Operation and maintenance method and system of multi-cloud management platform based on cloud edge collaboration

Abstract

The invention discloses an operation and maintenance method and system of a multi-cloud management platform based on cloud edge collaboration, and the method comprises the following steps: S1, deploying cloud service and edge nodes, and initializing a cloud edge collaboration environment; s2, constructing an edge autonomous optimization network, adjusting and optimizing a task strategy, and predicting a task load of an edge node; s3, adjusting task priorities, and classifying task strategies; s4, constructing a task feature matrix, a cloud edge resource matrix and a task and resource matching model; s5, allocating resources to each task to generate a task resource allocation table; and S6, identifying operation abnormity through the time sequence abnormity detection model, and triggering task migration, resource switching or task redistribution when detecting that the edge node is abnormal. According to the method, the edge autonomous optimization network and the intelligent scheduling algorithm are utilized, dynamic task allocation and exception recovery are achieved, and the method has the advantages of being high in resource utilization rate, high in scheduling real-time performance and good in system stability.

Description

Technical Field

[0001] The present invention relates to the field of cloud management platform operation and maintenance, and in particular to an operation and maintenance method and system for a multi-cloud management platform based on cloud-edge collaboration. Background Art

[0002] With the rapid development of cloud computing and edge computing technologies, multi-cloud collaboration and edge intelligence have become important means to meet complex business needs. Traditional cloud computing relies on centralized data centers. Although it can provide powerful computing power and resource pools, it has limitations in real-time, network latency, and local data processing. Especially in scenarios that require low latency and high reliability, such as the Internet of Things, smart manufacturing, and the Internet of Vehicles, relying solely on cloud resources can no longer meet the needs of efficient operation and maintenance in a dynamic environment. To solve this problem, edge computing came into being, distributing some computing tasks to edge nodes close to the data source to achieve localized data processing and real-time response. However, the independent operation modes of cloud computing and edge computing also bring new challenges. How to achieve collaborative management of cloud resources and edge nodes in a multi-cloud management platform has become a hot topic in current technical research.

[0003] In the prior art, many multi-cloud management platforms have introduced the concept of edge computing, achieving collaboration between the cloud and the edge through distributed resource scheduling. However, these platforms usually adopt static task scheduling and resource allocation methods and lack dynamic optimization capabilities. Specifically, existing resource scheduling schemes are usually based on predefined rules or simple weighted algorithms, which are difficult to adapt to the changing resource status and task characteristics. This lack of intelligent scheduling mechanism may lead to low resource utilization, unreasonable task allocation, and decreased system performance. In addition, most of the existing anomaly detection methods are based on fixed thresholds or simple rule matching, which are prone to misjudgment or missed judgment in complex dynamic environments, and it is difficult to effectively ensure the stability of the system.

[0004] In terms of edge computing, existing technologies usually rely on centralized cloud instructions to control the task execution and resource management of edge nodes, and the edge nodes themselves lack autonomy. This highly cloud-dependent approach not only increases network communication overhead, but may also cause interruptions in task execution in the event of high latency or network disconnection. Even though some technologies have attempted to introduce the concept of edge autonomous optimization, most of their optimization methods remain at the level of simple local rule adjustments, making it difficult to achieve adaptive optimization and dynamic adjustments in complex task scenarios. In addition, edge nodes lack the ability to recover in abnormal situations, and existing recovery solutions usually rely on cloud instructions for global resource adjustments, lacking the edge nodes' own rapid response mechanisms, resulting in weak fault tolerance and robustness of the system.

[0005] Existing technologies have limited dynamic display capabilities for resource distribution and task routing in multi-cloud management platforms. They usually use static configuration files or offline generated routing tables, which are difficult to adapt to dynamically changing resource states and network topologies. This not only affects the accuracy of task scheduling, but also increases the management complexity of the system. In multi-cloud collaboration scenarios, due to the significant heterogeneity of task and resource characteristics, traditional methods are difficult to efficiently build a matching relationship between tasks and resources, resulting in poor efficiency and fairness in resource allocation. At the same time, the lack of refined processing of task priority and real-time performance means that high-priority tasks may not be completed in time due to scheduling delays, which seriously affects the service quality of critical tasks.

[0006] In addition, the modeling methods for anomaly detection in the existing technology mostly use traditional statistical methods or simple machine learning models, which cannot fully capture the complex time series characteristics in the operation status of edge nodes. For example, in the operation status of edge nodes, parameters such as CPU utilization, storage occupancy, and network latency may have nonlinear and time series dependencies, and these relationships are crucial for detecting abnormal events. Existing methods show great limitations in dealing with these complex relationships, resulting in low detection accuracy and difficulty in timely discovering potential risks in the system.

[0007] In summary, the existing technologies have the following defects in the cloud-edge collaborative operation and maintenance of multi-cloud management platforms: lack of intelligent optimization mechanism for dynamic resource scheduling and task allocation, insufficient autonomy of edge nodes, low accuracy and real-time performance of anomaly detection, and limited dynamic display capabilities of resource distribution and task routing. These problems seriously limit the adaptability and stability of the multi-cloud collaborative operation and maintenance system in a complex dynamic environment, so an innovative technical solution is urgently needed to solve the above problems. Summary of the invention

[0008] One purpose of the present invention is to propose an operation and maintenance method of a multi-cloud management platform based on cloud-edge collaboration. The present invention realizes dynamic matching and optimal allocation of tasks and resources by introducing an edge autonomous optimization network and an intelligent scheduling mechanism, thereby improving resource utilization and task execution efficiency. At the same time, through the accurate prediction of the task load of edge nodes through a bidirectional long short-term memory network, the task queue can be scientifically sorted and prioritized in dynamic task scenarios, ensuring the timely processing of high-priority tasks and reducing delays and error rates in the task scheduling process.

