A method and system for managing bad blocks in NAND Flash based on artificial intelligence

Abstract

This invention discloses an artificial intelligence-based NAND Flash bad block management method and system, belonging to the field of solid-state storage technology. The method includes the following steps: S1, collecting and preprocessing multi-dimensional data of bad pages based on preset parameters and preset indicators; S2, inputting the preprocessed multi-dimensional data into a spatiotemporal graph convolutional network model and outputting the failure probability of each page; S3, determining a hierarchical isolation strategy based on the failure probability obtained in step S2, followed by dynamic loading of error correction codes and preventive data migration. This invention improves storage utilization: actual tests show that in a block with a single bad page, more than 85% of normal pages can be recycled, increasing the effective capacity by 10 times compared to traditional solutions; by reducing the abandonment of the entire block due to local bad pages, the overall chip lifespan is extended by about 30%. It also has real-time risk management capabilities, with the AI ​​model achieving a 92% accuracy rate in predicting potential risk areas, reducing the secondary failure rate by 70% compared to static isolation solutions.

Description

Technical Field

[0001] This invention relates to the field of solid-state storage technology, and more specifically to a method and system for managing bad blocks in NAND Flash based on artificial intelligence. Background Technology

[0002] NAND Flash memory is widely used in various electronic devices. With the development of 3D NAND technology, its storage density and capacity have increased significantly. However, this has also brought a series of new challenges, especially in bad block management. In traditional NAND Flash memory, bad block management is a critical issue. With the introduction of 3D NAND technology, the number of pages within a single block has increased dramatically, for example, from 64 layers to more than 256 layers. This high-density storage structure leads to the following problems:

[0003] 1) Bad Page Diffusion Effect: When a single Page is damaged due to programming failure, read interference, or charge leakage, its physically adjacent Pages may fail rapidly in subsequent operations due to crosstalk or charge migration. This bad Page Diffusion Effect is particularly significant in high-density 3D NAND structures because the Pages are closer together and the coupling effect is stronger.

[0004] 2) Significant resource waste: Traditional bad block management strategies mark the entire block as unusable, causing thousands of normal pages to be discarded as well. For example, in a certain 512GB chip, a single block contains 1024 pages, and just one bad page results in the loss of 2MB of usable space, leading to a significant loss in yield.

[0005] 3) Limitations of static isolation: Existing technologies use fixed-range isolation (such as isolating N pages before and after a bad page), which cannot dynamically adapt to the actual impact range of different processes, data loads or aging stages, resulting in insufficient isolation (residual risk) or excessive isolation (waste of resources). Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for managing bad blocks in NAND Flash based on artificial intelligence.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] In a first aspect, the present invention discloses a NAND Flash bad block management method based on artificial intelligence, comprising the following steps:

[0009] S1. Based on preset parameters and preset indicators, multi-dimensional data is collected from bad Pages. A Page represents a page, which is the smallest unit for performing read operations. Then, the collected multi-dimensional data is preprocessed.

[0010] S2. Train the spatiotemporal graph convolutional network model. Load the pre-trained spatiotemporal graph convolutional network model and input the pre-processed multi-dimensional data into the spatiotemporal graph convolutional network model. The spatiotemporal graph convolutional network model extracts and transforms the pre-processed multi-dimensional data through graph convolutional layers and temporal convolutional layers, and outputs the failure probability of each page.

[0011] S3. Based on the failure probability obtained in step S2, determine the directly affected area and the potential risk area. For Pages in the directly affected area, take isolation measures and prohibit read and write operations. For Pages in the potential risk area, allow reading but restrict writing, and start real-time health monitoring. At the same time, reorganize the remaining normal Pages across block space to form virtual blocks to maximize the utilization of storage space. Then, perform dynamic loading of error correction codes and preventive data migration.

[0012] Based on the first aspect, step S1 specifically includes the following steps:

[0013] S11. Physical feature acquisition: Obtain and store the three-dimensional coordinates of the bad page and the coupling capacitance measurement value of the adjacent page to obtain physical feature data.

