Adaptive data encoding and decoding method and system with flash channel feature guidance

Abstract

The application provides a flash channel feature guided adaptive data encoding and decoding method and system, and relates to the technical field of flash storage. The method comprises the following steps: according to a noise distribution form, a target check matrix which matches the current noise distribution form is adaptively selected from a pre-constructed multi-level check matrix set; wherein different check matrices in the multi-level check matrix set are respectively adapted to different noise degradation levels and noise skewness types; according to the noise distribution form, a soft information asymmetric correction parameter which is adapted to the current noise distribution form is calculated in a mode based on an elliptical focal point distance fixed constraint and a hyperbolic asymptote constraint. The application can realize the asymmetric degradation feature of the perceived channel noise, adaptively match the check matrix and correct the soft information, so that the stability of the decoding performance is maintained.

Description

Technical Field

[0001] This invention relates to the field of flash memory storage technology, and in particular to an adaptive data encoding and decoding method and system guided by flash memory channel characteristics. Background Technology

[0002] As the number of stacked layers in 3D NAND flash memory increases and the number of bits per cell increases, flash memory storage density continues to grow. In actual use, such as repeated programming and erasing operations, long-term data residence without refresh, and interference from neighboring cells introduced by read operations, the threshold voltage distribution of the storage cells will gradually degrade, and the distribution pattern of channel noise may gradually evolve from approximately symmetrical to a certain degree of asymmetric shift. That is, the diffusion of noise probability density in the positive and negative directions is no longer balanced, and the tail on one side is relatively more severe.

[0003] Currently, most common flash low-density parity-check (LDPC) decoding schemes rely on the simplified assumption of symmetric and unbiased channel noise when calculating soft information such as the log-likelihood ratio, and employ relatively fixed calculation methods.

[0004] However, existing technologies suffer from the following drawbacks: when flash memory channels degrade due to long-term use and noise distribution exhibits asymmetric shifts, a systematic deviation may occur between the soft information derived based on the symmetry assumption and the actual channel state. This leads to the LDPC decoder receiving poor-quality soft information input in the initial iteration stage. Consequently, as the number of programming and erasing cycles accumulates, the original bit error rate may exceed the error correction margin of the LDPC code, causing errors that are difficult to correct. Moreover, the decoder often needs to increase the number of iterations to converge in order to compensate for the soft information deviation, resulting in increased read latency fluctuations and affecting the consistency of the storage system's service quality. These problems are particularly pronounced in application environments such as enterprise-level solid-state drives and data center storage, where both reliability and response latency are critical. Summary of the Invention

[0005] This invention provides an adaptive data encoding and decoding method and system guided by flash channel characteristics. It can sense the asymmetric degradation characteristics of channel noise, adaptively match the parity check matrix and correct soft information deviation, thereby maintaining stable decoding performance and consistent read latency during channel degradation.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a flash memory channel feature-guided adaptive data encoding and decoding method, the method comprising: The channel state information of the target word line within the flash memory block is obtained. The channel state information includes the physical layer position of the target word line within the memory block, the cumulative number of program-erase cycles, the duration of data not being refreshed, and the cumulative number of read interferences. Based on the channel state information, the channel degradation type is determined. Based on the channel degradation type, the corresponding noise distribution pattern is determined. Based on the noise distribution pattern, a target verification matrix that matches the current noise distribution pattern is adaptively selected from a pre-constructed multi-level verification matrix set; wherein, different verification matrices in the multi-level verification matrix set are adapted to different noise degradation levels and noise skewness types. Based on the noise distribution pattern, soft information asymmetric correction parameters adapted to the current noise distribution pattern are calculated using a method based on fixed elliptical focal spacing constraints and hyperbolic asymptote constraints. Among them, the fixed elliptical focal spacing constraints are used to characterize the overall noise energy boundary of the channel, and the hyperbolic asymptote constraints are used to characterize the direction and degree of noise asymmetric offset. The soft information asymmetric correction parameters include the asymmetric offset direction and the asymmetric offset intensity. Using the target parity check matrix and soft information asymmetric correction parameters as input, low-density parity check iterative decoding is performed on the read data of the target word line to obtain the error-corrected output data.

[0007] Secondly, the flash memory channel characteristic-guided adaptive data encoding and decoding system includes: The channel degradation identification module is used to acquire the channel state information of the target word line in the flash memory block, determine the channel degradation type based on the channel state information, and determine the corresponding noise distribution pattern based on the channel degradation type. The channel state information includes the physical layer position of the target word line in the memory block, the cumulative number of programming and erasing cycles, the duration of data not being refreshed, and the cumulative number of read interferences. The verification matrix selection module is used to adaptively select a target verification matrix that matches the current noise distribution pattern from a pre-built multi-level verification matrix set; wherein, different verification matrices in the multi-level verification matrix set are adapted to different noise degradation levels and noise skewness types. The asymmetric parameter calculation module is used to calculate soft information asymmetric correction parameters adapted to the current noise distribution pattern, based on the fixed elliptical focal spacing constraint and the hyperbolic asymptote constraint. The fixed elliptical focal spacing constraint is used to characterize the overall noise energy boundary of the channel, and the hyperbolic asymptote constraint is used to characterize the direction and degree of noise asymmetric offset. The soft information asymmetric correction parameters include the asymmetric offset direction and the asymmetric offset intensity. The low-density parity decoding module is used to perform low-density parity iterative decoding on the read data of the target word line, taking the target parity matrix and soft information asymmetric correction parameters as input, to obtain the error-corrected output data.

[0008] The above-described solution of the present invention has at least the following beneficial effects: By acquiring multi-dimensional channel state information that characterizes the actual degradation state of the channel and determining the noise distribution pattern, the asymmetric offset characteristics of channel noise can be perceived. By adaptively selecting a verification matrix that matches the current noise degradation level and noise skewness type from a multi-level verification matrix set, the decoder structure itself has the ability to adapt to channel degradation.

[0009] By using the fixed constraint of the elliptical focal spacing to characterize the noise energy boundary and the hyperbolic asymptote constraint to characterize the direction and degree of asymmetric offset, a geometric constraint basis that matches the actual noise distribution is provided for soft information correction. This allows the corrected soft information to more accurately reflect the asymmetric characteristics of channel degradation, thereby maintaining relatively stable decoding performance when the flash memory channel undergoes different degrees of degradation and reducing the impact of bit error rate fluctuations on read latency. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the adaptive data encoding and decoding method guided by flash channel characteristics provided in an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of an adaptive data encoding and decoding system guided by flash channel characteristics provided in an embodiment of the present invention.

[0012] Figure 3 This is a flowchart illustrating how an embodiment of the present invention calculates soft information asymmetric correction parameters that are adapted to the current noise distribution pattern using a method based on fixed constraints on the focal spacing of an ellipse and constraints on hyperbolic asymptotes. Detailed Implementation

[0013] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0014] like Figures 1 to 3As shown, embodiments of the present invention propose an adaptive data encoding and decoding method guided by flash channel characteristics. In this embodiment, the flash storage device can be a NAND flash solid-state drive, an embedded flash storage chip (such as an embedded multimedia card eMMC, a general-purpose flash storage UFS), or an enterprise-level storage system based on a flash array.

[0015] For ease of description, the following explanation uses the flash controller in a 3D NAND flash solid-state drive as an example to illustrate the method of this invention. A flash controller refers to a dedicated processor and its firmware in a flash storage device that is responsible for managing the flash array, performing data encoding and decoding, and conducting wear leveling and garbage collection.

[0016] The target word line refers to the word line targeted by the current read or write operation. A word line is a control line in a NAND flash memory block that connects the same row of memory cells. Each word line corresponds to one or more pages of data. A memory block contains multiple word lines. Different word lines have different physical layer positions within the memory block (such as near the source side, near the drain side, or located in the middle region), and therefore experience different electric field stresses and interference from neighboring cells during programming and erasing.

[0017] The encoding and decoding in the method of this invention adopt the following parity check matrix usage strategy to ensure algebraic compatibility of encoding and decoding: On the encoding side, the flash memory controller always uses the generator matrix corresponding to the basic parity check matrix (i.e., the first layer in Table 2) in the multi-level parity check matrix set to perform LDPC encoding, encodes the information bits into codewords and writes them into the target word line. The generator matrix is ​​derived from the basic parity check matrix by Gaussian elimination.

[0018] On the decoding side, the flash memory controller adaptively selects the target parity check matrix based on the actual channel degradation state at the read time. To ensure compatibility, the multi-level parity check matrix set adopts a nested construction: each higher-level parity check matrix is ​​formed by concatenating all parity rows of the lower-level parity check matrix with a set of newly added parity rows. That is, the row sets of each level of parity check matrix have a hierarchical inclusion relationship, and the row set of the lower level is a subset of the row set of the higher level. Thus, all parity rows of the basic parity check matrix used during encoding (level 1, with the fewest rows and the highest code rate) are included in any higher-level decoding parity check matrix. The codewords generated at the write time naturally satisfy all parity constraints of any level of decoding parity check matrix at the read time, and the algebraic constraint relationship between the encoding and decoding matrices always holds. This embodiment focuses on the complete adaptive process on the decoding side. The encoding strategy is executed according to the standard LDPC encoding process in the flash memory controller write process.

[0019] Step 1: Obtain the channel state information of the target word line within the flash memory block. The channel state information includes the physical layer position of the target word line within the memory block, the cumulative number of programming and erasing cycles, the duration of data not being refreshed, and the cumulative number of read interferences. Based on the channel state information, determine the channel degradation type. Based on the channel degradation type, determine the corresponding noise distribution pattern.

