A cross-platform RPA data inspection robot control system based on a federated architecture

Abstract

This invention discloses a cross-platform RPA data inspection robot control system based on a federated architecture, belonging to the field of robotic process automation technology. It includes a federated task scheduling center, an RPA robot cluster, a multi-source dynamic adapter, and a distributed intelligent diagnostic engine. The federated task scheduling center atomically decomposes inspection tasks based on a dynamic priority algorithm, generates balanced scheduling instructions based on the real-time load of each RPA robot, and assigns task queues. The RPA robot cluster is deployed in different network security domains, possessing local caching and network outage resumption capabilities. The inspection process is executed through the multi-source dynamic adapter, which integrates an instruction generalization engine to automatically correct operation paths. This invention achieves intelligent inspection of the entire chain of heterogeneous IT systems, improving inspection execution efficiency and cluster resource utilization, enhancing system robustness, possessing strong scenario adaptability, accurately locating the root cause of anomalies, and continuously optimizing diagnostic accuracy.

Description

Technical Field

[0001] This invention relates to the field of robotic process automation (RPA) technology, and in particular to a cross-platform RPA data inspection robot control system based on a federated architecture. Background Technology

[0002] With the deepening of digital transformation, the IT architecture of modern enterprises and government agencies is becoming increasingly complex. The parallel operation of multiple heterogeneous systems has become the norm, encompassing business applications, database servers, and data sharing platforms based on different operating systems and technical architectures. These systems are often deployed in isolated network security domains according to their functions and security levels. The stable operation of these cross-domain heterogeneous systems directly affects business continuity, thus placing extremely high demands on the comprehensiveness, real-time nature, and intelligence of operation and maintenance inspections. Traditional manual inspection methods rely on maintenance personnel logging into each system one by one, which is not only inefficient and prone to missed or incorrect inspections due to human error, but also difficult to achieve 24 / 7 uninterrupted monitoring, with significant delays in anomaly detection. Simple automated scripts or single-function monitoring tools, on the other hand, suffer from poor adaptability, high maintenance costs, and a lack of a unified scheduling framework, failing to meet the complex inspection needs of cross-domain heterogeneous scenarios.

[0003] To address the pain points of manual inspections, RPA technology, with its non-intrusive deployment and user-simulated operation characteristics, has been widely explored in the field of operations and maintenance automation. However, existing technologies still have many limitations. Patent application CN117172684A, entitled "A Process Automation Method Based on RPA," mentions using RPA robots to receive business requests, process structured data, and periodically execute inspection tasks to achieve automated batch data processing and improve efficiency. However, this technology focuses on automating specific business processes in the financial sector and lacks a scheduling mechanism designed for inspection scenarios involving heterogeneous systems across network security domains. It can only execute preset tasks according to fixed rules and cannot dynamically allocate inspection tasks based on the robot's real-time load, easily leading to both resource idleness and overload. Furthermore, it relies on fixed operation scripts; when the target system's interface elements are updated or its version iterates, the script may fail to execute due to the inability to locate the elements, lacking adaptive correction capabilities, resulting in high maintenance costs and difficulty in adapting to diverse IT system inspection needs.

[0004] The patent application document with publication number CN120011179B and title "Inspection Method for Multi-System Operation and Maintenance Using RPA Inspection Robots Based on Artificial Intelligence" mentions that RPA inspection robots connect to multiple target systems through preset interface protocols, automatically identify login information, collect page data, perform anomaly detection, and send alarms. However, this technology does not adopt a distributed architecture design, making it impossible to achieve collaborative work of robot clusters deployed across domains. It can only complete basic data collection and alarm interception within a single domain system, lacking in-depth analysis and root cause localization capabilities for abnormal data. Furthermore, it does not consider the impact of network fluctuations or interruptions on inspections, lacks local caching and network interruption resumption mechanisms, and is prone to task interruption and data loss. In addition, this technology does not solve the problem of operational adaptation after changes in the target system interface, still requiring manual script adjustments to adapt to the new interface. Its intelligence and robustness are insufficient, making it difficult to support the full-link inspection of large-scale cross-domain heterogeneous systems.

[0005] In summary, existing technologies suffer from three major shortcomings: First, they lack cross-domain scheduling capabilities and a dynamic load balancing mechanism based on the robot's real-time status, resulting in low cluster resource utilization. Second, they have poor scenario adaptability, relying on fixed operation paths, making them unable to handle system interface updates and lacking the ability to resume downloads after network outages, thus exhibiting insufficient robustness. Third, their anomaly diagnosis depth is insufficient, limited to data collection and simple alarms, failing to accurately pinpoint the root cause of anomalies and thus failing to meet the core operational requirements of "early detection and early handling." These issues make existing RPA inspection technologies ill-suited for the complex operational scenarios of cross-domain heterogeneous systems. There is an urgent need for a complete solution integrating federated scheduling, dynamic adaptation, and distributed intelligent diagnosis to overcome existing technological bottlenecks and achieve cross-platform, end-to-end, and intelligent RPA data inspection. Summary of the Invention

[0006] To address the shortcomings of existing RPA inspection technologies, such as lack of load balancing in cross-domain scheduling, poor adaptability to changes in interface elements, easy task interruption and data loss due to network outages, and anomaly diagnosis that only stays at the alarm level without root cause localization, this invention provides a cross-platform RPA data inspection robot control system based on a federated architecture.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] The purpose of this invention is to provide a cross-platform RPA data inspection robot control system based on a federated architecture, the system comprising:

[0009] The Federation Task Scheduling Center decomposes inspection tasks into atomic values ​​based on a dynamic priority algorithm and generates load-balanced scheduling instructions based on the real-time load status of each RPA robot, assigning task queues to the RPA robot cluster.

