted. A dynamic knowledge graph is constructed, and forward and backward propagation searches are performed in the knowledge graph through a bidirectional propagation module. Combined with real-time indicator intensity calculation, the alarm convergence range is automatically identified and merged for output.

Benefits of technology

It achieves accurate convergence of alarms in dynamic dependencies, reduces redundant alarms, improves alarm readability and fault location speed, and ensures the completeness and accuracy of convergence results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123074908A_ABST
    Figure CN123074908A_ABST
Patent Text Reader

Abstract

This invention discloses a knowledge graph-based intelligent operation and maintenance alarm system, relating to the field of IT operation and maintenance alarm processing technology. The system includes: a graph construction module, which constructs a dynamic knowledge graph containing entity nodes and relational edges. Entity nodes carry static and real-time indicator attributes of the operation and maintenance object, and relational edges represent dependencies and temporal relationships; a bidirectional propagation module, which uses the alarm entity mapping node as the starting node and performs forward and backward propagation searches along the relational edges to obtain upstream and downstream propagation paths; an intensity calculation module, which calculates the propagation intensity value of each node in the propagation path at the alarm timestamp based on its real-time indicator attributes; and a convergence output module, which marks nodes that meet the convergence conditions based on the propagation intensity value and merges and outputs the alarm information of the marked nodes. This system can accurately capture the alarm impact range and intensity, achieving efficient noise reduction and convergence.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of IT operations and maintenance alarm processing technology, specifically to a smart operations and maintenance alarm system based on knowledge graphs. Background Technology

[0002] A massive number of monitoring alerts are generated daily in IT operations and maintenance environments. Many of these alerts originate from the same root cause of a fault or are related, creating an alert storm. Traditional alert convergence methods primarily rely on predefined rules, fixed-time-window aggregation, or static topology relationships for alert compression. These methods obtain asset topology from configuration management databases and merge some alerts based on static connections between devices, or simply merge alerts based on their occurrence time. However, the dependencies between operational objects change dynamically during business operations. Multi-level call and impact chains form between application services, system platforms, and physical resources. A single fault can propagate bidirectionally upstream and downstream along these dependencies, causing related alerts at multiple levels. Static rules cannot adapt to this dynamic propagation characteristic, easily leading to incorrect convergence of alerts without causal relationships or separating alarms that should be related, resulting in unstable noise reduction effects. Topology-based alert analysis methods often only trace dependency chains in a single direction, making it difficult to simultaneously address root cause location and impact scope identification. When an application service malfunctions, its impact spreads to downstream services, with the root cause potentially originating from upstream system platforms or physical resources. Analyzing only one direction can lead to missing upstream root cause nodes or affected downstream nodes, resulting in incomplete convergence results. Furthermore, existing convergence criteria typically rely on fixed topology hop counts or simple attribute matching, lacking a quantitative measure of node anomaly severity. This makes it impossible to distinguish which nodes are truly affected in the propagation path and which are merely on the path but have normal metrics. This blurs the convergence boundary, resulting in either over-convergence that suppresses alarms from normal nodes, causing the loss of critical alarms, or insufficient convergence, requiring maintenance personnel to handle a large number of redundant alarms. Therefore, how to simultaneously capture both forward and backward propagation of alarms in a dynamic dependency environment and accurately determine the alarm convergence range based on the real-time status of each node is a pressing issue that current noise reduction and convergence technologies need to address. Summary of the Invention

[0003] This paper presents a knowledge graph-based intelligent operation and maintenance alarm system, which enables bidirectional propagation and search of alarms in dynamic dependencies and adaptive convergence based on real-time indicator strength, thereby improving the accuracy of alarm convergence and noise reduction.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a knowledge graph-based intelligent operation and maintenance alarm system, comprising: a graph construction module, used to construct a dynamic knowledge graph containing entity nodes and relational edges, wherein the entity nodes carry static attributes and real-time indicator attributes of operation and maintenance objects, and the relational edges represent the dependency relationships and temporal association relationships between operation and maintenance objects; an alarm mapping module, used to receive the original alarm stream, extract alarm entities from the original alarm stream, and map the alarm entities to the corresponding entity nodes in the dynamic knowledge graph, wherein the alarm entities contain alarm timestamps and alarm type identifiers; and a bidirectional propagation module, used to propagate the alarm entities to the corresponding entity nodes in the dynamic knowledge graph. Starting from a node, forward and backward propagation searches are performed along the relational edges in the dynamic knowledge graph. The forward propagation search traverses upstream nodes along the dependency relationships, and the backward propagation search traverses downstream nodes along the dependency relationships, obtaining upstream and downstream propagation paths respectively. An intensity calculation module is used to calculate the propagation intensity value of each node at the alarm timestamp based on the real-time indicator attributes of each node in the upstream and downstream propagation paths. A convergence output module is used to mark nodes that meet the convergence conditions in the upstream and downstream propagation paths based on the propagation intensity values, and merge and output the alarm information of the marked nodes.

[0005] Preferably, the knowledge graph construction module obtains the asset list of the operation and maintenance objects through the configuration management database, and divides the operation and maintenance objects into physical resource entities, system platform entities, and application service entities according to the object type identifier; determines the hierarchical relationship between each entity according to the object hierarchy identifier, which includes the superior-subordinate relationship between the physical layer, platform layer, and application layer; determines the call dependency relationship between each entity according to the network connection configuration and process communication configuration, which includes the service call direction and service call port; combines the hierarchical relationship and the call dependency relationship into the relationship edge, and associates the static attributes and real-time indicator attributes of the operation and maintenance objects with the corresponding entity nodes. The dynamic knowledge graph constructed in this way comprehensively depicts the three-layer topology of the IT environment's physical deployment, platform support, and business services, as well as the actual call links between services, providing an accurate and evolving digital base map for alarm propagation analysis.

