An AI-based intelligent inspection and maintenance system for communication networks

Abstract

The application belongs to the technical field of communication network operation and management, and discloses an AI-based communication network intelligent inspection and maintenance system, which interacts with the Web interface of a multi-manufacturer network management, carries out document object model analysis on a page, establishes an interface operation model irrelevant to manufacturer protocols, configures inspection items according to network element types and service scenarios, and controls a browser to complete cross-page operation to collect running data. The system fuses inspection data with historical data, work orders and equipment history, constructs a running state view organized according to network elements and time, and identifies abnormal types and levels by combining rules and classification models. Furthermore, the system generates a maintenance strategy containing inspection and change steps by relying on a maintenance operation template library, executes a maintenance task by combining test session simulation and rollback control, and improves automation and running reliability.

Description

Technical Field

[0001] This invention belongs to the field of communication network operation and maintenance technology, specifically an AI-based intelligent inspection and maintenance system for communication networks. Background Technology

[0002] As the scale of operator backbone networks, access networks, and enterprise private networks continues to expand, multiple manufacturers and models of network elements operate in parallel for extended periods. Daily inspections and maintenance are typically carried out using web-based network management systems provided by various manufacturers. Under the current model, frontline maintenance personnel need to log in to multiple network management interfaces, follow the maintenance procedures to click menus, fill in query conditions, switch pages, export performance reports and alarm lists, and then manually summarize them to form inspection records. After discovering anomalies, they need to refer to the operation manual to perform configuration changes and troubleshooting on the corresponding pages. This is a large workload and relies heavily on personal experience.

[0003] To improve efficiency, some networks use scripts or work order systems to semi-automate fixed inspection processes. However, scripts are often highly coupled with specific network management versions, page structures, and redirection paths. Once the interface is adjusted or the version is upgraded, they need to be re-recorded or modified, resulting in high maintenance costs. Operation and maintenance-related data are also scattered across multiple platforms such as performance monitoring systems, alarm systems, work order systems, equipment history databases, and data center environment monitoring systems. There is a lack of a unified status view centered on network elements and organized along a timeline, which is not conducive to the comprehensive analysis of the relationship between performance changes, alarm evolution, and maintenance operations. In terms of anomaly diagnosis and maintenance decisions, existing networks mostly rely on single threshold alarms or simple alarm rules, making it difficult to make comprehensive judgments by combining performance trends, historical work orders, and environmental data. The quantitative assessment of the impact of maintenance operations on services and the management of rollback paths are also relatively weak, which brings certain difficulties to the refined inspection and maintenance of large-scale communication networks. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-based intelligent inspection and maintenance system for communication networks to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI-based intelligent inspection and maintenance system for communication networks, suitable for centralized operation and maintenance management of communication networks including network elements from multiple manufacturers and of multiple models. This system interfaces with existing Web-based network management systems, and automatically completes inspection data collection, operational status analysis, and the execution and adjustment of maintenance operations without modifying the network element side interfaces. This reduces reliance on manual intervention and interface operations, and lowers maintenance costs and risks caused by changes in the network management interface.

[0006] To achieve the above objectives, the system provided by this invention establishes a communication connection with the Web network management system in the operator's existing network. The system is composed of functional modules for interface parsing, inspection configuration and execution, status fusion and analysis, anomaly diagnosis, maintenance strategy generation, and maintenance execution and knowledge update. The modules interact with each other through clearly defined data flow and control flow, forming a closed loop from data acquisition to maintenance strategy adjustment.

[0007] The interface parsing unit is used to obtain the web page output by the network management system, parse the page document object model, identify interface elements such as menus, buttons, input boxes, and result tables and their hierarchical relationships, and establish a mapping relationship between interface elements and specific operation types by combining the event response methods of interface elements, forming an interface operation model that is independent of the network element manufacturer's interface protocol. With the help of this model, the system can drive page operations in a unified way on network management interfaces of different manufacturers and different versions, without having to maintain scripts or interface adaptation logic separately for each network element.

[0008] The inspection task configuration unit configures the performance indicators, alarm items, and configuration items to be acquired according to different network element types and business scenarios. It then maps these inspection requirements to the combination of interface elements in the interface operation model to form a structured inspection item configuration. Based on the inspection item configuration, the inspection execution unit controls the browser session to complete operations such as login, menu navigation, condition filling, query triggering, and result page switching. It collects the running data of the target network element on multiple interconnected pages and organizes the collected results into inspection raw data corresponding to each network element, providing a unified data entry point for subsequent analysis.

[0009] The status fusion and analysis unit receives the raw inspection data and aligns it with historical inspection data, fault work order data, equipment history data and environmental monitoring data on the time axis and network element identification dimension to construct an operation status view that includes network element entities, status attributes and event records. Through this view, information scattered in different business systems is incorporated into the same data structure, which makes it easier to extract features that reflect changes in operation status and provides a complete data foundation for subsequent diagnosis.

[0010] Based on the operational status view, the anomaly diagnosis unit combines pre-configured judgment rules and a classification model trained based on historical fault samples to evaluate the operational status of each network element, outputting the anomaly level and one or more suspected fault types for each network element. The rule part is used to quickly screen obvious anomalies such as threshold exceeding limits and specific alarm combinations, while the classification model part uses multi-dimensional features to distinguish candidate fault types, thereby covering more fault scenarios while maintaining diagnostic stability.

[0011] Based on the diagnostic results provided by the anomaly diagnosis unit, the maintenance strategy generation unit selects maintenance operations corresponding to the suspected fault types from the maintenance operation template library compiled from historical maintenance tasks. The maintenance operations are mapped to a sequence of page operations that can be executed step by step on the network management page, and organized into maintenance strategies in the order of checking first and then changing. During the generation of maintenance strategies, the maintenance strategy generation unit can combine the service carrying information in the running status view to calculate the service impact level of different change operations, and mark the operations with higher impact levels as steps that require manual confirmation.

[0012] The maintenance execution and knowledge update unit is responsible for applying the maintenance strategy to the target network element. This unit first calls the interface operation model in an independent test session to simulate the page operation sequence included in the maintenance strategy, and checks whether the page jumps, interface element positioning, and key field display are consistent with the maintenance strategy when it was compiled. Only when the test session verification is successful will the maintenance strategy be executed on the actual network element in the formal session. During the formal execution, this unit monitors relevant performance indicators and alarm changes. When no expected improvement occurs within the preset observation time, or when a new high-priority alarm occurs, the configuration changes that occurred during the maintenance period are restored in sequence according to the rollback steps preset in the maintenance strategy, and the network element configuration is restored to the state before maintenance to reduce the impact of configuration adjustment failure.

[0013] During the long-term operation of the system, the maintenance execution and knowledge update unit performs statistical analysis on the execution logs of each inspection and maintenance task, including information such as inspection path length, number of page visits, task completion time, whether the maintenance task was successful, whether it was rolled back, and whether the same type of fault recurred within a preset time period after maintenance. Based on these statistical results, the system can adjust the priority of maintenance operation templates in the template library, and at the same time adjust the threshold used for feature determination in the state fusion and analysis unit and the classification threshold of the classification model in the anomaly diagnosis unit, so that the inspection path and maintenance strategy for the same type of fault are gradually optimized in actual operation, which is more in line with the actual operating environment and equipment characteristics of the communication network.

[0014] This invention, without relying on a unified interface from network element manufacturers, utilizes structured parsing and abstract operation models of the Web network management interface to achieve automatic inspection and maintenance of network elements from multiple manufacturers and models. Through multi-source data fusion and a rule-based diagnostic approach combined with a sample-trained classification model, it improves the accuracy and coverage of fault identification. Furthermore, by employing test session verification, a rollback mechanism, and parameter adjustment based on execution logs, it enhances the security and controllability of automatic maintenance operations in complex network environments, thereby reducing the burden of manual operations and mitigating the impact of interface changes on network operation and maintenance.

[0015] Furthermore, the page parsing unit performs structured parsing of the web page of the network management system according to the preset processing flow and maintains the interface operation model. Specifically, the page parsing unit first obtains the document object model tree corresponding to the current web page. Based on the type of the node and its hierarchical position in the tree, the page is divided into a navigation area, a condition input area, and a result display area. In each area, the page parsing unit identifies interface elements such as menu items, text input boxes, drop-down boxes, buttons, and result tables. Subsequently, the page parsing unit detects the event handling functions bound to each interface element node, identifies interface elements that will trigger page jumps, initiate data queries, or submit configuration changes, and records these interface elements as entry points for executable operations.

[0016] Based on this, the page parsing unit generates stable positioning information for each type of interface element. For example, it determines the positioning method according to the document object model tree path, attribute combination, or text features. Combined with the aforementioned event types, it establishes a mapping relationship between the interface element identifier and the corresponding operation type and writes it into the interface operation model. With the help of this interface operation model, the subsequent inspection execution unit and maintenance execution and knowledge update unit can directly call the corresponding interface elements when driving the browser to perform menu jumps, condition filling, and operation triggering, without relying on manually recorded and maintained operation scripts.

