SSD fault prediction method and device based on multi-task learning and medium

Abstract

The invention provides an SSD fault prediction method and device based on multi-task learning and a medium, and relates to the technical field of machine learning, and the method comprises the steps: carrying out the preprocessing and standardization of data in an obtained hard disk log, obtaining a plurality of initial SSDs and the initial data of each initial SSD, obtaining a target age group list and a key feature list corresponding to each target age group, obtaining a target SSD list corresponding to each target age group and a target key feature value list corresponding to each target SSD, obtaining a target multi-task learning model, obtaining a target stacking result, obtaining a first other classifier and a second other classifier, and obtaining a target stacking result; obtaining a first other prediction result; and obtaining a second other prediction result, and obtaining a final prediction result corresponding to each target SSD based on the target stacking result, the first other prediction result and the second other prediction result, thereby improving the accuracy of SSD fault prediction.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a SSD fault prediction method, device and medium based on multi-task learning. Background Art

[0002] With the widespread use of solid-state drives (SSDs) in cloud storage systems and data centers, the high performance and low power consumption of SSDs have gradually replaced traditional mechanical hard disks (HDDs). However, the frequent failure of SSDs has become an important factor restricting the stability and reliability of cloud storage systems. Statistics show that about 80% of hardware failures in data centers are directly related to hard disk failures, and the special failure modes of SSDs make it more complicated to predict their failures.

[0003] Existing hard disk failure prediction methods are usually based on the Self-Monitoring, Analysis and Reporting Technology (SMART) log of SSD devices, which provides a series of parameters reflecting the operating status of the hard disk, such as read and write error rate, temperature, reallocatable sector count, etc. Researchers generally use machine learning or deep learning technology to extract features from SMART logs and build prediction models. These methods usually use the following technical means: (1) Static feature selection method: Researchers analyze the statistical distribution and correlation of SMART attributes, select several key features, and build a classification model. Common feature selection methods include Pearson correlation coefficient, Spearman correlation coefficient, random forest, etc.

[0004] (2) Single model prediction: Relying on a single classification model to predict all hard disk data, most existing studies use traditional machine learning models (such as support vector machines (SVMs) and gradient boosting trees (GBDTs)) or deep learning models (such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs)) to predict hard disk failures.

[0005] (3) Laboratory data analysis: Many studies have conducted load tests on SSDs in a simulated environment, recorded SMART log data under working conditions, and used the data in the laboratory for model training and performance evaluation.

[0006] Although the above methods have improved the accuracy of hard disk failure prediction to a certain extent, they still face many challenges: (1) Feature Importance Drift Problem The fault features of SSDs change over time and with increasing service life, that is, the importance of features drifts as the hard drive ages and the working environment changes. Most existing methods are based on static feature selection strategies, which are difficult to adapt to this dynamic change, thereby reducing the long-term prediction performance of the model. The weights of fault features change with increasing service life; the reasons are as follows: the dominant factors of fault features may be very different at different stages of the hard drive life cycle. For example, the failure of new hard drives is mostly caused by manufacturing defects, while after long-term use, read and write errors and uncorrectable sector counts become the main failure factors. Data noise and outliers (such as intermittent errors) in long-term operation are not fully processed, which further reduces the accuracy of feature selection.

[0007] (2) Limitations of a single model; A single model has limited performance in dealing with complex failure modes, especially when dealing with diverse data distributions, and is prone to overfitting or underfitting. The reasons are as follows: Complex SSD failure modes (such as multi-stage failures and intermittent errors) require models to capture multi-level feature relationships, and a single model is difficult to cover all situations. Some machine learning models are sensitive to a small number of samples. For example, support vector machines (SVMs) are less robust to sample imbalance and are prone to misjudgment when the actual failure rate is extremely low. Summary of the invention

