Method of training artificial intelligence to estimate sensing voltages for storage device

This application claims the benefit of priority to Taiwan Patent Application No. 107134387, filed on Sep. 28, 2018. The entire content of the above identified application is incorporated herein by reference.

Some references, which may include patents, patent applications and various publications, may be cited and discussed in the description of this disclosure. The citation and/or discussion of such references is provided merely to clarify the description of the present disclosure and is not an admission that any such reference is “prior art” to the disclosure described herein. All references cited and discussed in this specification are incorporated herein by reference in their entireties and to the same extent as if each reference was individually incorporated by reference.

The present disclosure relates to a storage device, and more particularly to a method of training artificial intelligence to estimate sensing voltages for a storage device.

Memories are seeing widespread use in recent years. However, memories may be damaged by multiple times of erasing and writing data, resulting in an increased probability of error and a significantly reduced reliability of the non-volatile memory. Therefore, it is necessary to improve the reliability of the non-volatile memory by supplying design techniques such as error correction techniques, such that a lifetime of a product can be prolonged and an operation state of the product can be more stable.

An error correction module for correcting error data read by the non-volatile memory is disposed in a control circuit of the memory to eliminate error caused by external factors in the non-volatile memory, thereby prolonging the lifetime of the non-volatile memory. A common error correction coding technology, such as a Bose-Chaudhuri-Hocquenghem (BCH) coding technology, is capable of fast computation and has a correction capability that increases with the increase of the number of redundant bits. However, the BCH coding technology cannot keep up with the manufacturing technologies of the non-volatile memory which have been greatly improved and cannot provide sufficient correction capability. Therefore, a Low Density Parity Code (LDPC) error correction technology widely used in the field of communication and having a strong correction capability is now being used in data storage.

In response to the above-referenced technical inadequacies, the present disclosure provides a method of training artificial intelligence to estimate sensing voltages for a storage device including a plurality of memory units each storing one or more bit values. The method includes the following steps: (a) supplying a plurality of initial sensing voltages having different voltage values to each of the memory units; (b) defining a plurality of storing states including a strong correct region, a weak correct region, a strong error region and a weak error region; (c) comparing the initial sensing voltages with the threshold voltage of each of the memory units, and accordingly classifying each of the memory units into the strong correct region, the weak correct region, the strong error region or the weak error region; (d) calculating a strong correct ratio of the number of the memory units classified in the strong correct region to the number of the memory units classified in the strong correct region and the weak correct region; (e) calculating a strong error ratio of the number of the memory units classified in the strong error region to the number of the memory units classified in the strong error region and the weak error region; (f) calculating the number of the memory units classified in the weak correct region and the weak error region to obtain a histogram parameter; (g) inputting the strong correct ratio, the strong error ratio and the histogram parameter to an artificial intelligence neural network system; and (h) using machine learning to adjust the initial sensing voltages based on the strong correct ratio, the strong error ratio and the histogram parameter to generate a plurality of practical sensing voltages replacing the initial sensing voltages, and then returning to step (a).

As described above, the present disclosure provides the method of training artificial intelligence to estimate the sensing voltages for the storage device, by which the initial sensing voltages are adjusted to generate the practical sensing voltages to be supplied to the memory units. The memory units are more precisely classified into the storing states including the strong correct, the weak correct, the strong error and the weak error, etc., based on the practical sensing voltages. It is worth noting that the number of the memory units classified in the strong error region based on the practical sensing voltage can be significantly reduced. Therefore, the success probability of decoding the bit values stored in the memory units by executing a probability decoding algorithm such as a low density parity check code or other algorithms by the decoder can be increased.

These and other aspects of the present disclosure will become apparent from the following description of the embodiment taken in conjunction with the following drawings and their captions, although variations and modifications therein may be affected without departing from the spirit and scope of the novel concepts of the disclosure.

The present disclosure will become more fully understood from the following detailed description and accompanying drawings.

FIG. 1 is a flowchart of a method of training artificial intelligence to estimate sensing voltages for a storage device according to a first embodiment of the present disclosure.

FIG. 2 is a flowchart of a method of training artificial intelligence to estimate sensing voltages for a storage device according to a second embodiment of the present disclosure.

