A deep learning-based machine room fault disposal decision support system

Abstract

The application relates to the technical field of operation and maintenance management, and discloses a machine room fault disposal decision support system based on deep learning, which comprises a global state sensing and fault prediction module, a cause-effect diagnosis and decision generation module, an autonomous decision and closed-loop execution module, and a dynamic optimization and knowledge evolution module. Through deep learning and multivariate anomaly detection, the application realizes the transformation from passive alarm to active early warning, and can discover implicit faults earlier; by using cause-effect discovery and graph reasoning technology, the root cause can be quickly located, and the average repair time can be greatly shortened; the generated disposal scheme is based on multi-objective optimization sorting, and an explanation based on counterfactual reasoning is provided, so that the administrator can understand and trust the recommendation of the system, and intelligent decision-making of man-machine cooperation is realized; and an equal energy efficiency index is directly integrated into a reward function of the decision model, so that the fault disposal strategy automatically tends to be optimal in energy efficiency while ensuring stability.

Description

Technical Field

[0001] This invention relates to the field of data center operation and maintenance management technology, specifically a data center fault handling decision support system based on deep learning. Background Technology

[0002] With the widespread adoption of cloud computing, big data, and artificial intelligence technologies, the scale and complexity of modern data centers are growing exponentially. These data centers contain a large number of servers, network devices, storage devices, and power and environmental facilities such as air conditioning and UPS systems, with complex coupling and dependencies among these devices. Once a failure occurs, its rapid propagation and wide-ranging impact pose a serious threat to business continuity and stability. Currently, traditional data center operation and maintenance management mainly faces the following challenges: 1. Delayed fault detection: Most monitoring systems rely on alarms based on static thresholds, which cannot effectively predict hidden faults or detect abnormal correlations between multiple indicators. Alarms are often triggered only after the fault has already affected the business. 2. Difficulty in root cause identification: When a failure occurs, maintenance personnel need to manually analyze the failure propagation chain from massive and heterogeneous monitoring data. Locating the root cause is time-consuming and difficult, resulting in a high mean time to repair (MTTR). 3. Decision-making relies heavily on personal experience: The handling solutions depend heavily on the personal experience of operations and maintenance experts, lacking systematic decision support. When faced with multiple trade-offs (such as repair speed, business impact, and cost), it is difficult to quickly generate and evaluate the optimal solution. 4. High execution risk: Automated operation and maintenance scripts lack sufficient security verification and context awareness capabilities, and blind execution may cause secondary failures such as the "avalanche effect".

[0003] 5. Disconnect between energy efficiency management and fault handling: Traditional energy-saving strategies and fault handling systems are independent of each other, making it difficult to optimize energy efficiency during the fault handling process while ensuring equipment stability.

[0004] For example, a Chinese patent discloses an intelligent decision support system for equipment operation and maintenance based on large-scale model inference (publication number CN120105294A). This patented technology acquires equipment operation data in real time through a data acquisition module, uses a large-scale pre-trained deep learning model to infer and analyze the health status of the equipment, and predicts the remaining service life and various fault types. The system combines multi-task learning and multi-objective optimization algorithms to intelligently optimize operation and maintenance decisions, generate personalized maintenance suggestions, and help operation and maintenance personnel effectively reduce equipment downtime, lower maintenance costs, and extend equipment life.

[0005] However, this patent also has drawbacks, such as low fault diagnosis accuracy, inability to achieve intelligent decision-making through human-machine collaboration, resulting in poor fault handling performance; and inability to achieve green, energy-saving and environmentally friendly operation of the computer room. Summary of the Invention

