Disk fault prediction method and device, readable storage medium and program product

Abstract

The invention belongs to the technical field of computers, and discloses a disk fault prediction method and device, a readable storage medium and a program product, and the method comprises the steps: obtaining a plurality of pieces of detection data information, and a disk health score corresponding to each piece of detection data information; determining the attribution of each piece of detection data information by using the disk health score to obtain prediction data and direct estimation data; wherein the prediction difficulty of the prediction data is higher than that of the direct estimation data; determining a first fault prediction result by using a disk health score corresponding to the detection data information in the direct estimation data; and inputting the prediction data into the target model for fault prediction to obtain a second fault prediction result. The method has the technical effects that not only can the disk fault prediction efficiency be improved, but also the fault prediction accuracy can be improved.

Description

Technical Field

[0001] The present application relates to the field of computer application technology, and in particular to a disk failure prediction method, device, readable storage medium and program product. Background Art

[0002] Data centers provide powerful computing and storage capabilities for services in various industries. Data centers have a huge number of storage disks. Therefore, it is increasingly important to ensure the reliability of hard disks, reduce server downtime caused by hard disk failures, and reduce the losses caused by them.

[0003] At present, there are two main types of solutions to disk failures. One is to use software technology to provide data redundancy and other functions to prevent data loss caused by failures. The other is to monitor indicators related to the health status of the hard disk and use a certain strategy to analyze them to obtain the health status of the hard disk. The second type of solution mainly includes two types of strategies. One is to formulate relevant hard disk diagnostic rules based on existing expert knowledge and perform predictive analysis based on the real-time status of the hard disk, but this method often cannot accurately obtain prediction results. The other is based on machine learning algorithms, which train hard disk data to obtain the corresponding prediction model. This method has a relatively large model calculation overhead and is relatively more time-consuming.

[0004] In summary, how to make disk failure prediction more efficient and accurate is a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0005] The purpose of this application is to provide a disk fault prediction method, device, readable storage medium and program product to achieve efficient and accurate fault prediction of the disk.

[0006] In order to solve the above technical problems, this application provides the following technical solutions:

[0007] A disk failure prediction method, comprising:

[0008] Get multiple pieces of test data information and the disk health score corresponding to each piece of test data information;

[0009] The disk health score is used to determine the attribution of each piece of detection data information, and the predicted data and the direct estimated data are obtained; wherein the prediction difficulty of the predicted data is higher than that of the direct estimated data;

[0010] Determine a first fault prediction result by using the disk health score corresponding to the detection data information in the direct estimation data;

[0011] The prediction data is input into the target model for fault prediction to obtain a second fault prediction result.

[0012] Preferably, it also includes:

[0013] Comparing the disk replacement record with the second fault prediction result to obtain a comparison result;

[0014] In the process of optimizing the target model by using the training data, rewards and punishments are given based on the comparison results.

[0015] Preferably, in the process of optimizing the target model by using the training data, rewards and punishments are performed based on the comparison results, including:

[0016] Determining reward and punishment values ​​using the comparison results;

[0017] In the process of optimizing the target model by using the training data, the reward and penalty values ​​are used to indicate the optimization direction of the target model.

[0018] Preferably, determining the reward or punishment value using the comparison result includes:

[0019] In the case that the comparison result corresponds to that the disk replacement record matches the second fault prediction result, determining a preset reward value as the reward or penalty value;

[0020] In the case that the comparison result corresponds to the disk replacement record and the second fault prediction result does not match, a preset penalty value is determined as the reward and penalty value.

[0021] Preferably, when the comparison result corresponds to that the disk replacement record matches the second fault prediction result, determining a preset reward value as the reward and penalty value includes:

[0022] In the case that the comparison result corresponds to that the disk replacement record matches the second fault prediction result, if the disk replacement record is replacement of a disk, a preset first reward value is determined as the reward or penalty value;

[0023] In the case where the comparison result corresponds to that the disk replacement record matches the second fault prediction result, if the disk replacement record is that the disk is not replaced, determining the preset second reward value as the reward and penalty value;

[0024] Wherein, the first reward value is greater than the second reward value.

[0025] Preferably, in the process of optimizing the target model by using the training data, using the reward and penalty value to indicate the optimization direction of the target model includes:

[0026] In the process of optimizing the target model using the training data, substituting the reward and punishment value into the advantage function to calculate the difference between the expected reward and the predicted reward;

[0027] Based on the difference, the target model is updated.

[0028] Preferably, the prediction data is input into a target model for fault prediction to obtain a second fault prediction result, including:

[0029] Adding manufacturer information and disk attribute information to each piece of detection data information belonging to the prediction data;

[0030] After completing the data addition, each piece of detection data information is converted into a vector;

[0031] Group the obtained vectors to obtain a vector group;

[0032] The vector group is input into the target model for fault prediction to obtain the second fault prediction result; wherein the target model is a model that has learned the fault prediction capabilities of disks of different manufacturers and different attributes.

[0033] Preferably, inputting the vector group into the target model for fault prediction to obtain the second fault prediction result includes:

[0034] Inputting the vector group into the target model, and using the information cross-learning module in the target model to perform cross-learning between different feature dimensions in the vector group to obtain effective features;

[0035] Using the time series information module in the strategy network to sort out the effective features to obtain statistical characteristics;

[0036] The statistical characteristics are normalized using a normalization layer in the strategy network to obtain the second fault prediction result.

[0037] Preferably, the effective features are sorted using the time sequence information module in the strategy network to obtain statistical characteristics, including:

[0038] Determine the dependency relationship between different positions in the effective features using the encoder in the temporal information module to obtain a global feature representation;

[0039] The residual network in the temporal information module is used to perform residual processing on the global feature representation to obtain the statistical characteristics.

[0040] Preferably, before using the disk health score to determine the attribution of each piece of detection data information and obtaining the predicted data and the direct estimated data, the method further includes:

[0041] Searching for invalid indicators in each piece of detection data information, and when an invalid indicator is found, adding one to the invalid index of the piece of detection data information; wherein each piece of detection data information includes a plurality of indicators;

[0042] If the invalidity index exceeds a preset threshold, the detection data information is deleted.

[0043] Preferably, if the invalidity index exceeds a preset threshold, deleting the piece of detection data information includes:

[0044] If the invalidity index exceeds a preset threshold, after obtaining the disk information corresponding to the piece of detection data information, the piece of detection data information is deleted;

[0045] The disk information is used to output prompt information of data collection abnormality.

