A disk failure prediction method based on feature selection and GS-LSTM

Abstract

The present invention relates to a disk fault prediction based on feature selection and GS-LSTM. In order to solve the problem of serious imbalance between positive and negative samples of disks, the SMOTEENN algorithm is used to balance the training samples. At the same time, the present invention adopts a grid search algorithm to optimize the hyperparameters of the LSTM model, which overcomes the low efficiency of the hyperparameter selection method of the existing method and improves the prediction effect of the model. Since the present invention adopts PCCs to select features of the SMART attributes of the disk, it has a stronger fault characterization ability than the features adopted by the existing method. At the same time, a prediction model of the grid optimization algorithm is designed to construct the model on demand, and a better prediction model is provided for different disk models, making the model more generalizable. It overcomes the problems of the existing method that does not rely on professional knowledge to select SMART attribute features and the low efficiency of industrial applications and prediction models. While maintaining a low false detection rate, the model has a high prediction accuracy and fault detection rate.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence and fault prediction, and relates to a disk fault prediction method based on feature selection and GS-LSTM. The disk fault prediction is mainly completed based on data generated by the SMART attributes of the disk, and can be used in the field of storage technology. Background Art

[0002] In today's big data era, large-scale, massive data storage systems are deployed throughout major data centers. Disks, as the primary storage medium, ensure stable and reliable data access, directly impacting the reliability of the entire system. While the theoretical probability of a single disk failure may be less than 1%, the actual annual failure rate of disks in data centers can exceed 10%. Disk failure not only disrupts the services it carries, but the loss of data stored on the disks can also cause immeasurable losses to businesses and individuals. The Ponemon Institute reports that the maximum cost of data center downtime increased from $1 million in 2010 to $2.4 million in 2016, with disk failure being the primary cause. Using methods such as failure prediction to ensure timely disk replacement and maintenance helps ensure data security and reduce data center operating costs. Therefore, disk failure prediction holds significant research value.

[0003] Most existing disk failure prediction methods are based on neural networks. The basic idea is to use historical data generated by the disk's SMART attributes as input to train the neural network to obtain the optimal prediction model; when performing fault prediction, the data generated by the real-time disk's SMART attributes is input to obtain the predicted disk status.

[0004] Xu et al. proposed in the paper Xu, Chang, Wang, Gang, Liu, Xiaoguang, Guo, Dongdong, Liu, Tie-Yan. Health Status Assessment and Failure Prediction for Hard Drives with Recurrent Neural Networks[J]. IEEE Transactions on Computers, 2016, 65(11). to define the health of the disk by dividing the remaining life of the disk, and then transform the disk failure prediction into a multi-classification problem. A recurrent neural network was used to establish a model to evaluate the health status of the disk based on the gradually changing sequential SMART attributes. Lima et al. proposed in the document Lima, FDdos S., Amaral, GMR, Leite, LGde M., Gomes, JPP, & Machado, J.de C. (2017). Predicting Failures in Hard Drives with LSTM Networks. 2017 Brazilian Conference on Intelligent Systems (BRACIS). doi: 10.1109 / bracis.2017.72 to discretize the remaining life attribute of the disk to define the health of the disk, treat the problem as a multi-label classification task, and use LSTM networks to predict long-term and short-term disk failures. Wang et al. proposed a multi-instance long-term data classification method based on long short-term memory network and attention mechanism for prediction in the literature Wang Guochao, Wang Yu, Sun Xiaojie. Multi-Instance Deep Learning Based on Attention Mechanism for Failure Prediction of Unlabeled Hard Disk Drives[J]. IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, 2021, 70. They regarded long time series HDD data as instance packages, divided them into multiple instances in the sub-concept layer through multi-instance learning, and then studied the relationship between instances and package labels to achieve fault prediction.

[0005] In the process of disk failure prediction using the above methods, most studies on the selection of disk features are based on expert knowledge or all disk features, which leads to low efficiency in industrial applications and prediction models. Summary of the Invention

[0006] Technical problems to be solved

[0007] In order to avoid the shortcomings of the existing technology, the present invention proposes a disk failure prediction based on feature selection and GS-LSTM, which adopts a feature selection method based on correlation analysis and has stronger fault characterization capabilities than the features used in the existing methods. To address the problem of severe imbalance between positive and negative samples of the disk, the SMOTEENN algorithm is used to balance the training samples. At the same time, the present invention adopts a grid search algorithm to optimize the hyperparameters of the LSTM model, overcoming the low efficiency of the hyperparameter selection method of the existing method and improving the prediction effect of the model.

