A cloud computing node fault prediction method based on improved graph neural network

Abstract

The present invention discloses a cloud computing node fault prediction method based on an improved graph neural network, S1, collecting operation data from each node in a cloud computing environment; S2, forming a training data set; S3, constructing a relationship graph of cloud computing nodes based on the preprocessed training data set; S4, using the relationship graph of cloud computing nodes to construct and extract features of an improved graph neural network model; S5, using historical fault data and real-time monitoring data to train the improved graph neural network model; S6, deploying the trained improved graph neural network model to a cloud computing environment; S7, generating fault warning information when a potential node fault is detected; S8, based on the fault warning information; S9, recording and analyzing the data and results in the fault prediction and warning process. The present invention proposes an efficient and accurate cloud computing node fault prediction method, which improves the stability and service reliability of the cloud computing system.

Description

Technical Field

[0001] The present invention relates to the field of cloud computing technology, and in particular to a cloud computing node fault prediction method based on an improved graph neural network. Background Art

[0002] With the rapid development of cloud computing technology, cloud computing has become an important part of modern information technology. Cloud computing integrates a large amount of computing resources through virtualization technology and provides them to users through the network. However, the problem of computing node failure in cloud computing environment still exists. Node failure may cause the interruption of the entire cloud service, seriously affecting the stability of the system and user experience. Therefore, how to effectively predict and prevent cloud computing node failure has become an important issue in the operation and maintenance of cloud computing systems.

[0003] Traditional cloud computing node fault prediction methods are mainly based on statistical analysis and machine learning techniques, such as decision trees, support vector machines, and random forests. However, these methods have certain limitations in dealing with complex relationships and dependencies between cloud computing nodes. The relationships between nodes in a cloud computing environment are complex and dynamically changing. Node failures are not only affected by their own status, but are also closely related to the status and performance of other nodes. Traditional methods are difficult to accurately capture these complex dependencies, resulting in insufficient accuracy in fault prediction and a high false alarm rate.

[0004] Existing fault prediction methods still have the problem of insufficient real-time performance. Many traditional methods can only perform offline analysis and cannot meet the requirements of real-time prediction and early warning. This means that when a fault occurs, the system cannot provide timely warning and response, thus affecting the stability of cloud computing services and user experience. Real-time monitoring and prediction of cloud computing node failures is crucial to reducing system downtime and improving service reliability.

[0005] In recent years, with the development of deep learning technology, graph neural networks, as a deep learning model that can process graph structure data, have gradually attracted attention. Graph neural networks are suitable for processing data types such as network structures, social networks, and molecular graphs, and can effectively capture complex relationships and dependencies between nodes. Applying graph neural networks to cloud computing node fault prediction, using their powerful feature extraction and relationship modeling capabilities, is expected to significantly improve the accuracy and efficiency of fault prediction. However, existing graph neural network models still have some shortcomings when applied to cloud computing environments.

[0006] The existing graph neural network models have high computational complexity when processing large-scale cloud computing node data, making them difficult to apply in real time to actual systems. Secondly, the existing graph neural network models fail to fully consider the dynamic changes of node status and performance indicators during feature extraction and relationship modeling, resulting in inaccurate fault identification and prediction. Finally, the existing graph neural network models lack an effective real-time update mechanism during fault prediction, making it difficult to adapt to the real-time changes of node status and performance indicators in cloud computing environments. Summary of the invention

[0007] One object of the present invention is to propose a cloud computing node fault prediction method based on an improved graph neural network. The present invention proposes an efficient and accurate cloud computing node fault prediction method, which improves the stability and service reliability of the cloud computing system.

[0008] A cloud computing node fault prediction method based on an improved graph neural network according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect operation data from each node in the cloud computing environment, including the connection relationship between nodes, node status, CPU utilization, memory usage, network throughput, and disk read and write rates;

[0010] S2. Preprocess the collected operation data, including data cleaning to remove noise and outliers, data standardization to ensure the consistency of data dimensions of different indicators, and data completion to fill in missing data to form a training data set;

[0011] S3, constructing a relationship graph of cloud computing nodes based on the preprocessed training data set, where the nodes represent the various computing nodes in the cloud computing, the edges represent the connection relationship between the nodes, and assigning corresponding feature vectors to the nodes and edges in the graph according to the states and performance indicators of the nodes;

[0012] S4. Use the relationship graph of cloud computing nodes to improve the construction and feature extraction of graph neural network models;

[0013] S5. Use historical fault data and real-time monitoring data to train the improved graph neural network model, wherein the training process includes batch processing of training data, parameter optimization of the improved graph neural network model, and multiple iterations;

[0014] S6. Deploy the trained improved graph neural network model to the cloud computing environment, process the real-time monitoring data online, update the feature vectors of nodes and edges in real time, and perform fault prediction through the improved graph neural network model.

[0015] S7. When a potential node failure is detected, fault warning information is generated, including the node identification, fault type and possibility of the failure, and the operation and maintenance personnel are notified in real time through the early warning system;

[0016] S8. Based on fault warning information, provide specific maintenance suggestions, including node restart, resource reallocation, and load balancing adjustment, to help operation and maintenance personnel take timely measures to prevent the occurrence or spread of faults;

[0017] S9. Record and analyze the data and results in the fault prediction and early warning process, and continuously improve the graph neural network model.