[0009] According to an embodiment of the present invention, an operation and maintenance method of a multi-cloud management platform based on cloud-edge collaboration includes the following steps:

[0010] S1. Deploy cloud services and edge nodes in the multi-cloud management platform, initialize the cloud-edge collaborative environment, and configure the computing resources, storage resources, and network resource pools of edge nodes;

[0011] S2. Build an edge autonomous optimization network, use the PPO algorithm to train the edge autonomous optimization network, adjust and optimize the task strategy, and predict the task load of edge nodes through a bidirectional long short-term memory network;

[0012] S3, dynamically adjust task priorities and classify task strategies based on the predicted task load and local resource status of edge nodes;

[0013] S4. Based on the adjusted task priorities, a task feature matrix and a cloud-edge resource matrix are constructed, and a task and resource matching model is established by combining the task feature matrix and the cloud-edge resource matrix;

[0014] S5. Allocate resources to each task according to the task and resource matching model, generate a task resource allocation table, and optimize and adjust the task resource allocation table;

[0015] S6. Monitor the operating status of edge nodes in real time, identify operating anomalies through the timing anomaly detection model, and trigger task migration, resource switching, or task reallocation when edge node anomalies are detected.

[0016] Optionally, the S1 specifically includes:

[0017] S11. Create a cloud control module and an edge management module in the multi-cloud management platform. The cloud control module manages the configuration, distributed deployment and resource association of cloud resources with edge nodes. The cloud resources include computing resources, storage resources and network resources. The edge management module monitors the resource status, task allocation and operation status of edge nodes.

[0018] S12, initializing edge nodes, connecting the edge nodes to the multi-cloud management platform, configuring basic environment parameters of each edge node, and assigning a unique node identifier, wherein the basic environment parameters include network bandwidth, storage capacity, and computing units;

[0019] S13, performing distributed deployment on the cloud service, dividing the cloud resources into multiple virtual resource pools, each of which includes multiple computing instances, storage units, and network interfaces;

[0020] S14. Based on a preset resource allocation rule, the computing instance in the virtual resource pool is bound to the edge node, and a cloud-edge resource association relationship is established. The resource allocation rule includes node location, task delay requirement, and network transmission efficiency.

[0021] S15. Configure monitoring services for edge nodes, monitor the resource usage, network connection status and task allocation of edge nodes in real time through the cloud control module, and generate node status logs;

[0022] S16. Enable a task distribution interface for the edge node, wherein the task distribution interface supports uploading tasks and returning results, including a receiving interface and a response interface for task data streams;

[0023] S17, establishing a task routing table, mapping the connection relationship between the cloud and each edge node into node routing information, the task routing table including node identification, current available bandwidth, task transmission delay and node status identification;

[0024] S18. After completing the initialization configuration of cloud services and edge nodes, a resource distribution map of cloud-edge collaboration is generated, and the resource distribution map is synchronized to the multi-cloud management platform in real time through the cloud control module. The resource distribution map includes the computing power, storage capacity and network connection status of each node.

[0025] Optionally, the S2 specifically includes:

[0026] S21, constructing an edge autonomous optimization network, and adopting the PPO algorithm for training, wherein the edge autonomous optimization network is composed of the PPO algorithm and a bidirectional long short-term memory network, wherein the PPO algorithm optimizes the task strategy, and the bidirectional long short-term memory network predicts the task load of the edge node;

[0027] S22, initializing the parameters of the PPO algorithm, the parameters of the PPO algorithm include state space, action space, reward function and objective function, the state space includes resource utilization of edge nodes, task queue length, network delay and task completion time, the action space includes task execution, task push, resource allocation adjustment and node sleep;

[0028] S23. Construct reward function:

[0029]

[0030] Among them, r t represents the reward value for performing a specific action a at time step t t The performance change of the back edge node, α, γ and δ represent weight parameters, U t Indicates resource utilization, L t Indicates network delay, L threshold represents the delay threshold, β represents the delay sensitivity coefficient, P t represents the task priority loss rate, which is the ratio of high-priority tasks not processed by edge nodes, E t represents the node energy consumption;

[0031] S24, optimization strategy function π θ , sample the action a of the current strategy t Generate samples and use the objective function to maximize the cumulative reward value r tTo update the mission strategy:

[0032] L PPO (θ) = E t [min(r t (θ)·A t ,clip(r t (θ),1-∈,1+∈)·A t )];

[0033]

[0034] Among them, L PPO (θ) represents the objective function of the PPO algorithm, E t represents the expected value at time step t, r t (θ) represents the importance sampling ratio, which represents the ratio of the current strategy to the old strategy, and clip represents the clipping function, which is used to limit r t The range of (θ), ∈ represents the cutoff range, which is used to limit the change range of the strategy update, A t represents the advantage function, which is used to estimate the performance difference between the current action and the benchmark action, γ represents the discount factor, which is used to weigh the importance of current rewards and future rewards, represents the state value at time step t, θ represents the parameters of the current strategy, and π θ Indicates the current strategy, represents the old policy, a t represents the action of the current strategy, s t Represents the node resource status at time step t;

[0035] S25, collecting load data of historical tasks of edge nodes and preprocessing the load data, wherein the load data includes CPU usage, memory usage, storage remaining amount and task queue length;

[0036] S26. Construct a bidirectional long short-term memory network, design two forward and backward long short-term memory networks to capture the forward and reverse time series characteristics of the task load, use the mean square error as the loss function, and use the Adam optimization algorithm to iteratively train the parameters of the bidirectional long short-term memory network until the value of the loss function converges;

[0037] S27. Deploy the trained bidirectional long short-term memory network to the task scheduling module of the edge node, receive the real-time task load sequence as input, and output the task load for several time periods in the future.

