Model generation method, fault prediction method, device, equipment, medium and product

Abstract

The present application discloses a model generation method, a fault prediction method, an apparatus, a device, a medium and a product, which belong to the field of storage technology. It includes: collecting a set of unit state parameters corresponding to a sample storage unit in a plurality of consecutive programming and erasing cycles; splicing the unit state parameter set based on the time sequence of the programming and erasing cycles corresponding to the unit state parameter set, and generating a sample feature sequence; inputting the sample feature sequence into a time series prediction model to obtain a target result output by the time series prediction model; calculating a target loss value according to the target result and the fault label value corresponding to the sample storage unit, and adjusting the model parameters of the time series prediction model based on the target loss value to obtain a target neural network model. A sample feature sequence with time continuity can be generated, and the time series prediction model can extract the timing characteristics of the storage unit state change, thereby improving the accuracy of the target neural network model in predicting the fault of the storage unit.

Description

Technical Field

[0001] The present application belongs to the field of storage technology, and specifically relates to a model generation method, a fault prediction method, a device, a equipment, a medium and a product. Background Art

[0002] With the rapid development of computer technology, the importance of data is increasing day by day. More and more data is generated and more and more memories are put into use. However, memories have a lifespan. When a memory fails, it will pose a greater threat to data security. Therefore, it is more important to predict the failure of storage units.

[0003] In the related art, the state parameters of the storage unit on the time slice are usually collected during the operation of the storage unit, and the state parameters are input into a pre-set conditional model, neural network model or other fault prediction model to determine whether the storage unit has a potential fault. For example, the failure probability of the memory is determined based on the historical cumulative write volume of the storage unit. The higher the historical cumulative write volume, the higher the failure probability.

[0004] However, due to the differences in the physical properties and working conditions of different storage units, the accuracy of using the above fault detection method to predict the fault of the storage unit is relatively poor. Summary of the invention

[0005] The purpose of the embodiments of the present application is to provide a model generation method, a fault prediction method, an apparatus, a device, a medium and a product, which can solve the problem of poor accuracy in fault prediction of storage units in related technologies.

[0006] In a first aspect, an embodiment of the present application provides a model generation method, the method comprising:

[0007] Collecting a set of cell state parameters corresponding to a plurality of consecutive programming and erasing cycles of a sample storage cell; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles;

[0008] Based on the time sequence of programming and erasing cycles corresponding to the cell state parameter set, the cell state parameter set is spliced ​​to generate a sample feature sequence;

[0009] Inputting the sample feature sequence into a time series prediction model to obtain a target result output by the time series prediction model;

[0010] A target loss value is calculated according to the target result and the fault label value corresponding to the sample storage unit, and the model parameters of the time series prediction model are adjusted based on the target loss value to obtain a target neural network model; wherein the target neural network model is used to output the failure probability of the target storage unit according to the target feature sequence of the target storage unit.

[0011] Optionally, the method further comprises:

[0012] Acquire a set of cell state parameters corresponding to each programming and erasing cycle of the sample storage cell in a complete life cycle;

[0013] Determine from the complete life cycle a first programming and erasing cycle in which the cell state parameter set satisfies a preset state condition;

[0014] The plurality of consecutive program-erase cycles are determined from a second program-erase cycle preceding the first program-erase cycle.

[0015] Optionally, the cell state parameter set includes a first state parameter and a second state parameter, and the step of splicing the cell state parameter set based on a time sequence of programming and erasing cycles corresponding to the cell state parameter set to generate a sample feature sequence includes:

[0016] Normalizing the first state parameter and the second state parameter corresponding to the program-erase cycle respectively to obtain a first normalized result corresponding to the first state parameter and a second normalized result corresponding to the second state parameter; wherein the normalization intervals of the first normalized result and the second normalized result are the same;

[0017] Splicing the first normalized result and the second normalized result to obtain a time step feature corresponding to the program-erase cycle;

[0018] Based on the time sequence of the programming and erasing cycles, the time step features corresponding to the respective programming and erasing cycles are concatenated to obtain the sample feature sequence.

[0019] Optionally, the splicing of the first normalized result and the second normalized result to obtain the time step feature corresponding to the program-erase cycle includes:

[0020] Performing feature crossover on the first normalized result and the second normalized result to obtain a crossover feature corresponding to the program-erase cycle; wherein the feature crossover includes at least one of a product crossover, an additive crossover, and a polynomial crossover;

[0021] The first normalization result, the second normalization result and the cross feature are concatenated to obtain a time step feature corresponding to the program-erase cycle.

[0022] Optionally, the time series prediction model includes a feature dimensionality enhancement layer and a time series prediction network layer, the feature dimensionality enhancement layer is used to enhance the feature dimension of the sample feature sequence and output a high-dimensional feature sequence corresponding to the sample feature sequence; the time series prediction network layer is used to output the target result according to the high-dimensional feature sequence; wherein the feature dimension of the high-dimensional feature sequence is greater than the feature dimension of the sample feature sequence.

[0023] Optionally, the feature dimension enhancement layer is constructed based on a convolutional neural network, and / or the time series prediction network layer is constructed based on a long short-term memory network.

[0024] Optionally, the convolutional neural network includes multiple hidden layers, each of which is composed of a first number of one-dimensional convolution kernels. The multiple hidden layers are used to perform feature dimension upgrading on the sample feature sequence and output the high-dimensional feature sequence with the feature dimension being the first number.

[0025] Optionally, the time series prediction network layer includes a multi-layer long short-term memory network, the long short-term memory network includes a plurality of cells, and the cells are composed of a forget gate, an input gate and an output gate.

[0026] Optionally, the target result includes failure sub-probabilities corresponding to target numbers of program-erase cycles to be predicted.

[0027] Optionally, the time series prediction model includes a fully connected layer consisting of a normalization layer and the target number of neurons, the neurons are used to output the linear transformation results corresponding to the programming and erasure cycle to be predicted, and the normalization layer is used to calculate the target natural index of the linear transformation results corresponding to the target programming and erasure cycle to be predicted, and the ratio of the target natural index of the linear transformation results corresponding to each programming and erasure cycle to be predicted to obtain the failure sub-probability corresponding to the target programming and erasure cycle to be predicted.

[0028] In a second aspect, an embodiment of the present application provides a fault prediction method, the method comprising:

[0029] Collecting a set of cell state parameters corresponding to a target storage cell in a plurality of consecutive programming and erasing cycles; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles;

[0030] Based on the time sequence of programming and erasing cycles corresponding to the cell state parameter set, the cell state parameter set is spliced ​​to generate a target feature sequence;

[0031] The target feature sequence is input into a target neural network model to obtain the failure probability of the target storage unit output by the target neural network model; wherein the target neural network model is obtained based on the above-mentioned model generation method.

[0032] Optionally, the failure probability includes failure sub-probabilities corresponding to a target number of program-erase cycles to be predicted.

[0033] Optionally, the method further comprises:

[0034] Collecting a set of cell state parameters corresponding to the programming and erasing cycles to be predicted;

[0035] Determine a target fault label value corresponding to the target feature sequence based on a set of cell state parameters corresponding to the program-erase cycle to be predicted and a preset state condition;

[0036] The target neural network model is trained based on the target feature sequence and the target fault label value.

[0037] Optionally, the method further comprises:

[0038] When the failure probability is greater than or equal to a first probability and less than a second probability, reducing the erasing frequency of the target storage unit; wherein the first probability is less than the second probability;

[0039] When the failure probability is greater than or equal to the second probability, the cold data stored in the memory where the target storage unit is located is migrated to the target storage unit.

[0040] Optionally, migrating the cold data stored in the memory where the target storage unit is located to the target storage unit includes:

[0041] When the failure probability is less than a third probability, migrating the cold data stored in the memory where the target storage unit is located to the target storage unit; wherein the third probability is greater than the second probability;

[0042] When the failure probability is greater than or equal to the third probability, the data stored in the target storage unit is migrated to other storage units.

[0043] In a third aspect, an embodiment of the present application provides a model generation device, the device comprising:

[0044] A collection module, used for collecting a set of cell state parameters corresponding to a sample storage cell in a plurality of consecutive programming and erasing cycles; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles;

[0045] A splicing module, used for splicing the cell state parameter set based on the time sequence of programming and erasing cycles corresponding to the cell state parameter set to generate a sample feature sequence;

[0046] An input module, used to input the sample feature sequence into a time series prediction model to obtain a target result output by the time series prediction model;

[0047] A training module is used to calculate a target loss value based on the target result and the fault label value corresponding to the sample storage unit, and adjust the model parameters of the time series prediction model based on the target loss value to obtain a target neural network model; wherein the target neural network model is used to output the failure probability of the target storage unit according to the target feature sequence of the target storage unit.

[0048] Optionally, the device further comprises:

[0049] A set acquisition module, used to acquire a set of cell state parameters corresponding to each programming and erasing cycle of the sample storage cell in a complete life cycle;

[0050] A first programming and erasing cycle module, configured to determine, from the complete life cycle, a first programming and erasing cycle in which the cell state parameter set satisfies a preset state condition;

[0051] The multiple consecutive program-erase cycle determination module is used to determine the multiple consecutive program-erase cycles from a second program-erase cycle before the first program-erase cycle.

[0052] Optionally, the unit state parameter set includes a first state parameter and a second state parameter, and the splicing module includes:

[0053] A normalization submodule, used to normalize the first state parameter and the second state parameter corresponding to the program-erase cycle respectively, to obtain a first normalization result corresponding to the first state parameter and a second normalization result corresponding to the second state parameter; wherein the normalization intervals of the first normalization result and the second normalization result are the same;

[0054] A result splicing submodule, used for splicing the first normalized result and the second normalized result to obtain a time step feature corresponding to the programming and erasing cycle;

[0055] The sample feature sequence splicing submodule is used to splice the time step features corresponding to each of the programming and erasing cycles based on the time sequence of the programming and erasing cycles to obtain the sample feature sequence.

[0056] Optionally, the result splicing submodule includes:

[0057] A feature crossover unit, configured to perform feature crossover on the first normalized result and the second normalized result to obtain a crossover feature corresponding to the program-erase cycle; wherein the feature crossover includes at least one of a product crossover, an additive crossover and a polynomial crossover;

[0058] A time step feature unit is used to concatenate the first normalization result, the second normalization result and the cross feature to obtain a time step feature corresponding to the programming and erasing cycle.

