SPI NAND flash memory life prediction method, device, equipment and storage medium

Abstract

The present invention relates to the field of flash memory technology, and discloses a life prediction method, device, equipment and storage medium for SPI NAND flash memory. The method includes: obtaining a storage parameter data set of multi-dimensional performance parameters, and preprocessing the storage parameter data set; extracting multi-level performance characteristics of the cleaned data set, and optimizing the characteristics of the performance characteristic set by adaptive selection; constructing an RBER prediction model for multi-period life prediction based on the performance optimization feature subset, and obtaining multiple programming / erase cycle data of the flash memory to be predicted; predicting the programming / erase cycle data using the prediction model of each period; quantifying the uncertainty of the initial life prediction result, obtaining the confidence interval and interval prediction result, and integrating the initial life prediction result, the confidence interval and the interval prediction result to obtain the final life prediction report. The present application realizes the consideration of multi-period influencing factors when predicting the life of the target flash memory, thereby improving the accuracy of the prediction result.

Description

Technical Field

[0001] The present invention relates to the field of flash memory technology, and in particular to a life prediction method, device, equipment and storage medium for a SPI NAND flash memory. Background Art

[0002] SPI NAND flash memory is widely used in embedded systems and consumer electronics due to its high storage density, low cost and simple interface design. However, like all NAND flash memory technologies, SPI NAND flash memory also faces the problem of limited program / erase (P / E) cycle life. Since each P / E cycle causes a small loss of flash memory cells, this loss will accumulate over time and eventually cause the storage cells to fail. In this regard, accurately predicting the life of SPI NAND flash memory is of great significance for maintaining data security, planning equipment replacement cycles and reducing operation and maintenance costs.

[0003] At present, the prediction of SPI NAND flash memory lifespan attempts to establish a physical model of flash memory aging, or use basic machine learning algorithms for lifespan prediction, etc. However, due to the complexity of the flash memory aging mechanism and multiple influencing factors, these models are often oversimplified and difficult to accurately reflect the actual situation. In addition, machine learning algorithms are generally unable to effectively handle the high dimensionality and nonlinear characteristics of flash memory data. In other words, the existing prediction of SPI NAND flash memory lifespan is difficult to consider multiple influencing factors in multiple periods, resulting in low accuracy of the prediction results. Summary of the invention

[0004] The main purpose of the present invention is to solve the problem that it is difficult to consider multi-period influencing factors when predicting the life of SPI NAND flash memory, resulting in low accuracy of the prediction result.

[0005] A first aspect of the present invention provides a life prediction method for an SPI NAND flash memory, the life prediction method for the SPI NAND flash memory comprising: obtaining a storage parameter data set of multi-dimensional performance parameters of a flash memory type corresponding to the SPI NAND flash memory, and performing multi-stage preprocessing on the storage parameter data set to obtain a cleaned storage parameter data set; performing multi-level performance feature extraction on the cleaned storage parameter data set to obtain a performance feature set, and performing feature optimization of adaptive selection on the performance feature set to obtain a performance optimization feature subset; constructing an RBER prediction model for multi-period life prediction based on the performance optimization feature subset, and obtaining multiple program / erase cycle data of a SPI NAND flash memory to be predicted within a preset prediction time period; using the RBER prediction model for life prediction in each period to predict the remaining number of program / erase cycles and expected failure time of the program / erase cycle data to obtain an initial life prediction result; quantifying the uncertainty of the initial life prediction result to obtain a confidence interval for life prediction and a corresponding interval prediction result, and integrating the initial life prediction result, the confidence interval and the interval prediction result to obtain a final life prediction report for the SPI NAND flash memory.

[0006] Optionally, in a first implementation of the first aspect of the present invention, the storage parameter data set is subjected to multi-stage preprocessing to obtain a cleaned storage parameter data set, including: performing standard deviation analysis on the RBER and voltage values ​​in the storage parameter data set, and based on the results of the standard deviation analysis, identifying and marking the storage parameter data that deviates from a preset numerical range to obtain an abnormal marked data set, and performing curve fitting and interpolation on the short-term missing RBER data in the abnormal marked data set to obtain a complete RBER time series; based on the complete RBER time series, calculating the RBER change rate of the abnormal marked data set, and based on the RBER change rate, The long-term missing voltage data in the constant-label data set are forward-filled and sequence-smoothed to obtain a complete voltage time series; the complete RBER time series and the complete voltage time series are time-axis aligned and deformed to obtain a synchronized multidimensional time series, and the RBER data and voltage data corresponding to the multidimensional time series are multi-scale decomposed and reconstructed to obtain a smooth data curve; based on the extreme values ​​of the RBER curve and the voltage data curve in the smooth data curve, the smooth data curve is interval-mapped to obtain a standardized parameter value, and the number of programming / erase cycles in the standardized parameter value is nonlinearly transformed to obtain a cleaned storage parameter data set.

[0007] Optionally, in a second implementation of the first aspect of the present invention, the multi-level performance feature extraction of the cleaned storage parameter data set to obtain a performance feature set includes: performing local statistical analysis on RBER data in the cleaned storage parameter data set to obtain RBER time domain statistical features, and performing statistical analysis on program / erase operation time interval data in the cleaned storage parameter data set to obtain program / erase operation mode features, and performing feature association on the RBER time domain statistical features and the program / erase operation mode features to obtain comprehensive operation features; based on the comprehensive operation features and the storage parameter data set, identifying and marking storage cells corresponding to the first error rate and the second error rate to obtain error cell marking data, and extracting the error cell marking data based on the error cell marking data. The error spatial distribution characteristics in the column; based on the error spatial distribution characteristics, the RBER time domain statistical characteristics are converted to the frequency domain and the characteristics are jointly analyzed to obtain the RBER frequency domain characteristics, and the RBER frequency domain characteristics and the voltage data in the cleaned storage parameter data set are jointly analyzed in time and frequency to obtain the time-frequency joint characteristics; using the time-frequency joint characteristics and the threshold voltage distribution data in the cleaned storage parameter data set, the morphological parameters of the threshold voltage distribution are calculated to obtain the threshold voltage distribution characteristics, and based on the threshold voltage distribution characteristics, the read-write voltage and erase voltage data in the cleaned storage parameter data set are differentially analyzed in time series to obtain the voltage drift characteristics; based on the comprehensive operation characteristics, the error spatial distribution characteristics, the time-frequency joint characteristics and the voltage drift characteristics, a performance feature set is generated.

