A method, device, equipment, medium and product for predicting disk failure

Abstract

The application provides a disk failure prediction method and device, the method comprising: determining a to-be-detected disk, obtaining static parameters of the to-be-detected disk, and first historical state data of the to-be-detected disk in a first preset time period and second historical state data of the to-be-detected disk in a second preset time period, the first preset time period being located in the second preset time period; inputting the first historical state data, the second historical state data, and the static parameters into a long-short-term failure prediction model to output a failure prediction result of the to-be-detected disk; wherein the long-short-term failure prediction model adopts an architecture of parallel processing of a first branch and a second branch, the first branch adopts a structure of DANet combined with BiGRU and is used for processing the first historical state data, and the second branch adopts a structure of ACNet combined with BiGRU and is used for processing the second historical state data. Thus, the accuracy and reliability of disk failure prediction can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, medium and product for predicting disk failures. Background Technology

[0002] With the advent of the big data era, the total amount of global data is rapidly increasing, and large data centers are playing an increasingly important role in daily life. Against this backdrop, hard disks, as the primary carrier of information storage, play an indispensable role. However, despite the high security of individual hard disks, disk failures still occur frequently in large data centers with tens of thousands of disks. These failures not only threaten the reliability and availability of data center systems but can also lead to the loss of important data, causing huge losses to enterprises and users.

[0003] Traditional passive fault tolerance mechanisms typically rely on redundancy techniques such as replication or erasure coding to handle disk failures. However, this undoubtedly increases the system's construction cost and reduces its service capacity when handling failures. Therefore, proactive fault tolerance mechanisms—that is, actively predicting disk failures—are particularly important. By providing early warnings of disk failures and migrating dangerous data in a timely manner, not only can backup redundancy overhead be reduced, but the risk of data loss can also be effectively mitigated, significantly improving the performance of the storage system. Utilizing statistical or machine learning methods to build predictive models based on disk SMART (Self-Monitoring, Analysis and Reporting Technology) data provides new possibilities for fault prediction.

[0004] However, in practical applications, the scarcity of faulty samples is a pressing technical problem that needs to be addressed. Compared to healthy disks, the proportion of faulty disks is extremely low, resulting in training data failing to cover all possible fault types and patterns. Furthermore, minor noise in the data (i.e., randomly occurring faults with no actual causal relationship to the indicators) is often amplified along with valid faulty samples during class balancing, leading to an increased false positive rate in the model. Adding to the complexity, different brands of disks exhibit significant differences in their SMART performance metrics; the same SMART metric may represent different meanings across different brands. This results in huge variations in the prediction accuracy and recall rates of neural network models for different brands of disks, reducing the accuracy and reliability of disk predictions.

[0005] Therefore, improving the accuracy and reliability of disk failure prediction has become an urgent technical problem to be solved. Summary of the Invention

[0006] This application provides a method, apparatus, device, medium, and product for predicting disk failures, in order to solve the technical problem of how to improve the accuracy and reliability of disk failure prediction in related technologies.

[0007] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a method for predicting disk failures, the method comprising: The disk to be tested is determined, and the static parameters of the disk to be tested are obtained, as well as the first historical state data of the disk to be tested within a first preset time period and the second historical state data within a second preset time period, wherein the first preset time period is located within the second preset time period. The first historical state data, the second historical state data, and the static parameters are input into the long-term and short-term fault prediction model, and the fault prediction result of the disk to be detected is output. The long-term and short-term fault prediction model adopts an architecture that processes the first branch and the second branch in parallel. The first branch uses a dual attention network (DANet) combined with a bidirectional gated recurrent unit (BiGRU) to process the first historical state data. The second branch uses an asymmetric convolutional network (ACNet) combined with the BiGRU to process the second historical state data.

[0008] Optionally, both the first historical status data and the second historical status data include: SMART self-monitoring, analysis and reporting technology data and input / output (I / O) related indicator data; inputting the first historical status data, the second historical status data, and the static parameters into the long-term and short-term fault prediction model, and outputting the fault prediction result of the disk to be tested includes: The IO-related indicator data is preprocessed to obtain preprocessed IO-related indicator data; In the SMART data, SMART data with a correlation greater than a first preset threshold to the fault of the disk to be tested are filtered out. Missing values ​​and outliers in the SMART data with a correlation greater than the first preset threshold to the fault of the disk to be tested are processed to obtain processed SMART data. Calculate the statistical characteristics of the processed SMART data, and determine the statistical characteristics, the processed SMART data, and the preprocessed IO-related index data as standardized historical state data. The standardized historical state data is divided into standardized first historical state data and standardized second historical state data. The standardized first historical state data, the standardized second historical state data, and the static parameters are input into the long-term and short-term fault prediction model, and the fault prediction result of the disk to be tested is output.

[0009] Optionally, before inputting the standardized first historical state data, the standardized second historical state data, and the static parameters into the long-term and short-term fault prediction model and outputting the fault prediction result of the disk to be tested, the method further includes: Obtain the cumulative power-on time and power-on cycle count of the disk to be tested; The disk to be tested is determined to be in a healthy state if it meets any of the following conditions: The cumulative energizing time is less than the second preset threshold; The power-on cycle count value is less than a third preset threshold; The cumulative power-on time is less than the second preset threshold, and the power-on cycle count is less than the third preset threshold.

[0010] Optionally, the DANet includes a position attention module and a channel attention module; The DANet is used to input the standardized first historical state data into the position attention module, calculate the spatial similarity matrix between any two positions in the feature map of the standardized first historical state data through a self-attention mechanism, and perform weighted aggregation on the features in the feature map based on the spatial similarity matrix to obtain spatially enhanced features. The DANet is further used to input the feature map into the channel attention module, calculate the dependency relationship between different channels through the self-attention mechanism, weight the features of different channels based on the dependency relationship to obtain channel enhancement features, and fuse the spatial enhancement features and the channel enhancement features to obtain the first output feature map.

[0011] Optionally, the ACNet is used to extract features from the input sample historical state data using 1×3 and 3×1 asymmetric convolutional kernels during the training phase of the long-term and short-term fault prediction model, and to re-integrate the 1×3 and 3×1 asymmetric convolutional kernels into a 3×3 convolutional kernel during the inference phase of the long-term and short-term fault prediction model to extract features from the standardized second historical state data and the static parameters to obtain primary features. The ACNet is also used to input the primary features into the binary neural tree structure module, analyze the primary features through a routing function to obtain the analysis results, route the primary features to the corresponding expert sub-network based on the analysis results, and obtain the second output feature map after processing by the corresponding expert sub-network.

