# US 2021/0334684 A1 · Method of building and operating decoding status and prediction system

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2021/0334684 A1
- Document ID: US20210334684A1
- Status: Published application
- Kind: A1
- Application number: 17/366,315
- Filed: 2021-07-02
- Published: 2021-10-28
- Earliest claimed: 2017-11-24
- Assignees: Solid State Storage Technology Corporation
- Applicants: Solid State Storage Technology Corporation
- Inventors: Yen-Chin LIAO; Ching-Hui HUANG; Shih-Jia ZENG; Hsie-Chia CHANG
- Web page: https://patentgrep.com/patent/US20210334684A1

## CPC

- G06N 5/04: Inference or reasoning models
- G06N 20/20: Ensemble learning
- G06F 12/0246: {in block erasable memory, e.g. flash memory}
- G06N 5/01: Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
- G06N 7/01: Probabilistic graphical models, e.g. probabilistic networks
- G06N 20/00: Machine learning
- G11C 16/28: using differential sensing or reference cells, e.g. dummy cells
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G11C: STATIC STORES (semiconductor memory devices H10B)

## 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.

## Claims

### Claim 1

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.

### Claim 2 (depends on 1)

The operating method as claimed in claim 1, 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.

### Claim 3 (depends on 1)

The operating method as claimed in claim 1, 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.

### Claim 4 (depends on 3)

The operating method as claimed in claim 3, 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.

### Claim 5 (depends on 4)

The operating method as claimed in claim 4, wherein a sign of the sum of log probability ratios obtained by the K prediction models represents the prediction result.

### Claim 6 (depends on 4)

The operating method as claimed in claim 4, 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.

### Claim 7 (depends on 4)

The operating method as claimed in claim 4, 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.