[0010] The RPA robot cluster consists of one or more robots deployed in different network security domains. The robots have local caching and network interruption resumption capabilities. They are used to receive task queues issued by the federal task scheduling center and execute specific inspection processes through built-in multi-source dynamic adapters. When the network connection with the task scheduling center is interrupted, the robots continue to perform inspections based on the locally cached task queues and temporarily store the inspection results locally. After the network is restored, the results are automatically synchronized to the task scheduling center.

[0011] The multi-source dynamic adapter integrates an instruction generalization engine. The instruction generalization engine matches the DOM structure similarity of the target system's graphical interface and automatically corrects the operation path when interface elements change, enabling the RPA robot to dynamically simulate user behavior, log in, and operate the target system.

[0012] The instruction generalization engine includes a DOM structure parser, a similarity matcher, and a path corrector. The DOM structure parser is used to obtain the DOM structure of the current interface of the target system in real time. The similarity matcher is used to calculate the similarity between the DOM structure of the current interface and the pre-stored standard DOM structure. The path corrector is used to automatically generate a new operation path to replace the original operation instruction when the similarity is lower than a preset threshold.

[0013] Furthermore, it also includes a distributed intelligent diagnostic engine, which includes a first-level lightweight diagnostic unit deployed on the robot side and a second-level deep diagnostic unit deployed on the central side. The first-level lightweight diagnostic unit is used to quickly predict anomalies in real-time collected data based on a preset rule base, and uploads the suspected anomaly data and its context information to the second-level deep diagnostic unit after compression. The second-level deep diagnostic unit is used to decompress and accurately authenticate the received compressed data based on a time-series data analysis model to locate the root cause of the anomaly.

[0014] Furthermore, the federal task scheduling center specifically includes a task decomposition module, a load monitoring module, and an instruction generation module;

[0015] The task decomposition module is used to atomize complex inspection tasks into the smallest executable units based on the type and urgency of the inspection tasks, combined with a dynamic priority algorithm.

[0016] The load monitoring module is used to obtain the CPU utilization, memory usage and task queue length of each robot in the RPA robot cluster in real time, and to generate the real-time load status.

[0017] The instruction generation module is used to generate balanced scheduling instructions based on the load status; wherein the dynamic priority calculation is based on a three-dimensional scheduling model of trust, coupling, and prediction. The three-dimensional scheduling model includes a cross-domain trust model, a task coupling model, and a load prediction model. The expression for the cross-domain trust model is:

[0018] ;

[0019] The task coupling model expression is:

[0020] ;

[0021] The load forecasting model expression is:

[0022] ;

[0023] in Let be the cross-domain trust level between task i and robot j at time t, and let be the output value of the model. This represents the task platform compatibility between task i and robot j, with a value ranging from 0 to 1. It indicates the historical success rate statistics of task i and robot j on their respective platforms. Let be the real-time reliability coefficient of robot j at time t, dynamically calculated based on the robot's own health status reported by the first-level lightweight diagnostic unit. The cost of cross-domain communication between task i and robot j is a normalized value based on the network latency, bandwidth, and number of firewall crossings between the network security domain where the robot resides and the domain where the target system resides. This set of other atomic tasks that have data or logical dependencies on atomic task i is marked by the task decomposition module during the atomization process. Let θ be the planned execution time interval between task i and task k, and let θ be the coupling attenuation coefficient. The load prediction value for robot j at a future time Δt is obtained based on the time series analysis model. This represents the robot's maximum tolerable load threshold. The load prediction fit of robot j at time t reflects the degree to which the future load is adapted to task execution.

[0024] Furthermore, the local cache is used to store the queue of tasks to be executed and the inspection results that have been executed but not synchronized; the network interruption resume capability includes a network status monitoring unit and a data synchronization unit. The data synchronization unit automatically starts the interruption resume mechanism after the network is restored to ensure data integrity.

[0025] Furthermore, the first-level lightweight diagnostic unit includes a data acquisition module, a rule base, an anomaly prediction module, and a data compression module. The data acquisition module is used to collect logs, screenshots, and performance indicators during the inspection process in real time. The rule base pre-stores quick judgment rules based on expert experience or historical data. The anomaly prediction module is used to mark data as suspected anomalies if it triggers a rule in the rule base. The data compression module is used to compress and package the suspected anomaly data and its context information to reduce network transmission overhead.

[0026] Furthermore, the second-level deep diagnostic unit includes a data decompression module, a time-series analysis model, and a root cause localization module; the data decompression module is used to receive and decompress compressed data from the first-level diagnostic unit; the time-series analysis model is a machine learning model trained based on historical data, used to analyze the time-series characteristics of abnormal data; the root cause localization module, combined with contextual information and model output results, locates the root cause and scope of impact of the anomaly.