[0008] Optionally, in a third implementation of the first aspect of the present invention, the feature optimization of adaptively selecting the performance feature set to obtain a performance optimized feature subset includes: calculating the correlation coefficient between each feature in the performance feature set and the RBER data to obtain an initial feature importance ranking, and based on the initial feature importance ranking, selecting features with correlation coefficients greater than a preset correlation threshold to obtain an initial feature subset; calculating the nonlinear correlation between each feature in the initial feature subset and the number of programming / erase cycles to obtain an updated feature importance ranking, and based on the updated feature importance ranking, performing secondary feature selection on the initial feature subset to obtain a related feature set; applying regularization constraints to the related feature set to obtain a sparse feature subset, and arranging and smoothing the features in the sparse feature subset in order of the number of programming / erase cycles to construct a smoothed feature importance score; The feature importance scores are predicted in time series to obtain predicted feature importance values ​​for programming / erase cycles in future time periods, and based on the predicted feature importance values, the sparse feature subsets are feature divided and grouped to obtain a plurality of candidate feature subsets; feature optimization search and adjustment selection are performed on the candidate feature subsets to obtain a performance optimized feature subset that balances the number of features and prediction performance.

[0009] Optionally, in a fourth implementation of the first aspect of the present invention, the RBER prediction model for multi-period life prediction is constructed based on the performance optimization feature subset, including: respectively using the performance optimization feature subset to construct a time-dependent short-term prediction model, using the short-term prediction model to perform a short-term analysis on the recent storage data in the storage parameter data set to obtain a short-term RBER prediction result, and using the performance optimization feature subset to construct a nonlinear regression medium-term prediction model, using the medium-term prediction model to perform a medium-term analysis on the medium-term storage data in the storage parameter data set to obtain a medium-term RBER prediction result, and using the performance optimization feature subset to construct a deep learning long-term prediction model, and using the long-term prediction model to perform a medium-term analysis on the medium-term storage data in the storage parameter data set to obtain a medium-term RBER prediction result. The long-term storage data in the storage parameter data set is subjected to long-term analysis to obtain long-term RBER prediction results, and an integrated learning supplementary prediction model is constructed using an optimized feature subset, and the full-cycle storage data in the storage parameter data set is subjected to periodic analysis using the supplementary prediction model to obtain an auxiliary RBER prediction result; the short-term RBER prediction result, the medium-term RBER prediction result, the long-term RBER prediction result and the auxiliary RBER prediction result are weightedly combined and bias corrected to obtain a corrected fusion prediction result, and based on the corrected fusion prediction result, weights are dynamically assigned to the prediction results of each prediction model on different time scales to the corrected fusion prediction result to construct an RBER prediction model for multi-period life prediction.

[0010] Optionally, in a fifth implementation of the first aspect of the present invention, the RBER prediction model for life prediction in each period is used to predict the remaining number of programming / erase cycles and expected failure time of the programming / erase cycle data to obtain an initial life prediction result, including: dividing the programming / erase cycle data into a plurality of overlapping programming / erase cycle data segments in each period, and using the RBER prediction model for life prediction in each period to predict each of the programming / erase cycle data segments to obtain an RBER prediction value for each period; detecting whether the RBER prediction value of each period is greater than a preset period failure threshold, and based on the detection result, calculating the remaining number of programming / erase cycles corresponding to the RBER prediction value that is not greater than the preset period failure threshold; using the preset programming / erase The operation time interval statistical model generates the same number of simulated programming / erase operation time intervals as the remaining programming / erase cycle number, and uses the simulated programming / erase operation time intervals to divide and average the remaining programming / erase cycle number to obtain a preliminary expected failure time; uses a preset temperature acceleration model to adjust the preliminary expected failure time to obtain a preliminary expected failure time after temperature correction, and uses preset threshold voltage distribution data and a changing trend of the programming / erase cycle number to adjust the remaining programming / erase cycle number to obtain a corrected remaining programming / erase cycle number, and generates an initial life prediction result based on the preliminary expected failure time after temperature correction and the corrected remaining programming / erase cycle number.

[0011] Optionally, in a sixth implementation method of the first aspect of the present invention, the uncertainty quantification of the initial life prediction result to obtain a confidence interval for the life prediction and a corresponding interval prediction result includes: performing random perturbation analysis and statistical analysis on the initial life prediction result to obtain a point prediction result, and performing regression calculation on the point prediction result to obtain a confidence interval for the life prediction; and selecting a predicted value corresponding to the confidence interval from the initial life prediction result to obtain an interval prediction result.

[0012] A second aspect of the present invention provides a life prediction device for an SPI NAND flash memory, the life prediction device for the SPI NAND flash memory comprising: a preprocessing module, used to obtain a storage parameter data set of multi-dimensional performance parameters corresponding to a flash memory type of the SPI NAND flash memory, and perform multi-stage preprocessing on the storage parameter data set to obtain a cleaned storage parameter data set; a feature optimization module, used to perform multi-level performance feature extraction on the cleaned storage parameter data set to obtain a performance feature set, and perform adaptively selected feature optimization on the performance feature set to obtain a performance optimized feature subset; a model construction module, used to construct an RBER prediction model for multi-period life prediction based on the performance optimized feature subset, and obtain the SPI to be predicted. The invention discloses a plurality of program / erase cycle data of a NAND flash memory within a preset prediction time period; an initial prediction module, which is used to predict the remaining program / erase cycle times and expected failure time of the program / erase cycle data by using the RBER prediction model of the life prediction in each period, and obtain an initial life prediction result; a final prediction module, which is used to quantify the uncertainty of the initial life prediction result, obtain a confidence interval of the life prediction and a corresponding interval prediction result, and integrate the initial life prediction result, the confidence interval and the interval prediction result to obtain a final life prediction report of the SPI NAND flash memory.

[0013] A third aspect of the present invention provides a life prediction device for an SPI NAND flash memory, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the life prediction device for the SPI NAND flash memory performs each step of the above-mentioned life prediction method for the SPI NAND flash memory.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the various steps of the above-mentioned method for predicting the life of an SPI NAND flash memory.

