Hard disk fault trend prediction method based on Internet of Things storage node cluster

Abstract

The invention discloses a hard disk fault trend prediction method based on an Internet of Things storage node cluster, and relates to the technical field of machine learning models.The method comprises the steps that firstly, SMART health state time sequences of all hard disks in a hard disk cluster are collected in real time; secondly, performing differential operation, linear regression and statistical deviation analysis on original sequences such as remapping sector counting and a track seeking error rate, and deeply extracting a plurality of quantized dynamic characteristic indexes such as bad track diffusion risk degree and track seeking system decline slope; and combining with a neighborhood stress index to serve as an enhanced feature to be input into a machine learning prediction model, and generating a fault prediction result which comprehensively considers individual states and environmental influences. According to the method, prediction is improved from single-point monitoring to an individual-environment collaborative analysis dimension, correlation faults caused by vibration conduction, local overheating or batch defects can be effectively warned, and the accuracy and foresight of prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning models, in particular to a hard disk failure trend prediction method based on an Internet of Things storage node cluster. BACKGROUND

[0002] With the rapid development of cloud computing, big data, artificial intelligence and the Internet of Things (IoT), the total amount of global data is growing at an exponential rate. To meet this demand, large data centers and cloud service providers generally use high-density storage servers and JBOD expansion cabinets to integrate as many mechanical hard disks (HDDs) as possible in limited rack space. Due to its huge capacity and low unit storage cost, mechanical hard disks have so far remained the cornerstone of mass cold and warm data storage.

[0003] However, this "ultra-high density" physical deployment method, while improving storage efficiency, also introduces unprecedented reliability challenges. Traditional hard disk failure prediction methods mainly rely on the following aspects: Internal parameter monitoring based on SMART (Self-Monitoring, Analysis and Reporting Technology): This is the most basic failure warning mechanism. By reading the internal parameters recorded by the hard disk firmware, such as read error rate, seek error rate, remapping sector count, power-on time, etc., to assess the "health" of the hard disk. However, the limitations of SMART lie in its "individuality" and "hysteresis". It can only reflect the wear state of the hard disk itself and cannot perceive the impact of the external environment. In addition, a large number of studies and practices have shown that many hard disks do not show obvious abnormalities in their SMART parameters before sudden failure, resulting in failed warnings.

[0004] External state analysis based on system logs and I / O behavior: This method analyzes I / O timeouts, read / write retries, and check errors recorded by the operating system or RAID controller to determine whether the hard disk is in a sub-healthy state. The drawback of this method is its "passivity". When the system records these errors, it usually means that data access has been substantially affected, and the failure has already occurred or is occurring, making it difficult to play a role in advance prediction.

[0005] In recent years, the industry has begun to use large-scale hard disk cluster historical operation and maintenance data to build a failure prediction model through machine learning algorithms. These models analyze tens of thousands of hard disk SMART data and failure records to try to find highly correlated parameter combinations and evolution patterns. Although this method improves the accuracy of prediction to some extent, it has a fundamental blind spot: it treats each hard disk as an independent statistical unit that is not affected by its neighbors. But in high-density chassis, hard disks form a close microenvironment, and the operation of a hard disk will conduct physical effects outward, forming a neighborhood effect, for example, when multiple hard disks perform high-intensity random read and write at the same time, the movement of their head arms will produce micro-vibrations. These vibrations are conducted to adjacent hard disks through the rigid structure of the server chassis, backplane, etc., forming "vibration crosstalk." This continuous external micro-vibration will interfere with the precise positioning of the hard disk head and accelerate its mechanical wear. A hard disk with high load will generate a lot of heat, and if the heat dissipation is not good, it will directly increase the working temperature of its adjacent hard disks, and long-term high temperature is a key factor in accelerating the aging of electronic components and mechanical parts.

[0006] The fundamental defect of the prior art is that all mainstream prediction methods ignore the fact that in ultra-high-density racks, the running state of a hard disk as a physical entity is directly physically affected by its neighbors. In a chassis containing dozens of hard disks, when multiple hard disks perform high-intensity random read and write operations at the same time, the high-speed reciprocating motion of their head arms will produce tiny, high-frequency vibrations. These vibrations are conducted to adjacent hard disks through the rigid structure of the server chassis, backplane, etc., forming a kind of "vibration crosstalk" or "resonance effect." This continuous external micro-vibration will interfere with the precise positioning of the hard disk head, forcing its servo system to make additional compensation, thereby increasing the seek time, reducing performance, and accelerating the mechanical wear of the head arm assembly. This is a long-term, cumulative damage process that does not immediately manifest as a serious error in SMART data, but it is a key, long-neglected "invisible killer" that leads to premature failure of hard disks. SUMMARY

[0007] To overcome the shortcomings of the prior art, the present application provides a hard disk failure trend prediction method based on an Internet of Things storage node cluster to solve the problems mentioned in the background art.

[0008] To achieve the above purpose, the present application is implemented by the following technical scheme: a hard disk failure trend prediction method based on an Internet of Things storage node cluster, comprising the following steps: S1, acquire and load the topology mapping data of the physical relationship of each hard disk in the hard disk cluster, and collect the real-time SMART health status data of the jth hard disk in the hard disk cluster, process it through the preset machine learning model, quantify the internal physical wear and performance degradation of the jth hard disk, and generate the first evaluation value representing its own health status; S2, according to the topology mapping data, identify a plurality of hierarchical neighborhood sets with a preset correlation strength for the jth hard disk, and correct the neighborhood effect of the fusion hard disk cluster, specifically including: constructing a neighborhood stress index by analyzing the health status and historical failure data of the physically adjacent hard disks of the jth hard disk; according to the neighborhood stress index, correcting the first evaluation value to obtain the second evaluation value, and classifying the second evaluation value to obtain the first classification result, so as to establish a key monitoring hard disk group; S3, for the key monitoring hard disk group, collect the servo error data of the head seeking during high-frequency read-write operation, identify and quantify the resonance peak value at a specific frequency caused by adjacent hard disk vibration conduction through frequency spectrum analysis of the servo error data, thereby constructing a vibration resonance degradation degree representing the influence of physical micro-vibration, and classifying to obtain the second classification result; S4, for the key monitoring hard disk group, collect the average seek time when executing random read-write tasks, construct a seek performance degradation coefficient for representing mechanical arm aging and response delay by analyzing the dispersion and fluctuation trend of the average seek time relative to the reference value, and classifying to obtain the third classification result.

[0009] Preferably, the real-time SMART health status data of any hard disk in the hard disk cluster is acquired, and a long short-term memory network LSTM model is constructed to train and process the real-time SMART health status data, thereby generating the first evaluation value representing the current hard disk health status; The process of constructing a long short-term memory network LSTM model to train and process the real-time SMART health status data includes: The time series of the collected SMART health status data of the hard disk is input into the long short-term memory network LSTM model as input data; The time series of the SMART health status data of the hard disk includes: The time series of the remapping sector count value, the time series of the seek error rate value, the time series of the underlying read error rate value, the time series of the spin retry count value, and the time series of the internal operating temperature; According to the time series of the SMART health status data of the hard disk, a set of health status characteristic indexes of the jth hard disk in the time period t is calculated, including: bad track diffusion risk degree, seek system degradation slope, read instability index, motor start difficulty frequency and peak stress temperature. The Min-Max-Normalization method is applied to normalize the health status characteristic indicators, which are mapped to the interval [0, 1], and the first evaluation value of the jth hard disk is finally obtained through weighted summation.

[0010] In the time period t, the specific steps for obtaining the bad track diffusion risk degree, the seek system degradation slope, the read instability index, the motor startup difficulty frequency and the peak stress temperature of the jth hard disk are as follows: For the time series of the remapping sector count value, differential operation is performed to determine the maximum instantaneous growth rate thereof within the monitoring period, the maximum instantaneous growth rate is defined as the bad track diffusion risk degree, and the specific time point at which the maximum instantaneous growth rate occurs is recorded; For the time series of the seek error rate value, data within a sliding time window of a preset length is intercepted, linear regression analysis is performed on the data within the sliding time window to calculate the slope of the regression line, and the slope is defined as the seek system degradation slope; For the time series of the underlying read error rate value, the moving average value thereof within a preset time span is calculated to obtain a running baseline, and then the current real-time error rate value is compared with the running baseline to calculate the statistical deviation degree therebetween, and the statistical deviation degree is defined as the read instability index; For the time series of the spin retry count value, within a fixed monitoring time window, the total number of events of jump growth of the value is counted, and then divided by the total length of the monitoring time window to perform time normalization to obtain the motor startup difficulty frequency; Based on the specific time point at which the maximum instantaneous growth rate occurs, the temperature value corresponding to the specific time point is extracted from the time series of the internal operating temperature, the extracted temperature value is subtracted by a preset standard operating temperature to obtain a difference value, and the difference value is defined as the peak stress temperature.

