A memory fault prediction method and device

Abstract

The present disclosure discloses a memory fault prediction method and device, which relates to the field of artificial intelligence technology. The method includes obtaining operating status data; the operating status data is used to indicate: n types of attributes of the memory of a target device within a first time period; performing feature extraction on the operating status data to obtain state feature data corresponding to the operating status data; using the state feature data, determining a first prediction result; the first prediction result is used to indicate: the number of correctable errors that occur in the memory within a second time period; and using the state feature data, determining a second prediction result; the second prediction result is used to indicate: the probability of uncorrectable errors that occur in the memory within a third time period. The present disclosure achieves the technical effect of accurately predicting the number of correctable errors that occur in the memory and the probability of uncorrectable errors that occur in the memory.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to a memory fault prediction method and device. Background Art

[0002] Currently, in servers and other devices, as memory frequency continues to increase and particle density continues to increase, the load on memory during operation is also increasing, which increases the possibility of memory failure.

[0003] Memory failures are primarily categorized into two types: correctable errors (CE) and uncorrectable errors (UCE). CEs are errors that can be automatically corrected by hardware or software mechanisms, such as single-bit errors. These errors typically do not immediately cause a system crash; error-correcting code (ECC) technology enables automatic detection and correction. UCEs, on the other hand, are errors that cannot be corrected by existing mechanisms and can easily cause system downtime or reboots, such as storage unit failures and multi-bit errors. Memory UCE errors are one of the main causes of system downtime.

[0004] Among them, if the number of correctable memory errors and the probability of uncorrectable memory errors in the future can be predicted in advance, the impact of memory failures on system operations can be reduced by isolating faulty memory in advance.

[0005] Therefore, how to accurately predict the number of correctable errors that occur in memory and the probability of uncorrectable errors that occur in memory is a problem that needs to be solved at present. Summary of the Invention

[0006] The present disclosure provides a memory fault prediction method and apparatus to at least solve the problem in the related art that the number of correctable errors occurring in the memory and the probability of uncorrectable errors occurring in the memory cannot be accurately predicted.

[0007] The present disclosure provides a memory fault prediction method, including: obtaining operating status data; the operating status data is used to indicate: n properties of the memory of the target device within a first time period; n is a positive integer. Feature extraction is performed on the operating status data through the shared layer of the dual-task neural network model to obtain state feature data corresponding to the operating status data. A first prediction result is determined by using the state feature data through the first task layer of the dual-task neural network model; the first prediction result is used to indicate: the number of correctable errors that occur in the memory within a second time period; the second time period is after the first time period. A second prediction result is determined by using the state feature data through the second task layer of the dual-task neural network model; the second prediction result is used to indicate: the probability of uncorrectable errors occurring in the memory within a third time period; the third time period is after the first time period.

[0008] The present disclosure also provides a model training method, comprising: obtaining x training samples and labels corresponding to the x training samples; wherein the x training samples include n attributes of a target device's memory within x different time periods, where n is a positive integer; and the labels corresponding to each training sample include the number of correctable errors occurring in the memory within a first preset time period after the time period corresponding to the training sample, and whether an uncorrectable error occurred in the memory within a second preset time period after the time period corresponding to the training sample. A dual-task neural network model is trained for supervised learning using the x training samples and the labels corresponding to the x training samples. The dual-task neural network model comprises a shared layer, a first task layer, and a second task layer; wherein the shared layer is configured to extract features from the n attributes of the memory within the first time period to obtain state feature data; the first task layer is configured to determine a first prediction result using the state feature data; and the second task layer is configured to determine a second prediction result using the state feature data; wherein the first prediction result indicates the number of correctable errors occurring in the memory within the second time period; and the second prediction result indicates the probability of an uncorrectable error occurring in the memory within a third time period; and the second and third time periods are subsequent to the first time period.

[0009] The present disclosure also provides a memory fault prediction device, including: an acquisition unit, used to acquire operating status data; the operating status data is used to indicate: n properties of the memory of the target device within a first time period; n is a positive integer; a processing unit, used to extract features of the operating status data through a shared layer of a dual-task neural network model, and obtain state feature data corresponding to the operating status data; the processing unit is also used to determine a first prediction result using the state feature data through the first task layer of the dual-task neural network model; the first prediction result is used to indicate: the number of correctable errors that occur in the memory within a second time period; the second time period is after the first time period; the processing unit is also used to determine a second prediction result using the state feature data through the second task layer of the dual-task neural network model; the second prediction result is used to indicate: the probability of an uncorrectable error occurring in the memory within a third time period; the third time period is after the first time period.

[0010] The present disclosure also provides a model training device, comprising: an acquisition unit configured to acquire x training samples and labels corresponding to the x training samples; wherein the x training samples include n attributes of a target device's memory within x different time periods, where n is a positive integer; and the labels corresponding to each training sample include the number of correctable errors occurring in the memory within a first preset time period following the time period corresponding to the training sample, and whether an uncorrectable error occurred in the memory within a second preset time period following the time period corresponding to the training sample. A training unit configured to perform supervised learning on a dual-task neural network model using the x training samples and the labels corresponding to the x training samples. The dual-task neural network model comprises: a shared layer, a first task layer, and a second task layer; wherein the shared layer is configured to extract features from the n attributes of the memory within the first time period to obtain state feature data; the first task layer is configured to determine a first prediction result using the state feature data; and the second task layer is configured to determine a second prediction result using the state feature data; wherein the first prediction result indicates the number of correctable errors occurring in the memory within the second time period; and the second prediction result indicates the probability of an uncorrectable error occurring in the memory within a third time period; and the second and third time periods are subsequent to the first time period.

[0011] The present disclosure also provides an electronic device, comprising: a memory for storing a computer program; a processor for implementing the steps of any one of the above-mentioned memory fault prediction methods when executing the computer program, or a processor for implementing the steps of any one of the above-mentioned model training methods when executing the computer program.

[0012] The present disclosure also provides a computer-readable storage medium, in which a computer program is stored, wherein when the computer program is executed by a processor, the steps of any one of the above-mentioned memory fault prediction methods are implemented, or when the computer program is executed by a processor, the steps of any one of the above-mentioned model training methods are implemented.

