An intelligent disaster recovery management service system

Abstract

The application relates to the technical field of disaster recovery management, in particular to an intelligent disaster recovery management service system, which collects resource running state data of each regional node of a cloud platform in real time in a distributed manner through a data acquisition module, including network connection and storage state. The data is cleaned, normalized and abnormally detected through a preprocessing module to generate resource availability evaluation data, providing accurate disaster recovery management basis. Based on the resource availability evaluation data, a disaster recovery process arrangement module dynamically adjusts the key node configuration in the switching scheme by using a Transformer model and an attention mechanism, automatically arranges the disaster recovery switching process and generates multiple switching schemes. A decision control module makes the optimal disaster recovery decision by analyzing the generated scheme and combining a graph neural network, ensuring that the system can quickly recover when the cloud platform fails, realizing real-time response to faults and resource changes in the cloud platform, automatic generation and adjustment of the disaster recovery switching scheme, and improvement of disaster recovery efficiency and system availability.

Description

Technical Field

[0001] This invention relates to the field of disaster recovery management technology, and more particularly to an intelligent disaster recovery management service system. Background Technology

[0002] The stability and availability of cloud platforms are crucial for the continuous operation of enterprise businesses. However, due to their distributed architecture, dynamic resource scheduling, and large-scale data centers, cloud platforms are susceptible to various internal and external factors, leading to problems such as network failures, storage failures, and system outages. These failures can severely impact the normal operation of enterprises.

[0003] Traditional disaster recovery management methods typically rely on static, pre-defined disaster recovery processes, which are insufficient to cope with the complex and ever-changing failure scenarios in modern cloud platforms. With the continuous development of disaster recovery technology, data-driven approaches are gradually becoming a trend. In particular, real-time collection, analysis, and processing of resource operational status data from nodes within the cloud platform enables more accurate and efficient disaster recovery management. While existing disaster recovery management systems have achieved a degree of automation, most still suffer from the following shortcomings: First, the switching of disaster recovery solutions lacks dynamic adjustment capabilities, making it difficult to adapt to complex and changing failure modes; second, the processing methods for resource operational status data are relatively simple, lacking in-depth analysis and prediction of anomalies; and third, the orchestration of disaster recovery processes is relatively fixed, failing to flexibly adjust recovery strategies based on real-time data, resulting in unsatisfactory response speed and recovery effectiveness. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides an intelligent disaster recovery management service system. This system trains a Transformer model on the resource operation status data of several nodes in different regions of a cloud platform to output a disaster recovery switching scheme, and then selects the optimal scheme through a graph neural network for disaster recovery management and control.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] An intelligent disaster recovery management service system includes: a data acquisition module, a preprocessing module, a disaster recovery process orchestration module, and a decision control module connected in sequence;

[0007] The data acquisition module is used to collect resource operation status data of several nodes in different regions of the cloud platform in a distributed manner. The resource operation status data includes network connection status data and storage status data.

[0008] The preprocessing module is used to perform data cleaning, normalization, and anomaly detection on the resource operation status data to generate resource availability assessment data.

[0009] The disaster recovery process orchestration module is used to train the Transformer model based on resource availability assessment data, dynamically adjust the key node configuration of the switching scheme through the attention mechanism, automatically orchestrate the disaster recovery switching process of nodes in different regions, and generate several switching schemes.

[0010] The decision control module is used to predict potential failure risks based on several switching schemes generated by the disaster recovery process orchestration module, generate feasibility scores, and select switching schemes for cloud platform control based on the feasibility scores.

[0011] Furthermore, the network connection status data includes network latency and bandwidth consumption, and the storage status data includes node CPU utilization, memory usage data, and storage usage data.

[0012] Furthermore, the preprocessing module is used to perform the following steps:

[0013] The resource operation status data is denoised using a median filtering algorithm and normalized using a linear normalization method.

[0014] The isolated forest algorithm is used to identify outliers in the cleaned and normalized data and mark potential outlier data points.

[0015] Principal component analysis is used to integrate resource operation status data from different regions and nodes to generate resource availability assessment data.

[0016] Furthermore, the disaster recovery switch includes switching computing nodes, adjusting computing resources, storage resources, and network resources.

[0017] Furthermore, training the Transformer model based on resource availability assessment data includes the following steps:

[0018] The resource availability assessment data is divided into fixed-length time series segments using the sliding window algorithm, and data augmentation is performed to generate training samples.

