Memory failure prediction method and apparatus, computing chip, and computing device

Abstract

The present disclosure relates to a memory failure prediction method, device, computing chip and computing device, and belongs to the technical field of computers. The method predicts the probability of a target failure of a memory by using a memory failure prediction model, and comprehensively considers error information, error position information and running environment information and other factors during the running of the memory in the prediction process, thereby improving the accuracy of memory failure prediction.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to a memory fault prediction method, apparatus, computing chip, and computing device. Background Technology

[0002] Dynamic Random Access Memory (DRAM), NAND Flash, and other types of memory can be used as memory in computing systems. With the continuous evolution of semiconductor technology, the capacity and integration of memory have been significantly improved. As a result, memory failure problems have become increasingly serious, becoming a key factor affecting the reliability and data integrity of computing systems.

[0003] Errors occurring during memory operation (such as correctable errors) and the memory operating environment can both lead to memory failures. Therefore, during memory operation, it's possible to collect information about memory errors (i.e., error messages) and the memory operating environment (i.e., operating environment information). Based on this information, the probability of memory failure can be predicted. For example, currently, simple data models (such as linear models or shallow decision trees) are typically used to process the error and operating environment information during memory operation to predict the probability of memory failure.

[0004] However, due to the relatively simple data model used, it is difficult to effectively capture the complex relationships between multiple features of error information and runtime environment information. The feature dimensions of error information and runtime environment information are also relatively few, resulting in inaccurate memory fault prediction using the data model. Summary of the Invention

[0005] This disclosure provides a method, apparatus, computing chip, and computing device for memory fault prediction, which can improve the accuracy of memory fault prediction. The technical solution of this disclosure is as follows: According to a first aspect of the present disclosure, a memory fault prediction method is provided, the method comprising: Obtain error information, error location information, and runtime environment information during the first memory operation; Feature extraction is performed on the error information, the error location information, and the operating environment information to obtain multi-dimensional features. The multi-dimensional features include at least one feature of the error information, at least one feature of the error location information, and at least one feature of the operating environment information. Based on the memory failure prediction model and the multidimensional features, a first probability and a first risk contribution of each feature in the multidimensional features are determined. The first probability is the probability of the first memory experiencing a target failure, predicted by the memory failure prediction model based on the multidimensional features. The first risk contribution of any feature is the risk contribution of the feature to the target failure, predicted by the memory failure prediction model based on the multidimensional features.

[0006] Optionally, determining the first probability and the first risk contribution of each feature in the multidimensional features based on the memory fault prediction model and the multidimensional features includes: The memory fault prediction model processes the multidimensional features of the third sample to obtain multiple second probabilities and multiple second risk contribution degrees for each feature in the multidimensional features. The second probability is the probability of the first memory experiencing the target fault, predicted based on some features in the multidimensional features. Each second risk contribution degree of the feature is the risk contribution degree of the feature to the target fault, predicted based on some features in the multidimensional features. The memory fault prediction model is used to process the multiple second probabilities to obtain the first probability; The memory fault prediction model processes multiple second risk contribution values ​​for each feature in the multidimensional features to obtain the first risk contribution value for each feature.

[0007] Optionally, the first probability is the average or weighted sum of the plurality of second probabilities, and the first risk contribution of any feature is the average or weighted sum of the plurality of second probabilities of the feature.

[0008] Optionally, the memory fault prediction model is a random forest model, which includes at least one first decision tree and at least one second decision tree; The first decision tree corresponds to a split feature in the multidimensional features. The first decision tree is constructed based on the split feature and is used to analyze the impact of the split feature on the target fault. The second decision tree corresponds to multiple split features in the multidimensional features. The second decision tree is constructed based on the multiple split features. The second prediction tree is used to analyze the combined impact of the multiple split features on the target fault.

[0009] Optionally, the step of processing the multidimensional features of the third sample through the memory fault prediction model to obtain multiple second probabilities and multiple second risk contribution values ​​of each feature in the multidimensional features includes: The first decision tree is used to process the multidimensional features of the third sample to obtain a second probability and a second risk contribution distribution. The second probability is the probability of the first memory causing the target failure, which is predicted based on the splitting feature corresponding to the first decision tree. The second risk contribution distribution includes the second risk contribution of each feature in the multidimensional features. The second risk contribution of any feature is the risk contribution of the feature to the target failure under the influence of the splitting feature corresponding to the first decision tree. The second decision tree is used to process the multidimensional features of the third sample to obtain another second probability and another second risk contribution distribution. The other second probability is the probability of the first memory experiencing the target failure, predicted based on multiple split features corresponding to the first decision tree. The other second risk contribution distribution includes the second risk contribution of each feature in the multidimensional features. The second risk contribution of any feature is the risk contribution of the feature to the target failure under the influence of multiple split features corresponding to the first decision tree.

[0010] Optionally, the random forest model includes multiple first decision trees, each first decision tree corresponding to at least one splitting feature, and the method further includes: Based on the multiple second risk contribution distributions obtained from the multiple first decision trees, the third risk contribution of the at least one splitting feature is determined. The third risk contribution of any splitting feature refers to the overall risk contribution of the splitting feature to the target failure without the influence of other features. Output the third risk contribution of the at least one splitting feature.

[0011] Optionally, the dimensions of the multidimensional feature are related to the memory type of the first memory.

[0012] Optionally, the method further includes: Based on the first risk contribution of each feature in the multidimensional features, a fault warning is given for the first memory.

[0013] Optionally, the method further includes: A heatmap is displayed based on the first risk contribution of each feature in the multidimensional features, and the heatmap is used to reflect the first risk contribution of each feature in the multidimensional features.

[0014] Optionally, the method further includes: Multiple first samples are obtained. The first samples include error information, error location information, and runtime environment information of the second memory during operation. The second memory and the first memory have the same memory type. Based on the multiple first samples, the preset model is trained to obtain the memory fault prediction model.

[0015] Optionally, the method further includes: Multiple second samples are obtained, including error information, error location information, and runtime environment information of the third memory during operation. The third memory and the first memory have the same memory type. Based on the multiple second samples, the memory fault prediction model is incrementally trained.

[0016] According to a second aspect of the present disclosure, a memory fault prediction apparatus is provided, the memory fault prediction apparatus including a functional module for implementing the memory fault prediction method provided in any possible implementation of the first aspect described above.

[0017] According to a second aspect of the present disclosure, a memory fault prediction apparatus is provided, the apparatus comprising: The first acquisition unit is used to acquire error information, error location information, and runtime environment information of the first memory during operation; An extraction unit is used to extract features from the error information, the error location information, and the operating environment information to obtain multi-dimensional features, wherein the multi-dimensional features include at least one feature of the error information, at least one feature of the error location information, and at least one feature of the operating environment information. The first determining unit is configured to determine a first probability and a first risk contribution of each feature in the multidimensional features based on the memory fault prediction model and the multidimensional features. The first probability is the probability of the first memory experiencing a target fault as predicted by the memory fault prediction model based on the multidimensional features, and the first risk contribution of any feature is the risk contribution of the feature to the target fault as predicted by the memory fault prediction model based on the multidimensional features.

[0018] According to a third aspect of the present disclosure, a computing chip is provided, the computing chip comprising: One or more processors; One or more memories for storing the one or more processor-executable instructions; The one or more processors are configured to perform the memory fault prediction method in any of the possible implementations of the above-described aspects.

[0019] According to a fourth aspect of the present disclosure, a computing device is provided, the computing device including a memory and a computing chip provided in the third aspect.

[0020] According to a fifth aspect of the present disclosure, a computer-readable storage medium is provided such that, when at least one instruction in the computer-readable storage medium is executed by one or more processors of a computing chip, the computing chip is enabled to perform the fault prediction method in any possible implementation of the first aspect described above.

[0021] According to a sixth aspect of the present disclosure, a computer program product is provided, including one or more instructions that can be executed by one or more processors of a computing chip, enabling the computing chip to perform the fault prediction method in any possible implementation of the first aspect described above.

[0022] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0023] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0024] Figure 1 This is a schematic diagram illustrating a computing device according to an exemplary embodiment; Figure 2 This is a flowchart illustrating a model training method according to an exemplary embodiment; Figure 3 This is a schematic diagram illustrating the training of a random forest model according to an exemplary embodiment; Figure 4 This is a flowchart illustrating a memory fault prediction method according to an exemplary embodiment; Figure 5 This is a flowchart illustrating a memory fault prediction system according to an exemplary embodiment; Figure 6 This is a structural block diagram of a memory fault prediction device according to an exemplary embodiment. Detailed Implementation

[0025] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0026] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0027] It should be noted that all information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in this application have been authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the logs involved in this application were all obtained with full authorization.

[0028] In response, it is used to indicate the conditions or states on which the operation performed depends. When the conditions or states on which it depends are met, one or more operations performed can be performed in real time or with a set delay. Unless otherwise specified, there is no restriction on the order in which the multiple operations are performed.

[0029] In some embodiments, the meaning of A and / or B includes three cases: A and B, and A and B.

[0030] This application provides a memory fault prediction method, which is applied to computing devices. Figure 1 This is a schematic diagram illustrating a computing device according to an exemplary embodiment. Figure 1 The computing device 100 shown refers to a device with computing capabilities. The computing device 100 can be a server or a computing device similar to a server. The computing device can also be a terminal, including but not limited to personal computers (such as laptops, tablets, desktop computers, ultrabooks, etc.), devices with network product interface designs and functions, mobile phones, smart home devices, wearable devices, artificial intelligence devices, smart cars, smart instruments, Internet of Things terminals, or extended reality devices, etc.