[0017] Considering that the page structure of the network management system may change during version upgrades or interface adjustments, the page parsing unit is also used to compare and update the page structure. When a change in the page structure is detected, the page parsing unit compares the types, hierarchical positions, and text content of nodes in the document object model tree of the old and new versions, redetermines the navigation area, condition input area, and result display area, and updates the positioning information and mapping relationship of relevant interface elements. In this way, without using fixed operation scripts, the consistency between the interface operation model and the actual page structure is maintained. In this way, when there are multiple vendors and multiple versions of network management interfaces and the interface structure changes, the system can still maintain the continuity of the inspection and maintenance process, reduce the workload of script maintenance, and reduce the risk of inspection or maintenance interruption due to page changes.

[0018] Furthermore, the inspection execution unit is used to automatically execute cross-page inspection processes across multiple interconnected pages of the network management system based on the interface operation model and inspection item configuration. Specifically, when the inspection execution unit starts an inspection task for a certain type of network element, it establishes a corresponding session context for that type of network element to record information such as the page identifier of the current browser page, the content of the entered query conditions, and the time of the most recent submission of the query operation. By saving the session context, the inspection execution unit can determine whether the current operation is on the expected page in subsequent steps, and retry or terminate the current operation based on the recorded content when an anomaly occurs.

[0019] After submitting a query request, the inspection execution unit checks the page loading status within the configured waiting time and, in conjunction with the interface operation model provided by the interface parsing unit, verifies whether the interface structure parsed by the current page is consistent with the expected page structure. If the page loading is not completed after the waiting time has expired, or if the interface structure parsed by the current page is inconsistent with the structure recorded in the interface operation model, the inspection execution unit marks the operation as a failure and re-executes the corresponding query or page jump operation according to the pre-configured number of retries. In this way, in the event of slow server response or abnormal page jump, retries can be performed automatically without manual intervention.

[0020] When the number of retries for the same inspection item reaches the configured limit and the corresponding page operation still cannot be successfully completed, the inspection execution unit will no longer try. Instead, it will record the collection result corresponding to the inspection item as a failure and attach failure reason information to the original inspection data. The failure reason includes at least one of page timeout, page structure mismatch, or unavailable interface elements. When processing the original inspection data, the subsequent state fusion and analysis unit can distinguish between cases where no data was collected and cases where abnormal data was collected based on the failure mark and failure reason. This reduces the interference of abnormal pages on the overall operation status analysis. Through the above mechanism, this invention detects and controls page loading anomalies and page structure changes during cross-page automatic inspection, improving the stability of the inspection process and the reliability of the inspection result data.

[0021] Furthermore, the state fusion and analysis unit is used to construct an operational status view after obtaining multi-source operational data, to uniformly represent the operational status of network elements and support subsequent analysis. Specifically, the state fusion and analysis unit represents each physical network element or logical network element instance as a network element entity node, represents the performance indicators, alarm records, and configuration change information collected by the network element at different times as multiple status attribute nodes, and represents the fault work order records and maintenance operation records related to the network element as event nodes. Through the above representation method, the current status, historical changes, and faults and maintenance behaviors that have occurred can be reflected simultaneously in the same data structure.

[0022] After the nodes are established, the state fusion and analysis unit establishes time relationship edges between network element entity nodes and corresponding state attribute nodes according to time sequence. These edges represent the performance indicators, alarm status, and configuration status of the network element at a certain moment. At the same time, it establishes influence relationship edges between state attribute nodes and corresponding event nodes. These edges represent the correspondence between a fault event or maintenance event and the changes in state before and after the event. Through this structure, it is possible to distinguish whether a certain state is the basic state before the event or the result state after the event, providing a basis for subsequent trend analysis and correlation analysis.

[0023] To ensure the reliability of data in the operational status view, the status fusion and analysis unit adds a data source identifier and timestamp information to each status attribute node when writing indicator values ​​or configuring status. When inconsistent status attribute values ​​from different data sources are detected for the same network element within the same preset time period, the status fusion and analysis unit marks the status attribute node as an untrusted node and will not use the node's data in subsequent feature calculation and fault diagnosis. Through the above processing, in scenarios with multiple monitoring systems or data synchronization delays, inconsistent data can be avoided from being directly used for feature extraction, reducing the impact of abnormal data on operational status analysis and fault diagnosis results, and improving the credibility of effective data in the operational status view.

[0024] Furthermore, based on the operational status view, the anomaly diagnosis unit performs phased diagnostic processing on the network element's operational status. First, the anomaly diagnosis unit performs rule screening on the performance indicators, alarm information, and environmental monitoring data in the operational status view to determine the set of candidate fault types that need further analysis. Specifically, when a performance indicator exceeds the threshold configured for that indicator for a consecutive preset number of sampling times, the anomaly diagnosis unit adds the fault type corresponding to that performance indicator to the candidate fault type set. When a pre-configured alarm combination pattern (including multiple specific alarms appearing simultaneously or sequentially within a preset time sequence) is detected in the operational status view, the fault type corresponding to that alarm combination pattern is added to the candidate fault type set. When changes in temperature, humidity, or power supply parameters recorded in the environmental monitoring data meet pre-configured environmental anomaly conditions, the fault type corresponding to those environmental anomaly conditions is added to the candidate fault type set. Through the above rule screening, the anomaly diagnosis unit can first determine a portion of candidate fault types that meet preset anomaly characteristics from all network elements and all fault types, reducing the number of objects for subsequent model judgment.

[0025] After obtaining the set of candidate fault types, the anomaly diagnosis unit uses a classification model trained based on historical fault samples to further determine each candidate fault type. To this end, the anomaly diagnosis unit extracts data features related to the candidate fault types from the operation status view. The statistical features of performance indicators within a preset time window (including mean, maximum value, and change range, etc.), the time sequence features of alarm occurrence (including the order of alarm occurrence and duration, etc.), and the environmental parameter change features (including the change trends of temperature, humidity, and power supply parameters, etc.) are used as input features of the classification model. The classification model outputs the credibility of each candidate fault type based on the input features. The anomaly diagnosis unit sorts or filters the candidate fault types according to the credibility, and takes the fault types with credibility higher than the configured threshold as the suspected fault types in the diagnosis results.

[0026] By combining rule-based screening with a sample-based classification model, the anomaly diagnosis unit in this embodiment can quickly screen for obvious anomalies by using threshold exceeding limits, alarm combination patterns, and abnormal environmental conditions. On the other hand, it can comprehensively determine candidate fault types by using statistical features, time sequence features, and environmental change features, thereby reducing false alarm rate and false negative rate and improving the accuracy and stability of fault identification results. The above diagnosis process uses the running status view as a unified data input, and the input-output relationship between the rule screening stage and the model determination stage is clear, which is convenient for implementation and maintenance in the system.

[0027] Furthermore, the maintenance strategy generation unit is used to automatically generate maintenance strategies for the target network element based on the maintenance operation template library after receiving the suspected fault type output by the anomaly diagnosis unit. The maintenance operation template library is formed by organizing historical maintenance task records. After each maintenance task is completed, the system extracts the actual page operation steps executed during the maintenance process and organizes these steps into maintenance operation templates. For each maintenance operation template, the system records at least the following information: the operation type of the maintenance operation (e.g., parameter check, log view, configuration change, restart operation, etc.), the corresponding interface element identifier, the execution order of each operation in the same maintenance process, and the suspected fault type to which the maintenance operation is applicable. In this way, the experience of previous manual maintenance can be precipitated into structured templates, which can be directly reused when similar faults occur in the future, reducing the reliance on personal experience to arrange maintenance steps.

[0028] After identifying the suspected fault type of the target network element, the maintenance strategy generation unit selects the corresponding maintenance operation template from the template library based on the suspected fault type. The selected templates are further divided into inspection-type and change-type operation templates. Inspection-type operation templates mainly include operations that do not change the network element configuration or service carrying status, such as reading status, viewing alarms, and exporting logs. Change-type operation templates include operations that modify the network element configuration or affect the service forwarding status. The maintenance strategy generation unit arranges all inspection-type operation templates before all change-type operation templates and generates a list of operation steps in the maintenance strategy according to the execution order recorded in the templates. This ensures that the maintenance process completes necessary status checks and precondition confirmations before performing any configuration changes, which helps to identify situations that do not meet the change conditions early on and avoids invalid or even risky configuration operations.

[0029] While generating maintenance strategies, the maintenance strategy generation unit also assesses the business impact of each type of change maintenance operation. To this end, the maintenance strategy generation unit uses the current service type (e.g., voice service, data service, leased line service), service traffic, and service level carried by the target network element as input parameters to calculate an impact score for each type of change maintenance operation. The score reflects the scope of service interruption and the degree of service quality fluctuation that the operation may cause. When the impact score of a certain type of change maintenance operation is higher than the configured threshold, the maintenance strategy generation unit marks the operation step as a step that requires manual confirmation in the maintenance strategy. Before actual execution, maintenance personnel are required to conduct secondary confirmation or execute the operation during off-peak hours. Through the above generation method, the maintenance strategy generation unit in this embodiment can achieve operation sequence control of checking first and then changing based on the reuse of historical maintenance experience, and explicitly mark high-risk change steps, thereby improving the degree of maintenance automation while reducing the risk of impact on critical services.