[0046] Preferably, obtaining multiple pieces of detection data information and a disk health score corresponding to each piece of detection data information includes:

[0047] Using a disk self-test tool to test the disk, a plurality of test data information is obtained; wherein a piece of test data information includes a plurality of indicator data;

[0048] Obtaining the indicator threshold set corresponding to each piece of detection data information, and comparing the current state value of each indicator in the detection data information with the corresponding threshold in the indicator threshold set;

[0049] According to the status comparison result, the disk health score corresponding to the detection data information is determined.

[0050] A disk failure prediction device, comprising:

[0051] A data acquisition unit, used to acquire multiple pieces of detection data information and a disk health score corresponding to each piece of detection data information;

[0052] A data screening unit, used to determine the attribution of each piece of detection data information by using the disk health score, and obtain predicted data and direct estimated data; wherein the prediction difficulty of the predicted data is higher than that of the direct estimated data;

[0053] A first fault prediction unit, configured to determine a first fault prediction result by using a disk health score corresponding to the detection data information in the direct estimation data;

[0054] The second fault prediction unit is used to input the prediction data into the target model for fault prediction to obtain a second fault prediction result.

[0055] An electronic device, comprising:

[0056] Memory for storing computer programs;

[0057] A processor is used to implement the steps of the above disk failure prediction method when executing the computer program.

[0058] A readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the disk failure prediction method are implemented.

[0059] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the steps of the above disk failure prediction method.

[0060] Apply the method provided in the embodiment of the present application to obtain multiple detection data information and the disk health score corresponding to each detection data information; use the disk health score to determine the attribution of each detection data information to obtain predicted data and direct estimation data; wherein, the prediction difficulty of the predicted data is higher than that of the direct estimation data; use the disk health score corresponding to the detection data information in the direct estimation data to determine the first fault prediction result; input the predicted data into the target model for fault prediction to obtain the second fault prediction result.

[0061] In the present application, when multiple detection data information and the disk health score corresponding to each detection data information are obtained. For the disk health score, the score size is related to the prediction difficulty. For example, the more extreme the score, the more accurate the fault prediction result. Therefore, the attribution of each detection data information can be determined based on the disk health score, thereby obtaining prediction data and direct estimation data. After completing the attribution division, the prediction difficulty of the obtained prediction data is higher than that of the direct estimation data. Therefore, the detection data information in the direct estimation data can be directly based on the corresponding disk health score to obtain a first fault prediction result. For prediction data with high prediction difficulty, the target model is used to perform fault prediction to obtain a second fault prediction result. In other words, for multiple detection data information to be predicted, the prediction difficulty can be screened and classified through the disk health score. For the part belonging to the direct estimation data, the disk fault prediction is directly based on the disk health score, and the first fault prediction result can be quickly obtained. For the part belonging to the prediction data, the fault prediction is performed with the help of the target model to obtain the second fault prediction result. Because the target model prediction is not simply based on the disk health score, but is directly based on the specific content of the detection data information, and the target model has been trained in advance, it can have a higher accuracy rate than fault prediction directly based on the disk health score.

[0062] Technical effect of the present application: In general, relatively accurate data can be obtained based on the disk health score. Therefore, by dividing the detection data information into different categories based on the disk health score, a large amount of direct estimation data and a small amount of prediction data can be obtained. For a large amount of direct estimation data, an accurate first fault prediction result can be quickly determined based on the disk health score, and for a small amount of prediction data for which it is difficult to obtain an accurate result based on the disk health score, fault prediction is performed based on the target model to obtain an accurate second fault prediction result. It can be seen that the present application can not only improve the efficiency of disk fault prediction, but also improve the accuracy of fault prediction.

[0063] Correspondingly, the embodiments of the present application also provide a disk failure prediction device, equipment, readable storage medium and computer program product corresponding to the above-mentioned disk failure prediction method, which have the above-mentioned technical effects and are not repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0065] Figure 1 This is a flowchart of a disk failure prediction method in an embodiment of the present application;

[0066] Figure 2 A schematic diagram of a strategy network in a target model in an embodiment of the present application;

[0067] Figure 3 This is a schematic diagram of the structure of a disk failure prediction device in an embodiment of the present application;

[0068] Figure 4 A schematic diagram of a disk failure prediction system in an embodiment of the present application;

[0069] Figure 5 This is a schematic diagram of the structure of an electronic device in an embodiment of the present application;

[0070] Figure 6 This is a schematic diagram of the specific structure of an electronic device in an embodiment of the present application. DETAILED DESCRIPTION

[0071] In order to enable those skilled in the art to better understand the present application, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0072] Please refer to Figure 1 , Figure 1 This is a flowchart of a disk failure prediction method in an embodiment of the present application, the method comprising the following steps:

[0073] S101. Obtain multiple pieces of detection data information and a disk health score corresponding to each piece of detection data information.

[0074] In the present application, the detection data information and its corresponding disk health score can be obtained by directly reading from the storage medium; the detection data information and its corresponding disk health score can also be obtained by receiving data; multiple detection data information can also be obtained first, and then scored by a scoring strategy to obtain the disk health score; multiple detection data information can also be collected by calling some detection tools, and after obtaining the multiple detection data information, they can be scored to obtain the disk health score.

[0075] In order to make disk failure prediction more reliable, a piece of detection data information may include multiple index data about the disk. In the acquired detection data information, one disk may correspond to one or more detection data information. One piece of detection data information corresponds to one disk.

[0076] In a specific implementation of the present application, multiple pieces of detection data information and a disk health score corresponding to each piece of detection data information are obtained, including:

[0077] Step 1: Use a disk self-test tool to test the disk and obtain multiple test data information; wherein one test data information includes multiple indicator data;

[0078] Step 2: Obtain the indicator threshold set corresponding to each piece of detection data information, and compare the current state value of each indicator in the detection data information with the corresponding threshold in the indicator threshold set;

[0079] Step 3: According to the status comparison result, determine the disk health score corresponding to the detection data information.

[0080] For ease of description, the above three steps are combined for explanation below.

[0081] In this embodiment, the disk self-check tool may be a tool capable of collecting index data of the disk. The following description takes the disk self-check tool as SMART (Self-Monitoring Analysis And Reporting Technology, storage device health status monitoring and reporting technology) as an example.

