A method and system for predicting NAND flash memory data errors
Through the two-level channel model combining hard judgment and soft judgment information, the problem of insufficient prediction of NAND flash data error model in the prior art is solved, more accurate life prediction and ECC optimization are achieved, and the reliability of data storage is improved.
Patent Information
- Application Number
- CN202311380568.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-24
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2043-10-24
AI Technical Summary
The existing flash data error model has shortcomings in predicting NAND flash data errors, especially using the upper board test data, while the random data generated by the general model lacks the error characteristics of NAND flash memory, resulting in the difference between the prediction results and the actual results. farther.
A two-level data error model is adopted, where the first-level channel model is used to calculate the hard judgment error rate, and the second-level channel model is used to generate soft judgment information and adjust it according to the voltage distribution map to more accurately fit the error behavior of NAND flash memory. .
This method can more accurately predict the life of NAND flash memory, provide a more realistic error model, optimize ECC algorithm, improve ECC design, and improve data storage reliability and prediction accuracy.
Smart Images

Figure CN117130822B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of flash memory storage technology, and in particular to a method and system for predicting NAND flash memory data errors. Background Art
[0002] In recent years, flash storage has developed rapidly, and a variety of flash memory products have appeared in the market, including SLC (single-layer storage cell), MLC (double-layer storage cell), TLC (triple-layer storage cell) and QLC (quadruple-layer storage cell). The storage page capacity of each flash memory product is also different, including 1KB (1024 bytes), 2KB (2048 bytes), 4KB (4096 bytes) and 8KB (8192 bytes) and 16KB (16384 bytes) with higher efficiency, etc. Another major difference is that the capacity of the OOB (out of band) data area will be quite different. Generally speaking, the larger the page capacity, the larger the capacity of the OOB data area allocated to the page. Most of the capacity of the OOB data area will be used to store the check codeword of the ECC (Error Checking and Correction) algorithm in the storage controller.
[0003] Flash memory has the characteristic of storing data without volatile power failure. Its internal storage structure is metal layer-oxide layer-semiconductor-field effect transistor (MOSFET), and the FG (Floating Gate) inside is the actual data storage unit. Data is stored in the flash memory in the form of EC (Electrical Charge). The data in the storage unit is judged as 0 or 1 based on the amount of charge, and the amount of charge depends on the voltage applied to the CG (Control Gate), which controls the charge to be charged into the storage unit or the release of charge.
[0004] For some NAND flash memory products, such as SSD, SD, USB flash drives and other storage products, a single storage unit of the flash memory may not be charged or released as expected due to aging of the field effect tube, product use environment temperature, voltage instability, strong magnetic field and other reasons, resulting in errors in some fixed or random locations of the data stored on the storage product. In order to ensure the security of data stored in NAND flash memory, it is necessary to predict the life of NAND flash memory, and in the process of using NAND flash memory, it is also necessary to configure an adaptive ECC algorithm. All of these require a channel model that can truly and accurately reflect NAND flash memory data errors.
[0005] In the existing flash memory data error models, the data either comes from the data collected by NAND flash on-board testing, or from the random data generated by applying some general models (such as BSC, AWGNC channel model, etc.). However, both methods have some problems. Using the data from the on-board test often requires a lot of experiments and takes a long time to collect the required original data; and the method of generating random numbers using a general model can quickly obtain the data required for simulation, but it lacks some error characteristics of NAND flash itself. Such data error models are often far from the actual results of NAND flash. Summary of the invention
[0006] In order to overcome the deficiencies of the above technologies, the present invention provides a method and system for predicting NAND flash memory data errors.
[0007] The method and system for predicting NAND flash memory data errors provided by the present invention include a very practical two-level data error model that is very easy to implement on computers and FPGAs, wherein the first-level channel model determines the hard decision error rate of the data unit, and the second-level channel model generates the soft decision information provided by the read-adjustment. The method and system described in the present invention can not only predict the NAND flash memory error rate, making the life prediction of the NAND flash memory more consistent with the actual NAND flash memory, but also provide a NAND flash memory error model for the ECC scheme, thereby optimizing the ECC algorithm and improving the ECC design scheme.
