Flash memory characteristic prediction model training method and flash memory characteristic prediction method and device

Abstract

The invention provides a flash memory characteristic prediction model training method and device and a flash memory characteristic prediction method and device, and relates to the technical field of artificial intelligence. The graph structure data comprises a plurality of nodes and edges between the nodes, the nodes are determined based on first target information of at least one type of flash memory, and the first target information comprises at least one of the following items: manufacturing process information, electrical attribute information and working condition information; the edges represent the interaction relationship among the nodes; and training an initial flash memory characteristic prediction model based on the graph structure data to obtain a target flash memory characteristic prediction model. According to the method, the target flash memory characteristic prediction model is obtained through training by using the multi-dimensional target information of various types of flash memories and the graph structure data constructed by the interaction relationship between the multi-dimensional target information, so that the accurate prediction of the flash memory characteristics can be realized, and the prediction accuracy of the flash memory characteristics is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a flash memory characteristic prediction model training method, a flash memory characteristic prediction method and a flash memory characteristic prediction device. Background Art

[0002] With the widespread application of solid-state drives (SSDs), NAND Flash is the core storage medium, and its performance directly affects the overall performance of SSDs. However, with the continuous advancement of NAND flash technology, the physical size of storage units continues to shrink, making the reliability of NAND flash increasingly prominent. Therefore, how to accurately predict the characteristics of NAND flash and optimize the design and use strategy of storage systems has become one of the current research hotspots.

[0003] Traditional NAND flash memory characteristic prediction methods mainly rely on empirical models and simplified mathematical models. Although these methods can reflect the characteristics of NAND flash memory to a certain extent, they cannot fully consider the complex factors in actual working conditions, such as temperature, voltage fluctuations, usage patterns, etc., so the prediction accuracy is limited. In addition, with the continuous development of NAND flash memory technology, the limitations of traditional methods have become increasingly prominent and cannot meet the growing demand for data storage.

[0004] In the prior art, some methods based on machine learning algorithms to predict NAND flash memory characteristics have been proposed. These methods mainly use machine learning models such as neural networks, support vector machines, and random forests to model and predict key parameters such as threshold voltage distribution, erase times, and data retention time of NAND flash memory. However, some methods only focus on single characteristics such as NAND flash memory signal detection or life prediction, and fail to fully consider multiple key characteristics of NAND flash memory; although some methods use neural network models, the model structure is relatively simple and cannot fully capture the complex nonlinear relationship in NAND flash memory characteristics. These prior arts still have some shortcomings, which makes the accuracy of flash memory characteristic prediction relatively low. Summary of the invention

[0005] The present invention provides a flash memory characteristic prediction model training method, a flash memory characteristic prediction method and a flash memory characteristic prediction device, which are used to solve the problem of low accuracy of flash memory characteristic prediction.

[0006] In a first aspect, the present invention provides a flash memory characteristic prediction model training method, comprising: Constructing graph structure data; the graph structure data includes a plurality of nodes and edges between the nodes, the nodes are determined based on first target information of at least one model of flash memory, the first target information includes at least one of the following: manufacturing process information, electrical property information and working condition information; the edges represent the interaction relationship between the nodes; Based on the graph structure data, an initial flash memory characteristic prediction model is trained to obtain a target flash memory characteristic prediction model.

[0007] According to a flash memory characteristic prediction model training method provided by the present invention, the step of constructing graph structure data includes: Obtaining the first target information of each model of flash memory; Performing a characteristic test on the first target information of each model of flash memory to obtain characteristic parameters and characteristic test results during the characteristic test; Normalizing the first target information, the characteristic parameter, and the characteristic test result respectively to obtain normalized first target information, normalized characteristic parameter, and normalized characteristic test result; The graph structure data is constructed based on the normalized first target information, the normalized characteristic parameters and the normalized characteristic test results; the node is any one of the normalized first target information, the normalized characteristic parameters and the normalized characteristic test results.

[0008] According to a flash memory characteristic prediction model training method provided by the present invention, the initial flash memory characteristic prediction model includes a graph attention network module and a full connection layer; The initial flash memory characteristic prediction model is trained based on the graph structure data to obtain a target flash memory characteristic prediction model, including: Step A: inputting the graph structure data into the graph attention network module in the initial flash memory characteristic prediction model, and obtaining the target weight between any node in each of the nodes output by the graph attention network module and other nodes in the neighborhood; Step B: inputting each of the target weights into the full-link layer in the initial flash memory characteristic prediction model to obtain the characteristic parameter prediction value of each model of flash memory output by the full-link layer; Step C: Determine the target flash memory characteristic prediction model based on the characteristic parameter prediction values ​​of each model of flash memory and the actual values ​​of the characteristic parameters of each model of flash memory.

[0009] According to a flash memory characteristic prediction model training method provided by the present invention, the target flash memory characteristic prediction model is determined based on the characteristic parameter prediction value of each model of flash memory and the true value of the characteristic parameter of each model of flash memory, including: Calculating a loss function based on the predicted values ​​of the characteristic parameters of each model of flash memory and the true values ​​of the characteristic parameters of each model of flash memory; If the cumulative value of the loss function does not tend to be stable or the number of training times does not reach the maximum number of training times, update the parameters in the initial flash memory characteristic prediction model, and repeat the above steps A to C until the cumulative value of the loss function tends to be stable or the number of training times reaches the maximum number of training times; The target flash memory characteristic prediction model is determined based on the parameters in the initial flash memory characteristic prediction model that is finally updated.

