Mechanical hard disk health assessment method, system and storage medium based on evolutionary graph

Abstract

The present invention relates to a mechanical hard disk health assessment method, system and storage medium based on an evolutionary graph. The method comprises: obtaining original data sets of multiple mechanical hard disks based on their full-life SMART data and corresponding full-life health labels; obtaining an importance ranking of the mechanical hard disk's SMART attributes based on the original data sets, obtaining target SMART attributes that meet preset ranking requirements based on the importance ranking, obtaining target SMART data and corresponding target data sets for all target SMART attributes; obtaining a mechanical hard disk health assessment model by self-learning the evolutionary correlation between target SMART attributes based on a multi-scale evolutionary graph network, target SMART attributes, target SMART data and target data sets; and evaluating the health of the mechanical hard disk to be tested based on the mechanical hard disk health assessment model. The present invention can make the mechanical hard disk health assessment accurate and reliable.

Description

Technical Field

[0001] The present invention relates to the field of data storage technology, and in particular to a mechanical hard disk health assessment method, system and storage medium based on an evolutionary graph. Background Art

[0002] The era of massive data has arrived, and data storage has shifted from local storage to the cloud. The reliability of large-scale data storage centers cannot be ignored. Mechanical hard drives are a vital component of modern data centers, and their data access stability directly impacts the reliability of data storage systems.

[0003] Currently, the mechanisms for improving the reliability of data storage systems are divided into passive fault tolerance and active fault tolerance. Passive fault tolerance means that when a mechanical hard disk fails, the security of the data is ensured by means of data backup or erasure coding. Since a large amount of data needs to be backed up, the passive fault tolerance mechanism seriously increases the operational burden of the data center. Active fault tolerance uses the various hard disk attribute data provided by Self-Monitoring Analysis and Reporting Technology (SMART) to detect the health status of the mechanical hard disk. Most of the existing mechanical hard disk health assessment methods use artificially set fixed thresholds for fault judgment and early warning. Although this method is simple and direct, it has low accuracy and poor generalization, and cannot quantify the health of the mechanical hard disk, which seriously affects the accuracy of the mechanical hard disk health assessment. Summary of the Invention

[0004] Therefore, the technical problem to be solved by the present invention is that in the related technology, fault judgment and early warning are performed by artificially setting fixed thresholds to achieve mechanical hard disk health assessment, which has low accuracy and poor generalization, and cannot quantify the health of the mechanical hard disk, seriously affecting the accuracy of the assessment.

[0005] To solve the above technical problems, the present invention provides a mechanical hard disk health assessment method based on an evolutionary graph, the method comprising:

[0006] Obtain the original data sets of multiple mechanical hard disks based on their lifecycle SMART data and corresponding lifecycle health labels.

[0007] Obtaining the importance ranking of the SMART attributes of the mechanical hard disk according to the original data set, obtaining the target SMART attributes that meet the preset ranking requirements based on the importance ranking, and obtaining the target SMART data and corresponding target data sets of all target SMART attributes;

[0008] Based on the multi-scale evolutionary graph network, target SMART attributes, target SMART data, and target dataset, the correlation between the evolution of target SMART attributes is self-learned to obtain a health assessment model for mechanical hard drives.

[0009] Evaluate the health of the hard disk to be tested based on the health evaluation model of the hard disk.

[0010] Optionally, obtaining the original data sets of the plurality of mechanical hard disks according to the full-life SMART data and corresponding full-life health labels of the plurality of mechanical hard disks includes:

[0011] Collect the full-life SMART data of multiple mechanical hard disks until they fail, and build the corresponding full-life health labels for multiple mechanical hard disks;

[0012] Combine the full-life SMART data and the full-life health label to obtain the full-life raw data of multiple mechanical hard drives;

[0013] Based on the full lifecycle raw data, the original data sets of multiple mechanical hard drives are constructed.

[0014] Optionally, obtaining an importance ranking of the SMART attributes of the mechanical hard disk according to the original data set, obtaining target SMART attributes that meet preset ranking requirements according to the importance ranking, and obtaining target SMART data and corresponding target data sets for all target SMART attributes include:

[0015] An initial ensemble tree model is constructed, and a training data set as input data of the initial ensemble tree model is obtained based on the original data set; wherein the training data set is the normalized values ​​of each SMART attribute selected from the original data set;

[0016] Construct a regression tree learning objective function, and iteratively train the initial ensemble tree model based on the training data set to obtain a mature ensemble tree model;

[0017] According to the number of times each SMART attribute appears in the mature ensemble tree model, the importance ranking of all SMART attributes of the mechanical hard disk is obtained;

[0018] Sort by importance and select the target SMART attributes that meet the preset sorting requirements from all SMART attributes;

[0019] Normalization processing is performed according to the target SMART attributes to obtain target SMART data, a health label corresponding to the target SMART data is obtained according to the original data set, and the target SMART data and the health label are combined to obtain a target data set with all target SMART attributes.

[0020] Optionally, the self-learning of the evolutionary correlation between target SMART attributes based on the multi-scale evolutionary graph network, target SMART attributes, target SMART data, and target data set to obtain a health assessment model for a mechanical hard disk includes:

[0021] Construct a multi-scale evolutionary graph network, combine target SMART attributes, target SMART data and target dataset, and self-learn the evolutionary correlation between target SMART attributes to extract multi-scale temporal representation of target SMART data;

[0022] The multi-scale evolutionary graph network is iteratively updated based on the multi-scale time series representation of the target SMART data to obtain the health assessment model of the mechanical hard disk.

