Server fault prediction method and device, electronic equipment and storage medium

Abstract

Embodiments of the invention provide a server fault prediction method and apparatus, an electronic device and a storage medium. The method comprises the steps of obtaining a server historical data set; extracting data features and truth values from the historical data set of the server; performing noise adding operation on the data features to generate noise-added data features; performing noise adding operation on the truth value to generate a noise-added truth value; generating an initial prediction result through the noise-added data feature and the noise-added true value; determining a mean square error for expressing training loss based on an initial prediction result and the true value, and determining a prediction model based on the mean square error; and when real-time data of a target server is obtained, generating a fault prediction result for the target server through the prediction model based on the real-time data. The fault of the server is predicted in the mode, so that early warning of the fault of the server is realized, and a decision basis is provided for operation and maintenance personnel.

Description

Technical Field

[0001] The present invention relates to the technical field of server fault prediction, and in particular to a server fault prediction method, a server fault prediction device, an electronic device and a computer-readable storage medium. Background Art

[0002] With the explosive growth of Internet business, the number of servers supporting its operation is also rising sharply. Although servers play an important role in the development of science and technology with their powerful GPU computing power and excellent compatibility, old equipment and high-load operation are very likely to cause various failures. Once a sudden failure or downtime occurs, it may not only lead to production interruption, but also cause huge economic losses.

[0003] Related technologies often detect and repair server failures after the fact, that is, after the failure occurs, the cause is analyzed manually and measures are taken. Alternatively, companies will regularly perform routine maintenance on servers, but this not only affects production efficiency, but also has high costs. Summary of the invention

[0004] The embodiments of the present invention provide a server fault prediction method, device, electronic device and computer-readable storage medium to overcome the above problems or at least partially solve the above problems.

[0005] The embodiment of the present invention discloses a server failure prediction method, comprising:

[0006] Get the server historical data set;

[0007] Extracting data features and true values ​​from the server historical data set;

[0008] Performing a noise addition operation on the data feature to generate a noisy data feature;

[0009] Performing a noise addition operation on the true value to generate a noisy true value;

[0010] Generate an initial prediction result by using the noise-added data features and the noise-added true value;

[0011] Determine a mean square error for expressing training loss based on the initial prediction result and the true value, and determine a prediction model based on the mean square error;

[0012] When the real-time data of the target server is acquired, a fault prediction result for the target server is generated through the prediction model based on the real-time data.

[0013] The embodiment of the present invention also discloses a server failure prediction device, comprising:

[0014] A historical data set acquisition module is used to acquire the server historical data set;

[0015] A data feature and true value extraction module, used to extract data features and true values ​​from the server historical data set;

[0016] A first noise adding module, used for performing a noise adding operation on the data feature to generate a noise adding data feature;

[0017] A second noise adding module, used for performing a noise adding operation on the true value to generate a noisy true value;

[0018] An initial prediction result generating module, used for generating an initial prediction result through the noise-added data feature and the noise-added true value;

[0019] A prediction model determination module, used to determine a mean square error for expressing training loss based on an initial prediction result and the true value, and determine a prediction model based on the mean square error;

[0020] The fault prediction result generating module is used to generate a fault prediction result for the target server based on the real-time data through the prediction model when the real-time data of the target server is acquired.

[0021] The embodiment of the present invention further discloses an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0022] The memory is used to store computer programs;

[0023] The processor is used to implement the method described in the embodiment of the present invention when executing the program stored in the memory.

[0024] The embodiment of the present invention further discloses a computer-readable storage medium having instructions stored thereon, which, when executed by one or more processors, enables the processors to execute the method described in the embodiment of the present invention.

[0025] The embodiments of the present invention include the following advantages:

[0026] The embodiment of the present invention obtains a server historical data set; extracts data features and true values ​​from the server historical data set; performs a noise operation on the data features to generate a noisy data feature; performs a noise operation on the true value to generate a noisy true value; generates an initial prediction result through the noisy data feature and the noisy true value; determines a mean square error for expressing the training loss based on the initial prediction result and the true value, and determines a prediction model based on the mean square error; when the real-time data of the target server is obtained, generates a fault prediction result for the target server through the prediction model based on the real-time data. The server fault is predicted in the above manner, and an early warning of the server fault is achieved, providing a decision-making basis for operation and maintenance personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flow chart of the steps of a server failure prediction method provided in an embodiment of the present invention;

[0028] Figure 2 It is a flow chart of a server failure prediction method provided in an embodiment of the present invention;

[0029] Figure 3 is a schematic diagram of a flow chart of a calculation model loss provided in an embodiment of the present invention;

[0030] Figure 4 is a structural block diagram of a server fault prediction device provided in an embodiment of the present invention;

[0031] Figure 5 is a hardware structure block diagram of an electronic device provided in an embodiment of the present invention;

[0032] Figure 6 is a schematic diagram of a computer-readable medium provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] Reference Figure 1 , shows a flow chart of the steps of a server failure prediction method provided in an embodiment of the present invention, which may specifically include the following steps:

[0035] Step 101, obtaining a server historical data set;

[0036] Step 102, extracting data features and true values ​​from the server historical data set;

[0037] Step 103, performing a noise addition operation on the data feature to generate a noisy data feature;

[0038] Step 104, performing a noise addition operation on the true value to generate a noisy true value;

[0039] Step 105, generating an initial prediction result through the noise-added data feature and the noise-added true value;

[0040] Step 106, determining a mean square error for expressing training loss based on the initial prediction result and the true value, and determining a prediction model based on the mean square error;

[0041] Step 107: When the real-time data of the target server is obtained, a fault prediction result for the target server is generated through the prediction model based on the real-time data.

[0042] The embodiment of the present invention can obtain a server historical data set.

[0043] For example, historical server operation data can be collected and organized, including various performance indicators such as central processing unit CPU usage, memory usage, disk I / O, network traffic, and server failure records, thereby providing sufficient data samples for subsequent model training.