[0059] Optionally, the time series prediction model includes a feature dimensionality enhancement layer and a time series prediction network layer, the feature dimensionality enhancement layer is used to enhance the feature dimension of the sample feature sequence and output a high-dimensional feature sequence corresponding to the sample feature sequence; the time series prediction network layer is used to output the target result according to the high-dimensional feature sequence; wherein the feature dimension of the high-dimensional feature sequence is greater than the feature dimension of the sample feature sequence.

[0060] Optionally, the feature dimension enhancement layer is constructed based on a convolutional neural network, and / or the time series prediction network layer is constructed based on a long short-term memory network.

[0061] Optionally, the convolutional neural network includes multiple hidden layers, each of which is composed of a first number of one-dimensional convolution kernels. The multiple hidden layers are used to perform feature dimension upgrading on the sample feature sequence and output the high-dimensional feature sequence with the feature dimension being the first number.

[0062] Optionally, the time series prediction network layer includes a multi-layer long short-term memory network, the long short-term memory network includes a plurality of cells, and the cells are composed of a forget gate, an input gate and an output gate.

[0063] Optionally, the target result includes failure sub-probabilities corresponding to target numbers of program-erase cycles to be predicted.

[0064] Optionally, the time series prediction model includes a fully connected layer consisting of a normalization layer and the target number of neurons, the neurons are used to output the linear transformation results corresponding to the programming and erasure cycle to be predicted, and the normalization layer is used to calculate the target natural index of the linear transformation results corresponding to the target programming and erasure cycle to be predicted, and the ratio of the target natural index of the linear transformation results corresponding to each programming and erasure cycle to be predicted to obtain the failure sub-probability corresponding to the target programming and erasure cycle to be predicted.

[0065] In a fourth aspect, an embodiment of the present application provides a fault prediction device, the device comprising:

[0066] A target acquisition module, used for acquiring a set of cell state parameters corresponding to a target storage cell in a plurality of consecutive programming and erasing cycles; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles;

[0067] A target feature sequence module, used to splice the cell state parameter set based on the time sequence of programming and erasing cycles corresponding to the cell state parameter set to generate a target feature sequence;

[0068] A target input module is used to input the target feature sequence into a target neural network model to obtain the failure probability of the target storage unit output by the target neural network model; wherein the target neural network model is obtained based on the above-mentioned model generation method.

[0069] Optionally, the failure probability includes failure sub-probabilities corresponding to a target number of program-erase cycles to be predicted.

[0070] Optionally, the device further comprises:

[0071] A cell state parameter set submodule, used for collecting a cell state parameter set corresponding to the programming and erasing cycle to be predicted;

[0072] A target fault label value submodule, used to determine a target fault label value corresponding to the target feature sequence based on a set of cell state parameters corresponding to the program-erase cycle to be predicted and a preset state condition;

[0073] The model optimization submodule is used to train the target neural network model based on the target feature sequence and the target fault label value.

[0074] Optionally, the device further comprises:

[0075] An erasing frequency module, used to reduce the erasing frequency of the target storage unit when the failure probability is greater than or equal to a first probability and less than a second probability; wherein the first probability is less than the second probability;

[0076] A cold data migration module is used to migrate the cold data stored in the memory where the target storage unit is located to the target storage unit when the failure probability is greater than or equal to the second probability.

[0077] Optionally, the cold data migration module includes:

[0078] A first migration submodule, configured to migrate the cold data stored in the memory where the target storage unit is located to the target storage unit when the failure probability is less than a third probability; wherein the third probability is greater than the second probability;

[0079] The second migration submodule is used to migrate the data stored in the target storage unit to other storage units when the failure probability is greater than or equal to the third probability.

[0080] In a fifth aspect, an embodiment of the present application provides an electronic device, which includes a processor and a memory, wherein the memory stores programs or instructions that can be run on the processor, and when the programs or instructions are executed by the processor, the above-mentioned model generation method or fault prediction method is implemented.

[0081] In a sixth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a program or instruction is stored, and the program or instruction, when executed by a processor, implements the above-mentioned model generation method or fault prediction method.

[0082] In a seventh aspect, an embodiment of the present application provides a computer program product, including a computer program or instructions, which implement the above-mentioned model generation method or fault prediction method when executed by a processor.

[0083] In an embodiment of the present application, a model generation method is provided, including collecting a set of cell state parameters corresponding to a sample storage unit in a plurality of consecutive programming and erasing cycles; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles; based on the time sequence of the programming and erasing cycles corresponding to the cell state parameter set, splicing the cell state parameter set to generate a sample feature sequence; inputting the sample feature sequence into a time series prediction model to obtain a target result output by the time series prediction model; calculating a target loss value according to the target result and the fault label value corresponding to the sample storage unit, and adjusting the model parameters of the time series prediction model based on the target loss value to obtain a target neural network model; wherein the target neural network model is used to output the failure probability of the target storage unit according to the target feature sequence of the target storage unit. Based on the multi-dimensional state parameters generated by the sample storage unit in a plurality of consecutive programming and erasing cycles, a sample feature sequence with time continuity can be generated, and input into the time series prediction model for model training, so that the time series prediction model can not only extract the time series characteristics of the state change of the storage unit, but also extract the correlation between the multiple state parameters of the storage unit in each time step, which is helpful to improve the accuracy of the trained target neural network model in predicting the failure of the storage unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 is a flowchart of the steps of a model generation method provided in an embodiment of the present application;

[0085] Figure 2 is a flowchart of another model generation method provided in an embodiment of the present application;

[0086] Figure 3 It is a structural diagram of a time series prediction model provided in an embodiment of the present application;

[0087] Figure 4 It is a schematic diagram of a long short-term memory network provided in an embodiment of the present application;

[0088] Figure 5 is a flowchart of the steps of a fault prediction method provided by an embodiment of the present application;

[0089] Figure 6 is a block diagram of a model generation device provided in an embodiment of the present application;

[0090] Figure 7 is a block diagram of a fault prediction device provided in an embodiment of the present application;

[0091] Figure 8 It is a structural schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0092] The following will be combined with the drawings in the embodiments of the present application to clearly describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments in the present application belong to the scope of protection of this application.

[0093] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of one type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0094] The model generation method provided in the embodiment of the present application is described in detail below through specific embodiments and their application scenarios in conjunction with the accompanying drawings.

[0095] Reference Figure 1 , Figure 1 A flow chart of the steps of a model generation method provided in an embodiment of the present application is shown as follows: Figure 1 As shown, the specific steps include:

[0096] Step 101 : collecting sets of cell state parameters corresponding to a plurality of consecutive programming and erasing cycles of a sample storage cell; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles.

[0097] In the embodiment of the present application, the sample storage unit refers to a storage medium or storage area that can count program-erase cycles. Specifically, the sample storage unit may include storage media such as solid-state hard disks, flash memory particles, USB flash drives, etc., and may also include storage areas of a specific level in the storage medium, such as flash memory pages, flash memory blocks, flash memory segments, etc., which are not specifically limited in the embodiment of the present application.

[0098] Program / Erase Cycle (P / E Cycle) refers to a complete erase operation and a subsequent programming (writing) operation experienced by a storage unit. For example, if a storage unit is a flash block A, and the size of the flash block is 128KB, when writing 128KB of new data to the flash block A, the flash block A is first reset to the default initial state (usually all bits are set to 1) through an erase operation, and then the above 128KB of new data is written. At this point, each bit of the flash block A has undergone an erase and programming process, and the flash block A undergoes a program erase cycle.

[0099] In an embodiment of the present application, a programming-erasure cycle may include multiple programming-erasure cycles. The smaller the number of programming-erasure cycles included in a programming-erasure cycle, the greater the corresponding model training cost and reasoning overhead. The larger the number of cycles, the smaller the corresponding model training cost and reasoning overhead. The number of programming-erasure cycles included in a programming-erasure cycle can be flexibly set according to actual business needs, and the embodiment of the present application does not make any specific limitation.

[0100] For example, if a program-erase cycle is set to include 10 program-erase cycles, when a sample memory cell is in the 105th program-erase cycle in its life cycle, the sample memory cell has experienced 10 complete program-erase cycles in total and is currently experiencing the 11th program-erase cycle in its life cycle.

[0101] In the embodiment of the present application, the cell state parameter set may include any multiple cell state parameters related to the life of the sample storage cell, and the above cell state parameters may include but are not limited to average erase time, maximum erase time, average write time, maximum write time, average read time, maximum read time, original error bit number, number of write failures, number of read failures, etc., which are not specifically limited in the embodiment of the present application. It should be noted that the above cell state parameters represent the state parameters of the sample storage cell in the programming and erasing cycle, and each sample storage cell can generate a set of cell state parameter sets in a programming and erasing cycle. For example, the average read time of the sample storage cell A in a certain programming and erasing cycle represents: the average value of the read time of multiple data read operations of the sample storage cell A in the programming and erasing cycle.

[0102] The starting program-erase cycle of multiple consecutive program-erase cycles can be the first program-erase cycle in the life cycle of the storage unit, or the Nth program-erase cycle in the life cycle of the storage unit; the ending program-erase cycle of multiple consecutive program-erase cycles can be the 1+Mth program-erase cycle in the life cycle of the storage unit, or the N+Mth program-erase cycle, or the last program-erase cycle in the life cycle of the storage unit. The number of cycles included in the multiple consecutive program-erase cycles can be flexibly set by technicians, for example, it can be set to 10 or 50, and the embodiments of the present application do not make specific limitations.

[0103] In the embodiment of the present application, the programming and erasure cycles in the multiple consecutive programming and erasure cycles corresponding to different sample storage units may be different. The starting programming and erasure cycle of the multiple consecutive programming and erasure cycles may be set as the first programming and erasure cycle in the life cycle of the sample storage unit, and the ending programming and erasure cycle may be set as the last programming and erasure cycle in the life cycle of the sample storage unit, that is, the last programming and erasure cycle experienced by the sample storage unit before it is in a fault state. Since the lifespans of different sample storage units may be different, the number of programming and erasure cycles in the multiple consecutive programming and erasure cycles corresponding to different sample storage units may also be different.

[0104] In an embodiment of the present application, the sample storage cell can be continuously detected during its life cycle, and multiple cell state parameters of the sample storage cell can be collected in each programming and erasing cycle experienced by the sample storage cell to obtain a set of cell state parameters corresponding to the sample storage cell in multiple consecutive programming and erasing cycles.