[0008] Technical Solution

[0009] A disk failure prediction method based on feature selection and GS-LSTM is characterized by the following steps:

[0010] Step 1: Use Person correlation analysis to select the SMART attributes of the disk:

[0011] Combine the SMART attribute data of each of the n failed disks t days before and on the day of failure into t+1 groups, and combine the t+1 group data of n failed disks into n*(t+1) groups; where the k SMART attributes of the failed disk are x i (i=1,2,…,k);

[0012] Randomly combine n normal disks into t+1 groups based on the SMART attribute data of each disk, and combine the t+1 group data of n faulty disks into n*(t+1) groups; the k SMART attributes of the normal disks are x j (j=1,2,…,k);

[0013] The k SMART attributes are R m If there are k variables in the space, the correlation coefficient between the SMART attributes of the faulty disk and the SMART attributes of the normal disk is:

[0014]

[0015] Where R(i,j) is the Pearson correlation coefficient PCCs,

[0016] Calculate the Pearson correlation coefficient of the two sequences, compare the Pearson correlation coefficient of each SMART attribute, and use the SMART attributes with 0.00≤|R|<0.20 as the fault characteristics;

[0017] Step 2: Divide all fault feature datasets into training sets and test sets;

[0018] Step 3: Use the SMOTEENN algorithm to balance the positive and negative samples in the training set to form a training set with balanced positive and negative samples;

[0019] Step 4: Use the balanced training set of positive and negative samples as the input of the GS-LSTM fault prediction model. Make predictions based on the balanced training set of positive and negative samples and use the GS algorithm to optimize the model's hyperparameters to obtain the optimal LSTM fault prediction model.

[0020] The GS-LSTM fault prediction model adopts a long short-term memory network and two hyperparameters: the number of LSTM network layers γ and the initial learning rate η.

[0021] The objective function of grid search is: minN(K t ,F t )

[0022] satisfy

[0023]

[0024] Where K t is the prediction sequence corresponding to the training set, F t For the test set, step γ and step η They are respectively expressed as search steps. First, we need to preset the value range of the number of LSTM network layers γ and the learning rate η. In order to speed up the optimization of the model, η is controlled in a certain value range, γ max It is necessary to control the value within a small range, and then select the appropriate search step size, train the LSTM model, and obtain the optimal number of network layers and learning rate;

[0025] Step 5: The test set is used as input to the optimized LSTM time series prediction model, and the output is the disk failure prediction result.

[0026] The disk failure prediction results obtained in step 5 are used to create a classification result confusion matrix. Based on the matrix, the model is evaluated using the three indicators of Accuracy, FAR, and FDR:

[0027] Accuracy: Accuracy is the proportion of correct predictions among all predictions and is also the most commonly used classification performance indicator.

[0028]

[0029] FAR: False positive rate, that is, the proportion of a hard drive that is actually good but predicted by the model to be bad:

[0030]

[0031] FDR: Fault Detection Rate, that is, the proportion of hard drives that are actually bad but predicted by the model:

[0032]

[0033] The classification result confusion matrix is:

[0034]

[0035] 70% of all fault feature data sets are training sets, and 30% are test sets.

[0036] The step 3 uses the SMOTEENN algorithm to balance the positive and negative samples of the training set: first, generate faulty disk samples through SMOTE oversampling, then use the ENN method to deeply clean the generated data, and finally form category-balanced data.

[0037] The sampling process is as follows:

[0038] Step (1): Determine the sampling rate N according to the number of normal state samples and fault state samples;

[0039] Step (2): For each sample X in the fault set i , i∈(1,2,…,T), find its K nearest neighbor samples;

[0040] Step (3): From X i Among the K nearest neighbor samples, randomly select M nearest neighbor samples, denoted as X ij , j=1,2,…,M;

[0041] Step (4): For each X ij The sample is randomly linearly interpolated to construct a new sample X new ;

[0042] X new =X i +rand(0,1)×|X ij -(X i )|

[0043] Step (5): Add the synthesized new samples to the unbalanced sample set to balance the samples and form positive and negative balanced samples;

[0044] Step (6): Use the nearest neighbor rule ENN to predict each sample in the positive and negative balanced samples. If the predicted fault state label does not match the actual label, delete the sample to obtain a training set with a positive and negative sample balance.

[0045] The long short-term memory network model is constructed by using the method disclosed in the document “Hochreiter, S., Schmidhuber, J., 1997. Longshort-term memory. Neural Computation 9, 1735–1780.”