[0027] Furthermore, it also includes an early warning and report generation module, which is deployed on the central side and connected to the second-level deep diagnostic unit. It is used to receive the abnormal root cause location results, generate early warning information of the corresponding level according to the preset early warning rules and push it to the administrator terminal, and automatically generate a formatted inspection report based on the inspection process data and diagnostic results.

[0028] Furthermore, it also includes a data perception and acquisition module, which works in deep collaboration with the RPA robot cluster to capture multimodal raw data of the target system in real time according to predefined acquisition points during the robot's operation process. The module also performs data cleaning, format normalization, structured extraction, and key indicator calculation on the raw data stream, ultimately outputting a unified, high-quality structured inspection dataset.

[0029] Furthermore, it also includes a feedback learning module, which is connected to the early warning and report generation module and the distributed intelligent diagnostic engine respectively. It is used to collect feedback from operation and maintenance personnel on the confirmation and processing of alarms and diagnostic results, and to use the feedback data to optimize the judgment rules and time series analysis model of the distributed intelligent diagnostic engine, reduce false alarms and false negatives, and continuously improve the diagnostic accuracy and scenario adaptability of the system.

[0030] Furthermore, during the process of performing atomized decomposition of the inspection task, the task decomposition module simultaneously marks the data dependencies and logical dependencies between each atomic task, providing a basis for calculation of the task coupling degree model. At the same time, it determines the planned execution sequence of each atomic task based on the dependencies, ensuring the complete execution of the inspection task.

[0031] Compared with existing technologies, the beneficial effects of this invention are:

[0032] This invention is based on a federated architecture to complete the overall system design, realizing fully automated, normalized, end-to-end, and intelligent data inspection of heterogeneous IT systems within an enterprise. It can adapt to the deployment requirements of different operating systems, technical architectures, and network security domains, significantly improving the execution efficiency and standardization of enterprise IT operation and maintenance inspection work, and reducing the workload of manual operation and maintenance and the risks of human operation.

[0033] This invention utilizes the dynamic priority scheduling mechanism of the federated task scheduling center, combined with the real-time load status of each RPA robot, to complete the atomic decomposition and balanced allocation of inspection tasks. This solves the single-point overload problem that occurs in the traditional RPA scheduling mode, significantly improves the overall task processing capability and hardware resource utilization of the robot cluster, and ensures the orderly and efficient execution of multiple concurrent inspection tasks.

[0034] The multi-source dynamic adapter integrated in this invention is equipped with an instruction generalization engine. It can automatically correct the operation path when the interface elements change based on the DOM structure similarity matching of the target system's graphical interface. It can complete the adaptive adjustment of the inspection process without manual intervention, which solves the problem of traditional RPA scripts failing to execute due to interface updates. It greatly improves the system's adaptability to heterogeneous systems and compatibility with complex scenarios.

[0035] The RPA robot in this invention has local caching and network interruption resume capabilities. Even in scenarios with network fluctuations or interruptions in connection with the scheduling center, it can still continuously complete the entire process inspection operation based on the locally cached task queue. After the network is restored, it automatically completes the synchronous upload of inspection results, solving the problem of task execution interruption and inspection data loss caused by network instability in cross-domain deployment, and greatly enhancing the overall robustness of the system.

[0036] This invention adopts a two-level diagnostic architecture of a distributed intelligent diagnostic engine. The lightweight diagnostic unit on the robot side can quickly predict anomalies in inspection data and compress and upload suspected abnormal data. The deep diagnostic unit on the central side can accurately authenticate and locate the root cause of abnormal data. While improving the response speed and accuracy of anomaly detection, it significantly reduces the transmission resource consumption of cross-domain networks.

[0037] This invention also possesses comprehensive early warning and report generation capabilities. It can automatically generate early warning information of corresponding levels based on abnormal diagnosis results and push it to the corresponding management personnel. At the same time, it can automatically generate formatted inspection reports based on the full amount of inspection data. With the feedback learning module, it can continuously optimize diagnostic rules and analysis models through the processing feedback of maintenance personnel, reduce false alarms and missed alarms, and continuously improve the diagnostic accuracy and scenario adaptability of the system, providing comprehensive and reliable technical support for the stable and secure operation of enterprise IT systems. Attached Figure Description

[0038] Figure 1 This is a block diagram of the cross-platform RPA data inspection robot control system based on federated architecture proposed in this invention.

[0039] Figure 2 This is a schematic diagram of the application method flow in the example embodiment;

[0040] Figure 3 This is a timing flowchart of the cross-platform RPA data inspection robot control system based on federated architecture proposed in this invention;

[0041] Figure 4 A flowchart of task decomposition and dynamic scheduling in the Federal Mission Scheduling Center;

[0042] Figure 5 Flowchart for instruction generalization and path correction of a multi-source dynamic adapter;

[0043] Figure 6 A bar chart comparing the core performance indicators of different inspection schemes;

[0044] Figure 7 Line graph comparing the load balancing effects of robot clusters. 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] Reference Figures 1 to 7 This invention discloses a cross-platform RPA data inspection robot control system based on a federated architecture. The system includes a federated task scheduling center, an RPA robot cluster, a multi-source dynamic adapter, and a distributed intelligent diagnostic engine.

