Server hardware predictive maintenance method and system based on federated learning and edge computing

Abstract

The application discloses a server hardware predictive maintenance method and system based on federated learning and edge computing, relates to the technical field of server hardware maintenance, and discloses the server hardware predictive maintenance method and system based on federated learning and edge computing. Through collection and processing of sensor data on the edge side, combination of a federated learning training model and dynamic updating, dynamic adaptation to server load changes, reduction of data transmission volume, protection of data privacy, improvement of real-time performance and accuracy of predictive maintenance, and reduction of false alarm and missed alarm risks can be achieved.

Description

Technical Field

[0001] This application relates to the field of server hardware maintenance technology, and in particular to a method and system for predictive maintenance of server hardware based on federated learning and edge computing. Background Technology

[0002] With the continuous expansion of modern data centers, real-time monitoring and preventative maintenance of server hardware failures have become core challenges in IT operations and maintenance management. Traditional monitoring systems generally employ static threshold alarm mechanisms, triggering alarms by pre-setting fixed performance thresholds (such as temperature, memory error rate, etc.). While these systems are simple in structure, static thresholds cannot dynamically adapt to performance fluctuations of servers under different load conditions, leading to frequent false alarms under high load and an increased risk of missed alarms under low load. Passive fault detection can only initiate recovery processes after a hardware failure occurs, completely lacking the ability to proactively warn of potential faults, making preventative maintenance difficult. Centralized data processing architectures require transmitting massive amounts of raw sensor data to a central server, not only causing excessive consumption of network bandwidth resources but also exposing sensitive operational data to the transmission link, leading to data privacy leaks and security risks.

[0003] In recent years, the development of machine learning algorithms and edge computing architectures has provided potential solutions to the above problems. However, in existing technologies, if the prediction model is deployed only at the edge, the limited amount of data on a single node and its uneven distribution result in insufficient training data for the model, making it difficult to guarantee prediction accuracy. If the dataset is centrally uploaded to the cloud for training, although the overall performance of the model can be improved, the risk of privacy leakage during data transmission increases, and the processing latency of the cloud cannot meet the needs of real-time maintenance.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a method and system for predictive maintenance of server hardware based on federated learning and edge computing, which aims to improve the real-time performance and accuracy of predictive maintenance.

[0006] To achieve the above objectives, this application proposes a predictive maintenance method for server hardware based on federated learning and edge computing, the method comprising: At the edge of multiple server nodes, raw sensor data of the server nodes is collected and preprocessed to obtain node preprocessed data. Based on the node preprocessed data, fault risk assessment is performed using a pre-trained prediction model local at the edge, generating node fault prediction data. Based on the node fault prediction data and the current server node load status data, determine whether the local triggering conditions are met. When the local triggering conditions are met, send the node preprocessed data to the cloud server. On the cloud server side, the node preprocessing data from multiple server nodes is received, and the prediction model is trained by federated learning based on all the received node preprocessing data to generate cloud global model parameters. The cloud-based global model parameters are distributed to the edge side of each server node to update the local pre-trained prediction model on the edge side. Predictive maintenance operations are performed on the server hardware based on the node failure prediction data.

[0007] In one embodiment, the step of generating node fault prediction data by performing fault risk assessment using a pre-trained prediction model local at the edge based on the node preprocessing data includes: Multi-dimensional temporal features are extracted from the preprocessed data of the nodes to generate temporal feature vector data; The time-series feature vector data is input into a pre-trained prediction model, which includes a pre-trained time-series neural network. The time-series feature vector data is fused and analyzed using the pre-trained time-series neural network to obtain fused feature data; Based on the fused feature data, fault probability values ​​and fault type data are generated, and the fault probability values ​​and fault type data are combined into the node fault prediction data.

[0008] In one embodiment, the step of extracting multi-dimensional temporal features from the node preprocessing data and generating temporal feature vector data includes: Temperature change rate characteristics are calculated from the preprocessed data of the nodes to obtain temperature time series data; Memory error statistical characteristics are calculated from the node preprocessing data to obtain memory error timing data; The hard disk comprehensive health characteristics are calculated from the node preprocessing data to obtain hard disk health time-series data; The temperature time-series data, the memory error time-series data, and the hard disk health time-series data are weighted and combined according to a preset weight ratio to generate the time-series feature vector data.

[0009] In one embodiment, the step of calculating the overall hard disk health characteristics from the node preprocessing data to obtain hard disk health time-series data includes: Obtain hard disk remapping sector statistics from the node preprocessing data, and calculate remapping health value data based on the preset first weighting coefficient; Obtain the unprocessed sector statistics of the hard disk from the node preprocessing data, and calculate the unprocessed health value data based on the preset second weighting coefficient; Obtain the cyclic redundancy check error statistics of the hard disk from the node preprocessing data, and calculate the check health value data based on the preset third weighting coefficient; The remapped health value data, the health value data to be processed, and the verification health value data are weighted and summed according to the first weight coefficient, the second weight coefficient, and the third weight coefficient to obtain the hard disk health time series data; wherein, the ratio of the first weight coefficient, the second weight coefficient, and the third weight coefficient is a preset fixed ratio.

[0010] In one embodiment, the step of generating node fault prediction data by performing fault risk assessment using a pre-trained prediction model local at the edge based on the node preprocessed data further includes: Based on the node failure prediction data and historical prediction accuracy data, calculate the prediction confidence assessment value data; The predicted confidence assessment value is compared with a preset confidence threshold value. When the predicted confidence assessment value is lower than the confidence threshold value, a multi-model collaborative prediction mechanism is activated. The node fault prediction data is corrected by the multi-model collaborative prediction mechanism to obtain corrected node fault prediction data. The steps of performing predictive maintenance operations on server hardware based on the node fault prediction data include: performing predictive maintenance operations on server hardware based on the corrected node fault prediction data.

[0011] In one embodiment, the step of determining whether a local triggering condition is met based on the node fault prediction data and the current server node load status data, and sending the node preprocessed data to the cloud server when the local triggering condition is met, includes: Analyze the historical trend of fault probability in the node fault prediction data to obtain trend analysis data; Based on the load status data, calculate the dynamic trigger threshold data; Compare the current fault probability data in the node fault prediction data with the dynamic trigger threshold data; When the current fault probability data is greater than the dynamic trigger threshold data, it is determined that the local trigger condition is met; When the local triggering conditions are met, the preprocessed data of the node is sent to the cloud server.

[0012] In one embodiment, the step of calculating the dynamic trigger threshold data in conjunction with the load status data includes: Obtain the current load ratio data from the load status data; Multiply the current load ratio data by a preset first adjustment coefficient data to obtain the load adjustment factor data; The upward trend strength data of the failure probability is analyzed from the trend analysis data, and the upward trend strength data is multiplied by a preset second adjustment coefficient data to obtain the trend adjustment factor data. The preset base threshold data is linearly combined with the load adjustment factor data and the trend adjustment factor data to obtain the dynamic trigger threshold data; wherein, the dynamic trigger threshold data increases with the increase of the current load ratio data and increases with the increase of the upward trend strength data.