[0013] The present disclosure also provides a computer program product, including a computer program, which implements the steps of any of the above-mentioned memory fault prediction methods when the computer program is executed by a processor, or implements the steps of any of the above-mentioned model training methods when the computer program is executed by a processor.

[0014] In the method disclosed above, a dual-task neural network model is first constructed. This dual-task neural network model uses n attributes (referred to as operating status data) of the target device's memory during a previous period (referred to as the first period) to predict the number of correctable errors occurring in the memory during a subsequent period (referred to as the second period), obtaining a first predetermined result. It also predicts the probability of uncorrectable errors occurring in the memory during a subsequent period (referred to as the third period), obtaining a second predicted result. Considering the correlation between the number of CE errors and the probability of UCE errors, the feature information used to predict the number of CE errors and the feature information used to predict the probability of UCE errors are also correlated. Therefore, in this dual-task neural network model, a shared layer can first be used to extract features from the operating status data of the memory during the first period to obtain state feature data. Then, in the first task layer, this state feature data is used to predict the number of CE errors in the second period (i.e., determine a first predicted result). In the second task layer, this state feature data is used to predict the probability of UCE errors in the third period (i.e., determine a second predicted result). This effectively leverages the correlation between the number of CE errors and the probability of UCE errors, resulting in more accurate predictions of the number of CE errors and the probability of UCE errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present disclosure, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0016] Figure 1 This is one of the structural diagrams of a dual-task neural network model provided by an embodiment of the present disclosure;

[0017] Figure 2 The second structural diagram of a dual-task neural network model provided by an embodiment of the present disclosure;

[0018] Figure 3 One of the flow charts of a model training method provided in an embodiment of the present disclosure;

[0019] Figure 4 A second flow chart of a model training method provided in an embodiment of the present disclosure;

[0020] Figure 5 The third flowchart of a model training method provided in an embodiment of the present disclosure;

[0021] Figure 6 This is a flowchart of a memory fault prediction method provided by an embodiment of the present disclosure;

[0022] Figure 7 A second flowchart of a memory fault prediction method provided by an embodiment of the present disclosure;

[0023] Figure 8 A third flow chart of a memory fault prediction method provided in an embodiment of the present disclosure;

[0024] Figure 9 A regression prediction scatter plot for predicting the number of CE occurrences provided by an embodiment of the present disclosure;

[0025] Figure 10 A schematic diagram of the statistical results of MSE and MAE values ​​under three types of algorithms provided in an embodiment of the present disclosure;

[0026] Figure 11 Schematic diagram of the statistical results of precision, recall and F1 score under three algorithm types provided in the embodiment of the present disclosure;

[0027] Figure 12 A schematic diagram of the structure of a memory fault prediction device provided by an embodiment of the present disclosure;

[0028] Figure 13 A schematic diagram of the structure of a model training device provided in an embodiment of the present disclosure;

[0029] Figure 14 A schematic structural diagram of an electronic device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present disclosure in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of them. Based on the embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present disclosure.

[0031] It should be noted that, in the description of this disclosure, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. The terms "first," "second," and the like in this disclosure are used to distinguish similar objects, and are not used to describe a particular order or precedence.

[0032] In order to enable those skilled in the art to better understand the present disclosure, the present disclosure is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0033] In order to accurately predict the number of correctable errors and the probability of uncorrectable errors in memory, the embodiment of the present disclosure provides a technical solution. In this technical solution, a dual-task neural network model can be constructed. Figure 1 As shown, the dual-task neural network model 10 can use n properties (called operating status data) of the memory of the target device in a previous period of time (called the first time period) to predict the number of correctable errors that occur in the memory in a subsequent period of time (called the second time period) to obtain a first preset result, and predict the probability of uncorrectable errors that occur in the memory in a subsequent period of time (called the third time period) to obtain a second prediction result.

[0034] Among them, considering that there is a correlation between the number of CE occurrences in memory and the probability of UCE occurrence, the feature information used to predict the number of CE occurrences is also correlated with the feature information used to predict the probability of UCE occurrence. Figure 1 As shown, in this dual-task neural network model 10, shared layer 101 can first extract features from the operational status data stored in a first time period to obtain status feature data. This status feature data is then used in first task layer 102 to predict the number of CE occurrences in a second time period (i.e., determine a first prediction result), and in second task layer 103, this status feature data is used to predict the probability of UCE occurrence in a third time period (i.e., determine a second prediction result). This fully exploits the correlation between the number of CE occurrences and the probability of UCE occurrence, resulting in more accurate predictions of the number of CE occurrences and the probability of UCE occurrence.

[0035] In some implementations, such as Figure 2As shown, the shared layer 101 may include: a feature extraction layer 1011 constructed using a long short-term memory network (LSTM) and a shared feature layer 1012 constructed using a fully connected network (FCN).

[0036] The feature extraction layer 1011 is used to extract features from the running state data input into the dual-task neural network model 10. The shared feature layer 1012 is used to fuse the feature data output by the feature extraction layer 1011 to obtain state feature data.

[0037] In some implementations, the first task layer may be a machine learning model constructed using a fully connected network, and the second task layer may be a machine learning model constructed using a probabilistic neural network.

[0038] The following introduces the model training method of the dual-task neural network model 10 provided in the embodiment of the present application with reference to examples. Specifically, the execution subject of the model training method can be a model training device. When the model training device is running, it can be used to execute all or part of the steps in the model training method provided in the embodiment of the present application. In particular, in actual application, the functions of the model training device can be implemented by a personal computer (including desktop computers, laptop computers, handheld computers and notebook computers, etc.), or electronic devices such as smart phones and servers; or, the functions of the above-mentioned model training device can also be implemented by some hardware / software devices in the above-mentioned electronic devices. The embodiment of the present application does not impose any special restrictions on the specific form of the model training device.

[0039] like Figure 3 As shown, the model training method provided in the embodiment of the present application may include:

[0040] S201: The model training device obtains x training samples and labels corresponding to the x training samples.

[0041] The x training samples include: n attributes of the target device's memory in x different time periods, where n is a positive integer.

[0042] In some implementations, the n attributes may specifically include one or more of the following: operating voltage, ambient temperature, operating frequency, power, data refresh rate, read and write delays, the number of row failures that occur in the memory in the current time period, the number of column failures that occur in the memory in the current time period, the number of correctable errors that occur in the memory in the current time period, and the number of uncorrectable errors that occur in the memory in the current time period.

