Enterprise-level solid state disk fault early warning method based on multi-instance adversarial learning

Abstract

The application discloses a kind of enterprise-level solid state disk fault early warning methods of multi-instance confrontation learning, based on the most relevant monitoring index, using multi-instance learning method long time sequence SMART data is converted into the instance of multiple solid state disk state degradation data, bagging analysis is carried out to multiple instances, realize solid state disk long time sequence fault information fusion;Using the alignment known solid state disk fault feature distribution and unknown solid state disk fault feature distribution of confrontation domain adaptation network based on multi-instance learning, the known solid state disk fault feature knowledge is migrated to unknown solid state disk data.Through self-adaptive fault early warning model, the health status of solid state disk at the current time is determined.The application can learn multiple solid state disk multiple fault feature knowledge, realize online fault early warning based on solid state disk state monitoring, can guarantee in zero false alarm, obtain high level of fault detection rate.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis and reliability engineering, and specifically relates to an enterprise-level solid-state drive fault early warning method based on multi-instance adversarial learning. Background Technology

[0002] Solid-state drives (SSDs) generally refer to storage devices composed of arrays of solid-state electronic storage chips. Structurally, they consist of two parts: a control unit integrated with random access memory (RAM) and storage units integrated with flash memory chips. SSDs are identical to traditional hard disk drives (HDDs) in terms of interface specifications and definitions, functions, and usage. However, their structure differs from HDDs due to the absence of mechanical components. Relying on the characteristics of their internal solid-state electronic storage chips, SSDs offer advantages such as high read / write speeds, a wide operating temperature range, portability, low power consumption, shock resistance, and high stability. Consequently, they are widely used in military, automotive, industrial control, video surveillance, network monitoring, network terminals, power, medical, aviation, and navigation equipment. With the continuous increase in data volume and the growing demand for high-speed, stable cloud services, data security issues are becoming increasingly prominent. This not only relates to individual interests and corporate operations but also directly impacts national security. New SSD manufacturing processes and microelectronics technologies have enabled the continuous miniaturization and reduction of the spacing between internal storage units, allowing for the application of stacked arrays of these units. The ever-increasing demands for performance and storage capacity also bring challenges to data storage security. For key components of data storage, reliability analysis and health management methods for solid-state drives (SSDs) have become new research hotspots.

[0003] Currently, most research in the field of solid-state drive (SSD) failure prediction is based on SMART (Self-Monitoring, Analysis, and Reporting Technology) data. Existing SSD failure prediction models built around SMART data, based on statistical and machine learning methods, largely rely on abundant datasets, including data from SSDs marked as faulty and those still functioning normally. Furthermore, the established evaluation metrics are limited to similar types of SSDs and cannot effectively guide users in maintaining storage devices. Therefore, it is necessary to explore research on SSD failure prediction and reliability assessment using methods different from existing statistical and machine learning approaches.

[0004] In practical engineering, thanks to the rapid development of deep learning, deep learning-based multi-instance fault diagnosis technology and cross-domain fault diagnosis methods based on transfer learning are highly promising tools in the field of deep learning. From the classic MMD domain probability distribution alignment to the more popular domain adversarial domain adaptation method, it has been successfully applied to target localization, tracking, speech recognition, speech enhancement and denoising, system identification parameter estimation, and fault detection of dynamic systems. Domain adversarial domain adaptation is a domain adaptation learning method based on adversarial generative networks, which can transfer knowledge from a source domain with rich labeled data to a target domain lacking labeled information for model training. It is very suitable for solving the fault diagnosis problem of new hard drives lacking diagnostic data. It is worth noting that in recent years, although deep learning-based fault diagnosis methods have been increasingly applied to the fault diagnosis of extremely complex large-scale systems such as mobile robots, gas turbine engines, and high-speed trains, research on fault prediction for solid-state drives is relatively limited. Therefore, for relatively complex systems such as solid-state drives, attempting to use particle filtering methods for fault diagnosis is an effective way to compensate for the shortcomings of statistical and machine learning methods, but it is also full of many unknown challenges. Summary of the Invention

