Prediction system, prediction method and program

Abstract

To provide a method of predicting an operating life and a fault of a storage device with high accuracy by using time-sequential S.M.A.R.T data.SOLUTION: A prediction system comprises: means which, using first training data which are time-sequential S.M.A.R.T. data collected by a plurality of kinds of storage devices, stores a general-purpose prediction model constructed by learning a relation between transition of a value included in the first training data and a residual durable period of the storage devices, and a specific prediction model constructed by learning a relation between transition of a value included in second training data and the residual durable period of the storage device of the type related to the second training data by using the second training data which are the time-sequential S.M.A.R.T. data of each type of the storage device; and means which predicts the residual durable period of the storage device to be predicted, on the basis of the time-sequential S.M.A.R.T. data collected by the storage device to be predicted, and the specific prediction model related to the storage device of the same type as the storage device to be predicted.SELECTED DRAWING: Figure 6

Description

[Technical Field]

[0001] The present disclosure relates to a prediction system, a prediction method, and a program. [Background technology]

[0002] Storage devices (hereinafter referred to as drives) such as hard disk drives (HDDs) and solid state drives (SSDs) are generally guaranteed to operate for a specified period (e.g., five years). However, depending on the environment and conditions of use, the drive may fail even during the warranty period. For this reason, operation and maintenance systems are often designed with the assumption that failures will occur. If the accuracy of predicting drive failure times and lifespans can be improved, unexpected maintenance due to sudden drive failures can be reduced. In other words, improving the accuracy of predicting RUL (Remaining Useful Life) contributes to improving operational stability, which is one of the most important expectations for drives.

[0003] SMART (Self-Monitoring, Analysis, and Reporting Technology) data is known as data used to monitor drive operating status and analyze failure causes. SMART is a self-diagnostic function built into drive products such as HDDs and SSDs. The SMART function allows the drive to self-diagnose various test items in real time and quantify its own condition. The results are then recorded as SMART data. SMART data includes information such as spin-up time, operating time, number of motor rotations / stops, number of bad sectors replaced, and total amount of written data. SMART data is periodically collected time-series data and can be used to capture phenomena that change over time. For example, it can be used to predict drive failure timing and lifespan. Patent Document 1 discloses a technology for evaluating drive health and predicting failures by comparing multiple parameters included in SMART data with thresholds set for each parameter.

[0004] When utilizing SMART data, it is common to use publicly available, general-purpose SMART data. General-purpose SMART data includes data on various types of drives and various vendors. The content, definitions, and indicators of SMART data generally differ depending on the type of drive and vendor. Therefore, even if general-purpose SMART data is used to identify the cause of a failure, evaluate the health status, predict failures, etc., there is a high possibility that sufficient accuracy will not be achieved. Patent Document 1 also does not disclose a solution to this problem. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Patent No. 7281854 Summary of the Invention [Problem to be solved by the invention]

[0006] One of the objectives is to provide a method for predicting the lifespan and failure of storage devices with high accuracy using SMART data. [Means for solving the problem]

[0007] According to one aspect of the present disclosure, a prediction system includes a means for storing a general-purpose prediction model constructed by using first training data, which is time-series SMART data collected from multiple types of storage devices, to learn the relationship between the trends in values ​​contained in the first training data and the remaining useful life of the storage devices, and a specialized prediction model constructed by using second training data, which is time-series SMART data for each type of storage device, to learn the relationship between the trends in values ​​contained in the second training data and the remaining useful life of storage devices of the type related to the second training data; and a means for predicting the remaining useful life of a storage device to be predicted based on the time-series SMART data collected from the storage device to be predicted and the specialized prediction model for storage devices of the same type as the storage device to be predicted.

[0008] According to one aspect of the present disclosure, the prediction method includes: a general-purpose prediction model constructed by using first training data, which is time-series SMART data collected from multiple types of storage devices, to learn the relationship between the trends in values ​​contained in the first training data and the remaining useful life of the storage devices; and a specialized prediction model constructed by using second training data, which is time-series SMART data for each type of storage device, to learn the relationship between the trends in values ​​contained in the second training data and the remaining useful life of the type of storage device related to the second training data; and obtains the specialized prediction model for a storage device of the same type as the storage device to be predicted from a means for storing the general-purpose prediction model; and predicts the remaining useful life of the storage device to be predicted based on the obtained specialized prediction model and the time-series SMART data collected from the storage device to be predicted.

