Information system operation and maintenance intelligent scheduling management system and method

Abstract

The invention relates to the technical field of information system operation and maintenance management, in particular to an intelligent scheduling management system and method for information system operation and maintenance. The system comprises a fault intelligent sensing module, a multi-source data fusion processing module, an intelligent scheduling decision module, an automatic execution engine and a dynamic optimization module. According to the method, through multi-dimensional fault feature collection and fusion diagnosis, a fault recovery strategy library based on a knowledge graph and deep reinforcement learning is constructed, an improved federated learning framework is adopted to realize standardized integration and quality optimization of multi-source heterogeneous data, and a scheduling priority is dynamically adjusted in combination with a real-time operation and maintenance load. And a sensing-diagnosis-decision-making-execution-optimization closed-loop operation and maintenance system is formed. According to the method, the traditional operation and maintenance bottleneck is broken through, the automatic self-healing rate of complex faults is improved, the data integration accuracy is high, medical and traffic industry customization adaptation capacity is achieved, and the special requirements of medical data privacy protection, traffic system millisecond response and the like are met.

Description

Technical Field

[0001] This invention relates to the field of information system operation and maintenance management technology, specifically to an intelligent scheduling and management system and method for information system operation and maintenance. Background Technology

[0002] With the deepening of digital transformation, information systems in key sectors such as finance, energy, government affairs, healthcare, and transportation are continuously expanding in scale and becoming increasingly complex in architecture. The widespread deployment of heterogeneous devices, multi-source data, and distributed applications presents severe challenges to operation and maintenance management. Currently, there are two core issues in information system operation and maintenance, which exhibit particular pain points in the healthcare and transportation industries: First, the self-healing capabilities of traditional systems are limited, and automated recovery strategies cover insufficient scenarios. The healthcare and transportation industries have even stricter requirements for response timeliness. Traditional maintenance relies on pre-set scripts or manual intervention, which can only handle simple faults. For complex scenarios such as "multi-node linkage failures in electronic medical record systems" and "interruptions in imaging equipment storage links" in the healthcare industry, and "synchronous failures at multiple intersections of main road signal control" and "interruptions in real-time traffic data transmission" in the transportation industry, there is a lack of effective automated diagnosis and recovery mechanisms. Statistics show that the average response time for information system failures in the healthcare industry exceeds 15 minutes, and the manual handling delay for signal failures in the transportation industry exceeds 5 minutes, both far exceeding industry safety standards (medical ≤ 5 minutes, transportation ≤ 1 minute), easily leading to serious consequences such as medical data loss and traffic congestion. Furthermore, traditional automated strategies cannot be adjusted to meet the specific priority requirements of "life support equipment priority" in healthcare systems and "peak-hour traffic priority" in transportation systems, resulting in poor adaptability.

[0003] Secondly, data quality and integration challenges are prominent. Medical and transportation data possess unique attributes (medical data privacy, transportation data real-time requirements), and the problem of multi-source heterogeneity is more pronounced. Medical operation and maintenance data includes electronic medical records (structured), image files (unstructured), and equipment logs (semi-structured). Subject to the Data Security Law and the Personal Information Protection Law, data transmission requires strict privacy protection. Transportation operation and maintenance data includes signal control commands (binary), traffic flow monitoring data (CSV), and road condition video streams (unstructured), requiring data integration latency of ≤100ms. Existing centralized data integration solutions cannot meet the privacy requirements of medical data, and the real-time processing capabilities of transportation data are insufficient. Data missing rates (medical 8%-20%, transportation 10%-18%) and noise levels (medical 5%-12%, transportation 8%-15%) are higher than in other industries, leading to difficulties in correlation analysis and affecting the accuracy of operation and maintenance decisions.

[0004] Existing technologies, while attempting to incorporate AI to enhance the intelligence of operations and maintenance (O&M), suffer from significant shortcomings: solutions for the healthcare industry lack collaborative design for privacy protection and priority scheduling of life support equipment; solutions for the transportation industry fail to address the integrated O&M issues of signal control and traffic flow data. Furthermore, existing solutions often fragment fault handling and data integration, failing to form a customized closed-loop system that meets the high reliability, real-time performance, and privacy requirements of the healthcare and transportation industries. Therefore, a customized intelligent O&M scheduling solution is urgently needed to overcome these technological bottlenecks. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent scheduling and management system and method for information system operation and maintenance. By optimizing core algorithm parameters and setting industry-customized modules, it solves the problems of slow fault self-healing response, weak data privacy protection, and insufficient real-time performance in the medical and transportation industries, while meeting the general operation and maintenance needs of multiple fields.

[0006] The technical solution adopted by this invention to solve its technical problem is: an intelligent scheduling and management system for information system operation and maintenance, comprising: a fault intelligent perception module, used to collect information system operation status data, log data, and fault alarm data through multi-dimensional monitoring nodes, and extract key fault features based on an attention mechanism-based feature extraction network (Transformer encoder layer set to 6 layers, self-attention head set to 8); a multi-source data fusion processing module, which adopts an improved federated learning framework (edge ​​node aggregation frequency 15min / time, setting 2-level edge aggregators), performs localized preprocessing of heterogeneous data sources through edge nodes, standardizes data format based on a dynamic weight mapping algorithm (confidence scoring cycle 5min / time), and repairs missing data values ​​and filters noisy data by combining a Bayesian inference model (prior probability update frequency 30min / time); and an intelligent scheduling decision module, which constructs... A decision-making model integrating knowledge graphs and deep reinforcement learning is constructed. The knowledge graph stores fault types, recovery strategies, and system topology relationships. The reinforcement learning agent (using the PPO algorithm with a training batch size of 64) interacts with the operation and maintenance environment to dynamically adjust scheduling priorities and recovery strategy combinations. An automated execution engine parses the scheduling instructions output by the decision-making module, calls the preset automated script library, API interface, and resource scheduling components, and executes fault recovery operations and data integration tasks. A dynamic optimization module collects real-time operation and maintenance execution effect data and optimizes the decision-making model parameters and data fusion weights based on the gradient descent algorithm (learning rate η=0.008, momentum parameter 0.9) to update the knowledge graph and automated strategy library. All modules achieve data interaction and collaborative work through a distributed message queue (12 Kafka partitions, 3 replicas), supporting horizontal scaling and cross-platform deployment.

[0007] Specifically, the multi-dimensional monitoring nodes of the fault intelligent perception module include a system-level monitoring unit, an application-level monitoring unit, and a network-level monitoring unit. The system-level monitoring unit collects data at a frequency of 1 second / time, the application-level monitoring unit at a frequency of 2 seconds / time, and the network-level monitoring unit at a frequency of 500 ms / time. The monitoring frequency for imaging equipment in the medical industry is 500 ms / time, and the monitoring frequency for signal control in the transportation industry is 300 ms / time. The feature extraction network strengthens the weight allocation of fault-related features through a self-attention mechanism, and the feature extraction accuracy is ≥92%.

[0008] Specifically, the improved federated learning framework of the multi-source data fusion processing module protects data privacy through a differential privacy algorithm (ε=1.2, noise intensity 0.05). Localized preprocessing includes data format verification (verification rules support XMLSchema / JSON Schema), redundant data removal (based on MD5 deduplication), and semantic annotation (using medical / transportation industry-specific dictionaries). In the medical industry, edge nodes are deployed by department, and in the transportation industry, edge nodes are deployed by road segment.

[0009] Specifically, the knowledge graph of the intelligent scheduling decision module includes an entity layer, a relation layer, and an attribute layer. The entity layer includes fault entities, resource entities, and strategy entities. Extended entities in the medical industry include "electronic medical record system fault entity," "imaging equipment resource entity," and "ICU equipment resource entity." Extended entities in the transportation industry include "signal control fault entity," "traffic flow data entity," and "intersection signal entity." The relation layer defines the association rules between entities and adds "medical data privacy association rules" and "traffic signal linkage rules." The attribute layer stores the characteristic parameters and constraints of the entities.