FIG. 3 is a flowchart of a method of training artificial intelligence to estimate sensing voltages for a storage device according to a third embodiment of the present disclosure.

FIG. 4 is a flowchart of a method of training artificial intelligence to estimate sensing voltages for a storage device according to a fourth embodiment of the present disclosure.

FIG. 5 is a flowchart of a method of training artificial intelligence to estimate sensing voltages for a storage device according to a fifth embodiment of the present disclosure.

FIG. 6 is a flowchart of a method of training artificial intelligence to estimate sensing voltages for a storage device according to a sixth embodiment of the present disclosure.

FIG. 7 is a flowchart of a method of training artificial intelligence to estimate sensing voltages for a storage device according to a seventh embodiment of the present disclosure.

FIG. 8 is a graph of the number of memory units of the storage device to which initial sensing voltages estimated by using the method of training artificial intelligence are supplied verse threshold voltages of the memory units according to the embodiments of the present disclosure.

FIG. 9 is a graph of the number of the memory units of the storage device to which practical sensing voltages estimated by using the method of training artificial intelligence are supplied verse the threshold voltages of the memory units according to the embodiments of the present disclosure.

The present disclosure is more particularly described in the following examples that are intended as illustrative only since numerous modifications and variations therein will be apparent to those skilled in the art. Like numbers in the drawings indicate like components throughout the views. As used in the description herein and throughout the claims that follow, unless the context clearly dictates otherwise, the meaning of “a”, “an”, and “the” includes plural reference, and the meaning of “in” includes “in” and “on”. Titles or subtitles can be used herein for the convenience of a reader, which shall have no influence on the scope of the present disclosure.

The terms used herein generally have their ordinary meanings in the art. In the case of conflict, the present document, including any definitions given herein, will prevail. The same thing can be expressed in more than one way. Alternative language and synonyms can be used for any term(s) discussed herein, and no special significance is to be placed upon whether a term is elaborated or discussed herein. A recital of one or more synonyms does not exclude the use of other synonyms. The use of examples anywhere in this specification including examples of any terms is illustrative only, and in no way limits the scope and meaning of the present disclosure or of any exemplified term. Likewise, the present disclosure is not limited to various embodiments given herein. Numbering terms such as “first”, “second” or “third” can be used to describe various components, signals or the like, which are for distinguishing one component/signal from another one only, and are not intended to, nor should be construed to impose any substantive limitations on the components, signals or the like.

Reference is made to FIGS. 1 and 8 . FIG. 1 is a flowchart of a method of training artificial intelligence to estimate sensing voltages for a storage device according to a first embodiment of the present disclosure; FIG. 8 is a graph of the number of memory units of the storage device to which initial sensing voltages estimated by using the method of training artificial intelligence are supplied verse threshold voltages of the memory units according to the embodiments of the present disclosure.

As shown in FIG. 1 , the method of training artificial intelligence to estimate the sensing voltages for the storage device includes steps S 101 to S 115 . The storage device includes a plurality of memory units, and each of the plurality of memory units stores one or more bit values each of which is logic “0” or “1”.

In step S 101 , the initial sensing voltages are supplied to the memory units and have different voltage values.

For example, as shown in FIG. 8 , three initial sensing voltages Vt 1 , Vt 2 , Vt 3 are supplied to the memory units. The initial sensing voltage Vt 3 is larger than the initial sensing voltage Vt 1 . The initial sensing voltage Vt 2 is smaller than the initial sensing voltage Vt 1 . It should be understood that the number of the supplied initial sensing voltages can be adjusted according to actual requirement, for example, according to the number of regions of storing states into which the memory units are classified, but the present disclosure is not limited thereto.

In step S 103 , the plurality of storing states including a strong correct (SC) region, a weak correct (WC) region, a strong error (SE) region and a weak error (WE) region are defined.

In step S 105 , the threshold voltage of the memory unit is compared with the initial sensing voltages, and accordingly the memory unit is classified into the strong correct region, the weak correct region, the strong error region or the weak error region.

As shown in FIG. 8 , an entire region formed by two curves respectively represent the bit values Bit of logic “1” and “0” is divided into the regions of the storing states based on the initial sensing voltages Vt 1 , Vt 2 , Vt 3 .