[0082] According to the set frequency (Freq1), the relevant SMART information (detection data information) of the disk is continuously collected from the server, and the 10 indicator data with the highest correlation with the disk failure can be selected, for example: (001) Raw ReadErrorRate, (004) Start / Stop Count, (005) Relocated Sector Count, (009) Power-OnTime Count, (010) Spin up Retry Count, (011) Calibration Retry Count, (012) PowerCycle Count, (194) Temperature, (199) ULTRAATA CRC Error Rate, (200) Write ErrorRate.

[0083] Among them, 001: underlying data read error rate, Raw Read Error Rate;

[0084] 004: Start / Stop Count;

[0085] 005: Relocated Sector Count;

[0086] 009: Power-On Time Count (POH);

[0087] 010: Spinup Retry Count (i.e. the number of hard disk spindle motor start retries), Spinup Retry Count;

[0088] 011: Disk calibration retry count, Calibration Retry Count;

[0089] 012: Power Cycle Count;

[0090] 194: Temperature;

[0091] 199: parity error rate, ULTRAATACRC Error Rate;

[0092] 200: Write error rate, Write ErrorRate.

[0093] In practical applications, you can change indicators or increase the weight of an indicator according to actual conditions.

[0094] When scoring, a heuristic preliminary disk fault diagnosis strategy (HS, Heuristic Strategy) can be used to score the collected data. Specifically: Set the initial score Score1 (disk health score) to 0. According to the disk model, you can select the indicator threshold set of the corresponding model from the expert knowledge base, and then compare the current state value of the indicator with the threshold of the indicator in the knowledge base (if the model threshold is missing, a similar model of the same manufacturer can be used as a reference). If the current value of the indicator is not within the healthy threshold range, Score1 is determined to be minus the specified value, such as 5 (of course, other values ​​can also be used in actual applications); if the current value of the indicator is within the healthy range, Score1 is added to the specified value, such as 10 (of course, other values ​​can also be used in actual applications). Compare all indicator data in a detection data information with the corresponding threshold, and add or subtract the disk health score to obtain the final disk health score. The calculation formula for the disk health score (Score) is as follows:

[0095]

[0096] Among them, score i is the score of the i-th indicator data, S i It is an addition or subtraction operator, and the addition or subtraction is determined based on the comparison between the indicator data and the threshold.

[0097] S102: Use the disk health score to determine the ownership of each piece of detection data information, and obtain predicted data and direct estimated data.

[0098] Among them, the difficulty of predicting data is higher than that of directly estimated data. Directly estimated data and predicted data are equivalent ion sets.

[0099] In this embodiment, the detection data information that needs to be used for fault prediction by the target model is used as prediction data, and the detection data information that can directly draw accurate conclusions based on the disk health score is used as direct estimation data. Both the prediction data and the direct estimation data can be used as training data for further optimization training of the target model.

[0100] The more extreme the disk health score is, the more accurate the fault prediction result determined based on the disk health score is. For example, the higher the value of the disk health score exceeding the health threshold, the higher the accuracy of judging the disk as healthy; correspondingly, the lower the value of the disk health score below the health threshold, the higher the accuracy of judging the disk as faulty.

[0101] Based on this, in this embodiment, the prediction difficulty of each detection data information can be determined based on the disk health score, and the detection data information with a high accuracy of fault judgment directly determined based on the score can be determined as direct estimation data, and the detection data information with a low accuracy of fault judgment determined based on the score can be determined as prediction data.

[0102] Therefore, after obtaining the disk health score, the ownership of each piece of detection data information can be determined based on the mapping relationship between the disk health score and the difficulty of judging disk failure. For example, the disk health score range can be determined for direct estimation data and prediction data respectively, and based on the disk health score, the ownership of each piece of detection data information can be directly delineated.

[0103] In a specific implementation of the present application, the health level of the disk is determined based on the disk health score;

[0104] According to the correspondence between the disk health level and various types of data, the ownership of each detection data information is determined to obtain training data, prediction data and direct estimation data.

[0105] You can define the health level of the disk, which is divided into four levels and the corresponding processing strategies.

[0106] Level 1: The disk is damaged, which means that normal reading and writing cannot be performed. The disk is marked as damaged, and the damage time and SMART information in the period before the damage are reported;

[0107] Level 2: high-risk disk, generates a high-risk disk alarm, records the disk information at this moment and in the past period, and reports it to the data center;

[0108] Level 3, general disk, generates a general disk warning, records the disk information at this moment and in the past period of time, and reports it to the data center.

[0109] Level 4, healthy disk, no alarm is generated.

[0110] The main purpose of collecting the SMART information of the problem disk in the previous period is to learn the subsequent algorithm model. Reporting to the data center requires the reported data to be used as training data.

[0111] According to the mapping relationship that the more normal the disk index data is, the higher the disk health score is, in this embodiment, the relationship between the score and the disk level is determined as follows:

[0112] The disk health score is above 90 and the health level is 4. This type of disk data information participates in the subsequent model training, but the model does not need to make a decision on its health level, that is, the health status can be determined by the heuristic strategy. In this way, the health data can be effectively filtered and the decision-making pressure of the machine learning algorithm can be reduced.

[0113] The disk health score is between 90 and 70, and the health level is 3. This type of disk data information is involved in subsequent model training and model decision-making of subsequent models.

[0114] The disk health score is between 70 and 50, and the health level is 2. This type of disk data information is involved in subsequent model training and model decision-making of subsequent models.

[0115] The disk health score is below 50, and the health level is 1. This type of disk data information participates in subsequent model training, but does not need to participate in the model decision of subsequent models (the level can be adjusted according to the score classification).

[0116] That is to say, the disk data information corresponding to health levels 1 and 4 belongs to direct estimation data, the disk data information corresponding to health levels 2 and 3 belongs to prediction data, and the disk data information corresponding to health levels 1 to 4 belongs to training data. In addition, the disk data information corresponding to the health levels 1 to 3 in the previous period can also be classified as training data.

[0117] In a specific implementation of the present application, the attribution of each piece of detection data information is determined by using the disk health score, and before obtaining the predicted data and the direct estimated data, the following is also included:

[0118] Step 1: search for invalid indicators in each piece of detection data information, and if an invalid indicator is found, increase the invalidity index of the detection data information by one; wherein each piece of detection data information includes multiple indicators;

[0119] Step 2: If the invalidity index exceeds the preset threshold, the detection data information is deleted.

[0120] For ease of description, the above steps are combined for explanation below.