[0020] A set of channel status registers is maintained by the flash memory controller to record the real-time channel status information of each word line within each memory block in the flash memory array. This channel status information refers to a multi-dimensional set of parameters reflecting the current degradation level of the flash memory channel where the target word line is located. When a data read operation needs to be performed on the target word line, the flash memory controller reads the following information corresponding to that target word line from the channel status registers: Physical layer location refers to the index of the target word line within its memory block. In a memory block containing a first predetermined number of word lines, the word line numbers increase sequentially from the source side to the drain side. Physical layer location information is used to characterize the region where the target word line is located in the memory block's spatial structure. Edge word lines (several word lines close to the source or drain side) and middle word lines have different initial channel quality characteristics due to manufacturing process variations and differences in electric field distribution.

[0021] The cumulative program-erase cycle count represents the total number of programming and erasing operations performed on the memory block since it was manufactured and put into use. Each time a complete programming operation is performed on all word lines of the memory block, and a block erase operation is executed, the flash memory controller increments the cumulative program-erase cycle count counter by 1. As the cumulative program-erase cycle count increases, trap charges gradually form in the tunneling oxide layer of the memory cells, causing the threshold voltage distribution to broaden and the mean to shift.

[0022] The duration of data not being refreshed is the actual time elapsed since the last data write or refresh operation was performed on the target word line. It is calculated by recording the completion time of each write operation using a system clock timer, and then calculating the difference between the current time and the most recent write time during subsequent reads. A longer duration of non-refreshing indicates a greater degree of charge loss due to leakage within the memory cell, and a more significant shift in the mean of the threshold voltage distribution towards lower voltage levels.

[0023] The cumulative read interference count is the total number of times other word lines within the memory block containing the target word line have been read since the last block erase operation. Each time a read operation is performed on any word line other than the target word line, a read path is established by applying a read voltage to the word line being read and simultaneously applying a conduction voltage to the target word line. This conduction voltage creates a weak programming effect on the memory cells of the target word line; that is, repeatedly applying the conduction voltage injects a small amount of charge into the memory cells of the target word line. As the cumulative count increases, the threshold voltage distribution shifts towards higher voltages.

[0024] The acquired four-dimensional channel state information vector is input into a pre-established channel degradation classifier. The channel degradation classifier outputs the current channel degradation type of the target word line. The channel degradation type refers to the classification label of the comprehensive degradation state of the flash memory channel after experiencing different degrees and modes of degradation. The process of establishing the channel degradation classifier is as follows: During the factory testing phase of flash memory storage devices, accelerated aging tests are performed on multiple batches of sample storage blocks under controlled conditions. Different combinations of cumulative program erase cycles (e.g., 100 times, 500 times, etc.), different durations of continuous data not being refreshed (e.g., 1 hour, 24 hours, etc.), and different cumulative read interference counts (e.g., 10,000 times, 100,000 times, etc.) are used to collect threshold voltage distribution data of sample word lines at each physical layer location under each combination condition.

[0025] For the threshold voltage distribution collected under each combination of conditions, the following two-dimensional features are extracted as a quantitative indicator of channel degradation, specifically described as follows: The distribution spread, or standard deviation of the threshold voltage distribution, quantifies the dispersion of the threshold voltage values ​​of each memory cell around the nominal center voltage value of its quantization level. It is calculated as follows: for all sample memory cells at a given quantization level, calculate the squared deviation of the threshold voltage value of each sample from the nominal center voltage value of that level; take the mean of the squared deviations to obtain the variance; and then take the arithmetic square root of the variance to obtain the distribution spread, with the same dimensions as the voltage value. A larger distribution spread indicates more severe noise diffusion and lower channel reliability. The distribution spread is used for subsequent noise degradation level determination.

[0026] Distribution skewness, or the dimensionless value of the third central moment of the threshold voltage distribution, is used to measure the degree of asymmetry of the distribution relative to its mean. It is calculated as follows: for all sample storage cells at a given quantization level, calculate the cubic deviation of the threshold voltage value of each sample from the nominal central voltage value. Take the mean of the cubic deviations to obtain the original third moment. Then, take the ratio of the original third moment to the cube of the distribution spread to obtain the dimensionless distribution skewness value.

[0027] A positive distribution skewness indicates that the tail of the threshold voltage distribution extends more on the high-voltage side than on the low-voltage side, typically corresponding to read interference-type degradation. A negative distribution skewness indicates that the tail on the low-voltage side extends more, typically corresponding to charge loss-type degradation. The larger the absolute value of the distribution skewness, the stronger the asymmetry. The distribution skewness is used to subsequently determine the type of noise skewness.

[0028] Based on the numerical range of the distribution width, the noise degradation level is divided into multiple discrete levels. In this embodiment, a four-level degradation level classification is adopted, as shown in Table 1 below: Table 1 Classification of Four Degradation Levels

[0029] Based on the sign and absolute value of the distribution skewness, noise skewness types are divided into three categories: left-skewed type, i.e., the distribution skewness is negative, indicating that the left tail of the threshold voltage distribution is relatively stronger, corresponding to the shift towards lower voltage caused by charge loss in the storage cell; symmetrical type, i.e., the absolute value of the distribution skewness is less than the first preset threshold, indicating that the threshold voltage distribution is approximately symmetrical from left to right; and right-skewed type, i.e., the distribution skewness is positive, indicating that the right tail of the threshold voltage distribution is relatively stronger, corresponding to the shift towards higher voltage caused by read interference.

[0030] The classic value of the first preset threshold is 0.05. This is based on the fact that, in statistics, the theoretical skewness of a normal (symmetric) distribution is 0. However, due to the limited sample size in actual testing, even if the underlying distribution is strictly symmetric, the sample skewness calculated from the limited sample will exhibit a small, non-zero value due to random fluctuations in sampling. Through threshold voltage acquisition and statistical analysis of a large number of flash memory storage units in their factory condition (i.e., without channel degradation), the measured absolute value of the skewness fluctuates between approximately 0.00 and 0.05 at a 95% confidence level.

[0031] Therefore, the first preset threshold is set to 0.05. When the absolute value of the skewness is less than 0.05, it is considered that the observed distribution pattern does not deviate significantly from the symmetrical distribution in a statistical sense. Therefore, the small skewness cannot be reliably attributed to the actual degradation of the channel, and it is classified as symmetrical.

[0032] Using the six-tuples (physical layer location, cumulative number of programmed erase cycles, duration of data not being refreshed, cumulative number of read interferences, distribution spread, and distribution skewness) extracted from the aforementioned accelerated aging test data as training samples, a mapping relationship is established from four-dimensional channel state information to channel degradation types (including noise degradation level and noise skewness type). In actual deployment, the above mapping relationship is stored as a set of multi-dimensional lookup tables. These multi-dimensional lookup tables are data structures that use multiple dimensions of input parameters as indexes and pre-calculated discretized output results as cell content. Their specific structure is as follows: Using physical layer location as the first dimension, the memory block is divided into three segments based on the location of the word line: source-side edge region, middle region, and drain-side edge region. Using the interval of cumulative program erase cycle count as the second dimension, the memory block is divided into multiple segments based on the order of magnitude of the cycle count. Using the interval of data not being refreshed for a continuous period of time as the third dimension, the memory block is divided into multiple segments based on the order of magnitude of the refresh duration. Using the interval of cumulative read interference count as the fourth dimension, the memory block is divided into multiple segments based on the order of magnitude of the read interference count.

[0033] The number of intervals in each dimension determines the total number of cells in the lookup table. Each cell stores a tuple containing a first field (noise degradation level) and a second field (noise skewness type). The values ​​of these tuples are pre-calibrated after statistical analysis of the combined conditions through the aforementioned accelerated aging test during the factory manufacturing stage.

[0034] In actual operation, the process of performing a table lookup operation on the target word line is as follows: Determine which segment of the first dimension the target word line belongs to based on its number within the storage block; read the current value of the cumulative programming erase cycle count counter of the target word line and determine which segment of the second dimension the value falls into; calculate the duration of the data not being refreshed by calculating the difference between the current time and the most recent write time, and determine which segment of the third dimension the duration falls into.

[0035] Read the current value of the cumulative number of read interference counts of the target word line and determine which interval of the fourth dimension the value falls into. Use the segment / interval number located in each of the above four dimensions as the four-dimensional index to locate the unique cell in the multidimensional lookup table and read the tuple stored in the cell. The tuple is the current channel degradation type, which includes the noise degradation level and the noise skewness type.

[0036] Based on the determined channel degradation type (including noise degradation level and noise skewness type), the corresponding noise distribution pattern is extracted from the pre-stored noise distribution pattern template library. The noise distribution pattern refers to a data structure that describes the probability density distribution characteristics of the threshold voltage of each memory cell on the target word line at multiple quantization levels, and specifically includes the following: For a four-level flash memory cell that stores 2 bits per cell, there are 4 quantization levels, corresponding to logic states 11 (erase state), 10, 00, and 01 respectively; each quantization level has its own nominal voltage center value.

[0037] The distribution width of each quantization level is the standard deviation of the threshold voltage distribution corresponding to that quantization level.

[0038] The distribution skewness value of each quantization level is the third central moment of the threshold voltage distribution corresponding to that quantization level.

[0039] The noise distribution pattern template library is pre-established during the equipment manufacturing stage, and the establishment process is as follows: For each degradation level (4 levels) and skewness type (3 categories), there are a total of 12 combinations. During accelerated aging testing, a fine-grained scan of the sample word lines under each combination is performed across the entire voltage range, with a step size of 5mV to 10mV, to reconstruct the complete threshold voltage distribution curve. The above three parameters are statistically analyzed for each distribution curve to form a noise distribution pattern record. All records of the 12 combinations are compiled into a template library and stored in the firmware in the form of a structured data table: the degradation level number is used as the first index, the skewness type number is used as the second index, and each cell stores one record.