[0105] Step 102: based on the time sequence of programming and erasing cycles corresponding to the cell state parameter set, the cell state parameter set is spliced ​​to generate a sample feature sequence.

[0106] In the embodiment of the present application, after obtaining the cell state parameter set corresponding to the sample storage cell, the cell state parameter sets can be spliced ​​according to the time sequence of the programming and erasing cycles corresponding to each cell state parameter set to obtain the sample feature sequence corresponding to the sample storage cell. The time sequence of the programming and erasing cycles can indicate the order in which each programming and erasing cycle occurs.

[0107] Specifically, the cell state parameter sets may be directly concatenated according to the time sequence of their corresponding programming and erasing cycles to obtain a sample feature sequence whose sequence elements are the cell state parameter sets.

[0108] For example, a sample storage cell collects a total of cell state parameter sets corresponding to three consecutive programming and erasing cycles, each of which contains four cell state parameters, namely, average erase time, average write time, average read time, and original number of error bits. The above three cell state parameter sets can be expressed as: cell state parameter set 1 (average erase time 1, average write time 1, average read time 1, original number of error bits 1), cell state parameter set 2 (average erase time 2, average write time 2, average read time 2, original number of error bits 2), cell state parameter set 3 (average erase time 3, average write time 3, average read time 3, original number of error bits 3), cell state parameter set 4 (average erase time 4, average write time 4, average read time 4, original number of error bits 4). The above cell state parameter sets 1 to 3 can be directly spliced ​​to obtain a sample feature sequence [(cell state parameter set 1), (cell state parameter set 2), (cell state parameter set 3), (cell state parameter set 4)].

[0109] It should be noted that, when the unit state parameter set contains only one parameter dimension, each element in the above sample feature sequence contains one parameter feature. At this time, the above sample feature sequence can be regarded as a one-dimensional vector; when the unit state parameter contains multiple parameter dimensions, each element in the above sample feature sequence contains multiple parameter features. At this time, the above sample feature sequence can be regarded as a two-dimensional matrix.

[0110] Step 103: input the sample feature sequence into a time series prediction model to obtain a target result output by the time series prediction model.

[0111] In an embodiment of the present application, the above-mentioned sample feature sequence can be input into a pre-built time series prediction model, and the time series prediction model processes the sample feature sequence and outputs the target result. Among them, the above-mentioned time series prediction model can include but is not limited to a recurrent neural network (RNN), a long short-term memory network (LSTM), a gated recurrent unit (GRU), a transformer model (Transformer), an autoregressive integrated moving average-long short-term memory network hybrid model (ARIMA-LSTM), and other neural network models that can process time series data with time dependence. The embodiment of the present application does not specifically limit the time series prediction model here.

[0112] The target result output by the time series prediction model may include the failure probability of the sample storage unit in the P+Kth programming and erasing cycle, and the value of the failure probability may be [0, 1], where 0 may represent a failure probability of 0% and 1 may represent a failure probability of 100%. Wherein, P represents the programming and erasing cycle order corresponding to the last unit state parameter set in the sample feature sequence input to the time series prediction model, and K may be any positive integer, that is, the failure probability of any number of programming and erasing cycles after the sample feature sequence of the sample storage unit may be predicted by the time series prediction model.

[0113] Exemplarily, if a sample feature sequence of a sample storage unit is generated based on a set of cell state parameters corresponding to the 1st to 100th programming and erasing cycles in the life cycle of the sample storage unit, then P=100. The sample feature sequence can be input into a time series prediction model, and the target result output by the time series model can include the failure probability corresponding to the 100th+K programming and erasing cycle of the sample storage unit, where K can be flexibly set by a technician, for example, K can be equal to 1, or can be equal to 2, etc., and the embodiments of the present application do not make specific limitations.

[0114] Step 104, calculate the target loss value according to the target result and the fault label value corresponding to the sample storage unit, and adjust the model parameters of the time series prediction model based on the target loss value to obtain a target neural network model; wherein the target neural network model is used to output the failure probability of the target storage unit according to the target feature sequence of the target storage unit.

[0115] In the embodiment of the present application, the fault label value corresponding to the sample storage unit can be determined according to whether the sample storage unit is in a fault state during the P+Kth programming and erasing cycle. Specifically, if the sample storage unit actually enters a fault state during the P+Kth programming and erasing cycle, its fault label value can be determined to be 1, otherwise, its fault label value is 0.

[0116] It should be noted that whether the sample storage unit is in a faulty state can be determined based on indicators such as the number of consecutive read failures, the number of consecutive write failures, the number of consecutive erase failures, and the data bit conversion rate. For example, if a sample storage unit fails to write data, and the data cannot be written correctly after more than 10 retries, it can be determined that the sample storage unit is in a faulty state. It can also be determined based on the faulty state marked by the storage controller of the sample storage unit. For example, in the case where the sample storage unit is a flash memory block, the bad block table (Bad Block Table, BBT) of the memory where the sample storage unit is located can be read to determine whether the sample storage unit is in a faulty state. In addition, technicians can also use other methods to determine whether the sample storage unit is in a faulty state, which is not specifically limited in the embodiments of the present application.

[0117] Further, in the embodiment of the present application, based on the cell state parameter sets corresponding to different multiple consecutive programming and erasing cycles of the same sample storage unit, multiple sample feature sequences corresponding to the sample storage unit can be generated, and for each sample feature sequence, a corresponding fault label value can be generated, thereby improving the diversity of training data. For example, if a sample storage unit is in a faulty state in the 100th programming and erasing cycle of its complete life cycle, in the case of K=1, sample feature sequence 1 can be generated based on the cell state parameter set corresponding to the 59th to 99th programming and erasing cycles of the sample storage unit, and the fault label value corresponding to the sample feature sequence 1 is determined to be 1; sample feature sequence 2 can also be generated based on the cell state parameter set corresponding to the 49th to 89th programming and erasing cycles of the sample storage unit, and the fault label value corresponding to the sample feature sequence 2 is determined to be 0, and so on, so that multiple sets of training data can be generated for a sample storage unit, the richness of the sample data can be improved, and the model training effect can be improved.

[0118] In the embodiment of the present application, the target loss value can be determined based on loss functions such as mean squared error loss (MSE), mean absolute error (MAE), binary cross entropy loss (BCE Loss), cross entropy loss (CE Loss), etc., and the embodiment of the present application does not make specific limitations. Specifically, the target result determined based on the sample feature sequence and the fault label value corresponding to the sample feature sequence can be input into the above loss function to obtain the target loss value output by the loss function.

[0119] Exemplarily, the target loss value may be a mean square error loss between the fault label value and the target result, and the mean square error loss may be calculated based on the following formula:

[0120]

[0121] in, represents the mean square error loss, Indicates the target result, Indicates the fault tag value, Indicates expected value.

[0122] The above time series prediction model can be trained by using a training set consisting of sample feature sequences and fault label values, using error back propagation and optimization algorithms (such as gradient descent), adjusting model parameters (weights) until the training goal is achieved, and obtaining a target neural network model. The above training goals may include reaching a predetermined number of training rounds, convergence of the loss function, and verification set performance meeting predetermined requirements, etc., which are not specifically limited in the embodiments of the present application.

[0123] The sample data composed of the above sample feature sequence and fault label value can also be divided into a training set, a validation set and a test set. The time series prediction model is trained by the training set, and the time series prediction model in the training process is verified by the validation set. If the verification result shows that the training goal is achieved, the time series prediction model can be fixed to obtain the target neural network model, and the target neural network model can be tested by the test set. For example, 60% of the sample data can be set as a training set, 20% of the sample data can be set as a validation set, and 20% of the sample data can be set as a test set.

[0124] After obtaining the target neural network model, the target neural network model can be deployed in the storage controller of the target storage unit, or in the operating system of the target device where the target storage unit is located, and the failure probability of the storage unit can be detected by the target neural network model. Among them, the target neural network model can output the failure probability of the target storage unit in the subsequent programming and erasing cycle according to the target feature sequence of the target storage unit. Among them, the target storage unit represents a storage unit that uses the model for fault prediction during the model application process. The target device may include but is not limited to a server (such as a physical machine in a server cluster), an industrial computer, a personal computer, a mobile terminal (such as a mobile phone, a laptop computer, etc.), etc., and the embodiments of the present application do not make specific limitations.

[0125] In summary, a model generation method provided by an embodiment of the present application includes collecting a set of cell state parameters corresponding to a sample storage unit in a plurality of consecutive programming and erasing cycles; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles; based on the time sequence of the programming and erasing cycles corresponding to the cell state parameter set, splicing the cell state parameter set to generate a sample feature sequence; inputting the sample feature sequence into a time series prediction model to obtain a target result output by the time series prediction model; calculating a target loss value according to the target result and the fault label value corresponding to the sample storage unit, and adjusting the model parameters of the time series prediction model based on the target loss value to obtain a target neural network model; wherein the target neural network model is used to output the failure probability of the target storage unit according to the target feature sequence of the target storage unit. Based on the multi-dimensional state parameters generated by the sample storage unit in a plurality of consecutive programming and erasing cycles, a sample feature sequence with time continuity can be generated, and input into the time series prediction model for model training, so that the time series prediction model can not only extract the time series characteristics of the state change of the storage unit, but also extract the correlation between the multiple state parameters of the storage unit in each time step, which is helpful to improve the accuracy of the trained target neural network model in predicting the failure of the storage unit.

[0126] Reference Figure 2 , Figure 2 A flowchart of another model generation method provided in an embodiment of the present application is shown. Figure 2 As shown, the specific steps include:

[0127] Step 201 : collecting sets of cell state parameters corresponding to a plurality of consecutive programming and erasing cycles of a sample storage cell; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles.

[0128] This step can refer to step 101 and will not be described in detail in this embodiment of the present application.

[0129] Optionally, step 201 may include:

[0130] Sub-step 2011, obtaining a set of cell state parameters corresponding to each programming and erasing cycle of the sample storage cell in a complete life cycle.

[0131] In the embodiment of the present application, a set of cell state parameters corresponding to all programming and erasing cycles experienced by the sample storage cell in its complete life cycle can be obtained, wherein the complete life cycle of the sample storage cell refers to the complete process experienced by the sample storage cell from the beginning of use until a failure occurs.