[0043] For example, taking a time period of 10 hours, the x training samples may include: training sample s1, training sample s2, training sample s3, ..., training sample sx. Training sample s1 includes: n attributes of memory detected from time t to time t+10 hours, training sample s2 includes: n attributes of memory detected from time t+10 hours to time t+20 hours, training sample s3 includes: n attributes of memory detected from time t+20 hours to time t+30 hours, and the same applies to other training samples.

[0044] In addition, the label corresponding to each training sample includes: the number of correctable errors that occur in the memory within a first preset time period after the time period corresponding to the training sample.

[0045] For example, continuing with the above example, taking the first preset time period as 1 hour, the label of the above training sample s1 includes: the number of correctable errors that occurred in the memory within 1 hour from time t to time t+10h; the label of the above training sample s2 includes: the number of correctable errors that occurred in the memory within 1 hour from time t+10h to time t+20h, and the labels of other training samples can be deduced in the same way.

[0046] In addition, the label corresponding to each training sample further includes: the probability of an uncorrectable error occurring in the memory within a second preset time period after the time period corresponding to the training sample.

[0047] For example, continuing with the above example, taking the second preset time period as 1 hour, the label of the above training sample s1 includes: the probability of an uncorrectable error occurring in the memory within 1 hour after the period from time t to time t+10h; the label of the above training sample s2 includes: the probability of an uncorrectable error occurring in the memory within 1 hour after the period from time t+10h to time t+20h, and the labels of other training samples can be deduced similarly.

[0048] It is understood that, in actual application, the first preset time period and the second preset time period may be of the same length. For example, in the above example, the first preset time period and the second preset time period are both 1 hour. In addition, the first preset time period and the second preset time period may also have different lengths. The values ​​of the first preset time period and the second preset time period may be determined as needed and are not limited in this regard in the embodiments of the present disclosure.

[0049] S202 , the model training device performs supervised learning training on the dual-task neural network model 10 using the x training samples and the labels corresponding to the x training samples.

[0050] In some implementations, such as Figure 4As shown, S202 may specifically include:

[0051] S2021. The model training device performs data dimensionality reduction on each of the x training samples to obtain x reduced-dimensionality training samples.

[0052] S2022. The model training device performs supervised learning training on the dual-task neural network model using the x dimensionality reduction training samples and the labels corresponding to the x training samples.

[0053] Among them, by performing data dimensionality reduction on x training samples, the amount of data in the training samples can be simplified and the information density of the training samples can be increased, which helps to improve the efficiency of model training.

[0054] Specifically, such as Figure 5 As shown, S2021 may include:

[0055] S20211, the model training device uses x training samples to calculate the correlation coefficient matrix corresponding to n attributes .

[0056] For example, the data included in the x training samples can be decentralized. Specifically, the mean of the values ​​of the y-th attribute in the x training samples can be calculated, and then the mean value of the y-th attribute in each training sample in the x training samples can be subtracted from the mean value to obtain the decentralized data.

[0057] The data set included in the decentralized x training samples can be expressed as the following formula 1:

[0058] Formula 1

[0059] in, represents the data included in x training samples, represents the value of the nth attribute arranged in the order of x training samples. For example, ,in, Indicates time period The value of the nth attribute in the corresponding training sample. For example, Indicates the memory operating voltage set, Indicates memory ambient temperature set, etc.

[0060] Furthermore, the correlation coefficient matrix R corresponding to n attributes can be calculated using the following formula 2:

[0061] Formula 2

[0062] Among them, cov means calculating covariance.

[0063] S20212. The model training device determines n eigenvectors of the correlation coefficient matrix and n eigenvalues ​​corresponding to the n eigenvectors.

[0064] The calculation process of the eigenvectors and eigenvalues ​​of the correlation coefficient matrix R can refer to the contents of the relevant technology and will not be described in detail here.

[0065] Among them, n eigenvalues ​​can be expressed as: .in, .

[0066] The n eigenvectors corresponding to the n eigenvalues ​​can be expressed as .in, .

[0067] After determining n eigenvectors ( ), each of the x training samples can be projected onto the feature space corresponding to the n feature vectors to obtain a new data set. As shown in the following formula 3:

[0068] Formula 3

[0069] in, .

[0070] S20213. The model training device determines p eigenvectors corresponding to p target eigenvalues ​​from n eigenvectors.

[0071] Among them, the p target eigenvalues ​​are: the first p eigenvalues ​​among the n eigenvalues, sorted from large to small in value.

[0072] In some designs, the first p eigenvalues ​​may specifically be: the first p eigenvalues ​​whose cumulative information contribution rate is greater than a contribution rate threshold among the n eigenvalues ​​sorted from largest to smallest in numerical value.

[0073] Specifically, the cumulative information contribution rate It can be calculated by the following formula:

[0074] Formula 4

[0075] Among them, the above-mentioned first p eigenvalues ​​can be specifically: so that the information cumulative contribution rate The first p eigenvalues ​​that are greater than a contribution rate threshold (for example, the contribution rate threshold can be 0.9, 0.95, or 0.99).

[0076] S20214. The model training device projects each of the x training samples to the feature space corresponding to the p feature vectors to obtain x dimensionality-reduced training samples.

[0077] Each dimensionality reduction training sample contains p items of feature data.

[0078] Specifically, the data set included in the x dimensionality reduction training samples can be expressed as the following formula 5:

[0079] Formula 5

[0080] in, Represents the data included in x dimensionality reduction training samples. The calculation process of can refer to the calculation process of formula 3 above. The value of the pth feature data is arranged in the order of x dimensionality reduction training samples. For example, ,in, Indicates time period The value of the p-th feature data in the corresponding training sample.

[0081] In the above implementation, we consider: the information cumulative contribution rate corresponding to the current p eigenvalues When it is close to 1, it indicates that the data set In the example, the feature data corresponding to the first p eigenvalues ​​(i.e. ) already contains most of the feature information of the original data set, and then the feature data corresponding to the np eigenvalues ​​(i.e. ) contains little useful feature information and can be regarded as noise. In this case, we can choose feature data (i.e. ) instead of the original dataset to train the model, thereby achieving data dimensionality reduction and denoising.

[0082] In some implementations, the loss function used to train the dual-task neural network model 10 is ,satisfy:

[0083] Loss Function The first function With the second function The weighted summation result of .