[0008] Terminology explanation:
[0009] 1. BER: Bit Error Rate, bit error probability.
[0010] 2. BSC: Binary Symmetric Channel.
[0011] 3. AWGNC: Additive White Gaussian Noise Channel, additive white Gaussian noise channel.
[0012] 4. LLR: Loglikelihood Ratio, log-likelihood ratio.
[0013] 5. ECC: Error Checking and Correcting, error checking and correction.
[0014] 6. P / E: Program / Erase.
[0015] 7. PMF: probability mass function.
[0016] 8. CDF: Cumulative Distribution Function.
[0017] 9. RNG: Random Number Generator, random data generator.
[0018] The technical solution adopted by the present invention to overcome the technical problems is:
[0019] A method for predicting NAND flash memory data errors comprises the following steps:
[0020] S1. For each data unit, obtain the BER and voltage distribution information of each data unit through NAND flash memory on-board testing;
[0021] S2, using the BER obtained in step S1 as a training sample of the primary channel model, obtaining the primary channel model through training, and calculating the hard decision error rate of each data unit through the primary channel model;
[0022] S3, using the voltage distribution information in step S1 as a training sample of the secondary channel model, obtaining the secondary channel model through training, and using the hard decision error rate of each data unit obtained in step S2 as the hard decision read error rate of all positions in the data unit in the secondary channel model, and obtaining the soft decision information of each data unit through the secondary channel model calculation;
[0023] S4. According to the actual situation of the voltage distribution diagram, the secondary channel model is adjusted, and the information of the transformation between non-adjacent states in the actual voltage distribution is added to obtain soft decision information that meets the characteristics of the NAND flash memory.
[0024] Further, in step S1, the data unit is at least a block of a NAND flash memory or a page of a NAND flash memory or a codeword of an ECC algorithm.
[0025] Furthermore, step S1 specifically includes:
[0026] S1.1. Find several good blocks in the NAND flash memory of the object to be tested according to the bad block table of the manufacturer;
[0027] S1.2. Determine the test parameters, which at least include the P / E wear point and the data storage time point;
[0028] S1.3, according to the P / E wear point, determine the target P / E value of each good block, and perform artificial wear on all good blocks;
[0029] S1.4, data retention time test: use a temperature box to accelerate the data retention time test, and set the operating time and temperature of the temperature box according to the test parameters set in step S1.2;
[0030] S1.5. At each data storage time point, read the data in all good blocks, compare the write data and the read data, and obtain the BER of each data unit at each data storage time point; through multiple adjustment readings, then collect the data under each voltage level, and compare and calculate with the write data, obtain the voltage distribution information of each data unit at each data storage time point.
[0031] Furthermore, step S2 specifically includes:
[0032] S2.1, using the BER obtained in step S1 as a training sample of the primary channel model, and obtaining the primary channel model through training;
[0033] S2.2, converting the PMF obtained after modeling the primary channel model into an inverse CDF, the generated inverse CDF maps the uniformly distributed values in the interval [0, 1] into the corresponding error rate;
[0034] S2.3. Use RNG to generate uniformly distributed values in the interval [0, 1], and finally obtain the hard decision error rate of the data unit.
[0035] Furthermore, step S3 specifically includes:
[0036] S3.1, using the voltage distribution information in step S1 as a training sample of the secondary channel model, and obtaining the secondary channel model through training;
[0037] S3.2. Determine the number of reads in the soft decision according to the actual application, wherein the number of reads is less than or equal to the maximum voltage threshold adjustment number set by the NAND flash memory manufacturer;
[0038] S3.3, after the secondary channel modeling, the probability of dividing the voltage distribution map into several regions by multiple readings is obtained, the probability of each region of the soft decision is scaled according to the hard decision error rate of each data unit obtained in step S2, and the LLR value corresponding to each region is calculated;
[0039] S3.4. Use the hard decision error rate of each data unit in step S2 to obtain the flip information of each bit of the data unit, and then use RNG to generate a uniformly distributed value in the interval [0, 1]. According to the probability and LLR value of each area of the soft decision obtained in step S3.3, finally obtain the soft decision information of the data unit.