[0038] Optionally, the S3 specifically includes:

[0039] S31. Obtain the load data output by the bidirectional long short-term memory network and define the resource status indicator function RS tRepresents the resource load at time step t:

[0040]

[0041] Among them, RS t represents the resource status indicator function, w1, w2, w3 and w4 represent weight parameters, k represents the storage space adjustment parameter, U CPU Indicates CPU usage, U MEM Indicates memory usage, S FREE Indicates the remaining storage capacity, Q LEN Indicates the length of the task queue;

[0042] S32. Define the task priority calculation function to dynamically adjust the task priority in the task queue:

[0043]

[0044] Among them, P represents the task priority calculation function, T RT Indicates the real-time requirements of the task, D SIZE Indicates the amount of data, T LIMIT Indicates the time limit for completion;

[0045] S33. Based on the output of the task priority calculation function, the task strategy is classified, and the classification includes local execution, push to the cloud and delayed processing. If the priority calculation function value is greater than the high threshold of the priority, the task is executed preferentially at the local edge node. If the priority calculation function value is less than the low threshold of the priority, the task is delayed. If the priority calculation function value is between the high threshold and the low threshold of the priority, the task is pushed to the cloud for processing.

[0046] Optionally, the S4 specifically includes:

[0047] S41. Construct a task feature matrix and a cloud edge resource matrix. The task features include the real-time nature of the task, computational complexity, data volume, task arrival time, and priority. The cloud edge resources include computing resources, storage capacity, network bandwidth, and resource availability.

[0048] T={(C i ,T RT,i ,D SIZE,i ,P i ,T ARR,i )|i=1,2,,N};

[0049] R={(E CPU,j ,E MEM,j ,B j ,L j )|j=1,2,,M};

[0050] Among them, T represents the task feature matrix, R represents the cloud-edge resource matrix, and C i represents the computational complexity, T RT,i Indicates the real-time nature of the task, D SIZE,i Indicates the amount of data, P i Indicates the task priority, T ARR,i represents the task arrival time, E CPU,j represents computing resources, E MEM,j Indicates storage capacity, B j Indicates network bandwidth, L j represents resource availability, N represents the total number of tasks, and M represents the total number of edge nodes;

[0051] S42. Establish a task and resource matching model and calculate the matching degree between tasks and resources:

[0052]

[0053] Among them, S(i,j) represents the matching function, w1, w2, w3, w4, w5 and w6 represent the weight coefficients of the matching function, indicating the influence of different task characteristics on the matching degree;

[0054] S43. According to the value of the matching degree model, a matching relationship matrix S between tasks and resources is constructed:

[0055] S={S(i,j)|i=1,2,,N; j=1,2,,M};

[0056] Among them, S(i,j) represents the matching degree between task index i and resource node index j.

[0057] Optionally, the S5 specifically includes:

[0058] S51. Sort each row in the matching relationship matrix S to find the best match between the current task and the resource node, that is, S(i,j) with the largest matching value:

[0059] j * = argmax j S(i,j),j=1,2,…,n;

[0060] Among them, j * Indicates the best resource node index matching the i-th task, argmax j S(i,j) means finding the index of the maximum value of S(i,j);

[0061] S52. Constructing a task resource allocation table:

[0062] A={(i,j * )|i=1,2,…,m};

[0063] Among them, A represents the task resource allocation table, (i,j * ) indicates that task i is assigned to resource node j * ;

[0064] S53, for each pair of allocation relations (i, j * ), re-evaluate the rationality of the allocation by optimizing the objective function:

[0065] O(i,j * )=λ1·P(i)+λ2·A r (j * )+λ3·C r (j * )+λ4·B r (j * );

[0066] Among them, O(i,j * ) represents the optimized allocation score, λ1, λ2, λ3 and λ4 represent the weight coefficients of the matching function, and represent the optimization weight coefficients, which are used to balance the influence of priority and resource attributes. r (j * ) represents the jth * The computing power of resource nodes, B r (j * ) represents the jth * The network bandwidth of resource nodes, A r (j * ) represents the jth * The current availability of resource nodes;

[0067] S54, if the assignment relationship of task i (i, j * ) score is low, then rematch the resource nodes and look for feasible alternative nodes to satisfy:

[0068] O(i,j′)>O(i,j * );

[0069] Among them, O(i,j′) represents the new assigned score.

[0070] Optionally, the S6 specifically includes:

[0071] S61, monitor the running status of the edge node in real time, collect multi-dimensional monitoring data, the multi-dimensional monitoring data includes CPU usage, storage occupancy, network delay and task execution status, and pre-process the multi-dimensional monitoring data;

[0072] S62. Use a gated temporal convolutional network in combination with an autoencoder to construct a time series anomaly detection model, wherein the gated temporal convolutional network includes causal convolution, dilated convolution, and a gating mechanism, and the autoencoder consists of an encoder and a decoder;

[0073] S63. Use the gated temporal convolutional network to extract multi-scale temporal features from the pre-processed multi-dimensional monitoring data and output the temporal embedding features:

[0074] H t =σ(W f *S t +b f )⊙tanh(W g *S t +b g );

[0075] Among them, H t represents the temporal embedding feature at time step t, W f and W g Represent the weights of the gate and main convolution kernel, b f and b g represents the corresponding bias term, σ represents the Sigmoid activation function, ⊙ represents the point-by-point product, S t represents the input sequence;