[0018] Optionally, the S1 specifically includes:

[0019] S11. Collecting operation data from each node in the cloud computing environment, including the connection relationship between nodes, node status, CPU utilization, memory usage, network throughput, and disk read and write rates;

[0020] S12. The collected operation data is represented as a data set, where each sample in the data set includes the following elements:

[0021] Connection relationship between nodes: use adjacency matrix A, where A ij =1 means there is a connection between node i and node j, A ij =0 means there is no connection;

[0022] Node state: Use state vector S, where S i Indicates the state of node i;

[0023] CPU utilization: Use the CPU utilization vector C, where C i Indicates the CPU utilization of node i;

[0024] Memory usage: Use the memory usage vector M, where M i Indicates the memory usage of node i;

[0025] Network throughput: Use the network throughput vector N, where N i represents the network throughput of node i;

[0026] Disk read / write rate: Use the disk read / write rate vector D, where D i Indicates the disk read and write rate of node i;

[0027] S13. Build and run data set D set , run the dataset D set Each sample in is composed of the above elements, expressed as:

[0028] Dset ={(A,S,C,M,T,D)}.

[0029] Optionally, the S2 specifically includes:

[0030] S21, collect the running data set D set Perform data cleaning and use the outlier detection algorithm to identify and remove outliers to obtain the running data set D clean For the noise data, the filtering algorithm is applied to smooth it and form the cleaned running data set d clean ;

[0031] S22, the cleaned running data set D clean Perform data standardization:

[0032] Calculate the mean and standard deviation of each indicator, standardize each data sample x, and obtain the standardized running data set D norm ;

[0033] S23, the standardized running data set D norm Perform data completion. For samples with missing values, use linear interpolation to complete them and obtain the completed running data set D complete ;

[0034] S24, forming a training data set D train , training data set D train Contains cleaned, standardized and completed data:

[0035] D train =D complete .

[0036] Optionally, the S3 specifically includes:

[0037] S31, based on the preprocessed training data set D train Construct a relationship graph of cloud computing nodes, where nodes represent individual computing nodes in cloud computing and edges represent connection relationships between nodes;

[0038] S32. Generate an adjacency matrix A using an adaptive graph construction algorithm, where A ij Dynamically adjust according to the correlation between node i and node j, the correlation is calculated by the similarity of node status and performance indicators:

[0039]

[0040] Among them, X ikrepresents the value of node i on feature k. The features include node status, CPU utilization, memory usage, network throughput, and disk read and write rate. σ is the normalization function, Z is the normalization constant, and σ k is the standard deviation of feature k;

[0041] S33. Assign corresponding feature vectors to the nodes and edges in the graph according to the node status and performance indicators:

[0042] Node feature vector F, where F i =[S i ,C i ,M i ,N i ,D i ],S i represents the state of node i, C i represents the CPU utilization of node i, M i represents the memory usage of node i, N i represents the network throughput of node i, D i Indicates the disk read and write rate of node i;

[0043] The edge feature vector E is generated based on the correlation and interaction between nodes i and j:

[0044]

[0045] Among them, h is the activation function, W k and b k are weights and biases, X ik and X jk Respectively represent the values ​​of node i and node j on feature k;

[0046] S34. Dynamically update the relationship graph in combination with real-time monitoring data, ensure that the relationship graph G can reflect the dynamic changes of the cloud computing environment by adjusting the feature vectors of nodes and edges in real time, monitor the changes of node status and performance indicators in real time, update the node feature vector F and edge feature vector E, dynamically adjust the adjacency matrix A, and adjust the connection relationship between nodes to update according to the latest data:

[0047]

[0048] Among them, t represents the time step, η is the learning rate, L is the loss function, ΔX ik Represents the change of feature k from time t to t+1;

[0049] S35. Construct the final relationship graph G of cloud computing nodes, expressed as:

[0050] G = (A, F, E);

[0051] Among them, A is the adaptively generated adjacency matrix, F is the set of node feature vectors, and E is the set of edge feature vectors.

[0052] Optionally, the S4 specifically includes:

[0053] S41. Using the relationship graph G = (A, F, E) of cloud computing nodes to construct an improved graph neural network model, where A is an adaptively generated adjacency matrix, F is a node feature vector set, and E is an edge feature vector set;

[0054] S42. Build a model with a multi-level graph neural network structure:

[0055] Hierarchical graph convolutional layers aggregate node features at multiple levels, with each layer aggregating different features to capture the complex relationships between nodes:

[0056]

[0057] Among them, H (l) represents the node feature matrix of the lth layer, is the weight matrix of the lth layer, σ is the activation function, GConv represents the graph convolution operation, F k is the kth eigenvector, K is the number of features;

[0058] Multi-scale attention mechanism, introducing attention mechanisms of different time windows to capture the dynamic changes of node status and performance indicators:

[0059]

