METHOD OF BUILDING AND OPERATING DECODING STATUS AND PREDICTION SYSTEM

Info

Publication number: 20190164068
Type: Application
Filed: Feb 1, 2018
Publication Date: May 30, 2019
Inventors: Yen-Chin LIAO (Taipei City), Ching-Hui Huang (Taipei City), Shih-Jia Zeng (Taipei City), Hsie-Chia Chang (Taipei City)
Application Number: 15/885,907

Abstract

A method of building a decoding status prediction system is provided. Firstly, plural read records are collected during read cycles of a flash memory. Then, the plural read records are classified into read records with a first read result and read records with a second read result. Then, a first portion of the read records with the first read result are divided into K0 groups according to a clustering algorithm, and a second portion of the read records with the second read result are divided into K1 groups according to the clustering algorithm. Then, the read records of the K0 groups and the K1 groups are used to train prediction models. Consequently, K0×K1 prediction models are generated. Then, the prediction models are combined as a prediction database.

Description

Description

This application claims the benefit of People's Republic of China Patent Application No. 201711191037.7, filed Nov. 24, 2017, the subject matter of which is incorporated herein by reference.

FIELD OF THE INVENTION

The present invention relates to a data analyzing method, and more particularly to a building method and an operating method of a decoding status prediction system.

BACKGROUND OF THE INVENTION

As is well known, flash memories are widely used in various electronic devices. After the flash memory has been written and erased many times, the characteristics of the flash memory are gradually suffered from degradation. Under this circumstance, the read data are possibly erroneous during the decoding process.

If the read data of the flash memory are erroneous, a read retry operation has to be performed on the flash memory. While the read retry operation is performed, the read parameters (e.g., a read voltage) for the flash memory are changed. Due to the proper read parameters, the read data can be successfully decoded and the read data are accurately outputted.

However, even if the read parameters for the flash memory are changed many time during the read retry process, the read data are possibly unable to be successfully decoded. That is, since the read data are erroneous during the decoding process, the flash memory cannot output the corrected read data. Under this circumstance, a read failure problem occurs.

SUMMARY OF THE INVENTION

An embodiment of the present invention provides a method of building a decoding status prediction system. Firstly, plural read records are collected during read cycles of a flash memory. Then, the plural read records are classified into read records with a first read result and read records with a second read result. Then, a first portion of the read records with the first read result are divided into K0 groups according to a clustering algorithm, and a second portion of the read records with the second read result are divided into K1 groups according to the clustering algorithm. Then, the prediction models are trained with one to one combination of the read records of the K0 groups and the K1 groups. Consequently, K0×K1 prediction models are generated. Then, these prediction models are collected as a prediction database.

Another embodiment of the present invention provides an operating method of a prediction system for use in a flash memory. The prediction system includes a prediction database. The prediction database includes K0×K1 prediction models that are obtained by training with one to one combination of data in K0 groups and K1 groups. Firstly, K0 central points of the K0 groups are calculated, and K1 central points of the K1 groups are calculated. Then, a reference point is received. Then, K0 distances between the reference point and the central points of the K0 groups are calculated, and K1 distances between the reference point and the K1 central points of the K1 groups are calculated. Then, K prediction models are selected from the prediction database according to the (K0+K1) distances. Then, the reference point is inputted into the K prediction models. Consequently, a prediction result is obtained.

Numerous objects, features and advantages of the present invention will be readily apparent upon a reading of the following detailed description of embodiments of the present invention when taken in conjunction with the accompanying drawings. However, the drawings employed herein are for the purpose of descriptions and should not be regarded as limiting.

BRIEF DESCRIPTION OF THE DRAWINGS

The above objects and advantages of the present invention will become more readily apparent to those ordinarily skilled in the art after reviewing the following detailed description and accompanying drawings, in which:

FIGS. 1A, 1B and 1C schematically illustrate the K0 group and the K1 group that are obtained according to the clustering algorithm;

FIG. 2 is a flowchart illustrating a method of building a decoding status prediction system according to an embodiment of the present invention;

FIGS. 3A and 3B illustrate a selecting process of the prediction model; and

FIG. 4 schematically illustrates plural test results of the prediction system.

DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

The present invention provides a method of building a decoding status prediction system. During read cycles of the flash memory, plural parameters for each read data and the read result are collected as read records. Then, the collected read records are divided into plural groups. Then, a machine learning technology is used to train and build a prediction model. Finally, all of the prediction models are collected as a prediction database, the prediction system of the present invention is built. The method of building the status prediction system will be described as follows.

When a host issues a command to read the data of the flash memory, plural parameters for reading each data and the read result are collected as a read record. For example, the plural parameters include a status parameter of the flash memory, an environment parameter, and so on.

For example, each read record contains three parameters. That is, each read data corresponds to a three-dimensional (x₁, x₂, x₃) read record. The first read record is y₁(x₁₁, x₂₁, x₃₁). The second read record is y₂(x₁₂, x₂₂, x₃₂). The m-th read record is y_m(x_1m, x_2m, x_3m). The rest may be deduced by analogy.

In the above read records, the values y₁˜y_mindicate the read results of the read data. For example, in case that y₁=0, the first data is successfully decoded. In case that y₂=1, the second data is not successfully decoded. In other words, among the m read records y₁˜y_m, those who have the read results “0” indicate that the corresponding data are successfully decoded, and the others, which read results are “1”, indicate that the corresponding data are not successfully decoded. If the data is successfully decoded, the read data can be acquired. Whereas, if the data is not successfully decoded, the read data cannot be acquired.

In the read record, x₁, x₂and x₃represent some parameters for reading the data. For example, x₁is a block erase count, x₂is a read voltage, and x3 is an environment temperature. If the first read record y₁(x₁₁, x₂₁, x₃₁)=y₁(50, 12.5, 35)=0, which means the first read data is successfully decoded under the condition that the block erase count is 50, the read voltage is 12.5V and the environment temperature is 35° C. If the second read record y₂(x₁₂, x₂₂, x₃₂)=y₂(105, 11.5, 45)=1, which means the second read data is not successfully decoded under the condition that the block erase count is 105, the read voltage is 11.5V and the environment temperature is 45° C. The rest may be deduced by analogy.

As mentioned above, the plural parameters for each read data and the read result are collected as the read record when the host issues the command to read the data from the flash memory. In each read record, the number of parameters is the dimension of the read record. It is noted that the dimension of the read record is not restricted. That is, if the read record has more parameters, the read record has a higher dimension. For example, we can have additional 3 parameters x₄, x₅, and x₆in the read record. The parameter x₄indicates the data program time, the parameter x₅indicates the address information, and the parameter x₆indicates the data retention time.

For clarification, a two-dimensional read record will be used to describe the method of building the decoding status prediction system. Of course, a read record with a higher dimension is suitably applied to the building method of the present invention.

Firstly, a software program collects plural read records when issuing the read recommends to the flash memory. For example, the software program collects m read records about the read cycle or the read command. Then, the m read records are classified into two categories according to the read results. For example, the m read records are classified into the read records with the read result “0” and the read records with the read result “1”.

Generally, the number of the read records with the read result “0” is much more than the number of the read records with the read result “1”. Then, a portion of the read records with the read result “0” are obtained according to a random under-sampling method. Consequently, the number of the read records with the read result “0” is equal to the number of the read records with the read result “1”.

Then, the entire or a portion of the read records with the read result “0” are divided into K0 groups according to a clustering algorithm, and the entire or a portion of the read records with the read result “1” are divided into K1 groups according to the clustering algorithm. For example, according to a k-means clustering algorithm, the read records with the read result “0” are divided into the K0 groups and the read records with the read result “1” are divided into the K1 groups.