[0005] To address the shortcomings and challenges of existing technologies, this invention aims to provide an enterprise-level solid-state drive (SSD) fault prediction method based on multi-instance adversarial learning. This method extracts fault features from the richly labeled SMART dataset using multi-instance learning, then transfers this extracted feature knowledge to a novel SSD dataset lacking fault label data using a domain adversarial adaptation method. A fault prediction model is then constructed in the target domain using this feature knowledge. This model, combined with a multi-point threshold alarm mechanism, determines whether a hard drive is about to fail. Finally, the effectiveness of the method is demonstrated by statistically analyzing the fault prediction accuracy across all tested hard drives. This method solves both the challenge of extracting SSD fault features and the problem of lacking diagnostic knowledge in novel SSDs, thus enabling a cross-domain SSD fault prediction model.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] An enterprise-grade solid-state drive (SSD) fault prediction method based on multi-instance adversarial learning includes the following steps:

[0008] The most relevant monitoring indicators for SSD fault early warning were selected by filtering the SMART parameter set of SSD operating data.

[0009] The most relevant monitoring index data is transformed into multiple instances of solid-state drive (SSD) state degradation data using a multi-instance learning method. The multiple instances are then bagged and analyzed to achieve long-term series fault information fusion of SSDs, resulting in long-sequence SSD fusion information.

[0010] An adversarial domain adaptation network based on multi-instance learning is used to align the known and unknown solid-state drive (SSD) fault feature distributions. The knowledge of known SSD fault features is then transferred to the unknown SSD data to obtain fault warning transfer features. Based on the obtained fault warning transfer features, an adaptive fault warning model is obtained. The current health status of the SSD is determined using the adaptive fault warning model based on the long sequence of SSD fusion information.

[0011] Furthermore, the specific steps include:

[0012] 1) Filter the most relevant monitoring indicators for SSD fault early warning from the SMART parameter set of SSD operating data;

[0013] 2) The most relevant monitoring indicators for solid-state drive (SSD) failure early warning are divided into instances to obtain K instance samples within the same SSD's lifespan. These K instance samples are then packaged into a set, known as instance bag X. j ={x1,x2,…x K}, where j is the instance bag count, and the solid-state drive instance bags X = {X1, X2, ..., X} j} and solid-state drive labels Y = {y1, y2, ..., y j The training set is constructed; based on the training set, long-sequence solid-state drive fusion information is obtained.

[0014] 3) Based on the adversarial transfer model of solid-state drive feature knowledge and the fusion information of long-sequence solid-state drives, an adaptive fault warning model suitable for different individual solid-state drives is obtained. Based on the adaptive fault warning model suitable for different individual solid-state drives, the known solid-state drive fault feature knowledge is transferred to the unknown solid-state drive data to determine the health status of the hard drive at the current moment and obtain the fault warning of a single solid-state drive.

[0015] Furthermore, the types of hard disk failures include data saving errors, write interference errors and read interference errors, and P / E cycle errors.

[0016] Furthermore, the most relevant monitoring metrics for SSD fault prediction include read error rate, reallocation sector count, available reserved block count, wear leveling count, reallocation event count, and uncorrectable error count.

[0017] Furthermore, the long-sequence solid-state drive fusion information based on the training set is obtained through the following process:

[0018] Supervised learning algorithms such as multi-instance learning convolutional neural networks are used to learn from the training set data to obtain long-sequence solid-state drive fusion information based on the training set.

[0019] Furthermore, the long-sequence solid-state drive fusion information based on the training set is calculated using the following formula:

[0020]

[0021] Where S∈[0,1], 0 represents a predicted failure of the SSD requiring replacement, 1 represents the current healthy state of the SSD, X represents the entire instance bag representing the lifespan data of all SSDs, K represents the number of instance bags (i.e., the number of SSDs), i represents the instance count, and x i This represents the i-th instance.

[0022] Furthermore, the adversarial transfer model of solid-state drive feature knowledge is obtained through the following process: using an adversarial domain adaptation network to align the probability distribution of known solid-state drive fault feature datasets with the probability distribution of unknown solid-state drive fault feature datasets, an adversarial transfer model of solid-state drive feature knowledge is established.

[0023] Furthermore, the overall loss function L of the solid-state drive feature knowledge adversarial transfer model. global (θ fe ,θ lc ,θ dd )as follows:

[0024] L global (θ fe ,θ lc ,θ dd )=L(θ fe )+L(θ lc )+λL adv

[0025] In the formula, θ fe For the feature extractor parameters, θ lc For the label classifier parameters, θ dd Let L(θ) be the initial inter-domain trainer parameters. fe L(θ) represents the feature extractor loss. lc L is the loss of the label classifier. adv Let λ be the domain adversarial loss and λ(·) be the gradient flip function.