[0010] According to a flash memory characteristic prediction model training method provided by the present invention, the graph attention network module includes at least one graph attention layer; The step of inputting the graph structure data into the graph attention network module in the initial flash memory characteristic prediction model to obtain the target weight between any node in each of the nodes output by the graph attention network module and other nodes in the neighborhood includes: Inputting the graph structure data into a first graph attention layer of the at least one graph attention layer included in the graph attention network module, obtaining a first weight between any node in each of the nodes output by the first graph attention layer and other nodes in the neighborhood; the first weight is used to update the graph structure data; Inputting the updated graph structure data into a second graph attention layer in the at least one graph attention layer, obtaining a second weight between any node in the nodes output by the second graph attention layer and other nodes in the neighborhood; the first weight is used to update the updated graph structure data; The execution is repeated until the finally updated graph structure data is input into the last graph attention layer of the at least one graph attention layer, and the target weight between any node in the nodes output by the last graph attention layer and other nodes in the neighborhood is obtained.

[0011] In a second aspect, the present invention further provides a flash memory characteristic prediction method, comprising: Acquire second target information of the target flash memory, wherein the second target information includes at least one of the following: manufacturing process information, electrical property information, and working condition information; The second target information is input into a target flash memory characteristic prediction model to obtain characteristic parameters of the target flash memory output by the target flash memory characteristic prediction model; the target flash memory characteristic prediction model is trained based on the flash memory characteristic prediction model training method described in any one of the first aspects.

[0012] In a third aspect, the present invention further provides a flash memory characteristic prediction model training device, comprising: A construction module, configured to construct graph structure data; the graph structure data includes a plurality of nodes and edges between the nodes, the nodes are determined based on first target information of at least one model of flash memory, the first target information includes at least one of the following: manufacturing process information, electrical property information, and working condition information; the edges represent the interaction relationship between the nodes; The training module is used to train the initial flash memory characteristic prediction model based on the graph structure data to obtain a target flash memory characteristic prediction model.

[0013] In a fourth aspect, the present invention further provides a flash memory characteristic prediction device, characterized in that it comprises: A first acquisition module is used to acquire second target information of a target flash memory, wherein the second target information includes at least one of the following: manufacturing process information, electrical property information, and working condition information; A prediction module is used to input the second target information into a target flash memory characteristic prediction model to obtain characteristic parameters of the target flash memory output by the target flash memory characteristic prediction model; the target flash memory characteristic prediction model is trained by the flash memory characteristic prediction model training device according to any one of the third aspects.

[0014] In a fifth aspect, the present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements a flash memory characteristic prediction model training method as described in any one of the first aspect above, or implements a flash memory characteristic prediction method as described in the second aspect above.

[0015] In a sixth aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a flash memory characteristic prediction model training method as described in any one of the first aspects above, or implements a flash memory characteristic prediction method as described in the second aspect above.

[0016] In a seventh aspect, the present invention further provides a computer program product, comprising a computer program, which, when executed by a processor, implements the flash memory characteristic prediction model training method as described in any one of the first aspects above, or implements the flash memory characteristic prediction method as described in the second aspect above.

[0017] The flash memory characteristic prediction model training method, flash memory characteristic prediction method and device provided by the present invention construct graph structure data; the graph structure data includes multiple nodes and edges between the nodes, the nodes are determined based on the first target information of at least one model of flash memory, the first target information includes at least one of the following: manufacturing process information, electrical property information and working condition information; the edges represent the interaction relationship between the nodes; based on the graph structure data, the initial flash memory characteristic prediction model is trained to obtain the target flash memory characteristic prediction model. By using the graph structure data constructed by multi-dimensional target information of various models of flash memory and the interaction relationship between them, the trained target flash memory characteristic prediction model can realize the accurate prediction of flash memory characteristics and improve the accuracy of flash memory characteristic prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 This is one of the flow charts of the flash memory characteristic prediction model training method provided by the present invention.

[0020] Figure 2 It is a schematic diagram of the graph structure data provided by the present invention.

[0021] Figure 3 It is a structural schematic diagram of the graph attention layer provided by the present invention.

[0022] Figure 4 This is the second flow chart of the flash memory characteristic prediction model training method provided by the present invention.

[0023] Figure 5 It is a flow chart of the flash memory characteristic prediction method provided by the present invention.

[0024] Figure 6 It is a structural schematic diagram of the flash memory characteristic prediction model training device provided by the present invention.

[0025] Figure 7 It is a structural schematic diagram of a flash memory characteristic prediction device provided by the present invention.

[0026] Figure 8 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] Combine the following Figure 1-Figure 5 The flash memory characteristic prediction model training method and the flash memory characteristic prediction method of the present invention are described.

[0029] Figure 1 FIG. 1 is one of the flow charts of the flash memory characteristic prediction model training method provided by the present invention, such as Figure 1 As shown, the method includes the following steps 101-102.