[0040] In actual operation, based on the determined channel degradation type (including degradation level number and skewness type number), these two numbers are used as two-dimensional indexes to directly locate and read the corresponding noise distribution pattern record in the template library for use in subsequent steps. The acquisition of the above noise distribution pattern is based on the factory-calibrated discrete anchor points for table lookup matching, and does not involve real-time statistics or model fitting during online operation. As a supplement, the 12 combinations are discrete anchor points; if the actual degradation state falls between adjacent anchor points, the above three parameters of the adjacent anchor points can be linearly interpolated according to the actual distribution width, that is, the intermediate value is calculated proportionally between the corresponding index values ​​of the adjacent anchor points based on the continuous index value of the current degradation state, so as to cover the continuous degradation state with a finite discrete template. This interpolation operation is an optional post-processing after table lookup and does not change the table lookup matching architecture of the main path.

[0041] Step 2: Based on the noise distribution pattern, adaptively select a target verification matrix that matches the current noise distribution pattern from the pre-constructed multi-level verification matrix set; wherein, different verification matrices in the multi-level verification matrix set are adapted to different noise degradation levels and noise skewness types.

[0042] A multi-level parity check matrix set, i.e., a hierarchical set containing several low-density parity check (LDPC) matrices, is a data structure pre-built and stored in the firmware storage area of ​​the flash controller or in an external non-volatile storage medium during the flash controller firmware compilation stage or device initialization stage. Low-density parity check refers to a linear block error correction code, in which the density of non-zero elements in the check matrix is ​​much lower than the density of zero elements, and error correction is achieved by iteratively passing soft information between check nodes and variable nodes. The construction process of the multi-level parity check matrix set is as follows: First, a basic parity check matrix is ​​determined, serving as the root node of the parity check matrix set. The basic parity check matrix is ​​generated using a quasi-cyclic low-density parity check construction method. This method involves dividing the parity check matrix into several equal-sized square sub-matrix blocks, each with the same number of rows and columns. Typical sub-matrix block sizes are 32×32, 64×64, or 128×128. Each sub-matrix block is either an all-zero matrix or a cyclic permutation matrix obtained by cyclically right-shifting the identity matrix by a certain number of positions.

[0043] The cyclic permutation matrix is ​​generated as follows: Take the identity matrix (a square matrix with all elements on the main diagonal being 1 and all other elements being 0), and cyclically shift each row of the identity matrix to the right by a predetermined offset. The offset determines the distribution pattern of non-zero elements in this sub-matrix block, ranging from 0 to the size of the sub-matrix block minus 1. When the offset is 0, the sub-matrix block is the identity matrix itself; when the offset is any other value, the sub-matrix block is the corresponding cyclic shift matrix.

[0044] The advantage of using the quasi-cyclic construction method is that the parity check matrix does not need to be stored in the form of a complete matrix, but only the offset bits of each sub-matrix block need to be stored (if it is an all-zero matrix, it is marked with a special value such as negative 1), thus saving storage space. In the hardware decoder implementation, message routing between variable nodes and parity check nodes can be efficiently completed using the cyclic structure of sub-matrix blocks with the help of simple circuits such as barrel shift registers, avoiding the complex interconnection network required for a fully random matrix.

[0045] The code rate of the basic parity-check matrix is ​​chosen to be a relatively high rate suitable for scenarios with good channel conditions, such as a code rate between 0.90 and 0.93. The code rate refers to the ratio of the number of information bits to the total number of bits after encoding. The number of rows in the basic parity-check matrix represents the number of parity-check equations, and the number of columns represents the length of the encoded codeword.

[0046] Starting with a basic parity-check matrix, a progressive expansion operation is performed on the basic parity-check matrix to generate parity-check matrices adapted to higher noise degradation levels. The progressive expansion operation includes the following: While keeping the number of columns constant, the number of rows in the parity-check matrix is ​​increased, i.e., the number of parity-check equations is increased, thereby reducing the equivalent code rate and enhancing error correction capability. For each additional degradation level, the number of parity-check rows increases by a preset second expansion ratio, where the classic value for the second expansion ratio is 15% to 25%, meaning 15% to 25% more new parity-check rows are added to the previous level's number of parity-check rows. The newly added parity-check rows are generated according to the same structured construction rules as the original parity-check matrix, ensuring row redistribution between the new and original parity-check rows, i.e., the number of non-zero elements in each row, to maintain statistical consistency.

[0047] Taking the fourth-level degradation level as an example, the fourth-level check matrix generated through the above progressive expansion operation is shown in Table 2 below: Table 2. Level 4 Verification Matrix:

[0048] Within the same matrix level, for each noise skewness type, the column redistribution of the parity check matrix (i.e., the number of non-zero elements in each column) is adjusted differentially to generate a candidate parity check matrix that fits the skewness type. Specifically: for left-skewed noise types, the connectivity of variable nodes corresponding to the quantization level of the erase state (low voltage region) in the parity check matrix is ​​appropriately increased, that is, the density of non-zero elements is increased at the corresponding column positions, so that these variable nodes can obtain error correction information from more parity check equations.

[0049] For right-biased noise, appropriately increase the connectivity of variable nodes corresponding to the quantization level in the programming state (high voltage region); for symmetrical noise, maintain a balanced connectivity of variable nodes corresponding to each quantization level. These adjustments do not change the number of rows or the total number of non-zero elements in the parity check matrix; they only redistribute the density of non-zero elements among the columns.

[0050] The noise degradation level and noise skewness type are directly extracted from the output noise distribution pattern. These two quantities have already been determined in the channel degradation classification stage mentioned above; here, only the corresponding fields need to be read from the noise distribution pattern data structure.

[0051] Using the noise degradation level as an index, the corresponding matrix level is located in a multi-level parity check matrix set. The index refers to direct addressing based on the numerical code of the degradation level. For example, when the noise degradation level is level 3, the system directly locates level 3 in the multi-level parity check matrix set. This is implemented as follows: the flash memory controller firmware maintains an array of parity check matrix pointers. The array index is the noise degradation level number, and each array element is a pointer to the starting address of the corresponding level's parity check matrix. Using the current noise degradation level number as the array index, the corresponding pointer is retrieved, thereby obtaining all candidate parity check matrices for that level.

[0052] The noise skewness type is directly extracted from the output noise distribution pattern, i.e., one of left-skewed, symmetrical, or right-skewed. Given a determined noise skewness type, the following matching operation is performed among multiple candidate parity-check matrices within the same matrix level: Each candidate parity-check matrix carries a skewness type tag in its metadata, identifying the noise skewness type to which the matrix fits. The metadata tags of all candidate parity-check matrices at the current level are traversed, and the parity-check matrix whose skewness type tag precisely matches the current noise skewness type is selected and marked as the target parity-check matrix. This marking operation involves writing the starting storage address and dimension information (number of rows, number of columns) of the target parity-check matrix into the decoding parameter register inside the flash memory controller, for direct reference by the low-density parity-check decoding module in subsequent steps.

[0053] Through the above two-level matching mechanism, the first level matches the matrix hierarchy based on the degradation level, and the second level matches the specific matrix based on the skewness type. This achieves the adaptive binding of the verification matrix with the current channel degradation characteristics, ensuring that the verification constraints used by the subsequent decoder are consistent with the actual distribution characteristics of channel noise at the structural level.

[0054] Step 3: Based on the noise distribution pattern, calculate the soft information asymmetric correction parameters that are adapted to the current noise distribution pattern using a method based on the fixed elliptical focal spacing constraint and the hyperbolic asymptote constraint. The fixed elliptical focal spacing constraint is used to characterize the overall noise energy boundary of the channel, and the hyperbolic asymmetric constraint is used to characterize the direction and degree of noise asymmetric offset. The soft information asymmetric correction parameters include the asymmetric offset direction and the asymmetric offset intensity.

[0055] The core logic of this step is to map the asymmetric degradation characteristics of flash memory channel noise to an elliptic-hyperbolic joint geometric model for quantitative description. This elliptic-hyperbolic joint geometric model is defined on a two-dimensional plane with the threshold voltage value as the horizontal axis and the probability density value as the vertical axis; this plane is the voltage-probability density plane for flash memory channel noise analysis. In this two-dimensional plane, the two foci of the ellipse are located on the horizontal axis (voltage axis), and the two branches of the hyperbola extend towards the two foci respectively. The asymptotes in the two-dimensional plane have precise directional and slope geometric meanings. The ellipse, with its fixed distance between its two foci, characterizes the overall energy diffusion range of channel noise in the positive and negative voltage directions: the larger the energy boundary, the more severe the degradation, and the larger the focal distance. The hyperbola, with its asymptote direction and slope, characterizes the direction and degree of the noise probability density distribution deviating from the center of symmetry; the greater the skewness, the greater the slope of the asymptote.

[0056] By jointly solving the elliptic constraint and the hyperbolic constraint, the direction parameter and intensity parameter required for asymmetric correction of the log likelihood ratio of each quantization level are obtained. The specific process is as follows.

[0057] The interrelationships between parameters in the elliptic-hyperbolic joint geometric model are as follows: the two foci of the ellipse are determined by the noise degradation level, i.e., through the mapping from the noise diffusion range to the focal distance. These two foci also serve as the shared foci of the hyperbola, geometrically coupling the elliptic and hyperbolic constraints together. When the noise degradation level increases, the focal distance increases (elliptic constraint changes), and the basic direction of the hyperbola automatically adjusts due to the change in the focal position. When the noise skewness changes, the slope of the hyperbola's asymptote changes accordingly (hyperbolic constraint changes), thus affecting the asymmetric offset intensity obtained based on the slope mapping.

[0058] The two achieve parameter linkage through the geometric link of sharing a focal point. There is no need to perform additional joint optimization or iterative training process during the online operation of the flash controller. Instead, the entire mapping relationship (degradation level to focal spacing, skewness to slope, slope to offset intensity) is pre-established at the factory through the independent calibration process described in the aforementioned sub-steps. During online operation, only table lookup and algebraic operations are required to complete the end-to-end conversion from noise distribution pattern to soft information asymmetric correction parameters.

