Hard disk fault prediction method and system based on multi-task learning

Abstract

The invention relates to a hard disk fault prediction method and system based on multi-task learning, and the method comprises the steps: obtaining the hard disk information of a plurality of hard disks, and obtaining a training sample set according to the hard disk information; based on the representation learning network and the metric learning network, training through the training sample set to obtain a classification and regression dual-task hard disk fault prediction model; when fault prediction needs to be carried out on the target hard disk, taking hard disk information of the target hard disk as a to-be-predicted sample, and forming a sample input pair by the to-be-predicted sample and training samples in the training sample set; and inputting the sample input pair into the classification and regression dual-task hard disk fault prediction model to obtain a fault prediction result of the target hard disk. According to the method, the characterization learning network and the measurement learning network are combined, the characterization learning network learns characterization features of the hard disks so as to better capture potential fault signs, similarity and difference between the hard disks of the same manufacturer and the same model are learned and measured through the measurement learning network, and the risk of hard disk faults is predicted more accurately.

Description

Technical Field

[0001] The present invention relates to the field of storage technology, and in particular to a hard disk failure prediction method and system based on multi-task learning. Background Art

[0002] With the development of artificial intelligence, the total amount of global data storage has grown exponentially, which has put forward higher requirements on the reliability and availability of storage systems. Hard disks are currently the most widely used storage devices. Hard disk failure is a common problem in computer systems, which may lead to data loss and system downtime, causing serious losses to enterprises and individuals. Therefore, it is necessary to perceive the health status of hard disks in advance and provide early warnings before hard disk failures occur, so as to help maintenance personnel replace or perform other repair operations on disks that are about to fail in advance based on the prediction results of the model, thereby effectively avoiding the impact of equipment instability on business when failure is imminent, alleviating the pressure on operation and maintenance personnel, and reducing the investment in data repair.

[0003] At present, most studies use basic machine learning models, such as decision trees and random forests, which are easy to train but have poor generalization. Some studies use deep learning methods, such as BP neural networks and recurrent neural networks, and have achieved good results in the hard disk failure prediction task of some signals. However, the prediction effect is poor when facing a small sample hard disk data set, and it is impossible to use a general model to adapt to hard disks of different manufacturers and models. Summary of the invention

[0004] The present invention provides a hard disk failure prediction method and system based on multi-task learning, which combines a representation learning network and a metric learning network to train a classification and regression dual-task hard disk failure prediction model. The representation learning network learns the representation features of the hard disk to better capture potential failure signs, and uses the metric learning network to learn and measure the similarities and differences between hard disks of the same manufacturer and model, so as to more accurately predict the risk of hard disk failure.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: In a first aspect, a hard disk failure prediction method based on multi-task learning is provided, comprising: Obtain hard disk information of multiple hard disks, and obtain a training sample set according to the hard disk information; Based on the representation learning network and the metric learning network, a classification and regression dual-task hard disk failure prediction model is obtained through training sample sets; When it is necessary to predict the failure of the target hard disk, the hard disk information of the target hard disk is used as the sample to be predicted, and the sample to be predicted and the training sample in the training sample set form a sample input pair; The samples are input into the dual-task hard disk failure prediction model of classification and regression to obtain the failure prediction result of the target hard disk.

[0006] Furthermore, based on the representation learning network and the metric learning network, a classification and regression dual-task hard disk failure prediction model is obtained through training of the training sample set, including: A representation learning network is used to capture the hard disk general information of all training samples in the training sample set. The hard disk general information includes hard disk manufacturer information, hard disk model information and hard disk fault information. The hard disk fault information is used as a label for binary classification, and a fault classification task model is trained. Based on the triplet sample search strategy and the hard disk general information, triplet learning is performed on all training samples to construct a training sample pool; The metric learning network is used to perform metric learning on the training sample pool, and the triplet loss is optimized to train the fault regression task model. Based on the fault classification task model and the fault regression task model, a classification and regression dual-task hard disk fault prediction model is constructed.