The curve representing the bit value of logic “1” is used for the classification of the memory unit that intends to store the bit value of logic “1”. For example, the memory unit is classified into a strong correct region SC 1 , a weak correct region WC 1 , a strong error region SE 1 or a weak error region WE 1 .

On the other hand, the curve representing the bit value Bit of logic “0” is used for the classification of the memory unit that intends to store the bit value of logic “0”. For example, the memory unit is classified into a strong correct region SC 0 , a weak correct region WC 0 , a strong error region SE 0 or a weak error region WE 0 .

In step S 107 , a strong correct ratio (SCR) of the number of the memory units in the strong correct region to the number of the memory units in the strong and weak correct regions is calculated, which is expressed by the following equation:

wherein SCR represents the strong correct ratio, SC represents the number of the memory units in the strong correct region, and WC represents the number of the memory units in the weak correct region.

The memory unit intending to store the bit value Bit of logic “1” may be classified into the strong correct region SC 1 or the weak correct region WC 1 as shown in FIG. 8 . An area of the strong correct region SC 1 and an area of the weak correct region WC 1 are respectively calculated and then summed up. Finally, the strong correct ratio of the area of the strong correct region SC 1 to the sum area of the strong correct region SC 1 and the weak correct region WC 1 is calculated.

In addition, the memory unit intending to store the bit value Bit of logic “0” may be classified into the strong correct region SC 0 or the weak correct region WC 0 as shown in FIG. 8 . An area of the strong correct region SC 0 and an area of the weak correct region WC 0 are respectively calculated and then summed up. Finally, the strong correct ratio of the area of the strong correct region SC 0 to the sum area of the strong correct region SC 0 and the weak correct region WC 0 is calculated.

In step S 109 , a strong error ratio (SER) of the number of the memory units in the strong error region to the number of the memory units in the strong and weak error regions is calculated, which is expressed by the following equation:

wherein SCR represents the strong error ratio, SE represents the number of the memory units in the strong error region, and WE represents the number of the memory units in the weak error region.

The memory unit intending to store the bit value Bit of logic “1” may be classified into the strong error region SE 1 or the weak error region WE 1 as shown in FIG. 8 . An area of the strong error region SE 1 and an area of the weak error region WE 1 are respectively calculated and then summed up. Finally, the strong error ratio of the area of the strong error region SE 1 to the sum area of the strong error region SE 1 and the weak error region WE 1 is calculated.

The memory unit intending to store the bit value Bit of logic “0” may be classified into the strong correct region SC 0 or the weak correct region WC 0 as shown in FIG. 8 . An area of the strong error region SE 0 and an area of the weak error region WE 0 are respectively calculated and then summed up. Finally, the strong error ratio of the area of the strong error region SE 0 to the sum area of the strong error region SE 0 and the weak error region WE 0 is calculated.

In step S 111 , the number of the memory units classified into the weak correct region and the number of the memory units classified into the weak error region are summed up to obtain a histogram parameter.

For example, the area of the weak correct region WC 1 corresponding to the curve representing the bit value of logic “1” as shown in FIG. 8 is calculated. The area of the weak error region WE 0 corresponding to the curve representing the bit value of logic “0” as shown in FIG. 8 is calculated. The area of the weak correct region WC 1 and the area of the weak error region WE 0 are summed up to obtain a first sub-histogram parameter. That is, the first sub-histogram parameter is equal to a sum of the number of the memory units classified in the weak correct region WC 1 and the number of the memory units classified in the weak error region WE 0 . In addition, the area of the weak correct region WC 0 corresponding to the curve representing the bit value of logic “0” as shown in FIG. 8 is calculated. The area of the weak error region WE 1 corresponding to the curve representing the bit value of logic “1” as shown in FIG. 8 is calculated. The area of the weak correct region WC 0 and the area of the weak error region WE 1 are summed up to obtain a second sub-histogram parameter.

Alternatively, the area of the weak correct region WC 1 and the area of the weak error region WE 1 which correspond to the curve representing the bit value of logic “1” as shown in FIG. 8 are summed up to obtain the first sub-histogram parameter HM 1 . In addition, the area of the weak correct region WC 0 and the area of the weak error region WE 0 which correspond to the curve representing the bit value of logic “0” as shown in FIG. 8 are summed up to obtain the second sub-histogram parameter.

