Hard disk fault prediction method based on multivariate data

Abstract

The invention discloses a hard disk failure prediction method based on multivariate data, which comprises the following steps of: constructing a multi-level cache structure based on a high-frequency access data sequence, and obtaining a cache priority score by combining access temperature change and read-write error counting; migrating the historical parameter sequence to a solid state disk cache layer based on the cache priority score to form optimized data distribution configuration; encrypting sensitive information in the optimized data distribution configuration, setting a hierarchical permission, and recording an audit log to generate an access behavior tracking sequence; aggregating multi-data center gradient updating by adopting a distributed learning method based on the access behavior tracking sequence to obtain global fault prediction parameters; adjusting a cache dynamic rule based on the global fault prediction parameter to determine an optimized feature data acquisition path; triggering an encryption cycle based on the feature data acquisition path to obtain updated sensitive protection configuration; and fusing the updated sensitive protection configuration into a distributed learning aggregation process to output a hard disk fault prediction result.

Description

Technical Field

[0001] This invention belongs to the field of hard disk fault prediction technology, and in particular relates to a hard disk fault prediction method based on multivariate data. Background Technology

[0002] Hard disk failure prediction, as a key technology for ensuring data security and system reliability, has been widely integrated into the daily maintenance of various large-scale storage systems. With the rapid increase in data volume, especially in complex environments such as cloud platforms and data centers, the demand for high real-time performance and high-precision prediction from storage systems is becoming increasingly urgent. Hard disk failure prediction technology has gradually evolved from a traditional model relying on static historical data to an intelligent analysis model based on dynamic, multi-dimensional data (such as real-time monitoring indicators, access patterns, and hardware logs). However, existing prediction solutions often focus on a single data source or a single technical means. Although they can achieve certain results in some scenarios, they often suffer from common problems such as low data access efficiency, unreasonable resource allocation, and insufficient sensitive protection when dealing with multi-source, heterogeneous, and massive amounts of data. These issues make it difficult to meet the comprehensive requirements of modern distributed storage systems in terms of performance, security, and efficiency.

[0003] Current hard drive failure prediction methods face two major technical bottlenecks when handling complex and diverse data: First, the mismatch between data access efficiency and storage resource allocation. Existing systems lack dynamic importance assessment and hierarchical management mechanisms for multi-source data, resulting in high-value, frequently accessed data failing to be quickly scheduled to high-speed storage layers, while low-value data occupies valuable memory or high-speed storage resources. This resource mismatch not only reduces the real-time response capability of the prediction model but also causes data access latency under high-load scenarios, thus missing the optimal window for failure warning. Second, the disconnect between data sensitivity protection mechanisms and the prediction process. Most existing solutions do not adequately protect sensitive information (such as user behavior logs and system configuration parameters) during data collection, transmission, and aggregation. Especially during cross-data center collaborative training, there is a lack of mechanisms to organically integrate encryption and desensitization, access control, and distributed learning, making it risky for data leakage during the prediction process and failing to meet increasingly stringent data security compliance requirements. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a hard disk fault prediction method based on multivariate data, thereby resolving the issues present in the prior art.

[0005] To achieve the above objectives, the present invention provides a hard disk fault prediction method based on multivariate data, comprising:

[0006] Cluster analysis is performed on the collected records obtained from the storage device to obtain high-frequency access data sequences;

[0007] A multi-level cache structure is constructed based on the high-frequency access data sequence, and the cache priority score is determined according to the data access temperature change and read / write error count.

[0008] Based on the cache priority score, the historical parameter sequence is migrated to the solid-state drive cache layer to obtain an optimized data distribution configuration;

[0009] The sensitive information in the optimized data distribution configuration is encrypted to obtain a de-identified multi-dimensional data set;

[0010] Hierarchical access permissions are set for the de-identified multi-source data set, and audit logs are recorded to obtain access behavior tracking sequences;

[0011] Based on the access behavior tracking sequence, a distributed learning method is used to aggregate gradient updates to obtain global fault prediction parameters;

[0012] Based on the global fault prediction parameters, adjust the cache dynamic rules to determine the optimized feature data acquisition path;

[0013] Based on the feature data acquisition path, an encryption processing loop is triggered to update the sensitive protection configuration;

[0014] The updated sensitive protection configuration is integrated into the distributed learning aggregation process to obtain hard disk failure prediction results.

[0015] Optionally, the process of obtaining high-frequency access data sequences based on the collected records includes:

[0016] The data distribution characteristics were obtained by grouping the collected records using the K-means clustering method.