[0009] According to one aspect of the present disclosure, a computer is caused to acquire from a means for storing: a general-purpose prediction model constructed by using first training data, which is time-series SMART data collected from multiple types of storage devices, to learn the relationship between the trends in values ​​contained in the first training data and the remaining useful life of the storage devices; and a specialized prediction model constructed by using second training data, which is time-series SMART data for each type of storage device, to learn the relationship between the trends in values ​​contained in the second training data and the remaining useful life of the type of storage device related to the second training data; and to perform a process of predicting the remaining useful life of the storage device to be predicted based on the acquired specialized prediction model and the time-series SMART data collected from the storage device to be predicted. [Effects of the Invention]

[0010] According to the present disclosure, it is possible to use SMART data to predict the lifespan and failure of storage devices with high accuracy. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram illustrating an example of the configuration of a prediction system according to an embodiment. [Figure 2] FIG. 10 is a diagram illustrating a process for constructing a prediction model according to an embodiment. [Figure 3] 10 is a flowchart illustrating an example of a prediction process according to the embodiment. [Figure 4] 10 is a flowchart illustrating an example of a risk assessment of an SDD according to an embodiment. [Figure 5] 10 is a flowchart illustrating an example of a maintenance process according to the embodiment. [Figure 6] FIG. 2 is a second diagram illustrating an example of the configuration of the prediction system according to the embodiment. [Figure 7] 1 is a flowchart illustrating an example of a prediction method according to an embodiment. [Figure 8] FIG. 2 is a diagram illustrating an example of a hardware configuration of the prediction system according to the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0012] Hereinafter, a prediction system according to each embodiment of the present disclosure will be described with reference to the drawings. In the drawings used in the following description, the description of parts not related to the present disclosure may be omitted or not shown. The same or equivalent parts in all drawings will be denoted by the same reference numerals, and common descriptions may be omitted.

[0013] First Embodiment (System Configuration) 1 is a first block diagram showing an example of the configuration of a prediction system according to the first embodiment. The prediction system 100 includes a storage device 1, a management server 2, and a cloud service 3.

[0014] (Storage Device Configuration) The storage device 1 is a storage device such as a primary storage or archive storage, and has functions such as data accumulation, management, and notification to users. The storage device 1 includes a drive connection unit 11 and a controller 12. The drive connection unit 11 has a number of drive interfaces and has the function of connecting various drives (for example, HDD1 to HDD3 and SDD1 to SDD2 shown in the figure) to the controller 12. A plurality of HDDs and SSDs connected to the drive connection unit 11 may be recognized as a collection of multiple drives. The drives in the drive connection unit 11 may include drives of various vendors, types, and capacities.

[0015] The controller 12 includes a data collection unit 121, a communication unit 122, and a drive operation unit 123. The controller 12 includes hardware such as a CPU (Central Processing Unit), memory, and a communication interface (not shown), and is equipped with management software that provides management functions for the storage device 1. The data collection unit 121, the communication unit 122, and the drive operation unit 123 are realized by this hardware and management software.

[0016] The data collection unit 121 is connected to the drive connection unit 11 and periodically collects basic information such as the vendor, model number (identification number), capacity, and administrator of the drive in operation, as well as SMART data. The frequency of collecting SMART data can be adjusted according to the required prediction accuracy.

[0017] The communication unit 122 is communicably connected to the management server 2 through a communication interface (not shown) of the controller 12. The communication unit 122 transmits data such as SMART data from the storage device 1 to the management server 2. The communication unit 122 receives risk analysis data and drive risk management policies transmitted from the management server 2 to the storage device 1.

[0018] The drive operation unit 123 acquires the risk analysis data and the drive risk management policy from the communication unit 122 and automatically performs backup operations and RAID (Redundant Array of Inexpensive Disks) rebuild operations on drives with a high risk of failure. The drive risk management policy specifies the values ​​required for performing a backup or rebuild, based on various index values ​​and failure risks indicating the drive status contained in the risk analysis data. Alternatively, the drive operation unit 123 may acquire command information for a backup or rebuild from the management server 2 and perform an automatic backup or rebuild based on this command information.

[0019] (Administration Server Configuration) The management server 2 is a server connected to one or more storage devices 1, and has functions such as monitoring the storage devices 1, analyzing time-series SMART data, and determining failure risks. The management server 2 includes a controller 21 and a model management unit 22.

[0020] The controller 21 includes a risk determination module 211, a drive information management unit 212, and an SSD power-off time statistics unit 213. The controller 21 includes hardware such as a CPU, memory, an inference machine learning accelerator, and a communication interface (not shown), and is equipped with management software that provides risk determination and machine learning model management functions. The risk determination module 211, the drive information management unit 212, and the SSD power-off time statistics unit 213 are realized by this hardware and management software.

[0021] The risk determination module 211 includes a data management unit 211a, an RUL prediction unit 211b, and a risk determination unit 211c. The data management unit 211a acquires basic information (such as the drive identification number) and time-series SMART data of the target drive from the drive information management unit 212. Then, the data management unit 211a acquires the specialized prediction model 222 or the general-purpose prediction model 221 corresponding to the target drive from the model management unit 22.

[0022] The RUL prediction unit 211b uses an inference machine learning accelerator to predict the RUL of the prediction target drive based on the time-series SMART data and RUL prediction model acquired from the data management unit 211a.