In step S 113 , the two strong correct ratios that respectively correspond to the bit values Bit of logic “0” and “1” as shown in FIG. 8 , the two strong error ratios that respectively correspond to the bit values Bit of logic “0” and “1” as shown in FIG. 8 , the first sub-histogram parameter and second sub-histogram parameter are inputted to an artificial intelligence neural network system (AI-NN).

In step S 115 , the initial sensing voltages are adjusted based on the above strong correct ratio, strong error ratio and histogram parameter to generate practical sensing voltages by using machine learning.

Reference is made to FIG. 2 , which is a flowchart of a method of training artificial intelligence to estimate sensing voltages for a storage device according to a second embodiment of the present disclosure. As shown in FIG. 2 , the method of training artificial intelligence to estimate the sensing voltages for the storage device includes steps S 201 to S 221 . The storage device includes the plurality of memory units and each of the plurality of memory units stores one or more bit values each of which is logic “0” or “1”.

In step S 201 , the initial sensing voltages are supplied to the memory units and have different voltage values.

In step S 203 , the bit values stored in the memory unit are decoded by executing a decoding program based on the initial sensing voltages by the decoder.

In step S 205 , it is determined whether the bit value stored in the memory unit is successfully decoded by the decoder or not. If the bit value stored in the memory unit is not successfully decoded, steps S 207 to S 219 are sequentially performed. If the bit value stored in the memory unit is successfully decoded, step S 221 is performed.

In step S 207 , the memory unit is reclassified into the strong correct region, the weak correct region, the strong error region or the weak error region.

In step S 209 , the strong correct ratio of the number of the memory units in the strong correct region to the number of the memory units in the strong and weak correct regions is calculated.

In step S 211 , the strong error ratio of the number of the memory units in the strong error region to the number of the memory units in the strong and weak error regions is calculated.

In step S 213 , the number of the memory units classified into the weak correct region and the number of the memory units classified into the weak error region are summed up to obtain the histogram parameter.

In step S 215 , the strong correct ratio, the strong error ratio and the histogram parameter are inputted to the artificial intelligence neural network system.

In step S 217 , the initial sensing voltages are adjusted based on the strong correct ratio, the strong error ratio and the histogram parameter to generate the practical sensing voltages by using machine learning.

In step S 219 , the practical sensing voltages are supplied to the memory units and have the different voltage values. Then, in step S 203 , the bit values stored in the memory unit are decoded by executing a decoding program based on the practical sensing voltages replacing the initial sensing voltages by the decoder.

In step S 221 , the sensing voltages by which the bit values stored in the memory unit are successfully decoded are recorded.

Reference is made to FIG. 3 , which is a flowchart of a method of training artificial intelligence to estimate sensing voltages for a storage device according to a third embodiment of the present disclosure. As shown in FIG. 3 , the method of training artificial intelligence to estimate the sensing voltages for the storage device includes steps S 301 to S 325 . The storage device includes the plurality of memory units and each of the plurality of memory units stores the one or more bit values each of which is logic “0” or “1”.

In step S 301 , the initial sensing voltages are supplied to each of the memory units and have the different voltage values.

In step S 303 , the threshold voltage of the memory unit is compared with the initial sensing voltages, and accordingly the memory unit is classified into the strong correct region, the weak correct region, the strong error region or the weak error region.

In step S 305 , the bit values stored in the memory unit are decoded by the decoder.

In step S 307 , it is determined whether the memory units are correctly classified into the storing states based on the initial sensing voltages or not, according to a success probability of decoding the bit values by the decoder. If the memory units are correctly classified into the storing states, the initial sensing voltages are recorded and the memory units are still classified based on the initial sensing voltages. If the memory units are not classified correctly into the storing states, steps S 309 to S 323 are sequentially performed.

In step S 309 , the strong correct ratio of the number of the memory units in the strong correct region to the number of the memory units in the strong and weak correct regions is calculated.

In step S 311 , the strong error ratio of the number of the memory units in the strong error region to the number of the memory units in the strong and weak correct regions is calculated.

In step S 313 , the number of the memory units classified into the weak correct region and the number of the memory units classified into the weak error region are summed up to obtain the histogram parameter.

In step S 315 , a process environment variable associated with a process in which the storage device accesses the bit values is obtained.

