AI-based server health prediction system

Abstract

The invention relates to the technical field of server management, and discloses an AI-based server health prediction system, which is characterized in that a data processing unit collects server operation index data and abnormal event data, generates an index sequence and an event sequence, performs consistency verification, and finally obtains a monitoring credibility score. A health prediction unit generates a configuration fingerprint, constructs a component coupling graph, and updates edge weights within a sliding time window. And when the monitoring credibility score is not lower than a consistency verification threshold value, the input graph time sequence prediction model outputs a health prediction result. And the operation and maintenance decision-making unit generates a maintenance work order, a maintenance schedule and spare part scheduling according to the health prediction result and in combination with confidence information, and performs complementary sampling diagnosis and delays maintenance work order issuing when the confidence information is lower than a maintenance threshold value. The calibration unit is used for updating calibration parameters and gating parameters of the graph timing prediction model based on the maintenance execution backfill data. According to the invention, the risks of false report, false maintenance and non-planned shutdown are reduced, and the operation and maintenance scheduling efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of server management technology, and more specifically, to an AI-based server health prediction system. Background Technology

[0002] In data center and enterprise IT scenarios, servers typically operate in clusters for extended periods, handling various workloads such as training, inference, databases, and virtualization. To ensure business continuity and resource utilization, the operations and maintenance (O&M) team needs to continuously collect server performance metrics and anomaly events. Based on this data, they can conduct health status assessments, predict fault risks, and make O&M management decisions oriented towards maintenance windows, spare parts preparation, and scheduling coordination. This reduces unplanned downtime and maintenance costs, and improves service level assurance capabilities.

[0003] Existing operation and maintenance (O&M) solutions have begun to incorporate artificial intelligence for health scoring and fault prediction. For example, patent document CN114139747A provides an intelligent O&M system framework that includes a system health scoring module and a fault prediction module. However, in the data processing stage of server O&M management, which aims at management decisions, the following problems still exist: Data collected from multiple sources is prone to misalignment and conflict in terms of time reference and event sequence, and there is a lack of consistency verification and monitoring credibility score generation mechanisms for indicator sequences and event sequences. This leads to prediction results being used for work order dispatch even when data is distorted, triggering alarm storms or erroneous maintenance. Simultaneously, there is a lack of configuration fingerprint-driven model gating mechanisms to address configuration differences across different machine models and redundancy strategies, making it difficult to stably migrate prediction capabilities. Furthermore, in O&M management outputs such as maintenance work orders, maintenance scheduling, and spare parts dispatch, there is a lack of incorporating confidence information and monitoring credibility scores into maintenance thresholds to form a constraint loop. This makes it difficult to translate predictions into executable scheduling and spare parts preparation decisions, thus failing to meet the needs of O&M management and resource scheduling.

[0004] Therefore, it is necessary to design an AI-based server health prediction system to address the problems existing in current technologies. Summary of the Invention

[0005] In view of this, the present invention proposes an AI-based server health prediction system, which aims to solve the problems of easy misalignment and conflict between multi-source operating indicators and abnormal events in terms of time base and event sequence, lack of consistency verification and monitoring credibility constraints leading to unreliable health prediction, and distortion of work orders and scheduling dispatch, resulting in alarm storms and resource mismatch.

[0006] This invention proposes an AI-based server health prediction system, comprising: The system comprises a data processing unit, a health prediction unit, an operation and maintenance decision-making unit, and a calibration unit; among which, The data processing unit is used to collect server operation indicator data and abnormal event data, perform unified time base processing on the server operation indicator data and the abnormal event data to generate indicator sequences and event sequences, and perform consistency verification on the indicator sequences and the event sequences to finally obtain the monitoring reliability score. The health prediction unit is used to generate configuration fingerprints based on the asset list, construct a component coupling graph, and update edge weights within a sliding time window. When the monitoring confidence score is not lower than the consistency verification threshold, the indicator sequence, the event sequence, the configuration fingerprint, and the component coupling graph are input into the graph time-series prediction model of the configuration fingerprint gating, and the health prediction result is output. When the monitoring confidence score is lower than the consistency verification threshold, a consistency verification failure indication is output to the operation and maintenance decision unit, and supplementary sampling diagnosis is triggered. The edges of the component coupling graph include thermal coupling edges, power supply coupling edges, interconnection coupling edges, and redundant replacement edges. The health prediction result includes server health index, component failure risk probability, and confidence information. The operation and maintenance decision unit is used for server operation and maintenance management. Based on the health prediction results and confidence information, it generates maintenance work orders, maintenance schedules and spare parts scheduling. When the confidence information or the monitoring confidence score is lower than the maintenance threshold, it performs supplementary sampling diagnosis and postpones the issuance of maintenance work orders. The maintenance threshold is not lower than the consistency verification threshold. The calibration unit is used to update the calibration parameters and gating parameters of the graph time series prediction model based on maintenance execution backfill data.

[0007] Furthermore, the data processing unit collects server operation indicator data and abnormal event data, generates indicator sequences and event sequences after unifying the time base, and performs consistency verification on the indicator sequences and event sequences. When finally obtaining the monitoring reliability score, the process includes: The data processing unit collects the server operation index data and the abnormal event data, and adds a collection source identifier and a collection time identifier to each data record; The server operation index data is resampled into an index sequence using a unified time base, and the abnormal event data is organized into an event sequence in chronological order. The consistency of the indicator sequence and the event sequence is checked. The consistency check includes time sequence check, missing sampling check, duplicate event check and cross-source state contradiction check. The cross-source state contradiction check is used to identify contradictory records based on the logical constraints between the in-situ state, running state and alarm state of the same component. The monitoring reliability score is generated based on the number of contradictory records identified in the consistency check, the number of missing samples, and the number of duplicate events.

[0008] Furthermore, when the health prediction unit generates a configuration fingerprint based on the asset list, it includes: The health prediction unit extracts model information, key component identification information, firmware version information, component connection relationship information, and redundancy strategy information from the asset list. The configuration fingerprint is obtained by generating configuration records according to a preset field order and performing summary encoding. The configuration fingerprint is associated with and saved with the configuration version; when the configuration fingerprint changes, the calibration unit is triggered to update the gating parameters.

[0009] Furthermore, when the health prediction unit constructs the component coupling graph and updates the edge weights within the sliding time window, it includes: The health prediction unit establishes a component coupling diagram using key components as nodes. To establish thermally coupled edges, power supply coupled edges, interconnection coupled edges, and redundant replacement edges for node pairs that meet the preset coupling judgment conditions; Within the sliding time window, the correlation stability of node pairs is calculated based on the index sequence, and the event linkage consistency of node pairs is calculated based on the event sequence. The edge weights of the corresponding edges are updated according to the correlation stability and event linkage consistency; the edge weights of the updated edges are checked for stability, and the edge weights of the edges that do not meet the preset stability conditions are kept as the edge weights before the update.

[0010] Furthermore, the graph time-series prediction model for configuring fingerprint gating includes: Feature encoder, graph temporal inferencer, gated fusion unit and confidence information calibration module; The feature encoder is used to encode the index sequence and the event sequence to obtain time-series features; The graph temporal inferencer is used to perform temporal inference on the component coupling graph based on edge weights and attention weights to obtain a risk propagation characterization, wherein the attention weights are calculated by the graph temporal inferencer based on the temporal features; The gated fusion unit is used to determine the gate weights of multiple prediction experts based on the configuration fingerprint and fuse them to obtain a health prediction result. The confidence information calibration module is used to calibrate the confidence information based on the monitoring confidence score.

[0011] Furthermore, when the operation and maintenance decision unit generates the maintenance work order based on the health prediction result, it includes: The operation and maintenance decision unit determines the components to be maintained based on the component failure risk probability and the server health index; determines the redundancy margin level based on the redundancy replacement relationship; synthesizes the redundancy margin level, the component failure risk probability, and the confidence information into a work order priority according to a preset priority mapping rule; and generates the maintenance work order, which includes a server identifier, a component identifier to be maintained, a work order priority, and a suggested maintenance window.

[0012] Furthermore, when the operation and maintenance decision-making unit generates the maintenance schedule and the spare parts scheduling, it includes: The operation and maintenance decision unit obtains the maintenance work order and preset maintenance window constraints, spare parts inventory constraints, and personnel shift constraints; determines the false alarm handling cost coefficient based on the confidence information, and constructs a decision objective together with the expected downtime loss; through optimization, under the conditions of satisfying the maintenance window constraints, the spare parts inventory constraints, and the personnel shift constraints, outputs the maintenance schedule, and outputs the spare parts demand list corresponding to the maintenance schedule as the spare parts scheduling.

[0013] Furthermore, when the confidence information or the monitoring confidence score is lower than the maintenance threshold, the operation and maintenance decision unit performs supplementary sampling diagnosis and postpones the issuance of the maintenance work order. The supplementary sampling diagnosis includes at least one of the following: Increase the sampling frequency of the server operation index data, increase the collection range of the abnormal event data, or realign the index sequence and the event sequence and recalculate the monitoring confidence score; after the confidence information and the monitoring confidence score both meet the maintenance threshold, issue the maintenance work order and update the maintenance schedule and the spare parts scheduling.

[0014] Furthermore, when the calibration unit updates the calibration parameters of the graph time series prediction model based on the maintenance execution backfill data, it includes: The calibration unit acquires maintenance execution backfill data corresponding to the maintenance work order. The maintenance execution backfill data includes fault confirmation results, maintenance action results, component replacement results, and recovery time results. The calibration unit compares the maintenance execution backfill data with the health prediction results according to the server identifier and the identifier of the component to be maintained, and determines the prediction consistency status. The calibration parameters of the confidence information are adjusted according to the prediction consistency status, and the adjusted calibration parameters are associated with the configuration version and saved. The prediction consistency status includes consistency, false alarms, and missed alarms.

