Flash memory error information detection method and device
Through multi-dimensional modeling and dynamic subpage division, combined with adaptive pre-checking and parallel detection, the problems of low detection efficiency and insufficient accuracy in the existing technology are solved, efficient and accurate flash error detection is achieved, and storage is extended System life and performance improvements.
Patent Information
- Application Number
- CN202411112844.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-14
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-08-14
AI Technical Summary
The existing flash error information detection technology cannot flexibly adjust based on the dynamic characteristics of flash memory and the differences in data, resulting in low detection efficiency, insufficient accuracy and large system overhead.
Through multi-dimensional modeling, the initial comprehensive error prediction model and error feature library are constructed, and the adaptive pre-checking strategy and parallel error detection are assigned according to the risk rating to achieve differentiated processing and efficient detection.
Improves detection efficiency and accuracy, reduces system burden, extends the service life of flash storage systems, and improves overall performance and reliability.
Smart Images

Figure CN118642889B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of flash memory error information detection, and in particular to a flash memory error information detection method and device. Background Art
[0002] With the continuous development of 3D NAND flash memory technology, flash memory storage systems play an increasingly important role in data-intensive applications. However, with the increase in storage capacity and the shrinking of manufacturing processes, the reliability of flash memory faces severe challenges. Factors such as program / erase (P / E) cycles, data retention time, and read interference can cause charge leakage or accumulation in flash memory cells, thereby causing bit flip errors. These errors not only affect the integrity of the data, but also lead to system performance degradation and shortened life. Therefore, accurate and efficient detection and correction of these errors are crucial to ensure the reliability and performance of flash memory storage systems.
[0003] Existing flash memory error information detection technology usually uses fixed error detection and correction codes (such as LDPC codes) to uniformly process all data. Although this method is simple and direct, it ignores the dynamics and diversity of flash memory errors and cannot adopt differentiated detection strategies for errors of different types and degrees. In addition, as the capacity of flash memory increases, comprehensive error detection and correction of all data will result in significant performance overhead, especially in scenarios with frequent read operations. This not only reduces the response speed of the system, but also increases power consumption. Therefore, there is an urgent need for an error detection method that can dynamically adapt to changes in flash memory characteristics and can perform intelligent and differentiated processing based on data characteristics and error risks. Summary of the invention
[0004] The main purpose of the present invention is to solve the problems that the existing flash memory error information detection technology usually adopts a unified error detection and correction strategy, which cannot be flexibly adjusted according to the dynamic characteristics of the flash memory and the differences in data, resulting in low detection efficiency, insufficient accuracy, high system overhead and poor adaptability.
[0005] A first aspect of the present invention provides a flash memory error information detection method, the flash memory error information detection method comprising:
[0006] Perform multi-dimensional modeling on the preset flash memory parameter sample set to obtain an initial comprehensive error prediction model, an initial error feature library and a sub-page division strategy;
[0007] Performing sub-page division and quality assessment on the flash memory data according to the initial comprehensive error prediction model, the initial error feature library and the sub-page division strategy to obtain a sub-page quality score;
[0008] Perform risk rating and pre-check strategy allocation on the sub-pages according to the initial comprehensive error prediction model, the initial error feature library and the sub-page quality score, and obtain the sub-page pre-check result;
[0009] A subpage to be detected is selected according to the subpage pre-check result, and parallel error detection and LDPC-based error correction are performed on the subpage to be detected to obtain error detection and correction results.
[0010] In combination with the first aspect, in a first implementation of the first aspect of the present invention, the multi-dimensional modeling of the preset flash memory parameter sample set to obtain an initial comprehensive error prediction model, an initial error feature library and a sub-page division strategy includes:
[0011] Collecting data on the preset flash memory parameter sample set to obtain a multi-dimensional parameter sample including program / erase cycle number, data retention time, flash memory level, ambient temperature and read interference;
[0012] A mathematical model is constructed based on the multi-dimensional parameter sample to obtain a relationship equation between the original bit error rate RBER and each parameter;
[0013] Performing curve fitting on the relationship equation to obtain an initial comprehensive error prediction model;
[0014] Calculate the RBER values under different parameter combinations according to the initial comprehensive error prediction model to obtain an RBER distribution graph;
[0015] Threshold division is performed based on the RBER distribution graph to obtain an initial error feature library and a sub-page division strategy.
[0016] In combination with the first aspect, in a second implementation of the first aspect of the present invention, the relationship equation is the following formula:
[0017] ;
[0018] Wherein, RBER represents the raw bit error rate, a is the proportionality coefficient, PE is the number of program / erase cycles, b is the exponent of PE, c is the time impact factor, t is the data retention time, d is the interaction impact factor of PE and retention time, e is the level impact factor, L is the flash memory level, f is the temperature impact factor, T is the ambient temperature, g is the read disturbance impact factor, and R is the read disturbance.
[0019] In combination with the first aspect, in a third implementation of the first aspect of the present invention, the threshold division is performed based on the RBER distribution graph to obtain an initial error feature library and a sub-page division strategy, including:
[0020] Performing statistical analysis on the RBER distribution graph to obtain a cumulative distribution function of the RBER value;
[0021] According to the cumulative distribution function, calculating the 33.3% quantile and the 66.7% quantile of the RBER value, and obtaining a first RBER threshold point and a second RBER threshold point respectively;
[0022] Based on the first RBER threshold point and the second RBER threshold point, the RBER distribution is divided into a low-risk area, a medium-risk area and a high-risk area, wherein the RBER value is less than the first RBER threshold point for the low-risk area, the RBER value is between the first RBER threshold point and the second RBER threshold point for the medium-risk area, and the RBER value is greater than the second RBER threshold point for the high-risk area;
[0023] Perform feature analysis on each risk area, extract the error type, error distribution and error frequency of the area, and obtain the error pattern description of each area;
[0024] According to the RBER range of each risk area, the sub-page size rule is determined, where the sub-page size of the low-risk area is S1, the sub-page size of the medium-risk area is S2, and the sub-page size of the high-risk area is S3, and S1>S2>S3;
[0025] Perform pre-inspection complexity analysis on each risk area and determine the pre-inspection digit rule. The number of pre-inspection digits in low-risk areas is N1, the number of pre-inspection digits in medium-risk areas is N2, and the number of pre-inspection digits in high-risk areas is N3. <N2<N3;
[0026] Associating the error pattern description, the subpage size rule and the pre-check bit number rule to construct an error feature mapping table;
[0027] According to the error feature mapping table, a multi-layer decision tree is constructed, wherein the root node of the decision tree is the RBER value, the intermediate nodes are the risk area and the sub-page size, and the leaf nodes are the number of pre-check bits and the error mode description;
[0028] Based on the multi-layer decision tree, an initial error feature library and a sub-page division strategy are generated.
[0029] In combination with the first aspect, in a fourth implementation of the first aspect of the present invention, the sub-page division and quality assessment of the flash memory data are performed according to the initial comprehensive error prediction model, the initial error feature library and the sub-page division strategy to obtain the sub-page quality score, including:
[0030] Dividing the flash memory data according to the sub-page division strategy to obtain a plurality of sub-pages;
[0031] Perform parameter extraction on each sub-page to obtain a parameter set including program / erase cycle number, data retention time and flash memory hierarchy;
[0032] Substituting the parameter set into the initial comprehensive error prediction model to obtain a predicted RBER value for each subpage;
[0033] Calculate the risk level of each subpage according to the predicted RBER value of each subpage and the RBER threshold in the initial error feature library to obtain an initial quality score;
[0034] Based on a preset sampling strategy, select some sub-pages from the multiple sub-pages as samples;
[0035] Perform a read operation on the selected sample subpage to obtain actual bit flip data and obtain an actual RBER value of the sample subpage;
[0036] Calculate the prediction deviation according to the actual RBER value of the sample subpage and the corresponding predicted RBER value to obtain a global correction coefficient;
[0037] Performing a weighted combination of the initial quality score and the global correction coefficient to obtain a comprehensive quality score for each subpage;
[0038] Sorting all sub-pages according to the comprehensive quality score to obtain a sub-page quality ranking list;
[0039] Based on the sub-page quality ranking list, the sub-page quality scores are normalized to obtain sub-page quality scores.