FIGS. 1A, 1B and 10 schematically illustrate the K0 group and the K1 group that are obtained according to the clustering algorithm.

In the two dimensional (X₁-X₂) plane of FIG. 1A, each circular dot represents one read record with the read result “0”. For example, K0=3. According to the clustering algorithm, plural read records with the read result “0” are divided into three groups K0₁, K0₂and K0₃. Moreover, the central points of the three groups K0₁, K0₂and K0₃are calculated according to the read records of the corresponding groups.

In the two dimensional (X₁-X₂) plane of FIG. 1B, each square mark represents one read record with the read result “1”. For example, K1=3. According to the clustering algorithm, plural read records with the read result “1” are divided into three groups K1₁, K1₂and K1₃. Moreover, the central points of the three groups K1₁, K1₂and K1₃are calculated according to the read records of the corresponding groups.

Generally, after K0 and K1 are determined, the clustering algorithm is subsequently performed. After the clustering algorithm is completed, the K0 groups and the K1 groups are obtained. It is noted that the numbers of K0 and K1 are not restricted. That is, the numbers of K0 and K1 may be determined according to the practical requirements.

Please refer to FIG. 10. The central point of the group K0₁is at (a1, b1). The central point of the group K0₂is at (a2, b2). The central point of the group K0₃is at (a3, b3). The central point of the group K1₁is at (c1, d1). The central point of the group K1₂is at (c2, d2). The central point of the group K1₃is at (c3, d3).

After the (K0+K1) groups are determined according to the clustering algorithm, the machine learning technology is used to train the prediction models with the combination data from one of the three groups K0₁, K0₂and K0₃and one of the groups K1₁, K1₂and K1₃. Consequently, a prediction model is built. In other words, K0×K1 prediction models are generated according to the K0 groups and the K1 groups. For example, if K0 is 3 and K1 is 3, a total of 9 prediction models are generated.

Generally, after one prediction model is generated, the prediction model can be used to predict the read action of the flash memory. For example, after arbitrary parameters x1 and x2 are inputted into the prediction model, the probability (P0) of the read result “0” and the probability (P1) of the read result “1” are outputted from the prediction model. If the probability (P0) of the read result “0” is higher than the probability (P1) of the read result “1”, the prediction model predicts that the data would be successfully decoded when the flash memory is operated under the parameters x1 and x2. If the probability (P0) of the read result “0” is lower than the probability (P1) of the read result “1”, the prediction model predicts that the data would not be decoded successfully when the flash memory is operated under the parameters x1 and x2.

For example, after the entire or a portion of the read records in the group K0₁and the entire or a portion of the read records in the group K1₁are used to train a prediction model according to the machine learning technology, the prediction model M₁₁is generated. Similarly, after the entire or a portion of the read records in the group K0₁and the entire or a portion of the read records in the group K1₂are used to train a prediction model according to the machine learning technology, the prediction model M₁₂is generated. The rest may be deduced by analogy. In other words, K0×K1 prediction models are generated according to the K0 groups and the K1 groups. The K0×K1 prediction models are collected as a prediction database. Preferably, the prediction model is a binary classifier such as a random forest classifier.

FIG. 2 is a flowchart illustrating a method of building a decoding status prediction system according to an embodiment of the present invention.

Firstly, during a software program issues read cycles to the flash memory, plural parameters for each read data and the read result are collected as a read record. Consequently, the software program collects plural read records about the flash memory.

Then, a classifying process is performed. Consequently, the collected read records are classified into the read records with a first read result (i.e., the successfully-decoded read records) and the read records with a second read result (i.e., the unsuccessfully-decoded read records).

Then, a clustering process is performed. Consequently, the entire or a portion of the read records with the first read result are divided into K0 groups, and the entire or a portion of the read records with the second read result are divided into K1 groups.