[0015] Furthermore, when the calibration unit updates the gating parameters of the graph time series prediction model based on the maintenance execution backfill data, it includes: The calibration unit collects maintenance execution backfill data according to the configuration fingerprint; for each configuration fingerprint, it determines the prediction error type based on the health prediction result and the corresponding maintenance execution backfill data, and updates the gating weights of multiple prediction experts to form gating parameters accordingly; the gating parameters are associated with and saved with the configuration fingerprint; when it is detected that the configuration fingerprint has not been associated with and saved with the gating parameters, a preset basic gating parameter is selected as the initial gating parameter based on the model information and redundancy strategy information. The prediction error types include false alarms, false misses, and premature and delayed errors, where premature and delayed errors are used to characterize the deviation direction of the predicted fault occurrence time from the fault confirmation time in the maintenance execution backfill data.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: The data processing unit performs unified time base processing and consistency verification on operational indicator data and abnormal event data, forming a quantifiable monitoring reliability score. This prevents health prediction from blindly outputting based on distorted data, suppressing false alarms and alarm storms caused by misaligned, missing, and contradictory events from the source. The health prediction unit generates configuration fingerprints from the asset list and constructs a component coupling graph that can be dynamically updated with a sliding time window by combining thermal coupling, power supply coupling, interconnection coupling, and redundancy replacement relationships. Then, through a graph time-series prediction model gated by configuration fingerprints, it outputs health prediction results that simultaneously possess server health index, component failure risk probability, and confidence information, thereby maintaining stability under different machine models and redundancy strategies. The predictive capability is transferable and calibrable. When the monitoring confidence score is lower than the consistency verification threshold, the system proactively outputs a consistency verification failure indication to the operation and maintenance decision-making unit and triggers supplementary sampling diagnosis. The operation and maintenance decision-making unit further constrains the issuance of work orders with a maintenance threshold not lower than the consistency verification threshold. When the confidence information or monitoring confidence score is insufficient, the work order is delayed and supplementary sampling diagnosis is performed. This upgrades maintenance work orders, maintenance schedules, and spare parts scheduling from predictive outputs to executable management decisions constrained by confidence and uncertainty, reducing the risk of erroneous maintenance and resource mismatch. The calibration unit uses maintenance execution backfill data to update the calibration parameters and gating parameters of the graph time series prediction model, realizing closed-loop optimization of prediction, work order dispatch, and execution results, and continuously improving the stability and economy of operation and maintenance management decisions. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a functional block diagram of an AI-based server health prediction system provided in an embodiment of the present invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] See Figure 1 As shown, this application proposes an AI-based server health prediction system, including: The data processing unit is used to collect server operation indicator data and abnormal event data, perform unified time base processing on the server operation indicator data and the abnormal event data to generate indicator sequences and event sequences, and perform consistency verification on the indicator sequences and event sequences to finally obtain the monitoring reliability score.

[0020] The health prediction unit generates configuration fingerprints based on the asset inventory, constructs a component coupling graph, and updates edge weights within a sliding time window. When the monitoring confidence score is not lower than the consistency verification threshold, the indicator sequence, event sequence, configuration fingerprint, and component coupling graph are input into the graph time-series prediction model of the configuration fingerprint gating, and the health prediction result is output. When the monitoring confidence score is lower than the consistency verification threshold, a consistency verification failure indication is output to the operation and maintenance decision unit, and supplementary sampling diagnosis is triggered. The edges of the component coupling graph include thermal coupling edges, power supply coupling edges, interconnection coupling edges, and redundant replacement edges. The health prediction result includes server health index, component failure risk probability, and confidence information.

[0021] The operation and maintenance decision unit is used for server operation and maintenance management. Based on the health prediction results and confidence information, it generates maintenance work orders, maintenance schedules and spare parts scheduling. When the confidence information or the monitoring confidence score is lower than the maintenance threshold, it performs supplementary sampling diagnosis and postpones the issuance of maintenance work orders. The maintenance threshold is not lower than the consistency verification threshold.

[0022] The calibration unit is used to update the calibration and gating parameters of the time series prediction model of the graph based on the maintenance backfill data.

[0023] Specifically, server performance metrics data refers to various performance parameters collected from server hardware, operating system, and applications, such as CPU utilization, memory usage, disk I / O, network throughput, and temperature. This data is typically recorded in time series format, reflecting the server's operational status at a specific point in time. Abnormal event data refers to information about abnormal events generated during server operation, such as alarms, error logs, and fault records. This data usually includes information such as event type, occurrence time, and affected components, used to indicate potential or already occurring anomalies. Metric sequences and event sequences refer to server performance metrics data and abnormal event data, respectively, processed using a unified time base. Metric sequences are continuous numerical time series, while event sequences are discrete event-based time series; together, they constitute the raw input for server health status. The monitoring reliability score is a quantitative assessment of the quality of the metric and event sequence data. This score reflects the reliability of the raw data in terms of time sequence, completeness, and consistency, used to determine whether the data is suitable for subsequent health prediction. Configuration fingerprints are digest encodings of static information such as server hardware configuration, firmware version, component connection relationships, and redundancy policies. This fingerprint uniquely identifies a server's specific configuration and serves as the basis for model gating, adapting to prediction needs under different configurations. A component coupling graph is a graph structure built with key server components as nodes and physical or logical coupling relationships between components as edges. This graph characterizes the mutual influence between components, such as thermal coupling edges, power supply coupling edges, interconnection coupling edges, and redundant replacement edges, and is the foundation for risk propagation reasoning in graph time-series prediction models. A configuration fingerprint-gated graph time-series prediction model is a prediction model that combines configuration fingerprint information and graph time-series reasoning capabilities. This model can dynamically adjust the weights of internal prediction experts based on the server's specific configuration fingerprint, thereby more accurately predicting the server's health status and component failure risks. Health prediction results include a server health index, component failure risk probability, and confidence information. The server health index is a comprehensive score of the server's overall health status. The component failure risk probability is the likelihood of a specific component failing within a future period. Confidence information is a quantitative assessment of the reliability of the prediction results. Maintenance work orders, maintenance schedules, and spare parts scheduling are outputs of the operations and maintenance decision-making unit. A maintenance work order is a maintenance task instruction for a specific server and component. Maintenance scheduling is the time arrangement for maintenance tasks. Spare parts scheduling is the plan for preparing the necessary spare parts for maintenance tasks. Maintenance execution backfill data refers to the information actually recorded after the maintenance task is executed, such as fault confirmation results, maintenance action results, component replacement results, and recovery time results. This data is used to calibrate the predictive model, forming a closed loop between prediction and actual maintenance.

[0024] Understandably, in some embodiments of the present invention, the data processing unit continuously collects server operation indicator data and abnormal event data, and generates indicator sequences and event sequences after processing on a unified time base. The indicator sequences and event sequences are then subjected to consistency verification to obtain a monitoring reliability score. When the monitoring reliability score is not lower than the consistency verification threshold, the health prediction unit generates a configuration fingerprint based on the asset list, constructs a component coupling graph with key components as nodes, and updates the edge weights within a sliding time window. The edges include thermally coupled edges, power supply coupled edges, interconnection coupled edges, and redundant replacement edges. The update of the edge weights is calculated based on the correlation stability of the node pairs between the indicator sequences and the event sequences and the consistency of event linkage. The updated edge weights are used when the edge weight stability test meets the preset stability conditions, and the updated edge weights are retained when the edge weight stability test does not meet the preset stability conditions. Subsequently, the health prediction unit inputs the indicator sequence, the event sequence, the configuration fingerprint, and the component coupling graph into the graph time-series prediction model of the configuration fingerprint gating to output a health prediction result. The health prediction result includes the server health index, component failure risk probability, and confidence information, and the confidence information is calibrated by the confidence information calibration module under the constraint of the monitoring confidence score. When the monitoring confidence score is lower than the consistency verification threshold, the health prediction unit outputs a consistency verification failure indication to the operation and maintenance decision unit, which then triggers supplementary sampling diagnosis and instructs accordingly. The data processing unit performs supplementary sampling diagnosis, which includes increasing the sampling frequency of server operation indicator data, increasing the collection range of abnormal event data, and / or realigning indicator sequences and event sequences and recalculating monitoring confidence scores. The operation and maintenance decision unit uses a maintenance threshold as the threshold for issuing maintenance work orders, and the maintenance threshold is not lower than the consistency verification threshold. When confidence information is insufficient or the monitoring confidence score is lower than the maintenance threshold, the issuance of maintenance work orders is delayed while maintaining the supplementary sampling diagnosis process. After the confidence information and the monitoring confidence score meet the maintenance threshold, maintenance work orders, maintenance schedules, and spare parts scheduling are generated. The operation and maintenance decision unit determines the false alarm handling cost coefficient based on the confidence information and constructs a decision objective together with the expected downtime loss, aiming to achieve the goal within the maintenance window. Under constraints of bundles, spare parts inventory, and personnel shifts, the system outputs a maintenance schedule and a spare parts requirement list corresponding to the maintenance schedule. After maintenance is completed, the calibration unit obtains the maintenance execution backfill data and compares it with the health prediction results according to the server identifier and the identifier of the component to be maintained to determine the prediction consistency status. The prediction consistency status includes consistency, false alarms, and missed alarms. Based on the prediction consistency status, the prediction error type is determined, which includes false alarms, missed alarms, early alarms, and late alarms. Then, the calibration parameters and gating parameters of the graph time series prediction model are updated and associated with the configuration version and configuration fingerprint for storage, so as to form a closed-loop optimization process for health prediction and server operation and maintenance management constrained by monitoring confidence scores and confidence information.

[0025] The embodiments of this application provide an AI-based server health prediction system.

[0026] The data processing unit is configured to collect server performance metrics and anomaly event data. One implementation involves the data processing unit periodically polling a monitoring agent on the server to obtain performance metrics data. Anomaly event data can be obtained by parsing server log files or receiving messages pushed by the alarm system. Another implementation involves the data processing unit subscribing to data streams in a message queue to receive server performance metrics and anomaly event data in real time.

[0027] The data processing unit is configured to perform unified time-base processing on the server's operational metric data and the abnormal event data to generate metric sequences and event sequences. The data processing unit can perform linear interpolation or nearest-neighbor interpolation on the collected operational metric data, resampling it to a preset unified time interval to form the metric sequence. For abnormal event data, it can be sorted according to its occurrence timestamp to form an event sequence. As an optional implementation, the data processing unit can use methods such as sliding window averaging or median filtering to smooth the metric data, eliminating noise and aligning it to a unified time base. The event data can simply be sorted in ascending order by timestamp.