[0040] In combination with the first aspect, in a fifth implementation of the first aspect of the present invention, the risk rating and pre-check strategy allocation for the sub-pages are performed according to the initial comprehensive error prediction model, the initial error feature library and the sub-page quality score to obtain the sub-page pre-check result, including:
[0041] According to the sub-page quality scores and multiple preset RBER thresholds in the initial error feature library, the sub-pages are dynamically divided into risk levels to obtain multiple risk level groups;
[0042] Based on the initial comprehensive error prediction model, combined with the importance and access frequency of the sub-pages, a comprehensive risk index is calculated for each risk level group to obtain the expected error probability of each group;
[0043] According to the expected error probability of each group and the pre-check strategy rules in the initial error feature library, an appropriate pre-check method and number of bits are selected for each risk level group to obtain an adaptive pre-check strategy table;
[0044] Based on the adaptive pre-check strategy table, a corresponding pre-check method and number of bits are assigned to each sub-page to obtain a sub-page pre-check configuration list;
[0045] Generate a corresponding error detection code for each sub-page according to the sub-page pre-check configuration list to obtain a sub-page error detection code set;
[0046] Operate the sub-page error detection code set with the original sub-page data to obtain a pre-check result sequence;
[0047] Analyze the pre-check result sequence, mark the sub-pages with negative pre-check results as error-free, and mark the sub-pages with positive pre-check results as potential errors, to obtain a sub-page error mark list;
[0048] Further verification is performed on subpages with positive pre-check results to reduce false positives and update the subpage error tag list;
[0049] According to the updated sub-page error mark list, the number of available processing units of the flash memory controller, and the size of the memory buffer, the sub-pages marked as potential errors are prioritized and added to the list of sub-pages that require further LDPC decoding, and the sub-pages marked as error-free are added to the list of sub-pages that do not require decoding;
[0050] The sub-page list that needs further LDPC decoding is combined with the sub-page list that does not need decoding to obtain a sub-page pre-check result.
[0051] In combination with the first aspect, in a sixth implementation of the first aspect of the present invention, selecting a subpage to be detected according to the subpage pre-check result, performing parallel error detection and LDPC-based error correction on the subpage to be detected, and obtaining an error detection and correction result includes:
[0052] According to the sub-page pre-check result, combined with the number of available processing units and the size of the memory buffer of the flash memory controller, the sub-pages to be checked are prioritized and allocated in parallel to obtain multiple parallel processing queues;
[0053] Performing LDPC hard decision decoding on the subpages in the parallel processing queue to obtain a first decoding result;
[0054] According to the first decoding result, soft information calculation and soft decision iterative decoding are performed on the sub-pages that failed to be decoded to obtain a second decoding result;
[0055] Adaptively adjusting the read voltage of the sub-page that still fails in the second decoding result, wherein the adaptive read voltage adjustment includes dynamically selecting an optimal read voltage according to an error mode and historical decoding performance of the sub-page;
[0056] Re-read the failed sub-page using the adjusted read voltage to obtain updated soft information;
[0057] According to the updated soft information, a final soft decision iterative decoding is performed on the remaining undecoded sub-pages to obtain a final decoding result;
[0058] Marking the subpages that still fail in the final decoding result as unrecoverable errors, and recording their location information;
[0059] Data is reorganized for all decoding results, subpages that do not need to be detected, and subpages with unrecoverable errors to obtain complete error detection and correction results.
[0060] In combination with the first aspect, in a seventh implementation of the first aspect of the present invention, performing soft information calculation and soft decision iterative decoding on the sub-page that failed to be decoded according to the first decoding result to obtain a second decoding result includes:
[0061] Reading the sub-pages that failed decoding multiple times, each time using a preset increasing or decreasing reading voltage threshold sequence, to obtain multiple groups of bit sequences;
[0062] According to the plurality of groups of bit sequences, a log likelihood ratio (LLR) of each bit is calculated using a probability statistics method to obtain an initial LLR value sequence;
[0063] Initializing an LDPC check matrix using the initial LLR value sequence to obtain an initialized LDPC decoder;
[0064] Performing a horizontal scanning step on the initialized LDPC decoder to update the check node information and obtain an updated check node message;
[0065] Perform a vertical scanning step according to the updated check node message, update the variable node information, and obtain an updated variable node message;
[0066] Calculating a soft decision result based on the updated variable node message to obtain a decoding result of the current iteration;
[0067] Determine whether the decoding result of the current iteration meets the stopping criteria, wherein the stopping criteria include reaching the maximum number of iterations or the bit error rate is lower than a preset threshold; if not, return to the horizontal scanning step to continue iterating; if satisfied, output the decoding result of the current iteration;
[0068] The output decoding result is verified, and the sub-pages that pass the verification are marked as decoding success, and the sub-pages that fail the verification are marked as decoding failure, to obtain the second decoding result.
[0069] In combination with the first aspect, in an eighth implementation of the first aspect of the present invention, after selecting a subpage to be detected according to the subpage pre-check result, performing parallel error detection and LDPC-based error correction on the subpage to be detected, and obtaining error detection and correction results, the flash memory error information detection method includes:
[0070] extracting new error patterns and error statistics based on the error detection and correction results;
[0071] Based on the new error mode and error statistical information, dynamically updating the initial error feature library to obtain an updated error feature library;
[0072] Using the latest error features in the updated error feature library, incremental learning is performed on the initial comprehensive error prediction model to obtain an updated comprehensive error prediction model;
[0073] The updated error feature library and the updated comprehensive error prediction model are used in a subsequent flash memory error information detection process.
[0074] A second aspect of the present invention provides a flash memory error information detection device, the flash memory error information detection device comprising:
[0075] A modeling module, used for performing multi-dimensional modeling on a preset flash memory parameter sample set to obtain an initial comprehensive error prediction model, an initial error feature library and a sub-page division strategy;
[0076] A partitioning and evaluation module, used to perform sub-page partitioning and quality evaluation on the flash memory data according to the initial comprehensive error prediction model, the initial error feature library and the sub-page partitioning strategy, and obtain a sub-page quality score;
[0077] A risk rating module, used to perform risk rating and pre-check strategy allocation on sub-pages according to the initial comprehensive error prediction model, the initial error feature library and the sub-page quality score, and obtain a sub-page pre-check result;
[0078] The error detection module is used to select a sub-page to be detected according to the sub-page pre-check result, perform parallel error detection and LDPC-based error correction on the sub-page to be detected, and obtain error detection and correction results.
[0079] The flash memory error information detection method proposed in the present invention effectively solves the problems of low detection efficiency, insufficient accuracy and high system overhead in the prior art through multi-dimensional modeling, dynamic sub-page division and adaptive pre-check strategy. First, multi-dimensional modeling takes into account multiple factors such as the number of program / erase cycles, data retention time, flash memory level, ambient temperature and read interference, and constructs a more accurate error prediction model, which enables error detection to adapt to the dynamic characteristics of flash memory. The dynamic sub-page division and quality assessment mechanism allow the system to flexibly adjust the granularity and intensity of detection according to the characteristics of the data and the predicted error risk, thereby avoiding excessive detection of low-risk data while ensuring detection accuracy, significantly improving detection efficiency and reducing unnecessary system overhead.
[0080] Secondly, this method further optimizes the detection process through adaptive pre-check strategy and parallel error detection technology. The adaptive pre-check strategy dynamically allocates pre-check methods and bits according to the risk level of the sub-page, realizes differentiated processing, effectively reduces the comprehensive detection of low-risk data, and greatly reduces the system burden. Parallel error detection and LDPC-based error correction make full use of hardware resources, improve the speed of detection and correction, especially in large-capacity flash memory and high-frequency reading scenarios, significantly improving the response speed of the system. The comprehensive application of these technical features not only improves the accuracy and efficiency of flash memory error detection, but also realizes the dynamic adaptability and intelligence of the detection process, effectively prolongs the service life of the flash memory storage system, improves the overall performance and reliability of the system, and reduces power consumption, laying the foundation for the widespread application of flash memory storage systems in data-intensive applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0081] Figure 1 A schematic diagram of an embodiment of a semiconductor chip flip-chip method according to an embodiment of the present invention;
[0082] Figure 2 FIG. 1 is a schematic diagram of an embodiment of a semiconductor chip flip-chip device in an embodiment of the present invention. DETAILED DESCRIPTION
[0083] Embodiments of the present invention provide a flash memory error information detection method and device. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0084] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of a flash memory error information detection method in an embodiment of the present invention includes:
[0085] S101, performing multi-dimensional modeling on a preset flash memory parameter sample set to obtain an initial comprehensive error prediction model, an initial error feature library, and a sub-page division strategy;
[0086] In one embodiment of the present invention, the multi-dimensional modeling of the preset flash memory parameter sample set to obtain an initial comprehensive error prediction model, an initial error feature library and a sub-page division strategy includes: collecting data on the preset flash memory parameter sample set to obtain a multi-dimensional parameter sample including program / erase cycle times, data retention time, flash memory level, ambient temperature and read interference; constructing a mathematical model based on the multi-dimensional parameter samples to obtain a relationship equation between the original bit error rate RBER and each parameter; curve fitting the relationship equation to obtain an initial comprehensive error prediction model; calculating the RBER values under different parameter combinations based on the initial comprehensive error prediction model to obtain an RBER distribution graph; performing threshold division based on the RBER distribution graph to obtain an initial error feature library and a sub-page division strategy.