[0030] Step 101, construct graph structure data; the graph structure data includes multiple nodes and edges between each of the nodes, the nodes are determined based on first target information of at least one model of flash memory, the first target information includes at least one of the following: manufacturing process information, electrical property information and working condition information; the edges represent the interaction relationship between each of the nodes.

[0031] It should be noted that the flash memory characteristic prediction model training method provided by the present invention can be applied to flash memory characteristic prediction scenarios, such as NAND flash memory characteristic prediction. The executor of the method can be a flash memory characteristic prediction model training device, such as an electronic device, or a control module in the flash memory characteristic prediction model training device for executing the flash memory characteristic prediction model training method.

[0032] Specifically, the manufacturing process information is the structure of the flash memory or the material of the flash memory, the electrical property information is the optimal read voltage axis information, and the working conditions are the working time, temperature, and number of erases and writes of the flash memory. The graph structure data includes multiple nodes and edges between the nodes, wherein the nodes are determined based on the first target information of at least one model of flash memory, and the first target information includes at least one of the following: manufacturing process information, electrical property information, and working condition information; the edges represent the interaction relationship between the nodes.

[0033] Step 102: Based on the graph structure data, an initial flash memory characteristic prediction model is trained to obtain a target flash memory characteristic prediction model.

[0034] Specifically, based on the graph structure data, the initial flash memory characteristic prediction model is trained to obtain the target flash memory characteristic prediction model.

[0035] The flash memory characteristic prediction model training method provided by the present invention constructs graph structure data; the graph structure data includes multiple nodes and edges between the nodes, the nodes are determined based on the first target information of at least one model of flash memory, the first target information includes at least one of the following: manufacturing process information, electrical property information and working condition information; the edges represent the interaction relationship between the nodes; based on the graph structure data, the initial flash memory characteristic prediction model is trained to obtain the target flash memory characteristic prediction model. By using the graph structure data constructed by using the multi-dimensional target information of various models of flash memory and the interaction relationship between them, the target flash memory characteristic prediction model obtained by training can realize the accurate prediction of flash memory characteristics and improve the accuracy of flash memory characteristic prediction.

[0036] Optionally, a specific implementation of step 101 includes: Acquire the first target information of each of the models of flash memory; perform a characteristic test on the first target information of each of the models of flash memory to obtain characteristic parameters and characteristic test results in the characteristic test process; normalize the first target information, the characteristic parameters and the characteristic test results respectively to obtain the normalized first target information, the normalized characteristic parameters and the normalized characteristic test results; construct the graph structure data based on the normalized first target information, the normalized characteristic parameters and the normalized characteristic test results; the node is any one of the normalized first target information, the normalized characteristic parameters and the normalized characteristic test results.

[0037] Specifically, first target information of various types of flash memory (such as NAND) is collected through data disclosed by various manufacturers, and the first target information includes at least one of the following: manufacturing process information, electrical property information and working condition information; then, characteristic tests such as timing and optimal read voltage axis are performed on the first target information of each type of flash memory to obtain characteristic parameters and characteristic test results of each type of flash memory during the characteristic test in the actual working environment, and the characteristic parameters include at least one of the following: read and write times, temperature, and data retention time; the characteristic test results include at least one of the following: write time and number of errors.

[0038] The first target information, characteristic parameters and characteristic test results are preprocessed respectively to clean out abnormal values ​​and inconsistent data to ensure the accuracy of subsequent analysis; the first target information, characteristic parameters and characteristic test results are normalized respectively to obtain the normalized first target information, normalized characteristic parameters and normalized characteristic test results to improve the efficiency and accuracy of flash memory characteristic prediction model training; graph structure data is constructed based on the normalized first target information, normalized characteristic parameters and normalized characteristic test results; wherein a node is any one of the normalized first target information, normalized characteristic parameters and normalized characteristic test results.

[0039] Figure 2 is a schematic diagram of the graph structure data provided by the present invention, such as Figure 2 As shown in the figure, the nodes are working conditions, manufacturing process, electrical properties, characteristic test results and characteristic parameters, which are connected by edges. The edges represent the interaction between working conditions, manufacturing process, electrical properties, characteristic test results and characteristic parameters, but there are no edges between characteristic parameters. Finally, feature embedding is performed on each node, and the characteristic parameters are converted into high-dimensional vector representations to facilitate the processing of the initial flash memory characteristic prediction model.

[0040] Optionally, the initial flash memory characteristic prediction model includes a graph attention network module and a full-link layer; the specific implementation of the above step 102 includes: Step A: Input the graph structure data into the graph attention network module in the initial flash memory characteristic prediction model to obtain the target weight between any node in the nodes output by the graph attention network module and other nodes in the neighborhood.

[0041] It should be noted that, in the training phase of the initial flash memory characteristic prediction model, the weight and bias of the initial flash memory characteristic prediction model must first be initialized.

[0042] Specifically, the graph attention network module is used to determine the target weight between any node in each node and other nodes in the neighborhood, so as to realize the weighted aggregation of neighboring nodes. The graph structure data is input into the graph attention network module in the initial flash memory characteristic prediction model, and the target weight between any node in each node and other nodes in the neighborhood output by the graph attention network module can be obtained.

[0043] Step B: Input each of the target weights into the full-link layer in the initial flash memory characteristic prediction model to obtain the characteristic parameter prediction value of each model of flash memory output by the full-link layer.