[0013] In one embodiment, on the cloud server side, the steps of receiving node preprocessed data from multiple server nodes, and training the prediction model using federated learning based on all received node preprocessed data to generate cloud-based global model parameters include: Receive preprocessed data from each server node to generate training dataset data for multiple nodes; Local model training is performed on the training dataset data of each node to generate node local model parameter data; The node-local model parameter data of each node is randomly masked to obtain privacy-preserving model parameter data; The privacy protection model parameter data of all nodes are weighted and fused to generate the cloud-based global model parameters.

[0014] In one embodiment, the step of performing random masking on the node-local model parameter data of each node to obtain privacy-preserving model parameter data includes: Generate random mask sequence data, wherein each mask element in the random mask sequence data is generated independently according to a preset probability distribution; The random mask sequence data is multiplied element-wise with the node local model parameter data to obtain the parameter data after masking. Based on the feature importance data of the node local model parameter data, the generation probability of the random mask sequence data is dynamically adjusted to obtain the adjusted mask sequence data. The adjusted mask sequence data is then applied again to the node's local model parameter data to generate the privacy-preserving model parameter data.

[0015] Furthermore, to achieve the above objectives, this application also proposes a server hardware predictive maintenance system based on federated learning and edge computing. The server hardware predictive maintenance system based on federated learning and edge computing includes: a memory, a processor, and a server hardware predictive maintenance program based on federated learning and edge computing stored in the memory and executable on the processor. The server hardware predictive maintenance program based on federated learning and edge computing is configured to implement the steps of the server hardware predictive maintenance method based on federated learning and edge computing.

[0016] The predictive maintenance method and system for server hardware based on federated learning and edge computing proposed in this application can dynamically adapt to changes in server load, reduce data transmission volume, protect data privacy, and improve the real-time performance and accuracy of predictive maintenance by collecting and processing sensor data at the edge, combining federated learning to train and dynamically update the model, thereby reducing the risk of false alarms and missed alarms. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating an embodiment of the predictive maintenance method for server hardware based on federated learning and edge computing provided in this application. Figure 2 This is a schematic diagram of a structural embodiment of the server hardware predictive maintenance system based on federated learning and edge computing provided in this application.

[0020] Explanation of icon numbers: 10. Memory; 20. Processor.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of this application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0023] It should be understood that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] In existing technologies, server hardware fault monitoring systems suffer from high false alarm rates when facing dynamically changing load conditions, particularly regarding preset threshold alarms. Furthermore, their passive fault detection mechanisms cannot provide early warnings or preventative maintenance for potential faults. In addition, centralized data collection and processing models not only impose significant bandwidth pressure but may also raise data privacy and security concerns. Current machine learning and edge computing solutions still have shortcomings in terms of data volume, real-time performance, and privacy protection.

[0025] Based on this, embodiments of this application provide a predictive maintenance method for server hardware based on federated learning and edge computing, referring to... Figure 1 The server hardware predictive maintenance method based on federated learning and edge computing includes steps S100 to S600, wherein: Step S100: At the edge of multiple server nodes, collect the raw sensor data of the server nodes and preprocess the raw sensor data to obtain node preprocessed data. Step S200: Based on the node preprocessing data, perform fault risk assessment through a pre-trained prediction model on the edge side to generate node fault prediction data. Step S300: Based on the node fault prediction data and the current server node load status data, determine whether the local triggering condition is met. When the local triggering condition is met, send the node preprocessed data to the cloud server. Step S400: On the cloud server, the node preprocessing data from multiple server nodes is received, and the prediction model is trained by federated learning based on all the received node preprocessing data to generate cloud global model parameters. Step S500: The cloud-based global model parameters are distributed to the edge side of each server node to update the local pre-trained prediction model on the edge side. Step S600: Perform predictive maintenance operations on the server hardware based on the node fault prediction data.

[0026] In this embodiment, federated learning is a distributed machine learning paradigm that allows multiple clients to train models locally and upload their local model parameters or gradients to a central server for aggregation, thereby generating a global model without directly sharing the original training data. This mechanism helps protect data privacy and reduces data transmission bandwidth requirements. Edge computing is a distributed computing paradigm that moves computation and data storage to the network edge near the data source. By performing data processing and analysis on edge devices close to the data source, latency can be reduced, bandwidth can be saved, and real-time response capabilities can be improved. A server node refers to a single physical or virtual server unit that constitutes a data center or distributed computing environment. Each server node typically contains hardware resources such as processors, memory, storage, and network interfaces, and runs an operating system and various applications. Raw sensor data refers to unprocessed initial data directly collected by internal or external sensors of the server hardware, such as temperature, voltage, current, fan speed, hard drive SMART information, etc.

[0027] In this embodiment, node preprocessing data refers to data obtained after cleaning, denoising, format conversion, missing value imputation, and normalization of raw sensor data. This data is more suitable as input for machine learning models. The pre-trained prediction model refers to a machine learning model deployed locally on the edge side that has undergone preliminary training. This model can perform preliminary fault risk assessment based on node preprocessing data and generate node fault prediction data. Node fault prediction data refers to the risk assessment results of potential future failures of server nodes output by the local pre-trained prediction model on the edge side, typically including information such as failure probability and possible failure types. Local triggering conditions refer to the conditions at the server node edge side for determining whether local data needs to be uploaded to the cloud server for federated learning training. These conditions typically consider factors such as fault risk, data freshness, and load status. The cloud server refers to the core server located in the data center, responsible for receiving local model parameters or preprocessing data from multiple edge server nodes and performing operations such as federated learning aggregation, global model updates, and model parameter distribution. Cloud global model parameters refer to the prediction model parameters representing global knowledge generated on the cloud server after aggregating multiple edge-side local model parameters through federated learning. This parameter is sent to the edge side to update the local model, thereby improving the overall prediction capability.

[0028] In this embodiment, the server hardware predictive maintenance method based on federated learning and edge computing first performs data acquisition and preprocessing at the edge of multiple server nodes. Specifically, raw data can be periodically acquired from various sensors on the server nodes, such as temperature, fan speed, and hard drive SMART information obtained by reading system logs or hardware interfaces. This raw data is then pre-processed, for example, by removing outliers using basic statistical methods or by standardizing the data format, thereby obtaining node preprocessed data. For example, a fixed sampling frequency can be set, collecting readings from all sensors every minute and storing these readings directly as preprocessed data.

[0029] Furthermore, based on the preprocessed data from this node, a fault risk assessment is performed using a pre-trained prediction model located locally at the edge, generating node fault prediction data. This pre-trained prediction model can be a basic linear regression model or a decision tree model, pre-deployed at the edge and initially trained using a small amount of local data. When the node preprocessed data is received, the model analyzes the data according to its internal logic, such as calculating the average or trend of certain indicators, and outputs a numerical value representing the fault risk; this value is the node fault prediction data. For example, a threshold can be set; when a sensor reading continuously exceeds this threshold for a certain period of time, the model outputs fault prediction data indicating a high risk.

[0030] Based on this, the method determines whether a local triggering condition is met based on node failure prediction data and the current server node's load status data. This local triggering condition can be a preset, fixed rule; for example, when the failure probability in the node failure prediction data exceeds a fixed threshold, or when the server node's CPU utilization exceeds a certain fixed percentage, the local triggering condition is considered met. When the local triggering condition is met, the node's preprocessed data will be sent to the cloud server. For example, it can be set that when the predicted failure probability is greater than 0.7, regardless of the load, the preprocessed data for the most recent period will be uploaded to the cloud.