[0044] The embodiment of the present invention can extract data features and true values ​​from the server historical data set to extract features meaningful for fault prediction from the original data and determine the corresponding fault label (true value).

[0045] In the specific implementation, meaningful features for fault prediction can be extracted from historical data, such as the peak value of the CPU load, the frequency of memory leaks, etc. At the same time, determine which data indicates whether the server has failed as the true value of the model, thereby reducing data redundancy, improving model training efficiency, and providing discriminative features for the model.

[0046] The embodiment of the present invention can perform a noise addition operation on the data features to generate noisy data features, thereby increasing the diversity of the data and improving the robustness of the model so that it can better adapt to different data distributions.

[0047] Random noise is added to the original data features to simulate data fluctuations in the real environment, enhance the robustness of the model, avoid model overfitting, and improve the generalization ability of the model.

[0048] In the embodiment of the present invention, a noise adding operation can be performed on the true value to generate a noisy true value, thereby simulating the uncertainty existing in the real scene and improving the tolerance of the model to noise.

[0049] In a specific implementation, an embodiment of the present invention increases the difficulty of training by adding noise to the true value labels, forcing the model to learn deeper features and enhancing the robustness of the model to data disturbances.

[0050] In the embodiment of the present invention, an initial prediction result can be generated by using the noisy data features and the noisy true value, so as to use the noisy data to train a model and obtain an initial prediction result.

[0051] The noisy data features and true values ​​are input into the trained model to obtain the initial prediction results, providing a basis for subsequent model optimization.

[0052] In an embodiment of the present invention, a mean square error for expressing training loss can be determined based on the initial prediction result and the true value, and a prediction model can be determined based on the mean square error, so as to minimize the mean square error and continuously adjust the model parameters so that the difference between the prediction result of the model and the true value is minimized.

[0053] The mean square error between the predicted result and the true value is calculated and used as the loss function. The model parameters are adjusted through the optimization algorithm to minimize the error between the predicted result and the true value, thereby optimizing the prediction performance of the model.

[0054] According to an embodiment of the present invention, when real-time data of a target server is acquired, a fault prediction result for the target server can be generated through the prediction model based on the real-time data, so as to apply the trained model to actual scenarios and perform fault prediction on the target server.

[0055] The newly collected server operation data is input into the trained model to obtain predictions of possible future failures of the server, thereby providing early warning of server failures and reducing losses.

[0056] The embodiment of the present invention obtains a server historical data set; extracts data features and true values ​​from the server historical data set; performs a noise operation on the data features to generate a noisy data feature; performs a noise operation on the true value to generate a noisy true value; generates an initial prediction result through the noisy data feature and the noisy true value; determines a mean square error for expressing the training loss based on the initial prediction result and the true value, and determines a prediction model based on the mean square error; when the real-time data of the target server is obtained, generates a fault prediction result for the target server through the prediction model based on the real-time data. The server fault is predicted in the above manner, and an early warning of the server fault is achieved, providing a decision-making basis for operation and maintenance personnel.

[0057] Furthermore, the following beneficial effects can be achieved:

[0058] Improve fault prediction accuracy: By introducing noise, optimizing loss functions, and other means, the generalization ability and robustness of the model can be improved, thereby improving prediction accuracy.

[0059] Reduce operation and maintenance costs: discover potential failures in advance, reduce losses caused by failures, and improve system availability.

[0060] Improve operation and maintenance efficiency: Automate the fault prediction process, reduce manual intervention, and improve operation and maintenance efficiency.

[0061] Enhance system reliability: Through real-time monitoring and prediction of server status, take measures in advance to ensure stable operation of the system.

[0062] The solution builds a complete server failure prediction system by combining data enhancement, feature extraction, model training and prediction technologies. The system has high prediction accuracy, robustness and interpretability, providing strong technical support for ensuring the stable operation of the server system.

[0063] Based on the above embodiment, a variant embodiment of the above embodiment is proposed. It should be noted that in order to make the description concise, only the differences from the above embodiment are described in the variant embodiment.

[0064] In practical applications, historical data sets may include training sets and test sets; however, in order to ensure the learning effect of the server failure prediction model on abnormal data, in an optional embodiment of the present invention, all historical data sets containing abnormal data such as server failures may be divided into test sets. In the data preprocessing stage, we use the improved Z-score method to ensure data quality for server data missing or abnormal problems caused by network fluctuations or bandwidth limitations.

[0065] In the model part, a prediction model including a data encoder and a feature decoder can be designed. The data encoder is responsible for feature extraction of training data, converting high-dimensional raw data into low-dimensional, more representative features for subsequent model learning.

[0066] The purpose is to:

[0067] Clear data division: All samples containing abnormal data are divided into the test set to verify the model's ability to identify and predict abnormal data.

[0068] Data preprocessing: The improved Z-score method is used to process abnormal data to ensure the quality and reliability of the data.

[0069] Model design: Design a data encoder to extract data features and provide a basis for subsequent predictions.

[0070] Beneficial effects include:

[0071] Improve model robustness: By including all abnormal data in the test set, the performance of the model in the face of abnormal situations can be more comprehensively evaluated, thus improving the robustness of the model.

[0072] Enhance model generalization ability: The improved Z-score method can effectively handle abnormal data, enabling the model to better adapt to data distribution in different scenarios and improve the generalization ability of the model.

[0073] Reduce model complexity: The data encoder converts high-dimensional data into low-dimensional features, reducing the complexity of the model and improving training efficiency.

[0074] Improve prediction accuracy: By extracting representative features, the model can better capture the underlying patterns in the data, thereby improving the accuracy of fault prediction.

[0075] Optionally, before the step of extracting data features and true values ​​from the server historical data set, the method further includes:

[0076] Performing a data cleaning operation on the historical data set;

[0077] Determining a constant factor for expressing the distribution state of the historical data set after the data cleaning operation;

[0078] Determine the median of the historical data set after the data cleaning operation, and the absolute deviation from the median;

[0079] performing a normalization operation on the historical data set subjected to the data cleaning operation based on the constant factor, the median and the absolute deviation;

[0080] The historical data set that has undergone the normalization operation is divided into a plurality of subsequences according to time windows.