[0030] Furthermore, the maintenance execution and knowledge update unit is used to verify the executability of the maintenance strategy before it is issued, and to update the maintenance-related parameters and maintenance operation templates during the long-term operation of the system. Specifically, after obtaining the maintenance strategy output by the maintenance strategy generation unit, the maintenance execution and knowledge update unit first calls the interface operation model in an independent test session to perform process simulation on the page operation sequence recorded in the maintenance strategy: page jump, interface element positioning and operation triggering are executed in sequence according to the order specified by the maintenance strategy. After each operation is completed, the maintenance execution and knowledge update unit verifies the existence of the target page identifier, interface element positioning result and expected key fields in the result page according to the current browser state, in order to determine whether the maintenance strategy is applicable under the current network management page structure. When any verification result does not meet the expectations, the maintenance execution and knowledge update unit terminates the subsequent operation of the test session and marks the maintenance strategy as unavailable under the current network management page version. The maintenance strategy will not be executed on the actual network element in the subsequent formal session, thereby avoiding erroneous operations caused by network management interface adjustments or interface element changes.

[0031] After the maintenance strategy is verified through a test session and executed in a formal session, the maintenance execution and knowledge update unit also summarizes and analyzes the execution logs of multiple inspection and maintenance tasks. Upon completion of each inspection and maintenance task, this unit records operational data related to that task, including inspection path length, page access count, task completion time, whether the maintenance task was ultimately successful, whether a rollback operation was performed, and whether similar faults recurred within a preset observation period after maintenance. Subsequently, based on the statistical data, the maintenance execution and knowledge update unit adjusts the order of maintenance operation templates in the maintenance operation template library, ensuring that they are correctly ranked in historical executions. Maintenance operation templates with higher success rates, fewer rollbacks, and lower fault recurrence rates have higher priority in subsequent maintenance strategy generation. Simultaneously, based on execution results, the threshold values ​​used to determine state anomalies in the state fusion and analysis unit are revised, and the classification thresholds of the classification model in the anomaly diagnosis unit are adjusted. This ensures that diagnostic and maintenance-related parameters are continuously updated according to network operation conditions, better aligning with current network and service characteristics. Through test session verification and parameter updates based on execution logs, the maintenance execution and knowledge update unit reduces the risk of maintenance strategy failures due to interface changes while improving the applicability of maintenance operation templates and diagnostic parameters in actual operation and maintenance scenarios.

[0032] Furthermore, when the maintenance execution and knowledge update unit determines that a rollback needs to be performed based on the rollback trigger conditions set in the maintenance strategy, the unit executes the corresponding page operation sequence sequentially according to the rollback steps pre-recorded in the maintenance strategy to restore the configuration of the target network element item by item. Specifically, before the maintenance task begins, the maintenance execution and knowledge update unit reads the current configuration of the target network element from the network management system, uses the values ​​of each configuration item as the baseline configuration for this maintenance task, and associates each configuration change with the corresponding baseline configuration in the maintenance strategy. After triggering the rollback operation, the maintenance execution and knowledge update unit sequentially calls the interface operation model to complete page navigation, interface element positioning, and operation triggering, restoring each configuration change that occurred during the maintenance task to the value recorded in the baseline configuration. After each rollback operation is completed, the current configuration value is read for verification to confirm that the restoration result is consistent with the baseline configuration. If the restoration of a certain configuration item fails during the rollback process, the reason for the failure is recorded, and subsequent rollback steps can be stopped at that configuration item to avoid introducing new anomalies due to erroneous restoration.

[0033] After the rollback operation is completed, the maintenance execution and knowledge update unit adds an event node corresponding to this rollback operation to the running status view. The event node records the rollback trigger time, the target network element involved, and the associated maintenance task identifier. At the same time, the performance indicators, alarm status, and configuration status collected after the rollback are completed are written into new status attribute nodes. These status attribute nodes are associated with the rollback event nodes through time relationship edges. In this way, when analyzing the running history of the same network element, the running status view can distinguish the pre-maintenance status, post-maintenance status, and post-rollback status, and determine whether the network status has recovered to the pre-maintenance level after a maintenance execution failure and rollback is triggered. In this embodiment, the event nodes, status attribute nodes, and the result of whether the rollback is successful are also included as part of the execution log for subsequent statistical analysis of the maintenance task effect. This is used in conjunction with the aforementioned maintenance strategy generation and parameter adjustment process to improve the reliability of the maintenance strategy in the actual network environment.

[0034] The AI-based intelligent inspection and maintenance system for communication networks described in this invention realizes automatic inspection and maintenance of communication networks according to a preset method and process. First, the system establishes a connection with the existing Web network management system. The page parsing unit performs structured parsing on the Web page output by the network management system, identifies interface elements such as menus, buttons, input boxes, and result tables and their hierarchical relationships, and establishes a mapping relationship between interface elements and specific operation types based on the event response methods of interface elements, forming an interface operation model, which provides a foundation for subsequent automated page operations.

[0035] After obtaining the interface operation model, the system configures the performance indicators, alarm items and configuration items to be collected by the inspection task configuration unit according to the different network element types and business scenarios. These inspection requirements are mapped to the combination of interface elements corresponding to the interface operation model to form the inspection item configuration. The inspection execution unit controls the browser session according to the interface operation model and the inspection item configuration, and completes operations such as login, menu jump, condition filling, query triggering and result page switching in a predetermined order. The system collects the running data of the target network element from multiple interconnected pages and organizes the collected results into inspection raw data corresponding to each network element.

[0036] Subsequently, the status fusion and analysis unit aligns the raw inspection data with historical inspection data, fault work order data, equipment history data, and environmental monitoring data along the time axis and network element identification dimension to construct an operational status view containing network element entities, status attributes, and event records. Based on this, it calculates feature data reflecting changes in operational status. The anomaly diagnosis unit takes the operational status view and its features as input. On the one hand, it uses pre-configured threshold rules, alarm combination modes, and environmental anomaly conditions to screen for situations that meet preset anomaly conditions, forming a set of candidate fault types. On the other hand, it inputs the statistical features of performance indicators, alarm time sequence features, and environmental parameter change features related to the candidate fault types into a classification model trained based on historical fault samples to calculate the credibility of each candidate fault type, thereby giving the anomaly level and one or more suspected fault types for each network element.

[0037] Based on the diagnostic results output by the anomaly diagnosis unit, the maintenance strategy generation unit selects maintenance operations corresponding to the suspected fault types from the maintenance operation template library formed by historical maintenance tasks. These maintenance operations are mapped to executable page operation sequences on the network management page and organized into maintenance strategies according to the principle of checking first and then changing. For change-type operations with significant business impact, the maintenance strategy generation unit also calculates an impact score based on the target network element's business type, business traffic, and service level. Operations with scores exceeding the configured threshold are marked as steps requiring manual confirmation in the maintenance strategy. Before the formal issuance of the maintenance strategy, the maintenance execution and knowledge update unit first calls the interface operation model in an independent test session to simulate the page operation sequence corresponding to the maintenance strategy. After each operation, it verifies the existence of page identifiers, interface element location results, and key fields. Only after the test session verification is successful is the maintenance strategy executed on the actual network element in the formal session.

[0038] During the formal execution of the maintenance strategy, the maintenance execution and knowledge update unit continuously monitors relevant performance indicators and alarm status. When the preset performance improvement conditions are not met within the preset observation time, or when a new high-priority alarm occurs, the interface operation model is invoked according to the pre-set rollback steps in the maintenance strategy. The configuration changes that occurred during the maintenance task are restored item by item to the configuration values ​​before the maintenance started. The corresponding rollback event and the status attributes after the rollback are recorded in the running status view. After each inspection and maintenance task is completed, the maintenance execution and knowledge update unit writes information such as inspection path length, page access count, task completion time, whether the maintenance task was successful, whether a rollback was performed, and whether the same type of fault recurred within the preset time period after maintenance into the execution log. Based on this, the unit adjusts the sorting of maintenance operation templates in the maintenance operation template library, as well as the threshold values ​​used to determine state anomalies in the status fusion and analysis unit and the classification thresholds of the classification model in the anomaly diagnosis unit. This ensures that the subsequent inspection process and maintenance strategy are continuously adjusted and optimized in the actual network environment.

[0039] The beneficial effects of this invention are as follows:

[0040] 1. This invention uses document object model parsing to analyze network management web pages, extracting interface elements such as navigation areas, condition input areas, and result display areas. It establishes an interface operation model independent of manufacturer protocols and records page identifiers and query conditions based on session context. This allows the inspection process to be independent of fixed script coordinates and jump paths, enabling the reuse of existing models even during network management version upgrades or page structure adjustments. This improves the adaptability and maintainability of inspection tasks across manufacturers and multiple versions. Furthermore, during cross-page inspections, it monitors page loading status and verifies page structure consistency. It records and marks the reasons for failures in inspections that cannot be completed, distinguishing between data collection failures and data anomalies, reducing human error and missed inspections, and decreasing the workload of frontline maintenance personnel who repeatedly log in and click.

[0041] 2. This invention unifies the representation of performance indicators, alarm records, configuration changes, fault work orders, and maintenance records into network element entity nodes, status attribute nodes, and event nodes, and establishes time relationship edges and influence relationship edges. This allows for a centralized representation of the evolution of network element operating status on the time axis, facilitating the backtracking of performance changes, alarm evolution, and the relationship before and after maintenance operations by time windows. Based on this, consistency checks are performed on status attribute values ​​from different acquisition channels, and inconsistent attribute nodes are marked as untrustworthy and removed during feature calculation, improving the data quality used for diagnosis. Furthermore, it combines rule screening based on performance thresholds, alarm combinations, and abnormal environmental conditions with credibility calculation based on classification models. While retaining existing operation and maintenance rule configuration habits, it introduces a comprehensive discrimination capability for multi-dimensional features, which helps reduce the probability of false alarms and missed alarms in complex network scenarios.