[0019] Based on the training samples, a mean squared error loss function is defined, and the Transformer model is trained using the backpropagation algorithm of the Adam optimizer.

[0020] Apply a multi-head self-attention mechanism to adjust the weight configuration of nodes.

[0021] Furthermore, the automated orchestration of the disaster recovery switching process for nodes in different regions includes the following steps:

[0022] Input real-time resource availability assessment data into the Transformer model;

[0023] The input data is used to extract features through a multi-head self-attention mechanism, identify the correlation between nodes and the importance of key nodes, and generate context-related feature representations.

[0024] Based on the extracted feature representations, several disaster recovery switching schemes are matched and output.

[0025] Furthermore, the formula for the Transformer model is as follows:

[0026]

[0027] Where Y is the output of the Transformer model; X = [L, B, U, M, S]; L is network latency; B is bandwidth consumption; U is CPU utilization; M is memory usage; S is storage usage; W Q To query the weight matrix; W K W is the key weight matrix; V Value weight matrix; d k is the dimension of the key vector; softmax is the normalization function.

[0028] Furthermore, the decision control module is used to perform the following steps:

[0029] The disaster recovery switching scheme generated by the disaster recovery process orchestration module is converted into a matrix representation, including the construction of a switching path matrix and a resource allocation matrix;

[0030] Using the switching path matrix and resource allocation matrix as input, a feasibility score is output through a pre-trained graph neural network model;

[0031] Disaster recovery switching schemes are ranked according to their feasibility scores, and the scheme with the highest feasibility score is selected as the implementation scheme.

[0032] The implementation plan will be applied to the cloud platform for control and scheduling.

[0033] Furthermore, the graph neural network model is trained through the following steps:

[0034] The switching path matrix and resource allocation matrix are used as input data for the graph structure. Each node in the graph structure is preset as a computing node in the cloud platform, and the edges are preset as the connection relationships between nodes. Each node and edge is assigned a corresponding feature vector, including network latency, bandwidth consumption, CPU utilization, memory usage, and storage usage.

[0035] Feasibility scores are assigned to disaster recovery switching schemes, cross-entropy loss function is selected, and graph neural network model is trained using backpropagation algorithm;

[0036] The performance of the graph neural network model is evaluated using cross-validation.

[0037] Furthermore, the formula for the graph neural network is as follows:

[0038]

[0039] Among them, H (l) H is the node feature matrix of the l-th layer; (l+1) H is the node feature matrix of the (l+1)th layer; (0) The initial layer includes the switching path matrix and resource allocation matrix; A is the normalized adjacency matrix; W (l) Let be the trainable weight matrix of the l-th layer; σ is the activation function.

[0040] The beneficial effects of this invention are as follows: This invention collects resource operation status data of nodes in various regions of the cloud platform in real time through a data acquisition module, including network connection status and storage status data. This data undergoes cleaning, normalization, and anomaly detection by a preprocessing module to generate resource availability assessment data, providing accurate foundational data for subsequent disaster recovery management. Using this assessment data, the disaster recovery process orchestration module, based on Transformer model training, dynamically adjusts the configuration of key nodes in the switching scheme through an attention mechanism, automatically orchestrating the disaster recovery switching process to generate multiple switching schemes. The decision control module analyzes the generated switching schemes and uses algorithms such as graph neural networks to make the optimal disaster recovery switching decision, ensuring that the system can quickly and accurately restore normal operation when a cloud platform failure occurs. Through this intelligent disaster recovery management approach, this solution can respond in real time to faults and resource status changes in the cloud platform, automatically generate and adjust disaster recovery switching schemes, achieving data-driven dynamic optimization and greatly improving the disaster recovery capability and system availability of the cloud platform. Compared with traditional methods, it achieves stronger adaptability and flexibility, can cope with more complex and changing fault scenarios, and provides more efficient and accurate disaster recovery management services. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the structure of an intelligent disaster recovery management service system according to the present invention.

[0042] Figure 2 This is a flowchart illustrating the steps involved in automating the disaster recovery switching process for nodes in different regions in this invention. Detailed Implementation

[0043] Please see Figure 1-2 As shown, the present invention relates to an intelligent disaster recovery management service system, comprising: a data acquisition module, a preprocessing module, a disaster recovery process orchestration module, and a decision control module connected in sequence;

[0044] The data acquisition module is used to collect resource operation status data of several nodes in different regions of the cloud platform in a distributed manner. The resource operation status data includes network connection status data and storage status data.

