SSD life intelligent prediction and management method and system

Abstract

The invention relates to an SSD intelligent life prediction and management method and system, and belongs to the field of reliability design of nonvolatile memories. According to the method, data in a storage block is divided into three activeness types according to reading times and data retention time, a reading mode prediction module, a data type conversion prediction module and a service life prediction module are provided, and dynamic fitting calculation is performed based on historical data to update prediction module parameters and corresponding RBER. And judging whether the updated RBER exceeds an early warning threshold or not, if so, sending an early warning, and adaptively updating and reading the early warning threshold in combination with the error correction capability and the fitting parameter. According to the method, through multi-dimensional data fusion, data activeness identification, behavior mode prediction and self-adaptive threshold adjustment, accurate prediction and active intervention management of the service life of the storage block of the NAND flash memory are realized, and the reliability, the data security and the operation efficiency of an SSD storage system are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of memory, and in particular to a method and system for intelligently predicting and managing the life of an SSD. Background Art

[0002] With the continuous evolution of NAND flash technology, its high storage density, low cost, and high throughput have made it a core storage medium in data centers, solid-state drives (SSDs), and embedded systems. In recent years, with the widespread adoption of cloud computing, big data analytics, and large-scale artificial intelligence (AI) model inference (such as GPT and DeepSeek), market demand for NAND flash has grown rapidly. In high-frequency access scenarios such as data center log storage and AI inference parameter loading, NAND flash memory must handle hundreds of thousands of read and write operations per second, posing unprecedented challenges to its internal management mechanisms. However, due to the physical limitations of NAND flash memory, two scenarios can occur: first, long-term frequent read operations can induce read disturb, causing the threshold voltage of unread cells to drift; second, long-term inactivity can cause lateral diffusion, resulting in a shift in the threshold voltage of adjacent cells. These two phenomena can gradually accumulate and ultimately lead to uncorrectable bit errors, seriously threatening storage system reliability and data security.

[0003] Existing technologies for SSD lifespan prediction and health management have significant shortcomings. First, most lifespan prediction methods rely solely on single, in-depth metrics, relying solely on P / E (Program / Erase) counts or simple RBER (Raw Bit Error Rate) thresholds. These methods fail to fully consider the combined impact of complex factors such as data retention time, read counts, and read patterns. This results in insufficient prediction accuracy and makes it difficult to provide accurate early warnings before actual failures occur. Second, existing solutions generally lack a granular understanding of the actual activity level of stored data (cold, hot, and warm), leading to crude management strategies that fail to optimize for different data characteristics, resulting in wasted resources and low efficiency. Furthermore, existing SSD health management mechanisms are mostly reactive, intervening only when obvious signs of failure appear. They lack proactive early warning and prevention capabilities, exposing data to a high risk of loss and limiting the controller's ability to adjust parameters to extend lifespan or ensure data integrity. Finally, SSDs generally use fixed early warning thresholds that cannot adapt to the dynamic and nonlinear aging characteristics of NAND flash memory, easily leading to false positives or missed warnings and lacking flexibility and precision in storage block management. These shortcomings collectively limit the long-term reliability and data security of SSDs and increase operating and maintenance costs. Therefore, there is an urgent need for a smarter, more accurate, more proactive and adaptive SSD intelligent life prediction and management method and system. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides an intelligent prediction and management method and system for SSD lifespan, which achieves accurate prediction of NAND flash memory performance degradation and lifespan by integrating multi-dimensional key indicators such as read count, data retention time, and PE count and performing dynamic fitting calculations.

[0005] In order to solve the above technical problems, the present invention adopts a technical solution: a method for intelligent prediction and management of SSD lifespan, comprising the following steps: S01: The SSD controller receives an operation request for a target NAND flash memory block and determines the type of the operation request. If the operation request is for data writing, the controller starts recording the retention time of the data in the storage block after the data writing is completed. If the operation request is for data reading, the controller executes step S02. S02: If data reading is successful, execute step S03; if data reading fails, mark the storage block as a bad block; S03, real-time statistics and determine whether the number of reads of the current storage block exceeds the warning threshold, if so, send a warning signal, if not, execute step S04; S04: Determine the data type being read. If the data type is warm, execute step S05; if the data type is cold, execute step S06; if the data type is hot, execute step S08; S05. Determine the reading mode of the temperature data. If a reading mode exists, fit a model describing the relationship between the P / E times, the reading mode, and the RBER to obtain the current P / E times and the RBER under the reading mode. Then go to step S09. If no reading mode exists, execute step S08. S06: Determine whether there is a cold-hot data conversion. If so, execute step S07. If not, perform regular inspection and execute step S08. S07, dynamically fitting the model describing the relationship between the number of cold data reads and the RBER to obtain the RBER under the current number of reads, and then going to step S09; S08. Dynamically fit a life prediction model. The life prediction model is used to describe the relationship between cold data retention time and RBER during cold data inspection, or the relationship between the number of reads and RBER in a hot data / warm data scenario where no read mode exists. The RBER under the current number of reads is obtained through fitting. S09: Determine whether the RBER exceeds the warning threshold. If it exceeds, send a warning signal; if it does not exceed, end.