[0087] Specifically, the flash memory error information detection method of the present invention first conducts comprehensive and systematic data collection on a preset flash memory parameter sample set. This process involves multi-dimensional, multi-scenario testing and analysis of 3D NAND flash memory to obtain multi-dimensional sample data including key parameters such as program / erase cycle count (PE), data retention time (t), flash memory level (L), ambient temperature (T) and read interference (R). Specifically, a series of complex experimental schemes were designed to simulate the performance of flash memory under various actual usage conditions.
[0088] First, by performing different numbers of program / erase cycles on the flash memory blocks (such as 1000 times, 5000 times, 10000 times, etc.), different degrees of flash memory wear are simulated. At the same time, the written data is retained for different lengths of time, such as 1 day, 1 week, and 1 month, to study the impact of data retention time on the error rate. For the multi-layer structure of 3D NAND flash memory, the bottom layer, middle layer, and top layer are tested separately to analyze the impact of the layer on the error rate. In addition, in order to capture the impact of ambient temperature on the error rate, tests are performed under different temperature conditions (such as -20°C, 25°C, and 85°C), which can simulate the performance of flash memory in various extreme environments. For example, long-term data retention tests at a high temperature of 85°C can accelerate the observation of the generation process of data retention errors.
[0089] In order to study the impact of read disturbance on error rate, a specific read mode is designed. For example, the adjacent pages of a certain page (such as the upper and lower adjacent word lines) are frequently read to observe the change in the error rate of the target page. This test can simulate the impact of frequent reading of certain hot data on the surrounding cold data in actual use.
[0090] In each test, the raw bit error rate (RBER) was recorded as the dependent variable. Through this comprehensive data collection process, a large-scale dataset containing multiple independent variables such as PE, t, L, T, and R, as well as the corresponding RBER, was obtained.
[0091] Next, based on the collected multi-dimensional parameter samples, a mathematical model is constructed to obtain a relationship equation between RBER and each parameter. In an embodiment of the present invention, the relationship equation is the following formula:
[0092] .
[0093] Among them, RBER represents the raw bit error rate, which is the dependent variable and represents the probability of an error in a flash memory cell under given conditions. a is the proportionality coefficient, which is used to adjust the size of the overall error rate. PE is the number of program / erase cycles, which represents the number of times the flash memory cell is written and erased. b is the exponent of PE, which reflects the nonlinear effect of PE on RBER. c is the time influence factor, which adjusts the influence of data retention time on RBER. t is the data retention time, which represents the length of time the data is stored in the flash memory. d is the interaction factor between PE and retention time, which reflects the interaction between PE and t. e is the level influence coefficient, which adjusts the influence of the flash memory level on RBER. L is the flash memory level, which represents the vertical position in the 3D NAND flash memory structure. f is the temperature influence coefficient, which adjusts the influence of temperature on RBER. T is the ambient temperature, which represents the temperature condition when the flash memory is working. g is the read interference influence coefficient, which adjusts the influence of read interference on RBER. R is read interference, which represents the interference caused by frequent reading of adjacent pages.
[0094] Then, a curve fitting is performed on this relationship equation to obtain the initial comprehensive error prediction model. This step uses the nonlinear least squares method, specifically the Levenberg-Marquardt algorithm. First, the collected data is preprocessed, including normalization to balance the scale differences of different parameters. Then, the optimal parameter value is found through iterative calculation to minimize the sum of squared errors between the model prediction value and the actual observation value. In order to avoid overfitting, the k-fold cross-validation method (such as k=10) is used to randomly divide the data set into 10 parts, 9 of which are used for training and 1 for validation, and repeated 10 times to ensure the stability and generalization ability of the model.
[0095] After having the initial comprehensive error prediction model, the RBER values under different parameter combinations are calculated to obtain the RBER distribution map. This process is implemented by the grid search method. Specifically, a series of discrete values are set for PE, t, L, T, and R. For example, PE ranges from 100 to 10,000 with a step size of 100, t is divided on a logarithmic scale from 1 day to 1 year, L ranges from 1 to 64 (assuming a 64-layer structure), T ranges from -20°C to 85°C, and R ranges from 0 to 100 (indicating the relative intensity of the read interference). Then, for all combinations of these parameters, the RBER values are calculated using the fitted model. In this way, a multi-dimensional RBER distribution map is obtained.
[0096] Finally, threshold division is performed based on the RBER distribution map to obtain the initial error feature library and sub-page division strategy. The cumulative distribution function of RBER is calculated, and the quantiles of 33.3% and 66.7% are selected as thresholds to divide the RBER distribution into three risk areas: low, medium, and high. For each risk area, its features are extracted to form the initial error feature library. For example, low-risk areas may mainly appear in the case of low PE, short t, bottom layer, normal temperature, and low read interference, and the error type is mainly random bit flipping; while high-risk areas may be concentrated in the case of high PE, long t, top layer, high temperature, and high read interference, and the error type may include more burst errors or continuous errors.
[0097] Based on these risk areas, a sub-page division strategy is formulated: low-risk areas use 4KB sub-page size and 16-bit CRC check; medium-risk areas use 2KB sub-page and 32-bit CRC; high-risk areas use 1KB sub-page and 64-bit CRC. This differentiated strategy ensures detection accuracy and improves system efficiency.
[0098] Through the above detailed steps, a comprehensive and accurate flash memory error prediction and feature analysis system is established, laying a solid foundation for subsequent dynamic error detection and correction. This method not only considers the complex characteristics and various influencing factors of 3D NAND flash memory, but also can adapt to different usage scenarios and environmental conditions, significantly improving the reliability and performance of the flash memory storage system.
[0099] Further, based on the RBER distribution map, threshold division is performed to obtain an initial error feature library and a sub-page division strategy, including: performing statistical analysis on the RBER distribution map to obtain the cumulative distribution function of the RBER value; according to the cumulative distribution function, calculating the 33.3% quantile and 66.7% quantile of the RBER value to obtain the first RBER threshold point and the second RBER threshold point respectively; based on the first RBER threshold point and the second RBER threshold point, dividing the RBER distribution into a low-risk area, a medium-risk area, and a high-risk area, where the RBER value less than the first RBER threshold point is the low-risk area, the RBER value between the first RBER threshold point and the second RBER threshold point is the medium-risk area, and the RBER value greater than the second RBER threshold point is the high-risk area; performing feature analysis on each risk area to extract the error type, error distribution, and error frequency of the area to obtain the error pattern description of each area; according to the RBER range of each risk area, determining the sub-page size rule, where the sub-page size of the low-risk area is S1, the sub-page size of the medium-risk area is S2, and the sub-page size of the high-risk area is S3, and S1 > S2 > S3; performing pre-check complexity analysis on each risk area to determine the pre-check bit number rule, where the pre-check bit number of the low-risk area is N1, the pre-check bit number of the medium-risk area is N2, and the pre-check bit number of the high-risk area is N3, and N1 < N2 < N3; associating the error pattern description, sub-page size rule, and pre-check bit number rule to construct an error feature mapping table; according to the error feature mapping table, constructing a multi-layer decision tree, where the root node of the decision tree is the RBER value, the intermediate nodes are the risk area and the sub-page size, and the leaf nodes are the pre-check bit number and the error pattern description; based on the multi-layer decision tree, generating an initial error feature library and a sub-page division strategy.
[0100] Specifically, first, perform statistical analysis on the RBER distribution map to calculate the cumulative distribution function of the RBER value. This process involves sorting all the calculated RBER values and then calculating the cumulative probability corresponding to each RBER value. For example, if there are 10000 RBER values, the cumulative probability corresponding to the 1000th value after sorting is 10%. In this way, a complete cumulative distribution function curve can be obtained, which describes the distribution characteristics of the RBER value.
[0101] Next, based on this cumulative distribution function, calculate the 33.3% quantile and 66.7% quantile of the RBER value, which are used as the first RBER threshold point and the second RBER threshold point respectively. The selection of these two threshold points aims to roughly divide the RBER distribution into three regions. Specifically, methods such as binary search can be used to find the RBER values corresponding to these two percentages on the cumulative distribution function. For example, if the RBER value corresponding to the 33.3% quantile is 1e-5 and the value corresponding to the 66.7% quantile is 1e-4, then these two values become the first and second RBER threshold points respectively.
[0102] Based on these two threshold points, divide the RBER distribution into three risk regions: low-risk region (RBER < 1e-5), medium-risk region (1e-5 ≤ RBER < 1e-4), and high-risk region (RBER ≥ 1e-4). This division method allows for differentiated management strategies for flash memory regions with different risk levels.