[0081] In a specific implementation, the embodiment of the present invention adopts an improved Z-score method to handle server data loss or abnormality problems caused by network jitter or bandwidth limitation. This method calculates the difference between each data point and the median, and then divides it by the median absolute deviation (MAD) to obtain a standardized value. A constant factor α is introduced to adjust the distribution of the standardized data.

[0082] For example, the preprocessing operation can be performed on the historical data set in the following manner:

[0083] Data cleaning: remove missing values, outliers, and duplicate data.

[0084] Data normalization: The data were normalized using the modified Z-score method.

[0085] Compute the median and absolute deviation from the median of a data set.

[0086] Calculate the normalized value for each data point according to the formula.

[0087] For example, a constant factor for expressing the distribution state of the historical data set subjected to the data cleaning operation may be determined;

[0088] Determine the median of the historical data set after the data cleaning operation, and the absolute deviation from the median;

[0089] Inputting the historical data set that has undergone the data cleaning operation, the constant factor, the median, and the absolute deviation into Formula 1 to perform a normalization operation;

[0090] Formula 1:

[0091]

[0092] Where α is a constant factor of the data set, which is related to the distribution of the server data set. median(x i ) represents the median of the data, MAD is the absolute deviation of the median, x i Represents the historical data set that will be subjected to the data cleaning operation.

[0093] Data windowing: The historical data set that has undergone the normalization operation is divided into multiple subsequences according to time windows, which are subsequently used as inputs to the encoder model.

[0094] Optionally, the step of extracting data features and true values ​​from the server historical data set includes:

[0095] Determine the number of layers of the encoder model and the number of neurons in each encoder layer to construct the encoder model;

[0096] Inputting the plurality of subsequences into the encoder model, and controlling the encoder model to generate data features;

[0097] A true value is extracted based on the historical data set that has been subjected to the data cleaning operation.

[0098] Exemplarily, data features and true values ​​may be extracted in the following manner.

[0099] Build the encoder model:

[0100] Encoder selection: You can choose a classic convolutional neural network (such as LeNet, VGG, ResNet) or a transformer as the encoder.

[0101] Feature extraction: The main purpose of the encoder is to extract features from the input data to provide a basis for subsequent predictions.

[0102] Network structure: For convolutional neural networks, skip connections can effectively solve the gradient vanishing problem and improve the performance of the model. Transformer captures the long-distance dependencies between input data through the self-attention mechanism.

[0103] Optionally, ResNeXt101 is a variant of the ResNet network and belongs to the family of deep residual networks. It introduces the concept of grouped convolution based on ResNet, improves the performance of the model by increasing the width of the network while maintaining the complexity of the model. If ResNeXt101 is used as the feature extraction network, the ResNeXt101 network introduces the idea of ​​grouped convolution based on the ResNet network, draws on the stacking idea of ​​VGG and the ideas of segmentation, transformation, and merging in Inception, and adds a new dimension C to widen the network structure. The newly added dimension C makes the network more scalable. When performing feature extraction, it improves the accuracy of model feature extraction without increasing the complexity of the model. The core formula of ResNeXt101 is:

[0104]

[0105] The formula shows that the output of ResNeXt is obtained by adding the outputs of multiple groups. Each group learns a part of the features of the input data, and by adding these features, a richer feature representation can be obtained. k It refers to a function that projects the data of the hidden layer into a lower-dimensional space and can be transformed arbitrarily. Since each transformation is performed in a low-dimensional environment, the features are summarized through the final merging operation, and the loss of two-dimensional image features is minimized, ensuring that the model learns deeper features and the accuracy of fault prediction.

[0106] Alternatively, Visual Transformer (ViT) is an image processing model based on the Transformer architecture, which is widely used for feature extraction. ViT divides the image into a series of patches and treats them as a sequence, and then uses the self-attention mechanism of the Transformer to capture the global dependencies between patches.

[0107] If the ViT encoder is used as the feature extraction network, Visual Transformer (ViT) is an image processing model based on the Transformer architecture, which divides the image into a series of small patches, and then inputs these small patches as a sequence into the Transformer for processing. The core idea of ​​ViT is to use the Transformer's self-attention mechanism to capture the global dependencies in the image.

[0108] Self-attention is the core of the Transformer, which works by computing the relationship between all elements (i.e. image patches) in a sequence. In ViT, the self-attention layer computes the attention scores between each image patch and all other patches, which indicate how much attention should be given to other patches when generating the representation of the current patch. After being processed by multiple Transformer layers, the representation of each image patch is integrated and improved to include global context information.

[0109]

[0110] For each element i, three different vectors are generated, namely the query vector O i , key vector K i Sum value vector V i .

[0111] For each element in the sequence, we first calculate the attention scores between it and all other elements in the sequence. These scores reflect the correlation between different elements. Then, we perform a weighted summation of the value vectors of other elements according to these scores to obtain the output representation of the current element.

[0112] Output vector O i It captures the contribution of all elements in the sequence to the current element i. The self-attention mechanism can be extended to multiple layers, and the output of each layer can be used as the input of the next layer, allowing the model to capture the global dependencies of the sequence at different levels. Compared with traditional convolutional neural networks, this method reduces the number of parameters and the computational complexity of the model.

[0113] The true value plays a vital role in machine learning tasks, especially supervised learning tasks. It is equivalent to the "answer" of model learning. The model continuously adjusts parameters to minimize the difference between the predicted result and the true value. In the server failure prediction task, the true value usually represents whether the server fails or the specific type of failure.

[0114] Exemplarily, the source and extraction method of the true value include:

[0115] Fully automated annotation: For some clearly defined fault types, automatic annotation can be performed using a rule engine or a simple machine learning model.

[0116] Real-time monitoring data:

[0117] Alarm system: When an abnormality occurs on the server, the alarm system will issue an alarm, which can be used as part of the true value.