[0084] Among them, the first function It is used to reflect the accuracy of the output results of the first task layer. The second function Used to reflect the accuracy of the output results of the second task layer.

[0085] For example, the loss function It can be expressed as the following formula 6:

[0086] Formula 6

[0087] in, The first function The weight value of For the second function In addition, and Represents the uncertainty parameter of the task, used to ensure the first function and the second function The magnitude of is relatively close. In practical application, and Can be a preset value, or and Can be optimized together with the weights and biases of neurons in a neural network.

[0088] In the above implementation, in order to make the learning process of the model fit both the prediction task of the number of CE occurrences in memory and the prediction task of the probability of UCE occurrence in memory, the first function With the second function The weighted sum of the results is used as the loss function for model training, thus ensuring the training effect of the two tasks.

[0089] In some designs, considering that for the prediction task of the number of CE occurrences in future memory, the mean square error can be used as its loss function. Therefore, the first function above It can be expressed as the following formula 7:

[0090] Formula 7

[0091] Where x is the number of training samples; It indicates the predicted value of the number of CEs in the memory output by the model when the input is the t-th training sample; Indicates the number of times CE occurs in the memory of the label of the t-th training sample.

[0092] In addition, in some designs, considering that for the prediction task of the probability of UCE in future memory, binary cross entropy can be used as its loss function. Therefore, the second function is listed It can be expressed as the following formula:

[0093] Style 8

[0094] Where x is the number of training samples; It represents the predicted value of the probability of UCE occurring in the memory output by the model when the input is the tth training sample; Indicates whether the memory in the label of the t-th training sample has the true situation of UCE.

[0095] In some designs, it is considered that when the model is overly biased towards fitting a certain subtask, the effects of other subtasks are often negatively affected, resulting in a relative decline in their performance. Therefore, in order to prevent the model's learning process from being dominated by a single subtask and thus causing learning bias, the weight values ​​of the first function and the second function in the loss function of this disclosure can be adjusted as follows:

[0096] During the j-th model optimization, the first function Corresponding weight value The second function is positively correlated with the degree of loss reduction of the output result of the first task layer within the preset training process. Corresponding weight value It is positively correlated with the degree of loss reduction of the output result of the second task layer within the preset training process. The preset training process includes the model training process from the j-1th model optimization to the jth model optimization.

[0097] The above design takes into account that: during the model training process from the j-1th model optimization to the jth model optimization, if the loss of a subtask decreases significantly, it means that the subtask has more accurately learned the characteristics and patterns in the data during training. In this case, the weight value of the corresponding function can be increased during the jth model optimization to focus more on the training of this subtask in the subsequent training process. If the rate of decrease of the subtask loss slows down, it means that the learning efficiency of the subtask has decreased. In this case, the weight value of the corresponding function can be reduced during the jth model optimization to focus more on the training of other subtasks in the subsequent training process.

[0098] In some other designs, if the loss of the output result of the first task layer increases during the preset training process, the weight value corresponding to the first function is set to zero during the j-th model optimization. Alternatively, if the loss of the output result of the second task layer increases during the preset training process, the weight value corresponding to the second function is set to zero during the j-th model optimization.

[0099] The above design takes into account that if the loss of a subtask increases, it indicates that the optimization process of the loss function of this subtask may have reached a local minimum. Therefore, by resetting the weight of this subtask to zero, the training step will only rely on the gradients of other subtasks whose losses are still decreasing, which helps the gradient escape the local minimum of this subtask.

[0100] The following example combines the above first function With the second function The weight value adjustment method is described in detail:

[0101] Assume that For function The weight value of , where the function Can be the first function or the second function Any one of the following. Represents the function of the j-th model optimization The weight value of . Represents the function of the j-th model optimization The value of can be determined using the following formula Value:

[0102] Formula 9

[0103] Among them, the auxiliary parameters Satisfies the following formula 10:

[0104] Style 10

[0105] Auxiliary loss function Satisfies the following formula 11:

[0106] Formula 11

[0107] It can be seen that when using the following formula to determine In the case of the value of , on the one hand, when the loss of subtask i increases (i.e. , ),but The value of is 0. On the other hand, when the loss of subtask i is small, The value of is positively correlated with the degree of smaller losses.

[0108] The following introduces a method for memory fault prediction using the dual-task neural network model 10 with reference to an example. Specifically, the execution subject of the memory fault prediction method can be a memory fault prediction device. When the memory fault prediction device is running, it can be used to execute all or part of the steps in the memory fault prediction method provided in the embodiment of the present application. In particular, in actual application, the function of the memory fault prediction device can be implemented by a personal computer (including desktop computers, laptop computers, handheld computers, and notebook computers, etc.), or a smart phone, server, and other electronic devices; or, the function of the above-mentioned memory fault prediction device can also be implemented by some hardware / software devices in the above-mentioned electronic devices. The embodiment of the present application does not impose any special restrictions on the specific form of the memory fault prediction device.

[0109] like Figure 6 As shown, the memory fault prediction method provided in the embodiment of the present application may include:

[0110] S301. A memory fault prediction device obtains operating status data.

[0111] The operating status data is used to indicate n properties of the memory of the target device within a first time period, where n is a positive integer.

[0112] In some implementations, the n attributes may specifically include one or more of the following ten attributes: operating voltage, ambient temperature, operating frequency, power, data refresh rate, read and write delay, the number of row failures that occur in the memory in the current time period, the number of column failures that occur in the memory in the current time period, the number of correctable errors that occur in the memory in the current time period, and the number of uncorrectable errors that occur in the memory in the current time period.

[0113] In some implementations, after obtaining n attributes of the target device's memory within the first time period, data dimensionality reduction can be performed to simplify the data volume, and then the simplified data can be used for prediction, which can improve the accuracy of the prediction results. Figure 7 As shown, S301 may specifically include:

[0114] S3011. The memory fault prediction device obtains n items of status data.

[0115] The n items of status data are respectively used to indicate n properties of the memory of the target device within the first time period.

[0116] For example, the n items of status data include: operating voltage, ambient temperature, operating frequency, power, data refresh rate, read and write delay, the number of row failures that occur in the memory in the current time period, the number of column failures that occur in the memory in the current time period, the number of correctable errors that occur in the memory in the current time period, and the number of uncorrectable errors that occur in the memory in the current time period. These ten items of status data.