[0017] Based on the data distribution characteristics, an access pattern clustering method is used to construct an overhead model to obtain the computational overhead distribution;

[0018] High-frequency sequences are generated based on the aforementioned computational overhead distribution using a sequence extraction method.

[0019] A fault warning sequence is obtained by associating system attributes based on the high-frequency sequence;

[0020] The high-frequency access data sequence is determined based on the fault warning sequence.

[0021] Optionally, the process of constructing a multi-level cache structure based on the high-frequency access data sequence and determining cache priority scores includes:

[0022] A multi-level cache structure is constructed by acquiring memory layer data based on the high-frequency access data sequence.

[0023] Based on the multi-level cache structure, dynamic adjustment rules are determined, and the temperature change is allocated to the nearest time window to obtain the temperature change allocation result.

[0024] Based on the temperature change allocation result, the read / write error count is obtained, and the error counts exceeding the counting threshold are allocated to the nearest time window to obtain the error count allocation result.

[0025] Based on the error count allocation results, the access patterns in the access logs are integrated to determine a preliminary priority score;

[0026] The cache priority score is obtained by adjusting the score weights based on the initial priority score and the preset error recovery mechanism.

[0027] Optionally, the process of encrypting sensitive information in the optimized data distribution configuration to obtain a de-identified multi-dimensional data set includes:

[0028] Sensitive data list is obtained by identifying sensitive information from the optimized data distribution configuration based on preset sensitivity quantification analysis;

[0029] The sensitive data list is encrypted using the AES encryption algorithm and a dynamic key generation method to obtain the encrypted output;

[0030] Based on the encrypted output, a mask replacement method is used for desensitization, and access control is set based on role allocation to obtain a preliminary desensitized set;

[0031] Based on the initial de-identified dataset, a final de-identified multivariate dataset is obtained using compliance simulation verification and data flow tracing methods.

[0032] Optionally, the process of aggregating gradient updates to obtain global fault prediction parameters based on the access behavior tracking sequence using a distributed learning method includes:

[0033] Based on the access behavior tracking sequence, collaborative training requirements are extracted;

[0034] Based on the aforementioned collaborative training requirements, local gradient updates are transmitted and aggregated across multiple data centers to obtain preliminary global model parameters.

[0035] Based on the preliminary global model parameters, a convolutional neural network is used to process the access behavior tracking sequence to obtain updated fault prediction parameters;

[0036] Based on the updated fault prediction parameters, obtain collaborative information on training requirements;

[0037] A long short-term memory network is used to fuse the collaborative information of the training requirements to obtain global parameter adjustment values;

[0038] Based on the global parameter adjustment values, a data center aggregation method is adopted, and the aggregation stability is judged by calculating the gradient deviation value to obtain an enhanced fault prediction parameter set;

[0039] Based on the enhanced fault prediction parameter set, abnormal event prediction details are extracted to obtain global fault prediction parameters.

[0040] Optionally, the process of adjusting the cache dynamic rules based on the global fault prediction parameters to determine the optimized feature data acquisition path includes:

[0041] A parameter adjustment mechanism is obtained based on the global fault prediction parameters;

[0042] The dynamic adjustment rules in the multi-level cache structure are updated based on the parameter adjustment mechanism, and the cache data is monitored based on the adjusted rules to obtain the cache data consistency verification results.

[0043] Based on the cached data consistency verification results, a real-time optimization process is performed, and the fault parameters are analyzed to obtain the optimized calculation basis.

[0044] Based on the optimized calculation basis, the feature data extraction path is determined, and the extracted values ​​that conform to the optimized calculation basis are used to generate extracted data.

[0045] Based on the extracted data, the multi-level caching structure is used to determine the optimized feature data acquisition path.

[0046] Optionally, the process of triggering an encryption processing loop based on the feature data acquisition path to obtain updated sensitive protection configuration includes:

[0047] Obtain change indicators based on cloud scenario load change monitoring;

[0048] The adjusted path is obtained by adjusting the feature data acquisition path based on the change index.

[0049] Based on the adjusted path, a controlled encryption processing loop is triggered, and an iterative encryption method is used to process sensitive data to obtain the encryption result.

[0050] Based on the encryption result, the updated sensitive data is obtained by verifying and dynamically updating the configuration.

[0051] The updated sensitive protection configuration is obtained by adjusting the protection configuration based on the updated sensitive data.