[0121] Considering that in actual applications, the collection of a certain indicator may be abnormal due to various reasons, such as the collection of a certain indicator is empty or the collection fails. When there are too many abnormalities in the collection of indicators in a piece of detection data information, it means that the detection data information does not have the function of disk health status assessment, and it can be deleted at this time. Specifically, an invalid index (Num1) can be set for each piece of detection data information, and its initial value is 0. After each unnecessary indicator with abnormal collection is found, it is added by 1. If the invalid index exceeds the preset threshold value such as 3 (of course, other values ​​can also be used in actual applications), the detection data information can be deleted.

[0122] The main function of Num1 here is to observe whether the characteristics of the sampling data are sufficient to represent the health status of the disk. If Num1 is greater than 3, it is considered that the information cannot assist in accurate prediction.

[0123] In a specific implementation of the present application, if the invalidity index exceeds a preset threshold, the detection data information is deleted, including:

[0124] If the invalid index exceeds the preset threshold, after obtaining the disk information corresponding to the detection data information, the detection data information is deleted;

[0125] Use disk information to output prompt information of data collection abnormalities.

[0126] That is to say, when a piece of detection data information is found to be collected abnormally, its corresponding disk information can be obtained first, and then deleted. In addition, with the help of the disk information, a prompt message of data collection abnormality can be output externally to facilitate the subsequent location of the disk with abnormal collection.

[0127] S103: Determine a first fault prediction result by using the disk health score corresponding to the detection data information in the direct estimation data.

[0128] For the detection data information in the direct estimation data, the first fault prediction result can be obtained directly based on the corresponding disk health score. For example, if the score is greater than a certain value, it is determined that the disk is healthy and does not need to be replaced. If the score is less than a certain value, it is determined that the disk is faulty and needs to be replaced.

[0129] The first fault prediction result may be a judgment conclusion on whether the disk is faulty, and may also include a suggestion on whether the disk should be replaced.

[0130] S104: Input the prediction data into the target model for fault prediction to obtain a second fault prediction result.

[0131] In an embodiment, a target model may be pre-trained or trained during use, and the target model may perform fault prediction based on input prediction data to obtain a second fault prediction result.

[0132] In this embodiment, the specific structure and training process of the target model are not limited.

[0133] In a specific implementation of the present application, the prediction data is input into the target model for fault prediction to obtain a second fault prediction result, including:

[0134] Step 1: Add manufacturer information and disk attribute information to each piece of detection data information belonging to the prediction data;

[0135] Step 2: After completing the data addition, convert each piece of detection data information into a vector;

[0136] Step 3: Group the obtained vectors to obtain a vector group;

[0137] Step 4: Input the vector group into the target model for fault prediction to obtain a second fault prediction result; wherein the target model is a model that has learned the fault prediction capabilities of disks of different manufacturers and different attributes.

[0138] For ease of description, the above steps are combined for explanation below.

[0139] In this embodiment, a reinforcement learning model can be used as the core model of secondary training (optimization training).

[0140] In order to facilitate the target model to predict disk failures based on the detection data information, manufacturer information and disk attribute information can be added to each detection data information in the prediction data. Then, each detection data information after the data is added is converted into a vector. In order to ensure the optimal efficiency of the target model, the obtained vectors can be grouped, and the number of vectors in each group of vectors is controlled within Size1 (size). The Size1 can be selected based on the processing of different numbers of vector groups and the statistics of their corresponding performance. In this way, the number of states that can be observed each time is limited to a certain range, such as the number of vectors in each group is Size1, which is the maximum number of SMART information. If the number of vectors at the current moment is less than Size1, empty nodes (all-zero vectors) are used to fill this group of vectors; if it is greater than Size1, the first Size1 inputs are selected for training according to the first-come, first-served principle, and subsequent information waits.

[0141] After completing the vector grouping, the vector group can be directly input into the target model for fault prediction to obtain a second fault prediction result. It should be noted that the target model is a model that has learned the fault prediction capabilities of disks of different manufacturers and different attributes.

[0142] In a specific implementation of the present application, the above step 4, inputting the vector group into the target model for fault prediction to obtain a second fault prediction result, includes:

[0143] Step 1: Input the vector group into the target model, and use the information cross-learning module in the target model to perform cross-learning between different feature dimensions in the vector group to obtain effective features;

[0144] Step 2: Use the time series information module in the strategy network to sort out the effective features and obtain statistical characteristics;

[0145] Step 3: Use the normalization layer in the strategy network to normalize the statistical characteristics to obtain the second fault prediction result.

[0146] Among them, step 2 uses the time series information module in the strategy network to organize the effective features and obtain statistical characteristics, including:

[0147] Using the encoder in the temporal information module, the dependencies between different positions in the effective features are determined to obtain the global feature representation.

[0148] The residual network in the temporal information module is used to perform residual processing on the global feature representation to obtain statistical characteristics.

[0149] For ease of description, the above steps are combined for explanation below.

[0150] In this embodiment, the model algorithm architecture of the target model includes a network model, which is a policy network and a value network. The policy network is used to learn the current SMART sequence information and perform training and decision-making. The value network adjusts the policy network according to the decision of the policy network and the reward returned by the environment, and guides the agent to update its next action in the correct direction. The input of the model is a group of SMART information vectors, each of which contains Size1 information. In order to better enable the agent to learn the time-series change characteristics of the SMART information status of the disk before failure, an encoder module can be introduced in the policy network. The settings of the two networks are described in detail below.

[0151] Strategy network Figure 2 First, the information of the input vector group and the updated environmental information after the decision is made, Info = {i0,i1,…,i A}, Info∈R A*B,The vector group set is identified as Info, i represents a single state, A is the length of the vector group, and B is the number of features of the vector, that is, the number of attributes of SMART information. The internal structure of the policy network includes an information cross-learning module and a temporal information module.

[0152] Among them, the information cross-learning module adopts a two-layer MLP (multi-layer perceptron) structure, with 32 neurons in each layer. It performs cross-learning between different dimensions of feature vectors, learns effective features and passes them to the next module.

