Memory anomaly prediction method and device based on multi-algorithm decision and electronic equipment

Abstract

The invention discloses a memory anomaly prediction method based on multi-algorithm decision, and belongs to the technical field of fault prediction. The method comprises the steps that data information generated in the running process of a server is obtained, preprocessing and feature extraction are conducted on error data, a feature set is obtained, and the feature set comprises counting dimension features, component dimension features, statistical dimension features and equipment configuration dimension features; the feature set is input into a trained basic prediction model, a first prediction probability, a second prediction probability and a third prediction probability are obtained, and the basic prediction model comprises a random forest sub-model, a lightweight gradient elevator sub-model and an extreme gradient lifting sub-model; obtaining a first prediction probability value based on the first prediction probability, the second prediction probability and the third prediction probability; inputting the feature set into a depth time sequence extraction network, and obtaining target features based on a feature masking method; inputting the target features into a gating circulation unit to obtain a fourth prediction probability value; obtaining a decision matrix based on the first prediction probability value and the fourth prediction probability value; and inputting the decision matrix into the trained decision optimizer logistic regression model to obtain a memory anomaly prediction result. According to the method, the accuracy and reliability of memory anomaly prediction are improved.

Description

Technical Field

[0001] This application belongs to the technical field of fault prediction, and particularly relates to a method for predicting memory anomalies based on multi-algorithm decision-making. Background Art

[0002] Today, with the rapid development of the digital economy, servers, as the core of enterprise IT infrastructure, the health of their memory systems is crucial for the stable operation of business. Memory failures can not only cause system crashes and data loss but also result in significant economic losses. Therefore, achieving early warning of memory failures has become a key issue in improving system reliability.

[0003] Existing memory fault prediction methods mainly rely on machine learning techniques, but still face many technical challenges. First, the fault samples are scarce and severely unbalanced in distribution, resulting in limited model training effects. Second, a single prediction model has insufficient generalization ability when dealing with complex multi-dimensional feature data and is difficult to adapt to the dynamic changes in the server operating environment. In addition, the mainstream centralized prediction architecture needs to transmit massive amounts of data to the central server for processing, significantly increasing the system's computational load and storage pressure, and it is difficult to meet the real-time warning requirements.

[0004] The application of related technologies in the field of memory fault prediction still has certain limitations. Existing methods are difficult to comprehensively cover complex and changeable fault modes, have poor adaptability to dynamic environments, and existing memory fault prediction methods generally have problems such as high data redundancy, insufficient prediction accuracy, and weak real-time response capabilities. Summary of the Invention

[0005] This application aims to at least solve one of the technical problems existing in the prior art. For this reason, this application proposes a method for predicting memory anomalies based on multi-algorithm decision-making, which improves the accuracy and reliability of memory anomaly prediction.

[0006] In a first aspect, this application provides a method for predicting memory anomalies based on multi-algorithm decision-making, and the method includes:

[0007] Obtain the data information generated during the operation of the server, preprocess and extract features from the data information to obtain a feature set, and the feature set includes count dimension features, component dimension features, statistical dimension features, and device configuration dimension features;

[0008] Input the feature set into a trained basic prediction model to obtain a first prediction probability, a second prediction probability, and a third prediction probability, and the basic prediction model includes a random forest sub-model, a light gradient boosting machine sub-model, and an extreme gradient boosting sub-model;

[0009] Based on the first prediction probability, the second prediction probability, and the third prediction probability, obtain a first prediction probability value;

[0010] Input the feature set into a deep temporal extraction network to obtain target features based on the feature masking method;

[0011] Input the target features into a gated recurrent unit to obtain a fourth prediction probability value;

[0012] Based on the first prediction probability value and the fourth prediction probability value, obtain a decision matrix;

[0013] Input the decision matrix into a trained decision optimizer logistic regression model to obtain a memory anomaly prediction result.

[0014] According to an embodiment of the present application, the step of inputting the feature set into a trained basic prediction model to obtain a first prediction probability, a second prediction probability, and a third prediction probability includes:

[0015] Input the feature set into the random forest sub-model to obtain the first prediction probability;

[0016] Input the feature set into the light gradient boosting sub-model to obtain the second prediction probability;

[0017] Input the feature set into the extreme gradient boosting sub-model to obtain the third prediction probability.

[0018] According to an embodiment of the present application, the step of obtaining a first prediction probability value based on the first prediction probability, the second prediction probability, and the third prediction probability includes:

[0019] Based on the first prediction probability, the second prediction probability, and the third prediction probability, obtain the first prediction probability value according to the following formula:

[0020] P fusion (X) = α RF ·P RF (X) + α LGB ·P LGB (X) + α XGB ·P XGB (X)

[0021] Wherein, P fusion (X) is the first prediction probability value, α RF is the influence factor of the random forest sub-model, P RF (X) is the first prediction probability, α LGB is the influence factor of the light gradient boosting sub-model, P LGB (X) is the second prediction probability, α XGB is the influence factor of the extreme gradient boosting sub-model, P XGB (X) is the third prediction probability.

[0022] According to an embodiment of the present application, inputting the feature set into the deep temporal extraction network to obtain target features based on the feature masking method includes:

[0023] Inputting the feature set into the deep temporal extraction network to obtain a reconstructed feature matrix;

[0024] Evaluating the reconstructed feature matrix based on the feature masking method to obtain the importance score of each feature in the reconstructed feature matrix;

[0025] Sorting each feature in the reconstructed feature matrix based on the importance score of each feature, and taking the top N features as the target features;

[0026] where N is a positive integer greater than zero.

[0027] According to an embodiment of the present application, inputting the target features into the gated recurrent unit to obtain the fourth prediction probability value includes:

[0028] Inputting the target features into the gated recurrent unit, and obtaining the fourth prediction probability value according to the following formula:

[0029] P GRU (X) = softmax(W o ·h t + b o )

[0030] where P GRU (X) is the fourth prediction probability value, W o is the weight of the output layer, b o is the bias of the output layer, and h t is the hidden state of the gated recurrent unit.

[0031] According to an embodiment of the present application, inputting the decision matrix into the trained decision optimizer logistic regression model to obtain the memory anomaly prediction result includes:

[0032] Inputting the decision matrix into the trained decision optimizer logistic regression model, and obtaining the memory anomaly prediction probability according to the following formula:

[0033]

[0034] where P final (X) is the memory anomaly prediction probability, W is the weight of the decision optimizer logistic regression model, b is the bias term of the decision optimizer logistic regression model, and M is the transpose of the decision matrix;

[0035] Based on the memory anomaly prediction probability and a preset threshold, obtain the memory anomaly prediction result.