[0010] Specifically, the script library of the automated execution engine supports multiple languages ​​such as Python, Shell, and PowerShell. The built-in fault recovery templates include four core scenarios: service restart, resource expansion, configuration rollback, and data synchronization. The medical industry-specific templates include "image storage service restart", "electronic medical record data synchronization", and "DICOM file synchronization". The transportation industry-specific templates include "signal control command retransmission", "vehicle flow data link repair", and "signal phase adjustment". It supports medical device manufacturers' proprietary APIs (GE, Siemens imaging equipment interfaces) and traffic signal controller protocols (NTCIP).

[0011] An intelligent scheduling and management method for information system operation and maintenance, implemented using the aforementioned intelligent scheduling and management system for information system operation and maintenance, includes the following steps: S1, Fault perception and feature extraction: Real-time collection of information system CPU utilization, memory usage, disk I / O, network bandwidth, application logs, and alarm information through multi-dimensional monitoring nodes. The data collection frequency is 1 second / time for the system layer, 2 seconds / time for the application layer, and 500 ms / time for the network layer. In the medical industry, key data collection includes CPU load of imaging equipment, number of connections to the medical record database, and PACS system response time. In the transportation industry, key data collection includes signal controller status, traffic flow detector data, and intersection passage time. An attention mechanism is utilized. The feature extraction network (6 layers of Transformer encoder) extracts key fault features and generates a fault feature vector with a dimension of 128. The feature vector for the medical industry contains 20 industry-specific features, and the feature vector for the transportation industry contains 25 industry-specific features. S2, Multi-source data fusion processing: Edge nodes receive data from heterogeneous data sources (databases, log systems, monitoring platforms) and perform localized preprocessing (format validation, redundancy removal, semantic annotation). For the medical industry, DICOM file format validation and patient privacy information anonymization are performed. For the transportation industry, binary-to-JSON conversion of signal commands and deduplication of traffic flow data are performed. A dynamic weight mapping algorithm (reliability scoring cycle of 5 minutes / time) is used to combine different formats... Data is mapped to a unified standard. The healthcare industry adds a "patient privacy compliance" scoring indicator, and the transportation industry adds a "data real-time performance" scoring indicator. A Bayesian inference model (prior probability updated every 30 minutes) is used to repair missing data and filter noisy data. A secure data aggregation is achieved based on an improved federated learning framework (edge ​​aggregation every 15 minutes). The healthcare industry uses a two-tiered encryption transmission system (department-hospital level), while the transportation industry uses a two-tiered aggregation system (segment-region level). S3, Intelligent Scheduling Decision: Fault feature vectors are input into a decision model that integrates knowledge graphs and deep reinforcement learning. The knowledge graph matches candidate recovery strategies, and the reinforcement learning agent (PPO algorithm) is combined with the system... Real-time load (resource scheduling is triggered when CPU utilization > 85%), resource availability (non-urgent tasks are restricted when memory availability < 15%), and fault urgency are considered. In the healthcare industry, priority is given to ICU equipment maintenance resources. In the transportation industry, priorities are adjusted based on real-time traffic data. The scheduling priority is dynamically adjusted, and the optimal recovery strategy combination and data integration solution are output. The reward function for deep reinforcement learning is set as: R = α × S + β × T - γ × C - δ × P. In the healthcare industry, P is the privacy risk coefficient (P ∈ [0,1], P = 0 when compliant), and in the transportation industry, P is the real-time performance coefficient (P ∈ [0,1], P = 1 when latency exceeds limits), with α = 0.5, β = 0.3, γ = 0.15, and δ = 0.05. And α+β+γ+δ=1; S4. Automated Execution: The automated execution engine parses decision instructions, calls corresponding scripts, API interfaces, or resource components, and executes fault recovery operations and data integration tasks. Execution latency is ≤300ms in the healthcare industry and ≤100ms in the transportation industry; S5. Dynamic Optimization: Collects execution result data (fault recovery success rate, data integration accuracy, and maintenance time), and optimizes decision model parameters and data fusion weights based on the gradient descent algorithm (η=0.008, momentum 0.9). Parameter optimization is triggered once after every 20 maintenance tasks in the healthcare industry and once after every 10 maintenance tasks in the transportation industry. Simultaneously, the confidence of association rules in the knowledge graph is updated based on industry expert feedback, and the entity relationships and strategy base of the knowledge graph are updated; S6. Iterative Cycle: Repeat steps S1-S5 to achieve continuous optimization and adaptive adjustment of maintenance scheduling.

[0012] Specifically, in step S2, the data format standardization adopts the JSON-LD unified data semantic description format. The medical industry extended JSON-LD context includes "DICOM standard terminology", and the transportation industry extended JSON-LD context includes "traffic signal control terminology".

[0013] Specifically, in step S3, fault classification scheduling divides faults into four levels: emergency fault (P1), important fault (P2), general fault (P3), and minor fault (P4). The medical industry adds a "life support system fault (P0)" level (response time ≤ 1 min), and the transportation industry adds a "main road signal fault (P0)" level (response time ≤ 30 s). Different levels correspond to different scheduling priorities and response times.

[0014] Specifically, in step S5, feedback from medical industry experts includes doctors' suggestions for the operation and maintenance of the medical record system, feedback from transportation industry experts includes traffic police's adjustment needs for signal control, and the confidence update of knowledge graph association rules is based on the validity verification of the feedback results.

[0015] Specifically, in step S2, the desensitization of patient privacy information in the medical industry is carried out according to the HIPAA standard, and the signal command transmission in the transportation industry adopts the MQTT protocol to achieve low-latency data interaction.

[0016] The beneficial effects of this invention are: By collaborating on multiple core technologies, a technology system combining general-purpose and industry-customized approaches is formed: Attention mechanism feature extraction combined with industry-specific monitoring frequency: In response to the special characteristics of fault features in the medical and transportation industries, high-frequency monitoring and feature weight enhancement are used to solve the problems of inaccurate and insufficient comprehensiveness in the extraction of industry-specific fault features, and the feature extraction accuracy is improved compared with general solutions. Improved collaboration between federated learning and industry data privacy protection: The healthcare industry combines differential privacy with department-level edge aggregation, while the transportation industry combines low-latency transmission with segment-level aggregation. This simultaneously resolves the contradiction between low efficiency in multi-source data integration and industry privacy / real-time requirements. Healthcare data privacy compliance reaches 100%, and transportation data integration latency is reduced by 50% compared to general solutions. Knowledge graph industry extension and enhanced learning reward function customization collaboration: Through industry-specific entities, rules and reward coefficients, it achieves deep adaptation between decision-making models and industry operation and maintenance logic, solves the problem of insufficient industry risk control in traditional decision-making solutions, achieves 100% compliance rate for P0 level fault response in the medical industry, and improves signal synchronization rate in the transportation industry; Automated execution industry templates plus dynamic optimization frequency adaptation and collaboration: Based on industry fault characteristics, customized execution templates and optimization frequencies solve the problems of high execution latency and low optimization efficiency of general operation and maintenance solutions. The interruption rate of emergency image transmission in the medical industry is reduced to below 0.1%, and the delay of fault handling during the morning rush hour in the transportation industry is ≤15s. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] Figure 1 : Overall system architecture diagram of this invention; Figure 2 : Flowchart of the method of this invention; Figure 3 : Schematic diagram of the feature extraction network structure of the fault intelligent perception module; Figure 4 : Schematic diagram of multi-source data fusion processing flow; Figure 5 : Schematic diagram of the dual-drive model structure of the intelligent scheduling decision module; Figure 6 Healthcare system deployment architecture diagram (divided by department); Figure 7 : Schematic diagram of edge nodes in the transportation industry (by transportation area); Figure 8 : Diagram of medical data privacy protection process (including desensitization and encrypted transmission); Figure 9 Traffic signal fault decision logic diagram (including strategy combination selection). Detailed Implementation

[0019] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0020] like Figure 1 , Figure 3 , Figure 5 , Figure 6 and Figure 7 As shown, the intelligent scheduling and management system for information system operation and maintenance of the present invention includes the following modules: The intelligent fault perception module adopts a three-layer monitoring architecture of "system-application-network," covering 12 core indicators including CPU, memory, disk, network, application logs, and alarm information. The monitoring frequency is adjusted according to industry characteristics: 500ms / time for medical imaging equipment and ICU life support equipment, and 300ms / time for traffic signal controllers and traffic flow detectors. The feature extraction network uses a 6-layer Transformer encoder (8 self-attention heads), strengthening industry-specific fault feature weights through a self-attention mechanism. For the medical industry, it emphasizes features such as "image storage error codes," "abnormal medical record data transmission," and "monitor data interruption," while for the traffic industry, it emphasizes features such as "signal command timeout," "traffic flow data jumps," and "link bandwidth saturation." The feature extraction accuracy is improved to over 95%, providing precise input for fault diagnosis.