[0076] S64, inputting the time series embedding feature into the autoencoder, compressing the high-dimensional time series embedding feature into a low-dimensional potential representation, and decoding the low-dimensional potential representation back into reconstructed data of the original input dimension through the decoder of the autoencoder;

[0077] S65. Compare the input data sequence and the reconstructed data generated by the autoencoder, and calculate the reconstruction error:

[0078]

[0079] Where E represents the reconstruction error, t represents the time step, and x i represents the input data sequence, represents the reconstructed data generated by the autoencoder;

[0080] S66, comparing the reconstruction error with a preset threshold, and determining it as abnormal when the reconstruction error is greater than the preset threshold;

[0081] S67. When an abnormal edge node is detected, an autonomous recovery mechanism is triggered according to the task queue and resource allocation status. The autonomous recovery mechanism includes task migration, resource switching and task reallocation. The task migration means migrating high-priority tasks on the abnormal node to other normal nodes. The resource switching means enabling redundant resource nodes to take over the tasks of the abnormal node. The task reallocation means recalculating the priority of unfinished tasks and reallocating them to suitable resource nodes.

[0082] An operation and maintenance system of a multi-cloud management platform based on cloud-edge collaboration according to an embodiment of the present invention includes:

[0083] The cloud management module is used to manage the configuration, distributed deployment and resource association of cloud resources with edge nodes;

[0084] The edge management module is used to monitor the resource status, task allocation and operation status of edge nodes;

[0085] The task scheduling module is used to generate the matching relationship between tasks and resources based on task characteristics and resource status, and optimize the task resource allocation strategy;

[0086] Edge autonomy module, which is used to realize task priority classification and dynamic policy adjustment of edge nodes by optimizing network and load prediction;

[0087] Anomaly detection module, which is used to monitor the status of edge nodes in real time, detect anomalies and make relevant adjustments;

[0088] Resource distribution map generation module, used to generate and synchronize cloud-edge collaborative resource distribution maps and display resource status;

[0089] Task management module, which is used to record task life cycle data, update task status and support scheduling and recovery operations;

[0090] The communication interface module is used to realize task data transmission and status feedback between the cloud and edge nodes.

[0091] The beneficial effects of the present invention are:

[0092] First of all, the present invention effectively realizes the precise matching and optimal allocation of tasks and resources by adopting the edge autonomous optimization network and dynamic task scheduling mechanism. The introduction of the bidirectional long short-term memory network enables the edge nodes to accurately predict the task load, and combines the task priority dynamic adjustment strategy to ensure the real-time and high efficiency of task scheduling, and reduce task delays and resource waste.

[0093] Secondly, in view of the high complexity of tasks and strong heterogeneity of resources in a multi-cloud environment, the present invention dynamically calculates the optimal allocation plan for tasks and resources through multi-dimensional modeling of the task feature matrix and the resource status matrix, combined with the matching function. Unlike traditional static rules or simple weighted algorithms, the present invention can adjust the allocation strategy in real time to adapt to changes in task characteristics and resource status, ensuring the efficient operation of the system in complex scenarios. Through the optimized task resource allocation table and scheduling plan, the present invention improves resource utilization and task execution efficiency, and provides a new technical path for intelligent task management in a multi-cloud environment.

[0094] Again, in terms of anomaly detection, the present invention uses a gated temporal convolutional network combined with an autoencoder to construct a time series anomaly detection model, which can extract deep time series features from multi-dimensional monitoring data and accurately determine anomalies by reconstructing errors. Compared with traditional rule matching or threshold detection methods, the accuracy and robustness of this model in processing complex time series data are significantly improved. When an anomaly occurs, the autonomous recovery mechanism designed by the present invention can quickly initiate task migration, resource switching or task reallocation, ensuring the high fault tolerance and stability of the system in emergencies, and reducing the risk of task interruption and data loss.

[0095] Finally, the present invention intuitively displays the distribution status of cloud-edge resources and the task execution path by dynamically generating and synchronizing resource distribution maps in real time, combined with the update mechanism of the task routing table. The real-time and visualization of the resource distribution map not only reduces the complexity of operation and maintenance management, but also significantly improves the transparency and decision-making efficiency of the system. Combining the capabilities of cloud-edge and edge collaborative optimization, the present invention provides comprehensive data support and intelligent operation tools for operation and maintenance managers.

[0096] In summary, the present invention has achieved improvements in resource utilization efficiency, task scheduling real-time and system stability through the combination of intelligent scheduling, precise anomaly detection and efficient recovery mechanism. It has shown strong adaptability and efficiency in dynamic and complex environments, and can provide advanced technical support and solutions for operation and maintenance management in multi-cloud collaborative scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] 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:

[0098] Figure 1 A flowchart of an operation and maintenance method of a multi-cloud management platform based on cloud-edge collaboration proposed by the present invention;

[0099] Figure 2This is a flowchart for the construction of a task feature matrix and a resource status matrix and an optimization of the task-resource matching relationship of an operation and maintenance method of a multi-cloud management platform based on cloud-edge collaboration proposed by the present invention. DETAILED DESCRIPTION

[0100] 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.

[0101] refer to Figure 1 and Figure 2 , an operation and maintenance method of a multi-cloud management platform based on cloud-edge collaboration, comprising the following steps:

[0102] S1. Deploy cloud services and edge nodes in the multi-cloud management platform, initialize the cloud-edge collaborative environment, and configure the computing resources, storage resources, and network resource pools of edge nodes;

[0103] S2. Build an edge autonomous optimization network, use the PPO algorithm to train the edge autonomous optimization network, adjust and optimize the task strategy, and predict the task load of edge nodes through a bidirectional long short-term memory network;

[0104] S3, dynamically adjust task priorities and classify task strategies based on the predicted task load and local resource status of edge nodes;

[0105] S4. Based on the adjusted task priorities, a task feature matrix and a cloud-edge resource matrix are constructed, and a task and resource matching model is established by combining the task feature matrix and the cloud-edge resource matrix;

[0106] S5. Allocate resources to each task according to the task and resource matching model, generate a task resource allocation table, and optimize and adjust the task resource allocation table;

[0107] S6. Monitor the operating status of edge nodes in real time, identify operating anomalies through the timing anomaly detection model, and trigger task migration, resource switching, or task reallocation when edge node anomalies are detected.

