A NAND flash memory reliability management method and system based on a bad block space diffusion dynamics model

Abstract

This invention discloses a NAND flash memory reliability management method and system based on a bad block spatial diffusion dynamics model, belonging to the field of semiconductor memory reliability management technology. The method uses confirmed bad blocks as diffusion sources and establishes a bad block spatial diffusion probability model using weighted physical distance between blocks, diffusion attenuation length varying with junction temperature according to an Arrhenius relation, a temperature change rate driving term, and a wear susceptibility function. A real-time risk score is calculated for each block, and based on this, safe zones, warning zones, and danger zones are divided, performing routine detection, encrypted detection and write restrictions, preventative migration, and pre-isolation respectively. Combined with temperature shock protection and online calibration of model parameters, bad block management shifts from reactive replacement to proactive handling, improving data reliability under high temperature and temperature shock conditions and reducing detection bandwidth overhead.

Description

Technical Field

[0001] This invention relates to the field of semiconductor memory reliability engineering technology, and in particular to a NAND flash memory reliability management method and system based on a bad block space diffusion dynamics model. Background Technology

[0002] NAND flash memory is widely used in automotive electronics, industrial control, edge computing, and high-reliability storage devices. These scenarios often involve high temperatures, rapid temperature rises and falls, thermal cycling, and long-term high-intensity erase and write operations, which accelerate the degradation of the tunneling oxide layer, charge leakage, and threshold voltage drift, resulting in increased original bit error rate and more new bad blocks. Automotive applications also require devices to maintain data reliability in a wide temperature range of -40°C to 105°C or even longer, placing far greater pressure on bad block management than in consumer scenarios.

[0003] Current bad block management generally treats each block as an independent failure unit: after a bad block is detected during runtime, it is replaced with a spare block, supplemented by wear leveling to distribute the write operation. This approach can only address the problem retrospectively. Building on this, some improved solutions have begun to predict failures based on the bit error rate trend of individual blocks and migrate data in advance. However, the prediction target remains an isolated single block, failing to utilize information about the physical location of the bad block. Device failure analysis has long shown that due to the localization of manufacturing defects and the propagation of thermal stress along the wafer plane, bad blocks often exhibit spatial clustering in the block array—after a block fails, the probability of blocks in its physical neighborhood failing in the subsequent period is significantly higher than that of distant blocks, and this clustering is asymmetrical in the wordline and bitline directions. Furthermore, existing solutions for temperature handling are mostly at the global level, such as reducing the frequency due to overheating or increasing the overall refresh rate, failing to pinpoint the impact of temperature shocks to the physical neighborhood most likely to experience cascading degradation. In other words, the spatial correlation and temperature dynamics of bad blocks are essentially idle in existing bad block management. This invention aims to incorporate these two types of information into a unified quantitative model for utilization. Summary of the Invention

[0004] To address the shortcomings of existing bad block management methods, which treat each block as an independent failure unit, only allow for post-incident replacement, fail to characterize the aggregation and diffusion of bad blocks in physical space, and cannot proactively adjust strategies in areas of temperature shock and high wear, this invention provides a NAND flash memory reliability management method and system based on a bad block spatial diffusion dynamics model. The method includes: acquiring the physical row and column coordinates of the NAND flash memory block array, as well as chip junction temperature, temperature change rate, temperature change acceleration, P / E cycle count of each block, original bit error rate, bad block status, and page failure information; using confirmed bad blocks as diffusion sources, constructing a bad block spatial diffusion probability model based on weighted physical distance, temperature-related diffusion attenuation length, temperature change rate driving term, and wear susceptibility function; calculating a real-time bad block risk score for each block and dividing each block into a safe zone, a warning zone, and a danger zone; performing differentiated detection, data placement restrictions, and preventative migration for different zones, and combining temperature shock protection and online calibration of model parameters to transform bad block management from post-incident replacement to proactive handling.

[0005] The parameters obtained include the physical row and column coordinates of the NAND flash memory block array, the chip junction temperature T, the rate of temperature change dT / dt, and the acceleration due to temperature change d. 2 T / dt 2 The number of P / E cycles, raw bit error rate (RBER), bad block status, and page failure information for each block.