[0014] S12. Electrical feature acquisition: Measure and record the threshold voltage distribution distortion of bad pages and the voltage offset when reading fails, and obtain electrical feature data.

[0015] S13. Environmental Feature Acquisition: Record and save temperature and power supply voltage fluctuations during operation to obtain environmental feature data;

[0016] S14. Timing correlation analysis: Collect the operation history of the Block before and after the bad page, including the number of erase / write operations, the frequency of read operations, and the programming order of adjacent pages to obtain timing correlation data; Block represents a block, which consists of multiple pages and is the smallest addressable unit of the erase operation;

[0017] S15. Preprocess the collected physical characteristic data, electrical characteristic data, environmental characteristic data, and time-series correlation data, including cleaning and normalization, to remove outliers and noise data.

[0018] Based on the first aspect, the training of the spatiotemporal graph convolutional network model in step S2 specifically includes the following steps:

[0019] S211. Training data acquisition: In a controllable experimental environment, pre-defined types of bad pages are generated, including programming failure bad pages, read interference bad pages, charge leakage bad pages, repeated erase and write aging bad pages, and temperature-sensitive bad pages. The diffusion path and influence range of the pre-defined types of bad pages are recorded as the first training sample data.

[0020] S212. By collecting data on the spread of bad pages caused by the natural aging of terminal devices in actual use environments, the first training sample data is supplemented to obtain the second training sample data.

[0021] S213. The training loss function of the spatiotemporal graph convolutional network model is the mean squared error loss function. Where N is the number of samples in the second training sample data. Indicates the actual failure situation. This represents the failure probability predicted by the model, used to measure the difference between the predicted failure probability and the actual failure situation.

[0022] S214. The spatiotemporal graph convolutional network model includes an input layer, a graph convolutional layer, a temporal convolutional layer, and an output layer. The parameters of the spatiotemporal graph convolutional network model are updated and optimized using the Adam optimization algorithm.

[0023] Based on the first aspect, step S2 further includes the following steps:

[0024] S221. Model the preprocessed multi-dimensional data into a three-dimensional mesh diagram, with each Page as a node in the three-dimensional mesh diagram. The node attributes include the Page's health status and electrical parameters. The edge weights reflect the coupling strength between Pages.

[0025] S222. Receive a three-dimensional mesh graph through the input layer, including the attribute information of all nodes and the edge weight matrix between nodes;

[0026] S223. Perform graph convolution operation on the received 3D mesh graph through the graph convolution layer. The node features are aggregated and updated to obtain spatial features, where This represents the feature matrix of the l-th layer node. This represents the adjacency matrix after adding self-joins. This represents the corresponding degree matrix. Let σ represent the weight matrix of the l-th layer, and σ be the activation function.

[0027] S224. The temporal convolutional layer is used to model the state changes of the Page at different time steps, further improving the accuracy of the model's prediction of the impact range of bad pages;

[0028] S225. Based on the model's prediction target, the output layer outputs the failure probability of each Page within a preset time, dividing the Page into directly affected areas and potential risk areas.

[0029] Based on the first aspect, step S3 includes the following steps:

[0030] S31. For areas directly affected, immediately isolate them and prohibit reading and writing; for potentially risky areas, allow reading but restrict writing, and enable real-time health monitoring.

[0031] S32. Construct a secondary mapping layer from logical address to physical address, and combine green area Pages from different Blocks into new virtual blocks. During the reorganization process, first select Pages with similar programming / erase cycles for combination.

[0032] S33. Each virtual block retains 5% to 10% of spare pages. When a potentially risky page is upgraded to a directly affected area, the replacement mechanism is automatically triggered, and the spare page is used to replace it.

[0033] Based on the first aspect, the dynamic loading of error correction codes in step S3 includes: using the default BCH code for the green area Page, upgrading the potential risk area Page to the LDPC code, and enabling the progressive adjustment algorithm for the read voltage.

[0034] Based on the first aspect, the preventive data migration described in step S3 includes: automatically migrating data to the green area when the predicted monthly failure rate of the Page in the potential risk area exceeds 15%.