[0023] The risk determination unit 211c analyzes the prediction results from the RUL prediction unit 211b, performs risk analysis in accordance with a drive risk management policy set in advance by the user, and transmits the analysis results to the communication unit 214. The drive risk management policy set in advance by the user is stored in the risk determination unit 211c. The drive risk management policy includes the conditions for implementing automatic backup operations, but these implementation conditions can be set arbitrarily by the user.

[0024] The drive information management unit 212 acquires time-series SMART data of the drives it manages via the communication unit 214 of the management server 2, and accumulates and stores the SMART data.

[0025] The SSD power-off time statistics unit 213 includes a power-off data management unit 213a and a risk determination unit 213b. The power-off data management unit 213a records and stores the power-on timestamp of each SSD drive. To enable information sharing between other facilities and multiple locations, the power-off data management unit 213a uploads the power-on timestamp to the cloud database 31. The risk determination unit 213b periodically monitors the power-off time of the managed drives, and if it determines that there is a possibility of data loss based on the drive risk management policy, it notifies the storage device 1 via the communication unit 214 that there is an SSD drive that may be at risk of data loss. The drive risk management policy sets the length of time that an SSD drive must be powered off before it is determined that there is a possibility of data loss.

[0026] The model management unit 22 is a storage location for RUL prediction models for predicting drive failure timing and lifespan. The RUL prediction models include a general-purpose prediction model 221 and a specialized prediction model 222 specialized for a particular product. The general-purpose prediction model 221 is a model constructed by using published time-series SMART data for various drives (each vendor, each model, each capacity) to learn, through machine learning and deep learning, the relationship between the trends in various indicator values ​​indicating the drive status included in the SMART data and the drive's lifespan and the time until failure. The specialized prediction model 222 is a model constructed by using time-series SMART data for the same drive as the drive installed in the storage device 1 to learn the relationship between the trends in various indicator values ​​included in the SMART data and the drive's lifespan and the time until failure. SMART data can be classified into, for example, (1) publicly available time-series SMART data regarding drive failures, (2) SMART data collected periodically until a drive fails, and (3) SMART data collected in real time during operation. Data (1) is used to train the general-purpose prediction model 221, and data (2) and (3) are used to train the specialized prediction model 222. The model management unit 22 stores at least the general-purpose prediction model 221. The general-purpose prediction model 221 is acquired from the cloud database 31. Furthermore, when predicting the RUL of a drive, the model management unit 22 communicates with the cloud database 31 to check whether a specialized prediction model 222 corresponding to the drive to be predicted installed in the storage device 1 exists. If a specialized prediction model 222 exists, the model management unit 22 acquires and stores the specialized prediction model 222 from the cloud database 31.

[0027] The cloud service 3 is a service provided by a vendor of the storage device 1 or the like. The cloud service 3 includes a cloud database 31 and a prediction model learning unit 32. The cloud service 3 has functions for building and sharing RUL prediction models, collecting and managing data on SSD power-on times, and collecting and managing time-series SMART data for various drives. The cloud database 31 stores the RUL prediction models (general-purpose prediction model 221, drive-specific specialized prediction model 222), SSD power-on time data, SMART data, and the like. The predictive model learning unit 32 uses time-series SMART data on drive failures obtained from multiple information sources to build a RUL predictive model. The model building process will now be described with reference to FIG.

[0028] (Building a RUL prediction model) Next, a two-stage process for constructing an RUL prediction model will be described with reference to Fig. 2. As shown in Fig. 2, in this embodiment, construction of an RUL prediction model is divided into two stages: a "general-purpose prediction model" and a "specialized prediction model." Construction of the RUL prediction model is performed by a prediction model learning unit 32 of a cloud service 3 provided by a vendor of the storage device 1 or the like.

[0029] (A) Construction of a general-purpose prediction model The prediction model training unit 32 collects time-series SMART data (data (1) above) related to drive failures of various types and capacities from data sources published by vendors and other organizations, and creates a pre-training dataset DSA. The prediction model training unit 32 uses the dataset DSA to train a neural network and construct a general-purpose prediction model MA (corresponding to the general-purpose prediction model 221 in Figure 1). For example, when time-series SMART data collected from the drive to be predicted is input to the general-purpose prediction model MA, the general-purpose prediction model MA is trained to predict and output the drive's lifespan and time to failure (remaining useful life: RUL). By learning from a large amount of failure time-series data, basic RUL prediction capabilities for various types of drives are developed. The neural network employed may have architectures such as, but not limited to, LSTM, CNN, RNN, and GAN. The general-purpose prediction model MA may be retrained and updated as new pre-training datasets DSA are collected, or may be retrained and updated periodically.