[0036] According to an embodiment of the present application, the training method of the basic prediction model and the decision optimizer logistic regression model includes:

[0037] Define the search space of the basic prediction model based on the probability selection method of the guiding factor;

[0038] Obtain combined parameters based on the search space, and train the basic prediction model according to the first data set;

[0039] Based on the output of the basic prediction model and the output of the gated recurrent unit, obtain a second data set;

[0040] Train the decision optimizer logistic regression model according to the second data set, and gradually guide the decision optimizer logistic regression model and the basic prediction model to converge to the optimal solution based on the dynamic guiding factor attenuation mechanism, optimizing the prediction performance index and the difference gain.

[0041] In a second aspect, the present application provides a memory anomaly prediction device based on multi-algorithm decision-making. The device includes:

[0042] An acquisition module, configured to acquire data information generated during the operation of the server, perform preprocessing and feature extraction on the data information to obtain a feature set, where the feature set includes a count dimension feature, a component dimension feature, a statistical dimension feature, and a device configuration dimension feature;

[0043] A processing module, configured to input the feature set into a trained basic prediction model to obtain a first prediction probability, a second prediction probability, and a third prediction probability. The basic prediction model includes a random forest sub-model, a light gradient boosting machine sub-model, and an extreme gradient boosting sub-model;

[0044] Based on the first prediction probability, the second prediction probability, and the third prediction probability, obtain a first prediction probability value;

[0045] Input the feature set into a deep time series extraction network, and obtain target features based on the feature masking method;

[0046] Input the target features into a gated recurrent unit to obtain a fourth prediction probability value;

[0047] Based on the first prediction probability value and the fourth prediction probability value, obtain a decision matrix;

[0048] A prediction module, configured to input the decision matrix into a trained decision optimizer logistic regression model to obtain a memory anomaly prediction result.

[0049] In a third aspect, the present application provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the method for predicting memory anomalies based on multi-algorithm decision as described in the first aspect above is implemented.

[0050] In a fourth aspect, the present application provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method for predicting memory anomalies based on multi-algorithm decision as described in the first aspect above is implemented.

[0051] In a fifth aspect, the present application provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the method for predicting memory anomalies based on multi-algorithm decision as described in the first aspect.

[0052] In a sixth aspect, the present application provides a computer program product, including a computer program. When the computer program is executed by a processor, the method for predicting memory anomalies based on multi-algorithm decision as described in the first aspect above is implemented.

[0053] Additional aspects and advantages of the present application will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present application.

[0054] The method for predicting memory anomalies based on multi-algorithm decision provided by the present invention has the following beneficial effects compared with the prior art:

[0055] (1) By obtaining the data information generated during the operation of the server, preprocessing and feature extraction of these data, combining multiple basic prediction models such as random forest, light gradient boosting machine, and extreme gradient boosting, and deep learning methods such as deep time series extraction network and gated recurrent unit, the present invention utilizes the advantages of different algorithms in specific features and data patterns, can more comprehensively mine the complex relationships between memory fault features, fuses the results of the basic prediction models and the target features to further enhance the ability to capture the complex patterns of memory faults. Finally, the decision optimizer logistic regression model is used to optimize the fusion result through linear weighting and non-linear transformation, improving the accuracy and real-time performance of memory fault prediction.

[0056] (2) The present invention synergistically optimizes the prediction performance and the model difference degree through a dynamic guiding factor attenuation mechanism, realizes the mining of feature time series correlation by combining a gated recurrent unit, generates optimal combined parameters based on a probabilistic search space, and improves the synergy effect and generalization ability of multi-algorithm fusion. Through the iterative collaborative training of a decision optimizer and a basic prediction model, a dynamic balance between the algorithm accuracy and diversity is achieved. Taking the prediction performance and the model difference degree as optimization objectives, and combining the dynamic guiding factor attenuation mechanism, multi-objective collaborative optimization is realized. An adaptive weight allocation mechanism is introduced. In the initial stage of optimization, emphasis is placed on improving the difference degree of the basic prediction algorithm, exploring solutions with higher diversity to enhance the generalization ability of the system. In the later stage of optimization, the algorithm gradually tends to improve the prediction performance. By dynamically adjusting the attenuation rate, the influencing factors can be flexibly adjusted according to the optimization stage, balancing the relationship between the prediction performance and the difference degree. The dynamic update strategy of the guiding factor can reflect the change of the effect in real time during the optimization process, improving the accuracy and robustness of memory fault prediction in a complex operating environment, and effectively reducing the resource consumption and convergence time of model training.

[0057] (3) The present invention inputs a feature set into a random forest, a light gradient boosting machine, and an extreme gradient boosting sub-model to obtain a first prediction probability, a second prediction probability, and a third prediction probability respectively. The probability output method can better reflect the confidence degree of the category. By combining the output results of multiple prediction models, the accuracy and stability of the prediction results are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The above and / or additional aspects and advantages of the present application will become apparent and be readily understood from the following description of the embodiments in conjunction with the accompanying drawings, where:

[0059] Figure 1 is a schematic flow chart of a memory anomaly prediction method based on multi-algorithm decision provided by an embodiment of the present application;

[0060] Figure 2 is a schematic structural diagram of a deep time series extraction network provided by an embodiment of the present application;

[0061] Figure 3 is a schematic flow chart of finding an optimal model through a swarm optimization algorithm provided by an embodiment of the present application;

[0062] Figure 4 is a schematic structural diagram of a memory fault prediction model provided by an embodiment of the present application;

[0063] Figure 5 is a schematic flow chart of memory fault model training and verification provided by an embodiment of the present application;

[0064] Figure 6It is a schematic structural diagram of a memory anomaly prediction device based on multi - algorithm decision - making provided by an embodiment of the present application;

[0065] Figure 7 It is a schematic structural diagram of an electronic device provided by an embodiment of the present application. Detailed implementation manners

[0066] Next, the technical solutions in the embodiments of the present application will be clearly described with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of them. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art belong to the scope of protection of the present application.

[0067] Terms such as "first" and "second" in the specification and claims of the present application are used to distinguish similar objects, rather than to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of the same category, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims means at least one of the connected objects, and the character " / " generally indicates an "or" relationship between the associated objects before and after.

[0068] Next, with reference to the accompanying drawings, the memory anomaly prediction method based on multi - algorithm decision - making, the memory anomaly prediction device based on multi - algorithm decision - making, the electronic device, and the readable storage medium provided by the embodiments of the present application will be described in detail through specific embodiments and their application scenarios.

[0069] Among them, the memory anomaly prediction method based on multi - algorithm decision - making can be applied to a terminal, and specifically can be executed by hardware or software in the terminal.

[0070] The terminal includes, but is not limited to, portable communication devices such as mobile phones or tablets with a touch - sensitive surface (for example, a touch - screen display and / or a touchpad). It should also be understood that in some embodiments, the terminal may not be a portable communication device, but a desktop computer with a touch - sensitive surface (for example, a touch - screen display and / or a touchpad).

[0071] In the following various embodiments, a terminal including a display and a touch - sensitive surface is described. However, it should be understood that the terminal may include one or more other physical user interface devices such as a physical keyboard, a mouse, and a joystick.