[0044] Specifically, the full-link layer is used to output the characteristic parameter prediction values ​​of each model of flash memory. Each target weight is input into the full-link layer in the initial flash memory characteristic prediction model to obtain the characteristic parameter prediction values ​​of each model of flash memory output by the full-link layer.

[0045] Step C: Determine the target flash memory characteristic prediction model based on the characteristic parameter prediction values ​​of each model of flash memory and the actual values ​​of the characteristic parameters of each model of flash memory.

[0046] Specifically, based on the predicted values ​​of the characteristic parameters of each model of flash memory and the true values ​​of the characteristic parameters of each model of flash memory, a target flash memory characteristic prediction model can be further determined.

[0047] Optionally, the graph attention network module includes at least one graph attention layer; the graph structure data is input into the graph attention network module in the initial flash memory characteristic prediction model, and the target weight between any node in each of the nodes output by the graph attention network module and other nodes in the neighborhood is obtained, including: (1) Inputting the graph structure data into a first graph attention layer of the at least one graph attention layer included in the graph attention network module, and obtaining a first weight between any node in the nodes output by the first graph attention layer and other nodes in the neighborhood; the first weight is used to update the graph structure data.

[0048] Specifically, the graph attention network module includes at least one graph attention layer, and each graph attention layer is in series, that is, multiple graph attention layers are stacked to form the main body of the network. Each graph attention layer calculates the weights between nodes through the attention mechanism to achieve weighted aggregation of neighboring nodes. The graph attention layer includes steps such as linear transformation, attention coefficient calculation, and feature aggregation to assign different weights to different nodes in the neighborhood; among them, the graph attention layer adopts the attention mechanism, uses the "attention head" structure to calculate the attention scores between nodes, measures the similarity and importance between nodes through methods such as dot product attention, and updates the feature representation of the node according to the attention score, so that the node can capture richer contextual information, thereby assigning different weights to different nodes in the neighborhood.

[0049] Figure 3 It is a schematic diagram of the structure of the graph attention layer provided by the present invention, such as Figure 3As shown in the figure, X0, X1, X2, X3 and X4 represent nodes, X0 is the central node, X1, X2, X3 and X4 are other nodes in the neighborhood of X0, a0,1 represents the weight between X1 and X0, a0,2 represents the weight between X2 and X0, a0,3 represents the weight between X3 and X0, and a0,4 represents the weight between X4 and X0; among them, a0,1 is the high-dimensional vector of nodes represented by nodes X1 and X0, which is obtained by measuring the similarity and importance between nodes through methods such as dot product attention.

[0050] In the model training phase, the weights between the nodes in the graph structure data need to be initialized. The graph structure data is input into the first graph attention layer of at least one graph attention layer included in the graph attention network module, and the first weight between any node in each node output by the first graph attention layer and other nodes in the neighborhood can be obtained; wherein the first weight is used to update the graph structure data.

[0051] (2) Inputting the updated graph structure data into a second graph attention layer in the at least one graph attention layer to obtain a second weight between any node in the nodes output by the second graph attention layer and other nodes in the neighborhood; the first weight is used to update the updated graph structure data.

[0052] Specifically, after the first weight between any node in each node output by the first graph attention layer and other nodes in the neighborhood is obtained, the graph structure data is updated based on the first weight to obtain the updated graph structure data. The updated graph structure data is then input into the second graph attention layer in at least one graph attention layer to obtain the second weight between any node in each node output by the second graph attention layer and other nodes in the neighborhood; the first weight is used to update the updated graph structure data.

[0053] (3) Repeat the process until the updated graph structure data is input into the last graph attention layer of the at least one graph attention layer, and the target weight between any node in the nodes output by the last graph attention layer and other nodes in the neighborhood is obtained.

[0054] Specifically, the execution is repeated until the finally updated graph structure data is input into the last graph attention layer in at least one graph attention layer, and the target weight between any node in each node output by the last graph attention layer and other nodes in the neighborhood can be obtained.

[0055] Optionally, the determining the target flash memory characteristic prediction model based on the characteristic parameter prediction values ​​of each model of flash memory and the actual values ​​of the characteristic parameters of each model of flash memory includes: (a) Calculating a loss function based on the predicted values ​​of the characteristic parameters of each model of the flash memory and the true values ​​of the characteristic parameters of each model of the flash memory.

[0056] Specifically, the true value of the characteristic parameter of each model of flash memory is the characteristic parameter obtained during the characteristic test process, and the mean square error is used as a loss function to calculate the difference between the predicted value of the characteristic parameter and the true value of the characteristic parameter.

[0057] (b) When the cumulative value of the loss function does not tend to be stable or the number of training times does not reach the maximum number of training times, update the parameters in the initial flash memory characteristic prediction model, and repeat the above steps A to C until the cumulative value of the loss function tends to be stable or the number of training times reaches the maximum number of training times.