[0015] The above-mentioned SPI NAND flash memory life prediction method, device, equipment and storage medium. In the embodiment of the present invention, a storage parameter data set of multi-dimensional performance parameters corresponding to the flash memory type of the SPI NAND flash memory is obtained, and the storage parameter data set is preprocessed in multiple stages to obtain a cleaned storage parameter data set; multi-level performance characteristics are extracted from the cleaned storage parameter data set to obtain a performance characteristic set, and the performance characteristic set is optimized by adaptive selection to obtain a performance optimization characteristic subset; based on the performance optimization characteristic subset, an RBER prediction model for multi-period life prediction is constructed, and multiple programming / erase cycle data of the SPI NAND flash memory to be predicted within a preset prediction time period are obtained; the RBER prediction model for life prediction in each period is used to predict the remaining programming / erase cycle number and expected failure time of the programming / erase cycle data to obtain an initial life prediction result; the uncertainty of the initial life prediction result is quantified to obtain a confidence interval of life prediction and a corresponding interval prediction result, and the initial life prediction result, the confidence interval and the interval prediction result are integrated to obtain a final life prediction report of the SPI NAND flash memory. Compared with the prior art, the present application obtains a storage parameter data set of multi-dimensional performance parameters of the corresponding flash memory type, and performs preprocessing, multi-level performance feature extraction and adaptive selection feature optimization on the storage parameter data set to obtain a performance optimization feature subset, and then uses the performance optimization feature subset to construct an RBER prediction model for short-term, medium-term and long-term life prediction. After obtaining multiple program / erase cycle data of the flash memory to be tested in the recent period, the multi-period RBER prediction model is used to predict the flash memory life for the program / erase cycle data, and then the uncertainty of the initial life prediction results is quantified and predicted and integrated to obtain the final life prediction report of the SPI NAND flash memory. This realizes the consideration of the influence of multi-period influencing factors when predicting the life of the target flash memory, and improves the accuracy of the flash memory life prediction results.

[0016] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of a first embodiment of a method for predicting the life of a SPI NAND flash memory in an embodiment of the present invention;

[0019] Figure 2A schematic diagram of an embodiment of a life prediction device for a SPI NAND flash memory in an embodiment of the present invention;

[0020] Figure 3 Schematic diagram of an embodiment of a life prediction device for a SPI NAND flash memory in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0023] To facilitate understanding of this embodiment, the specific process of the embodiment of the present invention is described below. Figure 1 , a first embodiment of the life prediction method of the SPI NAND flash memory in the embodiment of the present invention includes:

[0024] 101. Obtain a storage parameter data set of multi-dimensional performance parameters of a flash memory type corresponding to the SPI NAND flash memory, and perform multi-stage preprocessing on the storage parameter data set to obtain a cleaned storage parameter data set;

[0025] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0026] AI basic technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. AI software technologies mainly include computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0027] In this embodiment, the multi-dimensional performance parameters here include the number of program / erase (P / E) cycles, raw bit error rate (RBER), read and write voltage, erase voltage, threshold voltage distribution, P / E operation time interval, read retry number, programming time and temperature, etc.; the storage parameter data here refers to the historical storage data accumulated during the use of the SPI NAND flash memory and various flash memory related operation data. Standard deviation analysis is performed on the RBER and voltage values ​​in the storage parameter data set, and based on the results of the standard deviation analysis, storage parameter data that deviates from a preset numerical range is identified and marked to obtain an abnormal marked data set, and curve fitting and interpolation are performed on the short-term missing RBER data in the abnormal marked data set to obtain a complete RBER time series; based on the complete RBER time series, the RBER change rate of the abnormal marked data set is calculated, and based on the RBER change rate, forward data filling and sequence smoothing are performed on the long-term missing voltage data in the abnormal marked data set to obtain a complete voltage time series; time axis alignment and deformation matching are performed on the complete RBER time series and the complete voltage time series to obtain a synchronized multi-dimensional time series, and multi-scale decomposition and reconstruction are performed on the RBER data and voltage data corresponding to the multi-dimensional time series to obtain a smooth data curve; based on the extreme values ​​of the RBER curve and the voltage data curve in the smooth data curve, interval mapping is performed on the smooth data curve to obtain a standardized parameter value, and a nonlinear transformation is performed on the number of programming / erase cycles in the standardized parameter value to obtain a cleaned storage parameter data set.

[0028] In practical applications, a multi-dimensional performance parameter storage parameter data set corresponding to the NAND flash memory type is obtained. This data set includes key parameters such as the number of program / erase (P / E) cycles, raw bit error rate (RBER), read and write voltage, erase voltage, threshold voltage distribution, and P / E operation time interval. Then, a standard deviation analysis is performed on the RBER and voltage values ​​in the data set. By calculating the mean and standard deviation of the RBER and voltage values, a preset normal value range is determined, so that data points falling outside this range are regarded as potential outliers and are marked, thereby forming an abnormal marked data set; and for the short-term missing RBER data in the abnormal marked data set, a curve fitting is performed on the existing RBER data points, and then the fitted curve model is used to estimate the missing data points to obtain a complete RBER time series; based on the complete RBER time series, the rate of change of RBER (that is, it reflects the changing trend of RBER over time) is calculated, and based on this RBER change rate, the long-term missing voltage data in the abnormal marked data set is processed, that is, the last known voltage value is used to fill the subsequent missing data points, and then the filled data series is smoothed to eliminate the unnatural data jumps that may be caused by the filling, thereby obtaining a complete voltage time series; and in order to ensure the temporal consistency of the RBER and voltage data, the complete RBER time series and the complete voltage The time series are aligned on the time axis and deformed and matched. By processing the nonlinear deformation of the two time series on the time scale, the two series are completely synchronized in time, and a synchronized multi-dimensional time series is obtained. The RBER data and voltage data in the multi-dimensional time series are decomposed and reconstructed on a multi-scale basis using wavelet transform or empirical mode decomposition, so as to decompose the original signal into components of different scales, and then a smooth data curve is reconstructed (wherein this smooth data curve retains the main trend of the original data and removes high-frequency noise, making the data more stable and reliable). Then, the extreme points of the RBER curve and the voltage data curve are first found, and then the entire data range is interval mapped based on these extreme points, so that all data values ​​are mapped into a standardized range (usually between 0 and 1), so that data of different scales can be directly compared to obtain standardized parameter values, and the P / E cycle number in the standardized parameter value is nonlinearly transformed (such as logarithmic transformation or power law transformation) to capture the nonlinear relationship between the P / E cycle number and flash memory performance degradation, and finally a cleaned storage parameter data set is obtained. This dataset not only removes outliers and noise and fills in data gaps, but also achieves synchronization and standardization of multi-dimensional data. It also provides a reliable foundation for the construction and training of subsequent SPI NAND flash memory life prediction models, helping to improve the accuracy and reliability of predictions.