Then, a training process is performed. After the read records in each of the K0 groups and the read records in each of the K1 groups are combined to train prediction models, the K0×K1 prediction models are generated. Moreover, the K0×K1 prediction models are collected as a prediction database. Consequently, the prediction system of the present invention is built.

After the prediction system is built, a predicting process can be performed. Since the prediction database contains plural prediction models, the prediction system can select K predetermined models from the prediction database according to the inputted reference point. Moreover, the prediction system predicts whether the reference point would be successfully decoded according to the selected K predetermined models. The operating principles will be described in more details as follows.

FIGS. 3A and 3B illustrate a selecting process of the prediction system. As shown in FIG. 3A, a reference point (e1, f1) is inputted into the prediction system. Then, the distances between the reference point (e1, f1) and the center points of the (K0+K1) groups are calculated. The distance between the reference point (e1, f1) and the center point (a1, b1) of the group K0₁is r1. The distance between the reference point (e1, f1) and the center point (a2, b2) of the group K0₂is r2. The distance between the reference point (e1, f1) and the center point (a3, b3) of the group K0₃is r3. The distance between the reference point (e1, f1) and the center point (c1, d1) of the group K1₁is r4. The distance between the reference point (e1, f1) and the center point (c2, d2) of the group K1₂is r5. The distance between the reference point (e1, f1) and the center point (c3, d3) of the group K1₃is r6.

In the example of FIG. 3A, r4<r1<r3<r2<r5<r6. In case that the number K of the predetermined models is 1, the group K1₁corresponding to the distance r4 and the group K0₁corresponding to the distance r1 are selected. The prediction model M₁₁, which is trained by using the read records in the group K1₁and the read records in the group K0₁, is selected. Then, the reference point (e1, f1) is predicted according to the prediction model M₁₁.

In case that the number K of the predetermined models is 2, there are two prediction models to be selected from the prediction database. The first one is the prediction model M₁₁, which is described in previous paragraph, the other one is predication model M₃₁, which is trained by using the read records in the group K1₁and the read records in the group K0₃. Then, the reference point (e1, f1) is predicted according to the prediction model M₁₁and the prediction model M₃₁.

From the above description, the prediction system calculates the distances between the reference point and the center points of all groups. The prediction models corresponding to the shortest distances have the higher priority to be selected from the prediction database as the K prediction models. The reference point is predicted according to the selected prediction models. In case that the number K of the predetermined models is larger than 2, the similar method is employed to select the prediction models.

Moreover, the sign (+/−) of the sum of the K log probability ratios is used to indicate the prediction result. That is, the log-likelihood ratio (LLR) may be expressed as:

$L L R = \sum_{K} \log \frac{P 0_{K}}{P 1_{K}}$

For example, the number K of the predetermined models is 1, and the prediction model M₁₁is used to predict the reference point (e1, f1). After the reference point (e1, f1) is inputted into the prediction model M₁₁, the probability (P0) of the read result “0” is 0.8 and the probability (P1) of the read result “1” is 0.2. Consequently, the log-likelihood ratio (LLR) may be calculated as:

$L L R = \sum_{K} \log \frac{P 0_{K}}{P 1_{K}} = \log \frac{0.8}{0.2} = + 0.602$

Since the value of LLR is positive, the prediction system predicts that the data would be successfully decoded when the flash memory is operated under the condition of the reference point (e1, f1). In other words, the read data is accurately acquired.

For example, the number K of the predetermined models is 2, and the prediction model M₁₁and M₃₁are used to predict the reference point (e1, f1). After the reference point (e1, f1) is inputted into the prediction model M₁₁, the probability (P0) of the read result “0” is 0.8 and the probability (P1) of the read result “1” is 0.2. After the reference point (e1, f1) is inputted into the prediction model M₃₁, the probability (P0) of the read result “0” is 0.72 and the probability (P1) of the read result “1” is 0.28. Consequently, the log-likelihood ratio (LLR) may be calculated as:

$L L R = \sum_{K} \log \frac{P 0_{K}}{P 1_{K}} = \log \frac{0.8}{0.2} + \log \frac{0.72}{0.28} = + 1.12$

Since the value of LLR is positive, the prediction system predicts that the data would be successfully decoded when the flash memory is operated under the condition of the reference point (e1, f1). In other words, the read data is accurately acquired.