[0012] Optionally, the long-term and short-term fault prediction model further includes a scale-transformed convolutional layer, and the BiGRU includes a forward recurrent neural network GRU and a backward GRU. The scale-transformation convolutional layer is used to process the second output feature map to obtain a third output feature map, the scale of which is the same as that of the first output feature map. The long-term and short-term fault prediction model is further used to concatenate the third output feature map with the first output feature map to obtain a fourth output feature map, and input the fourth output feature map into the BiGRU. The forward GRU and the backward GRU respectively capture the temporal dependencies of the features in the fourth output feature map to obtain the corresponding hidden states. The hidden states of the forward GRU and the hidden states of the backward GRU are concatenated to obtain the final hidden state. After the final hidden state is processed by the hidden layer, the random deactivation layer and the fully connected layer, the fault prediction result of the disk to be detected is output.

[0013] Optionally, the fully connected layer is used to perform matrix multiplication of its own feature matrix with a preset set of post-domain weight parameters to amplify the weights of the feature dimensions in the feature matrix that are associated with preset key SMART indicators and key IO indicators.

[0014] Optionally, the activation function of the long-short-term fault prediction model is the Leak ReLU activation function; the loss function of the long-short-term fault prediction model is the log-hyperbolic cosine loss function.

[0015] Secondly, embodiments of this application provide a disk failure prediction device, the device comprising: The determination module is used to determine the disk to be tested, obtain the static parameters of the disk to be tested, and the first historical state data of the disk to be tested within a first preset time period and the second historical state data within a second preset time period, wherein the first preset time period is located within the second preset time period. The execution module is used to input the first historical state data, the second historical state data and the static parameters into the long-term and short-term fault prediction model, and output the fault prediction result of the disk to be detected. The long-term and short-term fault prediction model adopts an architecture that processes the first branch and the second branch in parallel. The first branch uses a dual attention network (DANet) combined with a bidirectional gated recurrent unit (BiGRU) to process the first historical state data. The second branch uses an asymmetric convolutional network (ACNet) combined with the BiGRU to process the second historical state data.

[0016] Thirdly, embodiments of this application provide a network device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of the disk failure prediction method as described in the first aspect.

[0017] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the disk failure prediction method as described in the first aspect.

[0018] Fifthly, embodiments of this application provide a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the disk failure prediction method as described in the first aspect.

[0019] In this embodiment, fault prediction can be performed by identifying the disk to be tested and obtaining its static parameters and historical state data over different time periods, using a long-term and short-term fault prediction model. In this model, the parallel processing architecture of the first and second branches enables the model to fully utilize the data features of different time periods (long-term and short-term). The first branch utilizes a dual-attention network (DANet) combined with a bidirectional gated recurrent unit (BiGRU) to focus on the rapid changes in short-term data, effectively capturing important patterns and time dependencies in the first historical state data, thereby enhancing sensitivity to short-term faults. Simultaneously, the second branch employs an asymmetric convolutional network (ACNet) combined with BiGRU to focus on the long-term trends and periodic changes in the second historical state data, aiming to extract long-term features from the second historical state data and ensure good identification capabilities for fault patterns over a wider time range. Through the synergistic effect of this dual network structure, the model can more comprehensively understand the operating state of the disk to be tested, improving the accuracy and reliability of fault prediction.

[0020] In summary, long-term and short-term fault prediction models can more effectively capture potential fault signals when processing data from different time periods, thereby significantly improving the accuracy and reliability of fault prediction. Attached Figure Description

[0021] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a disk failure prediction method provided in this application embodiment; Figure 2A structural block diagram of a long-term and short-term fault prediction model provided in this application embodiment; Figure 3 A structural block diagram of a disk failure prediction device provided in an embodiment of this application; Figure 4 This is a structural block diagram of a network device provided in an embodiment of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Figure 1 This application illustrates a disk failure prediction method according to an embodiment of the present application, such as... Figure 1 As shown, the method includes: Step S101: Determine the disk to be tested, obtain the static parameters of the disk to be tested, and the first historical state data of the disk to be tested within a first preset time period and the second historical state data within a second preset time period. The first preset time period is located within the second preset time period; Step S102: Input the first historical state data, the second historical state data, and the static parameters into the long-term and short-term fault prediction model, and output the fault prediction result of the disk to be tested. The long-term and short-term fault prediction model adopts an architecture that processes the first branch and the second branch in parallel. The first branch uses a dual attention network DANet combined with a bidirectional gated recurrent unit BiGRU to process the first historical state data, and the second branch uses an asymmetric convolutional network ACNet combined with BiGRU to process the second historical state data.

[0024] It should be noted that, Figure 1The illustrated scheme describes a method for predicting disk failures. First, the disk to be tested is identified, and its static parameters and two types of historical state data are obtained: one type is the first set of historical state data within a shorter, first preset time period (e.g., 7 days), and the other type is the second set of historical state data within a longer, second preset time period (e.g., 15 days). Then, these static parameters, the first historical state data, and the second historical state data are input into a specially designed long-term / short-term failure prediction model, which ultimately outputs the disk failure prediction result. The core of this long-term / short-term failure prediction model lies in its architecture of parallel processing of the first and second branches. The first branch is specifically designed to process the first historical state data reflecting recent state changes. It combines DANet (Dual Attention Network) and BiGRU (Bidirectional Gated Recurrent Unit) to precisely capture key features and temporal patterns in short-term data. Meanwhile, the second branch processes second-historical state data containing long-term degradation trends. Its structure consists of ACNet (Asymmetric Convolutional Network) and BiGRU, focusing on extracting robust failure symptoms from a broader time span. Through this parallel two-branch design, the model can synergistically utilize subtle anomalies in recent disk operation and long-term accumulated health degradation information, thereby achieving a more comprehensive and accurate assessment of disk failure risk.

[0025] Therefore, the long-term and short-term fault prediction model achieves multi-dimensional and multi-time-scale in-depth analysis of disk health status by tailoring different feature extraction components for data at different time scales (short-term data and long-term data) (the first branch uses DANet combined with BiGRU, and the second branch uses ACNet combined with BiGRU), aiming to improve the accuracy and reliability of fault prediction.

[0026] In one possible implementation, both the first and second historical state data include: SMART (Self-Monitoring, Analysis, and Reporting Technology) data and input / output (IO) related indicator data. Inputting the first and second historical state data, along with static parameters, into a short-term and long-term fault prediction model to output the fault prediction result for the disk under test includes: preprocessing the IO-related indicator data to obtain preprocessed IO-related indicator data; filtering out SMART data whose correlation with the fault of the disk under test is greater than a first preset threshold from the SMART data, processing missing and outlier values ​​in the SMART data with a correlation greater than the first preset threshold to obtain processed SMART data; calculating the statistical characteristics of the processed SMART data, and determining the statistical characteristics, processed SMART data, and preprocessed IO-related indicator data as standardized historical state data, wherein the standardized historical state data is divided into standardized first historical state data and standardized second historical state data; inputting the standardized first historical state data, standardized second historical state data, and static parameters into the short-term and long-term fault prediction model to output the fault prediction result for the disk under test.

[0027] It's important to note that IO-related metrics refer to performance and behavioral indicators related to disk input / output operations. They directly reflect the real-time performance and stress state of the disk under actual workloads. Specifically, these metrics typically include, but are not limited to, the following types: read / write throughput, read / write latency / response time, utilization, etc.