[0072] The memory anomaly prediction method based on multi-algorithm decision provided by the embodiments of the present application. The execution subject of the memory anomaly prediction method based on multi-algorithm decision can be an electronic device or a functional module or functional entity in the electronic device that can implement the memory anomaly prediction method based on multi-algorithm decision. The electronic devices mentioned in the embodiments of the present application include, but are not limited to, mobile phones, tablet computers, computers, cameras, wearable devices, etc. Hereinafter, taking the electronic device as the execution subject as an example, the memory anomaly prediction method provided by the embodiments of the present application will be described.

[0073] The following are several specific problems existing in the related technical solutions:

[0074] (1) Insufficient modeling of time series features. Existing models are difficult to capture the temporal dynamic associations of memory operation data, resulting in limited prediction accuracy.

[0075] Traditional memory fault prediction methods mainly rely on static feature analysis and lack in-depth mining of time series features in operation data. Especially in scenarios with significant temporal correlations such as server memory faults, existing models are difficult to effectively capture the dynamic coupling relationship between short-term sudden anomaly signals and long-term degradation trends. Due to the lack of a multi-dimensional time window analysis mechanism, there is an obvious lag in the early identification of precursor features of fault occurrence, resulting in insufficient timeliness and accuracy of prediction results. More seriously, interference factors such as load fluctuations and temperature changes in the dynamic operating environment will further exacerbate the problem of time series feature drift, making traditional static modeling methods face severe generalization challenges.

[0076] (2) Simplification of optimization objectives. Existing methods overly pursue a single performance index, restricting the synergistic effect of multi-algorithms.

[0077] Current optimization strategies based on machine learning generally have design defects in the objective function, overly focusing on a single index such as prediction accuracy and ignoring the key impact of model diversity on the effect of ensemble learning. This one-sided optimization leads to serious performance homogenization during the multi-algorithm fusion process and is difficult to form effective complementary advantages.

[0078] (3) Imbalanced allocation of computing resources, that is, the centralized architecture is difficult to meet the real-time warning requirements.

[0079] The centralized computing mode adopted by mainstream prediction systems has significant bottlenecks. The transmission and processing of massive memory operation data generate huge time delays, resulting in low real-time parsing efficiency of key fault features. Especially when processing high-frequency time series data, the computing load of the central server increases exponentially, seriously affecting the timeliness of warning responses. More seriously, this architecture lacks an intelligent cooperation mechanism for edge computing nodes and cannot perform dynamic resource scheduling according to the real-time state of the server cluster, resulting in a serious mismatch between the allocation of computing resources and the requirements of fault warning.

[0080] Figure 1 is a schematic flowchart of the memory anomaly prediction method based on multi-algorithm decision provided by an embodiment of the present application. As Figure 1 shown, the memory anomaly prediction method based on multi-algorithm decision includes: step 110, step 120, and step 130.

[0081] Step 110: Obtain the data information generated during the operation of the server, preprocess and extract features from the data information to obtain a feature set, where the feature set includes count dimension features, component dimension features, statistical dimension features, and device configuration dimension features;

[0082] It is easy to understand that the electronic device obtains the data information generated during the operation of the server, and preprocesses and extracts features from the data information.

[0083] Optionally, in the data information preprocessing stage, to improve the quality of the data information and perform effective cleaning, the linear interpolation method can be used to fill in the missing values. For example, by analyzing the distribution characteristics of adjacent data points and calculating the estimated value of the missing position, the data filling rate can reach 98%. For the data information with the key feature missing more than 30%, the elimination process can be adopted, and the elimination ratio is 2%.

[0084] In one embodiment, for the outlier detection of the data information, the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) density clustering algorithm can be selected. For example, setting the neighborhood radius ε = 0.5 and the minimum number of data information MinPts = 5 as the clustering parameters, about 1.5% of the outlier data information can be identified and eliminated through density analysis, which can improve the consistency of the data quality.

[0085] Furthermore, after preprocessing the data information, feature extraction is performed on the preprocessed data information to obtain a feature set, where the feature set includes count dimension features, component dimension features, statistical dimension features, and device configuration dimension features.

[0086] The count dimension features reveal the occurrence pattern of errors by setting a feature time window, counting the occurrence frequencies of various errors such as error correction code errors, hard errors, soft errors, read errors, and cleaning errors, and calculating the time intervals of these errors at the same time.

[0087] The component dimension features extract the abnormal feature information of each component by quantitatively analyzing the operating states of hardware units (including sockets, channels, groups, rows, columns, and cells) within the time window.

[0088] The statistical dimension features quantify the central tendency and fluctuation range of the error distribution by focusing on the error correction code errors of six types of key components and calculating statistics such as the mean, median, and standard deviation.

[0089] The device configuration dimension features construct a complete feature system by combining the information of the dynamic random access memory model, the number of memory modules, and the server manufacturer, and converting this non-numerical information into a numerical form through one-hot encoding technology.

[0090] Step 120: Input the feature set into the trained basic prediction model to obtain the first prediction probability, the second prediction probability, and the third prediction probability. The basic prediction model includes a random forest sub-model, a lightweight gradient boosting sub-model, and an extreme gradient boosting sub-model.

[0091] Based on the first prediction probability, the second prediction probability, and the third prediction probability, obtain the first prediction probability value.

[0092] Input the feature set into the deep time series extraction network to obtain the target feature based on the feature masking method.

[0093] Input the target feature into the gated recurrent unit to obtain the fourth prediction probability value.

[0094] Based on the first prediction probability value and the fourth prediction probability value, obtain the decision matrix.

[0095] Furthermore, input the feature set into the trained basic prediction model for modeling analysis. The basic prediction model includes a random forest sub-model, a lightweight gradient boosting sub-model, and an extreme gradient boosting sub-model. Input the feature set into the random forest sub-model to obtain the first prediction probability, input the feature set into the lightweight gradient boosting sub-model to obtain the second prediction probability, input the feature set into the extreme gradient boosting sub-model to obtain the third prediction probability, and fuse the first prediction probability, the second prediction probability, and the third prediction probability according to the weighted fusion formula to obtain the first prediction probability value.

[0096] Meanwhile, input the feature set into the deep time series extraction network for feature learning. Figure 2It is a schematic structural diagram of the deep time series extraction network provided by an embodiment of the present application. The deep time series extraction network includes a feature extraction layer, an intermediate compression layer, and a feature reconstruction layer. The feature extraction layer adopts a three-layer time series convolution structure to gradually extract the time series correlation patterns of the feature set, and at the same time realizes the gradual compression of the feature dimension, reducing the 16-dimensional feature to 8 dimensions and 4 dimensions in sequence. The intermediate compression layer is responsible for further condensing the feature vector. The feature reconstruction layer restores the original feature dimension through three-layer transposed convolution operations, ensuring the integrity of the feature expression, and optimizing the feature extraction effect through the contrast learning method. Finally, a reconstructed feature matrix is output. The shape of the reconstructed feature matrix is the same as that of the input feature set, representing an efficient expression of the core pattern and time series correlation of the original feature.