[0108] In this implementation manner, the S1 specifically includes:

[0109] S11. Create a cloud control module and an edge management module in the multi-cloud management platform. The cloud control module manages the configuration, distributed deployment and resource association of cloud resources with edge nodes. The cloud resources include computing resources, storage resources and network resources. The edge management module monitors the resource status, task allocation and operation status of edge nodes.

[0110] S12, initializing edge nodes, connecting the edge nodes to the multi-cloud management platform, configuring basic environment parameters of each edge node, and assigning a unique node identifier, wherein the basic environment parameters include network bandwidth, storage capacity, and computing units;

[0111] S13, performing distributed deployment on the cloud service, dividing the cloud resources into multiple virtual resource pools, each of which includes multiple computing instances, storage units, and network interfaces;

[0112] S14. Based on a preset resource allocation rule, the computing instance in the virtual resource pool is bound to the edge node, and a cloud-edge resource association relationship is established. The resource allocation rule includes node location, task delay requirement, and network transmission efficiency.

[0113] S15. Configure monitoring services for edge nodes, monitor the resource usage, network connection status and task allocation of edge nodes in real time through the cloud control module, and generate node status logs;

[0114] S16. Enable a task distribution interface for the edge node, wherein the task distribution interface supports uploading tasks and returning results, including a receiving interface and a response interface for task data streams;

[0115] S17, establishing a task routing table, mapping the connection relationship between the cloud and each edge node into node routing information, the task routing table including node identification, current available bandwidth, task transmission delay and node status identification;

[0116] S18. After completing the initialization configuration of cloud services and edge nodes, a resource distribution map of cloud-edge collaboration is generated, and the resource distribution map is synchronized to the multi-cloud management platform in real time through the cloud control module. The resource distribution map includes the computing power, storage capacity and network connection status of each node.

[0117] In this implementation, S2 specifically includes:

[0118] S21, constructing an edge autonomous optimization network, and adopting the PPO algorithm for training, wherein the edge autonomous optimization network is composed of the PPO algorithm and a bidirectional long short-term memory network, wherein the PPO algorithm optimizes the task strategy, and the bidirectional long short-term memory network predicts the task load of the edge node;

[0119] S22, initializing the parameters of the PPO algorithm, the parameters of the PPO algorithm include state space, action space, reward function and objective function, the state space includes resource utilization of edge nodes, task queue length, network delay and task completion time, the action space includes task execution, task push, resource allocation adjustment and node sleep;

[0120] S23. Construct reward function:

[0121]

[0122] Among them, r t represents the reward value for performing a specific action a at time step t t The performance change of the back edge node, α, γ and δ represent weight parameters, U t Indicates resource utilization, L t Indicates network delay, L threshold represents the delay threshold, β represents the delay sensitivity coefficient, P t represents the task priority loss rate, which is the ratio of high-priority tasks not processed by edge nodes, E t represents the node energy consumption;

[0123] S24, optimization strategy function π θ , sample the action a of the current strategy t Generate samples and use the objective function to maximize the cumulative reward value r t To update the mission strategy:

[0124] L PPO (θ) = E t [min(r t (θ)·A t ,clip(r t (θ),1-∈,1+∈)·A t )];

[0125]

[0126] Among them, L PPO (θ) represents the objective function of the PPO algorithm, E t represents the expected value at time step t, r t (θ) represents the importance sampling ratio, which represents the ratio of the current strategy to the old strategy, and clip represents the clipping function, which is used to limit r t The range of (θ), ∈ represents the cutoff range, which is used to limit the change range of the strategy update, A t represents the advantage function, which is used to estimate the performance difference between the current action and the benchmark action, γ represents the discount factor, which is used to weigh the importance of current rewards and future rewards, represents the state value at time step t, θ represents the parameters of the current strategy, and π θ Indicates the current strategy, represents the old policy, a t represents the action of the current strategy, s t Represents the node resource status at time step t;

[0127] S25, collecting load data of historical tasks of edge nodes and preprocessing the load data, wherein the load data includes CPU usage, memory usage, storage remaining amount and task queue length;

[0128] S26. Construct a bidirectional long short-term memory network, design two forward and backward long short-term memory networks to capture the forward and reverse time series characteristics of the task load, use the mean square error as the loss function, and use the Adam optimization algorithm to iteratively train the parameters of the bidirectional long short-term memory network until the value of the loss function converges;

[0129] S27. Deploy the trained bidirectional long short-term memory network to the task scheduling module of the edge node, receive the real-time task load sequence as input, and output the task load for several time periods in the future.

[0130] In this implementation, S3 specifically includes:

[0131] S31. Obtain the load data output by the bidirectional long short-term memory network and define the resource status indicator function RS t Represents the resource load at time step t:

[0132]

[0133] Among them, RS t represents the resource status indicator function, w1, w2, w3 and w4 represent weight parameters, k represents the storage space adjustment parameter, U CPU Indicates CPU usage, U MEM Indicates memory usage, S FREE Indicates the remaining storage capacity, Q LEN Indicates the length of the task queue;

[0134] S32. Define the task priority calculation function to dynamically adjust the task priority in the task queue:

[0135]

[0136] Among them, P represents the task priority calculation function, T RT Indicates the real-time requirements of the task, D SIZE Indicates the amount of data, T LIMIT Indicates the time limit for completion;

[0137] S33. Based on the output of the task priority calculation function, the task strategy is classified, and the classification includes local execution, push to the cloud and delayed processing. If the priority calculation function value is greater than the high threshold of the priority, the task is executed preferentially at the local edge node. If the priority calculation function value is less than the low threshold of the priority, the task is delayed. If the priority calculation function value is between the high threshold and the low threshold of the priority, the task is pushed to the cloud for processing.

