Equipment fault prediction model determination method and device, equipment, storage medium and product

Abstract

The invention provides an equipment fault prediction model determination method and device, equipment, a storage medium and a product. The method comprises the steps that a target neural network model used for equipment fault prediction and a disk data sample set are obtained, the disk data sample set comprises a plurality of disk data samples, and the disk data samples comprise disk state data of a disk and fault labels corresponding to the disk state data; the disk state data comprises feature data in multiple dimensions; inputting target feature data of each disk data sample into the target neural network model to obtain a contribution degree corresponding to the target feature data; constructing a contribution degree matrix according to the contribution degree corresponding to each piece of target feature data; and according to the contribution degree matrix, converting a target neural network model into a decision tree model taking the target feature data as nodes.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, device, equipment, storage medium and product for determining a device fault prediction model. Background Art

[0002] In the development of artificial intelligence technology, the neural network model is the current mainstream artificial intelligence model. The neural network model is a complex network system formed by a large number of simple processing units, that is, neurons that are widely connected to each other. It reflects many basic characteristics of human brain functions and is a highly complex nonlinear dynamic learning system. Neural networks have large-scale parallel, distributed storage and processing, self-organization, self-adaptation and self-learning capabilities, and are particularly suitable for processing imprecise and fuzzy information processing problems that require considering many factors and conditions at the same time. However, because there are a large number of neurons in the neural network model, it is basically impossible to explain and understand the internal decision-making process of the model, and it is impossible to effectively modify the model.

[0003] Neural network models have a wide range of applications. For example, they can be used to predict equipment faults. Although neural network models have a strong ability to learn data knowledge, they are black boxes and it is difficult to explain the internal decision-making process of the model. Therefore, even if deficiencies in the model's prediction results are found, effective intervention cannot be made, and it is difficult to accurately modify the model based on external intervention. Summary of the invention

[0004] The embodiments of the present application provide a method, apparatus, device, storage medium and product for determining an equipment fault prediction model, which can facilitate the modification of the model structure, thereby improving the accuracy of the model.

[0005] In a first aspect, the present application provides a method for determining an equipment failure prediction model, comprising:

[0006] Obtaining a target neural network model for device fault prediction and a disk data sample set, wherein the disk data sample set includes multiple disk data samples, and the disk data samples include disk status data of the disk and a fault label corresponding to the disk status data; the disk status data includes feature data in multiple dimensions;

[0007] Inputting the target feature data of each disk data sample into the target neural network model to obtain the contribution degree corresponding to the target feature data;

[0008] Constructing a contribution matrix according to the contribution corresponding to each of the target feature data;

[0009] According to the contribution matrix, the target neural network model is converted into a decision tree model with the target feature data as nodes.

[0010] In some possible implementations, inputting the target feature data of each sample in the disk data sample set into the target neural network model to obtain the contribution corresponding to the target feature data includes:

[0011] Inputting target feature data of each sample in the disk data sample set into the target neural network model to obtain a prediction score of the target feature data, wherein the prediction score indicates a predicted failure probability;

[0012] Calculate the entropy gain value of the feature corresponding to the target feature data according to the prediction score;

[0013] The entropy gain value is used as the contribution of the target feature data to the prediction result.

[0014] In some possible implementations, converting the target neural network model into a decision tree model with feature data as nodes according to the contribution matrix includes:

[0015] According to the order of the decision tree model from high to low levels, the nodes on each level are determined in turn until the preset stop condition is met:

[0016] According to the contribution matrix, the feature with the greatest contribution is selected from the initial feature set composed of feature data of multiple dimensions as a node of the current level;

[0017] According to the contribution segmentation threshold corresponding to the feature with the largest contribution, the features in the first target feature set are divided into a first subset and a second subset respectively; the first target feature set is a set of features remaining after removing the feature with the largest contribution from the initial feature set;

[0018] For both the first subset and the second subset, the following operations are performed:

[0019] Selecting the feature with the largest contribution from the target subset as the node of the next level of the current level, the target subset is the first subset or the second subset;

[0020] The second target feature set is updated to the initial feature set, and the feature with the largest contribution is selected as the node of the current level from the initial feature set composed of multiple dimensional feature data according to the contribution matrix; the second target set is the set of features remaining after removing the feature with the largest contribution from the target subset.

[0021] In some possible implementations, before inputting the target feature data of each sample in the disk data sample set into the target neural network model to obtain a prediction score of the target feature data, wherein the prediction score represents the predicted failure probability, the method further includes:

[0022] Target feature data is extracted from each sample in the disk data sample set based on a preset extraction rule, and the amount of the target feature data is less than the amount of feature data in the sample.

[0023] In some possible implementations, before obtaining the target neural network model for device fault prediction and the disk data sample set, the method further includes:

[0024] Acquire a training sample set; the training sample set includes a plurality of training samples, each of the training samples includes disk status data of a disk and a fault label corresponding to the disk status data;

[0025] For each training sample, perform the following steps:

[0026] Inputting the disk status data into a target neural network model to determine a prediction result of a disk failure condition;

[0027] Determining a loss function value of the target neural network model according to the prediction results of each of the training samples and the corresponding fault labels;

[0028] In the case where the loss function value of the target neural network model does not satisfy the preset training stop condition, the model parameters of the target neural network model are adjusted, and the step of returning to inputting the disk status data into the target neural network model to determine the prediction result of the disk failure condition is performed;

[0029] When the loss function value of the target neural network model meets the preset training stop condition, the training is stopped to obtain the trained target neural network model.

[0030] In some possible implementations, obtaining a training sample set includes:

[0031] Acquire initial disk status data and a fault label corresponding to the initial disk status data;

[0032] In the case where there are missing values ​​in the initial disk status data, the missing values ​​are filled in by a preset interpolation method to obtain the corrected first disk status data;

[0033] In the case that the initial disk status data contains a value beyond a preset range, the value beyond the preset range is replaced by a preset boundary value to obtain a corrected second disk status data;

[0034] Using the first disk status data or the second disk status data and the corresponding fault label as training samples;

[0035] A training sample set is constructed according to the training samples.

[0036] In some possible implementations, after converting the target neural network model into a decision tree model with feature data as nodes according to the contribution matrix, the method further includes:

[0037] Acquire actual status information of the target disk and actual fault information of the target disk;

[0038] Inputting the actual status information of the target disk into the decision tree model to obtain predicted fault information of the target disk;

[0039] Determining the accuracy of the decision tree model according to the difference between the predicted fault information and the actual fault information;

[0040] When the accuracy is lower than a threshold, the decision tree model is updated in response to an update operation of the decision tree model by a user.

[0041] In some possible implementations, the updating of the decision tree model includes: modifying the decision tree model based on target modification information, incrementally training the target neural network model, and fully training the target neural network model.

[0042] In a second aspect, the present application provides a device for determining a device failure prediction model, the device comprising:

[0043] A first acquisition module is used to acquire a target neural network model for device fault prediction and a disk data sample set, wherein the disk data sample set includes multiple disk data samples, and the disk data samples include disk status data of the disk and a fault label corresponding to the disk status data; the disk status data includes feature data in multiple dimensions;

[0044] A contribution determination module, used for inputting the target feature data of each disk data sample into the target neural network model to obtain the contribution corresponding to the target feature data;

[0045] A construction module, used to construct a contribution matrix according to the contribution corresponding to each of the target feature data;

[0046] A conversion module is used to convert the target neural network model into a decision tree model with the target feature data as nodes according to the contribution matrix.

