CN-121983105-A - Storage medium multi-mode voltage prediction method based on CNN and related equipment

CN121983105ACN 121983105 ACN121983105 ACN 121983105ACN-121983105-A

Abstract

The invention relates to the technical field of computer storage, and discloses a storage medium multi-mode voltage prediction method and device based on CNN, a computer readable storage medium and SSD equipment. According to the method, the nonlinear coupling regression model reflecting the nonlinear coupling relation between the P/E times, the data retention time, the temperature, the read target position and the read voltage is established, the nonlinear relation between the P/E times, the data retention time, the temperature, the read target position and the read voltage and the interaction characteristics between the capture parameters are fitted, the CNN is utilized for fine mapping, the accuracy of voltage prediction is remarkably improved, the CNN can dynamically adjust the predicted voltage according to the real-time P/E times, the data retention time, the temperature and the read target position, the accurate voltage prediction under different environments and different use scenes can be realized, and the problems of low voltage prediction precision and poor dynamic adaptability in the traditional voltage prediction method are effectively solved.

Inventors

LU FUBO
Gou Rongsong
LIU HONG

Assignees

成都芯忆联信息技术有限公司

Dates

Publication Date: 20260505
Application Date: 20260122

Claims (10)

1. The storage medium multi-mode voltage prediction method based on CNN is characterized by comprising the following steps: acquiring a plurality of key parameters affecting the read voltage, wherein the key parameters comprise P/E times, data retention time, temperature and a read target position; establishing a nonlinear coupling regression model reflecting nonlinear coupling relations between the plurality of key parameters and the read voltage; Based on the nonlinear coupling regression model, carrying out standardization processing and interaction feature calculation on the plurality of key parameters, and carrying out feature coding on the standardized plurality of key parameters and interaction features together to generate a multi-channel feature tensor; inputting the multi-channel characteristic tensor into a pre-trained convolutional neural network, processing the multi-channel characteristic tensor by using the convolutional neural network, and outputting a predicted read voltage value.
2. The CNN-based storage medium multi-modal voltage prediction method according to claim 1, wherein the nonlinear coupled regression model is expressed by the following formula: Wherein, the Representing the effect of the P/E times N on the read voltage, Indicating the effect of the data retention time T on the read voltage at temperature T, Representing read target location With respect to the effect of the read voltage, Indicating the effect of other factors on the read voltage during use of the storage medium, The method accords with the distribution of the n-grams.
3. The CNN-based storage medium multi-modal voltage prediction method of claim 2, wherein the The calculation formula of (2) is as follows: ; Wherein, the Representing the dielectric oxide damage coefficient under N times of P/E abrasion; Representing the damage coefficient of the dielectric oxide layer affected by temperature; Representing the minimum ability of electrons to pass through the floating gate layer; Representing a capability scaling factor; t represents temperature.
4. The CNN-based storage medium multi-modal voltage prediction method of claim 2, wherein the The calculation formula of (2) is as follows: ; Wherein, the Representing the minimum ability of electrons to pass through the floating gate layer; Representing a capability scaling factor; Representing an electron leakage coefficient; indicating the dielectric oxide layer multiple damage constant.
5. The CNN-based storage medium multi-modal voltage prediction method of claim 2, wherein the The calculation formula of (2) is as follows: ; Wherein, the Representing the medium space influence coefficients, i=1, 2,3,4,5; A word line representing the read target location, Bit lines representing read target locations.
6. The CNN-based storage medium multi-modal voltage prediction method according to claim 1, wherein the performing normalization processing and interaction feature calculation on the plurality of key parameters based on the nonlinear coupling regression model, performing feature encoding on the normalized plurality of key parameters and the calculated interaction features together, and generating a multi-channel feature tensor specifically includes the following steps: Respectively carrying out standardization treatment on the P/E times N, the data retention time T and the temperature T based on the mean value and the variance to obtain standardized P/E times Normalized data retention time Temperature after normalization ; Mapping read target locations to two-dimensional coordinates And normalize word lines of the read target positions And a bit line for reading the target location Return to the interval of [0,1]; calculating interaction characteristics among parameters, wherein the interaction characteristics comprise 、、 And ; Will respectively 、、、、、、、 And Expanding the characteristic matrix into 8 multiplied by 8 to obtain 9 characteristic matrices; splicing the 9 feature matrixes into a multi-channel feature tensor 。
7. The CNN-based storage medium multi-modal voltage prediction method according to claim 1, wherein the convolutional neural network includes a first convolutional layer, a second convolutional layer, a global pooling layer, a third convolutional layer, a maximum pooling layer and a full connection layer which are sequentially connected, the convolutional neural network is used to process the multi-channel characteristic tensor, and a predicted read voltage value is output, and the method specifically includes the following steps: convolving the multi-channel feature tensor with a first convolution layer, as shown in the following equation: ; Wherein, the Representing the input multi-channel feature tensor, A first convolution layer is shown consisting of 16 3*3 convolutions plus a ReLu activation function, Representing an output result of the first convolution layer; And convolving the output result of the first convolution layer again through the second convolution layer, wherein the output result is shown as the following formula: ; Wherein, the A second convolution layer consisting of 32 3*3 convolutions plus a ReLu activation function is shown, Representing an output result of the second convolution layer; And carrying out global pooling on the output result of the second convolution layer through a global pooling layer, wherein the global pooling is shown as the following formula: ; Wherein, the Representing a global pooling function, Representing a global pooling result; and convolving the global pooling result again through a third convolution layer, wherein the convolution is as follows: ; Wherein, the A third convolution layer consisting of 64 3*3 convolutions plus a ReLu activation function is shown, Representing the output result of the third convolution layer; and carrying out maximum pooling on the output result of the third convolution layer through a maximum pooling layer, wherein the maximum pooling is shown as the following formula: ; Wherein, the Representing the maximum pooling function, Representing a maximum pooling result; and carrying out feature fusion on the global pooling result through the full connection layer, and outputting a final reading voltage value.
8. A CNN-based storage medium multi-modal voltage prediction apparatus, comprising: The data acquisition module is used for acquiring a plurality of key parameters affecting the read voltage, wherein the key parameters comprise P/E times, data retention time, temperature and a read target position; The model building module is used for building a nonlinear coupling regression model reflecting nonlinear coupling relations between the plurality of key parameters and the read voltage; The multi-mode feature coding module is used for carrying out standardization processing and interaction feature calculation on the plurality of key parameters based on the nonlinear coupling regression model, carrying out feature coding on the plurality of standardized key parameters and interaction features together, and generating a multi-channel feature tensor; and the read voltage prediction module is used for inputting the multi-channel characteristic tensor into a pre-trained convolutional neural network, processing the multi-channel characteristic tensor by using the convolutional neural network and outputting a predicted read voltage value.
9. A computer readable storage medium having a computer program stored therein, which when executed by a processor, implements the steps of the CNN-based storage medium multi-modal voltage prediction method according to any one of claims 1 to 7.
10. An SSD device comprising the computer readable storage medium of claim 9.