[0035] Secondly, this invention discloses an artificial intelligence-based NAND Flash bad block management system, used in the artificial intelligence-based NAND Flash bad block management method described in the first aspect, comprising:

[0036] The bad page feature acquisition module is used to collect multi-dimensional data and preprocess the collected multi-dimensional data;

[0037] The AI ​​dynamic impact prediction model module is used to train and optimize the spatiotemporal graph convolutional network model. Then, through the trained spatiotemporal graph convolutional network model, features are extracted and transformed from the preprocessed multi-dimensional data, and the failure probability of each page is output.

[0038] The local isolation and virtual block reorganization module determines the hierarchical isolation strategy based on the failure probability output by the model. At the same time, it reorganizes the remaining normal pages across the block space to form virtual blocks.

[0039] An adaptive error correction enhancement module is used for dynamic loading of error correction codes and preventative data migration.

[0040] The beneficial effects of this invention are:

[0041] 1) This invention improves storage utilization. Actual tests show that in a block with a single bad page, more than 85% of the normal pages can be recycled, which is 10 times more effective than traditional solutions.

[0042] 2) By reducing the amount of chip scrap caused by local bad pages, this invention can extend the overall lifespan of the chip by about 30%.

[0043] 3) This invention has the function of real-time risk management. The AI ​​model has a prediction accuracy of 92% for potential risk areas, which reduces the secondary failure rate by 70% compared with the static isolation solution. Attached Figure Description

[0044] Figure 1 This is a schematic diagram illustrating the steps of an artificial intelligence-based NAND Flash bad block management method according to an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of the structure of an artificial intelligence-based NAND Flash bad block management system according to an embodiment of the present invention. Detailed Implementation

[0046] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] In NAND Flash bad block management, bad block marking and isolation are crucial steps. Existing technologies typically employ two methods: 1) Boot Block Recording: Bad block information is recorded in the NAND Flash's boot block or a dedicated bad block marking area. Each block may have a bad block flag indicating whether it is a bad block. For example, some status bytes are stored in the block header to indicate whether the block has been marked as bad. 2) Bad Block Table: A bad block table is maintained externally to the NAND Flash or in the controller's memory. This table records the addresses or block indices of all blocks marked as bad. The controller checks this table each time it performs a read or write operation on the NAND Flash to avoid accessing bad blocks.

[0048] Bad block isolation and replacement: Once a bad block is identified, measures must be taken to isolate and replace it. Common bad block replacement strategies include: 1) Block replacement: Whenever a bad block is found, the system selects a new free block to replace it. During the replacement process, valid data in the bad block can be migrated to the new block (data migration). In specific implementation, the system needs to maintain a block mapping table to record the mapping relationship between each logical block address (LBA) and physical block address. 2) Write log or write mapping table: To ensure that no data loss occurs after bad block replacement, the system should use a write log or mapping table to record the mapping between logical addresses and physical addresses. Logical addresses are mapped to physical addresses. If a physical address points to a bad block, the new location can be found by looking up the mapping table. 3) Wear leveling: To avoid premature damage caused by frequent writes to certain blocks, bad block management strategies should include wear leveling algorithms (such as Wear Leveling). Wear leveling extends the lifespan of the entire Flash memory by balancing the number of erases and writes to different blocks.

[0049] Bad block management strategies include: 1) Bad block detection frequency: The bad block detection frequency can be adjusted according to the specific requirements of the system. For some high-reliability applications, the bad block detection frequency can be increased, while for general consumer-grade storage, stable operation can be maintained through periodic scanning and online detection. 2) Free block management: To improve performance, the system needs to maintain a list of free blocks. This list records all free blocks that can be used for new data writing. When a bad block is discovered and replaced, a new free block should be selected from this list. 3) Adaptive bad block management: Modern NAND Flash controllers typically adopt adaptive bad block management strategies based on the trend of bad block occurrence. For example: 1. Implement stricter isolation strategies for older bad blocks. 2. Increase the isolation effect of newly emerging bad blocks through encryption technology or other means to prevent them from affecting other parts of the system.