[0031] In this embodiment, the cloud server then receives preprocessed data from multiple server nodes. The cloud server aggregates all received preprocessed data and uses this data to train the prediction model using federated learning. This training process can employ a basic federated averaging algorithm, where the cloud server receives local model parameters (not raw data) from each edge node, performs a simple arithmetic average on these parameters, and generates global model parameters in the cloud. For example, each edge node can independently train a small neural network and upload its weight parameters to the cloud, where the cloud server averages these weights. Subsequently, these global model parameters are distributed to the edge sides of each server node to update the local pre-trained prediction model. Upon receiving the global model parameters from the cloud, the local model on the edge side can directly replace its original local model parameters with these global model parameters, thus updating the model. For example, the prediction model on the edge side can be completely replaced with the latest global model distributed from the cloud.

[0032] In this embodiment, predictive maintenance operations on the server hardware are ultimately performed based on the node failure prediction data. These maintenance operations may include generating an alarm notification to remind maintenance personnel to manually check the server node, or triggering a pre-defined automated script, such as performing a hard drive health scan or restarting a service. For example, when the node failure prediction data indicates a high risk of hard drive failure, the system can automatically generate a work order requiring maintenance personnel to check the hard drive status.

[0033] In this embodiment, by performing preliminary fault risk assessment at the edge and selectively uploading data to the cloud for federated learning training based on local triggering conditions, the problems of high false alarm rates and inability to provide early warnings in traditional monitoring systems are effectively solved. Simultaneously, this method avoids the bandwidth pressure and data privacy and security risks associated with centralized transmission of large amounts of raw data, enabling accurate prediction and timely maintenance of server hardware failures, thus improving operational efficiency and system stability.

[0034] In one feasible implementation, the step of generating node fault prediction data by performing fault risk assessment through a pre-trained prediction model local at the edge based on the node preprocessed data includes: extracting multi-dimensional temporal features from the node preprocessed data to generate temporal feature vector data; inputting the temporal feature vector data into a pre-trained prediction model, the pre-trained prediction model including a pre-trained time series neural network; performing fusion analysis on the temporal feature vector data through the pre-trained time series neural network to obtain fused feature data; generating fault probability values ​​and fault type data based on the fused feature data; and combining the fault probability values ​​and the fault type data into the node fault prediction data.

[0035] In this embodiment, extracting multi-dimensional temporal features from the preprocessed node data to generate temporal feature vector data refers to identifying and quantifying various time-varying indicators that reflect hardware operating status and potential fault trends from the pre-processed server node sensor data. These features can be physical quantities (such as temperature, voltage, current, fan speed), performance indicators (such as CPU utilization, memory utilization, disk I / O), error counts (such as memory errors, disk bad sector counts), etc. The extraction process may involve statistical analysis (such as mean, variance, trend, peak value), signal processing techniques (such as Fourier transform to extract frequency components), or domain expert knowledge, aiming to transform the raw data into a structured, meaningful temporal feature vector to facilitate model learning. For example, temperature change rate, memory error growth trend, and hard disk read / write latency fluctuation can be calculated as temporal features.

[0036] In this embodiment, the time-series feature vector data is input into a pre-trained prediction model, which includes a pre-trained time-series neural network. This pre-trained prediction model is a model that has undergone initial training on a large amount of historical server operation and failure data, possessing the ability to identify common hardware failure modes. The core component of this model is the pre-trained time-series neural network, a neural network structure specifically designed to process sequential data (such as time-series feature vector data). Common time-series neural networks include recurrent neural networks (RNNs), long short-term memory networks (LSTMs), gated recurrent units (GRUs), or variants of the Transformer. These networks can effectively capture the time dependencies, long-term patterns, and complex relationships between different features in the data, thereby better understanding the dynamic evolution of hardware states. Inputting the time-series feature vector data into such a model means that the model will utilize the knowledge learned during the pre-training phase to perform in-depth analysis of these features.

[0037] In this embodiment, the time-series feature vector data is fused and analyzed using the pre-trained time-series neural network to obtain fused feature data. Fusion analysis refers to the pre-trained time-series neural network performing in-depth processing and abstraction on the input time-series feature vector data. Multiple layers of the neural network progressively extract and combine high-level semantic information from the original time-series features, such as identifying the correlation between different sensor data, the evolution trend of fault modes, and the duration of abnormal behavior. This process transforms the original multi-dimensional time-series features into more representative and discriminative fused feature data. The fused feature data is typically the output of the hidden layers of the neural network; it condenses the information most relevant to fault risk assessment from the original data, providing a more refined and effective input for subsequent fault prediction.

[0038] In this embodiment, fault probability values ​​and fault type data are generated based on the fused feature data, and the fault probability values ​​and fault type data are combined to form the node fault prediction data. After obtaining the fused feature data, the output layer of the pre-trained prediction model performs classification or regression prediction based on these fused features. For classification tasks, the model outputs the probability distribution of different fault types (e.g., CPU fault, memory fault, hard disk fault, power supply fault, etc.), where the highest probability value is the fault probability value, and the corresponding category is the fault type data. For regression tasks, the model may directly output a continuous value representing the degree of fault risk. These outputs together constitute the node fault prediction data, which not only quantifies the probability of server hardware failure but also indicates the possible location or nature of the fault, providing specific guidance for subsequent predictive maintenance operations.

[0039] In this embodiment, through the above technical solution, this application can perform a more refined and intelligent assessment of server hardware failure risks. By extracting multi-dimensional temporal features from node preprocessing data, the system can capture the dynamic behavior and interdependencies between various sensor readings, which are key information for identifying subtle precursors to hardware failures. The use of pre-trained time-series neural networks enables the model to deeply learn and automatically discover complex patterns and temporal correlations that are difficult to detect using traditional methods. The fusion analysis step further refines these patterns into highly discriminative features, making the generated failure probability values ​​and failure type data more accurate and instructive. This detailed predictive capability improves the effectiveness of predictive maintenance, enabling maintenance personnel to intervene in a targeted manner before actual failures occur, thereby effectively reducing server downtime and extending hardware lifespan. Compared to only performing general failure risk assessments, this solution provides more accurate and actionable predictive results.

[0040] In one feasible implementation, the step of extracting multi-dimensional temporal features from the node preprocessing data and generating temporal feature vector data includes: calculating temperature change rate features from the node preprocessing data to obtain temperature temporal data; calculating memory error statistical features from the node preprocessing data to obtain memory error temporal data; calculating hard disk comprehensive health features from the node preprocessing data to obtain hard disk health temporal data; and weighting and combining the temperature temporal data, the memory error temporal data, and the hard disk health temporal data according to a preset weight ratio to generate the temporal feature vector data.