[0023] Optionally, constructing a multi-scale evolutionary graph network, combining target SMART attributes, target SMART data, and target dataset, and self-learning the evolutionary correlation between target SMART attributes to extract a multi-scale temporal representation of the target SMART data, includes:

[0024] Construct a multi-scale evolutionary graph network, which includes a temporal convolution module, an evolutionary graph self-learning module, and an evolutionary graph convolution module;

[0025] Input the target SMART data into the temporal convolution module to extract the multi-scale temporal representation of the target SMART data along the time dimension;

[0026] The output data of the temporal convolution module is used as the input data of the evolutionary graph self-learning module, which automatically extracts the graph structure representation of the evolution of each target SMART attribute over time and outputs the evolved adjacency matrix.

[0027] The adjacency matrix output by the evolutionary graph self-learning module and the multi-scale temporal representation output by the temporal convolution module are used as input data of the evolutionary graph convolution module, and a new multi-scale temporal representation that takes into account the correlation between the SMART attributes of each target is obtained.

[0028] Optionally, the temporal convolution module includes a dilated convolution layer and a gating layer;

[0029] The target SMART data is input into the time convolution module, and the multi-scale temporal representation of the target SMART data is extracted along the time dimension, including:

[0030] The target SMART data is divided into multiple time series according to the SMART attributes. The time series of each target SMART attribute is used as the input data of the dilated convolution layer. The dilated convolution operation is performed using convolution kernels of different sizes to obtain time series representations at different time scales.

[0031] The temporal representations of different time scales output by the dilated convolutional layer are used as the input data of the gating layer to obtain the multi-scale temporal representations of each target SMART attribute output by the temporal convolution module.

[0032] Optionally, the evolutionary graph self-learning module includes a graph evolution stage division layer and a graph self-learning layer;

[0033] The output data of the temporal convolution module is used as the input data of the evolutionary graph self-learning module to automatically extract the graph structure representation of the evolution of each target SMART attribute over time and output the evolutionary adjacency matrix, including:

[0034] The multi-scale temporal representation of the output data of the temporal convolution module is divided into several segments along the time dimension through the graph evolution stage partitioning layer, and the temporal representation of each segment is aggregated into one dimension using an aggregation function as the attribute representation of each evolution stage;

[0035] The graph learning layer automatically learns the graph structure representation of each evolutionary stage based on the attribute representation of each evolutionary stage. The graph structure representation of each evolutionary stage is obtained from the graph structure representation of the previous evolutionary stage.

[0036] The graph structure representations of each evolutionary stage are connected pairwise, and the adjacency matrix of each evolutionary stage is derived using a multi-layer perceptron, so as to obtain the adjacency matrix evolved through the output of the evolutionary graph self-learning module.

[0037] Optionally, the adjacency matrix output by the evolutionary graph self-learning module and the multi-scale temporal representation output by the temporal convolution module are used as input data of the evolutionary graph convolution module to calculate and obtain a new multi-scale temporal representation that takes into account the correlation between each target SMART attribute, including:

[0038] The adjacency matrix output by the evolutionary graph self-learning module and the multi-scale temporal representation output by the temporal convolution module are input as input data to the evolutionary graph convolution module;

[0039] Through the evolutionary graph convolution module, the graph structure representation provided by the adjacency matrix is ​​used to perform graph convolution operations on the multi-scale temporal representation of each target SMART attribute, and a new multi-scale temporal representation that takes into account the correlation between each target SMART attribute is obtained.

[0040] In addition, the present invention also proposes a mechanical hard disk health assessment system based on an evolutionary graph, the system comprising:

[0041] The SMART data acquisition module is used to obtain the original data sets of multiple mechanical hard disks based on their full-life SMART data and corresponding full-life health labels;

[0042] A target data acquisition module is used to obtain the importance ranking of the SMART attributes of the mechanical hard disk according to the original data set, obtain the target SMART attributes that meet the preset ranking requirements according to the importance ranking, and obtain the target SMART data and corresponding target data sets of all target SMART attributes;

[0043] The evaluation model acquisition module is used to self-learn the evolutionary correlation between target SMART attributes based on the multi-scale evolutionary graph network, target SMART attributes, target SMART data, and target dataset, and obtain the health evaluation model of the mechanical hard disk;

[0044] The health evaluation module is used to evaluate the health of the mechanical hard disk to be tested according to the health evaluation model of the mechanical hard disk.

[0045] In addition, the present invention also proposes a computer-readable storage medium, which stores computer execution instructions. When the computer execution instructions are executed by a processor, they are used to implement all or part of the method steps of the mechanical hard disk health assessment method based on the evolutionary graph as described above.

[0046] The technical solution provided by the present invention has the following advantages:

[0047] The proposed evolutionary graph-based mechanical hard drive health assessment method models the time-varying correlations between the hard drive's SMART data, simultaneously mining multi-scale temporal representations from the raw data and aggregating these representations based on the interdependencies between the SMART attributes. This method accurately assesses and quantifies the hard drive's health. This method offers the advantages of higher accuracy, improved stability, and greater robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0049] Figure 1 This is a schematic diagram of the steps of the mechanical hard disk health assessment method based on the evolution graph according to an embodiment of the present invention;

[0050] Figure 2 is the attribute importance score value of each SMART attribute in the mechanical hard disk health assessment method based on the evolution graph according to an embodiment of the present invention;

[0051] Figure 3This is a graph showing the health evaluation results of a mechanical hard disk using the evolutionary graph-based mechanical hard disk health evaluation method according to an embodiment of the present invention;

[0052] Figure 4 This is a performance comparison chart of a mechanical hard disk health assessment method based on an evolutionary graph according to an embodiment of the present invention;

[0053] Figure 5 This is a simplified block diagram illustrating the structure of the mechanical hard disk health assessment system based on the evolution graph according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The technical solutions of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments. It should be noted that the embodiments of the present invention and the features therein may be combined with each other unless there is a conflict.