[0040] Furthermore, step S4 specifically includes:
[0041] S4.1, using the NAND flash memory board test data used in step S1, obtain a voltage distribution diagram of the storage unit;
[0042] S4.2, calculating the probability of transition between non-adjacent states in the voltage distribution diagram, and setting it as the probability of "retained storage unit";
[0043] S4.3. According to the probability of the "reserved storage unit", these bits in the data unit are forcibly assigned to the LLR value of the area most prone to error, that is, the maximum LLR value.
[0044] The present invention also discloses a system for predicting NAND flash memory data errors, comprising:
[0045] A NAND test calculation unit, for each data unit, is used to obtain BER and voltage distribution information of each data unit through NAND flash memory on-board testing;
[0046] A primary channel model is used to use the acquired BER as a training sample of the primary channel model, obtain the primary channel model through training, and calculate the hard decision error rate of each data unit through the primary channel model;
[0047] A secondary channel model is used to use the acquired voltage distribution information as a training sample of the secondary channel model, obtain the secondary channel model through training, and use the acquired hard decision error rate of each data unit as the hard decision read error rate of all positions in the data unit in the secondary channel model, and obtain the soft decision information of each data unit through the secondary channel model calculation;
[0048] The fine-tuning unit is used to adjust the secondary channel model according to the actual situation of the voltage distribution diagram, add the information of the transformation between non-adjacent states in the actual voltage distribution, and obtain soft decision information that meets the characteristics of the NAND flash memory.
[0049] The beneficial effects of the present invention are:
[0050] The present invention utilizes a two-level channel model to generate hard decision error rate and soft decision information respectively, which can better fit the actual error behavior of a real NAND flash memory. It not only counts the average error rate, but also takes into account the structure and error characteristics of the NAND flash memory, further subdivides the error rate statistics, and counts the error rate of each page, and even subdivides it to the error rate distribution of a codeword of the ECC algorithm. It can better simulate the real NAND flash memory errors and understand the error behavior of the NAND flash memory. It can not only more accurately predict the life of the NAND flash memory and provide life warning information, but also provide a NAND flash memory error model for the ECC solution, optimize the ECC algorithm, and then improve the ECC design solution. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 This is a flow chart of the method for predicting NAND flash memory data errors described in Embodiment 1 of the present invention.
[0052] Figure 2 It is a schematic diagram of the comparison between the model based on non-central chi-square distribution described in Example 1 of the present invention and NAND flash memory test data.
[0053] Figure 3 This is a schematic diagram of the measured voltage distribution of a certain TLC NAND flash memory with 3 bits per cell that follows the Gray code mapping of [111, 110, 100, 000, 010, 011, 001, 101] according to Embodiment 1 of the present invention.
[0054] Figure 4 This is a comparison diagram of LDPC algorithm simulations under different channel models described in Example 1 of the present invention. DETAILED DESCRIPTION
[0055] In order to facilitate those skilled in the art to better understand the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The following is only exemplary and does not limit the protection scope of the present invention.
[0056] Example 1
[0057] like Figure 1 As shown, a method for predicting NAND flash memory data errors according to the present invention comprises the following steps:
[0058] S1. For each data unit, obtain the BER and voltage distribution information of each data unit through NAND flash memory on-board testing;
[0059] S2, using the BER obtained in step S1 as a training sample of the primary channel model, obtaining the primary channel model through training, and calculating the hard decision error rate of each data unit through the primary channel model;
[0060] S3, using the voltage distribution information in step S1 as a training sample of the secondary channel model, obtaining the secondary channel model through training, and using the hard decision error rate of each data unit obtained in step S2 as the hard decision read error rate of all positions in the data unit in the secondary channel model, and obtaining the soft decision information of each data unit through the secondary channel model calculation;
[0061] S4. According to the actual situation of the voltage distribution diagram, the secondary channel model is adjusted, and the information of the transformation between non-adjacent states in the actual voltage distribution is added to obtain soft decision information that meets the characteristics of the NAND flash memory.