[0021] Multi-source data fusion processing module: Based on an improved federated learning framework, a "edge preprocessing-hierarchical aggregation" data processing architecture is constructed, setting up a two-level edge aggregator: for the medical industry, edge nodes are deployed in zones according to departments such as internal medicine, surgery, and radiology; for the transportation industry, edge nodes are deployed in zones according to urban main roads and secondary roads, with an aggregation frequency of 15 minutes / time, balancing data real-time performance and processing efficiency. A differential privacy algorithm (ε=1.2, noise intensity 0.05) is used to protect medical data privacy, avoiding the risk of leakage caused by centralized transmission of raw data; a dynamic weight mapping algorithm adds industry-specific scoring indicators, introducing "patient privacy compliance" (based on data anonymization integrity assessment) for the medical industry and "data real-time performance" (based on data collection and transmission delay assessment) for the transportation industry, updating the data source credibility score every 5 minutes, and combining JSON-LD to achieve data semantic unification. Data quality optimization employs a Bayesian inference model (prior probability updated every 30 minutes), repairs missing values ​​based on industry data correlation patterns (e.g., in the medical industry, based on image data patterns of patients with the same disease, and in the transportation industry, based on traffic flow data patterns of adjacent intersections), and filters noisy data using the Isolation Forest algorithm (outlier detection threshold of 0.95). The accuracy rate of medical data missing value repair is ≥98%, and the latency of traffic data integration is ≤80ms.

[0022] The intelligent scheduling decision-making module constructs a dual-driven decision-making model combining a knowledge graph and deep reinforcement learning. The knowledge graph covers over 800 general fault types and over 500 industry-specific fault types (300+ in healthcare, 200+ in transportation). The entity layer includes industry-specific entities such as "ICU equipment resources" and "intersection signal entities," while the relationship layer defines industry rules such as "medical data privacy association" and "traffic signal linkage." Reinforcement learning employs the PPO algorithm (training batch size 64), aiming to maximize fault recovery success rate, minimize operation and maintenance time, minimize resource consumption, and minimize industry risk. The reward function introduces an industry-specific coefficient P: for the healthcare industry, P is the privacy risk coefficient (P=1 when data transmission is not anonymized, P=0 when compliant); for the transportation industry, P is the real-time performance coefficient (P=1 for response timeout, P=0 when meeting standards). Through real-time interaction with the operation and maintenance environment, the module dynamically adjusts scheduling priorities and strategy combinations. The response time for P0 level (life support equipment failure) in the medical industry is ≤1 minute, and the response time for P0 level (main road signal failure) in the transportation industry is ≤30 seconds. The automatic self-healing rate of complex faults has been increased to over 88%.

[0023] Automated execution engine: Supports multi-language scripts (Python / Shell / PowerShell) and standardized API interfaces. It includes 100+ general-purpose fault recovery templates and 50+ industry-specific templates (30+ for medical and 20+ for transportation). Medical templates include "DICOM file repair," "electronic medical record data rollback," and "monitor network restart," while transportation templates include "signal phase adjustment," "vehicle flow data link repair," and "intersection signal synchronization calibration." Execution latency is precisely controlled according to industry requirements: ≤300ms for medical (ensuring uninterrupted image transmission and life support equipment data), and ≤100ms for transportation (avoiding congestion caused by signal control synchronization failures). It also supports industry-specific device interface calls (such as GE imaging equipment API and Siemens signal controller NTCIP protocol interface) to meet personalized operation and maintenance needs.

[0024] Dynamic optimization module: Employing an online learning model, this module uses a gradient descent algorithm (learning rate η=0.008, momentum 0.9) to optimize decision model parameters and data fusion weights in real time. The optimization trigger frequency is set according to industry-specific fault characteristics: once every 20 maintenance tasks in the healthcare industry (adapting to low-frequency, high-risk faults), and once every 10 maintenance tasks in the transportation industry (adapting to high-frequency, real-time faults). The module updates the confidence level of knowledge graph association rules based on feedback from industry experts (healthcare doctors' maintenance suggestions for the medical record system and traffic police's adjustment needs for signal control), enabling continuous evolution of maintenance capabilities. Even 6 months after system deployment, the fault self-healing rate remains above 85%.

[0025] like Figure 2 , Figure 4 , Figure 8and Figure 9 As shown, the intelligent scheduling and management method for information system operation and maintenance described in this invention includes six steps: fault perception and feature extraction, multi-source data fusion processing, intelligent scheduling decision-making, automated execution, dynamic optimization, and iterative iteration, forming a closed-loop operation and maintenance system. Each step has its execution logic refined to meet industry needs. Step S1 (Fault Perception and Feature Extraction): Multi-dimensional monitoring nodes collect data at preset frequencies. In the medical industry, the focus is on collecting data such as PACS system storage I / O, HIS system database connection count, and ICU monitor data transmission status. In the transportation industry, the focus is on collecting data such as signal controller command execution status, traffic flow detector data update frequency, and intersection communication link bandwidth. The collected data is standardized (normalized to the [0,1] interval) and then input into the feature extraction network. The feature association weights are calculated using the Transformer encoder's self-attention mechanism, and the Top-20 key features are extracted to generate a 128-dimensional fault feature vector. The medical industry vector includes 20 exclusive features such as "image storage error code" and "medical record writing failure identifier," while the transportation industry vector includes 25 exclusive features such as "signal command timeout count" and "traffic flow data jump amplitude."

[0026] Step S2 (Multi-Source Data Fusion Processing): After receiving multi-source data, the edge nodes first perform industry-specific preprocessing (DICOM format verification and patient ID number anonymization in the medical industry; NTCIP protocol format verification and traffic flow data timestamp deduplication in the transportation industry). Then, based on a dynamic weight mapping algorithm (combined with industry credibility indicators), the heterogeneous data is mapped to the JSON-LD standard format. A Bayesian inference model is used to repair missing data (e.g., in the medical industry, missing image parameters are repaired based on patient age and symptoms; in the transportation industry, missing traffic speed data is repaired based on road segment length and speed limit). An isolated forest algorithm is used to filter noisy data (e.g., abnormal monitoring data caused by temporary network outages in the medical industry; abnormal traffic flow data caused by detector malfunctions in the transportation industry). Finally, data aggregation is completed through an improved federated learning framework: the medical industry uses a two-level encrypted transmission (AES-256 encryption) from "department edge nodes to hospital center nodes," while the transportation industry uses a two-level low-latency transmission (MQTT protocol) from "road segment edge nodes to regional center nodes," generating a standardized operation and maintenance dataset.