[0028] The data processing unit is configured to perform consistency checks on the indicator sequence and the event sequence, ultimately obtaining a monitoring reliability score. The data processing unit can perform preliminary integrity checks on the generated indicator and event sequences, such as checking for null values ​​or outliers. Then, based on preset rules, such as checking whether the event occurrence time falls within the indicator data collection time range, it can determine whether there is a significant time discrepancy between the two. Based on these check results, a reliability score can be calculated. Specifically, the data processing unit can design a simple rule engine to identify data jumps in the indicator sequence or duplicate records in the event sequence. Based on the number of identified issues, a weighted average can be calculated as the monitoring reliability score.

[0029] The health prediction unit is configured to generate configuration fingerprints based on the asset inventory. It can extract key information from the server's asset inventory, such as server model, CPU type, memory capacity, and hard drive configuration, and concatenate this information into a string. Then, it performs a hash operation on this string to generate a fixed-length hash value as the configuration fingerprint. Alternatively, the health prediction unit can predefine a list of configuration fields, such as "server model," "CPU model," "number of memory modules," and "operating system version." For each server in the asset inventory, the values ​​of the corresponding fields are extracted in the order of this list, encoded into binary strings or integer sequences, and finally combined to form the configuration fingerprint.

[0030] The health prediction unit is configured to construct a component coupling graph and update edge weights within a sliding time window. The unit can predefine a list of all critical components in the server, such as CPU, memory, hard drive, power supply, and network interface card (NIC), and treat these components as nodes in the graph. Then, based on preset physical connections or logical dependencies (e.g., data transfer between CPU and memory, power supply providing power to all components), it manually establishes edges between components and assigns initial weights. Specifically, the health prediction unit can maintain a component relationship template library containing typical coupling relationships between components across different machine models. Once the configuration fingerprint of a specific server is identified, the corresponding component coupling graph structure can be loaded from the template library. During operation, component-related metric data can be collected within a fixed-size sliding time window, such as the past 24 hours, and the edge weights can be adjusted based on the correlation of this data.

[0031] The health prediction unit is configured to input the indicator sequence, event sequence, configuration fingerprint, and component coupling graph into a graph time-series prediction model gated by configuration fingerprint when the monitoring confidence score is not lower than the consistency verification threshold, and output the health prediction result. When the monitoring confidence score is lower than the consistency verification threshold, it outputs a consistency verification failure indication to the operation and maintenance decision unit and triggers supplementary sampling diagnosis. The health prediction unit can set a fixed value as the consistency verification threshold. When the monitoring confidence score output by the data processing unit is higher than or equal to the threshold, the health prediction unit packages the collected indicator sequence, event sequence, configuration fingerprint, and currently constructed component coupling graph, and sends it to the pre-deployed graph time-series prediction model for processing. After the model runs, it returns the server health index, component failure risk probability, and confidence information. As an optional implementation, when the monitoring confidence score is lower than the preset threshold, the health prediction unit can immediately generate an alarm message indicating poor data quality and send it to the operation and maintenance decision unit. At the same time, instructions can be sent to the data acquisition system to request more frequent data sampling or more comprehensive event log collection on the current server in an attempt to obtain more reliable data.

[0032] The Operations and Maintenance (O&M) decision unit is configured for server O&M management, generating maintenance work orders, maintenance schedules, and spare parts allocation based on the health prediction results and confidence information. The O&M decision unit can receive health prediction results from the health prediction unit, including server health index, component failure risk probability, and confidence information. The O&M decision unit can set a threshold for the component failure risk probability. When the failure risk probability of a component exceeds this threshold, the O&M decision unit generates a preliminary maintenance work order, which includes the server identifier and the identifier of the component to be maintained. Then, the O&M decision unit can manually schedule maintenance and determine the required spare parts based on currently available maintenance resources and preset maintenance strategies. Specifically, the O&M decision unit can maintain a simple rule base, such as "if the CPU failure risk probability is higher than 0.8, generate an emergency work order." For maintenance scheduling, a First-Come, First-Served (FCFS) principle can be used to queue generated work orders in chronological order. Spare parts allocation can reserve a corresponding number of spare parts from the spare parts inventory based on the type of component to be maintained in the work order.

[0033] The operations and maintenance (O&M) decision unit is configured to perform supplementary sampling and diagnosis and postpone the issuance of maintenance work orders when the confidence information or the monitoring confidence score is lower than the maintenance threshold. This maintenance threshold is not lower than the consistency verification threshold. The O&M decision unit can set a maintenance threshold that is higher than or equal to the consistency verification threshold. When the confidence information in the health prediction result is lower than the maintenance threshold, or the monitoring confidence score provided by the data processing unit is lower than the maintenance threshold, the O&M decision unit will suspend the issuance of maintenance work orders. At this time, the O&M decision unit can send instructions to the data processing unit or the health prediction unit, requesting additional diagnosis, such as increasing the data sampling frequency or expanding the data collection range, to obtain more information to improve the confidence of the prediction. As an optional implementation, when the confidence information or the monitoring confidence score does not meet the maintenance threshold, the O&M decision unit can issue a warning to the O&M personnel, indicating that the reliability of the current prediction result is low and suggesting manual intervention for further inspection. Simultaneously, the originally planned maintenance work order is marked as "pending confirmation" and temporarily removed from the maintenance queue until the data quality or prediction confidence improves.

[0034] The calibration unit is configured to update the calibration and gating parameters of the graph time-series prediction model based on maintenance execution backfill data. The calibration unit can periodically retrieve maintenance execution backfill data from the operations and maintenance system, such as tables recording actual fault occurrence times, operations performed by maintenance personnel, and component replacements. The calibration unit can manually compare this backfill data with historical health prediction results. Based on these comparisons, the calibration unit can manually adjust some parameters within the model, such as adjusting the offset or threshold of the prediction results, to make the model output more consistent with reality. Specifically, the calibration unit can collect maintenance execution backfill data, such as fault confirmation results and maintenance action results. Then, the calibration unit can run a simple statistical analysis program to calculate the model's prediction accuracy, false positive rate, and false negative rate. Based on these statistics, the calibration unit can manually adjust the parameters used to generate confidence information in the model. For gating parameters, the calibration unit can manually adjust the weights of the corresponding prediction experts based on the differences in prediction performance under different configuration fingerprints.

[0035] Therefore, this system avoids erroneous maintenance and alarm storms caused by data distortion by implementing consistency verification and monitoring credibility score generation mechanisms. Through configuring fingerprint-driven model gating and dynamic component coupling graph updates, the system improves the adaptability and prediction accuracy of the prediction model under different server configurations. Furthermore, by introducing confidence information and monitoring credibility scores into the maintenance decision-making closed loop, the reliability and executability of maintenance work orders, scheduling, and spare parts dispatching are ensured, thereby optimizing server operation and maintenance management efficiency, reducing the risk of unplanned downtime, and improving service level assurance capabilities.

[0036] This application further proposes specific steps for the data processing unit to collect server operation indicator data and abnormal event data, generate indicator sequences and event sequences after unifying the time base, and perform consistency verification on the indicator sequences and event sequences to finally obtain the monitoring reliability score. Specifically, when collecting server operation indicator data and abnormal event data, the data processing unit attaches a collection source identifier and a collection time identifier to each data record. The collection source identifier is used to clarify the specific source of the data, such as different sensors, log files, monitoring agents, or third-party systems. This helps to trace the root cause when data anomalies occur and distinguish the characteristics of data from different sources. The collection time identifier records the precise time point when the data was collected, ensuring the time attribute of the data and serving as the basis for subsequent time series analysis and event correlation. These identifiers are usually embedded in each data record as metadata fields or stored through independent database fields.

[0037] To ensure the comparability and consistency of different types of data across time, the data processing unit resamples server operation indicator data into indicator sequences using a unified time benchmark, and organizes abnormal event data into event sequences in chronological order. The unified time benchmark can be a preset fixed time interval. By performing operations such as interpolation, averaging, or aggregation on the original indicator data, it is transformed into indicator sequences with the same sampling frequency. For abnormal event data, it is strictly sorted in ascending order based on its collection time identifier, forming an event sequence arranged chronologically. This is crucial for analyzing the order of events, their duration, and the causal relationships between events.

[0038] Consistency checks between indicator sequences and event sequences are a crucial step in ensuring data quality. Consistency checks include time sequence checks, missing sample checks, duplicate event checks, and cross-source state inconsistency checks. Time sequence checks aim to verify that the timestamps of data points in the sequence are strictly increasing, eliminating time-out issues caused by acquisition or transmission errors. Missing sample checks identify whether there are missing data points within a preset time interval in the indicator sequence. This can be done by comparing the actual number of sampled points with the theoretical number of sampled points, or by detecting timestamp jumps. Duplicate event checks detect whether there are events in the event sequence that are completely identical or occur repeatedly within a very short period, which may indicate false alarms or data transmission redundancy. Cross-source state inconsistency checks are a deeper level of logical checks used to identify contradictory records based on predefined logical constraints between the in-situ, running, and alarm states of the same component. For example, if a component's "in-situ state" shows "removed," but its "running state" shows "running," or its "alarm state" shows "normal," but its "running state" shows "faulty," these records are considered contradictory. This verification is usually achieved by establishing a component state machine model or rule engine to ensure that the descriptions of the same component state from different data sources are logically consistent.

[0039] After completing the above consistency checks, the data processing unit generates a monitoring reliability score based on the number of contradictory records identified, the number of missing samples, and the number of duplicate events. This score is a quantitative assessment of data quality, typically calculated using weighted summation, a penalty function, or a machine learning model. For example, different weights can be assigned to different types of contradictions or missing data. Then, the number of all detected issues is multiplied by its corresponding weight and summed. A normalization function is then used to map this sum to a range of 0 to 100, where a higher score indicates better data quality and higher reliability. This score will serve as a crucial basis for the subsequent health prediction unit to determine whether to use the current data for prediction and for the operation and maintenance decision-making unit to make maintenance decisions.