[0047] In a third aspect, the present application provides a device for determining a device fault prediction model, the device comprising: a processor, and a memory storing computer program instructions; the processor reads and executes the computer program instructions to implement the device fault prediction model determination method as described above.

[0048] In a fourth aspect, the present application provides a computer-readable storage medium having computer program instructions stored thereon, and when the computer program instructions are executed by a processor, the method for determining a device fault prediction model as described above is implemented.

[0049] In a fifth aspect, the present application provides a computer program product. When the instructions in the computer program product are executed by a processor of an electronic device, the electronic device executes the device fault prediction model determination method as described above.

[0050] The equipment fault prediction model determination method, apparatus, device, storage medium and product provided in the embodiments of the present application can convert the neural network model used for equipment fault prediction into a decision tree model with feature data as nodes based on the contribution of feature data in the training sample to the prediction result. The hierarchical structure of the decision tree model is relatively easy for users to perceive, and is no longer a "black box structure" like a neural network. Therefore, fault prediction based on the decision tree model can clearly display the processing process of the decision tree model, allowing users to more easily explain the decision-making process within the model, and can also more easily change the model structure, which is conducive to the modification of the model structure, thereby improving the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The present application can be better understood from the following description of the specific embodiments of the present application in conjunction with the accompanying drawings, in which:

[0052] Other features, objects and advantages of the present application will become more apparent by reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings, in which the same or similar reference numerals represent the same or similar features.

[0053] Figure 1 is a flow chart of a method for determining a device failure prediction model provided by an embodiment of the present application;

[0054] Figure 2 is a specific flow chart of converting a target neural network model into a decision tree model provided by another embodiment of the present application;

[0055] Figure 3 is a flow chart of a target neural network model training process provided by another embodiment of the present application;

[0056] Figure 4 is a schematic diagram of the system architecture of a data collection process provided by another embodiment of the present application;

[0057] Figure 5 is a schematic diagram of a decision tree model provided by an embodiment of the present application;

[0058] Figure 6 It is a structural schematic diagram of a device for determining a device fault prediction model provided by an embodiment of the present application;

[0059] Figure 7 It is a schematic diagram of the hardware structure of the equipment failure prediction model determination equipment provided in the embodiment of the present application. DETAILED DESCRIPTION

[0060] The features and exemplary embodiments of various aspects of the present application will be described in detail below. In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present application, rather than to limit the present application. For those skilled in the art, the present application can be implemented without the need for some of these specific details. The following description of the embodiments is only to provide a better understanding of the present application by illustrating the examples of the present application.

[0061] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "include..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0062] Recurrent Neural Network (RNN): A recurrent neural network is a type of recursive neural network that takes sequence data as input, recursively in the direction of sequence evolution, and all nodes are connected in a chain.

[0063] Self-Monitoring Analysis and Report Technology (SMART): A specification for an automatic hard disk status detection and early warning system. It can monitor and record the operation of the hard disk hardware, such as the hard disk head unit, the disk motor drive system, the hard disk internal circuit, and the disk surface media material, through the detection instructions preset in the hard disk hardware.

[0064] With the advent of the big data era, the total amount of global data has increased significantly, and large data centers have played an increasingly important role in people's daily lives. Disks, as the main storage carrier of information at present, play a role that cannot be ignored. Although individual disks are very secure, disk failures are still common in large data centers with thousands of disks. Disk failures will challenge the system reliability and availability of actual data centers, and may even lead to the loss of disk storage data, causing huge losses to enterprises and users. Traditional passive fault tolerance generally uses redundancy mechanisms such as replicas or erasure coding technology to deal with possible disk failures, which will undoubtedly increase the cost of system construction and reduce the system's service capabilities in the process of handling disk failures. Therefore, an active fault tolerance mechanism has emerged, which actively predicts disk failures and migrates dangerous data based on early warnings, which can not only reduce the system's backup redundancy overhead, but also reduce the occurrence of data loss, greatly improving the performance of the storage system. Through statistical or machine learning methods, disk SMART data is used to establish a model, which provides the possibility of disk failure prediction.

[0065] However, in the existing technology, although the deep learning model has a strong ability to learn data knowledge, it is a black box. If a deep learning model is used, even if the model prediction results are found to be insufficient, it is impossible to effectively intervene manually to correct the model. If a traditional data mining model is used, although manual intervention is possible, it is difficult to achieve the prediction effect and training potential of deep learning. There is a lack of methods for predicting disk failures that can accurately predict results and can be intervened by the outside world.

[0066] The inventors found through research that if the trained recurrent neural network model is converted into an equivalent decision tree model, and human experts perform pruning and adding branches and parameter adjustment based on disk failure experience, this improvement can incorporate human experience, correct model defects caused by the lack of fault types and scarcity of fault samples in training data, break the inherent black box barrier of deep neural networks, and achieve the purpose of human-machine collaboration and improve the effect of disk failure prediction.

[0067] In order to solve the problems of the prior art, the embodiments of the present application provide a method, device, equipment, storage medium and product for determining a device failure prediction model. The following first introduces the method for determining a device failure prediction model provided by the embodiments of the present application. The method for determining a device failure prediction model of the present application can be executed by the above-mentioned device for determining a device failure prediction model.

[0068] Figure 1 FIG. 1 is a flow chart of a method for determining a device fault prediction model provided by an embodiment of the present application. Figure 1 As shown, the method includes the following steps S101 to S104.

[0069] Step S101: Obtain a target neural network model for device fault prediction and a disk data sample set, wherein the disk data sample set includes multiple disk data samples, and the disk data samples include disk status data of the disk and a fault label corresponding to the disk status data; the disk status data includes feature data in multiple dimensions.

[0070] The target neural network model can be a pre-trained neural network model, which is usually used to process input data and make predictions. In this embodiment, it is designed to predict the failure state of the disk.

[0071] The disk status data may include characteristic data in multiple dimensions, such as the disk's read speed, write speed, temperature, working time, etc.

[0072] In the specific implementation, disk data samples are continuously collected in advance to form a disk data sample set, and the database stores the trained target neural network model for equipment failure prediction. When it is necessary to execute this solution, the above target neural network model and disk data sample set can be directly obtained.

[0073] Step S102: input the target feature data of each disk data sample into the target neural network model to obtain the contribution degree corresponding to the target feature data.

[0074] The target feature data may be feature data in each disk data sample, that is, each feature value describing the disk status.

[0075] The contribution degree can be the influence of each input feature in the neural network model on the final prediction result.

[0076] In the specific implementation, the feature data of each disk data sample is input into the neural network model. The neural network processes the feature data through its hierarchical structure and weights. By analyzing the internal mechanism of the neural network, such as the calculation of the prediction probability, the contribution of each feature to the prediction result can be calculated, and the contribution corresponding to the above target feature data can be obtained.

[0077] As another example, in a neural network, the contribution of each input feature can be calculated by different methods, such as analysis of gradients or activation values.

[0078] Step S103: construct a contribution matrix according to the contribution corresponding to each of the above target feature data.

[0079] The contribution matrix can be a matrix in which each element represents the contribution of the corresponding feature in the neural network prediction. These contributions can be relative prediction scores or normalized weights.

[0080] In a specific implementation, after the contribution corresponding to each of the above target feature data is obtained, all the contribution is organized into a matrix form to construct a contribution matrix.

[0081] Step S104: According to the above contribution matrix, the target neural network model is converted into a decision tree model with the above target feature data as nodes.