[0026] Furthermore, the health status of the hard drive at the current moment is determined using a threshold method.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] This invention constructs the most relevant monitoring index data based on the original SMART parameter set of solid-state drives (SSDs), which better reflects the dynamic abnormal behavior during SSD operation. This helps the fault prediction algorithm detect faults more efficiently, resulting in a higher fault detection rate. The multi-instance learning method employed in this invention effectively extracts features from SSD status monitoring parameters, significantly improving the accuracy of the fault prediction algorithm. The implementation process is simple and easy to master. The SSD feature knowledge adversarial transfer model for reliability assessment in this invention has simpler application conditions, and diagnostic knowledge can be transferred from existing fault data in SSDs, providing stronger guidance for storage device maintenance personnel. The hard drive fault early warning and reliability assessment method of this invention is simple, reliable, and easy to implement in engineering practice for online SSD fault detection.

[0029] Furthermore, the adversarial domain adaptation transfer method adopted in this invention, that is, using an adversarial domain adaptation network, aligns the probability distribution of known solid-state drive (SSD) fault feature datasets with the probability distribution of unknown SSD fault feature datasets to establish an adversarial transfer model of SSD feature knowledge. This method can better solve the fault prediction problem of new SSDs that lack labeled data, and its estimation accuracy is higher and it is less likely to produce false alarms of hard drive faults.

[0030] Furthermore, this invention uses a threshold method to determine the current health status of the hard drive, which can better adapt to the differences between different solid-state drives and make various reliability evaluation indicators better. Attached Figure Description

[0031] Figure 1 A flowchart for proposing an enterprise-level solid-state drive fault early warning and reliability assessment method based on multi-instance adversarial learning;

[0032] Figure 2 Schematic diagram for learning from multiple instances;

[0033] Figure 3 This is a schematic diagram of the feature extraction branch;

[0034] Figure 4 A statistical histogram of the accuracy of fault warnings for 100 solid-state drives;

[0035] Figure 5 The ROC curve of the model for the early warning results of 100 healthy hard drives. Detailed Implementation

[0036] The present invention will now be described in detail with reference to the accompanying drawings:

[0037] This invention selects the most relevant monitoring indicators for SSD fault early warning by using the SMART parameter set of SSD operating data. Based on these most relevant indicators, a multi-instance learning method is used to transform them into multiple instances of SSD state degradation data. These instances are then bagged for analysis to achieve long-term series fault information fusion. Adversarial transfer theory is introduced, and an adversarial domain adaptation network based on multi-instance learning is used to align the known and unknown SSD fault feature distributions, transferring known SSD fault feature knowledge to unknown SSD data. Based on the obtained fault early warning transfer features, an adaptive fault early warning model suitable for different SSD models is designed, and the current predicted value is used to determine the current SSD health status. This method can learn multiple fault feature knowledge of various SSDs, fully utilize all useful information, achieve online fault early warning based on SSD state monitoring, and ensure a high fault detection rate with zero false alarms. It also outputs a more indicative and instructive evaluation of the health status of individual SSDs.

[0038] For details, see Figure 1 The present invention provides an enterprise-level solid-state drive fault early warning method based on multi-instance adversarial learning, comprising the following steps:

[0039] 1) Based on the failure mechanism of solid-state drives (SSDs), the most relevant monitoring indicators for SSD fault early warning are selected from the collected SSD operating data SMART parameter set. The specific process is as follows:

[0040] First, since the main failure mode of solid-state drives (SSDs) is conventionally caused by the flash memory chips that store data, and the failure types of flash memory chips can be mainly divided into the following categories:

[0041] Data storage error: When a solid-state drive stores data, because the tunnel oxide layer and the charge trapping layer are not completely insulated, charge leakage occurs in the charge storage layer. This causes the charge level of the cell to decrease, resulting in its charge level shifting to the charge level of other data states. In subsequent identification, it is read as other data values, causing data errors.

[0042] Write interference errors and read interference errors: Due to the voltage applied to the word line and bit line forming tiny capacitances in the tunnel oxide layer and gate oxide layer, the tiny voltages can cause mutual inductance because the memory cell array is dense, resulting in a shift in the charge level of the adjacent memory cells and causing data errors.