[0028] SMART data is the disk's internal self-monitoring data, which mainly records the health parameters of the disk hardware (such as the number of bad sectors, head flight altitude, and reallocation sector count). The SMART data differs between different types of disks, such as HDDs (Hard Disk Drives) and SSDs (Solid State Drives).

[0029] Static parameters are basic disk information, such as IP (Internet Protocol) information, location information, interface type, brand / manufacturer information, model information, etc.

[0030] It should be noted that the above possible implementation methods detail the process of standardizing and preprocessing the original historical state data before inputting the data into the long-term and short-term fault prediction model. Both the first and second historical state data contain SMART data and IO-related indicator data. The preprocessing process first performs specific preprocessing on the input and output IO-related indicator data to obtain preprocessed IO-related indicator data. For SMART data, specific indicators with a correlation greater than a first preset threshold to the disk fault to be detected (i.e., indicators strongly correlated with disk faults) are first selected. Then, missing and outlier values ​​in this part of the SMART data are processed (e.g., zero-padding), resulting in processed SMART data. Based on this, the statistical characteristics (mean, variance, etc.) of the processed SMART data are further calculated, and these statistical characteristics, the processed SMART data, and the preprocessed IO-related indicator data are integrated to determine unified standardized historical state data. This standardized historical state data is thus divided into standardized first historical state data and standardized second historical state data. Finally, the preprocessed standardized first historical state data and standardized second historical state data, which conform to the input format of the long-term and short-term fault prediction model, along with the static parameters, are input into the long-term and short-term fault prediction model to obtain the final fault prediction results.

[0031] This article will use a specific application scenario as an example to introduce the process of collecting and preprocessing historical disk status data.

[0032] First, data collection is performed, specifically gathering SMART data from the disk, including low-level data read error rate, spindle start-up time, remapped sector count, seek error rate, serial port slowdown error count, temperature, offline uncorrectable sector count, and Ultra ATA (a parallel data transfer interface standard for connecting computer motherboards and storage devices such as hard drives) access verification error rate. Simultaneously, basic disk information is recorded, such as IP address, location information, interface type, brand / manufacturer information, and model information.

[0033] Next, feature selection is performed, choosing SMART parameters highly correlated with disk failures as features, such as SMART 5 (remapped sector count), SMART 9 (power-on time), SMART 187 (reported uncorrected errors), SMART 194 (temperature), SMART 197 (current number of sectors to be mapped), and SMART 198 (number of offline uncorrected sectors). For each feature, its statistical characteristics, such as mean and variance, are calculated to more comprehensively reflect the disk's condition.

[0034] It should also be noted that the features to be selected are different for different types of disks. The following comparison will be made using SSDs and HDDs as examples.

[0035] Based on the SMART performance metrics of SSD disks, indicators (dynamic data) that are strongly correlated with disk failures were selected, as shown in Table 1 below.

[0036] Table 1 5_RAW_VALUE Remapped sector count / Decommissioned block count Health status Disk health status 9_raw_value Hard drive power-on time 1_raw_value underlying data read error rate 12_raw_value Equipment switch count 188_raw_value Communication timeout 190_raw_value airflow temperature 194_raw_value temperature 195_raw_value Hardware ECC calibration 198_raw_value Offline uncorrectable sector count 235_raw_value Good block count 241_raw_value Write remaining lifetime 242_raw_value Total LBA reads 251_raw_value Minimum number of spare blocks remaining 233_raw_value Wear rate 177_raw_value Wear rate 202_raw_value Wear rate 245_raw_value Wear rate Based on the SMART performance metrics of HDD disks, indicators (dynamic data) that are strongly correlated with disk failures were selected, as shown in Table 2 below.

[0037] Table 2 5_RAW_VALUE Remapped sector count / Decommissioned block count 187_RAW_VALUE Uncorrectable error 197_RAW_VALUE Current unmapped sector count Health status Disk health status 1_RAW_VALUE underlying data read error rate 3_RAW_VALUE Disk boot time 4_RAW_VALUE Motor start and stop count 7_RAW_VALUE Seeking error rate 9_RAW_VALUE Hard drive power-on time 12_RAW_VALUE Equipment switch count 190_RAW_VALUE airflow temperature 192_RAW_VALUE Power off, head retracts, counting. 194_RAW_VALUE temperature 231_RAW_VALUE Temperature sensor The static data (static parameters) selected for both SSD and HDD are shown in Table 3 below.

[0038] Table 3 Vendor Manufacturers Product model User Capacity capacity Logical block size Logic block size Rotation Rate rotational speed Serial number Serial Number Device type Equipment type Transport protocol Transmission Protocol Next, the collected SMART data is parsed and converted into a format suitable for model training. Missing and outlier values ​​are filled with zeros. New data types include disk I / O metrics, which are server-level metrics and therefore need to be associated with the disk via IP address. These metrics include disk I / O wait time, disk write IOPS (Input / Output Operations Per Second), disk busy percentage, and server I / O wait time status percentages. Data is then labeled based on whether a disk failure has occurred, serving as a label for model training. For example, short-term predictions use historical data from the 7 days prior to a disk failure, while long-term predictions use historical data from the 15 days prior to a disk failure.

[0039] Finally, the SMART parameters, which are highly correlated with disk failures, along with their variance, mean, disk I / O-related indicators of the new data type, and static parameters, are fused to generate a new set of data, which serves as the input for the long-term and short-term failure prediction model.

[0040] It should be noted that HDDs and SSDs differ in their feature processing, primarily in the selection of SMART metrics. The rules for selecting these metrics are derived from expert experience and statistical analysis of historical data, with the goal of choosing metrics strongly correlated with disk failure. Specifically, the selected metrics differ for HDDs and SSDs; detailed information can be found in Tables 1, 2, and 3.

[0041] Therefore, the preparation of model input data is systematically defined. Its core lies in transforming the raw, potentially noisy, diverse, and heterogeneous data into a high-quality, standardized model input dataset (standardized historical state data) through a series of targeted data cleaning, filtering, and feature enhancement operations. This process aims to improve data quality and consistency, and lay a reliable data foundation for the effective analysis and accurate prediction of subsequent long-term and short-term fault prediction models.

[0042] In one possible implementation, before inputting standardized first historical state data, standardized second historical state data, and static parameters into the long-term and short-term fault prediction model and outputting the fault prediction result of the disk to be tested, the method further includes: obtaining the cumulative power-on time value and power-on cycle count value of the disk to be tested; determining the state of the disk to be tested as healthy if the disk to be tested meets any of the following conditions: the cumulative power-on time value is less than a second preset threshold; the power-on cycle count value is less than a third preset threshold; the cumulative power-on time value is less than the second preset threshold and the power-on cycle count value is less than the third preset threshold.