[0031] like Figure 1As shown, the computing device 100 includes a bus 101, a memory 102, a computing chip 103, and a communication interface 104. The various components in the computing system 100 communicate via the bus 101. The bus 101 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 1 The symbol is represented by only one line, but this does not mean that there is only one bus or one type of bus.

[0032] The memory 102 may include at least one type of memory, namely volatile memory and non-volatile memory. The volatile memory includes, but is not limited to, Random Access Memory (RAM), Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), and Double Data Rate Synchronous Dynamic Random Access Memory (DDR SDRAM). The non-volatile memory includes, but is not limited to, Read-Only Memory (ROM), NAND Flash, Hard Disk Drive (HDD), or Solid State Drive (SSD).

[0033] The computing chip 103 refers to a chip with computing capabilities. The computing chip 103 includes at least one processor 31 and a memory 32. The processor 31 can be a Central Processing Unit (CPU) or an X Processing Unit (XPU). The XPU includes, but is not limited to, a Graphics Processing Unit (GPU), a Neural Processing Unit (NPU), and a Data Processing Unit (DPU). The memory 32 is the RAM or cache within the computing chip 103.

[0034] The memory in memory 102 stores executable program code, and the processor 31 reads and executes the executable program code, causing the computing device 100 to perform the model training method and / or memory fault prediction method provided in this application.

[0035] In some embodiments, memory 102 is the main memory of a computing device. In some embodiments, memory 102 is the secondary memory of the computing device 100, which also includes main memory, including at least one of the volatile and non-volatile memory described above. It is used to provide data storage services for the computing device.

[0036] The aforementioned memory may include at least one memory chip, each memory chip includes at least one die, each die includes at least one logical bank, each logical bank is an array of storage cells consisting of multiple rows and columns of storage units, each row of storage units in the logical bank forms a storage row, and each column of storage units in the logical bank forms a storage column.

[0037] The physical bus connecting memory is called a memory channel. Memory can be connected to multiple memory channels, and each memory channel is used to communicate with at least one logical storage block in memory. Each memory channel includes an address bus, a data (DQ) bus, and a command bus.

[0038] Memory may malfunction during operation. Memory malfunction refers to the phenomenon that the memory cannot read and write data correctly according to the design specifications, resulting in data errors, access abnormalities, or system failure.

[0039] Memory failures are mainly divided into transient errors and permanent errors. Transient errors are usually caused by factors such as high-energy particle radiation, power supply noise, or signal interference, and manifest as single-bit flips or multi-bit errors. They are sporadic and non-repeatable. Permanent errors are mostly caused by physical aging or hardware damage such as manufacturing defects, electromigration, and dielectric breakdown. They often present repeatable error patterns with fixed addresses and fixed cells, such as row failures, column failures, and bit line short circuits.

[0040] Based on the error correction capabilities of Error Check and Correction Codes (ECC), the aforementioned errors can be further categorized into Correctable Errors (CE) and Uncorrectable Errors (UCE). Single-bit transient errors and recurring errors caused by early minor hardware damage typically fall under the category of CE, which can be automatically corrected by ECC logic and do not affect normal system operation. However, multi-bit burst errors, severe transient interference, and permanent hardware failures that have deteriorated often exceed the error correction range of ECC and thus form UCEs. UCEs can cause the system to be unable to obtain correct data, compromising the reliability of instruction execution and data processing, thereby triggering the abnormal protection mechanisms of computing devices, ultimately leading to system crashes, restarts, or operational failures.

[0041] Memory failures that can cause computing devices to crash, restart, or malfunction are called target failures. These target failures include, but are not limited to, UCEs (Unique Memory Errors) and permanent errors. Target failures can have a significant impact on the reliability and availability of computing devices. Therefore, it is crucial to predict target failures in memory in a timely manner and implement appropriate strategies to minimize their impact on computing devices.

[0042] Several factors can trigger memory failures, including physical location, operating environment, and workload. Regarding physical location, the distribution of errors across different dies, banks, memory queues, rows, and columns within different memory modules may be uneven; for example, certain banks or queues may be more prone to concentrated errors. Regarding the operating environment, changes in external conditions such as temperature, voltage, and electromagnetic interference can significantly affect the probability of memory errors; for instance, high temperatures or voltage fluctuations can increase the error probability. Regarding workload, high-frequency access and prolonged operation can accelerate the accumulation of errors in memory; in other words, heavy loads and long operating times can also accelerate error exposure.

[0043] Based on this, embodiments of this application can train a machine learning model using error information, error location information, operating environment information, and workload information during memory operation to obtain a memory prediction model. During training, based on the learning mechanism of the machine learning model, the correlation between the target fault and multi-dimensional features such as physical location features, operating environment features, and workload features is fully explored, improving the ability to identify complex error patterns (such as multiple error types), the ability to identify target faults, and the accuracy of prediction. Specifically, the memory prediction model is used to predict the probability of a target fault occurring in memory and the risk contribution of each feature in the multi-dimensional features to the target fault.

[0044] Below, in conjunction with Figure 2The process of training a memory prediction model will be introduced.

[0045] Figure 2 This is a schematic diagram illustrating a model training method according to an exemplary embodiment. The model training method is applied to a first computing device, which may be the computing device 100 described above. The method includes the following steps.

[0046] 201. The first computing device acquires the first training set.

[0047] The first training set includes multiple first samples, and step 201 is that the first computing device acquires multiple first samples.

[0048] Each first sample includes error information, error location information, operating environment information, and workload information of the second memory during operation. The second memory can be the memory of the first computing device, or the memory of another computing device.

[0049] In this application, each sample corresponds to a time period, and each sample includes error information, error location information, runtime environment information, and workload information of a certain memory during the operation of that time period.

[0050] This error message indicates an error that occurred in the memory during the specified time period. The error can be a recoverable error and / or a transient error, such as a 1-bit error, a 2-bit error, a multi-bit error (more than 2 bits), a burst error, or a recurring error. The error message includes the type of error that occurred, which includes, but is not limited to, 1-bit errors, 2-bit errors, multi-bit errors (more than 2 bits), burst errors, or recurring errors. The error message also includes the time information when the error occurred.

[0051] Error location information indicates the location (i.e., the error location) where the memory encountered the error during the operation of that time period. Error location information includes at least one of the following: the identifier of the wafer to which the error location belongs, the identifier of the logical memory, the identifier of the data bus, the identifier of the memory row, the identifier of the memory column, and the identifier of the memory channel.

[0052] The runtime environment information is used to indicate the runtime environment of the memory during that time period. This runtime environment information includes at least one of the memory's voltage and temperature.

[0053] The workload information indicates the load on the memory during the operation of the memory within a certain time period. The workload information includes at least one of the memory access frequency, working duration, and read / write counts during the time period, wherein the read / write counts can be the number of times the memory is read and written during the time period. In some embodiments, the workload information includes the workload information of each storage row and / or each storage column in the memory, and the workload information of any storage row (or storage column) includes at least one of the storage row (or storage column) access frequency, working duration, and read / write counts during the time period.

[0054] When the second memory is the memory of the first computing device, the first computing device can read multiple error logs and multiple runtime logs from the second memory. The error logs read can be fault logs from the past month, two months, or one year, without any specific limitation. Similarly, the runtime logs read can be runtime logs from the past month, two months, or one year, without any specific limitation.

[0055] The fault log typically records the location information of the error that occurred in the secondary memory (i.e., error location information), ECC reporting information, and a timestamp. The ECC reporting information includes the error bit length, error code, and error type. The runtime log typically records the secondary memory's operating environment information, workload information, and a timestamp.

[0056] The first computing device generates multiple first samples based on multiple error logs and multiple runtime logs in the second memory. For example, it obtains error information, error location information, runtime environment information, and workload information from error logs and runtime logs within the same time period, and generates first samples based on the obtained error information, error location information, runtime environment information, and workload information.

[0057] In some embodiments, the first training set further includes a label corresponding to each first sample. In this application, the label of any sample is used to indicate the error type or target probability range of the error indicated by the sample, wherein the target probability range is the range of probabilities that the error will cause a target failure in the second memory. The label corresponding to the first sample can be configured by the user or set by the first computing device based on certain rules.

[0058] In some embodiments, where the second memory is the memory of the first computing device, such as Figure 3 As shown, the first computing device also includes a data acquisition module, which is used to collect the operation logs and error logs of the second memory in real time during the operation of the second memory, and generate multiple first samples based on the collected operation logs and error logs, and then generate a first training set based on the multiple first samples.

[0059] In the case where the second memory is not the memory of the first computing device, for example, when the second memory is the memory of a computing device other than the first computing device, the first computing device obtains multiple error logs and multiple runtime logs of the second memory from the other computing device, and then obtains the first training set based on the multiple error logs and multiple runtime logs of the second memory.

[0060] 202. The first computing device performs data preprocessing on the first training set.

[0061] by Figure 3 For example, the first computing device also includes a data preprocessing module, which cleans the data of each first sample in the first training set to remove invalid or erroneous data from the first samples and handle missing values, thereby ensuring the accuracy of subsequent analysis and modeling.

[0062] For each first sample after data cleaning, the data preprocessing module extracts features to obtain multidimensional features. These multidimensional features include at least one feature of error information, at least one feature of error location information, at least one feature of operating environment information, and at least one feature of workload information. The error location information features include location features, indicating the location where the error occurred. These features include at least one of error type features, error statistical features, error pattern distribution features, and time series features. Error type features are characteristics of the error type; error statistical features are statistical characteristics of the errors, such as the number of errors occurring per unit time; error pattern distribution features indicate the distribution of various error types; and time series features, such as consecutive error intervals, indicate the interval between consecutive error occurrences. The operating environment information features include at least one of the memory's temperature range and voltage fluctuation range over a period of time. The workload information features include at least one of the memory's runtime and read / write ratio, where the read / write ratio is the proportion of memory reads and writes.

