A method for constructing a solid state drive status prediction model

Abstract

The present invention relates to the field of solid-state drive prediction technology, and in particular to a method for constructing a solid-state drive state prediction model. The method utilizes real sample data and simulated fault sample data for training. The sample data includes log data, image data, and basic information of the solid-state drive. This basic information, combined with time series data, can provide more contextual information for the model, thereby improving the accuracy and reliability of solid-state drive fault prediction.

Description

Technical Field

[0001] The present invention relates to the field of solid-state hard disk prediction technology, and in particular to a method for constructing a solid-state hard disk state prediction model. Background Art

[0002] With the rapid development of cloud storage technology and data centers, solid-state drives (SSDs) have gradually replaced traditional mechanical hard disk drives (HDDs) and become a core component of modern storage systems. SSDs, with their high performance, low power consumption, vibration resistance, and fast response time, are widely used in cloud storage, big data processing, and high-performance computing. However, while SSDs offer significant performance advantages over HDDs, their failure modes are complex, and their failure rates show a distinct upward trend with age. This makes SSD failure prediction a critical issue in modern storage systems. Patent document (CN113778766A) discloses a method for building a hard drive failure prediction model based on multidimensional features and its application. This document uses the hard drive's SMART information, firmware version information, and event log information from the device where the hard drive resides as feature data for hard drive failure prediction. Within the feature data, each numerical data item is cumulatively valued, and the model is trained and tested based on the temporal nature of the feature data. While this multi-dimensional feature information ensures model training effectiveness to a certain extent, it does not consider the impact of physical characteristics of the hard drive on its status, such as model, vendor, wear level, and usage time. These physical characteristics are crucial for determining whether an SSD is in a faulty state. For example, the age and P / E (Program / Erase) count of a hard drive are important factors in determining SSD health. Especially in high-load environments, the risk of hard drive failure may increase as these physical characteristics change. Furthermore, this document uses time-per-hour sampling for positive samples and expands the number of negative samples. However, this approach often leads to overfitting, data loss, or the omission of important features. Therefore, it is necessary to provide a fault prediction solution that can improve the accuracy and reliability of solid state drive fault prediction. Summary of the Invention

[0003] In view of the above technical problems, the technical solution adopted by the present invention is:

[0004] According to a first aspect of the present invention, a method for constructing a solid state drive state prediction model is provided, the method comprising the following steps:

[0005] S100, obtaining an initial feature data set; wherein each initial feature data set includes a status label and feature data of the corresponding solid-state drive within a set time period, and the feature data includes time series data, image data, and basic information; the time series data is a time series formed by the numerical values ​​of key parameters in the SMART log data, the image data is image data obtained based on the time series data, the basic information is used to characterize the physical properties of the solid-state drive, and the status label includes a first label characterizing that the solid-state drive is in a normal state and a second label characterizing that the solid-state drive is in a faulty state.

[0006] S200 , taking feature data with a state label of the first label in the initial feature data set as data to be processed, and generating corresponding adversarial sample data based on the data to be processed.

[0007] S300 , adding the generated adversarial sample data to the initial feature data set as a sample data set, and dividing the sample data set into a training set and a test set.

[0008] S300, taking the feature data in the sample data set as input information and the state label as label information, uses the training set to train the initial solid-state hard disk state prediction model to obtain the trained solid-state hard disk state prediction model, and uses the test set to test the trained solid-state hard disk state prediction model.

[0009] The present invention has at least the following beneficial effects:

[0010] The solid-state drive status prediction model construction method provided by an embodiment of the present invention uses real sample data and simulated fault sample data for training. The sample data includes log data, image data and basic information of the solid-state drive. These basic information are combined with time series data to provide more contextual information for the model, thereby improving the accuracy and reliability of solid-state drive fault prediction.

[0011] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present invention, nor is it intended to limit the scope of the present invention. Other features of the present invention will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0013] Figure 1This is a flowchart of a method for constructing a solid-state drive status prediction model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0015] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used herein in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0016] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of the steps can be performed in parallel, concurrently, or simultaneously. In addition, the order of the steps can be rearranged. A process can be terminated when its operation is completed, but can also have additional steps not included in the accompanying drawings. A process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0017] The embodiment of the present invention provides a method for constructing a solid state hard disk state prediction model. Figure 1 As shown, the method may include the following steps:

[0018] S100, obtaining an initial feature dataset.