[0052] Optionally, the process of integrating the update-sensitive protection configuration into the distributed learning aggregation process to obtain the hard disk failure prediction result includes:

[0053] Extract encryption parameters from the updated sensitive protection configuration, and construct a sensitive fusion and aggregation framework based on the encryption parameters;

[0054] The federated learning algorithm is used to process data sharing between nodes within the sensitive fusion aggregation framework to obtain distributed training parameters.

[0055] Based on the distributed training parameters, hard disk vibration and temperature-sensitive data features are extracted;

[0056] Based on the aforementioned sensitive data features, an anomaly pattern matching method is used to identify fault modes and obtain preliminary fault indicators.

[0057] An enhanced prediction model is obtained by integrating the preliminary fault indicators with the real-time calculation process of sensitive audit business between nodes.

[0058] The enhanced prediction model integrates data stream monitoring attributes to calculate hard disk failure prediction results.

[0059] Compared with the prior art, the present invention has the following advantages and technical effects:

[0060] This invention achieves simultaneous improvement in data access efficiency and security during hard disk fault prediction by constructing a dynamic caching optimization and sensitive protection collaborative mechanism for multi-source data. On one hand, the system dynamically adjusts multi-level caching strategies based on clustering analysis and access popularity, prioritizing high-frequency critical data to high-speed storage layers, significantly shortening the latency of feature extraction and model inference, and ensuring real-time prediction. On the other hand, by embedding encryption desensitization and hierarchical access control in each stage of data preprocessing, distributed aggregation, and model update, sensitive information is reliably protected throughout the process, while supporting compliant collaborative training across data centers. This solution effectively overcomes the problems of resource mismatch and weak sensitivity in existing technologies, comprehensively enhancing the system's data security and compliance capabilities while improving fault prediction accuracy and response speed. Attached Figure Description

[0061] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0062] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0063] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0064] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0065] Example 1

[0066] like Figure 1 As shown, this embodiment provides a hard disk fault prediction method based on multivariate data, including the following steps:

[0067] Step S101: Obtain massive collection records from the storage device in the hard disk failure prediction system, and use cluster analysis to analyze the computational overhead distribution of the collection records to obtain high-frequency access data sequences.

[0068] Further, step S101 includes: using the K-means clustering method to group the collected records to obtain data distribution characteristics; using the access pattern clustering method based on the data distribution characteristics to construct an overhead model to obtain the computational overhead distribution; using the sequence extraction method based on the computational overhead distribution to generate high-frequency sequences; using the high-frequency sequences to perform system attribute association to obtain fault warning sequences; and using the fault warning sequences to determine high-frequency access data sequences.

[0069] Furthermore, as a specific implementation of this embodiment, massive amounts of data are collected from the storage devices in the hard disk failure prediction system. These data include the number of read / write operations, temperature changes, and error rates of the hard disk. The K-means clustering method is used to initially group the data, and similar records are clustered together based on Euclidean distance calculation to obtain data distribution characteristics reflecting differences in hard disk data patterns. Based on the data distribution characteristics, an access pattern clustering method is used to construct an overhead model. By analyzing the access frequency and latency within each cluster, the average access time and peak load of each cluster are statistically analyzed to generate a computational overhead distribution displaying high-overhead areas. The Apriori algorithm is used to extract frequent itemsets from the computational overhead distribution, calculate the frequency of read / write operation sequences, and mark sequences exceeding a daily threshold of 1000 as high-frequency access sequences. System attribute association is performed on the high-frequency access sequences, associating the sequences with hard disk age and firmware version to generate high-risk warning sequences. The high-frequency access data sequences in hard disk failure prediction are determined through fault warning sequence optimization. The random forest algorithm is used to train the warning sequences, outputting potential fault sequences to build the prediction model.

[0070] Step S102: A multi-level cache structure is constructed in the memory layer using high-frequency access data sequences. Dynamic adjustment rules are determined to allocate data access temperature changes and read / write error counts to the most recent time window to obtain a cache priority score.

[0071] Further, step S102 includes: acquiring memory layer data based on high-frequency access data sequences to construct a multi-level cache structure; determining dynamic adjustment rules based on the multi-level cache structure, allocating access temperature changes to the most recent time window to obtain temperature change allocation results; acquiring read / write error counts based on the temperature change allocation results, allocating error counts exceeding the counting threshold to the most recent time window to obtain error count allocation results; integrating access patterns in the access log based on the error count allocation results to determine a preliminary priority score; and adjusting the score weights based on the preliminary priority score and a preset error recovery mechanism to obtain a cache priority score.