[0058] Specifically, determine whether the cumulative value of the loss function tends to be stable or whether the number of training times has reached the maximum number of training times. If the cumulative value of the loss function does not tend to be stable or the number of training times has not reached the maximum number of training times, use the gradient of the loss function to the model parameters for back propagation and guide the update of the network weights. Finally, use the Adam optimization algorithm to update the network weights according to the gradient information obtained by back propagation. This process is iterative, and each iteration includes one forward propagation and one backward propagation, followed by the update of the weights, thereby updating the parameters in the initial flash memory characteristic prediction model, and repeating the above steps A to C until the cumulative value of the loss function tends to be stable or the number of training times reaches the maximum number of training times. The maximum number of training times is pre-set, for example, the maximum number of training times is 100 times.

[0059] (c) determining the target flash memory characteristic prediction model based on the parameters in the initial flash memory characteristic prediction model that is finally updated.

[0060] Specifically, the parameters in the final updated initial flash memory characteristic prediction model are determined as the parameters of the target flash memory characteristic prediction model, thereby determining the target flash memory characteristic prediction model.

[0061] Figure 4 FIG. 2 is a flow chart of the flash memory characteristic prediction model training method provided by the present invention. Figure 4 As shown, the method includes steps 401 to 411.

[0062] Step 401, obtaining data: obtaining first target information of each model of flash memory; the first target information includes at least one of the following: manufacturing process information, electrical property information and working condition information.

[0063] Step 402: Characteristic test: Perform a characteristic test on the first target information of each model of flash memory to obtain characteristic parameters and characteristic test results during the characteristic test.

[0064] Step 403, data preprocessing: The first target information, characteristic parameters and characteristic test results are cleaned and normalized respectively to obtain normalized first target information, normalized characteristic parameters and normalized characteristic test results.

[0065] Step 404: construct graph structure data. Graph structure data is constructed based on the normalized first target information, the normalized characteristic parameters, and the normalized characteristic test results; the node is any one of the normalized first target information, the normalized characteristic parameters, and the normalized characteristic test results.

[0066] Step 405, determine the first weight output by the first graph attention layer. Input the graph structure data into the first graph attention layer of at least one graph attention layer included in the graph attention network module, and obtain the first weight between any node in each node output by the first graph attention layer and other nodes in the neighborhood; the first weight is used to update the graph structure data.

[0067] Step 406, determining a second weight output by the second graph attention layer. Input the updated graph structure data to the second graph attention layer in at least one graph attention layer to obtain a second weight between any node in each node output by the second graph attention layer and other nodes in the neighborhood; the first weight is used to update the updated graph structure data.

[0068] Step 407, determine the target weight output by the last graph attention layer. Repeat the process until the updated graph structure data is input to the last graph attention layer in at least one graph attention layer, and obtain the target weight between any node in each node output by the last graph attention layer and other nodes in the neighborhood.

[0069] Step 408: Determine the predicted value of the characteristic parameter. Input each target weight into the full-link layer in the initial flash memory characteristic prediction model to obtain the predicted value of the characteristic parameter of each model of flash memory output by the full-link layer.

[0070] Step 409, calculating a loss function: Based on the predicted values ​​of the characteristic parameters of each model of flash memory and the true values ​​of the characteristic parameters of each model of flash memory, a loss function is calculated.

[0071] Step 410, update and optimize the parameters of the model. If the cumulative value of the loss function does not tend to be stable or the number of training times does not reach the maximum number of training times, update the parameters in the initial flash memory characteristic prediction model, and repeat the above steps 405 to 409 until the cumulative value of the loss function tends to be stable or the number of training times reaches the maximum number of training times.

[0072] Step 411, determining a target flash memory characteristic prediction model: Determine a target flash memory characteristic prediction model based on the parameters in the final updated initial flash memory characteristic prediction model.

[0073] The flash memory characteristic prediction model training method provided by the present invention fully considers the multi-dimensional characteristics of different models of flash memory and the complex relationships between them by constructing graph structure data, and then trains the flash memory characteristic prediction model through the graph structure data, so that the flash memory characteristic prediction model can adaptively learn the characteristic differences between different models of flash memory, and then the trained flash memory characteristic prediction model can still maintain a high prediction performance when facing new types or new batches of flash memory. The improvement of this generalization ability helps to reduce the retraining and debugging costs of the model in different application scenarios, and can achieve accurate prediction of flash memory characteristics.

[0074] Figure 5 FIG. 1 is a flow chart of a flash memory characteristic prediction method provided by the present invention. Figure 5 As shown, the method includes steps 501 and 502.

[0075] Step 501, obtaining second target information of a target flash memory, wherein the second target information includes at least one of the following: manufacturing process information, electrical property information, and working condition information.

[0076] Specifically, the manufacturing process information is the structure of the flash memory or the material of the flash memory, the electrical property information is the optimal read voltage axis information, and the working conditions are information such as the working time, temperature, and number of erase and write times of the flash memory.

[0077] Step 502: input the second target information into a target flash memory characteristic prediction model to obtain characteristic parameters of the target flash memory output by the target flash memory characteristic prediction model.

[0078] The target flash memory characteristic prediction model is obtained by training based on the flash memory characteristic prediction model training method described in the aforementioned embodiments.

[0079] Specifically, the second target information is input into a pre-trained target flash memory characteristic prediction model to obtain characteristic parameters of the target flash memory output by the target flash memory characteristic prediction model, wherein the characteristic parameters of the target flash memory include at least one of the following: read and write times, temperature, and data retention time.

