A flash memory reliability characteristic parameter prediction method and system

Abstract

The present invention discloses a method and system for predicting flash memory reliability characteristic parameters, belonging to the field of memory technology. The method includes: S1, under different target operations, respectively calculating the change in the reliability characteristic parameters of each flash memory chip sample before and after executing the target operation, and then dividing the actions according to their distribution; S2, receiving the flash memory reliability status sent by the storage system using the flash memory chip to be tested; S3, based on the current prediction strategy, predicting the action corresponding to the flash memory reliability status; S4, comparing the predicted action in the previous round with the actual action to obtain a reward value, and updating the prediction strategy based on the reward value; wherein the actual action corresponds to the change in the reliability characteristic parameter obtained after the storage system executes the operation to be executed; S5, repeating S2 to S4 until the stop condition is reached. Therefore, the present invention can predict flash memory reliability characteristic parameters without the need for pre-training with a large amount of training data.

Description

Technical Field

[0001] The present invention belongs to the field of memory technology, and more specifically, relates to a method and system for predicting flash memory reliability characteristic parameters. Background Art

[0002] With the advancement of electronic information technology, digital storage has gradually replaced traditional paper media as the primary means of information storage. Non-volatile memory, as a type of storage medium, can persist information even after a power outage and is currently the primary medium for long-term information storage. In recent years, the development of non-volatile memory has been rapid, and its application has become increasingly widespread. Reliability is a key performance indicator for memory, and its reliability directly impacts data security. Therefore, research on reliability assurance methods is of great significance.

[0003] With existing flash memory storage technology, the oxide degradation and de-trapping effects of flash memory cells increase with the number of programming and erasure cycles, ultimately leading to various forms of cell failure. Furthermore, the reduction in the distance between memory cells and the introduction of a three-bit array structure have made the reliability of flash memory chips increasingly complex. Traditional reliability assurance methods are no longer able to meet the reliability requirements of flash memory chips. Improving flash memory reliability has become a key topic in current memory research. To overcome these new reliability issues, researchers have proposed machine learning-based flash memory reliability prediction methods that can predict reliability based on flash memory error characteristics. However, these methods have the following problems: they require a large amount of training data to ensure accuracy, and they consume a large amount of time and storage space to perform the predictions. Summary of the Invention

[0004] In response to the defects of the existing technology and the need for improvement, the present invention provides a flash memory reliability characteristic parameter prediction method and system, which can accurately predict the flash memory reliability characteristic parameters while only spending less prediction time, thereby achieving more efficient failure prediction.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for predicting flash memory reliability characteristic parameters, comprising the following steps:

[0006] S1. Under different target operations, respectively calculate the change in reliability characteristic parameters of each flash memory chip sample before and after executing the target operation, and then divide each of the changes into multiple values ​​or ranges based on their distribution, where each value or range corresponds to an action; the reliability characteristic parameters include one or more of the following: programming time, read time, erase time, programming current, read current, erase current, power, and raw bit error rate; and the target operations include one or more of the following: programming, erasing, reading, and saving data;

[0007] S2, receiving a flash memory reliability status sent by a storage system using the flash memory chip to be tested, wherein the flash memory reliability status includes a current reliability characteristic parameter value of the flash memory chip to be tested and an operation to be performed; the operation to be performed is one of the different target operations;

[0008] S3, predicting an action corresponding to the flash memory reliability state based on the current prediction strategy;

[0009] S4, comparing the predicted action in the previous round with the actual action to obtain a reward value, and updating the prediction strategy based on the reward value; wherein the actual action corresponds to a change in the reliability characteristic parameter obtained after the storage system executes the operation to be executed;

[0010] S5, repeat S2 to S4 until the stop condition is reached.

[0011] Furthermore, in S1, each of the changes is divided into multiple values ​​or ranges according to its distribution, wherein each value or range corresponds to an action, including:

[0012] The variation amounts are divided into a plurality of values ​​or ranges according to the probability of occurrence of the variation amounts, wherein each value or range corresponds to an action.

[0013] Furthermore, in S1, each of the variation amounts is divided into a plurality of values ​​or ranges according to the probability of occurrence of each variation amount, wherein each value or range corresponds to an action, including:

[0014] Set probability thresholds P1, P2, ..., P n-1 、P n , where P1>P2>…>P n-1 >P n , n is a positive integer; the probability of the change amount being greater than the probability threshold P1, less than or equal to P1 and greater than P2, ..., less than or equal to P n-1 and greater than P n , less than or equal to P n into one action; or,

[0015] Set probability thresholds Q1, Q2, ..., Q m-1 , Q m , where Q1>Q2>…>Q m-1 >Q m , m is a positive integer; the probability of the change amount being greater than the probability threshold Q1, less than or equal to Q1 and greater than Q2, ..., less than or equal to Q m-1 and greater than Q m , less than or equal to Q m The range of variation is divided into one action.