[0041] In this embodiment, temperature change rate characteristics are calculated from the preprocessed data of the nodes to obtain temperature time-series data. This step aims to identify trends and abnormal fluctuations in temperature over time from the temperature sensor data of the server nodes. The calculation of temperature change rate characteristics typically involves processing continuously collected temperature values ​​using methods such as differencing, moving averages, or regression analysis to obtain the temperature rise and fall amplitude per unit time. For example, the difference in temperature values ​​at adjacent time points can be calculated, or the slope of a temperature curve can be fitted within a certain time window. By focusing on the rate of temperature change rather than the absolute value, early warning signals such as heat dissipation system failures and component overheating can be captured more sensitively, thereby forming temperature time-series data reflecting dynamic temperature changes.

[0042] In this embodiment, memory error statistical characteristics are calculated from the node preprocessing data to obtain memory error timing data. This step focuses on the operational health of the server memory module. The calculation of memory error statistical characteristics typically includes counting or frequency analysis of events such as memory error correction code (ECC) errors, uncorrectable errors, and abnormal memory access latency. This error information can be obtained from operating system logs, hardware management interfaces (such as IPMI), or memory controller registers. By statistically analyzing the frequency, type, or rate of these error events over a continuous time period, memory error timing data reflecting the evolution of memory health over time can be generated, providing a basis for assessing the stability and reliability of the memory module.

[0043] In this embodiment, the comprehensive health characteristics of the hard drive are calculated from the node preprocessing data to obtain hard drive health time-series data. This step aims to comprehensively assess the health status of the server storage device. The calculation of comprehensive hard drive health characteristics typically involves collecting and analyzing the hard drive's SMART (Self-Monitoring, Analysis and Reporting Technology) attributes. These attributes include, but are not limited to, remapped sector count, unprocessed sector count, uncorrectable sector count, cyclic redundancy check (CRC) error count, hard drive temperature, and power-on time. By weighting, normalizing, or aggregating these key SMART attributes, a value that comprehensively reflects the overall health status of the hard drive can be obtained. Recording these comprehensive health values ​​at different time points forms the hard drive health time-series data.

[0044] Based on this, the temperature time-series data, the memory error time-series data, and the hard disk health time-series data are weighted and combined according to a preset weight ratio to generate the time-series feature vector data. This step aims to fuse server hardware health time-series data from different sources and of different types to form a unified time-series feature vector with higher information density. Weighted combination means assigning different weights to different features based on their importance in predicting server failures. For example, certain types of memory errors may more directly predict hardware failures than slight temperature fluctuations, and therefore can be assigned higher weights. These weight ratios can be preset and optimized based on historical data analysis, expert experience, or machine learning methods. Through this weighted combination, a multi-dimensional time-series feature vector data can be generated, which contains comprehensive health information of various key components of the server hardware at each time point, providing a comprehensive and structured input for subsequent prediction models.

[0045] In this embodiment, through the above technical solution, this application can systematically extract multi-dimensional time-series features crucial to server hardware health from node preprocessing data. Specifically, by calculating temperature change rate features, early signs of abnormal heat dissipation or component overheating within the server can be detected in a timely manner; by statistically analyzing memory error features, the stability and reliability of memory modules can be directly reflected; and by comprehensively evaluating various health indicators of the hard drive, the operating status of the storage device can be fully understood. These carefully selected and weighted time-series features, compared to generalized feature extraction methods, can more comprehensively and accurately characterize the potential failure modes and evolution trends of server hardware. Therefore, inputting these high-quality time-series feature vector data into the pre-trained prediction model improves the accuracy and sensitivity of fault risk assessment, enabling predictive maintenance operations to be performed more promptly and effectively, thereby effectively avoiding prediction bias caused by insufficient feature extraction and ensuring the stable operation of the server system and data security.

[0046] In one feasible implementation, the step of calculating the hard disk comprehensive health characteristics from the node preprocessing data to obtain hard disk health time series data includes: obtaining the hard disk remapping sector statistics data from the node preprocessing data, and calculating remapping health value data based on a preset first weighting coefficient. The hard disk sector statistics to be processed are obtained from the node preprocessing data, and the health value data to be processed is calculated based on the preset second weighting coefficient; the hard disk cyclic redundancy check error statistics are obtained from the node preprocessing data, and the check health value data is calculated based on the preset third weighting coefficient; the remapping health value data, the health value data to be processed, and the check health value data are weighted and summed according to the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient to obtain the hard disk health time series data; wherein, the ratio of the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient is a preset fixed ratio.

[0047] In this embodiment, obtaining remapped sector statistics data of the hard drive and calculating remapped health value data based on a preset first weighting coefficient refers to the system extracting the number of remapped sectors of the hard drive from the preprocessed data of the server node. A remapped sector is the result of the hard drive transferring data to a spare sector and marking the original sector as unusable after detecting a damaged sector. An increase in the number of remapped sectors is usually an important indicator of the aging of the hard drive's physical media. The first weighting coefficient is used to quantify the impact of the number of remapped sectors on the overall health of the hard drive. The remapped health value data is a quantitative indicator calculated based on the remapped sector statistics data and the first weighting coefficient, used to reflect the health level of the hard drive due to remapped sector problems. For example, the remapped sector statistics data can be obtained by reading the "Reallocated Sector Count" attribute in the hard drive's SMART (Self-Monitoring, Analysis and Reporting Technology) data. When calculating the remapped health value data, a baseline health value (such as 100) can be set, and the health value can be decreased proportionally to the first weighting coefficient as the number of remapped sectors increases.

[0048] In this embodiment, obtaining the statistical data of unprocessed sectors on the hard drive and calculating the unprocessed health value data based on a preset second weighting coefficient means that the system extracts the number of unprocessed sectors on the hard drive from the node preprocessing data. Unprocessed sectors refer to sectors on the hard drive that encountered problems during reading but have not yet been remapped. These sectors may be successfully read in subsequent operations, or they may eventually be marked as bad sectors and remapped. An increase in the number of unprocessed sectors usually indicates potential physical problems with the hard drive and is an early warning signal of impending hard drive failure. The second weighting coefficient is used to quantify the impact of the number of unprocessed sectors on the overall health of the hard drive. The unprocessed health value data is a quantitative indicator calculated based on the unprocessed sector statistical data and the second weighting coefficient. For example, the unprocessed sector statistical data can be obtained through the "Current PendingSector Count" attribute in the SMART data. The calculation method for the unprocessed health value data is similar to that for the remapped health value data, but the second weighting coefficient may be different to reflect the difference in the severity of failure between unprocessed sectors and remapped sectors.

[0049] In this embodiment, obtaining CRC error statistics from the hard drive and calculating the check health value data based on a preset third weighting coefficient refers to the system extracting CRC error statistics from the node preprocessing data. CRC errors typically occur during data transmission, indicating that errors occurred when data is read from the hard drive to memory or written from memory to the hard drive. Frequent CRC errors may indicate problems with the hard drive interface, data cable, or internal hard drive controller, and are an important indicator of compromised hard drive data integrity. The third weighting coefficient is used to quantify the impact of CRC errors on the overall health of the hard drive. The check health value data is a quantitative indicator calculated based on the CRC error statistics and the third weighting coefficient. For example, CRC error statistics can be obtained through the "UDMA CRC Error Count" attribute in SMART data. The check health value data can be calculated based on the cumulative number or frequency of CRC errors and adjusted by the third weighting coefficient.