[0118] Log analysis: By analyzing server logs, some fault features can be extracted as part of the true value.

[0119] The above methods can all be performed on historical data sets that have undergone data cleaning operations.

[0120] By preprocessing the historical data set and extracting data features and true values ​​in the above manner, the following beneficial effects can be achieved.

[0121] Summary of beneficial effects

[0122] Improve the expressiveness of the model:

[0123] Multi-head self-attention mechanism: It can capture long-distance dependencies in sequence data and the interactions between different features, thereby representing the data more comprehensively.

[0124] Feedforward fully connected layer: Introduces nonlinear transformations to enhance the expressiveness of the model, enabling it to learn more complex features.

[0125] Improve model training efficiency:

[0126] Residual connection: Alleviates the gradient vanishing problem of deep networks, allowing the model to be trained deeper and thus extract higher-level features.

[0127] Pre-trained models: You can use the powerful representation capabilities of pre-trained models to reduce the time of training models from scratch and improve training efficiency.

[0128] Enhance the robustness of the model:

[0129] Improved Z-score standardization: By standardizing the data, the impact of outliers on the model is reduced and the model's robustness to noise is improved.

[0130] Data preprocessing: missing values, outliers, etc. were processed to ensure the quality of input data and improve the stability of the model.

[0131] Reduce the risk of overfitting the model:

[0132] Pre-trained model: The pre-trained model is trained on a large amount of data, which can effectively prevent overfitting.

[0133] Regularization techniques (such as L1 / L2 regularization): can prevent overfitting by limiting the complexity of the model.

[0134] Optionally, the step of performing a noise addition operation on the data feature to generate a noise added data feature comprises:

[0135] Determine the dimension and noise intensity parameter of the data feature, and generate a random noise matrix obeying Gaussian distribution based on the dimension and the noise intensity parameter;

[0136] Superimposing the random noise matrix with the data feature to generate an initial noisy data feature;

[0137] Noise intensity scheduling is performed on the initial noise-added data feature to generate a noise-added data feature.

[0138] Exemplarily, the noisy data feature may be generated in the following manner.

[0139] Generate Gaussian noise: According to the dimension of data features and noise intensity parameter α t , generates a random noise matrix that follows a Gaussian distribution.

[0140] Noise superposition: Add the generated noise matrix to the data features to obtain noisy features.

[0141] Noise intensity scheduling: As training progresses, gradually reduce the noise intensity parameter α t .

[0142] z t =α t ·u t +z0

[0143] z t : Data features after adding noise at time t.

[0144] α t : The noise intensity coefficient at time t, which is a scalar that decreases with time.

[0145] u t : The original features at time t, that is, the data features without adding noise.

[0146] z0: A fixed noise floor, which can be regarded as the initial noise.

[0147] This formula means that at each time step t, we transform the original feature u t Multiply by a decay coefficient α t , plus a fixed noise floor z0, to get a noisy feature z t As t increases, αt As the value decreases gradually, the intensity of the noise also decreases.

[0148] By generating noise-added data features in the above manner, the beneficial effects that can be achieved include:

[0149] Improve generalization ability: By introducing noise, the model can learn more robust features and perform better when faced with never-seen data.

[0150] Prevent overfitting: The presence of noise makes it difficult for the model to overfit the training data, thereby improving the generalization ability of the model.

[0151] Enhance the robustness of the model to noise: In practical applications, data often contains noise. By adding noise during the training process, the robustness of the model to noise can be improved.

[0152] Explore a larger solution space: In the early stages of training, a larger noise intensity helps the model explore more solution space and find better local minima.

[0153] By adding Gaussian noise and combining it with noise intensity scheduling, a data enhancement method is implemented. This method is widely used in deep learning, especially in the fields of image and speech. It can effectively improve the performance of the model and make it more robust and generalizable in practical applications.

[0154] In an optional embodiment of the present invention, the step of performing a noise addition operation on the true value to generate a noisy true value comprises:

[0155] Converting a discrete label used to express the true value into a dense vector, and performing a normalization operation on the dense vector;

[0156] The normalized dense vector is mapped into a preset range to generate a noisy true value.

[0157] The embodiment of the present invention can convert discrete labels (ie, true values) into a continuous vector representation process, and introduces some noise to enhance the robustness of the model.

[0158] Exemplarily, the noisy true value may be generated in the following manner.

[0159] Embedding layer mapping:

[0160] Convert discrete labels to dense vectors: Use a learnable embedding layer to map discrete category labels into a high-dimensional continuous space. This is similar to how words are mapped to word vectors.

[0161] Introducing nonlinearity: The vectors learned by the embedding layer can capture the semantic relationship between categories.

[0162] Sigmoid Normalization:

[0163] Mapping vector elements to the interval [0,1]: Use the sigmoid function to normalize the output of the embedding layer so that the value of each element is between 0 and 1.

[0164] Zoom:

[0165] Adjust the vector range: Multiply the normalized vector by a scaling factor so that it is in the range [-scale, +scale]. This range can be adjusted according to the specific task.

[0166] The beneficial effects of generating the noise-added true value in the above manner are as follows:

[0167] Enhance the expressiveness of the model: Converting discrete labels into continuous vectors can provide the model with richer features, thereby improving the expressiveness of the model.

[0168] Improve the robustness of the model: By introducing noise, the model can be made more robust to perturbations of the input data, thereby improving the generalization ability of the model.

[0169] Alleviate the class imbalance problem: In multi-classification problems, the number of samples in different categories is often unbalanced. By mapping the labels to a continuous space, the class imbalance problem can be alleviated.

[0170] Convenient for subsequent calculations: Converting discrete labels into continuous vectors can facilitate subsequent calculations, such as calculating loss functions and performing gradient descent.

[0171] In an optional embodiment of the present invention, the step of generating an initial prediction result by using the noisy data feature and the noisy true value comprises:

[0172] The noisy data feature and the noisy true value are connected in series to generate a concatenated vector;

[0173] The concatenated vector is input into a decoder to control the decoder to output an initial prediction result; the decoder is generated by stacking a masked multi-head self-attention layer, an encoder-decoder attention layer, and a feedforward neural network.