[0006] Furthermore, in step S05, the relationship between the P / E times, the read mode and the RBER is described by a three-dimensional degradation relationship model, that is, the relationship between the read mode and the RBER under a certain P / E times, and the relationship between the read mode and the RBER satisfies , x represents the reading mode, the reading mode is represented by the number of continuous readings M, y represents the RBER corresponding to the reading mode, and a and b are fitting parameters; at each fixed P / E number, multiple sets of RBER corresponding to different reading modes are collected, and formula (1) is fitted to obtain the fitting parameters a and b under the P / E number. The reading mode-RBER fitting curves corresponding to different P / E numbers are plotted together in three-dimensional space to construct a complete P / E-reading mode-RBER three-dimensional degradation surface model; after obtaining the fitting parameters a and b under the current P / E number by fitting, the current continuous reading number is substituted into formula (1) to obtain the RBER under the current P / E number and reading mode.

[0007] Furthermore, in step S07, the data retention time and optimal reading times table is first searched, and the optimal reading times in the table are used as segmentation points for segment fitting. Both segment fittings are based on the formula where x represents the number of consecutive readings of cold data after a certain data retention time, y represents the corresponding RBER, and c, d, and e are fitting parameters. At each fixed P / E number and data retention time, RBER data under multiple reading times are collected, and the optimal number of readings corresponding to the current retention time is obtained by looking up the table as the segmentation point. Then, the data before and after the segmentation point are fitted using formula (2), and finally two sets of fitting parameters corresponding to the P / E number and data retention time are obtained. After obtaining the fitting parameters, the current number of consecutive readings is substituted into formula (2) to obtain the RBER corresponding to the current number of readings.

[0008] Furthermore, the table of data retention time and optimal read counts is pre-set data, obtained by performing continuous read tests on NAND particles with different wear times after different data retention times. The optimal read count is defined as the number of reads at which the RBER reaches the minimum value during the continuous read process. This table defines the optimal number of reads for data to be determined as cold data to be converted to hot data under a specific data retention time.

[0009] Furthermore, in step S08, the life prediction model expression is: In the cold data inspection process, x represents the data retention time of cold data, and y represents the RBER under the corresponding data retention time. In the hot data / warm data scenario where there is no read mode, x represents the actual number of reads, and y represents the RBER under the corresponding number of reads. f, g, and h are fitting parameters. The life prediction model is fitted based on historical data to obtain the fitting parameters f, g, and h. After obtaining the fitting parameters, the current number of continuous reads is substituted into formula (3) to obtain the RBER corresponding to the current data retention time or number of reads.

[0010] Furthermore, the initial value of the fitting parameter is 1. During the actual operation, the fitting calculation is performed periodically or event-driven based on the newly collected data to update the fitting parameters.

[0011] Furthermore, in step S05, the data type is determined according to the following criteria: If, within the set historical time T1, the number of reads of the storage block is less than the set value R1, the data retention time exceeds the set value T2, and the access frequency is lower than the set frequency F1, then it is determined to be cold data; If the number of reads of the storage block within the set historical time T1 is greater than the set value R2, it is determined to be hot data; Within the set historical time T1, if the number of reads of the storage block is between [R1, R2], it is determined to be warm data.

[0012] Furthermore, the method further includes step S10, calculating the number of reads in combination with the maximum RBER that can be corrected by the SSD system and the fitting parameters, and updating the read threshold of step S03.

[0013] Furthermore, the initial warning threshold of the number of reads is set to 3000 times.