[0011] Preferably, based on the topology mapping data, a plurality of hierarchical neighborhood sets with a preset association strength for the jth hard disk are identified, including: All hard disks sharing the same backplane ID or controller ID with the jth hard disk are identified to constitute a first hierarchical neighborhood set; All hard disks sharing the same chassis ID with the jth hard disk, but not belonging to the first hierarchical neighborhood set, are identified to constitute a second hierarchical neighborhood set; All hard disks sharing the same cabinet ID with the jth hard disk, but not belonging to the first hierarchical neighborhood set and the second hierarchical neighborhood set, are identified to constitute a third hierarchical neighborhood set.

[0012] Preferably, the historical first evaluation value time series and the historical failure records of all hard disks in the preset time window are obtained in the first level neighborhood set, the second level neighborhood set and the third level neighborhood set; Based on the historical first evaluation value time series and the historical failure records, a plurality of risk components are calculated; The plurality of risk components include: a current static risk component, a historical stability risk component and a dynamic deterioration trend component; The current static risk component is obtained by calculating the arithmetic mean of the current latest first evaluation value of all hard disks in the corresponding level neighborhood set; The level neighborhood set historical stability risk component is obtained by calculating the reciprocal of the mean time between failures (MTBF) or the annual failure rate (AFR) based on the historical failure records of the corresponding level neighborhood set in the preset statistical period; The dynamic deterioration trend component is obtained by performing trend analysis on the historical first evaluation value time series of each hard disk in the corresponding level neighborhood set to obtain the evaluation value growth slope, and calculating the average of the growth slopes of all hard disks; The current static risk component, the historical stability risk component and the dynamic deterioration trend component are respectively configured with preset component weight values, and the weighted sum is calculated to obtain the neighborhood stress index.

[0013] Preferably, the first evaluation value is corrected using the neighborhood stress index, and the product of the sum of the first evaluation value and the neighborhood stress index plus one is obtained as the second evaluation value; The second evaluation value is evaluated to obtain the first classification result, including: The preset first risk threshold and the second risk threshold, and the first risk threshold is greater than the second risk threshold; When the second evaluation value is lower than the second risk threshold, it indicates that the hard disk performance is qualified, and is marked as a safe level; When the second evaluation value is not lower than the second risk threshold but lower than the first risk threshold, a first risk level is generated; When the second evaluation value is not lower than the first risk threshold, a second risk level is generated; The hard disks of the first risk level and the second risk level are summarized and included in the key monitoring hard disk group.

[0014] Preferably, for each hard disk in the key monitoring hard disk group, the following steps are performed: During the execution of high-frequency read-write operations on each hard disk, the time series of servo error data of its head seeking is collected; The time series of servo error data is subjected to frequency spectrum analysis to generate a frequency response spectrum; In the frequency response spectrum, the peak power spectral density of the resonance peak value is calculated; In the frequency response spectrum, a resonance peak appearing at a specific frequency caused by physical vibration conduction of the adjacent hard disk is identified; The specific frequency includes a fundamental frequency corresponding to the spindle speed of the adjacent hard disk and one or more harmonic frequencies thereof; Based on the quantitative characteristics of the resonance peak, a vibration resonance degradation degree of the jth hard disk is constructed; the peak power spectral density is compared with a preset servo error baseline noise power spectral density without external vibration interference, and the ratio of the two is calculated, which is defined as the vibration resonance degradation degree.

[0015] Preferably, a third risk threshold and a fourth risk threshold are preset, wherein the fourth risk threshold is greater than the third risk threshold; The vibration resonance degradation degree is compared with the third risk threshold and the fourth risk threshold to obtain a second classification result, including: When the vibration resonance degradation degree is lower than the third risk threshold, the vibration impact level of the current hard disk is determined as the first slight impact; When the vibration resonance degradation degree is not lower than the third risk threshold but lower than the fourth risk threshold, the vibration impact level of the current hard disk is determined as the second moderate impact, and a third risk level is generated; When the vibration resonance degradation degree is not lower than the fourth risk threshold, the vibration impact level of the current hard disk is determined as the third serious impact, and a fourth risk level is generated; For the hard disk determined as the third risk level, a first workload scheduling optimization instruction is automatically triggered, which is used to: re-schedule 10%-15% of the I / O load allocated to the hard disk to the remaining hard disks in the first level neighborhood set whose second evaluation value is qualified; For the hard disk determined as the fourth risk level, a first priority alarm is generated, and a first data migration instruction is triggered, including: migrating the preset key data on the current hard disk to the hard disk in the third level neighborhood set whose second evaluation value is qualified, and marking the current hard disk as offline replacement.

[0016] Preferably, during the execution of the preset random read-write task on each hard disk, the average seek time is periodically collected to form a time sequence of the average seek time; The dispersion and fluctuation trend of the time sequence of the average seek time relative to the reference seek time are analyzed; Based on the analysis result, a seek performance attenuation coefficient for representing the aging and response delay of the mechanical arm is constructed; The seek performance attenuation coefficient is obtained in the following way: Obtaining the factory calibration seek time of the hard disk or its historical average seek time in the early stage of health state as a reference seek time; calculating the statistical standard deviation of the time series of the average seek time relative to the reference seek time to quantify the dispersion; Applying a moving average or linear regression analysis method to the time series of the average seek time to calculate the change slope of the average seek time with time to quantify the fluctuation trend; and performing weighted summation on the normalized statistical standard deviation and the change slope to obtain a seek performance decay coefficient; Classifying according to the seek performance decay coefficient to obtain a third classification result, including: Predefining a fifth risk threshold and a sixth risk threshold, wherein the sixth risk threshold is greater than the fifth risk threshold; When the seek performance decay coefficient is lower than the fifth risk threshold, determining the seek performance state of the current hard disk as a first slight decay; When the seek performance decay coefficient is not lower than the fifth risk threshold but lower than the sixth risk threshold, determining the seek performance state of the current hard disk as a second moderate decay, and generating a fifth risk level; When the seek performance decay coefficient is not lower than the sixth risk threshold, determining the seek performance state of the current hard disk as a third severe decay, and generating a sixth risk level; For the hard disk determined to be in the first slight decay state, starting baseline recording and high-frequency trend tracking of the seek performance of the current hard disk; For the hard disk determined to be in the fifth risk level, automatically triggering a second workload scheduling optimization instruction, including: re-scheduling 16%-30% of the I / O load allocated to the current hard disk to the remaining hard disks in the first-level neighborhood set and having qualified second evaluation values; For the hard disk determined to be in the sixth risk level, generating an alarm of a second priority, and triggering a second data migration instruction, including: migrating all data on the hard disk to the hard disks in the second-level neighborhood set and having qualified second evaluation values, and marking the current hard disk as to be actively replaced.

[0017] Preferably, after the first workload scheduling optimization instruction, the second workload scheduling optimization instruction, the first data migration instruction and the second data migration instruction are completed, the steps S1-S4 are repeated for iteration until the second evaluation values of all hard disks are qualified.

[0018] The application provides a hard disk failure trend prediction method based on an Internet of Things storage node cluster. (1) The hard disk failure trend prediction method based on the storage node cluster of the Internet of Things, by constructing a hierarchical neighborhood set, for the first time, the physical proximity and environmental correlation between hard disks are included in the analysis model. This overcomes the limitations of traditional technologies that only analyze individual hard disks in isolation, and can effectively identify and warn of group and infectious failures caused by factors such as vibration conduction and local hot spots, improving the system and forward-looking of the prediction.

[0019] (2) The hard disk failure trend prediction method based on the storage node cluster of the Internet of Things, by performing frequency spectrum analysis on the head servo error, this method can accurately identify and quantify the physical vibration crosstalk influence caused by adjacent hard disks, and convert the fuzzy "environmental influence" into precise numerical features, providing reliable data basis for early and accurate fault diagnosis and root location.

[0020] (3) The hard disk failure trend prediction method based on the storage node cluster of the Internet of Things, establishes an automatic and differentiated response strategy based on multiple risk thresholds. According to the severity of the vibration impact, the system can automatically trigger different levels of intervention measures such as load scheduling optimization or data migration, realizing the transformation from passive alarm to active and low-cost risk mitigation, and promoting the improvement of operation efficiency and resource utilization.