As shown in FIG. 3B, a reference point (e2, f2) is inputted into the prediction model. Then, the prediction model calculates the distances between the reference point (e2, f2) and the center points of the (K0+K1) groups. The distance between the reference point (e2, f2) and the center point (a1, b1) of the group K0₁is r1. The distance between the reference point (e2, f2) and the center point (a2, b2) of the group K0₂is r2. The distance between the reference point (e2, f2) and the center point (a3, b3) of the group K0₃is r3. The distance between the reference point (e2, f2) and the center point (c1, d1) of the group K1₁is r4. The distance between the reference point (e2, f2) and the center point (c2, d2) of the group K1₂is r5. The distance between the reference point (e2, f2) and the center point (c3, d3) of the group K1₃is r6.

In the example of FIG. 3B, r5<r3<r6<r4<r2<r1. In case that the number K of the predetermined models is 1, the group K1₂corresponding to the distance r5 and the group K0₃corresponding to the distance r3 are selected. The prediction model M₃₂, which is trained by using the read records in the group K1₂and the read records in the group K0₃is selected. Then, the reference point (e2, f2) is predicted according to the prediction model M₃₂.

In case that the number K of the predetermined models is 2, there are two prediction models to be selected from the prediction database. The first one is the prediction model M₃₂, which is described in the previous paragraph, the other one is prediction model M₃₃, which is trained by using the read records in the group K1₃and the read records in the group K0₃. Then, the reference point (e2, f2) is predicted according to the prediction model M₃₂and the prediction model M₃₃.

For example, the number K of the predetermined models is 1, and the prediction model M₃₂is used to predict the reference point (e2, f2). After the reference point (e2, f2) is inputted into the prediction model M₃₂, the probability (P0) of the read result “0” is 0.42 and the probability (P1) of the read result “1” is 0.58. Consequently, the log-likelihood ratio (LLR) may be calculated as:

$L L R = \sum_{K} \log \frac{P 0_{K}}{P 1_{K}} = \log \frac{0.42}{0.58} = - 0.140$

Since the value of LLR is negative, the prediction system predicts that the data would not be successfully decoded when the flash memory is operated under the condition of the reference point (e2, f2). In other words, the read data cannot be accurately acquired.

For example, the number K of the predetermined models is 2, and the prediction model M₃₂and M₃₃are used to predict the reference point (e2, f2). After the reference point (e2, f2) is inputted into the prediction model M₃₂, the probability (P0) of the read result “0” is 0.42 and the probability (P1) of the read result “1” is 0.58. After the reference point (e2, f2) is inputted into the prediction model M₃₃, the probability (P0) of the read result “0” is 0.38 and the probability (P1) of the read result “1” is 0.62. Consequently, the log-likelihood ratio (LLR) may be calculated as:

$L L R = \sum_{K} \log \frac{P 0_{K}}{P 1_{K}} = \log \frac{0.42}{0.58} + \log \frac{0.38}{0.62} = - 0.353$

Since the value of LLR is negative, the prediction system predicts that the data would not be successfully decoded when the flash memory is operated under the condition of the reference point (e2, f2). In other words, the read data cannot be accurately acquired.

From the above descriptions, the preset invention provides a decoding status prediction system. The decoding status prediction system is used for predicting whether data in the flash memory can be successfully decoded. If the data would not be successfully decoded when the flash memory is operated under the condition of the reference point, the flash memory enters a read retry operation is performed on the flash memory. Consequently, the performance of the flash memory is enhanced.