[0027] Step S3 (Intelligent Scheduling Decision): The decision model first uses the knowledge graph to perform entity matching and relationship reasoning to determine the fault type and candidate recovery strategies (e.g., in the medical industry, "PACS storage array failure" corresponds to three candidate strategies: "non-emergency image archiving," "storage expansion," and "temporarily restricting non-emergency uploads"; in the transportation industry, "signal command timeout" corresponds to three candidate strategies: "compressed command transmission," "link expansion," and "phase adjustment"). The reinforcement learning agent calculates the reward value of each strategy combination based on the real-time system status (number of emergency patients and ICU equipment load in the medical industry; traffic congestion index and peak hours in the transportation industry), selects the optimal solution, and assigns scheduling priorities (e.g., in the medical industry, P0-level faults prioritize the use of maintenance resources; in the transportation industry, faults on main roads during morning rush hour are prioritized). For extremely complex industry faults (no matching record in the knowledge graph), a reinforcement learning exploration mode (exploration factor 0.1) is initiated to dynamically generate new strategy combinations and record them in the knowledge graph.

[0028] Step S4 (Automated Execution): The automated execution engine parses the decision instructions, matches the corresponding industry templates or scripts, and calls the dedicated interface to execute the operation: When the medical industry executes "non-urgent image archiving", it first calls the storage management API to query the remaining capacity of off-site storage, then performs batch migration of DICOM files, and synchronously updates the image storage path of the HIS system; When the transportation industry executes "signal command compression", it calls the router API to enable the compression algorithm (compression rate 60%), and at the same time retransmits the signal command through the NTCIP protocol. During the execution process, progress data (such as archiving completion rate, command execution status) is fed back in real time. If the execution fails (such as image migration interruption in the medical industry, signal adjustment in the transportation industry is invalid), the backup strategy is automatically triggered (the medical industry switches to local storage expansion, and the transportation industry switches to signal primary and backup switch).

[0029] Step S5 (Dynamic Optimization): Collect execution result data and calculate core indicators (fault recovery success rate, data integration accuracy, operation and maintenance time, resource consumption, and industry risk indicators). For the medical industry, focus on analyzing "emergency image acquisition time" and "privacy compliance rate," while for the transportation industry, focus on analyzing "signal synchronization rate" and "congestion relief rate." Optimize decision model parameters (e.g., adjusting the weight of the industry coefficient P in the reward function) and data fusion weights based on the gradient descent algorithm (e.g., increasing the credibility weight of PACS data in the medical industry and signal data in the transportation industry). Simultaneously, verify the effectiveness of the strategy by incorporating feedback from industry experts (e.g., doctors confirming no omissions in medical record data rollback and traffic police confirming that signal adjustments alleviate congestion). Update the confidence of the knowledge graph association rules (confidence of effective strategies +0.1, confidence of ineffective strategies -0.05).

[0030] Step S6 (Iterative Loop): Repeat steps S1-S5 to achieve continuous optimization of operation and maintenance scheduling. The system automatically expands the monitoring nodes and strategy library according to changes in the industry information system architecture (such as the addition of imaging equipment in hospitals or the addition of main roads in cities), without the need for manual redeployment, and the adaptability and scalability are significantly improved.

[0031] Example 1: Operation and Maintenance Scenario of Information System in a Top-Tier Hospital in the Healthcare Industry This embodiment is applied to the information system of a tertiary hospital. The system includes a PACS imaging system (storing 1 million+ medical records, with 5000+ new records added daily), a HIS hospital information system (supporting 5000+ outpatient visits / day and 2000+ inpatient management), and an ICU life support equipment networking system (including 50+ Mindray / Philips monitors and 10+ ventilators). The operation and maintenance requirements are: fault response time of P0 level (life support equipment failure) ≤1min, P1 level (imaging / HIS system failure) ≤5min, data privacy complies with the "Medical Data Security Guidelines" and "HIPAA Standard", and the image transmission interruption rate is ≤0.5%.

[0032] System Deployment: The fault intelligent perception module deploys 200+ monitoring nodes, with ICU equipment (monitors, ventilators) monitored at a frequency of 500ms / time, PACS system (storage array, image transmission link) monitored at a frequency of 1s / time, HIS system (database, outpatient registration module) monitored at a frequency of 2s / time, and network layer (hospital intranet, equipment networking link) monitored at a frequency of 500ms / time. The multi-source data fusion processing module deploys 10 edge nodes, zoned according to departments such as internal medicine, surgery, obstetrics and gynecology, radiology, and ICU. Each edge node connects to the corresponding department's medical equipment (e.g., the radiology edge node connects to a GE 1.5T MRI and a Siemens CT scanner), supporting DICOM 3.0 and HL7. FHIR data standard; the intelligent scheduling decision module is deployed on the hospital's private cloud server (8 cores, 16GB configuration), the knowledge graph preloads 500+ medical fault types (including "PACS storage array insufficient capacity", "electronic medical record writing deadlock", "monitor network interruption"), 400+ recovery strategies, and the reinforcement learning model training dataset contains 3 years of hospital operation and maintenance history data (100,000+ fault cases); the automated execution engine integrates with the hospital's existing operation and maintenance platform (WINHIS operation and maintenance system of Winning Health), and the script library adds DICOM file repair scripts, electronic medical record data rollback scripts, and monitor network restart scripts, supporting GE imaging equipment API and Mindray monitor SDK interface calls; the dynamic optimization module interfaces with the hospital's information department expert system to receive operation and maintenance effect data (such as image transmission stability and medical record data integrity) from doctors and nurses.

[0033] Specific implementation process: 1. Fault Detection and Feature Extraction: At 10:00 AM on Monday (peak outpatient hours), the monitoring node collected data showing that the PACS system's storage array IO utilization reached 98% (threshold 85%), and the image transmission delay exceeded 5 seconds (threshold 2 seconds). The HIS system reported an alarm for "Outpatient patient image retrieval failure," and the PACS log showed "DICOM file write error (error code 0x80070005)." After standardizing the collected data, the feature extraction network calculated feature weights using the Transformer encoder's self-attention mechanism. The weights of four feature categories—"Storage IO saturation," "DICOM error code," "Image transmission delay," and "Outpatient retrieval failure alarm"—accounted for 75%. The Top-20 key features were extracted to generate a 128-dimensional fault feature vector (including 15 medical-specific features such as "Imaging equipment model" and "Patient ID anonymization identifier").

[0034] 2. Multi-source data fusion processing: The radiology edge node receives three types of data: PACS storage array data (binary format, including IO read / write speed and remaining capacity), HIS system associated data (MySQL format, including outpatient image retrieval records and emergency / non-emergency identification), and imaging equipment logs (TXT format, including DICOM file transfer logs). The edge node first performs industry-specific preprocessing: verifying the DICOM file format (compliant with the DICOM 3.0 standard), desensitizing patient ID numbers, names, and other private information (replacing them with internal hospital virtual IDs), and deduplicating duplicate image transfer logs based on MD5. A dynamic weighted mapping algorithm is used to calculate the data source credibility: PACS storage data credibility 0.95 (98% historical data integrity), HIS associated data credibility 0.9 (95% accuracy of outpatient records), and equipment log credibility 0.92 (2% log missing rate). Combined with JSON-LD (Extended DICOM Standard Context), heterogeneous data is mapped to a unified format. Data repair was performed using a Bayesian inference model: Based on the storage I / O data patterns of the same time period (10:00-10:30 AM) over the past week, two missing monitoring data points caused by network jitter were repaired; one abnormal image transmission delay data point (value reaching 30 seconds, far exceeding the normal range) caused by a temporary network outage of the CT equipment was filtered out using the Isolation Forest algorithm. Finally, aggregation was completed using an improved federated learning framework: The radiology edge node first encrypts the locally processed data (AES-256) and transmits it to the hospital-level central aggregator. The aggregator combines data from edge nodes of other departments (such as image access request data from internal medicine and surgery) to generate a standardized operation and maintenance dataset. Data integration took 60ms, with 100% privacy compliance.