[0006] The alarm mapping module parses each alarm record in the original alarm stream, extracting the alarm source identifier, alarm timestamp, and alarm type identifier. Using the alarm source identifier as the query key, it searches the dynamic knowledge graph for entity nodes with the same object identity. If found, the alarm entity is bound to that entity node, achieving immediate association between the alarm and the maintenance object. If not found, a temporary entity node is created in the dynamic knowledge graph based on the alarm source identifier, and this temporary entity node is used as the target node for alarm entity mapping. This prevents alarms from being missed due to changes in asset information or input delays, ensuring the completeness of alarm convergence.

[0007] The bidirectional propagation module obtains the entity type and alarm timestamp of the starting node. In a preferred implementation, when the entity type is an application service entity, forward propagation search is prioritized along the call dependency relationship to trace the underlying root cause of the service anomaly upwards. When the entity type is a physical resource entity, backward propagation search is prioritized along the hierarchical affiliation relationship to identify the upper-layer service scope affected by the underlying fault downwards. In the forward propagation search, the system traverses hop-by-hop along the relationship edges starting from the starting node, recording the node depth value and relationship edge direction at each hop. In the backward propagation search, the system traverses hop-by-hop in the opposite direction of the relationship edges, recording the reverse depth value and relationship edge reverse identifier at each hop. The forward propagation search terminates when the node depth value exceeds a preset forward hop count threshold, and the backward propagation search terminates when the reverse depth value exceeds a preset backward hop count threshold. This mechanism adaptively limits the search scope to the alarm root cause and the affected domain, avoiding the computational overhead and unwanted node contamination caused by a full graph traversal. Furthermore, forward and backward propagation searches can be executed in parallel, and the propagation search results of the intersection node can be merged when the upstream and downstream propagation paths intersect, thereby shortening the bidirectional traversal time and improving real-time analysis efficiency.

[0008] The intensity calculation module extracts a first indicator sequence within a first time window prior to the alarm timestamp from each node in the upstream propagation path, and a second indicator sequence within a second time window prior to the alarm timestamp from each node in the downstream propagation path, wherein the duration of the second time window is shorter than that of the first time window. It calculates the first cumulative change between adjacent sampled values ​​in the first indicator sequence and uses this first cumulative change as the first propagation contribution value for each node in the upstream propagation path. It also calculates the second cumulative change between sampled values ​​in the second indicator sequence and the sampled values ​​corresponding to the alarm timestamp, and uses this second cumulative change as the second propagation contribution value for each node in the downstream propagation path. The first and second propagation contribution values ​​for the same node are added together to obtain the propagation intensity value for that node. The upstream uses a longer window to capture trend-based cumulative anomalies, while the downstream uses a shorter window to capture strong symptoms of adjacent abrupt changes. The superposition of these two methods ensures that the propagation intensity value accurately reflects the degree of fault correlation of the node at the time of the alarm. Preferably, the duration of the first time window and the duration of the second time window can be dynamically adjusted according to the alarm level corresponding to the alarm type identifier. The higher the alarm level, the shorter the window duration, so that high-urgency alarms focus on recent indicator changes, and the intensity calculation results are more timely and sensitive.

[0009] In a preferred embodiment, after the propagation intensity value is calculated, propagation intensity attenuation verification is performed on the upstream and downstream propagation paths: The first path length of the upstream propagation path from the starting node to the upstreammost node, and the second path length of the downstream propagation path from the starting node to the downstreammost node are obtained; a first attenuation factor is calculated based on the first path length, which is inversely proportional to the first path length; a second attenuation factor is calculated based on the second path length, which is inversely proportional to the second path length; the propagation intensity value of each node in the upstream propagation path is multiplied by the first attenuation factor to obtain an updated first propagation intensity value; the propagation intensity value of each node in the downstream propagation path is multiplied by the second attenuation factor to obtain an updated second propagation intensity value. By attenuating at a distance inversely proportional to the propagation intensity, the propagation intensity of distant nodes is reasonably reduced, making the convergence determination more focused on the core area near the alarm source and suppressing noise signals at the end of long dependency chains.

[0010] When performing convergence marking, the convergence output module compares the propagation intensity values ​​of adjacent nodes sequentially upstream from the starting node along the upstream propagation path. When the propagation intensity value of an upstream node is less than that of its downstream adjacent node, the upstream node is marked as a decay node, and propagation further upstream from the decay node is terminated. Similarly, when performing convergence marking, the module compares the propagation intensity values ​​of adjacent nodes sequentially downstream from the starting node along the downstream propagation path. When the propagation intensity value of a downstream node is less than that of its upstream adjacent node, the downstream node is marked as a terminal decay node, and propagation further downstream from the terminal decay node is terminated. The node with the largest propagation intensity value among all nodes in the unterminated propagation directions is selected as the convergence center node. Nodes in the upstream and downstream propagation paths whose spectral distance from the convergence center node is less than a distance threshold are marked as convergence nodes. This method, based on intensity gradient comparison and center clustering, eliminates the need for manually setting multiple threshold levels and can automatically identify the convergence boundary of alarm storms and their impact on the cluster core.