[0007] Furthermore, based on the triplet sample search strategy and the hard disk general information, triplet learning is performed on all training samples to construct a training sample pool, including: Based on the sample search strategy of triples and the general information of hard disks, randomly select P triple training samples of different hard disk models from the same hard disk manufacturer from all training samples in the preset training batch; For the ternary training samples of each hard disk model, K positive samples and K negative samples are randomly selected as sample pairs of the ternary training samples; positive samples represent training samples with the same hard disk fault information as the ternary training samples, and negative samples represent training samples with different hard disk fault information from the ternary training samples; According to the ternary training samples and the corresponding sample pairs, a training sample pool is constructed.

[0008] Furthermore, a metric learning network is used to perform metric learning on the training sample pool, and the triplet loss is optimized to train the fault regression task model, including: The metric learning network is used to perform metric learning on the training sample pool to obtain the distance value between the ternary training sample and each training sample in the sample pair. According to the ternary training samples and distance values, triples are formed and the triplet loss is calculated; The fault regression task model is obtained based on the metric learning network and triplet loss training.

[0009] Furthermore, according to the ternary training samples and the distance values, triplets are formed and the triplet loss is calculated, including: According to the ternary training samples and distance values, the most difficult positive samples and the most difficult negative samples corresponding to the ternary training samples are selected to form a triplet; the most difficult positive sample is the positive sample with the farthest distance from the ternary training sample, and the most difficult negative sample is the negative sample with the closest distance to the ternary training sample; The triplet loss is calculated by the triplet loss function, and the triplet loss function expression is:

[0010] in, represents the triplet loss, represents the training sample set of failed hard disks, Represents the training sample set of healthy hard disks. When the ternary training sample a If the hard disk fails, p for training samples of failed hard disks, n for The training samples of healthy hard disks in m is the margin between the hardest positive sample and the hardest negative sample.

[0011] In the second aspect, a hard disk failure prediction system based on multi-task learning is provided, including: A sample acquisition module is used to acquire hard disk information of multiple hard disks and obtain a training sample set according to the hard disk information; A model training module is used to obtain a classification and regression dual-task hard disk failure prediction model based on a representation learning network and a metric learning network through training sample sets; The sample processing module is used to, when it is necessary to predict the failure of the target hard disk, use the hard disk information of the target hard disk as the sample to be predicted, and form a sample input pair with the sample to be predicted and the training sample in the training sample set; The fault prediction module is used to input the sample into the input classification and regression dual-task hard disk fault prediction model to obtain the fault prediction result of the target hard disk.

[0012] Furthermore, the model training module includes: Representation learning network unit, triple sample search strategy unit, metric learning network unit and model building unit; The representation learning network unit is used to capture the hard disk general information of all training samples in the training sample set by using the representation learning network. The hard disk general information includes hard disk manufacturer information, hard disk model information and hard disk fault information; the hard disk fault information is used as a label for binary classification, and a fault classification task model is obtained by training; A triplet sample search strategy unit is used to perform triplet learning on all training samples based on the triplet sample search strategy and hard disk general information to construct a training sample pool; A metric learning network unit is used to perform metric learning on the training sample pool using a metric learning network, and optimize the triplet loss to obtain a fault regression task model through training; The model building unit is used to build a classification and regression dual-task hard disk fault prediction model based on the fault classification task model and the fault regression task model.

[0013] Furthermore, The triplet sample search strategy unit is specifically used to randomly select P ternary training samples of different hard disk models of the same hard disk manufacturer from all training samples of a preset training batch based on the triplet sample search strategy and hard disk general information; for the ternary training samples of each hard disk model, randomly select K positive samples and K negative samples as sample pairs of the ternary training samples; the positive samples represent training samples with the same hard disk fault information as the ternary training samples, and the negative samples represent training samples with different hard disk fault information from the ternary training samples; a training sample pool is constructed based on the ternary training samples and the corresponding sample pairs.

[0014] Furthermore, The metric learning network unit is specifically used to perform metric learning on the training sample pool using the metric learning network, obtain the distance value between the ternary training sample and each training sample in the sample pair, form a triplet based on the ternary training sample and the distance value, and calculate the triplet loss; and obtain the fault regression task model based on the metric learning network and the triplet loss training.