[0047] The Federation Task Scheduling Center decomposes inspection tasks into atomic values ​​based on a dynamic priority algorithm and generates load-balanced scheduling instructions based on the real-time load status of each RPA robot, assigning task queues to the RPA robot cluster.

[0048] An RPA robot cluster consists of one or more robots deployed in different network security domains. The robots have local caching and network interruption resume capabilities. They are used to receive task queues issued by the federal task scheduling center and execute specific inspection processes through built-in multi-source dynamic adapters. When the network connection with the task scheduling center is interrupted, the robot continues to perform inspections based on the locally cached task queues, temporarily stores the inspection results locally, and automatically synchronizes them to the task scheduling center after the network is restored.

[0049] The multi-source dynamic adapter integrates an instruction generalization engine. The instruction generalization engine matches the DOM structure similarity of the target system's graphical interface and automatically corrects the operation path when interface elements change, enabling the RPA robot to dynamically simulate user behavior, log in, and operate the target system.

[0050] The distributed intelligent diagnostic engine includes a first-level lightweight diagnostic unit deployed on the robot side and a second-level deep diagnostic unit deployed on the central side. The first-level lightweight diagnostic unit is used to quickly predict anomalies in real-time collected data based on a preset rule base, and uploads the suspected anomaly data and its context information after compression to the second-level deep diagnostic unit. The second-level deep diagnostic unit is used to decompress and accurately authenticate the received compressed data based on a time-series data analysis model to locate the root cause of the anomaly.

[0051] In this invention, the federal task scheduling center specifically includes a task decomposition module, a load monitoring module, and an instruction generation module;

[0052] The task decomposition module is used to atomize complex inspection tasks into the smallest executable units based on the type and urgency of the inspection tasks, combined with a dynamic priority algorithm.

[0053] The load monitoring module is used to obtain the CPU utilization, memory usage and task queue length of each robot in the RPA robot cluster in real time, and to generate real-time load status.

[0054] The instruction generation module generates balanced scheduling instructions based on load status. Dynamic priority calculation is based on a three-dimensional scheduling model of trust, coupling, and prediction. This model includes a cross-domain trust model, a task coupling model, and a load prediction model. The expression for the cross-domain trust model is:

[0055] ;

[0056] The task coupling model expression is:

[0057] ;

[0058] The load forecasting model expression is:

[0059] ;

[0060] in Let be the cross-domain trust level between task i and robot j at time t, and let be the output value of the model. This represents the task platform compatibility between task i and robot j, with a value ranging from 0 to 1. It indicates the historical success rate statistics of task i and robot j on their respective platforms. Let be the real-time reliability coefficient of robot j at time t, dynamically calculated based on the robot's own health status reported by the first-level lightweight diagnostic unit. The cost of cross-domain communication between task i and robot j is a normalized value based on the network latency, bandwidth, and number of firewall crossings between the network security domain where the robot resides and the domain where the target system resides. This set of other atomic tasks that have data or logical dependencies on atomic task i is marked by the task decomposition module during the atomization process. Let θ be the planned execution time interval between task i and task k, and let θ be the coupling attenuation coefficient. The load prediction value for robot j at a future time Δt is obtained based on the time series analysis model. This represents the robot's maximum tolerable load threshold. The load prediction fit of robot j at time t reflects the degree to which the future load is adapted to task execution.

[0061] In this invention, the local cache is used to store the queue of tasks to be executed and the inspection results that have been executed but not synchronized; the ability to resume transmission after network interruption includes a network status monitoring unit and a data synchronization unit. The data synchronization unit automatically starts the breakpoint resume mechanism after the network is restored to ensure data integrity.

[0062] In this invention, the instruction generalization engine includes a DOM structure parser, a similarity matcher, and a path corrector; the DOM structure parser is used to obtain the DOM structure of the current interface of the target system in real time; the similarity matcher is used to calculate the similarity between the DOM structure of the current interface and the pre-stored standard DOM structure; the path corrector is used to automatically generate a new operation path to replace the original operation instruction based on the matching result when the similarity is lower than a preset threshold.

[0063] In this invention, the first-level lightweight diagnostic unit includes a data acquisition module, a rule base, an anomaly prediction module, and a data compression module. The data acquisition module is used to collect logs, screenshots, and performance indicators during the inspection process in real time. The rule base contains pre-stored quick judgment rules based on expert experience or historical data. The anomaly prediction module is used to mark data as suspected anomalies if it triggers a rule in the rule base. The data compression module is used to compress and package suspected anomaly data and its context information to reduce network transmission overhead.

[0064] In this invention, the second-level deep diagnostic unit includes a data decompression module, a time-series analysis model, and a root cause localization module. The data decompression module is used to receive and decompress compressed data from the first-level diagnostic unit. The time-series analysis model is a machine learning model trained based on historical data, used to analyze the time-series characteristics of abnormal data. The root cause localization module, in combination with contextual information and model output results, locates the root cause and scope of impact of the abnormality.

[0065] This invention also includes an early warning and report generation module, which is deployed on the central side and connected to the second-level deep diagnostic unit. It is used to receive the abnormal root cause location results, generate early warning information of the corresponding level according to the preset early warning rules and push it to the administrator terminal, and automatically generate a formatted inspection report based on the inspection process data and diagnostic results.