[0006] The purpose of this invention is to provide a data center fault handling decision support system based on deep learning to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based decision support system for handling data center faults includes a global state perception and fault prediction module, a causal diagnosis and decision generation module, an autonomous decision-making and closed-loop execution module, and a dynamic optimization and knowledge evolution module. The causal diagnosis and decision generation module comprises a causal root cause localization unit and a multi-objective trade-off handling unit. Causal root cause localization unit: Constructs a service dependency graph with temporal relationships using a causal discovery algorithm; When a fault occurs, analyzes abnormal behavior of indicators and simulates the fault propagation path through graph reasoning technology to quickly locate the root cause node; Multi-objective trade-off handling unit: It associates each diagnostic result with a multi-level set of handling options generated by expert experience and reinforcement learning, generates multiple decision options ranked by indicators, and explains the reasons for the ranking of options and the recommended options; The autonomous decision-making and closed-loop execution module consists of a hierarchical decision-making and safe execution unit and a digital twin-based energy efficiency optimization unit. Tiered decision-making and security execution unit: Based on the risk level of the data center failure or the application scenario level, and in conjunction with the security interlock mechanism, select the appropriate execution decision scheme; Based on digital twin energy efficiency optimization unit: Construct a simulation environment of digital twin of computer room, integrate energy saving and emission reduction indicators as one of the core optimization objectives into the reward function; and train the hierarchical decision execution unit to formulate a dynamic strategy with optimal energy efficiency under the premise of ensuring stable operation of equipment.

[0008] As a further aspect of the present invention: the global state perception and fault prediction module comprises a heterogeneous data acquisition and aggregation unit, a dynamic health assessment unit, and an early warning knowledge base driving unit, wherein, Heterogeneous data acquisition and aggregation unit: collects data on the performance indicators of computer room equipment, power and environmental data, network and security data, as well as application and business data; performs unified cleaning, labeling and storage of the collected data, and uses data lineage analysis technology to trace the source and transformation process of the stored data to ensure the reliability of subsequent analysis; Dynamic Health Assessment Unit: Utilizes a time-series deep learning model to analyze historical and real-time monitoring data, and combines it with a multivariate time-series anomaly detection model to construct a dynamic learning device health model, dynamically predicting the health of data center equipment; Security baseline construction unit: Analyze system logs and combine them with user and traffic behavior analysis to build a security baseline; establish normal behavior profiles for each user and each IP through unsupervised learning, thereby accurately identifying potential malicious activities and early signs of distributed denial-of-service attacks; Early warning knowledge base driven unit: Combines operation and maintenance experience with dynamic warning values, and constructs an early warning knowledge base based on quantitative risk assessment of model confidence; and proactively pushes emergency alert information to administrators; As a further aspect of the present invention: in the heterogeneous data acquisition and aggregation unit, data lineage analysis technology is used to trace the source and transformation process of the stored data to ensure the reliability of subsequent analysis; the specific steps are as follows: S1. Metadata Collection and Aggregation: Register all data sources in the system's data directory or data catalog; and collect multi-level metadata; S2. Lineage Discovery and Resolution: The data is parsed using SQL scripts, programming code, and workflow dependencies to construct multi-dimensional data lineage relationships. Lineage relationships from different sources are then cleaned, transformed, and standardized to form lineage model data. S3. Construction, storage and visualization of kinship maps: Model kinship model data into kinship maps; then integrate the kinship maps with the system's data catalog or data governance platform, and provide a visualization interface.

[0009] As a further aspect of the present invention: in the dynamic health assessment unit, the formula for calculating health level is as follows: ; In the above formula, , and These are the weighting coefficients for failure rate, performance health, and lifespan health, respectively. The weighting coefficients are dynamically adjusted based on equipment type and business importance through an online learning mechanism. Failure rate is the failure classification probability output by the equipment health model; performance health is calculated based on the distance between key performance indicators and failure warning thresholds; and lifespan health is the remaining service life / design life.

[0010] As a further aspect of the present invention: in the early warning knowledge base driving unit, the construction rules of the early warning knowledge base include static rules, dynamic thresholds, and scenario rules, wherein, Static rules are entered from the experience-based rules of operations and maintenance experts; Dynamic thresholds utilize unsupervised learning algorithms to learn a dynamic, personalized normal range for each device and each metric; The scenario rules are automatically generated by the system from historical failure cases, describing the normal change patterns of indicators under specific business scenarios, thus avoiding false alarms.

[0011] As a further aspect of the present invention: in the multi-objective trade-off handling unit, the decision-making scheme is based on the effect ranking and introduces the idea of ​​multi-objective optimization, while comprehensively considering the handling time, cost, risk and impact on the service level agreement. The explanatory power of the solution is achieved through counterfactual reasoning.

[0012] As a further aspect of the present invention: in the hierarchical decision-making and security execution unit, the risk level of data center faults is divided into low risk, medium risk, and high risk, and the application scenario level is divided into high-determinism scenarios, medium-determinism scenarios, and low-determinism scenarios, wherein, For low-risk or high-certainty scenarios, the system automatically executes the recommended decision-making scheme; For medium-risk or medium-certainty scenarios, a human-machine collaborative approach is adopted to execute decision-making schemes; For high-risk or low-certainty scenarios, a supplementary reporting method is used to implement decision-making plans.