[0015] Furthermore, The metric learning network unit is also used to select the most difficult positive sample and the most difficult negative sample corresponding to the ternary training sample according to the ternary training sample and the distance value to form a triplet; the most difficult positive sample is the positive sample with the farthest distance from the ternary training sample, and the most difficult negative sample is the negative sample with the closest distance to the ternary training sample; The triplet loss is calculated by the triplet loss function, and the triplet loss function expression is:

[0016] in, represents the triplet loss, represents the training sample set of failed hard disks, Represents the training sample set of healthy hard disks. When the ternary training sample a If the hard disk fails, p for training samples of failed hard disks, n for The training samples of healthy hard disks in m is the margin between the hardest positive sample and the hardest negative sample.

[0017] The beneficial effects achieved by the present invention are: Obtain the hard disk information of multiple hard disks, and obtain a training sample set based on the hard disk information; Based on the representation learning network and the metric learning network, obtain a classification and regression dual-task hard disk fault prediction model through training the training sample set; When it is necessary to predict the fault of the target hard disk, use the hard disk information of the target hard disk as the sample to be predicted, and form a sample input pair with the sample to be predicted and the training sample in the training sample set; Input the sample input pair into the classification and regression dual-task hard disk fault prediction model to obtain the fault prediction result of the target hard disk. Combining the representation learning network and the metric learning network, the representation learning network learns the representation features of the hard disk to better capture potential fault signs, and uses the metric learning network to learn and measure the similarities and differences between hard disks of the same manufacturer and model, so as to more accurately predict the risk of hard disk failure. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of the hard disk failure prediction method based on multi-task learning of the present invention; Figure 2 A schematic diagram of the training process of the classification and regression dual-task hard disk failure prediction model of the present invention; Figure 3 It is a structural diagram of the hard disk failure prediction system based on multi-task learning of the present invention; Figure 4 This is a structural diagram of the model training module of the present invention. DETAILED DESCRIPTION

[0019] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0020] like Figure 1 As shown, an embodiment of the present invention provides a hard disk failure prediction method based on multi-task learning, comprising: 101, obtaining hard disk information of multiple hard disks, and obtaining a training sample set according to the hard disk information; In this embodiment, the hard disk information generally includes information describing the identity of the hard disk, such as the manufacturer, model, etc.; in the present invention, the hard disk information also needs to include SMART data that is highly correlated with hard disk failures. SMART data is the full name of hard disk self-monitoring, analysis and reporting technology data, which is a data tool for recording the operating status of a hard disk drive (HDD) or a solid state drive (SSD). The hard disk SMART data can provide the health status, performance indicators and potential fault warnings of the hard disk, helping users or system administrators take measures before the hard disk has problems; In order to be applicable to hard disks of different manufacturers and models, during training, it is necessary to collect hard disk information of hard disks of as many manufacturers and models as possible on the market, and use the hard disk information of each hard disk as a training sample to form a training sample set.