[0055] In related technologies, most mechanical hard drive health assessment methods rely on manually set fixed thresholds for fault diagnosis and early warning. While this method is simple and direct, it suffers from low accuracy, poor generalization, and inability to quantify the health of a mechanical hard drive, severely impacting the accuracy of mechanical hard drive health assessments. To address these technical issues, the present invention proposes a mechanical hard drive health assessment method and system based on an evolutionary graph.

[0056] The present invention provides a mechanical hard disk health assessment method based on an evolutionary graph, which is used to assess the health of a mechanical hard disk. Specifically, Figure 1 As shown, the mechanical hard disk health assessment method based on the evolution graph may include the following steps:

[0057] S100, obtaining original data sets of multiple mechanical hard disks based on the full-life SMART data and corresponding full-life health labels of the multiple mechanical hard disks;

[0058] S200: Obtaining an importance ranking of the SMART attributes of the mechanical hard disk according to the original data set, obtaining target SMART attributes that meet preset ranking requirements according to the importance ranking, and obtaining target SMART data and corresponding target data sets for all target SMART attributes;

[0059] S300, based on the multi-scale evolutionary graph network, target SMART attributes, target SMART data, and target data set, self-learning the evolutionary correlation between each target SMART attribute to obtain a health assessment model for the mechanical hard disk;

[0060] S400: Evaluate the health of the mechanical hard disk to be tested according to a health evaluation model of the mechanical hard disk.

[0061] Based on raw, lifelong SMART data from mechanical hard drives, deep learning methods are used to analyze and mine the patterns inherent in each attribute data, assessing and quantifying the health of the mechanical hard drive, significantly improving the reliability and maintainability of the data storage system. Furthermore, the data of each SMART attribute in SMART technology is interdependent. The observed value of each SMART attribute depends not only on its historical value, but also on the current observed values ​​of other SMART attributes. This interdependence is not static and will gradually change as the health of the mechanical hard drive deteriorates.

[0062] The evolutionary graph-based mechanical hard drive health assessment method provided in this embodiment models the time-varying correlations between the SMART attributes of the mechanical hard drive based on the SMART data. It also mines multi-scale temporal representations in the raw data and aggregates these representations based on the interdependencies between the SMART attributes to accurately assess and quantify the health of the mechanical hard drive. This method has the ability to model the time-varying interdependencies between the SMART attributes, resulting in higher accuracy, greater stability, and greater robustness in mechanical hard drive health assessments.

[0063] Furthermore, in step S100, obtaining the original data sets of the plurality of mechanical hard disks according to the full life SMART data and the corresponding full life health labels of the plurality of mechanical hard disks may specifically include the following steps:

[0064] S110: Collecting SMART data of the entire life of multiple mechanical hard disks until a failure occurs, and constructing health labels corresponding to the entire life of the multiple mechanical hard disks.

[0065] Specifically, the full life SMART data of multiple mechanical hard disks can be collected in advance until they fail. Among them, S t Represents the sampled SMART data of the mechanical hard disk at sampling time t;

[0066] At the same time, the corresponding full-life health label can be constructed Among them C t Indicates the sampling health label at sampling time t; C t The calculation expression is as follows:

[0067] C t =t / L;

[0068] Where t represents the sampling time, i.e., the time from the hard disk being put into use to the time when the SMART data S is sampled. tThe usage time experienced; L is the total usage time of the mechanical hard disk from the time it was put into use to the time it failed, that is, the service life of the mechanical hard disk.

[0069] S120: Combine the full-life SMART data and the full-life health label to obtain the full-life original data of multiple mechanical hard disks.

[0070] The full life SMART data of each mechanical hard disk can be and health labels Combined to obtain the original data of the entire life of these mechanical hard drives

[0071] S130: Constructing original data sets of multiple mechanical hard disks based on the full lifecycle original data.

[0072] The full life data of each mechanical hard disk can be The original dataset D that makes up these mechanical hard drives O ={(S i ,C i )}; where S i Represents the SMART data of the mechanical hard disk at sampling time i; C i Indicates the health label at sampling time i.

[0073] Furthermore, in step S200, the importance ranking of the SMART attributes of the mechanical hard disk is obtained according to the original data set, the target SMART attributes that meet the preset ranking requirements are obtained according to the importance ranking, and the target SMART data and corresponding target data sets of all the target SMART attributes are obtained. Specifically, the following steps may be included:

[0074] S210: construct an initial ensemble tree model, and obtain a training data set as input data of the initial ensemble tree model based on the original data set; wherein the training data set is the normalized values ​​of each SMART attribute selected from the original data set.

[0075] Specifically, the original data set D of the mechanical hard disk can be O ={(S i ,C i )} select the normalized value (Normalized_value) of each SMART attribute as the training data set D X ={(S' i ,C i )},in F indicates the number of SMART attributes provided by the mechanical hard disk model;

[0076] Build the initial ensemble tree model and train the dataset D X ={(S'i ,C i )} as the input data of the initial ensemble tree model; and the mathematical expression of the initial ensemble tree model is as follows:

[0077]

[0078] Where, represents the set of regression trees; K represents the number of regression trees; q represents the structure of each regression tree; f(·) represents the input-output functional relationship of the regression tree; q(·) represents the functional relationship of each regression tree, which maps the sampled SMART data S to the leaf node corresponding to the prediction result; w q(S) represents the weight of the q(S)th leaf node; N T Indicates the number of regression tree child nodes.