[0103] Subsequently, conduct a detailed feature analysis for each risk region, extract the error types, error distributions, and error frequencies in that region, and obtain the error pattern descriptions for each region. For example, in the low-risk region, it may be found that the errors are mainly random bit flips, the error distribution is relatively uniform, and the error frequency is low. In the high-risk region, more burst errors or consecutive errors may be observed, the error distribution may be concentrated in certain specific positions, and the error frequency is significantly increased. These feature analysis results provide an important basis for subsequent error detection and correction strategies.
[0104] According to the RBER ranges of each risk region, this method further determines the sub-page size rules. The sub-page size in the low-risk region is set to S1 (e.g., 4KB), in the medium-risk region is S2 (e.g., 2KB), and in the high-risk region is S3 (e.g., 1KB), and S1 > S2 > S3. This dynamically adjusted sub-page size strategy can maximize storage efficiency while ensuring data reliability. For example, for the low-risk region, using a larger sub-page size can reduce metadata overhead and improve storage efficiency; for the high-risk region, a smaller sub-page size can provide finer-grained error isolation and management.
[0105] In addition, this method also conducts a pre-check complexity analysis for each risk region to determine the pre-check bit number rules. The pre-check bit number in the low-risk region is set to N1 (e.g., 16 bits), in the medium-risk region is N2 (e.g., 32 bits), and in the high-risk region is N3 (e.g., 64 bits), and N1 < N2 < N3. This progressive pre-check bit number setting can achieve a balance between detection accuracy and efficiency among different risk levels. For example, for the low-risk region, fewer pre-check bits can speed up the checking process; for the high-risk region, increasing the pre-check bit number can improve the accuracy of error detection.
[0106] Next, the method associates the error pattern description, subpage size rules, and pre-check bit number rules to build a comprehensive error feature mapping table. This mapping table combines various factors to provide a clear handling strategy for each situation. For example, for the case where the RBER is 5e-6 and the error type is random bit flip, the mapping table may specify the use of a 2KB subpage size and a 32-bit pre-check bit number.
[0107] Based on this error feature mapping table, this method constructs a multi-layer decision tree. The root node of the decision tree is the RBER value, which is the primary basis for judgment. The intermediate nodes include risk areas and subpage sizes, which guide the basic data management strategy. The leaf nodes are specific pre-check bit numbers and error mode descriptions, providing detailed error handling guidance. For example, a path of the decision tree may be: RBER value (root node) Medium-risk area (intermediate node) 2KB subpage size (mid-node) 32-bit pre-check Random bit flipping pattern (leaf nodes).
[0108] Finally, based on this multi-layer decision tree, this method generates an initial error signature library and a sub-page partitioning strategy. The error signature library contains various possible error conditions and their corresponding processing methods, while the sub-page partitioning strategy provides guidance on how to dynamically adjust the sub-page size under different circumstances. These two outputs provide a comprehensive error management framework for the flash memory controller, enabling it to adopt the most suitable detection and correction strategy based on the real-time RBER value and error characteristics.
[0109] Through this series of complex and precise steps, the method of the present invention establishes an adaptive and efficient flash memory error management system. This system can dynamically adjust the management strategy according to different error risks, while ensuring data reliability, maximizing storage efficiency and system performance. For example, for a new flash memory block, it may be mainly in a low-risk area at the beginning, using larger sub-pages and fewer pre-check bits. With the increase of usage time and the accumulation of P / E cycles, some areas may enter medium-risk or high-risk areas, and the system will automatically adjust to smaller sub-pages and more pre-check bits to cope with the increased error risk. This dynamic adaptability enables the method of the present invention to maintain optimal performance and reliability throughout the life cycle of the flash memory.
[0110] S102, performing sub-page division and quality assessment on the flash memory data according to the initial comprehensive error prediction model, the initial error feature library and the sub-page division strategy to obtain a sub-page quality score;
[0111] In one embodiment of the present invention, the sub-page division and quality evaluation of the flash memory data according to the initial comprehensive error prediction model, the initial error feature library and the sub-page division strategy to obtain the sub-page quality score includes: dividing the flash memory data according to the sub-page division strategy to obtain a plurality of sub-pages; extracting parameters for each sub-page to obtain a parameter set including the number of program / erase cycles, data retention time and flash memory level; substituting the parameter set into the initial comprehensive error prediction model to obtain the predicted RBER value of each sub-page; calculating the risk level of the sub-page according to the predicted RBER value of each sub-page and the RBER threshold in the initial error feature library to obtain the initial quality score. Scoring; based on a preset sampling strategy, select some sub-pages from the multiple sub-pages as samples; perform a read operation on the selected sample sub-page to obtain actual bit flip data and obtain an actual RBER value of the sample sub-page; calculate the prediction deviation based on the actual RBER value of the sample sub-page and the corresponding predicted RBER value to obtain a global correction coefficient; perform a weighted combination of the initial quality score and the global correction coefficient to obtain a comprehensive quality score for each sub-page; sort all sub-pages according to the comprehensive quality score to obtain a sub-page quality sorting list; based on the sub-page quality sorting list, normalize the sub-page quality score to obtain a sub-page quality score.
[0112] Specifically, in the process of implementing the flash memory error information detection method, the flash memory data is first divided according to the previously formulated sub-page division strategy to obtain multiple sub-pages. This step is to manage and detect errors in the flash memory more finely. For example, for a 16KB flash memory page, it may be divided into 4 4KB sub-pages, 8 2KB sub-pages, or 16 1KB sub-pages according to different risk levels. This flexible division method allows the system to adopt different management strategies for areas with different risk levels.
[0113] Next, parameter extraction is performed on each subpage to obtain a parameter set including the number of program / erase cycles, data retention time, and flash memory level. These parameters directly affect the reliability and error rate of the flash memory. For example, for a specific subpage, the following parameter set may be obtained: the number of program / erase cycles is 5000, the data retention time is 30 days, and the flash memory level is 20th layer. These parameters reflect the usage history and physical location of the subpage and are key factors in predicting its error rate.
[0114] Subsequently, these parameter sets are substituted into the previously established initial comprehensive error prediction model to obtain the predicted RBER value for each subpage. Taking the above parameter set as an example, substituting it into the model may result in a predicted RBER value of 1e-5. This predicted value reflects the expected error rate level of this subpage based on the current parameters.
[0115] According to the predicted RBER value of each subpage and the RBER threshold in the initial error feature library, the risk level of the subpage is calculated to obtain an initial quality score. Continuing the above example, if 1e-5 falls in the medium risk area, then this subpage may be assigned a medium initial quality score, such as 70 points (assuming the full score is 100 points).
[0116] In order to verify and optimize the prediction model, this method selects some sub-pages from multiple sub-pages as samples based on a preset sampling strategy. This sampling strategy may consider factors such as different risk levels and different flash memory tiers to ensure the representativeness of the samples. For example, 10% of the sub-pages may be randomly selected from each risk level as samples.
[0117] Perform a read operation on the selected sample subpage to obtain the actual bit flip data and get the actual RBER value of the sample subpage. This step is to compare the predicted value with the actual value to evaluate the accuracy of the prediction model. For example, for the subpage with a predicted RBER of 1e-5, the actual RBER value may be 1.2e-5 after actual reading.
[0118] Based on the actual RBER values of the sample subpages and the corresponding predicted RBER values, the prediction deviation is calculated to obtain the global correction factor. This step aims to adjust the prediction model through actual observation data to improve its accuracy. For example, if it is found that the predicted values are generally 20% lower than the actual values, a global correction factor of 1.2 may be obtained.
[0119] The initial quality score is weighted and combined with the global correction factor to obtain the comprehensive quality score of each subpage. This step combines the model prediction and actual observation results to obtain a more accurate quality assessment. For example, a subpage originally scored 70 points may become 60 points after correction.
[0120] All subpages are sorted according to the comprehensive quality score to obtain a subpage quality sort list. This list reflects the relative quality of all subpages and provides an important reference for subsequent storage management and error detection. For example, the subpages with the highest quality may be used to store important data first, while subpages with lower quality may need to be checked and refreshed more frequently.
[0121] Finally, based on the sub-page quality ranking list, the sub-page quality scores are normalized to obtain the sub-page quality scores. The purpose of normalization is to convert scores of different scales into a unified standard range for easy comparison and use. For example, all scores can be mapped between 0 and 1, where 1 represents the highest quality and 0 represents the lowest quality.
[0122] Through this series of steps, the method of the present invention realizes a refined quality assessment of flash memory subpages. This assessment not only takes into account theoretical predictions, but also incorporates actual observation data, so it can more accurately reflect the actual state of the flash memory. This dynamic and adaptive assessment method enables the system to timely identify potential high-risk areas and take corresponding measures (such as adjusting storage strategies, increasing inspection frequency, etc.) to prevent data errors and losses. At the same time, by identifying low-risk areas, the system can optimize resource utilization and improve overall storage efficiency. The advantage of this method is that it can continuously optimize storage strategies and error detection schemes throughout the life cycle of the flash memory, thereby maximizing the performance and life of the flash memory storage system while ensuring data reliability.