[0011] When the convergence output module merges and outputs the alarm information of the marked nodes, it obtains the node identity identifiers of all marked nodes and the alarm record corresponding to each marked node; it uses the node identity identifier of the convergence center node as the primary key for merging alarms and uses the alarm record of the convergence center node as the main alarm record; it attaches the alarm records corresponding to other marked nodes (excluding the convergence center node) to the associated alarm list of the main alarm record. This associated alarm list contains the alarm type identifier and alarm timestamp of each associated alarm; it sorts the associated alarm list according to the order of the alarm timestamps, and combines the sorted associated alarm list with the main alarm record to form a converged alarm package. As a result, the operation and maintenance personnel receive a merged notification with the root cause or impact center as the main alarm and surrounding associated alarms in chronological order. A large number of redundant and derived alarms are converged and compressed, the number of alarms is significantly reduced, and the complete fault propagation context and chronological evidence are preserved, effectively improving alarm readability and fault location speed.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: The bidirectional propagation module starts with the entity node mapped to the alarm entity and simultaneously performs forward and backward propagation searches within a dynamic knowledge graph containing dependencies and temporal relationships. Forward propagation traverses upstream nodes along dependencies, while backward propagation traverses downstream nodes, obtaining the upstream and downstream propagation paths respectively. This bidirectional traversal mechanism comprehensively depicts the causal chain and impact range of an alarm event within the operations and maintenance system, without losing upstream root cause nodes or downstream affected nodes. Unlike convergence methods that only analyze along a single direction or rely solely on static topology, bidirectional propagation covers multi-level dependencies from physical resources and system platforms to application services along the detection path. When the fault originates from the underlying infrastructure, backward propagation can capture upper-layer application anomalies; when the fault originates from service components, forward propagation can trace the underlying supporting objects, ensuring a complete convergence perspective. This solution dynamically adapts to updates in the knowledge graph's relationship edges, eliminating the need for predefined propagation directions and preventing alarm relationship breaks due to changes in dependencies. The intensity calculation module utilizes the real-time indicator attributes of each node in the upstream and downstream propagation paths to calculate the propagation intensity value of each node at the alarm timestamp. It extracts the first indicator sequence within a first time window before the alarm timestamp from each node in the upstream path, calculating the first cumulative change between adjacent sampled values ​​as the first propagation contribution value. Similarly, it extracts the second indicator sequence within a second time window before the alarm timestamp from each node in the downstream path, calculating the second cumulative change between the sampled value and the sampled value corresponding to the alarm timestamp as the second propagation contribution value. The first and second propagation contribution values ​​of the same node are then added to obtain the propagation intensity value. This intensity value reflects the actual fluctuation of the node's indicators before and after the alarm and its contribution to the propagation path, rather than solely relying on graph topological distance or static attribute matching. Based on the propagation intensity value, convergence marking and boundary determination automatically identify the truly affected nodes in the propagation path, ensuring that nodes with significantly abnormal indicators are included in the convergence range, while path nodes with stable indicators are excluded, achieving precise adaptive delineation of the convergence boundary. Meanwhile, different time windows of varying durations are used to calculate the contributions of upstream and downstream, adapting to the different time characteristics of abnormal performance in upstream and downstream areas, thereby further improving the accuracy of intensity calculation. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0014] Figure 1 This is a schematic diagram of the structure of an IT operations and maintenance alarm noise reduction and convergence system based on knowledge graphs; Figure 2This is a flowchart of the alarm entity mapping to the dynamic knowledge graph; Figure 3 These are the depth distribution curves of the search nodes during forward and backward propagation; Figure 4 It is a curve comparing the propagation intensity value before and after it decreases due to the path length. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] See Figure 1 This invention provides a knowledge graph-based intelligent operation and maintenance alarm system, including a graph construction module, an alarm mapping module, a bidirectional propagation module, an intensity calculation module, and a convergence output module. The graph construction module constructs a dynamic knowledge graph containing entity nodes and relational edges. Entity nodes contain static attributes and real-time indicator attributes of operation and maintenance objects, while relational edges contain dependencies and temporal relationships between operation and maintenance objects. The alarm mapping module receives the original alarm stream, extracts alarm entities from it, and maps the alarm entities to corresponding entity nodes in the dynamic knowledge graph. Each alarm entity contains an alarm timestamp and an alarm type identifier. The bidirectional propagation module starts with the entity node mapped to the alarm entity and performs forward and backward propagation searches along relational edges in the dynamic knowledge graph. The forward propagation search traverses upstream nodes along dependencies, and the backward propagation search traverses downstream nodes along dependencies, obtaining the upstream and downstream propagation paths, respectively. The intensity calculation module calculates the propagation intensity value of each node at the alarm timestamp based on the real-time indicator attributes of each node in the upstream and downstream propagation paths. The convergence output module marks the nodes that meet the convergence conditions in the upstream and downstream propagation paths based on the propagation intensity value, and merges and outputs the alarm information of the marked nodes.

[0017] Example 1: In practice, the process of constructing a dynamic knowledge graph containing entity nodes and relation edges is as follows.

[0018] The asset list of operation and maintenance objects is obtained through the configuration management database, which stores the configuration information of managed and verified operation and maintenance objects. The asset list records the object type identifier, object identity identifier, and object level identifier for each operation and maintenance object in the form of a data table. The object type identifier is a preset category code used to distinguish the abstraction level to which the operation and maintenance object belongs; the object identity identifier is a unique identifier for the operation and maintenance object; and the object level identifier is the hierarchical number of the logical level of the operation and maintenance object in the IT environment.

[0019] Based on the object type identifier of each operation and maintenance object in the asset inventory, the operation and maintenance objects are divided into physical resource entities, system platform entities, and application service entities. Physical resource entities correspond to operation and maintenance objects with the first type code, system platform entities correspond to operation and maintenance objects with the second type code, and application service entities correspond to operation and maintenance objects with the third type code. Each type of entity forms an independent entity node in the dynamic knowledge graph, and the node category attribute of the entity node is assigned the corresponding entity type name.

[0020] Based on the object hierarchy identifier of each operation and maintenance object in the asset inventory, the hierarchical relationship between entities is determined. This hierarchical relationship includes the superior-subordinate relationship between the physical layer, platform layer, and application layer. Physical layer entity nodes are connected to platform layer entity nodes through hierarchical relationship edges, and platform layer entity nodes are connected to application layer entity nodes through hierarchical relationship edges. The direction of the hierarchical relationship edges is from lower-level entity to upper-level entity. The basis for determining the hierarchical relationship is as follows: in the asset inventory, if the object hierarchy identifier of the first operation and maintenance object represents the physical layer, and the object hierarchy identifier of the second operation and maintenance object represents the platform layer, and the physical location where the second operation and maintenance object runs has a deployment correspondence with the physical resources provided by the first operation and maintenance object, then a hierarchical relationship edge is established from the entity node corresponding to the first operation and maintenance object to the entity node corresponding to the second operation and maintenance object.