[0046] The specific steps of optimizing the LSTM step by the GS algorithm in step 4 are as follows:

[0047] Step (4.1) determines the range of hyperparameters, the initial learning rate η∈[0.0001,0.001], the number of network layers γ∈[1,5], and selects an appropriate search step size, step η Set to 0.0001, step γ Set to 1 to divide the hyperparameter space into grids and form grid nodes N i ;

[0048] Step (4.2) calculates each grid node N i The objective function value under ;

[0049] Step (4.3) If all grid nodes have been searched, compare each grid node N i , according to the objective function value, output the optimal hyperparameters η and γ.

[0050] Beneficial effects

[0051] The present invention proposes a disk failure prediction method based on feature selection and GS-LSTM. To address the problem of severe imbalance between positive and negative disk samples, the SMOTEENN algorithm is used to balance the training samples. At the same time, the present invention adopts a grid search algorithm to optimize the hyperparameters of the LSTM model, overcoming the low efficiency of the hyperparameter selection method of the existing method and improving the prediction effect of the model.

[0052] Because this method uses PCCs to select features for disk SMART attributes, it possesses stronger fault characterization capabilities than existing methods. Furthermore, to better adapt to workplace scenarios, the invention employs a grid-optimized prediction model to construct models on demand, resulting in better prediction models for different disk models and enhanced model generalization. This overcomes the existing methods' lack of reliance on specialized knowledge for SMART attribute feature selection, resulting in low industrial application and prediction model efficiency. This model achieves high prediction accuracy and fault detection rates while maintaining a low false positive rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 : Disk failure prediction process based on feature selection and GS-LSTM

[0054] Figure 2 : Linear correlation between properties of failed and non-failed disks

[0055] Figure 3 : LSTM prediction model

[0056] Figure 4 : Training and prediction accuracy

[0057] Figure 5 : The loss during training DETAILED DESCRIPTION

[0058] The present invention will now be further described with reference to the embodiments and accompanying drawings:

[0059] The technical solution adopted in the present invention is:

[0060] Step 1: Feature selection. Use Person correlation analysis to select the SMART attributes of the disk.

[0061] Select n faulty disks, intercept their SMART attribute data t days before the failure and on the day of the failure, and merge them into t+1 groups; merge the t+1 group data of n faulty disks into n*(t+1) groups. Then, for normal disks, randomly select n disks and merge the t+1 group data into n*(t+1) groups. The n*(t+1) group data of k SMART attribute variables of the faulty disk and the normal disk are recorded as x 1,m ,x 2,m ,...,x k,m (m=1,2,...,n*(t+1)), so that k SMART attributes can be regarded as R m k variables in the space, then the SMART attribute x of the failed disk i (i=1,2,…,k) and the SMART attribute x of a normal diskj The correlation coefficient between (j=1,2,…,k) can be defined as follows:

[0062]

[0063] Where R(i,j) is the Pearson correlation coefficient (PCCs),

[0064] Fault feature selection is performed based on the correlation coefficient. The specific steps are as follows:

[0065] Step (1): Select the disk SMART attribute and obtain the n*16-dimensional data sequence of the faulty disk and the normal disk under this attribute;

[0066] Step (2): Calculate the Pearson correlation coefficient of the two series;

[0067] Step (3): Change the SMART attributes in sequence and repeat steps (1) and (2);

[0068] Step (4): Through step (2), compare the Pearson correlation coefficient of each SMART attribute, and select attributes with very weak or no correlation as fault features according to Table 1.

[0069] The statistical significance represented by the correlation coefficient values ​​is shown in Table 1.

[0070] Table 1 Empirical interpretation of the correlation coefficients

[0071]

[0072] The lower the correlation, the easier it is for the SMART attribute to distinguish between a normal disk and a faulty disk. Therefore, only the SMART attributes with the lowest correlation are retained as fault features.

[0073] Step 2: Select SMART attribute features with low correlation in the disk dataset. Divide the dataset, select 70% as the training set and the remaining 30% as the test set.

[0074] Step 3: Balance the positive and negative samples in the training set. Since the majority of samples are from healthy disks, while the number of samples from faulty disks is small, the data is extremely unbalanced. Using this unbalanced training set directly to train the model would result in significant deviations in model accuracy. Therefore, the SMOTEENN algorithm is used to balance the positive and negative samples, ultimately creating a balanced training set.