[0021] (4) The hard disk failure trend prediction method based on the storage node cluster of the Internet of Things, combines SMART data reflecting the state of the hard disk itself with neighborhood risk indicators representing external environmental influences, and builds a multi-dimensional comprehensive health assessment model. This "individual-environment" collaborative analysis model can more comprehensively capture early warning signs of failure than single data source prediction methods, reducing false positive and false negative rates, and improving the accuracy and reliability of the overall prediction model. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 The method steps of the present application are shown in the figure.

[0023] Figure 2 The S1-S2 flowchart of the present application is shown.

[0024] Figure 3 The S3-S4 flowchart of the present application is shown. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0026] Embodiment 1 Referring to Figures 1 to 3 The application provides a hard disk failure trend prediction method based on an Internet of Things storage node cluster, comprising the following steps: S1, acquiring and loading the topology mapping data of the physical relationship of each hard disk in the hard disk cluster, and collecting the real-time SMART health state data of the jth hard disk in the hard disk cluster, processing through a preset machine learning model to quantify the internal physical wear and performance degradation, and generating a first evaluation value representing the health state of the hard disk itself; S2, according to the topology mapping data, identifying a plurality of hierarchical neighborhood sets with a preset correlation strength for the jth hard disk, and correcting the risk of the “neighborhood effect” of the fusion hard disk cluster, specifically including: constructing a neighborhood stress index by analyzing the health state and historical failure data of the physically adjacent hard disks of the jth hard disk; according to the neighborhood stress index, the first evaluation value is corrected to obtain a second evaluation value, and the second evaluation value is risk graded to obtain a first grading result, and a key monitoring hard disk group is established; S3, for the key monitoring hard disk group, collecting the servo error data of the head seeking during high-frequency read-write operation, identifying and quantifying the resonance peak value at a specific frequency caused by vibration conduction of adjacent hard disks through frequency spectrum analysis of the servo error data, thereby constructing a vibration resonance degradation degree representing the influence of physical micro-vibration, and grading to obtain a second grading result; S4, for the key monitoring hard disk group, collecting the average seek time when performing random read-write tasks, constructing a seek performance degradation coefficient for representing mechanical arm aging and response delay by analyzing the dispersion and fluctuation trend of the average seek time relative to the reference value, and grading to obtain a third grading result.

[0027] The real-time SMART health state data of any hard disk in the hard disk cluster is acquired, and a long short-term memory network LSTM model is constructed to train and process the real-time SMART health state data, generating a first evaluation value representing the health state of the hard disk itself, denoted as the first evaluation value; The training and processing process of the long short-term memory network LSTM model includes: The time series of the collected SMART health state data of the hard disk is taken as input data to the long short-term memory network LSTM model; The long short-term memory network (LSTM) model is a kind of machine learning model, and its core operation process aims to automatically learn and extract complex, nonlinear time-dependent relationships from original, multi-dimensional SMART health state time series data, to realize accurate prediction of the future health trend of the hard disk.

[0028] The model operates as follows: First, time-series data of five key SMART attributes (remapping sector count, seek error rate, etc.) are constructed into a multivariate sequence and input into the LSTM network. At each time step, the LSTM unit receives the current SMART data vector. Its internal "forget gate" determines and selectively forgets historical information that is no longer important in long-term memory (cell state) based on the new input data and the state at the previous time step.

[0029] Meanwhile, the "input gate" determines which new information at the current moment (e.g., a sudden seek error) is important and updates it to the cell state, thus forming a new long-term memory. Finally, based on the updated cell state, the "input gate" generates a hidden state for the current time step, which contains both an understanding of the current situation and serves as a short-term memory for the next time step.

[0030] By repeatedly performing this process over time, the LSTM model can capture complex patterns spanning multiple time steps, such as "bad sector growth is often accompanied by fluctuations in read error rate." After training, the model's final output layer can transform these learned deep temporal features into a quantified, forward-looking hard drive health assessment score or the probability of failure within a specific future time window.

[0031] The time series of SMART health status data for the hard drive includes: Time series of remapped sector counts, seek error rates, underlying read error rates, rotational retry counts, and internal operating temperatures; Based on the time series of SMART health status data of the hard drive, the risk of bad sector propagation, seek system degradation slope, read instability index, motor start-up difficulty frequency and peak stress temperature of the j-th hard drive within the time period t are calculated. Within time period t, the specific steps for obtaining the bad sector propagation risk, seek system degradation slope, read instability index, motor start-up difficulty frequency, and peak stress temperature of the j-th hard drive are as follows: Differential operations are performed on the time series of remapped sector count values ​​to determine the maximum instantaneous growth rate within the monitoring period. The maximum instantaneous growth rate is defined as the risk level of bad sector propagation, and the specific time point at which the maximum instantaneous growth rate occurs is recorded. For the time series of seek error rate values, data within a preset length sliding time window is extracted, and linear regression analysis is performed on the data within the sliding time window to calculate the slope of the regression line, and the slope is defined as the seek system decay slope. For the time series of the underlying read error rate value, calculate its moving average value within a preset time span to obtain a running baseline, then compare the current real-time error rate value with the running baseline to calculate the degree of statistical deviation between the two, and define the degree of statistical deviation as the read instability index; For the time series of the rotation retry count value, within a fixed monitoring time window, count the total number of times of the value jumping up and divide it by the total length of the monitoring time window to obtain the motor start difficulty frequency Based on the specific time point at which the maximum instantaneous growth rate occurs, extract the temperature value corresponding to the specific time point from the time series of the internal operating temperature, subtract the preset standard operating temperature from the extracted temperature value to obtain a difference value, and define the difference value as the peak stress temperature.

[0032] Apply the Min-Max-Normalization method to normalize the bad track spread risk, track system degradation slope, read instability index, motor start difficulty frequency, and peak stress temperature of the jth hard disk within the time period t to the [0, 1] interval, and finally obtain the first evaluation value of the jth hard disk by weighted summation. The SMART health status data sampling of the hard disk is shown in the following Table 1: Table 1: SMART data sampling example of hard disk SN-XYZ001 at 10 time points

[0033] Simulate the process of gradually deteriorating hard disk status: bad tracks start to appear and grow, seek error rate continues to rise, motor occasionally starts difficultly, and temperature peaks when bad tracks grow fastest.

[0034] Calculate five core health features, including: bad track spread risk (maximum instantaneous growth rate): calculate the point-by-point difference value of the remapped sector count: [0, 0, 0, 1, 2, 1, 0, 0, 0]; the maximum instantaneous growth rate occurs at tk=5, with a value of 2. Result: bad track spread risk = 2.0; record: the specific time point tpeak=5 at which the maximum instantaneous growth rate occurs.

[0035] Track system degradation slope: intercept the seek error rate data in the sliding window (tk=5 to 9): [20, 22, 25, 26, 29]; perform linear regression analysis on these 5 points to calculate the slope of the regression line. For simplicity, the regression line slope is (29-20) / (9-5)=9 / 4=2.25; the track system degradation slope = 2.25 is obtained; Read instability index (statistical deviation); calculate 3-point moving average of underlying read error rate. For example, at tk=8, the data point is 15, the two points before it are 9 and 11. Run baseline (tk=8) = (9+11+15) / 3 ~ 11.67; statistical deviation = current real-time value - run baseline = 15-11.67 = 3.33; result: read instability index = 3.33; Motor start difficulty frequency: observe the time series of spin retry count: [0, 0, 0, 0, 0, 1, 1, 1, 2, 2]; there are two events of numerical jump growth: tk=5, from 0 to 1. tk=8, from 1 to 2. Result: motor start difficulty frequency = 2; Peak stress temperature: based on tpeak=5, extract the corresponding temperature value from the internal operating temperature sequence: 48°C. Difference = extracted temperature value - standard operating temperature = 48-40 = 8; result: peak stress temperature = 8.0; apply the Min-Max-Normalization method to the above data as follows Table 2: Table 2: Normalization example table of five core health features calculated based on Table 1

[0036] Based on the data in Table 2, the first evaluation value is obtained by weighted summation: (0.200x0.4) + (0.542x0.2) + (0.167x0.15) + (0.400x0.15) + (0.520x0.1) = 0.0800 + 0.1084 + 0.0251 + 0.0600 + 0.0520 = 0.3255; The higher the first evaluation value, the greater the risk, and the bad track spread risk degree, the seek system recession slope, the read instability index, the motor start difficulty frequency, and the peak stress temperature of the jth hard disk, all five features have the same positive correlation risk attribute. Among them, the bad track spread risk degree has the physical meaning of: the growth rate of remapping sectors (bad tracks), the faster the growth rate, the more serious the damage to the disk medium, and the higher the risk. The seek system recession slope has the physical meaning of: the deterioration trend of seek error rate over time. The larger the slope, the faster the mechanical arm system declines, and the higher the risk. The read instability index has the physical meaning of: the degree of deviation of the underlying read error rate from the normal baseline. The larger the deviation, the more unstable the reading, and the higher the risk. The motor start difficulty frequency has the physical meaning of: the number of events of spindle motor startup failure and retry. The more the number, the more obvious the motor or circuit failure, and the higher the risk. The peak stress temperature has the physical meaning of: the additional thermal stress that the hard disk bears when the bad track grows the fastest; the larger the temperature difference, the greater the pressure the hard disk bears at the time, and the higher the risk.