[0008] In view of the above technical problems, the technical solution adopted by the present invention is: a SSD fault prediction method based on multi-task learning, the method comprising the following steps: S100, preprocessing and standardizing the data in the acquired hard disk log to obtain a plurality of initial SSDs and initial data of each initial SSD; S200, obtaining a target age group list and a key feature list corresponding to each target age group, wherein the target age group list includes a target age groups corresponding to different preset intervals of SSD usage time, and the key feature list includes a plurality of key features for predicting SSD failures, wherein the key features are obtained based on a hard disk log of the SSD; S300, based on the usage duration included in the initial data, obtain a target SSD list corresponding to each target age group and a target key feature value list set corresponding to the target SSD list; the target SSD list includes a plurality of target SSDs, and the target SSDs in the same target SSD list are initial SSDs belonging to the same target age group, and the target key feature value list set includes a plurality of target key feature value lists, and the target key feature value list includes target key feature values ​​corresponding to a plurality of specified features of the target SSD; wherein the specified feature is a key feature corresponding to the target age group to which the target SSD belongs; S400, inputting the target key feature value list sets of the target SSD lists corresponding to the a target age groups into the a target sub-learning models included in the target multi-task learning model, respectively, to obtain the fault prediction results output by each target sub-learning model; and inputting the a fault prediction results as meta-features into the target meta-model to perform fault prediction again, to obtain the target stacking result; S500, obtaining a first other classifier and a second other classifier, inputting the initial data of the target SSD into the first other classifier for fault prediction, and obtaining a first other prediction result; and inputting the initial data of the target SSD into the second other classifier for fault prediction, and obtaining a second other prediction result; S600 , obtaining a final prediction result corresponding to each target SSD based on the target stacking result, the first other prediction result and the second other prediction result.

[0009] According to another aspect of the present invention, a non-transitory computer-readable storage medium is provided, in which at least one instruction or at least one program is stored, and the at least one instruction or the at least one program is loaded and executed by a processor to implement the aforementioned method.

[0010] According to yet another aspect of the present invention, an electronic device is provided, comprising a processor and the aforementioned non-transitory computer-readable storage medium.

[0011] The present invention has at least the following beneficial effects: In summary, the data in the acquired hard disk log is preprocessed and standardized, a number of initial SSDs and the initial data of each initial SSD are obtained, a target age group list and a key feature list corresponding to each target age group are obtained, based on the usage time of the target SSD, a target SSD list corresponding to each target age group and a target key feature value list set corresponding to the target SSD list are obtained, the target key feature value list sets of the target SSD lists corresponding to a target age groups are respectively input into a target sub-learning models included in the target multi-task learning model, and the fault prediction results output by each target sub-learning model are obtained; and the a fault prediction results are input into the target meta-model as meta-features to perform fault prediction again, and the target age group list is obtained. The target stacking result is obtained, the first other classifier and the second other classifier are obtained, the initial data of the target SSD is input into the first other classifier for fault prediction, and the first other prediction result is obtained; and the initial data of the target SSD is input into the second other classifier for fault prediction, and the second other prediction result is obtained. Based on the target stacking result, the first other prediction result and the second other prediction result, the final prediction result corresponding to each target SSD is obtained. The present invention classifies the hard disk by using different key features according to the usage time, so as to improve the accuracy of SSD fault prediction; the present invention performs fault prediction through multiple models, effectively integrates the prediction results of the multi-task model and the additional classifier, significantly reduces the overfitting risk of a single model, and improves the robustness of the fault prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] 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.

[0013] Figure 1 A flowchart of a SSD fault prediction method based on multi-task learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0015] The embodiment of the present invention provides a SSD fault prediction method based on multi-task learning, such as Figure 1 As shown, the method comprises the following steps: S100, preprocessing and standardizing the data in the acquired hard disk log to obtain a plurality of initial SSDs and initial data of each initial SSD.

[0016] Specifically, the data in the hard disk log may be SMART log data; the preprocessing includes deleting attribute data with no recorded value or fixed value; the preprocessing also includes deleting data corresponding to preset initial features with low correlation, and the preset initial features with low correlation may be determined according to actual user needs to ensure that data with strong correlation with SSD failure is screened out. In one embodiment of the present invention, the quartile method is used to screen out irrelevant abnormal data points, such as negative values ​​or abnormally large values ​​caused by acquisition errors.