[0153] The time series information module can use an encoder and a residual network. The encoder is mainly used to learn the dependencies between different positions in the sequence, thereby forming a global feature representation. The residual is the residual operation of the front and back vectors of the encoder layer, which can encode the statistical characteristics of the entire data set and avoid bias towards a specific token in the sequence. That is, the role of the residual network is to ensure that the previous information will not be forgotten during the training process. The role of the Class_Token (category feature) in it is to take the information of the 0th dimension as the feature information extracted by the network, encode the statistical characteristics of the entire data set, and avoid bias towards a specific token (feature) in the sequence. The final normalization layer (which contains a normalization function and a Softmax function) will output a set of probabilities about the vector group, based on which it is chosen whether to replace the disk, that is, the second fault prediction result is obtained. The probability calculation formula is as follows:

[0154]

[0155] Among them, yi represents the i-th category, K represents the total number of categories, e represents the exponential function of a single category, the denominator on the right side of the equal sign is the sum of all category exponents, ensuring that the total probability is 1, thereby converting it into a valid probability distribution, and o(yi) represents the calculated probability corresponding to the category.

[0156] In a specific implementation of the present application, the target model may be further optimized and trained in the use environment to further improve its prediction accuracy. The optimization training process includes:

[0157] Step 1: Compare the disk replacement record with the second fault prediction result to obtain a comparison result.

[0158] Step 2: When optimizing the target model using training data, rewards and punishments are given based on the comparison results.

[0159] For ease of description, the above two steps are combined for explanation below.

[0160] A disk status update module can be set up. If the administrator replaces a disk, the disk status update module can delete the disk's alarm information from the alarm list and re-add it to the environment, collect information and wait for the fault prediction model to predict the health level of the new disk. The disk replacement record can be obtained based on the disk status update module.

[0161] After obtaining the disk replacement record, the second fault prediction result can be compared with the disk replacement record to obtain a comparison result. The comparison result can indicate whether the model prediction is accurate. Generally, if the prediction is accurate, the second fault prediction result is a disk failure and needs to be replaced, then the disk replacement record will correspond to the replacement disk, and the second fault prediction result is that the disk is normal and does not need to be replaced, then the disk replacement record will correspond to the unreplaced disk. The corresponding prediction is inaccurate, which means that the second fault prediction result does not match the disk replacement record.

[0162] It should be noted that the second fault prediction result is a partial prediction result, and the disk replacement record corresponds to the global replacement record. Therefore, when comparing, only the partial relevant disk replacement records corresponding to the second fault prediction result can be compared. For the partial disk replacement records not involved in the second fault prediction result, the accuracy of the model prediction is not compared.

[0163] In order to optimize the target model, the target model can be trained regularly or irregularly, and the training can be performed in an offline training manner.

[0164] In this embodiment, since the training data is the most comprehensive, the training data can be used to optimize the target model. In actual scenarios, in most cases, the disk will not have problems, and the proportion of different prediction results in the training samples is not balanced. In order to enable the target model to effectively learn its judgment ability for low-probability events, in this embodiment, rewards and penalties are required based on the comparison results to guide the optimization direction of the model.

[0165] Specifically, based on the comparison results and combined with the reward function, the optimization training direction of the target model can be guided.

[0166] In a specific implementation of the present application, in the process of optimizing the target model using the training data, rewards and punishments are given based on the comparison results, including:

[0167] Step 1: Use the comparison results to determine the reward and punishment values;

[0168] Step 2: When optimizing the target model using the training data, the reward and penalty values ​​are used to indicate the optimization direction of the target model.

[0169] Among them, the reward and punishment values ​​are determined by using the comparison results, including:

[0170] Step 1: when the disk replacement record corresponding to the comparison result matches the second fault prediction result, a preset reward value is determined as a reward or penalty value;

[0171] Step 2: When the disk replacement record corresponding to the comparison result does not match the second fault prediction result, the preset penalty value is determined as the reward and penalty value.

[0172] Wherein, step 1, when the disk replacement record corresponding to the comparison result matches the second fault prediction result, determining the preset reward value as the reward and penalty value, includes:

[0173] In the case that the disk replacement record corresponding to the comparison result matches the second fault prediction result, if the disk replacement record is to replace the disk, the preset first reward value is determined as the reward or penalty value;

[0174] In the case that the disk replacement record corresponding to the comparison result matches the second fault prediction result, if the disk replacement record is that the disk has not been replaced, the preset second reward value is determined as the reward or penalty value;

[0175] The first reward value is greater than the second reward value.

[0176] For ease of description, the above steps are combined for explanation below.

[0177] In order to better guide the agent to make decisions based on SMART and other information, a reward function is proposed in this embodiment to reward or punish each decision of the agent according to the function process. The target model can be trained offline, using SMART information data from a certain period of time in the past, and starting from a certain period of time in the past, the disk is trained based on SMART information.

[0178] It should be noted that due to the differences in SMART information of different models, each agent can only be trained for the SMART information of one model. Each time, the agent predicts the health status of the disk in the future based on the continuous SMART information of the server in a period of time and then takes action. The decision of the current agent and the corresponding reward: whether to replace the current disk within the prediction time period. If the disk predicts a disk failure within the prediction time period and generates a replacement action, the agent is given a larger reward R1 (first reward value); if the failure to occur is not predicted within the expected time period, the action is not to replace, or if there is no failure within the expected time period but the action is to replace the disk, the agent is given a larger penalty P1 (penalty value). In the continuously collected data set, most of the SMART information should indicate a healthy state, that is, the disk is generally normal. In this case, if the agent generates an action of not replacing the disk, it is given a smaller reward R2 (second reward value), so as to guide the agent to learn all health states and solve the sparse reward problem.

[0179] The structure of the value network in the target model uses a four-layer MLP. The value network trains itself by collecting rewards from the policy network to predict the expected reward for a given sequence.

[0180] In a specific implementation of the present application, in the process of optimizing the target model using the training data, the optimization direction of the target model is indicated by using the reward and punishment values, including:

[0181] Step 1: When optimizing the target model using training data, substitute the reward and punishment values ​​into the advantage function to calculate the difference between the expected reward and the predicted reward.

[0182] Step 2: Update the target model based on the difference.

[0183] For ease of description, the above steps are combined for explanation below.

[0184] In the value network, the calculation formula of the advantage function is as follows:

[0185]

[0186] Where i indicates that the currently executed action is the i-th one, i' indicates that the currently executed action is the i'th one, i'>i, a i represents the i-th action, x i Indicates the state corresponding to executing the i-th action, A t (x i ,a i ) is to execute action a in state xi iThe difference between the predicted reward value and the expected reward value corresponding to the time, r i is the single reward for the ith action, Indicates that action a is executed in the current state i The predicted reward value obtained after φ (x i ) is the state value function of the value network, representing the current state x i and the expected reward value for the current network parameters.