[0062] like Figure 1 As shown, the specific process of the method for predicting NAND flash memory data errors described in this embodiment is as follows:
[0063] This embodiment does not limit the type of flash memory particles, that is, SLC (single-layer storage cell), MLC (double-layer storage cell), TLC (triple-layer storage cell) and QLC (quadruple-layer storage cell) are all applicable; similarly, there is no requirement for the sample size of collected data. Generally speaking, the larger the sample size, the more accurate the result after channel fitting will be.
[0064] Step S1: For each data unit, obtain the BER and voltage distribution information of each data unit through NAND flash memory on-board testing.
[0065] Specifically, for each data unit, in this embodiment, the data unit can be a data block of any length, for example, a block of NAND flash memory, a page of NAND flash memory, or a codeword of the ECC algorithm, and of course it is not limited to the three listed. Among them, a codeword of the ECC algorithm is 1KB, 2KB or 4KB of data, a page of NAND flash memory is 4KB, 8KB or 16KB of data, and a block of NAND flash memory includes multiple pages, which may be different for different NAND flash memories. Then, for each data unit, the BER (Bit Error Rate, bit error probability) and voltage distribution information are obtained through the test data of the NAND flash memory.
[0066] In the embodiment of the present invention, a small amount of NAND flash memory test data is used, such as one block per P / E point. Specifically, the method steps for obtaining BER through the on-board test of NAND flash memory are as follows.
[0067] First, according to the bad block table provided by the manufacturer, several good blocks in the NAND flash memory of the object to be tested are found in the test disk.
[0068] Determine the test parameters. The test parameters described in this embodiment at least include P / E wear points and data retention time points.
[0069] According to the P / E wear point, the target P / E value of each good block is determined, and all good blocks are artificially worn to wear to the target P / E value. The artificial wear is to erase and write the block, and the number of wear times is the corresponding target P / E value. For example, if the target P / E value of a block is 1000, then the block needs to be manually erased 1000 times.
[0070] Data retention time test: Because the data retention time test generally takes a long time, it is necessary to accelerate the data retention time test. The accelerated test is generally carried out through a temperature chamber. After acceleration, the data retention time test time will be shortened, which can save a lot of time. According to the test parameters set above, the operation time and temperature of the temperature chamber are set. For example, the normal data retention time is generally 3 months, but in actual operation, the NAND flash memory chip will not be placed for 3 months before reading the data. Instead, it is accelerated by using a temperature chamber. It is assumed that it takes 3 months at 25°C (room temperature), and the Ea value is 1. It only takes 2 hours and 45 minutes to place it in an 80°C temperature chamber, and it only takes 24 hours to place it in a 60°C temperature chamber. The corresponding temperature is set according to the actual situation to achieve the required time.
[0071] In this embodiment, the calculation formula of the acceleration factor used in the acceleration experiment is:
[0072] (1)
[0073] In formula (1), is the acceleration factor; It is the activation energy. The activation energy of different products is different and is provided by the flash memory manufacturer. is the Boltzmann constant, whose value is 8.617385*10 -5 ; It is the temperature value under non-accelerated condition under use conditions. The temperature value here is absolute temperature value. (Kelvin) is the unit; It is the temperature value under accelerated test conditions. The temperature value here is the absolute temperature value. (Kelvin) as the unit.
[0074] S2. Use the BER obtained in step S1 as a training sample of the primary channel model, obtain the primary channel model through training, and calculate the hard decision error rate of each data unit through the primary channel model.
[0075] Specifically, step S2 includes the following:
[0076] The BER obtained in step S1 is used as a training sample of the primary channel model. After training, the parameters closest to the training sample are obtained, and then the primary channel model is obtained.
[0077] The PMF (probability mass function) obtained after modeling the primary channel model is converted into an inverse CDF (cumulative distribution function). The generated inverse CDF maps the uniformly distributed values in the interval [0, 1] to the corresponding error rate.
[0078] In this embodiment, the inverse CDF function is in the following form:
[0079] (2)
[0080] In formula (2), The number of bit errors is The probability of a data unit, Indicates the maximum number of error bits in a data unit. Indicates that the number of bit errors is greater than The probability of a data unit.