[0021] The call dependencies between entities are determined based on the network connection configuration and process communication configuration of the operation and maintenance objects. The network connection configuration contains TCP connection establishment information between operation and maintenance objects, specifically including the source IP address, source port number, destination IP address, and destination port number. The process communication configuration contains remote call records between application processes, specifically including the caller process identifier, the callee process identifier, and the communication protocol type. The directed connection between the source IP address and the destination IP address is extracted from the network connection configuration. The operation and maintenance object entity node corresponding to the source IP address is designated as the caller, and the operation and maintenance object entity node corresponding to the destination IP address is designated as the callee, establishing a call dependency edge. The direction of the call dependency edge is from the caller entity to the callee entity. The caller process identifier and the callee process identifier are extracted from the process communication configuration. The operation and maintenance object entity node to which the caller process identifier belongs is designated as the caller, and the operation and maintenance object entity node to which the callee process identifier belongs is designated as the callee, establishing a call dependency edge. The call dependency relationship includes the service call direction and the service call port. The service call direction is from the caller to the callee, and the service call port is the port number opened by the callee. The service call port is stored as an attribute of the call dependency edge.

[0022] Hierarchical affiliation and call dependency relationships are combined into relation edges in a dynamic knowledge graph. Relationship edges are represented in data storage as triples, including the source entity node identifier, relation type, and destination entity node identifier. Relationship types include hierarchical affiliation type and call dependency type. Hierarchical affiliation type relation edges store upstream and downstream node identifiers, and have a directional attribute from lower to higher levels. Call dependency type relation edges store the caller node identifier and the callee node identifier, along with a service call port attribute.

[0023] The static attributes and real-time metric attributes of the operation and maintenance objects are associated with the corresponding entity nodes. Static attributes are obtained from the asset list in the configuration management database, including the number of CPU cores, memory capacity, disk size, operating system version, application name, and version number. Real-time metric attributes are obtained through an interface with the operation and maintenance monitoring system, receiving real-time performance data of the operation and maintenance objects in each sampling period, including CPU utilization, memory utilization, disk I / O wait time, network throughput, and application response time. Internally, entity nodes store static attributes in key-value pairs and real-time metric attributes in a time-series data structure. Each record in the time-series data structure contains a sampling timestamp and the corresponding metric value. When the entity nodes of the dynamic knowledge graph receive new real-time metric attribute data, they update the corresponding time-series data structure, enabling the knowledge graph to reflect the real-time status of the operation and maintenance objects.

[0024] Example 2: In specific implementation, please refer to Figure 2 The implementation method for receiving the original alarm stream, extracting alarm entities from the original alarm stream, and mapping the alarm entities to the corresponding entity nodes in the dynamic knowledge graph is as follows.

[0025] The alarm mapping module receives raw alarm streams generated by the operation and maintenance monitoring system via a message queue or log collection agent. Each alarm record in the raw alarm stream is semi-structured text or JSON format data, containing an alarm source identifier field, an alarm timestamp field, and an alarm type identifier field. Each alarm record in the raw alarm stream is parsed, using regular expressions or a JSON parser to extract the alarm source identifier, alarm timestamp, and alarm type identifier. The alarm source identifier is a unique name for the operation and maintenance object that generated the alarm in the monitoring system; the alarm timestamp is the UTC time in milliseconds at the time the alarm was triggered; and the alarm type identifier is a preset alarm category code.

[0026] Using the extracted alarm source identifier as the query key, the system retrieves entity nodes with the same object identity identifier from the dynamic knowledge graph. In the entity node storage structure of the dynamic knowledge graph, the object identity identifier is a unique identifier for the operation and maintenance object obtained from the configuration management database, and it uses the same naming convention as the alarm source identifier. The retrieval process uses the node attribute index of the graph database to perform an equality match between the alarm source identifier and the object identity identifier attribute value of the entity node. If the string value of the alarm source identifier is completely identical to the object identity identifier attribute value of an entity node, then an entity node with the same object identity identifier has been retrieved.

[0027] If an entity node with the same object identity is found, the alarm entity corresponding to the alarm record is bound to the found entity node. An alarm entity is a data object composed of the alarm timestamp, alarm type identifier, and alarm source identifier from the alarm record. The binding method involves appending an alarm record entry to the attributes of the found entity node. This alarm record entry includes an alarm timestamp field, an alarm type identifier field, and a reference to the alarm entity. The entity node also retains its original object identity, static attributes, and real-time metric attributes.

[0028] If no entity node with the same object identity is found, a temporary entity node is created in the dynamic knowledge graph based on the alarm source identifier. The creation process for a temporary entity node is as follows: a new entity node is created in the graph database. The alarm source identifier is assigned to the object identity attribute of the temporary entity node, and the entity type attribute is set to a temporary type identifier. The remaining static attributes and real-time metric attributes are initialized to null values. Then, the alarm entity corresponding to the alarm record is bound to the temporary entity node in the same way as binding to an existing entity node. The temporary entity node serves as the target node for alarm entity mapping in the dynamic knowledge graph, performing subsequent processing, but it does not participate in the automatic construction of hierarchical relationships and call dependencies. When the configuration management database subsequently records the corresponding operation and maintenance object information, and the object identity of the temporary entity node matches the object identity in the configuration management database, the entity type attribute of the temporary entity node is updated to the corresponding entity type, and static attributes and real-time metric attributes are added. The dynamic knowledge graph synchronously establishes relationship edges between this node and other nodes.

[0029] Example 3: In its implementation, the bidirectional propagation module uses the entity node mapped to the alarm entity as the starting node and performs forward and backward propagation searches along the relational edges within the dynamic knowledge graph. The bidirectional propagation module retrieves the entity type attribute of the starting node and the alarm timestamp carried by the alarm entity from the dynamic knowledge graph. The entity type attribute is assigned a value during the construction of the dynamic knowledge graph, and its values ​​include physical resource entities, system platform entities, and application service entities. The alarm timestamp is the time value at which the alarm was triggered.