[0187] The advantage function can guide the agent not to favor high-frequency small reward events and increase the probability of low-frequency large reward events. In addition, during the update process, the estimated long-term return is not directly used to update the network, but the difference between the cumulative reward of environmental feedback and the estimated reward is used to update the network. This method can effectively improve training efficiency and prevent the update range from being too small or too large.

[0188] In a specific implementation of the present application, it also includes:

[0189] Alarm processing is performed according to the first fault prediction result and / or the second fault prediction result.

[0190] Specifically, an alarm module can be set up. Function of the alarm module: After the prediction result of the above health level or decision action is generated, it will be processed by the alarm module. If the output result of the fault prediction module is to replace the disk, an alarm will be generated immediately, prompting the administrator to replace the corresponding disk immediately and add the disk to the high-risk disk sequence; if the disk is not replaced, no alarm will be generated. When issuing an alarm, the first fault prediction result and the second fault prediction result can be processed indiscriminately. Of course, different differentiated alarms can also be issued. That is, when there is a prediction result of disk failure in the first fault prediction result and the second fault prediction result, an alarm can be issued.

[0191] Apply the method provided in the embodiment of the present application,

[0192] Acquire multiple pieces of detection data information and a disk health score corresponding to each piece of detection data information; use the disk health score to determine the attribution of each piece of detection data information, and obtain predicted data and direct estimation data; wherein, the prediction difficulty of the predicted data is higher than that of the direct estimation data; use the disk health score corresponding to the detection data information in the direct estimation data to determine the first fault prediction result; input the predicted data into the target model for fault prediction, and obtain the second fault prediction result.

[0193] In the present application, when multiple detection data information and the disk health score corresponding to each detection data information are obtained. For the disk health score, the score size is related to the prediction difficulty. For example, the more extreme the score, the more accurate the fault prediction result. Therefore, the attribution of each detection data information can be determined based on the disk health score, thereby obtaining prediction data and direct estimation data. After completing the attribution division, the prediction difficulty of the obtained prediction data is higher than that of the direct estimation data. Therefore, the detection data information in the direct estimation data can be directly based on the corresponding disk health score to obtain a first fault prediction result. For prediction data with high prediction difficulty, the target model is used to perform fault prediction to obtain a second fault prediction result. In other words, for multiple detection data information to be predicted, the prediction difficulty can be screened and classified through the disk health score. For the part belonging to the direct estimation data, the disk fault prediction is directly based on the disk health score, and the first fault prediction result can be quickly obtained. For the part belonging to the prediction data, the fault prediction is performed with the help of the target model to obtain the second fault prediction result. Because the target model prediction is not simply based on the disk health score, but is directly based on the specific content of the detection data information, and the target model has been trained in advance, it can have a higher accuracy rate than fault prediction directly based on the disk health score.

[0194] Technical effect of the present application: In general, relatively accurate data can be obtained based on the disk health score. Therefore, by dividing the detection data information into different categories based on the disk health score, a large amount of direct estimation data and a small amount of prediction data can be obtained. For a large amount of direct estimation data, an accurate first fault prediction result can be quickly determined based on the disk health score, and for a small amount of prediction data for which it is difficult to obtain an accurate result based on the disk health score, fault prediction is performed based on the target model to obtain an accurate second fault prediction result. It can be seen that the present application can not only improve the efficiency of disk fault prediction, but also improve the accuracy of fault prediction.

[0195] Corresponding to the above method embodiment, the embodiment of the present application further provides a disk failure prediction device. The disk failure prediction device described below and the disk failure prediction method described above can refer to each other.

[0196] See also Figure 3 As shown, the device includes the following units:

[0197] The data acquisition unit 101 is used to acquire multiple pieces of detection data information and a disk health score corresponding to each piece of detection data information;

[0198] The data screening unit 102 is used to determine the ownership of each detection data information by using the disk health score to obtain the predicted data and the direct estimated data;

[0199] A first fault prediction unit 103, configured to determine a first fault prediction result by using a disk health score corresponding to the detection data information in the direct estimation data;

[0200] The second fault prediction unit 104 is used to input the prediction data into the target model for fault prediction to obtain a second fault prediction result.

[0201] Using the device provided in the embodiment of the present application, multiple detection data information and the disk health score corresponding to each detection data information are obtained; the disk health score is used to determine the attribution of each detection data information to obtain predicted data and direct estimation data; wherein, the prediction difficulty of the predicted data is higher than that of the direct estimation data; the disk health score corresponding to the detection data information in the direct estimation data is used to determine the first fault prediction result; the predicted data is input into the target model for fault prediction to obtain the second fault prediction result.

[0202] In the present application, when multiple detection data information and the disk health score corresponding to each detection data information are obtained. For the disk health score, the score size is related to the prediction difficulty. For example, the more extreme the score, the more accurate the fault prediction result. Therefore, the attribution of each detection data information can be determined based on the disk health score, thereby obtaining prediction data and direct estimation data. After completing the attribution division, the prediction difficulty of the obtained prediction data is higher than that of the direct estimation data. Therefore, the detection data information in the direct estimation data can be directly based on the corresponding disk health score to obtain a first fault prediction result. For prediction data with high prediction difficulty, the target model is used to perform fault prediction to obtain a second fault prediction result. In other words, for multiple detection data information to be predicted, the prediction difficulty can be screened and classified through the disk health score. For the part belonging to the direct estimation data, the disk fault prediction is directly based on the disk health score, and the first fault prediction result can be quickly obtained. For the part belonging to the prediction data, the fault prediction is performed with the help of the target model to obtain the second fault prediction result. Because the target model prediction is not simply based on the disk health score, but is directly based on the specific content of the detection data information, and the target model has been trained in advance, it can have a higher accuracy rate than fault prediction directly based on the disk health score.

[0203] Technical effect of the present application: In general, relatively accurate data can be obtained based on the disk health score. Therefore, by dividing the detection data information into different categories based on the disk health score, a large amount of direct estimation data and a small amount of prediction data can be obtained. For a large amount of direct estimation data, an accurate first fault prediction result can be quickly determined based on the disk health score, and for a small amount of prediction data for which it is difficult to obtain an accurate result based on the disk health score, fault prediction is performed based on the target model to obtain an accurate second fault prediction result. It can be seen that the present application can not only improve the efficiency of disk fault prediction, but also improve the accuracy of fault prediction.