[0080] Optionally, the characteristic parameters of the target flash memory can also be output in the form of a visual chart or report, so that the user can intuitively understand the current status of the target flash memory (the characteristic parameters currently output) and the predicted trend, which helps the user to better formulate data storage and management strategies and optimize the allocation and use of storage resources.

[0081] Optionally, accurate prediction of the characteristics of the target flash memory provides strong support for the optimization of storage technology. Manufacturers can use the predicted characteristic parameters to improve the design and production process of flash memory, thereby improving its performance and reliability. At the same time, researchers can also conduct more in-depth storage technology research and development based on the predicted characteristic parameters.

[0082] The flash memory characteristic prediction method provided by the present invention obtains the second target information of the target flash memory, wherein the second target information includes at least one of the following: manufacturing process information, electrical property information, and working condition information; the second target information is input into the target flash memory characteristic prediction model to obtain the characteristic parameters of the target flash memory output by the target flash memory characteristic prediction model. Since the multi-dimensional characteristics of different types of flash memory and the complex relationship between them are considered when training the target flash memory characteristic prediction model, the flash memory characteristic prediction model can adaptively learn the characteristic differences between different types of flash memory, thereby enabling the trained flash memory characteristic prediction model to achieve accurate prediction of flash memory characteristics and improve the read and write efficiency and stability of SSD.

[0083] The flash memory characteristic prediction model training device provided by the present invention is described below. The flash memory characteristic prediction model training device described below and the flash memory characteristic prediction model training method described above can be referenced to each other.

[0084] Figure 6 Schematic diagram of the structure of the flash memory characteristic prediction model training device provided by the present invention. Figure 6 As shown, the flash memory characteristic prediction model training device 600 includes: a construction module 601 and a training module 602; wherein,

[0085] A construction module 601 is used to construct graph structure data; the graph structure data includes a plurality of nodes and edges between the nodes, the nodes are determined based on first target information of at least one model of flash memory, the first target information includes at least one of the following: manufacturing process information, electrical property information and working condition information; the edges represent the interaction relationship between the nodes; The training module 602 is used to train the initial flash memory characteristic prediction model based on the graph structure data to obtain a target flash memory characteristic prediction model.

[0086] The flash memory characteristic prediction model training device provided by the present invention constructs graph structure data; the graph structure data includes multiple nodes and edges between the nodes, the nodes are determined based on the first target information of at least one model of flash memory, the first target information includes at least one of the following: manufacturing process information, electrical property information and working condition information; the edges represent the interaction relationship between the nodes; based on the graph structure data, the initial flash memory characteristic prediction model is trained to obtain the target flash memory characteristic prediction model. By using the graph structure data constructed by using the multi-dimensional target information of various models of flash memory and the interaction relationship between them, the target flash memory characteristic prediction model obtained by training can realize the accurate prediction of flash memory characteristics and improve the accuracy of flash memory characteristic prediction.

[0087] Optionally, the construction module 601 is specifically used to: Obtaining the first target information of each model of flash memory; Performing a characteristic test on the first target information of each model of flash memory to obtain characteristic parameters and characteristic test results during the characteristic test; Normalizing the first target information, the characteristic parameter, and the characteristic test result respectively to obtain normalized first target information, normalized characteristic parameter, and normalized characteristic test result; The graph structure data is constructed based on the normalized first target information, the normalized characteristic parameters and the normalized characteristic test results; the node is any one of the normalized first target information, the normalized characteristic parameters and the normalized characteristic test results.

[0088] Optionally, the initial flash memory characteristic prediction model includes a graph attention network module and a full-link layer; the training module 602 is specifically used to: Step A: inputting the graph structure data into the graph attention network module in the initial flash memory characteristic prediction model, and obtaining the target weight between any node in each of the nodes output by the graph attention network module and other nodes in the neighborhood; Step B: inputting each of the target weights into the full-link layer in the initial flash memory characteristic prediction model to obtain the characteristic parameter prediction value of each model of flash memory output by the full-link layer; Step C: Determine the target flash memory characteristic prediction model based on the characteristic parameter prediction values ​​of each model of flash memory and the actual values ​​of the characteristic parameters of each model of flash memory.

[0089] Optionally, the training module 602 is further used to: Calculating a loss function based on the predicted values ​​of the characteristic parameters of each model of flash memory and the true values ​​of the characteristic parameters of each model of flash memory; If the cumulative value of the loss function does not tend to be stable or the number of training times does not reach the maximum number of training times, update the parameters in the initial flash memory characteristic prediction model, and repeat the above steps A to C until the cumulative value of the loss function tends to be stable or the number of training times reaches the maximum number of training times; The target flash memory characteristic prediction model is determined based on the parameters in the initial flash memory characteristic prediction model that is finally updated.

[0090] Optionally, the graph attention network module includes at least one graph attention layer; the training module 602 is further used for: Inputting the graph structure data into a first graph attention layer of the at least one graph attention layer included in the graph attention network module, obtaining a first weight between any node in each of the nodes output by the first graph attention layer and other nodes in the neighborhood; the first weight is used to update the graph structure data; Inputting the updated graph structure data into a second graph attention layer in the at least one graph attention layer, obtaining a second weight between any node in the nodes output by the second graph attention layer and other nodes in the neighborhood; the first weight is used to update the updated graph structure data; The execution is repeated until the finally updated graph structure data is input into the last graph attention layer of the at least one graph attention layer, and the target weight between any node in the nodes output by the last graph attention layer and other nodes in the neighborhood is obtained.