[0019] In an embodiment of the present invention, each initial feature data set includes the status label and feature data of the corresponding solid-state drive within a set time period, that is, an initial feature data set is the data of a solid-state drive within a set time period. The feature data includes time series data, image data, and basic information. The time series data is a time series formed by the values ​​of key parameters in the SMART log data.

[0020] In an embodiment of the present invention, the time series data of each solid-state drive may be data obtained by smoothing the SMART log data monitored during a set time period. Those skilled in the art will appreciate that any smoothing of the SMART log data monitored during a set time period to obtain the corresponding time series data falls within the scope of the present invention.

[0021] In an embodiment of the present invention, the length of the set time period can be set based on actual needs, such as several months. Time series data is data after normalization of the original data. The information recorded in the SMART log data includes: disk_id, date, model, label, smart_n. Among them, disk_id is the unique identification number / serial number of the hard disk, date is the generation date of the log data, model is the model of the hard disk, label is a label, indicating the status of the hard disk, and the value is 0 or 1. 0 indicates that the hard disk is running healthily, that is, in normal state, and 1 indicates a hard disk failure. The n in smart_n is a numerical value used to represent different SMART indicators. In an illustrative embodiment, the SMART indicators can be shown in Table 1 below:

[0022] Table 1

[0023]

[0024] In an embodiment of the present invention, the image data is obtained based on time series data. The image data can be obtained using existing methods for converting time series data into images. For example, the Gram angular field method is used to convert time series data in SMART log data into an image representation. The Gram angular field method calculates the angular representation of the time series and maps the numerical value at each time point to the pixel value of the image, effectively capturing long-term trends and patterns in the time series. This conversion not only preserves the changing trends of the time series data but also enables the extraction of potential fault characteristics through image processing techniques.

[0025] In an embodiment of the present invention, the basic information is used to characterize the physical characteristics of the solid-state drive. The basic information may include the hard drive model, supplier, wear level, and power-on time. Traditional methods mostly rely on numerical data in the SMART log, ignoring other key information of the hard drive (such as hard drive model, supplier, usage time, etc.), which are also crucial for SSD failure prediction. The present invention converts the basic information of the hard drive (such as hard drive model, supplier, wear level, power-on time, etc.) into text features. After standardizing these metadata, they are input into the model to ensure that these key information are effectively integrated into the prediction model. The textualized basic information of the hard drive combined with other numerical features can provide more contextual information for the model.

[0026] In an embodiment of the present invention, the status tag includes a first tag indicating that the solid state drive is in a normal state and a second tag indicating that the solid state drive is in a faulty state. The first tag can be represented by 0, and the second tag can be represented by 1.

[0027] Furthermore, in an embodiment of the present invention, the key parameters can be obtained by the following steps:

[0028] S1, based on an original data set, the original data set includes the status label of the solid-state drive and SMART log data within a set time period, using N similarity calculation methods to obtain the similarity between Z parameters in the SMART log data and the status of the solid-state drive, to obtain Z similarity sets; N ≥ 2.

[0029] In an embodiment of the present invention, similarity calculation methods may include, for example, a Pearson correlation coefficient calculation method, a Spearman correlation coefficient calculation method, a J-Index coefficient calculation method, and an XGBoost model. The XGBoost model provides a decision tree-based feature importance calculation method. First, the original data set is used as a training set to train the XGBoost model. Then, the gain value of each parameter is extracted as the similarity between the parameter and the state of the solid-state drive.

[0030] S2, get the first weight W of the u-th parameter among the Z parameters u =∑ N v=1 P uv , get Z first weights; P uv is the position of the uth parameter in the vth similarity set. For example, the position of the uth parameter in the vth similarity set is 3, then P uv = 3. The value of u ranges from 1 to Z, and the value of v ranges from 1 to N.

[0031] S3, sort the Z first weights in descending order, and select n1 parameters corresponding to the first n1 weights of the sorted Z first weights as candidate parameters, where n1 < Z. The candidate parameters can be set based on actual needs and can be obtained by the following steps:

[0032] Step 1: Use the original data corresponding to the current intermediate parameter set to train the current AI model to obtain the current state prediction result of the solid-state drive; the initial value of the current intermediate parameter set is the first parameter among the Z intermediate parameters, the Z intermediate parameters are the Z parameters corresponding to the sorted Z first weights, and the current AI model is the initialized AI model;

[0033] Step 2: Based on the current state prediction result of the solid-state drive and the corresponding actual result, obtain the prediction accuracy of the current AI model and add it to the current accuracy set. If the difference between any two prediction accuracy rates of the last Q prediction accuracy rates in the current accuracy set is less than the set difference, the parameters of the current intermediate parameter set are used as candidate parameters. Otherwise, the intermediate parameters after the last parameter in the current intermediate parameter set among the Z intermediate parameters are added to the current intermediate parameter set, and step 1 is executed.