[0043] P / E Cycle Error: As the number of P / E cycles increases, the wear and tear on the storage cells of a solid-state drive (SSD) intensifies. Wear and tear refers to the reduced ability of the charge trapping layer to retain charge and the presence of electrical stress in the tunnel oxide layer. This electrical stress accelerates the rate of charge leakage in the storage cells, and the cells cannot return to the lowest potential after erasure, making it difficult to erase and write data normally.

[0044] In summary, the original SMART parameters related to the above-mentioned fault mechanisms are first used as the parameter set for the solid-state drive fault early warning method.

[0045] Secondly, the correlation between the selected SMART parameter set and the solid-state drive failure characteristics is evaluated using the random forest method to obtain the most relevant monitoring indicators. This process includes the following steps:

[0046] Each known sample with different SMART parameters is divided into a training set and a test set. A model capable of identifying faults is trained on each training set. The test set is then used for evaluation, and the SMART group with the highest accuracy is considered the fault characteristic. The most relevant set of detection parameters for SSD fault warning is then selected, including: Read Error Rate, Reallocated Sector Count, Available Reserved Blocks Count, Wear Leveling Count, Reallocation Event Count, and Uncorrectable Sector Count.

[0047] 2) Based on the most relevant monitoring indicators selected for solid-state drive (SSD) failure early warning, which are long-term series sets, the most relevant monitoring indicators are divided into instances to obtain K instance samples within the same SSD lifecycle. Let i be the instance count, i = 1 - K. Pack the K instance samples into a set, called instance bag X. j ={x1,x2,…x K} j represents the instance bag count. The most relevant monitoring metrics for each hard drive are considered as an instance bag, where each instance corresponds to a sequence of monitoring parameters for a specific time period on a particular solid-state drive. Solid-state drive status is divided into impending failure and healthy states, and the solid-state drive status label is denoted as Y = {y1, y2, ..., y...}. j}, where the state y∈{0,1}, 0 represents a fault, and 1 represents a healthy state. Solid-state drive instance bag X={X1,X2,…X j} and Y = {y1, y2, ..., y jTogether, they constitute the training set.

[0048] See Figure 2 In theory, based on the selected training set, a multi-instance learning algorithm is used to learn the training set data and achieve the fusion of long-sequence time-series fault information to obtain long-sequence solid-state drive fusion information.

[0049] See Figure 3 Supervised learning algorithms such as multi-instance learning convolutional neural networks are used to learn from the training set data. Convolutional pooling operations are used to extract data features with the support of an attention mechanism, and long-sequence solid-state drive fusion information is obtained. Where S∈[0,1], 0 represents a predicted failure of the SSD requiring replacement, 1 represents the current healthy state of the SSD, X represents the entire instance bag representing the lifespan data of all SSDs, K represents the number of instance bags (i.e., the number of SSDs), i represents the instance count, and x i This represents the i-th segment in the lifespan data of the i-th instance, i.e., the all-solid-state drive.

[0050] 3) Introducing the adversarial domain adaptation algorithm, an adversarial domain adaptation network is used to align the probability distributions of known SSD fault feature datasets with those of unknown SSD fault feature datasets, establishing an adversarial transfer model for SSD feature knowledge. The specific process is as follows:

[0051] When t=1, the initialization step is:

[0052] Initialize the inter-domain trainer G dd (θ dd ) and label classifier G lc (θ lc The parameters of the label classifier θ are set. lc With the initial inter-domain trainer parameters θ dd The set of known solid-state drive (SSD) fault features containing labeled information in the training set is summarized into the source domain. For source domain samples, For source domain labels, n s For source domain sample counting, the set of unknown solid-state drive fault features lacking labeling information is summarized into the target domain. For the target domain sample, n T Count the samples in the target domain.

[0053] When t = 2:N, the training steps consist of forward propagation and backward propagation in each step:

[0054] Forward propagation:

[0055] From D S D TExtract the samples required for the current training step from G. dd (θ dd ) and G lc (θ lc The domain discriminant prediction value d is calculated for each sample. i Results of label classifier Where, θ fe These are the parameters for the feature extractor.

[0056] Backpropagation:

[0057] Calculate the feature extractor loss L(θ) respectively fe ):

[0058]

[0059] In the formula, N is the total number of source domain instance bags, and j is the source domain instance bag count. For the example bag label input, This is the output instance bag label.

[0060] Label classifier loss L(θ) lc ):

[0061]

[0062] In the formula, K is the total number of instance bags in the target domain, and l is the count of instance bags in the target domain. For instance tags, Let I be the source domain sample, and let I be the label indicator function.