[0042] 3. This invention constructs a maintenance operation template library based on historical maintenance task records and organizes inspection and change operations according to suspected fault types. When generating maintenance strategies, inspection steps are uniformly arranged before change steps, ensuring that the maintenance process meets the basic requirement of checking before changing at the template level, reducing the risk of directly changing configurations. Combining the service type, service traffic, and service level of the target network element, a service impact score is calculated for change-type maintenance operations. Steps with scores higher than a threshold are marked with manual confirmation, enabling the maintenance strategy to distinguish the degree of impact of different service importance. In the execution phase, the page operation sequence is simulated and verified through test sessions. In the formal session, the rollback is performed according to the rollback step sequence and combined with the pre-maintenance configuration snapshot. The sorting of maintenance operation templates and related threshold settings are adjusted based on the execution log statistics, allowing the maintenance strategy and diagnostic parameters to be gradually adjusted according to the running effect, which helps to improve the reliability of maintenance execution and shorten fault handling time. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the core workflow of the system of this invention;

[0044] Figure 2 This is a flowchart of the abnormal diagnosis process of the present invention. Detailed Implementation

[0045] 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, and 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.

[0046] like Figures 1 to 2 As shown, this embodiment of the invention provides an AI-based intelligent inspection and maintenance system for communication networks, deployed in the integrated network maintenance environment of an operator. It communicates with existing web-based network management systems via standard network interfaces, enabling centralized operation and maintenance management of communication networks (e.g., backbone networks, transmission networks, wired broadband, and enterprise private lines) including network elements from multiple manufacturers and models. The system includes a page parsing unit, an inspection task configuration unit, an inspection execution unit, a status fusion and analysis unit, an anomaly diagnosis unit, a maintenance strategy generation unit, and a maintenance execution and knowledge update unit. Each unit interacts with data through internal communication interfaces to jointly complete the automatic inspection and maintenance process of network element operating status.

[0047] Specifically, the page parsing unit is used to receive the Web page output by the network management system, parse the document object model of the page, identify the interface elements such as menus, buttons, input boxes, and result tables and their hierarchical relationships, and establish a mapping relationship between interface elements and specific operation types in combination with the event response methods of interface elements, forming an interface operation model that is independent of the network element manufacturer's interface protocol. Based on this interface operation model, subsequent units can drive page operations on different manufacturers' and different versions of network management interfaces in a unified manner without modifying the existing network element interfaces, reducing the reliance on manually recorded scripts and one-by-one adaptation;

[0048] The inspection task configuration unit configures the performance indicators, alarm items, and configuration items to be collected according to different network element types and business scenarios (such as backbone transmission equipment, access equipment, and government and enterprise leased line equipment). It maps these inspection requirements to a combination of interface elements corresponding to the interface operation model, forming a structured inspection item configuration. The inspection execution unit controls the browser session based on the interface operation model and the inspection item configuration, and completes operations such as login, menu selection, query condition filling, query triggering, and result page switching in sequence according to the predetermined process. It collects the running data of the target network element from at least two related pages, and organizes the collected data into the inspection raw data corresponding to each network element, providing a unified data source for subsequent analysis.

[0049] The status fusion and analysis unit is used to process multi-source operational data in a unified manner. It aligns the raw inspection data with historical inspection data, fault work order data, equipment history data, and environmental monitoring data on the time axis and network element identification dimension to construct an operational status view that includes network element entities, status attributes, and event records. Based on the operational status view, the anomaly diagnosis unit performs rule-based judgment on the operational status according to pre-configured threshold rules, alarm combination modes, and environmental anomaly conditions to filter out network elements that need to be focused on and candidate fault types. On the other hand, it uses a classification model trained based on historical fault samples to calculate the statistical features of performance indicators, alarm time sequence features, and environmental parameter change features extracted from the operational status view, comprehensively judges the anomaly level of the target network element, and provides one or more suspected fault types.

[0050] After obtaining the anomaly diagnosis results, the maintenance strategy generation unit selects maintenance operations corresponding to the anomaly level and suspected fault type from the maintenance operation template library compiled from historical maintenance tasks. These maintenance operations are mapped to a sequence of page operations that can be executed step by step on the network management page, and organized into a maintenance strategy according to the principle of checking first and then changing. This ensures that status verification and precondition confirmation are completed first during the maintenance process, reducing the risk of directly changing the configuration. The maintenance execution and knowledge update unit is used to simulate the process of the page operation sequence corresponding to the maintenance strategy in the test session by calling the interface operation model in the test session before the formal execution of the maintenance strategy. After each operation, the page identifier, interface element location result, and key fields are verified to see if they exist. Only if the test session verification is successful is the maintenance strategy executed on the actual network element in the formal session. During the formal execution, this unit monitors the relevant performance indicators and alarm changes in the operation status view. When the preset rollback conditions are met, the configuration changes that occurred during the maintenance period are restored in sequence according to the rollback steps pre-recorded in the maintenance strategy, and the rollback process and the status after the rollback are recorded in the operation status view and execution log.

[0051] During long-term system operation, the maintenance execution and knowledge update unit performs statistical analysis on the execution logs of multiple inspection and maintenance tasks. This analysis includes indicators such as the success rate of maintenance operation templates in different network scenarios, whether rollbacks were performed, and whether similar faults recurred within a preset time period after maintenance. Based on these indicators, the priority of maintenance operation templates in the maintenance operation template library is adjusted, giving higher priority to maintenance operation templates with stable historical performance when generating subsequent maintenance strategies. Simultaneously, based on the execution results, the threshold values ​​used to determine state anomalies in the state fusion and analysis unit and the classification thresholds of the classification model in the anomaly diagnosis unit are updated. This allows diagnostic parameters and maintenance strategies to be adjusted according to actual network operation conditions, maintaining relatively stable diagnostic and maintenance effects even when network topology and service load change.

[0052] In one specific embodiment, the page parsing unit is used to convert the operator's existing Web network management interface into an interface operation model that is independent of the vendor interface, so as to support the automatic operation of the page by the subsequent inspection execution unit and maintenance execution and knowledge update unit. The page parsing unit is deployed on a server with browser controls or browser drivers, establishes a session with the Web front-end of the network management system, and after the target page is loaded, parses the document object model (DOM) of the current page and generates the interface operation model.

[0053] Specifically, the page parsing unit first obtains the DOM tree structure of the current web page from the browser controls. Each node in the DOM tree contains at least a node type (e.g., div, a, input, table, tr, td), text content, a set of attributes (e.g., id, class, name, onclick), and its hierarchical position in the tree. Based on the node type, hierarchical position, and style information, the page parsing unit divides the page into three functional areas: a navigation area, a conditional input area, and a result display area. Node sets located at the top or side of the page, containing menu items or tree structures, and capable of triggering the display of different page content are identified as the navigation area; node sets containing text input boxes, dropdown lists, radio buttons, checkboxes, etc., used to set query conditions or parameters are identified as the conditional input area; and node sets containing tables, lists, or graphics used to display query results are identified as the result display area. The results of the area division are recorded in the interface operation model in the form of a mapping relationship between area identifiers and corresponding node sets.

[0054] After completing the area division, the page parsing unit analyzes the event handling functions bound to each node in the DOM tree. For nodes with onclick, onchange, and onsubmit event attributes, the page parsing unit determines whether the node will perform a page jump, initiate a data query, or submit a configuration change when triggered by parsing the event attribute values ​​and associated script functions. For example, if the onclick logic of a button node contains code that sends a query request to a specific URL and refreshes the table in the results area, the button is marked as a data query operation type; if the logic contains operations that modify the page location or expand the collapsible menu, the node is marked as a page jump operation type; if the logic contains code that sends a configuration change request to the backend, the node is marked as a configuration submission operation type. For nodes whose events are bound only by external scripts and are difficult to judge based on attribute values ​​alone, the page parsing unit performs a trial click on the suspicious node in a controlled test session and detects changes in the page URL, DOM structure, or results area content to help determine the actual operation type of the node.

[0055] After identifying the interface elements with operational significance, the page parsing unit generates a stable positioning path for each type of interface element and establishes a correspondence between interface element identifiers and operation types in the interface operation model. The positioning path can use XPath expressions, CSS selectors, or custom paths based on node hierarchy and attribute combinations. To improve the stability of positioning, the page parsing unit prioritizes using relatively fixed identifier attributes (such as id, name, and business-meaning class) in the page when generating positioning paths, and avoids simply relying on absolute hierarchy numbers by combining relative positional relationships. When an interface element lacks a stable identifier, it combines features such as its region, the text content of its sibling nodes, and column headers as the positioning basis. Finally, the interface operation model records at least: interface element identifier, corresponding page identifier, region type, positioning path, and operation type (e.g., page jump, data query, configuration submission, text entry, dropdown selection). This allows subsequent modules to complete page positioning and operation triggering based solely on the interface element identifier when an operation needs to be performed, without relying on operation scripts arranged in a fixed order.