[0174] Exemplarily, the initial prediction result may be generated in the following manner.

[0175] The processed true value and data features are concatenated in a preset order into a new concatenated vector, and the concatenated vector is input into the decoder.

[0176] The decoder is composed of multiple decoder layers stacked together, each of which performs the following operations:

[0177] Masked multi-head self-attention layer: focuses on the input before the current position, preventing the model from seeing future information, thereby ensuring the causality of the generation process.

[0178] Encoder-Decoder Attention Layer: Establishes a connection between the decoder and the encoder so that the decoder can utilize the feature information extracted from the encoder.

[0179] Feedforward neural network: Perform nonlinear transformation on features to increase the expressive power of the model.

[0180] Optionally, the decoder is responsible for restoring the features generated by the encoder to the original data form. It can be a convolutional neural network (CNN), a multi-layer perceptron (MLP), or a Transformer.

[0181] MLP: Maps input to output through fully connected layers and activation functions.

[0182] CNN: Use convolution operations to extract local features, and then perform classification or regression through fully connected layers.

[0183] Transformer: Uses self-attention mechanism to capture long-distance dependencies between elements in a sequence and generate a target sequence.

[0184] To improve the performance of the decoder, the following techniques can be used:

[0185] Depthwise separable convolution: reduces the number of parameters and improves computational efficiency.

[0186] Mish activation function: enhances the nonlinear expression ability of the model.

[0187] Layer normalization and residual connections: Stabilize the training process and prevent gradients from disappearing or exploding.

[0188] Through these technologies, an efficient and accurate decoder can be built to achieve effective restoration and prediction of data.

[0189] By generating the initial prediction results in the above manner, the following beneficial effects can be achieved:

[0190] Fusion of contextual information: By concatenating the true value and the feature, the decoder can simultaneously utilize the contextual information and feature information of the sequence to generate more accurate prediction results.

[0191] Improve the expressiveness of the model: The combination of the multi-head self-attention mechanism and the feedforward neural network enables the model to capture long-distance dependencies in sequence data and perform complex feature transformations.

[0192] Enhance the generalization ability of the model: By introducing the encoder-decoder attention mechanism, the model can better handle input sequences of different lengths and improve the generalization ability of the model.

[0193] Generate high-quality output sequences: The decoder gradually generates the output sequence in an autoregressive manner, which can generate more coherent and high-quality output.

[0194] In an optional embodiment of the present invention, the step of determining a mean square error for expressing training loss based on the initial prediction result and the true value comprises:

[0195] Determining an error prediction value for the initial prediction result;

[0196] The error prediction value is input into the following formula to calculate the mean square error;

[0197]

[0198] Where f(t) is the true value, is the error prediction value.

[0199] Exemplarily, the initial prediction result of the model is compared with the true value to obtain the error prediction value; the mean square error (MSE) is used as the loss function to quantify the size of the prediction error; through the back propagation algorithm, the gradient of the loss function is passed to the various parameters of the model, and the parameters are updated according to the gradient descent method.

[0200] MSE: Mean Squared Error. It is a commonly used loss function to measure the difference between the predicted value and the true value.

[0201] f(t): represents the real input data at time t, that is, the true value.

[0202] : Represents the output data predicted by the model at time t, that is, the error prediction value.

[0203] : It means summing the square of the difference between the predicted value and the true value at each time step.

[0204] By calculating the mean square error in the above manner, the following beneficial effects can be achieved:

[0205] Quantifying model performance: The mean square error can quantitatively measure the prediction effect of the model. The smaller the value, the higher the prediction accuracy of the model.

[0206] Guide model optimization: By calculating the gradient of the loss function, we can know which parameters of the model need to be adjusted, so as to achieve model optimization.

[0207] Realize model learning: The back propagation algorithm enables the model to learn patterns from data and continuously improve prediction accuracy.

[0208] In order to enable those skilled in the art to better understand the embodiments of the present invention, a complete example is used below to illustrate the embodiments of the present invention.

[0209] refer to Figure 2 , Figure 2 It is a flow chart of a server failure prediction method provided in an embodiment of the present invention;

[0210] In actual applications, with the rapid development of network technology, network services have gradually skyrocketed, and the number of servers has also increased year by year. Servers rely on powerful GPU (Graphics Processing Unit) computing power and excellent compatibility to occupy a place in the road of technological development, but old and heavily loaded servers are prone to various failures. Sudden failures and downtime may cause major production accidents and increase the silent costs of enterprises. Traditional server fault prediction and maintenance are to manually analyze the fault type and solve it after the server fails; or set a fixed period for maintenance, which not only affects production efficiency but also has high costs.

[0211] Related technologies are based on the refrigerator fault prediction method of convolutional neural network. After obtaining the refrigerator operation data, the data is processed based on the trained convolutional neural network to determine the refrigerator operation status, and then the refrigerator status is fed back to the user platform. However, in the field of cloud computing, the amount of data generated during server operation is huge, and the convolutional neural network is less effective in processing sequence data with long-term dependencies, and the accuracy of fault prediction cannot be guaranteed.

[0212] The related technology has the following disadvantages:

[0213] (1) Common machine learning methods are unable to learn, train, and test large-scale data sets, resulting in extremely low accuracy and efficiency in server failure prediction;

[0214] (2) RNN (Recurrent Neural Network) can only process input data one by one, resulting in high latency when processing large amounts of server data and the inability to provide fault prediction results in a timely manner.

[0215] In view of the above shortcomings, a server failure prediction method based on diffusion model is proposed to design the technical field of server failure prediction.

[0216] (1) The RNN serial processing method cannot process the large amount of data sets generated by the server in a timely manner. By introducing the Transformer and using matrix operations, the efficiency of fault prediction can be improved.