[0204] In a specific implementation of the present application, it also includes:

[0205] The optimization training unit is used to compare the disk replacement record with the second fault prediction result to obtain a comparison result; in the process of optimizing the target model using the training data, rewards and punishments are given based on the comparison result.

[0206] In a specific implementation of the present application, the optimization training unit is specifically used to determine the reward and punishment value using the comparison result;

[0207] In the process of optimizing the target model using training data, the reward and penalty values ​​are used to indicate the optimization direction of the target model.

[0208] In a specific implementation of the present application, the optimization training unit is specifically used to determine the preset reward value as the reward and penalty value when the comparison result corresponds to the disk replacement record and the second fault prediction result matches;

[0209] In the case that the disk replacement record corresponding to the comparison result does not match the second fault prediction result, the preset penalty value is determined as the reward and penalty value.

[0210] In a specific implementation of the present application, the optimization training unit is specifically used to determine the preset first reward value as the reward and penalty value if the disk replacement record is a disk replacement when the comparison result corresponds to a match between the disk replacement record and the second fault prediction result;

[0211] In the case that the disk replacement record corresponding to the comparison result matches the second fault prediction result, if the disk replacement record is that the disk has not been replaced, the preset second reward value is determined as the reward or penalty value;

[0212] The first reward value is greater than the second reward value.

[0213] In a specific implementation of the present application, the optimization training unit is specifically used to substitute the reward and punishment values ​​into the advantage function to calculate the difference between the expected reward and the predicted reward during the optimization training of the target model using the training data;

[0214] Based on the difference, the target model is updated.

[0215] In a specific implementation of the present application, the second fault prediction unit is specifically used to add manufacturer information and disk attribute information to each piece of detection data information belonging to the prediction data;

[0216] After completing the data addition, each piece of detection data information is converted into a vector;

[0217] Group the obtained vectors to obtain a vector group;

[0218] The vector group is input into the target model for fault prediction to obtain a second fault prediction result; wherein the target model is a model that has learned the fault prediction capabilities of disks of different manufacturers and different attributes.

[0219] In a specific implementation of the present application, the second fault prediction unit is specifically used to input the vector group into the target model, and use the information cross-learning module in the target model to perform cross-learning between different feature dimensions in the vector group to obtain effective features;

[0220] Use the time series information module in the strategy network to organize the effective features and obtain statistical characteristics;

[0221] The statistical characteristics are normalized using the normalization layer in the strategy network to obtain the second fault prediction result.

[0222] In a specific implementation of the present application, the second fault prediction unit is specifically used to determine the dependency relationship between different positions in the effective features using the encoder in the timing information module to obtain a global feature representation;

[0223] The residual network in the temporal information module is used to perform residual processing on the global feature representation to obtain statistical characteristics.

[0224] In a specific implementation of the present application, it also includes:

[0225] The data cleaning unit is used to determine the ownership of each piece of detection data information by using the disk health score, and before obtaining the predicted data and the direct estimated data, search for invalid indicators in each piece of detection data information, and when the invalid indicators are found, increase the invalid index of the detection data information by one; wherein each piece of detection data information includes multiple indicators;

[0226] If the invalidity index exceeds the preset threshold, the detection data information is deleted.

[0227] In a specific implementation of the present application, the data cleaning unit is specifically used to delete the piece of detection data information after obtaining the disk information corresponding to the piece of detection data information if the invalid index exceeds a preset threshold;

[0228] Use disk information to output prompt information of data collection abnormalities.

[0229] In a specific implementation of the present application, the data acquisition unit is specifically used to detect the disk using a disk self-test tool to obtain multiple detection data information; wherein one detection data information includes multiple indicator data;

[0230] Obtain the indicator threshold set corresponding to each piece of detection data information, and compare the current state value of each indicator in the detection data information with the corresponding threshold in the indicator threshold set;

[0231] According to the status comparison result, the disk health score corresponding to the detection data information is determined.

[0232] Please refer to Figure 4 In practical applications, a disk failure prediction system corresponding to the above disk failure prediction method can also be built, and the system includes:

[0233] The data collection and screening unit collects SMART information related data and performs preliminary screening of the information in units of disks according to the heuristic rule method in the first embodiment, thereby reducing the decision difficulty of subsequent models and the amount of calculation.

[0234] The fault prediction unit adopts the disk fault prediction algorithm based on the first embodiment as the core, sets a reward function, and predicts the fault of a disk based on the SMART information within a period of time according to the model, so as to guide the model to make better decisions.

[0235] The alarm monitoring unit generates corresponding alarms based on the current disk status information and the health level prediction generated by the model, and reminds the administrator to handle the corresponding disk in time.

[0236] The disk status update unit responds to disk updates. If the administrator updates the disk, the alarm list is updated and the corresponding information of the new disk is added to the prediction module to predict the new health level.

[0237] In summary, this application proposes a prediction technology that combines expert knowledge strategy with reinforcement learning (RL) algorithm, combines the advantages of both heuristic rules and machine learning algorithms, improves the accuracy of decision-making, reduces the decision-making pressure of the algorithm model, and improves the decision-making efficiency of the fault prediction platform.

[0238] A more detailed classification of disk health is proposed, which divides the health level according to the scores of heuristic rules, thereby providing more accurate suggestions for fault prediction.

[0239] Introducing encoder processing into the training model can improve the model's ability to predict sequential text information.

[0240] The technical solution provided by this application can be applied not only to disk failure prediction (DFP, Disk Failure Prediction), but also to the field of server component failure prediction. It is necessary to replace the disk detection data information with the component detection data information, collect the corresponding component information, and then train the network parameters based on the above-mentioned incentive function and target model, so as to achieve fast and accurate prediction of server component failure.

[0241] Corresponding to the above method embodiment, the embodiment of the present application further provides an electronic device. The electronic device described below and the disk failure prediction method described above can refer to each other.

[0242] See also Figure 5 As shown, the electronic device includes:

[0243] A memory 332, for storing computer programs;

[0244] The processor 322 is used to implement the steps of the disk failure prediction method of the above method embodiment when executing a computer program.