[0043] It should be noted that this possible implementation describes a pre-judgment mechanism introduced before formally inputting standardized first historical state data, standardized second historical state data, and static parameters into the long-term and short-term fault prediction model for complex calculations. This mechanism first obtains two basic indicators that intuitively reflect the basic wear and tear of the disk—the cumulative power-on time and the power-on cycle count of the disk under test. Then, a simple judgment logic is established: if the disk under test meets either of the following conditions—that the cumulative power-on time is less than a second preset threshold and the power-on cycle count is less than a third preset threshold—or both conditions are met simultaneously, then the disk under test is directly determined to be in a healthy state, thus eliminating the need for further in-depth analysis using the long-term and short-term fault prediction model.

[0044] In specific application scenarios, the system uses two parameters—Power-On Time Count (ID09) and Power Cycle Count (ID12)—to determine whether a disk is healthy (with fewer than 20 power-on cycles). This method can initially screen out 35% of healthy disks, reducing the model prediction time by one-third for scenarios with a large number of disks. The remaining 65% of disk data can be input into the long-term and short-term fault prediction model.

[0045] Therefore, by utilizing two simple and reliable physical lifetime indicators—cumulative power-on time and power-on cycle count—disks that are clearly in the early stages of their lifespan and exhibit extremely low wear can be quickly identified and directly classified as healthy. This significantly optimizes resource allocation, avoids unnecessary complex model calculations for low-risk disks, and thus improves overall detection efficiency.

[0046] In one possible implementation, DANet includes a position attention module and a channel attention module; DANet is used to input standardized first historical state data into the position attention module. It calculates the spatial similarity matrix between any two positions in the feature map of the standardized first historical state data through a self-attention mechanism. Based on the spatial similarity matrix, the features in the feature map are weighted and aggregated to obtain spatially enhanced features. DANet is also used to input feature maps into the channel attention module, calculate the dependencies between different channels through a self-attention mechanism, and weight the features of different channels based on the dependencies to obtain channel-enhanced features; the spatial enhancement features and channel enhancement features are fused to obtain the first output feature map.

[0047] It should be noted that this possible implementation details the specific working mechanism of the DANet used in the first branch of the long-term and short-term fault prediction model. This DANet comprises two core components: a location attention module and a channel attention module. The processing flow is as follows: First, standardized first historical state data is input into the location attention module. This module uses a self-attention mechanism to calculate the spatial similarity matrix between all pairs of points in the data feature map, and then weights and integrates the features at each location within the feature map based on this matrix, thereby generating spatially enhanced features richer in spatial context information. Subsequently, the original feature map is also fed in parallel into the channel attention module. This module also uses a self-attention mechanism to analyze the interdependencies between different feature channels, and weights the features of each channel based on these dependencies, ultimately obtaining channel-enhanced features that highlight key channel information. Finally, the dual attention network fuses the spatially enhanced features and channel-enhanced features generated separately above to form the output of the first branch, i.e., the first output feature map.

[0048] Thus, by simulating two different dimensions of focus (spatial correlation between feature units and dependency between feature channels), the model's ability to capture key information in the data was significantly improved, enabling subsequent models to focus more on feature patterns that are highly correlated with faults, laying a solid foundation for accurate prediction.

[0049] In one possible implementation, ACNet is used to extract features from the input sample historical state data during the training phase of the long-term short-term fault prediction model using 1×3 and 3×1 asymmetric convolutional kernels. During the inference phase of the long-term short-term fault prediction model, the 1×3 and 3×1 asymmetric convolutional kernels are re-fused into a 3×3 convolutional kernel to extract features from the standardized second historical state data and static parameters, resulting in primary features. ACNet is also used to input the primary features into the binary neural tree structure module, analyze the primary features through a routing function, obtain the analysis results, and route the primary features to the corresponding expert subnetwork based on the analysis results. After processing by the corresponding expert subnetwork, the second output feature map is obtained.

[0050] It should be noted that this possible implementation describes the working mechanism of the asymmetric convolutional network ACNet used in the second branch of the long-short-term fault prediction model and its subsequent processing. During the training phase of the long-short-term fault prediction model, this asymmetric convolutional network uses 1×3 and 3×1 asymmetric convolutional kernels to extract features from the input sample historical state data, thereby enhancing the model's feature learning ability. Subsequently, during the inference phase of the long-short-term fault prediction model, it reparameterizes and fuses the 1×3 and 3×1 asymmetric convolutional kernels used in the training phase into a single 3×3 convolutional kernel. This fused convolutional kernel is then used to efficiently extract features from the standardized second historical state data and static parameters, thus obtaining the primary features.

[0051] Subsequently, the primary feature is input into a binary neural tree structure module. This module analyzes the primary feature through its internal routing function and obtains the analysis results. Based on these analysis results, the primary feature is directed (routed) to the corresponding expert subnetwork that specializes in processing specific feature patterns for in-depth processing. Finally, the second output feature map is obtained after processing by the corresponding expert subnetwork.

[0052] Thus, asymmetric convolutional networks enhance the richness of feature extraction without increasing the computational burden of inference through the strategy of "enhancing during training and fusing during inference." Meanwhile, the binary neural tree structure module introduces a conditional computation mechanism that can intelligently assign features to different expert subnetworks for specialized processing, thereby enhancing the model's feature extraction and discrimination capabilities for different long-term degradation patterns. Together, they provide richer feature representations for the final accurate fault prediction.

[0053] In one possible implementation, the long-short-term fault prediction model further includes a scale-transformation convolutional layer. The BiGRU comprises a forward recurrent neural network GRU and a backward GRU. The scale-transformation convolutional layer processes the second output feature map to obtain a third output feature map, the scale of which is the same as that of the first output feature map. The long-short-term fault prediction model also concatenates the third output feature map with the first output feature map to obtain a fourth output feature map. The fourth output feature map is then input into the BiGRU, where the forward GRU and backward GRU capture the temporal dependencies of the features in the fourth output feature map to obtain the corresponding hidden states. The hidden states of the forward GRU and the backward GRU are concatenated to obtain the final hidden state. The final hidden state is then processed through a hidden layer, a random deactivation layer, and a fully connected layer to output the fault prediction result of the disk to be detected.

[0054] It should be noted that the long-short-term fault prediction model includes a scale-transformation convolutional layer, and the BiGRU used in the model is composed of a forward recurrent neural network GRU and a backward recurrent neural network GRU. Specifically, the scale-transformation convolutional layer is responsible for scaling the second output feature map from the second branch, generating a third output feature map with the same scale as the first output feature map generated by the first branch, thus achieving dimensionality matching between feature maps. Subsequently, the long-short-term fault prediction model concatenates and fuses the third output feature map, representing long-term trend information, with the first output feature map, representing short-term sensitive features, to form a fourth output feature map containing more comprehensive information. This fourth output feature map is input into the BiGRU, where its internal forward and backward GRUs capture the temporal dependencies in the feature sequence from both directions, and each outputs its corresponding hidden state. The model then concatenates this hidden state from the forward GRU with the hidden state from the backward GRU again, converging bidirectional temporal context information to obtain the final hidden state. Finally, the final hidden state will be transformed nonlinearly through a regular hidden layer, a random deactivation layer to prevent overfitting, and a fully connected layer to map the results, ultimately outputting the fault prediction result of the disk to be detected.