[0035] 3. Intelligent Scheduling Decision: Fault feature vectors are input into the decision model. The knowledge graph uses entity matching (matching "Storage IO saturation" and "DICOM write error" with "Insufficient PACS storage array capacity" fault types) and relational reasoning ("Outpatient peak hours" → "Emergency image priority" rule) to output three candidate recovery strategies: Strategy A "Storage array expansion" (requires calling the storage management API to add 10TB of capacity, taking approximately 5 minutes), Strategy B "Archiving non-emergency images to off-site storage" (archiving non-emergency patient images from 3 months ago to the hospital's off-site disaster recovery center, taking approximately 40 seconds), and Strategy C "Temporarily restricting non-emergency image uploads" (only retaining image upload permissions for emergency and ICU departments, taking approximately 10 seconds). The reinforcement learning agent calculates the reward value based on real-time status: Currently, during the outpatient peak hours, emergency patients account for 15% (approximately 750 people), and the remaining off-site storage capacity is 50TB. Expanding storage would require consuming the hospital's core network bandwidth (potentially affecting monitor data transmission). Reward function calculation: For strategy B+C, R = 0.5 × 1 (estimated self-healing success rate 100%) + 0.3 × 0.9 (time 40s, optimization rate 93%) - 0.15 × 0.1 (low resource consumption) - 0.05 × 0 (privacy compliance) = 0.755; For strategy A, R = 0.5 × 1 + 0.3 × 0.3 (time 5min, optimization rate 30%) - 0.15 × 0.8 (high resource consumption) - 0.05 × 0 = 0.51. Ultimately, strategy B+C is selected, with scheduling priority set to P1 (critical fault), and a response time requirement of ≤5min.

[0036] 4. Automated Execution: The automated execution engine parses the decision instructions and executes them in two steps: First, it calls the hospital's off-site storage management API to execute a non-emergency image archiving script, filtering non-emergency patient images from 3 months ago (totaling 20TB, involving 30,000+ cases), and transmitting them to the off-site disaster recovery center via a dedicated hospital fiber optic link at a transmission rate of 1GB / s, taking 40 seconds. After archiving, the remaining capacity of the PACS storage array increased from 5% to 25%. Second, it calls the HIS system's permission management interface to execute a non-emergency image upload restriction script, temporarily closing image upload permissions for general outpatient departments (internal medicine, surgery, obstetrics and gynecology), retaining only upload permissions for emergency, ICU, and pediatrics departments, taking 8 seconds. Real-time progress feedback was provided during execution: the archiving progress was updated every 10 seconds, and the permission adjustment results were synchronized to the hospital's information department monitoring platform in real time. No interruptions occurred in emergency image uploads (a total of 120 emergency images were successfully transmitted).

[0037] 5. Dynamic Optimization: Data collection and execution results show a 100% fault recovery success rate (PACS storage IO utilization reduced to 65%, image transmission latency restored to 1.2s, and 100% success rate in retrieving outpatient images), maintenance time of 48s (meeting P1 level response time limits), low resource consumption cost (no core network bandwidth consumed), and 100% privacy compliance (archived images have all been anonymized). The decision-making model is optimized based on the gradient descent algorithm (η=0.008, momentum 0.9): the weight of "emergency priority" in the reward function is increased from 0.1 to 0.15 to enhance emergency resource guarantee capabilities during peak hours; the PACS data credibility scoring algorithm is adjusted, adding the "remaining storage capacity percentage" indicator (weight 0.2). Based on feedback from IT experts (doctors confirmed that emergency image transmission was uninterrupted and archived images were retrieved normally), the knowledge graph association rules were updated: the association confidence of the strategy between "insufficient PACS storage capacity + peak outpatient hours" and "non-emergency image archiving + access restrictions" was increased from 0.85 to 0.99. At the same time, an "emergency image archiving exemption" rule was added (to prevent emergency images from being mistakenly archived).

[0038] Implementation Results: In this scenario, the automated self-healing rate of the medical information system reached 90%, with a 100% self-healing rate for P0-level faults (such as ICU monitor network interruption) and a response time of ≤45s. The self-healing rate for P1-level faults was 85%, with a response time of ≤3 minutes, representing a reduction of over 80% compared to traditional manual maintenance (response time 15 minutes). The success rate of PACS image transmission increased from 92% to 99.8%, and the image transmission interruption rate decreased to 0.1% (only one case due to equipment hardware failure), with no medical data leakage incidents. The average acquisition time for emergency images decreased from 3 minutes to 45 seconds, improving physician diagnostic efficiency by 75% and patient satisfaction by 20%. During the six months of system operation, over 1200 faults were handled, with only 120 requiring manual intervention (all hardware replacement-related faults), reducing the workload of maintenance personnel by 60%.

[0039] Example 2: Operation and Maintenance Scenario of Signal Control System for Urban Main Roads in the Transportation Industry This embodiment is applied to the main road signal control system of a first-tier city, covering 20 intersections (including 5 traffic hub intersections: the railway station intersection, the municipal government intersection, the commercial center intersection, etc.). It adopts the Siemens SC2000 signal controller (supporting the NTCIP1202 protocol) and is equipped with Hikvision ITS traffic flow monitoring radar (4 units per intersection, monitoring range 300m). The operation and maintenance requirements are: fault response time limit P0 level (main road signal failure) ≤30s, P1 level (traffic flow data interruption) ≤1min, signal synchronization rate ≥95%, average traffic speed on the main road during the morning peak (7:30-8:30) ≥30km / h, and traffic flow data transmission accuracy ≥98%.

[0040] System Deployment: The fault intelligent perception module deploys 150+ monitoring nodes, with 3 monitoring units deployed at each intersection: a signal controller monitoring unit (monitoring frequency 300ms / time, collecting command execution status, phase duration, and signal synchronization status), a traffic flow monitoring unit (monitoring frequency 500ms / time, collecting traffic flow, vehicle speed, and queue length), and a communication link monitoring unit (monitoring frequency 200ms / time, collecting bandwidth utilization, transmission delay, and packet loss rate); the multi-source data fusion processing module deploys 5 edge nodes, zoned according to the city's five major traffic areas (east, west, south, north, and central), with each edge node connecting to the signal controllers and traffic flow radars at 4 intersections, supporting NTCIP and MQTT protocols; the intelligent dispatch decision module is deployed at... The municipal traffic police command center cloud platform (16 cores, 32GB configuration) has a knowledge graph preloaded with 400+ traffic fault types (including "signal command execution timeout", "traffic flow data transmission interruption", and "intersection signal phase loss"), 300+ recovery strategies, and a reinforcement learning model training dataset containing one year of urban traffic operation and maintenance data (50,000+ fault cases and 100 million+ traffic flow data entries). The automated execution engine integrates with the existing platform of the traffic police command center (Hisense Intelligent Traffic Management System), and the script library has added scripts for signal phase adjustment, traffic flow data link repair, and intersection signal synchronization calibration, supporting Siemens signal controller API and Hikvision radar SDK interface calls. The dynamic optimization module interfaces with the traffic police branch command platform to receive real-time traffic condition feedback (such as congestion alarms and accident information).

[0041] Specific implementation process: 1. Fault Detection and Feature Extraction: At 7:45 AM on Tuesday morning, during the morning rush hour, the monitoring node collected feedback from the signal controller at the commercial center intersection (hub intersection) indicating "command execution timeout" (five consecutive phase adjustment commands failed to respond). Traffic flow monitoring radar data showed that the eastbound traffic flow jumped by more than 30% (from 200 vehicles / 5min to 260 vehicles / 5min), with a queue length of 200m (threshold 150m). The communication link monitoring unit showed that the dedicated fiber optic bandwidth occupancy rate between the intersection and the command center reached 99% (threshold 80%), and the transmission delay exceeded 200ms (threshold 50ms). After standardizing the data, the feature extraction network strengthened the features of "command timeout," "traffic flow jump," "bandwidth saturation," and "transmission delay" through a self-attention mechanism (with a weighting of 80%), extracting the Top-20 key features to generate a 128-dimensional fault feature vector (including 20 traffic-specific features such as "intersection type," "peak hour sign," and "current signal phase value").