[0030] (B) Building a product-specific forecasting model Vendors and the like conduct durability tests on various drives before shipping them and record the time-series SMART data (data (2) above) up to the time of drive failure in the cloud database 31. The prediction model training unit 32 uses this to create a verification dataset SDS1 for each type of drive. The prediction model training unit 32 also collects the time-series SMART data (data (3) above) up to the time of drive failure reported by users and stored in the cloud database 31, and creates a performance dataset SDS2. The prediction model training unit 32 combines two datasets, SDS1 and SDS2, collected for the same type of drive to create a fine-tuning dataset DSB dedicated to each type of drive. The DSB does not necessarily have to be composed of both SDS1 and SDS2. Depending on the data collection situation, a DBS may be composed of either SDS1 or SDS2 alone. The prediction model training unit 32 constructs a specialized prediction model MB (corresponding to the specialized prediction model 222 in FIG. 1 ) by adding, for example, a neural network layer based on the general-purpose prediction model MA and performing additional training (e.g., transfer learning and fine-tuning) using the fine-tuning dataset DSB. For example, when a specialized prediction model MB receives time-series SMART data for a drive corresponding to this model MB, it outputs a prediction of the drive's lifespan and time to failure (remaining useful life). The specialized prediction model MB has high RUL prediction accuracy because it has been retrained with a dataset for a specific product in addition to the general-purpose prediction capabilities of the general-purpose prediction model MA. The specialized prediction model MB may be retrained and updated as new validation datasets SDS1 and new performance datasets SDS2 are collected, or it may be retrained and updated periodically.

[0031] (operation) (RUL prediction and automatic data integrity processing) Next, the process flow for predicting the RUL of a drive and automatic data preservation will be described with reference to FIG. 3. When predicting the RUL of an operating drive, the controller 12 of the storage device 1 acquires basic information (such as the identification number) and time-series SMART data of the drive to be predicted (step S01) and sends them to the management server 2 (step S02). In the management server 2, the controller 21 receives the basic information and time-series SMART data of the drive to be predicted from the communication unit 122, stores the received data in the drive information management unit 212, and also performs product identification of the drive to identify the drive type, etc. The controller 21 passes information such as the identified drive type to the model management unit 22. Based on the drive type, etc. received from the controller 21, the model management unit 22 checks whether a specialized prediction model 222 for the same type of drive is stored in the cloud database 31 (step S03).

[0032] If a specialized prediction model 222 corresponding to the drive to be predicted exists (step S03; Yes), the model management unit 22 acquires and stores the specialized prediction model 222. The data management unit 211a acquires the specialized prediction model 222 from the model management unit 22 and acquires time-series SMART data of the drive to be predicted from the drive information management unit 212. Then, the RUL prediction unit 211b of the risk determination module 211 inputs the time-series SMART data into the specialized prediction model 222 to predict a highly accurate RUL, and the risk determination unit 211c performs risk determination based on the RUL prediction result (step S04). For example, the RUL prediction unit 211b predicts the time until a failure occurs or the end of the life (remaining useful life). The risk determination unit 211c evaluates the predicted remaining useful life with reference to the drive risk management policy. Various evaluation methods may be used. For example, the drive risk management policy may have evaluation criteria set to classify the risk into high (urgent), medium, or low according to the remaining useful life, and the risk determination unit 211c may evaluate the risk of the drive to be predicted as high (urgent), medium, or low based on this evaluation criteria and the predicted remaining useful life. Also, for example, if the predicted RUL value becomes equal to or less than the set value of the risk management policy (for example, 30% of the value at the start of drive use), it may be determined that there is a risk, and if it becomes equal to or less than 10% of the value at the start of drive use, it may be determined that there is a risk of data loss.

[0033] If the specialized prediction model 222 does not exist (step S03; No), the model management unit 22 acquires and stores the general-purpose prediction model 221. The data management unit 211a acquires the general-purpose prediction model 221 from the model management unit 22, and acquires time-series SMART data of the drive to be predicted from the drive information management unit 212. Then, the RUL prediction unit 211b of the risk determination module 211 inputs the time-series SMART data into the general-purpose prediction model 221 and predicts the RUL, and the risk determination unit 211c performs risk determination based on the RUL prediction result (step S05). The method of risk determination is the same as in step S04.

[0034] The controller 21 then determines the drive type (step S06). If the target drive is an SSD drive (step S06; No), the power-off data management unit 213a records a power-on timestamp by referencing the chronological SMART data and uploads the recorded power-on timestamp to the cloud database 31. This allows information on the power-off timestamp to be shared among equipment at multiple locations. The risk determination unit 213b also investigates the power-off time of the SSD drive and determines the risk (step S07). Specifically, the risk determination unit 213b refers to the power-on timestamps recorded and stored in the cloud database 31 and the power-off data management unit 213a. For example, if an SSD drive whose power-on timestamp has not been updated for more than six months (i.e., has not been powered for more than six months) is detected, the risk determination unit 213b determines that there is a risk of data loss. Risk assessment of the target SSD drive will now be described with reference to FIG. 4.

[0035] Next, the controller 21 notifies the drive administrator of the risk analysis results and the SMART data (step S08). For example, the communication unit 214 transmits the processing results of steps S04, S05, and S07 and the SMART data used for risk determination to a terminal of the drive administrator (not shown). At this time, the controller 21 may notify whether the RUL prediction was made using the general-purpose prediction model 221 or the specialized prediction model 222.