[0006] The bad block spatial diffusion probability model is P_spread(j|i)=min{1,α·exp(-d(i,j) / λ_T)·(1+β·|dT / dt|)·f(N_PE(j))}; this model simultaneously introduces physical distance attenuation, temperature-related diffusion attenuation length, temperature change rate driving term, and wear susceptibility function to represent the influence of a confirmed bad block on the degradation probability of neighboring blocks; where the diffusion attenuation length λ_T and the physical distance d(i,j) use the same length dimension measured in block spacing, and since exp(-d(i,j) / λ_T), α and f(N_PE(j)) are all not greater than 1, only the temperature change rate driving term may make the product greater than 1, so by truncating to min{1,·}, P_spread(j|i) is always a valid probability value in the interval [0,1].

[0007] The real-time bad block risk score S_risk(j) is calculated based on the spatial distribution of the surrounding confirmed bad blocks, the severity of each bad block, and the RBER trend of the block itself. Each block is divided into a safe zone, a warning zone, and a danger zone. The safe zone performs routine detection, the warning zone increases the detection frequency and restricts the writing of important data, and the danger zone triggers the preventive migration of valid data and performs pre-isolation.

[0008] Furthermore, the physical distance d(i,j) is a weighted Manhattan distance, and w_WL and w_BL are determined by statistically analyzing the occurrence frequency of actual bad block pairs in the Wordline and Bitline directions, so that the model can express the anisotropy of bad block propagation within the NAND array.

[0009] Furthermore, the model parameters α, β, γ, and λ_0 are calibrated online based on the new bad block event, the confirmation time, the junction temperature at the time of confirmation, and the distribution of bad blocks in the neighborhood, so that the diffusion probability model is updated with the actual degradation state of the device.

[0010] Furthermore, the system also determines the temperature shock level based on the temperature change rate and temperature change acceleration, and performs emergency scanning, write pause, and priority migration of dangerous areas under temperature shock conditions.

[0011] Compared to existing technologies, the key difference of this invention lies in the fact that the target of bad block management changes from isolated individual blocks to a group of blocks with spatial structure and temperature dynamics. The spatial diffusion probability model converts the location, junction temperature, temperature change rate, and neighborhood wear state of confirmed bad blocks into quantitative risks for each block. The three-zone division then concentrates limited detection bandwidth and migration resources on the highest-risk physical neighborhood—the denser detection locations are precisely the most likely to have problems, rather than uniformly increasing the full-disk scan frequency. Preventative migration, combined with FTL mapping updates, ensures that valid data leaves high-risk areas before a block truly fails. Temperature shock protection combines dT / dt and d... 2 T / dt 2 Direct access to risk scoring allows the system to automatically tighten its strategy during the most intense periods of thermal cycling; online calibration ensures that model parameters evolve along the actual aging trajectory of individual devices, preventing a disconnect between factory parameters and actual operating conditions. All these components share the output of the same diffusion model, supporting each other rather than simply existing side-by-side. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the overall architecture of the NAND flash memory reliability management system based on the bad block space diffusion dynamics model proposed in this invention.

[0013] Figure 2 This is a schematic diagram illustrating the construction and calculation process of the bad block spatial diffusion probability model proposed in this invention;

[0014] Figure 3 This is a heat map of spatial risk and a schematic diagram of the three-zone division of the NAND flash memory array proposed in this invention;

[0015] Figure 4 This is a schematic diagram of the adaptive detection scheduling and preventive data migration process based on risk scoring proposed in this invention;

[0016] Figure 5 This is a schematic diagram of the temperature change rate shock protection and online model calibration process proposed in this invention;

[0017] Figure 6 This is a schematic diagram of the anisotropic characteristics and weighted distance model of bad block spatial diffusion proposed in this invention.

[0018] In the diagram: 1. NAND flash memory array; 2. Temperature sensing module; 3. Diffusion dynamics modeling engine; 4. Adaptive detection scheduler; 5. Preventive data migration engine; 6. Model calibration module; 7. Temperature shock protection module; 8. NAND status monitoring module. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments are used to help understand the technical solution and are not intended to limit the scope of protection; firmware implementation details not described in detail in the embodiments are performed according to conventional storage controller technology.