[0037] Traditional methods typically monitor whether the raw or normalized values ​​of SMART parameters exceed manufacturer-preset, fixed thresholds. For example, an alarm might only be triggered when the "remapped sector count" exceeds a certain value. This method is essentially a static snapshot, only able to identify damage that has already occurred and accumulated to a certain extent. It suffers from significant lag and cannot dynamically track the evolution of the fault, thus often missing the optimal early warning opportunity.

[0038] The difference between this invention and traditional methods lies in the following: taking "bad sector propagation risk" as an example, it focuses not on the total number of bad sectors, but on their maximum instantaneous growth rate within the monitoring period. The risk hidden behind a hard drive's bad sectors slowly increasing from 0 to 10 is vastly different from its sudden surge from 0 to 10 in a single day. The former might be normal wear and tear, while the latter could be an abnormal precursor to catastrophic deterioration of the disk media. Similarly, the "seek system decay slope" quantifies the deterioration trend of the seek error rate through linear regression analysis. This dynamic analysis based on differentiation and slope can capture the early acceleration stage of a fault's transformation from quantitative to qualitative change, advancing the warning time from "after the fault occurs" to "before the fault accelerates its deterioration," achieving true predictivity rather than "post-event confirmation."

[0039] Secondly, the "peak stress temperature" approach doesn't view temperature in isolation, but rather correlates the "specific point in time when bad sectors grow the fastest" with the "internal operating temperature" at that moment. This establishes a strong causal relationship: it reveals whether the hard drive's most vulnerable part (the disk media) has suffered substantial damage when subjected to maximum physical stress (manifested as temperature peaks). This is far more insightful than simply looking at the highest temperature or the total number of bad sectors. This method integrates previously isolated data points into a deep feature reflecting the hard drive's response behavior under specific operating conditions, making the initial assessment value not just a score, but a diagnostic report of the hard drive's multi-dimensional internal health status. In summary, the first evaluation value generation method, by introducing dynamic trend analysis, in-depth mining of physical meaning, and multi-dimensional feature fusion, helps to overcome the lag, one-sidedness, and unreliability of existing technologies that rely on static and isolated SMART data, thereby improving the prediction accuracy of hard disk cluster failures.

[0040] Example 2 Please refer to Figures 1-2 Based on topological mapping data, identify multiple hierarchical neighborhood sets with preset association strengths for the j-th hard disk, including: Identify all hard drives that share the same backplane ID or controller ID with the jth hard drive to form a first-level neighborhood set; if the power filter capacitor on this backplane is aging, or a component is short-circuited, it will cause the power supply voltage to be unstable or generate ripple. This unstable "dirty electricity" will simultaneously affect all hard drives connected to this backplane, and it is extremely easy to induce them to collectively have read / write errors or even be damaged. Sharing a backplane means sharing the power supply path on the same circuit board. Any electrical problem of the backplane (such as unstable voltage, aging capacitor) will indiscriminately affect all hard drives on it, causing collective failure; their data streams usually converge to the same SAS expander chip on the backplane. The performance bottleneck or failure of the current chip will directly paralyze the data IO of all hard drives. Fixed on the same rigid metal structure. Abnormal vibration of any hard drive in the first-level neighborhood set will be directly and efficiently conducted to the neighbors through the backplane, interfering with their read / write heads and forming mechanical resonance; The controller (such as an HBA card or a RAID card) is the hub of commands and data. Firmware defects, overheating, or hardware failures of the controller will directly cause all hard drives under its management to simultaneously go offline or behave abnormally; the first-level neighborhood set represents a direct, strong coupling causal relationship. A serious problem (electrical, mechanical, logical) of one member can directly trigger the failure of other members. Therefore, the risk signal from this level has the highest weight; Identify all hard drives that share the same backplane ID or controller ID with the jth hard drive to form a first-level neighborhood set; if the power filter capacitor on this backplane is aging, or a component is short-circuited, it will cause the power supply voltage to be unstable or generate ripple. This unstable "dirty electricity" will simultaneously affect all hard drives connected to this backplane, and it is extremely easy to induce them to collectively have read / write errors or even be damaged. Sharing a backplane means sharing the power supply path on the same circuit board. Any electrical problem of the backplane (such as unstable voltage, aging capacitor) will indiscriminately affect all hard drives on it, causing collective failure; their data streams usually converge to the same SAS expander chip on the backplane. The performance bottleneck or failure of the current chip will directly paralyze the data IO of all hard drives. Fixed on the same rigid metal structure. Abnormal vibration of any hard drive in the first-level neighborhood set will be directly and efficiently conducted to the neighbors through the backplane, interfering with their read / write heads and forming mechanical resonance;

[0041] Identify all hard disks sharing the same cabinet ID with the jthhard disk, but not belonging to the first level neighborhood set and the second level neighborhood set, to form the third level neighborhood set; the regional infrastructure is shared, and the risk is the correlation of the environmental background value. This is the weakest level of association. Different chassis are physically isolated, but these hard disks are placed in the same cabinet; the power of the entire cabinet comes from the same PDU. Overload or failure of the PDU will affect all devices in the cabinet, but this is a higher level of failure point; in the same vertical physical space, the cabinet level cooling environment is shared. For example, the temperature at the top of the cabinet is usually higher than that at the bottom, and this "thermal stratification" will affect the reference operating temperature of all chassis in the current cabinet; the entire cabinet will be affected by the low-frequency vibration of the data center floor, and this vibration will be transmitted to all chassis in the cabinet; the third level neighborhood set represents the macro, background level risk correlation. These factors are more like setting a common "environmental pressure baseline" for all members, and their influence is chronic and widespread, but much weaker and less direct than the previous two levels.

[0042] Obtain the historical first evaluation value time series and historical failure records of all hard disks in the first level neighborhood set, the second level neighborhood set and the third level neighborhood set within a preset time window; Based on the historical first evaluation value time series and the historical failure records, calculate a plurality of risk components; The plurality of risk components include: a current static risk component, a historical stability risk component and a dynamic deterioration trend component; The current static risk component is obtained by calculating the arithmetic mean of the current latest first evaluation values of all hard disks in the corresponding level neighborhood set; The historical stability risk component is obtained by calculating the reciprocal of the mean time between failures MTBF or the annual failure rate AFR based on the historical failure records of the corresponding level neighborhood set within a preset statistical period; The dynamic deterioration trend component is obtained by performing trend analysis on the historical first evaluation value time series of each hard disk in the corresponding level neighborhood set to obtain the evaluation value growth slope, and calculating the average of the growth slopes of all hard disks; The current static risk component, the historical stability risk component and the dynamic deterioration trend component are respectively configured with a preset component weight value, and the weighted sum is calculated to obtain the neighborhood stress index. The identification process is as shown in Table 3: Table 3: Topology mapping data example table

[0043] Perform failure prediction on the target hard disk. When the target hard disk is SNABCD03, then: First level neighborhood set (N1): Identify all hard drives that share the same backplane ID (backplane 1A) or controller ID (controller 1A) as the target hard drive SNABCD03. In the execution process, find hard drives that meet the conditions: SNABCD02: backplane ID = backplane 1A, controller ID = controller 1A. Match.

[0044] SNABCD04: backplane ID = backplane 1A, controller ID = controller 1A. Match.

[0045] SNABCD05: backplane ID = backplane 1A, controller ID = controller 1A. Match.

[0046] The backplane ID and controller ID of all other hard drives are not backplane 1A or controller 1A; Identification result: First level neighborhood set (N1) = {SNABCD02, SNABCD04, SNABCD05}; This set represents the "fate community" most closely connected to SNABCD03. They share the same power supply path, the same data link controller, and the same rigid physical structure. Electrical noise from the backplane, firmware defects from the controller, or severe vibration from any of the hard drives will affect all other members in this set in the highest priority and most direct way; Second level neighborhood set (N2): Identification rule: Identify all hard drives that share the same chassis ID (chassis 101) as the target hard drive SNABCD03 and are not in the first level neighborhood set (N1). First, find all hard drives that share the chassis ID chassis 101: {SNABCD02, SNABCD03, SNABCD04, SNABCD05}.