[0072] Furthermore, as a specific implementation of this embodiment, frequently accessed data blocks are extracted from the access logs of the storage device using a high-frequency access sequence, and temperature sensor readings and error check codes are loaded into the memory layer. A multi-level cache structure including L1, L2, and L3 cache levels is constructed. Real-time vibration data is placed in the L1 cache using a hash mapping algorithm, and historical error logs are distributed to the L3 cache, determining the basic hierarchical distribution. A multi-level cache structure is used to determine dynamic adjustment rules, monitoring the changes in the number of accesses to data blocks within a 5-minute time window. When the number of accesses increases from 10 to 50, it is considered a temperature increase, and the data block is promoted from the L3 cache to the L2 cache, obtaining the temperature change allocation result. The error register value is read from the underlying hardware interface based on the temperature change allocation result. When the read / write error count reaches 25, its logs are integrated into a 1-hour window, generating an error count allocation result that includes bit flips and check failure classifications. Based on the error count allocation result, the timestamps and operation types in the access log file are parsed, the frequency of write-dominated patterns is calculated, and an initial priority score of 80 points is assigned. The RAID redundancy check recovery mechanism is obtained from the initial priority score. When the mechanism is activated, the weight of successful error recovery is increased from 0.3 to 0.5, resulting in the final cache priority score.

[0073] In step S103, if the cache priority score exceeds the preset threshold, the medium-priority historical parameter sequence is migrated to the solid-state drive cache layer to obtain the optimized data distribution configuration.

[0074] Furthermore, as a specific implementation of this embodiment, the current priority score is extracted from the historical parameter sequence through an access frequency monitoring mechanism. The historical parameter sequence contains data access frequency and timestamp metadata over a past period. A weighted summation of historical records is calculated using an access frequency formula, with the access frequency multiplied by a decay factor to ensure higher weight for recent accesses, resulting in a numerical priority score index. When the priority score reaches 85 points, the migration process is initiated, copying the moderately popular comment data sequence (scores between 50 and 79) from memory to the solid-state drive (SSD) layer. The migration process updates the index table to ensure seamless switching of access paths. After migration, the memory layer retains high-priority data, while the SSD layer adds a medium-priority sequence. The data layout of different levels is analyzed based on the post-migration sequence distribution. When uneven data distribution is observed in the SSD layer, data blocks are repartitioned, and user data related to geographical location is concentrated in sub-levels. By adjusting the scheme and integrating the historical parameter sequence with the new distribution scheme, high-frequency transaction data is fixed in memory, low-frequency historical data is moved to the hard drive, and a mapping table configuration specifying the location and access rules for each data block is output.

[0075] Step S104: The sensitive information is encrypted using the optimized data distribution configuration to obtain a desensitized multi-dimensional data set.

[0076] Furthermore, the S104 process includes: identifying sensitive information from the optimized data distribution configuration based on preset sensitivity quantification analysis to obtain a sensitive data list; encrypting the sensitive data list using the AES encryption algorithm and dynamic key generation method to obtain an encrypted output; performing desensitization processing using a mask replacement method based on the encrypted output, and setting access control based on role allocation to obtain a preliminary desensitized set; and obtaining a final desensitized multivariate data set using compliance simulation verification and data flow tracing methods based on the preliminary desensitized set.

[0077] Furthermore, as a specific implementation of this embodiment, sensitive information is extracted from the data distribution configuration through a preset sensitivity quantification analysis. Statistical calculations and correlation analyses are performed on the data fields, and frequently occurring name fields are marked as highly sensitive tags, forming a sensitive data list containing patient IDs and medical history descriptions. The sensitive data list is grouped and encrypted using the AES encryption algorithm, converting patient names into unreadable ciphertext blocks. Simultaneously, a 256-bit dynamic encryption key is generated based on timestamps and device IDs. When the security level assessment score exceeds 0.8, information anonymization is implemented, replacing specific ID numbers with randomly generated pseudo-IDs, resulting in an encrypted output containing encrypted data blocks and anonymized fields. Sensitive fields are masked by asterisks to cover specific disease names in diagnostic records. Access permissions are assigned based on roles, allowing doctors to view only the anonymized medical history, while administrators have decryption permissions. Multivariate set integration parameters are determined, and a preliminary anonymized set is calculated by merging encrypted data blocks from multiple patients. A compliance simulation is used to verify GDPR compliance by simulating nurses accessing patient records, while data flow tracking records the transmission logs from the server to the client, resulting in the final anonymized multivariate data set.

[0078] Step S105: Obtain the de-identified multi-dimensional data set, set hierarchical access permissions under the controlled framework, and record the audit log if the access request matches the permissions to obtain the access behavior tracking sequence.