[0036] Furthermore, if the risk determination process in steps S04, S05, and S07 determines that there is a risk (step S09; No), the details of the risk are notified to the user. For example, if the predicted RUL value falls below a value set in the risk management policy (e.g., 30% of the value at the start of drive use), the communication unit 214 notifies the user of the details of the risk by sending "risk of failure" and the predicted time until failure (remaining useful life) to the controller 12 of the storage device 1 (step S011). Furthermore, if the predicted RUL value falls below 10% of the value at the start of drive use, the communication unit 214 instructs the controller 12 of the storage device 1 to perform an automatic backup to prevent data loss (step S011). Specific processing of the automatic backup will be described later with reference to FIG. 5.

[0037] Furthermore, even if it is determined that there is no risk (step S09; Yes), the communication unit 214 notifies the user of the risk analysis result by transmitting the risk analysis result to the storage device 1 (step S010). For example, the data collection unit 121 records the risk analysis result transmitted in steps S10 and S11, periodically generates a deterioration trend analysis report, and transmits it to a user terminal (not shown) via the communication unit 122.

[0038] (SSD risk assessment process) Next, the flow of risk assessment of SDD will be described with reference to FIG. The controller 21 acquires the drive identification number and SMART data (step S11). Next, the controller 21 determines whether the target drive is an SSD (step S12). If the target drive is not an SSD (step S12; No), the process of FIG. 4 ends. If the target drive is an SSD (step S12; Yes), the power-off data management unit 213a references the acquired SMART data and records the power-on timestamp of the SSD. The risk determination unit 213b also references the latest power-on timestamp recorded in the power-off data management unit 213a and past power-on timestamps recorded in the cloud database 31 to check the last power-on time of all recorded SSDs (step S13). If the SSD has not been powered for six months or more (step S14; Yes), the risk determination unit 213b determines that there is an SSD that may have lost data or that may have lost data, and notifies the storage device 1 side in which the SSD that has been powered off for a long period of time is installed of this risk (step S15). For example, the identification information of the SSD drive and the possibility of data loss are notified to the terminal (not shown) of the administrator of such storage device 1. If the SSD power-off time is less than six months (step S14; No), the risk determination unit 213b determines that there is no risk of data loss and notifies the storage device 1 of this determination result (step S16). The power-off data management unit 213a uploads the power-on timestamp recorded this time and updates the power-on timestamp in the cloud database 31. This allows the power-on status of the SSD drive installed in a certain storage device 1 to be shared with other storage devices 1. For example, the management server 2 can grasp the power-on status of the SSD drives for all storage devices 1 connected to the management server 2, so it can notify the administrator or user of a storage device 1 that has been power-off for a long time (which the administrator or user may not be aware of) of the possibility of data loss.

[0039] (Conservation processing) Next, an example of a storage device-side protection process for the risk analysis result will be described with reference to FIG. 5. The controller 12 acquires the risk analysis result from the management server 2 via the communication unit 122 (step S21). Next, the controller 12 determines whether there is a possibility of data loss (step S22). The risk analysis result acquired from the management server 2 includes the processing result described in FIG. 4. The controller 12 makes this determination based on whether the risk analysis result includes information indicating a possibility of data loss. If there is no possibility of data loss (step S22; No), the process proceeds to step S24. If there is a possibility of data loss (step S22; Yes), the controller 21 notifies the user that data may have been lost (step S23). For example, the controller 21 notifies the user by transmitting information indicating a possibility of data loss to a user terminal (not shown) or the like.

[0040] Next, the controller 12 determines whether the predicted RUL is 30% or less of the value at the start of drive use (step S24). If it is 30% or less (step S24; Yes), the controller 12 determines whether the predicted RUL is 10% or less of the value at the start of drive use (step S25). If it is 10% or less (step S25; Yes), the controller 12 uses the drive operation unit 123 to perform an automatic backup of the drive whose predicted RUL is 10% or less (step S26). At this time, a rebuild may be performed in addition to the automatic backup. After the automatic backup in step S26 is completed, or if step S24 or step S25 is No, the controller 12 notifies the user of the risk analysis results and advice on how to deal with the problem (step S27). For example, if the predicted RUL is between 10% and 30%, the controller 12 may notify the user that the risk of failure is increasing and advise the user to perform manual backup or arrange for a spare drive as advice on how to deal with the problem.

[0041] (effect) As described above, according to this embodiment, a general-purpose prediction model 221 constructed using time-series SMART data and a specialized prediction model 222 constructed using SMART data for each product are prepared. If a specialized prediction model 222 corresponding to the drive to be predicted exists, the specialized prediction model 222 is used to predict the RUL of the drive. This enables highly accurate drive health checks and prediction of the drive's remaining useful life (RUL). This reduces data loss due to unexpected drive failure and unnecessary drive replacement costs due to misjudgment of drive life. Furthermore, even if a specialized prediction model 222 does not exist, the general-purpose prediction model 221 can be used to predict the RUL.