[0132] Exemplarily, several sample memories with the same specifications and production batches can be placed in an actual business environment, and the unit state parameters such as the average erase time, average write time, average read time and original error bit number of sample storage cells (such as flash memory blocks) in the sample memories can be detected and obtained according to the program-erase cycle T=30 program-erase cycles as a unit cycle until the sample storage cells are marked as being in a faulty state.

[0133] It should be noted that, when counting the cell state parameters of the above-mentioned sample storage unit, the sub-unit state parameters corresponding to the sub-units of the sample storage unit can be counted, and then the state parameters of the sample storage unit can be determined based on the sub-unit state parameters of the sample storage unit. By way of example, in the case where the sample storage unit is a flash memory block, the sub-unit state parameters corresponding to each upper page (upper page) in the flash memory block (such as the average sub-unit read time and the average sub-unit write time) can be counted, and then the counted sub-unit state parameters can be averaged according to the number of sub-units in the flash memory block to obtain the cell state parameters of the flash memory block. For example, the average read time of each sub-unit of the flash memory block can be averaged to obtain the average read time of the flash memory block.

[0134] Sub-step 2012, determining, from the complete life cycle, a first programming-erasing cycle in which the cell state parameter set satisfies a preset state condition.

[0135] In the embodiment of the present application, the first programming and erasing cycle in which the cell state parameter set meets the preset state condition can be determined from the complete life cycle of a sample storage unit, and then multiple consecutive programming and erasing cycles corresponding to the sample storage unit can be determined from the second programming and erasing cycle before the first programming and erasing cycle that meets the preset state condition. The preset state condition can be flexibly set according to the parameter type included in the cell state parameter set, and the embodiment of the present application does not specifically limit it.

[0136] Exemplarily, when the unit state parameter set includes an average read time and an average write time, a preset state condition may be: the average read time is greater than or equal to 1 millisecond and the average write time is greater than or equal to 2 milliseconds.

[0137] As the sample storage unit is continuously used, its unit state parameter set will continue to deteriorate. When the unit state parameter set meets the preset state condition for the first time during the programming and erasing cycle, it can be considered that the sample storage unit is not stable and is more likely to fail. For data security, it is not suitable to continue using it at this time. The sample storage unit can be considered to be in a failed state. The first programming and erasing cycle in which the unit state parameter set meets the preset state condition is the first programming and erasing cycle.

[0138] Sub-step 2013, determining the plurality of consecutive program-erase cycles from the second program-erase cycle before the first program-erase cycle.

[0139] In an embodiment of the present application, after determining the first programming-erasure cycle in the complete life cycle of the sample storage unit, multiple consecutive second programming-erasure cycles can be selected from all second programming-erasure cycles before the first programming-erasure cycle in the complete life cycle to obtain multiple consecutive programming-erasure cycles of the sample storage unit. Accordingly, the fault label value corresponding to the consecutive programming-erasure cycle can be determined based on the relationship between the programming-erasure cycle predicted by the time series prediction model and the above-mentioned first programming-erasure cycle. If the programming-erasure cycle predicted by the time series prediction model is the first programming-erasure cycle or the programming-erasure cycle after the first programming-erasure cycle, the fault label value corresponding to the consecutive programming-erasure cycle is 1. If the programming-erasure cycle predicted by the time series prediction model is the programming-erasure cycle before the first programming-erasure cycle, the fault label value corresponding to the consecutive programming-erasure cycle is 0.

[0140] In the embodiment of the present application, the cell state parameter set corresponding to each programming and erasing cycle of the sample storage unit in the complete life cycle can be obtained; the first programming and erasing cycle in which the cell state parameter set meets the preset state condition is determined from the complete life cycle; and multiple consecutive programming and erasing cycles are determined from the second programming and erasing cycle before the first programming and erasing cycle. It is possible to determine whether the sample storage unit is in a faulty state based on the cell state parameter set, so that it can still be determined to be in a faulty state when the sample storage unit has not actually failed, so that the trained model can predict the probability of potential failures that may exist in the storage unit, further reducing the probability of missing potential failures, and helping to further improve data storage security.

[0141] Step 202, normalize the first state parameter and the second state parameter corresponding to the programming and erasing cycle respectively to obtain a first normalized result corresponding to the first state parameter and a second normalized result corresponding to the second state parameter; wherein the normalization intervals of the first normalized result and the second normalized result are the same.

[0142] In the embodiment of the present application, the unit state parameter set may include multiple unit state parameters, for example, may include a first state parameter and a second state parameter, wherein the parameter types of the first state parameter and the second state parameter are different. In addition, the unit state parameter set may also include a third state parameter, a fourth state parameter, etc. of different parameter types, which will not be described in detail in the embodiment of the present application.

[0143] The same normalization interval can be used to normalize the first state parameters and the second state parameters corresponding to the programming and erasing cycles, respectively, to obtain a first normalized result corresponding to the first state parameter and a second normalized result corresponding to the second state parameter, so that the normalized results of different parameter types can have a unified scale of calculation specifications, which is conducive to improving the convergence speed of the model during the training process.

[0144] Exemplarily, the first state parameter may be an average read time, and the second state parameter may be a raw number of error bits. Both the average read time and the raw number of error bits may be normalized to a normalization interval (0, 1).

[0145] Optionally, due to sensor errors and other reasons, abnormal data may exist in the above-mentioned unit state parameters. In order to reduce the impact of abnormal data on model training, the unit state parameter set may also be filtered or corrected for abnormal data.

[0146] Specifically, in one embodiment, an abnormal state parameter may be determined from a set of unit state parameters, for example, a first abnormal state parameter may be determined from a first state parameter, a second abnormal state parameter may be determined from a second state parameter, and then the abnormal state parameter may be removed from the set of unit state parameters. The abnormal state parameter may represent a unit state parameter whose parameter value exceeds a parameter value limit interval of a corresponding parameter category, and different parameter categories may correspond to different parameter value limit intervals.

[0147] In another embodiment, the abnormal state parameter can be corrected based on the adjacent state parameter of the abnormal state parameter, wherein the adjacent state parameter of the abnormal state parameter can represent the cell state parameter of the same parameter type in the cell state parameter set of one or more consecutive programming and erasing cycles adjacent to the programming and erasing cycle corresponding to the abnormal state parameter. Specifically, the abnormal state parameter can be replaced according to the mean value of the adjacent state parameter of the abnormal state parameter, so as to ensure the integrity of the generated sample feature sequence and improve the model training effect.

[0148] Step 203, splicing the first normalized result and the second normalized result to obtain a time step feature corresponding to the program-erase cycle; wherein the first normalized result and the second normalized result are spliced ​​according to a preset splicing order.

[0149] In an embodiment of the present application, the first normalized result and the second normalized result corresponding to the same programming and erasing cycle can be spliced ​​to obtain the time step feature corresponding to the programming and erasing cycle. Among them, the time step feature represents the corresponding feature at each time step (Time Step) in the time series analysis. In an embodiment of the present application, each programming and erasing cycle can represent a time step. Exemplarily, if the first normalized result corresponding to a programming and erasing cycle is 0.5 and the second normalized result is 0.8, the time step feature (0.5, 0.8) can be obtained after splicing. Among them, the expression form of the time step feature may include but is not limited to a numerical sequence, a vector, etc., and the embodiment of the present application is not specifically limited. It should be noted that the first normalized result and the second normalized result can be spliced ​​in a preset splicing order to ensure that in the time step features corresponding to each programming and erasing cycle, the arrangement order of the parameter types corresponding to each normalized result is the same.

[0150] Optionally, step 203 may include:

[0151] Sub-step 2031, performing feature crossover on the first normalized result and the second normalized result to obtain a crossover feature corresponding to the program-erase cycle; wherein the feature crossover includes at least one of product crossover, additive crossover and polynomial crossover.

[0152] In the embodiment of the present application, the first normalized result and the second normalized result corresponding to the same program-erase cycle can be cross-characterized to obtain the cross-characteristics corresponding to the program-erase cycle. The cross-characteristics may include but are not limited to product cross-characterization, addition cross-characterization, polynomial cross-characterization, etc., which are not specifically limited in the embodiment of the present application.

[0153] Sub-step 2032, concatenating the first normalized result, the second normalized result and the cross feature to obtain a time step feature corresponding to the program-erase cycle.

[0154] In the embodiment of the present application, the first normalized result, the second normalized result, and the crossover feature corresponding to the same program-erase cycle can be spliced ​​to obtain the time step feature corresponding to the program-erase cycle. For example, if the first normalized result corresponding to a program-erase cycle is 0.5 and the second normalized result is 0.8, the first normalized result and the second normalized result are firstly cross-producted to obtain the crossover feature 0.4, and then the first normalized result, the second normalized result, and the crossover feature are spliced ​​to obtain the time step feature (0.5, 0.8, 0.4) corresponding to the program-erase cycle.

[0155] In the embodiment of the present application, more complex interactions between features can be captured through feature crossover, and new crossover features can be generated for the programming and erasing cycles, thereby improving the feature dimension of the time step feature, providing more useful information for the model, and thus improving the accuracy of the model.

[0156] Step 204 , based on the time sequence of the programming and erasing cycles, the time step features corresponding to the respective programming and erasing cycles are spliced ​​together to obtain the sample feature sequence.

[0157] After obtaining the time step features corresponding to each programming and erasing cycle, the time step features corresponding to each programming and erasing cycle can be spliced ​​based on the time sequence of the programming and erasing cycles to obtain a sample feature sequence. This allows the sample feature sequence to contain features of multiple feature types, and the features contained therein are in the same normalized interval, making it easier for the model to capture the correlation between features of different feature types, which helps to improve the accuracy of the model.

[0158] Step 205: input the sample feature sequence into a time series prediction model to obtain a target result output by the time series prediction model.

[0159] This step can refer to step 103, and will not be described in detail in this embodiment of the present application.

[0160] Optionally, the time series prediction model includes a feature dimensionality enhancement layer and a time series prediction network layer, the feature dimensionality enhancement layer is used to enhance the feature dimension of the sample feature sequence and output a high-dimensional feature sequence corresponding to the sample feature sequence; the time series prediction network layer is used to output the target result according to the high-dimensional feature sequence; wherein the feature dimension of the high-dimensional feature sequence is greater than the feature dimension of the sample feature sequence.