[0063] Domain confrontation loss L adv :

[0064]

[0065] In the formula, d i For weights.

[0066] And construct the overall loss function:

[0067] L global (θ fe ,θ lc ,θ dd )=L(θ fe )+L(θ lc )+λL adv

[0068] Where λ(·) is the gradient flip function.

[0069] The loss function is optimized and the parameters are updated in reverse:

[0070]

[0071]

[0072]

[0073] The updated parameters are saved into the model and the parameters of each layer of the model, except for the last output layer, are frozen to form a solid-state drive feature knowledge adversarial transfer model.

[0074] 4) Based on the obtained solid-state drive feature knowledge, the adversarial migration model and the long sequence solid-state drive fusion information are used to design an adaptive fault warning model suitable for different individual hard drives, and the threshold method is used to determine the health status of the hard drive at the current moment to obtain fault warnings for individual solid-state drives.

[0075] After performing fault warning processing on multiple hard drive samples using the above method, the relevant indicators used to evaluate hard drive reliability were statistically analyzed. The specific process is as follows:

[0076] (1) The statistical characteristics (mean and standard deviation) of the source domain solid-state disk dataset and the target domain solid-state disk dataset are learned by the solid-state disk feature knowledge adversarial transfer model. The parameters are stored in the solid-state disk feature knowledge adversarial transfer model. Combined with the long sequence solid-state disk fusion information, an adaptive fault warning model suitable for different disk individuals is designed when the target solid-state disk labeling parameters are insufficient.

[0077] (2) After performing fault warning processing on multiple hard drive samples using the above method, statistical analysis was conducted on relevant indicators used to evaluate hard drive reliability, including:

[0078] Classification accuracy (ACC):

[0079] Receptivity curve (ROC):

[0080] Solid State Drive Failure Warning Indicator: P out =Model(D in )

[0081] Where, n r n represents the number of solid-state drives (SSDs) whose fault predictions were correct. f This represents the total number of solid-state drives (SSDs). The closer the ROC curve is to the top left corner, the better the model's performance. (D) in This indicates the current monitoring data for the solid-state drive input model status, P. out This represents the output value of the adaptive fault warning model, and its value ranges from P. out= [0,1], the closer the value is to 1, the closer the solid-state drive is to failure. The health status warning of the solid-state drive is completed by setting the alarm threshold. Let 0 be a negative example and 1 be a positive example. Then TP represents the proportion of positive examples correctly predicted as positive examples, FP represents the proportion of negative examples incorrectly predicted as positive examples, FN represents the proportion of positive examples incorrectly predicted as negative examples, and TN represents the proportion of negative examples correctly predicted as negative examples.

[0082] By implementing the above three-part process, the results of the embodiments show that the enterprise-level solid-state drive fault warning and reliability assessment method based on multi-instance adversarial learning proposed in this invention can ensure a high level of fault warning accuracy. In addition, it can also obtain a reasonable model stability for server operators, which can fully guide them to carry out large-scale replacement of storage spare parts.

[0083] Example

[0084] An enterprise-level SSD fault prediction and reliability assessment method based on multi-instance adversarial learning was implemented on the daily SMART monitoring parameters collected from 100 SSDs in a data center over the past year. The results of this assessment method are as follows: Figure 4 and Figure 5 As shown, the MALDAN method is the method proposed in this invention, while common methods in related fields such as CNN, DDC, DANN, MIL, and MIAL are set as control groups.

[0085] from Figure 4 It can be seen that the enterprise-level solid-state drive fault warning and reliability assessment method based on multi-instance adversarial learning proposed in this invention has a higher accuracy rate than similar methods.

[0086] from Figure 5 It can be seen that the enterprise-level solid-state drive fault warning and reliability assessment method based on multi-instance adversarial learning proposed in this invention has a better fault warning effect than other models.

[0087] This invention addresses the practical problems of enterprise-level solid-state drive (SSD) fault prediction engineering. First, it selects raw SMART data from the SSD based on its main failure modes and chooses the parameter set most relevant to the fault. Second, based on the selected training set, it uses a multi-instance learning algorithm to learn from the training data and obtain fault feature prediction knowledge. Third, it uses an adversarial transfer domain adaptation method to transfer the fault prediction knowledge to a novel SSD dataset lacking labeled data and performs fault prediction to obtain model prediction accuracy. Finally, it sets an adaptive fault judgment threshold based on fault statistical characteristics (mean, variance) and statistically evaluates relevant reliability indicators. This method is simple to implement, highly effective, and provides an effective approach for enterprise-level SSD fault prediction.