[0063] For example, the first computing device performs feature extraction on the error information in the first sample to obtain at least one feature of the error information; performs feature extraction on the error location information in the first sample to obtain at least one feature of the error location information; performs feature extraction on the operating environment information in the first sample to obtain at least one feature of the operating environment information; and performs feature extraction on the workload information in the first sample to obtain at least one feature of the workload information.

[0064] It should be understood that each feature in the multidimensional features is a feature of one dimension. The embodiments of this application fully consider the features of various dimensions of memory when constructing the memory fault prediction model, thereby making the memory fault prediction model constructed based on these multidimensional features more accurate. This multidimensional feature fusion can more comprehensively capture the complex patterns of memory errors, and training the memory fault prediction model based on such multidimensional features can significantly improve the model's ability to represent errors.

[0065] After obtaining the multidimensional features of each first sample, the data preprocessing module performs feature encoding on the multidimensional features of each first sample to obtain the feature vector of each first sample. The feature vector is a quantized multidimensional feature. For example, for any multidimensional features of a first sample, the error location information feature among the multiple multidimensional features is hash-encoded to obtain the feature vector corresponding to the error location information. This feature vector is the identifier of the error location. Each feature of the error information among the multiple multidimensional features is one-hot encoded to obtain the feature vector corresponding to the error information. Each feature of the runtime environment information among the multiple multidimensional features is standardized to obtain the feature vector corresponding to the runtime environment information. Each feature of the workload information among the multiple multidimensional features is standardized to obtain the feature vector corresponding to the workload information. The feature vector corresponding to the error location information, the feature vector corresponding to the runtime environment information, and the feature vector corresponding to the workload information are concatenated to obtain the feature vector of the first sample.

[0066] 203. The first computing device trains the preset model based on the first training set after data preprocessing to obtain a memory fault prediction model.

[0067] The preset model can be a machine learning model with an ensemble learning mechanism (i.e., an ensemble learning model). For example, the preset model is a random forest model. The preset model can also be other ensemble learning models besides random forest models, such as Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Adaptive Boosting (AdaBoost) model, and Categorical Boosting (CatBoost), etc.

[0068] A memory fault prediction model is a mathematical or logical framework describing the probability distribution and patterns of errors occurring in memory. This application utilizes an ensemble learning model to construct this model to predict memory erroneous behavior under different conditions. The memory fault prediction model is used to predict the probability of a target memory fault occurring based on multidimensional features of the input, and to predict the risk contribution of each feature in the multidimensional features to the target fault. Each feature has a certain risk of causing a target memory fault, or in other words, each feature has a certain influence on the occurrence (or initiation) of the target fault. The risk contribution of any feature to the target fault indicates the degree of influence of that feature on the target fault event, or the magnitude of the risk that the feature causes the target fault. The target fault event refers to the event in which the target memory fault occurs. For ease of description, the risk contribution of any feature to the target fault is simply referred to as the feature's risk contribution.

[0069] Taking the default model as a random forest model as an example, the random forest model includes multiple decision trees. The model parameters of the random forest model include the number of decision trees (n_estimators), the maximum tree depth (max_depth), the size of the feature subset (max_features), and the weights of each feature in the feature subset. The value of n_estimators ranges from 100 to 500, but n_estimators can also have other value ranges. The value of max_depth ranges from 8 to 15, but max_depth can also have other value ranges. max_features = sqrt(n), where n is the total number of features, that is, the total dimension of the features.

[0070] The first computing device uses a random forest model to evaluate the contribution of each feature in the aforementioned multidimensional features, obtaining the contribution of each feature. The contribution of each feature indicates its contribution to reducing the uncertainty of fault classification during the entire decision tree partitioning process of the random forest model. For example, the first computing device inputs the multidimensional features (such as feature vectors) of each first sample into the random forest model. The random forest model evaluates the contribution of each feature in the multidimensional features based on each first sample, obtaining the contribution of each feature and outputting the contribution of each feature. For instance, the contribution of each feature can be determined using the Gini index in the decision tree of the random forest model. The Gini index is a commonly used node splitting metric in the decision tree of the random forest model, used to measure the contribution of a feature to improving class purity when partitioning samples. Based on the Gini index reduction brought by this feature in each decision tree, the importance of the feature can be further calculated, and the feature importance can be used as the feature's contribution. By quantifying the contribution of each feature, data support is provided for the model to accurately locate error patterns (such as the location, number, and type of error addresses).

[0071] The first computing device sorts the multidimensional features based on the contribution of each feature, obtaining a sorted multidimensional feature set. The X features with the highest contribution from this sorted set are identified as key features. Based on the X key features of multiple first samples, the multiple first samples themselves, and their labels, the random forest model is retrained. The model parameters are then adjusted according to the X key features to obtain the trained random forest model. Here, X is an integer greater than 1, and X is less than the total dimension of the multidimensional features. The output of the random forest model includes the predicted probability of a target failure occurring in the second memory location and the risk contribution of each feature in the multidimensional feature set.

[0072] Each decision tree in a random forest model consists of multiple nodes, categorized as root, intermediate, and leaf nodes. Specifically, each decision tree includes one root node, at least one intermediate node, and multiple leaf nodes. Nodes in a decision tree can form decision paths; each decision tree includes at least one decision path, and each decision path includes the root node, at least one intermediate node, and one leaf node. Different decision paths share the same root node, but different intermediate or leaf nodes exist within each path. The features used to partition samples and split nodes during decision tree construction are called splitting features. Each decision tree corresponds to at least one splitting feature. The number and types of splitting features may differ between different decision trees. Each decision tree is constructed based on its corresponding at least one splitting feature. Nodes used for sample partitioning in a decision tree are called splitting nodes, such as the root node and some intermediate nodes. A splitting node contains a splitting feature corresponding to the decision tree, a splitting threshold for that feature, and a splitting criterion. The splitting criterion (i.e., the splitting rule) is a judgment condition composed of the splitting feature and its corresponding splitting threshold. For example, the four features in the multidimensional features—temperature, number of CEs, storage row error density, and voltage drop—are all splitting features corresponding to a certain decision tree. This decision tree includes splitting nodes 1 to 4. In splitting node 1, the splitting feature is temperature, the splitting threshold for temperature is 85 degrees Celsius (°C), and the splitting criterion is: temperature > 85°C. In splitting node 2, the splitting feature is number of CEs, the splitting threshold for number of CEs is 100, and the splitting criterion is: number of CEs > 100. In splitting node 3, the splitting feature is storage row error density, the splitting threshold for storage row error density is 0.8, and the splitting criterion is: storage row error density > 0.8. In splitting node 4, the splitting feature is voltage drop, the splitting threshold for voltage drop is 50 millivolts (mV), and the splitting criterion is: voltage drop > 50 mV.

[0073] A random forest model consists of multiple decision trees, including at least one first decision tree and at least one second decision tree. The first decision tree corresponds to one splitting feature, while the second decision tree corresponds to multiple splitting features. In other words, the first decision tree is a decision tree corresponding to one splitting feature, and the second decision tree is a decision tree corresponding to multiple splitting features. The first decision tree is used to analyze the impact of a single feature on the target failure, and the second decision tree is used to analyze the impact of multiple features combined on the target failure. The single feature is the splitting feature corresponding to the first decision tree, and the multiple features are the multiple splitting features corresponding to the second decision tree.

[0074] The training process of random forest will be introduced below.

[0075] During training, for each decision tree, the computing device randomly draws samples with replacement from the first training set to obtain a sample subset corresponding to each decision tree. This sample subset includes multiple first samples, which are a subset of samples from the first training set. For any first decision tree, all features in the multidimensional features other than the splitting feature corresponding to that first decision tree are called non-splitting features corresponding to that first decision tree. In the sample subset corresponding to that first decision tree, the same non-splitting feature has the same feature value across all first samples, while the splitting feature has different feature values ​​across all first samples (or, it is permissible for the splitting feature to have the same feature value in a small number of first samples). Before training the first decision tree, the computing device can, based on the splitting feature corresponding to that first decision tree, select multiple first samples from the first training set that have the same splitting feature value but different feature values ​​for other features, and combine these multiple first samples to form the first training set corresponding to the first decision tree.

[0076] Because the same non-split feature has the same feature value across all first samples in the sample subset corresponding to the first decision tree, the impact of each non-split feature on the target fault is relatively small when constructing the first decision tree. This allows the constructed first decision tree to analyze the impact of the split feature (i.e., a single feature) corresponding to the first decision tree on the target fault. For any second decision tree, in the sample subset corresponding to the second decision tree, each feature of the multidimensional feature has a different feature value across all first samples, or in the sample subset corresponding to the second decision tree, each feature of the multidimensional feature has a different feature value across a large number of first samples, allowing each feature to have the same feature value across a small number of first samples.