[0013] As a further aspect of the present invention: the dynamic optimization and knowledge evolution module consists of a model dynamic optimization unit, an intelligent data lake warehouse unit, and an intelligent assessment and carbon footprint audit unit, wherein, The model dynamic optimization unit records the actual effect of each action and uses the feedback information to fine-tune the diagnostic and decision-making model. It also drives online learning, enabling the model to quickly adapt to changes in the data center environment without forgetting old knowledge. Furthermore, it establishes a fault handling case library and records the entire process of each successful diagnosis-decision-execution-feedback in a structured manner.

[0014] As a further aspect of the present invention: In the dynamic optimization and knowledge evolution module, the intelligent data lake warehouse unit adopts a hybrid architecture of data lake and data warehouse, storing the raw data in the data lake for in-depth mining; and storing the processed feature data and graph data in the time series database and graph database respectively to support high-performance real-time query and complex analysis.

[0015] As a further aspect of the present invention: In the dynamic optimization and knowledge evolution module, the intelligent assessment and carbon footprint audit unit assesses the equipment in the data center and audits the energy efficiency of the data center in conjunction with the carbon footprint calculation model, generating a multi-dimensional energy consumption and carbon emission report.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention achieves a shift from passive alarm to proactive early warning through deep learning and multivariate anomaly detection, enabling the earlier detection of hidden faults and preventing problems before they occur. By utilizing causal discovery and graph reasoning techniques, we can clearly reveal the path of fault propagation, quickly locate the root cause, and significantly shorten the average repair time. The generated disposal plan is based on multi-objective optimization ranking and provides an explanation based on counterfactual reasoning, enabling administrators to understand and trust the system's recommendations, thus realizing intelligent decision-making through human-machine collaboration. Through a hierarchical decision-making mechanism and safety interlocks, the safety of automated operations is ensured, forming a complete closed loop of "perception-diagnosis-decision-execution-optimization" and reducing the risk of human error. By directly integrating energy efficiency indicators into the reward function of the decision-making model, the fault handling strategy automatically tends towards optimal energy efficiency while ensuring stability, thus achieving a balance between business continuity and green energy conservation. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of a deep learning-based decision support system for handling computer room faults. Detailed Implementation

[0018] Please see Figure 1In this embodiment of the invention, a deep learning-based decision support system for handling data center faults includes a global state perception and fault prediction module, a causal diagnosis and decision generation module, an autonomous decision-making and closed-loop execution module, and a dynamic optimization and knowledge evolution module. The global state perception and fault prediction module comprises a heterogeneous data acquisition and aggregation unit, a dynamic health assessment unit, and an early warning knowledge base driving unit. Specifically, the heterogeneous data acquisition and aggregation unit collects data center equipment performance indicators, power environment data, network and security data, and application and business data. For example, it collects IT equipment performance data such as CPU, memory, disk I / O, and network traffic of servers via Agent or SNMP protocol; it collects power environment data such as air conditioners, UPS, power distribution cabinets, temperature and humidity sensors, and water leakage detectors via a power environment monitoring system; and it collects network equipment traffic, NetFlow / sFlow data, etc. Network and security data such as firewall logs, system security logs, and application logs; application and business data such as website access logs, business throughput, and transaction response time; the collected data is uniformly cleaned, labeled, and stored, and data lineage analysis technology is used to trace the source and transformation process of the stored data to ensure the reliability of subsequent analysis; Dynamic health assessment unit: using time-series deep learning models to analyze historical and real-time monitoring data, and combining multivariate time-series anomaly detection models to build dynamic learning device health models to dynamically predict the health of data center equipment; among them, time-series deep learning models such as LSTM or Transformer; multivariate time-series anomaly detection models (such as VAE-based or GAN-based models) not only analyze individual indicators, but also pay attention to the abnormal correlation between multiple indicators, which can detect hidden faults earlier; such as low CPU utilization but extremely high system load.

[0019] The formula for calculating health is as follows: .