[0075] First, we use SMOTE oversampling to generate faulty disk samples. Then, we use the ENN method to perform deep cleaning on the generated data, and finally form class-balanced data. The sampling process is as follows:

[0076] Step (1): Determine the sampling rate N according to the number of normal state samples and fault state samples;

[0077] Step (2): For each sample X in the fault set i , i∈(1,2,…,T), find its K nearest neighbor samples;

[0078] Step (3): From X i Among the K nearest neighbor samples, randomly select M nearest neighbor samples, denoted as X ij , j=1,2,…,M;

[0079] Step (4): For each X ij Perform random linear interpolation on the sample to construct a new sample X new , the method is as follows;

[0080] X new =X i +rand(0,1)×|X ij -(X i )|

[0081] Step (5): Add the synthesized new samples to the unbalanced sample set to balance the samples and form positive and negative balanced samples;

[0082] Step (6): Use the nearest neighbor rule (ENN) to predict each sample in the positive and negative balanced samples. If the predicted fault state label does not match the actual label, delete the sample. Finally, a training set with balanced positive and negative samples is obtained.

[0083] In step 4, the GS-LSTM fault prediction model is established by using the training set with balanced positive and negative samples generated in step 3 as input. The long short-term memory network is used to make predictions based on the training set with balanced positive and negative samples. The GS algorithm is used to optimize the model's hyperparameters to obtain the optimal fault prediction model.

[0084] (4a) Using the method disclosed in the document “Hochreiter, S., Schmidhuber, J., 1997. Long short-term memory. Neural Computation 9, 1735–1780.”, a long short-term memory network model was constructed.

[0085] (4b) For multi-layer LSTM networks, a random dropout layer (Dropout) is added to the network layer to randomly disconnect the connections between network layers and effectively prevent overfitting.

[0086] (4c) A grid search optimization method is used to optimize the two hyperparameters of the LSTM network layer number γ and the initial learning rate η. The principle of the grid search algorithm (GS) is to divide the valid range of the variable into a grid, traverse all the values ​​in the grid, find the objective function value that meets the constraints, and select the variable value with the smallest prediction error as the optimal hyperparameter.

[0087] The objective function of grid search is

[0088] minN(K t ,F t )

[0089] satisfy

[0090]

[0091] Where K t is the prediction sequence corresponding to the training set, F t For the test set, step γ and step η They are respectively expressed as search steps. First, we need to preset the value range of the number of LSTM network layers γ and the learning rate η. In order to speed up the optimization of the model, η is controlled in a certain value range, γ max It is necessary to control the value within a small range, and then select the appropriate search step size, train the LSTM model, and take the optimal number of network layers and learning rate.

[0092] The specific steps of the GS algorithm to optimize the LSTM step are:

[0093] (4c1) Determine the range of hyperparameters, initial learning rate η∈[0.0001,0.001], number of network layers γ∈[1,5], and select an appropriate search step size, step η Set to 0.0001, step γ Set to 1 to divide the hyperparameter space into grids and form grid nodes N i ;

[0094] (4c2) Calculate each grid node N i The objective function value under ;

[0095] (4c3) If all grid nodes have been searched, compare each grid node N i , according to the objective function value, output the optimal hyperparameters η and γ.

[0096] Step 5: Obtain the optimized LSTM time series prediction model, use the test set as the output of the model, obtain the prediction results, and evaluate the model using the three indicators of Accuracy, FAR, and FDR.

[0097] The confusion matrix of the classification results is given in Table 2.

[0098] Table 2 Classification results confusion matrix

[0099]

[0100] According to this matrix, the three evaluation indicators are calculated as follows:

[0101] Accuracy: Accuracy is the proportion of correct predictions among all predictions and is the most commonly used classification performance indicator.

[0102]

[0103] FAR: False positive rate, that is, the ratio of a hard drive that is actually good but predicted by the model to be bad.

[0104]

[0105] FDR: Fault detection rate, that is, the ratio of a hard drive that is actually bad but is predicted by the model to be bad.

[0106]

[0107] The disk failure prediction process based on feature selection and GS-LSTM is as follows Figure 1 As shown:

[0108] The present invention is further illustrated by the following simulation experiments, from which the effect compared with the prior art can be seen:

[0109] 1. Simulation conditions

[0110] This paper simulates the problem using PyCharm 2021, developed by JetBrains, on an AMD Ryzen 7 5800H with Radeon Graphics 3.20GHz CPU, an NVIDIA GeForce RTX 3050 GPU, and Windows 11. The data is from a publicly available disk dataset from Backblze.