[0034] In an embodiment of the present invention, the initial value of the current accuracy set is empty. The AI ​​model can be an existing neural network model. The set difference value can be an empirical value, for example, a value infinitely close to 0. In one exemplary embodiment, n1=20.

[0035] S4, using the original data corresponding to the n1 candidate parameters as training data, training the random forest model to obtain the second weights of the n1 candidate parameters.

[0036] S5 , sorting the n1 second weights in descending order, and taking n2 candidate parameters corresponding to the first n2 weights among the sorted n1 second weights as the key parameters, n2<n1.

[0037] In the embodiment of the present invention, n2 may be an empirical value. In an illustrative embodiment, n2=15.

[0038] S200 , taking feature data with a state label of the first label in the initial feature data set as data to be processed, and generating corresponding adversarial sample data based on the data to be processed.

[0039] In S200, adversarial sample data may be generated based on a trained adversarial sample generation model. The data input into the trained adversarial sample generation model includes image data and text fusion data in the data to be processed, where the text fusion data is data obtained by concatenating the time series data and text data in the data to be processed.

[0040] In an embodiment of the present invention, the adversarial sample generation model includes a first feature encoding module, a second feature encoding module, a feature merging module, a generator and a discriminator.

[0041] The first feature encoding module is configured to convert the text data in the received text fusion data into numerical features, thereby obtaining a text fusion data feature matrix, which is then input into the feature merging module. The second feature encoding module is configured to convert the received image data into numerical features, thereby obtaining an image data feature matrix, which is then input into the feature merging module. In one exemplary embodiment, the second feature encoding module may be a Gaussian mixture model.

[0042] The feature merging module is used to multiply the received text fusion data feature matrix and the image data feature matrix to obtain a merged matrix, and input it to the generator.

[0043] The generator is configured to add random noise to the received merged matrix and, based on the merged matrix with the added random noise, generate adversarial sample data, which is then input to the discriminator. In an embodiment of the present invention, the random noise is sampled from a standard normal distribution. The random noise vector is gradually converted into a simulated fault image through multiple layers of transposed convolution operations. The generated fault image is not only random but also influenced by the fault patterns learned during training, thereby more closely resembling the distribution of real fault images.

[0044] In an embodiment of the present invention, the generator may include a noise adding module, a fully connected layer, a first deconvolution layer, a second deconvolution layer and a third deconvolution layer connected in sequence.

[0045] The discriminator is used to determine the authenticity of the received adversarial sample data and obtain a corresponding discriminant value. The discriminator's goal is to identify whether the input data is from a real fault sample and provide feedback to the generator to help it continuously improve the authenticity of the generated samples. If it can determine that the real sample is real data, it will output 1; if it can determine that the generated sample is generated, it will output 0.

[0046] In an embodiment of the present invention, the discriminator includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer and a fully connected layer connected in sequence.

[0047] Furthermore, during the training process of the adversarial sample generation model, when the initial score between the adversarial sample data and the real sample data is greater than the set score threshold, and the initial distance between the adversarial sample data and the real sample data is less than the set distance threshold, the model training ends.

[0048] S300 , adding the generated adversarial sample data to the initial feature data set as a sample data set, and dividing the sample data set into a training set and a test set.

[0049] S300, taking the feature data in the sample data set as input information and the state label as label information, uses the training set to train the initial solid-state hard disk state prediction model to obtain the trained solid-state hard disk state prediction model, and uses the test set to test the trained solid-state hard disk state prediction model.

[0050] Those skilled in the art know that any method of using the training set to train the initial solid-state drive state prediction model to obtain the trained solid-state drive state prediction model, and any method of using the test set to test the trained solid-state drive state prediction model falls within the scope of protection of the present invention.

[0051] In this embodiment of the present invention, a focus loss function can be used as the loss function for the SSD status prediction model. Focus loss is used to focus on difficult-to-classify fault samples, reducing false positives and improving the recall rate of fault predictions. Furthermore, the hyperparameters of the SSD status prediction model can be tuned using methods such as grid search or Bayesian optimization to ensure that the model is trained under the optimal configuration, thereby improving model performance.