[0014] The present invention also discloses an intelligent prediction and management system for SSD life, including a data retention management module, a read management module, a data type determiner, a read mode prediction module, a life prediction module, a data type conversion prediction module, a risk warning module, and a model parameter unit; The data retention management module is used to record the retention time of the data in the target storage block after the data is written, and record the time point of each read; The read management module is used to count and determine in real time whether the number of reads of the current storage block exceeds the warning threshold after the data is successfully read. If it exceeds the warning threshold, the risk warning module is triggered; if it does not exceed the warning threshold, the data type determiner is triggered; The data type judger judges the data type of the data being read based on the number of reads and the data retention time. The data types include cold data, warm data and hot data. If it is cold data, it determines whether there is a cold-hot data conversion. If so, it triggers the data type conversion prediction module. If not, it conducts regular inspections and triggers the life prediction module. If it is warm data, it determines whether there is a reading mode. If so, it triggers the reading mode prediction module. If not, it triggers the life prediction module. If it is hot data, it triggers the life prediction module. The reading mode prediction module is used to fit the model describing the relationship between P / E times, reading mode and RBER, and obtain the current P / E times and RBER under the reading mode; The data type conversion prediction module is used to dynamically fit the model describing the relationship between the number of cold data reads and RBER, and obtain the RBER under the current number of reads. The life prediction module is used to dynamically fit the life prediction model. The life prediction model is used to describe the relationship between the cold data retention time and RBER during cold data inspection, or the relationship between the number of reads and RBER in hot data / warm data scenarios without read mode. The RBER under the current number of reads is obtained through fitting. The RBER generated by the reading mode prediction module, the data type conversion prediction module, and the life prediction module is compared with the RBER warning threshold. If the RBER generated by the reading mode prediction module, the data type conversion prediction module, and the life prediction module is greater than the RBER warning threshold, the risk warning module is triggered. If the RBER generated by the reading mode prediction module, the data type conversion prediction module, and the life prediction module is not greater than the RBER warning threshold, the reading threshold of the number of readings is updated in combination with the maximum RBER that the system can correct and the fitting parameters; The risk warning module is used to send warning signals according to trigger signals. The model parameter unit is used to store fitting parameters and update the fitting parameters in combination with the collected data.

[0015] Beneficial Effects of the Invention: This invention overcomes the limitations of traditional SSD lifespan prediction methods, which rely solely on a few simple indicators, suffer from insufficient prediction accuracy, and lack proactive intervention capabilities. The system innovatively incorporates real-time sensing of key parameters such as storage block read counts, data retention time, and read patterns. Based on this, it categorizes data activity into three levels: cold, warm, and hot, enabling dynamic assessment of data usage characteristics and storage block health. Furthermore, the lifespan prediction module dynamically models and fits the relationship between retention time, read counts, and RBER for different data types (e.g., cold data inspection and hot data usage scenarios). By leveraging the controller's ECC error correction capabilities, this module accurately predicts the remaining lifespan of storage blocks. Furthermore, two auxiliary modules are introduced: a read pattern prediction module, targeting warm data, constructs a three-dimensional relationship between P / E times, read patterns, and RBER. This is used to assess reliability evolution under varying read intensities and proactively predict reliability risks caused by read patterns. A data type conversion prediction module, targeting cold data, combines a "data retention time and optimal read count table" for piecewise fitting, providing a quantitative basis for dynamic read control and extending NAND flash memory lifespan. The present invention not only transforms the storage system from passive fault response to active risk perception and strategy optimization, but also improves the reliability, data security and system operation efficiency of SSD. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of the method described in Example 1; Figure 2 This is a functional block diagram of the system described in Example 2; Figure 3 It is a schematic diagram of the corresponding relationship between the data retention time and the optimal number of reads in a NAND storage block; Figure 4 This is a schematic diagram of the prediction results of the data type conversion module; Figure 5 This is a schematic diagram of the prediction results of the reading mode module; Figure 6 This is a schematic diagram of the prediction results of the life prediction module; Figure 7 This is a schematic diagram of the prediction accuracy loss of the data model conversion module; Figure 8 This is a schematic diagram of the prediction accuracy loss of the reading mode module; Figure 9 This is a schematic diagram of the prediction accuracy loss of the life prediction module. DETAILED DESCRIPTION

[0017] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 This method discloses an intelligent prediction and management method for SSD lifespan. This method realizes accurate prediction of NAND flash memory performance degradation and lifespan by integrating multi-dimensional key indicators such as the number of reads, data retention time, and the number of PEs and performing dynamic fitting calculations. In addition, the present invention also introduces a read mode prediction module and a data type conversion prediction module to intelligently predict future storage behaviors, and can realize active and adaptive risk warning and management based on dynamically adjusted thresholds. The present invention aims to significantly improve the reliability, data security and overall operating efficiency of the SSD storage system, and effectively reduce the risk of data loss. The present invention significantly improves the long-term reliability, data integrity and overall service life of NAND flash memory, thereby better supporting emerging application scenarios such as AI reasoning that have strict requirements on storage system performance and reliability.

[0019] This method adds multiple modules to the SSD controller, namely, data retention management module, read management module, risk warning module, read mode prediction module, life prediction module, data type determiner, data type conversion prediction module, and model parameter unit. Each module interacts with each other through a preset interface to ensure that the system can read data efficiently. Figure 1 As shown, the method includes the following steps: S01. When the SSD controller receives an operation request from the host for a target NAND flash memory block, the instruction analysis and task scheduling unit determines the operation request type. If the operation request is a write operation request, the instruction analysis and task scheduling unit directs the task flow to the write operation path and executes step S02 to initiate data retention management. If the operation request is a data read request, the instruction analysis and task scheduling unit directs the task flow to the data read path and executes step S03 to preliminarily determine the read status.