[0021] 102. Based on the representation learning network and the metric learning network, a classification and regression dual-task hard disk failure prediction model is obtained through training sample sets; In this embodiment, the specific classification and regression dual-task hard disk failure prediction model training process is as follows: Figure 2 As shown, the steps include: 201, a representation learning network is used to capture the hard disk common information of all training samples in the training sample set; Representation Learning is a key concept in deep learning, which involves learning useful representations or features from raw data in order to better express the intrinsic structure and pattern of the data. The representation learning network is specifically a Long Short-Term Memory (LSTM) network. The LSTM network is a special recurrent neural network (RNN) that aims to solve the gradient vanishing or gradient exploding problems that occur when ordinary RNNs process long sequence data. By introducing memory units and gating mechanisms, the LSTM network can capture dependencies in long time series, and is particularly suitable for processing time series, natural language processing, and speech recognition tasks. Therefore, a representation learning network is used to capture the hard disk general information of all training samples in the training sample set. The hard disk general information includes hard disk manufacturer information, hard disk model information and hard disk fault information; 202, using the hard disk fault information as a label for binary classification, and training to obtain a fault classification task model; The hard disk fault information can determine whether the corresponding hard disk is a faulty hard disk or a healthy hard disk. Therefore, it is suitable for binary classification. Combined with the representation learning network training in step 201, a fault classification task model is obtained. 203, based on the triplet sample search strategy and the hard disk general information, triplet learning is performed on all training samples to construct a training sample pool; Assume that the batch size of the preset training batch is b , indicating that the preset training batch contains b training samples, based on the sample search strategy of triples and the general information of hard disks, randomly select P triple training samples of different hard disk models of the same hard disk manufacturer X from all the training samples in the preset training batch; For the ternary training samples of each hard disk model, K positive samples and K negative samples are randomly selected as sample pairs of the ternary training samples; positive samples represent training samples with the same hard disk fault information as the ternary training samples, and negative samples represent training samples with different hard disk fault information from the ternary training samples; assuming that the ternary training sample is a faulty hard disk, the training samples of the faulty hard disk in the sample pair are marked as Y=1, which indicates a positive sample; conversely, the training samples of the healthy hard disk in the sample pair are marked as Y=0, which indicates a negative sample; similarly, if the ternary training sample is a healthy hard disk, the training samples of the healthy hard disk in the sample pair are marked as Y=1, which indicates a positive sample; the training samples of the faulty hard disk in the sample pair are marked as Y=0, which indicates a negative sample; According to triple learning, the number of samples in each preset training batch is b* P * 2K, a training sample pool is constructed; it can be seen that the use of the triple sample search strategy effectively expands the number of samples in the original training batch, which can alleviate the problem of poor prediction effect when there are fewer sample hard disk data sets in the prior art; 204, using a metric learning network to perform metric learning on the training sample pool, and optimizing the triplet loss, and training to obtain a fault regression task model; The metric learning network also uses an LSTM network. Unlike the representation learning network, the metric learning network aims to learn the similarity of two different hard disk time series. The hard disk failure prediction task is no longer considered a classification task, but a clustering or sorting task. The metric learning network is used to perform metric learning on the training sample pool to obtain the distance value between the ternary training sample and each training sample in the sample pair. According to the ternary training samples and distance values, the most difficult positive samples and the most difficult negative samples corresponding to the ternary training samples are selected to form a triplet; the most difficult positive sample is the positive sample with the farthest distance from the ternary training sample, and the most difficult negative sample is the negative sample with the closest distance to the ternary training sample; The triplet loss is calculated by the triplet loss function, and the triplet loss function expression is:

[0022] in, represents the triplet loss, represents the training sample set of failed hard disks, Represents the training sample set of healthy hard disks. When the ternary training sample a If the hard disk fails, p for training samples of failed hard disks, n for The training samples of healthy hard disks in m is the margin between the hardest positive sample and the hardest negative sample; The distance between two failed hard disks of different models from the same manufacturer will be smaller than the distance between a failed hard disk and a healthy hard disk of the same model. In this way, the failure modes of hard disks from the same manufacturer can be learned and distinguished from healthy hard disks. The fault regression task model is obtained based on the metric learning network and triplet loss training.

[0023] 205. Based on the fault classification task model and the fault regression task model, a classification and regression dual-task hard disk fault prediction model is constructed.

[0024] The models formed in the above steps 202 and 204 are combined to obtain a classification and regression dual-task hard disk failure prediction model.

[0025] 103, when it is necessary to predict the failure of the target hard disk, the hard disk information of the target hard disk is used as a sample to be predicted, and the sample to be predicted and a training sample in the training sample set form a sample input pair; 104 , input the sample into the dual-task hard disk failure prediction model of input classification and regression to obtain the failure prediction result of the target hard disk.

[0026] In this embodiment, the process of performing prediction includes: (1) For all training samples marked as faulty hard disks of the same manufacturer and hard disk model as the sample to be predicted, form sample input pairs; if any sample in the fault prediction result is predicted to be similar, then mark the sample to be predicted as a faulty hard disk. Otherwise, proceed to the next step (2); (2) In view of the sparsity of the healthy disk sample status, in order to improve the decision efficiency, first randomly select 10% healthy hard disks of the same manufacturer and the same hard disk model, and then randomly collect these healthy hard disk samples at intervals of one month, and form input sample pairs with the samples to be predicted; if all these sample pairs are predicted to be similar, then mark the sample to be predicted as a healthy hard disk. Otherwise, execute the next step (3); (3) Compare multiple fault samples of different hard disk models from the same manufacturer with the sample to be predicted; if more than half of the prediction results are similar, mark the sample to be predicted as a faulty hard disk; (4) Compare multiple fault samples of the same hard disk model from the same manufacturer with the sample to be predicted; if the disk model to which the sample to be predicted belongs does not exist in the sample pool, directly execute (3).