[0020] Reference Figures 1-6 A method and system for NAND flash memory reliability management based on a bad block space diffusion dynamics model is disclosed. The system includes a NAND flash memory array, a temperature sensing module, a diffusion dynamics modeling engine, an adaptive detection scheduler, a preventative data migration engine, a model calibration module, a temperature shock protection module, and a NAND status monitoring module. In this invention, the NAND status monitoring module collects block physical coordinates, P / E cycle counts, raw bit error rate (RBER), bad block status, and page failure information. The temperature sensing module collects the chip junction temperature T and calculates dT / dt and d... 2 T / dt 2 The diffusion dynamics modeling engine calculates P_spread(j|i) and S_risk(j) based on the above parameters, and then the adaptive detection scheduler and the preventive data migration engine execute the partitioning disposal strategy.

[0021] The criteria for confirming a block as a bad block are any of the following: The read operation fails despite ECC error correction, or the number of bits required for error correction exceeds the upper limit of ECC error correction capability; the raw bit error rate (RBER) exceeds the preset bad block judgment threshold (RBER_bad) for M consecutive tests; the read retries exceed the preset upper limit and still cannot be read correctly; or the erase or programming operation returns a failure status. A block that meets any of these conditions is marked as a confirmed bad block and added to the bad block set Ω_bad. Simultaneously, its confirmation timestamp, physical coordinates, junction temperature T at the time of confirmation, and temperature change rate dT / dt are recorded.

[0022] In one specific embodiment, the initial parameters of the model are α=0.01, β=0.1s / K, γ=1.5, λ_0=10 block spacing, reference temperature T_0=298.15K, equivalent activation energy E_a=0.1eV, f_0=0.1, w_WL=0.68, w_BL=0.32. When the junction temperature T = 358.15 K (85 °C), λ_T is calculated to be approximately 19 block spacings using λ_T = λ_0·exp[-(E_a / k_B)·(1 / T-1 / T_0)]. For adjacent blocks (d(i,j) = 1), |dT / dt| = 2 K / s, and N_PE(j) / N_PE_max = 0.8, the wear susceptibility f = 0.1 + 0.9 × 0.8^1.5 ≈ 0.744, so P_spread = 0.01 × exp(-1 / 19) × (1 + 0.1 × 2) × 0.744 ≈ 0.0085, and truncation is not triggered. When the diffusion source is adjacent to multiple bad blocks or when temperature shocks cause a significant increase in the temperature change rate driving term, min{1,·} truncation takes effect to ensure the probability is valid. The initial values ​​of parameters such as E_a and λ_0 can be obtained by fitting the device through accelerated aging tests, and calibrated online during operation as described in claim 9.

[0023] RBER_trend(j) is calculated as follows: The raw bit error rate (BER) of Block j is sampled every preset number of erase / write cycles or time intervals. The most recent N_sample samples are used to form a sequence. A least-squares linear regression is performed with the sample number as the independent variable and RBER as the dependent variable. The slope of the regression line is taken as the raw trend value. This slope is then normalized by dividing it by a preset upper limit and truncated to the [0,1] interval to obtain RBER_trend(j). In a specific embodiment, N_sample is 8, and the sampling interval is once every 1000 P / E cycles.

[0024] The severity weight w_severity(i) of bad blocks is determined according to the failure level: only those with an original bit error rate exceeding the threshold are given a smaller weight, those with read retry failures are given a medium weight, and those with erase failures or programming failures are given a larger weight. Each weight is normalized to the (0,1] interval, so that bad blocks with more severe failures contribute more to the risk of neighboring blocks.