[0020] S02. For a write operation request, the data retention management module is activated. Once the data is successfully written to the target NAND flash memory block, the data retention time tracker in the data retention management module begins accurately timing, recording the retention time of the data in the memory block since the write completion, and simultaneously recording the time of each read. The recorded data retention time and read time are crucial. They not only serve as important parameters for the subsequent life prediction module to evaluate the health status of the memory block, but also provide key judgment basis for the data type determiner to quickly identify the data activity type and the data type conversion prediction module.

[0021] In response to a data read request, the SSD controller performs the physical operation of reading data from the specified NAND flash memory block. During this process, the success of the read operation is monitored in real time. For example, this is determined by analyzing the status code returned by the NAND chip and checking the ECC (error correction code) results (e.g., whether uncorrectable errors occurred). If the read fails, indicating that the block may have a serious defect, the system immediately executes step S04 to mark it as bad. If the read succeeds, indicating that the current read operation has completed successfully, the instruction analysis and task scheduling unit directs the task flow to a more in-depth health assessment phase, instructing the system to execute step S05 to activate the read management module.

[0022] S04: If the target storage block fails to be read, it will be marked as a bad block. Once marked, the storage block will be logically isolated and the SSD controller will no longer attempt to write any new data to it. The bad block management mechanism of the SSD system will be activated to protect data from potential damage.

[0023] S05. Whether the system proceeds through the data retention management path in step S02 or the path after a successful read in step S03, it will enter the jurisdiction of the read management unit. The core responsibility of the read management unit is to maintain real-time and accurate statistics on the cumulative number of reads of the target storage block. Based on this, the read management unit continuously determines whether the cumulative number of reads of the current storage block has exceeded a preset "read warning threshold." This warning threshold is initially set to 3000 times and is a key parameter pre-set by the SSD controller based on the expected lifespan and reliability standards of the NAND flash memory. This setting is primarily based on the number of reads indicated in the flash memory manufacturer's datasheet, the recommended safe upper limit, and reliability test data of different flash memory chips in actual application scenarios. If the number of reads exceeds the warning threshold, step S06 is executed. If the number of reads does not exceed the warning threshold, the system directs the task flow further, instructing the system to proceed to step S07 to determine the data type and conduct a more detailed assessment.

[0024] S06. The risk warning unit is immediately activated and sends an early warning signal to the SSD controller, prompting that the storage block has a potential risk.

[0025] S07. Using the number of reads provided by the read management unit and the data retention time provided by the data retention time tracker, the data type determiner comprehensively determines the activity of the data in the storage block and classifies it as cold data, warm data, or hot data.

[0026] The basis for data type judgment is: within the set historical time T1, if the number of reads of the storage block is less than the set value R1, the data retention time exceeds the set value T2, and the access frequency is lower than the set frequency F1, it is judged as cold data; within the set historical time T1, if the number of reads of the storage block is greater than the set value R2, it is judged as hot data; within the set historical time T1, if the number of reads of the storage block is between [R1, R2], it is judged as warm data.

[0027] In this embodiment, the judgment criteria are as follows: If the storage block has had very few or zero reads over the past seven days, and the data retention period exceeds 90 days, indicating extremely low access frequency, likely archival or historical data, it is considered cold data, and the system will proceed directly to step S08 to determine whether the data type is hot or cold. For warm data, which is between hot and cold data, meaning both read frequency and data retention period are moderate, the system will proceed to step S10 to determine and predict the warm data pattern. If the storage block has had a very high read count over the past seven days, indicating frequent access and typically data currently being actively used by users, it is considered hot data, and the system will proceed to step S11.

[0028] S08: To accurately identify trends in data activity transitioning from cold data to hot data, the system continuously monitors data access patterns within storage blocks. When data is suddenly and significantly and continuously read after a long period of inactivity, the system determines that there is a trend of cold data transitioning to hot data. If this trend is identified, the system further records the number of consecutive reads after the data retention period and executes step S09. If there is no trend of transitioning from cold to hot data, meaning the data remains cold, the system will perform regular inspections (e.g., every seven days) to continuously monitor changes in activity. This ensures that the long-term storage status of inactive data is continuously tracked and potential changes are promptly detected, and then executes step S11.