[0079] Furthermore, as a specific implementation of this embodiment, a de-identified data set is obtained by integrating data from multiple sources, including the hospital's electronic medical record system, laboratory testing system, and patient self-reported health data. Direct identifying information such as patient names and ID numbers is removed, while medical indicator data such as blood pressure and blood sugar levels are retained. Under a controlled framework, hierarchical access permissions are set for the roles of doctors, nurses, administrators, and researchers. The access permission management module dynamically allocates access permission ranges based on role responsibilities and data sensitivity. Access requests initiated by doctors when logging into the medical platform to view patient medical records are acquired. The doctor's identity information is verified through an identity authentication and permission verification system, and the match between role permissions and the requested data range is checked. When permissions match successfully, operation details such as access time, accessed object, and operation type are recorded, forming a preliminary audit log completed by an independent audit module. All access records of doctors within a day are analyzed from the audit log. Time series correlation is used to determine whether the operation follows the conventional path of first viewing basic patient information and then accessing medical records. Behavioral path matching is used to compare actual operations with preset behavioral templates. When it is found that doctors frequently access data of patients they are not responsible for, an anomaly detection is performed through a behavioral scoring mechanism. The system performs threshold comparisons for abnormal behavior. When an operation score exceeds a threshold, an alert is generated and the administrator is notified. Based on historical behavior data and preset rules, the system monitors and analyzes in real time using an anomaly detection engine. It integrates information from multiple sources based on access behavior tracing sequences, and uses a data analysis engine to correlate the time, data source, and reason for the violation of abnormal operations to generate a complete tracking report.

[0080] Step S106: Extract collaborative training requirements from the access behavior tracking sequence, and use a distributed learning method to aggregate gradient updates across multiple data centers to obtain global fault prediction parameters.

[0081] Further, step S106 includes: extracting collaborative training requirements based on the access behavior tracking sequence; transmitting and aggregating local gradient updates across multiple data centers based on the collaborative training requirements to obtain preliminary global model parameters; processing the access behavior tracking sequence using a convolutional neural network based on the preliminary global model parameters to obtain updated fault prediction parameters; obtaining collaborative information on training requirements based on the updated fault prediction parameters; fusing the collaborative information on training requirements using a long short-term memory network to obtain global parameter adjustment values; using data center aggregation based on the global parameter adjustment values, determining aggregation stability by calculating gradient deviation values ​​to obtain an enhanced fault prediction parameter set; and extracting abnormal event prediction details based on the enhanced fault prediction parameter set to obtain global fault prediction parameters.

[0082] Furthermore, as a specific implementation of this embodiment, a cross-regional data analysis platform collects user login, query, and transaction operation records at different time periods to form complete access sequence tracking data. Based on the access sequence tracking analysis of branch offices' insufficient data or model training needs in risk identification, collaborative training requirements for sharing access behavior data are extracted. Local models are trained in the data centers of each branch office, generating local gradient update values ​​reflecting abnormal login attempt patterns based on user access behavior data. These local gradient update values ​​are sent to a central server in encrypted form for aggregation, combining data from all branches to form preliminary global model parameters. Access sequence tracking features, including login time intervals and operation frequencies, are input into a convolutional neural network. Behavioral pattern features are extracted through multi-layer convolution operations, outputting abnormal behavior pattern identification results. When the abnormal behavior pattern identification results exceed a preset threshold, multi-center collaborative data fusion is triggered to analyze whether a user's repeated attempts at high-value transactions within a short period constitute fraudulent behavior, determining updated fault prediction parameters. Risk collaboration information from multiple branch offices is input into a long short-term memory network to analyze long-term and short-term behavioral trends. Combining information on abnormal account logins and similar IP address attack attempts, global parameter adjustment values ​​under a real-time anomaly response mechanism are output. The stability of the global model parameters is evaluated by calculating the gradient deviation values ​​of each data center. When the data updates of a certain branch office deviate significantly from those of other centers, data acquisition anomalies are analyzed and the aggregation weights are adjusted to obtain an enhanced set of fault prediction parameters. Based on the enhanced parameter set, the prediction of abnormal events is refined, and more accurate prediction parameters are generated for high-frequency abnormal login behaviors by combining historical data and real-time monitoring.

[0083] Step S107: Adjust the dynamic adjustment rules in the multi-level cache according to the global fault prediction parameters to determine the real-time optimized feature data acquisition path.