[0117] S3012. The memory fault prediction device performs data dimensionality reduction on the n items of status data to obtain operation status data.

[0118] In some implementations, principal component analysis (PCA) can be used to reduce the dimension of n items of state data. Specifically, Figure 8 As shown, S3012 may include:

[0119] S30121. The memory fault prediction device obtains a correlation coefficient matrix corresponding to n types of attributes.

[0120] For example, the correlation coefficient matrix R corresponding to n attributes can be calculated using the above formula 2.

[0121] Formula 2

[0122] S30122. The memory fault prediction device determines n eigenvectors of the correlation coefficient matrix and n eigenvalues ​​corresponding to the n eigenvectors.

[0123] S30123. The memory fault prediction device determines p eigenvectors corresponding to the p target eigenvalues ​​from the n eigenvectors.

[0124] Among them, the p target eigenvalues ​​are: the first p eigenvalues ​​among the n eigenvalues, sorted from large to small in value.

[0125] For example, the p target eigenvalues ​​can be calculated according to the above formula 4 so that the information cumulative contribution rate is The first p eigenvalues ​​that are greater than a contribution rate threshold (for example, the contribution rate threshold can be 0.9, 0.95, or 0.99).

[0126] S30124. The memory fault prediction device projects the n items of status data onto the feature space corresponding to the p feature vectors to obtain the operating status data containing the p items of feature data.

[0127] For example, we can first decentralize n items of state data to obtain a data set :

[0128] .

[0129] in, Indicates the decentralized value of the n-th state data.

[0130] Specifically, the decentralized value of the nth item of status data can be obtained by subtracting the mean of the status data from the original value of the nth item of status data (wherein the mean can be obtained by collecting multiple samples and then averaging the multiple samples).

[0131] In addition, the p eigenvectors include: .in, .

[0132] Then project the n items of status data into the feature space corresponding to p eigenvectors to obtain the running status data ,Right now .

[0133] in, .

[0134] in, That is, the p-item characteristic data in the operating status data.

[0135] S302. The memory fault prediction device extracts features from the operating status data through the shared layer of the dual-task neural network model to obtain status feature data corresponding to the operating status data.

[0136] Specifically, the memory fault prediction device can input the operating status data into the shared layer 101 of the dual-task neural network model 10 to obtain the status feature data corresponding to the operating status data output by the shared layer 101.

[0137] S303. The memory fault prediction device determines a first prediction result by using the state feature data through the first task layer of the dual-task neural network model.

[0138] The first prediction result is used to indicate: the number of correctable errors that occur in the memory within a second time period; and the second time period is after the first time period.

[0139] Specifically, the memory fault prediction device may input the state feature data corresponding to the operating state data output by the sharing layer 101 into the first task layer 102 to obtain the first prediction result output by the first task layer 102 .

[0140] S304. The memory fault prediction device determines a second prediction result by using the state feature data through the second task layer of the dual-task neural network model.

[0141] The second prediction result is used to indicate the probability of an uncorrectable error occurring in the memory within a third time period, which is after the first time period.

[0142] Specifically, the memory fault prediction device may input the state feature data corresponding to the operating state data output by the sharing layer 101 into the second task layer 103 to obtain the second prediction result output by the second task layer 103 .

[0143] In some implementations, the second prediction result includes prediction information indicating whether an uncorrectable error will occur in the memory within a third time period.

[0144] For example, the second task layer 103 can first obtain a prediction score between 0 and 1 based on the state feature data corresponding to the operating state data output by the shared layer 101. This prediction score indicates the probability of an uncorrectable memory error occurring within the third time period. The second task layer 103 can then determine the aforementioned prediction information based on the prediction score. For example, when the prediction score is greater than a score threshold, the prediction information obtained by the second task layer 103 indicates that an uncorrectable memory error will occur within the third time period. When the prediction score is less than or equal to the score threshold, the prediction information obtained by the second task layer 103 indicates that an uncorrectable memory error will not occur within the third time period.

[0145] The technical solution provided by the present disclosure constructs a dual-task neural network model. The dual-task neural network model 10 can use n attributes (referred to as operating status data) of the target device's memory during a previous period (referred to as a first period) to predict the number of correctable errors that will occur in the memory during a subsequent period (referred to as a second period), obtaining a first predetermined result, and also predict the probability of uncorrectable errors that will occur in the memory during a subsequent period (referred to as a third period), obtaining a second predicted result. Furthermore, the dual-task neural network model 10 can include a shared layer 101, a first task layer 102, and a second task layer 103. The shared layer 101 can first extract features from the operating status data of the memory during the first period to obtain state feature data. Then, the first task layer 102 uses this state feature data to predict the number of CE errors during the second period (i.e., determine a first predicted result), and the second task layer 103 uses this state feature data to predict the probability of UCE errors during the third period (i.e., determine a second predicted result). In this way, the correlation between the number of CE occurrences and the probability of UCE occurrence can be fully utilized, making the predicted number of CE occurrences and the probability of UCE occurrence more accurate.

[0146] Specifically, Figure 9 (a) is a regression prediction scatter plot of the number of CE occurrences when the second task layer (hereinafter referred to as "CE" type) is removed based on the dual-task neural network model 10 provided by the embodiment of the present disclosure; Figure 9 (b) is a regression prediction scatter plot of the number of CE occurrences when the weight values ​​of the first function and the second function in the loss function are set to fixed values ​​(hereinafter referred to as "CE+UCE" type) based on the dual-task neural network model 10 provided in an embodiment of the present disclosure; Figure 9 (c) is a regression prediction scatter plot of the number of CE occurrences when the weight values ​​of the first and second functions in the dynamic adjustment loss function are added to the dual-task neural network model 10 provided in the embodiment of the present disclosure (hereinafter referred to as "CE+UCE+WeightAdj" type). It can be seen that Figure 9 The predicted points in (c) are more closely distributed near the y = x reference line, while the predicted points of the other models show a clear trend of discrete distribution. Therefore, it can be seen that the accuracy of the model prediction can be further improved by utilizing the additional supervision information provided by the UCE classification task in the multi-task learning framework and by dynamically adjusting the weights of the first and second functions in the loss function.