[0091] Figure 7 FIG. 1 is a schematic diagram of the structure of the flash memory characteristic prediction device provided by the present invention. Figure 7 As shown, the flash memory characteristic prediction device 700 includes: an acquisition module 701 and a prediction module 702; wherein,

[0092] An acquisition module 701 is used to acquire second target information of a target flash memory, where the second target information includes at least one of the following: manufacturing process information, electrical property information, and working condition information; Prediction module 702, used for inputting the second target information into a target flash memory characteristic prediction model to obtain characteristic parameters of the target flash memory output by the target flash memory characteristic prediction model; the target flash memory characteristic prediction model is based on Figure 6 The flash memory characteristic prediction model is trained using the flash memory characteristic prediction model training method shown.

[0093] The flash memory characteristic prediction device provided by the present invention obtains the second target information of the target flash memory, wherein the second target information includes at least one of the following: manufacturing process information, electrical property information, and working condition information; and inputs the second target information into the target flash memory characteristic prediction model to obtain the characteristic parameters of the target flash memory output by the target flash memory characteristic prediction model. Since the multi-dimensional characteristics of flash memories of different models and the complex relationships between them are considered when training the target flash memory characteristic prediction model, the flash memory characteristic prediction model can adaptively learn the characteristic differences between flash memories of different models, thereby enabling the trained flash memory characteristic prediction model to achieve accurate prediction of flash memory characteristics and improve the read and write efficiency and stability of SSDs.

[0094] Figure 8 is a schematic diagram of the physical structure of an electronic device provided by the present invention, such as Figure 8 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830 and a communication bus 840, wherein the processor 810, the communication interface 820 and the memory 830 communicate with each other through the communication bus 840. The processor 810 may call the logic instructions in the memory 830 to execute the flash memory characteristic prediction model training method, the method comprising: constructing graph structure data; the graph structure data comprises a plurality of nodes and edges between the nodes, the nodes are determined based on the first target information of at least one model of flash memory, the first target information comprises at least one of the following: manufacturing process information, electrical property information and working condition information; the edges represent the interaction relationship between the nodes; based on the graph structure data, the initial flash memory characteristic prediction model is trained to obtain the target flash memory characteristic prediction model.

[0095] The processor 810 can call the logic instructions in the memory 830 to execute the flash memory characteristic prediction method, which includes: obtaining second target information of the target flash memory, the second target information including at least one of the following: manufacturing process information, electrical property information and working condition information; inputting the second target information into the target flash memory characteristic prediction model to obtain the characteristic parameters of the target flash memory output by the target flash memory characteristic prediction model; the target flash memory characteristic prediction model is trained based on the flash memory characteristic prediction model training method described in any of the aforementioned embodiments.

[0096] In addition, the logic instructions in the above-mentioned memory 830 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0097] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the flash memory characteristic prediction model training method provided by the above-mentioned methods, the method including: constructing graph structure data; the graph structure data includes multiple nodes and edges between each of the nodes, the nodes are determined based on first target information of at least one model of flash memory, the first target information includes at least one of the following: manufacturing process information, electrical property information and working condition information; the edges represent the interaction relationship between each of the nodes; based on the graph structure data, the initial flash memory characteristic prediction model is trained to obtain a target flash memory characteristic prediction model.

[0098] When the computer program is executed by the processor, the computer can execute the flash memory characteristic prediction method provided by the above methods, the method comprising: obtaining second target information of the target flash memory, the second target information comprising at least one of the following: manufacturing process information, electrical property information and working condition information; inputting the second target information into a target flash memory characteristic prediction model to obtain characteristic parameters of the target flash memory output by the target flash memory characteristic prediction model; the target flash memory characteristic prediction model is trained based on the flash memory characteristic prediction model training method described in any of the above embodiments.

[0099] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the flash memory characteristic prediction model training method provided by the above-mentioned methods, the method comprising: constructing graph structure data; the graph structure data comprises multiple nodes and edges between each of the nodes, the nodes are determined based on first target information of at least one model of flash memory, the first target information comprises at least one of the following: manufacturing process information, electrical property information and working condition information; the edges represent the interaction relationship between each of the nodes; based on the graph structure data, the initial flash memory characteristic prediction model is trained to obtain a target flash memory characteristic prediction model.

[0100] When the computer program is executed by the processor, it is implemented to execute the flash memory characteristic prediction method provided by the above-mentioned methods, and the method includes: obtaining second target information of the target flash memory, and the second target information includes at least one of the following: manufacturing process information, electrical property information and working condition information; inputting the second target information into a target flash memory characteristic prediction model to obtain the characteristic parameters of the target flash memory output by the target flash memory characteristic prediction model; the target flash memory characteristic prediction model is trained based on the flash memory characteristic prediction model training method described in any of the above-mentioned embodiments.