[0082] A decision tree model can be a prediction model based on a tree structure, which is used to make decisions and classifications based on input features.

[0083] In the specific implementation, the contribution matrix is ​​used to determine the branch position and node division rules of each feature in the decision tree based on the preset feature selection strategy, and to determine the structure of the decision tree and the specific representation of each node. Usually, features with high contribution are selected as split nodes of the decision tree, thus forming a tree structure for predicting the possibility of future disk failures. In this way, the target neural network model is converted into a decision tree model with the above target feature data as nodes.

[0084] As another example, the decision rule of a specific node may be to select the most influential feature as the node of the decision tree according to the feature contribution in the contribution matrix, so as to split the data set.

[0085] The above implementation method provided in the embodiment of the present application can convert the neural network model used for equipment fault prediction into a decision tree model with feature data as nodes based on the contribution of the feature data in the training sample to the prediction result. The hierarchical structure of the decision tree model is easier for users to perceive, and is no longer a "black box structure" like a neural network. Therefore, fault prediction based on the decision tree model can clearly display the processing process of the decision tree model, allowing users to more easily explain the decision-making process within the model, and can also more easily modify the model structure, thereby improving the accuracy of the model.

[0086] In order to accurately obtain the contribution of the target feature data to the prediction result, in some embodiments, the above step S102 may include steps A1 to C1:

[0087] Step A1: input the target feature data of each sample in the above disk data sample set into the above target neural network model to obtain the prediction score of the above target feature data, and the above prediction score indicates the predicted failure probability.

[0088] The prediction score can be an estimate of the probability of disk failure under the corresponding feature combination. The higher the value, the more likely the model thinks that the disk is likely to fail in its current state.

[0089] In a specific implementation, the target feature data of each sample is extracted from the disk data sample set as the input of the neural network model. The target feature data is input into the trained neural network model. After the target feature data of each sample in the above disk data sample set is input into the above target neural network model, the model will convert the feature data into a prediction score through its internal weights and structure, and obtain the prediction score of the above target feature data.

[0090] Step B1: Calculate the entropy gain value of the feature corresponding to the target feature data according to the prediction score.

[0091] In the specific implementation, the entropy gain value of each feature can be determined by calculating the change in the prediction score. Generally speaking, the larger the entropy gain, the higher the contribution of the feature to the prediction result. Information entropy can be used to measure the uncertainty or confusion of the prediction score, while entropy gain measures the reduction of information entropy after adding a specific feature, that is, the impact on the prediction result.

[0092] Step C1: taking the entropy gain value as the contribution of the target feature data to the prediction result.

[0093] In the specific implementation, the above entropy gain value is used as the contribution of the target feature data to the prediction result, which can reflect the influence of the feature on the output of the neural network model, that is, the fault prediction score. A higher entropy gain value means that the feature has greater distinguishing and predictive power for fault prediction.

[0094] The above implementation method provided by the embodiment of the present application inputs the target feature data of each sample in the above disk data sample set into the above target neural network model, thereby obtaining a prediction score indicating the predicted failure probability. Based on this prediction score, the entropy gain value of the corresponding feature is calculated and used as the contribution of the target feature data to the prediction result. The contribution of the target feature data to the prediction result can be accurately obtained.

[0095] In order to obtain a reasonable decision tree model, in some embodiments, reference Figure 2 The above step S104 may include steps S1041 to S1045:

[0096] According to the order of the decision tree model from high to low levels, the nodes on each level are determined in turn until the preset stop condition is met:

[0097] Step S1041: According to the contribution matrix, the feature with the greatest contribution is selected from the initial feature set composed of feature data of multiple dimensions as the node of the current level.

[0098] In the specific implementation, the higher the contribution value in the contribution matrix, the greater the influence of the feature on the output of the prediction model. Therefore, according to the above contribution matrix, the feature with the highest contribution can be found from the initial feature set composed of multiple dimensional feature data. This feature is selected as the node of the current level.

[0099] Step S1042: Divide the features in the first target feature set into a first subset and a second subset according to the contribution segmentation threshold corresponding to the feature with the largest contribution; the first target feature set is a set of features remaining after removing the feature with the largest contribution from the initial feature set.

[0100] In a specific implementation, after removing the feature with the largest contribution from the initial feature set, the remaining features form the first target feature set. Using the contribution segmentation threshold corresponding to the feature with the largest contribution, the features in the first target feature set are divided into two subsets, namely, the first subset and the second subset, according to the relationship between the contribution corresponding to the feature and the contribution segmentation threshold.

[0101] For example, as another example, the first subset includes samples whose contribution of the feature is greater than the threshold, and the second subset includes samples whose contribution of the feature is less than the threshold.

[0102] For the first subset and the second subset, the following operations are performed:

[0103] Step S1043: Selecting the feature with the greatest contribution from the target subset as the node of the next level of the current level, the target subset being the first subset or the second subset.

[0104] In the specific implementation, the first subset or the second subset is selected as the target subset. Generally speaking, the above subsets need to be selected as the target subset for iteration. After the target subset is determined, the feature that can best distinguish the samples in the target subset is determined according to the contribution value recorded in the contribution matrix. From the target subset, the feature with the largest contribution is found according to the contribution matrix. This feature is selected as the node of the next level.

[0105] Step S1044: updating the second target feature set to the initial feature set, where the second target set is a set of features remaining after removing the feature with the maximum contribution from the target subset.

[0106] Step S1045: Determine whether the current decision tree meets the preset stop condition, if yes, stop iteration and obtain the decision tree model. If no, return to step S1041 and continue iteration until the preset stop condition is met.

[0107] In a specific implementation, after removing the feature with the maximum contribution from the target subset, the remaining features are combined into a second target set. The second target feature set is updated to the initial feature set, and then the process returns to step S1041 to iterate until a preset stop condition is met.

[0108] The above implementation method provided in the embodiment of the present application is to select the feature with the largest contribution from the initial feature set as the node of the current level in each layer in a continuous iterative manner according to the order of the levels of the decision tree model from high to low, and then divide the first subset and the second subset, and select the node of the next level from the above subset. In this way, the nodes on each level can be determined in turn to obtain a reasonable decision tree model.

[0109] In order to speed up the processing, in some implementations, before the above A1, the above method may further include:

[0110] Target feature data is extracted from each sample in the disk data sample set based on a preset extraction rule, and the amount of the target feature data is less than the amount of feature data in the sample.

[0111] In the specific implementation, extraction rules are pre-set, which contain specific requirements on how to select target feature data from each sample. First, the disk data sample set is loaded, and then according to the preset rules, part of the target feature data is extracted from the feature data of each sample. Due to the limitation of the extraction rules, the number of target feature data is less than the number of original feature data.

[0112] As another example, possible extraction rules include selecting specific columns or fields, or applying certain transformations or filtering conditions. For example, target features can be selected based on the type, importance, relevance, etc. of the feature.

[0113] In the above implementation provided by the embodiment of the present application, target feature data is extracted from each sample in the above disk data sample set according to preset extraction rules, and the number of extracted target feature data is less than the number of feature data in the above samples, which can reduce the amount of data that the target neural network model needs to process and speed up the processing.

[0114] In order to obtain a target neural network model with accurate prediction results, in some embodiments, reference Figure 3 Before the above S101, the above method may further include steps S001 to S006:

[0115] Step S001: Acquire a training sample set; the training sample set includes a plurality of training samples, each of which includes disk status data of a disk and a fault label corresponding to the disk status data.