[0060] in, represents the attention weight between node i and node j in time window t, a is a trainable weight vector, and Respectively represent the feature vectors of node i and node j in time window t, represents the performance index similarity vector of node i and node j in time window t, || represents the vector concatenation operation;

[0061] S43, through the hierarchical graph convolution layer and multi-scale attention mechanism, the features of nodes and edges are extracted and aggregated to form a multi-level node feature representation H' and edge feature representation E';

[0062] S44. Input the extracted multi-level feature representation into the recurrent neural network layer to capture the fault features in the time series:

[0063] h t =RNN(x t ,h t-1 );

[0064] Among them, h t is the hidden state at time step t, x t is the input feature vector at time step t;

[0065] S45. Input the output feature representation of the recursive neural network layer into the multi-layer fully connected network to perform fault prediction:

[0066] y=σ(W out h T +b out );

[0067] Among them, y is the prediction result, W out is the weight matrix of the fully connected layer, b out is the bias vector, σ is the activation function, h T is the hidden state at the final time step T.

[0068] Optionally, the S5 specifically includes:

[0069] S51. Use historical fault data and real-time monitoring data to train the improved graph neural network model;

[0070] S52, batch processing is performed on the training data, and the historical fault data and the real-time monitoring data are divided into several batches, and each batch of data sets is represented as {(A (b) ,F (b) ,E (b) ,y (b) )}, where b represents the batch number, A (b) is the adjacency matrix, F (b) is the node feature matrix, E (b) is the edge feature matrix, y (b) is the corresponding fault label;

[0071] S53. Use batch gradient descent algorithm to optimize model parameters:

[0072] Forward propagation, calculating the output of each batch of data by improving the graph neural network model

[0073]

[0074] in, To predict the results, is the hidden state at the final time step T, W out is the weight matrix of the fully connected layer, b out is the bias vector;

[0075] Calculate the loss and use the loss function L to calculate the prediction result with the actual fault label The error between:

[0076]

[0077] Among them, L (b) is the loss value of the bth batch, N is the batch size, and are the actual value and the predicted value respectively;

[0078] Back propagation, calculating the gradient of the loss function with respect to the model parameters And update the model parameters according to the gradient:

[0079]

[0080] Among them, θ is the model parameter and η is the learning rate;

[0081] S54. Repeat the S51-S53 training process for all batches of data until the model converges to obtain an optimized improved graph neural network model.

[0082] Optionally, the S6 specifically includes:

[0083] S61. Deploy the trained improved graph neural network model to the cloud computing environment. The model is represented as GNN. model , and configure an online processing module for real-time data processing;

[0084] S62, collect data such as node status, CPU utilization, memory usage, network throughput, and disk read and write rate in the cloud computing environment in real time, and build a real-time data set D real ;

[0085] According to the real-time data set D real Dynamically update node feature vector F real and edge eigenvector E real , where the node feature vector F real It is expressed as:

[0086] F real =[S real ,C real ,M real ,N real ,D real ];

[0087] Among them, S real Indicates the real-time status of the node, C real Indicates the real-time CPU utilization of the node, M real Indicates the real-time memory usage of the node, N real represents the real-time network throughput of the node, D real Indicates the real-time disk read and write rate of the node;

[0088] Edge eigenvector E real Calculated based on real-time dependencies between nodes:

[0089]

[0090] in, represents the edge feature vector between node i and node j, h is the activation function, W k is the weight of feature k, and Respectively represent the values ​​of node i and node j on feature k, b k is the bias, m is the number of features;

[0091] S63, update the node feature vector F real and edge eigenvector E real Input to GnN model Calculate the failure probability P of each node fault :

[0092] P fault =σ(W out H' real +b out );

[0093] Among them, P fault represents the probability of node failure, H' real is the real-time node feature representation extracted by the graph convolution layer and the attention mechanism, H' real is the weight matrix of the fully connected layer, b out is the bias vector;

[0094] S64. Generate fault warning information based on the prediction results, including the node identification, fault type and possibility of fault occurrence of the node that may fail, and notify the operation and maintenance personnel in real time through the early warning system.

[0095] The beneficial effects of the present invention are:

[0096] (1) The present invention effectively captures the complex dependencies and dynamic changes between cloud computing nodes by constructing an improved graph neural network model including a multi-level graph convolution layer and a multi-scale attention mechanism. The multi-level graph convolution layer can aggregate node features at multiple levels, with each layer aggregating different features, and can fully capture the complex relationships between nodes; the multi-scale attention mechanism can dynamically capture the changes in node status and performance indicators by introducing an attention mechanism with different time windows, significantly improving the accuracy of fault prediction.

[0097] (2) The present invention designs a real-time update mechanism, which ensures that the relationship graph can reflect the dynamic changes of the cloud computing environment in real time by dynamically adjusting the feature vectors of nodes and edges, thereby achieving real-time prediction and early warning of node failures.

[0098] (3) The improved graph neural network model proposed in the present invention uses historical fault data and real-time monitoring data for model training, combines batch processing and batch gradient descent algorithm to optimize model parameters, and improves the robustness and predictive ability of the model.