[0045] The preprocessing module is used to perform data cleaning, normalization, and anomaly detection on the resource operation status data to generate resource availability assessment data.

[0046] The disaster recovery process orchestration module is used to train the Transformer model based on resource availability assessment data, dynamically adjust the key node configuration of the switching scheme through the attention mechanism, automatically orchestrate the disaster recovery switching process of nodes in different regions, and generate several switching schemes.

[0047] The decision control module is used to predict potential failure risks based on several switching schemes generated by the disaster recovery process orchestration module, generate feasibility scores, and select switching schemes for cloud platform control based on the feasibility scores.

[0048] It's important to note that, firstly, the data acquisition module is responsible for collecting resource operational status data from various regional nodes within the cloud platform in real-time and in a distributed manner, including network connectivity and storage status data. This data originates from multiple physical or virtual nodes, which may be distributed across different geographical locations or data centers globally. Network connectivity data typically includes information such as network latency, bandwidth utilization, and packet loss rate between nodes; storage status data covers metrics such as disk space utilization, I / O load, and storage response time. This data comprehensively reflects the operational health of the cloud platform, especially network and storage resources, which are the most critical resources within the cloud platform. Subsequently, the preprocessing module cleans, normalizes, and detects anomalies in the collected resource operational status data. Data cleaning removes noise and redundant information to ensure the accuracy of subsequent analysis; normalization converts data of different dimensions into a unified standard scale for comprehensive evaluation; anomaly detection identifies abnormal fluctuations or potential failure points in the data, such as sudden network outages or insufficient storage capacity. After these processes, the resource availability assessment data generated by the system provides an accurate and reliable basis for subsequent disaster recovery decisions. After obtaining resource availability assessment data, the disaster recovery process orchestration module comes into play. This module trains a Transformer model based on the assessment data and dynamically adjusts the configuration of key nodes in the failover plan through an attention mechanism. As a deep learning model, the Transformer model, trained on historical disaster recovery data, can predict the resource availability trends of the cloud platform under different failure scenarios, thereby automating the orchestration of the disaster recovery failover process. The disaster recovery process orchestration module generates multiple failover plans based on the assessment data and prioritizes them. Each failover plan includes specific disaster recovery failover paths and steps, such as transferring services from a failed node to a standby node or performing load balancing adjustments across data centers. The generation of these failover plans fully considers the current system state, failure type, and the availability of disaster recovery resources. Finally, the decision control module, based on the multiple failover plans generated by the disaster recovery process orchestration module and combined with algorithms such as graph neural networks (GNNs), selects the optimal disaster recovery failover plan for implementation. Graph neural networks can calculate the most suitable resource failover path under a given failure condition by modeling the complex dependencies between nodes. This process not only considers resource availability but also takes into account factors such as the specific location of the faulty node, the fault level, and the load capacity of disaster recovery resources, ensuring that the system can resume normal operation in the shortest possible time under the influence of multiple factors. The decision control module also continuously monitors the status of the cloud platform and adjusts the disaster recovery strategy in real time to cope with possible new faults or resource changes, ensuring the long-term stable operation of the system.

[0049] Furthermore, the network connection status data includes network latency and bandwidth consumption, and the storage status data includes node CPU utilization, memory usage data, and storage usage data.

[0050] It's important to note that network connectivity status data includes network latency and bandwidth consumption. Network latency refers to the time delay in data transmission between nodes, typically measured in milliseconds (ms). High network latency can lead to poor communication between nodes, impacting application performance on the cloud platform, especially in real-time data processing scenarios requiring low latency. Furthermore, bandwidth consumption measures data traffic usage, usually expressed as the amount of data transmitted per unit of time (e.g., Mbps or Gbps). An abnormally high bandwidth consumption may indicate network overload, leading to bandwidth resource constraints and consequently affecting overall service availability and network response speed. Therefore, network latency and bandwidth consumption are crucial indicators for assessing the health of a cloud platform, helping to identify potential network bottlenecks and performance degradation risks in advance. Storage status data includes node CPU utilization, memory usage, and storage usage. CPU utilization reflects the load on the node's central processing unit; high CPU utilization usually indicates strained computing resources, potentially leading to decreased node processing power and impacting the overall computing efficiency of the cloud platform. Memory usage indicates the node's memory usage; excessive memory usage can cause memory overflow or frequent memory swapping, resulting in performance issues. Storage usage data focuses on the usage of node disk space, especially in cloud platform applications involving large-scale data storage, where insufficient disk space may lead to storage system crashes or data loss.