[0025] The risk thresholds T_warn and T_danger are determined as follows: During the early operation phase of the device, the risk score S_risk distribution of all normal blocks is statistically analyzed, and a high quantile (e.g., the 95th quantile) is taken as T_warn. A higher quantile (e.g., the 99th quantile) or the average S_risk of historically confirmed bad blocks within the observation window prior to confirmation is taken as T_danger. Alternatively, the threshold to ensure the number of dangerous blocks does not exceed the migration capacity can be determined by considering the number of available spare blocks and the migration bandwidth budget. T_warn and T_danger are periodically adjusted based on online calibration results and spare block consumption, and T_warn is always satisfied. <T_danger。

[0026] In one specific embodiment, the NAND flash memory is an automotive storage chip with a block array size of 1024 rows × 512 columns. After system startup, the physical topology parameters are read and the initial model parameters α=0.01, β=0.1s / K, γ=1.5, λ_0=10, w_WL=0.68, and w_BL=0.32 are loaded. Among them, w_WL and w_BL are determined by the historical failure statistics of the same type of device: the previous failure analysis counted a total of 1240 pairs of bad blocks, of which about 32% were adjacent along the Wordline direction and about 68% were adjacent along the Bitline direction. After normalizing the reciprocals of the proportions in the two directions, w_WL=0.68 and w_BL=0.32 are obtained. The more obvious the bad block clustering, the shorter the calculated distance, and the slower the decay of the diffusion term along the direction. During operation, when a block is identified as a bad block, the system calculates the diffusion probability of neighboring blocks based on the bad block as the diffusion source, and obtains a real-time bad block risk score by combining the RBER trend of each block. If the block falls into the warning zone, the detection frequency is increased and the writing of critical data is restricted. If the block falls into the danger zone, its valid data is migrated to a block in the safe zone with a lower risk score and fewer P / E cycles.

[0027] The specific control process for preventative migration and pre-isolation is as follows: First, when Block j is judged to be in the danger zone, the preventative data migration engine selects the Block with the lowest S_risk value and the fewest P / E cycle counts in the safe zone as the target Block; Second, the data of the valid pages in Block j is read page by page, verified by ECC, and then written to the target Block; Third, after the migration is completed, the logical address to physical address mapping table of the flash translation layer (FTL) is updated so that the relevant logical pages point to the target Block; Fourth, Block j is marked as pre-isolated, so that it no longer accepts new data writes, but it is still retained in the diffusion source set Ω_bad to participate in the neighborhood risk scoring calculation; Fifth, the S_risk of each Block in the neighborhood of Block j is recalculated and the partitioning result is updated. If the target Block write fails during the migration process, the target Block is added to the bad block set and a new migration target is selected.

[0028] In another specific embodiment, when the chip junction temperature rapidly rises from room temperature to the high-temperature range, the temperature sensing module detects that dT / dt exceeds the temperature surge threshold, and the system immediately performs rapid BER sampling on the warning and danger zones; when d 2 T / dt 2 If the threshold is exceeded further, the system suspends non-critical write operations and prioritizes migrating valid data in the danger zone blocks. This process does not simply increase the frequency of full-disk scans, but rather incorporates the risk increment caused by temperature shocks into the spatial diffusion model and risk score, thereby prioritizing the protection of physical neighborhoods more likely to experience cascading degradation.

[0029] In model calibration, whenever a new bad block is identified, the system selects candidate blocks within a neighborhood radius R centered on that bad block. Those that degenerate into bad blocks within the observation window W are marked as positive samples (y=1), and those that do not degenerate are marked as negative samples (y=0). These candidate blocks, along with their identification time, physical coordinates, junction temperature at identification time, temperature change rate, and the distribution of existing bad blocks in the neighborhood, are written into the calibration dataset. When the number of samples reaches a preset threshold, a log-likelihood function L(α,β,γ,λ_0)=Σ_k[y_k·lnP_spread,k+(1-y_k)·l The model uses a gradient ascent method [n(1-P_spread,k)] to update parameters with a learning rate η and a maximum of I_max iterations. The iteration stops when the log-likelihood increment between two consecutive iterations is less than the convergence threshold ε or the number of iterations reaches I_max. Constraints α∈(0,1], λ_0>0, and γ≥0 are applied to the parameters during iteration. The updated parameters are used for subsequent risk scoring calculations, allowing the model to adjust according to the actual aging state of the NAND device. This implementation can be carried out in firmware, storage controller, or flash memory conversion layer, or it can be executed by a dedicated reliability management module.