[0099] (4) The present invention uses an adaptive graph construction algorithm to generate an adjacency matrix. The algorithm can adjust the connection relationship between nodes in real time according to changes in node status and performance indicators, ensuring that the relationship graph can accurately reflect the actual situation in the cloud computing environment, thereby improving the accuracy and real-time performance of fault prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0101] Figure 1 This is a flow chart of a cloud computing node fault prediction method based on an improved graph neural network proposed in the present invention. DETAILED DESCRIPTION

[0102] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0103] refer to Figure 1 , a cloud computing node fault prediction method based on improved graph neural network, comprising the following steps:

[0104] S1. Collect operation data from each node in the cloud computing environment, including the connection relationship between nodes, node status, CPU utilization, memory usage, network throughput, and disk read and write rates;

[0105] S2. Preprocess the collected operation data, including data cleaning to remove noise and outliers, data standardization to ensure the consistency of data dimensions of different indicators, and data completion to fill in missing data to form a training data set;

[0106] S3, constructing a relationship graph of cloud computing nodes based on the preprocessed training data set, where the nodes represent the various computing nodes in the cloud computing, the edges represent the connection relationship between the nodes, and assigning corresponding feature vectors to the nodes and edges in the graph according to the states and performance indicators of the nodes;

[0107] S4. Use the relationship graph of cloud computing nodes to improve the construction and feature extraction of graph neural network models;

[0108] S5. Use historical fault data and real-time monitoring data to train the improved graph neural network model, wherein the training process includes batch processing of training data, parameter optimization of the improved graph neural network model, and multiple iterations;

[0109] S6. Deploy the trained improved graph neural network model to the cloud computing environment, process the real-time monitoring data online, update the feature vectors of nodes and edges in real time, and perform fault prediction through the improved graph neural network model.

[0110] S7. When a potential node failure is detected, fault warning information is generated, including the node identification, fault type and possibility of the failure, and the operation and maintenance personnel are notified in real time through the early warning system;

[0111] S8. Based on fault warning information, provide specific maintenance suggestions, including node restart, resource reallocation, and load balancing adjustment, to help operation and maintenance personnel take timely measures to prevent the occurrence or spread of faults;

[0112] S9. Record and analyze the data and results in the fault prediction and early warning process, and continuously improve the graph neural network model.

[0113] In this implementation, S1 specifically includes:

[0114] S11. Collecting operation data from each node in the cloud computing environment, including the connection relationship between nodes, node status, CPU utilization, memory usage, network throughput, and disk read and write rates;

[0115] S12. The collected operation data is represented as a data set, where each sample in the data set includes the following elements:

[0116] Connection relationship between nodes: use adjacency matrix A, where A ij =1 means there is a connection between node i and node j, A ij =0 means there is no connection;

[0117] Node state: Use state vector S, where S i Indicates the state of node i;

[0118] CPU utilization: Use the CPU utilization vector C, where C i Indicates the CPU utilization of node i;

[0119] Memory usage: Use the memory usage vector M, where M i Indicates the memory usage of node i;

[0120] Network throughput: Use the network throughput vector N, where N i represents the network throughput of node i;

[0121] Disk read / write rate: Use the disk read / write rate vector D, where D i Indicates the disk read and write rate of node i;

[0122] S13. Build and run data set D set , run the dataset D set Each sample in is composed of the above elements, expressed as:

[0123] D set ={(A,S,C,M,T,D)}.

[0124] In this implementation, S2 specifically includes:

[0125] S21, collect the running data set D set Perform data cleaning and use the outlier detection algorithm to identify and remove outliers to obtain the running data set D clean For the noise data, the filtering algorithm is applied to smooth it and form the cleaned running data set D clean ;

[0126] S22, the cleaned running data set D clean Perform data standardization:

[0127] Calculate the mean and standard deviation of each indicator, standardize each data sample x, and obtain the standardized running data set D norm ;

[0128] S23, the standardized running data set D norm Perform data completion. For samples with missing values, use linear interpolation to complete them and obtain the completed running data set D complete ;

[0129] S24, forming a training data set D train , training data set D train Contains cleaned, standardized and completed data:

[0130] D train =D complete .

[0131] In this implementation, S3 specifically includes:

[0132] S31, based on the preprocessed training data set D trainConstruct a relationship graph of cloud computing nodes, where nodes represent individual computing nodes in cloud computing and edges represent connection relationships between nodes;

[0133] S32. Generate an adjacency matrix A using an adaptive graph construction algorithm, where A ij Dynamically adjust according to the correlation between node i and node j, the correlation is calculated by the similarity of node status and performance indicators:

[0134]

[0135] Among them, X ik represents the value of node i on feature k. The features include node status, CPU utilization, memory usage, network throughput, and disk read and write rate. σ is the normalization function, Z is the normalization constant, and σ k is the standard deviation of feature k;

[0136] S33. Assign corresponding feature vectors to the nodes and edges in the graph according to the node status and performance indicators:

[0137] Node feature vector F, where F i =[S i ,C i ,M i ,N i ,D i ],S i represents the state of node i, C i represents the CPU utilization of node i, M i Indicates the memory usage of node i, M i represents the network throughput of node i, D i Indicates the disk read and write rate of node i;