[0123] S103, performing risk rating and pre-check strategy allocation on the sub-pages according to the initial comprehensive error prediction model, the initial error feature library and the sub-page quality scores, and obtaining sub-page pre-check results;
[0124] In one embodiment of the present invention, the sub-pages are risk graded and pre-check strategy assigned according to the initial comprehensive error prediction model, the initial error feature library and the sub-page quality score to obtain the sub-page pre-check result, including: dynamically dividing the sub-pages into risk levels according to the sub-page quality score and multiple preset RBER thresholds in the initial error feature library to obtain multiple risk level groups; based on the initial comprehensive error prediction model, in combination with the importance and access frequency of the sub-pages, a comprehensive risk index is calculated for each risk level group to obtain the expected error probability of each group; according to the expected error probability of each group and the pre-check strategy rules in the initial error feature library, an appropriate pre-check method and number of bits are selected for each risk level group to obtain an adaptive pre-check strategy table; based on the adaptive pre-check strategy table, a corresponding pre-check method and number of bits are assigned to each sub-page to obtain a sub-page pre-check configuration list; according to the sub-page pre-check configuration list Table, generate corresponding error detection code for each subpage, and obtain subpage error detection code set; operate the subpage error detection code set with original subpage data to obtain pre-check result sequence; analyze the pre-check result sequence, mark subpages with negative pre-check results as error-free, and mark subpages with positive pre-check results as potential errors, and obtain subpage error mark list; further verify subpages with positive pre-check results to reduce false positives, and update the subpage error mark list; according to the updated subpage error mark list, the number of available processing units of the flash memory controller and the size of the memory buffer, prioritize the subpages marked as potential errors and add them to the subpage list that needs further LDPC decoding, and add the subpages marked as error-free to the subpage list that does not need decoding; combine the subpage list that needs further LDPC decoding and the subpage list that does not need decoding to obtain the subpage pre-check result.
[0125] Specifically, in the process of implementing the flash memory error information detection method, the subpages are first divided into dynamic risk levels according to the subpage quality scores and multiple preset RBER thresholds in the initial error feature library to obtain multiple risk level groups. The purpose of this step is to classify the subpages more finely based on the latest quality assessment results. For example, subpages may be divided into three risk level groups: low, medium, and high, where subpages with quality scores between 0.8-1.0 are classified as low risk groups, subpages between 0.5-0.8 are classified as medium risk groups, and those below 0.5 are classified as high risk groups.
[0126] Next, based on the initial comprehensive error prediction model, combined with the importance and access frequency of the subpages, a comprehensive risk index is calculated for each risk level group to obtain the expected error probability of each group. This step takes into account the actual usage of the subpages, because even subpages with similar quality may have different actual risks due to the different importance and access frequency of the stored data. For example, a subpage that stores critical system data and is frequently accessed may be assigned a higher comprehensive risk index even if its quality score is higher.
[0127] According to the expected error probability of each group and the pre-check strategy rules in the initial error feature library, the appropriate pre-check method and bit number are selected for each risk level group to obtain an adaptive pre-check strategy table. The purpose of this step is to select the most suitable error detection method for sub-pages of different risk levels. For example, a simple parity check and 16-bit CRC check may be selected for the low-risk group, a 32-bit CRC check may be used for the medium-risk group, and a more complex 64-bit CRC check or a lightweight LDPC code may be used for the high-risk group.
[0128] Based on the adaptive pre-check strategy table, each sub-page is assigned a corresponding pre-check method and bit number to obtain a sub-page pre-check configuration list. This step concretizes the strategy to each sub-page, ensuring that each sub-page has its own exclusive pre-check configuration. For example, a 4KB sub-page classified as medium risk may be configured to use a 32-bit CRC check.
[0129] According to the sub-page pre-check configuration list, the corresponding error detection code is generated for each sub-page to obtain the sub-page error detection code set. This step actually generates the check code for error detection. Continuing with the above example, the system will generate a 32-bit CRC check code for the 4KB sub-page.
[0130] The sub-page error detection code set is operated with the original sub-page data to obtain the pre-check result sequence. This step actually performs the error detection process. For example, the system calculates the CRC value of the sub-page data and compares it with the previously generated check code.
[0131] The pre-check result sequence is analyzed, and the sub-pages with negative pre-check results are marked as error-free, and the sub-pages with positive pre-check results are marked as potential errors, to obtain a sub-page error mark list. This step preliminarily identifies sub-pages that may have errors. For example, if the calculated CRC value does not match the stored check code, the sub-page will be marked as a potential error.
[0132] Subpages with positive pre-check results are further verified to reduce false positives and update the subpage error mark list. This step is intended to improve the accuracy of error detection and avoid unnecessary error correction operations. For example, the system may re-read subpages marked as potential errors or use more complex verification methods for secondary checks.
[0133] Based on the updated subpage error tag list, the number of available processing units of the flash controller, and the memory buffer size, the subpages marked as potential errors are prioritized and added to the list of subpages that require further LDPC decoding, and the subpages marked as error-free are added to the list of subpages that do not require decoding. This step takes into account the limitations of system resources and prioritizes the subpages that require error correction. For example, if the system has only four LDPC decoders, the four subpages with the highest risk may be selected for LDPC decoding first.
[0134] Finally, the sub-page list that needs further LDPC decoding is combined with the sub-page list that does not need decoding to obtain the sub-page pre-check result. This final pre-check result contains the status of each sub-page and the subsequent processing plan.
[0135] Through this series of steps, the method of the present invention realizes a highly adaptive flash memory error detection system. The system can dynamically adjust the detection strategy according to the real-time quality and importance of the sub-page, while ensuring data reliability and maximizing system performance. For example, for a high-risk sub-page storing important data, the system may use a more complex pre-check method and arrange LDPC decoding immediately when a potential error is detected; while for a low-risk sub-page storing temporary data, the system may only perform a simple check, and even if a potential error is detected, the processing may be temporarily postponed. This flexible strategy enables the system to maximize the effect of error detection and correction with limited resources, while minimizing the impact on the overall system performance. In addition, this method also has good scalability, can be adjusted according to actual application scenarios and hardware capabilities, and is suitable for flash memory storage systems of various specifications and uses.
[0136] S104 , selecting a sub-page to be detected according to the sub-page pre-check result, performing parallel error detection and LDPC-based error correction on the sub-page to be detected, and obtaining error detection and correction results.
[0137] In one embodiment of the present invention, the sub-page to be detected is selected according to the sub-page pre-check result, and the sub-page to be detected is parallelized and corrected based on LDPC to obtain error detection and correction results, including: according to the sub-page pre-check result, in combination with the number of available processing units and the size of the memory buffer of the flash memory controller, the sub-page to be detected is prioritized and parallelly allocated to obtain multiple parallel processing queues; the sub-pages in the parallel processing queues are subjected to LDPC hard decision decoding to obtain a first decoding result; according to the first decoding result, soft information calculation and soft decision iterative decoding are performed on the sub-pages that failed to be decoded to obtain a second decoding result; The method comprises the steps of: performing adaptive read voltage adjustment on the sub-pages that still fail in the second decoding result, wherein the adaptive read voltage adjustment includes dynamically selecting an optimal read voltage according to the error pattern and historical decoding performance of the sub-page; re-reading the failed sub-pages using the adjusted read voltage to obtain updated soft information; performing a final soft decision iterative decoding on the remaining undecoded sub-pages according to the updated soft information to obtain a final decoding result; marking the sub-pages that still fail in the final decoding result as unrecoverable errors, and recording their location information; and reorganizing data on all decoding results, sub-pages that do not need to be detected, and sub-pages with unrecoverable errors to obtain a complete error detection and correction result.
[0138] Specifically, first, based on the sub-page pre-check results, combined with the number of available processing units and the size of the memory buffer of the flash controller, the sub-pages to be detected are prioritized and allocated in parallel to obtain multiple parallel processing queues. This step is intended to maximize the use of system resources and improve the efficiency of error detection and correction. For example, assuming that the flash controller has 4 available LDPC decoders and a large enough memory buffer, the system may create 4 parallel processing queues, each of which is assigned a group of sub-pages with similar priorities. High-priority sub-pages, such as those that store critical data or have a higher risk of errors, are assigned to the processing queue first.
[0139] Next, LDPC hard decision decoding is performed on the sub-pages in these parallel processing queues to obtain the first decoding results. Hard decision decoding is a relatively fast but less accurate decoding method that directly maps the read voltage value to 0 or 1. For example, for a 4KB sub-page, LDPC hard decision decoding may only take a few microseconds. This step can quickly correct minor errors and is sufficient for most sub-pages with low error rates.