[0097] Further, through the feature masking method, the contribution degree of each feature in the reconstructed feature matrix to the reconstruction error is evaluated one by one, its importance score is calculated, and the features are sorted according to importance to obtain the target matrix. The target matrix is input into the gated recurrent unit to obtain the fourth prediction probability value.

[0098] The first prediction probability and the fourth prediction probability value are fused by weighting to obtain a 1×2-dimensional decision matrix. Each dimension corresponds to different levels of information after fusion, including the prediction information of the basic prediction model and the in-depth analysis result of the gated recurrent unit on the feature set.

[0099] Step 130: Input the decision matrix into the trained decision optimizer logistic regression model to obtain the memory anomaly prediction result.

[0100] Finally, input the decision matrix into the trained decision optimizer logistic regression model. The decision optimizer logistic regression model further explores the potential associations between the features in the decision matrix through linear weighting and non-linear transformation to obtain the memory anomaly prediction probability. According to the magnitude relationship between the memory anomaly prediction probability and the preset threshold, the memory anomaly prediction result is obtained.

[0101] According to the memory anomaly prediction method based on multi-algorithm decision provided by an embodiment of the present application, by acquiring the data information generated during the operation of the server, preprocessing and feature extraction are performed on these data, combined with multiple basic prediction models such as random forest, light gradient boosting machine, and extreme gradient boosting, and deep learning methods such as deep time series extraction network and gated recurrent unit. The advantages of different algorithms in specific features and data patterns are utilized to more comprehensively explore the complex relationships between memory fault features. The fusion of the results of the basic prediction model and the target features further enhances the ability to capture the complex patterns of memory faults. Finally, the decision optimizer logistic regression model is used to optimize the fusion result through linear weighting and non-linear transformation, improving the accuracy and real-time performance of memory fault prediction.

[0102] In some embodiments, inputting the feature set into the trained basic prediction model to obtain the first prediction probability, the second prediction probability, and the third prediction probability includes:

[0103] Inputting the feature set into the random forest sub-model to obtain the first prediction probability;

[0104] Inputting the feature set into the light gradient boosting machine sub-model to obtain the second prediction probability;

[0105] Inputting the feature set into the extreme gradient boosting sub-model to obtain the third prediction probability.

[0106] It is easy to understand that when the feature set is input into the trained basic prediction model, the random forest sub-model, the light gradient boosting machine sub-model, and the extreme gradient boosting sub-model respectively perform modeling analysis on the feature set. The specific process is as follows:

[0107] Inputting the feature set into the random forest sub-model, the first prediction probability is obtained through the following formula:

[0108]

[0109] where T is the number of decision trees in the random forest sub-model, P t (X) is the prediction probability of the t-th decision tree for X, and P RF (X) is the first prediction probability.

[0110] Inputting the feature set into the light gradient boosting machine sub-model, the second prediction probability is obtained through the following formula:

[0111]

[0112] where M is the number of weak learners in the light gradient boosting machine sub-model, g i (X) is the output of the i-th weak learner, and P LGB (X) is the second prediction probability.

[0113] Inputting the feature set into the extreme gradient boosting sub-model, the third prediction probability is obtained through the following formula:

[0114]

[0115] where N is the number of base models, h j (X) is the output of the j-th base model, and P XGB (X) is the third prediction probability.

[0116] In this embodiment, the feature set is input into the random forest, light gradient boosting machine, and extreme gradient boosting sub-models to obtain the first prediction probability, the second prediction probability, and the third prediction probability respectively. The probability output method can better reflect the confidence level of the category. By combining the output results of multiple prediction models, the accuracy and stability of the prediction results are improved.

[0117] In some embodiments, obtaining the first prediction probability value based on the first prediction probability, the second prediction probability, and the third prediction probability includes:

[0118] Based on the first prediction probability, the second prediction probability, and the third prediction probability, the first prediction probability value is obtained according to the following formula:

[0119] P fusion (X) = α RF ·P RF (X) + α LGB ·P LGB (X) + α XGB ·P XGB (X)

[0120] Wherein, P fusion (X) is the first prediction probability value, α RF is the influence factor of the random forest sub-model, P RF (X) is the first prediction probability, α LGB is the influence factor of the light gradient boosting machine sub-model, P LGB (X) is the second prediction probability, α XGB is the influence factor of the extreme gradient boosting sub-model, P XGB (X) is the third prediction probability.

[0121] It is easy to understand that after obtaining the first prediction probability, the second prediction probability, and the third prediction probability, the three prediction probabilities are weighted and integrated through influence factor allocation to generate the first prediction probability value. α RF , α LGB , α XGB are the influence factors of the random forest sub-model, the light gradient boosting machine sub-model, and the extreme gradient boosting sub-model respectively, and satisfy α RF + α LGB + α XGB = 1.

[0122] In this embodiment, the first prediction probability value is obtained through weighted fusion calculation based on the first prediction probability, the second prediction probability, and the third prediction probability, balancing the prediction characteristics of the random forest, the light gradient boosting machine, and the extreme gradient boosting sub-models, effectively reducing the decision bias of a single algorithm, improving the accuracy and stability of the prediction, and optimizing the prediction performance.

[0123] In some embodiments, inputting the feature set into the deep temporal extraction network and obtaining target features based on the feature masking method includes:

[0124] Inputting the feature set into the deep temporal extraction network to obtain a reconstructed feature matrix;

[0125] Evaluating the reconstructed feature matrix based on the feature masking method to obtain the importance score of each feature in the reconstructed feature matrix;

[0126] Sorting each feature in the reconstructed feature matrix based on the importance score of each feature, and taking the top N features as the target features;

[0127] where N is a positive integer greater than zero.

[0128] In one embodiment, the feature set is input into the deep temporal extraction network to obtain a reconstructed feature matrix. The value of the feature is quantified by setting specific features to zero and analyzing the impact on the model reconstruction result. The compressed features are sorted according to the importance score, and the top 5 features are selected as the output result of the deep temporal extraction network. The 5 features are the number of read errors, the number of cleaning errors, the number of channel errors, the number of cell errors, and the number of row errors within the feature window (5 minutes). These high-value features effectively capture the key temporal patterns of memory operation, including the temporal dependencies of error events, the changing trends of error counts, and the evolution characteristics of hardware states over time, providing high-quality feature support for the prediction model.

[0129] In this embodiment, by inputting the feature set into the deep temporal extraction network, the deep temporal extraction network extracts the temporal correlation patterns in the input features through temporal convolution, and effectively extracts the key temporal features by combining the compression and reconstruction mechanisms. It can capture instantaneous anomalies and long-term trend changes, generate key features with high contributions, evaluate the importance of each feature in the reconstructed feature matrix using the feature masking method, sort to obtain the target features, improve the accuracy and stability of prediction, and optimize the prediction performance.