[0066] This invention also includes a data perception and acquisition module, which works in deep collaboration with the RPA robot cluster to capture multimodal raw data of the target system in real time according to predefined acquisition points during the robot's operation process. The module also performs data cleaning, format normalization, structured extraction, and key indicator calculation on the raw data stream, ultimately outputting a unified, high-quality structured inspection dataset.

[0067] This invention also includes a feedback learning module, which is connected to the early warning and report generation module and the distributed intelligent diagnostic engine. The feedback learning module is used to collect feedback from operation and maintenance personnel on the confirmation and processing of alarms and diagnostic results, and to use the feedback data to optimize the judgment rules and time series analysis model of the distributed intelligent diagnostic engine, reduce false alarms and false negatives, and continuously improve the diagnostic accuracy and scenario adaptability of the system.

[0068] In this invention, the task decomposition module simultaneously marks the data and logical dependencies between each atomic task during the atomization decomposition of the inspection task, providing a basis for the calculation of the task coupling degree model. At the same time, it determines the planned execution sequence of each atomic task based on the dependencies, ensuring the complete execution of the inspection task.

[0069] Example 1: Implementation of RPA Data Inspection in Multi-Domain IT Systems of Large Manufacturing Enterprises

[0070] This embodiment applies to the IT operations and maintenance scenario of a large manufacturing enterprise. This enterprise deploys multiple heterogeneous IT systems, including a production data server running Linux, office automation applications running Windows, a web-based supply chain management system, and a database server. These systems are deployed in different network security domains—office domain, production domain, and development / testing domain—and are isolated from each other. The enterprise's operations and maintenance team needs to perform 24 / 7 automated status checks on these systems, including login checks, service status monitoring, resource utilization collection, and log anomaly analysis.

[0071] This embodiment deploys a cross-platform RPA data inspection robot control system based on a federated architecture. The system architecture fully follows the system design scheme. The task scheduling center is deployed in the data center of the enterprise headquarters as a central control node. The RPA robot cluster consists of three RPA robots, deployed in different network security domains, with at least one robot deployed in each domain to ensure network reachability. Each robot is equipped with an RPA execution environment and has local caching and network interruption resume modules. A multi-source dynamic adapter is integrated into each RPA robot as the core engine for operating the target system. The distributed intelligent diagnostic engine includes a first-level lightweight diagnostic unit deployed on each robot side and a second-level deep diagnostic unit deployed on the central side. The early warning and report generation module is deployed on the central side, connected to the second-level diagnostic unit, and communicates with the administrator terminal via email, internal enterprise communication tools, and other means. The system synchronously sets up a data perception and acquisition module and a feedback learning module.

[0072] The operations and maintenance administrator defines inspection tasks through the management interface of the task scheduling center. The task decomposition module within the task scheduling center first parses the task into a series of atomic executable units, simultaneously marking the data and logical dependencies between each atomic task during the decomposition process. The load monitoring module of the task scheduling center monitors the load status of each RPA robot in real time, including CPU utilization, memory usage, and the current task queue length. The scheduling center calculates task priorities based on a three-dimensional scheduling model of trust-coupling-prediction, and generates balanced scheduling instructions based on the real-time load status of each robot, assigning task queues to the corresponding robots.

[0073] After receiving the task queue, the RPA robot executes the inspection process sequentially. During execution, if the network connection between the robot's domain and the scheduling center is unexpectedly interrupted, the robot's local caching module automatically detects this and temporarily stores the subsequent execution results in the local database, continuing to complete all atomic tasks. Once the network is restored, the data synchronization unit automatically starts and uploads the temporarily stored inspection results to the task scheduling center.

[0074] When the robot performs inspection operations, the instruction generalization engine in the multi-source dynamic adapter runs synchronously. The DOM structure parser captures the complete DOM tree of the current interface of the target system. The similarity matcher compares the DOM tree with the pre-stored standard interface DOM structure and calculates the similarity. When the similarity is lower than the preset threshold, the path corrector dynamically generates a new operation path based on the matching result, automatically corrects the subsequent operation steps, and completes the login and inspection operations of the target system.

[0075] During the inspection process, the data perception and acquisition module works in collaboration with the RPA robot. As the robot operates, it captures multimodal raw data from the target system in real time according to predefined acquisition points. The module then performs data cleaning, format normalization, structured extraction, and key indicator calculation on the raw data stream, outputting a unified structured inspection dataset. The first-level lightweight diagnostic unit performs rapid anomaly prediction on the real-time acquired data based on a preset rule base. If the data triggers a rule in the rule base, it is marked as a suspected anomaly. The suspected anomaly data and its context information are compressed and packaged, then uploaded to the second-level deep diagnostic unit.

[0076] The second-level deep diagnostic unit receives compressed data, decompresses it, and then inputs it into the time-series analysis model. After analysis, the model confirms the anomaly, locates the root cause of the anomaly based on the context, and pushes the diagnostic results to the early warning and report generation module. The early warning and report generation module generates corresponding level early warning information according to preset early warning rules and pushes it to the terminal of the on-duty maintenance engineer. Simultaneously, after the inspection is completed, it automatically summarizes all robot inspection data, diagnostic results, and anomaly handling records to generate a formatted inspection report. The feedback learning module collects feedback from maintenance personnel regarding the confirmation and handling of alarms and diagnostic results. This feedback data is used to optimize the intelligent diagnostic engine's judgment rules and models, reducing false alarms and missed alarms.