[0140] Then, based on the first decoding result, soft information calculation and soft decision iterative decoding are performed on the sub-pages that failed to be decoded to obtain the second decoding result. Soft information calculation involves estimating the reliability of each bit, rather than simply judging it as 0 or 1. Soft decision iterative decoding uses this reliability information to improve the accuracy of decoding through multiple iterations. Although this process takes longer than hard decision decoding, it can correct more errors. For example, a sub-page that failed in hard decision decoding may be successfully decoded after 10 soft decision iterations.
[0141] Adaptive read voltage adjustment is performed on subpages that still fail in the second decoding result. The core of this step is to dynamically select the optimal read voltage based on the error pattern and historical decoding performance of the subpage. The threshold voltage of the flash memory cell drifts over time and use, so a fixed read voltage may no longer be applicable. For example, if the system finds that the errors of a subpage are mainly concentrated in a specific voltage range, it may adjust the read voltage of that range accordingly. This adaptive adjustment can significantly improve the accuracy of the read.
[0142] Re-read the failed subpages using the adjusted read voltage to get updated soft information. This step gives these difficult-to-decode subpages a "second chance". For example, a subpage that could not be read correctly before may get more accurate soft information after using the new read voltage, thereby increasing the possibility of successful decoding.
[0143] According to the updated soft information, the remaining undecoded subpages are subjected to a final soft decision iterative decoding to obtain the final decoding result. This is the last step in the system's attempt to correct the error. At this stage, the system may increase the number of iterations or use a more complex decoding algorithm to do its best to recover the data.
[0144] The subpages that still failed in the final decoding result are marked as unrecoverable errors, and their location information is recorded. This information is very important for subsequent data management and system maintenance. For example, if too many unrecoverable subpages accumulate in a block, the system may decide to eliminate this block in the next garbage collection.
[0145] Finally, data is reassembled for all decoded results, subpages that do not need to be detected, and subpages with unrecoverable errors to obtain complete error detection and correction results. This step combines all processing results together to form a complete view. For example, for a flash page containing 100 subpages, the final result may show that 90 subpages are error-free, 8 subpages are successfully corrected, and 2 subpages contain unrecoverable errors.
[0146] Through this series of complex and sophisticated steps, the method of the present invention realizes a highly optimized flash memory error detection and correction system. The advantage of this system is that it can handle errors with a step-by-step strategy based on the severity of the error and the availability of system resources. From fast hard decision decoding, to more accurate soft decision decoding, to adaptive read voltage adjustment, the system corrects errors as much as possible at each step while avoiding unnecessary complex operations. This method can not only effectively handle various types and degrees of errors, but also achieve a good balance between performance and reliability.
[0147] For example, for an SSD storage system in a large data center, this approach can significantly improve data reliability while maintaining efficient read and write performance. Even under sustained high load, the system is able to quickly identify and correct most errors, using more computing resources only when dealing with the most difficult errors. This not only extends the service life of the SSD, but also ensures the overall performance and reliability of the storage system. In addition, by recording and analyzing the patterns of unrecoverable errors, this method also provides valuable information for preventive maintenance of SSDs, allowing storage system administrators to take action before serious problems occur. Overall, this approach represents an important advancement in flash memory error management technology, paving the way for building more reliable and efficient large-scale storage systems.
[0148] Furthermore, the method of performing soft information calculation and soft decision iterative decoding on the sub-page that failed to be decoded according to the first decoding result to obtain the second decoding result includes: reading the sub-page that failed to be decoded multiple times, each time using a preset increasing or decreasing reading voltage threshold sequence to obtain multiple groups of bit sequences; calculating the log likelihood ratio (LLR) of each bit using a probability statistics method according to the multiple groups of bit sequences to obtain an initial LLR value sequence; initializing the LDPC check matrix using the initial LLR value sequence to obtain an initialized LDPC decoder; performing a horizontal scanning step on the initialized LDPC decoder to update the check node information to obtain an updated a check node message; performing a vertical scanning step according to the updated check node message, updating the variable node information, and obtaining an updated variable node message; calculating a soft decision result based on the updated variable node message, and obtaining a decoding result of the current iteration; judging whether the decoding result of the current iteration satisfies a stopping criterion, wherein the stopping criterion includes reaching a maximum number of iterations or a bit error rate being lower than a preset threshold; if not satisfied, returning to the horizontal scanning step to continue iterating, and if satisfied, outputting the decoding result of the current iteration; verifying the output decoding result, marking the sub-page that passed the verification as a successful decoding, and marking the sub-page that failed the verification as a decoding failure, and obtaining a second decoding result.
[0149] Specifically, at the critical stage of the flash memory error information detection method, the system reads the subpage that failed to be decoded multiple times, each time using a preset increasing or decreasing read voltage threshold sequence to obtain multiple groups of bit sequences. This step is intended to obtain richer information to improve the accuracy of subsequent soft decision decoding. For example, for a 4KB subpage that failed to be decoded, the system may use 7 different read voltage thresholds for reading, each reading to obtain a 32768-bit bit sequence, and finally obtain 7 groups of bit sequences. This multiple reading method can better capture the actual state of the flash memory cell, especially for those cells near the threshold voltage boundary.
[0150] Next, based on these multiple sets of bit sequences, the system uses probability statistics to calculate the log-likelihood ratio (LLR) of each bit to obtain the initial LLR value sequence. The LLR value reflects the possibility of each bit being 0 or 1, and is a form of soft information. For example, if a bit is read as 1 5 times and 0 2 times in 7 readings, the LLR value of this bit will tend to be a positive value representing "1". This step converts hard information into soft information, laying the foundation for subsequent soft decision decoding.
[0151] Then, the system uses this initial LLR value sequence to initialize the LDPC check matrix and obtain the initialized LDPC decoder. The LDPC check matrix is the core of the LDPC code, which defines the rules for encoding and decoding. The initialization process actually inputs soft information into the LDPC decoder to prepare for iterative decoding.
[0152] Next, the system performs a horizontal scan step on the initialized LDPC decoder, updates the check node information, and obtains the updated check node message. Horizontal scanning involves calculating the result of each check equation and updating the reliability information of all bits related to the equation. This process can be likened to solving a complex Sudoku game. Each horizontal scan is like checking whether each row meets the rules and updating the judgment of the numbers in each grid.
[0153] According to the updated check node message, the system performs a vertical scanning step, updates the variable node information, and obtains the updated variable node message. Vertical scanning looks at the problem from another angle. It comprehensively considers the results of all check equations that affect each bit and further updates the reliability information of each bit. Continuing the Sudoku analogy, this is equivalent to checking whether each column meets the rules and updating the judgment of the number in each grid again.
[0154] Based on the updated variable node message, the system calculates the soft decision result and obtains the decoding result of the current iteration. The soft decision result is an estimate of the most likely value (0 or 1) of each bit, but unlike the hard decision, it also retains the confidence information of this estimate.
[0155] Next, the system determines whether the decoding result of the current iteration meets the stopping criteria, where the stopping criteria include reaching the maximum number of iterations or the bit error rate is lower than a preset threshold. This step is to strike a balance between decoding quality and decoding time. For example, the system may set the maximum number of iterations to 50, or require the bit error rate to be lower than 1e-6. If the stopping criteria are not met, the system will return to the horizontal scanning step to continue iterating; if they are met, the decoding result of the current iteration is output.
[0156] Finally, the system verifies the output decoding results, marks the subpages that pass the verification as decoding success, and marks the subpages that fail the verification as decoding failure, and obtains the second decoding result. This verification process involves calculating the checksum or using other error detection codes to ensure the correctness of the decoding results.
[0157] Through this series of complex steps, the method of the present invention realizes a highly optimized soft decision LDPC decoding process. The advantage of this process is that it can make full use of the soft information obtained by multiple reads, greatly improving the success rate of decoding. For example, for a subpage that fails in hard decision decoding, this soft decision method may be able to successfully recover the data even if there are a considerable number of bit errors.
[0158] In practical applications, this method is crucial to improving the reliability and data integrity of flash storage systems. For example, in an SSD storage system in a large data center, even a small percentage of unrecoverable errors can lead to serious data loss. By using this advanced soft-decision LDPC decoding method, the system can significantly reduce the occurrence of unrecoverable errors, thereby improving the reliability of the entire storage system.
[0159] In addition, the adaptive nature of this method enables it to effectively handle errors of varying severity. For subpages with less severe errors, only a few iterations may be needed to successfully decode; while for subpages with more severe errors, the system can perform more iterations to do its best to recover the data. This flexibility enables the system to achieve a good balance between decoding performance and processing time, ensuring the quality of data recovery while avoiding unnecessary waste of computing resources.