[0111] The methods compared in the experiment are as follows:

[0112] A disk failure prediction method based on convolution-LSTM, denoted as C-LSTM in the experiment, is referenced as ShiJunjie, DuJing, Ren Yingwen, Li Boyu, Zou Jinwei, Zhang Anyi. Convolution-LSTM-Based Mechanical Hard Disk Failure Prediction by SensoringS.MARTIndicators[J]. Journal of Sensors, 2022, 2022.

[0113] Multi-Instance Deep Learning Based on Attention Mechanism for Failure Prediction of Unlabeled Hard Disk Drives[J].IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, 2021, 70.

[0114] A method for health status assessment and failure prediction of hard drives based on recurrent neural networks, denoted as RNN in the experiment, is referenced by Xu, Chang, Wang, Gang, Liu, Xiaoguang, Guo, Dongdong, Liu, Tie-Yan. Health Status Assessment and Failure Prediction for Hard Drives with Recurrent Neural Networks[J]. IEEE Transactions on Computers, 2016, 65(11).

[0115] 2. Simulation content

[0116] 2.1 Example 1:

[0117] 2.1.1 Experimental Dataset

[0118] This dataset, sourced from Backblze's data from the first quarter of 2020, part of the second quarter, and all failed disks in 2020, contains the largest number of disk samples, 47,428, belonging to the ST4000DM000 model. This sample set was chosen for this study. 70% of the samples were used as the training set, and the remaining 30% were used as the test set.

[0119] 2.1.2 Feature Selection

[0120] We selected 170 failed disks from the dataset. For each disk, we collected SMART attribute data from the 15 days before and the day of failure, and combined them into 16 groups. We then combined the 16 groups of data from all 170 failed disks to form 2720 groups. We then randomly selected 170 healthy disks and combined the 16 groups of data from each disk to form 2720 groups of data from healthy disks. The 2720 groups of data on the 24 SMART attribute variables of the failed and healthy disks are denoted as x. 1m ,x 2m ,…,x 24m ,m=1,2,...,2720, so the 24 attributes can be regarded as R m 24 variables in the space, then the SMART attribute of the failed disk x i (i=1,2,...,24) and the SMART attribute x of a normal disk j The correlation between (j=1,2,…,24) is the correlation coefficient. The experimental results are as follows: Figure 2 shown.

[0121] from Figure 2 The six features with the smallest correlation coefficients are selected, as shown in Table 3.

[0122] Table 3 SMART attributes selected based on relevance

[0123]

[0124] 2.1.3 Balancing the Dataset

[0125] The SMOTEENN algorithm was chosen to balance the positive and negative samples in the training set. Table 4 shows that after using the SOMTEENN algorithm, the number of faulty samples increased significantly due to oversampling and was basically balanced with the number of normal samples, which laid a good foundation for subsequent model training.

[0126] Table 4 Comparison of sample numbers between the original dataset and the balanced dataset

[0127]

[0128] 2.1.4 Construction of prediction model and optimization

[0129] Through GS algorithm optimization, the optimal parameters are as follows: the number of network layers is 3, and the learning rate is 0.0003.

[0130] The LSTM prediction model finally established is as follows Figure 3 As shown in the figure, a 3-layer LSTM structure is adopted, the learning rate is 0.0003, the Dropout regularization parameter is 0.3, the sequence segmentation scale is 5, Adamax is selected as the model optimizer, and BCELoss is used as the model loss function.

[0131] 2.1.4 Experimental Results Analysis

[0132] The LSTM model is trained with a batch size of 64 and a maximum number of iterations of 200. The training and prediction accuracy are shown in Figure 2. Figure 4 As shown in , we can see that the prediction accuracy quickly reaches the highest value, and the decline curve of the training loss is as follows Figure 5 shown.

[0133] According to the specific embodiment of the present invention, Accuracy, FAR, and FDR are calculated and compared with RNN, LSTM, C-LSTM, and MIL-LSTM. The results are shown in Table 5.

[0134] Table 5 Statistics of model prediction results

[0135]

[0136] As can be seen from Table 5, the features obtained by selecting the SMART attributes of the disk using PCCs in the present invention have stronger fault characterization capabilities. GS is then used to optimize the hyperparameters of the LSTM model. The three indicators of the new model all reach the best, verifying the advanced nature of the present invention.