[0079] S220. Construct a regression tree learning objective function, and iteratively train the initial ensemble tree model based on the training data set to obtain a mature ensemble tree model.

[0080] Construct a regression tree learning objective function to learn each regression tree in the initial ensemble tree model; and the mathematical expression of the regression tree learning objective function is as follows:

[0081]

[0082] Where l(·) is a differentiable loss function that measures the predicted health value and sample health label C i The difference; Ω(f) = γN T +0.5λ||w|| 2 is the model complexity penalty term, which is used to smooth the final learned weights of the model to prevent overfitting and limit the total number of leaf nodes;

[0083] The initial ensemble tree model is iterated K times according to the training data set, and a regression tree model is learned in each iteration; the regression tree learning objective function of the kth iteration is:

[0084]

[0085] After iterative training, a mature ensemble tree model φ(·) is finally obtained.

[0086] S230 , obtaining an importance ranking of all SMART attributes of the mechanical hard disk according to the number of times each SMART attribute appears in the mature integrated tree model.

[0087] Specifically, the number of times each SMART attribute appears in the trained mature ensemble tree model φ(·) can be calculated as the importance score value of the SMART attribute. All SMART attributes provided by the mechanical hard disk are sorted according to the importance score value to obtain the importance ranking of all SMART attributes of the mechanical hard disk.

[0088] S240: According to the importance ranking, select a target SMART attribute that meets a preset ranking requirement from all SMART attributes. The preset ranking requirement may be that the importance score of the SMART attribute must be within a certain numerical range of the importance ranking (such as the top 40%, the top 50%, or the top 60%, etc.).

[0089] Specifically, according to the importance ranking of the SMART attributes of the mechanical hard disk, SMART attributes whose importance scores are in the top 50% of the importance ranking among all SMART attributes may be selected as target SMART attributes.

[0090] S250 , performing normalization processing according to the target SMART attributes to obtain target SMART data, obtaining a health label corresponding to the target SMART data according to the original data set, and combining the target SMART data and the health label to obtain a target data set of all target SMART attributes.

[0091] Specifically, normalization processing may be performed according to the target SMART attribute, that is, the normalized value of the target SMART attribute is obtained to obtain the target SMART data;

[0092] The original dataset D of the mechanical hard disk based on O ={(S i ,C i )}, obtain the health label corresponding to the target SMART data, and construct a target dataset that combines the target SMART data and the health label to obtain all the target SMART attributes as follows:

[0093] D={(X i ,Y i )},in, represents the target SMART data with a window width of P, N represents the number of target SMART attributes used to construct the data set, represents the normalized value of each target SMART attribute at sampling time t; Indicates that the health label of the last sampling time of each time window is used as the target SMART data X i Health label.

[0094] Furthermore, in step S300, based on the multi-scale evolutionary graph network, the target SMART attributes, the target SMART data, and the target data set, the correlation between the evolution of the target SMART attributes is self-learned to obtain a health assessment model for the mechanical hard disk, which may specifically include the following steps:

[0095] S310, constructing a multi-scale evolutionary graph network, combining target SMART attributes, target SMART data, and target data sets, and self-learning the evolutionary correlation between each target SMART attribute to extract a multi-scale temporal representation of the target SMART data;

[0096] S320. Iteratively update the multi-scale evolutionary graph network according to the multi-scale time series representation of the target SMART data to obtain a health assessment model for the mechanical hard disk.

[0097] Furthermore, in step S310, a multi-scale evolutionary graph network is constructed, and the target SMART attributes, target SMART data, and target data set are combined to self-learn the evolutionary correlation between the target SMART attributes to extract the multi-scale temporal representation of the target SMART data. The following steps may be further included:

[0098] S312, construct a multi-scale evolutionary graph network, which includes a time convolution module, an evolutionary graph self-learning module, and an evolutionary graph convolution module. The input data X of each layer of the multi-scale evolutionary graph network is the output of the previous layer, and the input data of the first layer is the target SMART data X. i .

[0099] S314: Input the target SMART data into the time convolution module, and extract the multi-scale time series representation of the target SMART data along the time dimension.

[0100] S316. Using the output data of the time convolution module as the input data of the evolutionary graph self-learning module, automatically extracting the graph structure representation of the evolution of each target SMART attribute over time, and outputting the evolved adjacency matrix.

[0101] S318. The adjacency matrix output by the evolutionary graph self-learning module and the multi-scale temporal representation output by the temporal convolution module are used as input data of the evolutionary graph convolution module, and a new multi-scale temporal representation that takes into account the correlation between the SMART attributes of each target is obtained by calculation.

[0102] Furthermore, the temporal convolution module includes a dilated convolution layer and a gating layer. Moreover, in step S314, the target SMART data is input into the temporal convolution module, and a multi-scale temporal representation of the target SMART data is extracted along the time dimension, which may further include the following steps:

[0103] S3142. Divide the target SMART data into multiple time series according to the SMART attributes, use the time series of each target SMART attribute as the input data of the dilated convolution layer, use convolution kernels of multiple different sizes to perform dilated convolution operations, and obtain time series representations of different time scales.

[0104] Specifically, the input data, i.e. the target SMART data, can be divided into N time series X according to the SMART attributes. (1) ,X (2) ,…X (N) , the time series of each target SMART attribute As the input data of the dilated convolution layer, dilated convolution operations are performed using convolution kernels of various sizes to capture temporal representations of different time scales at the same time;

[0105] Among them, the calculation expression of the dilated convolution operation is as follows:

[0106]

[0107] Where, Represents the a-th number of the convolution operation output; Indicates size is 1×k j One-dimensional convolution kernel; d t is the expansion rate.