[0017] Specifically, the standardization process is normalization process. Those skilled in the art know that any method of normalizing data in the prior art belongs to the protection scope of the present invention and will not be described in detail here. The data in the hard disk log obtained is normalized to eliminate the numerical differences between different hard disk models and improve the comparability of the data.

[0018] S200, obtaining a target age group list and a key feature list corresponding to each target age group, wherein the target age group list includes a target age groups corresponding to different preset intervals of SSD usage time, and the key feature list includes several key features for predicting SSD failures, and the key features are obtained based on the hard disk log of the SSD.

[0019] Among them, the target age group list K = {K1, K2, ..., K r , …, K a}, K r is the rth target age group, i.e., the preset range of the usage time of the rth SSD, where r ranges from 1 to a, and a is the number of target age groups; K r Corresponding key feature list AK r ={AK r1 , A.K. r2 , …, AK ry , …, AK rp(r)}, A.K. ry is the yth key feature corresponding to the rth target age group, y ranges from 1 to p(r), and p(r) is the number of key features corresponding to the rth target age group.

[0020] S300, based on the usage time included in the initial data, obtain a target SSD list corresponding to each target age group and a target key feature value list set corresponding to the target SSD list; the target SSD list includes several target SSDs and the target SSDs in the same target SSD list are initial SSDs belonging to the same target age group, and the target key feature value list set includes several target key feature value lists, and the target key feature value list includes target key feature values ​​corresponding to several specified features of the target SSD; wherein the specified feature is a key feature corresponding to the target age group to which the target SSD belongs.

[0021] Among them, the target SSD list E corresponding to the rth target age group r ={E r1 , E r2 ,…,E rt ,…,E rk（r）}, E rt is the tth target SSD of the rth target age group, t ranges from 1 to k(r), k(r) is the number of target SSDs of the rth target age group, E rt Corresponding target key feature value list AE rt ={AE rt1 , A.E. rt2 , …, AE rty , …, AE rtp(r)}, AE rty is the value of the yth key feature corresponding to the tth SSD of the rth age group.

[0022] S400, input the target key feature value list sets of the target SSD list corresponding to the a target age groups into the a target sub-learning models included in the target multi-task learning model respectively, and obtain the fault prediction results output by each target sub-learning model; and input the a fault prediction results as meta-features into the target meta-model to perform fault prediction again, and obtain the target stacking result.

[0023] Specifically, change E1 to E a Input them into a target sub-learning models respectively, and obtain the target prediction result list F={F1, F2, …, F r , …, F a}, F r YesE r The fault prediction result input to the rth target self-learning model.

[0024] S500, obtain a first other classifier and a second other classifier, input the initial data of the target SSD into the first other classifier for fault prediction, and obtain a first other prediction result; and input the initial data of the target SSD into the second other classifier for fault prediction, and obtain a second other prediction result.

[0025] Among them, the first other classifier is the target GBDT classifier, and the second other classifier is the target XGBoost classifier.

[0026] Specifically, an initial GBDT classifier is constructed, and the sample data of the sample SSD and the actual failure results of the sample SSD are used to train the initial GBDT classifier to obtain a target GBDT classifier; an initial XGBoost classifier is constructed, and the sample data of the sample SSD and the actual failure results of the sample SSD are used to train the initial XGBoost classifier to obtain a target XGBoost classifier.

[0027] S600 , obtaining a final prediction result corresponding to each target SSD based on the target stacking result, the first other prediction result and the second other prediction result.

[0028] Specifically, based on the target stacking result, the first other prediction result and the second other prediction result, a voting decision is made to obtain the fault label of each target SSD, thereby obtaining the final prediction result corresponding to each target SSD, thereby improving the robustness and comprehensive performance of the SSD fault prediction method based on multi-task learning of the present invention.

[0029] The fault tag of the target SSD is "1" or "0". Optionally, when the fault tag is "1", the target SSD is in a fault state, and when the fault tag is "0", the target SSD is in a non-fault state.