[0081] In an embodiment of the present invention, the primary channel model uses a model based on non-central chi-square distribution. Figure 2 The comparison between the model based on the noncentral chi-square distribution and the NAND flash memory test data is shown in Figure 2 In the figure, the curve "chModel" represents the channel model of the present invention, and "NAND measured" represents the test result of NAND flash memory on the board. Figure 2 It can be seen that the model based on the non-central chi-square distribution is very consistent with the measurement data of this embodiment, so the model based on the non-central chi-square distribution is very suitable for the measurement data of this embodiment.
[0082] The PMF (probability mass function) used by the primary channel model has the following form:
[0083] (3)
[0084] In formula (3), , represents the degree of freedom parameter, represents the noncentrality parameter; Indicates the measured value of PMF or the value after smoothing. is a scaling factor. Represents the scale parameter, and its value range is 0~1. is a constant with a value range of .
[0085] Then, a RNG (random number generator) is used to generate uniformly distributed values in the interval [0, 1], and finally the hard decision error rate of the data unit is obtained.
[0086] The test data of each P / E wear point + data retention time is trained and modeled, and finally a quantized inverse CDF (cumulative distribution function) lookup table is generated, so that the uniform distribution can be converted into error probability when the channel model is used. The error distribution of the data generated by the first-level channel model of the present invention is about 69% consistent with the NAND flash memory test results compared with the BSC channel. As shown in Table 1, the error distribution of the data generated by the first-level channel model of the present invention and the BSC channel model is compared. It can be seen that the average error deviation of the data generated by the first-level channel model of the present invention is 6.9%, while the average error deviation of the data generated by the BSC channel model is 75.9%, a difference of 69%.
[0087] Table 1 Comparison of error distribution of the primary channel model in the present invention and the data generated by the BSC channel
[0088]
[0089] Step S3, using the voltage distribution information in step S1 as a training sample of the secondary channel model, obtaining the secondary channel model through training, and using the hard decision error rate of each data unit obtained in step S2 as the hard decision reading error rate of all positions in the data unit in the secondary channel model, and obtaining the soft decision information of each data unit through the secondary channel model calculation.
[0090] Specifically, this embodiment uses the soft decision information of LDPC to illustrate. For the soft decision information of LDPC with n times of adjustment and multiple readings, the BER of each data unit will be linked to n+1 thresholds, and these thresholds will divide the [0, 1] interval into n+2 regions, which are also defined by the secondary channel model. Specifically, the soft decision information of LDPC needs to be read several times to obtain. Considering the entire range of voltage distribution of storage cells with a given state, the hard decision reading divides it into two regions, and the other multiple adjustment readings further divide these two regions into several regions. The probability of each region depends only on the integral or summation of the PDF (probability density function) or PMF (probability mass function) in the region. The number of reads here can be set according to actual needs, and the range is the maximum voltage threshold adjustment number set by the NAND flash memory manufacturer. Taking the soft decision information of LDPC read by two times as an example, the soft decision information required by LDPC not only includes the probabilities corresponding to the four regions divided by the voltage distribution curve by two times, but also includes the LLR values corresponding to each region. Once the probability of each region is determined, the corresponding LLR values can be calculated. Here, the probability of each region can be determined by the voltage distribution curve information of the upper board test data at the given P / E wear point + data storage time point. For different voltage threshold offsets, the probability of the corresponding region can be determined, and then the probability is scaled according to the hard decision error rate determined by the primary channel model. Assuming that the hard decision error rate of the primary channel model is p, the probability of each region determined by multiple reads of the secondary channel model, if it is judged that the current bit in the data block needs to be flipped, then the probability of [0, 1] is reduced to [0, p], if it is judged that the current bit in the data block does not need to be flipped, then the probability of [0, 1] is reduced to [0, 1-p]. Here, it is assumed that all positions in each data unit have the same hard decision read error rate, and the size of this value is determined by the primary channel model. A RNG (random number generator) is used to generate uniformly distributed values in the interval [0, 1], and finally the soft decision information of the data unit is obtained.