[0147] In addition, the mean absolute error (MAE) shown in Formula 12 and the mean squared error (MSE) shown in Formula 13 are used to measure the prediction accuracy of the number of memory CEs.

[0148] Formula 12

[0149] Style Thirteen

[0150] in, represents the number of samples, represents the true value, Represents the predicted value.

[0151] Figure 10 The MSE and MAE values ​​for the three algorithms, CE, CE+UCE, and CE+UCE+WeightAdj, are shown in Table 1.

[0152]

[0153] It can be seen that the "CE+UCE+WeightAdj" method proposed in this invention significantly outperforms other comparison models in terms of both MAE and MSE indicators. Its MAE value (3.1257) is 39.9% lower than that of the single-task CE model (5.2115), and its MSE value (15.5323) is 58.0% lower than that of the single-task model (37.0351). It is worth noting that when a dual-task structure is adopted but dynamic weight adjustment is not introduced (CE+UCE), its MSE value (45.3586) is actually higher than that of the single-task model. This is because when multi-task learning is done without weight adjustment, the two tasks interfere with each other, resulting in a decrease in regression performance. Dynamic weight adjustment helps balance tasks and improve overall performance. Therefore, the synergistic effect of multi-task learning needs to rely on a reasonable weight distribution mechanism to be effectively exerted.

[0154] In addition, the precision (Precision) shown in Formula 14, the recall (Recall) shown in Formula 15, and the F1 score (F1-Score) shown in Formula 16 are used to measure the prediction accuracy of whether memory UCE occurs.

[0155] Formula 14

[0156] Formula 15

[0157] Formula 16

[0158] in, represents the true positive examples, that is, the number of samples correctly predicted by the model as positive. Represents false positives, that is, the number of samples that the model incorrectly predicts as positive. represents false negative examples, i.e., samples that are incorrectly predicted as negative by the model.

[0159] Figure 11 Table 2 shows the statistical results of precision, recall and F1 score in three cases: when the first task layer is removed on the basis of the dual-task neural network model 10 provided in the embodiment of the present disclosure (hereinafter referred to as the "UCE" type), when the weight values ​​of the first function and the second function in the loss function are set to fixed values ​​on the basis of the dual-task neural network model 10 provided in the embodiment of the present disclosure (hereinafter referred to as the "CE+UCE" type), and when the weight values ​​of the first function and the second function in the loss function are dynamically adjusted on the basis of the dual-task neural network model 10 provided in the embodiment of the present disclosure (hereinafter referred to as the "CE+UCE+WeightAdj" type). The specific values ​​are shown in Table 2.

[0160]

[0161] As can be seen, in the UCE classification task, the "UCE" model has low precision but high recall, indicating that the model tends to predict many samples as positive, resulting in a high number of false positives. The "CE+UCE+WeightAdj" model achieves the best performance in terms of precision (0.7143), recall (0.7143), and F1 score (0.7143). Compared with the "UCE" model, the F1 score improves by 71.4%, with the improvement in precision (142.9%) being particularly significant. This demonstrates that the joint training framework effectively alleviates the high false positive prediction problem in the "UCE" model. The classification prediction result graph shows that the true positive (TP) region of the proposed method has significantly expanded, while the false positive (FP) region has significantly shrunk, indicating that the model can more accurately distinguish between normal memory and faulty memory that is about to experience UCE.

[0162] It is noteworthy that while the recall rate of the "CE+UCE" model remains high (0.8571), its precision rate (0.4286) is still significantly lower than that of the full model. This confirms the necessity of the dynamic weight adjustment mechanism. By automatically balancing the loss weights of the two tasks during training, it not only retains the high sensitivity of UCE predictions but also significantly improves the specificity of predictions.

[0163] In summary, when performing CE regression and UCE classification simultaneously, there is a significant synergistic enhancement effect between the two tasks. In the full model, the MAE of memory CE prediction is 39.9% lower than that of the single-task model, while the F1 score of UCE prediction is 71.4% higher. The dual-task model (CE+UCE) without dynamic weights even achieves a higher MSE (45.3586) for memory CE prediction than the single-task model (37.0351), indicating that fixed weights can lead to unbalanced training when the loss scales of the two tasks differ significantly.

[0164] Based on the same inventive concept, as an implementation of the above method, an embodiment of the present application further provides a memory fault prediction device, which corresponds to the above method embodiment. For ease of reading, this embodiment will no longer repeat the details of the above method embodiment one by one, but it should be clear that the memory fault prediction device in this embodiment can correspond to the implementation of part or all of the contents of the above method embodiment.

[0165] The present invention provides a memory fault prediction device. Figure 12 A structural diagram of the memory fault prediction device is shown in FIG. Figure 12 As shown, the memory fault prediction device 40 includes:

[0166] The acquisition unit 401 is configured to acquire operation status data; the operation status data is used to indicate: n attributes of the memory of the target device within a first time period; n is a positive integer;

[0167] The processing unit 402 is configured to extract features from the running status data through a shared layer of the dual-task neural network model to obtain state feature data corresponding to the running status data;

[0168] The processing unit 402 is further configured to determine a first prediction result using the state feature data through a first task layer of the dual-task neural network model; the first prediction result is configured to indicate a number of correctable errors that occurred in the memory during a second time period; the second time period being after the first time period;

[0169] The processing unit 402 is also used to determine a second prediction result using the state feature data through the second task layer of the dual-task neural network model; the second prediction result is used to indicate: the probability of an uncorrectable error occurring in the memory within a third time period; the third time period is after the first time period.

[0170] In some implementations, the acquiring unit 401 is configured to acquire the operating status data, including:

[0171] An acquiring unit 401 is configured to acquire n items of status data; the n items of status data are respectively used to indicate n properties of the memory of the target device within a first time period;

[0172] The acquisition unit 401 is configured to perform data dimensionality reduction on n items of status data to obtain operation status data.

[0173] In some implementations, the acquiring unit 401 is configured to perform data dimensionality reduction on n items of status data to obtain operating status data, including:

[0174] An acquisition unit 401 is used to acquire a correlation coefficient matrix corresponding to n attributes;

[0175] An acquisition unit 401 is configured to determine n eigenvectors of a correlation coefficient matrix and n eigenvalues ​​corresponding to the n eigenvectors;

[0176] The acquisition unit 401 is configured to determine p eigenvectors corresponding to p target eigenvalues ​​from the n eigenvectors; the p target eigenvalues ​​are: the first p eigenvalues ​​in the n eigenvalues ​​sorted from largest to smallest in numerical value;

[0177] The acquisition unit 401 is configured to project n items of status data onto a feature space corresponding to p feature vectors to obtain operating status data containing the p items of feature data.