[0138] In this implementation manner, the S4 specifically includes:

[0139] S41. Construct a task feature matrix and a cloud edge resource matrix. The task features include the real-time nature of the task, computational complexity, data volume, task arrival time, and priority. The cloud edge resources include computing resources, storage capacity, network bandwidth, and resource availability.

[0140] T={(C i ,T RT,i ,D SIZE,i ,P i ,T ARR,i )|i=1,2,,N};

[0141] R={(E CPU,j ,E MEM,j ,B j ,L j )|j=1,2,,M};

[0142] Among them, T represents the task feature matrix, R represents the cloud-edge resource matrix, and C i represents the computational complexity, T RT,i Indicates the real-time nature of the task, D SIZE,i Indicates the amount of data, P i Indicates the task priority, T ARR,i represents the task arrival time, E CPU,j represents computing resources, E MEM,j Indicates storage capacity, B j Indicates network bandwidth, L j represents resource availability, N represents the total number of tasks, and M represents the total number of edge nodes;

[0143] S42. Establish a task and resource matching model and calculate the matching degree between tasks and resources:

[0144]

[0145] Among them, S(i,j) represents the matching function, w1, w2, w3, w4, w5 and w6 represent the weight coefficients of the matching function, indicating the influence of different task characteristics on the matching degree;

[0146] S43. According to the value of the matching degree model, a matching relationship matrix S between tasks and resources is constructed:

[0147] S={S(i,j)|i=1,2,,N; j=1,2,,M};

[0148] Among them, S(i,j) represents the matching degree between task index i and resource node index j.

[0149] In this implementation manner, S5 specifically includes:

[0150] S51. Sort each row in the matching relationship matrix S to find the best match between the current task and the resource node, that is, S(i,j) with the largest matching value:

[0151] j * = argmax j S(i,j),j=1,2,…,n;

[0152] Among them, j * Indicates the best resource node index matching the i-th task, argmax j S9i,j) means finding the index of the maximum value of S(i,j);

[0153] S52. Constructing a task resource allocation table:

[0154] A={(i,j * )|i=1,2,…,m};

[0155] Among them, A represents the task resource allocation table, (i,j * ) indicates that task i is assigned to resource node j * ;

[0156] S53, for each pair of allocation relations (i, j * ), re-evaluate the rationality of the allocation by optimizing the objective function:

[0157] O(i,j * )=λ1·P(i)+λ2·A r (j * )+λ3·C r (j * )+λ4·B r (j * );

[0158] Among them, O(i,j * ) represents the optimized allocation score, λ1, λ2, λ3 and λ4 represent the weight coefficients of the matching function, and represent the optimization weight coefficients, which are used to balance the influence of priority and resource attributes. r (j* ) represents the jth * The computing power of resource nodes, B r (j * ) represents the jth * The network bandwidth of resource nodes, A r (j * ) represents the jth * The current availability of resource nodes;

[0159] S54, if the assignment relationship of task i (i, j * ) score is low, then rematch the resource nodes and look for feasible alternative nodes to satisfy:

[0160] O(i,j′)>O(i,j * );

[0161] Among them, O(i,j′) represents the new assigned score.

[0162] In this implementation manner, S6 specifically includes:

[0163] S61, monitor the running status of the edge node in real time, collect multi-dimensional monitoring data, the multi-dimensional monitoring data includes CPU usage, storage occupancy, network delay and task execution status, and pre-process the multi-dimensional monitoring data;

[0164] S62. Use a gated temporal convolutional network in combination with an autoencoder to construct a time series anomaly detection model, wherein the gated temporal convolutional network includes causal convolution, dilated convolution, and a gating mechanism, and the autoencoder consists of an encoder and a decoder;

[0165] S63. Use the gated temporal convolutional network to extract multi-scale temporal features from the pre-processed multi-dimensional monitoring data and output the temporal embedding features:

[0166] H t =σ(W f *S t +b f )⊙tanh(W g *S t +b g );

[0167] Among them, H t represents the temporal embedding feature at time step t, W f and W g Represent the weights of the gate and main convolution kernel, b f and b g represents the corresponding bias term, σ represents the Sigmoid activation function, ⊙ represents the point-by-point product, S t represents the input sequence;

[0168] S64, inputting the time series embedding feature into the autoencoder, compressing the high-dimensional time series embedding feature into a low-dimensional potential representation, and decoding the low-dimensional potential representation back into reconstructed data of the original input dimension through the decoder of the autoencoder;

[0169] S65. Compare the input data sequence and the reconstructed data generated by the autoencoder, and calculate the reconstruction error:

[0170]

[0171] Where E represents the reconstruction error, t represents the time step, and x i represents the input data sequence, represents the reconstructed data generated by the autoencoder;

[0172] S66, comparing the reconstruction error with a preset threshold, and determining it as abnormal when the reconstruction error is greater than the preset threshold;

[0173] S67. When an abnormal edge node is detected, an autonomous recovery mechanism is triggered according to the task queue and resource allocation status. The autonomous recovery mechanism includes task migration, resource switching and task reallocation. The task migration means migrating high-priority tasks on the abnormal node to other normal nodes. The resource switching means enabling redundant resource nodes to take over the tasks of the abnormal node. The task reallocation means recalculating the priority of unfinished tasks and reallocating them to suitable resource nodes.