[0077] A subset of samples extracted for each decision tree is used as independent training data for that decision tree to ensure the diversity between decision trees. The computing device trains a single decision tree based on the corresponding sample subset. For example, for any decision tree, the computing device inputs the corresponding sample subset of the decision tree into the root node. The root node determines at least one usable splitting feature from at least one splitting feature corresponding to the decision tree. Based on each usable splitting feature and all the first samples input into the root node, a target splitting feature and the corresponding splitting threshold are determined. Here, the target splitting feature is a usable splitting feature, and the target splitting feature and the corresponding splitting threshold maximize the purity gain. The purity gain refers to the increase in sample purity after the samples in the node are split. The root node constructs a splitting criterion based on the target splitting feature and the corresponding splitting threshold. According to this criterion, the first sample in the root node is split into two new sample subsets, thus dividing the samples in the root node and generating two child nodes. These two new sample subsets are then input into these two child nodes, causing node splitting at the root node. Both child nodes of the root node are intermediate nodes. Following the sample splitting method in the root node, the intermediate nodes continuously split the input sample subsets, further subdividing the sample space. For any intermediate node on any decision path in the decision tree, when the intermediate node meets the stopping growth condition, it is designated as a leaf node on that decision path. The stopping growth condition includes, but is not limited to, the number of samples in the intermediate node (i.e., the current node) being less than a preset number of samples; the decision tree depth reaching the maximum tree depth; the purity gain from further splitting being less than the gain threshold; or, there are no remaining usable splitting features. Each leaf node can input multiple first samples. Each leaf node determines the memory fault probability and misclassification information based on the multiple input first samples and their labels. The memory failure probability is the probability of a target memory failure predicted based on some features in this multidimensional feature set (such as the splitting feature corresponding to the decision tree). Error classification information includes error labels corresponding to multiple error types. These labels indicate whether an error of the corresponding error type has occurred in the memory. The first label indicates that an error of the corresponding error type has occurred, while the second label indicates that an error of the corresponding error type has not occurred. These error types include, but are not limited to, 1-bit errors, 2-bit errors, multi-bit errors (more than 2 bits), burst errors, or recurring errors, allowing for flexible handling of new error types (such as bit flips caused by Rowhammer attacks).

[0078] Each leaf node is also used to store its own determined memory failure probability and error classification information.

[0079] Each leaf node is also used to generate a risk contribution distribution based on the decision path to which it belongs. The risk contribution distribution includes the risk contribution of each feature in the multidimensional features. The risk contribution of any feature is based on the number of splits of the feature on the decision path to which the leaf node belongs, the change in failure probability, the feature importance, the SHapley Additive exPlanations (SHAP) value, and the path contribution. The number of splits is based on the number of times the feature splits on the decision path. The change in failure probability indicates the change in failure probability caused by the splits of the feature on the decision path. The SHAP value is the precise quantitative contribution of the feature to the final predicted memory failure probability in the decision path traversed by a single sample. The path contribution refers to the influence weight of the feature on the final output memory failure probability of the decision path. It should be understood that when the feature is a splitting feature, the number of splits and the change in failure probability corresponding to the feature are both greater than 0. When the feature is not a splitting feature in the decision path (i.e., the feature is a non-splitting feature), the number of splits and the change in failure probability corresponding to the feature are both equal to 0. Therefore, in the risk contribution distribution generated at the same leaf node, the risk contribution of any splitting feature is greater than the risk contribution of a non-splitting feature. For example, when the decision tree is a first decision tree, in the risk contribution distribution generated at the same leaf node, the risk contribution of the splitting feature corresponding to this decision tree is significantly greater than the risk contribution of other features. When the decision tree is a second decision tree, in the risk contribution distribution generated at the same leaf node, the risk contribution of multiple splitting features corresponding to this decision tree is greater than the risk contribution of other features.

[0080] It should be understood that during the construction of the first decision tree, when each splitting node determines a usable splitting feature from the splitting features corresponding to the first decision tree, it can directly use one of the classification features corresponding to the first decision tree as a usable classification feature, which can be achieved by using a random selection function that overrides its default. When selecting a usable splitting feature from the splitting features corresponding to the first decision tree using the random selection function, since there is only one splitting feature corresponding to the first decision tree, no matter how many times the random selection function is used, the only usable splitting feature selected will be the one corresponding to the first decision tree. The first decision tree splits and grows based on this one splitting feature, thus becoming a tree that can analyze the independent influence of this one splitting feature. Furthermore, in the training subset input to the first decision tree, all features except the splitting feature corresponding to the first decision tree have the same feature value (i.e., the feature values ​​of other features are constant). During the construction of the first decision tree, the splitting feature corresponding to the first decision tree plays a dominant role, while the other features have very little effect and can be basically ignored. The risk contribution of each feature in the risk contribution distribution output by the first decision tree is the risk contribution of that feature to the target failure under the influence of the splitting feature corresponding to the first decision tree. Therefore, the risk contribution distribution output by the first decision tree is the influence of the splitting feature (i.e., a single feature) corresponding to the first decision tree analyzed by the first decision tree on the target failure. The first decision tree can analyze the influence of the splitting feature (i.e., a single feature) corresponding to the first decision tree on the target failure. During the construction of the second decision tree, the same feature in the training subset input to the first decision tree has a large number of different feature values, and the nodes are split using multiple split features corresponding to the second decision tree. Therefore, during the construction of the second decision tree, the multiple split features corresponding to the second decision tree play a dominant role, while other features play an auxiliary role. The risk contribution of each feature in the risk contribution distribution output by the second decision tree is the risk contribution of that feature to the target failure under the influence of the multiple split features corresponding to the first decision tree. Therefore, the risk contribution distribution output by the second decision tree is the combined impact of the multiple split features (i.e., multiple features) on the target failure as analyzed by the second decision tree. Thus, the second decision tree can analyze the combined impact of the multiple split features (i.e., multiple features) corresponding to the first decision tree on the target failure.

[0081] Each leaf node also outputs its own determined memory failure probability, misclassification information, and risk contribution distribution. A decision tree is complete when each decision path in the tree has a leaf node, and the leaf nodes can output memory failure probability, misclassification information, and risk contribution distribution.

[0082] Repeat the above sampling and decision tree construction process to generate a specified number of decision trees, forming a random forest model. Each decision tree has independent feature partitioning logic and leaf node probability distribution.

[0083] The first computing device evaluates the prediction accuracy of the trained random forest model using a validation sample set. If the prediction accuracy of the trained random forest model is greater than a preset threshold, the training of the random forest model is considered complete. The trained random forest model is an in-memory prediction model.

[0084] In some embodiments, when the memory fault prediction model is a random forest model, after the random forest model is trained, a control module is set in the random forest model. The control model is used to integrate the output results of each decision tree to obtain the final output result of the random forest model and output the final output result.

[0085] Optionally, the control module is further configured to, for any feature among the multidimensional features, statistically analyze the splitting thresholds corresponding to that feature when used as a splitting feature in all decision trees, obtaining multiple thresholds corresponding to that feature, and selecting multiple candidate thresholds from these multiple splitting thresholds. The candidate thresholds can be splitting thresholds whose total occurrence frequency in all decision trees exceeds a preset frequency (i.e., high-frequency thresholds), or splitting thresholds that cause the memory failure probability change to exceed a preset change (i.e., high-risk change thresholds). The memory failure probability change indicates the degree to which the predicted memory failure probability changes. When the feature value reaches the high-frequency threshold or the high-risk change threshold, the risk of a target memory failure increases. Based on this, the control module determines the target value range corresponding to the feature based on the multiple candidate thresholds and outputs the target value range corresponding to the feature. The minimum value in the target value range is the minimum threshold among the multiple candidate thresholds, and the maximum value in the target value range is the maximum threshold among the multiple candidate thresholds. The target value range corresponding to the feature is the range of values ​​within which the feature increases the probability of a target failure; that is, when the feature value is within the target value range, the probability of a target failure increases.

[0086] Similarly, the control module can determine and output the target value range for each feature in the multi-dimensional features. By outputting the target value range for each feature, users can maintain memory of the same type as the second memory based on the target value range for each feature, improving memory maintenance efficiency and reducing the risk of memory failure. For example, if the feature value of a certain feature of the memory falls within the target value range corresponding to that feature, the user can maintain the memory to ensure that the feature value of that memory does not fall within the target value range, thereby reducing the risk of memory failure.

[0087] by Figure 3As shown, after the first computing device preprocesses the sample data in the first training set, it constructs a random forest prediction model (i.e., a random forest model for memory fault prediction). Then, it trains the constructed random forest prediction model based on the first training set. During the training process, the decision trees of the random forest model are pruned, and the hyperparameters of the random forest model are optimized. Finally, the trained random forest model serves as the memory fault prediction model. The hyperparameters of the random forest model include the model parameters mentioned above, as well as other parameters besides those listed above.

[0088] The inherent parallelization of the random forest model allows multiple decision trees to be trained independently, making full use of distributed computing resources, quickly processing terabyte-level error logs, and significantly shortening model training and evaluation time. This enables this application to efficiently process massive amounts of memory error data.

[0089] The above solution can extract key features, automatically extracting key features such as error type, frequency, and location distribution from massive amounts of memory error data, without manual intervention. This improves the efficiency and accuracy of data processing and reduces human error.

[0090] It should be understood that the implementation process of step 203 is the process of training a preset model based on multiple first samples to obtain a memory fault prediction model.

[0091] In some embodiments, the memory prediction model is an online model. The first computing device can periodically acquire the latest samples and fine-tune the model parameters based on these samples, thereby dynamically optimizing the memory prediction model and enabling it to adapt to new data. This achieves continuous optimization of model parameters, adapting to changes in memory error types during long-term operation, ensuring the timeliness and accuracy of the memory fault prediction model, and ensuring that prediction accuracy improves over time. Since the model parameters are fine-tuned, model maintenance costs can be reduced. The implementation method will be described below in conjunction with the processes shown in steps 204 and 206.