[0030] To verify the effectiveness of the method, an accelerated aging control test was conducted using eight automotive-grade 3D TLC NAND flash chips from the same batch: four chips used traditional bad block management (runtime detection, spare block replacement, and wear leveling), while the other four chips had the partition management of this invention superimposed on the same firmware. The test conditions alternated between an 85°C constant temperature range and a -40°C to 105°C temperature cycling range, with continuous random write load applied up to 120% of the nominal P / E lifetime. At the end of the test, the control group had a total of 7 uncorrectable error events, while the test group had 1. In the test group, 81% of the bad blocks that were finally identified had been classified into the danger zone and completed effective data migration before identification, with an average early warning window of approximately 1100 P / E cycles. Since BER sampling was concentrated in the warning and danger zones, the average detection bandwidth overhead of the entire disk decreased by approximately 37% compared to the equal-frequency scanning scheme, and the additional write amplification introduced by preventive migration was approximately 0.04. It is worth noting that the differences between the two groups mainly occur in the temperature cycling segment, while the differences are not significant in the isothermal segment. This is consistent with the design expectation of the temperature change rate driving term in the model, and also indicates that the benefits of this method are concentrated in the working conditions with drastic temperature dynamics.

[0031] The above embodiments are merely preferred embodiments of the present invention. Those skilled in the art can make adjustments or equivalent substitutions to the parameter values, threshold settings, distance weights, calibration cycles, migration target selection methods, or system module deployment locations without departing from the concept of the present invention; all such adjustments or substitutions should fall within the protection scope of the present invention.

Claims

1. A NAND flash memory reliability management method based on a bad block space diffusion dynamics model, characterized in that, The method includes: acquiring physical topology parameters and environmental parameters of the NAND flash memory, wherein the physical topology parameters include at least the physical row and column coordinates of each block in the block array, and the environmental parameters include at least the chip junction temperature T and the temperature change rate dT / dt; constructing a bad block spatial diffusion probability model based on the physical topology parameters and environmental parameters, wherein the bad block spatial diffusion probability model is P_spread(j|i)=min{1,α·exp(-d(i,j) / λ_T)·(1+β·|dT / dt|)·f(N_PE(j))}; where P_spread(j|i) is the conditional probability that after Block i is identified as a bad block, its adjacent Block j will degenerate into a bad block within a preset observation window, and d(i,j) is the probability of Block i and Block j being bad blocks. The physical distance between j, λ_T is the temperature-related diffusion decay length and uses the same length dimension measured by the block spacing as d(i,j), α is the basic diffusion coefficient with a value range of (0,1], β is the temperature change rate sensitivity coefficient, f(N_PE(j)) is the wear susceptibility function based on the number of P / E cycles of block j with a value range of (0,1], and min{·} indicates that the truncation operation is performed to limit the conditional probability within the interval [0,1]; a real-time bad block risk score S_risk(j) is calculated for each block in the NAND flash memory according to the bad block space diffusion probability model; based on the risk score, all blocks in the NAND flash memory are divided into safe areas, warning areas and dangerous areas, and differentiated detection frequencies, data placement strategies and preventive migration strategies are adopted for blocks in different areas.

2. The method according to claim 1, characterized in that, The physical distance d(i,j) is a weighted Manhattan distance, defined as d(i,j) = w_WL·|r_i-r_j| + w_BL·|c_i-c_j|; where (r_i,c_i) and (r_j,c_j) are the row and column coordinates of Block i and Block j in the NAND array, respectively, w_WL is the weighting coefficient in the Wordline direction, and w_BL is the weighting coefficient in the Bitline direction; w_WL and w_BL are determined by statistically analyzing and normalizing the physical locations of bad block pairs observed during the operation of NAND flash memory, and the sum of w_WL and w_BL is 1.

3. The method according to claim 1, characterized in that, The temperature-dependent diffusion decay length λ_T is determined by the formula λ_T=λ_0·exp[-(E_a / k_B)·(1 / T-1 / T_0)]; where λ_0 is the reference diffusion decay length at the reference absolute temperature T_0 in units of block spacing, E_a is the equivalent activation energy characterizing the temperature sensitivity of NAND oxide degradation, k_B is the Boltzmann constant, T is the absolute temperature value corresponding to the chip junction temperature, and T_0 is the preset reference absolute temperature; the diffusion decay length λ_T increases monotonically with the junction temperature T, and takes the value λ_0 when T equals T_0.