[0040] In some embodiments of this application, the monitoring reliability score is normalized to the [0,1] interval and, together with the consistency verification threshold and the maintenance threshold, constitutes a hierarchical reliability constraint. The consistency verification threshold controls whether the health prediction result is output externally, and the maintenance threshold controls whether a maintenance work order is issued, and the maintenance threshold is not lower than the consistency verification threshold. When the edge weights are updated within the sliding time window, a stability test is performed on the updated edge weights. The stability test can be based on the condition that "the weight change exceeds a preset change threshold and maintains the same direction of change for at least M consecutive sliding time windows." The operation and maintenance decision-making side can determine the false alarm handling cost coefficient based on confidence information. The false alarm handling cost coefficient is a mapping parameter that decreases monotonically with respect to confidence information. It is used to balance the ineffective maintenance costs caused by false alarms and the expected downtime losses caused by missed alarms when optimizing the solution. At the same time, in order to facilitate closed-loop calibration and gating updates, the predicted consistency status is divided into consistent, false alarm, and missed alarm according to the comparison results of maintenance execution backfill data and health prediction results. The prediction error types include at least false alarm, missed alarm, early and late, where early and late are used to characterize the deviation direction of the predicted fault occurrence time relative to the fault confirmation time.

[0041] Through the aforementioned technical solution, the data processing unit adds a source identifier and a time identifier to each data record when collecting server operation indicator data and abnormal event data. It also resamples and reorganizes the data using a unified time benchmark. This lays a solid foundation for subsequent data analysis and verification, ensuring the temporal integrity and traceability of the data. Furthermore, by introducing refined consistency verification mechanisms such as time sequence verification, missing sampling verification, duplicate event verification, and cross-source state contradiction verification, this application can comprehensively and accurately identify various potential problems in the data. In particular, cross-source state contradiction verification can discover logical conflicts that are difficult to detect with a single data source, improving the quality and reliability of the original data. Finally, the monitoring credibility score generated based on these detailed verification results can more realistically and objectively reflect the credibility of the current data, providing a more accurate basis for the health prediction unit to determine whether to perform predictions and for the operation and maintenance decision-making unit to make maintenance decisions. This avoids misjudgments or omissions caused by low-quality data, thereby improving the accuracy and decision-making efficiency of the entire server health prediction system.

[0042] This application further proposes a specific method for the health prediction unit to generate configuration fingerprints based on an asset inventory, including: the health prediction unit extracts model information, key component identification information, firmware version information, component connection relationship information, and redundancy strategy information from the asset inventory. Configuration records are formed according to a preset field order and digest encoding is performed to obtain the configuration fingerprint. The configuration fingerprint is associated with and saved with the configuration version. When the configuration fingerprint changes, the calibration unit is triggered to update the gating parameters.

[0043] Specifically, the health prediction unit extracts server model information, key component identification information, firmware version information, component connection relationship information, and redundancy strategy information from the asset inventory. The asset inventory typically originates from a Configuration Management Database (CMDB) or automated asset discovery tools, containing detailed hardware configurations, software versions, network topologies, and other information about the servers. Server model information refers to the server's manufacturer and model number, crucial for understanding the server's overall architecture and performance characteristics. Key component identification information refers to the unique identifiers or serial numbers of the server's core components (such as CPUs, memory modules, hard drives, network interface cards, power modules, etc.), used for accurate component identification and tracking. Firmware version information refers to the version numbers of the firmware or microcode running on these key components; firmware versions directly impact the functionality, performance, and stability of the components. Component connection relationship information describes the physical or logical connections between various components within the server, such as PCIe bus connections, SATA / SAS connections, network port connections, etc. These relationships form the basis for constructing the component coupling diagram. Redundancy strategy information refers to the redundancy configurations adopted by the server, such as RAID levels, dual power supply configurations, multi-path network connections, etc. These strategies determine the system's fault tolerance capability in the event of partial component failures. The health prediction unit can automatically and comprehensively obtain these configuration data by integrating with or querying these asset management systems.

[0044] Based on this, the health prediction unit forms configuration records according to a preset field order and performs digest encoding to obtain a configuration fingerprint. The preset field order refers to the predefined, fixed order in which the extracted configuration information is organized into configuration records. For example, it can be specified that the machine model information is listed first, followed by CPU information, then memory, storage, network, etc., and the attributes within each component (such as model, version, serial number) also have a fixed order. This fixed order ensures that even if the configuration content is the same, as long as the order is consistent, the generated configuration record will be consistent, thus guaranteeing the determinism of the digest encoding. The configuration record is a structured data representation, which can be JSON, XML, or other text formats. Digest encoding processes this structured configuration record using a one-way hash function (such as SHA-256, MD5, etc.) to generate a fixed-length, unique string, i.e., the configuration fingerprint. This fingerprint can concisely represent the current configuration state of the server; any minor configuration change will lead to a change in the fingerprint.

[0045] Simultaneously, the health prediction unit associates and stores configuration fingerprints with configuration versions. Configuration versions are a mechanism for historically tracking and managing server configuration status. Each change in the configuration fingerprint signifies a configuration change, generating a new configuration version and associating the new fingerprint with that version number. This association is typically implemented in a database or dedicated configuration management system for easy subsequent querying, comparison, and backtracking. When the configuration fingerprint changes, it triggers the calibration unit to update the gating parameters. This means the system can detect server configuration changes in real time and automatically notify the calibration unit. Upon receiving the trigger signal, the calibration unit adjusts the gating parameters in the graph time-series prediction model based on the new configuration fingerprint. The gating parameters control the weights of different prediction experts (potentially corresponding to different configuration types or failure modes) during the prediction process, ensuring the model adapts to new configuration environments and maintains prediction accuracy.

[0046] Through the above technical solution, the health prediction unit can comprehensively and accurately extract key configuration information of servers from the asset list, form configuration records according to a preset standardized order, and then generate unique configuration fingerprints through digest encoding. This approach not only ensures the standardization and traceability of configuration information, but also achieves effective management of server configuration history by associating and saving configuration fingerprints with configuration versions. When server configuration changes, the automatic update of configuration fingerprints can promptly trigger the calibration unit to adjust the gating parameters of the graph time-series prediction model, enabling the prediction model to quickly adapt to the new configuration environment and avoiding the problem of decreased prediction accuracy due to configuration changes. This improves the adaptability and prediction accuracy of the health prediction system to complex and dynamically changing server environments, providing a solid foundation for subsequent health prediction and operation and maintenance decisions.

[0047] This application further proposes a health prediction unit that constructs a component coupling graph and updates edge weights within a sliding time window. Specifically, the prediction unit establishes a component coupling graph using key components as nodes. A component coupling graph is a graph structure used to represent the relationships between key components within a server, where each node represents an important hardware or software component in the server, such as a central processing unit, memory module, storage device, power supply unit, or network interface card. By abstracting these key components into graph nodes, a foundation is provided for subsequent analysis of their complex interactions. The health prediction unit can extract key component identification information from the asset inventory and assign a unique node identifier to each key component, while storing the component's static attributes, such as type, model, and physical location.

[0048] To meet preset coupling criteria, node pairs are assigned thermal coupling edges, power supply coupling edges, interconnect coupling edges, and redundancy replacement edges. These edges represent different types of coupling relationships between nodes (critical components), and the preset coupling criteria define these relationships. Thermal coupling edges indicate physical proximity relationships where components transfer heat or influence each other, such as components sharing heat dissipation resources. Power supply coupling edges indicate power supply dependencies between components, such as a power module providing power to multiple components. Interconnect coupling edges indicate logical connections between components for data or control signals, such as the data bus connection between a processor and memory. Redundancy replacement edges indicate functional backup or substitution relationships between components, such as storage devices in dual power modules or redundant arrays. The health prediction unit can determine and establish these different types of coupling edges based on component connection information, physical layout information, power topology information, and redundancy strategy information from the asset inventory, combined with preset rules or expert knowledge.

[0049] Within a sliding time window, the correlation stability of node pairs is calculated based on the index sequence, and the event linkage consistency of node pairs is calculated based on the event sequence. The sliding time window is used to capture the dynamic changes in the coupling strength between components, ensuring the timeliness of the analysis. Correlation stability reflects the degree of synchronous change and stability of the operating indicators (such as temperature, load, throughput, etc.) of two components over a period of time. High correlation stability may mean that they are functionally closely cooperating or sharing resources. Event linkage consistency reflects the co-occurrence or causal relationship of abnormal events (such as alarms, error logs) of two components over time. High event linkage consistency may mean that a failure in one component will cause an anomaly in the other component. For index sequences, time series analysis methods, such as Pearson correlation coefficient or mutual information, can be used to calculate the index correlation of node pairs within each sliding time window, and further analyze the fluctuation of the correlation coefficient within the window to evaluate stability. For event sequences, the frequency of simultaneous or successive occurrences of abnormal events of two components within a short period of time can be statistically analyzed, for example, by using conditional probability to measure the linkage strength between events.

[0050] The edge weights are updated based on the correlation stability and event linkage consistency. Edge weights are key information in the component coupling graph, quantifying the strength of the coupling relationship between components. By dynamically updating edge weights, the component coupling graph can reflect real-time changes in server operating status, enabling the health prediction model to capture more accurate inter-component interactions. The health prediction unit can design a fusion function to combine the calculated correlation stability and event linkage consistency to generate a comprehensive edge weight. For example, a weighted average or rule-based logical judgment can be used to update the weights.

[0051] The updated edge weights undergo a stability check. Edge weights that fail to meet the preset stability conditions are retained as their original weights. This stability check prevents edge weights from being frequently or inaccurately updated due to short-term, accidental fluctuations, which could introduce noise or destabilize the model. A change in edge weights is considered real and meaningful only when it reaches a certain level of significance or persistence. The stability check can include setting a threshold, requiring updates only when the difference between the updated and original edge weights exceeds this threshold. Alternatively, it can require the edge weights to exhibit a consistent trend over multiple consecutive sliding time windows. If the updated weights fail these stability checks, the system reverts to using the edge weights from the previous stable state to ensure model robustness.