[0027] The beneficial effects of the embodiments of the present invention are: Obtain the hard disk information of multiple hard disks, and obtain a training sample set based on the hard disk information; based on the representation learning network and the metric learning network, obtain a classification and regression dual-task hard disk fault prediction model through training the training sample set; when it is necessary to predict the fault of the target hard disk, use the hard disk information of the target hard disk as the sample to be predicted, and form a sample input pair with the sample to be predicted and the training sample in the training sample set; input the sample input pair into the classification and regression dual-task hard disk fault prediction model to obtain the fault prediction result of the target hard disk. The representation learning network and the metric learning network are combined. The representation learning network learns the representation features of the hard disk to better capture potential fault signs, and uses the metric learning network to learn and measure the similarities and differences between hard disks of the same manufacturer and model to more accurately predict the risk of hard disk failure; The triple sample search strategy is adopted to effectively expand the number of samples in the original training batch, which can alleviate the problem of poor prediction effect when there are fewer sample hard disk data sets in the existing technology.

[0028] In combination with the hard disk failure prediction method based on multi-task learning described in the above embodiments, a hard disk failure prediction system based on multi-task learning is described below through embodiments.

[0029] like Figure 3 As shown, an embodiment of the present invention provides a hard disk failure prediction system based on multi-task learning, including: The sample acquisition module 301 is used to acquire hard disk information of multiple hard disks and obtain a training sample set according to the hard disk information; A model training module 302 is used to obtain a classification and regression dual-task hard disk failure prediction model through training sample sets based on a representation learning network and a metric learning network; The sample processing module 303 is used to, when it is necessary to predict the failure of the target hard disk, use the hard disk information of the target hard disk as the sample to be predicted, and form a sample input pair with the sample to be predicted and the training sample in the training sample set; The fault prediction module 304 is used to input the sample into the classification and regression dual-task hard disk fault prediction model to obtain the fault prediction result of the target hard disk.

[0030] Combination Figure 3 The embodiment shown is preferably as Figure 4 As shown, in some embodiments of the present invention, the model training module 302 includes: Representation learning network unit 401, triple sample search strategy unit 402, metric learning network unit 403 and model building unit 404; The representation learning network unit 401 is used to capture the hard disk general information of all training samples in the training sample set by using the representation learning network, where the hard disk general information includes hard disk manufacturer information, hard disk model information and hard disk fault information; the hard disk fault information is used as a label for binary classification, and a fault classification task model is obtained by training; A triplet sample search strategy unit 402, used to perform triplet learning on all training samples based on the triplet sample search strategy and hard disk general information to construct a training sample pool; The metric learning network unit 403 is used to perform metric learning on the training sample pool using the metric learning network, and optimize the triplet loss to obtain the fault regression task model through training; The model building unit 404 is used to build a classification and regression dual-task hard disk fault prediction model based on the fault classification task model and the fault regression task model.

[0031] Combination Figure 4 The embodiments shown are preferably some embodiments of the present invention. The ternary sample search strategy unit 402 is specifically used to randomly select P ternary training samples of different hard disk models of the same hard disk manufacturer from all training samples of a preset training batch based on the ternary sample search strategy and the general information of the hard disk; for the ternary training samples of each hard disk model, randomly select K positive samples and K negative samples as sample pairs of the ternary training samples; the positive samples represent training samples with the same hard disk fault information as the ternary training samples, and the negative samples represent training samples with different hard disk fault information from the ternary training samples; a training sample pool is constructed based on the ternary training samples and the corresponding sample pairs.