[0217] (2) The deep structure of the diffusion model enables it to learn more complex and abstract feature representations. By stacking multiple diffusion layers, the network can gradually learn the high-level features of the data, thereby achieving a higher level of representation and reasoning capabilities.

[0218] The training process is mainly divided into two modules: data encoder and feature decoder.

[0219] The data encoder encodes the historical data to obtain the feature F, and uses these features as conditions to gradually refine the prediction from the noise features. The historical data truth value pairs are noised through the Markov chain rule and time step to obtain the noise feature Zt. The noise feature Zt is then connected in series with F, and the dimensions of the two are consistent. If they are inconsistent, they are reshaped to the same dimension.

[0220] Finally, it is sent to the feature decoder to obtain the predicted result. The feature decoder is composed of a six-layer Transformer encoder. In addition, the feature encoder can also choose other convolutional networks or Transformer structure networks.

[0221] Testing process: During testing, Zt is not obtained from the true value, but from random noise. You only need to pass the test data through the data encoder to obtain the feature F, then connect it in series with the random noise Zt, and send it to the feature decoder to get the final prediction result. Note: The feature decoder can choose multiple inferences here. For example, the prediction result of the first step is continued to be input into the feature decoder for secondary prediction. Without occupying additional video memory, the prediction result will be further refined.

[0222] Exemplarily, a method for predicting server failures based on a diffusion model includes the following steps:

[0223] Step 1: Collect the server historical data set and divide it into training set and test set for training the prediction model;

[0224] Step 2: Preprocess the historical server data set and input it into the data encoder model to extract data features F. Here, the data encoder can choose some pre-trained models with better performance to process, and can choose not to update parameters, reduce the number of model parameters, and improve efficiency;

[0225] In step 2, the data encoder model selects a learnable encoder layer by default, where the encoding part is composed of multiple encoder layers stacked together, and each encoder layer consists of two sub-layer connection structures. The first sub-layer connection structure includes a multi-head self-attention sub-layer, a normalization layer, and a residual connection, and the second sub-layer connection structure includes a feedforward fully connected sub-layer, a normalization layer, and a residual connection. In addition, you can also choose a pre-trained model with better performance for processing, and you can choose not to update the parameters to reduce the number of model parameters and improve efficiency. Finally, the encoded feature is named F as the input for the next step.

[0226] After preprocessing, the historical server data set is input into the data encoder model to extract data features F. Here, the data encoder can select some pre-trained models with better performance for processing, and can choose not to update parameters to reduce the number of model parameters and improve efficiency; the data preprocessing stage is mainly aimed at the processing means for server data loss or abnormal problems caused by network jitter or limited bandwidth. This article uses the modified Z-score method to ensure data quality for special working scenarios of the server. The calculation formula is:

[0227]

[0228] Where α is a constant factor of the data set, which is related to the distribution of the server data set. median(x i ) represents the median of the data, and MAD is the absolute deviation from the median.

[0229] Step 3: Add noise to the data feature F to obtain F';

[0230] In step 3, noise is added to the data feature F, which is to obtain the noise feature by adding Gaussian noise. The intensity of the noise is determined by the parameter α. t To control, a monotonically decreasing schedule is adopted in different time steps t∈[0,T].

[0231] z t =α t ·u t +z0

[0232] Step 4: Add noise to the true value of the data G to obtain G';

[0233] In step 4, adding noise to the true value of the data G is to use a learnable embedding layer to project the discrete labels into a high-dimensional continuous space and normalize them using the sigmoid function. Finally, the label range is normalized and scaled to be within the range of [-scale, +scale].

[0234] Step 5: Concatenate the true value G' after adding noise with the data feature F' and send them to the feature decoder for prediction to obtain the initial prediction result;

[0235] refer to Figure 3 , Figure 3 is a schematic diagram of a flow chart of a calculation model loss provided in an embodiment of the present invention;

[0236] The server failure prediction model consists of two parts: the encoding stage and the decoding stage. The operations of the encoder and decoder in the model are both implemented using neural networks. In order to ensure that the encoder has good feature extraction capabilities, a classic and effective feature extraction network can be selected to implement it, such as a convolutional neural network or a Transformer (the usage of the two will be listed below).

[0237] The encoding stage mainly encodes the input data, predicts the amplitude of future data based on historical data, and obtains the prediction through the encoder's prediction reconstruction operation. The encoder can choose a pre-trained model to reduce training overhead. If a convolutional neural network type encoder is selected, the neurons in each layer of the encoder use jump connections to ensure that the problems of gradient explosion and gradient disappearance are avoided in deeper neural networks. The jump connection method can largely solve the problem of inaccurate feature extraction and ensure the accuracy of model prediction. The main purpose of the encoder is to model the relationship between input error data, learn the characteristics of input error data, and continuously encode the characteristics of the hidden layer to obtain reliability features with strong representation capabilities of input error data.

[0238] The convolutional neural network type encoder includes multiple groups of convolution and ReLu activation functions; the encoder takes the collected and preprocessed two-dimensional image data as input and obtains data features through the convolutional neural network; here, the classic convolutional neural network can be used to encode the features, such as LeNet network, VGG16, ResNet series network, etc.

[0239] The Transformer type encoder contains multiple groups of attention and LayerNorm normalization functions; the input data is input into the self-attention in the form of tokens to learn the correlation between the input error data; multiple layers of attention can be used here, and the default setting is six layers.

[0240] If ResNeXt101 is used as the feature extraction network, the ResNeXt101 network introduces the idea of ​​group convolution on the basis of the ResNet network, draws on the stacking idea of ​​VGG and the ideas of segmentation, conversion, and merging in Inception, and adds a new dimension C to widen the network structure. The newly added dimension C makes the network more scalable, and when performing feature extraction, it improves the accuracy of model feature extraction without increasing the complexity of the model. The core formula of ResNeXt101 is:

[0241]

[0242] Among them, H k It refers to a function that projects the data of the hidden layer into a lower-dimensional space and can be transformed arbitrarily. Since each transformation is performed in a low-dimensional environment, the features are summarized through the final merging operation, and the loss of two-dimensional image features is minimized, ensuring that the model learns deeper features and the accuracy of fault prediction.