[0174] An operation and maintenance system of a multi-cloud management platform based on cloud-edge collaboration, comprising:

[0175] The cloud management module is used to manage the configuration, distributed deployment and resource association of cloud resources with edge nodes;

[0176] The edge management module is used to monitor the resource status, task allocation and operation status of edge nodes;

[0177] The task scheduling module is used to generate the matching relationship between tasks and resources based on task characteristics and resource status, and optimize the task resource allocation strategy;

[0178] Edge autonomy module, which is used to realize task priority classification and dynamic policy adjustment of edge nodes by optimizing network and load prediction;

[0179] Anomaly detection module, which is used to monitor the status of edge nodes in real time, detect anomalies and make relevant adjustments;

[0180] Resource distribution map generation module, used to generate and synchronize cloud-edge collaborative resource distribution maps and display resource status;

[0181] Task management module, which is used to record task life cycle data, update task status and support scheduling and recovery operations;

[0182] The communication interface module is used to realize task data transmission and status feedback between the cloud and edge nodes.

[0183] Embodiment 1:

[0184] The scenario of this embodiment is in a smart manufacturing factory, where the factory's production process relies on the collaborative work of a large number of distributed devices, such as CNC machine tools, industrial robots, sensor networks, etc. The tasks generated by these devices include quality inspection tasks with high real-time requirements, image processing tasks with large computational loads, and low-priority log archiving tasks. The traditional cloud computing model is difficult to meet the needs of real-time response due to bandwidth limitations and network latency issues, while the model that relies entirely on edge computing cannot efficiently complete high-computation tasks due to limited resources. In addition, the operating status of the equipment is complex and dynamically changeable, and problems such as unreasonable task scheduling, resource overload, and equipment abnormalities may occur, seriously affecting production efficiency and system stability.

[0185] In this scenario, the operation and maintenance method of the multi-cloud management platform based on cloud-edge collaboration of the present invention is applied to the task scheduling and exception management of the production line. First, the factory deployed a multi-cloud management platform, and the cloud service was responsible for the unified management of computing resources, storage resources and network resources. At the same time, edge nodes were deployed on the production line for the processing of local real-time tasks. By building an edge autonomous optimization network, the edge node can optimize the task execution strategy according to the PPO algorithm and predict the task load using a bidirectional long short-term memory network. Based on task characteristics and resource status, the system dynamically adjusts the task priority and generates a task resource allocation table in real time. For abnormal situations in the production line, such as equipment overload or network delay exceeding the standard, the system quickly identifies the anomaly through a timing anomaly detection model and triggers an autonomous recovery mechanism, including task migration and resource switching, to ensure the efficient operation of the production line. In the intelligent manufacturing factory, the system processed 500 tasks respectively, and the task types included quality inspection tasks with high real-time requirements (120), image processing tasks with large computational load (250), and low-priority log archiving tasks (130).

[0186] Table 1 Test data comparison table

[0187] Test indicators Traditional methods Method of the present invention Optimization effect Average task scheduling time (seconds) 3.0 2.1 30% reduction Edge node CPU utilization 65％ 85％ 20% improvement Cloud resource utilization 50％ 72％ 22% improvement Anomaly detection accuracy 90％ 98.5％ Improved by 8.5% Mission success rate 92％ 96.5％ Improved by 4.5% Abnormal recovery time (seconds) 3.5 2.0 42.8% reduction Abnormal recovery rate 85％ 95％ Improve by 10%

[0188] In terms of task scheduling efficiency, the average task scheduling time of the method of the present invention is reduced from 3.0 seconds of the traditional method to 2.1 seconds, and the optimization range reaches 30%. This result shows that through dynamic task allocation and resource optimization scheduling mechanism, the present invention can quickly complete resource allocation and task scheduling in high-concurrency task scenarios, effectively improving the overall operation efficiency.

[0189] In terms of resource utilization, the CPU utilization of edge nodes has increased from 65% of the traditional method to 85%, and the cloud resource utilization has increased from 50% to 72%. This significant increase in resource utilization is due to the improvements in the task characteristics and resource status modeling of the present invention, as well as the application of dynamic matching algorithms, which make resource allocation more reasonable and efficient. At the same time, this optimization also avoids the problems of idle resources and task accumulation, providing a better solution for resource management in a multi-cloud environment.

[0190] The present invention also demonstrates powerful capabilities in anomaly detection. The accuracy of anomaly detection has increased from 90% of the traditional method to 98.5%, an increase of 8.5 percentage points, indicating that the detection model combining the gated time convolutional network and the autoencoder adopted by the present invention can more accurately capture complex time series features and significantly reduce the occurrence of missed detections and false detections. In terms of anomaly recovery efficiency, the average recovery time is shortened from 3.5 seconds of the traditional method to 2.0 seconds, and the anomaly recovery rate is increased from 85% to 95%. This result proves that the autonomous recovery mechanism of the present invention can respond to sudden abnormal situations quickly and efficiently, ensuring the continuity of tasks and the stability of the system.

[0191] In addition, the present invention also improves the overall success rate of tasks from 92% of the traditional method to 96.5%, especially in the execution of high-priority tasks. This effect shows that through the dynamic adjustment of task priorities and scientific task sequencing, the present invention can ensure the efficient completion of key tasks and provide strong technical support for time-sensitive tasks.