[0047] Then, exclude all members that have already belonged to N1: {SNABCD02, SNABCD04, SNABCD05}.

[0048] At the same time, exclude the target hard drive itself SNABCD03. After the exclusion, the second level neighborhood set is empty.

[0049] Third level neighborhood set (N3): Identify all hard drives that share the same cabinet ID (cabinet A07) as the target hard drive SNABCD03 and are not in the first level neighborhood set (N1) and the second level neighborhood set (N2).

[0050] Execution process: First, find all hard drives that share the cabinet ID cabinet A07: {SNABCD02,...,SNABCD05} (from chassis 101); {SNEFGH15,SNEFGH16,SNEFGH17} (from chassis 102); {SNIJKL30,SNIJKL31} (from chassis 103); Then, exclude all members already belonging to N1 : {SNABCD02, SNABCD04, SNABCD05}.

[0051] Then, exclude all members already belonging to N2: (empty set).

[0052] Finally, exclude the target disk itself SNABCD03. The remaining disks form the N3 set. Result: Third level neighborhood set (N3) = {SNEFGH15, SNEFGH16, SNEFGH17, SNIJKL30, SNIJKL31}. The third level neighborhood set represents disks that share the macro environment (same chassis) with SNABCD03 but do not have direct physical or environmental medium coupling. They share power from the same PDU, are in similar chassis temperature stratification, and are subjected to the same low frequency vibrations from the data center floor. These factors provide a reference for the "environmental stress baseline" of SNABCD03.

[0053] Exclusion calculation. Because: suppose the backplane (N1) where the target disk j resides has an electrical failure, causing the health of its neighbors to drop dramatically. At the same time, the fan speed of the chassis (N2) where it resides also drops slightly. If not excluded, calculate N1 stress: the system will collect the poor state of N1 neighbors and derive an abnormally high N1 stress index. (Correct) Calculate N2 stress: the system will count all the disks in the chassis (including N1 members that have already been calculated). Since the N1 members are in an abnormally poor state, they will again severely "pollute" the N2 calculation results, causing the N2 stress index to also become abnormally high, resulting in a high N2 stress index, indicating an error state; Calculate N3 stress: for the same reason, the poor state of N1 members will pollute the N3 calculation for the third time, causing the N3 stress index to also be abnormally high, resulting in a high N3 stress index, indicating an error state; The system will report: "Target disk j is simultaneously subjected to three high-intensity risks from the backplane, chassis, and cabinet!" This is completely misleading. The operations personnel will not be able to determine the root cause of the problem and will not know whether to replace the backplane, repair the fan, or check the PDU.

[0054] Calculate N1 neighborhood stress index: the result is abnormally high, clearly pointing to a backplane / controller problem.

[0055] Compute N2 neighborhood stress index: Since N1 members are excluded, N2 computation is only based on those hard drives that share the same chassis but not the backplane. If these hard drives are in good condition, N2 stress index will be normal.

[0056] Compute N3 neighborhood stress index: Similarly, N3 computation excludes N1 and N2 members, resulting in a normal background stress value. Correct conclusion: System reports: “Target hard drive j is experiencing abnormally high first level (N1) stress, while other levels are normal”. The diagnosis is clear and unambiguous: the problem lies in the backplane or controller.

[0057] Traditional cluster analysis methods, even if considering physical location, often only come to a vague “correlation” conclusion, such as “hard drives in cabinet A07 have a higher failure rate”. This conclusion is almost useless to the operations personnel, because it cannot answer the most critical question: “Why?” The root cause could be the PDU of the cabinet, the cooling of a certain chassis, or a specific backplane? The operations personnel can only check one by one, which is time-consuming and labor-intensive, and is prone to misdiagnosis.

[0058] In this embodiment, the neighborhood is divided into N1 (strong causal coupling), N2 (environmental medium coupling), and N3 (macroscopic background coupling) three levels. More importantly, through exclusive calculation, the present application can decouple stress sources at different levels. As you mentioned, when the backplane (N1) has an electrical fault, the present application can accurately attribute the risk to N1, and will not “pollute” the evaluation of N2 and N3 with this strong signal. This makes the system no longer output a general alarm, but a conclusion with diagnostic value, such as: “Target hard drive is experiencing abnormally high first level neighborhood stress, and the root cause of the problem is likely to be in the backplane or controller”. This capability improves the efficiency of fault troubleshooting by several orders of magnitude.

[0059] Without the concept of hierarchical, all neighboring hard drives are treated equally, and the urgency and impact range of the threat cannot be distinguished. A neighbor hard drive temperature is slightly higher, and another neighbor hard drive is in a backplane with a large number of I / O errors, which may be blurredly counted as the same level of “neighborhood risk” in the traditional model.

[0060] The component weight values applied to the first level neighborhood set, the second level neighborhood set and the third level neighborhood set are different; wherein, in the weight configuration for calculating the neighborhood stress index of the first level neighborhood set, the weight value of the dynamic deterioration trend component is the highest; wherein, in the weight configuration for calculating the neighborhood stress index of the second level neighborhood set, the weight value of the current static risk component is the highest. The preferred weight configuration is as follows Table 4: Table 4: Preferred weight configuration example table

[0061] Among them, the technical principle description: the risk core of N1 is "sudden" and "strong infectivity". A historically stable controller (low Whistory) may suddenly crash due to firmware bug. At this time, the health degree of the hard disk it manages will decrease sharply, which is manifested as an abnormally high "dynamic deterioration trend slope". Assigning the highest weight to Wtrend is to make the system respond most sensitively to this fatal and accelerating risk. The weight of Whistory is the lowest, because for such a sudden event, the past "good record" has almost no reference value. The risk core of N2 is "environment" and "collectivity". For example, the failure of the case fan leads to poor heat dissipation, which will slowly and collectively raise the temperature of all hard disks in the case, which is reflected in the slow lifting of the overall mean value of their "first evaluation value". Wstatic (current health mean value) is the best indicator to capture this collective sub-health state. At this time, Wtrend and Whistory work together as auxiliary indicators to depict the long-term stability and trend of the environment. The risk core of N3 is "macro" and "long-term". For example, a certain cabinet has a high historical failure rate (AFR) of the devices inside it due to its location in the hot spot area of the data center. This risk is chronic and background. Whistory can most stably reflect this long-term "environmental pressure baseline". The weights of Wstatic and Wtrend are relatively low, because the N3 members are large and heterogeneous, and their short-term average health status and change trend are easily overwhelmed by various business noise, and the reference value is relatively limited.

[0062] The N1, N2 and N3 levels in this embodiment naturally correspond to the priority of risks.

[0063] N1 neighborhood stress index rises: represents serious problems within the "fate community", with high infectivity and suddenness, and should be given the highest alarm priority, and may need to be isolated or migrated immediately. The entire N1 set.

[0064] N2 neighborhood stress index rises: represents the deterioration of the case-level environment (such as heat dissipation), and the risk is progressive and collective, which should trigger a medium-priority alarm and schedule operations such as checking the case fan and cleaning the air duct.

[0065] N3 neighborhood stress index increases: represents the macro-environmental changes of the cabinet or data center level, the risk is chronic, background, can be used as a reference for long-term observation and capacity planning. This hierarchical risk assessment enables the operation system to automatically prioritize a large number of alarms, allowing the operation personnel to focus on the most deadly threats first, avoiding "alarm storm" and "operation fatigue". The hard disks in the N1 set share the most fragile lifeline (power supply, data link). When any member of the N1 set experiences a catastrophic failure of a specific type (for example, a SMART data instantaneous anomaly caused by a circuit short circuit), the application can immediately and unconditionally raise the risk level of all other apparently healthy hard disks in the entire N1 set to the highest. This is a forward-looking prediction based on the physical causal chain. It can trigger preventive data evacuation of the entire "community of destiny", effectively preventing "case-style" data loss caused by single-point failures such as backplane and controller (for example, multiple hard disks in a RAID group fail simultaneously), and enhancing the resilience and data security of the entire storage system.

[0066] Embodiment 3 Please refer to Figures 1-2 , using the neighborhood stress index to modify the first evaluation value, multiplying the sum of the first evaluation value and the neighborhood stress index plus one to obtain the second evaluation value; because the final risk of the hard disk is the nonlinear amplification result of its own inherent risk under external environmental pressure; evaluating the second evaluation value to obtain the first classification result, including: presetting a first risk threshold and a second risk threshold, and the first risk threshold is greater than the second risk threshold; when the second evaluation value is lower than the second risk threshold, it indicates that the hard disk performance is qualified, and is marked as a safe level; when the second evaluation value is not lower than the second risk threshold but lower than the first risk threshold, it indicates that the hard disk performance has a slight risk, and a first risk level is generated; when the second evaluation value is not lower than the first risk threshold, it indicates that the hard disk performance has a high-risk, and a second risk level is generated; summarizing the hard disks of the first risk level and the second risk level into a key monitoring hard disk group.