[0052] In an embodiment of the present invention, a solid-state drive (SSD) state prediction model includes a time series data feature encoding module, an image encoder, a text encoder, a feature fusion module, and a prediction output module. The time series data feature encoding module, the image encoder, and the text encoder are respectively connected to the feature fusion module, which is connected to the prediction output module. The time series data feature encoding module is used to perform feature encoding on the time series data to obtain a corresponding data feature vector as a first modal feature vector; the image encoder is used to perform feature encoding on the image data to obtain a corresponding image feature vector as a second modal feature vector; the text encoder is used to perform feature encoding on the basic information to obtain a corresponding text feature vector as a third modal feature vector; the feature fusion module is used to fuse the received first modal feature vector, the second modal feature vector, and the third modal feature vector to obtain a modal fusion feature vector, and input the result to the prediction output module. The prediction output module is used to perform linear transformation on the received modal fusion feature vector to obtain a corresponding linear transformation value, and then perform mapping processing on the linear transformation value to obtain a SSD state prediction value. The state prediction value is used to indicate whether the SSD is likely to fail.

[0053] Furthermore, the feature fusion module is specifically configured to perform the following operations:

[0054] S10, perform dot products on the received i-th modal eigenvector with the first transformation weight matrix, the second transformation weight matrix and the third transformation weight matrix corresponding to the i-th modal eigenvector, respectively, to obtain the first transformation eigenvector, the second transformation eigenvector and the third transformation eigenvector corresponding to the i-th modal eigenvector; the value of i is 1 to 3; the first transformation weight matrix, the second transformation weight matrix and the third transformation weight matrix are all learnable weight matrices, that is, they will be updated based on the loss of the solid-state hard disk state prediction model during the training process of the solid-state hard disk state prediction model.

[0055] S20, dot product the first transformed eigenvector corresponding to the i-th modal eigenvector with the second transformed eigenvector of the j-th modal eigenvector, and obtain the first correlation coefficient vector R1 corresponding to the i-th modal eigenvector i =(R1 i1 , R1 i2 , R1 i3 ); R1ir R i The correlation coefficient of the rth dimension in , where r ranges from 1 to 3.

[0056] S30, based on R1 i , get the second correlation coefficient vector R2 of the i-th modal eigenvector i =(R2 i1 , R2 i2 , R2 i3 ); R2 ir =R1 ir / d r , d r is the dimension of the second transformed eigenvector corresponding to the rth modal eigenvector.

[0057] S40, based on R2 i , get the third correlation coefficient vector R3 of the i-th modal eigenvector i =(R3 i1 , R3 i2 , R3 i3 ); R3 ir =exp(R2 ir ) / ∑ 3 h=1 exp(R2 ih ), R2 ih R2 i The hth correlation coefficient in , where h ranges from 1 to 3, and exp() is an exponential function.

[0058] S50, obtaining the output eigenvector F of the i-th modal eigenvector i =(F i1 , F i2 , F i3 ), F ir F i The rth output feature in F ir =R3 i1 ×F3 1r +R3 i2 ×F3 2r +R3 i3 ×F3 3r , F3 sr is the rth feature in the third transformed eigenvector corresponding to the sth modal eigenvector.

[0059] S60, add the i-th modal eigenvector and the corresponding output eigenvector to obtain the final eigenvector of the i-th modal eigenvector, which is used as the i-th modal final eigenvector; obtain the fusion feature matrix M, the size of M is q×3, the r-th row data of M is the i-th modal final eigenvector, and q is the dimension size of the final eigenvector of each modality.

[0060] S70, based on M, obtain the modal fusion feature vector FM = (FM1, FM2, ..., FM x ,……,FM q ), FM x is the xth feature in FM, x ranges from 1 to q, FM x is the average value of the x-th column data in M.

[0061] In an embodiment of the present invention, the prediction output module may be a fully connected layer. Specifically, the state prediction value satisfies the following conditions:

[0062] S=(1+e -y ) -1 ; S is the state prediction value, e is a natural number, y is the linear conversion value, y=FM×W+b, W is the learnable weight matrix, and b is the bias term.

[0063] The solid-state drive status prediction model construction method provided by an embodiment of the present invention uses real sample data and simulated fault sample data for training. The sample data includes log data, image data and basic information of the solid-state drive. These basic information are combined with time series data to provide more contextual information for the model, thereby improving the accuracy and reliability of solid-state drive fault prediction.

[0064] Another embodiment of the present invention provides a method for predicting a state of a solid-state drive, the method comprising:

[0065] Get the current characteristic data of the solid-state drive.