[0084] Further, step S107 includes: obtaining a parameter adjustment mechanism based on global fault prediction parameters; updating the dynamic adjustment rules in the multi-level cache structure based on the parameter adjustment mechanism, and monitoring cache data based on the adjusted rules to obtain cache data consistency verification results; performing a real-time optimization process based on the cache data consistency verification results, analyzing fault parameters to obtain the optimized calculation basis; determining the feature data extraction path based on the optimized calculation basis, and generating extracted data for extracted values ​​that conform to the optimized calculation basis; and determining the optimized feature data acquisition path based on the extracted data using a multi-level cache structure.

[0085] Furthermore, as a specific implementation of this embodiment, fault prediction parameters aggregated from multiple centers are collected from global fault prediction, including access sequence anomaly probability and response latency indicators. A parameter adjustment mechanism is used to analyze parameter deviation values ​​and generate adjustment instructions. The parameter adjustment mechanism updates the dynamic adjustment rules in the multi-layer cache structure, mapping the prediction parameters to rule updates. When the prediction parameters indicate high-frequency access anomalies, the adjustment rule is to increase the prefetch buffer size of the L2 cache. A monitoring agent is deployed in the cache hierarchy management for the adjusted rules, collecting cache hit rate and read / write latency indicators every second. When the hit rate is below 80%, a threshold exceedance judgment is triggered. An adaptive algorithm based on machine learning is integrated to combine the new rules with existing data. The cache data consistency verification result is determined by comparing the data hash values ​​between multi-layer caches. When L1 cache data is inconsistent with L3 cache data, it is marked as inconsistent. Based on the cache data consistency verification result, an optimization engine is launched to analyze fault parameters such as network jitter rate and node load. A path graph algorithm is used to evaluate the latency path from the source node to the target node to obtain the fault parameter analysis required for optimizing the path calculation. The optimized path weights and parameter thresholds are used as the basis for post-optimization calculations. This calculation basis is then integrated with the extracted paths to determine whether vehicle flow characteristics meet the basic thresholds within the intelligent transportation system. When the characteristics meet the basic thresholds, an abnormal flow sequence is generated as the extracted data. Data is then retrieved from the extracted data using a multi-layered caching structure, and the shortest path to directly obtain feature data from the L1 cache is selected based on optimization rules.

[0086] Step S108: If the real-time optimized feature data acquisition path detects a change in cloud scenario load, the controlled framework's encryption processing loop is triggered to obtain the updated sensitive data protection configuration.

[0087] Further, step S108 includes: obtaining change indicators based on cloud scenario load change monitoring; adjusting the feature data acquisition path based on the change indicators to obtain the adjusted path; triggering an encryption processing loop under the controlled framework based on the adjusted path, and using an iterative encryption method to process sensitive data to obtain encryption results; verifying and dynamically updating the configuration based on the encryption results to obtain updated sensitive data; and adjusting the protection configuration based on the updated sensitive data to obtain updated sensitive protection configuration.

[0088] Furthermore, as a specific implementation of this embodiment, real-time optimization features are obtained from the cloud scenario load. The average response time and data throughput during peak load periods are extracted by monitoring server CPU utilization, memory usage, and network traffic metrics. Load changes are monitored through data acquisition paths, and log analysis tools are used to track changes in user access patterns in real time, detecting increases in response latency caused by sudden traffic surges, and calculating the percentage increase in latency as a change indicator value. When the change indicator exceeds a 20% threshold, path adjustment and optimization are performed, analyzing bottlenecks in the current path, reallocating requests to backup paths, and determining the adjusted data routing scheme. An encryption loop is triggered based on the adjusted path. A modular software architecture, including a key management module and a data segmentation module, is used in the controlled framework processing to divide sensitive data into multiple 128-bit data blocks. In the encryption loop, a temporary key is generated for each data block, and the AES encryption algorithm is applied. Plaintext is converted to ciphertext through substitution, shifting, and obfuscation operations. A complete encryption result is obtained after 3 to 5 iterative loops. The integrity of the encrypted data is verified using a hash function based on the encryption result. When verification is successful, the configuration is dynamically updated to adjust the key rotation frequency, and the updated sensitive data is obtained. Based on the analysis of encryption results for potential vulnerabilities, the protection configuration is adjusted to enhance the access control list and update the firewall rules, thereby achieving an overall optimized configuration update for sensitive protection.

[0089] Step S109: The updated sensitive data protection configuration is integrated into the aggregation process of the distributed learning method to obtain the final real-time calculation result of hard disk failure prediction.