[0092] 204. The first computing device acquires the second training set.

[0093] The second training set includes multiple second samples and the label of each second sample. Each second sample corresponds to a time period, and the time period corresponding to the second sample is later than the time period corresponding to the first sample. Therefore, the second training set is a new training set relative to the first training set.

[0094] Each second sample includes error information, error location information, operating environment information, and workload information of the third memory during its operation. This operation refers to the running process of the third memory within the corresponding time period of the second sample. The third and second memory have the same memory type; they can be the same memory or different memory types. The error information, error location information, operating environment information, and workload information are as described above and will not be repeated here. The labels for the second samples are similar to those for the first samples and will not be repeated here. The data in the second samples comes from real-time data within the corresponding time period, such as data from the error log of the third memory at the corresponding time segment.

[0095] This step 204 is the same as step 201 above. Here, this application embodiment will not repeat this step 204. This step 204 is the process of the first computing device acquiring multiple second samples.

[0096] 205. The first computing device performs data preprocessing on the second training set.

[0097] Step 205 is similar to step 204 above. For example, the data preprocessing module in the first computing device cleans the data of each second sample in the second training set to remove invalid or erroneous data and handle missing values. For each cleaned second sample, the data preprocessing module extracts features to obtain multidimensional features of the second sample. The dimensions of the multidimensional features of the second sample may be the same as those of the first sample, or they may be greater than those of the first sample. This allows the memory prediction model to be updated for the new features, enabling the updated model to support fault prediction for the new features. The data preprocessing module encodes the multidimensional features of each second sample to obtain a feature vector for each second sample. This encoding method is similar to the method described above for encoding the multidimensional features of each first sample.

[0098] 206. The first computing device incrementally updates the memory fault prediction model based on the second training set after data processing, and obtains the updated memory fault prediction model.

[0099] Taking a random forest model as an example of a memory fault prediction model, the first computing device evaluates the contribution of each feature in the multidimensional features of the second sample using the memory fault prediction model, obtaining the contribution of each feature. This process is similar to the method described above for evaluating the contribution of each feature in the multidimensional features of the first sample. The Y features with the highest contribution in the multidimensional features of the second sample are identified as key features, and the weights corresponding to these Y key features in the model parameters of the memory fault prediction model are increased. The features other than the key features in the multidimensional features are called residual features, and the weights corresponding to these residual Y key features in the model parameters are decreased to achieve weight updates. Y is an integer greater than 1, and Y is less than the total dimension of the multidimensional features of the second sample. Alternatively, users can also select multiple features with the highest probability of causing the target fault from the multidimensional features of the second sample based on their work experience, and increase the weights corresponding to these multiple features in the model parameters to achieve weight updates.

[0100] The first computing device incrementally trains the memory fault prediction model with updated weights based on Y key features (or multidimensional features of multiple second samples) and the label of each second sample. This enables incremental training of the memory fault prediction model based on multiple second samples. During the incremental training process, the model parameters of the memory fault prediction model are updated. Taking a random forest model as an example, the first computing device incrementally updates the memory fault prediction model according to the random forest incremental learning mechanism. This application embodiment does not limit the specific update method.

[0101] In the model training method provided in this application embodiment, multiple factors are comprehensively considered, including error location information, error type, operating environment, and workload during memory operation. Furthermore, multi-dimensional features within each factor are considered to obtain multi-dimensional features during memory operation. This results in higher prediction accuracy for the memory fault prediction model trained based on these multi-dimensional features, leading to more accurate predictions of the probability of a target memory fault and the probability of each feature triggering a target fault, thus improving the overall accuracy of the memory fault prediction model. Moreover, when the memory fault prediction model is a random forest model, it can capture the multi-dimensional features that trigger memory faults. Through parallel training of multiple decision trees, the random forest model can effectively capture complex correlation patterns in memory operation state data (such as data in samples). By hierarchically partitioning the feature space using multiple decision trees, it can capture the non-linear relationships between features, thereby accurately predicting the probability of a target memory fault.

[0102] The memory types of the second and third memory are referred to as the target memory type, and the dimensions of the aforementioned multi-dimensional features are related to the target memory type. For example, different memory types may have different packaging forms, and the aforementioned multi-dimensional features also include the packaging features of the target memory type. These packaging features represent the packaging form of the target memory type; for example, if the target memory type is 3D memory, the packaging feature would be a feature that reflects a "3D stacking structure." Furthermore, different memory types may have different workloads, and the aforementioned workload information features are features that reflect the workload of the target memory type. The memory fault prediction model subsequently trained based on the multi-dimensional features related to the target memory type is more suitable for the target memory type and can predict faults in the target memory type. Therefore, for different memory types, features related to the corresponding memory type are adaptively selected, as described above. Figure 2 The training process shown can train a memory fault prediction model suitable for the corresponding memory type. This makes the training method adaptable to different types of memory architectures, giving it wide applicability in various application scenarios such as memory design optimization, testing and verification, and fault diagnosis.

[0103] The memory fault prediction model described above can be used to predict the probability of a target fault occurring in memory of the same type as the second memory. The following section combines... Figure 4 This section will introduce this implementation method.

[0104] Figure 4 This is a flowchart illustrating a memory fault prediction method according to an exemplary embodiment. The method is applied to a second computing device, which can be a first computing device or the aforementioned computing device 100, such as... Figure 4 As shown, the method includes the following steps.

[0105] 401. The second computing device obtains error information, error location information, and operating environment information of the first memory during operation.

[0106] The first memory, second memory, and third memory all have the same memory type, meaning the first memory is the target memory type. The first memory, second memory, and third memory can be the same memory or different memory types. The first memory can be the memory in the second computing device or the memory of other computing devices besides the second computing device.

[0107] The second computing device acquires a third sample, which is used in the prediction phase of the memory prediction model. This third sample contains the operating status information of the first memory, indicating its operating status during data processing. For example, the third sample corresponds to a time period and includes error information, error location information, operating environment information, and workload information of the first memory during that time period. The time period corresponding to the third sample is later than the time period of the first sample. The definitions of error information, error location information, operating environment information, and workload information are as described above and will not be repeated here.

[0108] The process by which the second computing device acquires the third sample is similar to the process by which the first computing device acquires the first sample. For example, if the first memory is the memory of the second computing device, the second computing device reads the error logs and operation logs of the first memory for a certain period of time. Based on the read error logs and operation logs, it generates a third sample so that the third sample corresponds to that period of time. This process is similar to the process by which the second computing device generates multiple first samples based on multiple error logs and multiple operation logs of the second memory.

[0109] by Figure 5 For example, the second computing device is equipped with a memory fault prediction system, which includes a data acquisition module. The data acquisition module performs real-time data monitoring on the error logs and operation logs of the first memory over a period of time, collects the error logs and operation logs of the first memory over a period of time, and generates a third sample based on the collected error logs and operation logs.

[0110] If the first memory is not the memory of the second computing device, for example, if the first memory is the memory of another computing device besides the second computing device, the second computing device obtains the error log and operation log of the first memory from the other computing device, and then generates a third sample based on the error log and operation log of the second memory.

[0111] 402. The second computing device performs feature extraction on the error information, the error location information, and the operating environment information to obtain multi-dimensional features, which include at least one feature of the error information, at least one feature of the error location information, and at least one feature of the operating environment information.

[0112] For example, the second computing device performs feature region extraction (i.e., feature extraction on the third sample) on the error information, error location information, operating environment information, and workload information in the third sample, obtaining multidimensional features of the third sample. These multidimensional features include at least one feature of the error information, at least one feature of the error location information, at least one feature of the operating environment information, and at least one feature of the workload information in the third sample. The multidimensional features of the third sample have the same feature dimensions as the multidimensional features of the first sample, and can be referenced from the multidimensional features of the first sample; therefore, they will not be elaborated further here.

[0113] The process of extracting features from the third sample to obtain the multidimensional features of the third sample is the same as the process of extracting features from the first sample to obtain the multidimensional features of the first sample, and will not be repeated here.

[0114] It should be understood that the multidimensional features used in training the memory fault prediction model have the same dimensions as the multidimensional features of the third sample. The multidimensional features used in training the memory fault prediction model are related to the target memory type, and correspondingly, the multidimensional features of the third sample are related to the target memory type.

[0115] 403. The second computing device determines a first probability and a first risk contribution of each feature in the multidimensional feature based on the memory failure prediction model and the multidimensional feature. The first probability is the probability of the first memory causing a target failure as predicted by the memory failure prediction model based on the multidimensional feature. The first risk contribution of any feature is the risk contribution of the feature predicted by the memory failure prediction model based on the multidimensional feature to the target failure.

[0116] The multidimensional feature refers to the multidimensional feature of the third sample mentioned above. The memory fault prediction model can be the memory fault prediction model trained by the first computing device based on steps 201 to 203, or it can be the updated memory fault prediction model obtained based on steps 204 to 206. The first computing device and the second computing device may be the same computing device or different computing devices. If the first computing device and the second computing device are the same computing device, the memory fault prediction model is trained by the second computing device. If the first computing device and the second computing device are different computing devices, the second computing device obtains the memory fault prediction model trained by the first computing device.

[0117] The second computing device processes the multidimensional features of the third sample using a memory fault prediction model to obtain multiple second probabilities and multiple second risk contributions for each feature in the multidimensional features. The second probability is the probability of a target memory fault occurring, predicted based on some features in the multidimensional features. Each second risk contribution for any feature is the risk contribution of that feature to the target fault, predicted based on some features in the multidimensional features. The second computing device processes the multiple second probabilities using the memory fault prediction model to obtain a first probability. The second computing device processes the multiple second risk contributions for each feature using the memory fault prediction model to obtain a first risk contribution for each feature.