[0029] 102. Perform multi-level performance feature extraction on the cleaned storage parameter data set to obtain a performance feature set, and perform feature optimization of adaptive selection on the performance feature set to obtain a performance optimization feature subset;

[0030] In this embodiment, local statistical analysis is performed on the RBER data in the cleaned storage parameter data set to obtain RBER time domain statistical features, and statistical analysis is performed on the programming / erasing operation time interval data in the cleaned storage parameter data set to obtain programming / erasing operation mode features, and feature association is performed on the RBER time domain statistical features and the programming / erasing operation mode features to obtain comprehensive operation features; based on the comprehensive operation features and the storage parameter data set, storage cells corresponding to the first error rate and the second error rate are identified and marked to obtain error cell marking data, and based on the error cell marking data, error spatial distribution features of the storage cells in the flash memory array are extracted; based on the error spatial distribution features, The RBER time-domain statistical characteristics are converted into frequency domain and jointly analyzed to obtain the RBER frequency domain characteristics, and the RBER frequency domain characteristics and the voltage data in the cleaned storage parameter data set are jointly analyzed in time and frequency to obtain the time-frequency joint characteristics; the morphological parameters of the threshold voltage distribution are calculated using the time-frequency joint characteristics and the threshold voltage distribution data in the cleaned storage parameter data set to obtain the threshold voltage distribution characteristics, and based on the threshold voltage distribution characteristics, the read-write voltage and erase voltage data in the cleaned storage parameter data set are differentially analyzed in time series to obtain the voltage drift characteristics; based on the comprehensive operation characteristics, the error space distribution characteristics, the time-frequency joint characteristics and the voltage drift characteristics, a performance feature set is generated. The correlation coefficient between each feature in the performance feature set and the RBER data is calculated to obtain an initial feature importance ranking, and based on the initial feature importance ranking, the features with correlation coefficients greater than a preset correlation threshold are selected to obtain an initial feature subset; the nonlinear correlation between each feature in the initial feature subset and the number of programming / erase cycles is calculated to obtain an updated feature importance ranking, and based on the updated feature importance ranking, secondary feature selection is performed on the initial feature subset to obtain a related feature set; regularization constraints are applied to the related feature set to obtain a sparse feature subset, and the features in the sparse feature subset are arranged and smoothed in order of the number of programming / erase cycles to construct a smoothed feature importance score; time series prediction is performed on the feature importance score to obtain a feature importance prediction value for programming / erase cycles in future time periods, and based on the feature importance prediction value, the sparse feature subset is feature-divided and grouped to obtain multiple candidate feature subsets; feature optimization search and adjustment selection are performed on the candidate feature subset to obtain a performance optimized feature subset that balances the number of features and prediction performance.

[0031] In practical applications, a fixed-size window (e.g., 100 P / E cycles) is selected, and the mean, standard deviation, skewness, and kurtosis of RBER are calculated within this window. The window slides along the time axis, each time moving a certain step (e.g., 10 P / E cycles). These statistics are calculated repeatedly to achieve local statistical analysis of RBER data, and the time domain statistical characteristics of RBER are obtained. In addition, the program / erase operation time interval data is statistically analyzed to calculate the average interval time (reflecting the frequency of use of the flash memory), the coefficient of variation (indicating the discrete degree of the interval time), and the autocorrelation coefficient (revealing the periodicity of the operation mode). These statistics constitute the program / erase operation mode characteristics, and the Pearson correlation is used to analyze the program / erase operation mode characteristics. The RBER time-domain statistical features and programming / erasing operation mode features are correlated and analyzed by methods such as error statistics or mutual information, so as to obtain a comprehensive operation feature (wherein this feature set reflects the relationship between the RBER change and the flash memory usage mode); based on the comprehensive operation feature and the storage parameter data set, two error rate thresholds are defined: the first error rate (such as 100%) and the second error rate (such as 50%). For each storage unit, the number of errors in the entire data set is counted. If the error rate exceeds the first threshold, it is marked as a "severe error unit"; if the error rate exceeds the second threshold but is lower than the first threshold, it is marked as a "moderate error unit", so that the error unit marking data is obtained, and the flash memory array is regarded as a two-dimensional plane through Moran's I, and each unit corresponds to a coordinate point, so as to calculate the spatial autocorrelation coefficient of the error unit, where a positive Moran's I indicates that the error unit tends to cluster, and a negative value indicates dispersion. In addition, the aggregation index of the error unit, such as the nearest neighbor index, can also be calculated. These indicators together constitute the error spatial distribution characteristics, and the error spatial distribution characteristics of the storage unit in the flash memory array are extracted; then based on the error spatial distribution characteristics, the RBER time domain statistical characteristics are transformed into the frequency domain by fast Fourier transform and the characteristics are jointly analyzed, and then the first 10 main frequency components are selected as the frequency domain characteristics of RBER, and then the RBER frequency domain characteristics and voltage data are jointly analyzed in time and frequency by continuous wavelet transform, where the continuous wavelet transform formula is:

[0032] ;

[0033] Where a is the scale parameter, b is the translation parameter, and ψ is the wavelet function, which provides the performance of the signal at different time and frequency scales, thereby obtaining the time-frequency joint feature; then, the time-frequency joint feature and the threshold voltage distribution data are used to calculate the morphological parameters of the threshold voltage distribution (including the mean, standard deviation, skewness and kurtosis of the distribution), and for multi-level cells (such as TLC), the intervals and overlaps between the thresholds need to be calculated. These parameters constitute the threshold voltage distribution characteristics (reflecting the stability and distinguishability of the flash memory cell state); and based on the threshold voltage distribution characteristics, the read and write voltage and erase voltage data are subjected to time series differential analysis, and the first-order derivatives (rate of change) and second-order derivatives (acceleration) of these voltage values ​​are calculated by the central difference method. These derivatives reflect the trend of the voltage value over time, and constitute the voltage drift characteristics; and then, by integrating the comprehensive operation characteristics, error space distribution characteristics, time-frequency joint features and voltage drift characteristics, a comprehensive performance feature set is finally generated, among which, this feature set covers SPI Multiple aspects of NAND flash memory performance degradation, including error rate variation, operating mode impact, spatial distribution characteristics, time-frequency characteristics, and voltage drift trends.

[0034] Then, the correlation coefficient between each feature in the performance feature set and the RBER data is calculated by the Pearson correlation coefficient or the Spearman rank correlation coefficient, and the features are initially ranked by importance. The larger the absolute value of the correlation coefficient, the more significant the impact of the feature on the RBER, and the initial feature importance ranking is obtained. Based on the initial feature importance ranking, features with correlation coefficients greater than a preset correlation threshold (for example, 0.5) are selected to form an initial feature subset to filter out features with a low correlation with the RBER and reduce the complexity of subsequent processing; and then the nonlinear correlation between each feature in the initial feature subset and the number of programming / erase cycles is calculated by mutual information (MI), where the mutual information formula is as follows:

[0035] ;