[0042] 2. Multi-source data fusion processing: Edge nodes in the urban area receive three types of data: signal controller data (binary command format, including phase adjustment commands and execution status codes), traffic flow monitoring data (CSV format, including traffic flow, vehicle speed, and queue length, updated every 5 seconds), and communication link data (JSON format, including bandwidth utilization, packet loss rate, and transmission latency). The edge nodes first perform industry-specific preprocessing: verifying the signal command format (compliant with the NTCIP 1202 protocol), deduplicating duplicate traffic flow data based on "timestamp + intersection ID + lane number," and semantically annotating the signal controller status codes (e.g., "0x0001" is annotated as "command executed successfully," and "0x0002" is annotated as "command timed out"). The reliability of data sources was calculated using a dynamic weight mapping algorithm: signal controller data reliability was 0.98 (historical command execution feedback accuracy was 99%), traffic flow data reliability was 0.92 (radar monitoring error rate was 5%), and link data reliability was 0.95 (bandwidth monitoring error rate was 2%). Combined with JSON-LD (Extended Traffic Signal Control Terminology Context), heterogeneous data was mapped to a unified format. A Bayesian inference model was used to repair the data: based on the traffic flow data patterns of adjacent intersections (the intersection east of the commercial center, "Department Store Intersection," and the intersection west of the commercial center, "Park Intersection"), three missing traffic flow data points caused by temporary radar malfunctions were repaired (supplementing eastbound speed and queue length data); two abnormal signal command data points caused by link interference were filtered using an isolated forest algorithm (command execution time was displayed as 10 seconds, far exceeding the normal 0.5 seconds). Finally, the aggregation was completed by improving the federated learning framework: the edge nodes in the city first compress the local data (compression rate of 60%) and transmit it to the aggregator of the city traffic data center via the MQTT protocol. The aggregator combines the data of edge nodes in other areas (such as the traffic flow scheduling needs of the east and west of the city) to generate a standardized operation and maintenance dataset. The data integration takes 60ms, which meets the real-time requirements.

[0043] 3. Intelligent Dispatch Decision: The fault feature vector is input into the decision model. The knowledge graph uses entity matching (matching "command timeout" and "bandwidth saturation" with the fault type "communication link saturation between signal controller and command center") and relational reasoning ("morning rush hour" → "main road traffic priority" rule) to output three candidate recovery strategies: Strategy A "Communication link bandwidth expansion" (requires calling the operator's API to add 100M bandwidth, taking approximately 5 minutes), Strategy B "Signal command compression transmission" (enabling command compression algorithm, compression rate 60%, taking approximately 2 seconds), and Strategy C "Temporarily adjust signal phase" (extending the green light duration from 60 seconds to 70 seconds for eastbound traffic, taking approximately 5 seconds). The reinforcement learning agent calculates the reward value based on real-time status: the current morning rush hour eastbound congestion index is 1.8 (threshold 1.2), the average traffic speed on the main road is 25km / h (not meeting the standard), and link expansion requires emergency communication resources (potentially affecting signal transmission at other intersections). Reward function calculation: For strategy B+C, R = 0.5 × 1 (estimated self-healing success rate 100%) + 0.3 × 0.98 (time 7s, optimization rate 89%) - 0.15 × 0.02 (low resource consumption) - 0.05 × 0 (real-time performance meets requirements) = 0.791; For strategy A, R = 0.5 × 1 + 0.3 × 0.3 (time 5min, optimization rate 30%) - 0.15 × 0.9 (high resource consumption) - 0.05 × 1 (real-time performance does not meet requirements) = 0.39. Ultimately, strategy B+C is selected, with scheduling priority set to P0 (emergency fault), and a response time requirement of ≤30s.

[0044] 4. Automated Execution: The automated execution engine parses the decision instructions and executes them in two steps: First, it calls the city's telecommunications operator's router API to execute a signal instruction compression script, enabling the LZ77 compression algorithm (compression rate 60%) to compress the signal control instructions from the original 1024 bytes to 410 bytes. This reduces link bandwidth utilization from 99% to 65% and transmission latency from 200ms to 60ms, taking 2 seconds. Second, it calls the Siemens SC2000 signal controller's NTCIP protocol interface to execute a phase adjustment script, extending the green light duration for eastbound traffic at the commercial center intersection from 60s to 70s. Simultaneously, it adjusts the green light durations at the adjacent "department store intersection" and "park intersection" (extending them by 5s each) to prevent traffic congestion from spreading, taking 5 seconds. During execution, the signal synchronization status is monitored in real time: the signal synchronization rate at the 20 intersections remains at 98%, and the queue length for eastbound traffic is updated every 10 seconds. After execution, the queue length is reduced from 200m to 120m within 5 minutes.

[0045] 5. Dynamic Optimization: Data collection and execution results show a 100% fault recovery success rate (signal command execution response time restored to 80ms, bandwidth utilization stabilized at 65%), maintenance time of 7 seconds (meeting P0 level response time limits), and an increase in average east-to-west traffic speed during morning peak hours from 25km / h to 32km / h (meeting standards). Traffic flow data transmission accuracy is 99.3%. The decision-making model is optimized based on the gradient descent algorithm (η=0.008, momentum 0.9): the weight of "link data credibility" is increased from 0.95 to 0.97 to enhance the impact of communication link status on decision-making; the weight of "real-time coefficient P" in the reward function is adjusted from 0.05 to 0.08 to further strengthen response time requirements. Based on feedback from the traffic police detachment (congestion at the commercial center intersection during the morning rush hour has eased, and no chain-reaction congestion has occurred), the knowledge graph association rules have been updated: the association confidence of the strategy of "morning rush hour + link saturation + traffic congestion" with "command compression + phase adjustment" has been increased from 0.8 to 0.98. At the same time, the rule of "signal coordination between adjacent intersections during peak hours" has been added (to avoid traffic accumulation caused by adjustments at a single intersection).

[0046] Implementation Results: In this scenario, the traffic signal control system achieved an 88% self-healing rate for automated faults. Specifically, the self-healing rate for P0-level faults (main road signal faults) was 100%, with response times ≤20s; the self-healing rate for P1-level faults (traffic flow data interruption) was 82%, with a response time ≤45s, representing a reduction of over 94% compared to traditional manual maintenance (response time 8 minutes). The average traffic speed on main roads during morning rush hour increased from 25km / h to 32km / h, and congestion duration decreased from 60 minutes to 35 minutes (a 40% reduction). The signal synchronization rate at 20 intersections remained stable above 98%. The accuracy of traffic flow data transmission improved from 85% to 99.3%, providing high-quality data support for the traffic guidance system. The average morning rush hour commute time for citizens was reduced by 15 minutes, and the traffic violation rate decreased by 10%. During the six months of system operation, over 800 faults were handled, with only 80 instances of manual intervention (all due to equipment hardware failures), reducing the workload of traffic police command center maintenance personnel by 70%.

[0047] Example 3: Operation and Maintenance Scenario of Core Business Systems in the Financial Industry (Banking) This embodiment is applied to the core business system of a provincial branch of a state-owned bank, including a comprehensive business system (supporting deposit, loan, and payment services, with an average of 10 million+ transactions per day), a credit card system (supporting bill processing for 5 million+ users), and a data center disaster recovery system (active-active architecture in different locations). The operation and maintenance requirements are: fault response time limit of P0 level (payment service interruption) ≤30s, P1 level (non-core business fault) ≤5min, data integration accuracy ≥99%, and transaction success rate ≥99.99%.

[0048] System Deployment: The fault intelligent perception module deploys 120+ monitoring nodes, with the core transaction server monitored 1 second / time, the payment interface monitored 500ms / time, and the database (Oracle RAC cluster) monitored 1 second / time; the multi-source data fusion processing module deploys 6 edge nodes, partitioned by business line (comprehensive business, credit card, fund clearing); the intelligent scheduling decision module preloads 300+ financial fault types into its knowledge graph; the automated execution engine has built-in templates for payment interface restart, database master-slave switching, etc.; the dynamic optimization module triggers parameter optimization once every 30 maintenance tasks.