[0056] Considering that the page structure of the operator's network management system may change during version upgrades or interface adjustments, the page parsing unit also implements page structure change detection and model update functions to ensure consistency between the interface operation model and the actual page. Specifically, after the page parsing unit completes the parsing of a page for the first time, it generates a structural summary of the page's DOM tree. The structural summary records at least the node type distribution of each functional area, the hierarchical position of key nodes, and several representative text features. It can also calculate a hash value for this summary information and save it as a feature of the current page version. In subsequent runs, whenever the same page is loaded, the page parsing unit re-obtains the DOM tree and generates a new structural summary. The page parsing unit determines whether the page structure has changed by comparing the differences in node types, hierarchical positions, and key text content between the old and new structural summaries. When the difference exceeds a preset threshold, or the original positioning path cannot correctly locate the corresponding element on the new page, the page parsing unit considers the page structure to have changed and triggers a re-parsing process: re-execute the region division, event detection, and positioning path generation, and write the new interface element positioning information and operation type into the interface operation model; at the same time, mark the positioning path of the corresponding page in the old version as invalid, only retain its record, and no longer use it for subsequent automatic operations. The preset threshold can be configured according to the operator's sensitivity to interface changes and the frequency of historical changes.

[0057] Throughout the parsing and updating process, the page parsing unit does not store the operation script sequence arranged in a fixed order. Instead, it stores the interface operation model that is relatively decoupled from the page structure. When the inspection execution unit and the maintenance execution and knowledge update unit are actually running, they query the interface operation model to obtain the location path and operation type of the current version page based on the interface element identifiers referenced in the inspection item configuration and maintenance strategy. Then, they generate specific operation steps in combination with the session context at that time. In this way, when the network management page adds fields, adjusts the layout, or changes the theme style, the page parsing unit can automatically update the interface operation model to ensure that the subsequent inspection and maintenance process continues to run. There is no need to frequently modify and replay the operation scripts as in the existing solution that relies on fixed scripts.

[0058] The page parsing unit in this embodiment has the following characteristics:

[0059] Based on the DOM tree structure, the page is divided into regions and elements are identified, so that the interface operation model is relatively decoupled from the specific HTML layout and display style. When the page layout is adjusted but the functional areas and element semantics remain unchanged, there is no need to rewrite the entire operation logic.

[0060] By analyzing event handling functions and testing exploratory operations in the session, we can identify operation types such as page jumps, data queries, and configuration submissions, enabling the same interface operation model to adapt to different implementation methods in multiple manufacturers' and versions of network management systems.

[0061] By comparing page structure summaries and detecting the validity of location paths, the interface operation model is automatically updated, enabling the system to restore its inspection and maintenance capabilities after network management version upgrades or interface changes, thus reducing manual intervention and script maintenance workload.

[0062] In one specific embodiment, the inspection execution unit is used to perform cross-page inspections between multiple interconnected pages of the network management system based on the interface operation model and inspection item configuration. To avoid losing status information during page jumps, when the inspection execution unit starts an inspection task for a certain type of network element, it establishes a session context data structure for that type of network element. The session context is indexed by the network element type and network element identifier, and records at least the page identifier of the current browser session, the filled-in query conditions, and the timestamp of the most recent query submission. It is updated after each page jump, condition filling, and query submission so that the session context reflects the page position and query condition status of the current inspection process.

[0063] After submitting a query request, the inspection execution unit enters the page loading monitoring phase. Based on the waiting time parameter configured by the operation and maintenance department, the inspection execution unit periodically checks the page loading status within this time period and determines whether the page has finished loading by combining information such as the loading flag of the browser control and whether the DOM nodes in the result display area are ready. At the same time, it calls the page parsing unit to perform a brief parsing of the DOM structure of the current page, compares the obtained page identifier and the structural information of the main functional areas with the structural summary recorded in the interface operation model. If the page has not finished loading after the waiting time has expired, or if the parsed interface structure is significantly inconsistent with the interface operation model, the inspection execution unit marks the operation as a failure and re-executes the corresponding page jump or query submission operation according to the pre-configured number of retries. During the retry, the inspection execution unit restores the operation scene using the page identifier and query conditions recorded in the session context to avoid repeated login or re-entry of conditions.

[0064] When the number of retries for the same inspection item reaches the configured limit and the corresponding page operation still cannot be successfully completed, the inspection execution unit stops further attempts on the inspection item, records the collection result corresponding to the inspection item as a failure status, and writes the failure reason information into the inspection raw data. The failure reason information includes at least the page identifier where the failure occurred, the operation type at the time of failure, the timestamp of the last attempt, and the failure reason category (page timeout, page structure mismatch, target interface element cannot be located, or server returned an error, etc.). When the status fusion and analysis unit processes the inspection raw data in the subsequent process, it can distinguish between the case of no data collection and the case of abnormal data collection based on the failure mark and failure reason.

[0065] In this embodiment, the inspection execution unit always performs page operations and structural checks based on the interface operation model during cross-page inspection, without relying on pre-recorded fixed script sequences. On the one hand, by recording page identifiers, query conditions, and operation times through session context, fine-grained control of the inspection status of individual network elements can be achieved in multi-page, multi-step inspection processes. On the other hand, by combining page loading status detection and interface structure consistency verification, and setting a limited number of automatic retries, the impact of network fluctuations, short-term server anomalies, or partial page adjustments on the inspection process is limited to a single inspection item. When certain inspection items cannot be completed, the reason for failure is explicitly written into the original inspection data, providing a basis for subsequent diagnosis. Compared with existing solutions that rely on sequential scripts and only provide task failure results after timeout, this embodiment can improve the stability and maintainability of the inspection process in a multi-vendor, multi-page Web network management environment.

[0066] In one specific embodiment, the state fusion and analysis unit is used to construct an operational status view based on multi-source operational data, and to uniformly represent the operational status and historical evolution process of each network element. Specifically, the state fusion and analysis unit uses the network element identifier in the network management system as the primary key, and represents each physical network element or logical network element instance as a network element entity node; for the same network element, the performance indicators, alarm records, and configuration change information collected from the original inspection data, historical inspection data, and equipment history data at different time points are represented as multiple state attribute nodes; maintenance task records from the fault work order system and the maintenance execution and knowledge update unit are represented as event nodes associated with the network element. Each state attribute node records at least the network element identifier, attribute type (performance indicator, alarm, or configuration change), attribute value, collection time, and data source identifier; each event node records at least the network element identifier, event type (fault work order or maintenance operation), event occurrence time, and event number.

[0067] After creating network element entity nodes, status attribute nodes, and event nodes, the state fusion and analysis unit establishes relationship edges between nodes according to time sequence. For the same network element, the network element entity node is connected to the status attribute node at the corresponding time point on the time axis to construct a time relationship edge, which is used to represent the status information of the network element at a certain moment. For fault work orders and maintenance operation records related to the network element, the state fusion and analysis unit establishes an influence relationship edge between the status attribute nodes adjacent to the event time according to the event occurrence time, which is used to represent the relationship between the state change and the event. The time relationship edge and the influence relationship edge can be recorded in the data structure through an adjacency list or similar structure, and a timestamp and relationship type are attached to the edge to support the backtracking of state changes by time window and the retrieval of related states by event.

[0068] To improve the reliability of data in the operational status view, the status fusion and analysis unit adds a data source identifier and timestamp to each status attribute node when writing data. This is used to distinguish between periodic inspection tasks, real-time alarm reports, device log analysis results, and other monitoring channels. Within a preset time period, the status fusion and analysis unit compares status attribute nodes of the same network element and the same attribute type. When the same status attribute value from different data sources is inconsistent within this time period, the corresponding status attribute node is marked as an untrusted node, and the untrusted flag and conflict source information are recorded in the node attribute. When calculating statistical features for anomaly diagnosis in subsequent calculations, the status fusion and analysis unit skips the status attribute nodes marked as untrusted and only uses the data marked as trustworthy for statistics, reducing the impact of data asynchrony or collection anomalies on the diagnostic results.

[0069] The operational status view constructed in the above manner includes the performance, alarms, and configuration changes of network elements on the time axis, as well as event information related to fault work orders and maintenance operations. It also records the data source and credibility at the node level. This embodiment adopts a data organization method of network element entity node - status attribute node - event node + relationship edge + untrusted marker. Compared with the method of recording monitoring data and work order records only in a flat table, it is easier to retrieve the network element status by time and event dimensions in subsequent anomaly diagnosis, and complete the filtering of abnormal data in the data fusion stage.

[0070] In one specific embodiment, the anomaly diagnosis unit performs rule screening and model evaluation on the operating status of network elements based on the operating status view to determine the anomaly level and candidate fault type. Specifically, the anomaly diagnosis unit first selects a preset time window corresponding to the target network element in the operating status view, obtains the performance index attribute node, alarm attribute node and associated environmental parameter attribute node of the network element within the time window, and performs candidate fault type screening on this basis.

[0071] In terms of performance metrics, the anomaly diagnosis unit extracts the sequence of measured values ​​of each pre-configured key performance metric from the running status view in chronological order within the sampling period. It then determines whether the metric continuously exceeds the threshold configured for the metric within a preset number of consecutive sampling moments. When a performance metric is detected to exceed the corresponding threshold for Nseq consecutive sampling moments (Nseq is configured by the operations and maintenance personnel according to the monitoring period and business characteristics), the anomaly diagnosis unit adds the fault type associated with the performance metric to the candidate fault type set and internally records the name of the performance metric that triggered the fault type and the time interval of the exceedance.