[0051] Furthermore, the preprocessing module is used to perform the following steps:

[0052] The resource operation status data is denoised using a median filtering algorithm and normalized using a linear normalization method.

[0053] The isolated forest algorithm is used to identify outliers in the cleaned and normalized data and mark potential outlier data points.

[0054] Principal component analysis is used to integrate resource operation status data from different regions and nodes to generate resource availability assessment data.

[0055] In some embodiments, firstly, for resource operation status data collected from various nodes, including indicators such as network latency, bandwidth consumption, CPU utilization, memory usage, and storage usage, the preprocessing module performs noise reduction processing using a median filtering algorithm. Median filtering is a non-linear filtering method that can effectively remove spike noise or abnormal fluctuations in the data. In cloud platform environments, due to the influence of various external factors (such as sudden traffic or equipment failures), resource status data often exhibits short-term drastic fluctuations. Median filtering can preserve the core trend of the data, filter out occasional noise, and ensure that subsequent processing is based on accurate data. Subsequently, the data undergoes linear normalization. Linear normalization is a method to transform data to a uniform dimension. Typically, it involves subtracting the minimum value in the dataset from each data point and then dividing by the difference between the maximum and minimum values ​​in the dataset, thereby mapping all data to the interval [0,1]. Normalization can eliminate the dimensional differences between different resource operation status data, allowing data of different indicators to be compared and analyzed on the same scale, avoiding unnecessary impact on subsequent analysis due to the large numerical range of certain indicators. After denoising and normalization, the preprocessing module further uses the Isolation Forest algorithm for outlier detection. The Isolation Forest algorithm is a tree-based anomaly detection method that rapidly detects outliers by randomly partitioning the data and constructing a tree structure. Its advantages lie in its efficient handling of large-scale datasets and its strong ability to identify outliers. Through the Isolation Forest algorithm, the preprocessing module automatically identifies anomalous data points that deviate from normal patterns. These data points may be caused by sensor malfunctions, network fluctuations, or other factors and require further processing or removal to prevent them from affecting the accuracy of system decisions. Finally, the preprocessing module uses Principal Component Analysis (PCA) to integrate resource operation status data from different regions and nodes. PCA is a commonly used data dimensionality reduction technique that maps high-dimensional data to a lower-dimensional space through linear transformation, retaining the most important information and removing redundant parts. In cloud platforms, resource operation status data typically involves multiple nodes and regions. This data has high dimensionality, and direct analysis may lead to excessive computation and difficulty in identifying key patterns. Principal component analysis (PCA) allows the preprocessing module to extract the main components from the data, generating resource availability assessment data containing core information. The resulting assessment data accurately reflects the resource health status of each node in the cloud platform, providing a basis for subsequent disaster recovery decisions.

[0056] Furthermore, the disaster recovery switch includes switching computing nodes, adjusting computing resources, storage resources, and network resources.

[0057] Specifically, the first step in disaster recovery failover is switching compute nodes. In a cloud platform environment, compute nodes handle the computational tasks of applications. When a compute node fails or experiences a performance bottleneck, the system needs to perform a failover to transfer tasks to other healthy compute nodes. To achieve this, the system first assesses the load, response time, and other performance metrics of each compute node based on resource availability assessment data generated by the preprocessing module. If a node's resource utilization is too high, causing its computing capacity to be insufficient, the system will automatically trigger a failover, switching the compute tasks of that node to a node with a lower load and more stable performance, ensuring that tasks are not interrupted. Secondly, disaster recovery failover also involves adjusting compute resources. In a cloud platform, compute resources typically consist of multiple virtual machines or containers, and the cloud platform can dynamically allocate compute resources based on the load. When a node's compute resources become bottlenecked (e.g., high CPU utilization), the system will adjust resource allocation through a failover. For example, the system may increase the number of virtual machines or expand the resources of existing virtual machines (e.g., the number of CPU cores or memory size) to balance the load and ensure that computing capacity meets application requirements. Similarly, adjusting storage resources is another crucial step in the disaster recovery failover process. In cloud platforms, storage resources are typically distributed, with multiple nodes sharing storage. In the event of a storage failure, such as a storage node crash or a storage device performance degradation, the system must quickly migrate data to healthy storage nodes. For example, the system can migrate data from a failed node to the disks of other nodes or a distributed storage system, ensuring continuous data availability and consistency. The system dynamically monitors the health status of storage nodes through resource availability assessment data and automatically adjusts storage resources. In cloud platforms, the stability of network connections is crucial to the overall system performance. If a network link fails, or bandwidth consumption at some nodes increases abnormally, impacting network performance, the system will switch over by adjusting network resources. For example, the system may reconfigure the network topology, change data flow paths, or dynamically adjust bandwidth allocation between nodes to ensure uninterrupted data transmission and minimize network latency. This process relies on real-time monitoring of network connection status data to promptly detect anomalies and make adjustments.