[0161] In an embodiment of the present application, the time series prediction model may include a feature dimensionality increase layer and a time series prediction network layer, wherein the feature dimensionality increase layer can be used to increase the dimension of the sample feature sequence and output a high-dimensional feature sequence, and the time series prediction network layer can be used to output the target result according to the high-dimensional feature sequence, and the feature dimension of the above-mentioned high-dimensional feature sequence is greater than the feature dimension of the sample feature sequence.

[0162] Specifically, the feature dimension enhancement layer can enhance the dimension of each time step feature contained in the sample feature sequence to obtain the time step high-dimensional feature corresponding to each time step feature, and then combine the time step high-dimensional features in the order of the time step features corresponding to the time step high-dimensional features to obtain the high-dimensional feature sequence corresponding to the sample feature sequence.

[0163] For example, if the length of the sample feature sequence is L, and the feature dimension of each time step feature is 4, the sample feature sequence of the input feature dimension-raising layer can be expressed as , the high-dimensional feature sequence output by the feature dimension-raising layer can be expressed as , where D>4. It should be noted that the feature dimension of the high-dimensional feature sequence can be a preset dimension, such as 7 dimensions; the feature dimension of the high-dimensional feature sequence can also be the sum of the feature dimension of the sample feature sequence and the lifting dimension. For example, when the feature dimension of the sample feature sequence is 4 and the lifting dimension is 2, the feature dimension of the high-dimensional feature sequence can be 4+2=6.

[0164] The above-mentioned feature dimension-raising layer may include but is not limited to a feature crossover layer, a feature transformation layer, etc., so that feature crossover or feature transformation can be performed on the feature values ​​in each time step feature in the sample feature sequence, and a new extended feature value is generated for each time step feature, and the extended feature value is spliced ​​into the time step feature to obtain the time step dimension-raising feature, thereby achieving dimensionality improvement of the sample feature sequence. Among them, the above-mentioned feature crossover may include but is not limited to product crossover, additive crossover or polynomial crossover, etc., and the above-mentioned feature transformation may include but is not limited to feature linear transformation, feature nonlinear transformation, feature discrete transformation, etc., which are not specifically limited in the embodiments of the present application.

[0165] In an embodiment of the present application, the output of the feature dimensionality upgrading layer can be connected to the input of the time series prediction network layer, and the high-dimensional feature sequence output by the feature dimensionality upgrading layer is input into the time series prediction network layer for prediction, which helps to improve the accuracy of the target result output by the time series prediction network layer.

[0166] Optionally, in an embodiment of the present application, in order to improve the effect of feature dimensionality increase, a feature dimensionality increase layer can be constructed based on a neural network, and a convolutional neural network (CNN) can be used to construct the above feature dimensionality increase layer. In order to improve the prediction accuracy, a long short-term memory network can be used to construct the above time series prediction network layer. The convolutional neural network can automatically extract hierarchical features from the data through multi-layer convolution operations. This automated feature extraction method reduces manual intervention and makes feature dimensionality increase more flexible and efficient.

[0167] Reference Figure 3 , Figure 3 A time series prediction model structure diagram provided by an embodiment of the present application is shown in FIG. Figure 3 As shown, the time series prediction model may include a convolutional neural network and a long short-term memory network, wherein the convolutional neural network receives a sample feature sequence as a model input, and the long short-term memory network outputs a target result.

[0168] Optionally, the convolutional neural network includes multiple hidden layers, each of which is composed of a first number of one-dimensional convolution kernels. The multiple hidden layers are used to perform feature dimension upgrading on the sample feature sequence and output the high-dimensional feature sequence with the feature dimension being the first number.

[0169] Exemplarily, the input data format of the convolutional neural network can be , where N represents the number of sample feature sequences in a single batch of training (batch size), L represents the number of time step features or sequence length of each sample feature sequence, and 4 represents the number of feature dimensions of each time step feature. A convolutional neural network can include an input layer, a hidden layer, and an output layer. The number of hidden layers can be multiple, and each hidden layer contains a set of D convolution kernels, and the size of each convolution kernel is , the input data is processed by multiple hidden layers, each of which can extract data features through convolution operations. In the hidden layer, one-dimensional convolution (Conv1D) operations can be used to process the input data. The main tasks of these convolution kernels are to extract feature information from two aspects: one is to capture the potential relationship between adjacent time steps, that is, to identify the correlation of continuous data points in the time dimension; the other is to analyze the internal relationship between different features in the same time step, that is, to understand the interaction between different types of unit state parameters.

[0170] Through the layer-by-layer processing of multiple hidden layers, these convolution kernels gradually merge and extract the deep features of the unit state parameters, and finally output the feature data after dimensionality increase. At this time, the format of the output data of the output layer can be , where D is the number of feature dimensions obtained by the convolution operation. The number of convolution kernels D contained in each hidden layer can be flexibly set according to the number of feature dimensions required for output, and is not specifically limited in the embodiments of the present application.

[0171] The above-mentioned convolutional neural network can automatically extract and fuse the temporal relationship between adjacent time steps and the correlation between features of different dimensions from the sample feature sequence. Through layer-by-layer convolution operations, the convolutional neural network can map the low-dimensional sample feature sequence to a higher-dimensional feature space, thereby enhancing the data's expressive power. It can provide richer information for the subsequent time series prediction network layer, allowing the time series prediction network layer to more effectively capture the dependencies in the time series, thereby improving the overall performance of the model.

[0172] Optionally, the time series prediction network layer includes a multi-layer long short-term memory network, the long short-term memory network includes a plurality of cells, and the cells are composed of a forget gate, an input gate and an output gate.

[0173] In an embodiment of the present application, the time series prediction network layer may be composed of multiple long short-term memory network layers, for example, 2 layers. Each long short-term memory network layer may include multiple long short-term memory networks, and each long short-term memory network may include multiple cells.

[0174] Reference Figure 4 , Figure 4 A schematic diagram of a long short-term memory network provided by an embodiment of the present application is shown. Figure 4 As shown, the long short-term memory network can include 3 cells: cell 1, cell 2 and cell 3. The inputs of cells 1 to 3 are , , The outputs are , , .in, Represents the time step characteristics at time t (tth programming and erasing cycle) in the input sequence, and each cell has a corresponding cell state ,A single cell can be composed of a forget gate, an input gate, and an output gate.,Relying on the memory relationship between the three gates, it can realize the memory of long-distance information and obtain the feature dependency in multiple consecutive programming and erasing cycles.,Through the above model structure, it is possible to accurately capture the timing characteristics of the input data and improve the model prediction accuracy.

[0175] Specifically, the forget gate can be used to control which information in the cell state needs to be forgotten, that is, to remove invalid redundant features so that the cell state Update reasonably during the update flow. and the output of the previous cell As input parameters, the corresponding forget gate parameters are obtained after sigmoid activation. The specific formula of this process is as follows:

[0176]

[0177] in, represents the sigmoid activation function, Indicates the current input Output with the previous moment The weight in the forget gate, Indicates the offset of the forget gate.

[0178] The input gate can use the input features to update the cell state , decides whether to write the feature into the memory unit. First, the long short-term memory network passes an input gate parameter To select the updated information, the input gate parameter It can be obtained through the sigmoid activation function. Then generate a candidate value , and perform nonlinear mapping on it through the tanh activation function. Finally, and Multiply to remember the current input feature information, and forget the above gate and the previous cell state Weighing is performed to obtain historical dependency characteristic information, and the cell state is updated as .

[0179] The output gate can weigh the characteristic information in the cell state, thereby determining which parts are important for the output of the current sequence step, and is responsible for controlling the amount of characteristic information output by the current cell to the outside. The calculation of the output gate can involve two steps. First, the activation weight of the output gate is calculated through the sigmoid activation function. , and then the cell state Through the tanh activation function, it is mapped to the interval [-1,1] Perform point multiplication with the activated cell state to obtain the final output result .

[0180] Step 206, calculate the target loss value according to the target result and the fault label value corresponding to the sample storage unit, and adjust the model parameters of the time series prediction model based on the target loss value to obtain a target neural network model; wherein the target neural network model is used to output the failure probability of the target storage unit according to the target feature sequence of the target storage unit.

[0181] Optionally, in an embodiment of the present application, the target result includes the sub-probabilities of failures corresponding to the target number of programming and erasure cycles to be predicted. That is, the target result may include the sub-probability of failures corresponding to one programming and erasure cycle to be predicted, or may include the sub-probabilities of failures corresponding to multiple programming and erasure cycles to be predicted. The above target number of programming and erasure cycles to be predicted may be continuous or discrete. The above target number may be flexibly set by the technician according to actual business needs. For example, the above target number may be 5, which is not specifically limited in the embodiment of the present application.

[0182] In an embodiment of the present application, the failure probability of a storage unit in multiple future predicted programming and erasing cycles can be predicted, so that a user can promptly discover the changing trend of the failure probability of the storage unit and can more accurately determine the time point when the storage unit may fail, thereby improving the accuracy of the user's judgment on the life of the storage unit.

[0183] Optionally, the time series prediction model includes a fully connected layer consisting of a normalization layer and the target number of neurons, the neurons are used to output the linear transformation results corresponding to the programming and erasure cycle to be predicted, and the normalization layer is used to calculate the target natural index of the linear transformation results corresponding to the target programming and erasure cycle to be predicted, and the ratio of the target natural index of the linear transformation results corresponding to each programming and erasure cycle to be predicted to obtain the failure sub-probability corresponding to the target programming and erasure cycle to be predicted.

[0184] In an embodiment of the present application, a fully connected layer can be set at the last output layer of the above-mentioned time series prediction model (e.g., a multi-layer long short-term memory network), and the fully connected layer includes a normalization layer and a target number of neurons, wherein the target number of neurons has a one-to-one correspondence with the target number of programming and erasing cycles to be predicted, and the fully connected layer can convert the final features extracted from the time series prediction model into a failure probability. Taking the time series prediction model as a multi-layer long short-term memory network as an example, the fully connected layer can convert the output of the previous layer of the long short-term memory network into Scale, where is the number of programming and erasing cycles to be predicted, that is, the fully connected layer can be set neurons, each neuron can correspond to a set of weights and biases, through The group weights and biases respectively perform linear transformation on the output of the previous layer of long short-term memory network, and the output corresponds to the failure probability at the programming and erasing cycle to be predicted.