[0030] When the obtained entity type is an application service entity, during the forward propagation search, the bidirectional propagation module prioritizes traversing along the call dependency edges; during the backward propagation search, it traverses along the reverse direction of the relationship edges. When the obtained entity type is a physical resource entity, during the backward propagation search, the bidirectional propagation module prioritizes traversing along the reverse direction of the hierarchical affiliation edges; during the forward propagation search, it traverses along the relationship edges. When the obtained entity type is a system platform entity, the bidirectional propagation module does not distinguish between directional priorities in the forward and backward propagation searches, and simultaneously traverses along relationship edges of all types.

[0031] In the forward propagation search, the bidirectional propagation module starts from the initial node and traverses adjacent entity nodes hop-by-hop along the directions indicated by the relation edges. During each hop, the bidirectional propagation module records the node depth and relation edge direction of the currently traversed node. The node depth represents the number of hops from the initial node to the current traversed node; the node depth of the initial node is set to 0, and increases by 1 for each hop. The relation edge direction indicates the direction in which the traversed edges extend outward from the initial node.

[0032] In the backpropagation search, the bidirectional propagation module starts from the initial node and traverses adjacent entity nodes hop-by-hop along the reverse direction of the relation edges. During each hop, the bidirectional propagation module records the reverse depth value and the reverse identifier of the relation edge at the currently traversed node. The reverse depth value represents the number of hops from the initial node to the current traversed node along the reverse direction of the relation edge; the reverse depth value of the initial node is set to 0, and increases by 1 for each reverse hop. The reverse identifier of the relation edge is used to indicate that the current traversal was achieved by reaching the node through the reverse direction of the relation edge.

[0033] The bidirectional propagation module maintains a forward hop count threshold and a backward hop count threshold during traversal. The forward hop count threshold is an integer greater than or equal to 1, and the backward hop count threshold is also an integer greater than or equal to 1. These thresholds are pre-configured by the system administrator based on the average service call path depth and physical layer depth of the operational environment and stored in the system configuration file. After each jump in the forward propagation search, it checks whether the node depth of the currently traversed node exceeds the preset forward hop count threshold. If the node depth exceeds the threshold, the forward propagation search along the current path is terminated, and traversal continues. After each jump in the backward propagation search, it checks whether the backward depth of the currently traversed node exceeds the preset backward hop count threshold. If the backward depth exceeds the threshold, the backward propagation search along the current path is terminated, and traversal continues. The bidirectional propagation module ultimately obtains the upstream propagation path and the downstream propagation path starting from the starting node. The upstream propagation path is composed of all nodes recorded during the forward propagation search in traversal order, and the downstream propagation path is composed of all nodes recorded during the backward propagation search in traversal order.

[0034] See Figure 3 In the graph, the horizontal axis represents the depth value of nodes in the dynamic knowledge graph, ranging from 0 to 20. The vertical axis represents the probability density corresponding to the node depth value, which is used to characterize the statistical properties of node distribution during forward and backward propagation search. In the legend, the blue solid line corresponds to the node depth distribution during forward propagation search, and the red dashed line corresponds to the node depth distribution during backward propagation search.

[0035] Judging from the trend of the curve, the forward propagation search node depth distribution reaches the maximum probability density of about 0.263 at a depth value of 1. This indicates that during the forward propagation process, the vast majority of nodes are concentrated in the first layer of nodes adjacent to the starting node. Subsequently, as the node depth increases, the probability density decreases rapidly and approaches zero after the depth value of 9. This indicates that the occurrence frequency of deeper nodes is low, and the forward propagation search mainly covers the shallow nodes near the starting node.

[0036] The backward propagation search node depth distribution reaches a peak probability density of approximately 0.174 at a depth value of 1. The peak value is lower than that of the forward propagation search, but the overall probability density decreases at a relatively moderate rate, and the curve shows a smooth decreasing trend. Furthermore, it maintains a relatively low probability density around a depth value of 10, indicating that the backward propagation search is relatively more extensive and in-depth when traversing deep nodes.

[0037] Example 4: In practice, the intensity calculation module calculates the propagation intensity value of each node at the alarm timestamp based on the real-time indicator attributes of each node in the upstream and downstream propagation paths. For each node in the upstream propagation path, the first indicator sequence within the first time window preceding the alarm timestamp is extracted from the time series of the node's real-time indicator attributes. The first indicator sequence contains multiple sampled values ​​of the real-time indicator attributes within the first time window, arranged in chronological order of sampling time. The duration of the first time window is a preset fixed value, stored in the system configuration parameters in milliseconds.

[0038] The intensity calculation module calculates the first cumulative change between adjacent sample values ​​in the first index sequence. The formula for calculating the first cumulative change is: in, This represents the cumulative amount of the first change. This represents the total number of sampled values ​​in the first index sequence. Indicates the first index sequence Each sample value, Indicates the first index sequence Each sample value, This represents the absolute value of the difference between two adjacent sampled values. The intensity calculation module uses the calculated first cumulative change as the first propagation contribution value of the corresponding node.

[0039] For each node in the downstream propagation path, the intensity calculation module extracts a second indicator sequence within a second time window preceding the alarm timestamp from the real-time indicator attribute time series of that node. The duration of the second time window is shorter than the duration of the first time window, and the duration of the second time window is also a preset fixed value. The intensity calculation module obtains the real-time indicator attribute sample value corresponding to the alarm timestamp, and records it as the baseline sample value. It calculates the absolute value of the difference between each sample value in the second indicator sequence and the baseline sample value, and sums all the absolute values ​​of the differences to obtain the second cumulative change. The intensity calculation module uses the second cumulative change as the second propagation contribution value of the corresponding node.