[0108] Moreover, the output of each dilated convolution operation is truncated to the same length as the output of the convolution operation of the maximum convolution kernel, and they are concatenated to obtain a multi-scale temporal representation of the output of the dilated convolution layer. The calculation expression is as follows:

[0109]

[0110] Where concat(·) represents the concatenation function; w represents the number of convolution kernels in the dilated convolution layer, and the size of the w-th convolution kernel is 1×k w ;P ξ Represents the dimension of the output of the dilated convolutional layer.

[0111] S3144. Using the temporal representations of different time scales output by the dilated convolutional layer as input data for the gating layer, a multi-scale temporal representation of each target SMART attribute output by the temporal convolution module is obtained.

[0112] Specifically, the multi-scale temporal representation output by the dilated convolutional layer is used as the input of the gating layer to control the amount of information passed from the temporal convolution module to the next module. Moreover, the calculation expression of the gating layer is as follows:

[0113] ξ=σ(ξ)⊙μ(ξ);

[0114] Where μ(·) represents the tanh activation function; σ(·) represents the sigmoid activation function, which serves as the gate to μ(ξ ( ' i) ) information is filtered; ⊙ represents the Hadamard product, which corresponds to element multiplication;

[0115] After the input data X is calculated through the above steps in the temporal convolution module, the multi-scale temporal representation of each target SMART attribute output by the temporal convolution module is obtained as follows:

[0116] Right now

[0117] Furthermore, the evolutionary graph self-learning module includes a graph evolution stage division layer and a graph self-learning layer. In step S316, the output data of the temporal convolution module is used as the input data of the evolutionary graph self-learning module, which automatically extracts the graph structure representation of the evolution of each target SMART attribute over time and outputs the evolved adjacency matrix. The following steps may be further included:

[0118] S3162. Divide the multi-scale temporal representation of the output data of the time convolution module into several segments along the time dimension through the graph evolution stage division layer, and use the aggregation function to aggregate the temporal representation of each segment into one dimension as the attribute representation of each evolution stage.

[0119] Specifically, the input data formed by the multi-scale temporal representation of the output data of the time convolution module is divided into layers by the graph evolution stage. The time dimension is divided into several segments, and the aggregation function is used to aggregate the time series representation of each segment into one dimension as the attribute representation of each evolution stage. The expression is as follows:

[0120]

[0121]

[0122] Where, AGG(·) represents an aggregation function, such as the mean function; d a Represents the time interval of each stage; M is the total number of segments in the evolutionary stage.

[0123] S3164. Automatically learn the graph structure representation of each evolutionary stage based on the attribute representation of each evolutionary stage through the graph learning layer. The graph structure representation of each evolutionary stage is obtained from the graph structure representation of the previous evolutionary stage.

[0124] Specifically, the graph learning layer can automatically learn the graph structure representation of each evolutionary stage based on the attribute representation of each evolutionary stage. The graph structure representation of each evolutionary stage is obtained from the graph structure representation of the previous evolutionary stage, that is, the graph structure representation of each stage is gradually evolved. The graph learning layer can represent the attribute representation of each evolutionary stage γ m and the graph structure representation α of the previous stage m-1 As the input sequence, we get the graph structure representation α of the current stage m The mathematical expression is as follows:

[0125] r m =σ(W r [γ m ,α m-1 ]+b r );

[0126] u m =σ(W u [γ m ,α m-1 ]+b u );

[0127] o m =σ(W o [γ m ,(r m ⊙α m-1 )]+b o );

[0128] α m =u m ⊙α m-1 +(1-u m )⊙o m ;

[0129] Where r m and u m Represent the forget gate and update gate respectively; ⊙ represents the Hadamard product, which corresponds to element multiplication; W r 、W u and W o represents the learnable weight matrix; b r 、b u and b o represents the bias term.

[0130] After the above steps, we finally get a series of evolving graph structure representations in Represents the graph structure representation of the m-th evolutionary stage.

[0131] S3166. Connect the graph structure representations of each evolutionary stage in pairs, apply a multi-layer perceptron to derive the adjacency matrix of each evolutionary stage, and use the output of the evolutionary graph self-learning module to evolve the adjacency matrix.

[0132] Specifically, the graph structure representations of each evolutionary stage are connected pairwise, and the adjacency matrix of each evolutionary stage is automatically derived using a multi-layer perceptron. The mathematical expression is as follows:

[0133]

[0134] M m,ij =MLP m (concat(α m,(i) ,α m,(j) ));

[0135]

[0136] Where, Represents the candidate adjacency matrix of the mth evolutionary stage The element in row i and column j; M m,ij Represents the element in row i and column j of the mask matrix; represents the adjacency matrix of the mth evolutionary stage;

[0137] The output of the evolution graph self-learning module is a series of adjacency matrices A={A1,A2,…,A m ,…A M}, where the adjacency matrix A m The adjacency matrix A from the previous stage m-1 and the current stage input data α m Joint decision, i.e. A m By A m-1 Evolved.

[0138] Furthermore, in step S318, the adjacency matrix output by the evolutionary graph self-learning module and the multi-scale temporal representation output by the temporal convolution module are used as input data of the evolutionary graph convolution module, and a new multi-scale temporal representation that considers the correlation between each target SMART attribute is obtained by calculation, which may further include the following steps:

[0139] S3182. Input the adjacency matrix output by the evolutionary graph self-learning module and the multi-scale temporal representation output by the time convolution module as input data into the evolutionary graph convolution module.

[0140] Specifically, the adjacency matrix output by the evolutionary graph self-learning module and the multi-scale temporal representation output by the temporal convolution module are combined into a As the input data of the information propagation layer of the evolutionary graph convolution module; first, the multi-scale temporal representation The graph evolution stages are divided according to the division method in step S3162. The mathematical expression is as follows:

[0141]

[0142] Where, d a Represents the time interval of each stage; M is the total number of segments in the evolutionary stage, M = P ξ / d a .