[0020] In the above formula, , and These are the weighting coefficients for failure rate, performance health, and lifespan health, respectively. These weighting coefficients are dynamically adjusted based on device type and business importance through an online learning mechanism, rather than being fixed values. Failure rate is the failure classification probability output by the device health model. Performance health is calculated based on the distance between key performance indicators and failure warning thresholds. Key performance indicators include CPU utilization, memory utilization, disk I / O throughput, and network traffic. Lifespan health is the remaining service life divided by the design life.

[0021] Security baseline construction unit: Parses system logs and combines them with user and traffic behavior analysis, such as using NLP models to analyze log sequences and using clustering-based behavior analysis models to analyze user and traffic behavior; constructs security baselines, such as normal log streams and network behavior baselines, and any behavior that significantly deviates from the baseline will be marked as abnormal; for example, detecting crawler behavior, low-frequency brute-force attacks, etc.; establishes normal behavior profiles for each user and each IP through unsupervised learning, thereby accurately identifying potential malicious activities and early signs of Distributed Denial of Service (DDoS) attacks.

[0022] The early warning knowledge base driving unit combines operational experience with dynamic alert values, constructs an early warning knowledge base based on quantitative risk assessment using model confidence levels, and proactively pushes emergency alert information to administrators. This information includes the event type, associated resources, risk level, evidence chain, and preliminary suggested contingency plans. Examples of event types include hardware alerts, performance alerts, and security threats. Associated resources clearly identify the fault prediction targets, such as server IP addresses and disk serial numbers. Risk levels are categorized as high, medium, and low based on model confidence levels and potential impact. The evidence chain lists key indicators and log fragments leading to the alert. Preliminary suggested contingency plans provide initial handling recommendations, such as checking disk SMART information, rather than simply generating threshold exceedance alerts. The rules for constructing the early warning knowledge base include static rules, dynamic thresholds, and scenario rules, among which: Static rules are entered from the experience-based rules of operations and maintenance experts, such as disk utilization exceeding 90% and continuously increasing. Dynamic thresholds utilize unsupervised learning algorithms (such as Isolation Forest and Autoencoder) to learn a dynamic, personalized normal range for each device and each metric, replacing fixed static thresholds.

[0023] The scenario rules are automatically generated by the system from historical failure cases, describing the normal change patterns of indicators under specific business scenarios (such as "Double Eleven promotion" and "month-end settlement"), thus avoiding false alarms.

[0024] Preferably, in the heterogeneous data acquisition and aggregation unit, data lineage analysis technology is used to trace the source and transformation process of the stored data to ensure the reliability of subsequent analysis; the specific steps are as follows: S1. Metadata Collection and Aggregation: Register all data sources in the system's data directory or data catalog; and collect multi-level metadata, such as structural metadata (including database name, table name, field name, data type, primary and foreign key constraints, etc.), business metadata (including field business definitions, business terms, data owner, data sensitivity labels), process metadata (ETL / ELT tasks, data development scripts, and BI and reporting tools), and operational metadata (data creation time, update time, access logs, lifecycle status, data size changes, etc.).

[0025] S2. Lineage Discovery and Resolution: The data is parsed using SQL scripts, programming code, and workflow dependencies to construct multi-dimensional data lineage relationships. The lineage relationships from different sources (SQL parsing, code parsing, workflow parsing) are then cleaned, transformed, and standardized to form lineage model data. For example, in SQL script parsing, an SQL parser (such as Apache Calcite) is used to perform lexical and syntactic analysis on the SQL code in ETL tasks or query scripts; establish the dependency relationship between the target table and the source table; and identify which fields in the source table a certain field in the target table is generated by and what kind of calculations (such as functions, expressions, join operations) from which fields in the source table. For example, in programming code parsing, by constructing an abstract syntax tree, we can analyze the data read and write operations in the code, thereby establishing the relationship between code logic and data entities; In workflow dependency resolution, for example, by resolving the upstream and downstream relationships between task nodes, a task-level lineage is constructed, meaning that the output of one task is the input of another task; this complements the data-level lineage, forming a complete processing chain.

[0026] S3. Construction, Storage, and Visualization of the Lineage Graph: The lineage model data is modeled into a lineage graph. In this graph, nodes represent various data entities, such as databases, tables, fields, BI reports, and data processing tasks; edges represent relationships between entities, such as originating from, written to, or generated by a task; edges can carry attributes, such as transformation logic and data update time; the lineage graph is then integrated with the system's data directory or data governance platform, and a visual interface is provided; when a user clicks on any table or field, the system can graphically display the complete lineage chain centered on that table: tracing its source upstream and exploring all its influence downstream.