[0059] This pre-calibrated-online lookup architecture ensures the real-time performance of the joint model. The computational latency of the elliptic-hyperbolic parameters depends only on the processing time of the lookup and a few basic arithmetic operations, and is compatible with the read latency budget of the flash controller.

[0060] Based on the noise degradation level, a fixed value for the focal distance of the ellipse is determined, and this fixed value is used as a fixed constraint for the focal distance of the ellipse. Based on the noise distribution pattern, the positions of the two foci of the ellipse are determined, as follows: First, based on the determined noise degradation level, the corresponding noise diffusion range is looked up in the pre-defined mapping relationship between degradation level and noise diffusion range. The noise diffusion range refers to the maximum achievable offset, measured in voltage units (e.g., millivolts), from the nominal center voltage of each memory cell on the target word line under the current degradation conditions. There is a positive correlation between noise diffusion range and noise degradation level; the higher the degradation level, the larger the noise diffusion range.

[0061] The above mapping relationship is pre-established during the equipment manufacturing stage by statistically analyzing the threshold voltage distribution of sample word lines under different degradation levels. The calibration process is as follows: for each degradation level, all sample word lines belonging to that level are selected. For each quantization level of each sample word line, the distribution width is extended by 3 times in both the positive and negative directions with the nominal center voltage value as a reference, which is taken as the noise diffusion boundary of that quantization level. The 3-fold extension range statistically covers approximately 99.7% of the sample points. The absolute value of the difference between the maximum value of the upper boundary and the minimum value of the lower boundary of all quantization levels is taken as the total noise diffusion voltage span of that word line.

[0062] The average of the total noise diffusion voltage span of all sample word lines under this level is taken to obtain the noise diffusion range calibration value corresponding to this degradation level; then the ratio of this calibration value to the basic scene calibration value is taken to obtain the normalized value of the ellipse focal distance. The three fields of degradation level, noise diffusion range calibration value, and ellipse focal distance normalized value are stored together in the noise diffusion range and degradation level mapping table.

[0063] The mapping table obtained through the above calibration process is shown in Table 3 below: Table 3. Mapping Table of Noise Propagation Range and Degradation Level:

[0064] In the mapping table above, the normalized unit of the elliptical focal distance refers to a dimensionless value normalized based on a reference voltage span (such as the noise diffusion range calibration value under the basic scenario). According to the current noise degradation level, the corresponding row is located in the mapping table, and the normalized value of the elliptical focal distance is read. This value is used as a fixed constraint on the elliptical focal distance; that is, under the current channel conditions, the distance between the two foci of the ellipse is a fixed value and does not change with subsequent calculations.

[0065] Based on the noise distribution pattern, the overall energy distribution center position of the channel noise is extracted. From the output noise distribution pattern, the nominal center voltage value and distribution width of each quantization level are extracted. Then, the following weighted average calculation is performed to obtain the overall energy distribution center position of the channel noise. The overall energy distribution center position refers to the comprehensive reference point obtained by weighting the nominal centers of each quantization level according to their respective reciprocals of the signal-to-noise ratio in the soft information domain (hereinafter referred to as the log-likelihood ratio domain), which is used to determine the symmetry center of the elliptical focal distribution.

[0066] The method for determining the location of the overall energy distribution center is as follows: For each quantization level, the ratio of 1 to the distribution width of that quantization level is taken to obtain the signal-to-noise ratio weight of that quantization level. The signal-to-noise ratio weight is used to measure the information contribution ratio of that quantization level in the overall noise distribution. The larger the distribution width (the stronger the noise), the smaller the weight of that quantization level.

[0067] The signal-to-noise ratio weights of each quantization level are accumulated level by level to obtain the total weight value.

[0068] For each quantization level, its nominal center voltage value is multiplied by the signal-to-noise ratio weight of that quantization level to obtain the weighted center value of that quantization level.

[0069] The weighted center values ​​of each quantization level are accumulated level by level to obtain the total weighted center value; the ratio of the total weighted center value to the total weight value is used to obtain the overall energy distribution center position.

[0070] Using the center of the overall energy distribution as the center of symmetry and the fixed constraint of the focal spacing of the ellipse as the constraint condition, the positions of the two foci of the ellipse are determined; wherein, the two foci are symmetrical about the center of the overall energy distribution, and the distance between the two foci is equal to the fixed value of the focal spacing of the ellipse. The specific implementation process is as follows: The overall energy distribution center obtained above is the center of symmetry. Using this center of symmetry as a reference, the voltage axis in the soft information domain (log-likelihood ratio domain) is offset by the same distance in both the positive and negative directions, such that the distance between the two offset points is exactly equal to the fixed value of the determined elliptical focal distance. Specifically: Take half of the fixed value of the focal spacing of the ellipse to obtain the focal half spacing.

[0071] Starting from the center of the overall energy distribution, the first focal position is obtained by shifting the focal half-spacing distance in the positive direction of the voltage axis.

[0072] Starting from the center position of the overall energy distribution, the second focal position is obtained by shifting the focal half-spacing distance in the negative direction of the voltage axis.

[0073] The two focal positions mentioned above are symmetrically distributed about the center of the overall energy distribution, and the distance between them is equal to the fixed value of the elliptical focal spacing. The fixed constraint of the elliptical focal spacing is manifested here as follows: regardless of subsequent changes in channel conditions, within the determination period of the current degradation level, the distance between the two focal positions remains constant, that is, the basic shape parameter of the ellipse, the focal length, is locked by the degradation level.

[0074] The two focal positions and the fixed value of the focal distance are output together as a complete description of the fixed focal distance constraint of the ellipse, which is used for the joint constraint calculation in subsequent steps.

[0075] Based on the noise skewness type and the positions of the two foci of the ellipse, the direction and slope of the hyperbola asymptote are determined as hyperbola asymptote constraints; whereby the direction of the hyperbola asymptote characterizes the direction of the noise asymmetric shift, and the slope of the hyperbola asymptote characterizes the degree of noise asymmetric shift.

[0076] Hyperbolic asymptote constraints are used to quantitatively describe the directional and intensity properties of noise deviating from a symmetrical distribution. The two asymptotes of the hyperbola extend towards the two foci, and their directions and slopes together determine the geometric characteristics of the asymmetrical expansion of the noise probability density distribution in both positive and negative directions.

[0077] Based on the determined noise skewness type, perform the following conditional branch judgment: When the noise bias type is left-biased, the bias direction points to the first focus of the ellipse; in this embodiment, the first focus is the focus located on the negative side of the voltage axis mentioned above, which corresponds to the physical meaning of the threshold voltage shifting towards a lower voltage direction (charge loss-type degradation). The bias direction is marked as biased towards the first focus.

[0078] When the noise skewness type is right-skewed, the bias direction points to the second focus of the ellipse; the second focus is the focus located on the positive side of the voltage axis mentioned above, which corresponds to the physical meaning of the threshold voltage shifting towards the high voltage direction (reading interference-type degradation). The bias direction is marked as biased towards the second focus.

[0079] When the noise skewness type is symmetrical, the bias direction is zero, meaning no component is generated biased towards either side of the focal point. The bias direction is marked as no bias.

[0080] The determination of the bias direction directly maps to the physical orientation of the noise asymmetric offset: whichever focal point the bias direction points to means that the noise probability density distribution on the voltage side of that focal point needs more soft information correction compensation. The soft information correction compensation refers to enhancing and adjusting the absolute value of the log-likelihood ratio on the biased side when the decoder calculates the log-likelihood ratio, so that the decoder can more reliably distinguish the storage state between adjacent quantization levels on that side.

[0081] Determine the basic direction of the asymptote of the hyperbola based on the positions of the two foci of the ellipse; the basic direction is consistent with the direction of the line connecting the two foci.

[0082] Determine the voltage coordinates at the first and second foci. The direction of the straight line segment connecting the two foci is taken as the fundamental direction of the hyperbola's asymptote. The fundamental direction is determined as follows: Subtracting the voltage coordinates of the first focal point from the voltage coordinates of the second focal point yields the direction vector of the line connecting the focal points. This direction vector determines the fundamental direction: when the voltage value at the second focal point is greater than that at the first focal point (i.e., the second focal point is on the positive voltage side), the fundamental direction points in the positive direction of the voltage axis; otherwise, it points in the negative direction.

[0083] Using the aforementioned basic direction as a reference, a bias direction is superimposed for correction to determine the direction of the hyperbola asymptote; wherein, this direction characterizes the direction of the noise asymmetric offset.

[0084] The final direction of the hyperbola's asymptote is obtained by superimposing the determined base direction with the determined offset direction. The superposition rules are as follows: When the bias direction is unbiased (i.e., the noise skewness type is symmetrical): the final direction is equal to the fundamental direction. At this time, the hyperbola degenerates into a standard form symmetrical about the two foci, and the asymptote direction is the direction of the line connecting the foci.

[0085] When the bias direction is towards the first focus (i.e., left-biased noise): the final direction, based on the base direction, deflects towards the side where the first focus is located by a predetermined first deflection angle. The classic value for the first deflection angle is between 5° and 15°. The magnitude of the first deflection angle is positively correlated with the noise bias; that is, the larger the absolute value of the bias, the larger the first deflection angle. The specific mapping relationship will be handled uniformly in subsequent steps.

[0086] When the bias direction is biased towards the second focus (i.e., right bias noise): the final direction is based on the basic direction, and is biased towards the side where the second focus is located by the first bias angle.

[0087] The final direction of the hyperbolic asymptote represents the direction of the noise asymmetry shift, indicating which side of the voltage axis the log-likelihood ratio correction should be biased towards in the soft information domain.

[0088] Based on the noise distribution pattern, the noise skewness is extracted; the greater the noise skewness, the greater the absolute value of the slope of the hyperbola asymptote; based on the noise skewness and the preset mapping relationship between skewness and slope, the slope of the hyperbola asymptote is determined; this slope characterizes the degree of noise asymmetric shift.