Moreover, after m read records are collected and the above method is implemented, the prediction system of the present invention is built. The m read records may be collected when the flash memory is online operated by the user. Similarly, other n read records are collected as test read records. That is, the prediction system may be tested according to the test read records. The n read records with the known results are used to test the prediction results of the prediction models in the prediction database. Generally, the ratio of the value m to the value n is about 7:3. It is noted that the ratio of value m to the value n is not restricted. A method of testing the prediction system will be illustrated as follows.

FIG. 4 schematically illustrates plural test results of the prediction system. For example, the flash memory collects 100182 read records. Each read record contains 31 parameters and 1 read result. That is, each read record has 31 dimensions.

Firstly, 70% of the collected 100182 read records are randomly selected for building the prediction system. Moreover, the other 30% of the collected 100182 read records are used for testing the prediction system. That is, about 70127 read records (i.e., m=70127) are used for building the prediction system, and about 30055 read records (i.e., n=30055) are used for testing the prediction system.

In an embodiment, four prediction systems are built by using the prediction system building system of the present invention. In the first prediction system, the entire or a portion of the read records with the first read result are divided into 6 groups (i.e., K0=6) and the entire or a portion of the read records with the second read result are divided into 6 groups (i.e., K1=6). Consequently, the prediction database of the first prediction system contains 36 prediction models. In the second prediction system, the entire or a portion of the read records with the first read result are divided into 12 groups (i.e., K0=12) and the entire or a portion of the read records with the second read result are divided into 6 groups (i.e., K1=6). Consequently, the prediction database of the second prediction system contains 72 prediction models. In the third prediction system, the entire or a portion of the read records with the first read result are divided into 6 groups (i.e., K0=6) and the entire or a portion of the read records with the second read result are divided into 12 groups (i.e., K1=12). Consequently, the prediction database of the third prediction system contains 72 prediction models. In the fourth prediction system, the entire or a portion of the read records with the first read result are divided into 12 groups (i.e., K0=12) and the entire or a portion of the read records with the second read result are divided into 12 groups (i.e., K1=12). Consequently, the prediction database of the third prediction system contains 144 prediction models.

Take the first prediction system for example. In case that 1 predetermined model (K=1) is selected from the prediction database to predict all of the n (=30055) test read records, the first prediction system predicts that the accuracy of successfully decoding the data is about 0.840 (i.e., Label-0) and predicts that the accuracy of unsuccessfully decoding the data is about 0.865 (i.e., Label-1).

In case that 2 predetermined models (K=2) is selected from the prediction database to predict all of the n (=30055) test read records, the first prediction system predicts that the accuracy of successfully decoding the data is about 0.878 (i.e., Label-0) and predicts that the accuracy of unsuccessfully decoding the data is about 0.811 (i.e., Label-1).

In case that 3 predetermined models (K=3) is selected from the prediction database to predict all of the n (=30055) test read records, the first prediction system predicts that the accuracy of successfully decoding the data is about 0.892 (i.e., Label-0) and predicts that the accuracy of unsuccessfully decoding the data is about 0.813 (i.e., Label-1).

In case that 4 predetermined models (K=4) is selected from the prediction database to predict all of the n (=30055) test read records, the first prediction system predicts that the accuracy of successfully decoding the data is about 0.923 (i.e., Label-0) and predicts that the accuracy of unsuccessfully decoding the data is about 0.751 (i.e., Label-1).

Obviously, the first prediction system has higher predicting accuracy. Consequently, the first prediction system can be effectively applied to the flash memory to enhance the performance of the flash memory. Similarly, the second prediction system, the third prediction system and the fourth prediction system also have the higher predicting accuracy. Consequently, the second prediction system, the third prediction system and the fourth prediction system can be effectively applied to the flash memory to enhance the performance of the flash memory.