[0050] In this embodiment, the remapped health value data, the health value data to be processed, and the verification health value data are weighted and summed according to a first weighting coefficient, a second weighting coefficient, and a third weighting coefficient to obtain hard disk health time-series data. The ratio of the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient is a preset fixed ratio. Weighted summation is a common method to combine multiple indicators into a single indicator. By assigning different weights to each indicator, its relative importance in the overall evaluation can be reflected. The hard disk health time-series data is a quantitative result that integrates the above three key health indicators. It can more comprehensively and precisely reflect the real-time health status of the hard disk and its trend over time. The preset fixed ratio ensures that the relative importance of different hard disk health indicators remains consistent in the comprehensive evaluation, thereby guaranteeing the stability and comparability of the evaluation results. For example, it can be achieved using the formula "Hard disk health timing data = W1 × remapping health value data + W2 × unprocessed health value data + W3 × verification health value data", where W1, W2, and W3 are the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively, and their proportions are determined in advance based on experience, expert knowledge, or historical data analysis.

[0051] In this embodiment, the assessment of the overall health characteristics of the hard drive is refined into a quantitative analysis of three key indicators: remapped sector statistics, unprocessed sector statistics, and cyclic redundancy check error statistics. By assigning preset weight coefficients to each indicator and performing weighted summation, multiple potential failure modes and degradation trends of the hard drive can be captured more comprehensively and precisely. This multi-dimensional, weighted combination assessment method improves the accuracy and sensitivity of hard drive health status assessment, enabling subsequent failure risk assessment to be based on more reliable hard drive health time-series data. This effectively avoids misjudgments or omissions caused by single or general assessments, thereby improving the accuracy and timeliness of predictive maintenance of server hardware.

[0052] In one feasible implementation, the step of generating node failure prediction data by performing fault risk assessment using a pre-trained prediction model locally on the edge side based on the node preprocessed data further includes: calculating prediction confidence assessment value data based on the node failure prediction data and historical prediction accuracy data; comparing the prediction confidence assessment value data with a preset confidence threshold data; and activating a multi-model collaborative prediction mechanism when the prediction confidence assessment value data is lower than the confidence threshold data; correcting the node failure prediction data through the multi-model collaborative prediction mechanism to obtain corrected node failure prediction data; and performing predictive maintenance operations on server hardware based on the node failure prediction data, including: performing predictive maintenance operations on server hardware based on the corrected node failure prediction data.

[0053] In this embodiment, after generating node fault prediction data, a prediction confidence assessment value is calculated based on the node fault prediction data and historical prediction accuracy data. This step aims to quantify the reliability of the current node fault prediction data (including fault probability values ​​and fault type data). Specifically, it can be combined with the performance of the pre-trained prediction model on historical data, such as prediction accuracy, recall, or F1 score under similar operating conditions or data distributions. For example, if the model's historical prediction accuracy for a specific fault type is low, the current prediction confidence assessment value for that type of fault will also be correspondingly lower. Furthermore, the entropy value of the predicted probability distribution output by the model can also be considered; a higher entropy value indicates higher uncertainty about the prediction result, resulting in a lower prediction confidence assessment value. In this way, the system can self-evaluate the quality of each prediction result.

[0054] In this embodiment, the predicted confidence assessment value is compared with a preset confidence threshold. This step is used to determine whether the current prediction result has reached the expected reliability level. The preset confidence threshold is a configurable parameter, which can be set based on factors such as the potential impact of server hardware failure, maintenance costs, and tolerance for prediction accuracy. When the calculated predicted confidence assessment value is lower than the confidence threshold, it indicates that the prediction result of a single pre-trained prediction model may have a large deviation or uncertainty. At this time, the system will automatically trigger a multi-model collaborative prediction mechanism.

[0055] In this embodiment, the multi-model collaborative prediction mechanism aims to improve the accuracy and robustness of predictions by integrating the advantages of multiple prediction models. This mechanism can be implemented in various ways. For example, multiple prediction models with different architectures (such as support vector machines, random forests, and other types of neural networks) or different training strategies can be deployed. When confidence levels are insufficient, these models can predict the same time-series feature vector data in parallel. Another approach is to use model ensemble techniques, such as voting, weighted averaging, or stacking, to fuse the prediction results of multiple models. Furthermore, a rule system based on expert knowledge can be introduced to correct the prediction results using predefined rules when model predictions are uncertain.

[0056] In this embodiment, the node fault prediction data is corrected using the multi-model collaborative prediction mechanism to obtain corrected node fault prediction data. The correction process is the core of the multi-model collaborative prediction mechanism. For example, if a voting method is used, when multiple models give different predictions for the fault type, the prediction result of the majority model can be adopted as the corrected fault type. If a weighted average method is used, different weights can be assigned to the prediction results of each model based on their historical performance or current prediction confidence, and then a weighted average can be performed to obtain the corrected fault probability value. This approach effectively reduces the risk caused by prediction bias from a single model, making the final prediction result closer to the actual situation.

[0057] In this embodiment, the step of performing predictive maintenance operations on server hardware based on the node fault prediction data includes: performing predictive maintenance operations on server hardware based on the corrected node fault prediction data. This step ensures that subsequent maintenance decisions are based on more reliable fault prediction data that has been collaboratively corrected by multiple models. This means that the system will plan and execute maintenance tasks based on more accurate fault probability values ​​and fault type data, such as issuing more precise warnings, automatically generating more reasonable maintenance work orders, scheduling more suitable maintenance resources, or suggesting the replacement of specific components.

[0058] In this embodiment, by introducing a prediction confidence assessment and a multi-model collaborative prediction mechanism, the system can dynamically identify and correct low-confidence fault predictions, thereby improving the accuracy and reliability of server hardware fault prediction. This not only avoids resource waste and unnecessary maintenance operations caused by false alarms, but also effectively reduces the risk of unexpected server downtime caused by missed alarms, thereby optimizing maintenance resource allocation and improving server operational stability and availability.

[0059] In one feasible implementation, the step of determining whether a local triggering condition is met based on the node fault prediction data and the current server node load status data, and sending the node preprocessing data to the cloud server when the local triggering condition is met, includes: analyzing the historical trend of fault probability in the node fault prediction data to obtain trend analysis data; calculating dynamic triggering threshold data by combining the load status data; comparing the current fault probability data in the node fault prediction data with the dynamic triggering threshold data; determining that the local triggering condition is met when the current fault probability data is greater than the dynamic triggering threshold data; and sending the node preprocessing data to the cloud server when the local triggering condition is met.

[0060] In this embodiment, trend analysis data is obtained by analyzing the historical trend of failure probability in node failure prediction data. This refers to identifying and quantifying the pattern of failure probability value of server nodes over a period of time. By performing time series analysis on historical failure probability data, such as using algorithms like moving average, exponential smoothing, and linear regression, indicators such as the average growth rate, maximum growth rate, and duration of continuous growth of failure probability can be calculated, thus forming trend analysis data. This data can reflect whether the failure risk is rising, falling, or remaining stable, providing an important basis for subsequent dynamic threshold calculation.