In step S 317 , the process environment variable, the strong correct ratio, the strong error ratio and the histogram parameter are inputted to the artificial intelligence neural network system.

In step S 319 , the practical sensing voltages replacing the initial sensing voltages are analyzed, based on the process environment variable, the strong correct ratio, the strong error ratio and the histogram parameter by using machine learning.

In step S 321 , the practical sensing voltages are supplied to the memory units and have the different voltage values from each other. The practical sensing voltage may be different from the initial sensing voltage.

In step S 323 , the threshold voltage of the memory unit is compared with the practical sensing voltages, and accordingly the memory unit is classified into the strong correct region, the weak correct region, the strong error region or the weak error region. If a classification result of the memory units based on the practical sensing voltages is matched with a decoding result of the decoder, the practical sensing voltages are recorded, and the memory units are still classified based on the practical sensing voltages.

However, it should be understood that the threshold voltages of the memory units are variable values. The changed threshold voltages may cause the memory units to be incorrectly classified based on the practical sensing voltages. For example, in step S 307 , it is determined that the memory units are wrongly classified into the storing state based on the practical sensing voltages according to the success probability of decoding the bit values by the decoder. Under this condition, steps S 307 to S 323 are performed again to generate new practical sensing voltages and the memory units are classified into the different storing states based on the new practical sensing voltages.

Reference is made to FIG. 4 , which is a flowchart of a method of training artificial intelligence to estimate sensing voltages for a storage device according to a fourth embodiment of the present disclosure. As shown in FIG. 4 , the method of training artificial intelligence to estimate the sensing voltages for the storage device includes steps S 401 to S 423 . The storage device includes the plurality of memory units and each of the plurality of memory units stores the one or more bit values each of which is logic “0” or “1”.

In step S 401 , the initial sensing voltages are supplied to the memory units and have different voltage values.

In step S 403 , the threshold voltage of the memory unit is compared with the initial sensing voltages, and accordingly the memory unit is classified into the strong correct region, the weak correct region, the strong error region or the weak error region.

In step S 405 , the number of the memory units classified in the strong error region based on the initial sensing voltages is calculated.

In step S 407 , it is determined whether the number of the memory units classified in the strong error region based on the initial sensing voltages is smaller than an error threshold or not. If the number of the memory units classified in the strong error region is not smaller than the error threshold, steps S 409 to S 421 are sequentially performed. If the number of the memory units classified in the strong error region is smaller than the error threshold, step S 423 is performed.

In step S 409 , the strong correct ratio of the number of the memory units in the strong correct region to the number of the memory units in the strong and weak correct regions is calculated.

In step S 411 , the strong error ratio of the number of the memory units in the strong error region to the number of the memory units in the strong and weak correct regions is calculated.

In step S 413 , the number of the memory units classified into the weak correct region and the number of the memory units classified into the weak error region are summed up to obtain the histogram parameter.

In step S 415 , the process environment variable associated with the process in which the storage device accesses the bit values is obtained.

In step S 417 , the process environment variable, the strong correct ratio, the strong error ratio and the histogram parameter are inputted to the artificial intelligence neural network system.

In step S 419 , the practical sensing voltages are analyzed based on the process environment variable, the strong correct ratio, the strong error ratio and the histogram parameter by using machine learning.

In step S 421 , the practical sensing voltages are supplied to the memory units and have different voltage values from each other. The practical sensing voltage may be different from the initial sensing voltage. Then, the initial sensing voltages are replaced with the practical sensing voltages and then steps S 403 to S 407 are performed.

In step S 423 , the initial sensing voltages or the practical sensing voltages by which the number of the memory units classified into the strong error region is smaller than the error threshold are recorded.

Reference is made to FIG. 5 , which is a flowchart of a method of training artificial intelligence to estimate sensing voltages for a storage device according to a fifth embodiment of the present disclosure. As shown in FIG. 5 , the method of training artificial intelligence to estimate the sensing voltages for the storage device includes steps S 501 to S 513 . The storage device includes the plurality of memory units each stores the one or more bit values each of which is logic “0” or “1”.

In step S 501 , the bit values stored in the memory unit are decoded by executing an initial decoding program based on the initial sensing voltages by the decoder.