[0055] Therefore, the scale alignment problem of feature maps of different branches is solved by scale transformation convolutional layers, the long-term and short-term information is effectively integrated by splicing operation, and the key dynamic patterns are extracted from the fused features by BiGRU’s powerful temporal modeling capability, thus ensuring that the model can make accurate fault prediction by comprehensively utilizing long-term and short-term information.

[0056] In one possible implementation, a fully connected layer is used to perform matrix multiplication of its own feature matrix with a preset set of post-domain weight parameters to amplify the weights of the feature dimensions in the feature matrix that are associated with preset key SMART and key IO metrics.

[0057] It's important to note that when a fully connected layer performs its standard linear transformation, its core operation involves matrix multiplication of its own feature matrix with a predefined set of post-domain weight parameters. Through the specific weight distribution set by these parameters, the values ​​corresponding to feature dimensions in the input feature matrix that are associated with predefined key SMART and IO metrics are significantly weighted, thereby greatly enhancing the contribution of these key feature dimensions in the final linear combination.

[0058] Therefore, by introducing the post-domain weight parameter group mechanism, the judgments of operation and maintenance experts on the importance of key SMART indicators and key IO indicators are directly encoded and integrated into the model's automated decision-making process in the form of weight amplification. This guides the model to pay more attention to those strong signal features that are highly correlated with faults and have been verified by domain knowledge. This not only improves the accuracy of model predictions but also enhances the interpretability of the model's decision-making process.

[0059] In one possible implementation, the activation function of the long-short-term fault prediction model is the Leak ReLU activation function of the leakage rectifier linear unit; the loss function of the long-short-term fault prediction model is the log-hyperbolic cosine loss function.

[0060] Based on the above introduction, the structure of the long-term and short-term fault prediction model shown in the embodiments of this application is now described in general, which can be referred to as follows. Figure 2 .

[0061] It can construct long-term and short-term fault prediction models, use convolutional neural networks to extract the local correlation of features, and then capture time-series information through bidirectional gated recurrent units (BiGRU).

[0062] The DANet network structure primarily consists of an attention module and a channel attention module, adaptively integrating local features and global dependencies. The Position Attention Module captures the spatial dependencies between any two locations in the feature map through a self-attention mechanism. Regardless of the distance between two locations, they are correlated if their features are similar. The module calculates the similarity matrix between pixels and aggregates features through weighted summation. The Channel Attention Module captures the dependencies between channels through a self-attention mechanism. It selectively emphasizes interdependent channels by integrating relevant features across all channel mappings. This mechanism enhances responses to specific categories by weighting channel features based on dependencies, highlighting important channel features. This mechanism better captures the feature differences between different brands and models of hard drives, thereby improving the accuracy of fault prediction. The outputs of the Position Attention Module and the Channel Attention Module are then summed to further improve the feature representation. This fusion approach helps to more accurately identify the characteristics of faulty hard drives.

[0063] ACNet significantly improves feature extraction capabilities and model efficiency through a unique convolutional kernel splitting and fusion strategy, combined with an attention mechanism and a binary neural tree structure. During training, the traditional square convolutional kernel (3×3) is split into multiple one-dimensional asymmetric convolutions (1×3 and 3×1). This splitting enhances the model's feature extraction capabilities, capturing richer local and global features. During inference, these asymmetric convolutional kernels are then fused back into the original square convolutional kernels without incurring additional computational overhead. This design retains the advantages of traditional convolutional kernels while improving the model's flexibility and efficiency. Simultaneously, the attention mechanism and binary neural tree structure are combined. Through a tree-like topology, attention convolutions and routing functions are applied to the edges and nodes of the tree to achieve coarse-to-fine feature learning. Different branches focus on different local regions, such as different brands or models of hard drives. In the tree structure, each branch dynamically adjusts the feature extraction weights through an attention mechanism, ensuring the model adaptively focuses on the most representative feature regions. This multi-branch interaction logic enables the model to better handle complex disk failure characteristics and improve prediction accuracy.

[0064] In specific convolutional layers of the model, features are concatenated by channel stacking to further optimize feature representation and provide richer input for subsequent sequence processing modules.

[0065] Specifically, Convolutional Layer 2 outputs a feature map of the same size as the DANet output. Through feature concatenation, the output feature map of DANet is concatenated with the output feature map of Convolutional Layer 2 to form a new feature map, which is then input into the next BiGRU layer. This concatenation method integrates global attention features and local convolutional features, providing a more comprehensive feature representation for the sequence processing module. Convolutional Layer 4 outputs a feature map of the same size as the Convolutional Layer 1 output. Similarly, through feature concatenation, the output feature map of Convolutional Layer 4 is concatenated with the output feature map of Convolutional Layer 1 to form a new feature map, which is then input into the next hidden layer. This feature concatenation strategy further enhances the diversity and discriminative power of features, providing stronger support for subsequent long sequence processing.

[0066] BiGRU is an improved recurrent neural network architecture for processing sequential data. It's a bidirectional extension of the GRU (Gated Recurrent Unit) architecture, capable of processing sequence information simultaneously from both forward and backward directions, thus capturing richer contextual features. BiGRU consists of two independent GRU layers: one processes the sequence forward and the other backward. For the input sequence... X ={ x 1, x 2, …, x T The forward computation formula for BiGRU is: Forward hidden state: ; Backward hidden state: ; The final hidden state h t It is a concatenation of the forward and backward hidden states: .

[0067] BiGRU inherits the properties of GRU in mitigating gradient vanishing and exploding, and is capable of learning long-range dependencies, showing good performance in time series forecasting. In disk failure prediction tasks, this characteristic enables the model to better capture long-term trends and periodic changes in time series data, thereby improving prediction accuracy.

[0068] The main function of hidden layers is to address the long-term dependency problem encountered when processing long sequence data. By introducing three gating mechanisms (forget gate, input gate, and output gate) and a unit state, the retention and forgetting of information are effectively controlled, thus preserving important information in long sequence data. Random deactivation layers are used to temporarily remove a portion of neurons and their connections during training, setting their outputs to 0, thereby reducing dependencies between neurons and preventing the model from becoming overly complex and overfitting. Fully connected layers: In fully connected layers, the connection weights between each input node and output node are initially randomly assigned. During training, these weights are continuously adjusted using the backpropagation algorithm to minimize prediction error. Specifically, the backpropagation algorithm calculates the partial derivative of the loss function with respect to the weights, and then updates the weights based on this partial derivative and the learning rate. This process is repeated until a preset number of iterations is reached or the loss function reaches an acceptable range.

[0069] It is also possible to set post-domain parameters to increase the learning weight of key indicators by multiplying the feature matrix of the fully connected layer by the weighted weight parameters w[w1,w2,w3,......,wn].