[0050] Traditional bad block management schemes have the following main limitations: 1) Resource waste: Once a bad page is detected, the entire block is marked as unusable, resulting in the discarding of a large number of normal pages. 2) Static isolation: Using a fixed-range isolation strategy, it cannot dynamically adapt to the actual impact range of different processes, data loads, or aging stages. 3) Poor adaptability: It cannot effectively cope with the bad page diffusion effect in high-density 3D NAND structures, leading to a decrease in storage reliability. Existing technologies cannot finely isolate local bad pages and lack the ability to predict potential failures. To address the shortcomings of traditional bad block management schemes, this invention aims to significantly improve storage utilization, extend chip lifespan, and reduce storage space waste caused by local bad pages while ensuring data reliability.

[0051] This invention provides an AI-based NAND Flash bad block management method. The AI-based NAND Flash bad page management scheme dynamically predicts the impact range of bad pages using the ST-GCN (Spatiotemporal Graph Convolutional Network) model, accurately delineates risk areas using a hierarchical isolation strategy, and overcomes physical block limitations to reorganize virtual storage blocks. Combined with adaptive error correction and preventative data migration, it can reduce storage space waste caused by localized bad pages by more than 80% while ensuring data reliability, significantly improving the economic efficiency of high-density NAND Flash chips. The steps are illustrated in the diagram below. Figure 1 As shown, the specific steps include:

[0052] S1. Based on preset parameters and preset indicators, perform comprehensive multi-dimensional data collection on bad Pages. A Page represents a page and is the smallest unit for performing read operations. Then, preprocess the collected multi-dimensional data. The preset parameters and preset indicators are not limited to a specific value, but are values ​​preset according to the actual working environment. At the same time, before performing multi-dimensional data collection, it is also necessary to initialize the collection device and sensors to ensure the accuracy and stability of data collection.

[0053] S2. Train the spatiotemporal graph convolutional network model. Load the pre-trained spatiotemporal graph convolutional network model and input the pre-processed multi-dimensional data into the spatiotemporal graph convolutional network model. The spatiotemporal graph convolutional network model extracts and transforms the pre-processed multi-dimensional data through graph convolutional layers and temporal convolutional layers, and outputs the failure probability of each page.

[0054] S3. Based on the failure probability obtained in step S2, determine the directly affected area and the potential risk area. For Pages in the directly affected area, take isolation measures and prohibit read and write operations. For Pages in the potential risk area, allow reading but restrict writing, and start real-time health monitoring. At the same time, reorganize the remaining normal Pages across block space to form virtual blocks to maximize the utilization of storage space. Then, perform dynamic loading of error correction codes and preventive data migration.

[0055] Specifically, step S1 includes the following steps:

[0056] S11. Physical feature acquisition: Acquire the three-dimensional coordinates (layer / row / column) of the bad page within the block and the coupling capacitance measurement value of the adjacent pages to obtain physical feature data; the three-dimensional coordinates are used to accurately locate the physical position of the bad page, while the coupling capacitance measurement value reflects the capacitive coupling between pages, which helps to understand the potential impact of the bad page on other pages.

[0057] S12. Electrical feature acquisition: Obtain the threshold voltage distribution distortion of bad pages and the voltage offset when reading fails, and obtain electrical feature data. Threshold voltage distribution distortion can reflect the changes in the electrical performance of the page, while the voltage offset when reading fails directly indicates the abnormal situation that occurred in the page during the reading operation. This information is crucial for assessing the health status of the page.

[0058] S13. Environmental Feature Acquisition: Record and save temperature and power supply voltage fluctuations during operation to obtain environmental feature data; temperature changes may affect page performance and reliability, while power supply voltage fluctuations may interfere with page read and write operations. These environmental factors are of great significance for analyzing the formation and spread of bad pages.

[0059] S14. Timing correlation analysis: Collect the operation history of the Block before and after the bad page, including the number of erase / write operations, the frequency of read operations, and the programming order of adjacent Pages to obtain timing correlation data. A Block represents a block, which consists of multiple pages and is the smallest addressable unit of an erase operation. By analyzing this timing correlation data, the potential correlation between bad pages and operations can be discovered, providing strong support for predicting the spread trend of bad pages.