[0245] For details, please refer to Figure 6 , Figure 6 A specific structural diagram of an electronic device provided for this embodiment, the electronic device may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) (for example, one or more processors) and a memory 332, and the memory 332 stores one or more computer programs 342 or data 344. Among them, the memory 332 can be a temporary storage or a permanent storage. The program stored in the memory 332 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the data processing device. Furthermore, the processor 322 can be configured to communicate with the memory 332 to execute a series of instruction operations in the memory 332 on the electronic device 301.

[0246] The electronic device 301 may further include one or more power supplies 326 , one or more wired or wireless network interfaces 350 , one or more input and output interfaces 358 , and / or one or more operating systems 341 .

[0247] The steps in the disk failure prediction method described above can be implemented by the structure of an electronic device.

[0248] Corresponding to the above method embodiment, the embodiment of the present application further provides a readable storage medium. The readable storage medium described below and the disk failure prediction method described above can refer to each other.

[0249] A readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the disk failure prediction method of the above method embodiment are implemented.

[0250] The readable storage medium may specifically be a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or other readable storage medium that can store program codes.

[0251] Corresponding to the above method embodiment, the embodiment of the present application further provides a computer program product. The computer program product described below and the disk failure prediction method described above can refer to each other.

[0252] A computer program product includes a computer program / instruction. When the computer program / instruction is executed by a processor, the steps of the disk failure prediction method are implemented.

[0253] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0254] Those skilled in the art may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0255] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0256] Finally, it should be noted that, in this article, relationships such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms include, include or any other variations are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0257] Specific examples are used herein to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core idea of ​​the present application. At the same time, for those skilled in the art, according to the idea of ​​the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A disk failure prediction method, characterized in that: include: Get multiple pieces of test data information and the disk health score corresponding to each piece of test data information; The attribution of each piece of detection data information is determined by using the disk health score to obtain predicted data and direct estimated data; wherein the prediction difficulty of the predicted data is higher than the prediction difficulty of the direct estimated data; Determine a first fault prediction result by using the disk health score corresponding to the detection data information in the direct estimation data; The prediction data is input into the target model for fault prediction to obtain a second fault prediction result.

2. The method according to claim 1, characterized in that Also includes: Comparing the disk replacement record with the second fault prediction result to obtain a comparison result; In the process of optimizing the target model by using the training data, rewards and punishments are given based on the comparison results.

3. The method according to claim 2, characterized in that In the process of optimizing the target model by using the training data, rewards and punishments are given based on the comparison results, including: Determining reward and punishment values ​​using the comparison results; In the process of optimizing the target model by using the training data, the reward and penalty values ​​are used to indicate the optimization direction of the target model.

4. The method according to claim 3, characterized in that Determining reward and punishment values ​​using the comparison results includes: In the case that the comparison result corresponds to that the disk replacement record matches the second fault prediction result, determining a preset reward value as the reward or penalty value; In the case that the comparison result corresponds to the disk replacement record and the second fault prediction result does not match, a preset penalty value is determined as the reward and penalty value.

5. The method according to claim 4, characterized in that In the case that the comparison result corresponds to that the disk replacement record matches the second fault prediction result, determining a preset reward value as the reward and penalty value includes: In the case where the comparison result corresponds to that the disk replacement record matches the second fault prediction result, if the disk replacement record is replacement of a disk, a preset first reward value is determined as the reward or penalty value; In the case where the comparison result corresponds to that the disk replacement record matches the second fault prediction result, if the disk replacement record is that the disk is not replaced, determining the preset second reward value as the reward and penalty value; Wherein, the first reward value is greater than the second reward value.

6. The method according to claim 3, characterized in that In the process of optimizing the target model by using the training data, using the reward and penalty value to indicate the optimization direction of the target model includes: In the process of optimizing the target model using the training data, substituting the reward and punishment value into the advantage function to calculate the difference between the expected reward and the predicted reward; Based on the difference, the target model is updated.

7. The method according to claim 1, characterized in that Inputting the prediction data into the target model for fault prediction to obtain a second fault prediction result includes: Adding manufacturer information and disk attribute information to each piece of detection data information belonging to the prediction data; After completing the data addition, each piece of detection data information is converted into a vector; Group the obtained vectors to obtain a vector group; The vector group is input into the target model for fault prediction to obtain the second fault prediction result; wherein the target model is a model that has learned the fault prediction capabilities of disks of different manufacturers and different attributes.

8. The method according to claim 7, characterized in that Inputting the vector group into the target model for fault prediction to obtain the second fault prediction result includes: Inputting the vector group into the target model, and using the information cross-learning module in the target model to perform cross-learning between different feature dimensions in the vector group to obtain effective features; Using the time series information module in the strategy network to sort out the effective features to obtain statistical characteristics; The statistical characteristics are normalized using a normalization layer in the strategy network to obtain the second fault prediction result.

9. The method according to claim 8, characterized in that The effective features are sorted using the time series information module in the strategy network to obtain statistical characteristics, including: Determine the dependency relationship between different positions in the effective features using the encoder in the temporal information module to obtain a global feature representation; The residual network in the temporal information module is used to perform residual processing on the global feature representation to obtain the statistical characteristics.

10. The method according to claim 1, characterized in that Before using the disk health score to determine the ownership of each piece of detection data information and obtaining the predicted data and the direct estimated data, the following steps are also included: Searching for invalid indicators in each piece of detection data information, and when an invalid indicator is found, adding one to the invalid index of the piece of detection data information; wherein each piece of detection data information includes a plurality of indicators; If the invalidity index exceeds a preset threshold, the detection data information is deleted.

11. The method according to claim 10, characterized in that If the invalidity index exceeds a preset threshold, the detection data information is deleted, including: If the invalidity index exceeds a preset threshold, after obtaining the disk information corresponding to the piece of detection data information, the piece of detection data information is deleted; The disk information is used to output prompt information of data collection abnormality.

12. The method according to any one of claims 1 to 11, characterized in that: Get multiple test data information and the disk health score corresponding to each test data information, including: Using a disk self-test tool to test the disk, a plurality of test data information is obtained; wherein a piece of test data information includes a plurality of indicator data; Obtaining the indicator threshold set corresponding to each piece of detection data information, and comparing the current state value of each indicator in the detection data information with the corresponding threshold in the indicator threshold set; According to the status comparison result, the disk health score corresponding to the detection data information is determined.

13. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the disk failure prediction method according to any one of claims 1 to 11 when executing the computer program.

14. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the disk failure prediction method according to any one of claims 1 to 11 are implemented.

15. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the disk failure prediction method described in any one of claims 1 to 11 are implemented.

Images