[0091] Step S4: According to the actual situation of the voltage distribution diagram, the secondary channel model is adjusted, and the information of the transformation between non-adjacent states in the actual voltage distribution is added to obtain soft decision information that meets the characteristics of the NAND flash memory.
[0092] Since the method of the present invention extracts a simple secondary channel model from the voltage distribution of a complex storage unit, it may not be possible to obtain the transformation information between non-adjacent states of the voltage distribution, such as Figure 3In the example of the measured voltage distribution diagram of a certain TLC NAND flash memory with 3 bits per cell, which follows the Gray code mapping of [111, 110, 100, 000, 010, 011, 001, 101], the voltage of some storage cells drops, and the state "111" is changed to "101". Such irregular information cannot be simulated by the secondary channel model described in this embodiment. The behavior of such irregular information can be simulated by setting some "reserved storage cells". First, a range of probability or the number of "reserved storage cells" is selected, and then some bit errors are forcibly converted into "strongest" errors. Let's use the soft decision information of LDPC to illustrate. For the soft decision information of LDPC, it is forced to enter the error area of the maximum LLR magnitude. The number of "retained storage cells" can be determined by studying the voltage distribution of the storage cells. The specific method is: using the test data of the NAND flash memory, the voltage distribution diagram of the chip can be obtained, and the number of storage cells that change between two non-adjacent states can be counted, for example, the proportion of the number of state "111" changing to state "101" in the entire data. In the secondary channel model, the number of "retained error storage cells" can be determined based on this proportion, and this information can be reflected in the secondary channel model, making it more consistent with the actual NAND flash memory errors. Figure 4 As shown in FIG. 1 , using the two-stage channel model of the present invention, the results of LDPC algorithm simulation are more consistent with the actual measured results of NAND flash memory than those of BSC channel. Figure 4 In the figure, the vertical axis "SFR" stands for Sector Failure Rate, which indicates the proportion of data units that are not corrected by LDPC; the horizontal axis "BER" stands for Bit Error Rate, which indicates the bit error probability; the curve "LDPC,chModel" represents the channel model in the present invention; the curve "LDPC,BSC" represents the BSC channel model; the curve "LDPC,NAND measured" represents the NAND flash memory on-board test result.
[0093] The key point of the present invention is that a small amount of NAND flash memory test data can be used to fit the data error model of the two-level model, and a data error model that is more in line with the actual error behavior of the NAND flash memory is generated. The present invention uses a two-level channel model to generate hard decision error rate and soft decision information respectively, which can better fit the actual error behavior of the real NAND flash memory. It not only counts the average error rate, but also takes into account the structure and error characteristics of the NAND flash memory, and further subdivides the statistics of the error rate to count the error rate of each page, and even subdivides it to the error rate distribution of a codeword of the ECC algorithm. It can better simulate the real NAND flash memory errors and understand the error behavior of the NAND flash memory. It can not only more accurately predict the life of the NAND flash memory, but also provide life warning information, and provide a NAND flash memory error model for the ECC solution, optimize the ECC algorithm, and then improve the design of the ECC.
[0094] Example 2
[0095] This embodiment also discloses a system for predicting NAND flash memory data errors, including:
[0096] A NAND test calculation unit, for each data unit, is used to obtain BER and voltage distribution information of each data unit through NAND flash memory on-board testing;
[0097] A primary channel model is used to use the acquired BER as a training sample of the primary channel model, obtain the primary channel model through training, and calculate the hard decision error rate of each data unit through the primary channel model;
[0098] A secondary channel model is used to use the acquired voltage distribution information as a training sample of the secondary channel model, obtain the secondary channel model through training, and use the acquired hard decision error rate of each data unit as the hard decision read error rate of all positions in the data unit in the secondary channel model, and obtain the soft decision information of each data unit through the secondary channel model calculation;
[0099] The fine-tuning unit is used to adjust the secondary channel model according to the actual situation of the voltage distribution diagram, add the information of the transformation between non-adjacent states in the actual voltage distribution, and obtain soft decision information that meets the characteristics of the NAND flash memory.