[0130] In some embodiments, inputting the target features into the gated recurrent unit and obtaining the fourth prediction probability value includes:

[0131] Inputting the target features into the gated recurrent unit, and obtaining the fourth prediction probability value according to the following formula:

[0132] P GRU (X) = softmax(W o ·h t +b o )

[0133] Among them, P GRU (X) is the fourth prediction probability value, W o is the weight of the output layer, b o is the bias of the output layer, and h t is the hidden state of the gated recurrent unit.

[0134] It should be noted that the target feature is input into the gated recurrent unit, and the gated recurrent unit captures the dynamic pattern and long-term dependence of the target feature changing over time through its gating mechanism. The output of the gated recurrent unit is a soft probability value, that is, a comprehensive representation of the time series feature.

[0135] In this embodiment, by inputting the target feature into the gated recurrent unit and calculating the fourth prediction probability value using the hidden state and weight of the gated recurrent unit, the changing trend and relationship of the feature over time can be effectively reflected, the prediction ability for time series data is improved, and thus the accuracy and stability of memory fault prediction are improved.

[0136] In some embodiments, the inputting the decision matrix into the trained decision optimizer logistic regression model to obtain the memory anomaly prediction result includes:

[0137] Inputting the decision matrix into the trained decision optimizer logistic regression model, and obtaining the memory anomaly prediction probability according to the following formula:

[0138]

[0139] Among them, P final (X) is the memory anomaly prediction probability, W is the weight of the decision optimizer logistic regression model, b is the bias term of the decision optimizer logistic regression model, and M is the transpose of the decision matrix;

[0140] Based on the memory anomaly prediction probability and the preset threshold, the memory anomaly prediction result is obtained.

[0141] It is easy to understand that based on the first prediction probability value and the fourth prediction probability value, a decision matrix with a dimension of 1×2 is constructed, and each dimension corresponds to different levels of information after fusion. The calculation formula of the decision matrix is as follows:

[0142] M = [P fusion (X) P GRU (X)]

[0143] Among them, M is the decision matrix, P GRU (X) is the fourth prediction probability value, and P fusion (X) is the first prediction probability value.

[0144] It should be noted that the decision matrix not only contains the core prediction information of the basic algorithm, but also integrates the in-depth analysis results of the gated time unit for dynamic features, providing more comprehensive decision support for the prediction task.

[0145] Furthermore, the decision matrix is input into the trained decision optimizer logistic regression model to optimize the final prediction result. The decision optimizer logistic regression model further explores the potential associations between input features through linear weighting and non-linear transformation, generating the prediction probability of memory faults.

[0146] Finally, the decision optimizer logistic regression model combines the fusion results of multiple algorithms and the timing information provided by the gated recurrent unit to conduct a more comprehensive assessment of the memory fault state. The memory anomaly prediction result is obtained based on the comparison between the generated prediction probability and the preset threshold. When the prediction probability exceeds the preset threshold, it is determined that the memory may be about to fail; conversely, if the prediction probability does not exceed the preset threshold, it is determined that the memory is in a normal operating state.

[0147] In this embodiment, by inputting the decision matrix into the trained decision optimizer logistic regression model, the memory anomaly prediction probability is obtained, and whether the memory is abnormal is judged according to the preset threshold. The prediction of memory anomalies is realized, the accuracy of memory anomaly prediction is improved, potential memory problems are identified in advance, and the failure risk is reduced.

[0148] In some embodiments, the training methods of the basic prediction model and the decision optimizer logistic regression model include:

[0149] Define the search space of the basic prediction model based on the probability selection method of the guiding factor;

[0150] Obtain the combined parameters based on the search space, and train the basic prediction model according to the first data set;

[0151] Obtain the second data set based on the output of the basic prediction model and the output of the gated recurrent unit;

[0152] Train the decision optimizer logistic regression model according to the second data set, and gradually guide the decision optimizer logistic regression model and the basic prediction model to converge to the optimal solution based on the dynamic guiding factor attenuation mechanism, optimizing the prediction performance index and the difference gain.

[0153] In one embodiment, error information collected from more than 250,000 servers over an eight - month period can be collected as the first data set. The first data set includes three parts: memory error logs, memory fault lists, and memory configuration tables. Memory error logs can be collected by the mcelog tool, which is a Linux tool based on the machine check architecture and is used to monitor and record memory errors. By running the mcelog daemon process, this tool details all memory error event information, including server identification, memory module number, row number, column number, and the source of error detection (such as whether detected by the memory scrubber) or the type of memory read - write operation. The memory fault list is generated by the background monitoring daemon process, which is used to monitor system - level abnormal events (such as server crashes) and send system event logs to the centralized maintenance system. This system checks for any server failures through rule - based detection. The fault list records the server ID timestamp and the fault type. The memory configuration table records the server ID, memory model, server manufacturer, and the amount of memory.

[0154] It is worth noting that there may be a significant class imbalance in the first data set, that is, the number of negative samples is much larger than the number of positive samples. A down - sampling strategy can be used to adjust the sample distribution. By using the random sampling method, the number of negative samples is reduced so that the ratio of positive to negative samples is controlled at 1:50. On the premise of ensuring sufficient information of negative samples, the recognition ability of the model for the minority class is improved. This sampling ratio achieves a good compromise between balancing the sample distribution and retaining the data information volume, effectively improving the prediction performance of the model on the imbalanced data set.

[0155] Furthermore, in the first data set construction stage, first, the eight - month time - series data is segmented according to a fixed 5 - minute time window. The statistical data within each window interval constitutes an independent sample unit. The generated feature dimensions include a count dimension, a component dimension, a statistical dimension, and a device configuration dimension. To eliminate the scale differences between different features, all feature values can be normalized, and the numerical range is uniformly mapped to the [0,1] interval to improve the stability of training.

[0156] Optionally, during the division of the first data set, a time - sliding validation strategy can be adopted: in the first stage, data from January to May is selected to construct the training set, and data from June is used as the validation set; in the second stage, the time window is shifted backward, data from February to June is used for training, and data from July is used for validation; finally, data from March to July is used for training, and data from August is used for validation. This progressive validation method fully evaluates the model at different time spans and effectively tests the adaptability of the algorithm to the dynamic characteristics of time - series data.

[0157] It should be noted that during the training process of the basic prediction model and the decision optimizer logistic regression model, a swarm optimization algorithm is adopted. Based on the dynamic guiding factor attenuation mechanism, the prediction performance index and the difference gain are optimized, and the basic prediction model and the decision optimizer logistic regression model are gradually guided to converge to the optimal solution. Figure 3 It is a schematic flowchart of finding the optimal model through the swarm optimization algorithm provided by the embodiment of the present application. The specific process is as follows:

[0158] (1) Initialize the swarm optimization parameters, set the swarm optimization parameters, and the formula for the swarm optimization parameters is as follows:

[0159] Φ = {τ 0 , ρ 0 , n ants , n iter}

[0160] Among them, τ 0 is the guiding factor, ρ 0 is the attenuation rate, n ants is the number of search individuals, n iter is the maximum number of iteration rounds, and Φ represents the swarm optimization parameters.