[0036] Where X represents the feature, Y represents the number of programming / erase cycles, p(x, y) is their joint probability distribution, p(x) and p(y) are marginal probability distributions, respectively. The larger the mutual information value, the stronger the nonlinear correlation between the feature and the number of programming / erase cycles. Based on these nonlinear correlations, the feature importance ranking is updated, and the initial feature subset is subjected to secondary feature selection to obtain a relevant feature set to retain those features with a strong nonlinear relationship with the number of programming / erase cycles. These features may be ignored in linear correlation analysis, but are crucial for accurately predicting the life of flash memory. Regularization constraints are then applied to the relevant feature set (i.e., implemented by Lasso (L1 regularization) or Elastic Net (a combination of L1 and L2 regularization)) to obtain a sparse feature subset, and the features in the sparse feature subset are arranged in the order of the number of programming / erase cycles, and smoothed using the exponentially weighted moving average (EWMA) method. The exponentially weighted moving average formula is: , where S_t is the smoothed feature importance score, X_t is the original feature importance score, and α is the smoothing factor (0<α≤1). This smoothing process can reduce the short-term fluctuation of the feature importance score, highlight the long-term trend, and construct a smoothed feature importance score; in order to predict the feature importance of the program / erase cycle in the future period, a time series prediction model, such as the ARIMA (autoregressive integrated moving average) model, is used to perform time series prediction on the feature importance score to obtain the feature importance prediction value of the program / erase cycle in the future period, and based on the prediction value, hierarchical clustering or K-means clustering is used to perform feature partitioning and grouping on the sparse feature subsets, and features with similar feature importance prediction values ​​are grouped into the same group to obtain multiple candidate feature subsets; then, a genetic algorithm is used to perform feature optimization search and adjustment selection on the candidate feature subsets, that is, each candidate feature subset is encoded as a binary string, where 1 indicates that the feature is selected and 0 indicates that it is not selected, and a fitness function f is defined, which comprehensively considers the prediction performance (such as the mean square error of cross-validation) and the number of features: , where MSE is the mean square error, N_total is the total number of features, N_selected is the number of selected features, and λ is the balance parameter. Through the iterative evolution process (selection, crossover, mutation), an optimized feature subset that balances the number of features and prediction performance is obtained. This multi-stage, multi-angle feature selection process not only considers the correlation between the features and the target variable, but also considers the relationship between the features, the time series characteristics, and the contribution to the prediction performance. In this way, a feature subset that can fully express the performance changes of SPI NAND flash memory without causing overfitting is obtained.

[0037] 103. Based on the performance optimization feature subset, an RBER prediction model for multi-period life prediction is constructed, and multiple program / erase cycle data of the SPI NAND flash memory to be predicted within a preset prediction time period are obtained;

[0038] In this embodiment, the RBER prediction model for multi-period life prediction includes RBER prediction models for short-term, medium-term and long-term life prediction; respectively, a time-dependent short-term prediction model is constructed using a performance optimization feature subset, and a short-term prediction model is used to perform a short-term analysis on the recent storage data in the storage parameter data set to obtain a short-term RBER prediction result; a nonlinear regression medium-term prediction model is constructed using a performance optimization feature subset, and a medium-term prediction model is used to perform a medium-term analysis on the medium-term storage data in the storage parameter data set to obtain a medium-term RBER prediction result; and a deep learning long-term prediction model is constructed using a performance optimization feature subset, and the long-term prediction model is used to perform a medium-term analysis on the medium-term storage data in the storage parameter data set to obtain a medium-term RBER prediction result. The long-term storage data in the storage parameter data set are analyzed for a long time to obtain the long-term RBER prediction results, and an integrated learning supplementary prediction model is constructed using the optimized feature subset. The full-cycle storage data in the storage parameter data set is analyzed periodically using the supplementary prediction model to obtain the auxiliary RBER prediction results. The short-term RBER prediction results, the medium-term RBER prediction results, the long-term RBER prediction results and the auxiliary RBER prediction results are weightedly combined and bias corrected to obtain the corrected fusion prediction results. Based on the corrected fusion prediction results, the prediction results of each prediction model on different time scales are dynamically weighted to the corrected fusion prediction results to construct an RBER prediction model for multi-period life prediction.

[0039] In practical applications, time series analysis techniques such as ARIMA or LSTM networks are used to focus on the time correlation and short-term fluctuations in recent storage data to build a time-dependent short-term prediction model to analyze recent storage data and obtain short-term RBER prediction results (to capture sudden changes and short-term trends in flash memory performance). Support vector regression (SVR) or polynomial regression and other techniques are used to focus on the nonlinear relationship in medium-term storage data to build a medium-term prediction model of nonlinear regression. This model is used to capture the nonlinear changes of RBER over time and usage, thereby obtaining medium-term RBER prediction results (to predict performance changes of flash memory in a few months to a year). Deep neural networks or recurrent neural networks are used. This model can learn and capture complex patterns and long-term dependencies in long-term storage data to build a deep learning long-term prediction model. This model is used to analyze long-term storage data to obtain long-term RBER prediction results (to predict the performance of flash memory throughout its life cycle). can change); and in order to further improve the accuracy and robustness of the prediction, a variety of machine learning algorithms, such as random forests, gradient boosting trees, etc., are combined to analyze the full-cycle storage data and build an integrated learning supplementary prediction model, so that through this model, auxiliary RBER prediction results are obtained (providing valuable supplementary information for other models); and then after obtaining the prediction results of the four models, the four results are weighted according to the performance of each model on different time scales. For example, the short-term prediction model may get a higher weight when predicting the recent RBER, and the systematic errors that may exist in the model prediction are eliminated by using technologies such as Bayesian model averaging, and the corrected fusion prediction results are obtained. The performance of each prediction model on different time scales is dynamically evaluated by the sliding window technology, that is, the prediction accuracy of each model in the recent period is regularly evaluated, and their weights in the fusion prediction are dynamically adjusted accordingly, so as to build an RBER prediction model for multi-period life prediction that can be adaptively adjusted. In addition, multiple programming / erase cycle data (including RBER, voltage value, P / E cycle number and other key parameters) of the flash memory in the preset prediction time period (the most recent period of flash memory usage) are obtained. These data will serve as input to the constructed prediction model to analyze the future RBER change trend of the flash memory and thus predict its remaining life.

[0040] 104. Use the RBER prediction model for life prediction in each period to predict the remaining number of program / erase cycles and expected failure time of the program / erase cycle data to obtain an initial life prediction result;

[0041] In this embodiment, the program / erase cycle data is divided into a plurality of overlapping program / erase cycle data segments of each period, and the RBER prediction model of the life prediction of each period is used to predict each program / erase cycle data segment to obtain the RBER prediction value of each period; it is detected whether the RBER prediction value of each period is greater than the preset period failure threshold, and based on the detection result, the remaining number of program / erase cycles corresponding to the RBER prediction value not greater than the preset period failure threshold is calculated; the preset program / erase operation time interval statistical model is used to generate the same number of simulated program / erase operation cycles as the remaining number of program / erase cycles time interval, and using the simulated programming / erase operation time interval, divide and average the remaining programming / erase cycles to obtain a preliminary expected failure time; using a preset temperature acceleration model to adjust the preliminary expected failure time to obtain a preliminary expected failure time after temperature correction, and using the preset threshold voltage distribution data and the changing trend of the programming / erase cycle number to adjust the remaining programming / erase cycle number to obtain a corrected remaining programming / erase cycle number, and based on the preliminary expected failure time after temperature correction and the corrected remaining programming / erase cycle number, generate an initial life prediction result.