[0116] In a specific implementation, various disk status data and fault labels corresponding to the disk status data are collected in advance. These data are combined into training samples. Multiple training samples are then combined into a training sample set. When needed, the training sample set is loaded and obtained.

[0117] For each training sample, the following steps S002 to S006 are performed respectively:

[0118] Step S002: Input the above disk status data into the target neural network model to determine the prediction result of the disk failure situation.

[0119] In a specific implementation, the disk status data of the training sample obtained in step S001 is then input into a predefined target neural network model for prediction, thereby obtaining a prediction result of the disk failure condition.

[0120] Step S003: Determine the loss function value of the target neural network model according to the prediction results of each of the training samples and the corresponding fault labels.

[0121] In a specific implementation, based on a pre-selected loss function, such as mean square error, cross entropy, etc., according to the prediction results of each of the training samples and the corresponding fault labels, the loss function value of the target neural network model is calculated.

[0122] Step S004: Determine whether the loss function value of the target neural network model meets the preset training stop condition, if yes, execute S006, if not, execute S005.

[0123] Step S005: When the loss function value of the target neural network model does not satisfy the preset training stop condition, adjust the model parameters of the target neural network model and return to S002.

[0124] In a specific implementation, when the loss function value of the target neural network model does not meet the preset training stop condition, the parameters of the target neural network model are adjusted to optimize the prediction ability of the model, and then the disk status data is input into the target neural network model to determine the prediction result of the disk failure. A new round of training is performed again to continuously optimize the prediction ability of the model.

[0125] As another example, a back-propagation algorithm may be used in the process of adjusting the parameters of a target neural network model to update the weights and biases of the neural network through methods such as gradient descent.

[0126] Step S006: When the loss function value of the target neural network model satisfies the preset training stop condition, stop the training to obtain the trained target neural network model.

[0127] In a specific implementation, if the loss function value meets the preset training stop condition, the training process is stopped and the current target neural network model is used as the trained target neural network model.

[0128] As another example, the stopping condition of the training process may be that the loss function value reaches a certain threshold or the number of training rounds reaches a preset maximum value.

[0129] After obtaining the training sample set, the above implementation provided by the embodiment of the present application uses each training sample in the training sample set to train the target neural network model. During the training process, the model parameters of the above target neural network model are continuously adjusted until the loss function value of the above target neural network model meets the preset training stop condition, and the training is stopped to obtain the trained target neural network model. Thus, a target neural network model with accurate prediction results is obtained.

[0130] In order to ensure the accuracy of the training samples in the training sample set, in some implementations, the above S001 may include steps A4 to E4:

[0131] Step A4: Acquire initial disk status data and a fault label corresponding to the initial disk status data.

[0132] In a specific implementation, the initial disk status data and its corresponding fault label are obtained from a preset data source, such as a database, log file, etc. These data can represent various disk indicators, such as temperature, read and write rate, error count, etc., and whether the disk has failed.

[0133] Step B4: when there are missing values ​​in the initial disk status data, the missing values ​​are filled in by a preset interpolation method to obtain the corrected first disk status data.

[0134] In a specific implementation, the initial disk status data may be missing due to sensor failure, data transmission error, etc. In the case where there are missing values ​​in the above initial disk status data, a preset interpolation method, such as linear interpolation, mean interpolation, etc., is used to fill in the data points with missing values.

[0135] As another example, in the process of implementing the interpolation algorithm, the time series characteristics of the data can be considered for interpolation to ensure that the filled data can maintain reasonable continuity and authenticity.

[0136] Step C4: when there are values ​​exceeding the preset range in the initial disk status data, the values ​​exceeding the preset range are replaced by the preset boundary value to obtain the corrected second disk status data.

[0137] In a specific implementation, when there are values ​​beyond a preset range in the above initial disk status data, the data values ​​beyond the preset range are replaced with preset boundary values. For example, if a disk temperature is abnormally recorded as a negative value, it can be corrected to 0 or a preset minimum temperature value.

[0138] Step D4: Use the first disk status data or the second disk status data and the corresponding fault label as training samples.

[0139] In a specific implementation, after processing in step B4 and step C4, a complete and corrected data set is obtained. Then, each corrected disk status data point and its corresponding fault label are combined into a training sample.

[0140] Step E4: Construct a training sample set based on the above training samples.

[0141] In the specific implementation, the above training samples are combined to form the final training sample set, which will be used in the subsequent neural network model training process.

[0142] After obtaining the initial disk status data and the fault labels corresponding to the initial disk status data, the implementation method provided by the embodiment of the present application fills in the missing values ​​by a preset interpolation method to obtain the corrected disk status data when there are missing values ​​in the initial disk status data, and replaces the values ​​exceeding the preset range by a preset boundary value to obtain the corrected disk status data when there are values ​​exceeding the preset range in the initial disk status data. The corrected disk status data and the corresponding fault labels are then used as training samples, and a training sample set is constructed based on the training samples. The accuracy of the training samples in the training sample set can be guaranteed.

[0143] In order to ensure the adaptability of the decision tree model, in some embodiments, after the above S104, the above method may further include steps A5 to D5:

[0144] Step A5: Acquire the actual status information of the target disk and the actual fault information of the target disk.

[0145] In the specific implementation, after the target disk is determined, the current actual status information of the target disk is obtained through data sources, such as relevant real-time monitoring systems, log files, etc., including various indicators such as temperature, read and write speed, etc. The actual fault information of the target disk is also obtained, that is, whether the target disk has failed.

[0146] As another example, the state information and fault information of the target disk may be extracted from the data source by matching with the identifier corresponding to the target disk.

[0147] Step B5: Input the actual status information of the target disk into the decision tree model to obtain the predicted fault information of the target disk.

[0148] In the specific implementation, the actual status information of the target disk is input into the decision tree model, and the model will predict whether the target disk may fail in the future based on the input actual status information and its internal decision rules, and obtain the predicted failure information of the above target disk.

[0149] As another example, it may be necessary to perform preprocessing operations such as feature extraction and format conversion on the state information in order to match the model input requirements.

[0150] Step C5: Determine the accuracy of the decision tree model according to the difference between the predicted fault information and the actual fault information.

[0151] In a specific implementation, the predicted fault information of the decision tree model is compared with the actual fault information, and the accuracy of the decision tree model is determined based on the difference between the two to evaluate the performance of the model.

[0152] Step D5: When the accuracy is lower than the threshold, in response to the user's update operation on the decision tree model, the decision tree model is updated.

[0153] In a specific implementation, if the accuracy of the decision tree model is lower than a set threshold, an update operation is triggered. Based on the response to the user's update operation on the decision tree model, the decision tree model is updated according to the user's needs.

[0154] As another example, the update operation can be to collect a new training data set based on the latest actual status and fault information. The decision tree model is then retrained using the new data set to improve its prediction accuracy. Finally, the retrained model is deployed to the production environment to replace the old model, so that subsequent fault predictions can be performed based on the updated model.

[0155] After obtaining the actual status information of the target disk and the actual fault information of the target disk, the implementation method provided by the embodiment of the present application inputs the actual status information of the target disk into the decision tree model to obtain the predicted fault information of the target disk. Thus, the predicted fault information can be compared with the actual fault information, and the accuracy of the decision tree model can be determined based on the difference between the two. If the accuracy is lower than the threshold, the decision tree model is updated in response to the user's update operation on the decision tree model. In the case where the accuracy of the decision tree model is insufficient, the decision tree model can be updated to ensure the adaptability of the decision tree model.