[0032] Combination Figure 4 The embodiments shown are preferably some embodiments of the present invention. The metric learning network unit 403 is specifically used to perform metric learning on the training sample pool using the metric learning network, obtain the distance value between the ternary training sample and each training sample in the sample pair, form a triplet according to the ternary training sample and the distance value, and calculate the triplet loss; obtain the fault regression task model according to the metric learning network and the triplet loss training; The metric learning network unit 403 is further used to select the most difficult positive sample and the most difficult negative sample corresponding to the ternary training sample according to the ternary training sample and the distance value to form a triplet; the most difficult positive sample is the positive sample farthest from the ternary training sample, and the most difficult negative sample is the negative sample closest to the ternary training sample; The triplet loss is calculated by the triplet loss function, and the triplet loss function expression is:

[0033] in, represents the triplet loss, represents the training sample set of failed hard disks, Represents the training sample set of healthy hard disks. When the ternary training sample a If the hard disk fails, p for training samples of failed hard disks, n for The training samples of healthy hard disks in m is the margin between the hardest positive sample and the hardest negative sample.

[0034] In summary, the beneficial effects of the hard disk failure prediction system based on multi-task learning in the embodiment of the present invention are: It combines a representation learning network and a metric learning network. The representation learning network learns the representation features of the hard disk to better capture potential failure signs, and the metric learning network learns to measure the similarities and differences between hard disks of the same manufacturer and model to more accurately predict the risk of hard disk failure. The triple sample search strategy is adopted to effectively expand the number of samples in the original training batch, which can alleviate the problem of poor prediction effect when there are fewer sample hard disk data sets in the existing technology.

[0035] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete 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 disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0036] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, 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 flowchart and / or block diagram. 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.

[0037] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0038] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0039] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are included in the scope of the claims of the present invention to be approved.

Claims

1. A hard disk failure prediction method based on multi-task learning, characterized in that: include: Obtaining hard disk information of multiple hard disks, and obtaining a training sample set according to the hard disk information; Based on the representation learning network and the metric learning network, a classification and regression dual-task hard disk failure prediction model is obtained through training of the training sample set; When it is necessary to predict the failure of the target hard disk, the hard disk information of the target hard disk is used as a sample to be predicted, and the sample to be predicted and the training sample in the training sample set form a sample input pair; The sample input pair is input into the classification and regression dual-task hard disk fault prediction model to obtain a fault prediction result of the target hard disk.

2. The hard disk failure prediction method based on multi-task learning according to claim 1 is characterized in that: The dual-task hard disk failure prediction model of classification and regression is obtained by training the training sample set based on the representation learning network and the metric learning network, including: Using a representation learning network to capture hard disk general information of all training samples in the training sample set, the hard disk general information includes hard disk manufacturer information, hard disk model information and hard disk fault information; The hard disk fault information is used as a label for binary classification, and a fault classification task model is obtained by training; Based on the triplet sample search strategy and the hard disk general information, triplet learning is performed on all training samples to construct a training sample pool; A metric learning network is used to perform metric learning on the training sample pool, and the triplet loss is optimized to obtain a fault regression task model through training; Based on the fault classification task model and the fault regression task model, a classification and regression dual-task hard disk fault prediction model is constructed.

3. The hard disk failure prediction method based on multi-task learning according to claim 2 is characterized in that: The triplet-based sample search strategy and the hard disk general information perform triplet learning on all training samples to construct a training sample pool, including: Based on the sample search strategy of the triplet and the general information of the hard disk, randomly select P triplet training samples of different hard disk models of the same hard disk manufacturer from all training samples of the preset training batch; For the ternary training samples of each hard disk model, K positive samples and K negative samples are randomly selected as sample pairs of the ternary training samples; the positive samples represent training samples with the same hard disk fault information as the ternary training samples, and the negative samples represent training samples with different hard disk fault information from the ternary training samples; A training sample pool is constructed based on the ternary training samples and the corresponding sample pairs.

4. The hard disk failure prediction method based on multi-task learning according to claim 3 is characterized in that: The metric learning network is used to perform metric learning on the training sample pool, and the triplet loss is optimized to obtain a fault regression task model through training, including: A metric learning network is used to perform metric learning on the training sample pool to obtain a distance value between the ternary training sample and each training sample in the sample pair. According to the ternary training samples and the distance values, a triplet is formed, and a triplet loss is calculated; A fault regression task model is obtained according to the metric learning network and the triplet loss training.