[0101] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0102] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A flash memory characteristic prediction model training method, characterized in that: include: Constructing graph structure data; the graph structure data includes a plurality of nodes and edges between the nodes, the nodes are determined based on first target information of at least one model of flash memory, the first target information includes at least one of the following: manufacturing process information, electrical property information and working condition information; the edges represent the interaction relationship between the nodes; Based on the graph structure data, an initial flash memory characteristic prediction model is trained to obtain a target flash memory characteristic prediction model.

2. The flash memory characteristic prediction model training method according to claim 1, characterized in that: The constructing of graph structure data includes: Obtaining the first target information of each model of flash memory; Performing a characteristic test on the first target information of each model of flash memory to obtain characteristic parameters and characteristic test results during the characteristic test; Normalizing the first target information, the characteristic parameter, and the characteristic test result respectively to obtain normalized first target information, normalized characteristic parameter, and normalized characteristic test result; The graph structure data is constructed based on the normalized first target information, the normalized characteristic parameters and the normalized characteristic test results; the node is any one of the normalized first target information, the normalized characteristic parameters and the normalized characteristic test results.

3. The flash memory characteristic prediction model training method according to claim 1, characterized in that: The initial flash memory characteristic prediction model includes a graph attention network module and a fully connected layer; The initial flash memory characteristic prediction model is trained based on the graph structure data to obtain a target flash memory characteristic prediction model, including: Step A: inputting the graph structure data into the graph attention network module in the initial flash memory characteristic prediction model, and obtaining the target weight between any node in each of the nodes output by the graph attention network module and other nodes in the neighborhood; Step B: inputting each of the target weights into the full-link layer in the initial flash memory characteristic prediction model to obtain the characteristic parameter prediction value of each model of flash memory output by the full-link layer; Step C: Determine the target flash memory characteristic prediction model based on the characteristic parameter prediction values ​​of each model of flash memory and the actual values ​​of the characteristic parameters of each model of flash memory.

4. The flash memory characteristic prediction model training method according to claim 3, characterized in that: The step of determining the target flash memory characteristic prediction model based on the characteristic parameter prediction values ​​of each model of flash memory and the actual values ​​of the characteristic parameters of each model of flash memory comprises: Calculating a loss function based on the predicted values ​​of the characteristic parameters of each model of flash memory and the true values ​​of the characteristic parameters of each model of flash memory; If the cumulative value of the loss function does not tend to be stable or the number of training times does not reach the maximum number of training times, update the parameters in the initial flash memory characteristic prediction model, and repeat the above steps A to C until the cumulative value of the loss function tends to be stable or the number of training times reaches the maximum number of training times; The target flash memory characteristic prediction model is determined based on the parameters in the initial flash memory characteristic prediction model that is finally updated.

5. The flash memory characteristic prediction model training method according to claim 3, characterized in that: The graph attention network module includes at least one graph attention layer; The step of inputting the graph structure data into the graph attention network module in the initial flash memory characteristic prediction model to obtain the target weight between any node in each of the nodes output by the graph attention network module and other nodes in the neighborhood includes: Inputting the graph structure data into a first graph attention layer of the at least one graph attention layer included in the graph attention network module, obtaining a first weight between any node in each of the nodes output by the first graph attention layer and other nodes in the neighborhood; the first weight is used to update the graph structure data; Inputting the updated graph structure data into a second graph attention layer in the at least one graph attention layer, obtaining a second weight between any node in the nodes output by the second graph attention layer and other nodes in the neighborhood; the first weight is used to update the updated graph structure data; The execution is repeated until the finally updated graph structure data is input into the last graph attention layer of the at least one graph attention layer, and the target weight between any node in the nodes output by the last graph attention layer and other nodes in the neighborhood is obtained.

6. A flash memory characteristic prediction method, characterized in that: include: Acquire second target information of the target flash memory, wherein the second target information includes at least one of the following: manufacturing process information, electrical property information, and working condition information; The second target information is input into a target flash memory characteristic prediction model to obtain characteristic parameters of the target flash memory output by the target flash memory characteristic prediction model; the target flash memory characteristic prediction model is trained based on the flash memory characteristic prediction model training method described in any one of claims 1 to 5.

7. A flash memory characteristic prediction model training device, characterized in that: include: A construction module, configured to construct graph structure data; the graph structure data includes a plurality of nodes and edges between the nodes, the nodes are determined based on first target information of at least one model of flash memory, the first target information includes at least one of the following: manufacturing process information, electrical property information, and working condition information; the edges represent the interaction relationship between the nodes; The training module is used to train the initial flash memory characteristic prediction model based on the graph structure data to obtain a target flash memory characteristic prediction model.

8. A flash memory characteristic prediction device, characterized in that: include: An acquisition module, used for acquiring second target information of a target flash memory, wherein the second target information includes at least one of the following: manufacturing process information, electrical property information, and working condition information; A prediction module, used to input the second target information into a target flash memory characteristic prediction model to obtain the characteristic parameters of the target flash memory output by the target flash memory characteristic prediction model; the target flash memory characteristic prediction model is trained based on the flash memory characteristic prediction model training method described in any one of claims 1 to 5.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the flash memory characteristic prediction model training method as described in any one of claims 1 to 5 is implemented, or the flash memory characteristic prediction method as described in claim 6 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the flash memory characteristic prediction model training method as described in any one of claims 1 to 5 is implemented, or the flash memory characteristic prediction method as described in claim 6 is implemented.

Images