[0066] The acquired characteristic data is input into the solid-state drive state prediction model constructed by the aforementioned method to predict the current state of the solid-state drive.

[0067] An embodiment of the present invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are configured to execute the method described in the embodiment of the present invention.

[0068] An embodiment of the present invention further provides a computer-readable storage medium storing computer-executable instructions, wherein the computer instructions are used to execute the method described in the embodiment of the present invention.

[0069] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present invention can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in the present invention can be achieved. This is not limited herein.

[0070] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A method for constructing a solid state drive state prediction model, characterized in that: The method comprises the following steps: S100, obtaining an initial feature data set; wherein each initial feature data set includes a status label and feature data of the corresponding solid-state drive within a set time period, the feature data including time series data, image data, and basic information; the time series data is a time series formed by the values ​​of key parameters in the SMART log data, the image data is image data obtained based on the time series data, the basic information is used to characterize the physical properties of the solid-state drive, and the status label includes a first label indicating that the solid-state drive is in a normal state and a second label indicating that the solid-state drive is in a faulty state; S200, taking feature data with a state label of the first label in the initial feature data set as data to be processed, and generating corresponding adversarial sample data based on the data to be processed; S300, adding the generated adversarial sample data to the initial feature dataset as a sample dataset, and dividing the sample dataset into a training set and a test set; S300, using the feature data in the sample data set as input information and the state label as label information, training an initial solid-state drive state prediction model using the training set to obtain a trained solid-state drive state prediction model, and testing the trained solid-state drive state prediction model using the test set; In S200, adversarial sample data is generated based on the trained adversarial sample generation model, wherein the data input into the trained adversarial sample generation model includes image data and text fusion data in the data to be processed, the text fusion data being data obtained by concatenating time series data and text data in the data to be processed; the text data being basic information in text form; Among them, the adversarial sample generation model includes a first feature encoding module, a second feature encoding module, a feature merging module, a generator and a discriminator, wherein the first feature encoding module is used to convert the text data in the received text fusion data into numerical form features, obtain the text fusion data feature matrix, and input it to the feature merging module; the second feature encoding module is used to convert the received image data into numerical form features, obtain the image data feature matrix, and input it to the feature merging module; the feature merging module is used to multiply the received text fusion data feature matrix and the image data feature matrix to obtain a merged matrix, and input it to the generator; the generator is used to add random noise to the received merged matrix, and generate adversarial sample data based on the merged matrix with added random noise, and input it to the discriminator; the discriminator is used to judge the authenticity of the received adversarial sample data and obtain the corresponding discriminant value.

2. The method according to claim 1, characterized in that The solid-state hard disk state prediction model includes a time series data feature encoding module, an image encoder, a text encoder, a feature fusion module and a prediction output module, wherein the time series data feature encoding module, the image encoder and the text encoder are respectively connected to the feature fusion module, and the feature fusion module is connected to the prediction output module. The time series data feature encoding module is used to feature encode the time series data to obtain the corresponding data feature vector as the first modal feature vector; the image encoder is used to feature encode the image data to obtain the corresponding image feature vector as the second modal feature vector; the text encoder is used to feature encode the basic information to obtain the corresponding text feature vector as the third modal feature vector; the feature fusion module is used to fuse the received first modal feature vector, the second modal feature vector and the third modal feature vector to obtain a modal fusion feature vector, and input it to the prediction output module. The prediction output module is used to perform linear conversion on the received modal fusion feature vector to obtain the corresponding linear conversion value, and map the linear conversion value to obtain the state prediction value of the solid-state hard disk.