[0052] Through the above technical solution, the health prediction unit can dynamically construct and maintain a component coupling graph reflecting the complex interactions between internal server components. This solution not only considers the static physical and logical connections between components but also achieves real-time quantification and dynamic updating of the coupling strength between components by calculating the correlation stability of indicator sequences and the linkage consistency of event sequences within a sliding time window. This dynamically updated edge weight can more accurately capture changes in the server's operating state. For example, when the indicators of two components show abnormal synchronization or frequent event linkage, their coupling edge weights will be strengthened accordingly, enabling the subsequent graph time-series prediction model to more sensitively identify potential fault propagation paths and risks. Furthermore, the introduction of edge weight stability checks avoids erroneous updates caused by short-term data fluctuations, ensuring the accuracy of the component coupling graph and the robustness of the prediction model. This improves the accuracy and reliability of server health prediction, providing a more solid foundation for operational and maintenance decisions.

[0053] This application further proposes a graph temporal prediction model with configuration fingerprint gating. The model includes a feature encoder, a graph temporal inferencer, a gating fusion unit, and a confidence information calibration module. The feature encoder encodes indicator sequences and event sequences to obtain temporal features. The graph temporal inferencer performs temporal inference on the component coupling graph based on edge weights and attention weights to obtain a risk propagation representation, where the attention weights are calculated by the graph temporal inferencer based on the temporal features. The gating fusion unit determines the gating weights of multiple prediction experts based on the configuration fingerprint and fuses them to obtain a health prediction result. The confidence information calibration module calibrates the confidence information based on the monitoring confidence score.

[0054] Specifically, the configuration fingerprint-gated graph time-series prediction model is an intelligent model specifically designed for server health prediction. Its core lies in its ability to dynamically adjust its prediction logic based on the server's specific configuration (i.e., configuration fingerprint), thereby achieving highly customized and accurate predictions. This model integrates multiple data types (such as time-series data, graph structure data, and configuration data) and utilizes gating mechanisms to select or combine different prediction strategies to adapt to the complexity and diversity of server environments.

[0055] The feature encoder is responsible for transforming the raw server performance indicator sequence and anomalous event sequence into more informative and abstract temporal features. This is typically achieved using deep learning techniques, such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), gated recurrent units (GRUs), or Transformer-based encoders. These encoders capture the temporal dependencies, trends, and patterns in the sequence data, compressing the raw, high-dimensional, and potentially noisy temporal data into low-dimensional, semantically rich feature vectors, providing high-quality input for subsequent graph temporal inference.

[0056] Graph temporal inference performs reasoning based on the structure of the component coupling graph to understand the propagation mechanisms of faults or anomalies between components within the server. It considers not only the static coupling relationships between components (represented by edge weights, such as thermal coupling and power supply coupling) but also introduces dynamic attention weights. These attention weights are calculated by the graph temporal inferencer based on current temporal characteristics. They dynamically identify which connections between components or which components themselves are most critical for risk propagation at a specific point in time or under specific circumstances, making the inference process more targeted and flexible. In this way, the graph temporal inferencer can output a risk propagation representation, revealing potential fault propagation paths and their scope of impact.

[0057] The gating fusion unit is a key component for implementing "configuration fingerprint gating." It dynamically determines the gating weights of multiple pre-defined prediction experts based on the server's configuration fingerprint. Each prediction expert can be considered a sub-model optimized for a specific server configuration type or failure mode. The gating fusion unit intelligently activates, weights, or combines the outputs of these prediction experts based on the input configuration fingerprint, thereby generating the final health prediction result. This mechanism enables the model to provide highly customized predictions for servers of different models, hardware configurations, and software environments, avoiding the limitations of a single model applying a "one-size-fits-all" approach to all configurations.

[0058] The confidence information calibration module adjusts the confidence information accompanying the health prediction results. It uses the monitoring confidence score generated by the data processing unit as the calibration basis. If the monitoring confidence score is low, it indicates a problem with the quality or completeness of the input data; therefore, even if the model provides a prediction result, its confidence level should be reduced accordingly. This module adjusts the confidence information through calibration algorithms (e.g., methods based on Bayesian calibration, isotonic regression, or temperature scaling) to ensure that the final output confidence information accurately reflects the reliability of the prediction results, providing a more accurate risk assessment basis for operational decisions.

[0059] By introducing a feature encoder, a graph temporal inferencer, a gating fusion unit, and a confidence information calibration module, this application can encode server operation indicator sequences and abnormal event sequences into temporal features, and utilize component coupling graphs and dynamic attention mechanisms for risk propagation characterization, thereby gaining a deeper understanding of the interactions between components. In particular, the gating fusion unit dynamically adjusts the weights of prediction experts based on configuration fingerprints, enabling the prediction model to provide customized health prediction results for servers with different configurations, improving the accuracy and applicability of predictions. Simultaneously, the confidence information calibration module calibrates the prediction confidence information based on monitoring credibility scores, ensuring the reliability of prediction results and providing a more solid foundation for operational decisions. This addresses the technical challenges of multi-source heterogeneous data fusion, complex coupling relationship modeling, and ensuring the credibility of prediction results.

[0060] This application further proposes a specific method for the operation and maintenance decision-making unit to generate maintenance work orders based on health prediction results. The operation and maintenance decision-making unit determines the components to be maintained based on the component failure risk probability and the server health index. It determines the redundancy margin level based on redundancy substitution relationships. According to a preset priority mapping rule, the redundancy margin level, component failure risk probability, and confidence information are combined to form a work order priority. A maintenance work order is then generated, which includes a server identifier, a component identifier to be maintained, a work order priority, and a suggested maintenance window.

[0061] Specifically, after receiving the health prediction results, the operations and maintenance decision-making unit first determines the components to be maintained based on the component failure risk probability and the server health index. For example, a component failure risk probability threshold can be set; when the predicted risk probability of a component exceeds this threshold, it is marked as a component to be maintained. Simultaneously, the server health index can also be considered. For instance, when the overall server health index is below a certain level, even if the component risk probability does not reach an independent threshold, its internal components may still be included in the maintenance scope for preventative maintenance.

[0062] Subsequently, the operations and maintenance decision-making unit determines the redundancy margin level of the component to be maintained based on redundancy replacement relationships. Redundancy replacement relationships can be obtained from the asset inventory or analyzed through component coupling diagrams. For example, if a component to be maintained has multiple available hot or cold standby components, its redundancy margin level may be rated as high. Conversely, if the component is a single point of failure in the system, or its redundancy configuration is insufficient, its redundancy margin level will be rated as low. Redundancy margin levels can be divided into multiple levels, such as high, medium, and low, or expressed numerically to quantify their impact on system stability.

[0063] Based on this, the operation and maintenance decision-making unit takes redundancy margin level, component failure risk probability, and confidence information as inputs, and synthesizes work order priorities according to a preset priority mapping rule. This priority mapping rule can be a predefined function, a lookup table, or a decision tree model based on expert experience. For example, for high-risk, low-redundancy components with high prediction confidence information, their work order priority will be set to the highest. Conversely, for low-risk, high-redundancy components or components with low prediction confidence information, their work order priority may be relatively low. Confidence information plays a crucial adjusting role in this process; when the confidence information of the prediction result is low, even if the component failure risk probability is high, the system may appropriately lower the work order priority to await further diagnosis or information confirmation, avoiding unnecessary maintenance operations.

[0064] Finally, the operations and maintenance decision-making unit generates a maintenance work order. This work order is a structured data record that explicitly includes: a server identifier to uniquely identify the server requiring maintenance; a component identifier to precisely indicate the specific component requiring maintenance; a work order priority to guide operations and maintenance personnel in executing tasks according to importance and urgency; and a suggested maintenance window to provide time guidance for maintenance operations, facilitating operations and maintenance scheduling and resource coordination.

[0065] Through the aforementioned technical solutions, the operation and maintenance decision-making unit can systematically transform the server health index, component failure risk probability, and confidence information output by the health prediction unit into specific, executable maintenance work orders. This approach comprehensively considers component failure risk, overall server health status, system redundancy, and prediction confidence information, making maintenance decisions more comprehensive and scientific, avoiding the limitations of single-indicator decisions. In particular, by introducing redundancy margin levels, the system can prioritize identifying and addressing critical components or single points of failure that have a greater impact on system stability, thereby reducing potential system-level failure risks. The synthesis of work order priorities allows operation and maintenance personnel to clearly understand the urgency and importance of each maintenance task, optimizing the allocation and scheduling of maintenance resources, and improving operation and maintenance efficiency and resource utilization. Furthermore, the key information included in the maintenance work order, such as server identifier, component identifier to be maintained, work order priority, and suggested maintenance window, provides operation and maintenance personnel with clear and detailed guidance, reducing information search and decision-making time, ensuring the accuracy and timeliness of maintenance operations, and thus improving the overall efficiency of server operation and maintenance management.

[0066] This application further proposes an operation and maintenance decision-making unit that obtains maintenance work orders and presets maintenance window constraints, spare parts inventory constraints, and personnel shift constraints. False alarm handling cost coefficients are determined based on confidence information and used in conjunction with expected downtime losses to construct a decision objective. Through optimization, under the conditions of satisfying maintenance window constraints, spare parts inventory constraints, and personnel shift constraints, a maintenance schedule is output, along with a spare parts demand list corresponding to the maintenance schedule for spare parts scheduling.

[0067] Specifically, the operations and maintenance (O&M) decision-making unit first obtains maintenance work orders generated by the system, which contain key information such as the component to be maintained and its priority. Simultaneously, the O&M decision-making unit also obtains preset maintenance window constraints, which define the time periods during which maintenance operations are allowed. For example, to avoid impacting peak business hours, some maintenance tasks may only be performed at night or on weekends. Furthermore, the system considers spare parts inventory constraints, i.e., the types and quantities of currently available spare parts, to ensure the timely supply of spare parts required for maintenance tasks. Personnel shift constraints involve the availability of maintenance personnel, skill matching, and work schedule arrangements to ensure that suitable personnel are available to perform maintenance tasks.

[0068] Based on this, the operations and maintenance decision-making unit determines the false alarm handling cost coefficient according to the confidence information in the health prediction results. The confidence information reflects the reliability of the prediction results; the lower the confidence level, the greater the probability of a false alarm, and therefore the higher the corresponding false alarm handling cost coefficient, and vice versa. This false alarm handling cost coefficient, together with the expected downtime loss, constitutes the decision objective. Expected downtime loss refers to the potential business interruption and economic losses that may result if a server or component fails and causes downtime. By combining the false alarm handling cost with the expected downtime loss, the system can form a comprehensive optimization objective, such as minimizing total operations and maintenance costs or maximizing system availability.