[0029] S09, query the preset data retention time and optimal reading times table. Figure 3 , this graph defines the optimal number of reads at which data can be effectively determined to be converted from cold data to hot data under a specific data retention time. This table is key data preset in advance by the NAND controller system, and is obtained by the manufacturer through continuous reading tests on NAND particles with different wear experiences at different data retention times. The optimal number of reads is defined as the number of reads at which the RBER reaches the minimum value during the continuous reading process. If the number of continuous reads of the current record is ≥ the optimal number of reads, the data type conversion prediction module is started. This module adopts a piecewise function fitting strategy, using the "optimal number of reads" provided in the "data retention time and optimal number of reads table" as the segmentation point to model the relationship between the number of reads and RBER. Both functions are based on The equation is used for fitting. Here, x represents the number of consecutive reads of cold data after a certain hold time, y represents the corresponding RBER, and c, d, and e are fitting parameters. Under each fixed P / E number and data hold time condition, RBER sample data is collected under multiple read times, and the optimal number of reads corresponding to the current hold time is obtained by table lookup as the segmentation point. Subsequently, the data before and after the segmentation point are fitted with the above formula using the least squares method, ultimately obtaining two sets of fitting parameters corresponding to the specific hold time under this wear level (i.e., c1, d1, e1 and c2, d2, e2 for the two segments before and after). After obtaining the fitting parameters, the current number of consecutive reads is substituted into equation (2) to obtain the RBER corresponding to the current number of reads. Figure 4 Schematic diagram of the prediction results of the data type conversion prediction module in this embodiment, which obtains the relationship curve between RBER and reading times when the P / E times are 100 / 1000 / 2000.

[0030] When the data type conversion prediction module is initialized, the fitting parameters are all set to 1, and the parameters are continuously updated in subsequent use. This module can accurately characterize the degradation trend of cold data under different read intensities, especially capturing the inflection point of nonlinear changes before and after the optimal number of reads. This helps to improve the robustness of life prediction and provide a quantitative basis for dynamic read control, extending the life of NAND flash memory. If the number of consecutive reads currently recorded is less than the optimal number of reads, indicating that the conversion conditions have not yet been fully met, the data conversion prediction model will continue to wait until the number of consecutive reads reaches the optimal number of reads, and then fully start the data conversion prediction module and then execute step S13).

[0031] S10. Analyze the historical read behavior patterns of the warm data (e.g., access time interval, locality of accessed data blocks, etc.) to determine whether there is a recognizable read pattern (e.g., whether there is periodic access with an interval of n minutes). If the read pattern prediction model determines that a read pattern exists, the read pattern prediction module will be officially activated to further learn and attempt to predict the future read behavior pattern of the data block. The read pattern prediction module constructs a three-dimensional degradation relationship model between the P / E number, the read pattern, and the RBER. Among them, the relationship between the read pattern and the RBER satisfies the formula , where x represents the reading mode (usually M consecutive readings, i.e., M consecutive readings after the data is held for M minutes), y represents the RBER corresponding to the reading mode, and a and b are fitting parameters. At each fixed P / E number, the system collects multiple sets of RBER data corresponding to different reading modes. Least squares fitting is performed using this formula to obtain parameters a and b at that wear level. The reading mode-RBER fitting curves corresponding to different P / E numbers are plotted together in three-dimensional space to construct a complete three-dimensional degradation surface model of P / E-reading mode-RBER. After obtaining the fitting parameters a and b, the current consecutive reading number is substituted into formula (1) to obtain the RBER corresponding to the current reading number.

[0032] The read pattern prediction model characterizes the nonlinear impact of read patterns on reliability degradation under varying flash memory wear conditions, providing data support for read strategy adjustments. When the read pattern prediction model is initialized, all parameters are set to 1, and the parameters are continuously updated during subsequent use. This prediction helps the SSD controller optimize prefetch strategies, adjust cache management, or perform intelligent data rearrangement to improve performance, thus executing step S13. If the read pattern prediction module determines that no read pattern exists, indicating that read behavior is highly random, the module proceeds directly to step S11, where parameters are recorded and the lifespan prediction module is activated. Figure 5 FIG. 4 is a schematic diagram of prediction results of the read mode prediction module in this embodiment, which describes the three-dimensional degradation relationship between the P / E number, the read mode and the RBER.

[0033] S11, the life prediction module is based on the number of reads provided by the read management unit and the data retention time provided by the data retention time tracker. This step marks the full involvement of the life prediction module. The core function of this module is to systematically collect and record key parameters closely related to the health of the storage block: RBER, data retention time, number of reads, etc. When using the service life prediction module, the formula Fitting is performed. There are two scenarios. Scenario 1: During the cold data inspection process, x represents the cold data retention time, and y represents the RBER at the corresponding retention time. This scenario is used to evaluate the reliability change trend of long-term static data and is suitable for life monitoring of read-only data or archived data. Scenario 2: Hot data / warm data scenario with no read mode: x represents the actual number of reads generated, and y represents the corresponding RBER. This scenario is used to evaluate the degradation rate of hot data blocks under continuous reading and is suitable for life prediction and scheduling optimization of dynamic data. Parameters f, g, and h are all obtained from historical sampling data using the least squares fitting method. By constructing this unified functional form of degradation model, the module can flexibly evaluate the remaining life and degradation degree of the block in different application scenarios, providing the controller with accurate life warning and scheduling recommendations. After obtaining the fitting parameters, the current number of continuous reads is substituted into formula (3) to obtain the RBER corresponding to the current number of reads.