[0042] More specifically, the following effects can be obtained. (1) Predicting RUL for various types of drives There are many different types of drives (e.g., different HDD rotation speeds, SSD NAND chip types, and capacities). For example, HDDs and SSDs have different RUL trends, and the meaning of the drive health indicated by SMART data also differs. Furthermore, the definitions, standard values, and thresholds for SMART data do not match between drive vendors. As a result, a model trained on data from one drive (e.g., a 10TB HDD from Company A) has low reliability in predicting the RUL for a different type of drive (e.g., a 3.84TB SSD from Company B). Furthermore, analyzing only the drive SMART data at the time of analysis makes it impossible to predict trends in the drive's health.

[0043] In contrast, in this embodiment, two prediction models are constructed to improve the accuracy of RUL prediction for various types of drives. The first is a general-purpose prediction model 221, which is trained using SMART data from various drives (each vendor, each model, each capacity). The second is a specialized prediction model, based on the general-purpose prediction model 221, which is trained using SMART data from the same drive as the drive installed in the storage device 1, and tailored for the specific drive. This improves the accuracy of drive RUL prediction.

[0044] (2) The amount of training data Generally, training datasets are often composed of SMART data for the drives installed in the storage device 1 that is the target of prediction. Therefore, in environments where the storage device 1 is newly installed or has a small number of installed drives, the number of drive failures is small and it is difficult to collect SMART data leading up to a failure, and as a result, the accuracy of the RUL prediction is often insufficient.

[0045] In contrast, in this embodiment, publicly available time-series SMART data related to all kinds of drive failures is collected to create a pre-learning dataset. In this way, a general-purpose prediction model 221 can be constructed based on a large amount of SMART data, regardless of the operating time of the storage device 1 that is the target of failure prediction or the type and number of drives installed in the storage device 1. In addition, drive failure data of other users is also accumulated in the cloud database 31 and used as training data for constructing the general-purpose prediction model 221, allowing even new users to quickly predict RUL with high accuracy.

[0046] (3) Predicting failures due to long-term deterioration Drive failures are caused by two patterns: long-term degradation and short-term failure. In the case of long-term degradation, even if a certain amount of error data exists in the SMART data, the impact on the RUL is often not significant. Therefore, when it comes to failures due to long-term degradation, simply setting thresholds for each item in the SMART data tends to result in low RUL prediction accuracy.

[0047] The specialized prediction model 222 is a model trained with training data that includes a certain amount of error data as well as time-series SMART data that indicates long-term degradation leading to failure. By using the specialized prediction model to generate RUL predictions, more accurate predictions can be achieved.

[0048] (4) Predicting data loss due to long-term power outage of SSD In addition to failures during operation, SSDs are known to be susceptible to data loss if they are left without power for a long period of time (long-term power outage). Data can be lost even if the drive itself is not faulty. Generally, there is no prediction of data loss due to long-term power outages on SSDs.

[0049] In contrast, in this embodiment, when analyzing the RUL of an SSD drive, the management server 2 records a timestamp, uploads it to the cloud database 31, and shares it among the facilities. If an SSD drive whose timestamp has not been updated for a long time is detected (even if it is in another facility), the user of the SSD drive is notified. This makes it possible to prevent data loss due to a long period of power outage.

[0050] Second Embodiment FIG. 6 is a second diagram illustrating an example of the configuration of the prediction system according to the embodiment. The prediction system 800 includes a storage unit 801 that stores a general-purpose prediction model constructed by using first training data, which is time-series SMART data collected from multiple types of storage devices (drives), to learn the relationship between the transitions of values ​​included in the first training data and the remaining useful life of the storage device, and a specialized prediction model constructed by using second training data, which is time-series SMART data for each type of storage device, to learn the relationship between the transitions of values ​​included in the second training data and the remaining useful life of the storage device. The prediction system 800 also includes a storage unit 802 that predicts the remaining useful life of a storage device to be predicted based on the time-series SMART data collected from the storage device to be predicted and the specialized prediction model for a storage device of the same type as the storage device to be predicted. The drive in the first embodiment is an example of a storage device. The data management unit 211a, model management unit 22, and cloud database 31 in the first embodiment are examples of the storage unit 801. The RUL prediction unit 211b in the first embodiment is an example of the prediction unit 802.

[0051] FIG. 7 is a flowchart showing an example of a prediction method according to the embodiment. The prediction means 802 acquires from the storage means 801 a general-purpose prediction model constructed by using first training data, which is time-series SMART data collected from multiple types of drives, to learn the relationship between the transitions of values ​​included in the first training data and the remaining useful life of the drive, and a specialized prediction model constructed by using second training data, which is time-series SMART data for each type of drive, to learn the relationship between the transitions of values ​​included in the second training data and the remaining useful life of the drive (step S801). Next, the prediction means 802 predicts the remaining useful life of the drive to be predicted based on the acquired specialized prediction model and the time-series SMART data collected from the drive to be predicted.