[0156] In order to update the decision tree model from multiple aspects, in some embodiments, the updating of the decision tree model may include: modifying the decision tree model based on target modification information, incrementally training the target neural network model, and fully training the target neural network model.

[0157] In a specific implementation, modifying the decision tree model based on the target modification information may be modifying the decision tree model according to the user's requirements. Incremental training of the target neural network model may be retraining the target neural network model based on the added data, and then converting the retrained target neural network model to obtain a new decision tree model. Full training of the target neural network model may be training a new target neural network model based on new data, and then converting the target neural network model to obtain a new decision tree model.

[0158] As another example, collect data related to the target modification. For example, it can include user feedback, new business indicators, and performance evaluation results. This information will help determine where the model needs to be improved. Based on the collected information, define specific modification goals. For example, modify certain parameters of the model, adjust the feature selection strategy, or change the decision tree construction algorithm. For example, if it is found that the model's predictions for certain categories are not accurate enough, the goal can be defined as improving the recall or precision of these categories. Or it can be to determine the model parameters that need to be adjusted, such as the maximum depth of the tree, the minimum number of sample splits, etc. After defining the target modification, start implementing these modifications. Adjust the parameters of the decision tree algorithm, such as the maximum depth, the minimum number of sample splits, etc., or change the feature selection method, such as using different criteria such as information gain, gain rate, or Gini index.

[0159] In the above implementation provided by the embodiment of the present application, the method for updating the above decision tree model includes modifying the above decision tree model based on the target modification information, incrementally training the above target neural network model, and fully training the above target neural network model. The decision tree model can be updated from multiple aspects.

[0160] In one embodiment of the present application, in this embodiment, taking a disk array (Redundant Arrays of Independent Disks, RAID) as an example, Figure 4 As shown, there is a data acquisition device 201, which adopts a regular and equal-interval acquisition strategy, and collects and uploads data from the RAID monitoring system through an acquisition script. The RAID monitoring system collects data from the external disk array 202. The uploaded data is parsed and necessary data cleaning and preparation work is performed, including data verification, abnormal data processing, standardization processing, etc. For training data, it also includes the extraction of target feature values.

[0161] Among them, data verification is to identify whether there are missing values ​​or data beyond the domain in the data. For missing values, the quadratic smoothing difference method is used to fill in the missing values. For values ​​beyond the range, the corresponding boundary values ​​are used to replace them, and then the mean + / -3 standard deviations are used to replace all values ​​beyond the boundary (mean + / -3 standard deviations). When standardizing, the maximum and minimum values ​​can be used for normalization, and the preset maximum and minimum values ​​are used to replace the data that exceeds the preset range before, that is, the original mean + 3 standard deviations are the maximum value, and the mean -3 standard deviations are the minimum value. The intervals of each variable in the data obtained in this way are unified, and each variable will not be affected by certain extreme values, resulting in the data distribution outside the extreme values ​​being too concentrated.

[0162] After processing, the data is used to train the RNN model. First, random sampling is performed in the data, and part of the data is used as the training data set and the other part of the data is used as the validation data set. The training data will be input into the initial RNN model to train the RNN model weight parameters. Then, the RNN model is converted into an equivalent decision tree model using the conversion algorithm. Finally, expert analysis is performed to evaluate the decision tree model. Then, according to the evaluation results, manual corrections such as branching, pruning, and parameter adjustment are performed to incorporate expert knowledge and obtain the final decision tree model.

[0163] Specifically, the input for RNN model training is SMART multi-dimensional attribute data in the form of time series, that is, the SMART feature set extracted at each moment. That is, during the RNN model training process, all indicators at m moments are input to predict the probability of failure of the disk at the future m+1 moment. After the training is completed through continuous iteration and the trained RNN model is obtained, the RNN model is converted into a decision tree model. Suppose the selected training set S contains N prediction variables V i , i = 1,…, N; S contains M records in total, of which the jth record is represented by S j ,j=1,…,M; record S j For variable V i The contribution value is recorded as: c ij , and by c ij , i = 1, ..., N; j = 1, ..., M can form the feature contribution distribution matrix C. Specifically, the Local Interpretable Model-agnostic Explanations (LIME) algorithm can be used to calculate c ij . Before sampling using the LIME algorithm, the training data needs to be preprocessed and the samples need to be segmented into feature regions. The segmentation is achieved with the help of the dilated convolution method. In this way, the process of perturbing the data using the LIME algorithm is to randomly retain a part of the feature region and discard a part of the feature region. These perturbed samples are input into the RNN model to obtain the prediction score of the sample, that is, the probability of failure. The prediction score is the weight value of the feature region retained in the sample. Here, only one area of ​​the segmented sample is retained, and all other segmented areas are masked. The prediction score obtained is the contribution value of this feature, that is, the contribution value of the input variable of the sample in each row in the contribution distribution matrix.

[0164] In the process of generating decision nodes, the best variable is selected from the variable set as the decision variable of each node in the decision tree. The basis for selection is the variable V with the largest contribution value G(i). i As a decision variable, the specific contribution value is calculated as follows:

[0165]

[0166] Among them, G(i) represents in the current decision tree, if V is adopted i As a decision variable, the maximum entropy gain that can be brought. Here, the maximum entropy gain refers to the maximum entropy gain when V i is the decision variable, and different thresholds are taken as sample subsets S L , S R The partitioning condition is that among all G(i) calculated, the largest one is the one that is calculated. If V i is selected, then this threshold will also become the threshold of the decision node. i The value of G(i) is determined by dividing the sample subset S L , S R The threshold that can make G(i) reach the maximum value: argMax(G(i)) is the final segmentation threshold. L represents the subset of samples that belong to the tree node and meet the threshold, S R Represents the subset of samples that belong to the tree node and do not meet the threshold. The determination of the threshold is not only about the value itself, but also about the operator that meets the threshold, that is, whether greater than is considered "satisfied" or less than or equal to is considered "satisfied". The solution to this problem, together with the determination of the threshold itself, is included in the above-mentioned process of finding argMax(G(i)).

[0167] L is the left branch of the decision tree, and R is the right branch of the decision tree; |S L |Represents S L The number of samples contained in ; Represents S L The V contained in i The sum of the contributions of the variables.

[0168] Using a recursive method, we continuously L , S R The tree grows on the data subset until the predetermined depth is reached or no new node that meets the minimum requirement of G(i) is found. The final determination of the ownership of the leaf node is based on the above threshold determination process and the calculation formula of G(i). It can be seen that G(i) is positively correlated with the contribution value of the left branch and negatively correlated with the right branch. At the same time, it has been defined in the above steps that the contribution value c ij What we are concerned about is the probability of failure. Therefore, the leaf nodes on the left branch represent failures, and the leaf nodes on the right branch represent normality.

[0169] The following is an example of a disk prediction scenario. Assume that the training set contains the following 10 observation statistics, that is, N = 10, as shown in Table 1.