[0030] In summary, the data in the hard disk log is preprocessed and standardized, several initial SSDs and the initial data of each initial SSD are obtained, a target age group list and a key feature list corresponding to each target age group are obtained, and based on the usage time of the target SSD, a target SSD list corresponding to each target age group and a target key feature value list set corresponding to the target SSD list are obtained, and the target key feature value list sets of the target SSD lists corresponding to a target age groups are respectively input into a target sub-learning models included in the target multi-task learning model to obtain the fault prediction results output by each target sub-learning model; and the a fault prediction results are input into the target meta-model as meta-features to perform fault prediction again, and the target stacking results are obtained. The first other classifier and the second other classifier input the initial data of the target SSD into the first other classifier for fault prediction to obtain the first other prediction result; and input the initial data of the target SSD into the second other classifier for fault prediction to obtain the second other prediction result, based on the target stacking result, the first other prediction result and the second other prediction result, obtain the final prediction result corresponding to each target SSD. The present invention classifies the hard disk by using different key features according to the usage time, so as to improve the accuracy of SSD fault prediction; the present invention performs fault prediction through multiple models, effectively integrates the prediction results of the multi-task model and the additional classifier, significantly reduces the overfitting risk of a single model, and improves the robustness of the fault prediction results.

[0031] Furthermore, the present invention uses a multi-task learning model to jointly learn the failure characteristics of different hard disk models and age groups, making full use of the correlation between heterogeneous data, and effectively solving the problem of insufficient applicability of existing models in heterogeneous hard disk environments. The modular design of the architecture of the present invention makes it easy to integrate into existing cloud storage monitoring systems, and has high practical application value. The multi-task multi-layer stacking module adopts a distributed parallel computing framework to support the efficient processing of large-scale hard disk data, so that the present invention can run stably in an actual production environment. In addition, the method of the present invention does not rely on specific hardware or data sources, and can be flexibly deployed in a variety of data centers or storage environments, reducing implementation costs. By predicting potential hard disk failures in advance and taking preventive maintenance measures, the present invention can significantly reduce system downtime and data loss risks caused by hard disk failures.

[0032] Experimental results show that on the SSD dataset, the dynamic feature selection mechanism of the present invention improves the F1 score of the prediction model by more than 13% and the recall rate by 10%-15%. In addition, in a long-term hard disk environment, the present invention can significantly reduce the missed alarm rate and false alarm rate, ensuring the model's ability to dynamically capture hard disk failure characteristics.

[0033] On mixed data sets, the prediction accuracy of the present invention is more than 18% higher than that of traditional methods, especially when it involves hard disk data of multiple manufacturers and models, it still maintains high stability and accuracy. By sharing hidden layer features between tasks, the present invention enhances the model's adaptability to unknown hard disk model data and realizes efficient migration of prediction models.

[0034] In experiments, the false alarm rate of the present invention is reduced to less than 2%, while the existing model is usually between 3% and 5%. The voting integration mechanism effectively integrates the advantages of different classifiers, ensuring more stable prediction results and reducing extreme prediction situations caused by specific model deviations.

[0035] In the test scenario, the accuracy of the prediction of the present invention reduces the cost of system operation and maintenance by about 20%-30%. It can effectively reduce the emergency replacement and data recovery operations caused by sudden hard disk failures, and improve the overall operation efficiency and service quality of the data center.

[0036] Specifically, the target age group and the key feature list corresponding to the target age group are obtained through the following steps: S210, obtaining several sample SSDs and sample data of each sample SSD, preprocessing and standardizing the sample data, obtaining a sample feature value list of each sample SSD, wherein the sample feature value list includes values ​​of several preset initial features corresponding to the sample SSD, and the preset initial features are obtained based on the sample data. Specifically, the preset initial features include: usage time, hard disk model, read and write error rate, temperature, reallocatable sector count, etc. The preset initial features are features that reflect the operating status of the hard disk.

[0037] S220, obtaining a preset usage time period list, wherein the preset usage time period list includes a plurality of preset usage time periods, wherein time intervals of different preset usage time periods are the same.