[0185] In the embodiment of the present application, the fully connected layer can use the softmax function to output the probability, and the calculation formula of the softmax function can be as follows:

[0186]

[0187] in, Indicates The linear transformation result corresponding to the program-erase cycle to be predicted is: Indicates The failure probability corresponding to the program-erase cycle to be predicted is, is the number of neurons, that is, the length of the probability sequence contained in the model output , Represents a natural constant.

[0188] In an embodiment of the present application, the time series prediction model includes a fully connected layer consisting of a normalization layer and a target number of neurons, the neurons are used to output the linear transformation results corresponding to the programming and erasure cycles to be predicted, and the normalization layer is used to calculate the target natural index of the linear transformation results corresponding to the target programming and erasure cycles to be predicted, and the ratio between the accumulated values ​​of the natural indexes of the linear transformation results corresponding to each programming and erasure cycle to be predicted, and obtain the fault sub-probability corresponding to the target programming and erasure cycle to be predicted. By calculating the natural index of the linear transformation result and normalizing it, the output value can be smoothed, reducing the sensitivity of the model to abnormal data, and this smoothing process makes the fault sub-probability more stable and less susceptible to interference from noise or abnormal values ​​in the input data.

[0189] In summary, a model generation method provided by an embodiment of the present application includes collecting a set of cell state parameters corresponding to a sample storage unit in a plurality of consecutive programming and erasing cycles; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles; based on the time sequence of the programming and erasing cycles corresponding to the cell state parameter set, splicing the cell state parameter set to generate a sample feature sequence; inputting the sample feature sequence into a time series prediction model to obtain a target result output by the time series prediction model; calculating a target loss value according to the target result and the fault label value corresponding to the sample storage unit, and adjusting the model parameters of the time series prediction model based on the target loss value to obtain a target neural network model; wherein the target neural network model is used to output the failure probability of the target storage unit according to the target feature sequence of the target storage unit. Based on the multi-dimensional state parameters generated by the sample storage unit in a plurality of consecutive programming and erasing cycles, a sample feature sequence with time continuity can be generated, and input into the time series prediction model for model training, so that the time series prediction model can not only extract the time series characteristics of the state change of the storage unit, but also extract the correlation between the multiple state parameters of the storage unit in each time step, which is helpful to improve the accuracy of the trained target neural network model in predicting the failure of the storage unit.

[0190] Reference Figure 5 , Figure 5 A flowchart of a fault prediction method provided by an embodiment of the present application is shown. Figure 5 As shown, the specific steps include:

[0191] Step 301 : collecting cell state parameter sets corresponding to a plurality of consecutive programming and erasing cycles of a target storage cell; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles.

[0192] In an embodiment of the present application, a continuous state detection can be performed on a target storage unit that needs to be predicted for a fault, and a set of unit state parameters corresponding to a plurality of consecutive programming and erasing cycles of the target storage unit can be collected, and the above-mentioned programming and erasing cycle can include a plurality of programming and erasing cycles. The above-mentioned target storage unit represents a storage medium or storage area that can count programming and erasing cycles. Specifically, the target storage unit can include storage media such as solid-state hard disks, flash memory particles, and USB flash drives, and can also include storage areas of a specific level in the storage medium, such as flash memory pages, flash memory blocks, flash memory segments, etc., which are not specifically limited in the embodiments of the present application.

[0193] Step 302: based on the time sequence of programming and erasing cycles corresponding to the cell state parameter set, the cell state parameter set is spliced ​​to generate a target feature sequence.

[0194] After obtaining the cell state parameter sets corresponding to a plurality of consecutive programming and erasing cycles of the target storage cell, the cell state parameter sets can be spliced ​​based on the time sequence of the programming and erasing cycles corresponding to the cell state parameter sets to generate a target feature sequence.

[0195] Step 303, input the target feature sequence into the target neural network model to obtain the failure probability of the target storage unit output by the target neural network model; wherein the target neural network model is obtained based on the above-mentioned model generation method.

[0196] In an embodiment of the present application, a target neural network model can be trained and generated based on the method provided in the above model generation method embodiment. The target feature sequence of the target storage unit is input into the target neural network model to obtain the failure probability of the target storage unit output by the target neural network model in the subsequent Mth programming and erasing cycle.

[0197] Optionally, the failure probability includes failure sub-probabilities corresponding to the target number of predicted programming and erasing cycles, thereby improving the coverage time interval of the failure prediction, enabling the user to accurately determine the time point when the target storage unit may fail, and improving the accuracy of the failure prediction.

[0198] Optionally, during the use of the target neural network model, the target feature sequence is continuously collected, and prediction is performed based on the target feature sequence to obtain the corresponding failure sub-probabilities corresponding to the programming and erasing cycles to be predicted. In the embodiment of the present application, these collected target feature sequences can be fully utilized to continuously optimize the target neural network model for training. For details, refer to the following steps A1 to A3:

[0199] Step A1, collecting a set of cell state parameters corresponding to the program-erase cycle to be predicted.

[0200] When the programming and erasing cycle experienced by the target storage unit reaches the programming and erasing cycle to be predicted corresponding to the target feature sequence, the cell state parameter set corresponding to the programming and erasing cycle to be predicted can be collected. Exemplarily, if a target storage unit is experiencing the 100th programming and erasing cycle of its complete life cycle at time t0, then the target feature sequence is generated according to the cell state parameter set corresponding to its 80th to 100th programming and erasing cycle, and the target feature sequence is input into the target neural network model to obtain the next 10 consecutive programming and erasing cycles to be predicted at time t0 (101 to 110 programming and erasing cycles of its complete life cycle). At time t1, the target storage unit is experiencing the 101st programming and erasing cycle of its complete life cycle. At this time, the cell state parameter set of the 101st programming and erasing cycle can be obtained as the cell state parameter set of the above-mentioned programming and erasing cycle to be predicted, and so on, until at time t3, the target storage unit is experiencing the 110th programming and erasing cycle of its complete life cycle, and the cell state parameter sets corresponding to the above-mentioned 10 consecutive programming and erasing cycles to be predicted are obtained.

[0201] Step A2, determining a target fault label value corresponding to the target feature sequence based on the cell state parameter set corresponding to the to-be-predicted programming-erasing cycle and a preset state condition.

[0202] In an embodiment of the present application, the target fault label value corresponding to the target feature sequence can be determined based on the set of cell state parameters corresponding to the programming and erasure cycle to be predicted and the preset state conditions. Specifically, the set of cell state parameters corresponding to each programming and erasure cycle to be predicted can be compared with the preset state conditions respectively, and the fault label value corresponding to each programming and erasure cycle to be predicted can be determined according to the comparison result. If the set of cell state parameters corresponding to a programming and erasure cycle to be predicted does not meet the preset state conditions, the fault label value corresponding to the programming and erasure cycle to be predicted can be determined to be 0, otherwise it is 1. The above-mentioned target fault label value can include the fault label values ​​corresponding to each programming and erasure cycle to be predicted, so that in the subsequent training process, the fault label value of each programming and erasure cycle to be predicted and its corresponding fault sub-probability are used to jointly calculate the model loss.

[0203] Step A3: training the target neural network model based on the target feature sequence and the target fault label value.

[0204] In the embodiment of the present application, the target fault label value corresponding to the target feature sequence can be determined according to the above-mentioned set of cell state parameters corresponding to the predicted programming and erasing cycle and the preset state conditions, and a training set can be constructed to continuously train and optimize the target neural network model. The specific training process can be referred to the above-mentioned model generation method embodiment, and the embodiment of the present application will not be repeated.

[0205] By collecting the cell state parameter set corresponding to the predicted programming and erasing cycle; determining the target fault label value corresponding to the target feature sequence based on the cell state parameter set corresponding to the predicted programming and erasing cycle and the preset state conditions; and training the target neural network model based on the target feature sequence and the target fault label value. The target neural network model can be continuously optimized using the data collected during the model application, making it more suitable for the user's application environment and achieving better accuracy in the user's application environment.

[0206] Optionally, after obtaining the failure probability of the target storage unit, the working state of the target storage unit may be adjusted based on the failure probability, thereby extending the life of the target storage unit or improving the security of data in the target storage unit. For details, refer to the following steps B1 to B2:

[0207] Step B1, when the failure probability is greater than or equal to a first probability and less than a second probability, reducing the erasing frequency of the target storage unit; wherein the first probability is less than the second probability.

[0208] If the failure probability of the target storage unit is greater than the first probability and less than the second probability, it means that the target storage unit has certain hidden dangers of failure. The erase frequency of the target storage unit can be reduced to limit the workload of the target storage unit. For example, the erase frequency of the target storage unit can be limited to 1 time / hour.

[0209] It should be noted that, when the failure probability includes multiple failure sub-probabilities, the average value of all failure sub-probabilities can be calculated. When the average value is greater than or equal to the first probability and less than the second probability, the failure probability is determined to be greater than or equal to the first probability and less than the second probability. The failure sub-probability can also be compared with the first probability and the second probability to determine the first proportion of the failure sub-probabilities greater than or equal to the first probability and less than the second probability to all failure sub-probabilities. When the first proportion is greater than or equal to a preset proportion, the failure probability is determined to be greater than or equal to the first probability and less than the second probability.

[0210] Step B2: when the failure probability is greater than or equal to the second probability, migrate the cold data stored in the memory where the target storage unit is located to the target storage unit.

[0211] If the failure probability of the target storage unit is greater than the second probability, it means that the target storage unit has a short lifespan, and the cold data stored in the memory where the target storage unit is located can be migrated to the target storage unit, thereby avoiding the target storage unit from being frequently called, and further reducing the workload of the target storage unit. Among them, cold data refers to data with a low access frequency and is not frequently read or written, and the cold data in the memory where the target storage unit is located can be determined according to the access frequency of the data.

[0212] By reducing the erasing frequency of the target storage unit when the failure probability is greater than or equal to the first probability and less than the second probability, wherein the first probability is less than the second probability, and migrating the cold data stored in the memory where the target storage unit is located to the target storage unit when the failure probability is greater than or equal to the second probability, the operating frequency of the target storage unit can be reduced according to the failure probability of the target storage unit, thereby extending the service life of the target storage unit.

[0213] Optionally, step B2 may include:

[0214] Sub-step B21, when the failure probability is less than a third probability, migrating the cold data stored in the memory where the target storage unit is located to the target storage unit; wherein the third probability is greater than the second probability.