[0118] Optionally, the second computing device further processes the multidimensional features in the third sample using a memory fault prediction model to obtain multiple prediction error types. The prediction error type is the error type of the first memory that is predicted based on some features in the multidimensional features. The memory fault prediction model then processes the multiple prediction error types to obtain a target error type, which is the error type of the first memory that is predicted based on the multidimensional features.

[0119] In some embodiments, the first probability is the average of multiple second probabilities. The above-described processing of the multiple second probabilities using the memory fault prediction model to obtain the first probability includes: averaging the multiple second probabilities using the memory fault prediction model to obtain the first probability. In other embodiments, the first probability is a weighted sum of multiple second probabilities. The above-described processing of the multiple second probabilities using the memory fault prediction model to obtain the first probability includes: weighted summing the multiple second probabilities using the memory fault prediction model to obtain the first probability.

[0120] Since each second probability predicts the probability of the first memory causing a target failure based on some features in the multidimensional features, the first probability is obtained by averaging or weighted summing multiple second probabilities. This makes the first probability the probability of the first memory causing a target failure predicted by the various features in the multidimensional features, thus making the first probability the overall failure prediction result and improving the accuracy of the failure prediction result.

[0121] In some embodiments, the first risk contribution of any feature is the average of multiple second risk contributions of that feature. The above-mentioned processing of the multiple second risk contributions of each feature using a memory fault prediction model to obtain the second risk contribution of each feature includes: for any feature among the multidimensional features, averaging the multiple second risk contributions of that feature using the memory fault prediction model to obtain the first risk contribution of that feature. In other embodiments, the first risk contribution of any feature is a weighted sum of the multiple second risk contributions of that feature. The above-mentioned processing of the multiple second risk contributions of each feature using a memory fault prediction model to obtain the second risk contribution of each feature includes: for any feature, weighted summing the multiple second risk contributions of that feature using the memory fault prediction model to obtain the first risk contribution of that feature.

[0122] Since the multiple second risk contributions of any feature are predicted based on some features in the multidimensional features, the first risk contribution of any feature is obtained by averaging or weighted summing the multiple second risk contributions of any feature. Thus, the first risk contribution is the risk contribution of the feature to the target failure predicted by each feature in the multidimensional features, and the first probability becomes the overall risk contribution prediction result, thereby improving the accuracy of the feature's risk contribution.

[0123] The following section uses a random forest model as an example to introduce the process by which the memory fault prediction model generates error types, first probabilities, and the first risk contribution of each feature in the multidimensional features based on the multidimensional features of the third sample.

[0124] The second computing device processes the multidimensional features of the third sample through each decision tree in the random forest model, obtaining the second probability, second risk contribution distribution, and prediction error type output by each decision tree. That is, for any decision tree in the random forest model, the multidimensional features of the third sample are processed through that decision tree to obtain the second probability, second risk contribution distribution, and prediction error type output by that decision tree. The second risk contribution distribution includes the second risk contribution of each feature in the multidimensional features.

[0125] The following section describes the process of obtaining the second probability, the second risk contribution distribution, and the prediction error type of a decision tree output.

[0126] For any decision tree in a random forest model, the second computing device inputs the third sample and its multidimensional features into the decision tree. The third sample is continuously transmitted through the decision tree until it reaches a leaf node in each decision tree. The splitting node traversed by the third sample in the decision tree is called the target splitting node, and the leaf node in the decision tree where the third sample is obtained is called the target leaf node. The decision path in the same decision tree that includes all the target splitting nodes and target leaf nodes traversed by the third sample is called the target decision path.

[0127] The second computing device inputs the third sample and its multidimensional features into the root node of the decision tree. Any target split node (such as the root node and intermediate nodes) in the decision tree, based on its stored split criteria and the feature values ​​of the split features corresponding to the split criteria in the multidimensional features of the third sample, divides the third sample into a child node of the split node in the decision tree (i.e., a node split from the split node). The third sample and its multidimensional features are then transmitted to the child node. When the child node is a split node in the decision tree, after obtaining the input third sample and its multidimensional features, the child node becomes the new target split node. When the child node is a leaf node in the decision tree, after obtaining the input third sample and its multidimensional features, the child node becomes the target leaf node. At this time, the third sample is transmitted to the target leaf node along the target decision path in the decision tree. When the third sample is transmitted to the target leaf node along the target decision path in the decision tree, the target leaf node, in response to receiving the third sample and its multidimensional features, generates a second risk contribution distribution based on the target decision path, and outputs the second risk contribution distribution, a second probability, and error distribution information. The second risk contribution distribution includes the risk contribution of each feature in the multidimensional features of the third sample. The risk contribution of each feature in the second risk contribution distribution is the second risk contribution of that feature, and the second risk contribution of any feature is the risk contribution of that feature to the target failure under the influence of the splitting feature corresponding to the decision tree. The process of generating the second risk contribution distribution based on the target decision path is similar to the process of generating the risk contribution distribution based on the decision path to which the leaf node belongs. The second probability output by the target leaf node is the memory failure probability stored in the target leaf node, which is the probability of the first memory failing as predicted based on each splitting feature corresponding to the decision tree. The error classification information output by the target leaf node is the error classification information stored in the target leaf node, and the error type corresponding to the first label in this error classification information is also the predicted error type. The second risk contribution distribution, second probability, and misclassification information output by the target leaf node are the prediction results of the decision tree for the third sample, which are also the output results of the decision tree.

[0128] It should be understood that when the decision tree is a first decision tree, the second probability output by the first decision tree is the probability of the first memory memory experiencing the target failure predicted based on the splitting features corresponding to the first decision tree. The second risk contribution of any feature in the second risk contribution distribution output by the first decision tree is the risk contribution of that feature to the target failure under the influence of the splitting features corresponding to the first decision tree. The predicted error type in the error classification information output by the first decision tree is the error type predicted for the first memory memory under the influence of the splitting features corresponding to the first decision tree. When the decision tree is a second decision tree, the second probability output by the second decision tree is the probability of the first memory memory experiencing the target failure predicted based on multiple splitting features corresponding to the second decision tree. The second risk contribution of any feature in the second risk contribution distribution output by the second decision tree is the risk contribution of that feature to the target failure under the influence of multiple splitting features corresponding to the second decision tree. The predicted error type in the error classification information output by the second decision tree is the error type predicted for the first memory memory under the influence of multiple splitting features corresponding to the second decision tree.

[0129] Accordingly, the above-mentioned memory fault prediction model processes the multidimensional features of the third sample to obtain multiple second probabilities and multiple second risk contributions of each feature in the multidimensional features, including: The multidimensional features of the third sample are processed through the first decision tree to obtain a second probability and a second risk contribution distribution. The second probability is the probability of the first memory to fail based on the splitting feature corresponding to the first decision tree. The second risk contribution distribution includes the second risk contribution of each feature in the multidimensional features. The second risk contribution of any feature is the risk contribution of the feature to the target failure under the influence of the splitting feature corresponding to the first decision tree. The second decision tree is used to process the multidimensional features of the third sample to obtain another second probability and another second risk contribution distribution. The other second probability is the probability of the first memory to fail as predicted by multiple split features corresponding to the first decision tree. The other second risk contribution distribution includes the second risk contribution of each feature in the multidimensional features. The second risk contribution of any feature is the risk contribution of the feature to the target failure under the influence of multiple split features corresponding to the first decision tree.

[0130] In some embodiments, the target leaf nodes in each decision tree output their respective prediction results to the control module in the random forest model. That is, each decision tree outputs its own prediction results to the control module in the random forest model. The prediction results of each decision tree include a second probability, a second risk contribution degree, a second risk contribution degree distribution, a second probability, and error distribution information.

[0131] The control module in the random forest model can receive prediction results from multiple decision trees in the random forest. Thus, the control module can receive multiple second probabilities, multiple second risk contribution distributions, and multiple misclassification information. Different second probabilities come from different decision trees, different second risk contribution distributions come from different decision trees, and different misclassification information comes from different decision trees.

[0132] The control module calculates the average or weighted sum of the multiple second probabilities to obtain the first probability.

[0133] The control module generates a first risk contribution distribution based on multiple second risk contribution distributions. The first risk contribution distribution includes the first risk contribution of each feature in the multidimensional features of the third sample. Specifically, the first risk contribution of any feature is the average or weighted sum of the second risk contributions of that feature among the multiple second risk contribution distributions. For example, for any feature in the multidimensional features, the average or weighted sum of the second risk contributions of that feature in the multiple second risk contribution distributions is used to obtain the first risk contribution of that feature. Alternatively, for any feature in the multidimensional features, when that feature is a splitting feature corresponding to at least one first decision tree, the average or weighted sum of the second risk contribution distributions output by at least one first decision tree and the second risk contribution distributions output by each of the second decision trees is used to obtain the first risk contribution of that feature, thus ignoring the second risk contribution of that feature output by the first decision trees corresponding to other splitting features. After obtaining the first risk contribution of each feature in the multidimensional features, the first risk contribution of each feature in the multidimensional features is combined to form the first risk contribution distribution.

[0134] The control module counts the error types with the most corresponding first labels from the error classification information output by all decision trees (i.e., the predicted error types), and takes the error types with the most corresponding first labels as the target error types.