[0100] The implementation process of the functions and effects of each unit or model in the above system is specifically detailed in the implementation process of the corresponding steps in the above method, which will not be repeated here. For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can be referred to the partial description of the method embodiment.
[0101] The above only describes the basic principles and preferred embodiments of the present invention. Those skilled in the art may make many changes and improvements based on the above description, and these changes and improvements should fall within the protection scope of the present invention.
Claims
1. A method for predicting NAND flash memory data errors, characterized in that: The steps include: S1. For each data unit, obtain the BER and voltage distribution information of each data unit through NAND flash memory on-board testing; S2, using the BER obtained in step S1 as a training sample of the primary channel model, obtaining the primary channel model through training, and calculating the hard decision error rate of each data unit through the primary channel model; the primary channel model uses a model based on non-central chi-square distribution, and converts the PMF obtained after the primary channel model is modeled into an inverse CDF, and the generated inverse CDF maps the uniformly distributed value in the interval [0, 1] to the corresponding error rate; The inverse CDF function is of the following form: In formula (2), p(i) represents the probability of a data unit with bit error number i, N represents the maximum number of error bits of a data unit, and F(x) represents the probability of a data unit with bit error number greater than x; The PMF used by the primary channel model has the following form: In formula (3), f(·)~χ 2 (d, v), d represents the degree of freedom parameter, v represents the non-centrality parameter; p0(x) represents the measured value of PMF or the value after smoothing; is a proportional factor; s represents the scale parameter, ranging from 0 to 1; X is a constant, ranging from 0 <X<N; Then, a random number generator RNG is used to generate uniformly distributed values in the interval [0, 1], and finally the hard decision error rate of the data unit is obtained; S3, using the voltage distribution information in step S1 as a training sample of the secondary channel model, obtaining the secondary channel model through training, and using the hard decision error rate of each data unit obtained in step S2 as the hard decision read error rate of all positions in the data unit in the secondary channel model, and obtaining the soft decision information of each data unit through the secondary channel model calculation; step S3 specifically includes: S3.1, using the voltage distribution information in step S1 as a training sample of the secondary channel model, and obtaining the secondary channel model through training; S3.
2. Determine the number of reads in the soft decision according to the actual application, wherein the number of reads is less than or equal to the maximum voltage threshold adjustment number set by the NAND flash memory manufacturer; S3.3, after the secondary channel modeling, the probability of dividing the voltage distribution map into several regions by multiple readings is obtained, the probability of each region of the soft decision is scaled according to the hard decision error rate of each data unit obtained in step S2, and the LLR value corresponding to each region is calculated; S3.4, using the hard decision error rate of each data unit in step S2 to obtain the flip information of each bit of the data unit, and then using the random number generator RNG to generate a uniformly distributed value in the interval [0, 1], according to the probability and LLR value of each area of the soft decision obtained in step S3.3, finally obtaining the soft decision information of the data unit; S4, according to the actual situation of the voltage distribution diagram, the secondary channel model is adjusted, and the information of the transformation between non-adjacent states in the actual voltage distribution is added to obtain soft decision information that conforms to the characteristics of the NAND flash memory; step S4 specifically includes: S4.1, using the NAND flash memory board test data used in step S1, obtain a voltage distribution diagram of the storage unit; S4.2, calculating the probability of transition between non-adjacent states in the voltage distribution diagram, and setting it as the probability of "retained storage unit"; S4.
3. According to the probability of the "reserved storage unit", these bits in the data unit are forcibly assigned to the LLR value of the area most prone to error, that is, the maximum LLR value.
2. The method for predicting NAND flash memory data errors according to claim 1, characterized in that: In step S1, the data unit is at least a block of a NAND flash memory or a page of a NAND flash memory or a codeword of an ECC algorithm.
3. The method for predicting NAND flash memory data errors according to claim 1, characterized in that: Step S1 specifically includes: S1.
1. Find several good blocks in the NAND flash memory of the object to be tested according to the bad block table provided by the manufacturer; S1.