[0034] refer to Figure 6 , showing the predicted relationship between read count and RBER. When the prediction module is initialized, all parameters are set to 1 and are continuously updated during subsequent use. The lifespan prediction module also integrates other internal parameters related to the physical characteristics of NAND flash memory (such as the number of erase / write cycles). These parameters, accurately recorded by the lifespan prediction module, serve as the foundation for the dynamic fitting calculation in the subsequent step m).

[0035] S12: The model parameter unit collects the latest fitting parameters, and then executes step S13 to determine whether the updated RBER exceeds the warning threshold.

[0036] S13. After completing the fitting calculation and parameter update, each module evaluates its newly calculated or predicted RBER to determine whether it has reached or exceeded the preset "RBER warning threshold." The initial RBER warning threshold is set based on the maximum RBER value that can be corrected by the ECC in the NAND controller. If the RBER exceeds the warning threshold, indicating that the error rate of the storage block may have reached a dangerous level, the risk warning unit will be reactivated, immediately sending a higher-level warning signal to the SSD controller, and the system will continue to execute step S06. If the RBER does not exceed the preset RBER warning threshold, indicating that the current RBER is still within the acceptable range, step S14 is executed to adaptively update the reading warning threshold.

[0037] In step S14, the model parameter unit combines the maximum RBER that the SSD controller can correct and uses the corresponding formula in the life prediction module to calculate a more accurate and adaptive read warning threshold. This newly calculated threshold is then used to update the "read warning threshold" used in step S05. Through this dynamic adjustment, the system can more flexibly adjust the warning sensitivity based on the actual aging status of the storage block and feedback from the life prediction model, ensuring the timeliness and accuracy of warnings.

[0038] In the above description, RBER represents the raw bit error rate, which is the bit error rate compared with the expected data read from the flash memory before error correction code correction.

[0039] All of the above steps together form a complete, closed-loop process for intelligent NAND flash storage block management and prediction. The SSD controller continuously receives host operation requests and executes these steps in a continuous loop in the background, dynamically monitoring, assessing health, providing risk warnings, and adaptively managing each NAND flash storage block. This intelligent mechanism enables the system to proactively identify and address potential storage block issues, significantly improving the overall reliability of the SSD at the hardware level and ensuring the security of user data.

[0040] In the above description, the number of continuous reads refers to continuous reads performed after a long period of data retention. The actual number of reads includes continuous reads or non-continuous reads without a long period of data retention.

[0041] The prediction method proposed in this invention was applied to a 3D NAND flash memory and the results were as follows: like Figure 7 As shown in the figure, the data type conversion prediction module of the present invention performs well. Without distinguishing complex factors such as the number of P / E times, data retention time, and the number of consecutive reads, the prediction accuracy loss of 90% of the samples in the measured samples is only 2.56%, which verifies its accuracy in identifying data activity conversion. Figure 8 As shown in Figure 1, the prediction effect of the read mode prediction module of the present invention is shown. Without distinguishing the number of PEs, read modes, etc., the prediction accuracy loss of 90% of the samples in the measured samples is 14.98%. Figure 9 The figure below shows the prediction results of the lifespan prediction module of the present invention. Without distinguishing between the number of PEs and the number of reads, the prediction accuracy loss for 90% of the measured samples was 5.98%. The above description of not distinguishing between the number of PEs, data retention time, and number of consecutive reads means that the number of PEs, data retention time, and number of consecutive reads were all recorded independently during the experiment; however, in the prediction loss graph, the loss accuracy results calculated for different FE times, different data retention times, and different consecutive reads are integrated and superimposed on the same graph.

[0042] Further analysis shows that the present invention achieves full-link intelligence from prediction and warning to management by virtue of its high prediction accuracy and early warning capabilities, intelligent data management and optimization (such as data type judgment, reading mode and data type conversion prediction), and enhanced adaptability and robustness (through dynamic adjustment of thresholds and model parameters). It effectively reduces the risk of data corruption and loss, maximizes the usable life of the SSD, and thus comprehensively improves the reliability, data security and user experience of the SSD. In particular, it can better support emerging application scenarios such as AI reasoning that have strict requirements on storage performance and reliability.

[0043] Example 2 This embodiment discloses an intelligent prediction and management system for SSD lifespan, such as Figure 2 As shown, it includes a data retention management module, a read management module, a data type determiner, a read mode prediction module, a life prediction module, a data type conversion prediction module, a risk warning module, and a model parameter unit.

[0044] The data retention management module is used to record the retention time of the data in the target storage block after the data writing is completed, and record the time point of each reading.

[0045] The read management module is used to count and determine in real time whether the number of reads of the current storage block exceeds the warning threshold after the data is successfully read. If it exceeds the warning threshold, the risk warning module is triggered; if it does not exceed the warning threshold, the data type determiner is triggered.