[0067] In this embodiment, the first evaluation value of the hard disk itself is combined with the neighborhood stress index through the correction formula. Through the multiplication operation, the formula can reasonably reflect the amplification effect of external environmental pressure on the inherent risk of the hard disk, thereby obtaining a second evaluation value closer to the actual working condition.

[0068] Based on the second evaluation value, the hard disks that need to be focused on can be clearly screened out to form a clear monitoring group. This enables the operation and maintenance resources to be concentrated on handling the highest risk devices, improving the pertinence of fault warning and the efficiency of maintenance work.

[0069] Embodiment 4 Please refer to Figure 1 and Figure 3 For each hard disk in the key monitoring hard disk group, the following steps are performed: During the execution of high-frequency read-write operations on each hard disk, a time series of servo error data of its head seeking is collected; The time series of servo error data is subjected to spectral analysis to generate a frequency response spectrum; In the frequency response spectrum, the peak power spectral density of the resonance peak value is calculated; In the frequency response spectrum, resonance peak values that occur at specific frequencies caused by physical vibration conduction of adjacent hard disks are identified; the specific frequencies include the fundamental frequency corresponding to the spindle speed of the adjacent hard disk and its harmonic frequency (s) ; Specifically, the time-domain signal of the servo error data is processed using Fast Fourier Transform (FFT), and the energy peak value at the fundamental frequency (for example, 120 Hz for a 7200RPM hard disk) and its harmonic frequency determined by the spindle speed of the adjacent hard disk is identified, with the amplitude of the current peak value as the basis for quantifying the resonance peak value; Based on the quantification characteristics of the resonance peak value, the vibration resonance degradation degree of the jth hard disk is constructed; the peak power spectral density is compared with the preset servo error baseline noise power spectral density under no external vibration interference, and the ratio of the two is calculated, defined as the vibration resonance degradation degree.

[0070] A third risk threshold and a fourth risk threshold are preset, wherein the fourth risk threshold is greater than the third risk threshold; The vibration resonance degradation degree is compared with the third risk threshold and the fourth risk threshold to obtain a second classification result, including: When the vibration resonance degradation degree is lower than the third risk threshold, the vibration impact level of the current hard disk is determined as the first slight impact; When the vibration resonance degradation degree is not lower than the third risk threshold but lower than the fourth risk threshold, the vibration impact level of the current hard disk is determined as the second moderate impact, and a third risk level is generated; When the vibration resonance degradation degree is not lower than the fourth risk threshold, the vibration impact level of the current hard disk is determined as the third severe impact, and a fourth risk level is generated; For the hard disk determined as the third risk level, a first workload scheduling optimization instruction is automatically triggered, which is used to: reschedule 10%-15% of the I / O load allocated to the hard disk to the remaining hard disks in the first hierarchical neighborhood set, which have qualified second evaluation values, so as to reduce the high concurrent I / O access conflict between the hard disk and the adjacent hard disk generating the main vibration source; For the hard disk determined as the fourth risk level, a first priority alarm is generated, and a first data migration instruction is triggered, including: migrating the preset key data on the current hard disk to the hard disk in the third hierarchical neighborhood set, which has a qualified second evaluation value, and marking the current hard disk as offline replacement.

[0071] During the execution of the preset random read-write task on each hard disk, the average seek time is periodically collected to form a time sequence of the average seek time; The dispersion and fluctuation trend of the time sequence of the average seek time relative to the reference seek time are analyzed; Based on the analysis result, a seek performance decay coefficient for characterizing the aging and response delay of the mechanical arm is constructed. In the present application, the "mechanical arm" refers to the mechanical moving component inside the hard disk for moving the magnetic head to perform data read-write, which is called a magnetic head assembly or a magnetic head arm, and the core components thereof include a voice coil motor (VCM) for driving the movement of the magnetic head arm and a precision bearing for supporting the rotation thereof.

[0072] The seek performance decay coefficient is obtained in the following manner: The factory calibration seek time of the hard disk or the historical average seek time thereof in the early health state is obtained as the reference seek time; the statistical standard deviation of the time sequence of the average seek time relative to the reference seek time is calculated to quantify the dispersion; The moving average or linear regression analysis method is applied to the time sequence of the average seek time to calculate the change slope thereof with time growth, so as to quantify the fluctuation trend; the normalized statistical standard deviation and the change slope are weighted and summed to obtain the seek performance decay coefficient; The seek performance decay coefficient is graded to obtain a third grading result, including: A fifth risk threshold and a sixth risk threshold are preset, wherein the sixth risk threshold is greater than the fifth risk threshold; When the seek performance decay coefficient is lower than the fifth risk threshold, the seek performance state of the current hard disk is determined as a first slight decay; When the seek performance decay coefficient is not lower than the fifth risk threshold but lower than the sixth risk threshold, the seek performance state of the current hard disk is determined as a second moderate decay, and a fifth risk level is generated; When the seek performance decay coefficient is not lower than the sixth risk threshold, the seek performance state of the current hard disk is determined as a third severe decay, and a sixth risk level is generated; For the hard disk determined as the first slight attenuation state, the baseline record and high frequency trend tracking of the current hard disk seek performance are started; For the hard disk determined as the fifth risk level, the second workload scheduling optimization instruction is automatically triggered, including: 16%-30% of the I / O load allocated to the current hard disk is rescheduled to the remaining hard disks in the first level neighborhood set, which have qualified second evaluation values; For the hard disk determined as the sixth risk level, a second priority alarm is generated, and a second data migration instruction is triggered, including: all data on the hard disk is migrated to the hard disks in the second level neighborhood set, which have qualified second evaluation values, while the current hard disk is marked as to be actively replaced.

[0073] After the first workload scheduling optimization instruction, the second workload scheduling optimization instruction, the first data migration instruction and the second data migration instruction are completed, the steps S1-S4 are repeated for iteration until the second evaluation values of all hard disks are qualified. The threshold preferred value is as shown in Table 5: Table 5: Risk threshold preferred table Risk threshold Correlation indicator Preferred range Trigger action Second risk threshold Second evaluation value 1.10~1.25 Labeled "minor risk" First risk threshold Second evaluation value 1.40~1.60 Labeled "high risk" and included in intensive monitoring Third risk threshold Vibration resonance degradation degree 3.0~5.0 Decided as "moderate impact" and triggers first load scheduling Fourth risk threshold Vibration resonance degradation degree 8.0~12.0 Decided as "severe impact" and triggers first data migration Fifth risk threshold Seek performance attenuation coefficient 0.30~0.45 Decided as "moderate attenuation" and triggers second load scheduling Sixth risk threshold Seek performance attenuation coefficient 0.60~0.80 Decided as "severe attenuation" and triggers second data migration The second risk threshold preferred value range is 1.10-1.25; The setting basis is: the core of the second evaluation value is the first evaluation value x (1+neighbor stress index). In an ideal state, this value should be close to 1.0. This threshold is set between 1.10 and 1.25, which means that when the comprehensive risk (self-state superimposed environmental stress) of a hard disk exceeds its health baseline by 10% to 25%, it is considered to enter the "slight risk" state that needs attention. This range is obtained through statistical analysis of a large amount of hard disk life cycle data, which can effectively identify potential problems at an early stage, while avoiding excessive false positives due to normal data fluctuations.

[0074] The first risk threshold (triggering the "slight risk" to "high risk" limit): Preferred range: 1.40-1.60; Setting basis: when the comprehensive risk exceeds the baseline by 40% to 60%, it indicates that the health of the hard disk has deteriorated significantly, or the stress of the environment it is in has reached a dangerous level. According to historical failure data, the probability of failure of a hard disk reaching this interval within a future period (such as a month) will increase exponentially. Setting the threshold at this point is the best compromise between "early warning" and "urgency of action", ensuring that there is enough time window for deep diagnosis and intervention before failure occurs.

[0075] Threshold for "vibration resonance degradation" (external influence diagnosis); Third risk threshold (triggering the boundary from "minor impact" to "moderate impact"): Preferred range: 3.0~5.0 (based on power spectral density ratio); Setting basis: "Vibration resonance degradation degree" is defined as the ratio of resonance peak power spectral density to baseline noise. A ratio of 1 indicates no impact. When the ratio reaches 3.0 to 5.0, it indicates that the vibration energy conducted from adjacent hard drives is significantly higher than the background noise, enough to cause measurable interference to the servo system of the magnetic head, possibly leading to occasional read / write retries and performance degradation. At this time, triggering "load scheduling optimization" is a low-cost, high-benefit mitigation measure.