[0072] In terms of alarm information, the anomaly diagnosis unit restores the alarm occurrence of the network element within the time window in chronological order according to the alarm attribute node sequence in the operation status view, and matches it with the pre-configured alarm combination mode. The alarm combination mode is provided by the operation and maintenance rule configuration module. Each mode includes a set of alarm types and their allowed time sequence and time interval range. When the anomaly diagnosis unit detects that there is an alarm sequence in the operation status view that matches a certain alarm combination mode, it adds the fault type corresponding to the alarm combination mode to the candidate fault type set.

[0073] In terms of environmental parameters, the anomaly diagnosis unit extracts environmental monitoring data associated with the data center or cabinet where the target network element is located from the operation status view, including temperature, humidity, and power supply parameters. It calculates the changes of these parameters within a time window. For environmental anomaly conditions pre-configured by maintenance personnel (such as temperature continuously exceeding the threshold for a certain period of time, humidity significantly increasing in a short period of time, and power supply voltage dropping multiple times in a short period of time), when the environmental parameter changes are detected to meet the corresponding environmental anomaly conditions, the anomaly diagnosis unit adds the fault type corresponding to the environmental anomaly condition to the candidate fault type set. Through three categories of rules—performance threshold exceeding limits, alarm combination pattern matching, and environmental anomaly conditions—the anomaly diagnosis unit selects a set of candidate fault types from all predefined fault types for subsequent model evaluation.

[0074] After obtaining the set of candidate fault types, the anomaly diagnosis unit calls the pre-deployed classification model to calculate the credibility of each candidate fault type. To do this, the anomaly diagnosis unit extracts feature data related to the candidate fault types from the operation status view and organizes them into the input feature vector of the classification model. The input features include at least: statistical features of performance indicators (mean, extreme values, and change range within a time window), time sequence features of alarm occurrence (sequential relationship between different alarms, duration, number of repetitions, etc.), and environmental parameter change features (trend values ​​and fluctuation range of temperature, humidity, and power supply parameters, etc.). The anomaly diagnosis unit concatenates the above features into a fixed-dimensional feature vector in a preset order and inputs it into the classification model. The classification model outputs the credibility score of each candidate fault type and determines the final suspected fault type and corresponding anomaly level based on the score results.

[0075] Through the above implementation, when performing fault diagnosis, the anomaly diagnosis unit first uses threshold rules, alarm combination modes, and abnormal environmental conditions to filter candidate fault types from the operation status view, and then evaluates the candidate fault types based on the classification model. On the one hand, the rule screening stage uses existing operation and maintenance rules to quickly identify situations with obvious abnormal characteristics; on the other hand, the classification model performs comprehensive calculations on multi-dimensional features to distinguish different fault types with similar features. Compared with the solution that only relies on a single performance threshold or simple alarm rules, this embodiment improves the reliability of fault diagnosis results while ensuring controllable implementation complexity through the combination of "rule screening + model evaluation".

[0076] In one specific embodiment, the maintenance strategy generation unit is used to automatically generate a maintenance strategy for the target network element based on the maintenance operation template library after obtaining the suspected fault type output by the anomaly diagnosis unit. To this end, the system organizes the executed maintenance tasks into a maintenance operation template library during daily operation and maintenance. Specifically, for each historical maintenance task record, the maintenance execution and knowledge update unit extracts the actual page operation steps executed during the maintenance process from the execution log, organizes a continuous set of page operations into a maintenance operation template, and records at least the following in the template: the operation type of the maintenance operation, the corresponding interface element identifier, the execution order in the same maintenance process, and the applicable suspected fault type. The operation type is used to indicate the functional nature of the operation on the network management page, such as status check, log viewing, parameter query, configuration modification, or port restart; the corresponding interface element identifier corresponds to the interface element identifier in the interface operation model; the execution order is used to restore the order of operations within the same template; the applicable suspected fault type field is marked by the operation and maintenance personnel after the task is completed in combination with the fault cause, or it is automatically filled by the system according to the confirmed fault type. In the above way, the maintenance operation template library stores the operation steps in previous maintenance tasks in a structured manner, providing a reusable operation sequence for similar faults in the future;

[0077] When generating maintenance policies, the maintenance policy generation unit first retrieves a set of maintenance operation templates applicable to the suspected fault type from the maintenance operation template library based on the suspected fault type given by the anomaly diagnosis unit. Then, the maintenance policy generation unit classifies the retrieved maintenance operation templates into inspection operation templates and change operation templates according to the operation type field. Inspection operation templates are maintenance operations that do not change the network element configuration and service carrying status, such as reading the running status, viewing alarms, exporting logs, and querying the current configuration. Change operation templates are maintenance operations that modify the network element configuration or have a direct impact on service carrying, such as modifying parameters, closing or opening ports, switching routes, or issuing new configurations. The maintenance policy generation unit arranges all inspection operation templates before all change operation templates in the maintenance policy and sorts them within each category according to the execution order field recorded in the template, forming an operation sequence of inspection first and change later, so that status verification and precondition confirmation are completed first during the maintenance process.

[0078] When determining the impact of each change-type operation on the business in the maintenance strategy, the maintenance strategy generation unit obtains the current business type, business traffic, and service level of the target network element from the business support system or network element configuration. Using the above information as input parameters, it calculates the impact score for each change-type maintenance operation. Specifically, the maintenance strategy generation unit configures weight coefficients according to the business type, configures impact factors according to the business traffic and service level, and calculates the impact score in combination with the operation type and scope of the change operation. For each change-type maintenance operation, the maintenance strategy generation unit compares the calculated impact score with the impact threshold pre-configured by the operation and maintenance personnel. When the impact score of a certain change-type maintenance operation is higher than the configured threshold, the operation step is marked as a step that requires manual confirmation in the generated maintenance strategy. Before the maintenance execution and knowledge update unit actually executes the maintenance strategy, the system requires the operation and maintenance personnel to manually confirm the marked steps or select an appropriate execution time window to manually control high-impact operations.

[0079] Through the above implementation, the maintenance strategy generation unit, on the one hand, utilizes maintenance operation templates organized by suspected fault type in the maintenance operation template library to reduce the workload of manually writing maintenance processes for various faults; on the other hand, by dividing maintenance operations into inspection operations and change operations and arranging them in sequence, and by calculating the impact score of change operations and marking high-impact steps according to business type, business traffic and service level, the generated maintenance strategy can differentiate and control change risks based on the reuse of historical maintenance experience and the current business carrying capacity.

[0080] In one specific embodiment, the maintenance execution and knowledge update unit is used to simulate and verify the page operation process before the maintenance strategy is officially issued, and to update the maintenance operation template and diagnostic parameters based on the execution logs of inspection tasks and maintenance tasks during the long-term operation of the system.

[0081] Specifically, during process simulation in the test session, the maintenance execution and knowledge update unit generates the maintenance strategy based on the maintenance strategy output by the maintenance strategy generation unit, reads the page operation sequence recorded in the maintenance strategy, and drives the browser to complete page navigation, interface element location, and operation triggering in sequence. After each operation is executed, the maintenance execution and knowledge update unit obtains the page identifier and interface structure information of the current page from the page parsing unit, verifies whether the current page identifier is consistent with the expected page identifier in the maintenance strategy, and calls the interface operation model to locate the interface elements involved in this operation and checks whether the interface element location result is valid. For key fields that should appear on the result page (such as running status fields, confirmation prompt fields, execution result fields, etc.), the maintenance execution and knowledge update unit checks whether these fields exist on the current page. When any check fails, the maintenance execution and knowledge update unit terminates the subsequent operations of the test session, marks the maintenance strategy as unavailable in the current network management page version, records the failed step number and the reason for failure, and does not execute the maintenance strategy on the actual network element in the formal session.

[0082] For maintenance strategies that pass simulation in test sessions and are executed in formal sessions, the maintenance execution and knowledge update unit statistically analyzes the execution status of multiple inspection and maintenance tasks. For each inspection and maintenance task, the maintenance execution and knowledge update unit extracts operational metrics such as inspection path length, page access count, and task completion time from the execution process. The inspection path length can be determined according to the number of different pages accessed in the task, the page access count represents the total number of times page loading is triggered in the task, and the task completion time is determined by the task start time and task end time. At the same time, the maintenance execution and knowledge update unit records the execution result (success or failure) of each maintenance task, whether a rollback operation is performed, and whether the same type of fault reappears within the preset observation period after the maintenance is completed. The above information is stored in the form of an execution log, associated with the maintenance operation template identifier, maintenance strategy identifier, and target network element identifier.

[0083] After the system has been running for a period of time, the maintenance execution and knowledge update unit aggregates and statistically analyzes the execution log data obtained from the statistics, according to the maintenance operation templates. This yields indicators such as the number of calls, success rate, rollback ratio, and task completion time distribution for each maintenance operation template in different scenarios. Based on these statistical indicators, the maintenance execution and knowledge update unit adjusts the sorting of maintenance operation templates in the maintenance operation template library. This prioritizes maintenance operation templates with higher success rates and lower rollback ratios in multiple executions, while moving maintenance operation templates with lower success rates or those that frequently trigger rollbacks to the back of the list. These templates are then prioritized by the maintenance strategy generation unit when generating maintenance strategies in the future.