[0046] The data type judge determines the data type of the data being read based on the number of reads and the data retention time. The data types include cold data, warm data and hot data. If it is cold data, it determines whether there is a cold-hot data conversion. If so, the data type conversion prediction module is triggered. If not, regular inspection is performed and the life prediction module is triggered. If it is warm data, it determines whether there is a reading mode. If so, the reading mode prediction module is triggered. If not, the life prediction module is triggered. If it is hot data, the life prediction module is triggered.

[0047] The reading mode prediction module is used to fit the model describing the relationship between P / E times, reading mode and RBER, and obtain the current P / E times and RBER under the reading mode.

[0048] The read data type conversion prediction module is used to dynamically fit the model describing the relationship between the number of cold data reads and RBER to obtain the RBER under the current number of reads.

[0049] The life prediction module is used to dynamically fit the life prediction model. The life prediction model is used to describe the relationship between the cold data retention time and RBER during cold data inspection, or the relationship between the number of reads and RBER in hot data / warm data scenarios without a read mode. The RBER under the current number of reads is obtained through fitting.

[0050] The reading pattern prediction module, data type conversion prediction module, and life prediction module compare the generated RBER with the RBER warning threshold. If the RBER generated by the reading pattern prediction module, data type conversion prediction module, and life prediction module is greater than the RBER warning threshold, the risk warning module is triggered. If the RBER generated by the reading pattern prediction module, data type conversion prediction module, and life prediction module is not greater than the RBER warning threshold, the reading threshold of the number of reads is updated in combination with the maximum RBER that the system can correct and the fitting parameters.

[0051] The risk warning module is used to send warning signals according to trigger signals.

[0052] The model parameter unit is used to store fitting parameters and update the fitting parameters in combination with the collected data.

[0053] This system can implement the method described in Example 1. The specific content of the method is described in Example 1.

[0054] The above description is only the basic principle and preferred embodiments of the present invention. Improvements and substitutions made by those skilled in the art based on the present invention fall within the protection scope of the present invention.

Claims

1. A method for intelligent prediction and management of SSD lifespan, characterized by: The following steps are involved: S01: The SSD controller receives an operation request for a target NAND flash memory block and determines the type of the operation request. If the operation request is for data writing, the controller starts recording the retention time of the data in the storage block after the data writing is completed. If the operation request is for data reading, the controller executes step S02. S02: If data reading is successful, execute step S03; if data reading fails, mark the storage block as a bad block; S03, real-time statistics and determine whether the number of reads of the current storage block exceeds the warning threshold, if so, send a warning signal, if not, execute step S04; S04: Determine the data type being read. If the data type is warm, execute step S05; if the data type is cold, execute step S06; if the data type is hot, execute step S08; S05. Determine the reading mode of the temperature data. If a reading mode exists, fit a model describing the relationship between the P / E times, the reading mode, and the RBER to obtain the current P / E times and the RBER under the reading mode. Then go to step S09. If no reading mode exists, execute step S08. S06: Determine whether there is a cold-hot data conversion. If so, execute step S07. If not, perform regular inspection and execute step S08. S07, dynamically fitting the model describing the relationship between the number of cold data reads and the RBER to obtain the RBER under the current number of reads, and then going to step S09; S08. Dynamically fit a life prediction model. The life prediction model is used to describe the relationship between cold data retention time and RBER during cold data inspection, or the relationship between the number of reads and RBER in a hot data / warm data scenario where no read mode exists. The RBER under the current number of reads is obtained through fitting. S09: Determine whether the RBER exceeds the warning threshold. If it exceeds, send a warning signal; if it does not exceed, end.

2. The SSD lifespan intelligent prediction and management method according to claim 1, characterized in that: In step S05, the relationship between the P / E times, the read mode and the RBER is described by a three-dimensional degradation relationship model, that is, the relationship between the read mode and the RBER under a certain P / E times, and the relationship between the read mode and the RBER satisfies , x represents the reading mode, the reading mode is represented by the number of continuous readings M, y represents the RBER corresponding to the reading mode, and a and b are fitting parameters; at each fixed P / E number, multiple sets of RBER corresponding to different reading modes are collected, and formula (1) is fitted to obtain the fitting parameters a and b under the P / E number. The reading mode-RBER fitting curves corresponding to different P / E numbers are plotted together in three-dimensional space to construct a complete P / E-reading mode-RBER three-dimensional degradation surface model; after obtaining the fitting parameters a and b under the current P / E number by fitting, the current continuous reading number is substituted into formula (1) to obtain the RBER under the current P / E number and reading mode.