Claims

1. A disk failure prediction method based on feature selection and GS-LSTM, characterized by Here are the steps: Step 1: Use Person correlation analysis to select the SMART attributes of the disk: Combine the SMART attribute data of each of the n failed disks t days before and on the day of failure into t+1 groups, and combine the t+1 group data of n failed disks into n*(t+1) groups; where the k SMART attributes of the failed disk are x i , i=1,2,……k; Randomly combine n normal disks into t+1 groups based on the SMART attribute data of each disk, and combine the t+1 group data of n faulty disks into n*(t+1) groups; the k SMART attributes of the normal disks are x j , j=1,2,……k; The k SMART attributes are If there are k variables in the space, the correlation coefficient between the SMART attributes of the faulty disk and the SMART attributes of the normal disk is: Where, is the Pearson correlation coefficient PCCs, , ; Calculate the Pearson correlation coefficient of the two series and compare the Pearson correlation coefficient of each SMART attribute, with 0.00≤| SMART attributes with |<0.20 are used as fault features; Step 2: Divide all fault feature datasets into training sets and test sets; Step 3: Use the SMOTEENN algorithm to balance the positive and negative samples in the training set to form a training set with balanced positive and negative samples; Step 4: Use the balanced training set of positive and negative samples as the input of the GS-LSTM fault prediction model. Make predictions based on the balanced training set of positive and negative samples and use the GS algorithm to optimize the model's hyperparameters to obtain the optimal LSTM fault prediction model. The GS-LSTM fault prediction model adopts a long short-term memory network and the number of LSTM network layers and the initial learning rate Two hyperparameters: The objective function of the grid search is: satisfy Where, is the prediction sequence corresponding to the training set, is the test set, and They are respectively represented as search steps; first, the number of LSTM network layers must be preset and learning rate In order to speed up the optimization of the model, Controlled within a certain value range, It is necessary to control the value within a small range, and then select the appropriate search step size, train the LSTM model, and obtain the optimal number of network layers and learning rate; Step 5: The test set is used as input to the optimized LSTM time series prediction model, and the output is the disk failure prediction result.

2. The disk failure prediction based on feature selection and GS-LSTM according to claim 1 is characterized by: The disk failure prediction results obtained in step 5 are used to create a classification result confusion matrix. Based on the matrix, the model is evaluated using the three indicators of Accuracy, FAR, and FDR: Accuracy: Accuracy is the proportion of correct predictions among all predictions and is also the most commonly used classification performance indicator. FAR: False positive rate, that is, the proportion of a hard drive that is actually good but predicted by the model to be bad: FDR: Failure Detection Rate, that is, the proportion of hard drives that are actually bad but predicted by the model: The parameters of the classification result confusion matrix are: TP: Good market and predicted market, FP: Bad market and predicted market, TN: Good market and predicted market, FN: Bad market and predicted market.

3. The disk failure prediction based on feature selection and GS-LSTM according to claim 1 is characterized by: 70% of all fault feature data sets are training sets and 30% are test sets.

4. The disk failure prediction based on feature selection and GS-LSTM according to claim 1 is characterized by: The step 3 uses the SMOTEENN algorithm to balance the positive and negative samples of the training set: first, generate faulty disk samples through SMOTE oversampling, then use the ENN method to deeply clean the generated data, and finally form category-balanced data.

5. The disk failure prediction based on feature selection and GS-LSTM according to claim 4 is characterized in that: The sampling process is as follows: Step (1): Determine the sampling rate based on the number of normal state samples and fault state samples ; Step (2): For each sample in the fault set , , find its Nearest neighbor samples; Step (3): From of Among the nearest neighbor samples, randomly select M nearest neighbor samples, denoted as , ; Step (4): For each The samples are randomly linearly interpolated to construct new samples ; Step (5): Add the synthesized new samples to the unbalanced sample set to balance the samples and form positive and negative balanced samples; Step (6): Use the nearest neighbor rule ENN to predict each sample in the positive and negative balanced samples. If the predicted fault state label does not match the actual label, delete the sample to obtain a training set with a positive and negative sample balance.

6. The disk failure prediction based on feature selection and GS-LSTM according to claim 1 is characterized by: The specific steps of optimizing the LSTM step by the GS algorithm in step 4 are as follows: Step (4.1) determines the range of hyperparameters and the initial learning rate , number of network layers , choose an appropriate search step size, Set to 0.0001, Set to 1 to divide the hyperparameter space into grids and form grid nodes. ; Step (4.2) calculates each grid node The objective function value under ; Step (4.3) If all grid nodes have been searched, compare each grid node , according to the objective function value, output the optimal hyperparameters 、 .

Images