[0076] Fourth risk threshold (triggering the boundary from "moderate impact" to "severe impact"): Preferred value: 8.0~12.0 (based on power spectral density ratio); Setting basis: When the resonance peak energy reaches more than 8 times the baseline noise, the servo system will consume a lot of resources to correct in order to maintain the magnetic head on the correct track, resulting in severe performance degradation and increasing the risk of off-track writing. This is a clear danger signal, indicating that physical vibration interference has become the primary threat. Therefore, "data migration" and "alarm" of the highest priority must be triggered to avoid imminent data security risks.

[0077] Threshold for "seek performance attenuation coefficient" (internal aging diagnosis); Fifth risk threshold (triggering the boundary from "minor attenuation" to "moderate attenuation"): Preferred range: 0.30~0.45 (normalized coefficient value); Setting basis: "Seek performance attenuation coefficient" is a normalized comprehensive index, 0 represents the factory state. When the coefficient value reaches 0.30 to 0.45, it indicates that the response time and stability of the head arm have shown a clear and continuous degradation trend. This degradation has exceeded the normal wear range and has begun to affect the I / O delay of the upper application. At this time, starting "load scheduling optimization" can reduce the pressure on the aging mechanical arm and delay its complete failure process.

[0078] Sixth risk threshold (triggering the boundary from "moderate attenuation" to "severe attenuation"): Preferred range: 0.60~0.80 (normalized coefficient value); Setting basis: when the coefficient value exceeds 0.60, it usually means that the mechanical wear of the head arm has entered the end stage. At this time, the random read-write performance of the hard disk will decrease sharply, becoming a system bottleneck, and the risk of mechanical part jamming or failure is abnormally high. Defining this interval as "severe attenuation" and triggering "full data migration" and "active replacement" instructions is the most reliable window to protect data and services before the hard disk has irreversible physical failure.

[0079] After screening out the key monitoring hard disk group, a double-path, deep-level fault root cause diagnosis and automatic disposal process is started. The core principle of this process is to finely deconstruct the physical health status of the hard disk from two dimensions of external environmental influence and internal mechanical wear; specifically: Diagnosis principle of vibration resonance degradation degree: this path focuses on diagnosing performance problems caused by external physical environment, especially adjacent hard disks. When the hard disk head reads and writes on the track, its positioning accuracy requirement is extremely high, usually at the nanometer level. In high-density deployed servers, the periodic vibration generated by the spindle motor of adjacent hard disks (such as 7200RPM corresponding to 120Hz) will be conducted through the chassis structure, causing interference to the target hard disk. This interference will make the servo system of the head have to constantly make fine adjustments, resulting in servo errors. This invention uses fast Fourier transform (FFT) technology to convert the collected servo error time domain signal into a frequency domain spectrum. In the frequency domain spectrum, the vibration caused by a specific neighbor hard disk will appear as a clear energy peak at its spindle speed frequency (and harmonic frequency). By quantifying the power spectral density of this resonance peak and comparing it with the baseline noise level of the hard disk in the ideal state, the "vibration resonance degradation degree" can be accurately calculated. This degradation degree objectively quantifies the actual impact of external vibration on the current hard disk read-write stability.

[0080] Diagnosis principle of seek performance attenuation coefficient: this path focuses on diagnosing the physical aging of the hard disk's internal core mechanical component, the head arm (Actuator-Arm). The performance of the head arm directly determines the random read-write ability of the hard disk. With long-time running, its mechanical structure will wear out, causing slow response and declining positioning accuracy. This invention represents this aging by analyzing the average seek time, a key performance indicator. Its innovation lies in that it not only focuses on the average value of seek time, but also focuses on two key statistical characteristics of its time series: Dispersity (quantified by standard deviation): a healthy head arm has stable and consistent seek time. An aging head arm will have jitter and irregularity, causing the dispersity of seek time distribution to increase.

[0081] Trend of fluctuation (quantified by slope): As wearout aggravates, the overall response speed of the head arm will show a downward trend, which is manifested as a slow growth of the average seek time over time. By weighted sum of the two quantitative indicators, a "seek performance decay coefficient" is constructed, which comprehensively reflects the complete decay process of the head arm from "performance stability" to "irregular response" to "overall slow down".

[0082] In this embodiment, not only is the problem diagnosed, but a set of automatic and hierarchical treatment strategies closely coupled with the diagnosis results is also provided. For moderate risk, the system takes "performance optimization" measures to alleviate the problem (such as reducing vibration conflicts and reducing the stress of aging components) without interrupting service through small-scale load scheduling. For severe risk, the system takes "active avoidance" measures to completely eliminate hidden dangers before failure occurs through data migration and alarms. From risk identification, deep diagnosis to automatic treatment and effect verification, the entire process forms a complete closed loop. This design enables the storage system to have a certain degree of "self-healing" capability, actively identifies and alleviates potential problems, reduces the dependence on manual intervention, and improves the stability and service continuity of the entire system. The intelligent selection of migration targets (such as migrating vibration problems to N3 and migrating aging problems to N2) also makes the treatment measures more targeted.

[0083] The size of the threshold is set for ease of comparison, and the size of the threshold depends on how much sample data and the base number set by the person skilled in the art for each group of sample data; as long as it does not affect the proportional relationship between the parameters and the quantized values.