[0058] Furthermore, training the Transformer model based on resource availability assessment data includes the following steps:

[0059] The resource availability assessment data is divided into fixed-length time series segments using the sliding window algorithm, and data augmentation is performed to generate training samples.

[0060] Based on the training samples, a mean squared error loss function is defined, and the Transformer model is trained using the backpropagation algorithm of the Adam optimizer.

[0061] Apply a multi-head self-attention mechanism to adjust the weight configuration of nodes.

[0062] Specifically, resource availability assessment data is divided into fixed-length time series segments using a sliding window algorithm. The key to this step is leveraging the temporal correlation of the time series data to capture the dynamic trends of cloud platform resources. The sliding window algorithm divides continuous resource status data into multiple segments, each containing resource status information within a specific time range. By performing data augmentation on these segments, more representative training samples can be generated. Data augmentation can employ various methods, such as noise injection, time shifting, or numerical transformation based on the original data, to increase the model's robustness and generalization ability, thereby effectively avoiding overfitting. Next, based on the generated training samples, the system defines mean squared error (MSE) as the loss function, aiming to minimize the difference between the model's predictions and the actual values. The mean squared error loss function measures the magnitude of the error between the model's predictions and the actual resource status assessment data. To improve the model's learning efficiency, the Adam optimizer is used for backpropagation optimization. The Adam optimizer is a widely used deep learning optimization algorithm that, through adaptive adjustment of the learning rate, effectively avoids gradient vanishing or exploding problems during training, improving the model's convergence speed and accuracy. Furthermore, the multi-head self-attention mechanism of the Transformer model is applied to adjust node weight configurations. By computing multiple attention heads in parallel, the multi-head self-attention mechanism can weight input data in different subspaces, effectively capturing the interdependencies between different resources. In disaster recovery management, the resource states of nodes are interconnected; the failure of one node can have a cascading effect on other nodes. The multi-head self-attention mechanism can dynamically adjust the weight configuration of each node to flexibly respond to different failure scenarios, thereby achieving precise disaster recovery switching decisions. For example, when a node fails, the model can intelligently reallocate resources based on network latency, storage usage, and computing resource requirements among nodes, ensuring that the replacement node can take over computing and data storage tasks in the shortest possible time, thus guaranteeing high system availability.

[0063] Furthermore, the automated orchestration of the disaster recovery switching process for nodes in different regions includes the following steps:

[0064] Input real-time resource availability assessment data into the Transformer model;

[0065] The input data is used to extract features through a multi-head self-attention mechanism, identify the correlation between nodes and the importance of key nodes, and generate context-related feature representations.

[0066] Based on the extracted feature representations, several disaster recovery switching schemes are matched and output.

[0067] In some embodiments, real-time resource availability assessment data is first input into the Transformer model. This assessment data typically includes multi-dimensional resource operational status information, including but not limited to node CPU utilization, memory usage, storage status, and network latency. By inputting this data into the Transformer model, the system can leverage the model's powerful sequence modeling capabilities to capture resource change trends and interdependencies of each node at different points in time, thus providing high-quality input information for disaster recovery switching. Next, the Transformer model extracts features from the input data using a multi-head self-attention mechanism. The multi-head self-attention mechanism is a core component of the Transformer model; it can learn different features of the input data in parallel from multiple subspaces, thereby identifying complex patterns hidden in the data. In disaster recovery management systems, this mechanism can not only extract resource status information of each node but also identify potential correlations between nodes and the importance of critical nodes. For example, some nodes may have higher priority due to differences in geographical location or load status, or have a greater impact on the overall availability of the system at a specific time. Through the multi-head self-attention mechanism, the system can generate context-sensitive feature representations, accurately capturing the dynamic relationships and interdependencies between nodes, thus providing a more precise basis for disaster recovery decisions.