[0161] (2) Randomly select a path and construct the basic prediction model

[0162] First, define the search space of the basic prediction model. The specific process is as follows:

[0163] Define the search spaces of the random forest sub-model, the light gradient boosting machine sub-model, and the extreme gradient boosting sub-model respectively.

[0164] The formula for the search space of the random forest sub-model is as follows:

[0165] Θ RF = {n estimators , max_depth, min_samples_split}

[0166] Among them, n estimators is the number of decision trees in the random forest, max_depth is the maximum depth of the decision tree, min_samples_split is the minimum number of samples required for node splitting, and Θ RF is the search space of the random forest sub-model.

[0167] The formula for the search space of the light gradient boosting machine sub-model is as follows:

[0168] Θ LGB = {n estimators , max_depth, learning_rate}

[0169] where n estimators is the number of weak learners, max_depthwe is the maximum depth of the tree, learning_rate is the learning rate, and Θ LGB is the search space of the lightweight gradient boosting submodel.

[0170] The formula for the search space of the extreme gradient boosting submodel is as follows:

[0171] Θ XGB = {n estimators , max_depth, subsample}

[0172] where n estimators is the number of base learners, max_depth is the maximum depth of the tree, subsample is the proportion of data used for building each tree, and Θ XGB is the search space of the extreme gradient boosting submodel.

[0173] Furthermore, based on the probability selection method of the guiding factor, parameter combinations are selected in the search space of the basic prediction model, and the basic prediction model is trained according to the first dataset. The specific process is as follows:

[0174] Path selection is performed based on the guiding factor, and parameter combinations are selected in the search space through the probability distribution of the guiding factor. The formula for the path selection probability is as follows:

[0175]

[0176] where p i is the probability of selecting the i-th path, τ i is the guiding factor of the i-th path, indicating the priority of this path,

[0177] and n is the number of optional parameter combinations in the search space.

[0178] The random forest submodel constructs multiple decision trees by randomly selecting samples and features in the first dataset. Each decision tree outputs a prediction result through the voting method, and the model is optimized by combining parameter adjustment. The training formula of the random forest submodel is as follows:

[0179] P RF = RandomForest(n estimators , max_depth, min_samples_split)

[0180] where P RF is the output prediction probability of the random forest model, and n estimatorsis the number of decision trees, max_depth is the maximum depth of the tree, and min_samples_split is the minimum number of samples required to split an internal node.

[0181] The lightweight gradient boosting machine sub-model uses the histogram algorithm to split nodes, quickly locate the optimal split point, updates the model through gradient information in each iteration, and optimizes the overall prediction performance. The training formula of the lightweight gradient boosting machine sub-model is as follows:

[0182] P LGB = LightGBM(n estimators , max_depth, learning_rate)

[0183] where: P LGB is the output prediction probability of the lightweight gradient boosting machine sub-model, n estimators is the number of weak learners, max_depth is the maximum depth of the tree, and learning_rate is the learning rate.

[0184] The extreme gradient boosting sub-model utilizes an incremental tree structure. Each tree sequentially learns the residuals of the previous tree, controls overfitting through regularization constraints, and combines the subsample strategy to improve computational efficiency. The training formula of the extreme gradient boosting sub-model is as follows:

[0185] P XGB = XGBoost(n estimators , max_depth, subsample)

[0186] where, P XGB is the output prediction probability of the extreme gradient boosting sub-model, n estimators is the number of base learners, max_depth is the maximum depth of the tree, and subsample is the subsample ratio.

[0187] (3) Construct a decision optimizer logistic regression model

[0188] Meanwhile, based on the output of the basic prediction model and the output of the gated recurrent unit, a second data set is obtained, and the decision optimizer logistic regression model is trained according to the second data set. The training process is as follows:

[0189] The formula for the second data set is as follows:

[0190] X meta = [P RF , P LGB , P XGB , P GRU

[0191] where, P GRU ​(X) is the fourth predicted probability value, P RF (X) is the first predicted probability, P LGB (X) is the second predicted probability, P XGB (X) is the third predicted probability, X meta is the second data set.

[0192] The decision optimizer logistic regression model aggregates the outputs of the base prediction model and the gated recurrent unit as the input of the meta-classifier. Logistic regression adjusts the input features through linear weights, combines with the Sigmoid function to generate the final predicted probability, and optimizes the logistic regression parameters according to the validation set performance to improve the classification performance. The training formula of the decision optimizer logistic regression model is as follows:

[0193]

[0194] Among them, X meta is the second data set, is the output of the decision optimizer logistic regression model.

[0195] (4) Calculate the difference gain and the dynamic decay rate

[0196] The calculation formula of the difference gain is as follows:

[0197]

[0198] Among them, y i is the prediction result of the base learner i, y j is the prediction result of the base learner j, and H is the difference gain.

[0199] Furthermore, combining the F1 score and the difference gain, dynamically adjust the volatility rate. In the early stage of training, pay more attention to the difference gain. In the later stage, gradually tilt towards optimizing the F1 score. The formula of the dynamic decay rate is as follows:

[0200]

[0201] Among them, ΔF1 is the change in the current F1 score, ΔH is the change in the current difference gain, ρ 0 is the guiding factor, and t is the current iteration number.

[0202] (5) Update the guiding factor, and based on the dynamic decay rate, update the guiding factor values on the path. Combine the performance score and the dynamic decay rate to enhance the convergence characteristics of the optimal path. The formula for updating the decay rate is as follows:

[0203]

[0204] Among them, is the pheromone value on path i after the (t + 1)-th iteration, is the pheromone value on path i after the t-th iteration, ρ t is the dynamic evaporation rate, and ΔF1 is the increment of the F1 score of the prediction model corresponding to the current path i in this iteration.

[0205] (6) Repeat the execution of path selection, algorithm training, difference calculation, dynamic decay adjustment, and guiding factor update until the maximum number of iterations is reached or the performance index converges.

[0206] (7) Output the trained basic prediction model and the decision optimization logistic regression model.

[0207] In one embodiment, Figure 4 is a schematic structural diagram of the memory fault prediction model provided by the embodiment of the present application. As Figure 4 shown, the memory fault prediction model includes a basic prediction model, a deep time series extraction network, a gated recurrent unit, and a decision optimization logistic regression model. The basic prediction model includes a random forest sub-model, a lightweight gradient boosting sub-model, and an extreme gradient sub-model.

[0208] The memory fault prediction model is loaded into the model through a pre-saved file (such as.pkl or.h5) to make its parameters and structure consistent with the training stage. Subsequently, the actual running server data is input into the memory fault prediction model, and the server data is preprocessed and feature extracted to obtain a feature set. After the feature set is normalized or standardized, data that meets the input requirements of the memory fault prediction model is formed.