[0178] In some implementations, the shared layer includes: a feature extraction layer constructed using a long short-term memory network, and a shared feature layer constructed using a fully connected network;

[0179] The processing unit 402 is further configured to perform feature extraction on the operating status data through the shared layer of the dual-task neural network model to obtain status feature data corresponding to the operating status data, including:

[0180] The processing unit 402 is further configured to extract features from the operating status data through a feature extraction layer;

[0181] The processing unit 402 is further configured to fuse the feature data output by the feature extraction layer through the shared feature layer to obtain state feature data.

[0182] In some implementations, the first task layer is a machine learning model constructed using a fully connected network; the second task layer is a machine learning model constructed using a probabilistic neural network.

[0183] In some implementations, the n attributes include one or more of the following: operating voltage, ambient temperature, operating frequency, power, data refresh rate, read and write latency, the number of row failures that occur in the memory within the first time period, the number of column failures that occur in the memory within the first time period, the number of correctable errors that occur in the memory within the first time period, and the number of uncorrectable errors that occur in the memory within the first time period.

[0184] Based on the same inventive concept, as an implementation of the above method, the embodiment of the present application also provides a model training device, which corresponds to the above method embodiment. For ease of reading, this embodiment will no longer repeat the details of the above method embodiment one by one, but it should be clear that the model training device in this embodiment can correspond to the implementation of part or all of the contents of the above method embodiment.

[0185] The present application embodiment provides a model training device, Figure 13 A structural diagram of the model training device is shown in Figure 2. Figure 13 As shown, the model training device 50 includes:

[0186] An acquisition unit 501 is configured to acquire x training samples and labels corresponding to the x training samples; wherein the x training samples include: n attributes of the memory of the target device in x different time periods, where n is a positive integer; and the label corresponding to each training sample includes: the number of correctable errors that occurred in the memory in a first preset time period after the time period corresponding to the training sample, and whether an uncorrectable error occurred in the memory in a second preset time period after the time period corresponding to the training sample;

[0187] A training unit 502 is configured to perform supervised learning training on the dual-task neural network model using the x training samples and the labels corresponding to the x training samples;

[0188] Among them, the dual-task neural network model includes: a shared layer, a first task layer and a second task layer; wherein the shared layer is used to: extract features of n attributes of the memory within the first time period to obtain state feature data; the first task layer is used to: use the state feature data to determine a first prediction result; the second task layer is used to: use the state feature data to determine a second prediction result; wherein the first prediction result is used to indicate: the number of correctable errors that occur in the memory within the second time period; the second prediction result is used to indicate: the probability of uncorrectable errors that occur in the memory within the third time period; the second time period and the third time period are after the first time period.

[0189] The memory fault prediction device 40 and the model training device 50 provided in the embodiments of the present application can execute the method provided in any of the above embodiments. The implementation principles and technical effects are similar and will not be repeated here.

[0190] The embodiment of the present application also provides an electronic device, such as Figure 14As shown, the electronic device includes a memory 601 and a processor 602, wherein the memory 601 stores a computer program, and the processor 602 is configured to run the computer program to execute the steps in the above-mentioned memory fault prediction method embodiment, or the processor 602 is configured to run the computer program to execute the steps in the above-mentioned model training method embodiment.

[0191] An embodiment of the present disclosure also provides a computer-readable storage medium, which stores a computer program, wherein the computer program is configured to execute the steps of any one of the above-mentioned memory fault prediction method embodiments when running, or the computer program is configured to execute the steps of any one of the above-mentioned model training method embodiments when running.

[0192] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk, or an optical disk.

[0193] An embodiment of the present disclosure also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps in any of the above-mentioned memory fault prediction method embodiments, or when the computer program is executed by a processor, it implements the steps in any of the above-mentioned model training method embodiments.

[0194] An embodiment of the present disclosure also provides another computer program product, including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the computer program implements the steps in any of the above-mentioned memory fault prediction method embodiments, or when the computer program is executed by a processor, the computer program implements the steps in any of the above-mentioned model training method embodiments.

[0195] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0196] The method provided by the present disclosure is described in detail above. Specific examples are used herein to illustrate the principles and implementation methods of the present disclosure. The description of the above embodiments is only used to help understand the method of the present disclosure and its core ideas. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present disclosure, several improvements and modifications can be made to the present disclosure, and these improvements and modifications also fall within the scope of protection of the claims of the present disclosure.

Claims

1. A memory fault prediction method, characterized in that: include: Get running status data; The operating status data is used to indicate: n properties of the memory of the target device within a first time period; n is a positive integer; Performing feature extraction on the operating status data through a shared layer of a dual-task neural network model to obtain status feature data corresponding to the operating status data; Determining a first prediction result using the state feature data through the first task layer of the dual-task neural network model; the first prediction result is used to indicate: the number of correctable errors that occurred in the memory during a second time period; The second time period is after the first time period; Determining a second prediction result using the state feature data through the second task layer of the dual-task neural network model; the second prediction result is used to indicate: a probability of an uncorrectable error occurring in the memory within a third time period; the third time period being after the first time period; The loss function used in training the dual-task neural network model satisfies the following conditions: the loss function is a weighted sum of a first function and a second function; the first function is used to reflect the accuracy of the output result of the first task layer; and the second function is used to reflect the accuracy of the output result of the second task layer. During the j-th model optimization, the weight value corresponding to the first function is positively correlated with the degree of loss reduction of the output result of the first task layer within the preset training process, and the weight value corresponding to the second function is positively correlated with the degree of loss reduction of the output result of the second task layer within the preset training process; the preset training process includes the model training process from the j-1-th model optimization to the j-th model optimization.

2. The method according to claim 1, characterized in that The obtaining of the operating status data includes: Acquire n items of status data; the n items of status data are respectively used to indicate n properties of the memory of the target device within the first time period; Performing data dimensionality reduction on the n items of status data to obtain the operating status data.