[0068] Finally, based on the extracted feature representations, the system can match and output several disaster recovery switching schemes. These switching schemes are dynamically generated according to the current state and contextual features of the nodes, ensuring that the system can quickly make optimal disaster recovery decisions under different failure scenarios. For example, if a node in a certain region experiences high load or failure, the system will recommend migrating the computing tasks in that region to a region with lower load based on the feature representations generated by the model, while also considering the availability of network bandwidth and storage resources to optimize resource allocation and switching strategies. The output of these switching schemes not only ensures the automation of the disaster recovery process but also ensures that the system can respond quickly and take effective recovery measures when facing complex failure scenarios.

[0069] Furthermore, the formula for the Transformer model is as follows:

[0070]

[0071] Where Y is the output of the Transformer model; X = [L, B, U, M, S]; L is network latency; B is bandwidth consumption; U is CPU utilization; M is memory usage; S is storage usage; W Q To query the weight matrix; WK W is the key weight matrix; V Value weight matrix; d k is the dimension of the key vector; softmax is the normalization function.

[0072] Specifically, suppose a large-scale cloud computing platform consists of multiple geographically distributed regions, each containing several computing nodes. These nodes undertake different computing tasks, and their resource operating status X includes network latency (L), bandwidth consumption (B), CPU utilization (U), memory usage (M), and storage utilization (S). To ensure rapid and effective disaster recovery switching in the event of a failure, this system first collects the resource operating status data of each node in real time through a data acquisition module. In this embodiment, the Transformer model first converts the input resource operating status data into query, key, and value matrices using weight matrices. Subsequently, the model calculates the dot product of the query matrix and the key matrix and divides it by the square root of the key vector dimension to obtain a scaled attention score. These scores are normalized using the softmax function to obtain the attention weight of each node in the current switching scheme. These weights are then multiplied by the value matrix to generate the final output, which comprehensively considers the resource status and interrelationships between nodes.

[0073] Furthermore, the decision control module is used to perform the following steps:

[0074] The disaster recovery switching scheme generated by the disaster recovery process orchestration module is converted into a matrix representation, including the construction of a switching path matrix and a resource allocation matrix;

[0075] Using the switching path matrix and resource allocation matrix as input, a feasibility score is output through a pre-trained graph neural network model;

[0076] Disaster recovery switching schemes are ranked according to their feasibility scores, and the scheme with the highest feasibility score is selected as the implementation scheme.

[0077] The implementation plan will be applied to the cloud platform for control and scheduling.

[0078] In some embodiments, firstly, the decision control module converts the disaster recovery switching scheme from the disaster recovery process orchestration module into a matrix representation. This process includes constructing a switching path matrix and a resource allocation matrix. The switching path matrix represents the path selection and transfer relationships when performing disaster recovery switching between different nodes, typically represented in matrix form, where each element represents the connection state or switching condition between nodes. The resource allocation matrix reflects the computing, storage, network, and other resource requirements allocated to each node under different disaster recovery switching schemes. Converting the disaster recovery switching scheme into a matrix representation facilitates subsequent calculation and optimization using graph neural networks. Next, the switching path matrix and resource allocation matrix are used as inputs for further processing by a pre-trained graph neural network model. Graph neural networks (GNNs) can effectively capture the complex relationships between nodes and the mutual influence between resources. In this context, the role of the graph neural network model is to calculate the feasibility of each path and resource configuration based on the network topology of the switching paths and resource allocation using a backpropagation algorithm, and output a feasibility score. These feasibility scores reflect whether each disaster recovery switching scheme can operate efficiently under resource constraints during execution, and the recovery effect under different fault scenarios. Based on the feasibility score, the decision control module will rank all disaster recovery switchover options and prioritize the option with the highest feasibility score as the final implementation plan. This process ensures that in the event of a failure, the system can select the optimal switchover option for resource scheduling, minimizing cloud platform downtime and ensuring business continuity.