[0089] From the output noise distribution pattern, the distribution skewness value of each quantization level is extracted. These distribution skewness values ​​are then combined to obtain the noise skewness, which is a scalar value used to quantify the extent to which the noise distribution of the entire channel deviates from the symmetrical pattern.

[0090] The comprehensive calculation method for noise skewness is as follows: Take the absolute value of the distribution skewness of each quantization level.

[0091] The absolute deviation values ​​of each quantization level are accumulated level by level to obtain the total absolute deviation.

[0092] The average absolute skewness is obtained by taking the ratio of the total absolute skewness to the total number of quantization levels, and is used as the noise skewness.

[0093] The obtained noise skewness is input into a preset mapping relationship between skewness and slope to determine the slope of the hyperbola asymptote. This mapping relationship is established during the equipment manufacturing stage by performing hyperbola fitting calibration on sample data under different skewness conditions. The hyperbola fitting calibration refers to using the hyperbola equation to mathematically approximate the contour of the measured threshold voltage distribution on a two-dimensional plane of voltage and probability density, thereby extracting the geometric parameters associated with the noise asymmetric offset characteristics. The specific calibration process is as follows: For sample data preparation, sample word lines covering different noise skewnesses (from approximately 0.00 for near-symmetric to approximately 1.50 for extremely skewed) were selected from the accelerated aging test database. These were divided into multiple subsets based on their skewness values, with each subset containing a predetermined number of sample word lines (e.g., 50 to 100). For each sample word line, discrete data points were extracted on the voltage-probability density plane. A two-dimensional scatter plot was constructed with the threshold voltage value as the x-axis and the corresponding statistical probability density as the y-axis.

[0094] The hyperbola equation is parametrically represented, wherein the hyperbola is uniquely determined by its two foci and a vertex distance difference. In this embodiment, the two foci of the hyperbola are the same as the output first and second focal positions, which have been determined in the elliptic constraints; the hyperbola and the ellipse share the same pair of foci. The absolute value of the difference between the distance from any point on the hyperbola to the first focal position and the distance from that point to the second focal position is defined as a constant, denoted as the vertex distance difference. This constant determines the opening size of the hyperbola and the steepness of its asymptotes.

[0095] To begin constructing the objective function, for a given set of sample data points (voltage value-probability density value pairs), the goodness-of-fit metric is constructed as follows: For each data point, calculate the perpendicular distance from the data point to the geometric contour of the hyperbola. This is calculated by drawing a perpendicular line from the data point to the hyperbola and finding the length of the line segment from the foot of the perpendicular to the data point. Squaring the perpendicular distances for all data points and summing them up yields the residual sum of squares. The smaller the residual sum of squares, the better the hyperbola fits the given set of sample data points.

[0096] Searching for the optimal vertex distance difference. Keeping the two focal points fixed, within a preset range of vertex distance difference values, determined by the focal distance and the voltage span of the sample data, the vertex distance difference value is changed successively with a preset search step size (e.g., 1% of the focal distance). For each vertex distance difference value, a corresponding hyperbola geometric profile is generated according to the described hyperbola definition, and the corresponding residual sum of squares is calculated using the same method described above. After traversing all vertex distance difference values, the vertex distance difference value that minimizes the residual sum of squares is selected as the optimal vertex distance difference for that subset of sample data.

[0097] The slope of the asymptotes is extracted as follows, given a determined optimal vertex distance difference and two known focal positions: Half the distance between the first and second foci is taken as the focal half-distance, which is the linear eccentricity parameter of the hyperbola in geometry and is measured in voltage units (millivolts); half the distance difference between the optimal vertices is taken as the vertex half-distance difference, which is also measured in voltage units.

[0098] Calculate the square of the half-distance between the foci and the square of the half-distance between the vertices, and take the arithmetic square root of the difference to obtain the second intermediate quantity. Take the ratio of the second intermediate quantity to the half-distance between the vertices to obtain the absolute value of the slope of the asymptote. This absolute value of the slope is a dimensionless value, and its geometric meaning is the tangent of the angle of inclination of the hyperbola asymptote relative to the line connecting the foci.

[0099] The absolute value of the slope of an asymptote means the degree of inclination of the straight line (i.e., the asymptote) that a branch of a hyperbola approaches from the vertex to infinity relative to the line connecting the foci. The larger the absolute value of the slope of the asymptote, the larger the angle between the two branches of the hyperbola, which in turn means a stronger skewness in the noise distribution.

[0100] To establish a mapping pair, the average skewness (first field) of the sample data subset is paired with the absolute value of the asymptote slope extracted in the previous step (second field), forming a mapping pair. This process is repeated for all skewness subsets to obtain the complete mapping relationship.

[0101] The following table, Table 4, shows the skewness-slope mapping obtained through the above calibration process: Table 4. Skewness-Slope Mapping Table:

[0102] In the mapping table above, the larger the absolute value of the asymptote slope, the steeper the noise probability density distribution is in the deviation direction, i.e., the stronger the asymmetry. When the noise skewness type is left-skewed, the slope sign is negative; when it is right-skewed, the slope sign is positive; and when it is symmetrical, the slope is zero.

[0103] The slope sign is determined based on the current noise skewness type. The corresponding absolute slope value interval is located in the mapping table based on the noise skewness, and the median of this interval is taken as the final slope of the hyperbola asymptote. This slope, along with the determined final direction, is output as a complete description of the hyperbola asymptote constraint.

[0104] Using the two foci of the ellipse and the direction and slope of the hyperbola as constraints, soft information asymmetric correction parameters adapted to the current noise distribution pattern are calculated based on the noise distribution pattern. The soft information asymmetric correction parameters include the asymmetric offset direction and the asymmetric offset intensity.

[0105] At this point, the following geometric constraints have been obtained: the positions of the two foci of the ellipse and their fixed spacing, characterizing the overall noise energy boundary of the channel; and the direction and slope of the hyperbola's asymptote, characterizing the direction and extent of the noise asymmetric shift. This sub-step applies the two sets of geometric constraints jointly to the soft information domain to calculate the parameters required for asymmetric correction of the log-likelihood ratio for each quantization level.

[0106] Based on the positions of the two foci of the ellipse, the symmetry reference for soft information correction is determined; where the symmetry reference is the perpendicular bisector of the line connecting the two foci, and this symmetry reference corresponds to the zero-offset reference line in the soft information domain.

[0107] Given the determined first and second focal positions, calculate the midpoint of the line connecting the two focal positions. This midpoint is obtained by summing the voltage coordinates of the two focal points and dividing by 2. Then, draw a straight line perpendicular to the line connecting the focal points through this midpoint (represented as a point on the one-dimensional voltage axis and as a hyperplane in the multi-dimensional soft information domain). This perpendicular bisector is the symmetry reference.

[0108] The meaning of a symmetrical reference is: under ideal conditions of noise-free asymmetric offset, the log-likelihood ratio of each quantization level should be symmetrically distributed about this symmetrical reference. In the soft information domain, the symmetrical reference corresponds to the zero-offset reference line, that is, the correction of the log-likelihood ratio should be based on this line and offset to one side.

[0109] The elliptic focal position and symmetry reference are defined on the voltage axis (in millivolts), rather than directly in the log-likelihood ratio (LLR) domain. This is because the initial LLR of each memory cell is calculated based on the read voltage value, which is directly mapped to a specific position on the voltage axis. Therefore, determining the voltage offset direction—that is, determining which side of the symmetry reference the read voltage is on—and thus deciding the direction of the LLR correction, is logically consistent and feasible.

[0110] Based on the direction of the hyperbola asymptote, the asymmetric offset direction for soft information correction is determined; whereby the asymmetric offset direction points to the side of the noise probability density distribution that needs to be enhanced and compensated.

[0111] The direction of the hyperbola asymptote is directly determined as the asymmetric offset direction for soft information correction. The asymmetric offset direction points to the side where the noise probability density distribution is relatively sparse and requires enhanced compensation: when the noise is left-biased, the left tail of the voltage distribution is intensified, meaning that the noise on the low-voltage side is stronger than that on the high-voltage side. Therefore, the decoder needs a larger soft information amplitude on the low-voltage side to compensate for the uncertainty caused by the noise. That is, the asymmetric offset direction points to the low-voltage side, i.e., the first focal side; when the noise is right-biased, the opposite is true, and the asymmetric offset direction points to the high-voltage side, i.e., the second focal side.

[0112] Based on the slope of the hyperbola asymptote and the preset mapping relationship between the slope and the offset intensity, the asymmetric offset intensity of the soft information correction is determined; where the larger the absolute value of the slope, the greater the asymmetric offset intensity.

[0113] The determined slope (including sign) of the hyperbola asymptote is input into the preset mapping relationship between slope and offset intensity to determine the asymmetric offset intensity of soft information correction. The asymmetric offset intensity is a scalar value used to control the amplitude scaling factor when asymmetrically adjusting the log-likelihood ratio of each quantization level.

[0114] The above mapping relationship is established during the equipment manufacturing stage through the following calibration process: Select multiple slope levels covering different absolute values ​​of asymptotic slopes (from approximately 0.00 to approximately 1.50). For each slope level, select a batch of sample memory blocks on the flash memory test chip, and adjust the channels of each sample memory block so that the asymptotic slope falls into the degradation state of that slope level. Prepare a third predetermined number (e.g., 100 to 200) of sample word lines for each slope level.

[0115] For each sample word line at each slope level, a known test data pattern is first written, followed by a normal read operation to acquire the original decision voltage value of each memory cell. Simultaneously, the average number of iterations required to decode each sample word line using a standard LDPC decoder (without asymmetric correction) is recorded as the baseline number of iterations for that slope level.