[0061] In this embodiment, calculating the dynamic trigger threshold data by combining load status data refers to adaptively adjusting the data upload trigger threshold based on the actual operating status and fault risk trend of the server node. The dynamic trigger threshold data is a variable threshold that can be adjusted in real time based on trend analysis data and the current server load status data. For example, a weighted summation model can be designed to linearly combine a preset base threshold with a load factor determined based on load status (such as CPU utilization, memory usage, network bandwidth utilization, etc.) and a trend factor determined based on the strength of the upward trend in fault probability. This dynamic adjustment mechanism allows the threshold to reasonably reflect the current urgency and resource availability of the server, avoiding the limitations that may arise from fixed thresholds.

[0062] In this embodiment, comparing the current fault probability data in the node fault prediction data with the dynamic trigger threshold data is the core step in determining whether the local triggering conditions are met. The edge-side control module periodically obtains the latest fault probability value output by the pre-trained prediction model and compares it with the dynamically calculated trigger threshold data in real time. This comparison operation ensures that the triggering conditions are based on a comprehensive consideration of real-time risk and system status.

[0063] In this embodiment, when the current failure probability data is greater than the dynamic trigger threshold data, the local trigger condition is determined to be met. This means that when the current failure risk of a server node (represented by the failure probability data) exceeds the threshold dynamically adjusted based on historical trends and current load, the system considers that the node has a high potential failure risk or requires further learning and correction from the cloud model, thus fulfilling the condition for uploading data to the cloud. Sending the node preprocessed data to the cloud server when the local trigger condition is met is the final execution action of the trigger mechanism. Once the local trigger condition is determined to be met, the edge system activates its communication module, which is responsible for establishing a secure communication connection with the cloud server and packaging and encrypting the previously collected and preprocessed node preprocessed data before sending it to the cloud server over the network. This data will be used for federated learning training in the cloud to update the global prediction model.

[0064] In this embodiment, through the above technical solution, this application can achieve a more intelligent and adaptive data upload strategy. When the failure risk of server nodes shows an upward trend, or when the system load is low, even if the current failure probability value is relatively low, data upload may be triggered earlier, thereby allowing the cloud-based global model to learn potential failure modes in a timely manner, improving the timeliness and accuracy of prediction. Conversely, when the failure probability fluctuates but does not show a significant upward trend, or when the system load is high, the trigger threshold can be appropriately increased to avoid unnecessary frequent data uploads, thereby effectively reducing network bandwidth consumption and cloud computing resource consumption. This dynamic adjustment mechanism allows the edge side to flexibly decide when to send node preprocessed data to the cloud server based on actual operating conditions, thereby optimizing the utilization of system resources while ensuring the efficiency and accuracy of prediction model updates, avoiding the risk of false alarms or missed alarms caused by fixed thresholds, and improving the robustness and economy of the entire predictive maintenance system.

[0065] In one feasible implementation, the step of calculating the dynamic trigger threshold data in conjunction with the load status data includes: obtaining the current load ratio data from the load status data; multiplying the current load ratio data by a preset first adjustment coefficient to obtain load adjustment factor data; analyzing the upward trend strength data of the failure probability from the trend analysis data, multiplying the upward trend strength data by a preset second adjustment coefficient to obtain trend adjustment factor data; and linearly combining the preset base threshold data with the load adjustment factor data and the trend adjustment factor data to obtain the dynamic trigger threshold data; wherein the dynamic trigger threshold data increases with the increase of the current load ratio data and increases with the increase of the upward trend strength data.

[0066] In this embodiment, obtaining the current load ratio data from the load status data refers to real-time monitoring of the resource usage of server nodes at a certain moment and representing it as the percentage of used resources out of the total resources. This data can be obtained through API interfaces provided by the operating system or monitoring tools of the virtualization platform. For example, metrics such as CPU utilization, memory utilization, disk I / O rate, and network bandwidth usage can be obtained, and these metrics can be comprehensively evaluated or weighted averaged to obtain a ratio value representing the overall server load. For example, when CPU utilization reaches 80% or memory utilization reaches 90%, the current load ratio data can be considered high.

[0067] In this embodiment, the current load ratio data is multiplied by a preset first adjustment coefficient to obtain load adjustment factor data. The load adjustment factor data is used to quantify the impact of the current server load on the dynamic trigger threshold data. In this scheme, when the current load ratio data of the server increases, the dynamic trigger threshold data also increases, which means that a higher current failure probability data is required to meet the local triggering conditions. Therefore, the first adjustment coefficient data should be set to a positive value to ensure that the load adjustment factor data increases accordingly when the current load ratio data increases. For example, if the first adjustment coefficient data is set to 0.1, when the current load ratio data is 0.7, the load adjustment factor data is 0.07.

[0068] In this embodiment, the upward trend strength data of the failure probability is analyzed from the trend analysis data, and the upward trend strength data is multiplied by a preset second adjustment coefficient to obtain the trend adjustment factor data. The upward trend strength data reflects the urgency of the failure probability changing over time, and can be obtained by performing time series analysis on historical failure probability data, such as calculating the average growth rate, slope, or second derivative of the failure probability within a specific time window. In this scheme, when the upward trend strength data of the failure probability increases, the dynamic trigger threshold data also increases, making the local triggering conditions more stringent. Therefore, the second adjustment coefficient data should be set to a positive value to ensure that the trend adjustment factor data increases accordingly when the upward trend strength data increases. For example, if the second adjustment coefficient data is set to 0.05, when the upward trend strength data is 0.6, the trend adjustment factor data is 0.03.

[0069] In this embodiment, the preset base threshold data is linearly combined with the load adjustment factor data and the trend adjustment factor data to obtain the dynamic trigger threshold data. Linear combination is a method that comprehensively considers multiple influencing factors to determine the final threshold. The dynamic trigger threshold data can be obtained by adding the preset base threshold data with the calculated load adjustment factor data and trend adjustment factor data, i.e.: Dynamic trigger threshold data = Base threshold data + Load adjustment factor data + Trend adjustment factor data. Here, the base threshold data is a preset benchmark value, representing the trigger standard under normal and stable conditions.

[0070] In this embodiment, through the above technical solution, this application can achieve dynamic and refined management of local triggering conditions. By incorporating the current load ratio data of the server and the intensity of the upward trend of the failure probability data into the calculation of the dynamic triggering threshold, and increasing it as these factors increase, the system can more intelligently determine when to send node preprocessed data to the cloud server. This mechanism avoids frequent data uploads due to overly sensitive threshold settings when the server load is high or the failure trend is unclear, thereby effectively reducing the communication overhead between the edge side and the cloud and the computational burden on the cloud server. At the same time, through this more stringent triggering condition, it is ensured that the cloud model is updated only when the failure risk reaches a high level, so that resources are used more effectively and the operating efficiency and stability of the entire predictive maintenance system are improved.

[0071] In one feasible implementation, on the cloud server side, the steps of receiving node preprocessed data from multiple server nodes and performing federated learning training on the prediction model based on all received node preprocessed data to generate cloud-based global model parameters include: receiving node preprocessed data from each server node to generate training dataset data for multiple nodes; training a local model based on the training dataset data for each node to generate node-local model parameter data; performing random masking processing on the node-local model parameter data for each node to obtain privacy-preserving model parameter data; and weightedly fusing the privacy-preserving model parameter data from all nodes to generate the cloud-based global model parameters.