[0060] S15. Preprocess the collected physical characteristic data, electrical characteristic data, environmental characteristic data, and time-series correlation data, including cleaning and normalization operations, to remove outliers and noisy data. This ensures the accuracy and consistency of the data and prepares it for subsequent input into the model.

[0061] Specifically, the training of the spatiotemporal graph convolutional network model in step S2 includes the following steps:

[0062] S211. Training data acquisition: Under a controlled experimental environment, pre-defined types of bad pages are created, including programming failure bad pages, read interference bad pages, charge leakage bad pages, repeated erase / write aging bad pages, and temperature-sensitive bad pages. The diffusion path and influence range of the pre-defined types of bad pages are recorded as the first training sample data. In this embodiment, for example, for programming failure bad pages, incorrect programming operations are intentionally performed on a specific page to observe its impact on surrounding pages; for charge leakage bad pages, accelerated aging experiments are conducted to induce charge leakage in the storage unit, and the corresponding data changes are recorded. This experimental data can directly reflect the diffusion characteristics of bad pages under specific conditions, providing high-quality samples for model training.

[0063] S212. By collecting data on the spread of bad pages caused by the natural aging of terminal devices in actual use environments, the first training sample data is supplemented to obtain the second training sample data. These data cover the real aging process of NAND Flash memory under different use scenarios and working conditions, which can supplement the deficiencies of the first training sample data and make the model more universal and adaptable.

[0064] S213. The training loss function of the spatiotemporal graph convolutional network model is the mean squared error loss function. Where N is the number of samples in the second training sample data. Indicates the actual failure status (0 indicates normal, 1 indicates failure). This represents the failure probability predicted by the model, used to measure the difference between the predicted failure probability and the actual failure situation; by minimizing this loss function, the model can accurately predict the impact range of bad pages.

[0065] S214. The spatiotemporal graph convolutional network model (ST-GCN model) includes an input layer, a graph convolutional layer, a temporal convolutional layer, and an output layer. The parameters of the spatiotemporal graph convolutional network model are updated and optimized using the Adam optimization algorithm. The Adam algorithm combines the advantages of momentum and AdaGrad, and can adaptively adjust the learning rate during training to accelerate the convergence speed of the model, while avoiding getting trapped in local optima, which helps the model to better learn the complex patterns of bad page spread.

[0066] Specifically, step S2 also includes the following steps:

[0067] S221. The preprocessed multi-dimensional data is modeled as a 3D mesh graph, with each Page as a node in the 3D mesh graph. Node attributes include the Page's health status and electrical parameters. The health status can be obtained from the error detection mechanism, such as whether an uncorrectable error has occurred. Electrical parameters include threshold voltage distribution, voltage offset when a read fails, etc. These parameters reflect the current physical state and performance of the Page. The edge weights reflect the coupling strength between Pages, mainly calculated based on factors such as the physical distance between Pages and charge coupling effects. Adjacent Pages have a greater charge coupling effect due to their proximity, resulting in larger edge weights; while non-adjacent Pages with some correlation have relatively smaller edge weights. This weight setting reflects the degree of mutual influence between Pages.

[0068] S222. The input layer receives a three-dimensional mesh graph, including the attribute information of all nodes and the edge weight matrix between nodes. The input layer is responsible for integrating and pre-processing the complex NAND Flash storage structure data to prepare for subsequent convolution operations.

[0069] S223. Perform graph convolution operation on the received 3D mesh graph through the graph convolution layer. The node features are aggregated and updated to obtain spatial features, where This represents the feature matrix of the l-th layer node. This represents the adjacency matrix after adding self-joins. This represents the corresponding degree matrix. Let σ represent the weight matrix of the l-th layer, and σ be the activation function. This operation enables each node to aggregate information from its neighboring nodes, thereby learning a richer spatial feature representation. During the graph convolution operation, not only are the features of the node itself considered, but also the features of its neighboring nodes and edge weight information are combined to capture the spatial dependencies between Pages.