[0160] Overall, this soft-decision LDPC decoding method represents an important advancement in flash memory error correction technology. It not only improves the success rate of data recovery, but also provides greater design flexibility for flash storage systems, allowing the system to operate reliably under more severe conditions, thereby extending the service life of flash memory devices and improving the overall performance of storage systems.
[0161] In one embodiment of the present invention, after step S104, the following steps are included: extracting new error patterns and error statistical information based on the error detection and correction results; dynamically updating the initial error feature library based on the new error patterns and error statistical information to obtain an updated error feature library; using the latest error features in the updated error feature library to incrementally learn the initial comprehensive error prediction model to obtain an updated comprehensive error prediction model; and using the updated error feature library and the updated comprehensive error prediction model in subsequent flash memory error information detection processes.
[0162] Specifically, first, the system extracts new error patterns and error statistics based on the error detection and correction results. This step involves a comprehensive analysis of all subpages just processed, including the type, distribution, frequency and other characteristics of the errors. For example, the system may find that certain specific error patterns, such as continuous bit flips or errors concentrated in a certain voltage range, have become more common.
[0163] Then, based on these new error patterns and error statistics, the system dynamically updates the initial error signature library to obtain an updated error signature library. This process may include adding newly discovered error patterns, adjusting the weights of known error patterns, or deleting old patterns that are no longer relevant. For example, if the system finds that a previously rare error pattern has become more common due to flash aging, it will increase the weight of this error pattern in the signature library accordingly.
[0164] The system then uses the latest error features in the updated error feature library to perform incremental learning on the initial comprehensive error prediction model to obtain an updated comprehensive error prediction model. Incremental learning allows the model to adapt to new data without completely retraining. For example, if new error statistics show that RBER increases faster than expected with P / E cycles, the model may adjust its parameters accordingly to more accurately predict future error rates.
[0165] Finally, the system uses the updated error signature library and the updated comprehensive error prediction model for subsequent flash memory error information detection processes. This continuous self-optimization mechanism ensures that the system can continuously adjust and improve its error detection and correction strategies as the flash memory characteristics change. For example, in a long-running data center SSD, this approach can help the system adapt to the changes in the flash memory over time and always maintain the best error management performance, thereby extending the life of the SSD and maintaining data reliability.
[0166] The flash memory error information detection method proposed in the present invention effectively solves the problems of low detection efficiency, insufficient accuracy and high system overhead in the prior art through multi-dimensional modeling, dynamic sub-page division and adaptive pre-check strategy. First, multi-dimensional modeling takes into account multiple factors such as the number of program / erase cycles, data retention time, flash memory level, ambient temperature and read interference, and constructs a more accurate error prediction model, which enables error detection to adapt to the dynamic characteristics of flash memory. The dynamic sub-page division and quality assessment mechanism allow the system to flexibly adjust the granularity and intensity of detection according to the characteristics of the data and the predicted error risk, thereby avoiding excessive detection of low-risk data while ensuring detection accuracy, significantly improving detection efficiency and reducing unnecessary system overhead.
[0167] Secondly, this method further optimizes the detection process through adaptive pre-check strategy and parallel error detection technology. The adaptive pre-check strategy dynamically allocates pre-check methods and bits according to the risk level of the sub-page, realizes differentiated processing, effectively reduces the comprehensive detection of low-risk data, and greatly reduces the system burden. Parallel error detection and LDPC-based error correction make full use of hardware resources, improve the speed of detection and correction, especially in large-capacity flash memory and high-frequency reading scenarios, significantly improving the response speed of the system. The comprehensive application of these technical features not only improves the accuracy and efficiency of flash memory error detection, but also realizes the dynamic adaptability and intelligence of the detection process, effectively prolongs the service life of the flash memory storage system, improves the overall performance and reliability of the system, and reduces power consumption, laying the foundation for the widespread application of flash memory storage systems in data-intensive applications.
[0168] The above describes the flash memory error information detection method in the embodiment of the present invention. The following describes the flash memory error information detection device in the embodiment of the present invention. Figure 2 , an embodiment of a flash memory error information detection device in an embodiment of the present invention includes:
[0169] Modeling module 201, used to perform multi-dimensional modeling on a preset flash memory parameter sample set to obtain an initial comprehensive error prediction model, an initial error feature library and a sub-page division strategy;
[0170] A partitioning and evaluation module 202 is used to perform sub-page partitioning and quality evaluation on the flash memory data according to the initial comprehensive error prediction model, the initial error feature library and the sub-page partitioning strategy to obtain a sub-page quality score;
[0171] The risk rating module 203 is used to perform risk rating and pre-check strategy allocation on the sub-pages according to the initial comprehensive error prediction model, the initial error feature library and the sub-page quality score, and obtain the sub-page pre-check result;
[0172] The error detection module 204 is used to select a sub-page to be detected according to the sub-page pre-check result, perform parallel error detection and LDPC-based error correction on the sub-page to be detected, and obtain error detection and correction results.
[0173] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A flash memory error information detection method, characterized in that: include: Perform multi-dimensional modeling on the preset flash memory parameter sample set to obtain an initial comprehensive error prediction model, an initial error feature library and a sub-page division strategy; Performing sub-page division and quality assessment on the flash memory data according to the initial comprehensive error prediction model, the initial error feature library and the sub-page division strategy to obtain a sub-page quality score; Perform risk rating and pre-check strategy allocation on the sub-pages according to the initial comprehensive error prediction model, the initial error feature library and the sub-page quality score, and obtain the sub-page pre-check result; The method of performing risk rating and pre-check strategy allocation on sub-pages according to the initial comprehensive error prediction model, the initial error feature library and the sub-page quality score to obtain the sub-page pre-check result includes: dynamically dividing the sub-pages into risk levels according to the sub-page quality score and multiple preset RBER thresholds in the initial error feature library to obtain multiple risk level groups; calculating a comprehensive risk index for each risk level group based on the initial comprehensive error prediction model in combination with the importance and access frequency of the sub-pages to obtain the expected error probability of each group; selecting an appropriate pre-check method and number of bits for each risk level group according to the expected error probability of each group and the pre-check strategy rules in the initial error feature library to obtain an adaptive pre-check strategy table; allocating a corresponding pre-check method and number of bits to each sub-page based on the adaptive pre-check strategy table to obtain a sub-page pre-check configuration list; and allocating a corresponding pre-check method and number of bits to each sub-page according to the sub-page pre-check configuration list. The page generates a corresponding error detection code to obtain a sub-page error detection code set; the sub-page error detection code set is operated with the original sub-page data to obtain a pre-check result sequence; the pre-check result sequence is analyzed, and the sub-page with a negative pre-check result is marked as error-free, and the sub-page with a positive pre-check result is marked as a potential error, to obtain a sub-page error mark list; the sub-page with a positive pre-check result is further verified to reduce false positives, and the sub-page error mark list is updated; according to the updated sub-page error mark list, the number of available processing units of the flash memory controller and the size of the memory buffer, the sub-pages marked as potential errors are prioritized and added to the sub-page list requiring further LDPC decoding, and the sub-pages marked as error-free are added to the sub-page list not requiring decoding; the sub-page list requiring further LDPC decoding and the sub-page list not requiring decoding are combined to obtain a sub-page pre-check result; A subpage to be detected is selected according to the subpage pre-check result, and parallel error detection and LDPC-based error correction are performed on the subpage to be detected to obtain error detection and correction results.
2. The flash memory error information detection method according to claim 1, characterized in that: The multi-dimensional modeling of the preset flash memory parameter sample set is performed to obtain an initial comprehensive error prediction model, an initial error feature library and a sub-page division strategy, including: Collecting data on the preset flash memory parameter sample set to obtain a multi-dimensional parameter sample including program / erase cycle number, data retention time, flash memory level, ambient temperature and read interference; A mathematical model is constructed based on the multi-dimensional parameter sample to obtain a relationship equation between the original bit error rate RBER and each parameter; Performing curve fitting on the relationship equation to obtain an initial comprehensive error prediction model; Calculate the RBER values under different parameter combinations according to the initial comprehensive error prediction model to obtain an RBER distribution graph; Threshold division is performed based on the RBER distribution graph to obtain an initial error feature library and a sub-page division strategy.
3. The flash memory error information detection method according to claim 2, characterized in that: The relationship equation is the following formula: ; Wherein, RBER represents the raw bit error rate, a is the proportionality coefficient, PE is the number of program / erase cycles, b is the exponent of PE, c is the time impact factor, t is the data retention time, d is the interaction impact factor of PE and retention time, e is the level impact factor, L is the flash memory level, f is the temperature impact factor, T is the ambient temperature, g is the read disturbance impact factor, and R is the read disturbance.