[0072] In this embodiment, the preprocessed data is first received from each server node, and training datasets for multiple nodes are generated accordingly. Specifically, the cloud server establishes a secure communication channel to receive preprocessed data sent by each edge server node after meeting local triggering conditions. This data is organized on the cloud server into independent training datasets corresponding to each node, serving as input for subsequent local model training.

[0073] In this embodiment, secondly, local model training is performed based on the training dataset data of each node to generate node-local model parameter data. Each server node independently trains the prediction model using its local training dataset data. This training process typically involves loading initial global model parameters, and then iteratively learning on the local dataset using optimization algorithms (such as gradient descent) to adjust parameters such as model weights and biases, ultimately generating a set of node-local model parameter data that reflects the characteristics of the node's data.

[0074] In this embodiment, the local model parameter data of each node is further subjected to random masking to obtain privacy-preserving model parameter data. Before uploading the local model parameters to the cloud server, these parameters are randomly masked to enhance data privacy. This process can be achieved by adding random noise conforming to a specific distribution (such as a Laplace distribution or a Gaussian distribution) to the local model parameter data, or by employing techniques such as secure multi-party computation or homomorphic encryption, making it difficult for the cloud server to deduce the original sensitive data information from the model parameters of a single node.

[0075] In this embodiment, the privacy-preserving model parameter data of all nodes are finally weighted and fused to generate the cloud-based global model parameters. After collecting all privacy-preserving model parameter data that has undergone random masking, the cloud server performs a weighted average of these parameters according to a preset aggregation algorithm (such as a federated average algorithm). The weights can be determined based on the size or quality of each node's training dataset to ensure that nodes with greater data contributions occupy a more important position in the global model. In this way, the learning experience of all nodes is integrated to generate a more robust and generalized cloud-based global model parameter set.

[0076] In this embodiment, by employing the aforementioned technical solution, when performing federated learning training on the cloud server, the original node-local model parameter data is no longer directly aggregated. Instead, the local model parameter data of each node is first randomly masked to obtain privacy-preserving model parameter data. This approach effectively reduces the risk of sensitive node preprocessing data being leaked during the transmission and aggregation of local model parameters, thereby enhancing the privacy protection capabilities of the entire predictive maintenance system. By weighted fusion of the privacy-preserving model parameter data, the cloud server can generate cloud-based global model parameters that integrate the learning experience of all nodes, while ensuring data privacy and improving the security of federated learning. This enables the server hardware predictive maintenance method to better meet data privacy compliance requirements while utilizing distributed data for model training, thereby improving system reliability and user trust.

[0077] In one feasible implementation, the step of performing random masking processing on the node-local model parameter data of each node to obtain privacy-preserving model parameter data includes: generating random mask sequence data, wherein each mask element in the random mask sequence data is generated independently according to a preset probability distribution; performing element-wise multiplication of the random mask sequence data with the node-local model parameter data to obtain masked parameter data; dynamically adjusting the generation probability of the random mask sequence data according to the feature importance data of the node-local model parameter data to obtain adjusted mask sequence data; and applying the adjusted mask sequence data again to the node-local model parameter data to generate the privacy-preserving model parameter data.

[0078] In this embodiment, a random mask sequence data is generated. Each mask element in the random mask sequence data is generated independently according to a preset probability distribution. This random mask sequence data is crucial for protecting the privacy of node-local model parameter data. By generating a random sequence that matches the dimension of the model parameters, with each element generated independently, the randomness and unpredictability of the mask are ensured. For example, a Bernoulli distribution can be used to generate a mask of 0 or 1, allowing for the random discarding or retention of parameters; or a Gaussian or Laplace distribution can be used to generate continuous random numbers, achieving random perturbation of the parameters.

[0079] In this embodiment, the random mask sequence data is multiplied element-wise with the node's local model parameter data to obtain the masked parameter data. This step is the core operation of applying the random mask. Through element-wise multiplication, the values ​​in the random mask sequence directly affect the local model parameters, achieving random discarding or perturbation of the parameters.

[0080] To address the aforementioned balance issue, this application further dynamically adjusts the generation probability of the random mask sequence data based on the feature importance data of the node's local model parameter data, resulting in adjusted mask sequence data. Feature importance data reflects the contribution of model parameters to model performance. Methods for obtaining feature importance data include, but are not limited to: analyzing the gradient magnitude of parameters during local training, where a larger absolute gradient value indicates greater importance; or using sensitivity analysis to assess importance by changing individual parameters and observing changes in model performance; or utilizing model interpretability tools such as LIME and SHAP to analyze the model's dependence on input features. Based on this importance data, a mapping function can be designed to reduce the masking probability of highly important parameters, for example, by increasing the probability of generating 1 in the Bernoulli distribution; while increasing the masking probability of low-importance parameters. Finally, the adjusted mask sequence data is applied again to the node's local model parameter data to generate the privacy-preserving model parameter data. This step ensures that the final privacy-preserving model parameters are the result of a dynamically adjusted masking strategy, thereby achieving more refined privacy protection.

[0081] In this embodiment, through the above technical solution, this application can dynamically adjust the generation probability of random mask sequence data based on the feature importance of the node's local model parameter data, thereby achieving differentiated privacy protection. Specifically, for important parameters that contribute significantly to model performance, the risk of over-masking can be reduced, effectively preserving key information of the local model and avoiding a decline in global model performance due to privacy protection. Simultaneously, for less important parameters that contribute little to model performance, the probability of masking can be increased, enhancing privacy protection without affecting model accuracy and reducing potential privacy leakage risks. This adaptive masking strategy achieves a better balance between data privacy protection and model training accuracy, improving the practical application value and robustness of federated learning in server hardware predictive maintenance scenarios, and overcoming the performance loss or privacy deficiencies that may result from traditional fixed random masking strategies.

[0082] In the embodiments of this application, the server hardware predictive maintenance method based on federated learning and edge computing, by collecting and processing sensor data at the edge, combining federated learning to train and dynamically update the model, can dynamically adapt to changes in server load, reduce data transmission volume, protect data privacy, and improve the real-time performance and accuracy of predictive maintenance, thereby reducing the risk of false alarms and missed alarms.

[0083] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the predictive maintenance method for server hardware based on federated learning and edge computing in this application. Any simple modifications based on this technical concept are within the scope of protection of this application.

[0084] This application also provides a server hardware predictive maintenance system based on federated learning and edge computing, referenced... Figure 2 The server hardware predictive maintenance system based on federated learning and edge computing includes: a memory 10, a processor 20, and a server hardware predictive maintenance program based on federated learning and edge computing stored on the memory 10 and executable on the processor 20. The server hardware predictive maintenance program based on federated learning and edge computing is configured to implement the steps of the server hardware predictive maintenance method based on federated learning and edge computing.

[0085] The server hardware predictive maintenance system based on federated learning and edge computing provided in this application adopts the server hardware predictive maintenance method based on federated learning and edge computing in the above embodiments, which can improve the real-time performance and accuracy of predictive maintenance. Compared with the prior art, the beneficial effects of the server hardware predictive maintenance system based on federated learning and edge computing provided in this application are the same as those of the server hardware predictive maintenance method based on federated learning and edge computing provided in the above embodiments, and other technical features of the server hardware predictive maintenance system based on federated learning and edge computing are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0086] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0087] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. All equivalent structural transformations made under the technical concept of this application using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included within the scope of patent protection of this application.