[0069] Subsequently, the system optimizes the solution algorithm to find the optimal solution that achieves the decision objective, while strictly satisfying the aforementioned maintenance window constraints, spare parts inventory constraints, and personnel shift constraints. The optimization process may employ methods such as linear programming, integer programming, heuristic algorithms, or reinforcement learning to find the best balance between complex constraints and objectives.

[0070] Ultimately, the optimization solution outputs a detailed maintenance schedule, which specifies the execution time, location, and responsible personnel for each maintenance task. Simultaneously, the system also outputs a spare parts requirement list corresponding to this maintenance schedule, serving as a spare parts dispatch plan. This list details the types and quantities of spare parts required for each maintenance task under a specific maintenance schedule, guiding the procurement, allocation, and distribution of spare parts.

[0071] Through the above technical solution, the operations and maintenance decision-making unit can comprehensively consider multiple constraints and cost-effectiveness in actual operations and maintenance when formulating maintenance plans, rather than just predicting failure risks. This makes the generated maintenance schedules and spare parts allocation more practical, resource-available, and cost-effective. The system can weigh the cost of handling false alarms against expected downtime losses, avoiding resource waste or business interruptions caused by blind or insufficient maintenance. Therefore, this solution improves the intelligence level and decision-making quality of server maintenance management, ensures the continuity of critical business operations, and reduces overall operations and maintenance costs.

[0072] This application further proposes that when the confidence information or monitoring credibility score is lower than the maintenance threshold, the operation and maintenance decision-making unit should perform supplementary sampling diagnosis and postpone the issuance of maintenance work orders. Supplementary sampling diagnosis includes at least one of the following: increasing the sampling frequency of server operation indicator data, increasing the collection scope of abnormal event data, or realigning the indicator sequence and event sequence and recalculating the monitoring credibility score. After both the confidence information and the monitoring credibility score meet the maintenance threshold, a maintenance work order is issued and the maintenance schedule and spare parts scheduling are updated.

[0073] When the operation and maintenance decision-making unit determines that the confidence information or monitoring reliability score is below the maintenance threshold, the system will initiate supplementary sampling diagnosis and temporarily postpone the issuance of maintenance work orders. This measure aims to avoid making maintenance decisions based on insufficient data quality or predictive reliability. Specific measures for supplementary sampling diagnosis include, but are not limited to, at least one of the following: First, the sampling frequency of server performance metrics can be increased. For example, metrics such as CPU utilization and memory usage, which are usually collected on a minute-by-minute basis, can be temporarily increased to the second or even millisecond level during diagnostics. This allows for the capture of more subtle and instantaneous changes in system status, thereby providing richer and more accurate input data for health prediction models.

[0074] Secondly, the scope of abnormal event data collection can be expanded. This is not limited to the system's preset alarms or error logs, but can be extended to collecting a wider range of system logs, application anomaly information, network traffic anomaly patterns, etc., in order to discover potential abnormal events that may have been overlooked and are currently of low confidence, thereby comprehensively assessing the server's health status.

[0075] Furthermore, the indicator sequences and event sequences can be realigned, and the monitoring reliability score can be recalculated. During the collection and transmission of raw data, timestamp discrepancies or data packet loss may occur, causing indicators and events to not match precisely on the timeline. By employing more advanced time synchronization algorithms or association rules, these sequences can be finely aligned, and the integrity and consistency of the data can be reassessed, thereby obtaining a more accurate monitoring reliability score.

[0076] After completing the above-mentioned resampling diagnosis, the system will reassess the confidence information and monitoring reliability score of the health prediction results. Only when both values ​​meet the preset maintenance thresholds, indicating that the system has reached a sufficient level of trust in the prediction results and data quality, will the operation and maintenance decision-making unit issue the previously postponed maintenance work order. Meanwhile, since the resampling diagnosis process may take some time, the originally established maintenance schedule and spare parts scheduling may no longer be applicable. Therefore, the system will update the maintenance schedule and spare parts scheduling based on the latest time, resource availability, and maintenance work order priority to ensure the timeliness and effectiveness of maintenance work.

[0077] By triggering supplementary sampling and diagnosis and delaying the issuance of maintenance work orders when the confidence information or monitoring confidence score falls below the maintenance threshold, this application avoids blindly performing maintenance operations when data quality is questionable or prediction results are uncertain. Specifically, by increasing the sampling frequency of server operation indicator data, expanding the collection range of abnormal event data, or realigning indicator sequences and event sequences and recalculating monitoring confidence scores, the system can proactively acquire richer and more accurate data, thereby improving the reliability and confidence of health prediction results. Maintenance work orders are only issued and maintenance schedules and spare parts allocations are updated when both the confidence information and monitoring confidence score meet the maintenance threshold. This ensures that all maintenance decisions are based on highly reliable data and predictions, reducing the risk of false alarms and missed alarms, improving maintenance efficiency and resource utilization, and avoiding unnecessary downtime and cost waste.

[0078] This application further proposes calibration parameters for a calibration unit based on a timing prediction model of maintenance execution backfill data update graphs. Specifically, the calibration unit acquires maintenance execution backfill data corresponding to maintenance work orders. This backfill data includes fault confirmation results, maintenance action results, component replacement results, and recovery time results. The calibration unit compares the maintenance execution backfill data with the health prediction results according to server identifiers and component identifiers to determine the prediction consistency status. Based on the prediction consistency status, the calibration parameters of the confidence information are adjusted, and the adjusted calibration parameters are saved in association with the configuration version. The prediction consistency status includes consistent, false alarms, and missed alarms.

[0079] The maintenance execution backfill data acquired by the calibration unit is feedback information collected from the maintenance management system or manual records after the actual maintenance activities have been completed. This data provides objective evidence regarding whether predicted faults actually occurred, the effectiveness of maintenance operations, and resource consumption. The calibration unit can interface with Enterprise Resource Planning (ERP) systems, Computerized Maintenance Management Systems (CMMS), or dedicated maintenance work order management systems to acquire this data automatically or semi-automatically. Maintenance execution backfill data is typically stored in a structured format, such as database records or log files, and includes fault confirmation results (indicating whether the predicted fault was confirmed to exist during maintenance), maintenance action results (describing the specific maintenance operations performed on the components, such as repair, replacement, and adjustment), component replacement results (clearly indicating which components were replaced and the information of the replaced components), and recovery time results (recording the time required from the occurrence of the fault to the system returning to normal operation).

[0080] The calibration unit compares maintenance execution backfill data with health prediction results to determine the prediction consistency status. This process assesses the accuracy of the prediction by comparing the actual maintenance results with the system's previous health predictions. The calibration unit first matches the maintenance execution backfill data with the corresponding health prediction results (including component failure risk probability, server health index, and confidence information) based on the server identifier and the identifier of the component to be maintained. Then, according to preset logical rules, it determines the consistency between the prediction and the actual results. Prediction consistency status includes: Consistent, indicating that the predicted fault did indeed occur, and the maintenance operation successfully resolved the problem; False alarm, indicating that the system predicted a fault, but no corresponding fault was found during actual maintenance, or the fault severity was much lower than predicted; Missed alarm, indicating that the system did not predict the fault, but an unpredicted fault was found during actual maintenance.

[0081] The calibration parameters of the confidence information are adjusted based on the prediction consistency status and saved in association with the configuration version. This aims to dynamically adjust the model's evaluation method for its own prediction confidence based on actual maintenance feedback, making it more accurately reflect the real situation. When false positives are identified, the model's confidence in similar prediction scenarios can be reduced. When false negatives are identified, the model's sensitivity to potential risks can be increased, and the confidence calculation method can be adjusted. When predictions are consistent, the model's confidence in similar scenarios can be enhanced. This adjustment can be achieved by modifying the weights, thresholds, or function parameters within the confidence information calibration module. For example, methods such as Bayesian updates, regression analysis, or reinforcement learning can be used to optimize these parameters. The adjusted calibration parameters are saved in association with the current configuration version (represented by the configuration fingerprint), ensuring that when the server configuration changes, the corresponding calibration parameters can be loaded to maintain prediction accuracy, thereby enabling the model to adapt to the characteristics of different hardware configurations, software versions, or operating environments.

[0082] By introducing maintenance execution backfill data and comparing it with health prediction results, the system can identify the accuracy of predictions, including consistency, false alarms, and missed alarms. Based on these prediction consistency statuses, the calibration unit can dynamically adjust the calibration parameters of the confidence information. This closed-loop calibration mechanism based on actual maintenance feedback enables the system to continuously learn and optimize its assessment of the confidence level of prediction results, thereby improving the accuracy and reliability of the confidence information. More accurate confidence information allows the operation and maintenance decision-making unit to more accurately assess risks and uncertainties when generating maintenance work orders, maintenance schedules, and spare parts scheduling, avoiding unnecessary maintenance, reducing resource waste caused by false alarms, and promptly identifying potential problems, reducing losses caused by missed alarms, thereby optimizing resource allocation and improving overall operation and maintenance efficiency and server availability. In addition, saving the adjusted calibration parameters in association with the configuration version ensures the adaptability and continuous effectiveness of the calibration, maintaining prediction accuracy even when server configurations change, thereby enhancing the robustness and long-term stability of the system.

[0083] This application further proposes a specific method for the calibration unit to update the gating parameters of the time-series prediction model based on maintenance execution backfill data. The calibration unit aggregates the maintenance execution backfill data according to the configuration fingerprint. Specifically, the calibration unit can extract the server identifier from the maintenance execution backfill data and obtain the configuration fingerprint corresponding to the server by querying the asset list or configuration management database. Subsequently, all maintenance execution backfill data with the same configuration fingerprint are grouped or aggregated to form a dataset classified by configuration fingerprint. For example, a hash table or database index can be created, using the configuration fingerprint as the key, to store the corresponding list of maintenance execution backfill data.