Description

Storage medium multi-mode voltage prediction method based on CNN and related equipment Technical Field The present invention relates to the field of computer storage technologies, and in particular, to a storage medium multi-mode voltage prediction method and apparatus based on CNN, a computer readable storage medium, and an SSD device. Background With the exponential increase in storage density, a four-layer memory cell (NAND QLC) has become a core medium of the storage market by virtue of its high capacity, low cost. However, the increase of the storage density (from TLC to QLC, i.e. from three-layer memory cells to four-layer memory cells) results in an increase of the charge states to 16 charge states, which is significantly narrower than the conventional SLC/TLC, resulting in the problems of overlapping threshold voltage distributions and voltage drift, and further resulting in the problem of degradation of the reliability of the QLC storage medium, and further aggravates the problems of overlapping threshold voltage distributions and voltage drift due to the effect of multi-mode physical coupling effects such as process, charge leakage, read disturbance, and lifetime. In order to effectively solve the problems of overlapping and drifting of threshold voltage distribution, accurate prediction of QLC read voltage is required under different use scenes and environments. However, the conventional voltage prediction method (e.g., a method based on a fixed threshold or linear compensation) has the following drawbacks: 1. the voltage multi-mode problem processing capability is insufficient, and the nonlinear coupling relation between a plurality of factors with larger influence on the read voltage, such as NAND service life, temperature, data retention time (Data Rentention) and the like and the read voltage is not considered, so that the voltage prediction complexity is low and the stability is poor; 2. The dynamic adaptation capability is lacking, the NAND ageing, the temperature change and other dynamic factors cannot be responded in real time by relying on offline calibration. Therefore, the traditional voltage prediction method has the problems of low prediction precision and poor dynamic adaptability, cannot meet the accurate and self-adaptive prediction requirements of the QLC read voltage, and is difficult to cope with the problems of overlapping and drifting of the QLC threshold voltage distribution. Disclosure of Invention The technical problem to be solved by the invention is to provide a storage medium multi-mode voltage prediction method and device based on CNN, a computer readable storage medium and SSD equipment, so as to solve the problems of low voltage prediction precision and poor dynamic adaptability of the traditional voltage prediction method. The first aspect provides a storage medium multi-mode voltage prediction method based on CNN, which comprises the following steps of obtaining a plurality of key parameters affecting read voltage, wherein the key parameters comprise P/E times, data retention time, temperature and a read target position, establishing a nonlinear coupling regression model reflecting nonlinear coupling relation between the key parameters and the read voltage, carrying out standardization processing and interaction feature calculation on the key parameters based on the nonlinear coupling regression model, carrying out feature coding on the standardized key parameters and interaction features together to generate a multi-channel feature tensor, inputting the multi-channel feature tensor into a convolutional neural network trained in advance, processing the multi-channel feature tensor by using the convolutional neural network, and outputting a predicted read voltage value. The storage medium multi-mode voltage prediction device based on the CNN comprises a data acquisition module, a model building module, a multi-mode feature encoding module and a read voltage prediction module, wherein the data acquisition module is used for acquiring a plurality of key parameters affecting read voltage, the key parameters comprise P/E times, data retention time, temperature and a read target position, the model building module is used for building a nonlinear coupling regression model reflecting nonlinear coupling relation between the key parameters and the read voltage, the multi-mode feature encoding module is used for carrying out standardization processing and interaction feature calculation on the key parameters based on the nonlinear coupling regression model, carrying out feature encoding on the standardized key parameters and interaction features together to generate a multi-channel feature tensor, and the read voltage prediction module is used for inputting the multi-channel feature tensor into a pre-trained convolutional neural network, processing the multi-channel feature tensor by using the convolutional neural network and outputting a predicted read voltage value. In a t