[0042] In practical applications, the acquired programming / erase cycle data is divided into multiple overlapping data segments of each period. For example, by using sliding window technology, different window sizes are set to represent short-term, medium-term and long-term. For example, the short-term may be the latest 100 P / E cycles, the medium-term may be the latest 1000, and the long-term may include all historical data to capture the RBER change trend on different time scales, and use the previously constructed RBER prediction model for each period to predict these data segments. The short-term model may focus on the instantaneous fluctuation of RBER, the medium-term model may capture the nonlinear relationship between RBER and the number of P / E cycles, and the long-term model may identify the change trend of RBER growth rate. By predicting these overlapping data segments, a series of RBER prediction values ​​for each period are obtained; then, it is detected whether the RBER prediction value for each period exceeds the preset period fault threshold, where this threshold may be dynamically adjusted with the increase of time and the number of P / E cycles to reflect the characteristics of flash memory performance changing over time. That is, for each predicted RBER value R_i and the corresponding fault threshold T_i, check whether R_i<=T_i is satisfied. Based on this detection result, the last RBER prediction value that does not exceed the failure threshold is found, and the corresponding remaining P / E cycle number (i.e., the expected remaining life of the flash memory under the current performance standard) is calculated; then, a preset P / E operation time interval statistical model is used to generate the same number of simulation time intervals as the remaining P / E cycle number, wherein this statistical model is constructed based on historical data (for example, using a Gaussian mixture model or kernel density estimation to fit the actual time interval distribution to construct), and after generating the simulation time interval, the remaining P / E cycle number is divided by the sum of these time intervals and the average is taken to obtain a preliminary expected failure time estimate. However, the performance and life of the flash memory are also significantly affected by temperature. Therefore, the preliminary expected failure time is adjusted using a preset temperature acceleration model (a model based on the Arrhenius equation), and through this model, the preliminary expected failure time is corrected according to the expected operating temperature to obtain the preliminary expected failure time after temperature correction. In addition, the threshold voltage distribution of the flash memory will also change with the increase in the number of P / E cycles, which will affect the actual remaining number of P / E cycles. The remaining number of P / E cycles is adjusted using the preset threshold voltage distribution data and the changing trend of the P / E cycle number. That is, the adjustment uses polynomial fitting or a machine learning model to describe the changing trend of the threshold voltage distribution, and then the remaining number of P / E cycles is corrected based on this trend. The preliminary expected failure time after temperature correction and the corrected remaining number of P / E cycles are combined to generate an initial life prediction result (wherein, the result may include information such as the expected failure time point, the confidence interval of the remaining number of P / E cycles, and the expected remaining usage time calculated based on the current usage pattern).

[0043] 105. Quantify the uncertainty of the initial life prediction results to obtain the confidence interval of the life prediction and the corresponding interval prediction results, and integrate the initial life prediction results, confidence interval and interval prediction results to obtain the final life prediction report of SPINAND flash memory.

[0044] In this embodiment, random perturbation analysis and statistical analysis are performed on the initial life prediction results to obtain point prediction results, and regression calculation is performed on the point prediction results to obtain a confidence interval for the life prediction; the predicted values ​​corresponding to the confidence interval are selected from the initial life prediction results to obtain an interval prediction result.

[0045] In practical applications, the Monte Carlo simulation method is used to add random noise to the key parameters in the initial prediction results (such as RBER prediction value, remaining P / E cycle number, etc.), and the prediction process is repeated many times to obtain a series of simulated prediction results. These simulation results constitute a prediction distribution, which reflects the uncertainty and variability of the prediction results, and statistical analysis is performed on this prediction distribution to calculate the mean, median, standard deviation and other statistics of various indicators. Among them, the mean or median can be used as a point prediction result, representing the most likely life prediction value, thereby obtaining a point prediction result; then the point prediction result is regressed, such as the linear regression model is: , where y is the life prediction value and x is a related factor (such as the number of P / E cycles experienced). and is the regression coefficient, is the error term. The regression coefficient is estimated by the least squares method to obtain the regression equation, and the standard error of the predicted value is calculated based on the regression equation, and then the confidence interval under a specific confidence level (such as 95%) is obtained (wherein this confidence interval indicates that there is a 95% probability that the true life value falls within this interval), and then the predicted values ​​falling within this interval are selected from the initial life prediction results. These values ​​constitute the interval prediction results, where the interval prediction results provide a more conservative and robust life estimation range for risk assessment and conservative decision making; and then the initial life prediction results, confidence intervals and interval prediction results are integrated together to form a comprehensive SPI NAND flash memory life prediction report. This report may include the following key parts: 1) Point prediction results, providing the most likely lifespan estimate; 2) Confidence interval, indicating the range of uncertainty in the prediction; 3) Interval prediction results, providing a conservative range of lifespan estimates; 4) Statistical characteristics of the prediction distribution, such as skewness, kurtosis, etc., reflecting the distribution characteristics of the prediction results; 5) Sensitivity analysis results, indicating which factors have the greatest impact on lifespan prediction; 6) Performance indicators of the prediction model, such as root mean square error (RMSE), coefficient of determination (R²), etc., reflecting the accuracy and fit of the prediction model.

[0046] In addition, the report can also include visual elements, such as probability density function diagrams of prediction results, graphical representations of confidence intervals, and trend diagrams of RBER changes with P / E cycle times. These charts intuitively display the prediction results and related uncertainties. Finally, based on these analysis results, the report can also provide some practical suggestions on how to formulate storage system management strategies and when to consider replacing or upgrading storage devices.

[0047] In the embodiment of the present invention, a storage parameter data set of multi-dimensional performance parameters of the corresponding flash memory type is obtained, and the storage parameter data set is preprocessed, multi-level performance feature extraction and adaptive selection feature optimization are performed to obtain a performance optimization feature subset, and then the performance optimization feature subset is used to construct an RBER prediction model for short-term, medium-term and long-term life prediction, and then after obtaining multiple program / erase cycle data of the flash memory to be tested in the recent period, the multi-period RBER prediction model is used to predict the flash memory life for the program / erase cycle data, and then the uncertainty of the initial life prediction result is quantified and predicted and integrated to obtain the final life prediction report of the SPI NAND flash memory. This realizes the consideration of the influence of multi-period influencing factors when predicting the life of the target flash memory, and improves the accuracy of the flash memory life prediction result.