[0215] In an embodiment of the present application, a third probability greater than the above-mentioned second probability can also be set. If the failure probability is less than the third probability, that is, the failure probability is greater than or equal to the second probability and less than the third probability, the cold data stored in the target storage unit memory can be migrated to the target storage unit.

[0216] Furthermore, target cold data can be matched to the target storage unit from all cold data according to the access frequency of the cold data and the failure probability of the target storage unit, so that the higher the failure probability of the target storage unit, the lower the access frequency of the matched target cold data, thereby further extending the life of the target storage unit.

[0217] Sub-step B22, when the failure probability is greater than or equal to the third probability, migrating the data stored in the target storage unit to other storage units.

[0218] In an embodiment of the present application, if the failure probability of the target storage unit is greater than or equal to the third probability, it can be determined that the target storage unit is highly dangerous and may be damaged at any time, resulting in data loss. At this time, the data stored in the target storage unit can be migrated to other storage units to improve data security.

[0219] By migrating the cold data stored in the storage where the target storage unit is located to the target storage unit when the failure probability is less than the third probability, and migrating the data stored in the target storage unit to other storage units when the failure probability is greater than or equal to the third probability, the data of the target storage unit can be migrated when the failure probability of the target storage unit is high, thereby improving data security.

[0220] In summary, a fault prediction method provided by an embodiment of the present application includes collecting a set of cell state parameters corresponding to a sample storage unit in a plurality of consecutive programming and erasing cycles; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles; based on the time sequence of the programming and erasing cycles corresponding to the cell state parameter set, splicing the cell state parameter set to generate a sample feature sequence; inputting the sample feature sequence into a time series prediction model to obtain a target result output by the time series prediction model; calculating a target loss value according to the target result and the fault label value corresponding to the sample storage unit, and adjusting the model parameters of the time series prediction model based on the target loss value to obtain a target neural network model; wherein the target neural network model is used to output the fault probability of the target storage unit according to the target feature sequence of the target storage unit. The accuracy of fault prediction for storage units can be improved.

[0221] Reference Figure 6 , Figure 6 is a block diagram of a model generation device provided in an embodiment of the present application, such as Figure 6 As shown, the model generating device 60 includes:

[0222] The acquisition module 61 is used to acquire a set of cell state parameters corresponding to a plurality of consecutive programming and erasing cycles of a sample storage cell; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles;

[0223] A splicing module 62, configured to splice the cell state parameter set based on the time sequence of the programming and erasing cycles corresponding to the cell state parameter set to generate a sample feature sequence;

[0224] An input module 63 is used to input the sample feature sequence into a time series prediction model to obtain a target result output by the time series prediction model;

[0225] The training module 64 is used to calculate the target loss value according to the target result and the fault label value corresponding to the sample storage unit, and adjust the model parameters of the time series prediction model based on the target loss value to obtain a target neural network model; wherein the target neural network model is used to output the failure probability of the target storage unit according to the target feature sequence of the target storage unit.

[0226] Optionally, the device further comprises:

[0227] A set acquisition module, used to acquire a set of cell state parameters corresponding to each programming and erasing cycle of the sample storage cell in a complete life cycle;

[0228] A first programming and erasing cycle module, configured to determine, from the complete life cycle, a first programming and erasing cycle in which the cell state parameter set satisfies a preset state condition;

[0229] The multiple consecutive program-erase cycle determination module is used to determine the multiple consecutive program-erase cycles from a second program-erase cycle before the first program-erase cycle.

[0230] Optionally, the unit state parameter set includes a first state parameter and a second state parameter, and the splicing module includes:

[0231] A normalization submodule, used to normalize the first state parameter and the second state parameter corresponding to the program-erase cycle respectively, to obtain a first normalization result corresponding to the first state parameter and a second normalization result corresponding to the second state parameter; wherein the normalization intervals of the first normalization result and the second normalization result are the same;

[0232] A result splicing submodule, used for splicing the first normalized result and the second normalized result to obtain a time step feature corresponding to the programming and erasing cycle;

[0233] The sample feature sequence splicing submodule is used to splice the time step features corresponding to each of the programming and erasing cycles based on the time sequence of the programming and erasing cycles to obtain the sample feature sequence.

[0234] Optionally, the result splicing submodule includes:

[0235] A feature crossover unit, configured to perform feature crossover on the first normalized result and the second normalized result to obtain a crossover feature corresponding to the program-erase cycle; wherein the feature crossover includes at least one of a product crossover, an additive crossover and a polynomial crossover;

[0236] A time step feature unit is used to concatenate the first normalization result, the second normalization result and the cross feature to obtain a time step feature corresponding to the programming and erasing cycle.

[0237] Optionally, the time series prediction model includes a feature dimensionality enhancement layer and a time series prediction network layer, the feature dimensionality enhancement layer is used to enhance the feature dimension of the sample feature sequence and output a high-dimensional feature sequence corresponding to the sample feature sequence; the time series prediction network layer is used to output the target result according to the high-dimensional feature sequence; wherein the feature dimension of the high-dimensional feature sequence is greater than the feature dimension of the sample feature sequence.

[0238] Optionally, the feature dimension enhancement layer is constructed based on a convolutional neural network, and / or the time series prediction network layer is constructed based on a long short-term memory network.

[0239] Optionally, the convolutional neural network includes multiple hidden layers, each of which is composed of a first number of one-dimensional convolution kernels. The multiple hidden layers are used to perform feature dimension upgrading on the sample feature sequence and output the high-dimensional feature sequence with the feature dimension being the first number.

[0240] Optionally, the time series prediction network layer includes a multi-layer long short-term memory network, the long short-term memory network includes a plurality of cells, and the cells are composed of a forget gate, an input gate and an output gate.

[0241] Optionally, the target result includes failure sub-probabilities corresponding to target numbers of program-erase cycles to be predicted.

[0242] Optionally, the time series prediction model includes a fully connected layer consisting of a normalization layer and the target number of neurons, the neurons are used to output the linear transformation results corresponding to the programming and erasure cycle to be predicted, and the normalization layer is used to calculate the target natural index of the linear transformation results corresponding to the target programming and erasure cycle to be predicted, and the ratio of the target natural index of the linear transformation results corresponding to each programming and erasure cycle to be predicted to obtain the failure sub-probability corresponding to the target programming and erasure cycle to be predicted.

[0243] In summary, a model generation device provided in an embodiment of the present application can generate a sample feature sequence with time continuity based on the multi-dimensional state parameters generated by the sample storage unit in multiple consecutive programming and erasing cycles, and input the sample feature sequence into a time series prediction model for model training. This not only enables the time series prediction model to extract the timing characteristics of the storage unit state changes, but also can extract the correlation between the multiple state parameters of the storage unit in each time step, which helps to improve the accuracy of the trained target neural network model in predicting faults of the storage unit.

[0244] Reference Figure 7 , Figure 7 is a block diagram of a fault prediction device provided in an embodiment of the present application, such as Figure 7 As shown, the fault prediction device 70 includes:

[0245] The target acquisition module 71 is used to acquire a set of cell state parameters corresponding to a target storage cell in a plurality of consecutive programming and erasing cycles; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles;

[0246] A target feature sequence module 72, configured to concatenate the cell state parameter set based on the time sequence of programming and erasing cycles corresponding to the cell state parameter set to generate a target feature sequence;

[0247] The target input module 73 is used to input the target feature sequence into the target neural network model to obtain the failure probability of the target storage unit output by the target neural network model; wherein the target neural network model is obtained based on the above-mentioned model generation method.

[0248] Optionally, the failure probability includes failure sub-probabilities corresponding to a target number of program-erase cycles to be predicted.

[0249] Optionally, the device further comprises:

[0250] A cell state parameter set submodule, used for collecting a cell state parameter set corresponding to the programming and erasing cycle to be predicted;

[0251] A target fault label value submodule, used to determine a target fault label value corresponding to the target feature sequence based on a set of cell state parameters corresponding to the program-erase cycle to be predicted and a preset state condition;

[0252] The model optimization submodule is used to train the target neural network model based on the target feature sequence and the target fault label value.

[0253] Optionally, the device further comprises:

[0254] An erasing frequency module, used to reduce the erasing frequency of the target storage unit when the failure probability is greater than or equal to a first probability and less than a second probability; wherein the first probability is less than the second probability;

[0255] A cold data migration module is used to migrate the cold data stored in the memory where the target storage unit is located to the target storage unit when the failure probability is greater than or equal to the second probability.

[0256] Optionally, the cold data migration module includes:

[0257] A first migration submodule, configured to migrate the cold data stored in the memory where the target storage unit is located to the target storage unit when the failure probability is less than a third probability; wherein the third probability is greater than the second probability;

[0258] The second migration submodule is used to migrate the data stored in the target storage unit to other storage units when the failure probability is greater than or equal to the third probability.

[0259] In summary, a fault prediction device provided by an embodiment of the present application includes a target acquisition module for acquiring a set of cell state parameters corresponding to a target storage unit in a plurality of consecutive programming and erasing cycles; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles; a target feature sequence module for splicing a set of cell state parameters based on the time sequence of the programming and erasing cycles corresponding to the set of cell state parameters to generate a target feature sequence; a target input module for inputting the target feature sequence into a target neural network model to obtain the fault probability of the target storage unit output by the target neural network model; wherein the target neural network model is obtained based on the above-mentioned model generation method. It can improve the accuracy of fault prediction for storage units.

[0260] The model generation device or fault prediction device in the embodiment of the present application can be an electronic device, or a component in the electronic device, such as an integrated circuit or a chip. The electronic device can be a terminal, or it can be other devices other than a terminal. Exemplarily, the electronic device can be a GPU BOX, a mobile phone, a tablet computer, a laptop computer, a PDA, a vehicle-mounted electronic device, a mobile Internet device (Mobile Internet Device, MID), an augmented reality (augmented reality, AR) / virtual reality (virtual reality, VR) device, a robot, a wearable device, an ultra-mobile personal computer (ultra-mobile personal computer, UMPC), a netbook or a personal digital assistant (personal digital assistant, PDA), etc. It can also be a server, a network attached storage (Network Attached Storage, NAS), a personal computer (personal computer, PC), a television (television, TV), a teller machine or a self-service machine, etc., which is not specifically limited in the embodiment of the present application.