[0209] Furthermore, the preprocessed data is sequentially input into the loaded basic prediction model. Each sub-model of the basic prediction model independently completes the prediction and generates the first prediction probability, the second prediction probability, and the third prediction probability, and the first prediction probability value is obtained through a weighted fusion formula. At the same time, the feature set is input into the deep time series extraction network, and after obtaining the target feature, it is input into the gated recurrent unit. The gated recurrent unit captures the dynamic changes of the time feature through its gating mechanism and obtains the fourth prediction probability value.

[0210] Finally, a decision matrix is obtained through the first prediction probability value and the fourth prediction probability value, and is optimized through the decision optimization logistic regression model to generate a memory anomaly prediction probability. According to the comparison between the memory anomaly prediction probability and the preset threshold, when the memory anomaly prediction probability exceeds the threshold, it is determined that the memory is about to fail, otherwise it is determined that the memory is operating normally.

[0211] Figure 5 is a schematic flowchart of the training and verification of the memory fault model provided by the embodiment of the present application. As Figure 5As shown in the figure, the first data set is preprocessed and feature generated to obtain the first training set and the second test set. The basic prediction model is trained through sliding window cross - validation. The outputs of the basic prediction model and the gated recurrent unit model are used as the second data set, which is preprocessed to obtain the second training set and the second test set to train the decision optimizer logistic regression model, and the output is the first data set.

[0212] In this embodiment, the prediction performance and the model difference degree are co - optimized through the dynamic guiding factor decay mechanism, and the gated recurrent unit is combined to realize the mining of feature time - series correlation. The optimal combination parameters are generated based on the probabilistic search space, which improves the synergy effect and generalization ability of the multi - algorithm fusion. Through the iterative co - training of the decision optimizer and the basic prediction model, the dynamic balance between the algorithm accuracy and diversity is realized. Taking the prediction performance and the model difference degree as the optimization objectives and combining the dynamic guiding factor decay mechanism, multi - objective co - optimization is realized. The adaptive weight allocation mechanism is introduced. In the initial stage of optimization, it focuses on improving the difference degree of the basic prediction algorithm, exploring solutions with higher diversity to enhance the generalization ability of the system. In the later stage of optimization, the algorithm gradually tends to improve the prediction performance. By dynamically adjusting the decay rate, the influencing factor can be flexibly adjusted according to the optimization stage to balance the relationship between the prediction performance and the difference degree. The dynamic update strategy of the guiding factor can reflect the change of the effect in real time during the optimization process, improving the accuracy and robustness of the memory fault prediction in a complex operating environment, and effectively reducing the resource consumption and convergence time of model training.

[0213] In the embodiment of the present application, the memory anomaly prediction method based on multi - algorithm decision - making, the execution subject can be a memory anomaly prediction device based on multi - algorithm decision - making. In the embodiment of the present application, taking the memory anomaly prediction device based on multi - algorithm decision - making executing the memory anomaly prediction method based on multi - algorithm decision - making as an example, the memory anomaly prediction device based on multi - algorithm decision - making provided by the embodiment of the present application is described.

[0214] The embodiment of the present application also provides a memory anomaly prediction device based on multi - algorithm decision - making, as Figure 6 shown, the memory anomaly prediction device based on multi - algorithm decision - making includes: an acquisition module 610, a processing module 620, and a prediction module 630.

[0215] The acquisition module 610 is used to acquire the data information generated during the operation of the server, preprocess and extract features from the data information to obtain a feature set, and the feature set includes count - dimension features, component - dimension features, statistical - dimension features, and device - configuration - dimension features;

[0216] A processing module 620, configured to input the feature set into a trained basic prediction model to obtain a first prediction probability, a second prediction probability, and a third prediction probability, where the basic prediction model includes a random forest sub-model, a light gradient boosting machine sub-model, and an extreme gradient boosting sub-model;

[0217] Based on the first prediction probability, the second prediction probability, and the third prediction probability, obtain a first prediction probability value;

[0218] Input the feature set into a deep time series extraction network, and obtain target features based on the feature masking method;

[0219] Input the target features into a gated recurrent unit to obtain a fourth prediction probability value;

[0220] Based on the first prediction probability value and the fourth prediction probability value, obtain a decision matrix;

[0221] A prediction module 630, configured to input the decision matrix into a trained decision optimizer logistic regression model to obtain a memory anomaly prediction result.

[0222] According to the memory anomaly prediction device based on multi-algorithm decision provided by the embodiment of the present application, by acquiring the data information generated during the operation of the server, preprocessing and feature extraction of these data, combining multiple basic prediction models such as random forest, light gradient boosting machine, and extreme gradient boosting, and deep learning methods such as deep time series extraction network and gated recurrent unit, the advantages of different algorithms in specific features and data patterns are utilized, the complex relationships between memory fault features can be more comprehensively mined, the fusion of the results of the basic prediction model and the target features further enhances the ability to capture the complex patterns of memory faults. Finally, a decision optimizer logistic regression model is used to optimize the fusion result through linear weighting and non-linear transformation, improving the accuracy and real-time performance of memory fault prediction.

[0223] The memory anomaly prediction device based on multi-algorithm decision provided by the embodiment of the present application can implement Figures 1 to 5 each process implemented by the embodiment of the memory anomaly prediction method based on multi-algorithm decision. To avoid repetition, it will not be elaborated here.

[0224] In some embodiments, as Figure 7 shown, the embodiment of the present application further provides an electronic device 700, including a processor 701, a memory 702, and a computer program stored on the memory 702 and executable on the processor 701. When the program is executed by the processor 701, it implements each process of the embodiment of the memory anomaly prediction method based on multi-algorithm decision, and can achieve the same technical effect. To avoid repetition, it will not be elaborated here.

[0225] It should be noted that the electronic devices in the embodiments of the present application include the above-mentioned mobile electronic devices and non-mobile electronic devices.

[0226] The embodiments of the present application also provide a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements each process of the above-mentioned embodiment of the memory anomaly prediction method based on multi-algorithm decision-making, and can achieve the same technical effect. To avoid repetition, it will not be elaborated here.

[0227] Among them, the processor is the processor in the electronic device in the above-mentioned embodiment. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk or optical disc, etc.

[0228] The embodiments of the present application also provide a computer program product, including a computer program, which implements the above-mentioned memory anomaly prediction method based on multi-algorithm decision-making when executed by a processor.

[0229] Among them, the processor is the processor in the electronic device in the above-mentioned embodiment. The readable storage medium includes computer-readable storage media, such as computer read-only memory ROM, random access memory RAM, magnetic disk or optical disc, etc.

[0230] The embodiments of the present application further provide a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement each process of the above-mentioned embodiment of the memory anomaly prediction method based on multi-algorithm decision-making, and can achieve the same technical effect. To avoid repetition, it will not be elaborated here.

[0231] It should be understood that the chip mentioned in the embodiments of the present application may also be referred to as a system-on-chip, system chip, chip system or system-on-chip, etc.