[0016] Furthermore, the S3 includes:

[0017] Based on the current prediction strategy, the scores of each action corresponding to the flash memory reliability state are predicted, and the action with the largest score is selected as the predicted action; wherein the score of each action is the expected reward value obtained by selecting the action under the flash memory reliability state.

[0018] Furthermore, the S3 includes:

[0019] A random number is taken between 0 and 1. If the random number is less than or equal to a preset threshold, an action is randomly selected from the actions corresponding to the flash memory reliability state as the predicted action; otherwise, based on the current prediction strategy, the scores of the actions corresponding to the flash memory reliability state are predicted, and the action with the largest score is selected as the predicted action; wherein the score of each action is the expected reward value obtained by selecting the action under the flash memory reliability state.

[0020] Furthermore, the corresponding relationship between the flash memory reliability status and the predicted action is implemented in the form of a table, a function or an artificial neural network.

[0021] Furthermore, the update method of the prediction strategy is SARSA, Q-learning, dynamic programming, A3C or Monte Carlo.

[0022] In a second aspect, the present invention provides a flash memory reliability characteristic parameter prediction system, comprising: a computer-readable storage medium and a processor;

[0023] The computer-readable storage medium is used to store executable instructions;

[0024] The processor is configured to read the executable instructions stored in the computer-readable storage medium and execute the flash memory reliability characteristic parameter prediction method as described in the first aspect.

[0025] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0026] (1) The flash memory reliability characteristic parameter prediction method proposed in the present invention first calculates the change in reliability characteristic parameters of each flash memory chip sample before and after executing the target operation under different target operations, and then divides the actions corresponding to each target operation according to the distribution of the change. Then, the prediction strategy is updated by interacting with the storage system to predict the actions corresponding to the flash memory reliability state. Therefore, the present invention can predict flash memory reliability characteristic parameters without the need for large amounts of training data in advance.

[0027] (2) The correspondence between the flash memory reliability status and the predicted action in the present invention is implemented flexibly, and can be implemented through a table, a function, or an artificial neural network, which can better adapt to different application requirements.

[0028] (3) The prediction actions divided by the present invention are designed based on the flash memory error characteristics, which can more accurately predict the flash memory failure. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A flowchart of a method for predicting flash memory reliability characteristic parameters provided by an embodiment of the present invention is provided.

[0030] Figure 2 Schematic diagram of the artificial neural network structure in an embodiment of the present invention.

[0031] Figure 3 This is a flow chart of a prediction action selection strategy in an embodiment of the present invention.

[0032] Figure 4 This is a flowchart of the prediction strategy update in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0034] In the present invention, the terms "first", "second", etc. (if any) in the present invention and the drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0035] See Figure 1 , combined with Figures 2 to 4 The present invention provides a method for predicting flash memory reliability characteristic parameters. The method includes operations S1 to S5.

[0036] Operation S1, under different target operations, respectively calculate the change in the reliability characteristic parameters of each flash memory chip sample before and after executing the target operation, and then divide each of the changes into multiple values ​​or ranges according to their distribution, wherein each value or range corresponds to an action; the reliability characteristic parameters include one or more of the following: programming time, read time, erase time, programming current, read current, erase current, power and raw bit error rate; the target operation includes one or more of the following: programming, erasing, reading and saving data.

[0037] In this embodiment, each action is designed based on the reliability degradation mechanism and error characteristics of the flash memory. In this embodiment, by combining the reliability degradation mechanism to design experiments, error characteristics are obtained from the experimental data, and predictive actions are designed. The specific steps are as follows:

[0038] (1) Select a flash memory chip sample and randomly select 16 blocks from different planes of the chip. The number of each block must cover different positions, and the number of odd-numbered blocks and even-numbered blocks in the blocks selected from each plane is equal;

[0039] (2) Place the chip sample on the test equipment and perform 100 programming / erasing operations on the selected blocks. During the test, the parameters such as the flash memory chip operation time and the original bit error rate are recorded;

[0040] (3) Place the flash memory chip for 10 minutes and then place it in an incubator at 70°C for n hours. The value of n is determined based on the predicted demand. In this embodiment, the maximum value of n is 31.8 hours, which is equivalent to storing data at room temperature for one year (calculated according to the Arrhenius formula);

[0041] (4) Calculate the change in the flash memory chip operation time and raw bit error rate after erasing / programming and saving data for a period of time;

[0042] (5) Draw the distribution of the change value before and after every 100 programming / erase operations and every 2.6 hours of data storage; divide the actions corresponding to the target operation based on the change value distribution. Specifically, the change value with a high probability of occurrence is divided into one action, and the change value with a low probability of occurrence is divided into another action; for chip information with discrete values, a single value or multiple values ​​corresponds to one action; for chip information with continuous values, a range of values ​​corresponds to one action. Similarly, draw the distribution of the change value before and after different target operations such as 100, 200, 300, ..., 3000 programming / erase operations and holding for 2.6 hours, 5.2 hours, ..., 31.8 hours, and divide the actions corresponding to each target operation based on the change value distribution.