[0049] Specific implementation process: 1. Fault detection and feature extraction: When the payment interface response delay exceeds 3 seconds (threshold 500ms) and the transaction failure rate reaches 10% (threshold 0.1%), the feature extraction network strengthens the features of "interface delay" and "transaction failure" to generate fault feature vectors.

[0050] 2. Multi-source data fusion processing: Edge nodes receive payment logs and database connection data, integrate them through a dynamic weight mapping algorithm, repair two missing database connection data entries, and filter one abnormal transaction log entry.

[0051] 3. Intelligent scheduling decision: The knowledge graph matches the "payment interface connection pool exhaustion" fault, and the candidate strategies are "connection pool expansion" and "interface restart". After reinforcement learning calculates the reward value, the combination of "connection pool expansion + interface restart" is selected with priority P0.

[0052] 4. Automated execution: Call the database API to expand the connection pool (from 500 to 1000), restart the payment interface, which takes 25 seconds.

[0053] 5. Dynamic optimization: 100% success rate in fault recovery, optimized decision model parameters, and increased weighting of payment interface data credibility.

[0054] Implementation results: 92% fault self-healing rate, payment service interruption time ≤25s, transaction success rate 99.995%, data integration accuracy rate 99.5%, and 85% improvement in operation and maintenance efficiency compared to traditional methods.

[0055] Comparison Example To verify the technical superiority of the present invention, the following six sets of control examples were set up and compared with the above-mentioned Example 1 (medical), Example 2 (transportation), and Example 3 (finance) for a period of 3 months: Compare with Example 1: Traditional manual + fixed script operation and maintenance solution (medical industry) Technical solution: Manual inspection (2 hours / time) + fixed script (only supports HIS system restart and PACS service restart), data integration adopts centralized ETL (no privacy protection), fault decision-making relies on the experience of operation and maintenance personnel, and there is no dynamic optimization mechanism.

[0056] Test results: Fault self-healing rate 30% (can only handle 2 types of simple faults), P0 level fault response time 15min (far exceeds the 1min standard), high risk of data privacy leakage (transmission without anonymization); PACS system image transmission success rate 85%, emergency image acquisition delay exceeds 10min, affecting 10 emergency treatments; maintenance personnel handle an average of 20+ faults per day, with a heavy workload.

[0057] Compare with Example 2: General AI Operation and Maintenance Solution (No industry customization, medical industry) Technical solution: LSTM fault diagnosis model is adopted (without medical-specific feature enhancement), data integration adopts general federated learning (without privacy compliance optimization), the decision model has no industry rules, and the execution latency is uniformly set to 500ms.

[0058] Test results: Fault self-healing rate 60% (accuracy rate of complex fault diagnosis 55%), P0 level fault response time 2min (not up to standard), data privacy compliance 80% (some image data was not anonymized); image transmission success rate 92%, emergency image acquisition delay 2min, manual intervention is still required for 30% of faults.

[0059] Compare with Example 3: Traditional manual + timed monitoring and maintenance solution (transportation industry) Technical solution: It adopts timed monitoring (5 minutes / time) + manual remote operation (manually issuing instructions through the traffic police command platform), data integration adopts local storage (no edge processing), and signal synchronization relies on daily fixed time (2 am) calibration.

[0060] Test results: Fault self-healing rate 25% (can only handle signal restart at a single intersection), P0 level fault response time 8min (far exceeds the 30s standard), signal synchronization rate 85%; average traffic speed on main roads during morning rush hour 22km / h, congestion duration 80min, 2000+ vehicles stranded due to signal faults.

[0061] Compare with Example 4: General Intelligent Operation and Maintenance Solution (No industry customization, transportation industry) Technical solution: CNN fault feature extraction is adopted (without traffic-specific feature weights), data integration adopts centralized processing (transmission delay of 200ms), the decision model reward function has no real-time coefficient, and the execution delay is uniformly set to 200ms.

[0062] Test results: Fault self-healing rate 55% (signal phase adjustment accuracy 60%), P0 level fault response time 1 minute (not up to standard), signal synchronization rate 90%; average traffic speed on main roads during morning rush hour 28 km / h, congestion duration 50 minutes, manual intervention is still required for 40% of faults.

[0063] Compare with Example 5: Traditional fixed-script operation and maintenance solution (financial industry) Technical solution: Pre-set scripts are used (only handling service restarts and disk cleanup), data integration is done using the DataX tool (without quality optimization), and decision-making relies on human judgment.

[0064] Test results: Fault self-healing rate 35%, payment service interruption time 5 minutes, transaction success rate 99.9%, data integration accuracy 95%, and manual intervention rate 60%.

[0065] Compare with Example 6: Single Machine Learning Operation and Maintenance Solution (Financial Industry) Technical solution: The MLP fault prediction model is adopted, but the data integration has no dynamic weights and the decision-making has no strategy combination optimization.

[0066] Test results: Fault self-healing rate 70%, payment service interruption time 1 minute, transaction success rate 99.95%, data integration accuracy 98%, and manual intervention rate 30%.

[0067] Comparison Results Analysis As can be seen from the comparison, this invention demonstrates significant advantages in various industries: In the healthcare industry: the self-healing rate of faults increased by 60 percentage points compared to control case 1 and by 30 percentage points compared to control case 2; the response time for P0-level faults was shortened by 97% compared to control case 1 and by 75% compared to control case 2; data privacy compliance reached 100%, an improvement of 20 percentage points compared to control case 2; and the time for acquiring emergency images was shortened by 92% compared to control case 1 and by 79% compared to control case 2.

[0068] Transportation sector: Fault self-healing rate increased by 63 percentage points compared to control case 3 and by 33 percentage points compared to control case 4; P0 level fault response time was shortened by 96% compared to control case 3 and by 83% compared to control case 4; morning peak traffic speed increased by 28% compared to control case 3 and by 14% compared to control case 4; congestion duration decreased by 44% compared to control case 3 and by 29% compared to control case 4.

[0069] Financial industry: The fault self-healing rate increased by 57 percentage points compared to control case 5 and by 22 percentage points compared to control case 6; the payment service interruption time was shortened by 92% compared to control case 5 and by 75% compared to control case 6; the transaction success rate increased by 0.095 percentage points compared to control case 5 and by 0.045 percentage points compared to control case 6.

[0070] The system of this invention can use Java (core module) and Python (machine learning module) as development languages, and implement distributed deployment based on Spring Cloud microservice architecture. The database uses MySQL (structured data), Neo4j (knowledge graph), and Redis (caching). The edge nodes of the multi-source data fusion processing module can use Raspberry Pi (lightweight) or industrial servers (high performance), support Docker containerized deployment, and can elastically expand the number of nodes according to the industry scale (one edge node for every five departments in the medical industry, and one edge node for every four intersections in the transportation industry).

[0071] Deployment in each industry must comply with the corresponding industry standards: the healthcare industry must comply with the "Basic Functional Specifications for Hospital Information Systems" and the "Guidelines for Medical Data Security," and support HL7 FHIR and DICOM 3.0 standards; the transportation industry must comply with the "Technical Requirements for Urban Traffic Signal Control Systems" (GB / T 29870-2013), and support the NTCIP 1202 protocol and the GB / T 20999 traffic data interface standard; the financial industry must comply with the "Guidelines for Information Technology Risk Management of Banking Financial Institutions," and support financial transaction data encryption standards (such as SM4).

[0072] The scope of protection of this invention is not limited to the above embodiments. For those skilled in the art, various modifications and variations can be made to the system modules, algorithm parameters, and industry adaptation details without departing from the principle of this invention. All such modifications and variations should be considered within the scope of protection of this invention.