2. Determine the test parameters, which at least include the P / E wear point and the data storage time point; S1.3, according to the P / E wear point, determine the target P / E value of each good block, and perform artificial wear on all good blocks; S1.4, data retention time test: use a temperature box to accelerate the data retention time test, and set the operating time and temperature of the temperature box according to the test parameters set in step S1.2; S1.
5. At each data storage time point, read the data in all good blocks, compare the write data and the read data, and obtain the BER of each data unit at each data storage time point; through multiple adjustment readings, then collect the data under each voltage level, and compare and calculate with the write data, obtain the voltage distribution information of each data unit at each data storage time point.
4. The method for predicting NAND flash memory data errors according to claim 1, characterized in that: Step S2 specifically includes: S2.1, using the BER obtained in step S1 as a training sample of the primary channel model, and obtaining the primary channel model through training; S2.2, converting the PMF obtained after modeling the primary channel model into an inverse CDF, the generated inverse CDF maps the uniformly distributed values in the interval [0, 1] into the corresponding error rate; S2.
3. Use a random number generator RNG to generate uniformly distributed values in the interval [0, 1], and finally obtain the hard decision error rate of the data unit.
5. A system for predicting NAND flash memory data errors, characterized in that: include: A NAND test calculation unit, for each data unit, is used to obtain BER and voltage distribution information of each data unit through NAND flash memory on-board testing; The primary channel model is used to use the acquired BER as a training sample of the primary channel model, obtain the primary channel model through training, and calculate the hard decision error rate of each data unit through the primary channel model; the primary channel model uses a model based on non-central chi-square distribution, and the PMF obtained after the primary channel model is modeled is converted into an inverse CDF, and the generated inverse CDF maps the uniformly distributed value in the interval [0, 1] to the corresponding error rate; The inverse CDF function is of the following form: In formula (2), p(i) represents the probability of a data unit with bit error number i, N represents the maximum number of error bits of a data unit, and F(x) represents the probability of a data unit with bit error number greater than x; The PMF used by the primary channel model has the following form: In formula (3), f(·)~χ 2 (d, v), d represents the degree of freedom parameter, v represents the non-centrality parameter; p0(x) represents the measured value of PMF or the value after smoothing; is a proportional factor; s represents the scale parameter, ranging from 0 to 1; X is a constant, ranging from 0 <X<N; Then, a random number generator RNG is used to generate uniformly distributed values in the interval [0, 1], and finally the hard decision error rate of the data unit is obtained; A secondary channel model is used to use the acquired voltage distribution information as a training sample of the secondary channel model, obtain the secondary channel model through training, and use the acquired hard decision error rate of each data unit as the hard decision read error rate of all positions in the data unit in the secondary channel model, and obtain the soft decision information of each data unit through the secondary channel model calculation; Specifically: The voltage distribution information is used as a training sample of the secondary channel model, and the secondary channel model is obtained through training; Determine the number of reads in the soft decision according to the actual application, wherein the number of reads is less than or equal to the maximum voltage threshold adjustment number set by the NAND flash memory manufacturer; After the secondary channel modeling, the probability of dividing the voltage distribution map into several regions by multiple readings is obtained, the probability of each region of the soft decision is scaled according to the hard decision error rate of each data unit obtained, and the LLR value corresponding to each region is calculated; The hard decision error rate of each data unit is used to obtain the flip information of each bit of the data unit, and then a random number generator RNG is used to generate a uniformly distributed value in the interval [0, 1]. According to the probability and LLR value of each area of the soft decision, the soft decision information of the data unit is finally obtained; The fine-tuning unit is used to adjust the secondary channel model according to the actual situation of the voltage distribution diagram, add the information of the transformation between non-adjacent states in the actual voltage distribution, and obtain soft decision information that conforms to the characteristics of the NAND flash memory; specifically: Using the NAND flash memory on-board test data, the voltage distribution diagram of the storage unit is obtained; Calculate the probability of transition between non-adjacent states in the voltage distribution diagram and set it as the probability of "retained storage cell"; According to the probability of the "reserved storage unit", these bits in the data unit are forced to be assigned to the LLR value of the area most prone to error, that is, the maximum LLR value.