[0261] The model generation device or fault prediction device in the embodiment of the present application may be a device having an operating system. The operating system may be an Android operating system, a Linux operating system, a Windows operating system, etc., or other possible operating systems, which are not specifically limited in the embodiment of the present application.

[0262] The model generation device provided in the embodiment of the present application can achieve Figures 1 to 4 The fault prediction device provided in the embodiment of the present application can realize the various processes implemented by the method embodiment. Figure 5 To avoid repetition, the various processes implemented by the method embodiment are not described here.

[0263] Alternatively, if Figure 8As shown, an embodiment of the present application also provides an electronic device M00, including a processor M01 and a memory M02, wherein the memory M02 stores programs or instructions that can be executed on the processor M01, and when the program or instructions are executed by the processor M01, the above-mentioned embodiment or each step of the above-mentioned embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0264] It should be noted that the electronic devices in the embodiments of the present application include the mobile electronic devices and non-mobile electronic devices mentioned above.

[0265] An embodiment of the present application also provides a computer-readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the above embodiment or each process of the above embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0266] The processor is the processor in the electronic device described in the above embodiment. The computer readable storage medium includes a computer read-only memory ROM, a random access memory RAM, a magnetic disk or an optical disk.

[0267] An embodiment of the present application further provides a chip, which includes a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the above embodiment or each process of the above embodiment, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0268] It should be understood that the chip mentioned in the embodiments of the present application can also be called a system-level chip, a system chip, a chip system or a system-on-chip chip, etc.

[0269] An embodiment of the present application provides a computer program product, including a computer program or instructions. When the computer program or instructions are executed by a processor, the above embodiment or each process of the above embodiment is implemented and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0270] It should be noted that, in this article, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0271] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the relevant technology, can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, disk, CD), and includes a number of instructions for a terminal (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0272] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A model generation method, characterized in that: The method comprises: Collecting a set of cell state parameters corresponding to a plurality of consecutive programming and erasing cycles of a sample storage cell; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles; Based on the time sequence of programming and erasing cycles corresponding to the cell state parameter set, the cell state parameter set is spliced ​​to generate a sample feature sequence; Inputting the sample feature sequence into a time series prediction model to obtain a target result output by the time series prediction model; Calculating a target loss value according to the target result and the fault label value corresponding to the sample storage unit, and adjusting the model parameters of the time series prediction model based on the target loss value to obtain a target neural network model; wherein the target neural network model is used to output the failure probability of the target storage unit according to the target feature sequence of the target storage unit; The time series prediction model includes a feature dimension enhancement layer and a time series prediction network layer, wherein the feature dimension enhancement layer is used to enhance the feature dimension of the sample feature sequence and output a high-dimensional feature sequence corresponding to the sample feature sequence; the time series prediction network layer is used to output the target result according to the high-dimensional feature sequence; wherein the feature dimension of the high-dimensional feature sequence is greater than the feature dimension of the sample feature sequence; The method further comprises: Acquire a set of cell state parameters corresponding to each programming and erasing cycle of the sample storage cell in a complete life cycle; Determine a first programming and erasing cycle in which the cell state parameter set satisfies a preset state condition from the complete life cycle; The plurality of consecutive program-erase cycles are determined from a second program-erase cycle preceding the first program-erase cycle.

2. The method according to claim 1, characterized in that The cell state parameter set includes a first state parameter and a second state parameter, and the cell state parameter set is spliced ​​based on the time sequence of programming and erasing cycles corresponding to the cell state parameter set to generate a sample feature sequence, including: Normalizing the first state parameter and the second state parameter corresponding to the program-erase cycle respectively to obtain a first normalized result corresponding to the first state parameter and a second normalized result corresponding to the second state parameter; wherein the normalization intervals of the first normalized result and the second normalized result are the same; Splicing the first normalized result and the second normalized result to obtain a time step feature corresponding to the program-erase cycle; wherein the first normalized result and the second normalized result are spliced ​​according to a preset splicing order; Based on the time sequence of the programming and erasing cycles, the time step features corresponding to the respective programming and erasing cycles are concatenated to obtain the sample feature sequence.

3. The method according to claim 2, characterized in that The splicing of the first normalized result and the second normalized result to obtain the time step characteristics corresponding to the program-erase cycle includes: Performing feature crossover on the first normalized result and the second normalized result to obtain a crossover feature corresponding to the program-erase cycle; wherein the feature crossover includes at least one of a product crossover, an additive crossover, and a polynomial crossover; The first normalization result, the second normalization result and the cross feature are concatenated to obtain a time step feature corresponding to the program-erase cycle.

4. The method according to claim 3, characterized in that The feature dimension-enhancing layer is constructed based on a convolutional neural network, and / or the time series prediction network layer is constructed based on a long short-term memory network.

5. The method according to claim 4, characterized in that The convolutional neural network includes multiple hidden layers, each of which is composed of a first number of one-dimensional convolution kernels. The multiple hidden layers are used to perform feature dimension upgrade on the sample feature sequence and output the high-dimensional feature sequence with the feature dimension being the first number.

6. The method according to claim 4, characterized in that The time series prediction network layer includes a multi-layer long short-term memory network, and the long short-term memory network includes a plurality of cells, and the cells are composed of a forget gate, an input gate and an output gate.

7. The method according to claim 1, characterized in that The target result includes failure sub-probabilities corresponding to target numbers of program-erase cycles to be predicted.

8. The method according to claim 7, characterized in that The time series prediction model includes a normalization layer and a fully connected layer consisting of the target number of neurons, the neurons are used to output the linear transformation results corresponding to the programming and erasure cycle to be predicted, and the normalization layer is used to calculate the target natural index of the linear transformation results corresponding to the target programming and erasure cycle to be predicted, and the ratio between the target natural index cumulative value of the linear transformation results corresponding to each programming and erasure cycle to be predicted, so as to obtain the failure sub-probability corresponding to the target programming and erasure cycle to be predicted.

9. A fault prediction method, characterized in that: The method comprises: Collecting a set of cell state parameters corresponding to a target storage cell in a plurality of consecutive programming and erasing cycles; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles; Based on the time sequence of programming and erasing cycles corresponding to the cell state parameter set, the cell state parameter set is spliced ​​to generate a target feature sequence; The target feature sequence is input into a target neural network model to obtain the failure probability of the target storage unit output by the target neural network model; wherein the target neural network model is obtained based on the model generation method described in any one of claims 1 to 8.

10. The method according to claim 9, characterized in that The failure probability includes failure sub-probabilities corresponding to target numbers of program-erase cycles to be predicted.

11. The method according to claim 10, characterized in that The method further comprises: Collecting a set of cell state parameters corresponding to the programming and erasing cycles to be predicted; Determine a target fault label value corresponding to the target feature sequence based on a set of cell state parameters corresponding to the program-erase cycle to be predicted and a preset state condition; The target neural network model is trained based on the target feature sequence and the target fault label value.

12. The method according to claim 9, characterized in that The method further comprises: When the failure probability is greater than or equal to a first probability and less than a second probability, reducing the erasing frequency of the target storage unit; wherein the first probability is less than the second probability; When the failure probability is greater than or equal to the second probability, the cold data stored in the memory where the target storage unit is located is migrated to the target storage unit.

13. The method according to claim 12, characterized in that The step of migrating the cold data stored in the memory where the target storage unit is located to the target storage unit includes: When the failure probability is less than a third probability, migrating the cold data stored in the memory where the target storage unit is located to the target storage unit; wherein the third probability is greater than the second probability; When the failure probability is greater than or equal to the third probability, the data stored in the target storage unit is migrated to other storage units.

14. A model generation device, characterized in that: The device comprises: A collection module, used for collecting a set of cell state parameters corresponding to a sample storage cell in a plurality of consecutive programming and erasing cycles; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles; A splicing module, used for splicing the cell state parameter set based on the time sequence of programming and erasing cycles corresponding to the cell state parameter set to generate a sample feature sequence; An input module, used to input the sample feature sequence into a time series prediction model to obtain a target result output by the time series prediction model; A training module, used to calculate a target loss value according to the target result and the fault label value corresponding to the sample storage unit, and adjust the model parameters of the time series prediction model based on the target loss value to obtain a target neural network model; wherein the target neural network model is used to output the failure probability of the target storage unit according to the target feature sequence of the target storage unit; The time series prediction model includes a feature dimension enhancement layer and a time series prediction network layer, wherein the feature dimension enhancement layer is used to enhance the feature dimension of the sample feature sequence and output a high-dimensional feature sequence corresponding to the sample feature sequence; the time series prediction network layer is used to output the target result according to the high-dimensional feature sequence; wherein the feature dimension of the high-dimensional feature sequence is greater than the feature dimension of the sample feature sequence; The device also includes: A set acquisition module, used to acquire a set of cell state parameters corresponding to each programming and erasing cycle of the sample storage cell in a complete life cycle; A first programming and erasing cycle module, configured to determine, from the complete life cycle, a first programming and erasing cycle in which the cell state parameter set satisfies a preset state condition; The multiple consecutive program-erase cycle determination module is used to determine the multiple consecutive program-erase cycles from a second program-erase cycle before the first program-erase cycle.

15. A fault prediction device, characterized in that: The device comprises: A target acquisition module, used for acquiring a set of cell state parameters corresponding to a target storage cell in a plurality of consecutive programming and erasing cycles; wherein the programming and erasing cycle includes a plurality of programming and erasing cycles; A target feature sequence module, used to splice the cell state parameter set based on the time sequence of programming and erasing cycles corresponding to the cell state parameter set to generate a target feature sequence; A target input module is used to input the target feature sequence into a target neural network model to obtain the failure probability of the target storage unit output by the target neural network model; wherein the target neural network model is obtained based on the model generation method described in any one of claims 1 to 8.

16. An electronic device, characterized in that: The electronic device includes a processor and a memory, the memory stores programs or instructions that can be run on the processor, and when the programs or instructions are executed by the processor, the model generation method described in any one of claims 1 to 8, or the fault prediction method described in any one of claims 9 to 13 is implemented.

17. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a program or instruction, and when the program or instruction is executed by a processor, the model generation method described in any one of claims 1 to 8, or the fault prediction method described in any one of claims 9 to 13 is implemented.

18. A computer program product, characterized in that The method comprises a computer program or an instruction, which, when executed by a processor, implements the model generation method according to any one of claims 1 to 8, or the fault prediction method according to any one of claims 9 to 13.

Images