Claims

1. An intelligent scheduling management system for information system operation and maintenance, characterized in that, The system comprises: a fault intelligent perception module for collecting operation state data, log data and fault alarm data of the information system through multi-dimensional monitoring nodes, and extracting fault key features based on an attention mechanism feature extraction network; a multi-source data fusion processing module adopting an improved federated learning framework, performing localized preprocessing of heterogeneous data sources through edge nodes, realizing data format standardization based on a dynamic weight mapping algorithm, repairing data missing values and filtering noise data by combining a Bayesian inference model; an intelligent scheduling decision module for constructing a decision model combining a knowledge graph and deep reinforcement learning, the knowledge graph storing fault types, recovery strategies and system topology relationships, and the reinforcement learning agent dynamically adjusting scheduling priorities and recovery strategy combinations by interacting with the operation and maintenance environment; an automatic execution engine for analyzing scheduling instructions output by the decision module, calling a preset automatic script library, API interface and resource scheduling component, and executing fault recovery operations and data integration tasks; a dynamic optimization module for collecting real-time operation and maintenance execution effect data, optimizing decision model parameters and data fusion weights based on a gradient descent algorithm with a learning rate η=0.008 and a momentum parameter 0.9, and updating the knowledge graph and the automatic strategy library; The modules interact and work cooperatively through a distributed message queue, supporting horizontal expansion and cross-platform deployment. 2.The information system operation and maintenance intelligent scheduling management system of claim 1, wherein: The multi-dimensional monitoring nodes of the fault intelligent perception module include system layer monitoring units, application layer monitoring units and network layer monitoring units, the system layer monitoring units collect data at a frequency of 1s / time, the application layer monitoring units collect data at a frequency of 2s / time, and the network layer monitoring units collect data at a frequency of 500ms / time, the monitoring frequency of medical industry image devices is 500ms / time, and the monitoring frequency of traffic industry signal control is 300ms / time, the feature extraction network strengthens the weight distribution of fault related features through a self-attention mechanism, and the feature extraction accuracy is ≥92%. 3.The information system operation and maintenance intelligent scheduling management system of claim 1, characterized in that: The improved federated learning framework of the multi-source data fusion processing module protects data privacy through a differential privacy algorithm, the localized preprocessing includes data format verification, redundant data elimination and semantic annotation, the edge nodes are deployed by department in the medical industry and by road section in the traffic industry. 4.The information system operation and maintenance intelligent scheduling management system of claim 1, characterized in that: The knowledge graph of the intelligent scheduling decision module includes an entity layer, a relationship layer and an attribute layer, the entity layer includes fault entities, resource entities and strategy entities, the medical industry extends the entities to include "electronic medical record system fault entities", "image device resource entities" and "ICU device resource entities", the traffic industry extends the entities to include "signal control fault entities", "traffic flow data entities" and "intersection signal entities", the relationship layer defines the association rules between entities, adds "medical data privacy association rules" and "traffic signal linkage rules", and the attribute layer stores the feature parameters and constraint conditions of the entities. 5.The information system operation and maintenance intelligent scheduling management system of claim 1, characterized in that: The script library of the automation execution engine supports Python, Shell, and PowerShell multi-language scripts, built-in fault recovery templates include four core scenarios of service restart, resource expansion, configuration rollback, and data synchronization, medical industry exclusive templates include "image storage service restart", "electronic medical record data synchronization", and "DICOM file synchronization", traffic industry exclusive templates include "signal control instruction retransmission", "vehicle flow data link repair", and "signal phase adjustment", and support medical equipment manufacturer exclusive API and traffic signal controller protocol.

6. The information system operation and maintenance intelligent scheduling management method, the method is implemented by using the information system operation and maintenance intelligent scheduling management system of any one of claims 1-5, characterized in that, The method comprises the following steps: S1, fault perception and feature extraction: through multi-dimensional monitoring nodes, real-time collection of information system CPU utilization, memory occupancy, disk IO, network bandwidth, application logs, and alarm information, system layer data collection frequency 1s / time, application layer data collection frequency 2s / time, network layer data collection frequency 500ms / time, medical industry focuses on collecting image equipment CPU load, medical record database connection number, and PACS system response time, traffic industry focuses on collecting signal controller state, vehicle flow detector data, and intersection passing time, and a fault key feature vector with a dimension of 128 is generated by using an attention mechanism feature extraction network to extract fault key features, the medical industry feature vector contains 20 industry exclusive features, and the traffic industry contains 25 industry exclusive features; S2, multi-source data fusion processing: the edge node receives data of heterogeneous data sources, performs localized preprocessing, the medical industry completes DICOM file format verification and patient privacy information desensitization, the traffic industry completes signal instruction binary to JSON and vehicle flow data deduplication, maps different format data to a unified standard through a dynamic weight mapping algorithm, the medical industry increases the "patient privacy compliance" score index, the traffic industry increases the "data real-time" score index, adopts a Bayesian inference model to repair missing data and filter noise data, and completes data security aggregation based on an improved federal learning framework, the medical industry adopts "department-level-hospital-level" two-level encryption transmission, and the traffic industry adopts "road segment-level-regional-level" two-level aggregation; S3, intelligent scheduling decision: input the fault feature vector into the decision model of the fusion knowledge graph and deep reinforcement learning, the knowledge graph matches the candidate recovery strategy, the reinforcement learning agent combines the system real-time load, resource margin, and fault emergency degree, the medical industry prioritizes ICU device operation and maintenance resources, the traffic industry adjusts the priority in combination with real-time vehicle flow data, dynamically adjusts the scheduling priority, and outputs the optimal recovery strategy combination and data integration scheme; the reward function of deep reinforcement learning is set as: R=α×S+β×T-γ×C-δ×P, P is the privacy risk coefficient in the medical industry, P is the real-time coefficient in the traffic industry, α=0.5, β=0.3, γ=0.15, δ=0.05, and α+β+γ+δ=1; S4, automated execution: the automated execution engine parses the decision instructions, calls the corresponding scripts, API interfaces or resource components, executes the fault recovery operations and data integration tasks, and the execution delay of the medical industry is ≤300ms, and the execution delay of the traffic industry is ≤100ms; S5, dynamic optimization: collect execution result data, optimize decision model parameters and data fusion weights based on gradient descent algorithm, trigger parameter optimization once every 20 operation and maintenance tasks in the medical industry, trigger parameter optimization once every 10 operation and maintenance tasks in the traffic industry, and update the correlation rule confidence of the knowledge graph based on the feedback of industry experts, update the entity relationship and strategy library of the knowledge graph; S6, cyclic iteration: repeat steps S1-S5 to realize continuous optimization and adaptive adjustment of operation and maintenance scheduling.

7. The information system operation and maintenance intelligent scheduling management method of claim 6, characterized in that: In step S2, the data format standardization adopts JSON-LD unified data semantic description format, and the medical industry extends the JSON-LD context to include "DICOM standard terms", and the traffic industry extends the JSON-LD context to include "traffic signal control terms". 8.The information system operation and maintenance intelligent scheduling management method of claim 6, characterized in that: In step S3, the fault classification scheduling divides the fault into P1 emergency fault, P2 important fault, P3 general fault and P4 slight fault, and the medical industry adds "P0 life support system fault" level, and the traffic industry adds "P0 main road signal fault" level, different levels correspond to different scheduling priorities and response time limits. 9.The information system operation and maintenance intelligent scheduling management method of claim 6, characterized in that: In step S5, the medical industry expert feedback includes doctors' operation and maintenance suggestions for medical record systems, and the traffic industry expert feedback includes traffic police's adjustment requirements for signal control, and the knowledge graph correlation rule confidence is updated based on the effectiveness verification of the feedback results. 10.The information system operation and maintenance intelligent scheduling management method of claim 6, characterized in that: In step S2, the medical industry patient privacy information desensitization is executed according to the HIPAA standard, and the traffic industry signal instruction transmission adopts MQTT protocol to realize low delay data interaction.

Images