[0070] The key HDD metrics are: 5 Reallocated_Sector_Ct (number of sectors that have been remapped), 10 Spin_Retry_Count (number of spindle retry attempts), 187 Reported_Uncorrect (errors that cannot be corrected by hardware ECC), 198 Offline_Uncorrectable Sector Count (total number of uncorrectable errors that occurred while reading or writing sectors), and 197 Current Pending Sector Count (count of currently pending sectors. If an unstable sector is subsequently successfully read or written, that sector will no longer be included in the waiting list, and the data value will decrease).

[0071] The key metrics for SSDs are as follows: Percentage Used Endurance Indicator, Percentage used, and 207 Percentage of SSD Used Life (Hikvision) (representing the percentage of used life).

[0072] Furthermore, the activation function for the long-term and short-term fault prediction model can be the LeakReLU activation function to compensate for the phenomenon that some neurons may permanently die when the learning rate is too large, leading to the inability to update network parameters normally in the later stages. The LeakReLU activation function improves upon ReLU while retaining its advantages, solving the problem of ReLU killing some neurons. The formula is as follows: ; where α typically ranges from 0.01 to 0.2, representing the positive slope of the negative axis. Even with negative inputs, backpropagation is possible.

[0073] The loss function of the long-short-term fault prediction model is the log-hyperbolic cosine (logcosh) loss function. The log-hyperbolic cosine (logcosh) loss function is a smoother loss function than L2, using hyperbolic cosines to calculate the prediction error. Its advantage is that for very small errors, log(cosh(x)) is similar to... The results are very similar, and for large errors, they are very similar to |x|-log(2). The logcosh loss function can have the advantages of MSE without being too affected by outliers. The formula is: Where y is the true value, y p These are predicted values, where n is the sample size and y is the predicted value. i It is the true value of the i-th sample. It is the predicted value of the i-th sample.

[0074] It's important to note that the two models (DAnet and ACnet) process the data in parallel, focusing on short-term and long-term data features respectively. The DAnet model focuses on rapid changes in short-term data, while the ACnet model focuses on long-term trends and cyclical changes. Combining the features of both models' historical data (both short-term and long-term), and fusing the short-term features extracted by the DAnet model with the long-term features extracted by the ACnet model, extracts and integrates information from different branches to form a high-dimensional feature vector, which can improve the accuracy of fault prediction.

[0075] In one possible implementation, the model can be trained using collected labeled data. By adjusting model parameters such as the learning rate and the number of iterations, the model can achieve better performance on the training set. During training, methods such as cross-validation can be used to evaluate and optimize the model.

[0076] Furthermore, for short-term (e.g., within 7 days) fault prediction, a model with DAnet as the backbone network can be used, combined with a BiGRU model. The input is disk status data within a recent period, and the model outputs the probability that the disk will fail within 7 days.

[0077] For long-term fault prediction (e.g., within 15 days), in addition to considering temporal characteristics, it is also necessary to pay attention to the long-term changing trends of features. A combination of CAnet convolutional neural networks and BiGRU models can be used to analyze and predict long-term data.

[0078] The model can also be evaluated and optimized. In model evaluation and optimization, metrics such as accuracy, recall, and F1 score can be used to assess model performance. Special attention should be paid to the false positive rate and false negative rate to ensure the model's reliability in practical applications. Based on the evaluation results, the model can be adjusted and optimized. For example, the model's structure and parameters can be adjusted, and the amount of data increased. Online incremental training can be used to continuously update the model to adapt to changes in disk status.

[0079] Finally, the prediction results can be presented to users in a visual manner, including the probability of disk failure and the estimated time of failure. Based on the prediction results, disks that may fail can be maintained or replaced in a timely manner to ensure data security and system stability.

[0080] In this embodiment, fault prediction can be performed by identifying the disk to be tested and obtaining its static parameters and historical state data over different time periods, using a long-term and short-term fault prediction model. In this model, the parallel processing architecture of the first and second branches enables the model to fully utilize the data features of different time periods (long-term and short-term). The first branch utilizes a dual-attention network (DANet) combined with a bidirectional gated recurrent unit (BiGRU) to focus on the rapid changes in short-term data, effectively capturing important patterns and time dependencies in the first historical state data, thereby enhancing sensitivity to short-term faults. Simultaneously, the second branch employs an asymmetric convolutional network (ACNet) combined with BiGRU to focus on the long-term trends and periodic changes in the second historical state data, aiming to extract long-term features from the second historical state data and ensure good identification capabilities for fault patterns over a wider time range. Through the synergistic effect of this dual network structure, the model can more comprehensively understand the operating state of the disk to be tested, improving the accuracy and reliability of fault prediction.

[0081] In summary, long-term and short-term fault prediction models can more effectively capture potential fault signals when processing data from different time periods, thereby significantly improving the accuracy and reliability of fault prediction.

[0082] Figure 3 This application illustrates a disk failure prediction apparatus according to an embodiment of the present application, such as... Figure 3 As shown, device 30 includes: The determination module 301 is used to determine the disk to be tested, obtain the static parameters of the disk to be tested, and the first historical state data of the disk to be tested within a first preset time period and the second historical state data within a second preset time period, wherein the first preset time period is located within the second preset time period. The execution module 302 is used to input the first historical state data, the second historical state data and static parameters into the long-term and short-term fault prediction model, and output the fault prediction result of the disk to be detected. The long-term and short-term fault prediction model adopts an architecture that processes the first branch and the second branch in parallel. The first branch uses a dual attention network DANet combined with a bidirectional gated recurrent unit BiGRU to process the first historical state data, and the second branch uses an asymmetric convolutional network ACNet combined with BiGRU to process the second historical state data.

[0083] In one possible implementation, both the first historical state data and the second historical state data include: SMART data (Self-Monitoring, Analysis and Reporting Technology) and input / output (IO) related indicator data; the execution module 302 is also used to preprocess the IO related indicator data to obtain preprocessed IO related indicator data. In the SMART data, SMART data with a correlation greater than the first preset threshold to the fault of the disk to be tested is filtered out. Missing values ​​and outliers in the SMART data with a correlation greater than the first preset threshold to the fault of the disk to be tested are processed to obtain the processed SMART data. The statistical characteristics of the processed SMART data are calculated, and the statistical characteristics, processed SMART data, and preprocessed IO-related index data are determined as standardized historical state data. The standardized historical state data is divided into standardized first historical state data and standardized second historical state data. The standardized first historical state data, standardized second historical state data, and static parameters are input into the long-term and short-term fault prediction model, and the fault prediction results of the disk to be tested are output.

[0084] In one possible implementation, the execution module 302 is further configured to obtain the cumulative power-on time value and power-on cycle count value of the disk to be tested before inputting the standardized first historical state data, the standardized second historical state data and the static parameters into the long-term and short-term fault prediction model and outputting the fault prediction result of the disk to be tested. The disk to be tested is determined to be in a healthy state if it meets any of the following conditions: The cumulative power-on time is less than the second preset threshold; The power-on cycle count value is less than the third preset threshold; The cumulative power-on time is less than the second preset threshold, and the power-on cycle count is less than the third preset threshold.