[0084] The above formulas are obtained by collecting a large amount of data for software simulation and selecting a formula close to the true value. The coefficients in the formula are set by the person skilled in the art according to the actual situation. The above are only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can make equivalent replacements or changes to the technical solutions and inventive concepts of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1.A hard disk failure trend prediction method based on an Internet of Things storage node cluster, characterized in that, The method comprises the following steps: S1, acquiring and loading the topological mapping data of the physical relationship of each hard disk in the hard disk cluster, and collecting the real-time SMART health status data of the jth hard disk in the hard disk cluster, processing through a preset machine learning model to quantify the internal physical wear and performance degradation of the jth hard disk, and generating a first evaluation value representing the health status of the jth hard disk; S2, according to the topological mapping data, identifying a plurality of hierarchical neighborhood sets with a preset correlation strength of the jth hard disk, and correcting the neighborhood effect of the fusion hard disk cluster, specifically including: constructing a neighborhood stress index by analyzing the health status and historical failure data of the physically adjacent hard disk of the jth hard disk; according to the neighborhood stress index, the first evaluation value is corrected to obtain a second evaluation value, and the second evaluation value is risk graded to obtain a first grading result, so as to establish a key monitoring hard disk group; S3, for the key monitoring hard disk group, collecting the servo error data of the head tracking during high-frequency read-write operation, identifying and quantifying the resonance peak value at a specific frequency caused by adjacent hard disk vibration conduction through frequency spectrum analysis of the servo error data, thereby constructing a vibration resonance degradation degree representing the influence of physical micro-vibration, and grading to obtain a second grading result; S4, for the key monitoring hard disk group, collecting the average seek time when executing random read-write tasks, constructing a seek performance degradation coefficient for representing mechanical arm aging and response delay by analyzing the dispersion and fluctuation trend of the average seek time relative to the reference value, and grading to obtain a third grading result. 2.The hard disk failure trend prediction method based on the Internet of Things storage node cluster of claim 1, wherein, The real-time SMART health status data of any hard disk in the hard disk cluster is acquired, and a long short-term memory network (LSTM) model is constructed to train and process the real-time SMART health status data, thereby generating a first evaluation value representing the health status of the current hard disk; The training and processing of the real-time SMART health status data by the long short-term memory network (LSTM) model comprises: The time series of the collected SMART health status data of the hard disk is input into the long short-term memory network (LSTM) model as input data; The time series of the SMART health status data of the hard disk includes: The time series of the remapping sector count value, the time series of the seek error rate value, the time series of the underlying read error rate value, the time series of the spin retry count value, and the time series of the internal operating temperature; According to the time series of the SMART health status data of the hard disk, a set of health status characteristic indexes of the jth hard disk in the time period t is calculated, including: bad track diffusion risk degree, seek system degradation slope, read instability index, motor start difficulty frequency and peak stress temperature; The Min-Max-Normalization method is applied to normalize the health status characteristic indexes, which are mapped to the [0, 1] interval, and the first evaluation value of the jth hard disk is finally obtained by weighted summation. 3.The hard disk failure trend prediction method based on the Internet of Things storage node cluster of claim 2, characterized in that, The risk degree of the bad track diffusion, the decline slope of the track system, the read instability index, the frequency of the motor starting difficulty and the peak stress temperature of the jth hard disk in the time period t are specifically obtained as follows: For the time series of the remap sector count value, a differential operation is performed to determine the maximum instantaneous growth rate thereof in the monitoring period, the maximum instantaneous growth rate is defined as the risk degree of the bad track diffusion, and a specific time point at which the maximum instantaneous growth rate occurs is recorded; For the time series of the track error rate value, data in a sliding time window of a preset length is intercepted, linear regression analysis is performed on the data in the sliding time window to calculate the slope of the regression line, and the slope is defined as the decline slope of the track system; For the time series of the bottom read error rate value, a moving average value thereof in a preset time span is calculated to obtain a running baseline, then the current real-time error rate value is compared with the running baseline to calculate the degree of statistical deviation therebetween, and the degree of statistical deviation is defined as the read instability index; For the time series of the spin retry count value, in a fixed monitoring time window, the total number of events of the value thereof jumping up is counted, and then the total number of events is divided by the total length of the monitoring time window to obtain the frequency of the motor starting difficulty after time normalization; Based on the specific time point at which the maximum instantaneous growth rate occurs, a temperature value corresponding to the specific time point is extracted from the time series of the internal operating temperature, the extracted temperature value is subtracted by a preset standard operating temperature to obtain a difference value, and the difference value is defined as the peak stress temperature. 4.The hard disk failure trend prediction method based on the Internet of Things storage node cluster of claim 1, wherein, Based on the topology mapping data, a plurality of hierarchical neighborhood sets with a preset association strength of the jth hard disk are identified, including: All hard disks sharing the same backplane ID or controller ID with the jth hard disk are identified to constitute a first-level neighborhood set; All hard disks sharing the same chassis ID with the jth hard disk but not belonging to the first-level neighborhood set are identified to constitute a second-level neighborhood set; All hard disks sharing the same cabinet ID with the jth hard disk but not belonging to the first-level neighborhood set and the second-level neighborhood set are identified to constitute a third-level neighborhood set. 5.The hard disk failure trend prediction method based on the Internet of Things storage node cluster of claim 1, wherein, The historical first evaluation value time series and the historical failure records of all hard disks in the preset time window in the first-level neighborhood set, the second-level neighborhood set and the third-level neighborhood set are obtained; Based on the historical first evaluation value time series and the historical failure records, a plurality of risk components are calculated; The plurality of risk components include: a current static risk component, a historical stability risk component and a dynamic deterioration trend component; The current static risk component is obtained by calculating the arithmetic mean of the current latest first evaluation values of all hard disks in the corresponding hierarchical neighborhood set; The historical stability risk component is obtained by calculating the reciprocal of the mean time between failures MTBF or the annual failure rate AFR based on the historical failure records of the corresponding hierarchical neighborhood set in a preset statistical period. The dynamic deterioration trend component is obtained by performing trend analysis on the historical first evaluation value time series of each hard disk in the corresponding hierarchical neighborhood set to obtain an evaluation value growth slope, and calculating the average of the growth slopes of all hard disks; The current static risk component, the historical stability risk component and the dynamic deterioration trend component are respectively configured with a preset component weight value, and the weighted sum is calculated to obtain the neighborhood stress index. 6.The hard disk failure trend prediction method based on the Internet of Things storage node cluster of claim 5, wherein, The first evaluation value is corrected using the neighborhood stress index, and the product of the first evaluation value and the neighborhood stress index plus one is obtained to obtain the second evaluation value; The second evaluation value is evaluated to obtain the first classification result, including: The first risk threshold and the second risk threshold are preset, and the first risk threshold is greater than the second risk threshold; When the second evaluation value is lower than the second risk threshold, it indicates that the hard disk performance is qualified, and is marked as a safe level; When the second evaluation value is not lower than the second risk threshold but lower than the first risk threshold, a first risk level is generated; When the second evaluation value is not lower than the first risk threshold, a second risk level is generated; The hard disks of the first risk level and the second risk level are summarized and included in the key monitoring hard disk group. 7.The hard disk failure trend prediction method based on the Internet of Things storage node cluster of claim 1, wherein, For each hard disk in the key monitoring hard disk group, the following steps are performed: During the execution of high-frequency read-write operations on each hard disk, a time series of servo error data of its head seeking is collected; The time series of servo error data is subjected to frequency spectrum analysis to generate a frequency response spectrum; In the frequency response spectrum, the peak power spectral density of the resonance peak value is calculated; In the frequency response spectrum, the resonance peak value appearing at a specific frequency caused by the physical vibration conduction of adjacent hard disks is identified; The specific frequency includes the fundamental frequency corresponding to the spindle speed of the adjacent hard disk and its one or more harmonic frequencies; Based on the quantitative characteristics of the resonance peak value, the jth hard disk vibration resonance degradation degree is constructed; the peak power spectral density is compared with a preset servo error baseline noise power spectral density under no external vibration interference, and the ratio of the two is calculated, which is defined as the vibration resonance degradation degree. 8.The hard disk failure trend prediction method based on the Internet of Things storage node cluster of claim 7, wherein, The third risk threshold and the fourth risk threshold are preset, and the fourth risk threshold is greater than the third risk threshold; The vibration resonance degradation degree is compared with the third risk threshold and the fourth risk threshold to obtain the second classification result, including: When the vibration resonance degradation degree is lower than the third risk threshold, the vibration influence level of the current hard disk is determined as the first slight influence; When the vibration resonance degradation degree is not lower than the third risk threshold but lower than the fourth risk threshold, the vibration influence level of the current hard disk is determined as the second moderate influence, and a third risk level is generated; When the vibration resonance degradation degree is not lower than the fourth risk threshold, the vibration influence level of the current hard disk is determined as the third serious influence, and a fourth risk level is generated; For the hard disk determined as the third risk level, a first workload scheduling optimization instruction is automatically triggered, which is used to: 10%-15% of the I / O load allocated to the hard disk is rescheduled to the remaining hard disks in the first hierarchical neighborhood set whose second evaluation value is qualified; For the hard disk judged as the fourth risk level, a first priority alarm is generated, and a first data migration instruction is triggered, including: migrating the preset key data on the current hard disk to the hard disks in the third level neighborhood set whose second evaluation values are qualified, and marking the current hard disk as to be offline replaced. 9.The hard disk failure trend prediction method based on the Internet of Things storage node cluster of claim 1, wherein, During the execution of the preset random read-write task on each hard disk, the average seek time is periodically collected to form a time sequence of the average seek time; The dispersion and fluctuation trend of the time sequence of the average seek time relative to the reference seek time are analyzed; Based on the analysis result, a seek performance decay coefficient for characterizing the aging and response lag of the mechanical arm is constructed; The seek performance decay coefficient is obtained in the following manner: The factory calibration seek time of the hard disk or its historical average seek time in the early health state is obtained as the reference seek time; the statistical standard deviation of the time sequence of the average seek time relative to the reference seek time is calculated to quantify the dispersion; The moving average or linear regression analysis method is applied to the time sequence of the average seek time to calculate the change slope thereof with the increase of time, to quantify the fluctuation trend; the normalized statistical standard deviation and change slope are weighted and summed to obtain the seek performance decay coefficient; The seek performance decay coefficient is classified to obtain a third classification result, including: A fifth risk threshold and a sixth risk threshold are preset, wherein the sixth risk threshold is greater than the fifth risk threshold; When the seek performance decay coefficient is lower than the fifth risk threshold, the seek performance state of the current hard disk is determined as a first slight decay; When the seek performance decay coefficient is not lower than the fifth risk threshold but lower than the sixth risk threshold, the seek performance state of the current hard disk is determined as a second moderate decay, and a fifth risk level is generated; When the seek performance decay coefficient is not lower than the sixth risk threshold, the seek performance state of the current hard disk is determined as a third severe decay, and a sixth risk level is generated; For the hard disk judged as the first slight decay state, baseline recording and high-frequency trend tracking of the seek performance of the current hard disk are started; For the hard disk judged as the fifth risk level, a second workload scheduling optimization instruction is automatically triggered, including: 16%-30% of the I / O load allocated to the current hard disk is rescheduled to the remaining hard disks in the first level neighborhood set whose second evaluation values are qualified; For the hard disk judged as the sixth risk level, a second priority alarm is generated, and a second data migration instruction is triggered, including: all data on the hard disk is migrated to the hard disks in the second level neighborhood set whose second evaluation values are qualified, and the current hard disk is marked as to be actively replaced. 10.The hard disk failure trend prediction method based on the Internet of Things storage node cluster of claim 9, wherein, After the first workload scheduling optimization instruction, the second workload scheduling optimization instruction, the first data migration instruction and the second data migration instruction are completed, the steps S1-S4 are repeated for iteration until the second evaluation values of all hard disks are qualified.

Images