[0070] S224. After capturing spatial features, a temporal convolution operation is introduced to model the state changes of the Page at different time steps. Considering that the health status of the Page will change over time during the long-term use of NAND Flash memory, the temporal convolution layer can effectively capture this time-series dependency and further improve the accuracy of the model in predicting the impact range of bad pages.

[0071] S225. According to the prediction target of the model, the output layer outputs the failure probability of each Page within a preset time. In this embodiment, the preset time is 100 programming / erasing cycles. The Page is divided into a directly affected area (failure probability > 90%) and a potential risk area (failure probability between 30% and 90%), providing a basis for subsequent isolation and reorganization strategies.

[0072] Specifically, step S3 includes the following steps:

[0073] S31. For the directly affected area (red area), immediately isolate it and prohibit reading and writing to prevent the fault from spreading further and affecting other pages; for the potentially risky area (yellow area), allow reading but restrict writing, and enable real-time health monitoring. This can reduce the possibility of further deterioration of the potentially risky area without affecting data reading.

[0074] S32. Construct a secondary mapping layer from logical addresses to physical addresses, breaking the limitations of physical blocks. Pages in green areas (healthy or unused areas) from different blocks are combined into new virtual blocks. This helps optimize the layout of storage space and improve the utilization of storage resources. During the reorganization process, pages with similar programming / erase cycles are selected for combination first to prevent the new virtual blocks from failing again due to excessively rapid local aging. In this way, the lifespan of the new virtual blocks can be extended, improving the overall stability of the storage system.

[0075] S33. Each virtual block reserves 5% to 10% of spare pages. When a potentially risky page is upgraded to a directly affected area, the replacement mechanism is automatically triggered to replace it with a spare page, ensuring data reliability and the continuity of the storage system.

[0076] Specifically, the dynamic loading of error correction codes in step S3 includes: using the default BCH code for the green area Page, upgrading the potential risk area Page to the LDPC code, and enabling the progressive adjustment algorithm for the read voltage.

[0077] Specifically, the preventive data migration described in step S3 includes: automatically migrating data to green areas when the predicted monthly failure rate of Page in a potential risk area exceeds 15%.

[0078] This invention also discloses an artificial intelligence-based NAND Flash bad block management system, used in the aforementioned artificial intelligence-based NAND Flash bad block management method, the structural diagram of which is shown below. Figure 2 As shown, it includes:

[0079] The bad page feature acquisition module is used to collect multi-dimensional data and preprocess the collected multi-dimensional data;

[0080] The AI ​​dynamic impact prediction model module is used to train and optimize the spatiotemporal graph convolutional network model. Then, through the trained spatiotemporal graph convolutional network model, features are extracted and transformed from the preprocessed multi-dimensional data, and the failure probability of each page is output.

[0081] The local isolation and virtual block reorganization module determines the hierarchical isolation strategy based on the failure probability output by the model. At the same time, it reorganizes the remaining normal pages across the block space to form virtual blocks.

[0082] An adaptive error correction enhancement module is used for dynamic loading of error correction codes and preventative data migration.

[0083] This invention proposes an AI-based method and system for dynamic isolation and spatial reorganization of locally bad pages, overcoming the limitations of traditional bad block management schemes: 1) Dynamic impact prediction: Using AI models to analyze bad page characteristics (such as error type, physical coordinates, P / E cycle, etc.) in real time, predicting their potential impact range, rather than relying on fixed rules. This method can dynamically adapt to the actual impact range of different processes, data loads, or aging stages, effectively avoiding insufficient or excessive isolation; 2) Cross-block spatial reorganization: Recombining unaffected normal pages within multiple blocks into virtual blocks, breaking through physical block boundary limitations. This method can maximize the utilization of dispersed normal pages, significantly improving storage utilization; 3) Adaptive error correction enhancement: Dynamically allocating error correction code strength and read voltage compensation strategies based on the AI ​​risk assessment results of the page's location. This method can effectively improve data reliability and reduce the risk of data loss due to bad pages; 4) Closed-loop management system: Constructing a full-chain closed-loop management system of "prediction-isolation-reorganization-monitoring" to achieve adaptive optimization of storage resources. It can monitor storage status in real time, dynamically adjust management strategies, and ensure the long-term reliability of the storage system. This invention significantly improves storage utilization, extends chip lifespan, and substantially reduces storage space waste caused by localized bad pages, while ensuring data reliability, and has significant practical application value.