[0040] For nodes appearing in both the upstream and downstream propagation paths, the intensity calculation module adds the node's first propagation contribution value and second propagation contribution value to obtain the node's propagation intensity value. For nodes appearing only in the upstream propagation path, the first propagation contribution value is directly used as the propagation intensity value. For nodes appearing only in the downstream propagation path, the second propagation contribution value is directly used as the propagation intensity value.

[0041] The convergence output module marks nodes that meet the convergence condition in both the upstream and downstream propagation paths based on their propagation intensity values. The module arranges all nodes in the upstream propagation path in ascending order from the starting node, forming an upstream node sequence. Starting from the starting node, the module sequentially extracts adjacent pairs of nodes along the upstream node sequence. The node closer to the starting node is called the downstream adjacent node, and the node farther from the starting node is called the upstream node. The module compares the propagation intensity values ​​of the upstream node and the downstream adjacent node. When the propagation intensity value of the upstream node is less than that of the downstream adjacent node, the upstream node is marked as a decay node, and propagation further upstream from this decay node is terminated; that is, no further processing is performed on the remaining nodes in the upstream node sequence after this decay node.

[0042] The convergence output module arranges all nodes in the downstream propagation path in order from the starting node downstream, forming a downstream node sequence. Starting from the starting node, the convergence output module sequentially extracts adjacent pairs of nodes along the downstream node sequence. The node closer to the starting node is called the upstream adjacent node, and the node farther from the starting node is called the downstream node. The convergence output module compares the propagation intensity values ​​of the downstream node and the upstream adjacent node. When the propagation intensity value of the downstream node is less than that of the upstream adjacent node, the downstream node is marked as a terminal decay node, and propagation further downstream of this terminal decay node is terminated. That is, the remaining nodes in the downstream node sequence after this terminal decay node are no longer processed.

[0043] Among all nodes in the unterminated propagation direction, the convergence output module obtains the propagation intensity value of each node and selects the node with the largest propagation intensity value, marking this node as the convergence center node. When multiple nodes have equal propagation intensity values, and all are the maximum value, the node closest to the starting node in the graph is selected as the convergence center node.

[0044] The convergence output module pre-maintains a distance threshold, which is a positive integer configuration parameter. Using all nodes in the upstream and downstream propagation paths as candidate sets, the module calculates the graph distance from each node in the candidate set to the convergence center node. The graph distance is defined as the minimum number of relation edges traversed from that node to the convergence center node in the dynamic knowledge graph. When the graph distance of a node is less than the distance threshold, the convergence output module marks that node as a converged node. The convergence center node itself is also marked as a converged node.

[0045] The convergence output module treats all marked convergence nodes as marker nodes, obtaining the node identity identifiers of all marker nodes and the corresponding alarm records for each marker node. The node identity identifier is a unique identifier for an entity node in the dynamic knowledge graph. The convergence output module uses the node identity identifier of the convergence center node as the primary key for merging alarms and uses the alarm record corresponding to the convergence center node as the primary alarm record. The convergence output module creates a related alarm list, adding the alarm records corresponding to other marked nodes besides the convergence center node to the related alarm list. Each related alarm record contains an alarm type identifier and an alarm timestamp. The convergence output module sorts the alarm records in the related alarm list according to their alarm timestamps, with earlier timestamps appearing first. The sorted related alarm list and the primary alarm record are combined into a convergence alarm package, which contains the primary key, the primary alarm record, and the sorted related alarm list. The convergence output module sends the convergence alarm package as the final alarm convergence result to the operation and maintenance alarm display interface or downstream alarm processing system.

[0046] Example 5: In practice, after the propagation intensity value is calculated, the intensity calculation module performs a propagation intensity attenuation check on the upstream and downstream propagation paths.

[0047] The strength calculation module obtains the path structure data of the upstream propagation path from the bidirectional propagation module. It then reads the first path length from the starting node to the upstreammost node from this data. The first path length is defined as the total number of relation edges traversed in the upstream propagation path from the starting node along the relation edge direction to the upstreammost node. Similarly, the strength calculation module reads the second path length from the starting node to the downstreammost node from the path structure data of the downstream propagation path. The second path length is defined as the total number of relation edges traversed in the downstream propagation path from the starting node along the relation edge direction in the opposite direction to the downstreammost node.

[0048] The first attenuation factor is calculated based on the first path length, and it is inversely proportional to the first path length. The formula for calculating the first attenuation factor is: in, Indicates the first attenuation factor. Indicates the length of the first path. For integers greater than or equal to 0, Indicates to Take the natural logarithm. When the length of the first path... When the value is 0, Equal to 0, first attenuation factor The value is 1; it varies with the length of the first path. Increase, denominator increases, first decay factor The decrease reflects the fact that the attenuation effect increases with increasing path length.

[0049] The second attenuation factor is calculated based on the second path length, and it is inversely proportional to the second path length. The formula for calculating the second attenuation factor uses the same functional form as the first attenuation factor; the second attenuation factor is obtained by substituting the second path length into the calculation. Let the second attenuation factor be denoted as... The length of the second path is , It is an integer greater than or equal to 0.

[0050] The intensity calculation module multiplies the propagation intensity value of each node in the upstream propagation path by the first attenuation factor. The updated first propagation intensity value is obtained and replaced with the updated first propagation intensity value. The intensity calculation module multiplies the propagation intensity value of each node in the downstream propagation path by the second attenuation factor. The updated second propagation strength value is obtained and replaced with the updated second propagation strength value. For nodes that are simultaneously located in the upstream and downstream propagation paths, the propagation strength value is first updated by the first attenuation factor, then by the second attenuation factor, and the final propagation strength value is the result of the two updates.