[0085] In one possible implementation, DANet includes a position attention module and a channel attention module; DANet is used to input standardized first historical state data into the position attention module. It calculates the spatial similarity matrix between any two positions in the feature map of the standardized first historical state data through a self-attention mechanism. Based on the spatial similarity matrix, the features in the feature map are weighted and aggregated to obtain spatially enhanced features. DANet is also used to input feature maps into the channel attention module, calculate the dependencies between different channels through a self-attention mechanism, and weight the features of different channels based on the dependencies to obtain channel-enhanced features; the spatial enhancement features and channel enhancement features are fused to obtain the first output feature map.

[0086] In one possible implementation, ACNet is used to extract features from the input sample historical state data using 1×3 and 3×1 asymmetric convolutional kernels during the training phase of the long-short-term fault prediction model. During the inference phase of the long-short-term fault prediction model, the 1×3 and 3×1 asymmetric convolutional kernels are re-fused into a 3×3 convolutional kernel to extract features from the standardized second historical state data and static parameters to obtain primary features. ACNet is also used to input primary features into the binary neural tree structure module, analyze the primary features through the routing function, obtain the analysis results, and route the primary features to the corresponding expert subnetworks based on the analysis results. After processing by the corresponding expert subnetworks, the second output feature map is obtained.

[0087] In one possible implementation, the long-short-term fault prediction model also includes a scale-transformed convolutional layer, and BiGRU includes a forward recurrent neural network GRU and a backward GRU. The scale-transformation convolutional layer is used to process the second output feature map to obtain the third output feature map, and the scale of the third output feature map is the same as that of the first output feature map. The long-term and short-term fault prediction model is also used to concatenate the third output feature map with the first output feature map to obtain the fourth output feature map. The fourth output feature map is then input into a BiGRU, which captures the temporal dependencies of features in the fourth output feature map through forward GRU and backward GRU respectively to obtain the corresponding hidden states. The hidden states of forward GRU and backward GRU are concatenated to obtain the final hidden state. The final hidden state is then processed through a hidden layer, a random deactivation layer, and a fully connected layer to output the fault prediction result of the disk to be detected.

[0088] In one possible implementation, a fully connected layer is used to perform matrix multiplication of its own feature matrix with a preset set of post-domain weight parameters to amplify the weights of the feature dimensions in the feature matrix that are associated with preset key SMART and key IO metrics.

[0089] In one possible implementation, the activation function of the long-short-term fault prediction model is the Leak ReLU activation function of the leakage rectifier linear unit; the loss function of the long-short-term fault prediction model is the log-hyperbolic cosine loss function.

[0090] In summary, the scheme shown in this application constructs a unified network with "DAnet+CAnet+BiGRU" as its core: DAnet's channel-temporal dual attention simultaneously mines local subtle patterns of HDD mechanical wear and SSD flash memory failure within a 7-day window and outputs weighted spatial features; ACNet, through a unique convolutional kernel splitting and fusion strategy, combined with an attention mechanism and a binary neural tree structure, mines long-term trends and periodic changes in disk status data within a 15-day window, significantly improving feature extraction capabilities and model efficiency. The features mined by the two models are combined and directly fed into the bidirectional GRU, capturing the entire lifecycle degradation trajectory through forward-backward temporal memory. This allows heterogeneous SMART metrics to be mapped to a shared semantic space, achieving "one set of parameters, HDD / SSD compatibility," and improving disk failure prediction recall.

[0091] The method described in this application constructs a long-term and short-term fault detection model applicable to both HDD and SSD disks. Utilizing the DAnet convolutional neural network to extract local correlation features from disk data, it can capture subtle changes in disk usage patterns, improving feature extraction capabilities. By capturing temporal information using a bidirectional gated recurrent unit (BiGRU), it can better understand changes in disk usage patterns over time, which is very useful for predicting future performance trends. The hybrid model combining DAnet and BiGRU can process both spatial and temporal data simultaneously, resulting in a more comprehensive and accurate understanding and prediction of disk behavior patterns. This model can adapt to different types of disks (HDD and SSD) because they differ in data access patterns and performance characteristics. This adaptability makes the model more flexible and effective in practical applications. By comprehensively considering historical data and temporal information, the model can more accurately predict the future state of the disk and can handle data at different time scales (such as 7-day and 15-day historical data). This helps capture disk behavior patterns across different time ranges, and through feature fusion, it can jointly analyze the features of long-term and short-term changes to determine the disk failure probability, improving prediction accuracy. In addition, by combining various neural network structures, the model exhibits a certain degree of robustness to noise and outlier data, which helps improve the stability and reliability of predictions. In conclusion, constructing a hybrid long-term and short-term model based on CNN and BiGRU can not only improve the accuracy and reliability of disk predictions but also adapt to different types of disks and various application scenarios, demonstrating broad application prospects and technological advantages.

[0092] Furthermore, in practical application scenarios, the method shown in the embodiments of this application has the following advantages: More sensitive to features: The industry-standard single-channel attention CNN (Convolutional Neural Network) is almost unresponsive to fluctuations in SMART data within 2%. In the method shown in the embodiments of this application, DAnet dual attention increases the sensitivity from 0.62 to 0.86, and the AUC (Area Under the Curve) increases by 3.9% with the same number of parameters.

[0093] Longer-term sequential memory: The traditional one-way LSTM (Long Short-Term Memory) network has a seven-day early recall rate of 89%, which is improved to 96% after adding a lightweight BiGRU, with an inference latency of only 1.4ms, and the long-term dependency forgetting problem is solved in one go.

[0094] A single model is compatible with both HDD and SSD: Existing solutions require training two networks for different media. In this application, the embodiment uses "media type embedding vectors" to map heterogeneous attributes to a unified space. The difference between F1 (referring to F1-Score, which is the harmonic mean of precision and recall, and a commonly used indicator for comprehensively evaluating model performance) between HDD and SSD is ≤1%, and the production line deployment cost is reduced by 50%.

[0095] This application provides a network device 40, such as... Figure 4 As shown, the network device 40 includes a processor 401, a memory 402, and a program stored on the memory 402 and executable on the processor 401. When the program is executed by the processor 401, it implements the steps of a disk failure prediction method as shown in the above embodiment.