[0052] FIG. 8 is a diagram illustrating an example of a hardware configuration of the prediction system according to the embodiment. The computer 900 includes a CPU 901, a main memory device 902, an auxiliary memory device 903, an input / output interface 904, and a communication interface 905. The prediction systems 100 and 800 described above are implemented in the computer 900. The above-described functions are stored in the auxiliary memory device 903 in the form of a program. The CPU 901 reads the program from the auxiliary memory device 903, loads it into the main memory device 902, and executes the above-described processing in accordance with the program. The CPU 901 also allocates a storage area in the main memory device 902 in accordance with the program. The CPU 901 also allocates a storage area in the auxiliary memory device 903 for storing data being processed in accordance with the program.

[0053] Alternatively, a program for implementing all or part of the functions of the prediction systems 100 and 800 may be recorded on a computer-readable recording medium, and the program may be loaded into a computer system and executed to perform processing by each functional unit. The term "computer system" as used herein includes hardware such as an OS and peripheral devices. Furthermore, if a WWW system is used, the term "computer system" also includes the homepage provision environment (or display environment). Furthermore, the term "computer-readable recording medium" refers to portable media such as CDs, DVDs, and USBs, as well as storage devices such as hard disks built into the computer system. Furthermore, if the program is distributed to the computer 900 via a communication line, the computer 900 that receives the program may load the program into the main storage device 902 and execute the above-described processing. Furthermore, the program may be for implementing part of the above-described functions, or may be capable of implementing the above-described functions in combination with a program already stored in the computer system.

[0054] Although one embodiment of the present disclosure has been described in detail above with reference to the drawings, the specific configuration is not limited to the above, and various design modifications are possible within the scope of the gist of the present invention. Furthermore, one aspect of the present disclosure may be modified in various ways within the scope of the claims, and embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present disclosure. Furthermore, configurations in which elements described in the above embodiments and variations are substituted with elements that achieve the same effect are also included. Furthermore, each embodiment may be combined with other embodiments as appropriate.

[0055] Some or all of the above embodiments can be described as, but are not limited to, the following supplementary notes.

[0056] (Appendix 1) The prediction system includes: a means for storing a general-purpose prediction model constructed by using first training data, which is time-series SMART data collected from multiple types of storage devices, to learn the relationship between the trends in values ​​contained in the first training data and the remaining useful life of the storage devices; and a specialized prediction model constructed by using second training data, which is time-series SMART data for each type of storage device, to learn the relationship between the trends in values ​​contained in the second training data and the remaining useful life of storage devices of the type related to the second training data; and a means for predicting the remaining useful life of a drive to be predicted based on time-series SMART data collected from the drive to be predicted and the specialized prediction model for a drive of the same type as the drive to be predicted.

[0057] (Appendix 2) In the prediction system described in Appendix (1), if the storing means does not store the specialized prediction model for a drive of the same type as the drive to be predicted, the predicting means predicts the remaining useful life of the drive to be predicted based on time-series SMART data collected on the drive to be predicted and the general-purpose prediction model.

[0058] (Appendix 3) The specialized prediction model is a prediction system described in Appendix (1) or Appendix (2), which is a model constructed based on the general-purpose prediction model by additional learning using the second training data.

[0059] (Appendix 4) The prediction system according to any one of claims (1) to (3) further comprises a means for automatically backing up the storage device to be predicted when the remaining useful life of the storage device to be predicted falls below a predetermined threshold.

[0060] (Appendix 5) The prediction system described in Supplementary Notes (1) to (4) further includes a means for notifying of the possibility of data loss when the storage device to be predicted is an SSD and the power-off time of the drive exceeds a predetermined threshold.

[0061] (Appendix 6) The prediction system according to any one of claims (1) to (5) further comprises means for constructing the general-purpose prediction model and the specialized prediction model.

[0062] (Appendix 7) The prediction system according to any one of claims (1) to (6), wherein the first training data includes publicly available time-series SMART data relating to failures of storage devices.

[0063] (Appendix 8) A prediction system as described in Appendix (1) to Appendix (7), wherein the second training data includes SMART data collected periodically until the storage device fails and / or SMART data collected during operation of the storage device to be predicted.

[0064] (Appendix 9) This is a prediction method that acquires a specialized prediction model for a storage device of the same type as the storage device to be predicted from a means for storing: a general-purpose prediction model constructed by using first training data, which is time-series SMART data collected from multiple types of storage devices, to learn the relationship between the trends in values ​​contained in the first training data and the remaining useful life of the storage devices; and a specialized prediction model constructed by using second training data, which is time-series SMART data for each type of storage device, to learn the relationship between the trends in values ​​contained in the second training data and the remaining useful life of the storage device of the type related to the second training data; and predicts the remaining useful life of the storage device to be predicted based on the acquired specialized prediction model and the time-series SMART data collected from the storage device to be predicted.