[0138] The edge feature vector E is generated based on the correlation and interaction between nodes i and j:

[0139]

[0140] Among them, h is the activation function, W k and b k are weights and biases, X ik and X jk Respectively represent the values ​​of node i and node j on feature k;

[0141] S34. Dynamically update the relationship graph in combination with real-time monitoring data, ensure that the relationship graph G can reflect the dynamic changes of the cloud computing environment by adjusting the feature vectors of nodes and edges in real time, monitor the changes of node status and performance indicators in real time, update the node feature vector F and edge feature vector E, dynamically adjust the adjacency matrix A, and adjust the connection relationship between nodes to update according to the latest data:

[0142]

[0143] Among them, t represents the time step, η is the learning rate, L is the loss function, ΔX ik Represents the change of feature k from time t to t+1;

[0144] S35. Construct the final relationship graph G of cloud computing nodes, expressed as:

[0145] G = (A, F, E);

[0146] Among them, A is the adaptively generated adjacency matrix, F is the set of node feature vectors, and E is the set of edge feature vectors.

[0147] In this implementation, S4 specifically includes:

[0148] S41. Using the relationship graph G = (A, F, E) of cloud computing nodes to construct an improved graph neural network model, where A is an adaptively generated adjacency matrix, F is a node feature vector set, and E is an edge feature vector set;

[0149] S42. Build a model with a multi-level graph neural network structure:

[0150] Hierarchical graph convolutional layers aggregate node features at multiple levels, with each layer aggregating different features to capture the complex relationships between nodes:

[0151]

[0152] Among them, H (l) represents the node feature matrix of the lth layer, is the weight matrix of the lth layer, σ is the activation function, GConv represents the graph convolution operation, F k is the kth eigenvector, K is the number of features;

[0153] Multi-scale attention mechanism, introducing attention mechanisms of different time windows to capture the dynamic changes of node status and performance indicators:

[0154]

[0155] in, represents the attention weight between node i and node j in time window t, a is a trainable weight vector, and Respectively represent the feature vectors of node i and node j in time window t, represents the performance index similarity vector of node i and node j in time window t, || represents the vector concatenation operation;

[0156] S43, through the hierarchical graph convolution layer and multi-scale attention mechanism, the features of nodes and edges are extracted and aggregated to form a multi-level node feature representation H' and edge feature representation E';

[0157] S44. Input the extracted multi-level feature representation into the recurrent neural network layer to capture the fault features in the time series:

[0158] h t =RNN(x t ,h t-1 );

[0159] Among them, h t is the hidden state at time step t, x t is the input feature vector at time step t;

[0160] S45. Input the output feature representation of the recursive neural network layer into the multi-layer fully connected network to perform fault prediction:

[0161] y=σ(W out h T +b out );

[0162] Among them, y is the prediction result, W out is the weight matrix of the fully connected layer, b out is the bias vector, σ is the activation function, h T is the hidden state at the final time step T.

[0163] In this implementation, S5 specifically includes:

[0164] S51. Use historical fault data and real-time monitoring data to train the improved graph neural network model;

[0165] S52, batch processing is performed on the training data, and the historical fault data and the real-time monitoring data are divided into several batches, and each batch of data sets is represented as {(A (b) ,F (b) ,E (b) ,y (b) )}, where b represents the batch number, A (b) is the adjacency matrix, F (b) is the node feature matrix, E (b)is the edge feature matrix, y (b) is the corresponding fault label;

[0166] S53. Use batch gradient descent algorithm to optimize model parameters:

[0167] Forward propagation, calculating the output of each batch of data by improving the graph neural network model

[0168]

[0169] in, To predict the results, is the hidden state at the final time step T, W out is the weight matrix of the fully connected layer, b out is the bias vector;

[0170] Calculate the loss and use the loss function L to calculate the prediction result with the actual fault label The error between:

[0171]

[0172] Among them, L (b) is the loss value of the bth batch, N is the batch size, and are the actual value and the predicted value respectively;

[0173] Back propagation, calculating the gradient of the loss function with respect to the model parameters And update the model parameters according to the gradient:

[0174]

[0175] Among them, θ is the model parameter and η is the learning rate;

[0176] S54. Repeat the S51-S53 training process for all batches of data until the model converges to obtain an optimized improved graph neural network model.

[0177] In this implementation, S6 specifically includes:

[0178] S61. Deploy the trained improved graph neural network model to the cloud computing environment. The model is represented as GNN. model , and configure an online processing module for real-time data processing;

[0179] S62, collect data such as node status, CPU utilization, memory usage, network throughput, and disk read and write rate in the cloud computing environment in real time, and build a real-time data set D real ;

[0180] According to the real-time data set D real Dynamically update node feature vector F real and edge eigenvector E real , where the node feature vector F real It is expressed as:

[0181] F real =[S real ,C real ,M real ,N real ,D real ];

[0182] Among them, S real Indicates the real-time status of the node, C real Indicates the real-time CPU utilization of the node, M real Indicates the real-time memory usage of the node, N real represents the real-time network throughput of the node, D real Indicates the real-time disk read and write rate of the node;

[0183] Edge eigenvector E real Calculated based on real-time dependencies between nodes:

[0184]

[0185] in, represents the edge feature vector between node i and node j, h is the activation function, W k is the weight of feature k, and Respectively represent the values ​​of node i and node j on feature k, b k is the bias, m is the number of features;

[0186] S63, update the node feature vector F real and edge eigenvector E real Input to GNN model Calculate the failure probability P of each node fault :

[0187] P fault =σ(W out H' real +b out );

[0188] Among them, P fault represents the probability of node failure, H' real is the real-time node feature representation extracted by the graph convolution layer and the attention mechanism, H' real is the weight matrix of the fully connected layer, b out is the bias vector;

[0189] S64. Generate fault warning information based on the prediction results, including the node identification, fault type and possibility of fault occurrence of the node that may fail, and notify the operation and maintenance personnel in real time through the early warning system.

[0190] Embodiment 1:

[0191] A large cloud computing service provider has deployed thousands of computing nodes in its data center, which jointly undertake a large number of computing tasks. In order to ensure service continuity and high availability, the service provider needs to monitor and predict the failure of computing nodes in real time so that preventive maintenance can be performed before failure occurs. However, traditional fault prediction methods perform poorly in dealing with such a large and complex cloud computing environment, with low prediction accuracy and inability to achieve real-time warning.

[0192] In order to verify the effectiveness of the present invention, we selected the data center of the service provider as the application scenario. During the operation of the data center, we first collected operation data from each computing node, including the connection relationship between nodes, node status, CPU utilization, memory usage, network throughput, and disk read and write rate. Then, we preprocessed these data and constructed a relationship graph of cloud computing nodes.

[0193] In the preprocessing stage, we used data cleaning techniques to remove noise and outliers, performed data standardization, and filled missing data through interpolation. The processed data was used to build an improved graph neural network model with multi-level graph convolutional layers and multi-scale attention mechanisms. The training dataset contains data from the past 6 months, including several known node failure events. The model was optimized through batch processing and batch gradient descent algorithms. After multiple iterations of training, we deployed the trained model to the real-time monitoring system of the data center.

[0194] In actual applications, the real-time monitoring system continuously collects the operating data of each node and dynamically updates the feature vectors of nodes and edges. The improved graph neural network model uses these real-time data to predict the failure probability of each node. When the model detects that the failure probability of a node exceeds the preset threshold, the system immediately generates fault warning information and notifies the operation and maintenance personnel.

[0195] In order to evaluate the effectiveness of the present invention, we compared the performance of the present invention and traditional fault prediction methods in the same data center. The comparison data includes key indicators such as prediction accuracy, false alarm rate, and early warning response time. In the first month of operation of the data center, we recorded the following data:

[0196] Table 1 Comparison of three-month data between the present invention and the traditional method

[0197] month method Prediction accuracy False Positive Rate Warning response time First Month Traditional methods 78％ 15％ 20 minutes First Month Method of the present invention 92％ 5％ 5 minutes Second month Traditional methods 80％ 14％ 18 minutes Second month Method of the present invention 93％ 4％ 4 minutes Third month Traditional methods 79％ 13％ 19 minutes Third month Method of the present invention 94％ 3％ 3 minutes

[0198] It can be seen from the data in Table 1 above that the method of the present invention is significantly superior to traditional methods in terms of fault prediction accuracy, false alarm rate and warning response time. Specifically, the method of the present invention can more accurately capture the complex dependencies and dynamic changes between cloud computing nodes by improving the graph neural network model, thereby significantly improving the accuracy of fault prediction. At the same time, real-time data processing and dynamic update mechanisms enable the system to generate warning information in time before node failure occurs, effectively reducing system downtime and improving service reliability.

[0199] In addition, the method of the present invention also shows strong robustness and adaptability. Regardless of different operating environments or different fault types, the improved graph neural network model shows stable prediction performance. This advantage makes the method of the present invention have higher practical value and promotion potential in practical applications.

[0200] In summary, through practical application and testing in large-scale cloud computing data centers, the method of the present invention not only solves the shortcomings of traditional fault prediction methods in dealing with complex relationships and dependencies between cloud computing nodes, but also realizes real-time prediction and early warning of node failures, effectively improving the stability of the system and the reliability of services. These data and results fully demonstrate the effectiveness and innovation of the present invention.

[0201] The present invention effectively captures the complex dependencies and dynamic changes between cloud computing nodes by constructing an improved graph neural network model including a multi-level graph convolution layer and a multi-scale attention mechanism. The multi-level graph convolution layer can aggregate node features at multiple levels, with each layer aggregating different features, and can fully capture the complex relationships between nodes; the multi-scale attention mechanism can dynamically capture changes in node status and performance indicators by introducing attention mechanisms of different time windows, significantly improving the accuracy of fault prediction.

[0202] The present invention designs a real-time update mechanism, which ensures that the relationship graph can reflect the dynamic changes of the cloud computing environment in real time by dynamically adjusting the characteristic vectors of nodes and edges, thereby realizing real-time prediction and early warning of node failures.

[0203] The improved graph neural network model proposed in the present invention uses historical fault data and real-time monitoring data for model training, combines batch processing and batch gradient descent algorithm to optimize model parameters, and improves the robustness and predictive ability of the model.