[0077] Table 1. Performance Comparison between the System of this Embodiment and Traditional Inspection Solutions

[0078]

[0079] Table 1 shows the measured results from this embodiment, intuitively demonstrating the performance difference between this system and traditional inspection solutions. Traditional manual inspections are time-consuming per round, have severely delayed anomaly responses, and their fault location accuracy is greatly affected by personnel experience, making uninterrupted inspections impossible. Traditional fixed-script RPA systems have long adaptation cycles to new systems, their task execution success rate drops significantly after interface element changes, they cannot automatically correct operation paths, and their fault location capabilities are insufficient. This system is built based on a complete technical solution, significantly reducing inspection time and anomaly response time, improving fault location accuracy and task execution success rate, shortening the adaptation cycle to new systems, and fully matching the inspection needs of manufacturing enterprises' multi-domain heterogeneous systems.

[0080] Example 2: Implementation of Multi-System RPA Data Inspection in Government Service Center

[0081] This embodiment applies to an IT operations and maintenance scenario in a government service center. The center deploys multiple heterogeneous government service systems, including a government data sharing server running on a Linux operating system, office approval applications running Windows, an integrated government service platform based on a web architecture, and a database server. These systems are deployed in different network security domains—the government external network domain, the internal office network domain, and the data sharing domain—and are isolated from each other via network gateways. The operations and maintenance team needs to perform 24 / 7 automated status checks on these systems, including system login verification, service availability monitoring, server resource utilization collection, and approval log anomaly analysis.

[0082] This embodiment deploys a cross-platform RPA data inspection robot control system based on a federated architecture. The system architecture fully follows the system design scheme. The federated task scheduling center is deployed in the core computer room of the government service center as a central control node, with built-in task decomposition, load monitoring, and instruction generation modules. The RPA robot cluster consists of four RPA robots, deployed in different network security domains, with at least one robot deployed in each domain to ensure network reachability. Each robot is equipped with an RPA execution environment and has local caching and network interruption resumption modules. A multi-source dynamic adapter is integrated into each RPA robot as the core engine for operating the target system. The distributed intelligent diagnostic engine includes a first-level lightweight diagnostic unit deployed on each robot side and a second-level deep diagnostic unit deployed on the central side. The early warning and report generation module is deployed on the central side and connected to the second-level deep diagnostic unit, communicating with the administrator terminal through the government intranet communication channel. The system synchronously sets up a data perception and acquisition module and a feedback learning module.

[0083] Operations administrators define inspection tasks through the management interface of the federated task scheduling center. The task decomposition module, based on the type and urgency of the inspection tasks and combined with a dynamic priority algorithm, atomizes complex inspection tasks into the smallest executable units. During the decomposition process, the data and logical dependencies between each atomic task are simultaneously marked. The load monitoring module obtains the CPU utilization, memory usage, and task queue length of each robot in the RPA robot cluster in real time, generating the real-time load status of each robot. The instruction generation module calculates task priorities based on a three-dimensional scheduling model of trust-coupling-prediction, and generates balanced scheduling instructions based on the real-time load status of each robot, assigning task queues to the corresponding robots.

[0084] After receiving the task queue, the RPA robot executes the inspection process sequentially. During execution, if the network connection between the robot and the scheduling center is interrupted due to adjustments in the domain isolation policy, the robot's local cache module automatically detects the network status change, temporarily stores the inspection results of subsequent executions locally, and continues to complete all atomic tasks according to the locally cached task queue. After the network is restored, the data synchronization unit automatically starts the breakpoint resume mechanism and uploads the temporarily stored inspection results to the task scheduling center.

[0085] During the robot's inspection operation, the instruction generalization engine in the multi-source dynamic adapter works synchronously, the DOM structure parser obtains the DOM structure of the current interface of the government service system in real time, and the similarity matcher calculates the similarity between the DOM structure of the current interface and the pre-stored standard DOM structure. When the similarity is lower than the preset threshold, the path corrector automatically generates a new operation path based on the matching result, replaces the original operation instruction, automatically corrects the subsequent operation steps, and completes the login and inspection operation of the target system.

[0086] During the inspection process, the data perception and acquisition module works closely with the RPA robot. As the robot executes its operations, it captures multimodal raw data from the target system in real time at predefined collection points. The module then performs data cleaning, format normalization, structured extraction, and key indicator calculation on the raw data stream, outputting a unified structured inspection dataset. The first-level lightweight diagnostic unit performs rapid anomaly prediction on the real-time collected data based on a pre-defined rule base. When data triggers rules within the rule base, it is marked as a suspected anomaly. The data compression module compresses and packages the suspected anomaly data and its context information before uploading it to the second-level deep diagnostic unit.

[0087] The second-level deep diagnostic unit receives and decompresses data through the data decompression module, inputs the data into a time-series analysis model trained on historical data, analyzes the time-series characteristics of abnormal data, and the root cause localization module combines contextual information with the model output to accurately locate the root cause and scope of impact of the anomaly, pushing the diagnostic results to the early warning and report generation module. The early warning and report generation module generates corresponding level early warning information according to preset early warning rules and pushes it to the operations and maintenance administrator terminal. Simultaneously, after the daily inspection, it automatically summarizes all robot inspection data, diagnostic results, and processing records to generate a formatted inspection report. The feedback learning module collects feedback from operations and maintenance personnel regarding the confirmation and processing of alarms and diagnostic results, uses the feedback data to optimize the judgment rules and time-series analysis model of the distributed intelligent diagnostic engine, and continuously improves the system's diagnostic accuracy.