In step S 503 , an initial success probability of decoding the bit values stored in the memory unit by executing the initial decoding program is calculated.

In step S 505 , the bit values stored in the memory unit are decoded by executing a practical decoding program based on the practical sensing voltages by the decoder.

In step S 507 , a practical success probability of decoding the bit values stored in the memory unit by executing the practical decoding program is calculated.

In step S 509 , it is determined whether the practical success probability is larger than the initial success probability or not. If the practical success probability is not larger than the initial success probability, step S 511 is performed. If the practical success probability is larger than the initial success probability, step S 513 is performed.

In step S 511 , the new practical sensing voltages are generated and steps 505 to S 509 are performed again based on the new practical sensing voltages.

In step S 513 , the practical sensing voltages are recorded.

Reference is made to FIG. 6 , which is a flowchart of a method of training artificial intelligence to estimate sensing voltages for a storage device according to a sixth embodiment of the present disclosure. As shown in FIG. 6 , the method of training artificial intelligence to estimate the sensing voltages for the storage device includes steps S 601 to S 613 . The storage device includes the plurality of memory units and each of the plurality of memory units stores the one or more bit values each of which is logic “0” or “1”.

In step S 601 , the threshold voltage of the memory unit is compared with the initial sensing voltages, and accordingly the memory unit is classified into the strong correct region, the weak correct region, the strong error region or the weak error region.

In step S 603 , an initial error number of the memory units classified into the strong error region based on the initial sensing voltages is calculated.

In step S 605 , the threshold voltage of the memory unit is compared with the practical sensing voltages, and accordingly the memory unit is classified into the strong correct region, the weak correct region, the strong error region or the weak error region.

In step S 607 , a practical error number of the memory units classified into the strong error region based on the practical sensing voltages is calculated.

In step S 609 , it is determined whether the practical error number is smaller than the initial error number or not. If the practical error number is smaller than the initial error number, step S 611 is performed. If the practical error number is not smaller than the initial error number, step S 613 is performed.

In step S 611 , the memory units are still classified based on the practical sensing voltages.

In step S 613 , the practical sensing voltages are generated by using machine learning by the artificial intelligence neural network system.

Reference is made to FIGS. 7 to 9 . FIG. 7 is a flowchart of a method of training artificial intelligence to estimate sensing voltages for a storage device according to a seventh embodiment of the present disclosure; FIG. 8 is a graph of the number of the memory units of the storage device to which initial sensing voltages estimated by using the method of training artificial intelligence are supplied verse threshold voltages of the memory units according to the embodiments of the present disclosure; FIG. 9 is a graph of the number of the memory units of the storage device to which practical sensing voltages estimated by using the method of training artificial intelligence are supplied verse the threshold voltages of the memory units according to the embodiments of the present disclosure.

As shown in FIG. 7 , the method of training artificial intelligence to estimate the sensing voltages for the storage device includes steps S 701 to S 713 . The storage device includes the plurality of memory units and each of the plurality of memory units stores the one or more bit values each of which is logic “0” or “1”.

In step S 701 , the threshold voltage of the memory unit is compared with the initial sensing voltages, and accordingly the memory unit is classified into the strong correct region, the weak correct region, the strong error region or the weak error region.

As shown in FIG. 8 , the threshold voltage of the memory unit is compared with the initial sensing voltages Vt 1 , Vt 2 , Vt 3 , and accordingly the memory unit intending to store the bit value Bit of logic “1” is classified into the strong correct region SC 1 , the weak correct region WC 1 , the strong error region SE 1 or the weak error region WE 1 , according to a comparison result of the threshold voltage and the initial sensing voltages Vt 1 , Vt 2 , Vt 3 . In addition, the memory unit intending to store the bit value Bit of logic “0” is classified into the strong correct region SC 0 , the weak correct region WC 0 , the strong error region SE 0 or the weak error region WE 0 , according to a comparison result of the threshold voltage and the initial sensing voltages Vt 1 , Vt 2 , Vt 3 .

In step S 703 , the initial correct number of the memory units classified into the strong correct region based on the initial sensing voltages is calculated.

In step S 705 , the threshold voltage of the memory unit is compared with the practical sensing voltages, and accordingly the memory unit is classified into the strong correct region, the weak correct region, the strong error region or the weak error region.