[0143] S3184. Through the evolutionary graph convolution module, the graph structure representation provided by the adjacency matrix is ​​used to perform graph convolution operations on the multi-scale temporal representation of each target SMART attribute, so as to obtain a new multi-scale temporal representation that takes into account the correlation between each target SMART attribute.

[0144] Specifically, through the evolutionary graph convolution module, the temporal representation of each stage is The corresponding adjacency matrix A m Perform graph convolution operations to obtain a temporal representation of the integrated attribute correlation information. The mathematical expression of the graph convolution operation is as follows:

[0145]

[0146]

[0147] Where I represents the unit matrix; express The degree matrix is ​​a diagonal matrix, and the calculation expression of the i-th element on its diagonal is: W G Represents a self-learning weight matrix;

[0148] The temporal representation of each stage integrates the adjacency matrix of the corresponding stage to obtain the temporal representation that integrates the attribute correlation information

[0149] Furthermore, in S320 , the multi-scale evolutionary graph network is iteratively updated based on the multi-scale temporal representation of the target SMART data to obtain a health assessment model for the mechanical hard disk, which may specifically include the following steps:

[0150] S322. Using residual connections, the input of the multi-scale evolutionary graph network is integrated into the output. The calculation expression of the output is:

[0151] Z=h(X)+H;

[0152] Where X is the input data of the multi-scale evolutionary graph network; h(·) is a 1×1 convolution operation used to adjust the dimension;

[0153] S324, repeat the above step S310 once, that is, cross-stack an evolutionary graph learning structure, and record the output of the last layer of multi-scale evolutionary graph network as Z (l) ;

[0154] S326. The output information of each layer of the multi-scale evolutionary graph network is integrated through residual connections, and the health value of the mechanical hard disk is calculated using the fully connected network. The mathematical expression is as follows:

[0155]

[0156] Where h(·) is a 1×1 convolution operation used to adjust the dimension; Indicates the health assessment value of the mechanical hard disk;

[0157] S328. Based on the Adam optimization algorithm, repeat the above steps S310 and S320, set the number of iterations E, and iteratively update the parameters of the multi-scale evolutionary graph network to obtain the optimal health assessment model for the mechanical hard disk, that is, minimize the mean square error objective function:

[0158]

[0159] Where Y is the health label value of the mechanical hard disk.

[0160] Furthermore, in step S400, the health of the mechanical hard disk to be tested is evaluated according to the health evaluation model of the mechanical hard disk, which may specifically include:

[0161] According to the health evaluation model of the mechanical hard disk, the target SMART data of the mechanical hard disk The data is input into the optimal health assessment model to evaluate the health of the mechanical hard disk to be tested.

[0162] This method can directly extract multi-scale time-series representations from the SMART data of mechanical hard disks and model the evolving interdependencies of various SMART data attributes, thereby accurately mining degradation information related to the health of the mechanical hard disk. This overcomes the lack of existing methods in modeling evolving attribute correlations and enables the health assessment of mechanical hard disks based on SMART data.

[0163] Specifically, the effectiveness of the proposed method was verified using 657 sets of full-life data from a Seagate ST4000DM000 hard drive. The parameters of the evolutionary graph network were set as follows: the number of layers in the evolutionary graph network was 3; the number of mini-batch training was 16; and the number of iterations was selected as 100.

[0164] During the experimental verification, the importance of each attribute of the SMART data is first quantified and ranked. According to the attribute importance quantification results, the attributes used for network training are selected as shown in Table 1. The attribute importance scores of each attribute are as follows: Figure 2 shown.

[0165] Then, 80% of the hard drives (525 hard drive full life data) were selected to construct the training set, and the remaining 20% ​​of the hard drive full life data (132 hard drive full life data) were used as the test set to test the effectiveness of the method of the present invention. In order to more intuitively demonstrate the generalization performance of the method of the present invention, the test set hard drive full life data were grouped according to the hard drive life as shown in Table 2. From each group of data, a hard drive full life data was randomly selected to perform health assessment using the method of the present invention, and the results were visualized as shown in Table 2. Figure 3 As shown in (a), (b), (c), and (d).

[0166] from Figure 3 It can be seen that the health evaluation value of the mechanical hard disk by the method of the present invention is highly close to the actual health value, indicating that the method of the present invention can make an accurate assessment of the health of the mechanical hard disk. To further verify the superiority of the present invention, the method of the present invention is compared with the method based on optimized gradient enhancement (XGBoost) and the method based on long short-term memory network (LSTM). The three methods are evaluated using the scoring function and the root mean square error prediction performance index. The results (mean and variance) on the four groups of hard disk data in the test set are shown as follows: Figure 4 shown.

[0167] from Figure 4 It can be seen that the two performance index values ​​of the method of the present invention are both lower than those of the other two prediction methods, indicating that the health assessment accuracy of the mechanical hard disk of the method of the present invention is higher, the stability is better, and the robustness is stronger.

[0168] Table 1

[0169]

[0170] Table 2

[0171]

[0172] Through the above mechanical hard disk health assessment results and the performance comparison with the two methods, it can be found that the method of the present invention effectively improves the accuracy of mechanical hard disk health assessment by modeling the time-varying dependencies of various SMART attributes and integrating the correlation information of each attribute with multi-scale time series representation, thereby achieving more superior assessment performance.