4. The method according to claim 1, characterized in that, The wear susceptibility function f(N_PE(j)) is defined as f(N_PE(j))=f_0+(1-f_0)·(N_PE(j) / N_PE_max)^γ; wherein f_0 is a base susceptibility coefficient, N_PE(j) is the current number of P / E cycles of Block j, N_PE_max is the nominal maximum P / E cycle life of the NAND device, and γ is a wear acceleration index.

5. The method according to claim 1, characterized in that, The real-time bad block risk score S_risk(j) is calculated by the formula S_risk(j)=Σ_{i∈Ω_bad}P_spread(j|i)·w_severity(i)+μ·RBER_trend(j); wherein Ω_bad is a set of all currently confirmed bad blocks, w_severity(i) is the severity weight of bad block i, the value of which increases with the failure level of bad block i and is normalized to the interval (0,1], μ is an RBER trend weight coefficient, RBER_trend(j) is a value obtained after normalizing the least square linear regression slope of the raw bit error rate sequence of Block j in the last N_sample detection cycles through a preset slope range and truncating the normalized slope to the interval [0,1], N_sample is a positive integer not less than 3, and the detection cycle is a preset erase-write count interval or a preset time interval.

6. The method according to claim 1, characterized in that, The dividing all Blocks in the NAND flash into a safety zone, a warning zone and a danger zone comprises: setting risk thresholds T_warn and T_danger; classifying Blocks satisfying S_risk(j)<T_warn into the safety zone and adopting a conventional detection frequency; classifying Blocks satisfying T_warn≤S_risk(j)<T_danger into the warning zone, increasing the detection frequency and limiting the data placement type of the Blocks; classifying Blocks satisfying S_risk(j)≥T_danger into the danger zone and triggering preventive migration of valid data.

7. The method according to claim 6, characterized in that, The preventive migration strategy comprises: selecting a Block with a low S_risk value and a small number of P / E cycles from safety zone Blocks as a migration target; updating the risk scores of related Blocks in the bad block space diffusion probability model after migration is completed; maintaining a diffusion history log, wherein the diffusion history log records timestamps, physical coordinates, chip junction temperatures and temperature change rates of bad block confirmation events.

8. The method according to claim 1, characterized in that, It also includes a temperature change rate shock protection step: real-time calculation of the time change rate dT / dt of the chip junction temperature and the acceleration of temperature change d 2 T / dt 2 A global emergency scan is initiated when dT / dt exceeds the preset temperature shock threshold; when d 2 T / dt 2 When the preset temperature acceleration threshold is exceeded, non-critical write operations are suspended, and data migration of dangerous blocks is prioritized.

9. The method according to claim 1, characterized in that, The model also includes an online calibration step for model parameters: whenever a new bad block is confirmed, candidate blocks are selected within a preset neighborhood radius centered on the bad block. Candidate blocks that degenerate into bad blocks within a preset observation window are recorded as positive samples with a binary label y of 1, and those that do not degenerate are recorded as negative samples with a binary label y of 0. The physical coordinates, the junction temperature at the time of confirmation, and the distribution of existing bad blocks in the neighborhood of each candidate block are added to the calibration dataset. When the number of samples in the calibration dataset reaches a preset calibration threshold, a log-likelihood function L(θ) = Σ[y·lnP_spread(j|i) + (1-y)·ln(1-P_spread(j|i))] is constructed based on the Bernoulli likelihood of each sample, where the parameter vector θ = (α,β,γ,λ_0). The gradient ascent method is used to iteratively solve for the parameter θ that maximizes L(θ). The iteration stops when the log-likelihood increment between two adjacent iterations is less than a preset convergence threshold or the number of iterations reaches a preset upper limit, and the obtained parameter θ is used to update the bad block spatial diffusion probability model.

10. A NAND flash memory reliability management system based on a bad block space diffusion dynamics model, characterized in that, The system includes a NAND flash array, a temperature sensing module, a diffusion dynamics modeling engine, an adaptive detection scheduler, a preventive data migration engine, a model calibration module, a temperature shock protection module, and a NAND status monitoring module; the diffusion dynamics modeling engine is used to execute the method described in any one of claims 1 to 9, calculate the real-time bad block risk score for each block, and output a partition management strategy.

Images