[0048] The above describes the life prediction method of the SPI NAND flash memory in the embodiment of the present invention. The following describes the life prediction device of the SPI NAND flash memory in the embodiment of the present invention. Figure 2 , an embodiment of the life prediction device of SPI NAND flash memory in the embodiment of the present invention includes:

[0049] A preprocessing module 201 is used to obtain a storage parameter data set of multi-dimensional performance parameters of the flash memory type corresponding to the SPI NAND flash memory, and perform multi-stage preprocessing on the storage parameter data set to obtain a cleaned storage parameter data set;

[0050] The feature optimization module 202 is used to extract multi-level performance features from the cleaned storage parameter data set to obtain a performance feature set, and to optimize the performance feature set by adaptive selection to obtain a performance optimized feature subset;

[0051] A model building module 203 is used to build an RBER prediction model for multi-period life prediction based on the performance optimization feature subset, and obtain multiple program / erase cycle data of the SPI NAND flash memory to be predicted within a preset prediction time period;

[0052] An initial prediction module 204 is used to predict the remaining number of program / erase cycles and expected failure time of the program / erase cycle data using the RBER prediction model of the life prediction in each period to obtain an initial life prediction result;

[0053] The final prediction module 205 is used to quantify the uncertainty of the initial life prediction result, obtain the confidence interval of the life prediction and the corresponding interval prediction result, and integrate the initial life prediction result, the confidence interval and the interval prediction result to obtain the final life prediction report of the SPI NAND flash memory.

[0054] In the embodiment of the present invention, a storage parameter data set of multi-dimensional performance parameters of the corresponding flash memory type is obtained, and the storage parameter data set is preprocessed, multi-level performance feature extraction and adaptive selection feature optimization are performed to obtain a performance optimization feature subset, and then the performance optimization feature subset is used to construct an RBER prediction model for short-term, medium-term and long-term life prediction, and then after obtaining multiple program / erase cycle data of the flash memory to be tested in the recent period, the multi-period RBER prediction model is used to predict the flash memory life for the program / erase cycle data, and then the uncertainty of the initial life prediction result is quantified and predicted and integrated to obtain the final life prediction report of the SPI NAND flash memory. This realizes the consideration of the influence of multi-period influencing factors when predicting the life of the target flash memory, and improves the accuracy of the flash memory life prediction result.

[0055] above Figure 2 The life prediction device of the SPI NAND flash memory in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The life prediction device of the SPI NAND flash memory in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0056] Figure 3It is a structural diagram of a life prediction device for a SPI NAND flash memory provided by an embodiment of the present invention. The life prediction device 300 for the SPI NAND flash memory may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage devices) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 can be short-term storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the life prediction device 300 for the SPI NAND flash memory. Furthermore, the processor 310 may be configured to communicate with the storage medium 330 to execute a series of instruction operations in the storage medium 330 on the life prediction device 300 for the SPI NAND flash memory.

[0057] The life prediction device 300 of the SPI NAND flash memory may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The illustrated structure of the life prediction device for the SPI NAND flash memory does not constitute a limitation on the life prediction device for the SPI NAND flash memory, and may include more or less components than those illustrated, or a combination of certain components, or a different arrangement of components.

[0058] The present invention also provides a life prediction device for an SPI NAND flash memory, wherein the computer device includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes each step of the life prediction method for the SPI NAND flash memory in the above-mentioned embodiments.

[0059] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the various steps of the SPI NAND flash memory life prediction method.

[0060] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0061] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0062] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0063] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the life of a SPI NAND flash memory, characterized in that: The life prediction method of the SPI NAND flash memory comprises: Acquire a storage parameter data set of multi-dimensional performance parameters of a flash memory type corresponding to the SPI NAND flash memory, and perform multi-stage preprocessing on the storage parameter data set to obtain a cleaned storage parameter data set; Performing multi-level performance feature extraction on the cleaned storage parameter data set to obtain a performance feature set, and performing adaptively selected feature optimization on the performance feature set to obtain a performance optimized feature subset; Based on the performance optimization feature subset, an RBER prediction model for multi-period life prediction is constructed, and multiple program / erase cycle data of the SPI NAND flash memory to be predicted within a preset prediction time period are obtained; Using the RBER prediction model for life prediction in each period to predict the remaining number of program / erase cycles and expected failure time of the program / erase cycle data, to obtain an initial life prediction result; The uncertainty of the initial life prediction result is quantified to obtain a confidence interval of the life prediction and a corresponding interval prediction result, and the initial life prediction result, the confidence interval and the interval prediction result are integrated to obtain a final life prediction report of the SPI NAND flash memory.

2. The life prediction method of SPI NAND flash memory according to claim 1, characterized in that: The multi-stage preprocessing of the storage parameter data set to obtain a cleaned storage parameter data set includes: Performing standard deviation analysis on the RBER and voltage values ​​in the storage parameter data set, and based on the result of the standard deviation analysis, identifying and marking the storage parameter data that deviates from a preset numerical range to obtain an abnormal marked data set, and performing curve fitting and interpolation on the short-term missing RBER data in the abnormal marked data set to obtain a complete RBER time series; Based on the complete RBER time series, the RBER change rate of the abnormal mark data set is calculated, and based on the RBER change rate, the long-term missing voltage data in the abnormal mark data set is forward filled with data and the sequence is smoothed to obtain a complete voltage time series; Performing time axis alignment and deformation matching on the complete RBER time series and the complete voltage time series to obtain a synchronized multi-dimensional time series, and performing multi-scale decomposition and reconstruction on the RBER data and voltage data corresponding to the multi-dimensional time series to obtain a smooth data curve; Based on the extreme values ​​of the RBER curve and the voltage data curve in the smooth data curve, the smooth data curve is interval mapped to obtain a standardized parameter value, and the number of program / erase cycles in the standardized parameter value is nonlinearly transformed to obtain a cleaned storage parameter data set.

3. The life prediction method of SPI NAND flash memory according to claim 2, characterized in that: The multi-level performance feature extraction is performed on the cleaned storage parameter data set to obtain a performance feature set, including: Performing local statistical analysis on the RBER data in the cleaned storage parameter data set to obtain RBER time domain statistical features, and performing statistical analysis on the program / erase operation time interval data in the cleaned storage parameter data set to obtain program / erase operation mode features, and performing feature correlation on the RBER time domain statistical features and the program / erase operation mode features to obtain comprehensive operation features; Based on the comprehensive operation characteristics and the storage parameter data set, identifying and marking storage cells corresponding to the first error rate and the second error rate to obtain error cell marking data, and extracting error spatial distribution characteristics of the storage cells in the flash memory array based on the error cell marking data; Based on the error spatial distribution characteristics, the RBER time domain statistical characteristics are converted into frequency domain and jointly analyzed to obtain RBER frequency domain characteristics, and the RBER frequency domain characteristics and the voltage data in the cleaned storage parameter data set are jointly analyzed in time and frequency to obtain time-frequency joint characteristics; Calculating the morphological parameters of the threshold voltage distribution by using the time-frequency joint feature and the threshold voltage distribution data in the storage parameter data set after cleaning to obtain the threshold voltage distribution feature, and performing time series differential analysis on the read / write voltage and erase voltage data in the storage parameter data set after cleaning based on the threshold voltage distribution feature to obtain the voltage drift feature; A performance feature set is generated based on the comprehensive operation feature, the error spatial distribution feature, the time-frequency joint feature and the voltage drift feature.