[0084] For each configuration fingerprint, the prediction error type is determined based on the health prediction results and the corresponding maintenance execution backfill data. Prediction error types include false positives, false negatives, and premature or delayed predictions. Premature and delayed predictions characterize the direction of deviation between the predicted fault occurrence time and the fault confirmation time in the maintenance execution backfill data. Specifically, for each aggregated configuration fingerprint, the calibration unit compares the health prediction results (e.g., component failure risk probability, predicted fault time) of all servers under that configuration fingerprint with the actual maintenance execution backfill data (e.g., fault confirmation result, fault occurrence time). If the model predicts a fault but the actual maintenance execution backfill data shows no fault, it is considered a false positive. If the model does not predict a fault but the actual maintenance execution backfill data shows a fault, it is considered a false negative. For time deviation, the difference between the predicted time and the actual confirmation time can be calculated and compared with a preset time window to determine premature or delayed predictions. For example, if the predicted fault occurrence time is earlier than the actual fault confirmation time, it is premature. If the predicted fault occurrence time is later than the actual fault confirmation time, it is delayed.

[0085] This updates the gating weights of multiple prediction experts to form gating parameters. Gating parameters are a set of weights used to control how the outputs of different prediction experts (e.g., sub-models trained for different component types, failure modes, or data sources) are fused. When a certain type of prediction error is detected (e.g., too many false positives), the weights of the specific prediction expert causing that error can be reduced, or the weights of experts capable of correcting that error can be increased. This process can employ reinforcement learning, gradient descent, or other optimization algorithms to dynamically adjust these weights until the gating parameters converge or reach a preset performance metric.

[0086] The gating parameters are associated with and saved as configuration fingerprints. The calibration unit stores the updated gating parameters in a configuration database or key-value store, where the configuration fingerprint serves as a unique identifier (key) and the gating parameter is the corresponding value. This allows the health prediction unit to quickly retrieve and load the appropriate gating parameters when it needs to predict for a server with a specific configuration.

[0087] When a configuration fingerprint is detected as not being associated with and saved with gating parameters, a preset basic gating parameter is selected as the initial gating parameter based on the machine type information and redundancy policy information. When the health prediction unit encounters a new configuration fingerprint and no corresponding record is found in the associated saved gating parameters, the calibration unit matches the preset basic gating parameter based on the machine type information (e.g., server model, CPU type) and redundancy policy information (e.g., N+1 redundancy, dual-machine hot standby) contained in the configuration fingerprint. These basic gating parameters can be general parameters that are set by offline training on typical machine types and redundancy policies or by expert experience before the system goes live.

[0088] Through the above technical solution, the calibration unit can finely adjust the gating parameters of the graph time-series prediction model based on actual maintenance feedback from different server configurations. Specifically, by collecting maintenance execution backfill data according to configuration fingerprints and comparing health prediction results with actual maintenance data, the system identifies various prediction error types such as false alarms, missed alarms, premature alarms, and delayed alarms. The system can then selectively update the gating weights of multiple prediction experts. This dynamic adjustment mechanism of gating parameters based on actual operational feedback allows the prediction model to better adapt to the unique operating characteristics and failure modes of different server configurations, improving the accuracy and reliability of health prediction results. Furthermore, for configuration fingerprints encountered for the first time, the system can select preset basic gating parameters based on machine type information and redundancy strategy information, ensuring that newly configured servers can also obtain reasonable initial prediction capabilities. This avoids prediction blind spots caused by a lack of historical data, thereby improving the robustness and practicality of the entire system and ultimately optimizing server operation and maintenance management efficiency and resource scheduling.

[0089] The following example will provide a more detailed explanation of the above technical solution: In a large data center, various types of servers are deployed, such as server A and server B. These servers carry different business workloads, and their operating status needs to be continuously monitored and predicted to ensure business continuity.

[0090] First, the data processing unit continuously collects operational metrics data from server A, such as CPU utilization, memory usage, disk I / O, and network throughput, as well as abnormal event data, such as error messages, hardware alarms, and process crash records in the system logs. This data may originate from different monitoring agents or hardware sensors; therefore, each data record is appended with a source identifier and a collection time identifier. To ensure data consistency, the data processing unit resamples this raw data using a unified time base. For example, the CPU utilization data collected per second is resampled to a point per minute, forming a metric sequence. Simultaneously, abnormal event data is organized into an event sequence in chronological order.

[0091] After generating the indicator and event sequences, the data processing unit performs rigorous consistency checks. These checks include time sequence verification to ensure all data points and events are arranged chronologically; missing sample verification to identify any data interruptions or missing data in the indicator sequences; duplicate event verification to exclude abnormal events with duplicate records; and cross-source state inconsistency verification. For example, if the hardware monitoring system for server A's power module 100 reports it as "online," but the operating system log shows "power module 100 fault," the system will identify this as a contradictory record. Based on these verification results, the data processing unit generates a monitoring reliability score according to the number of identified contradictory records, the number of missing samples, and the number of duplicate events. This score quantifies the reliability of the current input data. Unlike existing technologies that may directly use unverified data for prediction, this system avoids prediction errors and mis-maintenance caused by data distortion by generating a monitoring reliability score, thus solving the problem of misalignment and conflict in multi-source data collection.

[0092] Next, the health prediction unit begins its work. It first extracts server A's model information, key component identification information (such as CPU, memory modules, hard drives, power module 100, etc.), firmware version information, component connection information (such as the connection between the CPU and memory, and the connection between power module 100 and the motherboard), and redundancy policy information (such as server A being configured with N+1 power redundancy) based on the data center's asset inventory. This information is arranged in a preset field order to form a configuration record, and then digested to generate a unique configuration fingerprint. This configuration fingerprint is associated with and stored with server A's current configuration version. When server A's configuration changes (e.g., memory modules are replaced), the configuration fingerprint is updated accordingly, triggering the calibration unit to update the gating parameters, ensuring the model can adapt to the new configuration.

[0093] Simultaneously, the health prediction unit establishes a component coupling graph using key components of server A (such as CPU, memory, power module 100, network card, etc.) as nodes. To meet preset coupling criteria, the system creates different types of edges for node pairs. For example, a thermal coupling edge is created between the CPU and the cooling fan. A power supply coupling edge is created between the power module 100 and the motherboard. An interconnect coupling edge is created between the CPU and memory. A redundant replacement edge is created between two redundant power modules 100 and 200. Within a sliding time window (e.g., the past 24 hours), the health prediction unit calculates the correlation stability of node pairs based on indicator sequences (e.g., the correlation between CPU temperature and fan speed) and the event linkage consistency of node pairs based on event sequences (e.g., the correlation between CPU error logs and memory error logs). Based on these calculation results, the system updates the edge weights of the corresponding edges. The updated edge weights also undergo stability checks; edge weights that do not meet preset stability conditions will remain as their previous weights to prevent short-term fluctuations from interfering with the model.

[0094] When the monitoring confidence score is not lower than a preset consistency verification threshold (e.g., 0.8), the health prediction unit inputs the indicator sequence, event sequence, configuration fingerprint, and component coupling graph into the configuration fingerprint-gated graph time-series prediction model. This model includes a feature encoder, a graph time-series inferencer, a gating fusion unit, and a confidence information calibration module. The feature encoder encodes the indicator sequence and event sequence to obtain time-series features. The graph time-series inferencer performs time-series inference on the component coupling graph based on the updated edge weights and attention weights calculated from the time-series features, thereby obtaining a risk propagation representation (e.g., how an anomaly in one component affects other related components). The gating fusion unit determines the gating weights of multiple prediction experts (sub-models optimized for different configurations) based on the configuration fingerprint of server A, and fuses the prediction results of these experts to finally output the health prediction result. The confidence information calibration module calibrates the confidence information in the prediction result based on the monitoring confidence score to ensure that the confidence reflects the reliability of the input data. The health prediction results include the server health index of server A (e.g., 0.75, indicating a fair health condition), the probability of component failure risk (e.g., the failure risk of power module 100 is 60%), and confidence information (e.g., the confidence level for the failure prediction of power module 100 is 80%).

[0095] If the monitoring reliability score falls below the consistency verification threshold, the health prediction unit will output a consistency verification failure indication to the operation and maintenance decision-making unit and trigger supplementary sampling diagnosis. For example, the system will increase the sampling frequency of server A's operational indicator data, expand the collection range of abnormal event data, and realign the indicator sequence and event sequence to recalculate the monitoring reliability score in an attempt to obtain more reliable data. This solves the problem in existing technologies where prediction results are still used for dispatching when data is distorted.

[0096] Subsequently, the operations and maintenance decision-making unit performs server operations and maintenance management based on the health prediction results. For example, based on the 60% failure risk probability of power module 100 and the server health index of 0.75, power module 100 is identified as a component requiring maintenance. Simultaneously, based on the N+1 power redundancy strategy of server A, the current redundancy margin level is determined. Then, according to a preset priority mapping rule, the redundancy margin level, component failure risk probability, and confidence information are combined to form a work order priority (e.g., "high priority"). Finally, a maintenance work order is generated, which includes the identifier of server A, the identifier of the component requiring maintenance, power module 100, the work order priority, and a suggested maintenance window (e.g., "within the next 24 hours").

[0097] When generating maintenance schedules and spare parts allocation, the operations and maintenance decision-making unit (O&M) obtains maintenance work orders and combines them with preset maintenance window constraints (e.g., the maintenance window for server A is every Tuesday from 2:00 AM to 4:00 AM), spare parts inventory constraints (e.g., whether there is a usable power module 100 in the data center spare parts warehouse), and personnel shift constraints. Simultaneously, the O&M decision-making unit determines the false alarm handling cost coefficient based on the confidence information of the prediction results (e.g., the higher the confidence level, the lower the false alarm cost), and constructs the decision objective together with the expected downtime losses. Through optimization, under the premise of satisfying all constraints, the system outputs a maintenance schedule (e.g., scheduling engineer B to replace power module 100 on server A at 2:00 AM on Tuesday) and outputs a spare parts demand list corresponding to this maintenance schedule as spare parts allocation (e.g., allocating a power module 100 from the spare parts warehouse).