[0027] The causal diagnosis and decision generation module consists of a causal root cause localization unit and a multi-objective trade-off unit. The causal root cause localization unit employs causal discovery algorithms (such as PC algorithm, Fast Causal algorithm, etc.). Inference or NOTEARS-based continuous optimization methods are used to construct a service dependency graph with temporal relationships. When a fault occurs, abnormal behavior of indicators is analyzed, and the fault propagation path is simulated through graph reasoning techniques (such as random walks and graph neural networks) to quickly locate the root cause node and significantly reduce repair time. A multi-objective trade-off handling unit associates each diagnostic result with a multi-level set of handling solutions generated by expert experience and reinforcement learning, generating multiple decision solutions ranked by indicators. The unit explains the ranking of solutions and the reasons for recommending solutions, such as the top-ranked solution effectively isolating the fault domain and minimizing the impact on core business. The decision solutions are ranked based on effectiveness and incorporate multi-objective optimization, comprehensively considering handling time, cost, risk, and impact on service level agreements (SLAs). For example, one solution may repair the fault the fastest but have a short-term impact on business, while another solution is imperceptible but takes longer. The explanatory power of the solutions is achieved through counterfactual reasoning. For example, the system might generate the following explanation: "Recommended solution A (restart service) because if the CPU load had been 10% lower at the time (counterfactual condition), there would have been a 95% probability that this fault would not have occurred."

[0028] Preferably, the autonomous decision-making and closed-loop execution module consists of a hierarchical decision-making and secure execution unit and a digital twin-based energy efficiency optimization unit. The hierarchical decision-making and secure execution unit selects the appropriate execution decision scheme based on the risk level of the data center failure or the application scenario level, combined with the security interlock mechanism. For example, according to the security interlock mechanism, before executing any automatic operation, the system will perform a final quick health check. If the conditions change abruptly, such as the target server suddenly losing power, the operation will be stopped and an alarm will be issued.

[0029] The risk levels of data center failures are categorized as low, medium, and high, while application scenario levels are categorized as high-determinism, medium-determinism, and low-determinism. For low-risk or high-determinism scenarios, the system automatically executes recommended decision-making solutions; for example, if a non-core service outage is detected, it is automatically restarted; if localized overheating is detected, the airflow and temperature of the designated area's intelligent air conditioning are automatically adjusted. For medium-risk or medium-determinism scenarios, a human-machine collaborative approach is used to execute decision-making solutions. If a surge in application load is detected, the system provides multiple contingency plans, including vertical scaling, horizontal scaling, and traffic shaping, along with their estimated costs and impacts, allowing administrators to choose which solution to implement. In human-machine collaborative scenarios, the system provides an A / B testing sandbox for decision-making solutions, allowing administrators to simulate the effects of different solutions in a digital twin environment before making a choice, significantly reducing decision-making risk. For high-risk or low-determinism scenarios, an auxiliary reporting approach is used to execute decision-making solutions. If a complex security attack chain is discovered, the system provides a comprehensive impact surface analysis and forensic report, which is then submitted to security experts for decision-making.

[0030] Based on digital twin energy efficiency optimization unit: Construct a simulation environment of digital twin of computer room, and incorporate energy saving and emission reduction indicators as one of the core optimization objectives into the reward function; The formula for calculating the reward function is as follows:

[0031] In the above formula, The total reward is the final score obtained by the reinforcement learning agent after taking a certain action (such as adjusting the air conditioner setting temperature); the agent's goal is to maximize the long-term cumulative reward. This is an energy efficiency incentive item, directly linked to PUE optimization; This is a stability bonus item to ensure that the equipment does not overheat or experience performance issues under energy-saving strategies; To constrain rewards, used to punish actions that violate physical or safety rules; These are weighting coefficients used to balance the importance of the three reward items; these coefficients need to be adjusted through domain knowledge and experimentation; typically, The stability weight will be set high to ensure that safety is prioritized.