3. The method according to claim 2, characterized in that The performing data dimensionality reduction on the n items of status data to obtain the operating status data includes: Obtain the correlation coefficient matrix corresponding to the n attributes; Determining n eigenvectors of the correlation coefficient matrix and n eigenvalues ​​corresponding to the n eigenvectors; Determine p eigenvectors corresponding to p target eigenvalues ​​from the n eigenvectors; the p target eigenvalues ​​are: the first p eigenvalues ​​of the n eigenvalues ​​sorted from largest to smallest in numerical order; The n items of status data are projected onto the feature space corresponding to the p feature vectors to obtain the operating status data containing the p items of feature data.

4. The method according to claim 1, wherein The shared layer includes: a feature extraction layer constructed using a long short-term memory network, and a shared feature layer constructed using a fully connected network; The step of extracting features from the running status data through the shared layer of the dual-task neural network model to obtain state feature data corresponding to the running status data includes: Performing feature extraction on the operating status data through the feature extraction layer; The feature data output by the feature extraction layer are fused through the shared feature layer to obtain the state feature data.

5. The method according to claim 1, characterized in that The first task layer is a machine learning model constructed using a fully connected network; the second task layer is a machine learning model constructed using a probabilistic neural network.

6. The method according to claim 1, characterized in that The n attributes include one or more of the following: operating voltage, ambient temperature, operating frequency, power, data refresh rate, read and write delay, the number of row failures that occur in the memory during the first time period, the number of column failures that occur in the memory during the first time period, the number of correctable errors that occur in the memory during the first time period, and the number of uncorrectable errors that occur in the memory during the first time period.

7. A model training method, characterized in that: The method comprises: Obtain x training samples and labels corresponding to the x training samples; wherein the x training samples include: n attributes of the memory of the target device in x different time periods, where n is a positive integer; and the label corresponding to each training sample includes: the number of correctable errors occurring in the memory in a first preset time period after the time period corresponding to the training sample, and whether an uncorrectable error occurs in the memory in a second preset time period after the time period corresponding to the training sample; Performing supervised learning training on the dual-task neural network model using the x training samples and the labels corresponding to the x training samples; The dual-task neural network model includes: a shared layer, a first task layer, and a second task layer; wherein the shared layer is used to extract features of the n attributes of the memory within a first time period to obtain state feature data; the first task layer is used to use the state feature data to determine a first prediction result; the second task layer is used to use the state feature data to determine a second prediction result; wherein the first prediction result is used to indicate the number of correctable errors that occurred in the memory within a second time period; the second prediction result is used to indicate the probability of uncorrectable errors occurring in the memory within a third time period; the second time period and the third time period are after the first time period; The loss function used in training the dual-task neural network model satisfies the following conditions: the loss function is a weighted sum of a first function and a second function; the first function is used to reflect the accuracy of the output result of the first task layer; and the second function is used to reflect the accuracy of the output result of the second task layer. During the j-th model optimization, the weight value corresponding to the first function is positively correlated with the degree of loss reduction of the output result of the first task layer within the preset training process, and the weight value corresponding to the second function is positively correlated with the degree of loss reduction of the output result of the second task layer within the preset training process; the preset training process includes the model training process from the j-1-th model optimization to the j-th model optimization.

8. The method according to claim 7, characterized in that If the loss of the output result of the first task layer increases during the preset training process, the weight value corresponding to the first function is set to zero during the j-th model optimization; Alternatively, if the loss of the output result of the second task layer increases during the preset training process, the weight value corresponding to the second function is set to zero during the j-th model optimization.

9. The method according to claim 7, characterized in that The method of performing supervised learning training on the dual-task neural network model using the x training samples and the labels corresponding to the x training samples includes: Performing data dimensionality reduction on each of the x training samples to obtain x reduced-dimensionality training samples; Supervised learning training is performed on the dual-task neural network model using the x dimensionality reduction training samples and the labels corresponding to the x training samples.

10. The method according to claim 9, characterized in that The performing data dimensionality reduction on each of the x training samples to obtain x dimensionality-reduced training samples includes: Using the x training samples, calculate the correlation coefficient matrix corresponding to the n attributes; Determining n eigenvectors of the correlation coefficient matrix and n eigenvalues ​​corresponding to the n eigenvectors; Determine p eigenvectors corresponding to p target eigenvalues ​​from the n eigenvectors; the p target eigenvalues ​​are: the first p eigenvalues ​​of the n eigenvalues ​​sorted from largest to smallest in numerical order; Each of the x training samples is projected onto the feature space corresponding to the p feature vectors to obtain x dimensionality-reduced training samples; wherein each of the dimensionality-reduced training samples contains p items of feature data.

11. A memory fault prediction device, characterized in that: include: An acquisition unit, used for acquiring operation status data; The operating status data is used to indicate: n properties of the memory of the target device within a first time period; a processing unit, configured to perform feature extraction on the operating status data through a shared layer of a dual-task neural network model to obtain status feature data corresponding to the operating status data; The processing unit is further configured to determine a first prediction result using the state feature data through a first task layer of the dual-task neural network model; the first prediction result is configured to indicate a number of correctable errors that occurred in the memory within a second time period; the second time period being after the first time period; The processing unit is further configured to determine a second prediction result using the state feature data through a second task layer of the dual-task neural network model; the second prediction result is configured to indicate a probability of an uncorrectable error occurring in the memory within a third time period; the third time period being after the first time period; The loss function used in training the dual-task neural network model satisfies the following conditions: the loss function is a weighted sum of a first function and a second function; the first function is used to reflect the accuracy of the output result of the first task layer; and the second function is used to reflect the accuracy of the output result of the second task layer. During the j-th model optimization, the weight value corresponding to the first function is positively correlated with the degree of loss reduction of the output result of the first task layer within the preset training process, and the weight value corresponding to the second function is positively correlated with the degree of loss reduction of the output result of the second task layer within the preset training process; the preset training process includes the model training process from the j-1-th model optimization to the j-th model optimization.

12. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the memory fault prediction method according to any one of claims 1 to 6 when executing the computer program; Alternatively, the processor is configured to implement the steps of the model training method as claimed in any one of claims 7 to 10 when executing the computer program.

13. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, the steps of the memory fault prediction method according to any one of claims 1 to 6 are implemented, or when the computer program is executed by a processor, the steps of the model training method according to any one of claims 7 to 10 are implemented.

Images