[0088] Table 2 Performance Comparison of the System in This Embodiment and Traditional Operation and Maintenance Solutions

[0089]

[0090] Table 2 shows the actual test results based on the system's technical solution deployment in this embodiment, clearly demonstrating the system's performance advantages in multi-domain e-government service scenarios. Traditional manual inspections are time-consuming per round, failing to meet the requirements of uninterrupted inspections in e-government systems, resulting in delayed anomaly responses and low fault location accuracy. Traditional fixed-script RPA systems face frequent version updates and interface element changes in e-government systems, leading to a significant drop in task execution success rate, long new system adaptation cycles, and a lack of anomaly root cause localization capabilities. This system, strictly adhering to the overall system technical solution without any additional content, significantly improves inspection execution efficiency and anomaly response speed, ensures task execution success rate, shortens new system adaptation cycles, and fully matches the inspection needs of heterogeneous multi-domain e-government service scenarios.

[0091] refer to Figure 4 This demonstrates how the task scheduling center atomizes inspection tasks and generates balanced scheduling instructions based on real-time robot load. The task decomposition module uses a dynamic priority algorithm to break down complex tasks into the smallest executable units based on task type and urgency, and marks dependencies. The load monitoring module collects real-time data on each robot's CPU, memory, and queue length. The instruction generation module calculates dynamic priorities based on a three-dimensional scheduling model of cross-domain trust, task coupling, and load prediction, generates balanced scheduling instructions, and distributes task queues to the RPA robot cluster, ensuring efficient cross-platform collaboration.

[0092] refer to Figure 5 This paper describes how the instruction generalization engine responds to changes in the target system interface. When the RPA robot executes a task, the DOM structure parser obtains the current interface DOM tree, and the similarity matcher calculates its similarity with the pre-stored standard DOM structure. If the similarity is lower than a preset threshold, the path corrector automatically generates a new operation path based on the matching result, replacing the original instruction. This allows the robot to dynamically simulate user behavior, adapt to changes in interface elements, and ensure the continuous execution of the inspection process.

[0093] refer to Figure 6 This invention visually demonstrates its comprehensive performance advantages, with data fusion averaging the measured results of two embodiments. Traditional manual inspection takes 270 minutes per round, anomaly response takes 1440 minutes, and fault location accuracy is only 57.5%; traditional fixed-script RPA task execution success rate is 67.5%, with poor adaptability; the system of this invention takes only 22 minutes per round of inspection, anomaly response within 3 minutes, fault location accuracy of 94.5%, and task execution success rate of 98.5%. This effect stems from the synergistic effect of federated scheduling, dynamic adaptation, and distributed diagnostics, solving the pain points of traditional solutions such as "low efficiency, slow response, and poor accuracy".

[0094] refer to Figure 7This invention verifies the load balancing advantages of the federated task scheduling center. Traditional scheduling lacks a dynamic allocation mechanism, resulting in a maximum load difference of 75% when there are 50 tasks, leading to "some robots being overloaded and others idle." This invention, through a three-dimensional scheduling model (trust-coupling-prediction) that atomically decomposes tasks and combines this with real-time dynamic load allocation, achieves a load difference of only 18% when there are 50 tasks, all below the 20% load balancing threshold. This effect ensures the orderly execution of multiple concurrent tasks, improves cluster resource utilization, solves the core problem of "load imbalance" in traditional RPA scheduling, and is suitable for cross-domain multi-robot collaborative scenarios.

[0095] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A cross-platform RPA data inspection robot control system based on a federated architecture, characterized in that, The system includes: The Federation Task Scheduling Center decomposes inspection tasks into atomic values ​​based on a dynamic priority algorithm and generates load-balanced scheduling instructions based on the real-time load status of each RPA robot, assigning task queues to the RPA robot cluster. The RPA robot cluster consists of one or more robots deployed in different network security domains. The robots have local caching and network interruption resumption capabilities, and are used to receive task queues issued by the federal task scheduling center and execute specific inspection processes through built-in multi-source dynamic adapters. The multi-source dynamic adapter integrates an instruction generalization engine. The instruction generalization engine matches the DOM structure similarity of the target system's graphical interface and automatically corrects the operation path when interface elements change. The RPA robot dynamically simulates user behavior, logs in, and operates the target system. The instruction generalization engine includes a DOM structure parser, a similarity matcher, and a path corrector. The DOM structure parser is used to obtain the DOM structure of the current interface of the target system in real time. The similarity matcher is used to calculate the similarity between the DOM structure of the current interface and the pre-stored standard DOM structure. The path corrector is used to automatically generate a new operation path to replace the original operation instruction based on the matching result when the similarity is lower than a preset threshold. The federal mission scheduling center specifically includes a task decomposition module, a load monitoring module, and an instruction generation module; The task decomposition module is used to atomize complex inspection tasks into the smallest executable units based on the type and urgency of the inspection tasks, combined with a dynamic priority algorithm. The load monitoring module is used to obtain the CPU utilization, memory usage and task queue length of each robot in the RPA robot cluster in real time, and to generate the real-time load status. The instruction generation module is used to generate balanced scheduling instructions based on the load status; wherein the dynamic priority calculation is based on a three-dimensional scheduling model of trust, coupling, and prediction. The three-dimensional scheduling model includes a cross-domain trust model, a task coupling model, and a load prediction model. The expression for the cross-domain trust model is: The task coupling model expression is: The load forecasting model expression is: in Let be the cross-domain trust level between task i and robot j at time t, and let be the output value of the model. This represents the task platform compatibility between task i and robot j, with a value ranging from 0 to 1. It indicates the historical success rate statistics of task i and robot j on their respective platforms. Let be the real-time reliability coefficient of robot j at time t, dynamically calculated based on the robot's own health status reported by the first-level lightweight diagnostic unit. The cost of cross-domain communication between task i and robot j is a normalized value based on the network latency, bandwidth, and number of firewall crossings between the network security domain where the robot resides and the domain where the target system resides. This set of other atomic tasks that have data or logical dependencies on atomic task i is marked by the task decomposition module during the atomization process. Let θ be the planned execution time interval between task i and task k, and let θ be the coupling attenuation coefficient. The load prediction value for robot j at a future time Δt is obtained based on the time series analysis model. This represents the robot's maximum tolerable load threshold. The load prediction fit of robot j at time t reflects the degree to which the future load is adapted to task execution.