[0084] For example, taking 256-layer 3D NAND Flash as an example, NAND Flash: NAND flash memory is a type of flash memory that uses a non-linear macrocell pattern internally, providing a cost-effective solution for realizing large-capacity solid-state memory. Currently, 3D TLC NAND has the largest shipment volume and the widest usage.

[0085] Bad page detection: The ECC module (error correction code module) detected an uncorrectable error in Page(125,34,78) and marked it as a bad page.

[0086] Feature extraction was performed, collecting the lateral crosstalk coefficient, vertical charge leakage rate, and variance of voltage fluctuations over the last 100 operations for the layer containing the Page.

[0087] AI predicts that the ST-GCN model output directly affects the same-layer Page (124~126, 33~35, 77~79), while the potential risk area is the adjacent Pages on the upper and lower layers.

[0088] Isolation and Reorganization: Red area: 24 Pages are isolated (directly affected area). Yellow area: 48 Pages are marked as restricted. Reorganization: The remaining 952 green Pages of the original Block are merged with the normal Pages of two other Blocks into 3 virtual blocks.

[0089] Dynamic monitoring: Each write operation to the yellow-area Page triggers a health assessment; if the failure rate reaches a threshold, data migration is performed.

[0090] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for managing bad blocks in NAND Flash based on artificial intelligence, characterized in that, Includes the following steps: S1. Based on preset parameters and preset indicators, multi-dimensional data is collected from bad Pages. A Page represents a page and is the smallest unit for performing read operations. Then, the collected multi-dimensional data is preprocessed. S2. Train the spatiotemporal graph convolutional network model. Load the pre-trained spatiotemporal graph convolutional network model and input the pre-processed multi-dimensional data into the spatiotemporal graph convolutional network model. The spatiotemporal graph convolutional network model extracts and transforms the pre-processed multi-dimensional data through graph convolutional layers and temporal convolutional layers, and outputs the failure probability of each page. S3. Based on the failure probability obtained in step S2, determine the directly affected area and the potential risk area. For Pages in the directly affected area, take isolation measures and prohibit read and write operations. For Pages in the potential risk area, allow reading but restrict writing, and start real-time health monitoring. At the same time, reorganize the remaining normal Pages across block space to form virtual blocks to maximize the utilization of storage space. Then, perform dynamic loading of error correction codes and preventive data migration. Step S1 specifically includes the following sub-steps: S11. Physical feature acquisition: Obtain and store the three-dimensional coordinates of the bad page and the coupling capacitance measurement value of the adjacent page to obtain physical feature data. S12. Electrical feature acquisition: Measure and record the threshold voltage distribution distortion of bad pages and the voltage offset when reading fails, and obtain electrical feature data. S13. Environmental Feature Acquisition: Record and save temperature and power supply voltage fluctuations during operation to obtain environmental feature data; S14. Timing correlation analysis: Collect the operation history of the Block before and after the bad page, including the number of erase / write operations, the frequency of read operations, and the programming order of adjacent pages to obtain timing correlation data; Block represents a block, which consists of multiple pages and is the smallest addressable unit for erase operations.

2. The NAND Flash bad block management method based on artificial intelligence according to claim 1, characterized in that, Step S1 also includes the following steps: S15. Preprocess the collected physical characteristic data, electrical characteristic data, environmental characteristic data, and time-series correlation data, including cleaning and normalization, to remove outliers and noise data.