[0043] Operation S2: receiving a flash memory reliability status sent by a storage system using the flash memory chip to be tested, wherein the flash memory reliability status includes a current reliability characteristic parameter value of the flash memory chip to be tested and an operation to be performed; the operation to be performed is one of the different target operations.

[0044] It is understandable that before the storage system uses the flash memory chip under test, it is necessary to first predict the reliability characteristic parameters of the flash memory chip under test, thereby providing a reliable reference for the storage system to determine whether to use the flash memory chip under test. For example, the storage system is about to use the flash memory chip under test to perform a target operation, and the current raw bit error rate of the flash memory chip under test is 30%, and the threshold is 35%. When it is predicted that the storage system will use the flash memory chip under test to perform the target operation, the raw bit error rate will change by 6%, which is greater than the threshold, and the flash memory chip under test may fail. At this point, the storage system should decide whether to continue using the flash memory chip under test or replace it with another flash memory chip to perform the target operation.

[0045] Therefore, before the storage system uses the flash memory chip to be tested, the current reliability characteristic parameter values ​​of the flash memory chip to be tested are collected and the operations to be performed are clarified in order to predict the flash memory reliability characteristic parameters.

[0046] Operation S3: predicting an action corresponding to the flash memory reliability status based on the current prediction strategy.

[0047] Specifically, the corresponding relationship between the flash memory reliability status and the predicted action is implemented in the form of a table, a function, or an artificial neural network.

[0048] In this embodiment, taking an artificial neural network as an example, the artificial neural network type is a feedforward neural network, the number of layers is 5, each layer has 10 units, and the activation function is RELU, such as Figure 2 As shown, its input is the flash memory reliability status and the output is the score of each action.

[0049] The prediction strategy can be a greedy algorithm or an e-greedy algorithm.

[0050] Taking the greedy algorithm as an example, S3 includes: based on the current prediction strategy, predicting the scores of each action corresponding to the flash memory reliability state, and selecting the action with the largest score as the predicted action; wherein the score of each action is the expected reward value obtained by selecting the action under the flash memory reliability state.

[0051] Taking the e-greedy algorithm as an example, S3 includes: taking a random number between 0 and 1, and if the random number is less than or equal to a preset threshold, randomly selecting an action from the actions corresponding to the flash memory reliability state as the predicted action; otherwise, based on the current prediction strategy, predicting the scores of the actions corresponding to the flash memory reliability state, and selecting the action with the largest score as the predicted action; wherein the score of each action is the expected reward value obtained by selecting the action under the flash memory reliability state. In this embodiment, Figure 3 As shown, the preset threshold is 0.1.

[0052] Operation S4, comparing the last round of predicted actions with the actual actions to obtain a reward value, and updating the prediction strategy based on the reward value; wherein the actual action corresponds to a change in the reliability characteristic parameter obtained after the storage system executes the operation to be executed.

[0053] In this embodiment, the prediction action of the previous round is judged to be consistent with the actual action. If it is consistent, the reward value is positive. If it is not consistent, the reward value is 0 or negative. The specific value of the reward value is determined according to the prediction requirements.

[0054] The prediction strategy update method is a strategy update method in the field of reinforcement learning. Its goal is to maximize the expected return value under different flash memory reliability states. Possible prediction strategy update methods include but are not limited to: SARSA, Q-learning, dynamic programming, A3C, and Monte Carlo. The expected return value is calculated as follows:

[0055]

[0056] Where S t Represents the flash memory reliability status at time t, A t represents the predicted action at time t, R is the reward value, k is the number of executions, γ is the discount coefficient, 0<γ≤1.

[0057] In this embodiment, the flowchart of the prediction strategy update is as follows: Figure 4 As shown, it is used to train the artificial neural network and update the weights of the artificial neural network. The training steps are as follows:

[0058] 1. Send training samples to the artificial neural network;

[0059] 2. Calculate the loss function of the artificial neural network;

[0060] 3. Update the neural network weights according to the gradient descent direction of the loss function;

[0061] 4. Determine whether to stop training. If so, stop training; otherwise, return to step 1.

[0062] In this embodiment, the loss function is as follows:

[0063] E π [(R+γ·max a' Q(S t+1 ,A';θ)-Q(S t ,A t ;θ)) 2 ]

[0064] In the loss function formula, R is the return value, γ is the discount coefficient, 0<γ≤1. t+1,A';θ) represents the artificial neural network output corresponding to action A' under the flash memory reliability state S at time t+1, and action A' is the predicted action selected by the e-greedy strategy. Q(S t ,A t θ) is the selected prediction action A under the flash memory reliability state S at time t t The corresponding artificial neural network output. R+γ·max a' Q(S t+1 ,A';θ) can be regarded as the actual action of the simulation.