[0135] The control module outputs a first probability, a first risk contribution distribution, and a target error type. These outputs all correspond to the third sample. Therefore, the first probability corresponding to the third sample is the probability of the first memory memory experiencing a target failure, predicted by the random forest model (i.e., the memory failure prediction model) based on the multi-dimensional features of the third sample. Similarly, the second risk contribution of any feature in the first risk contribution distribution corresponding to the third sample is the risk contribution of that feature to the target failure, predicted by the random forest model (i.e., the memory failure prediction model) based on the multi-dimensional features of the third sample. Finally, the target error type corresponding to the third sample is the error type of the first memory memory error predicted by the random forest model (i.e., the memory failure prediction model) based on the multi-dimensional features of the third sample.

[0136] In some embodiments, when the random forest model includes multiple first decision trees, each corresponding to a splitting feature, the splitting features corresponding to different first decision trees may be the same or different. Therefore, the multiple first decision trees may collectively correspond to at least one splitting feature. The control module obtains multiple second risk contribution distributions based on the multiple first decision trees, determines the third risk contribution of the at least one splitting feature, and outputs the third risk contribution of the at least one splitting feature. The third risk contribution of any splitting feature refers to its overall risk contribution to the target failure without the influence of other features. When the splitting feature corresponds to only one first decision tree, the third risk contribution of the splitting feature is the second risk contribution of the splitting feature in the second risk contribution distribution output by the corresponding first decision tree. When the splitting feature corresponds to multiple first decision trees, the multiple first decision trees corresponding to the splitting feature collectively output multiple second risk contribution distributions, and the third risk contribution of the splitting feature is the average or weighted sum of the second risk contributions of the splitting feature in the multiple second risk contribution distributions.

[0137] By outputting the third risk contribution of at least one splitting feature, users can focus on maintaining splitting features with higher third risk contributions when performing memory maintenance, based on the magnitude of the third risk contribution of at least one splitting feature, thereby improving memory maintenance efficiency.

[0138] When the memory fault prediction model is a random forest model, integrating the prediction results of multiple decision trees effectively reduces the sensitivity of a single model to noisy data and avoids overfitting, making it particularly suitable for scenarios with massive and highly noisy memory error data. Leveraging the parallel processing capabilities of random forests, multiple error dimensions (such as multidimensional features) are considered simultaneously for comprehensive evaluation. This improves the comprehensiveness and accuracy of predictions, enabling more precise identification of different types of memory error patterns (i.e., different error types).

[0139] In some embodiments, the computing device provides a fault warning for the first memory based on a first probability corresponding to a third sample. For example, if the first probability corresponding to any third sample is greater than or equal to a probability threshold, the second computing device sends an alarm message, for example, through an application programming interface (API). This alarm message indicates that the first memory is at risk of a target fault, and the alarm message includes the first probability corresponding to the third sample.

[0140] In some embodiments, the second computing device provides a fault warning for the first memory based on the first risk contribution of each feature in the multidimensional features of the third sample. For example, each feature in the multidimensional features corresponds to a contribution threshold, and the contribution thresholds corresponding to different features may be the same or different. If the first risk contribution of any feature is greater than or equal to the contribution threshold corresponding to that feature, the second computing device sends an alarm message indicating that the feature is a high-risk factor for causing a target fault, and the alarm message includes the first risk contribution of the feature.

[0141] In some embodiments, the computing device displays a heatmap based on the first risk contribution of each feature in the multidimensional features of the third sample, the heatmap being used to represent the first risk contribution of each feature in the multidimensional features.

[0142] For example, the second computing device acquires multiple third samples, extracts features from each third sample to obtain multidimensional features for each third sample, and for any third sample, determines a first risk contribution distribution corresponding to that third sample based on the memory fault prediction model and the multidimensional features of that third sample. The second computing device generates a heatmap based on the first risk contribution distribution corresponding to the multiple third samples, for example, using a matrix plotting library (Matplotlib) or an interactive visualization library (Plotly). After generating the heatmap, the computing device can display the heatmap or send the heatmap to a terminal for display. The heatmap includes multiple points, each with coordinates (x, y, z), where x represents the number of one feature in the multidimensional features, y represents the number of one third sample among the multiple third samples, and z represents the color value of the point. This intensity is used to characterize the first risk contribution of feature x predicted based on the third sample y. Therefore, this heatmap is a multidimensional distribution heatmap. It visually reflects which features among the multidimensional characteristics are high-risk factors causing the primary memory to fail, clearly demonstrating the vulnerability of the primary memory. This helps engineers quickly locate problematic areas of the memory, improving the efficiency of memory reliability management. It can also improve system maintenance efficiency and reduce the risk of data loss or system crashes due to memory failures.

[0143] In some embodiments, the computing device provides fault warnings for the first memory based on a heatmap. For example, if the first risk contribution represented by the color value corresponding to any feature in the heatmap is greater than or equal to the contribution threshold of that feature, the second computing device sends an alarm message, for example, via an API. This alarm message indicates that the feature is a high-risk factor causing a target fault, and the alarm message includes the first risk contribution of the feature.

[0144] by Figure 5For example, a memory fault prediction system also includes a fault warning model, a data processing module, and a real-time warning module. The data acquisition module sends the multi-dimensional features of the third sample to the fault warning model. The fault warning module inputs the multi-dimensional features of the third sample into a random forest prediction model (i.e., a memory fault preset model), and the random forest prediction model outputs the first probability, the first risk contribution distribution, and the target error type corresponding to the third sample. The fault warning module also sets warning rules; for example, if the first probability is greater than a probability threshold, a fault warning is issued for the first memory fault. The fault warning model sends the first probability, the first risk contribution distribution, and the target error type output by the random forest prediction model to the data processing module. The data processing module generates an error report based on the first probability, the first risk contribution distribution, and the target error type. The error report includes the first probability, the first risk contribution distribution, and the target error type. The data processing module generates a heatmap based on the first risk contribution distribution in the error report. Based on the heatmap, it performs error analysis on the first memory module, generating an error analysis report, which is then sent to the real-time early warning module. The error analysis report includes the first risk contribution of at least one fault feature. Here, a fault feature refers to one of the F features with the highest first risk contribution among the multi-dimensional features; the F features are those most likely to cause the target fault. F is an integer greater than or equal to 1. Based on the error analysis report, the real-time early warning module provides fault warnings for the second memory module. For example, if the first risk contribution of any fault feature is greater than or equal to the contribution threshold corresponding to that fault feature, an alarm message is sent, including the first risk contribution of that fault feature. The real-time early warning module supports firmware and software collaboration and also has error probability prediction capabilities. For example, the contribution threshold for each feature in the multi-dimensional features is predicted by the real-time early warning module. If the first risk contribution of any fault feature in the error analysis report is less than the contribution threshold for that fault feature, it indicates that the first risk contribution of that fault feature is inaccurate. The real-time early warning module then adjusts the contribution threshold for that fault feature so that the adjusted threshold is less than or equal to the first risk contribution of that fault feature, thereby dynamically adjusting the alarm threshold corresponding to the feature. The real-time early warning module also has reliability analysis capabilities, such as performing reliability scoring on the first memory module based on the error analysis report.

[0145] The output of memory failure prediction models can be used to generate detailed reliability analysis reports, providing crucial data support for memory quality control and lifespan management. These models can analyze error trends and predict potentially high-risk areas, allowing for preventative measures to be taken to replace or optimize potential failure points in advance, thereby improving system stability.

[0146] The memory fault prediction method provided in this application uses a memory fault prediction model to predict the probability of a target fault occurring in the memory. In the prediction process, it comprehensively considers multiple factors such as error information, error location information and operating environment information during memory operation, thereby improving the accuracy of memory fault prediction.

[0147] It should be understood that in the above method embodiments, each step performed by the decision tree or module in any model (such as a random forest model, a memory fault prediction model, or a preset model) is a step performed by that model.

[0148] Figure 6 This is a structural block diagram of a memory fault prediction device according to an exemplary embodiment, such as... Figure 6 As shown, the memory fault prediction device 600 includes: The first acquisition unit 601 is used to acquire error information, error location information and operating environment information of the first memory during the running process; The extraction unit 602 is used to extract features from the error information, the error location information and the operating environment information to obtain multi-dimensional features, wherein the multi-dimensional features include at least one feature of the error information, at least one feature of the error location information and at least one feature of the operating environment information. The first determining unit 603 is used to determine a first probability and a first risk contribution of each feature in the multidimensional features based on the memory fault prediction model and the multidimensional features. The first probability is the probability of the first memory having a target fault as predicted by the memory fault prediction model based on the multidimensional features. The first risk contribution of any feature is the risk contribution of the feature to the target fault as predicted by the memory fault prediction model based on the multidimensional features.

[0149] Optionally, the first determining unit 603 is used for: The memory fault prediction model processes the multidimensional features of the third sample to obtain multiple second probabilities and multiple second risk contribution degrees for each feature in the multidimensional features. The second probability is the probability of the first memory experiencing the target fault, predicted based on some features in the multidimensional features. Each second risk contribution degree of the feature is the risk contribution degree of the feature to the target fault, predicted based on some features in the multidimensional features. The memory fault prediction model is used to process the multiple second probabilities to obtain the first probability; The memory fault prediction model processes multiple second risk contribution values ​​for each feature in the multidimensional features to obtain the first risk contribution value for each feature.

[0150] Optionally, the first probability is the average or weighted sum of the plurality of second probabilities, and the first risk contribution of any feature is the average or weighted sum of the plurality of second probabilities of the feature.