[0170] Table 1

[0171] Variable Name Business Implications Project code <![CDATA[V1]]> Write Error Rate A01 <![CDATA[V2]]> Bottom-layer data read error rate A02 <![CDATA[V3]]> DMA cycle check error rate B01 <![CDATA[V4]]> Reallocation event count B02 <![CDATA[V5]]> Reallocate the number of sectors B03 <![CDATA[V6]]> Unsafe power outages C01 <![CDATA[V7]]> Spindle motor rotation synchronization retry rate D01 <![CDATA[V8]]> Seek Error Rate D02 <![CDATA[V9]]> Spindle motor rotation synchronization time D03 <![CDATA[V 10 ]]> Head lift times D04

[0172] The trained recurrent neural network model is converted into an equivalent decision tree, and the following can be obtained: Figure 5 The disk failure decision tree prediction model shown.

[0173] Afterwards, the obtained decision tree can be manually modified. For example, by adding decision branches, the fault mechanism or fault type not included in the training data is manually added. For example, after analyzing the decision tree, it is believed that the influence of the variable "number of head lifts" should also be considered, otherwise the recall rate will be reduced. Therefore, a decision node can be added, and the decision node of "D04 number of head lifts" can be added to increase the branch. Alternatively, by reducing decision branches, the corresponding branches formed by overfitting of noise data are removed. For example, if the first B02 on the left reallocates the number of events, it is already the best decision path for the branch. Continuing with the following D03 is the result of overfitting. Under this path, the existence of D03 does not improve the recall rate, but only reduces the recall rate. Therefore, the DO3 node and its subordinate decision nodes are removed and replaced with fault decision nodes to complete the pruning operation. In addition, the degree of consistency between fault prediction and actual conditions can be improved by fine-tuning the decision reference threshold on the tree node. For example, suppose the original threshold of the C01 unsafe power outage times node in the third layer of the figure above is greater than 20 times. Based on experience, it is considered that the setting is too low compared with the actual situation, so it is changed to a more reasonable value of greater than 50 times. And through global observation, no phenomenon causing decision conflicts is found (conflict means that there is already an upper-level node with C01 < 30 times on the decision path where the current node is located, which is mutually exclusive with C01 > 50), so the threshold is adjusted to achieve model tuning.

[0174] After that, the corrected decision tree model can be used to make actual fault predictions, and its prediction effect can be evaluated regularly. When the performance drops beyond the threshold, the model can be updated. For example, the system presets the recall rate of faulty samples to be >50% and the precision rate to be >90%. As long as one of them is below the threshold, the model is considered to need maintenance. The specific maintenance method can be to modify only the decision tree, or to re-formulate the training set to train the deep model as needed, transform the decision tree, and then manually optimize the decision tree. After the latest test set evaluation and continuous tuning, the recall rate can be >50% and the precision rate can be >90%, and then the current model can be replaced with the new model.

[0175] In the specific implementation process, the prediction model is first loaded, and the SMART data of the disk is collected at fixed intervals. If necessary, the data is preprocessed to make it consistent with the data prediction processing during training. Then the disk device is predicted, and the prediction conclusion and confidence level of whether a failure is about to occur are given.

[0176] At the same time, we collect the repair records of disk failures and confirm the actual occurrence of disk failures. We compare the actual disk failure cases with the prediction records of the corresponding period to determine the recall rate and accuracy. Then we compare the corresponding recall rate and accuracy with the set thresholds. The default thresholds are: recall rate>40%, accuracy>99%. As long as one of them is lower than the threshold, it is considered that the model needs maintenance.

[0177] If the model is judged to need maintenance, the model update attempt can be made in order from simple to complex. First, update the training set and validation set of the model so that they both contain the latest fault samples. Then the operation and maintenance experts adjust the threshold and prune the decision branches of the existing decision tree model. If the evaluation criteria can be met in the validation set, the update is completed, otherwise the next step of updating is to incrementally train the deep neural network model and convert it into a decision tree. After manual pruning and tuning, the evaluation is performed. If the evaluation criteria can be met in the validation set, the update is completed, otherwise the next step is to fully train the deep neural network model and convert it into a decision tree. After manual pruning and tuning, the evaluation is performed. If the evaluation criteria can be met in the validation set, the update is completed, otherwise the training parameters are adjusted and this step is repeated until the evaluation criteria can be met.

[0178] By converting the trained deep learning model into an equivalent decision tree model, while maintaining the prediction advantages of the deep learning model, human experts can easily add disk failure-related variables in the decision tree, that is, add branches, to add failures that are not expressed in the training data to the decision tree, and prevent possible incompatibility risks with the original model, thereby effectively improving the prediction ability; the negative impact of noise gain brought by adjusting extremely biased data can be avoided through pruning operations to reduce overfitting, thereby reducing the false alarm rate in disk failure prediction.

[0179] Based on the equipment failure prediction model determination method provided in the above embodiment, the present application also provides a specific implementation of the equipment failure prediction model determination device. Please refer to the following embodiment.

[0180] See first Figure 6 The device 600 for determining a device failure prediction model provided in the embodiment of the present application includes the following modules:

[0181] The first acquisition module 601 is used to obtain a target neural network model for equipment fault prediction and a disk data sample set, wherein the disk data sample set includes multiple disk data samples, and the disk data samples include disk status data of the disk and a fault label corresponding to the disk status data; the disk status data includes feature data in multiple dimensions.

[0182] The contribution determination module 602 is used to input the target feature data of each disk data sample into the above-mentioned target neural network model to obtain the contribution corresponding to the above-mentioned target feature data.

[0183] The construction module 603 is used to construct a contribution matrix according to the contribution corresponding to each of the above target feature data.

[0184] The conversion module 604 is used to convert the target neural network model into a decision tree model with the target feature data as nodes according to the contribution matrix.

[0185] The above implementation method provided in the embodiment of the present application can convert the neural network model used for equipment fault prediction into a decision tree model with feature data as nodes based on the contribution of the feature data in the training sample to the prediction result. The hierarchical structure of the decision tree model is easier for users to perceive, and is no longer a "black box structure" like a neural network. Therefore, fault prediction based on the decision tree model can clearly display the processing process of the decision tree model, allowing users to more easily explain the decision-making process within the model, and can also more easily modify the model structure, thereby improving the accuracy of the model.

[0186] As an implementation of the present application, the contribution determination module 602 may include:

[0187] The prediction score determination unit is used to input the target feature data of each sample in the above disk data sample set into the above target neural network model to obtain the prediction score of the above target feature data, and the above prediction score indicates the predicted failure probability.

[0188] A calculation unit is used to calculate the entropy gain value of the feature corresponding to the above-mentioned target feature data according to the above-mentioned prediction score.

[0189] The determination unit is used to use the above entropy gain value as the contribution of the above target feature data to the prediction result.

[0190] The above implementation method provided by the embodiment of the present application inputs the target feature data of each sample in the above disk data sample set into the above target neural network model, thereby obtaining a prediction score indicating the predicted failure probability. Based on this prediction score, the entropy gain value of the corresponding feature is calculated and used as the contribution of the target feature data to the prediction result. The contribution of the target feature data to the prediction result can be accurately obtained.

[0191] As an implementation of the present application, the conversion module 604 may include:

[0192] The selection unit is used to select the feature with the greatest contribution as the node of the current level from the initial feature set composed of feature data of multiple dimensions according to the contribution matrix.

[0193] A division unit is used to divide the features in the first target feature set into a first subset and a second subset according to the contribution segmentation threshold corresponding to the feature with the largest contribution; the first target feature set is a set of features remaining after removing the feature with the largest contribution from the initial feature set.

[0194] The selection unit is further used to select the feature with the greatest contribution from the target subset as the node of the next level of the above-mentioned current level, and the above-mentioned target subset is the first subset or the second subset.