[0090] Further, step S109 includes: extracting encryption parameters from the updated sensitive protection configuration; constructing a sensitive fusion aggregation framework based on the encryption parameters; using a federated learning algorithm to process data sharing between nodes under the sensitive fusion aggregation framework to obtain distributed training parameters; extracting hard disk vibration and temperature sensitive data features based on the distributed training parameters; identifying fault modes and obtaining preliminary fault indicators based on the sensitive data features using anomaly pattern matching; optimizing the real-time calculation process of sensitive audit business between nodes based on the preliminary fault indicators to obtain an enhanced prediction model; and integrating data flow monitoring attributes based on the enhanced prediction model to calculate the hard disk fault prediction result.

[0091] Furthermore, as a specific implementation of this embodiment, the updated sensitive data protection configuration is obtained. Encryption parameter definitions are parsed from the JSON-formatted configuration file, and automatically adjusted to use a 256-bit key configuration when the data encryption level is high. The encryption key and obfuscation factor are extracted from the configuration, and a parameter fusion algorithm is used to embed the key into gradient calculation. The encrypted parameters are combined with the initial data aggregation function to construct a sensitive fusion aggregation framework. A federated learning algorithm is used for the sensitive fusion aggregation framework to handle data sharing between nodes on a distributed AI training platform. Each node trains the model locally and only shares the encrypted model updates. A global model is initialized in a data center hard drive monitoring system. Each node uploads the encrypted parameter updates. When node A detects an abnormal vibration, it calculates its local gradient, encrypts it, and shares it. Distributed training parameters with a learning rate of 0.01 and a batch size of 32 are determined. Hard drive vibration and temperature sensitive data features are obtained from the distributed training parameters. Vibration frequency peaks are extracted using Fourier transform, and average temperature features are monitored. Temperatures exceeding 50 degrees Celsius are considered abnormal. Fault mode recognition is performed through abnormal pattern matching. Vibration patterns are compared with a historical fault database. When the matching degree is higher than 80%, preliminary fault indicators for high-risk vibration faults are generated. By integrating preliminary fault indicators with sensitive inter-node auditing procedures, the auditing module verifies the encryption integrity of each update and incorporates it into the indicators to optimize the real-time calculation process, adjusting aggregation weights to prioritize high-trust nodes. In a cloud computing environment, a real-time computing engine processes the data stream, resulting in an enhanced predictive model with a 20% improvement in fault probability. Based on this enhanced predictive model, monitoring attributes from the data stream obtained from real-time computation are integrated, and a Bayesian network is used to calculate prediction scores. When vibration and temperature attributes are abnormal, a real-time result with a fault probability of 0.85 is output.

[0092] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A hard disk fault prediction method based on multivariate data, characterized in that, Includes the following steps: Cluster analysis is performed on the collected records obtained from the storage device to obtain high-frequency access data sequences; A multi-level cache structure is constructed based on the high-frequency access data sequence, and the cache priority score is determined according to the data access temperature change and read / write error count. Based on the cache priority score, the historical parameter sequence is migrated to the solid-state drive cache layer to obtain an optimized data distribution configuration; The sensitive information in the optimized data distribution configuration is encrypted to obtain a de-identified multi-dimensional data set; Hierarchical access permissions are set for the de-identified multi-source data set, and audit logs are recorded to obtain access behavior tracking sequences; Based on the access behavior tracking sequence, a distributed learning method is used to aggregate gradient updates to obtain global fault prediction parameters; Based on the global fault prediction parameters, adjust the cache dynamic rules to determine the optimized feature data acquisition path; Based on the feature data acquisition path, an encryption processing loop is triggered to update the sensitive protection configuration; The updated sensitive protection configuration is integrated into the distributed learning aggregation process to obtain hard disk failure prediction results.

2. The hard disk fault prediction method based on multi-source data according to claim 1, characterized in that, The process of obtaining high-frequency access data sequences based on collected records includes: The data distribution characteristics were obtained by grouping the collected records using the K-means clustering method. Based on the data distribution characteristics, an access pattern clustering method is used to construct an overhead model to obtain the computational overhead distribution; High-frequency sequences are generated based on the aforementioned computational overhead distribution using a sequence extraction method. A fault warning sequence is obtained by associating system attributes based on the high-frequency sequence; The high-frequency access data sequence is determined based on the fault warning sequence.