[0032] The hierarchical decision-making execution units are trained to develop dynamic strategies that optimize energy efficiency while ensuring stable equipment operation. For example, cooling strategies can be dynamically adjusted based on server load and room temperature field, such as increasing chiller temperature settings or guiding air conditioning airflow to optimize data center power usage effectiveness (PUE).

[0033] Preferably, the dynamic optimization and knowledge evolution module consists of a model dynamic optimization unit, an intelligent data lake warehouse unit, and an intelligent assessment and carbon footprint audit unit. The model dynamic optimization unit records the actual effect of each action, including automatic or manual actions, and uses feedback information to fine-tune the diagnostic and decision-making model, driving online learning so that the model can quickly adapt to changes in the data center environment, such as newly installed equipment or changes in business models, without forgetting old knowledge; achieving personalized adaptation that becomes more accurate with use; and establishing a fault handling case library, structurally recording the entire process of each successful diagnosis-decision-execution-feedback, forming organizational assets. The intelligent data lake warehouse unit adopts a hybrid architecture of data lake and data warehouse, storing raw data in the data lake (e.g., based on HDFS) for deep mining; and storing processed feature data and graph data in time-series databases (e.g., InfluxDB, TDengine) and graph databases (e.g., Neo4j, etc.), respectively. NebulaGraph supports high-performance real-time queries and complex analyses; the intelligent assessment and carbon footprint audit unit assesses data center equipment, such as IT equipment resource utilization, fault prediction accuracy, and mean time to repair, and audits data center energy efficiency using a carbon footprint calculation model; it combines IT load and energy consumption data with the carbon emission factor of the regional power grid to calculate the carbon footprint of the data center and specific businesses; it generates energy consumption and carbon emission reports for equipment, data center, and time, providing data support for energy conservation and emission reduction strategies; and it generates multi-dimensional energy consumption and carbon emission reports.

[0034] The above description is merely a preferred embodiment 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 deep learning-based decision support system for handling data center faults, characterized in that, It includes a global state perception and fault prediction module, a causal diagnosis and decision generation module, an autonomous decision-making and closed-loop execution module, and a dynamic optimization and knowledge evolution module; among them, the causal diagnosis and decision generation module consists of a causal root cause localization unit and a multi-objective trade-off handling unit, wherein... Causal root cause localization unit: Constructs a service dependency graph with temporal relationships using a causal discovery algorithm; When a fault occurs, analyzes abnormal behavior of indicators and simulates the fault propagation path through graph reasoning technology to quickly locate the root cause node; Multi-objective trade-off handling unit: It associates each diagnostic result with a multi-level set of handling options generated by expert experience and reinforcement learning, generates multiple decision options ranked by indicators, and explains the reasons for the ranking of options and the recommended options; The autonomous decision-making and closed-loop execution module consists of a hierarchical decision-making and safe execution unit and a digital twin-based energy efficiency optimization unit. Tiered decision-making and security execution unit: Based on the risk level of the data center failure or the application scenario level, and in conjunction with the security interlock mechanism, select the appropriate execution decision scheme; Based on digital twin energy efficiency optimization unit: Construct a simulation environment of digital twin of computer room, integrate energy saving and emission reduction indicators as one of the core optimization objectives into the reward function; and train the hierarchical decision execution unit to formulate a dynamic strategy with optimal energy efficiency under the premise of ensuring stable operation of equipment.

2. The data center fault handling decision support system based on deep learning according to claim 1, characterized in that, The global state perception and fault prediction module consists of a heterogeneous data acquisition and aggregation unit, a dynamic health assessment unit, and an early warning knowledge base driving unit. Heterogeneous data acquisition and aggregation unit: collects data on the performance indicators of computer room equipment, power and environmental data, network and security data, as well as application and business data; performs unified cleaning, labeling and storage of the collected data, and uses data lineage analysis technology to trace the source and transformation process of the stored data to ensure the reliability of subsequent analysis; Dynamic Health Assessment Unit: Utilizes a time-series deep learning model to analyze historical and real-time monitoring data, and combines it with a multivariate time-series anomaly detection model to construct a dynamic learning device health model, dynamically predicting the health of data center equipment; Security baseline construction unit: Analyze system logs and combine them with user and traffic behavior analysis to build a security baseline; establish normal behavior profiles for each user and each IP through unsupervised learning, thereby accurately identifying potential malicious activities and early signs of distributed denial-of-service attacks; Early warning knowledge base driven unit: Combines operation and maintenance experience with dynamic warning values, and constructs an early warning knowledge base based on quantitative risk assessment of model confidence; and proactively pushes emergency alert information to administrators.