5. The hard disk failure prediction method based on multi-task learning according to claim 4 is characterized in that: The forming of a triplet according to the ternary training sample and the distance value and calculating the triplet loss comprises: According to the ternary training samples and the distance values, the most difficult positive sample and the most difficult negative sample corresponding to the ternary training samples are selected to form a triplet; the most difficult positive sample is the positive sample farthest from the ternary training samples, and the most difficult negative sample is the negative sample closest to the ternary training samples; The triplet loss is calculated by the triplet loss function, and the triplet loss function expression is: ； Among them, the represents the triple loss, represents a training sample set of faulty hard disks, Represents a training sample set of healthy hard disks. When the ternary training sample a If the hard disk fails, p For the The training sample of the failed hard disk in n For the The training samples of healthy hard disks in m is the margin between the most difficult positive sample and the most difficult negative sample.

6. A hard disk failure prediction system based on multi-task learning, characterized in that: include: A sample acquisition module, used to acquire hard disk information of multiple hard disks, and obtain a training sample set according to the hard disk information; A model training module, used for obtaining a classification and regression dual-task hard disk failure prediction model through training of the training sample set based on the representation learning network and the metric learning network; A sample processing module, used for, when it is necessary to predict the failure of a target hard disk, using the hard disk information of the target hard disk as a sample to be predicted, and forming a sample input pair with the sample to be predicted and a training sample in the training sample set; The fault prediction module is used to input the sample input pair into the classification and regression dual-task hard disk fault prediction model to obtain the fault prediction result of the target hard disk.

7. The hard disk failure prediction system based on multi-task learning according to claim 6 is characterized in that: The model training module includes: Representation learning network unit, triple sample search strategy unit, metric learning network unit and model building unit; The representation learning network unit is used to capture the hard disk general information of all training samples in the training sample set by using the representation learning network, wherein the hard disk general information includes hard disk manufacturer information, hard disk model information and hard disk fault information; the hard disk fault information is used as a label for binary classification, and a fault classification task model is obtained by training; The triplet sample search strategy unit is used to perform triplet learning on all training samples based on the triplet sample search strategy and the hard disk general information to construct a training sample pool; The metric learning network unit is used to perform metric learning on the training sample pool using a metric learning network, and optimize the triplet loss to obtain a fault regression task model through training; The model building unit is used to build a classification and regression dual-task hard disk fault prediction model based on the fault classification task model and the fault regression task model.

8. The hard disk failure prediction system based on multi-task learning according to claim 7 is characterized in that: The triple sample search strategy unit is specifically used to randomly select P triple training samples of different hard disk models of the same hard disk manufacturer from all training samples of a preset training batch based on the triple sample search strategy and the hard disk general information; For the ternary training samples of each hard disk model, K positive samples and K negative samples are randomly selected as sample pairs of the ternary training samples; the positive samples represent training samples with the same hard disk fault information as the ternary training samples, and the negative samples represent training samples with different hard disk fault information from the ternary training samples; a training sample pool is constructed based on the ternary training samples and the corresponding sample pairs.

9. The hard disk failure prediction system based on multi-task learning according to claim 8, characterized in that: The metric learning network unit is specifically used to perform metric learning on the training sample pool using a metric learning network to obtain a distance value between the ternary training sample and each training sample in the sample pair, form a triplet based on the ternary training sample and the distance value, and calculate the triplet loss; and obtain a fault regression task model based on the metric learning network and the triplet loss training.

10. The hard disk failure prediction system based on multi-task learning according to claim 9, characterized in that: The metric learning network unit is further used to select the most difficult positive sample and the most difficult negative sample corresponding to the ternary training sample according to the ternary training sample and the distance value to form a triplet; the most difficult positive sample is the positive sample farthest from the ternary training sample, and the most difficult negative sample is the negative sample closest to the ternary training sample; The triplet loss is calculated by the triplet loss function, and the triplet loss function expression is: ； Among them, the represents the triple loss, represents a training sample set of faulty hard disks, Represents a training sample set of healthy hard disks. When the ternary training sample a If the hard disk fails, p For the The training sample of the failed hard disk in n For the The training samples of healthy hard disks in m is the margin between the most difficult positive sample and the most difficult negative sample.

Images