[0204] The present invention adopts an adaptive graph construction algorithm to generate an adjacency matrix. The algorithm can adjust the connection relationship between nodes in real time according to changes in node status and performance indicators, ensuring that the relationship graph can accurately reflect the actual situation in the cloud computing environment, thereby improving the accuracy and real-time performance of fault prediction.

[0205] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A cloud computing node fault prediction method based on improved graph neural network, characterized in that: The steps include: S1. Collect operation data from each node in the cloud computing environment, including the connection relationship between nodes, node status, CPU utilization, memory usage, network throughput, and disk read and write rates; S2. Preprocess the collected operation data, including data cleaning to remove noise and outliers, data standardization to ensure the consistency of data dimensions of different indicators, and data completion to fill in missing data to form a training data set; S3, constructing a relationship graph of cloud computing nodes based on the preprocessed training data set, where the nodes represent the various computing nodes in the cloud computing, the edges represent the connection relationship between the nodes, and assigning corresponding feature vectors to the nodes and edges in the graph according to the states and performance indicators of the nodes; S4. Use the relationship graph of cloud computing nodes to improve the construction and feature extraction of graph neural network models; S5. Use historical fault data and real-time monitoring data to train the improved graph neural network model, wherein the training process includes batch processing of training data, parameter optimization of the improved graph neural network model, and multiple iterations; S6. Deploy the trained improved graph neural network model to the cloud computing environment, process the real-time monitoring data online, update the feature vectors of nodes and edges in real time, and perform fault prediction through the improved graph neural network model. S7. When a potential node failure is detected, fault warning information is generated, including the node identification, fault type and possibility of the failure, and the operation and maintenance personnel are notified in real time through the early warning system; S8. Based on fault warning information, provide specific maintenance suggestions, including node restart, resource reallocation, and load balancing adjustment, to help operation and maintenance personnel take timely measures to prevent the occurrence or spread of faults; S9. Record and analyze the data and results in the fault prediction and early warning process, and continuously improve the graph neural network model; The S4 specifically includes: S41. Using the relationship graph G = (A, F, E) of cloud computing nodes to construct an improved graph neural network model, where A is an adaptively generated adjacency matrix, F is a node feature vector set, and E is an edge feature vector set; S42. Build a model with a multi-level graph neural network structure: Hierarchical graph convolutional layers aggregate node features at multiple levels, with each layer aggregating different features to capture the complex relationships between nodes: Among them, H (l) represents the node feature matrix of the lth layer, is the weight matrix of the lth layer, σ is the activation function, GConv represents the graph convolution operation, F k is the kth eigenvector, K is the number of features; Multi-scale attention mechanism, introducing attention mechanisms of different time windows to capture the dynamic changes of node status and performance indicators: in, represents the attention weight between node i and node j in time window t, a is a trainable weight vector, and Respectively represent the feature vectors of node i and node j in time window t, represents the performance index similarity vector of node i and node j in time window t, || represents the vector concatenation operation; S43, through the hierarchical graph convolution layer and multi-scale attention mechanism, the features of nodes and edges are extracted and aggregated to form a multi-level node feature representation H' and edge feature representation E'; S44. Input the extracted multi-level feature representation into the recurrent neural network layer to capture the fault features in the time series: h t =RNN(x t ,h t-1 ); Among them, h t is the hidden state at time step t, x t is the input feature vector at time step t; S45. Input the output feature representation of the recursive neural network layer into the multi-layer fully connected network to perform fault prediction: y=σ(W out h T +b out ); Among them, y is the prediction result, W out is the weight matrix of the fully connected layer, b out is the bias vector, σ is the activation function, h T is the hidden state at the final time step T; The S5 specifically includes: S51. Use historical fault data and real-time monitoring data to train the improved graph neural network model; S52, batch processing is performed on the training data, and the historical fault data and the real-time monitoring data are divided into several batches, and each batch of data sets is represented as {(A (b) ,F (b) ,E (b) ,y (b) )}, where b represents the batch number, A (b) is the adjacency matrix, F (b) is the node feature matrix, E (b) is the edge feature matrix, y (b) is the corresponding fault label; S53. Use batch gradient descent algorithm to optimize model parameters: Forward propagation, calculating the output of each batch of data by improving the graph neural network model in, To predict the results, is the hidden state at the final time step T, W out is the weight matrix of the fully connected layer, b out is the bias vector; Calculate the loss and use the loss function L to calculate the prediction result with the actual fault label The error between: Among them, L (b) is the loss value of the bth batch, N is the batch size, and are the actual value and the predicted value respectively; Back propagation, calculating the gradient of the loss function with respect to the model parameters And update the model parameters according to the gradient: Among them, θ is the model parameter and η is the learning rate; S54, repeat the training process of S51-S53 for all batches of data until the model converges, and obtain the optimized improved graph neural network model; The S6 specifically includes: S61. Deploy the trained improved graph neural network model to the cloud computing environment. The model is represented as GNN. model , and configure an online processing module for real-time data processing; S62, collect node status, CPU utilization, memory usage, network throughput and disk read and write rate data in the cloud computing environment in real time, and build a real-time data set D real ; According to the real-time data set D real Dynamically update node feature vector F real and edge eigenvector E real , where the node feature vector F real It is expressed as: F real =[S real ,C real ,M real ,N real ,D real ]; Among them, S real Indicates the real-time status of the node, C real Indicates the real-time CPU utilization of the node, M real Indicates the real-time memory usage of the node, N real represents the real-time network throughput of the node, D real Indicates the real-time disk read and write rate of the node; Edge eigenvector E real Calculated based on real-time dependencies between nodes: in, represents the edge feature vector between node i and node j, h is the activation function, W k is the weight of feature k, and Respectively represent the values ​​of node i and node j on feature k, b k is the bias, m is the number of features; S63, update the node feature vector F real and edge eigenvector E real Input to GNN model Calculate the failure probability P of each node fault : P fault =σ(W out H′ real +b out ); Among them, P fault represents the probability of node failure, H′ real is the real-time node feature representation extracted by the graph convolution layer and the attention mechanism, W out is the weight matrix of the fully connected layer, b out is the bias vector; S64. Generate fault warning information based on the prediction results, including the node identification, fault type and possibility of fault occurrence of the node that may fail, and notify the operation and maintenance personnel in real time through the early warning system.