[0232] It should be noted that in this text, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device including a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including such element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in a reverse order according to the functions involved. For example, the described methods may be performed in an order different from that described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0233] From the description of the above embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions for causing a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the memory anomaly prediction method based on multiple algorithms in each embodiment of the present application.

[0234] In the description of the present application, the "first feature" and "second feature" may include one or more of such features.

[0235] In the description of the present application, the meaning of "a plurality" is two or more.

[0236] The embodiments of the present application have been described above in conjunction with the accompanying drawings. However, the present application is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Those of ordinary skill in the art, under the inspiration of the present application and without departing from the spirit and scope protected by the claims of the present application, can also make many forms, all of which fall within the protection scope of the present application.

[0237] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "schematic embodiments", "examples", "specific examples", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of this application. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0238] Although the embodiments of this application have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of this application, and the scope of this application is defined by the claims and their equivalents.

Claims

1. A memory anomaly prediction method based on multi-algorithm decision making, characterized in that: The method comprises: Acquire data information generated during the operation of the server, preprocess and extract features of the data information to obtain a feature set, wherein the feature set includes counting dimension features, component dimension features, statistical dimension features, and device configuration dimension features; Inputting the feature set into a trained basic prediction model to obtain a first prediction probability, a second prediction probability and a third prediction probability, wherein the basic prediction model includes a random forest sub-model, a lightweight gradient boosting machine sub-model and an extreme gradient boosting sub-model; Based on the first predicted probability, the second predicted probability and the third predicted probability, obtaining a first predicted probability value; Inputting the feature set into a deep temporal extraction network, and obtaining target features based on a feature masking method; Inputting the target feature into a gated recurrent unit to obtain a fourth predicted probability value; Based on the first predicted probability value and the fourth predicted probability value, obtaining a decision matrix; The decision matrix is ​​input into the trained decision optimizer logistic regression model to obtain the memory anomaly prediction result.

2. The memory anomaly prediction method based on multi-algorithm decision making according to claim 1 is characterized in that: The step of inputting the feature set into a trained basic prediction model to obtain a first prediction probability, a second prediction probability, and a third prediction probability includes: Inputting the feature set into the random forest sub-model to obtain the first prediction probability; Inputting the feature set into the lightweight gradient boosting sub-model to obtain the second prediction probability; The feature set is input into the extreme gradient boosting sub-model to obtain the third prediction probability.

3. The memory anomaly prediction method based on multi-algorithm decision making according to claim 1 is characterized in that: The obtaining a first predicted probability value based on the first predicted probability, the second predicted probability and the third predicted probability includes: Based on the first predicted probability, the second predicted probability and the third predicted probability, the first predicted probability value is obtained according to the following formula: P fusion (X)=a RF ·P RF (X)+a LGB ·P LGB (X)+a XGB ·P XGB (X) Among them, P fusion (X) is the first predicted probability value, α RF is the impact factor of the random forest sub-model, P RF (X) is the first predicted probability, α LGB is the impact factor of the lightweight gradient boosting sub-model, P LGB (X) is the second predicted probability, α XGB is the impact factor of the extreme gradient boosting sub-model, P XGB (X) is the third predicted probability.

4. The memory anomaly prediction method based on multi-algorithm decision making according to claim 1 is characterized in that: The step of inputting the feature set into a deep temporal extraction network and obtaining target features based on a feature masking method includes: Inputting the feature set into a deep temporal extraction network to obtain a reconstructed feature matrix; Evaluate the reconstructed feature matrix based on the feature masking method to obtain an importance score of each feature in the reconstructed feature matrix; Sorting each feature in the reconstructed feature matrix based on the importance score of each feature, and taking all the top N features in the sorting as target features; Wherein, N is a positive integer greater than zero.

5. The memory anomaly prediction method based on multi-algorithm decision making according to claim 1 is characterized in that: The step of inputting the target feature into a gated recurrent unit to obtain a fourth predicted probability value comprises: The target feature is input into the gated recurrent unit, and the fourth prediction probability value is obtained according to the following formula: P GRU (X)=softmax(W o ·h t +b o ) Among them, P GRU (X) is the fourth predicted probability value, W o is the weight of the output layer, b o is the bias of the output layer, h t is the hidden state of the gated recurrent unit.

6. The memory anomaly prediction method based on multi-algorithm decision making according to claim 1 is characterized in that: The step of inputting the decision matrix into a trained decision optimizer logistic regression model to obtain a memory anomaly prediction result includes: The decision matrix is ​​input into the trained decision optimizer logistic regression model, and the memory anomaly prediction probability is obtained according to the following formula: Among them, P final (X) is the predicted probability of memory anomaly, W is the weight of the decision optimizer logistic regression model, b is the bias term of the decision optimizer logistic regression model, and M is the transpose of the decision matrix; Based on the memory abnormality prediction probability and a preset threshold, the memory abnormality prediction result is obtained.

7. The memory anomaly prediction method based on multi-algorithm decision making according to claim 1 is characterized in that: The training method of the basic prediction model and the decision optimizer logistic regression model comprises: Defining a search space of the basic prediction model based on a probabilistic selection method of guiding factors; Obtaining a combination parameter based on the search space, and training the basic prediction model according to the first data set; Obtaining a second data set based on the output of the basic prediction model and the output of the gated recurrent unit; The decision optimizer logistic regression model is trained according to the second data set, and based on the dynamic guidance factor attenuation mechanism, the optimization prediction performance index and the difference gain, the decision optimizer logistic regression model and the basic prediction model are gradually guided to converge to the optimal solution.

8. A memory anomaly prediction device based on multi-algorithm decision making, implemented by the memory anomaly prediction method based on multi-algorithm decision making according to any one of claims 1 to 7, characterized in that: The device comprises: An acquisition module is used to acquire data information generated during the operation of the server, preprocess and extract features of the data information, and obtain a feature set, wherein the feature set includes counting dimension features, component dimension features, statistical dimension features, and device configuration dimension features; A processing module, configured to input the feature set into a trained basic prediction model to obtain a first prediction probability, a second prediction probability, and a third prediction probability, wherein the basic prediction model includes a random forest sub-model, a lightweight gradient boosting machine sub-model, and an extreme gradient boosting sub-model; Based on the first predicted probability, the second predicted probability and the third predicted probability, obtaining a first predicted probability value; Inputting the feature set into a deep temporal extraction network, and obtaining target features based on a feature masking method; Inputting the target feature into a gated recurrent unit to obtain a fourth predicted probability value; Based on the first predicted probability value and the fourth predicted probability value, obtaining a decision matrix; The prediction module is used to input the decision matrix into the trained decision optimizer logistic regression model to obtain a memory anomaly prediction result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the memory exception prediction method based on multi-algorithm decision-making as described in any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the memory exception prediction method based on multi-algorithm decision-making as described in any one of claims 1 to 7 is implemented.

Images