3. The SSD lifespan intelligent prediction and management method according to claim 1, characterized in that: In step S07, first query the data retention time and optimal reading times table, and use the optimal reading times in the table as the segmentation points for segment fitting. Both segment fittings are based on the formula Where x represents the number of consecutive readings of cold data after a certain data retention time, y represents the RBER corresponding to the number of readings, c, d and e are fitting parameters. At each fixed P / E number and data retention time, RBER data under multiple reading times are collected, and the optimal number of readings corresponding to the current retention time is obtained by looking up the table as the segmentation point. Then, the data before and after the segmentation point are fitted using formula (2), and finally two sets of fitting parameters corresponding to the P / E number and data retention time are obtained. After obtaining the fitting parameters, the current number of consecutive readings is substituted into formula (2) to obtain the RBER corresponding to the current number of readings.

4. The SSD lifespan intelligent prediction and management method according to claim 3, characterized in that: The table of data retention time and optimal read counts is preset data, obtained by performing continuous read tests on NAND chips with different wear times after different data retention times. The optimal read count is defined as the number of reads during the continuous read process at which the RBER reaches the minimum. This table defines the optimal number of reads required to determine that data is cold data and then converted to hot data under a specific data retention time.

5. The SSD lifespan intelligent prediction and management method according to claim 1, characterized in that: In step S08, the life prediction model expression is: In the cold data inspection process, x represents the data retention time of cold data, and y represents the RBER under the corresponding data retention time. In the hot data / warm data scenario where there is no read mode, x represents the actual number of reads, and y represents the RBER under the corresponding number of reads. f, g, and h are fitting parameters. The life prediction model is fitted based on historical data to obtain the fitting parameters f, g, and h. After obtaining the fitting parameters, the current number of continuous reads is substituted into formula (3) to obtain the RBER corresponding to the current data retention time or number of reads.

6. The method for intelligent prediction and management of SSD lifespan according to any one of claims 2, 3, and 4, characterized in that: The initial value of the fitting parameter is 1. During the actual operation, the fitting calculation is performed periodically or event-driven based on the newly collected data to update the fitting parameters.

7. The SSD lifespan intelligent prediction and management method according to claim 1, characterized in that: In step S05, the data type is determined according to the following criteria: If, within the set historical time T1, the number of reads of the storage block is less than the set value R1, the data retention time exceeds the set value T2, and the access frequency is lower than the set frequency F1, then it is determined to be cold data; If the number of reads of the storage block within the set historical time T1 is greater than the set value R2, it is determined to be hot data; Within the set historical time T1, if the number of reads of the storage block is between [R1, R2], it is determined to be warm data.

8. The SSD lifespan intelligent prediction and management method according to claim 1, characterized in that: The method further includes step S10 , calculating the number of reads in combination with the maximum RBER that can be corrected by the SSD system and the fitting parameters, and updating the read threshold of step S03 .

9. The SSD lifespan intelligent prediction and management method according to claim 1 or 8, characterized in that: The initial warning threshold for the number of reads is set to 3000 times.

10. An intelligent SSD life prediction and management system, characterized by: It includes data retention management module, reading management module, data type judge, reading mode prediction module, life prediction module, data type conversion prediction module, risk warning module, and model parameter unit; The data retention management module is used to record the retention time of the data in the target storage block after the data is written, and record the time point of each read; The read management module is used to count and determine in real time whether the number of reads of the current storage block exceeds the warning threshold after the data is successfully read. If it exceeds the warning threshold, the risk warning module is triggered; if it does not exceed the warning threshold, the data type determiner is triggered; The data type judger judges the data type of the data being read based on the number of reads and the data retention time. The data types include cold data, warm data and hot data. If it is cold data, it determines whether there is a cold-hot data conversion. If so, it triggers the data type conversion prediction module. If not, it conducts regular inspections and triggers the life prediction module. If it is warm data, it determines whether there is a reading mode. If so, it triggers the reading mode prediction module. If not, it triggers the life prediction module. If it is hot data, it triggers the life prediction module. The reading mode prediction module is used to fit the model describing the relationship between P / E times, reading mode and RBER, and obtain the current P / E times and RBER under the reading mode; The data type conversion prediction module is used to dynamically fit the model describing the relationship between the number of cold data reads and RBER, and obtain the RBER under the current number of reads. The life prediction module is used to dynamically fit the life prediction model. The life prediction model is used to describe the relationship between the cold data retention time and RBER during cold data inspection, or the relationship between the number of reads and RBER in hot data / warm data scenarios without read mode. The RBER under the current number of reads is obtained through fitting. The RBER generated by the reading mode prediction module, the data type conversion prediction module, and the life prediction module is compared with the RBER warning threshold. If it is greater than the RBER warning threshold, the risk warning module is triggered. If it is not greater than the RBER warning threshold, the reading threshold of the number of readings is updated based on the maximum RBER that the system can correct and the fitting parameters; The risk warning module is used to send warning signals according to trigger signals. The model parameter unit is used to store fitting parameters and update the fitting parameters in combination with the collected data.

Images