3. The hard disk fault prediction method based on multi-source data according to claim 1, characterized in that, The process of constructing a multi-level cache structure based on the high-frequency access data sequence and determining cache priority scoring includes: A multi-level cache structure is constructed by acquiring memory layer data based on the high-frequency access data sequence. Based on the multi-level cache structure, dynamic adjustment rules are determined, and the temperature change is allocated to the nearest time window to obtain the temperature change allocation result. Based on the temperature change allocation result, the read / write error count is obtained, and the error counts exceeding the counting threshold are allocated to the nearest time window to obtain the error count allocation result. Based on the error count allocation results, the access patterns in the access logs are integrated to determine a preliminary priority score; The cache priority score is obtained by adjusting the score weights based on the initial priority score and the preset error recovery mechanism.

4. The hard disk fault prediction method based on multi-source data according to claim 1, characterized in that, The process of encrypting sensitive information in the optimized data distribution configuration to obtain a de-identified multi-dimensional data set includes: Sensitive data list is obtained by identifying sensitive information from the optimized data distribution configuration based on preset sensitivity quantification analysis; The sensitive data list is encrypted using the AES encryption algorithm and a dynamic key generation method to obtain the encrypted output; Based on the encrypted output, a mask replacement method is used for desensitization, and access control is set based on role allocation to obtain a preliminary desensitized set; Based on the initial de-identified dataset, a final de-identified multivariate dataset is obtained using compliance simulation verification and data flow tracing methods.

5. The hard disk fault prediction method based on multivariate data according to claim 1, characterized in that, The process of obtaining global fault prediction parameters by aggregating gradient updates based on the access behavior tracking sequence using a distributed learning method includes: Based on the access behavior tracking sequence, collaborative training requirements are extracted; Based on the aforementioned collaborative training requirements, local gradient updates are transmitted and aggregated across multiple data centers to obtain preliminary global model parameters. Based on the preliminary global model parameters, a convolutional neural network is used to process the access behavior tracking sequence to obtain updated fault prediction parameters; Based on the updated fault prediction parameters, obtain collaborative information on training requirements; A long short-term memory network is used to fuse the collaborative information of the training requirements to obtain global parameter adjustment values; Based on the global parameter adjustment values, a data center aggregation method is adopted, and the aggregation stability is judged by calculating the gradient deviation value to obtain an enhanced fault prediction parameter set; Based on the enhanced fault prediction parameter set, abnormal event prediction details are extracted to obtain global fault prediction parameters.

6. The hard disk fault prediction method based on multi-source data according to claim 1, characterized in that, The process of adjusting the cache dynamic rules and determining the optimized feature data acquisition path based on the global fault prediction parameters includes: A parameter adjustment mechanism is obtained based on the global fault prediction parameters; The dynamic adjustment rules in the multi-level cache structure are updated based on the parameter adjustment mechanism, and the cache data is monitored based on the adjusted rules to obtain the cache data consistency verification results. Based on the cached data consistency verification results, a real-time optimization process is performed, and the fault parameters are analyzed to obtain the optimized calculation basis. Based on the optimized calculation basis, the feature data extraction path is determined, and the extracted values ​​that conform to the optimized calculation basis are used to generate extracted data. Based on the extracted data, the multi-level caching structure is used to determine the optimized feature data acquisition path.

7. The hard disk fault prediction method based on multivariate data according to claim 1, characterized in that, The process of updating the sensitive protection configuration by triggering an encryption processing loop based on the feature data acquisition path includes: Obtain change indicators based on cloud scenario load change monitoring; The adjusted path is obtained by adjusting the feature data acquisition path based on the change index. Based on the adjusted path, a controlled encryption processing loop is triggered, and an iterative encryption method is used to process sensitive data to obtain the encryption result. Based on the encryption result, the updated sensitive data is obtained by verifying and dynamically updating the configuration. The updated sensitive protection configuration is obtained by adjusting the protection configuration based on the updated sensitive data.

8. The hard disk fault prediction method based on multivariate data according to claim 1, characterized in that, The process of integrating the updated sensitive protection configuration into the distributed learning aggregation process to obtain hard disk fault prediction results includes: Extract encryption parameters from the updated sensitive protection configuration, and construct a sensitive fusion and aggregation framework based on the encryption parameters; The federated learning algorithm is used to process data sharing between nodes within the sensitive fusion aggregation framework to obtain distributed training parameters. Based on the distributed training parameters, hard disk vibration and temperature-sensitive data features are extracted; Based on the aforementioned sensitive data features, an anomaly pattern matching method is used to identify fault modes and obtain preliminary fault indicators. An enhanced prediction model is obtained by integrating the preliminary fault indicators with the real-time calculation process of sensitive audit business between nodes. The enhanced prediction model integrates data stream monitoring attributes to calculate hard disk failure prediction results.

Images