[0038] S230, dividing the sample SSD based on the preset usage time period list and the hard disk model of the sample SSD, and obtaining a sample age group list set, wherein the sample age group list set includes a plurality of sample age group lists, and the sample age group list includes a plurality of sample SSDs in the same preset usage time period and with the same hard disk model; S240, based on the sample age group list, obtain the importance of each preset initial feature in each sample age group, sort them from large to small according to the importance, mark the first b preset initial features as intermediate features of the sample age group, thereby obtaining an intermediate feature list corresponding to each sample age group, wherein the intermediate feature list includes several intermediate features.

[0039] Specifically, b can be determined according to actual needs.

[0040] S250, obtaining the similarity coefficient of the intermediate feature lists of any two adjacent sample age groups. If the similarity coefficient is less than a preset similarity threshold, the two sample age groups are respectively used as two target age groups, and the intermediate feature lists corresponding to the two sample age groups are respectively used as key feature lists of the corresponding target age groups.

[0041] Specifically, the Jaccard coefficient of the intermediate feature lists of any two adjacent sample age groups is used as the similarity coefficient.

[0042] S260: If the similarity coefficient is not less than the preset similarity threshold, the two sample age groups are merged as a target age group, and the intermediate feature lists of the two sample age groups are merged as the key feature list corresponding to the target age group. By merging the two sample age groups and merging the two intermediate key feature lists, redundancy is reduced and the accuracy of key feature selection is improved.

[0043] In summary, several sample SSDs and sample data of each sample SSD are obtained, the sample data are preprocessed and standardized, a sample feature value list of each sample SSD is obtained, a preset usage time interval is obtained, the sample SSD is divided based on the preset usage time interval and the hard disk model, a sample age group list set is obtained, and based on the sample age group list, the importance of each preset initial feature in each sample age group is obtained, the importance in each sample age group is sorted from large to small, and the first b preset initial features are marked as intermediate features of the sample age group, so as to obtain the intermediate feature list corresponding to each sample age group, and obtain the similarity coefficient of the intermediate feature lists of any two adjacent sample age groups. If the similarity coefficient is less than the preset similarity threshold, the two sample age groups are respectively used as two target age groups, and the intermediate feature list corresponding to the target age group is used as the key feature list; if the similarity coefficient is not less than the preset similarity threshold, the two sample age groups are merged as a target age group, and the intermediate feature lists of the two sample age groups are merged as the key feature list corresponding to the target age group. The present invention obtains the key features more accurately by screening the preset initial features based on the importance.

[0044] Further, S220 obtains the time interval of the preset usage time period through the following steps: S221, based on the sample data, obtain the probability density function of the sample data by using a kernel density estimation method. Specifically, the probability density function is a function corresponding to the probability density estimation curve.

[0045] S222, based on the sample data, obtain the sample mutation point of the sample data through a change point detection algorithm.

[0046] Specifically, those skilled in the art know that any change point detection algorithm in the prior art belongs to the protection scope of the present invention, for example, the change point detection algorithm is a method based on machine learning.

[0047] S223, based on the probability density function and the sample mutation point, determine the time interval of the preset usage time period. Specifically, combine the probability density function and the sample mutation point to analyze the specific changes in the distribution of the sample data before and after the change point; determine the time interval of the preset usage time period based on the changes in the sample data, for example, use the usage time interval between two adjacent change points as the time interval of the preset usage time period.

[0048] Further, S230, based on the sample age group list, obtaining the importance of each preset initial feature in each sample age group includes: S231, obtaining an importance calculation method list, wherein the importance calculation method list includes a plurality of importance calculation methods.

[0049] The importance calculation method includes: a Spearman correlation coefficient calculation method, a Pearson correlation coefficient calculation method. Specifically, the importance calculation method also includes: a J-Index coefficient calculation method, a random forest calculation feature contribution method.

[0050] S232, obtaining the actual fault result of the sample SSD, and based on the actual fault result of the sample SSD and the sample feature value list of the sample SSD corresponding to the sample age group list, using an importance calculation method to obtain the importance of the preset initial feature.

[0051] Specifically, the average of the importances obtained by the multiple importance calculation methods is used as the importance of a preset initial feature. The importance of the preset initial feature is obtained by multiple importance calculation methods to improve accuracy and reliability.