3. The method according to claim 2, characterized in that The feature fusion module is specifically used to perform the following operations: S10, performing dot products on the received i-th modal eigenvector with the first transformation weight matrix, the second transformation weight matrix, and the third transformation weight matrix corresponding to the i-th modal eigenvector, respectively, to obtain the first transformation eigenvector, the second transformation eigenvector, and the third transformation eigenvector corresponding to the i-th modal eigenvector; the value of i ranges from 1 to 3; the first transformation weight matrix, the second transformation weight matrix, and the third transformation weight matrix are all learnable weight matrices; S20, dot product the first transformed eigenvector corresponding to the i-th modal eigenvector with the second transformed eigenvector of the j-th modal eigenvector, and obtain the first correlation coefficient vector R1 corresponding to the i-th modal eigenvector i =(R1 i1 , R1 i2 , R1 i3 ); R1 ir is the correlation coefficient between the eigenvector of the i-th modal and the eigenvector of the r-th modal, and the value of r ranges from 1 to 3; S30, based on R1 i , get the second correlation coefficient vector R2 of the i-th modal eigenvector i =(R2 i1 , R2 i2 , R2 i3 ); R2 ir =R1 ir / d r , d r is the dimension of the second transformed eigenvector corresponding to the rth modal eigenvector; S40, based on R2 i , get the third correlation coefficient vector R3 of the i-th modal eigenvector i =(R3 i1 , R3 i2 , R3 i3 ); R3 ir =exp(R2 ir ) / ∑ 3 h=1 exp(R2 ih ), R2 ih R2 i The hth correlation coefficient in , where h ranges from 1 to 3, and exp() is an exponential function; S50, obtaining the output eigenvector F of the i-th modal eigenvector i =(F i1 , F i2 , F i3 ), F ir F i The rth output feature in F ir =R3 i1 ×F3 1r +R3 i2 ×F3 2r +R3 i3 ×F3 3r , F3 sr is the rth feature in the third transformed eigenvector corresponding to the sth modal eigenvector; S60, adding the ith modal eigenvector and the corresponding output eigenvector to obtain the final eigenvector of the ith modal eigenvector as the ith modal final eigenvector; obtaining The fusion feature matrix M has a size of q×3, the rth row of M is the final feature vector of the i-th mode, and q is the dimension of the final feature vector of each mode; S70, based on M, obtain the modal fusion feature vector FM = (FM1, FM2, ..., FM x ,……,FM q ), FM x is the xth feature in FM, x ranges from 1 to q, FM x is the average value of the x-th column data in M.

4. The method according to claim 2, characterized in that The state prediction value meets the following conditions: S=(1+e -y ) -1 ; S is the state prediction value, e is a natural number, y is the linear conversion value, y=FM×W+b, W is the learnable weight matrix, and b is the bias term.

5. The method according to claim 1, wherein The generator includes a noise addition module, a fully connected layer, a first deconvolution layer, a second deconvolution layer, and a third deconvolution layer connected in sequence, and the discriminator includes a first convolution layer, a second convolution layer, a third convolution layer, a fourth convolution layer, and a fully connected layer connected in sequence; Among them, during the training process of the adversarial sample generation model, when the initial score between the adversarial sample data and the real sample data is greater than the set score threshold, and the initial distance between the adversarial sample data and the real sample data is less than the set distance threshold, the model training ends.

6. The method according to claim 1, characterized in that The key parameters are obtained by the following steps: S1, based on an original data set, the original data set includes the state label of the solid state drive and SMART log data within a set time period, using N similarity calculation methods to obtain the similarity between Z parameters in the SMART log data and the state of the solid state drive, to obtain Z similarity sets; N ≥ 2; S2, get the first weight W of the u-th parameter among the Z parameters u =∑ N v=1 P uv , get Z first weights; P uv is the position of the uth parameter in the vth similarity set, where u ranges from 1 to Z and v ranges from 1 to N; S3, sorting the Z first weights in descending order, and taking n1 parameters corresponding to the first n1 weights among the sorted Z first weights as candidate parameters, where n1<Z; S4, using the original data corresponding to the n1 candidate parameters as training data, training the random forest model to obtain the second weights of the n1 candidate parameters; S5 , sorting the n1 second weights in descending order, and taking n2 candidate parameters corresponding to the first n2 weights among the sorted n1 second weights as the key parameters, n2<n1.

7. The method according to claim 6, characterized in that Candidate parameters are obtained through the following steps: Step 1: Use the original data corresponding to the current intermediate parameter set to train the current AI model to obtain the current state prediction result of the solid-state drive; the initial value of the current intermediate parameter set is the first parameter among the Z intermediate parameters, the Z intermediate parameters are the Z parameters corresponding to the sorted Z first weights, and the current AI model is the initialized AI model; Step 2: Based on the current state prediction result of the solid-state drive and the corresponding actual result, obtain the prediction accuracy of the current AI model and add it to the current accuracy set. If the difference between any two prediction accuracy rates of the last Q prediction accuracy rates in the current accuracy set is less than the set difference, the parameters of the current intermediate parameter set are used as candidate parameters. Otherwise, the intermediate parameters after the last parameter in the current intermediate parameter set among the Z intermediate parameters are added to the current intermediate parameter set, and step 1 is executed.

8. The method according to claim 1, characterized in that The basic information includes hard disk model, vendor, wear level, and boot time.

Images