[0065] (Appendix 10) The computer is caused to acquire from a means for storing a general-purpose prediction model constructed by using first training data, which is time-series SMART data collected from multiple types of storage devices, to learn the relationship between the trends in values ​​contained in the first training data and the remaining useful life of the storage device, and a specialized prediction model constructed by using second training data, which is time-series SMART data for each type of storage device, to learn the relationship between the trends in values ​​contained in the second training data and the remaining useful life of the type of storage device related to the second training data, and to execute a process of predicting the remaining useful life of the storage device to be predicted based on the acquired specialized prediction model and the time-series SMART data collected from the storage device to be predicted. [Explanation of symbols]

[0066] 100···Prediction System 1. Storage device 11 Drive connection 12 Controller 121 Data Collection Unit 122···Communications Department 123 Drive operation unit 2. Management Server 21 Controller 211 Risk Discrimination Module 212 Drive Information Management Unit 213...SSD non-power time statistics section 211a···Data Management Department 211b···RUL prediction section 211c Risk Identification Section 213a···No-power data management section 213b Risk Identification Section 22. Model Management Department 221···General-purpose prediction model 222···Specialized Prediction Model 3. Cloud services 31. Cloud Database 32. Prediction model training section 800···Prediction System 801...Storage means 802...Prediction Methods 900···Computer 901 CPU 902...Main memory 903...Auxiliary storage device 904 Input / Output Interface 905···Communication Interface

Claims

1. a means for storing a general-purpose prediction model constructed by using first training data, which is time-series SMART data collected from multiple types of storage devices, to learn the relationship between the transition of values ​​included in the first training data and the remaining useful life of the storage devices; and a specialized prediction model constructed by using second training data, which is time-series SMART data for each type of storage device, to learn the relationship between the transition of values ​​included in the second training data and the remaining useful life of the storage device of the type related to the second training data; means for predicting the remaining useful life of a storage device to be predicted based on time-series SMART data collected from the storage device to be predicted and the specialized prediction model for storage devices of the same type as the storage device to be predicted; A prediction system comprising:

2. If the storage means does not store the specialized prediction model related to a storage device of the same type as the storage device to be predicted, the predicting means predicts the remaining useful life of the storage device to be predicted based on time-series SMART data collected from the storage device to be predicted and the general-purpose prediction model; The prediction system of claim 1 .

3. The specialized prediction model is a model constructed by additional learning using the second training data based on the general-purpose prediction model. The prediction system according to claim 1 or claim 2.

4. means for automatically backing up the storage device that is the target of the prediction when the remaining useful life of the storage device that is the target of the prediction falls below a predetermined threshold; The prediction system according to claim 1 or claim 2, further comprising:

5. means for notifying the possibility of data loss when the storage device to be predicted is an SSD and the power-off time of the storage device exceeds a predetermined threshold; The prediction system according to claim 1 or claim 2, further comprising:

6. means for constructing the general-purpose prediction model and the specialized prediction model; The prediction system according to claim 1 or claim 2, further comprising:

7. The first training data includes publicly available time-series SMART data regarding failures of storage devices; A prediction system according to claim 1 or claim 2.

8. The second training data includes SMART data collected periodically until the storage device fails and / or SMART data collected while the storage device to be predicted is in operation. A prediction system according to claim 1 or claim 2.

9. a general-purpose prediction model constructed by using first training data, which is time-series SMART data collected from multiple types of storage devices, to learn the relationship between the transitions of values ​​included in the first training data and the remaining useful life of the storage devices; and a specialized prediction model constructed by using second training data, which is time-series SMART data for each type of storage device, to learn the relationship between the transitions of values ​​included in the second training data and the remaining useful life of the storage device of the type related to the second training data. The specialized prediction model is acquired for a storage device of the same type as the storage device to be predicted from a means for storing the general-purpose prediction model constructed by using first training data, which is time-series SMART data collected from multiple types of storage devices, to learn the relationship between the transitions of values ​​included in the first training data and the remaining useful life of the storage device of the type related to the second training data. predicting a remaining useful life of the storage device to be predicted based on the acquired specialized prediction model and time-series SMART data collected from the storage device to be predicted; Forecasting methods.

10. On the computer, a general-purpose prediction model constructed by using first training data, which is time-series SMART data collected from multiple types of storage devices, to learn the relationship between the transition of values ​​included in the first training data and the remaining useful life of the storage devices; and a specialized prediction model constructed by using second training data, which is time-series SMART data for each type of storage device, to learn the relationship between the transition of values ​​included in the second training data and the remaining useful life of the storage device of the type related to the second training data. The specialized prediction model for a drive of the same type as the storage device to be predicted is acquired from a means for storing the general-purpose prediction model constructed by using first training data, which is time-series SMART data collected from multiple types of storage devices, to learn the relationship between the transition of values ​​included in the first training data and the remaining useful life of the storage device of the type related to the second training data. A process of predicting the remaining useful life of the storage device to be predicted based on the acquired specialized prediction model and time-series SMART data collected from the storage device to be predicted; A program that executes the following.

Images