[0052] Furthermore, the following steps are included to obtain the target multi-task learning model: S001. Based on the sample data of the sample SSD, obtain an intermediate SSD list corresponding to each target age group and an intermediate key feature value list corresponding to each intermediate SSD, wherein the intermediate SSD list includes several intermediate SSDs corresponding to sample SSDs in the same target age group, and the intermediate key feature value list includes target key feature values ​​of several candidate features corresponding to the intermediate SSD; wherein the candidate features are key features corresponding to the target age group corresponding to the usage time of the intermediate SSD.

[0053] S002, constructing an initial multi-task learning model, wherein the initial multi-task learning model includes a initial sub-learning models.

[0054] S003, based on the actual failure results of the intermediate SSDs and the intermediate key feature value list corresponding to each intermediate SSD, train the initial multi-task learning model to obtain the target multi-task learning model.

[0055] Specifically, the present invention uses a target multi-task learning model, the core idea of ​​which is to capture the correlation between different tasks by sharing hidden parameters and improve the generalization ability of the model. The shared hidden layer structure extracts common features, while retaining the exclusive features of each age group according to the key feature list, and training multiple initial sub-learning models.

[0056] Specifically, during the training process, the cross entropy loss function and Adam optimizer are used to jointly optimize the parameters of all tasks, so that the shared hidden layer can fully learn the common features while retaining the specific feature representation of each task, thus completing the training of the multi-task learning model.

[0057] An embodiment of the present invention also provides a non-transitory computer-readable storage medium, which can be set in an electronic device to store at least one instruction or at least one program related to implementing a method in a method embodiment. The at least one instruction or the at least one program is loaded and executed by the processor to implement the method provided in the above embodiment.

[0058] An embodiment of the present invention further provides an electronic device, comprising a processor and the aforementioned non-transitory computer-readable storage medium.

[0059] Although some specific embodiments of the present invention have been described in detail by way of example, it should be understood by those skilled in the art that the above examples are only for illustration, not for limiting the scope of the present invention. It should also be understood by those skilled in the art that various modifications may be made to the embodiments without departing from the scope and spirit of the present invention.

Claims

1. A SSD fault prediction method based on multi-task learning, characterized in that: The method comprises the following steps: S100, preprocessing and standardizing the data in the acquired hard disk log to obtain a plurality of initial SSDs and initial data of each initial SSD; S200, obtaining a target age group list and a key feature list corresponding to each target age group, wherein the target age group list includes a target age groups corresponding to different preset intervals of SSD usage time, and the key feature list includes a plurality of key features for predicting SSD failures, wherein the key features are obtained based on a hard disk log of the SSD; S300, based on the usage duration included in the initial data, a target SSD list corresponding to each target age group and a target key feature value list set corresponding to the target SSD list are obtained, wherein the target SSD list includes a plurality of target SSDs and the target SSDs in the same target SSD list are initial SSDs belonging to the same target age group, and the target key feature value list set includes a plurality of target key feature value lists, and the target key feature value list includes target key feature values ​​corresponding to a plurality of specified features of the target SSD, wherein the specified feature is a key feature corresponding to the target age group to which the target SSD belongs; S400, inputting the target key feature value list sets of the target SSD lists corresponding to the a target age groups into the a target sub-learning models included in the target multi-task learning model, respectively, to obtain the fault prediction results output by each target sub-learning model; and inputting the a fault prediction results as meta-features into the target meta-model to perform fault prediction again, to obtain the target stacking result; S500, obtaining a first other classifier and a second other classifier, inputting the initial data of the target SSD into the first other classifier for fault prediction, and obtaining a first other prediction result; and inputting the initial data of the target SSD into the second other classifier for fault prediction, and obtaining a second other prediction result; S600 , obtaining a final prediction result corresponding to each target SSD based on the target stacking result, the first other prediction result and the second other prediction result.