[0243] If the ViT encoder is used as the feature extraction network, Visual Transformer (ViT) is an image processing model based on the Transformer architecture, which divides the image into a series of small patches, and then inputs these small patches as a sequence into the Transformer for processing. The core idea of ​​ViT is to use the self-attention mechanism of the Transformer to capture the global dependencies in the image. Self-attention is the core of the Transformer, which works by calculating the relationship between all elements (i.e. image patches) in a sequence. In ViT, the self-attention layer calculates the attention scores between each image patch and all other patches. These scores indicate how much attention should be given to other patches when generating the representation of the current patch. After being processed by multiple Transformer layers, the representation of each image patch is integrated and enhanced, containing global context information.

[0244]

[0245] For each element i, three different vectors are generated, namely the query vector O i , key vector K i Sum value vector V i .

[0246] Output vector O iIt captures the contribution of all elements in the sequence to the current element i. The self-attention mechanism can be extended to multiple layers, and the output of each layer can be used as the input of the next layer, allowing the model to capture the global dependencies of the sequence at different levels. Compared with traditional convolutional neural networks, this method reduces the number of parameters and the computational complexity of the model.

[0247] The decoder performs a decoding operation on the result of the above encoding, that is, a process from feature dimensionality reduction to feature restoration and dimensionality increase, and finally obtains output features of the same dimension as the input features; decodes the data obtained after noise addition to obtain the final prediction data of the model; finally, uses the prediction loss function to correct the prediction results and estimate the parameters, and updates the network gradient through back propagation, and so on, until the function converges to reach the final prediction result.

[0248] The decoder can choose a classic neural network to implement, such as a convolutional neural network or a Transformer (the usage of both will be listed below).

[0249] If the decoder uses an MLP network as a decoding network. MLP (Multilayer Perceptron) is a basic artificial neural network that consists of multiple layers, including an input layer, multiple hidden layers, and an output layer. MLP is one of the most commonly used models in supervised learning, mainly used for classification and regression tasks. Each neuron in MLP is connected to all neurons in the previous layer, and nonlinearity is introduced through the activation function. This fully connected structure enables MLP to learn complex input-output mappings. The activation function is a key component in MLP, which determines how neurons update their states based on input signals. Common activation functions include Sigmoid, ReLU (Rectified Linear Unit), tanh, etc. Each connection in MLP has a weight, and each neuron has a bias term. During training, these weights and biases are adjusted through the back-propagation algorithm to minimize the difference between the predicted output and the true label. The network performs accurate classification and recognition through multi-layer convolution and pooling operations, improving the accuracy and efficiency of image recognition.

[0250] The standard convolution in the decoder is replaced by depthwise convolution and pointwise convolution to reduce model parameters. Batch normalization and Mish activation function are performed after the convolution layer. In the depthwise convolution, a 3*3 convolution kernel is first used to convolve on the input channel. The number of features obtained is the same as the target number of features, but the number of channels does not reach the target number of output channels. Therefore, it is necessary to continue to use pointwise convolution to use 4 groups of 1*1 convolution kernels to increase the number of output channels to the target number of channels. The depthwise separable convolution operation greatly improves the accuracy of the model, reduces the number of parameters, simplifies the complexity of the model, and greatly improves the running speed of the model.

[0251] If the decoder uses the Transformer decoder as the decoding network. The decoder is designed to use the output of the encoder to generate the target sequence while allowing the model to consider previously generated elements during the generation process. The decoder uses a multi-head self-attention mechanism to process information at different positions in parallel and capture complex dependencies within the sequence. For each decoder layer, the self-attention layer calculates the attention score of the input sequence of the current layer and generates a contextual representation. After each self-attention layer, the decoder uses a feedforward network to further process the representation of each position. The feedforward network usually contains two linear transformations and a nonlinear activation function. In order to facilitate the training process and reduce the problem of gradient disappearance or explosion, the Transformer uses layer normalization on the output of each sub-layer and adds residual connections. The last layer of the decoder outputs a vector sequence that passes through a linear layer and a softmax layer or through an embedding layer and a linear layer to generate the final output.

[0252] The server fault prediction model is trained by collecting historical server data, and the prediction data is processed by the encoder. Here, the encoder can select a pre-trained model with better performance for processing, and then input it into the decoder for training to achieve fault identification and prediction. In the prediction stage, it can be sent to the decoder multiple times to improve the accuracy of server fault prediction and improve the availability and reliability of the system.

[0253] Step 6: Calculate the training loss between the initial prediction result and the true value G, that is, calculate the mean square error, and adjust the prediction model parameters to obtain the best prediction model;

[0254] Step 6 calculates the loss of the prediction error, that is, calculates the mean square error. The mean square error loss function was originally used to measure the distance from the sample point to the regression curve, and by minimizing the square loss, the sample point can better fit the regression curve. The smaller the value of the mean square error loss function (MSE), the better the accuracy of the sample data described by the prediction model. Here, the mean square error loss function (MSE) is used to calculate the loss between the predicted value and the true value:

[0255]

[0256] Among them, f(t) is the initial input data, It is the forecast data.

[0257] Step 7: Collect real-time data of server operation and input it into the prediction model to obtain the prediction result of server failure.

[0258] In step 7, the real-time data of the server operation is collected and input into the prediction model to obtain the prediction result of the server failure. Different from the training, in the prediction stage, the real-time data is only sent to the data encoder once, and the Gaussian noise is randomly given in series with it, and then sent to the feature decoder to obtain the mean square error. The mean square error is sorted from large to small, and the data with a large mean square error is identified as faulty data. Since the faulty server can be located based on the faulty data, it is possible to predict whether the server is faulty.

[0259] The encoding and decoding stages are mainly to train parameters through historical data sets, where the hyperparameters of the model include: number of trainings, learning rate, learning decay rate, adjustment parameters, training batch size, etc. In particular, the trained parameters are used as learning weights in the encoding and decoding stages.