[0084] Meanwhile, the maintenance execution and knowledge update unit also adjusts the threshold values ​​used to determine state anomalies in the state fusion and analysis unit and the classification thresholds of the classification model in the anomaly diagnosis unit based on the inspection results, maintenance results, and subsequent fault recurrence recorded in the execution log. When a certain threshold is set too low, causing a large number of alarms that have been confirmed by maintenance to have no actual faults to be triggered, the maintenance execution and knowledge update unit will appropriately increase the threshold of the corresponding indicator. When the deviation of an indicator for a certain type of fault before multiple maintenance is significantly greater than the current threshold setting, the maintenance execution and knowledge update unit will adjust the threshold or classification threshold accordingly to make subsequent diagnosis more sensitive to this type of anomaly. Through the above parameter adjustments based on the execution log, the judgment parameters used in the state fusion and analysis unit and the anomaly diagnosis unit can be updated according to the network operation characteristics and maintenance effects.

[0085] Through the above implementation, the maintenance execution and knowledge update unit verifies the page operation sequence through a test session before the maintenance strategy is executed, so as to avoid applying maintenance strategies that do not match the current network management page structure to the formal session; after the maintenance strategy is executed, the maintenance operation template sorting and diagnostic thresholds are adjusted through the execution log, so that the maintenance process and diagnostic parameters can be updated according to the actual operation.

[0086] In one specific embodiment, when pre-configured rollback conditions are met, the maintenance execution and knowledge update unit triggers a rollback operation. To ensure the recoverability of the rollback process, before the maintenance task begins, the maintenance execution and knowledge update unit reads the current configuration values ​​of the configuration items to be involved from the network management system according to the change-type operation steps recorded in the maintenance strategy, and generates an initial configuration snapshot in the format of "network element identifier – configuration item identifier – configuration value". This snapshot is then associated with and stored with the unique identifier of this maintenance task. When forming the maintenance strategy, the maintenance strategy generation unit simultaneously generates corresponding rollback steps for each change-type operation and records them in the rollback step sequence in the maintenance strategy, ensuring a one-to-one correspondence between the rollback step sequence and the configuration items in the initial configuration snapshot.

[0087] When a rollback operation is triggered, the maintenance execution and knowledge update unit reads the rollback step sequence recorded in the maintenance strategy and executes the corresponding page operations sequentially according to the sequence. Specifically, the maintenance execution and knowledge update unit drives the browser based on the interface operation model to complete the page jump, interface element positioning, and operation triggering corresponding to the rollback step. It restores each change made to network element configuration items during the maintenance task to the configuration values ​​recorded in the network management system before the start of the maintenance task. For each rollback step, the maintenance execution and knowledge update unit retrieves the baseline configuration value of the configuration item before the maintenance started by searching for the corresponding configuration item record in the initial configuration snapshot. This value is then used as the target configuration value for this rollback operation, rather than simply inferring the "reverse operation" based on the direction of the change, to avoid inconsistencies between the rollback result and the original configuration. When executing a rollback operation, if a configuration item has been modified multiple times during maintenance, the maintenance execution and knowledge update unit still uses the value in the initial configuration snapshot as the restoration target, ensuring that the configuration state after rollback is consistent with the state before maintenance.

[0088] After completing the page operations corresponding to the rollback sequence, the maintenance execution and knowledge update unit collects post-rollback operational status data through the inspection execution unit or network management interface, and incorporates the rollback process into the operational status view record. Specifically, the maintenance execution and knowledge update unit creates a rollback event node in the operational status view for the network element that triggered the rollback. This rollback event node records at least the network element identifier, rollback trigger time, associated maintenance task identifier, and rollback execution result. Simultaneously, within a preset time point or time window after the rollback is completed, the collected performance indicators, alarm status, and key configuration item status are written into new status attribute nodes, and these status attribute nodes are associated with the corresponding network element entity nodes and rollback event nodes through time relationships. Through this recording method, the operational status view can distinguish between the pre-maintenance status, post-maintenance status, and post-rollback status, providing a data foundation for subsequent evaluation of the effectiveness of maintenance strategies and threshold adjustments.

[0089] Through the above implementation, the maintenance execution and knowledge update unit, when performing rollback operations, relies on the rollback step sequence in the maintenance strategy and the configuration snapshot before maintenance to achieve item-by-item restoration by configuration item dimension. Simultaneously, rollback events and post-rollback status changes are uniformly incorporated into the operational status view management, making rollback behavior and effects traceable. Those skilled in the art can implement the rollback function of the maintenance execution and knowledge update unit in the operator's live network environment according to the rollback step sequence generation method, configuration snapshot recording method, and the method of recording event nodes and status attribute nodes in the operational status view described in the specification.

[0090] In one specific embodiment, the AI-based intelligent inspection and maintenance method for communication networks provided by the present invention is executed by the aforementioned system, and is jointly implemented by a page parsing unit, an inspection task configuration unit, an inspection execution unit, a state fusion and analysis unit, an anomaly diagnosis unit, a maintenance strategy generation unit, and a maintenance execution and knowledge update unit, specifically including the following steps:

[0091] First, after the network management system establishes a web session, the page parsing unit parses the web page output by the network management system, obtains the document object model tree structure of the page from the browser controls, identifies interface elements such as navigation areas, condition input areas, and result display areas and their hierarchical relationships, analyzes the event handling functions of the interface elements, determines the interface elements that can trigger page jumps, data queries, or configuration submissions, generates stable positioning paths for these interface elements, establishes the correspondence between interface element identifiers and operation types, and forms an interface operation model that is independent of specific vendor interface protocols.

[0092] Secondly, based on the network element type and service scenario configured on the operation and maintenance side, the inspection task configuration unit selects the performance indicators, alarm items and configuration items to be inspected from the predefined indicator library, maps the above indicators to the combination of interface elements in the interface operation model, and forms the inspection item configuration for different network element types. The inspection task configuration unit can set different inspection item sets and inspection cycles according to different scenarios such as core network, bearer network or access network, and save them in the form of task templates.

[0093] Then, the inspection execution unit controls the browser session to complete cross-page operations based on the interface operation model and inspection item configuration. The inspection execution unit establishes a session context for each type of network element, records the current page identifier, the entered query conditions, and the time of the most recent query submission. It drives page navigation, condition filling, and query triggering according to the combination of interface elements in the inspection item configuration. During the configured waiting time, it monitors the page loading status and verifies whether the current page structure is consistent with expectations based on the interface operation model. When the page loads successfully and the structure matches, the unit collects the operation data of the target network element from the result page to generate the raw inspection data. If an inspection item cannot be completed after multiple retries, the item is marked as failed, the reason for failure is recorded, and the item is written into the raw inspection data along with the successfully collected operation data.

[0094] Subsequently, the state fusion and analysis unit integrates the raw inspection data with historical inspection data, fault work order data, equipment history data, and environmental monitoring data. Using the network element identifier as the primary key, the unit represents each network element instance as a network element entity node, performance indicators, alarm records, and configuration changes collected at different time points as state attribute nodes, and fault work orders and maintenance task records as event nodes. It establishes time relationship edges between entity nodes and state attribute nodes, and influence relationship edges between state attribute nodes and event nodes. Simultaneously, it adds data source identifiers and timestamps to state attribute nodes, performs consistency checks on the same attribute values ​​from different sources within the same time period, marks inconsistent state attribute nodes as untrusted nodes, and removes them during subsequent feature calculations. This constructs an operational state view containing time evolution relationships and data credibility information, and on this basis, calculates performance indicator statistical characteristics, alarm time sequence characteristics, and environmental parameter change characteristics.

[0095] After completing the construction of the operation status view and feature calculation, the anomaly diagnosis unit uses the above view and features to perform a two-stage diagnosis. In the first stage, based on the preset performance threshold rules, alarm combination patterns and environmental anomaly conditions, the operation status view is screened according to rules. Fault types that meet the conditions of continuous over-limit, matching alarm combinations, or meeting environmental anomaly conditions are added to the candidate fault type set. In the second stage, for the candidate fault types, the performance index statistical features, alarm time sequence features, and environmental parameter change features within the corresponding time window are extracted from the operation status view to form the input feature vector of the classification model. The pre-trained classification model outputs the confidence score of each candidate fault type, and the final suspected fault type and corresponding anomaly level are determined based on the score results.

[0096] After obtaining the suspected fault type and anomaly level, the maintenance strategy generation unit selects the corresponding maintenance operation template from the maintenance operation template library. The maintenance operation template library is formed based on historical maintenance task records. Each template records at least the operation type, corresponding interface element identifier, execution order, and applicable suspected fault type of the maintenance operation. The maintenance strategy generation unit divides the templates into inspection operation templates and change operation templates according to the operation type. The inspection operation is arranged before the change operation, and sorted in the execution order within each category to form a maintenance operation sequence of inspection first and change later. At the same time, the maintenance strategy generation unit calculates the service impact score for each change maintenance operation based on the service type, service traffic, and service level of the target network element. Operation steps with scores higher than the configured threshold are marked as steps that require manual confirmation, and a maintenance strategy containing the operation sequence and manual confirmation mark is generated.