2. The SSD fault prediction method based on multi-task learning according to claim 1, characterized in that: Get the target age group and the list of key features corresponding to the target age group by following the steps below: S2 10. Obtain several sample SSDs and sample data of each sample SSD, preprocess and standardize the sample data, and obtain a sample feature value list of each sample SSD, wherein the sample feature value list includes values ​​of several preset initial features corresponding to the sample SSD, and the preset initial features are obtained based on the sample data; S220, obtaining a preset usage time period list, wherein the preset usage time period list includes a plurality of preset usage time periods, wherein time intervals of different preset usage time periods are the same; S230, dividing the sample SSD based on the preset usage time period list and the hard disk model of the sample SSD, and obtaining a sample age group list set, wherein the sample age group list set includes a plurality of sample age group lists, and the sample age group list includes a plurality of sample SSDs in the same preset usage time period and with the same hard disk model; S240, based on the sample age group list, obtaining the importance of each preset initial feature in each sample age group, sorting them in descending order according to the importance, marking the first b preset initial features as intermediate features of the sample age group, thereby obtaining an intermediate feature list corresponding to each sample age group, wherein the intermediate feature list includes a plurality of intermediate features; S250, obtaining a similarity coefficient of the intermediate feature lists of any two adjacent sample age groups, and if the similarity coefficient is less than a preset similarity threshold, taking the two sample age groups as two target age groups, and taking the intermediate feature lists corresponding to the two sample age groups as key feature lists corresponding to the target age groups; S260: If the similarity coefficient is not less than a preset similarity threshold, the two sample age groups are merged into a target age group, and the intermediate feature lists of the two sample age groups are merged into a key feature list corresponding to the target age group.

3. The SSD fault prediction method based on multi-task learning according to claim 2 is characterized in that: S220 obtains the time interval of the preset usage time period through the following steps: S221, based on the sample data, obtaining a probability density function of the sample data by a kernel density estimation method; S222, based on the sample data, obtaining a sample mutation point of the sample data by using a change point detection algorithm; S223: Determine the time interval of the preset usage time period based on the probability density function and the sample mutation point.

4. The SSD fault prediction method based on multi-task learning according to claim 2 is characterized in that: The Jaccard coefficient of the intermediate feature lists of any two adjacent sample age groups is taken as the similarity coefficient.

5. The SSD fault prediction method based on multi-task learning according to claim 2, characterized in that: In S230, based on the sample age group list, obtaining the importance of each preset initial feature in each sample age group includes: S231, obtaining a list of importance calculation methods, wherein the list of importance calculation methods includes a plurality of importance calculation methods; S232, obtaining the actual fault result of the sample SSD, and based on the actual fault result of the sample SSD and the sample feature value list of the sample SSD corresponding to the sample age group list, using an importance calculation method to obtain the importance of the preset initial feature.

6. The SSD fault prediction method based on multi-task learning according to claim 5, characterized in that: The importance calculation method includes: a Spearman correlation coefficient calculation method and a Pearson correlation coefficient calculation method.

7. The SSD fault prediction method based on multi-task learning according to claim 5, characterized in that: The following steps are also included to obtain the target multi-task learning model: S001, based on the sample data of the sample SSD, obtain an intermediate SSD list corresponding to each target age group and an intermediate key feature value list corresponding to each intermediate SSD, wherein the intermediate SSD list includes several intermediate SSDs corresponding to the sample SSDs in the same target age group, and the intermediate key feature value list includes target key feature values ​​of several candidate features corresponding to the intermediate SSD; wherein the candidate feature is a key feature corresponding to the target age group corresponding to the usage time of the intermediate SSD; S002, constructing an initial multi-task learning model, wherein the initial multi-task learning model includes a initial sub-learning models; S003, based on the actual failure results of the intermediate SSDs and the intermediate key feature value list corresponding to each intermediate SSD, train the initial multi-task learning model to obtain the target multi-task learning model.

8. The SSD fault prediction method based on multi-task learning according to claim 1, characterized in that: The first other classifier is the target GBDT classifier, and the second other classifier is the target XGBoost classifier.

9. A non-transitory computer-readable storage medium, wherein at least one instruction or at least one program is stored in the storage medium, characterized in that: The at least one instruction or the at least one program is loaded and executed by the processor to implement the SSD fault prediction method based on multi-task learning as described in any one of claims 1-8.

10. An electronic device, characterized in that: The invention comprises a processor and the non-transitory computer-readable storage medium as claimed in claim 9.

Images