[0195] An updating unit is used to update the second target feature set to the initial feature set, and return the above-mentioned feature with the largest contribution from the initial feature set composed of multiple dimensional feature data according to the above-mentioned contribution matrix as the node of the current level; the above-mentioned second target set is a set of features remaining after removing the feature with the largest contribution from the above-mentioned target subset.

[0196] The above implementation method provided in the embodiment of the present application is to select the feature with the largest contribution from the initial feature set as the node of the current level in each layer in a continuous iterative manner according to the order of the levels of the decision tree model from high to low, and then divide the first subset and the second subset, and select the node of the next level from the above subset. In this way, the nodes on each level can be determined in turn to obtain a reasonable decision tree model.

[0197] As an implementation of the present application, the prediction score determination unit may include:

[0198] The extraction subunit is used to extract target feature data from each sample in the above disk data sample set based on a preset extraction rule, and the amount of the above target feature data is less than the amount of feature data in the above sample.

[0199] In the above implementation provided by the embodiment of the present application, target feature data is extracted from each sample in the above disk data sample set according to preset extraction rules, and the number of extracted target feature data is less than the number of feature data in the above samples, which can reduce the amount of data that the target neural network model needs to process and speed up the processing.

[0200] As an implementation of the present application, the equipment failure prediction model device 600 may further include:

[0201] The second acquisition module is used to acquire a training sample set; the training sample set includes a plurality of training samples, and each of the training samples includes disk status data of a disk and a fault label corresponding to the disk status data.

[0202] The prediction result determination module is used to input the above disk status data into the target neural network model to determine the prediction result of the disk failure situation.

[0203] The loss function value determination module is used to determine the loss function value of the target neural network model according to the prediction results of each of the training samples and the corresponding fault labels.

[0204] The adjustment module is used to adjust the model parameters of the above-mentioned target neural network model when the loss function value of the above-mentioned target neural network model does not meet the preset training stop conditions, and return to the step of inputting the above-mentioned disk status data into the target neural network model to determine the prediction result of the disk failure situation.

[0205] The stop module is used to stop training when the loss function value of the target neural network model meets the preset training stop condition to obtain the trained target neural network model.

[0206] After obtaining the training sample set, the above implementation provided by the embodiment of the present application uses each training sample in the training sample set to train the target neural network model. During the training process, the model parameters of the above target neural network model are continuously adjusted until the loss function value of the above target neural network model meets the preset training stop condition, and the training is stopped to obtain the trained target neural network model. Thus, a target neural network model with accurate prediction results is obtained.

[0207] As an implementation of the present application, the second acquisition module may include:

[0208] The acquisition unit is used to acquire the initial disk status data and the fault label corresponding to the initial disk status data.

[0209] The interpolation unit is used to fill in the missing values ​​by a preset interpolation method when there are missing values ​​in the initial disk status data, so as to obtain the corrected first disk status data.

[0210] The replacement unit is used to replace the value beyond the preset range with a preset boundary value when there is a value beyond the preset range in the initial disk status data, so as to obtain the corrected second disk status data.

[0211] A determination unit is configured to use the first disk status data or the second disk status data and a corresponding fault label as a training sample.

[0212] The training unit is used to construct a training sample set according to the above training samples.

[0213] After obtaining the initial disk status data and the fault labels corresponding to the initial disk status data, the implementation method provided by the embodiment of the present application fills in the missing values ​​by a preset interpolation method to obtain the corrected disk status data when there are missing values ​​in the initial disk status data, and replaces the values ​​exceeding the preset range by a preset boundary value to obtain the corrected disk status data when there are values ​​exceeding the preset range in the initial disk status data. The corrected disk status data and the corresponding fault labels are then used as training samples, and a training sample set is constructed based on the training samples. The accuracy of the training samples in the training sample set can be guaranteed.

[0214] As an implementation of the present application, the equipment failure prediction model device 600 may further include:

[0215] The third acquisition module is used to acquire the actual status information of the target disk and the actual fault information of the target disk.

[0216] The input module is used to input the actual status information of the target disk into the decision tree model to obtain the predicted fault information of the target disk.

[0217] The accuracy determination module is used to determine the accuracy of the decision tree model according to the difference between the predicted fault information and the actual fault information.

[0218] The response module is used to update the decision tree model in response to the user's update operation on the decision tree model when the accuracy is lower than the threshold.

[0219] After obtaining the actual status information of the target disk and the actual fault information of the target disk, the implementation method provided by the embodiment of the present application inputs the actual status information of the target disk into the decision tree model to obtain the predicted fault information of the target disk. Thus, the predicted fault information can be compared with the actual fault information, and the accuracy of the decision tree model can be determined based on the difference between the two. If the accuracy is lower than the threshold, the decision tree model is updated in response to the user's update operation on the decision tree model. In the case where the accuracy of the decision tree model is insufficient, the decision tree model can be updated to ensure the adaptability of the decision tree model.

[0220] The various modules in the device for determining a equipment fault prediction model provided in the embodiment of the present application can implement the various steps in the above-mentioned method for determining a equipment fault prediction model and achieve the corresponding effects. For the sake of concise description, they will not be repeated here.

[0221] Figure 7 A schematic diagram of the structure of the equipment fault prediction model determination hardware provided in an embodiment of the present application is shown.

[0222] The device for determining a device fault prediction model may include a processor 701 and a memory 702 storing computer program instructions.

[0223] Specifically, the processor 701 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0224] The memory 702 may include a large capacity memory for data or instructions. By way of example and not limitation, the memory 702 may include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive or a combination of two or more of these. In appropriate cases, the memory 702 may include a removable or non-removable (or fixed) medium. In appropriate cases, the memory 702 may be inside or outside the integrated gateway disaster recovery device. In a specific embodiment, the memory 702 is a non-volatile solid-state memory.

[0225] The memory may include a read-only memory (ROM), a random access memory (RAM), a magnetic disk storage medium device, an optical storage medium device, a flash memory device, an electrical, optical or other physical / tangible memory storage device. Therefore, generally, the memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., a memory device) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the device fault prediction model determination method according to any one embodiment of the present disclosure.

[0226] The processor 701 implements any one of the equipment fault prediction model determination methods in the above embodiments by reading and executing computer program instructions stored in the memory 702 .

[0227] In one example, the device failure prediction model determination device may further include a communication interface 703 and a bus 710. Figure 7 As shown, the processor 701, the memory 702, and the communication interface 703 are connected via a bus 710 and communicate with each other.

[0228] The communication interface 703 is mainly used to implement communication between various modules, devices, units and / or equipment in the embodiments of the present application.

[0229] Bus 710 includes hardware, software or both, and the parts of online data flow billing equipment are coupled to each other. For example, but not limitation, bus may include accelerated graphics port (AGP) or other graphics bus, enhanced industrial standard architecture (EISA) bus, front-end bus (FSB), hypertransport (HT) interconnection, industrial standard architecture (ISA) bus, infinite bandwidth interconnection, low pin count (LPC) bus, memory bus, micro channel architecture (MCA) bus, peripheral component interconnection (PCI) bus, PCI-Express (PCI-X) bus, serial advanced technology attachment (SATA) bus, video electronics standard association local (VLB) bus or other suitable bus or two or more of these combinations. In appropriate cases, bus 710 may include one or more buses. Although the present application embodiment describes and shows a specific bus, the present application considers any suitable bus or interconnection.