Claims

1. A method for enterprise-level solid state disk failure warning based on multi-instance adversarial learning, characterized in that, Includes the following steps: The most relevant monitoring indicators for SSD fault early warning were selected by filtering the SMART parameter set of SSD operating data. The most relevant monitoring index data is transformed into multiple instances of solid-state drive (SSD) state degradation data using a multi-instance learning method. The multiple instances are bagged and analyzed to construct a training set. Based on the training set, long-term series fault information of SSDs is fused to obtain long-term SSD fusion information. An adversarial domain adaptation network based on multi-instance learning is used to align the known and unknown solid-state drive (SSD) fault feature distributions, and the known SSD fault feature knowledge is transferred to the unknown SSD data to obtain fault warning transfer features. Based on the obtained fault warning transfer features, an adaptive fault warning model is obtained. Based on the long sequence of SSD fusion information, the adaptive fault warning model is used to determine the current health status of the SSD. The most relevant monitoring metrics for solid-state drive (SSD) failure early warning include read error rate, reallocation sector count, available reserved block count, wear leveling count, reallocation event count, and uncorrectable error count. The long-sequence solid-state drive fusion information based on the training set is obtained through the following process: The training set data is learned by using a multi-instance learning convolutional neural network supervised learning algorithm to obtain long-sequence solid-state drive fusion information based on the training set. 2.The enterprise-level SSD fault early warning method based on multi-instance adversarial learning of claim 1, wherein, Specifically, the following steps are included: 1) Filter the most relevant monitoring indicators for SSD fault early warning from the SMART parameter set of SSD operating data; 2) The most relevant monitoring indicators for solid-state drive (SSD) failure early warning are divided into instances to obtain the results within the same SSD's lifespan. Each instance sample will Each instance sample is packaged into a collection, known as an instance bag. j is the count of the instance bags, which represents the number of solid-state drive instance bags. Solid State Drive Label Construct a training set; based on the training set, obtain long-sequence solid-state drive fusion information; 3) Based on the adversarial transfer model of solid-state drive feature knowledge and the fusion information of long-sequence solid-state drives, an adaptive fault warning model suitable for different individual solid-state drives is obtained. Based on the adaptive fault warning model suitable for different individual solid-state drives, the known solid-state drive fault feature knowledge is transferred to the unknown solid-state drive data to determine the health status of the hard drive at the current moment and obtain the fault warning of a single solid-state drive. 3.The enterprise-level SSD fault early-warning method based on multi-instance adversarial learning of claim 2, characterized in that, Solid-state drive (SSD) failure types include data saving errors, write interference errors and read interference errors, and P / E cycle errors.

4. The enterprise-level solid state disk fault early warning method based on multi-instance adversarial learning according to claim 2, characterized in that, The long-sequence solid-state drive fusion information based on the training set is calculated using the following formula: wherein , 0 represents that the prediction of the solid state drive is about to fail and needs replacement operation, 1 represents that the solid state drive is in a healthy state, X represents the entire instance bag of the solid state drive life data, K represents the number of instance bags, i.e., the number of solid state drives, i represents the instance count, x i represents the i-th instance.

5. The enterprise-level solid-state disk fault early warning method based on multi-instance adversarial learning according to claim 2, characterized in that, The adversarial transfer model of solid-state drive feature knowledge is obtained through the following process: the probability distribution of known solid-state drive fault feature datasets and unknown solid-state drive fault feature datasets are aligned using an adversarial domain adaptation network to establish the adversarial transfer model of solid-state drive feature knowledge.

6. The enterprise-level solid-state disk fault early warning method based on multi-instance adversarial learning according to claim 5, characterized in that, Solid state drive feature knowledge adversarial migration model overall loss function As follows: wherein are feature extractor parameters, are label classifier parameters, are initial inter-domain trainer parameters, is a feature extractor loss, is a label classifier loss, is a domain adversarial loss, is a gradient reversal function.

7. The enterprise-level solid state disk fault early warning method based on multi-instance adversarial learning according to claim 2, characterized in that, The threshold method is used to determine the current health status of the hard drive.

Images