Moreover, while the flash memory is operated online, the newly-collected read records may be fed back to the flash memory for training the prediction system. For example, whenever p read records are newly collected, the p read records are fed back to the train the prediction system and then the prediction system is updated. For example, in case that the prediction system has (K0+K1) groups, the 1000 newly-collected read records (i.e., p=1000) are allocated into the (K0+K1) groups. After the newly-collected read records are allocated into the (K0+K1) groups, the center points of the allocated into the (K0+K1) groups are changed. After the training process, the prediction models are updated. Consequently, the predicting accuracy of the updated prediction system is enhanced.

While the invention has been described in terms of what is presently considered to be the most practical and preferred embodiments, it is to be understood that the invention needs not be limited to the disclosed embodiment. On the contrary, it is intended to cover various modifications and similar arrangements included within the spirit and scope of the appended claims which are to be accorded with the broadest interpretation so as to encompass all such modifications and similar structures.

Claims

1. A method of building a decoding status prediction system, the method comprising steps of:

collecting plural read records during read cycles of a flash memory;

classifying the plural read records into read records with a first read result and read records with a second read result;

dividing a first portion of the read records with the first read result into K0 groups according to a clustering algorithm

dividing a second portion of the read records with the second read result into K1 groups according to the clustering algorithm;

training the prediction models with the read records of the K0 groups and the K1 groups, so that K0×K1 prediction models are generated; and

collecting all the prediction models as a prediction database.

2. The method as claimed in claim 1, wherein each of the plural read records contains plural parameters and one of the first read result and the second read result.

3. The method as claimed in claim 1, wherein the plural parameters include at least one of a block erase count, a read voltage, an environment temperature, a data program time, an address information and a data retention time of the flash memory, wherein the first read result indicates a successful decoding result, and the second read result indicates a failed decoding result.

4. The method as claimed in claim 1, wherein the clustering algorithm is a k-means clustering algorithm.

5. The method as claimed in claim 1, further comprising a step of selecting a first group from the K0 groups and selecting a second group from the K1 groups, wherein after the read records in the first group and the second group are used to train a prediction model, a first prediction model of the plural prediction models is obtained.

6. The method as claimed in claim 5, wherein the first prediction model is a binary classifier.

7. The method as claimed in claim 6, wherein the binary classifier is a random forest classifier.

8. An operating method of a prediction system for use in a flash memory, the prediction system comprising a prediction database, the prediction database containing K0×K1 prediction models that are obtained by using records in K0 groups and K1 groups to train the prediction models, the operating method comprising steps of:

calculating K0 central points of the K0 groups, and calculating K1 central points of the K1 groups;

receiving a reference point;

calculating K0 distances between the reference point and the central points of the K0 groups, and calculating K1 distances between the reference point and the K1 central points of the K1 groups;

selecting K prediction models from the prediction database according to the (K0+K1) distances; and

inputting the reference point into the K prediction models, so that a prediction result is obtained.

9. The operating method as claimed in claim 8, wherein the prediction models corresponding to the distances between the reference point and the (K0+K1) groups, the shorter of the distances the higher of the priority to be selected from the prediction database as the K prediction models.

10. The operating method as claimed in claim 8, wherein after the reference point is inputted into a first prediction model of the K prediction models, a probability of a first read result and a probability of a second read result are obtained.

11. The operating method as claimed in claim 10, wherein after the probability of the first read result is divided by the probability of the second read result, a probability ratio of the first prediction model is obtained.

12. The operating method as claimed in claim 11, wherein a sign of the sum of log probability ratios obtained by the K prediction models represents the prediction result.

13. The operating method as claimed in claim 11, further comprising steps of:

collecting plural read records during read cycles of the flash memory; and

allocating the read records into the corresponding (K0+K1) groups, so that the central points of the K0 groups and the central points of the K1 groups are changed.

14. The operating method as claimed in claim 11, further comprising steps of:

collecting plural read records and corresponding read result during read cycles of the flash memory;

allocating the read records into the corresponding (K0+K1) groups; and

using the read records of the K0 groups and the K1 groups to train predication models, so that K0×K1 updated prediction models are generated.