[0230] In addition, in combination with the method for determining a device fault prediction model in the above embodiment, the embodiment of the present application may provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when the computer program instructions are executed by a processor, any method for determining a device fault prediction model in the above embodiment is implemented.

[0231] An embodiment of the present application also provides a computer program product, including a computer program, which, when executed, implements any one of the methods for determining a device fault prediction model in the above embodiments.

[0232] It should be clear that the present application is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present application is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present application.

[0233] The functional blocks shown in the above-described block diagram can be implemented as hardware, software, firmware or a combination thereof. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present application are programs or code segments that are used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link by a data signal carried in a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, optical fiber media, radio frequency (RF) links, etc. The code segment can be downloaded via a computer network such as the Internet, an intranet, etc.

[0234] It should also be noted that the exemplary embodiments mentioned in this application describe some methods or systems based on a series of steps or devices. However, this application is not limited to the order of the above steps, that is, the steps can be performed in the order mentioned in the embodiment, or in a different order from the embodiment, or several steps can be performed simultaneously.

[0235] Aspects of the present disclosure are described above with reference to the flowchart and / or block diagram of the method, device (system) and computer program product according to the embodiment of the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device to produce a machine so that these instructions executed by the processor of the computer or other programmable data processing device enable the implementation of the function / action specified in one or more boxes of the flowchart and / or block diagram. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field programmable logic circuit. It can also be understood that each box in the block diagram and / or flowchart and the combination of boxes in the block diagram and / or flowchart can also be implemented by dedicated hardware that performs a specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions.

[0236] The above is only a specific implementation of the present application. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, modules and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. It should be understood that the protection scope of the present application is not limited to this. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in this application, and these modifications or replacements should be included in the protection scope of this application.

Claims

1. A method for determining a device failure prediction model, characterized in that: include: Obtaining a target neural network model for device fault prediction and a disk data sample set, wherein the disk data sample set includes multiple disk data samples, and the disk data samples include disk status data of the disk and a fault label corresponding to the disk status data; the disk status data includes feature data in multiple dimensions; Inputting the target feature data of each disk data sample into the target neural network model to obtain the contribution degree corresponding to the target feature data; Constructing a contribution matrix according to the contribution corresponding to each of the target feature data; According to the contribution matrix, the target neural network model is converted into a decision tree model with the target feature data as nodes.

2. The method for determining an equipment failure prediction model according to claim 1, characterized in that: The step of inputting the target feature data of each disk data sample into the target neural network model to obtain the contribution degree corresponding to the target feature data includes: Inputting target feature data of each sample in the disk data sample set into the target neural network model to obtain a prediction score of the target feature data, wherein the prediction score indicates a predicted failure probability; Calculate the entropy gain value of the feature corresponding to the target feature data according to the prediction score; The entropy gain value is used as the contribution of the target feature data to the prediction result.

3. The method for determining an equipment failure prediction model according to claim 2, characterized in that: The step of converting the target neural network model into a decision tree model with the target feature data as nodes according to the contribution matrix includes: According to the order of the decision tree model from high to low levels, the nodes on each level are determined in turn until the preset stop condition is met: According to the contribution matrix, the feature with the greatest contribution is selected from the initial feature set composed of feature data of multiple dimensions as a node of the current level; According to the contribution segmentation threshold corresponding to the feature with the largest contribution, the features in the first target feature set are divided into a first subset and a second subset respectively; the first target feature set is a set of features remaining after removing the feature with the largest contribution from the initial feature set; For both the first subset and the second subset, the following operations are performed: Selecting the feature with the largest contribution from the target subset as the node of the next level of the current level, the target subset is the first subset or the second subset; The second target feature set is updated to the initial feature set, and the feature with the largest contribution is selected as the node of the current level from the initial feature set composed of multiple dimensional feature data according to the contribution matrix; the second target set is the set of features remaining after removing the feature with the largest contribution from the target subset.

4. The method for determining an equipment failure prediction model according to claim 2, characterized in that: Before inputting the target feature data of each sample in the disk data sample set into the target neural network model to obtain the prediction score of the target feature data, the method further includes: Target feature data is extracted from each sample in the disk data sample set based on a preset extraction rule, and the amount of the target feature data is less than the amount of feature data in the sample.

5. The method for determining a device failure prediction model according to claim 1, characterized in that: Before obtaining the target neural network model and the disk data sample set for device fault prediction, the method further includes: Acquire a training sample set; the training sample set includes a plurality of training samples, each of the training samples includes disk status data of a disk and a fault label corresponding to the disk status data; For each training sample, perform the following steps: Inputting the disk status data into a target neural network model to determine a prediction result of a disk failure condition; Determining a loss function value of the target neural network model according to the prediction results of each of the training samples and the corresponding fault labels; In the case where the loss function value of the target neural network model does not satisfy the preset training stop condition, the model parameters of the target neural network model are adjusted, and the step of returning to inputting the disk status data into the target neural network model to determine the prediction result of the disk failure condition is performed; When the loss function value of the target neural network model meets the preset training stop condition, the training is stopped to obtain the trained target neural network model.

6. The method for determining a device failure prediction model according to claim 5, characterized in that: The step of obtaining a training sample set includes: Acquire initial disk status data and a fault label corresponding to the initial disk status data; In the case where there are missing values ​​in the initial disk status data, the missing values ​​are filled in by a preset interpolation method to obtain the corrected first disk status data; In the case that the initial disk status data contains a value beyond a preset range, the value beyond the preset range is replaced by a preset boundary value to obtain a corrected second disk status data; Using the first disk status data or the second disk status data and the corresponding fault label as training samples; A training sample set is constructed according to the training samples.

7. The method for determining a device failure prediction model according to claim 1, characterized in that: After converting the target neural network model into a decision tree model with the target feature data as nodes according to the contribution matrix, the method further includes: Acquire actual status information of a target disk and actual fault information of the target disk; Inputting the actual status information of the target disk into the decision tree model to obtain predicted fault information of the target disk; Determining the accuracy of the decision tree model according to the difference between the predicted fault information and the actual fault information; When the accuracy is lower than a threshold, the decision tree model is updated in response to an update operation of the decision tree model by a user.

8. The method for determining a device failure prediction model according to claim 7, characterized in that: The updating of the decision tree model includes: modifying the decision tree model based on target modification information, incrementally training the target neural network model, and fully training the target neural network model.

9. A device for determining a device failure prediction model, characterized in that: The device comprises: A first acquisition module is used to acquire a target neural network model for device fault prediction and a disk data sample set, wherein the disk data sample set includes multiple disk data samples, and the disk data samples include disk status data of the disk and a fault label corresponding to the disk status data; the disk status data includes feature data in multiple dimensions; A contribution determination module, used for inputting the target feature data of each disk data sample into the target neural network model to obtain the contribution corresponding to the target feature data; A construction module, used to construct a contribution matrix according to the contribution corresponding to each of the target feature data; A conversion module is used to convert the target neural network model into a decision tree model with the target feature data as nodes according to the contribution matrix.

10. A device failure prediction model determining device, characterized in that: The device includes: a processor, and a memory storing computer program instructions; the processor reads and executes the computer program instructions to implement the method for determining a device fault prediction model as described in any one of claims 1-8.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by a processor, the method for determining a device fault prediction model according to any one of claims 1 to 8 is implemented.

12. A computer program product, characterized in that When the instructions in the computer program product are executed by a processor of an electronic device, the electronic device executes the method for determining a device fault prediction model as described in any one of claims 1 to 8.

Images