Claims

1. A predictive maintenance method for server hardware based on federated learning and edge computing, characterized in that, The method includes: At the edge of multiple server nodes, raw sensor data of the server nodes is collected and preprocessed to obtain node preprocessed data. Based on the node preprocessed data, fault risk assessment is performed using a pre-trained prediction model local at the edge, generating node fault prediction data. Based on the node fault prediction data and the current server node load status data, determine whether the local triggering conditions are met. When the local triggering conditions are met, send the node preprocessed data to the cloud server. On the cloud server side, the node preprocessing data from multiple server nodes is received, and the prediction model is trained by federated learning based on all the received node preprocessing data to generate cloud global model parameters. The cloud-based global model parameters are distributed to the edge side of each server node to update the local pre-trained prediction model on the edge side. Predictive maintenance operations are performed on the server hardware based on the node failure prediction data.

2. The server hardware predictive maintenance method based on federated learning and edge computing as described in claim 1, characterized in that, The step of generating node fault prediction data by performing fault risk assessment using a pre-trained prediction model local at the edge based on the node preprocessed data includes: Multi-dimensional temporal features are extracted from the preprocessed data of the nodes to generate temporal feature vector data; The time-series feature vector data is input into a pre-trained prediction model, which includes a pre-trained time-series neural network. The time-series feature vector data is fused and analyzed using the pre-trained time-series neural network to obtain fused feature data; Based on the fused feature data, fault probability values ​​and fault type data are generated, and the fault probability values ​​and fault type data are combined into the node fault prediction data.

3. The server hardware predictive maintenance method based on federated learning and edge computing as described in claim 2, characterized in that, The steps for extracting multi-dimensional temporal features from the preprocessed data of the nodes and generating temporal feature vector data include: Temperature change rate characteristics are calculated from the preprocessed data of the nodes to obtain temperature time series data; Memory error statistical characteristics are calculated from the node preprocessing data to obtain memory error timing data; The hard disk comprehensive health characteristics are calculated from the node preprocessing data to obtain hard disk health time-series data; The temperature time-series data, the memory error time-series data, and the hard disk health time-series data are weighted and combined according to a preset weight ratio to generate the time-series feature vector data.

4. The server hardware predictive maintenance method based on federated learning and edge computing as described in claim 3, characterized in that, The steps for calculating the overall hard disk health characteristics from the node preprocessing data to obtain hard disk health time-series data include: Obtain hard disk remapping sector statistics from the node preprocessing data, and calculate remapping health value data based on the preset first weighting coefficient; Obtain the unprocessed sector statistics of the hard disk from the node preprocessing data, and calculate the unprocessed health value data based on the preset second weighting coefficient; Obtain the cyclic redundancy check error statistics of the hard disk from the node preprocessing data, and calculate the check health value data based on the preset third weighting coefficient; The remapped health value data, the health value data to be processed, and the verification health value data are weighted and summed according to the first weight coefficient, the second weight coefficient, and the third weight coefficient to obtain the hard disk health time series data; wherein, the ratio of the first weight coefficient, the second weight coefficient, and the third weight coefficient is a preset fixed ratio.

5. The server hardware predictive maintenance method based on federated learning and edge computing as described in claim 2, characterized in that, Based on the node preprocessed data, the step of generating node fault prediction data by performing fault risk assessment through a pre-trained prediction model locally on the edge side further includes: Based on the node failure prediction data and historical prediction accuracy data, calculate the prediction confidence assessment value data; The predicted confidence assessment value is compared with a preset confidence threshold value. When the predicted confidence assessment value is lower than the confidence threshold value, a multi-model collaborative prediction mechanism is activated. The node fault prediction data is corrected by the multi-model collaborative prediction mechanism to obtain corrected node fault prediction data. The steps of performing predictive maintenance operations on server hardware based on the node fault prediction data include: performing predictive maintenance operations on server hardware based on the corrected node fault prediction data.

6. The predictive maintenance method for server hardware based on federated learning and edge computing as described in claim 1, characterized in that, The step of determining whether the local triggering conditions are met based on the node fault prediction data and the current server node load status data, and sending the node preprocessed data to the cloud server when the local triggering conditions are met, includes: Analyze the historical trend of fault probability in the node fault prediction data to obtain trend analysis data; Based on the load status data, calculate the dynamic trigger threshold data; Compare the current fault probability data in the node fault prediction data with the dynamic trigger threshold data; When the current fault probability data is greater than the dynamic trigger threshold data, it is determined that the local trigger condition is met; When the local triggering conditions are met, the preprocessed data of the node is sent to the cloud server.

7. The server hardware predictive maintenance method based on federated learning and edge computing as described in claim 6, characterized in that, The steps for calculating the dynamic trigger threshold data based on the load status data include: Obtain the current load ratio data from the load status data; Multiply the current load ratio data by a preset first adjustment coefficient data to obtain the load adjustment factor data; The upward trend strength data of the failure probability is analyzed from the trend analysis data, and the upward trend strength data is multiplied by a preset second adjustment coefficient data to obtain the trend adjustment factor data. The preset base threshold data is linearly combined with the load adjustment factor data and the trend adjustment factor data to obtain the dynamic trigger threshold data; wherein, the dynamic trigger threshold data increases with the increase of the current load ratio data and increases with the increase of the upward trend strength data.

8. The predictive maintenance method for server hardware based on federated learning and edge computing as described in claim 1, characterized in that, On the cloud server side, the steps of receiving node preprocessed data from multiple server nodes, performing federated learning training on the prediction model based on all received node preprocessed data, and generating cloud-based global model parameters include: Receive preprocessed data from each server node to generate training dataset data for multiple nodes; Local model training is performed on the training dataset data of each node to generate node local model parameter data; The node-local model parameter data of each node is randomly masked to obtain privacy-preserving model parameter data; The privacy protection model parameter data of all nodes are weighted and fused to generate the cloud-based global model parameters.

9. The predictive maintenance method for server hardware based on federated learning and edge computing as described in claim 8, characterized in that, The steps for performing random masking on the node-local model parameter data of each node to obtain privacy-preserving model parameter data include: Generate random mask sequence data, wherein each mask element in the random mask sequence data is generated independently according to a preset probability distribution; The random mask sequence data is multiplied element-wise with the node local model parameter data to obtain the parameter data after masking. Based on the feature importance data of the node local model parameter data, the generation probability of the random mask sequence data is dynamically adjusted to obtain the adjusted mask sequence data. The adjusted mask sequence data is then applied again to the node's local model parameter data to generate the privacy-preserving model parameter data.

10. A predictive maintenance system for server hardware based on federated learning and edge computing, characterized in that, The server hardware predictive maintenance system based on federated learning and edge computing includes: a memory, a processor, and a server hardware predictive maintenance program based on federated learning and edge computing stored on the memory and executable on the processor. The server hardware predictive maintenance program based on federated learning and edge computing is configured to implement the steps of the server hardware predictive maintenance method based on federated learning and edge computing as described in any one of claims 1 to 9.

Images