3. The NAND Flash bad block management method based on artificial intelligence according to claim 2, characterized in that, The training of the spatiotemporal graph convolutional network model described in step S2 specifically includes the following steps: S211. Training data acquisition: In a controllable experimental environment, pre-defined types of bad pages are generated, including programming failure bad pages, read interference bad pages, charge leakage bad pages, repeated erase and write aging bad pages, and temperature-sensitive bad pages. The diffusion path and influence range of the pre-defined types of bad pages are recorded as the first training sample data. S212. By collecting data on the spread of bad pages caused by the natural aging of terminal devices in actual use environments, the first training sample data is supplemented to obtain the second training sample data. S213. The training loss function of the spatiotemporal graph convolutional network model is the mean squared error loss function. Where N is the number of samples in the second training sample data. Indicates the actual failure situation. This represents the failure probability predicted by the model, used to measure the difference between the predicted failure probability and the actual failure situation. S214. The spatiotemporal graph convolutional network model includes an input layer, a graph convolutional layer, a temporal convolutional layer, and an output layer. The parameters of the spatiotemporal graph convolutional network model are updated and optimized using the Adam optimization algorithm.

4. The NAND Flash bad block management method based on artificial intelligence according to claim 3, characterized in that, Step S2 also includes the following steps: S221. Model the preprocessed multi-dimensional data into a three-dimensional mesh diagram, with each Page as a node in the three-dimensional mesh diagram. The node attributes include the Page's health status and electrical parameters. The edge weights reflect the coupling strength between Pages. S222. Receive a three-dimensional mesh graph through the input layer, including the attribute information of all nodes and the edge weight matrix between nodes; S223. Perform graph convolution operation on the received 3D mesh graph through the graph convolution layer. The node features are aggregated and updated to obtain spatial features, where Show the feature matrix of the l-th layer node. This represents the adjacency matrix after adding self-joins. This represents the corresponding degree matrix. Let σ represent the weight matrix of the l-th layer, and σ be the activation function. S224. The temporal convolutional layer is used to model the state changes of the Page at different time steps, further improving the accuracy of the model's prediction of the impact range of bad pages; S225. Based on the model's prediction target, the output layer outputs the failure probability of each Page within a preset time, dividing the Page into directly affected areas and potential risk areas.

5. A NAND Flash bad block management method based on artificial intelligence according to claim 4, characterized in that, Step S3 includes the following steps: S31. For areas directly affected, immediately isolate them and prohibit reading and writing; for potentially risky areas, allow reading but restrict writing, and enable real-time health monitoring. S32. Construct a secondary mapping layer from logical address to physical address, and combine green area Pages from different Blocks into new virtual blocks. During the reorganization process, first select Pages with similar programming / erase cycles for combination. S33. Each virtual block retains 5% to 10% of spare pages. When a potentially risky page is upgraded to a directly affected area, the replacement mechanism is automatically triggered, and the spare page is used to replace it.

6. A NAND Flash bad block management method based on artificial intelligence according to claim 5, characterized in that, The dynamic loading of error correction codes in step S3 includes: using the default BCH code for the green area Page, upgrading the potentially risky area Page to the LDPC code, and enabling the progressive adjustment algorithm for the read voltage.

7. A NAND Flash bad block management method based on artificial intelligence according to claim 5, characterized in that, The preventative data migration described in step S3 includes automatically migrating data to green areas when the predicted monthly failure rate of a potential risk area exceeds 15%.

8. An AI-based NAND Flash bad block management system, used in any one of claims 1-7, characterized in that, include: The bad page feature acquisition module is used to collect multi-dimensional data and preprocess the collected multi-dimensional data; The AI ​​dynamic impact prediction model module is used to train and optimize the spatiotemporal graph convolutional network model. Then, through the trained spatiotemporal graph convolutional network model, features are extracted and transformed from the preprocessed multi-dimensional data, and the failure probability of each page is output. The local isolation and virtual block reorganization module determines the hierarchical isolation strategy based on the failure probability output by the model. At the same time, it reorganizes the remaining normal pages across the block space to form virtual blocks. The adaptive error correction enhancement module is used for dynamic loading of error correction codes and preventative data migration.

Images