[0098] It is worth noting that if the confidence information or monitoring reliability score of the prediction result is lower than the preset maintenance threshold (e.g., confidence score below 0.75 or monitoring reliability score below 0.85, and the maintenance threshold is not lower than the consistency verification threshold), the operation and maintenance decision unit will perform supplementary sampling diagnosis again and postpone the issuance of the maintenance work order. Only when both the confidence information and the monitoring reliability score meet the maintenance threshold will the maintenance work order be issued, and the maintenance schedule and spare parts scheduling will be updated. This mechanism solves the problem in existing technologies where predictions are difficult to translate into executable scheduling and spare parts preparation decisions. By incorporating the confidence information and monitoring reliability score into the maintenance threshold and forming a constraint closed loop, the reliability of operation and maintenance decisions is ensured.

[0099] Finally, the calibration unit is responsible for the continuous optimization of the model. After a maintenance work order is completed, maintenance personnel will back-update the maintenance execution data, including fault confirmation results (e.g., "Fault confirmed"), maintenance action results (e.g., "Replace power module 100"), component replacement results, and recovery time results. The calibration unit compares this maintenance execution back-up data with previous health prediction results, categorized by server identifier and component identifier to determine the prediction consistency status. For example, if a fault is predicted for power module 100 and is actually confirmed, it is considered "consistent." If a fault is predicted but not actually found, it is considered a "false alarm." If a fault is not predicted but actually occurs, it is considered a "missed alarm." Based on these prediction consistency statuses, the calibration unit adjusts the calibration parameters of the confidence information and saves the adjusted calibration parameters along with the configuration version.

[0100] Simultaneously, the calibration unit also aggregates maintenance execution backfill data according to configuration fingerprints. For a specific configuration fingerprint of server A, the calibration unit determines the prediction error type based on the health prediction results and the corresponding maintenance execution backfill data, such as "false alarm," "missed alarm," "early" (predicted fault occurrence time is earlier than the actual confirmation time), or "lagging" (predicted fault occurrence time is later than the actual confirmation time). Accordingly, the calibration unit updates the gating weights of multiple prediction experts to form gating parameters. These gating parameters are associated with and stored with the configuration fingerprint. When it is detected that a configuration fingerprint has not yet been associated with gating parameters, the system selects preset basic gating parameters as initial gating parameters based on machine type information and redundancy strategy information. This dynamic calibration mechanism solves the problem in existing technologies of lacking a configuration fingerprint-driven model gating mechanism, making it difficult to stably migrate prediction capabilities, and ensures that the system can continuously learn and adapt to constantly changing server configurations and operating environments.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An AI-based server health prediction system, characterized in that, include: The data processing unit is used to collect server operation indicator data and abnormal event data, perform unified time base processing on the server operation indicator data and the abnormal event data to generate indicator sequences and event sequences, and perform consistency verification on the indicator sequences and the event sequences to finally obtain the monitoring reliability score. The health prediction unit is used to generate configuration fingerprints based on the asset list, construct component coupling graphs, and update edge weights within a sliding time window. When the monitoring confidence score is not lower than the consistency verification threshold, the indicator sequence, the event sequence, the configuration fingerprint, and the component coupling graph are input into the graph time-series prediction model of the configuration fingerprint gating, and the health prediction result is output; when the monitoring confidence score is lower than the consistency verification threshold, a consistency verification failure indication is output to the operation and maintenance decision unit and a supplementary sampling diagnosis is triggered; wherein, the edges of the component coupling graph include thermal coupling edges, power supply coupling edges, interconnection coupling edges, and redundant replacement edges, and the health prediction result includes server health index, component failure risk probability, and confidence information; The operation and maintenance decision unit is used for server operation and maintenance management. Based on the health prediction results and confidence information, it generates maintenance work orders, maintenance schedules and spare parts scheduling. When the confidence information or the monitoring confidence score is lower than the maintenance threshold, it performs supplementary sampling diagnosis and postpones the issuance of maintenance work orders. The maintenance threshold is not lower than the consistency verification threshold. The calibration unit is used to update the calibration parameters and gating parameters of the graph time series prediction model based on maintenance backfill data.

2. The AI-based server health prediction system according to claim 1, characterized in that, The data processing unit collects server operation indicator data and abnormal event data, generates indicator sequences and event sequences after unifying the time base, and performs consistency verification on the indicator sequences and event sequences. When finally obtaining the monitoring reliability score, the following steps are included: The data processing unit collects the server operation index data and the abnormal event data, and adds a collection source identifier and a collection time identifier to each data record; The server operation index data is resampled into an index sequence using a unified time base, and the abnormal event data is organized into an event sequence in chronological order. The consistency of the indicator sequence and the event sequence is checked. The consistency check includes time sequence check, missing sampling check, duplicate event check and cross-source state contradiction check. The cross-source state contradiction check is used to identify contradictory records based on the logical constraints between the in-situ state, running state and alarm state of the same component. The monitoring reliability score is generated based on the number of contradictory records identified in the consistency check, the number of missing samples, and the number of duplicate events.

3. The AI-based server health prediction system according to claim 1, characterized in that, When the health prediction unit generates a configuration fingerprint based on the asset list, it includes: The health prediction unit extracts model information, key component identification information, firmware version information, component connection relationship information, and redundancy strategy information from the asset list. The configuration fingerprint is obtained by generating configuration records according to a preset field order and performing summary encoding. The configuration fingerprint is associated with and saved with the configuration version; when the configuration fingerprint changes, the calibration unit is triggered to update the gating parameters.

4. The AI-based server health prediction system according to claim 3, characterized in that, When the health prediction unit constructs the component coupling graph and updates the edge weights within the sliding time window, it includes: The health prediction unit establishes a component coupling diagram using key components as nodes. To establish thermally coupled edges, power supply coupled edges, interconnection coupled edges, and redundant replacement edges for node pairs that meet the preset coupling judgment conditions; Within the sliding time window, the correlation stability of node pairs is calculated based on the index sequence, and the event linkage consistency of node pairs is calculated based on the event sequence. The edge weights of the corresponding edges are updated according to the correlation stability and event linkage consistency; the edge weights of the updated edges are checked for stability, and the edge weights of the edges that do not meet the preset stability conditions are kept as the edge weights before the update.

5. The AI-based server health prediction system according to claim 4, characterized in that, The graph time-series prediction model configured with fingerprint gating includes: Feature encoder, graph temporal inferencer, gated fusion unit and confidence information calibration module; The feature encoder is used to encode the index sequence and the event sequence to obtain time-series features; The graph temporal inferencer is used to perform temporal inference on the component coupling graph based on edge weights and attention weights to obtain a risk propagation characterization, wherein the attention weights are calculated by the graph temporal inferencer based on the temporal features; The gated fusion unit is used to determine the gate weights of multiple prediction experts based on the configuration fingerprint and fuse them to obtain a health prediction result. The confidence information calibration module is used to calibrate the confidence information based on the monitoring confidence score.

6. The AI-based server health prediction system according to claim 5, characterized in that, When the operation and maintenance decision unit generates the maintenance work order based on the health prediction result, it includes: The operation and maintenance decision unit determines the components to be maintained based on the component failure risk probability and the server health index; determines the redundancy margin level based on the redundancy replacement relationship; synthesizes the redundancy margin level, the component failure risk probability, and the confidence information into a work order priority according to a preset priority mapping rule; and generates the maintenance work order, which includes a server identifier, a component identifier to be maintained, a work order priority, and a suggested maintenance window.

7. The AI-based server health prediction system according to claim 6, characterized in that, When the operation and maintenance decision-making unit generates the maintenance schedule and the spare parts scheduling, it includes: The operation and maintenance decision unit obtains the maintenance work order and preset maintenance window constraints, spare parts inventory constraints, and personnel shift constraints; determines the false alarm handling cost coefficient based on the confidence information, and constructs a decision objective together with the expected downtime loss; through optimization, under the conditions of satisfying the maintenance window constraints, the spare parts inventory constraints, and the personnel shift constraints, outputs the maintenance schedule, and outputs the spare parts demand list corresponding to the maintenance schedule as the spare parts scheduling.

8. The AI-based server health prediction system according to claim 7, characterized in that, When the confidence information or the monitoring confidence score is lower than the maintenance threshold, the operation and maintenance decision unit performs supplementary sampling diagnosis and postpones the issuance of the maintenance work order. The supplementary sampling diagnosis includes at least one of the following: Increase the sampling frequency of the server operation indicator data, increase the collection range of the abnormal event data, or realign the indicator sequence and the event sequence and recalculate the monitoring reliability score; Once the confidence information and the monitoring credibility score both meet the maintenance threshold, the maintenance work order is issued and the maintenance schedule and spare parts scheduling are updated.

9. The AI-based server health prediction system according to claim 8, characterized in that, When the calibration unit updates the calibration parameters of the graph time-series prediction model based on the maintenance execution backfill data, it includes: The calibration unit acquires maintenance execution backfill data corresponding to the maintenance work order. The maintenance execution backfill data includes fault confirmation results, maintenance action results, component replacement results, and recovery time results. The calibration unit compares the maintenance execution backfill data with the health prediction results according to the server identifier and the identifier of the component to be maintained, and determines the prediction consistency status. The calibration parameters of the confidence information are adjusted according to the prediction consistency status, and the adjusted calibration parameters are associated with the configuration version and saved. The prediction consistency status includes consistency, false alarms, and missed alarms.

10. The AI-based server health prediction system according to claim 9, characterized in that, When the calibration unit updates the gating parameters of the graph time series prediction model based on the maintenance execution backfill data, it includes: The calibration unit collects maintenance execution backfill data according to the configuration fingerprint; for each configuration fingerprint, it determines the prediction error type based on the health prediction result and the corresponding maintenance execution backfill data, and updates the gating weights of multiple prediction experts to form gating parameters accordingly; the gating parameters are associated with and saved with the configuration fingerprint; when it is detected that the configuration fingerprint has not been associated with and saved with the gating parameters, a preset basic gating parameter is selected as the initial gating parameter based on the model information and redundancy strategy information. The prediction error types include false alarms, false misses, and premature and delayed errors, where premature and delayed errors are used to characterize the deviation direction of the predicted fault occurrence time from the fault confirmation time in the maintenance execution backfill data.

Images