[0151] Optionally, the memory fault prediction model is a random forest model, which includes at least one first decision tree and at least one second decision tree; The first decision tree corresponds to a split feature in the multidimensional features. The first decision tree is constructed based on the split feature and is used to analyze the impact of the split feature on the target fault. The second decision tree corresponds to multiple split features in the multidimensional features. The second decision tree is constructed based on the multiple split features. The second prediction tree is used to analyze the combined impact of the multiple split features on the target fault.

[0152] Optionally, the first determining unit 603 is used to: The first decision tree is used to process the multidimensional features of the third sample to obtain a second probability and a second risk contribution distribution. The second probability is the probability of the first memory causing the target failure, which is predicted based on the splitting feature corresponding to the first decision tree. The second risk contribution distribution includes the second risk contribution of each feature in the multidimensional features. The second risk contribution of any feature is the risk contribution of the feature to the target failure under the influence of the splitting feature corresponding to the first decision tree. The second decision tree is used to process the multidimensional features of the third sample to obtain another second probability and another second risk contribution distribution. The other second probability is the probability of the first memory experiencing the target failure, predicted based on multiple split features corresponding to the first decision tree. The other second risk contribution distribution includes the second risk contribution of each feature in the multidimensional features. The second risk contribution of any feature is the risk contribution of the feature to the target failure under the influence of multiple split features corresponding to the first decision tree.

[0153] Optionally, the random forest model includes multiple first decision trees, each first decision tree corresponding to at least one splitting feature, and the memory fault prediction device 600 further includes: The second determining unit is used to determine the third risk contribution of the at least one splitting feature based on the multiple second risk contribution distributions obtained from the multiple first decision trees. The third risk contribution of any splitting feature refers to the overall risk contribution of the splitting feature to the target failure without the influence of other features. An output unit is used to output the third risk contribution of the at least one splitting feature.

[0154] Optionally, the dimensions of the multidimensional feature are related to the memory type of the first memory.

[0155] Optionally, the memory fault prediction device 600 further includes: The early warning unit is used to provide early warning of faults to the first memory based on the first risk contribution of each feature in the multidimensional features.

[0156] Optionally, the memory fault prediction device 600 further includes: The display unit is used to display a heatmap based on the first risk contribution of each feature in the multidimensional features, wherein the heatmap is used to reflect the first risk contribution of each feature in the multidimensional features.

[0157] Optionally, the memory fault prediction device 600 further includes: The second acquisition unit is used to acquire multiple first samples, the first samples including error information, error location information and operating environment information of the second memory during operation, and the second memory and the first memory have the same memory type; The training unit is used to train the preset model based on the multiple first samples to obtain the memory fault prediction model.

[0158] Optionally, the second acquisition unit is further configured to acquire multiple second samples, the second samples including error information, error location information and operating environment information of the third memory during operation, wherein the third memory and the first memory have the same memory type; The training unit is also used to incrementally train the memory fault prediction model based on the plurality of second samples.

[0159] Regarding the memory fault prediction device 600 in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments of the memory fault prediction method, and will not be elaborated here.

[0160] In some embodiments, this application also provides a computer-readable storage medium, which, when at least one instruction in the computer-readable storage medium is executed by one or more processors of a computing chip, enables the computing chip to perform at least one of the memory fault prediction method and model training method provided in this application.

[0161] In some embodiments, this application also provides a computer program product comprising one or more instructions that can be executed by one or more processors of a computing chip, enabling the computing chip to execute at least one of the memory fault prediction method and model training method provided in this application.

[0162] This application provides multiple possible implementation methods, and some of these implementation methods can be combined to obtain even more implementation methods.

[0163] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0164] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A memory fault prediction method, characterized in that, The method includes: Obtain error information, error location information, and runtime environment information during the first memory operation; Feature extraction is performed on the error information, the error location information, and the operating environment information to obtain multi-dimensional features. The multi-dimensional features include at least one feature of the error information, at least one feature of the error location information, and at least one feature of the operating environment information. Based on the memory failure prediction model and the multidimensional features, a first probability and a first risk contribution of each feature in the multidimensional features are determined. The first probability is the probability of the first memory experiencing a target failure, predicted by the memory failure prediction model based on the multidimensional features. The first risk contribution of any feature is the risk contribution of the feature to the target failure, predicted by the memory failure prediction model based on the multidimensional features.

2. The method according to claim 1, characterized in that, The determination of the first probability and the first risk contribution of each feature in the multidimensional features based on the memory fault prediction model and the multidimensional features includes: The memory fault prediction model processes the multidimensional features of the third sample to obtain multiple second probabilities and multiple second risk contribution degrees for each feature in the multidimensional features. The second probability is the probability of the first memory experiencing the target fault based on some features in the multidimensional features. Each second risk contribution degree of the feature is the risk contribution degree of the feature to the target fault based on some features in the multidimensional features. The memory fault prediction model is used to process the multiple second probabilities to obtain the first probability; The memory fault prediction model processes multiple second risk contribution values ​​for each feature in the multidimensional features to obtain the first risk contribution value for each feature.

3. The method according to claim 2, characterized in that, The first probability is the average or weighted sum of the plurality of second probabilities, and the first risk contribution of any feature is the average or weighted sum of the plurality of second probabilities of the feature.

4. The method according to claim 2, characterized in that, The memory fault prediction model is a random forest model, which includes at least one first decision tree and at least one second decision tree. The first decision tree corresponds to a split feature in the multidimensional features. The first decision tree is constructed based on the split feature and is used to analyze the impact of the split feature on the target fault. The second decision tree corresponds to multiple split features in the multidimensional features. The second decision tree is constructed based on the multiple split features. The second prediction tree is used to analyze the combined impact of the multiple split features on the target fault.

5. The method according to claim 4, characterized in that, The process of processing the multidimensional features of the third sample using the memory fault prediction model to obtain multiple second probabilities and multiple second risk contribution values ​​for each feature in the multidimensional features includes: The first decision tree is used to process the multidimensional features of the third sample to obtain a second probability and a second risk contribution distribution. The second probability is the probability of the first memory causing the target failure, which is predicted based on the splitting feature corresponding to the first decision tree. The second risk contribution distribution includes the second risk contribution of each feature in the multidimensional features. The second risk contribution of any feature is the risk contribution of the feature to the target failure under the influence of the splitting feature corresponding to the first decision tree. The second decision tree is used to process the multidimensional features of the third sample to obtain another second probability and another second risk contribution distribution. The other second probability is the probability of the first memory experiencing the target failure, predicted based on multiple split features corresponding to the first decision tree. The other second risk contribution distribution includes the second risk contribution of each feature in the multidimensional features. The second risk contribution of any feature is the risk contribution of the feature to the target failure under the influence of multiple split features corresponding to the first decision tree.

6. The method according to claim 5, characterized in that, The random forest model includes multiple first decision trees, each first decision tree corresponding to at least one splitting feature; the method further includes: Based on the multiple second risk contribution distributions obtained from the multiple first decision trees, the third risk contribution of the at least one splitting feature is determined. The third risk contribution of any splitting feature refers to the overall risk contribution of the splitting feature to the target failure without the influence of other features. Output the third risk contribution of the at least one splitting feature.

7. The method according to any one of claims 1-6, characterized in that, The dimensions of the multidimensional feature are related to the memory type of the first memory.

8. The method according to any one of claims 1-6, characterized in that, The method further includes: Based on the first risk contribution of each feature in the multidimensional features, a fault warning is given for the first memory.

9. The method according to any one of claims 1-6, characterized in that, The method further includes: A heatmap is displayed based on the first risk contribution of each feature in the multidimensional features, and the heatmap is used to reflect the first risk contribution of each feature in the multidimensional features.

10. The method according to any one of claims 1-6, characterized in that, The method further includes: Multiple first samples are obtained. The first samples include error information, error location information, and runtime environment information of the second memory during operation. The second memory and the first memory have the same memory type. Based on the multiple first samples, the preset model is trained to obtain the memory fault prediction model.

11. The method according to any one of claims 1-6, characterized in that, The method further includes: Multiple second samples are obtained, including error information, error location information, and runtime environment information of the third memory during operation. The third memory and the first memory have the same memory type. Based on the multiple second samples, the memory fault prediction model is incrementally trained.

12. A memory fault prediction device, characterized in that, The device includes: The first acquisition unit is used to acquire error information, error location information, and runtime environment information of the first memory during operation. An extraction unit is used to extract features from the error information, the error location information, and the operating environment information to obtain multi-dimensional features, wherein the multi-dimensional features include at least one feature of the error information, at least one feature of the error location information, and at least one feature of the operating environment information. The first determining unit is configured to determine a first probability and a first risk contribution of each feature in the multidimensional features based on the memory fault prediction model and the multidimensional features. The first probability is the probability of the first memory experiencing a target fault as predicted by the memory fault prediction model based on the multidimensional features, and the first risk contribution of any feature is the risk contribution of the feature to the target fault as predicted by the memory fault prediction model based on the multidimensional features.

13. A computing chip, characterized in that, include: One or more processors; One or more memories for storing the one or more processor-executable instructions; The one or more processors are configured to execute the instructions to implement the memory fault prediction method as described in any one of claims 1 to 11.

14. A computing device, characterized in that, The computing device includes a memory and the computing chip of claim 13.

15. A computer-readable storage medium, characterized in that, When at least one instruction in the computer-readable storage medium is executed by the processor of the computing chip, the computing chip is enabled to perform the memory fault prediction method as described in any one of claims 1 to 11.

16. A computer program product, characterized in that, It includes one or more instructions, which are executed by the processor of a computer chip, enabling the computing chip to perform the memory fault prediction method as described in any one of claims 1 to 11.

Images