[0173] The present invention also provides a mechanical hard disk health assessment system 100 based on an evolutionary graph, which is used to assess the health of a mechanical hard disk. Figure 5 As shown, the system includes:

[0174] The SMART data acquisition module 102 is used to acquire the original data sets of the multiple mechanical hard disks based on the full life SMART data of the multiple mechanical hard disks and the corresponding health tags;

[0175] The target data acquisition module 104 is configured to obtain an importance ranking of the SMART attributes of the mechanical hard disk according to the original data set, obtain target SMART attributes that meet preset ranking requirements according to the importance ranking, and obtain target SMART data and corresponding target data sets for all target SMART attributes;

[0176] An evaluation model acquisition module 106 is configured to self-learn the evolutionary correlation between target SMART attributes based on the multi-scale evolutionary graph network, target SMART attributes, target SMART data, and target dataset, and acquire a health evaluation model for a mechanical hard disk;

[0177] The health evaluation module 108 is configured to evaluate the health of the mechanical hard disk to be tested according to the health evaluation model of the mechanical hard disk.

[0178] The mechanical hard disk health assessment system 100 based on the evolutionary graph described in this embodiment corresponds to the above-mentioned mechanical hard disk health assessment method based on the evolutionary graph. The functions of each module in the mechanical hard disk health assessment system 100 based on the evolutionary graph in this embodiment are explained in detail in the corresponding method embodiment and will not be repeated here.

[0179] In addition, the present invention also proposes a computer-readable storage medium, which stores computer execution instructions. When the computer execution instructions are executed by a processor, they are used to implement all or part of the method steps of the mechanical hard disk health assessment method based on the evolutionary graph as described above.

[0180] The present invention implements all or part of the process in the above method, and can also be completed by instructing related hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of each of the above method embodiments. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device that can carry computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0181] Based on the same inventive concept, an embodiment of the present application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program running on the processor, and when the processor executes the computer program, all or part of the method steps in the above method are implemented.

[0182] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor. The processor is the control center of a computer device, connecting all parts of the entire computer device using various interfaces and lines.

[0183] The memory can be used to store computer programs and / or models. The processor implements various functions of the computer device by running or executing the computer programs and / or models stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required for a function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created based on the use of the mobile phone (such as audio data, video data, etc.). In addition, the memory can include a high-speed random access memory and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (SmartMedia Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0184] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, servers, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage) containing computer-usable program code.

[0185] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), servers, and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0186] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0187] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0188] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A mechanical hard disk health assessment method based on evolutionary graph, characterized in that: The method comprises: Obtain the original data sets of multiple mechanical hard disks based on their lifecycle SMART data and corresponding lifecycle health labels. Obtaining the importance ranking of the SMART attributes of the mechanical hard disk according to the original data set, obtaining the target SMART attributes that meet the preset ranking requirements based on the importance ranking, and obtaining the target SMART data and corresponding target data sets of all target SMART attributes; Based on the multi-scale evolutionary graph network, target SMART attributes, target SMART data, and target dataset, the correlation between the evolution of target SMART attributes is self-learned to obtain a health assessment model for mechanical hard drives. Evaluate the health of the hard disk to be tested based on the hard disk health assessment model; The method of self-learning the evolutionary correlation between target SMART attributes based on the multi-scale evolutionary graph network, target SMART attributes, target SMART data, and target data set to obtain a health assessment model for a mechanical hard disk includes: Construct a multi-scale evolutionary graph network, combine target SMART attributes, target SMART data and target dataset, and self-learn the evolutionary correlation between target SMART attributes to extract multi-scale temporal representation of target SMART data; The multi-scale evolutionary graph network is iteratively updated based on the multi-scale time series representation of the target SMART data to obtain a health assessment model for the mechanical hard disk. The multi-scale evolutionary graph network is constructed, and the target SMART attributes, target SMART data, and target data set are combined to self-learn the evolutionary correlation between the target SMART attributes to extract the multi-scale temporal representation of the target SMART data, including: Construct a multi-scale evolutionary graph network, which includes a temporal convolution module, an evolutionary graph self-learning module, and an evolutionary graph convolution module; Input the target SMART data into the temporal convolution module to extract the multi-scale temporal representation of the target SMART data along the time dimension; The output data of the temporal convolution module is used as the input data of the evolutionary graph self-learning module, which automatically extracts the graph structure representation of the evolution of each target SMART attribute over time and outputs the evolved adjacency matrix. The adjacency matrix output by the evolutionary graph self-learning module and the multi-scale temporal representation output by the temporal convolution module are used as input data of the evolutionary graph convolution module, and a new multi-scale temporal representation that takes into account the correlation between the SMART attributes of each target is obtained.

2. The method for evaluating the health of a mechanical hard disk based on an evolutionary graph according to claim 1, wherein: The method of obtaining the original data sets of the plurality of mechanical hard disks according to the full life SMART data and the corresponding full life health labels of the plurality of mechanical hard disks includes: Collect the life-cycle SMART data of multiple mechanical hard disks until they fail, and construct corresponding life-cycle health labels; Combine the full-life SMART data and the full-life health label to obtain the full-life raw data of multiple mechanical hard drives; Based on the full lifecycle raw data, the original data sets of multiple mechanical hard drives are constructed.