[0079] Furthermore, the graph neural network model is trained through the following steps:

[0080] The switching path matrix and resource allocation matrix are used as input data for the graph structure. Each node in the graph structure is preset as a computing node in the cloud platform, and the edges are preset as the connection relationships between nodes. Each node and edge is assigned a corresponding feature vector, including network latency, bandwidth consumption, CPU utilization, memory usage, and storage usage.

[0081] Feasibility scores are assigned to disaster recovery switching schemes, cross-entropy loss function is selected, and graph neural network model is trained using backpropagation algorithm;

[0082] The performance of the graph neural network model is evaluated using cross-validation.

[0083] Specifically, firstly, the switching path matrix and resource allocation matrix generated in the disaster recovery switching scheme are used as input to construct a graph structure data. Each node in the graph represents a computing node in the cloud platform, and nodes are connected by edges. The connection relationship of the edges represents the dependencies between nodes and the resource flow path. For each node and edge, the system assigns a corresponding feature vector. These feature vectors include multiple dimensions of data closely related to node performance, such as network latency, bandwidth consumption, CPU utilization, memory usage, and storage usage. The feature vector of each node reflects the current resource status of that node, while the feature vector of each edge can represent the transmission cost or connection quality between nodes. These data together constitute the input of the graph neural network model. Next, the system labels the disaster recovery switching scheme with a feasibility score. These feasibility scores reflect the rationality and feasibility of each disaster recovery switching scheme based on the actual operating status of the cloud platform and the efficiency of the switching path. To train the graph neural network model, the system selects the cross-entropy loss function as a metric to evaluate the difference between the model's prediction results and the true labels. The cross-entropy loss function guides the model to optimize by calculating the error between the actual feasibility score and the model's predicted score. Through backpropagation, the graph neural network model continuously adjusts its weights and parameters to reduce prediction errors and improve its accuracy in assessing the feasibility of disaster recovery switching solutions. Finally, cross-validation is used to evaluate the performance of the graph neural network model. During training, cross-validation helps the system evaluate the model's performance on different subsets of data, ensuring the model's generalization ability and stability. Cross-validation typically involves dividing the dataset into multiple subsets, using one subset for validation at a time, and using the remaining subsets for training. The evaluation results from all subsets are then averaged to obtain the overall performance metric of the model. This evaluation step allows the system to confirm whether the graph neural network model can stably make efficient disaster recovery switching decisions in real-world applications.

[0084] Furthermore, the formula for the graph neural network is as follows:

[0085]

[0086] Among them, H (l) H is the node feature matrix of the l-th layer; (l+1) H is the node feature matrix of the (l+1)th layer; (0) The initial layer includes the switching path matrix and resource allocation matrix; A is the normalized adjacency matrix; W (l) Let be the trainable weight matrix of the l-th layer; σ is the activation function.

[0087] It's important to note that the input to the graph neural network is constructed by transforming the switching path matrix and resource allocation matrix into a graph structure. In this graph structure, each node represents a computing node in a cloud platform, and edges represent the connections between nodes or the dependencies in resource scheduling. Each node and edge is assigned a corresponding feature vector, including key resource data such as network latency, bandwidth consumption, CPU utilization, memory usage, and storage usage. For training, the graph neural network utilizes these node and edge features for propagation and updates. Specifically, the model learns the relationships between nodes and resource allocation optimization strategies through a multi-layered information transmission mechanism (i.e., the propagation of node features in the graph). At each layer, the features of a node are updated based on the features of its neighboring nodes. The update process combines the adjacency matrix (representing the connections between nodes) and the node's feature vector, and propagates it to the next layer of nodes through a weighted sum. During training, the graph neural network is optimized by defining a loss function. Each disaster recovery switching scheme is labeled with a feasibility score, and the cross-entropy loss function is chosen to measure the error between the model output and the true label. Backpropagation is used to update the weight matrix in the model, enabling the model to continuously optimize its parameters and improve the accuracy of feasibility predictions for disaster recovery switching solutions. Through this training process, the graph neural network model can effectively identify potential dependencies between different nodes, integrate various resource data, and ultimately generate scores that reflect the feasibility of different disaster recovery switching solutions. After training, the model can rank different disaster recovery solutions based on these scores, select the disaster recovery switching solution with the highest feasibility score, and apply it to the cloud platform for scheduling and implementation, thereby ensuring that the system can efficiently and stably recover and operate in the event of a failure.