4. The flash memory error information detection method according to claim 2, characterized in that: The threshold division is performed based on the RBER distribution graph to obtain an initial error feature library and a sub-page division strategy, including: Performing statistical analysis on the RBER distribution graph to obtain a cumulative distribution function of the RBER value; According to the cumulative distribution function, calculating the 33.3% quantile and the 66.7% quantile of the RBER value, and obtaining a first RBER threshold point and a second RBER threshold point respectively; Based on the first RBER threshold point and the second RBER threshold point, the RBER distribution is divided into a low-risk area, a medium-risk area and a high-risk area, wherein the RBER value is less than the first RBER threshold point for the low-risk area, the RBER value is between the first RBER threshold point and the second RBER threshold point for the medium-risk area, and the RBER value is greater than the second RBER threshold point for the high-risk area; Perform feature analysis on each risk area, extract the error type, error distribution and error frequency of the area, and obtain the error pattern description of each area; According to the RBER range of each risk area, the sub-page size rule is determined, where the sub-page size of the low-risk area is S1, the sub-page size of the medium-risk area is S2, and the sub-page size of the high-risk area is S3, and S1>S2>S3; Perform pre-inspection complexity analysis on each risk area and determine the pre-inspection digit rule. The number of pre-inspection digits in low-risk areas is N1, the number of pre-inspection digits in medium-risk areas is N2, and the number of pre-inspection digits in high-risk areas is N3. <N2<N3; Associating the error pattern description, the subpage size rule and the pre-check bit number rule to construct an error feature mapping table; According to the error feature mapping table, a multi-layer decision tree is constructed, wherein the root node of the decision tree is the RBER value, the intermediate nodes are the risk area and the sub-page size, and the leaf nodes are the number of pre-check bits and the error mode description; Based on the multi-layer decision tree, an initial error feature library and a sub-page division strategy are generated.
5. The flash memory error information detection method according to claim 1, characterized in that: The sub-page division and quality assessment of the flash memory data are performed according to the initial comprehensive error prediction model, the initial error feature library and the sub-page division strategy to obtain the sub-page quality score, including: Dividing the flash memory data according to the sub-page division strategy to obtain a plurality of sub-pages; Perform parameter extraction on each sub-page to obtain a parameter set including program / erase cycle number, data retention time and flash memory hierarchy; Substituting the parameter set into the initial comprehensive error prediction model to obtain a predicted RBER value for each subpage; Calculate the risk level of each subpage according to the predicted RBER value of each subpage and the RBER threshold in the initial error feature library to obtain an initial quality score; Based on a preset sampling strategy, select some sub-pages from the multiple sub-pages as samples; Perform a read operation on the selected sample subpage to obtain actual bit flip data and obtain an actual RBER value of the sample subpage; Calculate the prediction deviation according to the actual RBER value of the sample subpage and the corresponding predicted RBER value to obtain a global correction coefficient; Performing a weighted combination of the initial quality score and the global correction coefficient to obtain a comprehensive quality score for each subpage; Sorting all sub-pages according to the comprehensive quality score to obtain a sub-page quality ranking list; Based on the sub-page quality ranking list, the sub-page quality scores are normalized to obtain sub-page quality scores.
6. The flash memory error information detection method according to claim 1, characterized in that: The step of selecting a subpage to be detected according to the subpage pre-check result, performing parallel error detection and LDPC-based error correction on the subpage to be detected, and obtaining an error detection and correction result includes: According to the sub-page pre-check result, combined with the number of available processing units and the size of the memory buffer of the flash memory controller, the sub-pages to be checked are prioritized and allocated in parallel to obtain multiple parallel processing queues; Performing LDPC hard decision decoding on the subpages in the parallel processing queue to obtain a first decoding result; According to the first decoding result, soft information calculation and soft decision iterative decoding are performed on the sub-pages that failed to be decoded to obtain a second decoding result; Adaptively adjusting the read voltage of the sub-page that still fails in the second decoding result, wherein the adaptive read voltage adjustment includes dynamically selecting an optimal read voltage according to an error mode and historical decoding performance of the sub-page; Re-read the failed sub-page using the adjusted read voltage to obtain updated soft information; According to the updated soft information, a final soft decision iterative decoding is performed on the remaining undecoded sub-pages to obtain a final decoding result; Marking the subpages that still fail in the final decoding result as unrecoverable errors, and recording their location information; Data is reorganized for all decoding results, subpages that do not need to be detected, and subpages with unrecoverable errors to obtain complete error detection and correction results.
7. The flash memory error information detection method according to claim 6, characterized in that: The step of performing soft information calculation and soft decision iterative decoding on the sub-pages that failed to be decoded according to the first decoding result to obtain the second decoding result includes: Reading the sub-pages that failed decoding multiple times, each time using a preset increasing or decreasing reading voltage threshold sequence, to obtain multiple groups of bit sequences; According to the plurality of groups of bit sequences, a log likelihood ratio (LLR) of each bit is calculated using a probability statistics method to obtain an initial LLR value sequence; Initializing an LDPC check matrix using the initial LLR value sequence to obtain an initialized LDPC decoder; Performing a horizontal scanning step on the initialized LDPC decoder to update the check node information and obtain an updated check node message; Perform a vertical scanning step according to the updated check node message, update the variable node information, and obtain an updated variable node message; Calculating a soft decision result based on the updated variable node message to obtain a decoding result of the current iteration; Determine whether the decoding result of the current iteration meets the stopping criteria, wherein the stopping criteria include reaching the maximum number of iterations or the bit error rate is lower than a preset threshold; if not, return to the horizontal scanning step to continue iterating; if satisfied, output the decoding result of the current iteration; The output decoding result is verified, and the sub-pages that pass the verification are marked as decoding success, and the sub-pages that fail the verification are marked as decoding failure, to obtain the second decoding result.
8. The flash memory error information detection method according to claim 1, characterized in that: After selecting a subpage to be detected according to the subpage pre-check result, performing parallel error detection and LDPC-based error correction on the subpage to be detected, and obtaining error detection and correction results, the flash memory error information detection method includes: extracting new error patterns and error statistics based on the error detection and correction results; Based on the new error mode and error statistical information, dynamically updating the initial error feature library to obtain an updated error feature library; Using the latest error features in the updated error feature library, incremental learning is performed on the initial comprehensive error prediction model to obtain an updated comprehensive error prediction model; The updated error feature library and the updated comprehensive error prediction model are used in a subsequent flash memory error information detection process.
9. A flash memory error information detection device, characterized in that: include: A modeling module, used for performing multi-dimensional modeling on a preset flash memory parameter sample set to obtain an initial comprehensive error prediction model, an initial error feature library and a sub-page division strategy; A partitioning and evaluation module, used to perform sub-page partitioning and quality evaluation on the flash memory data according to the initial comprehensive error prediction model, the initial error feature library and the sub-page partitioning strategy, and obtain a sub-page quality score; A risk rating module is used to perform risk rating and pre-check strategy allocation on sub-pages according to the initial comprehensive error prediction model, the initial error feature library and the sub-page quality score to obtain the sub-page pre-check result; wherein, the risk rating and pre-check strategy allocation on sub-pages according to the initial comprehensive error prediction model, the initial error feature library and the sub-page quality score to obtain the sub-page pre-check result includes: dynamically dividing the sub-pages into risk levels according to the sub-page quality score and multiple preset RBER thresholds in the initial error feature library to obtain multiple risk level groups; based on the initial comprehensive error prediction model, combined with the importance and access frequency of the sub-pages, calculating the comprehensive risk index for each risk level group to obtain the expected error probability of each group; according to the expected error probability of each group and the pre-check strategy rules in the initial error feature library, selecting an appropriate pre-check method and number of bits for each risk level group to obtain an adaptive pre-check strategy table; based on the adaptive pre-check strategy table, allocating a corresponding pre-check method to each sub-page and the number of bits, to obtain a sub-page pre-check configuration list; according to the sub-page pre-check configuration list, generate a corresponding error detection code for each sub-page to obtain a sub-page error detection code set; operate the sub-page error detection code set with the original sub-page data to obtain a pre-check result sequence; analyze the pre-check result sequence, mark the sub-page with a negative pre-check result as error-free, and mark the sub-page with a positive pre-check result as a potential error, to obtain a sub-page error mark list; further verify the sub-page with a positive pre-check result to reduce false positives, and update the sub-page error mark list; according to the updated sub-page error mark list, the number of available processing units of the flash memory controller and the size of the memory buffer, prioritize the sub-pages marked as potential errors and add them to the sub-page list that requires further LDPC decoding, and add the sub-pages marked as error-free to the sub-page list that does not require decoding; combine the sub-page list that requires further LDPC decoding and the sub-page list that does not require decoding to obtain a sub-page pre-check result; The error detection module is used to select a sub-page to be detected according to the sub-page pre-check result, perform parallel error detection and LDPC-based error correction on the sub-page to be detected, and obtain error detection and correction results.