[0051] In some embodiments, forward propagation search and backward propagation search are executed in parallel. The bidirectional propagation module internally includes a forward search thread and a backward search thread. The forward search thread is responsible for performing the forward propagation search, and the backward search thread is responsible for performing the backward propagation search. Both threads start simultaneously after the alarm mapping module completes alarm entity mapping. The forward and backward search threads share the same set of visited nodes, which records the identifiers of entity nodes traversed by both search threads. During the operation of the forward and backward search threads, when a search thread traverses an entity node, it checks whether the identifier of that entity node already exists in the set of visited nodes. If it does, it indicates that the other search thread has also traversed that entity node, and this entity node is the intersection node of the upstream and downstream propagation paths. The bidirectional propagation module merges the forward and backward propagation search results of the intersection node. The merging method is to add the node depth value of the intersection node in the upstream propagation path to its reverse depth value in the downstream propagation path to obtain a composite depth value, and use this composite depth value to mark the position of the intersection node in the merged propagation path.

[0052] In some embodiments, the duration of the first time window and the duration of the second time window used by the intensity calculation module are dynamically adjusted according to the alarm level corresponding to the alarm type identifier. The system has a preset alarm level mapping table, which stores the correspondence between each alarm type identifier and the alarm level. Alarm levels include severity, major, minor, and alert levels, with severity having the highest priority and alert having the lowest priority. Before extracting the first and second indicator sequences, the intensity calculation module first uses the alarm type identifier of the alarm entity as the query key to search for the corresponding alarm level in the alarm level mapping table. When the found alarm level is severity, the duration of the first time window is set to the first severity window duration value, and the duration of the second time window is set to the second severity window duration value. When the found alarm level is major, the duration of the first time window is set to the first major window duration value, and the duration of the second time window is set to the second major window duration value. When the alarm level is minor or alert, the durations of the first and second time windows are set to their corresponding preset duration values, respectively. The higher the alarm level, the shorter the duration of the first time window and the shorter the duration of the second time window. The duration of the first critical window is shorter than the duration of the first major window, and the duration of the second critical window is shorter than the duration of the second major window. The duration values ​​of these windows are stored in the system configuration parameters in milliseconds and are set by the operations and maintenance administrator according to the indicator sampling period and alarm response time requirements of the operations and maintenance environment.

[0053] See Figure 4In the figure, the horizontal axis represents the relative position of the nodes in the propagation path, where the starting node's position is 0, the upstream nodes are negative, and the downstream nodes are positive; the vertical axis represents the propagation intensity value corresponding to each node. The figure contains two curves: the blue solid line represents the propagation intensity value without attenuation factor update, and the red dashed line represents the propagation intensity value after attenuation factor update as described in Example 5.

[0054] As observed from the blue solid line curve, the propagation intensity value reaches its peak near the starting node, with a peak value of approximately 150. Subsequently, it shows a significant downward trend both upstream and downstream from the starting node, gradually decreasing to lower values. This reflects a regular change in the propagation intensity value as the distance from the starting node increases. The curve is symmetrical overall, indicating that the propagation intensity changes in a consistent manner along the upstream and downstream directions of the propagation path.

[0055] The red dashed curve lies entirely below the blue solid curve, reflecting the effective compression of the overall propagation intensity value after the attenuation factor update. Its peak value is significantly lower than that of the blue curve, approximately 20, and the intensity values ​​in both upstream and downstream directions are significantly reduced, exhibiting a smoother trend. This attenuation effect aligns with the processing logic in Example 5, which calculates the attenuation factor based on the propagation path length and multiplies the propagation intensity value of each node by the corresponding attenuation factor, thus demonstrating the characteristic of propagation intensity gradually decreasing with increasing path length.

[0056] Furthermore, the red dashed curve approaches lower and more stable values ​​at both ends of the node position, which are far from the starting node. This indicates that the node propagation intensity on the long-distance propagation path is further compressed after attenuation, reducing the interference of abnormal nodes on the overall propagation path and enhancing the accuracy and stability of alarm convergence.

[0057] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A knowledge graph-based intelligent operation and maintenance alarm system, characterized in that, include: The graph construction module constructs a dynamic knowledge graph containing entity nodes and relation edges. The entity nodes contain the static attributes and real-time indicator attributes of the operation and maintenance objects, and the relation edges contain the dependency relationships and temporal association relationships between the operation and maintenance objects. The alarm mapping module receives the original alarm stream, extracts alarm entities from the original alarm stream, and maps the alarm entities to the corresponding entity nodes in the dynamic knowledge graph. The alarm entity includes an alarm timestamp and an alarm type identifier. The bidirectional propagation module takes the entity node mapped by the alarm entity as the starting node, and performs forward propagation search and backward propagation search in the dynamic knowledge graph along the relation edge. The forward propagation search traverses the upstream node along the dependency relationship, and the backward propagation search traverses the downstream node along the dependency relationship, respectively obtaining the upstream propagation path and the downstream propagation path. The intensity calculation module calculates the propagation intensity value of each node at the alarm timestamp based on the real-time indicator attributes of each node in the upstream propagation path and the downstream propagation path. The convergence output module marks the nodes that meet the convergence conditions in the upstream and downstream propagation paths based on the propagation intensity value, and merges and outputs the alarm information of the marked nodes.

2. The intelligent operation and maintenance alarm system based on knowledge graphs according to claim 1, characterized in that, The construction of a dynamic knowledge graph containing entity nodes and relation edges includes the following steps: The asset list of operation and maintenance objects is obtained by configuring the management database. The asset list includes the object type identifier, object identity identifier, and object hierarchy identifier for each operation and maintenance object. Based on the object type identifier, the operation and maintenance objects are divided into physical resource entities, system platform entities, and application service entities; The hierarchical relationship between entities is determined based on the object hierarchy identifier, and the hierarchical relationship includes the superior-subordinate relationship between the physical layer, platform layer and application layer; The call dependencies between entities are determined based on the network connection configuration and process communication configuration of the operation and maintenance objects. The call dependencies include the service call direction and the service call port. The hierarchical affiliation relationship and the call dependency relationship are combined into the relationship edge, and the static attributes and real-time indicator attributes of the operation and maintenance object are associated with the corresponding entity nodes.