2. According to claim 1, a cloud computing node fault prediction method based on improved graph neural network is characterized in that: The S1 specifically includes: S11. Collecting operation data from each node in the cloud computing environment, including the connection relationship between nodes, node status, CPU utilization, memory usage, network throughput, and disk read and write rates; S12. The collected operation data is represented as a data set, where each sample in the data set includes the following elements: Connection relationship between nodes: use adjacency matrix A, where A ij =1 indicates that there is a connection between node i and node j, A ij =0 means there is no connection; Node state: Use state vector S, where S i Indicates the state of node i; CPU utilization: Use the CPU utilization vector C, where C i Indicates the CPU utilization of node i; Memory usage: Use the memory usage vector M, where M i Indicates the memory usage of node i; Network throughput: Use the network throughput vector N, where N i represents the network throughput of node i; Disk read / write rate: Use the disk read / write rate vector D, where D i Indicates the disk read and write rate of node i; S13. Build and run data set D set , run the dataset D set Each sample in is composed of the above elements, expressed as: D set ={(A,S,C,M,N,D)}。 3. The cloud computing node fault prediction method based on improved graph neural network according to claim 2 is characterized in that: The S2 specifically includes: S21, collect the running data set D set Perform data cleaning and use the outlier detection algorithm to identify and remove outliers to obtain the running data set D clean For the noise data, the filtering algorithm is applied to smooth it and form the cleaned running data set D clean ; S22, the cleaned running data set D clean Perform data standardization: Calculate the mean and standard deviation of each indicator, standardize each data sample x, and obtain the standardized running data set D norm ; S23, the standardized running data set D norm Perform data completion. For samples with missing values, use linear interpolation to complete them and obtain the completed running data set D complete ; S24, forming a training data set D train , training data set D train Contains cleaned, standardized and completed data: D train =D complete 。 4. The cloud computing node fault prediction method based on improved graph neural network according to claim 3 is characterized in that: The S3 specifically includes: S31, based on the preprocessed training data set D train Construct a relationship graph of cloud computing nodes, where nodes represent individual computing nodes in cloud computing and edges represent connection relationships between nodes; S32. Generate an adjacency matrix A using an adaptive graph construction algorithm, where A ij Dynamically adjust according to the correlation between node i and node j, the correlation is calculated by the similarity of node status and performance indicators: Among them, X ik represents the value of node i on feature k. The features include node status, CPU utilization, memory usage, network throughput, and disk read and write rate. σ is the normalization function, Z is the normalization constant, and σ k is the standard deviation of feature k; S33. Assign corresponding feature vectors to the nodes and edges in the graph according to the node status and performance indicators: Node feature vector F, where F i =[S i ,C i ,M i ,N i ,D i ],S i represents the state of node i, C i represents the CPU utilization of node i, M i represents the memory usage of node i, N i represents the network throughput of node i, D i Indicates the disk read and write rate of node i; The edge feature vector E is generated based on the correlation and interaction between nodes i and j: Among them, h is the activation function, W k and b k are weights and biases, X ik and X jk Respectively represent the values ​​of node i and node j on feature k; S34. Dynamically update the relationship graph in combination with real-time monitoring data, ensure that the relationship graph G can reflect the dynamic changes of the cloud computing environment by adjusting the feature vectors of nodes and edges in real time, monitor the changes of node status and performance indicators in real time, update the node feature vector F and edge feature vector E, dynamically adjust the adjacency matrix A, and adjust the connection relationship between nodes to update according to the latest data: Among them, t represents the time step, η is the learning rate, L is the loss function, ΔX ik Represents the change of feature k from time t to t+1; S35. Construct the final relationship graph G of cloud computing nodes, expressed as: G = (A, F, E); Among them, A is the adaptively generated adjacency matrix, F is the set of node feature vectors, and E is the set of edge feature vectors.

Images