2. The cross-platform RPA data inspection robot control system based on a federated architecture as described in claim 1, characterized in that, It also includes a distributed intelligent diagnostic engine, which comprises a first-level lightweight diagnostic unit deployed on the robot side and a second-level deep diagnostic unit deployed on the center side. The first-level lightweight diagnostic unit is used to quickly predict anomalies in real-time collected data based on a preset rule base, and uploads the suspected anomaly data and its context information to the second-level deep diagnostic unit after compression. The second-level deep diagnostic unit is used to decompress and authenticate the received compressed data based on a time-series data analysis model to locate the root cause of the anomaly.

3. The cross-platform RPA data inspection robot control system based on a federated architecture as described in claim 1, characterized in that, The local cache is used to store the queue of tasks to be executed and the inspection results that have been executed but not synchronized; the network interruption resume capability includes a network status monitoring unit and a data synchronization unit, and the data synchronization unit automatically starts the interruption resume mechanism after the network is restored.

4. The cross-platform RPA data inspection robot control system based on a federated architecture according to claim 2, characterized in that, The first-level lightweight diagnostic unit includes a data acquisition module, a rule base, an anomaly prediction module, and a data compression module. The data acquisition module is used to collect logs, screenshots, and performance indicators during the inspection process in real time. The rule base pre-stores quick judgment rules based on expert experience or historical data. The anomaly prediction module is used to mark data as suspected anomalies if it triggers a rule in the rule base. The data compression module is used to compress and package the suspected anomaly data and its context information to reduce network transmission overhead.

5. A cross-platform RPA data inspection robot control system based on a federated architecture as described in claim 2, characterized in that, The second-level deep diagnostic unit includes a data decompression module, a time-series analysis model, and a root cause localization module; the data decompression module is used to receive and decompress compressed data from the first-level diagnostic unit; the time-series analysis model is a machine learning model trained based on historical data, used to analyze the time-series characteristics of abnormal data; The root cause localization module, by combining contextual information and model output results, locates the root cause and scope of impact of the anomaly.

6. The cross-platform RPA data inspection robot control system based on a federated architecture according to claim 2, characterized in that, It also includes an early warning and report generation module, which is deployed on the central side and connected to the second-level deep diagnostic unit. It is used to receive the abnormal root cause location results, generate early warning information of the corresponding level according to the preset early warning rules and push it to the administrator terminal, and automatically generate a formatted inspection report based on the inspection process data and diagnostic results.

7. The cross-platform RPA data inspection robot control system based on a federated architecture according to claim 1, characterized in that, It also includes a data perception and acquisition module, which works in deep collaboration with the RPA robot cluster to capture multimodal raw data of the target system in real time according to predefined acquisition points during the robot's operation process. The module also performs data cleaning, format normalization, structured extraction and key indicator calculation on the raw data stream, and finally outputs a unified high-quality structured inspection dataset.

8. A cross-platform RPA data inspection robot control system based on a federated architecture as described in claim 6, characterized in that, It also includes a feedback learning module, which is connected to the early warning and report generation module and the distributed intelligent diagnostic engine. The feedback learning module is used to collect feedback from operation and maintenance personnel on the confirmation and processing of alarms and diagnostic results, and to use the feedback data to optimize the judgment rules and time series analysis model of the distributed intelligent diagnostic engine, reduce false alarms and false negatives, and continuously improve the diagnostic accuracy and scenario adaptability of the system.

9. A cross-platform RPA data inspection robot control system based on a federated architecture as described in claim 1, characterized in that, During the process of performing atomized decomposition of inspection tasks, the task decomposition module simultaneously marks the data dependencies and logical dependencies between each atomic task, providing a basis for calculation of the task coupling degree model. At the same time, it determines the planned execution sequence of each atomic task based on the dependencies, ensuring the complete execution of the inspection task.

Images