[0096] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of a disk failure prediction method as shown in the above embodiments, achieving the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0097] This application also provides a computer program product, including computer instructions. When executed by a processor, the computer instructions implement the steps of the disk failure prediction method shown in the above embodiments and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0098] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0099] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0100] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method of predicting disk failure, characterized by, The method includes: The disk to be tested is determined, and the static parameters of the disk to be tested are obtained, as well as the first historical state data of the disk to be tested within a first preset time period and the second historical state data within a second preset time period, wherein the first preset time period is located within the second preset time period. The first historical state data, the second historical state data, and the static parameters are input into the long-term and short-term fault prediction model, and the fault prediction result of the disk to be detected is output. The long-term and short-term fault prediction model adopts an architecture that processes the first branch and the second branch in parallel. The first branch adopts a structure of dual attention network DANet combined with bidirectional gated recurrent unit BiGRU to process the first historical state data. The second branch adopts a structure of asymmetric convolutional network ACNet combined with BiGRU to process the second historical state data. The DANet includes a location attention module and a channel attention module; The DANet is used to input standardized first historical state data into the location attention module, calculate the spatial similarity matrix between any two locations in the feature map of the standardized first historical state data through a self-attention mechanism, and perform weighted aggregation on the features in the feature map based on the spatial similarity matrix to obtain spatially enhanced features. The DANet is further used to input the feature map into the channel attention module, calculate the dependency relationship between different channels through a self-attention mechanism, weight the features of different channels based on the dependency relationship to obtain channel enhancement features, and fuse the spatial enhancement features and the channel enhancement features to obtain a first output feature map. The ACNet is used to extract features from the input sample historical state data using 1×3 and 3×1 asymmetric convolution kernels during the training phase of the long-term and short-term fault prediction model. During the inference phase of the long-term and short-term fault prediction model, the 1×3 and 3×1 asymmetric convolution kernels are re-fused into a 3×3 convolution kernel to extract features from the standardized second historical state data and the static parameters to obtain primary features. The ACNet is also used to input the primary features into the binary neural tree structure module, analyze the primary features through a routing function to obtain the analysis results, route the primary features to the corresponding expert sub-network according to the analysis results, and obtain the second output feature map after processing by the corresponding expert sub-network. Wherein, the standardized first historical state data is the data after standardizing the first historical state data, and the standardized second historical state data is the data after standardizing the second historical state data.

2. The method of claim 1, wherein, Both the first historical status data and the second historical status data include: SMART self-monitoring, analysis and reporting technology data and input / output (I / O) related indicator data; inputting the first historical status data, the second historical status data, and the static parameters into the long-term and short-term fault prediction model, the output fault prediction results of the disk to be tested include: The IO-related indicator data is preprocessed to obtain preprocessed IO-related indicator data; In the SMART data, SMART data with a correlation greater than a first preset threshold to the fault of the disk to be tested are filtered out. Missing values ​​and outliers in the SMART data with a correlation greater than the first preset threshold to the fault of the disk to be tested are processed to obtain processed SMART data. Calculate the statistical characteristics of the processed SMART data, and determine the statistical characteristics, the processed SMART data, and the preprocessed IO-related index data as standardized historical state data. The standardized historical state data is divided into standardized first historical state data and standardized second historical state data. The standardized first historical state data, the standardized second historical state data, and the static parameters are input into the long-term and short-term fault prediction model, and the fault prediction result of the disk to be tested is output.

3. The method of claim 2, wherein, Before inputting the standardized first historical state data, the standardized second historical state data, and the static parameters into the long-term and short-term fault prediction model and outputting the fault prediction result of the disk to be tested, the method further includes: Obtain the cumulative power-on time and power-on cycle count of the disk to be tested; The disk to be tested is determined to be in a healthy state if it meets any of the following conditions: The cumulative energizing time is less than the second preset threshold; The power-on cycle count value is less than a third preset threshold; The cumulative power-on time is less than the second preset threshold, and the power-on cycle count is less than the third preset threshold.

4. The method of claim 1, wherein, The long-term and short-term fault prediction model also includes a scale-transformation convolutional layer, and the BiGRU includes a forward GRU and a backward GRU; The scale-transformation convolutional layer is used to process the second output feature map to obtain a third output feature map, the scale of which is the same as that of the first output feature map. The long-term and short-term fault prediction model is further used to concatenate the third output feature map with the first output feature map to obtain a fourth output feature map, and input the fourth output feature map into the BiGRU, which captures the temporal dependencies of the features in the fourth output feature map through the forward GRU and the backward GRU respectively to obtain the corresponding hidden state; the hidden state of the forward GRU and the hidden state of the backward GRU are concatenated to obtain the final hidden state; After the final hidden state is processed through a hidden layer, a random deactivation layer, and a fully connected layer, the fault prediction result of the disk to be detected is output.

5. The method of claim 4, wherein, The fully connected layer is used to perform matrix multiplication of its own feature matrix with a preset set of post-domain weight parameters to amplify the weights of the feature dimensions in the feature matrix that are associated with preset key SMART indicators and key IO indicators.

6. The method according to any one of claims 1-5, characterized in that, The activation function of the long-term and short-term fault prediction model is the Leak ReLU activation function of the leakage rectifier linear unit; the loss function of the long-term and short-term fault prediction model is the log-hyperbolic cosine loss function.

7. A disk failure prediction apparatus characterized by comprising: The device includes: The determination module is used to determine the disk to be tested, obtain the static parameters of the disk to be tested, and the first historical state data of the disk to be tested within a first preset time period and the second historical state data within a second preset time period, wherein the first preset time period is located within the second preset time period. The execution module is used to input the first historical state data, the second historical state data and the static parameters into the long-term and short-term fault prediction model, and output the fault prediction result of the disk to be detected. The long-term and short-term fault prediction model adopts an architecture that processes the first branch and the second branch in parallel. The first branch adopts a structure of dual attention network DANet combined with bidirectional gated recurrent unit BiGRU to process the first historical state data. The second branch adopts a structure of asymmetric convolutional network ACNet combined with BiGRU to process the second historical state data. The DANet includes a location attention module and a channel attention module; The DANet is used to input standardized first historical state data into the location attention module, calculate the spatial similarity matrix between any two locations in the feature map of the standardized first historical state data through a self-attention mechanism, and perform weighted aggregation on the features in the feature map based on the spatial similarity matrix to obtain spatially enhanced features. The DANet is further used to input the feature map into the channel attention module, calculate the dependency relationship between different channels through a self-attention mechanism, weight the features of different channels based on the dependency relationship to obtain channel enhancement features, and fuse the spatial enhancement features and the channel enhancement features to obtain a first output feature map. The ACNet is used to extract features from the input sample historical state data using 1×3 and 3×1 asymmetric convolution kernels during the training phase of the long-term and short-term fault prediction model. During the inference phase of the long-term and short-term fault prediction model, the 1×3 and 3×1 asymmetric convolution kernels are re-fused into a 3×3 convolution kernel to extract features from the standardized second historical state data and the static parameters to obtain primary features. The ACNet is also used to input the primary features into the binary neural tree structure module, analyze the primary features through a routing function to obtain the analysis results, route the primary features to the corresponding expert sub-network according to the analysis results, and obtain the second output feature map after processing by the corresponding expert sub-network. Wherein, the standardized first historical state data is the data after standardizing the first historical state data, and the standardized second historical state data is the data after standardizing the second historical state data.

8. A network device, comprising: include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of a disk failure prediction method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of a disk failure prediction method as described in any one of claims 1 to 6.

10. A computer program product, characterised in that, It includes computer instructions that, when executed by a processor, implement the steps of a disk failure prediction method as described in any one of claims 1 to 6.

Images