[0116] For each slope level, a set of candidate offset intensity values ​​is preset. The candidate offset intensity values ​​start at 0.00 and increase sequentially with a preset offset intensity search step size (e.g., 0.05) until a preset upper limit for the offset intensity (e.g., 0.85) is reached. For each candidate offset intensity value, the following operations are performed: The candidate offset intensity value is applied to the log-likelihood ratio asymmetric correction unit, and asymmetric correction is performed on the initial soft information sequence of each sample word line according to the correction rules described in the preceding steps.

[0117] Using the corrected soft information sequence as input, perform LDPC iterative decoding (the decoder configuration is the same as in step 4).

[0118] Calculate the average number of decoding iterations for the sample word line under the current candidate offset strength.

[0119] For each slope level, among all candidate offset strength values ​​at that level, the candidate offset strength value that minimizes the average number of decoding iterations is selected as the optimal offset strength for that slope level. If multiple candidate offset strength values ​​all reach the same minimum number of iterations, the one with the smallest absolute value of the offset strength is selected, prioritizing a conservative correction approach to reduce the risk of over-correction.

[0120] Pair the absolute value of the asymptote slope of the slope setting (first field) with the optimal offset intensity determined above (second field) to form a mapping pair. Repeat the above steps for all slope settings to obtain the complete mapping relationship.

[0121] The following table, Table 5, shows the slope-offset intensity mapping obtained through the above calibration process: Table 5. Mapping table of slope and offset intensity:

[0122] In the mapping table above, the offset strength is a dimensionless scaling factor, ranging from 0 to 1. 0 indicates no asymmetric correction; a larger value indicates a greater asymmetric adjustment to the log-likelihood ratio. The offset strength has an additional sign attribute: the sign is the same as the slope, negative for left skew, positive for right skew, and 0 for symmetry. This sign is only used to indicate the direction of the asymmetric offset, i.e., which side needs enhanced compensation, and does not participate in the numerical scaling calculation. When actually substituting into the above steps for correction (1 ± offset strength), the absolute value of the offset strength is always used in the calculation.

[0123] In this mapping table, the corresponding interval is located based on the absolute value of the asymptote slope. The median of the interval is taken as the absolute value of the offset intensity. Then, the sign of the offset intensity is set to be the same as the sign of the slope to obtain the final offset intensity.

[0124] Based on the noise distribution pattern, the initial log-likelihood ratio of each quantization level is extracted.

[0125] From the output noise distribution, for each quantization level, the initial log-likelihood ratio under that noise distribution is extracted. The initial log-likelihood ratio is the natural logarithm of the ratio of the probability that a given read voltage value is programmed into an erase state to the probability that it is programmed into a programmed state, without asymmetric correction. A positive log-likelihood ratio indicates that the decoder tends to classify it as an erase state, a negative value indicates that it tends to classify it as a programmed state, and a larger absolute value indicates a higher confidence level.

[0126] For each quantization level and each possible read voltage condition, the initial log-likelihood ratio is calculated based on the distribution parameters of each quantization level stored in the noise distribution pattern, as follows: The input data is: Read the nominal center voltage and distribution width of the current quantization level, as well as the nominal center voltage and distribution width of the adjacent quantization levels, from the noise distribution pattern record.

[0127] The calculation rules and processing procedures are as follows: Determine the voltage window range by expanding the distribution width by a first predetermined multiple in both the positive and negative voltage directions, centered on the nominal center voltage value of the current quantization level (typically 3 times, covering approximately 99.7% of the threshold voltage distribution range), to obtain the upper and lower voltage values ​​of the voltage window used to calculate the initial log-likelihood ratio.

[0128] The voltage window is discretized by dividing the voltage window range from the lower limit to the upper limit by a preset voltage step size (e.g., 5mV) into a set of read voltage index points. Each index point corresponds to a specific read voltage value. The voltage indices are numbered sequentially starting from 0, and the total number of indices is equal to the width of the voltage window divided by the voltage step size plus 1.

[0129] For each discretized read voltage index point, the conditional probability value of the current quantization level and the conditional probability value of the adjacent quantization level are calculated. These conditional probability values ​​are not calculated analytically using a probability density function, but are based on the quantization level distribution parameters (nominal center voltage value and distribution width) already stored in the noise distribution pattern. They are equivalently described using a voltage deviation rule with the nominal center as a reference and the distribution width as a scale factor, as detailed below: The first voltage deviation is obtained by subtracting the nominal center voltage value of the current quantization level from the current read voltage value.

[0130] The first standardized deviation is obtained by taking the ratio of the absolute value of the first voltage deviation to the distribution width of the current quantization level. This value represents how many distribution width units are equivalent to the distance of the current read voltage from the nominal center of the current quantization level.

[0131] The current quantization level conditional probability value is determined based on the first standardized deviation and the preset mapping rule from the standardized deviation to the conditional probability value.

[0132] The mapping rule is as follows: when the standardized deviation is 0 (i.e., the read voltage is exactly at the nominal center), the conditional probability value is at its maximum value; as the standardized deviation increases, the conditional probability value monotonically decreases according to the decay trend of a bell curve; when the standardized deviation reaches a first predetermined multiple (3 times), the conditional probability value decays to less than about three-thousandths of the maximum value. In actual deployment, the above mapping from the standardized deviation to the conditional probability value is pre-calculated as a one-dimensional reference table. The one-dimensional reference table is indexed by the standardized deviation (from 0.00 to the first predetermined multiple, with a step size of 0.01), and each entry stores the corresponding relative value of the conditional probability, which is normalized with the maximum value at the nominal center being 1.

[0133] In the same way, the absolute value of the difference between the current reading voltage value and the nominal center voltage value of the adjacent quantization level is taken and then divided by the distribution spread of the adjacent quantization level to obtain the second standardized deviation. Based on the same mapping reference table from the standardized deviation to the conditional probability value, the conditional probability value of the adjacent quantization level is obtained.

[0134] The likelihood ratio is obtained by taking the ratio of the current quantization level conditional probability value to the conditional probability value of the adjacent quantization level; the natural logarithm of the likelihood ratio is then taken to obtain the initial log-likelihood ratio of the current quantization level relative to its adjacent levels at the current read voltage index point.

[0135] The above calculations are performed once for each quantization level (a total of 4 levels) and each discretization read voltage index point, and all calculation results are stored in the form of a lookup table. The specific structure and content of the lookup table are as follows: The lookup table uses the quantization level number as its first dimension, which contains four entries corresponding to the erase state (number 0), the first programming state (number 1), the second programming state (number 2), and the third programming state (number 3). Each quantization level entry is itself a sub-table, which uses the discretized index of the read voltage value as its second dimension. The second dimension contains approximately 100 to 400 index entries, depending on the voltage window width and voltage step size settings.

[0136] Each cell in the sub-table stores an initial log-likelihood ratio value, which is a signed floating-point number. The positive or negative sign indicates whether the log-likelihood ratio favors the erased state or the programmed state, and the absolute value indicates the strength of the confidence level. The entire lookup table is stored in the firmware storage area of ​​the flash memory controller and is loaded into the internal static random access memory of the flash memory controller once during the device initialization phase for high-speed table lookup operations in subsequent steps.

[0137] Using a symmetrical reference as a reference, along the asymmetric offset direction, the initial log-likelihood ratio of each quantization level is asymmetrically corrected according to the asymmetric offset intensity to obtain the corrected log-likelihood ratio; the asymmetric offset direction and asymmetric offset intensity are used together as soft information asymmetric correction parameters.

[0138] For each extracted quantization level's initial log-likelihood ratio, the following asymmetric correction operation is performed sequentially: For each initial log-likelihood ratio record (corresponding to a specific quantization level and read voltage value), perform the following calculation steps: Take the current log-likelihood ratio record corresponding to the read voltage value, subtract the voltage value corresponding to the symmetrical reference, and obtain the voltage offset. The positive or negative sign of the voltage offset indicates which side of the symmetrical reference the current read voltage value is located on.

[0139] Determine whether the sign of the voltage offset is consistent with the asymmetric offset direction. If the voltage offset and the asymmetric offset direction point to the same side (i.e., both positive or both negative), then the current log-likelihood ratio record is located on the side that needs enhanced compensation; otherwise, it is located on the opposite side.

[0140] For the log-likelihood ratio record located on the side requiring enhanced compensation, its absolute value is multiplied by (1 and the sum of the offset intensity) to obtain the absolute value of the corrected log-likelihood ratio, i.e., the magnitude of the log-likelihood ratio is amplified. The offset intensity is the dimensionless scaling factor of the output.

[0141] For the log-likelihood ratio record located on the opposite side, its absolute value is multiplied by (1 minus the offset intensity) to obtain the absolute value of the corrected log-likelihood ratio, that is, the magnitude of the log-likelihood ratio is reduced.

[0142] Keeping the original sign of the log-likelihood ratio unchanged, we combine the corrected absolute value with the unchanged sign to obtain the corrected log-likelihood ratio.

[0143] The effects of the above corrections are as follows: on the voltage side with stronger noise, the absolute value of the log-likelihood ratio is amplified, making the decoder's determination of the storage state on that side clearer (signal-to-noise ratio perception enhanced); on the voltage side with weaker noise, the absolute value of the log-likelihood ratio is reduced, avoiding excessive confidence from misleading the decoder's convergence.

[0144] The asymmetric offset direction and asymmetric offset intensity are packaged together as the final output of step 3, namely the soft information asymmetric correction parameter. This parameter will be passed to step 4 to perform asymmetric correction on the initial soft information sequence extracted in real time during the actual decoding process.

[0145] Step 4: Using the target parity check matrix and soft information asymmetric correction parameters as input, perform low-density parity check iterative decoding on the read data of the target word line to obtain the error-corrected output data.

[0146] Perform a read operation on the target word line to extract the raw decision voltage values ​​of each memory cell.