4. The life prediction method of SPI NAND flash memory according to claim 3, characterized in that: The feature optimization of adaptively selecting the performance feature set to obtain a performance optimized feature subset includes: Calculating the correlation coefficient between each feature in the performance feature set and the RBER data in the cleaned storage parameter data set to obtain an initial feature importance ranking, and based on the initial feature importance ranking, selecting features with correlation coefficients greater than a preset correlation threshold to obtain an initial feature subset; Calculating the nonlinear correlation between each feature in the initial feature subset and the number of programming / erase cycles to obtain an updated feature importance ranking, and performing secondary feature selection on the initial feature subset based on the updated feature importance ranking to obtain a related feature set; Applying regularization constraints to the relevant feature set to obtain a sparse feature subset, and arranging and smoothing the features in the sparse feature subset in order of the number of programming / erase cycles to construct a smoothed feature importance score; Performing time series prediction on the feature importance scores to obtain feature importance prediction values ​​of program / erase cycles in future time periods, and performing feature partitioning and grouping on the sparse feature subsets based on the feature importance prediction values ​​to obtain a plurality of candidate feature subsets; The candidate feature subsets are subjected to feature optimization search and adjustment selection to obtain a performance optimized feature subset that balances the number of features and the prediction performance.

5. The life prediction method of SPI NAND flash memory according to claim 1, characterized in that: The RBER prediction model for multi-period life prediction is constructed based on the performance optimization feature subset, including: Respectively constructing a time-dependent short-term prediction model using the performance optimization feature subset, performing a short-term analysis on the recent storage data in the storage parameter data set using the short-term prediction model to obtain a short-term RBER prediction result, and constructing a nonlinear regression medium-term prediction model using the performance optimization feature subset, performing a medium-term analysis on the medium-term storage data in the storage parameter data set using the medium-term prediction model to obtain a medium-term RBER prediction result, and constructing a deep learning long-term prediction model using the performance optimization feature subset, performing a long-term analysis on the long-term storage data in the storage parameter data set using the long-term prediction model to obtain a long-term RBER prediction result, and constructing an integrated learning supplementary prediction model using the performance optimization feature subset, performing a periodic analysis on the full-cycle storage data in the storage parameter data set using the supplementary prediction model to obtain an auxiliary RBER prediction result; The short-term RBER prediction results, the medium-term RBER prediction results, the long-term RBER prediction results and the auxiliary RBER prediction results are weighted combined and bias corrected to obtain a corrected fusion prediction result, and based on the corrected fusion prediction result, the prediction results of each prediction model on different time scales are dynamically evaluated, and the weight of each prediction model in the fusion prediction is dynamically adjusted according to the result of the dynamic evaluation to construct an RBER prediction model for multi-period life prediction.

6. The life prediction method of SPI NAND flash memory according to claim 1, characterized in that: The RBER prediction model for predicting the life of each period is used to predict the remaining number of program / erase cycles and the expected failure time of the program / erase cycle data to obtain an initial life prediction result, including: Dividing the program / erase cycle data into a plurality of overlapping program / erase cycle data segments in each period, and using the RBER prediction model for life prediction in each period to predict each program / erase cycle data segment to obtain an RBER prediction value in each period; Detecting whether the RBER prediction value of each period is greater than a preset period failure threshold, and calculating the remaining number of programming / erase cycles corresponding to the RBER prediction value not greater than the preset period failure threshold based on the detection result; Generate a number of simulated program / erase operation time intervals equal to the number of remaining program / erase cycles using a preset program / erase operation time interval statistical model, and divide and average the remaining program / erase cycle numbers using the simulated program / erase operation time intervals to obtain a preliminary expected failure time; The preliminary expected failure time is adjusted using a preset temperature acceleration model to obtain the preliminary expected failure time after temperature correction, and the remaining programming / erase cycles are adjusted using preset threshold voltage distribution data and a change trend of the programming / erase cycle number to obtain a corrected remaining programming / erase cycle number, and an initial life prediction result is generated based on the preliminary expected failure time after temperature correction and the corrected remaining programming / erase cycle number.

7. The life prediction method of SPI NAND flash memory according to claim 1, characterized in that: The uncertainty quantification of the initial life prediction result to obtain a confidence interval of the life prediction and a corresponding interval prediction result includes: Performing random perturbation analysis and statistical analysis on the initial life prediction results to obtain point prediction results, and performing regression calculation on the point prediction results to obtain a confidence interval for life prediction; The predicted value corresponding to the confidence interval is selected from the initial life prediction result to obtain an interval prediction result.

8. A life prediction device for SPI NAND flash memory, characterized in that: The life prediction device of the SPI NAND flash memory comprises: A preprocessing module, used to obtain a storage parameter data set of multi-dimensional performance parameters of the flash memory type corresponding to the SPI NAND flash memory, and perform multi-stage preprocessing on the storage parameter data set to obtain a cleaned storage parameter data set; A feature optimization module is used to extract multi-level performance features from the cleaned storage parameter data set to obtain a performance feature set, and to perform adaptively selected feature optimization on the performance feature set to obtain a performance optimized feature subset; A model building module, used to build an RBER prediction model for multi-period life prediction based on the performance optimization feature subset, and obtain multiple program / erase cycle data of the SPI NAND flash memory to be predicted within a preset prediction time period; An initial prediction module is used to predict the remaining number of program / erase cycles and expected failure time of the program / erase cycle data using the RBER prediction model of the life prediction in each period to obtain an initial life prediction result; The final prediction module is used to quantify the uncertainty of the initial life prediction result, obtain the confidence interval of the life prediction and the corresponding interval prediction result, and integrate the initial life prediction result, the confidence interval and the interval prediction result to obtain the final life prediction report of the SPI NAND flash memory.

9. A life prediction device for SPI NAND flash memory, characterized in that: The life prediction device of the SPI NAND flash memory comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instruction in the memory to enable the SPI NAND flash memory life prediction device to perform each step of the SPI NAND flash memory life prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instruction is executed by the processor, each step of the life prediction method of the SPI NAND flash memory as described in any one of claims 1 to 7 is implemented.

Images