3. The intelligent operation and maintenance alarm system based on knowledge graphs according to claim 2, characterized in that, The steps of receiving the original alarm stream, extracting alarm entities from the original alarm stream, and mapping the alarm entities to the corresponding entity nodes in the dynamic knowledge graph include the following: Each alarm record in the original alarm stream is parsed to extract the alarm source identifier, alarm timestamp, and alarm type identifier from the alarm record; Using the alarm source identifier as the query key, retrieve entity nodes with the same object identity identifier in the dynamic knowledge graph; If an entity node with the same object identity is found, the alarm entity corresponding to the alarm record is bound to the found entity node. If no entity node with the same object identity is found, a temporary entity node is created in the dynamic knowledge graph according to the alarm source identifier, and the temporary entity node is used as the target node for the alarm entity mapping.

4. The intelligent operation and maintenance alarm system based on knowledge graphs according to claim 3, characterized in that, The step of performing forward and backward propagation searches along the relational edges in the dynamic knowledge graph, starting from the entity node mapped by the alarm entity, includes the following steps: Obtain the entity type of the starting node and the alarm timestamp. The entity type includes physical resource entities, system platform entities, and application service entities. When the entity type is an application service entity, the forward propagation search is performed first along the call dependency relationship; When the entity type is a physical resource entity, the backward propagation search is performed preferentially along the hierarchical affiliation relationship; In the forward propagation search, starting from the initial node, the process traverses the relation edges hop by hop, recording the node depth value and relation edge direction of the current node during each hop. In the backpropagation search, starting from the starting node, the traversal is performed hop by hop in the opposite direction of the relation edge. During each hop, the reverse depth value of the current node and the reverse identifier of the relation edge are recorded. The forward propagation search is terminated when the node depth value exceeds a preset forward hop count threshold, and the backward propagation search is terminated when the backward depth value exceeds a preset backward hop count threshold.

5. The intelligent operation and maintenance alarm system based on knowledge graphs according to claim 4, characterized in that, The step of calculating the propagation intensity value of each node at the alarm timestamp based on the real-time indicator attributes of each node in the upstream propagation path and the downstream propagation path includes the following steps: Extract a first indicator sequence within a first time window prior to the alarm timestamp from each node of the upstream propagation path. The first indicator sequence contains multiple sampled values ​​of the real-time indicator attribute within the first time window. Extract a second indicator sequence within a second time window prior to the alarm timestamp from each node of the downstream propagation path, wherein the duration of the second time window is shorter than the duration of the first time window; Calculate the first cumulative change between adjacent sampled values ​​in the first index sequence, and use the first cumulative change as the first propagation contribution value of each node in the upstream propagation path; Calculate the second cumulative change between the sampled value in the second index sequence and the sampled value corresponding to the alarm timestamp, and use the second cumulative change as the second propagation contribution value of each node in the downstream propagation path; The propagation strength value of a node is obtained by adding the first propagation contribution value and the second propagation contribution value of the same node.

6. The intelligent operation and maintenance alarm system based on knowledge graphs according to claim 5, characterized in that, The step of marking nodes that meet the convergence condition in the upstream and downstream propagation paths based on the propagation intensity value includes the following steps: Along the upstream propagation path, the propagation intensity values ​​of adjacent nodes are compared sequentially from the starting node upstream. When the propagation intensity value of an upstream node is less than that of its downstream adjacent node, the upstream node is marked as an attenuation node, and propagation to the direction further upstream of the attenuation node is terminated. Along the downstream propagation path, the propagation intensity values ​​of adjacent nodes are compared sequentially from the starting node to the downstream direction. When the propagation intensity value of a downstream node is less than that of its upstream adjacent node, the downstream node is marked as an end attenuation node, and propagation to the direction further downstream of the end attenuation node is terminated. The node with the largest propagation intensity value among all nodes in the unterminated propagation directions is selected as the convergence center node. Nodes in the upstream and downstream propagation paths whose spectral distance from the convergence center node is less than a distance threshold are marked as convergence nodes.

7. The intelligent operation and maintenance alarm system based on knowledge graphs according to claim 6, characterized in that, The process of merging and outputting the alarm information of the marked nodes includes the following steps: Obtain the node identity of all marked nodes and the alarm record corresponding to each marked node; The node identity identifier of the convergence center node is used as the primary key for merging alarms, and the alarm record of the convergence center node is used as the main alarm record. The alarm records corresponding to the other marker nodes besides the convergence center node are attached to the associated alarm list of the main alarm record. The associated alarm list includes the alarm type identifier and alarm timestamp of each associated alarm. The associated alarm list is sorted according to the order of the alarm timestamps, and the sorted associated alarm list is combined with the main alarm record to form a converged alarm package.

8. The intelligent operation and maintenance alarm system based on knowledge graph according to claim 1, characterized in that, After the propagation intensity value is calculated, a propagation intensity attenuation check is performed on the upstream propagation path and the downstream propagation path. The attenuation check includes the following steps: Obtain the first path length from the starting node to the upstream node in the upstream propagation path, and the second path length from the starting node to the downstream node in the downstream propagation path; A first attenuation factor is calculated based on the first path length, and the first attenuation factor is inversely proportional to the first path length. The second attenuation factor is calculated based on the second path length, and the second attenuation factor is inversely proportional to the second path length. Multiply the propagation intensity value of each node in the upstream propagation path by the first attenuation factor to obtain the updated first propagation intensity value; The propagation intensity value of each node in the downstream propagation path is multiplied by the second attenuation factor to obtain the updated second propagation intensity value.

9. A knowledge graph-based intelligent operation and maintenance alarm system according to claim 4, characterized in that, The forward propagation search and the backward propagation search are executed in parallel, and the propagation search results of the intersection node are merged when the upstream propagation path and the downstream propagation path intersect.

10. A knowledge graph-based intelligent operation and maintenance alarm system according to claim 5, characterized in that, The duration of the first time window and the duration of the second time window are dynamically adjusted according to the alarm level corresponding to the alarm type identifier. The higher the alarm level, the shorter the duration of the first time window and the second time window.