3. The data center fault handling decision support system based on deep learning according to claim 2, characterized in that, In the heterogeneous data acquisition and aggregation unit, data lineage analysis technology is used to track the source and transformation process of the stored data to ensure the reliability of subsequent analysis; The specific steps are as follows: S1. Metadata Collection and Aggregation: Register all data sources in the system's data directory or data catalog; and collect multi-level metadata; S2. Lineage Discovery and Resolution: The data is parsed using SQL scripts, programming code, and workflow dependencies to construct multi-dimensional data lineage relationships. Lineage relationships from different sources are then cleaned, transformed, and standardized to form lineage model data. S3. Construction, storage and visualization of kinship maps: Model kinship model data into kinship maps; then integrate the kinship maps with the system's data catalog or data governance platform, and provide a visualization interface.

4. The data center fault handling decision support system based on deep learning according to claim 2, characterized in that, In the dynamic health assessment unit, the formula for calculating health status is as follows: ； In the above formula, , and These are the weighting coefficients for failure rate, performance health, and lifespan health, respectively. The weighting coefficients are dynamically adjusted based on equipment type and business importance through an online learning mechanism. Failure rate is the failure classification probability output by the equipment health model. Performance health is calculated based on the distance between key performance indicators and fault warning thresholds; lifespan health is calculated as remaining service life / design life.

5. The data center fault handling decision support system based on deep learning according to claim 2, characterized in that, In the early warning knowledge base driving unit, the construction rules of the early warning knowledge base include static rules, dynamic thresholds, and scenario rules, among which, Static rules are entered from the experience-based rules of operations and maintenance experts; Dynamic thresholds utilize unsupervised learning algorithms to learn a dynamic, personalized normal range for each device and each metric; The scenario rules are automatically generated by the system from historical failure cases, describing the normal change patterns of indicators under specific business scenarios, thus avoiding false alarms.

6. The data center fault handling decision support system based on deep learning according to claim 1, characterized in that, In the multi-objective trade-off handling unit, the decision-making scheme is based on the effect ranking and introduces the concept of multi-objective optimization, while comprehensively considering the handling time, cost, risk and impact on the service level agreement; The explanatory power of the solution is achieved through counterfactual reasoning.

7. The data center fault handling decision support system based on deep learning according to claim 1, characterized in that, In the hierarchical decision-making and security execution unit, the risk level of data center faults is divided into low risk, medium risk, and high risk, and the application scenario level is divided into high-determinism scenarios, medium-determinism scenarios, and low-determinism scenarios. For low-risk or high-certainty scenarios, the system automatically executes the recommended decision-making scheme; For medium-risk or medium-certainty scenarios, a human-machine collaborative approach is adopted to execute decision-making schemes; For high-risk or low-certainty scenarios, a supplementary reporting method is used to implement decision-making plans.

8. The data center fault handling decision support system based on deep learning according to claim 1, characterized in that, The dynamic optimization and knowledge evolution module consists of a model dynamic optimization unit, an intelligent data lake warehouse unit, and an intelligent assessment and carbon footprint audit unit. The model dynamic optimization unit records the actual effect of each action and uses the feedback information to fine-tune the diagnostic and decision-making model. It also drives online learning, enabling the model to quickly adapt to changes in the data center environment without forgetting old knowledge. Furthermore, it establishes a fault handling case library and records the entire process of each successful diagnosis-decision-execution-feedback in a structured manner.

9. A deep learning-based decision support system for handling computer room faults according to claim 2, characterized in that, In the dynamic optimization and knowledge evolution module, the intelligent data lake warehouse unit adopts a hybrid architecture of data lake and data warehouse, storing raw data in the data lake for in-depth mining. The processed feature data and graph data are stored in a time-series database and a graph database, respectively, to support high-performance real-time queries and complex analyses.

10. A deep learning-based decision support system for handling computer room faults according to claim 1, characterized in that, In the dynamic optimization and knowledge evolution module, the intelligent assessment and carbon footprint audit unit assesses the equipment in the data center and audits the energy efficiency of the data center in conjunction with the carbon footprint calculation model, generating a multi-dimensional energy consumption and carbon emission report.

Images