[0097] Finally, in the maintenance execution phase, the maintenance execution and knowledge update unit first simulates the page operation sequence corresponding to the maintenance strategy in the test session. The unit drives the browser to perform page redirection, interface element location, and operation triggering according to the operation sequence recorded in the maintenance strategy. After each step, it verifies the existence of the page identifier, interface element location result, and key fields of the result page. If any verification fails, the test session is terminated, and the maintenance strategy is marked as unavailable on the current network management page and not executed in the formal session. For maintenance strategies that pass verification in the test session, the maintenance execution and knowledge update unit executes the corresponding maintenance operations based on the interface operation model in the formal session, provided that the pre-configured rollback conditions are met. During this process, page operations are executed according to the rollback step sequence in the maintenance strategy. The configuration snapshot generated before maintenance begins is used to restore the configuration changes that occurred during maintenance to the configuration values ​​before maintenance. The running status data after rollback is collected through inspection or interface. At the same time, the maintenance execution and knowledge update unit records the execution logs of multiple inspection tasks and maintenance tasks, including inspection path length, number of page visits, task completion time, maintenance results, whether rollback was performed, and whether the same type of fault recurred during the observation period after rollback. Based on this, the success rate and rollback ratio of each maintenance operation template are calculated, the order of maintenance operation templates in the maintenance operation template library is adjusted, and the threshold values ​​in the status fusion and analysis unit and the classification thresholds of the classification model in the anomaly diagnosis unit are determined.

[0098] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0099] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based intelligent inspection and maintenance system for communication networks, characterized in that: The application is applied to the operation and maintenance of a communication network including multiple manufacturers and multiple types of network elements, and the system is in communication connection with a Web-based network management system, and the system comprises: a page analysis unit configured to perform document object model analysis on a Web page output by the network management system, identify interface elements and their hierarchical relationships in the page, establish a mapping relationship between the interface elements and operation types based on an event response mode of the interface elements, and generate an interface operation model independent of an interface protocol of a network element manufacturer; a patrol task configuration unit configured to map performance indicators, alarm items, and configuration items to be obtained to a combination of interface elements in the interface operation model according to a network element type and a business scenario, and form a patrol item configuration; a patrol execution unit configured to control a browser session to perform login, menu selection, condition filling, query triggering, and result page switching based on the interface operation model and the patrol item configuration, collect running data of a target network element from at least two pages associated with each other, and generate patrol raw data; a state fusion and analysis unit configured to perform data alignment on the patrol raw data, historical patrol data, fault work order data, equipment history data, and environmental monitoring data, and construct a running state view according to time and a network element identifier; an abnormality diagnosis unit configured to determine an abnormality level and a suspected fault type of the target network element based on a preset rule and a classification model trained according to historical fault samples on the basis of the running state view; a maintenance strategy generation unit configured to select a maintenance operation from a maintenance operation template library and map the maintenance operation to a page operation sequence according to the abnormality level and the suspected fault type, and generate a maintenance strategy including a check step and a change step; a maintenance execution and knowledge updating unit configured to verify the page operation sequence corresponding to the maintenance strategy in a test session, execute the maintenance strategy in a formal session after the verification is passed, monitor a running state change, execute a rollback operation when a rollback condition is met, adjust a priority of the maintenance operation template in the maintenance operation template library and threshold values in the state fusion and analysis unit and classification thresholds of the classification model in the abnormality diagnosis unit based on execution logs of multiple patrol tasks and maintenance tasks.

2. The AI-based intelligent inspection and maintenance system for communication networks according to claim 1, characterized in that: When generating the interface operation model, the page analysis unit performs the following steps: based on a document object model tree of a page, identifying a navigation area, a condition input area, and a result display area according to node types and hierarchical positions; by detecting event handling functions bound to each node, identifying interface elements that can trigger page jumping, data query, and configuration submission; for each type of interface element, recording a positioning path and an operation type thereof in the page, and establishing a corresponding relationship between an interface element identifier and an operation type in the interface operation model; when detecting a change in a structure of a network management page, the page analysis unit updates the positioning path and the corresponding relationship of the interface elements by comparing types, hierarchical positions, and text contents of nodes in new and old document object model trees, updates the interface operation model, and does not store a fixed operation script sequence. 3.The AI-based intelligent communication network inspection and maintenance system of claim 2, wherein: When performing cross-page patrol, the patrol execution unit: A session context is established for each type of network element, and the current page identifier, the filled-in query conditions, and the time of the last submitted query are recorded; After the query is submitted, the page loading state is detected within a configured waiting time. If the page is not loaded or the parsed interface structure does not match the interface operation model within the waiting time, the current operation is marked as failed, and the corresponding operation is re-executed according to the configured number of retries; If the corresponding patrol item cannot be completed after the configured number of retries, the collection result of the patrol item is recorded as a failure state, and the failure reason information is saved in the patrol original data.

4. The AI-based intelligent communication network inspection and maintenance system of claim 3, wherein: When constructing the running state view, the state fusion and analysis unit: regards each network element instance as a network element entity node, regards the performance indicators, alarm records, and configuration changes collected at different times as state attribute nodes, and regards the fault work order records and maintenance operation records as event nodes; establishes a time relationship edge between the network element entity node and the state attribute node according to the time sequence, and establishes an influence relationship edge between the state attribute node and the event node; attaches a data source identifier and a timestamp to each state attribute node. If the same state attribute value from different data sources of the same network element within the same preset time period is inconsistent, the state attribute node is marked as an untrusted node, and the data of the untrusted node is not used in subsequent feature calculation.

5. The AI-based intelligent inspection and maintenance system for communication networks according to claim 4, characterized in that: When screening the running state view, the anomaly diagnosis unit: adds the fault type corresponding to the performance indicator to the candidate fault type set when the performance indicator exceeds the corresponding threshold at a continuous preset number of sampling moments; adds the fault type corresponding to the preconfigured alarm combination mode to the candidate fault type set when the running state view contains the preconfigured alarm combination mode; adds the fault type corresponding to the preconfigured environmental anomaly condition to the candidate fault type set when the changes in temperature, humidity, or power supply parameters recorded in the environmental monitoring data satisfy the preconfigured environmental anomaly condition; When calculating the credibility of the candidate fault types using the classification model, the anomaly diagnosis unit takes the statistical features of the performance indicators, the time sequence features of the alarm occurrence, and the environmental parameter change features as the input features of the classification model.

6. The AI-based intelligent inspection and maintenance system for communication networks according to claim 5, characterized in that: When selecting a maintenance operation from the maintenance operation template library, the maintenance strategy generation unit: The maintenance operation template library is formed by historical maintenance task records. Each maintenance operation template record includes the operation type, the corresponding interface element identifier, the execution sequence, and the applicable suspected fault type of the maintenance operation; After determining the suspected fault type, the inspection operation templates and the change operation templates corresponding to the suspected fault type are selected from the maintenance operation templates corresponding to the suspected fault type. All inspection operation templates are arranged before all change operation templates; When calculating the impact level of the maintenance operation on the business, the business type, the business traffic, and the service level of the target network element are taken as input parameters. The impact score of each change operation is calculated, and the change operation with a score higher than a configured threshold is marked as a step that requires manual confirmation in the maintenance strategy.

7. The AI-based intelligent inspection and maintenance system for communication networks according to claim 6, characterized in that: The maintenance execution and knowledge updating unit checks whether the target page identifier, interface element positioning result and key field in the result page exist when the page jump, interface element positioning and operation triggering are sequentially executed according to the page operation sequence recorded in the maintenance strategy after each step is executed; When the check fails at any step, the subsequent operation of the test session is terminated, and the maintenance strategy is marked as unavailable under the current network management page, and the maintenance strategy is not executed in the formal session; the maintenance execution and knowledge updating unit records the inspection path length, page access times and task completion time of each inspection task and maintenance task, and the execution result of the maintenance task, whether the rollback operation is executed and whether the same type of fault occurs again within the preset observation period after the maintenance; According to the data obtained by the statistics, the sorting of the maintenance operation template in the maintenance operation template library and the threshold value in the state fusion and analysis unit are adjusted, and the classification threshold of the classification model is adjusted. When the rollback operation is triggered, the maintenance execution and knowledge updating unit executes the corresponding page operation sequence according to the rollback step sequence recorded in the maintenance strategy, restores the configuration changes occurring during the maintenance task to the configuration values recorded in the network management system before the maintenance task starts item by item, and records the event node corresponding to the rollback operation and the state attribute node collected after the rollback in the running state view. The system of any one of claims 1 to 8 is executed, comprising the following steps: 8.The AI-based intelligent communication network inspection and maintenance system of claim 7, wherein: Parsing the Web page of the network management system, identifying the interface elements and establishing the interface operation model; 9. An AI-based intelligent inspection and maintenance method for a communication network, characterized in that: Generating the inspection item configuration according to the type of the target network element and the business scenario; Based on the interface operation model and the inspection item configuration, the browser session is controlled to complete the cross-page operation, the target network element running data is collected, and the inspection raw data is generated; The inspection raw data is fused with the historical inspection data, fault work order data, equipment history data and environment monitoring data, aligned in time and network element identifier, the running state view is constructed and the features are calculated; Based on the running state view and the features, two-stage diagnosis is performed to obtain the abnormal level and suspected fault type of the target network element; The corresponding maintenance operation is selected from the maintenance operation template library, the maintenance operation is mapped to the page operation sequence, and the maintenance strategy is generated in combination with the running state view; The page operation sequence corresponding to the maintenance strategy is verified in the test session, the maintenance strategy is executed in the formal session after the verification is passed, the execution result is monitored, the rollback operation is executed when the rollback condition is met, and the sorting of the maintenance operation template, the state feature threshold and the classification threshold of the classification model are updated according to the execution log. ​ ​

Images