[0260] It should be noted that, for the sake of simplicity, the method embodiments are described as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because according to the embodiments of the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0261] Reference Figure 4 , shows a structural block diagram of a server fault prediction device provided in an embodiment of the present invention, which may specifically include the following modules:

[0262] The historical data set acquisition module 401 is used to acquire the server historical data set;

[0263] A data feature and true value extraction module 402, used to extract data features and true values ​​from the server historical data set;

[0264] A first noise adding module 403, configured to perform a noise adding operation on the data feature to generate a noise added data feature;

[0265] A second noise adding module 404, configured to perform a noise adding operation on the true value to generate a noisy true value;

[0266] An initial prediction result generating module 405, configured to generate an initial prediction result through the noise-added data features and the noise-added true value;

[0267] A prediction model determination module 406, configured to determine a mean square error for expressing a training loss based on an initial prediction result and the true value, and to determine a prediction model based on the mean square error;

[0268] The fault prediction result generating module 407 is used to generate a fault prediction result for the target server based on the real-time data through the prediction model when the real-time data of the target server is acquired.

[0269] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0270] In addition, an embodiment of the present invention further provides an electronic device, such as Figure 5 As shown, it includes a processor 501, a communication interface 502, a memory 503 and a communication bus 504, wherein the processor 501, the communication interface 502, and the memory 503 communicate with each other through the communication bus 504.

[0271] Memory 503, used for storing computer programs;

[0272] The processor 501 is configured to implement the server failure prediction method described in any one of the above embodiments when executing the program stored in the memory 503:

[0273] The communication bus mentioned in the above terminal can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0274] The communication interface is used for communication between the above terminal and other devices.

[0275] The memory may include a random access memory (RAM) or a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0276] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0277] like Figure 6 As shown, in another embodiment provided by the present invention, a computer-readable storage medium 601 is also provided, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer executes the server fault prediction method described in the above embodiment.

[0278] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

[0279] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the embodiments of the present invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0280] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0281] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0282] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0283] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0284] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, ROM, RAM, magnetic disks, or optical disks.

[0285] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A server failure prediction method, characterized in that: include: Get the server historical data set; Extracting data features and true values ​​from the server historical data set; Performing a noise addition operation on the data feature to generate a noisy data feature; Performing a noise addition operation on the true value to generate a noisy true value; Generate an initial prediction result by using the noise-added data features and the noise-added true value; Determine a mean square error for expressing training loss based on the initial prediction result and the true value, and determine a prediction model based on the mean square error; When the real-time data of the target server is acquired, a fault prediction result for the target server is generated through the prediction model based on the real-time data.

2. The method according to claim 1, characterized in that Before the step of extracting data features and true values ​​from the server historical data set, the method further includes: Performing a data cleaning operation on the historical data set; Determining a constant factor for expressing the distribution state of the historical data set after the data cleaning operation; Determine the median of the historical data set after the data cleaning operation, and the absolute deviation from the median; performing a normalization operation on the historical data set subjected to the data cleaning operation based on the constant factor, the median and the absolute deviation; The historical data set that has undergone the normalization operation is divided into a plurality of subsequences according to time windows.

3. The method according to claim 2, characterized in that The step of extracting data features and true values ​​from the server historical data set includes: Determine the number of layers of the encoder model and the number of neurons in each encoder layer to construct the encoder model; Inputting the plurality of subsequences into the encoder model, and controlling the encoder model to generate data features; A true value is extracted based on the historical data set that has been subjected to the data cleaning operation.

4. The method according to claim 1, characterized in that: The step of performing a noise addition operation on the data feature to generate a noise added data feature comprises: Determine the dimension and noise intensity parameter of the data feature, and generate a random noise matrix obeying Gaussian distribution based on the dimension and the noise intensity parameter; Superimposing the random noise matrix with the data feature to generate an initial noisy data feature; Noise intensity scheduling is performed on the initial noise-added data feature to generate a noise-added data feature.

5. The method according to claim 1, characterized in that The step of performing a noise addition operation on the true value to generate a noise added true value comprises: Converting a discrete label used to express the true value into a dense vector, and performing a normalization operation on the dense vector; The normalized dense vector is mapped into a preset range to generate a noisy true value.

6. The method according to claim 1, characterized in that The step of generating an initial prediction result by using the noise-added data features and the noise-added true value comprises: The noisy data feature and the noisy true value are connected in series to generate a concatenated vector; The concatenated vector is input into a decoder to control the decoder to output an initial prediction result; the decoder is generated by stacking a masked multi-head self-attention layer, an encoder-decoder attention layer, and a feedforward neural network.

7. The method according to claim 1, characterized in that The step of determining a mean square error for expressing training loss based on the initial prediction result and the true value comprises: Determining an error prediction value for the initial prediction result; The error prediction value is input into the following formula to calculate the mean square error; Where f(t) is the true value, is the error prediction value.

8. A server failure prediction device, characterized in that: include: A historical data set acquisition module is used to acquire the server historical data set; A data feature and true value extraction module, used to extract data features and true values ​​from the server historical data set; A first noise adding module, used for performing a noise adding operation on the data feature to generate a noise adding data feature; A second noise adding module, used for performing a noise adding operation on the true value to generate a noisy true value; An initial prediction result generating module, used for generating an initial prediction result through the noise-added data feature and the noise-added true value; A prediction model determination module, used to determine a mean square error for expressing training loss based on an initial prediction result and the true value, and determine a prediction model based on the mean square error; The fault prediction result generating module is used to generate a fault prediction result for the target server based on the real-time data through the prediction model when the real-time data of the target server is acquired.

9. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; The memory is used to store computer programs; The processor is used to implement the method according to any one of claims 1 to 7 when executing the program stored in the memory.

10. A computer-readable storage medium having instructions stored thereon, which, when executed by one or more processors, cause the processors to perform the method according to any one of claims 1 to 7.

Images