For example, in the above embodiments, the initial sensing voltages Vt 1 , Vt 2 , Vt 3 as shown in FIG. 8 may be adjusted into the practical sensing voltages Vt 1 ′, Vt 2 ′, Vt 3 ′ having different voltage values from the initial sensing voltages Vt 1 , Vt 2 , Vt 3 as shown in FIG. 9 . It is worth noting that the initial sensing voltage Vt 1 is adjusted into the practical sensing voltage Vt 1 ′, such that a dotted line representing the initial sensing voltage Vt 1 ′ as shown in FIG. 9 crisscrosses with an intersection point or its vicinity of the curve representing the bit value Bit of the logic “1” and the curve representing the bit value Bit of the logic “0”.

As shown in FIG. 9 , the threshold voltage of the memory unit is compared with the practical sensing voltages Vt 1 ′, Vt 2 ′, Vt 3 ′. The memory unit intending to store the bit value Bit of logic “1” is classified into a strong correct region SC 1 ′, a weak correct region WC 1 ′, a strong error region SE 1 ′ or a weak error region WE 1 ′, according to a comparison result of the threshold voltage and the initial sensing voltages Vt 1 ′, Vt 2 ′, Vt 3 ′. The memory unit intending to store the bit value Bit of logic “0” is classified into a strong correct region SC 0 ′, a weak correct region WC 0 ′, a strong error region SE 0 ′ or a weak error region WE 0 ′, according to a comparison result of the threshold voltage and the practical sensing voltages Vt 1 ′, Vt 2 ′, Vt 3 ′.

In step S 707 , a practical correct number of the memory units classified into the strong correct region based on the practical sensing voltages is calculated.

In step S 709 , it is determined whether the practical correct number is larger than the initial correct number or not. If the practical correct number is not larger than the initial correct number, step S 711 is performed. If the practical correct number is larger than the initial correct number, step S 713 is performed.

In step S 711 , the practical sensing voltages are regenerated by using machine learning by the artificial intelligence neural network system.

In step S 713 , the memory units are still classified based on the practical sensing voltages. As shown in FIG. 9 , the memory units are classified based on the practical sensing voltages Vt 1 ′, Vt 2 ′, Vt 3 ′.

The number of the memory units classified into the strong error region SE 1 ′ that is an area of the strong error region SE 1 ′ based on the practical sensing voltages Vt 1 ′, Vt 2 ′, Vt 3 ′ as shown in FIG. 9 is smaller than the number of the memory units classified into the strong error region SE 1 that is an area of the strong error region SE 1 based on the practical sensing voltages Vt 1 , Vt 2 , Vt 3 as shown in FIG. 8 .

In addition, the number of the memory units classified into the strong error region SE 0 ′ that is an area of the strong error region SE 0 ′ based on the practical sensing voltages Vt 1 ′, Vt 2 ′, Vt 3 ′ as shown in FIG. 9 is smaller than the number of the memory units classified into the strong error region SE 0 that is an area of the strong error region SE 0 based on the practical sensing voltages Vt 1 , Vt 2 , Vt 3 as shown in FIG. 8 .

In summary, the present disclosure provides the method of training artificial intelligence to estimate the sensing voltages for the storage device, by which the initial sensing voltages are adjusted to generate the practical sensing voltages to be supplied to the memory units. The memory units are more precisely classified into the storing states including the strong correct, the weak correct, the strong error and the weak error, etc., based on the practical sensing voltages. It is worth noting that the number of the memory units classified in the strong error region based on the practical sensing voltage can be significantly reduced. Therefore, the success probability of decoding the bit values stored in the memory units by executing a probability decoding algorithm such as a low density parity check code or other algorithms by the decoder can be increased.

The foregoing description of the exemplary embodiments of the disclosure has been presented only for the purposes of illustration and description and is not intended to be exhaustive or to limit the disclosure to the precise forms disclosed. Many modifications and variations are possible in light of the above teaching.

The embodiments were chosen and described in order to explain the principles of the disclosure and their practical application so as to enable others skilled in the art to utilize the disclosure and various embodiments and with various modifications as are suited to the particular use contemplated.

Alternative embodiments will become apparent to those skilled in the art to which the present disclosure pertains without departing from its spirit and scope.