3. The method for evaluating the health of a mechanical hard disk based on an evolutionary graph according to claim 1, wherein: The step of obtaining the importance ranking of the SMART attributes of the mechanical hard disk according to the original data set, obtaining the target SMART attributes that meet the preset ranking requirements according to the importance ranking, and obtaining the target SMART data and corresponding target data sets of all the target SMART attributes includes: An initial ensemble tree model is constructed, and a training data set as input data of the initial ensemble tree model is obtained based on the original data set; wherein the training data set is the normalized values ​​of each SMART attribute selected from the original data set; Construct a regression tree learning objective function, and iteratively train the initial ensemble tree model based on the training data set to obtain a mature ensemble tree model; According to the number of times each SMART attribute appears in the mature ensemble tree model, the importance ranking of all SMART attributes of the mechanical hard disk is obtained; Sort by importance and select the target SMART attributes that meet the preset sorting requirements from all SMART attributes; Normalization processing is performed according to the target SMART attributes to obtain target SMART data, a health label corresponding to the target SMART data is obtained according to the original data set, and the target SMART data and the health label are combined to obtain a target data set with all target SMART attributes.

4. The method for evaluating the health of a mechanical hard disk based on an evolutionary graph according to claim 1, wherein: The temporal convolution module includes dilated convolution layers and gating layers; The target SMART data is input into the time convolution module, and the multi-scale temporal representation of the target SMART data is extracted along the time dimension, including: The target SMART data is divided into multiple time series according to the SMART attributes. The time series of each target SMART attribute is used as the input data of the dilated convolution layer. The dilated convolution operation is performed using convolution kernels of different sizes to obtain time series representations at different time scales. The temporal representations of different time scales output by the dilated convolutional layer are used as the input data of the gating layer to obtain the multi-scale temporal representations of each target SMART attribute output by the temporal convolution module.

5. The method for evaluating the health of a mechanical hard disk based on an evolutionary graph according to claim 1, wherein: The evolutionary graph self-learning module includes a graph evolution stage division layer and a graph self-learning layer; The output data of the temporal convolution module is used as the input data of the evolutionary graph self-learning module to automatically extract the graph structure representation of the evolution of each target SMART attribute over time and output the evolutionary adjacency matrix, including: The multi-scale temporal representation of the output data of the temporal convolution module is divided into several segments along the time dimension through the graph evolution stage partitioning layer, and the temporal representation of each segment is aggregated into one dimension using an aggregation function as the attribute representation of each evolution stage; The graph learning layer automatically learns the graph structure representation of each evolutionary stage based on the attribute representation of each evolutionary stage. The graph structure representation of each evolutionary stage is obtained from the graph structure representation of the previous evolutionary stage. The graph structure representations of each evolutionary stage are connected pairwise, and the adjacency matrix of each evolutionary stage is derived using a multi-layer perceptron, so as to obtain the adjacency matrix evolved through the output of the evolutionary graph self-learning module.

6. The method for evaluating the health of a mechanical hard disk based on an evolutionary graph according to claim 1, wherein: The adjacency matrix output by the evolutionary graph self-learning module and the multi-scale temporal representation output by the temporal convolution module are used as input data of the evolutionary graph convolution module to calculate and obtain a new multi-scale temporal representation that takes into account the correlation between each target SMART attribute, including: The adjacency matrix output by the evolutionary graph self-learning module and the multi-scale temporal representation output by the temporal convolution module are input as input data to the evolutionary graph convolution module; Through the evolutionary graph convolution module, the graph structure representation provided by the adjacency matrix is ​​used to perform graph convolution operations on the multi-scale temporal representation of each target SMART attribute, and a new multi-scale temporal representation that takes into account the correlation between each target SMART attribute is obtained.

7. A mechanical hard disk health assessment system based on evolutionary graph, characterized in that: include: The SMART data acquisition module is used to obtain the original data sets of multiple mechanical hard disks based on their full-life SMART data and corresponding full-life health labels; A target data acquisition module is used to obtain the importance ranking of the SMART attributes of the mechanical hard disk according to the original data set, obtain the target SMART attributes that meet the preset ranking requirements according to the importance ranking, and obtain the target SMART data and corresponding target data sets of all target SMART attributes; The evaluation model acquisition module is used to self-learn the evolutionary correlation between target SMART attributes based on the multi-scale evolutionary graph network, target SMART attributes, target SMART data, and target dataset, and obtain the health evaluation model of the mechanical hard disk; A health assessment module is used to assess the health of the mechanical hard disk to be tested based on the health assessment model of the mechanical hard disk; The method of self-learning the evolutionary correlation between target SMART attributes based on the multi-scale evolutionary graph network, target SMART attributes, target SMART data, and target data set to obtain a health assessment model for a mechanical hard disk includes: Construct a multi-scale evolutionary graph network, combine target SMART attributes, target SMART data and target dataset, and self-learn the evolutionary correlation between target SMART attributes to extract multi-scale temporal representation of target SMART data; The multi-scale evolutionary graph network is iteratively updated based on the multi-scale time series representation of the target SMART data to obtain a health assessment model for the mechanical hard disk. The multi-scale evolutionary graph network is constructed, and the target SMART attributes, target SMART data, and target data set are combined to self-learn the evolutionary correlation between the target SMART attributes to extract the multi-scale temporal representation of the target SMART data, including: Construct a multi-scale evolutionary graph network, which includes a temporal convolution module, an evolutionary graph self-learning module, and an evolutionary graph convolution module; Input the target SMART data into the temporal convolution module to extract the multi-scale temporal representation of the target SMART data along the time dimension; The output data of the temporal convolution module is used as the input data of the evolutionary graph self-learning module, which automatically extracts the graph structure representation of the evolution of each target SMART attribute over time and outputs the evolved adjacency matrix. The adjacency matrix output by the evolutionary graph self-learning module and the multi-scale temporal representation output by the temporal convolution module are used as input data of the evolutionary graph convolution module, and a new multi-scale temporal representation that takes into account the correlation between the SMART attributes of each target is obtained.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement all or part of the method steps of the mechanical hard disk health assessment method based on the evolutionary graph as described in any one of claims 1 to 6.

Images