[0088] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An intelligent disaster recovery management service system, characterized in that, include: The data acquisition module, preprocessing module, disaster recovery process orchestration module, and decision control module are connected sequentially. The data acquisition module is used to collect resource operation status data of several nodes in different regions of the cloud platform in a distributed manner. The resource operation status data includes network connection status data and storage status data. The preprocessing module is used to perform data cleaning, normalization, and anomaly detection on the resource operation status data to generate resource availability assessment data. The disaster recovery process orchestration module is used to train the Transformer model based on resource availability assessment data, dynamically adjust the key node configuration of the switching scheme through the attention mechanism, automatically orchestrate the disaster recovery switching process of nodes in different regions, and generate several switching schemes. The decision control module is used to predict potential failure risks based on several switching schemes generated by the disaster recovery process orchestration module, generate a feasibility score through a graph neural network model, and select a switching scheme for cloud platform control based on the feasibility score. The automated orchestration of disaster recovery switching processes for nodes in different regions includes the following steps: Input real-time resource availability assessment data into the Transformer model; The input data is used to extract features through a multi-head self-attention mechanism, identify the correlation between nodes and the importance of key nodes, and generate context-related feature representations. Based on the extracted feature representation, several disaster recovery switching schemes are matched and output; The formula for the Transformer model is as follows: ； in, The output of the Transformer model; ; Due to network latency; For bandwidth consumption; CPU utilization; Memory usage; Storage usage; To query the weight matrix; This is the key weight matrix; Value weight matrix; The dimension of the key vector; This is the normalization function.

2. The intelligent disaster recovery management service system according to claim 1, characterized in that, The network connection status data includes network latency and bandwidth consumption, and the storage status data includes node CPU utilization, memory usage, and storage usage.

3. The intelligent disaster recovery management service system according to claim 1, characterized in that, The preprocessing module is used to perform the following steps: The resource operation status data is denoised using a median filtering algorithm and normalized using a linear normalization method. The isolated forest algorithm is used to identify outliers in the cleaned and normalized data and mark potential outlier data points. Principal component analysis is used to integrate resource operation status data from different regions and nodes to generate resource availability assessment data.

4. The intelligent disaster recovery management service system according to claim 1, characterized in that, The disaster recovery switch includes switching computing nodes, adjusting computing resources, storage resources, and network resources.

5. The intelligent disaster recovery management service system according to claim 1, characterized in that, Training the Transformer model based on resource availability assessment data includes the following steps: The resource availability assessment data is divided into fixed-length time series segments using the sliding window algorithm, and data augmentation is performed to generate training samples. Based on the training samples, a mean squared error loss function is defined, and the Transformer model is trained using the backpropagation algorithm of the Adam optimizer. Apply a multi-head self-attention mechanism to adjust the weight configuration of nodes.

6. The intelligent disaster recovery management service system according to claim 1, characterized in that, The decision control module is used to perform the following steps: The disaster recovery switching scheme generated by the disaster recovery process orchestration module is converted into a matrix representation, including the construction of a switching path matrix and a resource allocation matrix; Using the switching path matrix and resource allocation matrix as input, a feasibility score is output through a pre-trained graph neural network model; Disaster recovery switching schemes are ranked according to their feasibility scores, and the scheme with the highest feasibility score is selected as the implementation scheme. The implementation plan will be applied to the cloud platform for control and scheduling.

7. The intelligent disaster recovery management service system according to claim 1, characterized in that, The graph neural network model is trained through the following steps: The switching path matrix and resource allocation matrix are used as input data for the graph structure. Each node in the graph structure is preset as a computing node in the cloud platform, and the edges are preset as the connection relationships between nodes. Each node and edge is assigned a corresponding feature vector, including network latency, bandwidth consumption, CPU utilization, memory usage, and storage usage. Feasibility scores are assigned to disaster recovery switching schemes, cross-entropy loss function is selected, and graph neural network model is trained using backpropagation algorithm; The performance of the graph neural network model is evaluated using cross-validation.

8. The intelligent disaster recovery management service system according to claim 7, characterized in that, The formula for the graph neural network is as follows: ； in, For the first The node feature matrix of the layer; For the first The node feature matrix of the layer; The initial layer includes a switching path matrix and a resource allocation matrix; This is the normalized adjacency matrix; For the first The trainable weight matrix of the layer; This is the activation function.

Images