[0147] A read command is sent to the flash memory chip containing the target word line. The read circuit inside the flash memory chip sequentially applies a series of increasing read reference voltages to each memory cell on the target word line. In this embodiment, for a four-level cell flash memory, three read reference voltages are used to divide the entire voltage range into four quantization intervals, and the conduction state of each memory cell is detected. When the threshold voltage of the memory cell is lower than the applied read reference voltage, the memory cell is turned on; otherwise, the memory cell is turned off.

[0148] Based on the on / off response of the memory cell at each read reference voltage, the quantization interval to which the original decision voltage value of each memory cell belongs is determined, and the interval number is output as the original decision value. The original decision values ​​of all memory cells constitute a decision value sequence, and the sequence length is equal to the total number of memory cells contained in the target word line.

[0149] For each memory cell on the target word line, the flash memory controller calculates its initial log-likelihood ratio based on the quantization interval to which its original decision voltage value belongs and the noise distribution pattern corresponding to that memory cell. This calculation is performed by looking up a pre-established and stored initial log-likelihood ratio lookup table, rather than by performing real-time online calculation. The specific process is as follows: For each memory cell on the target word line, the current noise degradation level and noise skewness type are obtained from the output noise distribution pattern. Using these two parameters, the corresponding noise distribution pattern record is located in the noise distribution pattern template library. From this record, the nominal center voltage value and distribution width of each quantization level are read. Simultaneously, the storage base address of the initial log-likelihood ratio lookup table is determined. Different combinations of degradation levels and skewness types correspond to different lookup table instances, and each combination of lookup tables occupies a contiguous storage space in the firmware storage area.

[0150] For each memory cell, the quantization level number of the memory cell is determined based on the quantization interval into which the original decision voltage value of the memory cell falls. For a four-level cell flash memory, the value range of the quantization level number is 0, 1, 2, or 3.

[0151] Based on the original decision voltage value of the memory cell, find the discretized voltage index point that is closest to the original decision voltage value within the voltage window range corresponding to the quantization interval. The method for determining the discretized voltage index point is as follows: subtract the lower limit of the voltage window corresponding to the quantization level from the original decision voltage value to obtain the voltage offset; divide the voltage offset by the voltage step size (e.g., 5mV), and round the quotient to obtain the voltage index value of the memory cell at that quantization level.

[0152] Using the quantization level number as the first-dimensional index and the voltage index value as the second-dimensional index, a two-dimensional addressing is performed in the initial log-likelihood ratio lookup table instance corresponding to the noise degradation level and skewness type. First, the corresponding sub-table is located by the quantization level number, and then the specific cell in the sub-table is located by the voltage index value. The initial log-likelihood ratio value stored in that cell is then read.

[0153] The initial log-likelihood ratio read is written into the corresponding position of the initial soft information sequence according to the physical arrangement order of the storage cell on the target word line.

[0154] The hardware implementation of the above table lookup operation is as follows: the direct memory access engine inside the flash memory controller generates a two-dimensional table lookup address (quantization level number spliced ​​with voltage index value) sequentially for each memory cell using the memory cell number as a circular index. It then initiates a read request to the static random access memory through the internal bus of the flash memory controller, reads the corresponding initial log-likelihood ratio from the memory, and stores it in the initial soft information sequence buffer.

[0155] The initial log-likelihood ratios of all storage cells are arranged sequentially according to the physical arrangement of the storage cells on the target word line to form an initial soft information sequence. The initial soft information sequence is a one-dimensional real number sequence, and each element in the sequence is the initial log-likelihood ratio of the corresponding storage cell.

[0156] Using the output soft information asymmetric correction parameters, including the asymmetric offset direction and asymmetric offset intensity, the asymmetric correction operation is performed on each element in the generated initial soft information sequence according to the exact same correction rules as in the previous steps: For each element in the initial soft information sequence, where each element represents the initial log-likelihood ratio of a storage unit, the following judgment and correction are performed: The current confidence level is obtained by taking the absolute value of the current log-likelihood ratio elements.

[0157] Take the original decision voltage value of the current memory cell (extracted from the previous steps), subtract the voltage value corresponding to the symmetric reference, and obtain the voltage offset; determine whether the sign direction of the voltage offset is consistent with the asymmetric offset direction. If the voltage offset and the asymmetric offset direction point to the same side (i.e., both positive or both negative), then the memory cell is located on the side that needs enhanced compensation; otherwise, it is located on the opposite side.

[0158] If it is on the pointing side, calculate the product of the absolute value of the current log-likelihood ratio and (the sum of the absolute values ​​of 1 and the offset intensity). The result of this product, which has the same sign as the current log-likelihood ratio, is used as the corrected log-likelihood ratio.

[0159] If it is on the opposite side, calculate the product of the absolute value of the current log-likelihood ratio and (the difference between 1 and the absolute value of the offset intensity). The result of this product, which has the same sign as the current log-likelihood ratio, is used as the corrected log-likelihood ratio.

[0160] If the offset intensity is 0, that is, the noise skewness type is symmetric, no correction is performed, and the original log-likelihood ratio is directly retained.

[0161] The corrected log-likelihood ratios are used to form a corrected soft information sequence, which is then used as the initial input to the low-density parity-check decoder.

[0162] The corrected soft information sequence is used as the initial soft information value of the variable nodes, and the selected target parity check matrix is ​​used as the parity constraint to initiate the low-density parity check iterative decoding process. The decoder contains two types of functional nodes, variable nodes and parity check nodes, as well as a bipartite graph connecting the two types of nodes. The variable nodes are nodes in the decoder's bipartite graph that correspond one-to-one with each encoded codeword bit, and their function is to summarize the soft information from the channel and the parity check nodes and make bit decisions.

[0163] The verification node refers to the node in the bipartite graph of the decoder that corresponds one-to-one with each row of the verification matrix. Its function is to calculate and feedback the degree of satisfaction of the verification constraints based on the current soft information of multiple variable nodes associated with that row.

[0164] Each iteration executes the following two sub-phases sequentially, as detailed below: During the verification node update phase, for each non-zero element in the target verification matrix, assume that this position connects the i-th verification node and the j-th variable node. Collect the current soft information values ​​of all other variable nodes connected to the i-th verification node, excluding the j-th variable node. For each collected soft information value, take its hyperbolic tangent function value. The hyperbolic tangent function is the difference between the base-e exponential function value and the base-e negative exponential function value of the input value, divided by the sum of the two.

[0165] Multiply all the hyperbolic tangent function values ​​together to obtain the first product. Take the inverse hyperbolic tangent function value from the first product; where the inverse hyperbolic tangent function is the ratio of the sum of 1 and the input value to the difference between 1 and the input value, take the natural logarithm of this ratio, and then divide by 2. Use this inverse hyperbolic tangent result as the verification message transmitted from the i-th verification node to the j-th variable node.

[0166] During the variable node update phase, for each non-zero element in the target parity check matrix, assume that this position connects the i-th parity check node and the j-th variable node. Collect the parity messages from all other parity check nodes connected to the j-th variable node (excluding the i-th parity check node), as well as the initial soft information of the channel corresponding to the j-th variable node (i.e., the corrected soft information value). Accumulate all the message values ​​to obtain the updated soft information for the j-th variable node. Make a hard decision based on the sign of the updated soft information: if the updated soft information is positive, determine that the codeword bit corresponding to the variable node is 0; if it is negative, determine that it is 1.

[0167] After each iteration, the hard decision result and the target parity check matrix are used to perform a check detection: the hard decision codeword vector and the transpose of the target parity check matrix are multiplied by matrix (in the binary field). If the product result is a zero vector (i.e. all check equations are satisfied), then the decoding is successful, and the decoding result of this iteration is the output data after error correction.

[0168] After each iteration, check the satisfaction status of the validation constraints and execute the following branch decision: If the hard decision codeword of the current iteration satisfies all the parity equations of the target parity matrix, that is, the parity result is a zero vector, then the iteration is terminated, and the information bit part of the current hard decision codeword after removing the parity bits is used as the error-corrected output data and returned to the host or upper-layer firmware.

[0169] If the hard decision codeword of the current iteration does not satisfy all the check equations, and the current iteration count has not yet reached the preset iteration limit (the classic value of the iteration limit in this embodiment is 15 to 25 times), then jump to the aforementioned sub-step to continue executing the next round of iteration.

[0170] If the current iteration count has reached the iteration limit and the verification equations are still not fully satisfied, the iteration is terminated. The hard decision codeword obtained from the last iteration (or the optimal codeword selected after post-processing) is output as the final decoding result. At the same time, the decoding failure status is reported to the upper-layer firmware, and the upper-layer firmware initiates a higher-level error recovery mechanism, such as read retry, read voltage offset adjustment, or independent disk redundant array verification recovery.

[0171] like Figure 2 As shown, embodiments of the present invention also provide an adaptive data encoding and decoding system guided by flash channel characteristics, comprising: The channel degradation identification module is used to acquire the channel state information of the target word line in the flash memory block, determine the channel degradation type based on the channel state information, and determine the corresponding noise distribution pattern based on the channel degradation type. The channel state information includes the physical layer position of the target word line in the memory block, the cumulative number of programming and erasing cycles, the duration of data not being refreshed, and the cumulative number of read interferences. The verification matrix selection module is used to adaptively select a target verification matrix that matches the current noise distribution pattern from a pre-built multi-level verification matrix set; wherein, different verification matrices in the multi-level verification matrix set are adapted to different noise degradation levels and noise skewness types. The asymmetric parameter calculation module is used to calculate soft information asymmetric correction parameters adapted to the current noise distribution pattern, based on the fixed elliptical focal spacing constraint and the hyperbolic asymptote constraint. The fixed elliptical focal spacing constraint is used to characterize the overall noise energy boundary of the channel, and the hyperbolic asymptote constraint is used to characterize the direction and degree of noise asymmetric offset. The soft information asymmetric correction parameters include the asymmetric offset direction and the asymmetric offset intensity. The low-density parity decoding module is used to perform low-density parity iterative decoding on the read data of the target word line, taking the target parity matrix and soft information asymmetric correction parameters as input, to obtain the error-corrected output data.