[0065] S5, repeat S2 to S4 until the stop condition is reached.

[0066] It should be noted that when the storage system no longer needs to predict the flash memory reliability characteristic parameters, it is considered that the stop condition is met.

[0067] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting flash memory reliability characteristic parameters, characterized in that: The following steps are involved: S1. Under different target operations, respectively calculate the change in reliability characteristic parameters of each flash memory chip sample before and after executing the target operation, and then divide each of the changes into multiple values ​​or ranges based on their distribution, where each value or range corresponds to an action; the reliability characteristic parameters include one or more of the following: programming time, read time, erase time, programming current, read current, erase current, power, and raw bit error rate; and the target operations include one or more of the following: programming, erasing, reading, and saving data; S2, receiving a flash memory reliability status sent by a storage system using the flash memory chip to be tested, wherein the flash memory reliability status includes a current reliability characteristic parameter value of the flash memory chip to be tested and an operation to be performed; the operation to be performed is one of the different target operations; S3, predicting an action corresponding to the flash memory reliability state based on the current prediction strategy; S4, comparing the predicted action in the previous round with the actual action to obtain a reward value. If the predicted action in the previous round is consistent with the actual action, the reward value is a positive number; if the predicted action in the previous round is inconsistent with the actual action, the reward value is 0 or a negative number. The prediction strategy is updated based on the reward value. The prediction strategy update method is a strategy update method in the field of reinforcement learning, and the purpose is to maximize the expected reward value under different flash memory reliability states; wherein the actual action corresponds to the change in the reliability characteristic parameter obtained after the storage system executes the operation to be executed; S5, repeat S2 to S4 until the stop condition is reached.

2. The flash memory reliability characteristic parameter prediction method according to claim 1, characterized in that: In S1, each of the changes is divided into multiple values ​​or ranges according to its distribution, wherein each value or range corresponds to an action, including: The variation amounts are divided into a plurality of values ​​or ranges according to the probability of occurrence of the variation amounts, wherein each value or range corresponds to an action.

3. The flash memory reliability characteristic parameter prediction method according to claim 2, characterized in that: In S1, each of the variation amounts is divided into a plurality of values ​​or ranges according to the probability of occurrence of each variation amount, wherein each value or range corresponds to an action, including: Set probability thresholds P1, P2, ..., P n-1 、P n , where P1>P2>…>P n-1 >P n , n is a positive integer; the probability of the change amount being greater than the probability threshold P1, less than or equal to P1 and greater than P2, ..., less than or equal to P n-1 and greater than P n , less than or equal to P n into one action; or, Set probability thresholds Q1, Q2, ..., Q m-1 , Q m , where Q1>Q2>…>Q m-1 >Q m , m is a positive integer; the probability of the change amount being greater than the probability threshold Q1, less than or equal to Q1 and greater than Q2, ..., less than or equal to Q m-1 and greater than Q m , less than or equal to Q m The range of variation is divided into one action.

4. The flash memory reliability characteristic parameter prediction method according to any one of claims 1 to 3, characterized in that: The S3 includes: Based on the current prediction strategy, the scores of each action corresponding to the flash memory reliability state are predicted, and the action with the largest score is selected as the predicted action; wherein the score of each action is the expected reward value obtained by selecting the action under the flash memory reliability state.

5. The flash memory reliability characteristic parameter prediction method according to any one of claims 1 to 3, characterized in that: The S3 includes: A random number is taken between 0 and 1. If the random number is less than or equal to a preset threshold, an action is randomly selected from the actions corresponding to the flash memory reliability state as the predicted action; otherwise, based on the current prediction strategy, the scores of the actions corresponding to the flash memory reliability state are predicted, and the action with the largest score is selected as the predicted action; wherein the score of each action is the expected reward value obtained by selecting the action under the flash memory reliability state.

6. The flash memory reliability characteristic parameter prediction method according to claim 1, characterized in that: The corresponding relationship between the flash memory reliability status and the predicted action is implemented in the form of a table, a function or an artificial neural network.

7. The flash memory reliability characteristic parameter prediction method according to claim 1, characterized in that: The update method of the prediction strategy is SARSA, Q-learning, dynamic programming, A3C or Monte Carlo.

8. A flash memory reliability characteristic parameter prediction system, characterized in that: include: Computer-readable storage media and processor; The computer-readable storage medium is used to store executable instructions; The processor is configured to read the executable instructions stored in the computer-readable storage medium and execute the flash memory reliability characteristic parameter prediction method according to any one of claims 1 to 7.

Images