[0172] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0173] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A flash memory channel feature-guided adaptive data encoding and decoding method, characterized in that, The method includes: The channel state information of the target word line within the flash memory block is obtained. The channel state information includes the physical layer position of the target word line within the memory block, the cumulative number of program-erase cycles, the duration of data not being refreshed, and the cumulative number of read interferences. Based on the channel state information, the channel degradation type is determined. Based on the channel degradation type, the corresponding noise distribution pattern is determined. Based on the noise distribution pattern, a target verification matrix that matches the current noise distribution pattern is adaptively selected from a pre-constructed multi-level verification matrix set; wherein, different verification matrices in the multi-level verification matrix set are adapted to different noise degradation levels and noise skewness types. Based on the noise distribution pattern, soft information asymmetric correction parameters adapted to the current noise distribution pattern are calculated using a method based on fixed elliptical focal spacing constraints and hyperbolic asymptote constraints. Among them, the fixed elliptical focal spacing constraints are used to characterize the overall noise energy boundary of the channel, and the hyperbolic asymptote constraints are used to characterize the direction and degree of noise asymmetric offset. The soft information asymmetric correction parameters include the asymmetric offset direction and the asymmetric offset intensity. Using the target parity check matrix and soft information asymmetric correction parameters as input, low-density parity check iterative decoding is performed on the read data of the target word line to obtain the error-corrected output data.

2. The adaptive data encoding and decoding method guided by flash channel characteristics according to claim 1, characterized in that, Based on the noise distribution pattern, a target verification matrix that matches the current noise distribution pattern is adaptively selected from a pre-constructed multi-level verification matrix set, including: Based on the noise distribution pattern, determine the noise degradation level and noise skewness type of the channel where the target word line is located; Based on the noise degradation level, the matrix level corresponding to the noise degradation level is determined from a pre-constructed set of multi-level parity check matrices; wherein, the matrix level contains multiple candidate parity check matrices, and each candidate parity check matrix is ​​adapted to a different noise skewness type; Based on the noise skewness type, a suitable parity check matrix is ​​selected from the candidate parity check matrices contained in the matrix level as the target parity check matrix.

3. The adaptive data encoding and decoding method guided by flash channel characteristics according to claim 2, characterized in that, Based on the noise distribution pattern, soft information asymmetric correction parameters adapted to the current noise distribution pattern are calculated using a method based on the fixed constraint of elliptical focal spacing and hyperbolic asymptote constraint, including: Based on the noise degradation level, a fixed value for the focal distance of the ellipse is determined, and this fixed value is used as a fixed constraint for the focal distance of the ellipse; based on the noise distribution pattern, the positions of the two foci of the ellipse are determined. Based on the noise skewness type and the positions of the two foci of the ellipse, the direction and slope of the hyperbola asymptote are determined as hyperbola asymptote constraints; whereby the direction of the hyperbola asymptote characterizes the direction of the noise asymmetric shift, and the slope of the hyperbola asymptote characterizes the degree of the noise asymmetric shift. Using the two foci of the ellipse and the direction and slope of the hyperbola as constraints, soft information asymmetric correction parameters adapted to the current noise distribution pattern are calculated based on the noise distribution pattern. The soft information asymmetric correction parameters include the asymmetric offset direction and the asymmetric offset intensity.

4. The adaptive data encoding and decoding method guided by flash channel characteristics according to claim 3, characterized in that, Based on the noise degradation level, a fixed value for the elliptical focal distance is determined, and this fixed value is used as a fixed constraint for the elliptical focal distance. Based on the noise distribution pattern, determine the positions of the two foci of the ellipse, including: Based on the noise degradation level, the noise propagation range of the channel where the target word line is located is determined; where the higher the noise degradation level, the larger the noise propagation range; taking the noise propagation range as the geometric constraint corresponding to the fixed value of the elliptical focal spacing, the fixed value of the elliptical focal spacing is determined from the preset mapping relationship between degradation level and focal spacing; this value is used as the fixed constraint of the elliptical focal spacing. Based on the noise distribution pattern, extract the overall energy distribution center location of the channel noise; Using the center of the overall energy distribution as the center of symmetry and the fixed constraint of the focal spacing of the ellipse as the constraint condition, determine the positions of the two foci of the ellipse; wherein the two foci are symmetrical about the center of the overall energy distribution, and the distance between the two foci is equal to the fixed value of the focal spacing of the ellipse.

5. The adaptive data encoding and decoding method guided by flash channel characteristics according to claim 4, characterized in that, Based on the noise skewness type and the positions of the two foci of the ellipse, the direction and slope of the hyperbola's asymptotes are determined as constraints for the hyperbola's asymptotes, including: Based on the noise skewness type, determine the bias direction of the hyperbola asymptote; when the noise skewness type is left-biased, the bias direction points to the first focus of the two foci of the ellipse; when the noise skewness type is right-biased, the bias direction points to the second focus of the two foci of the ellipse; where the first focus is the focus located on the low-voltage side, corresponding to the direction of tail extension when charge loss causes the distribution to be biased to the left; the second focus is the focus located on the high-voltage side, corresponding to the direction of tail extension when read interference causes the distribution to be biased to the right. The basic direction of the asymptote of the hyperbola is determined by the positions of the two foci of the ellipse; the basic direction is consistent with the direction of the line connecting the two foci. Using the aforementioned basic direction as a reference, a bias direction is superimposed for correction to determine the direction of the hyperbola asymptote; wherein, this direction characterizes the direction of the noise asymmetric offset. Based on the noise distribution pattern, the noise skewness is extracted; the greater the noise skewness, the greater the absolute value of the slope of the hyperbola asymptote; based on the noise skewness and the preset mapping relationship between skewness and slope, the slope of the hyperbola asymptote is determined; this slope characterizes the degree of noise asymmetric shift.

6. The adaptive data encoding and decoding method guided by flash channel characteristics according to claim 5, characterized in that, Using the positions of the two foci of the ellipse and the direction and slope of the hyperbola as constraints, and based on the noise distribution pattern, calculate the soft information asymmetric correction parameters adapted to the current noise distribution pattern, including: Based on the positions of the two foci of the ellipse, the symmetry reference for soft information correction is determined; where the symmetry reference is the perpendicular bisector of the line connecting the two foci, and this symmetry reference corresponds to the zero offset reference line in the soft information domain. Based on the direction of the hyperbola asymptote, determine the asymmetric offset direction for soft information correction; whereby the asymmetric offset direction points to the side of the noise probability density distribution that needs to be enhanced and compensated. Based on the slope of the hyperbola asymptote and the preset mapping relationship between the slope and the offset intensity, the asymmetric offset intensity of the soft information correction is determined; where the larger the absolute value of the slope, the greater the asymmetric offset intensity. Based on the noise distribution pattern, the initial log-likelihood ratio of each quantization level is extracted; Using a symmetrical reference as a reference, along the asymmetric offset direction, the initial log-likelihood ratio of each quantization level is asymmetrically corrected according to the asymmetric offset intensity to obtain the corrected log-likelihood ratio; the asymmetric offset direction and asymmetric offset intensity are used together as soft information asymmetric correction parameters.

7. The adaptive data encoding and decoding method guided by flash channel characteristics according to claim 6, characterized in that, Using the target parity-check matrix and soft information asymmetric correction parameters as input, low-density parity-check iterative decoding is performed on the read data of the target word line to obtain the error-corrected output data, including: Perform a read operation on the target word line to extract the raw decision voltage value of each memory cell; Based on the original decision voltage value and channel state information, the initial log-likelihood ratio of each storage unit is calculated to form the initial soft information sequence. The initial soft information sequence is asymmetrically corrected by using the asymmetric offset direction and asymmetric offset intensity in the soft information asymmetric correction parameters, resulting in the corrected soft information sequence; wherein, the correction direction is determined by the asymmetric offset direction, and the correction magnitude is determined by the asymmetric offset intensity. Using the corrected soft information sequence as the initial input and the target parity matrix as the parity constraint, low-density parity check iterative decoding is performed; in each iteration, the parity node update and variable node update are alternately performed to obtain the decoding result of that iteration. If all verification constraints are satisfied, the current decoding result is used as the corrected output data; otherwise, the next iteration continues until the iteration limit is reached, and the final decoding result is output when the iteration limit is reached.

8. A flash memory channel feature-guided adaptive data encoding and decoding system, characterized in that, include: The channel degradation identification module is used to acquire the channel state information of the target word line in the flash memory block, determine the channel degradation type based on the channel state information, and determine the corresponding noise distribution pattern based on the channel degradation type. The channel state information includes the physical layer position of the target word line in the memory block, the cumulative number of programming and erasing cycles, the duration of data not being refreshed, and the cumulative number of read interferences. The verification matrix selection module is used to adaptively select a target verification matrix that matches the current noise distribution pattern from a pre-constructed multi-level verification matrix set; wherein, different verification matrices in the multi-level verification matrix set are adapted to different noise degradation levels and noise skewness types. The asymmetric parameter calculation module is used to calculate soft information asymmetric correction parameters adapted to the current noise distribution pattern, based on the fixed elliptical focal spacing constraint and the hyperbolic asymptote constraint. The fixed elliptical focal spacing constraint is used to characterize the overall noise energy boundary of the channel, and the hyperbolic asymptote constraint is used to characterize the direction and degree of noise asymmetric offset. The soft information asymmetric correction parameters include the asymmetric offset direction and the asymmetric offset intensity. The low-density parity decoding module is used to perform low-density parity iterative decoding on the read data of the target word line, taking the target parity matrix and soft information asymmetric correction parameters as input, to obtain the error-corrected output data.

Images