[0192] In summary, the present invention has achieved optimization in key indicators such as task scheduling efficiency, resource utilization, anomaly detection accuracy, and task success rate compared with traditional methods. These data fully verify the technical advantages and application value of the method of the present invention in complex dynamic environments, and can provide a reliable solution for operation and maintenance management in multi-cloud collaborative scenarios.

[0193] 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. An operation and maintenance method of a multi-cloud management platform based on cloud-edge collaboration, characterized in that: The steps include: S1. Deploy cloud services and edge nodes in the multi-cloud management platform, initialize the cloud-edge collaborative environment, and configure the computing resources, storage resources, and network resource pools of edge nodes; S2. Build an edge autonomous optimization network, use the PPO algorithm to train the edge autonomous optimization network, adjust and optimize the task strategy, and predict the task load of edge nodes through a bidirectional long short-term memory network; S3, dynamically adjust task priorities and classify task strategies based on the predicted task load and local resource status of edge nodes; S4. Based on the adjusted task priorities, a task feature matrix and a cloud-edge resource matrix are constructed, and a task and resource matching model is established by combining the task feature matrix and the cloud-edge resource matrix; S5. Allocate resources to each task according to the task and resource matching model, generate a task resource allocation table, and optimize and adjust the task resource allocation table; S6. Monitor the operating status of edge nodes in real time, identify operating anomalies through the timing anomaly detection model, and trigger task migration, resource switching, or task reallocation when edge node anomalies are detected.

2. According to claim 1, an operation and maintenance method of a multi-cloud management platform based on cloud-edge collaboration is characterized in that: The S1 specifically includes: S11. Create a cloud control module and an edge management module in the multi-cloud management platform. The cloud control module manages the configuration, distributed deployment and resource association of cloud resources with edge nodes. The cloud resources include computing resources, storage resources and network resources. The edge management module monitors the resource status, task allocation and operation status of edge nodes. S12, initializing edge nodes, connecting the edge nodes to the multi-cloud management platform, configuring basic environment parameters of each edge node, and assigning a unique node identifier, wherein the basic environment parameters include network bandwidth, storage capacity, and computing units; S13, performing distributed deployment on the cloud service, dividing the cloud resources into multiple virtual resource pools, each of which includes multiple computing instances, storage units, and network interfaces; S14. Based on a preset resource allocation rule, the computing instance in the virtual resource pool is bound to the edge node, and a cloud-edge resource association relationship is established. The resource allocation rule includes node location, task delay requirement, and network transmission efficiency. S15. Configure monitoring services for edge nodes, monitor the resource usage, network connection status and task allocation of edge nodes in real time through the cloud control module, and generate node status logs; S16. Enable a task distribution interface for the edge node, wherein the task distribution interface supports uploading tasks and returning results, including a receiving interface and a response interface for task data streams; S17, establishing a task routing table, mapping the connection relationship between the cloud and each edge node into node routing information, the task routing table including node identification, current available bandwidth, task transmission delay and node status identification; S18. After completing the initialization configuration of cloud services and edge nodes, a resource distribution map of cloud-edge collaboration is generated, and the resource distribution map is synchronized to the multi-cloud management platform in real time through the cloud control module. The resource distribution map includes the computing power, storage capacity and network connection status of each node.

3. According to claim 1, the operation and maintenance method of a multi-cloud management platform based on cloud-edge collaboration is characterized in that: The S2 specifically includes: S21, constructing an edge autonomous optimization network, and adopting the PPO algorithm for training, wherein the edge autonomous optimization network is composed of the PPO algorithm and a bidirectional long short-term memory network, wherein the PPO algorithm optimizes the task strategy, and the bidirectional long short-term memory network predicts the task load of the edge node; S22, initializing the parameters of the PPO algorithm, the parameters of the PPO algorithm include state space, action space, reward function and objective function, the state space includes resource utilization of edge nodes, task queue length, network delay and task completion time, the action space includes task execution, task push, resource allocation adjustment and node sleep; S23. Construct reward function: Among them, r t represents the reward value for performing a specific action a at time step t t The performance change of the back edge node, α, γ and δ represent weight parameters, U t Indicates resource utilization, L t Indicates network delay, L threshold represents the delay threshold, β represents the delay sensitivity coefficient, P t represents the task priority loss rate, which is the ratio of high-priority tasks not processed by edge nodes, E t represents the node energy consumption; S24, optimization strategy function π θ , sample the action a of the current strategy t Generate samples and use the objective function to maximize the cumulative reward value r t To update the mission strategy: L PPO (θ)=E t [min(r t (θ)·A t ,clip(r t (θ),1-∈,1+∈)·A t )]; Among them, L PPO (θ) represents the objective function of the PPO algorithm, E t represents the expected value at time step t, r t (θ) represents the importance sampling ratio, which represents the ratio of the current strategy to the old strategy, and clip represents the clipping function, which is used to limit r t The range of (θ), ∈ represents the cutoff range, which is used to limit the change range of the strategy update, A t represents the advantage function, which is used to estimate the performance difference between the current action and the benchmark action, γ represents the discount factor, which is used to weigh the importance of current rewards and future rewards, represents the state value at time step t, θ represents the parameters of the current strategy, and π θ Indicates the current strategy, represents the old policy, a t represents the action of the current strategy, s t Represents the node resource status at time step t; S25, collecting load data of historical tasks of edge nodes and preprocessing the load data, wherein the load data includes CPU usage, memory usage, storage remaining amount and task queue length; S26. Construct a bidirectional long short-term memory network, design two forward and backward long short-term memory networks to capture the forward and reverse time series characteristics of the task load, use the mean square error as the loss function, and use the Adam optimization algorithm to iteratively train the parameters of the bidirectional long short-term memory network until the value of the loss function converges; S27. Deploy the trained bidirectional long short-term memory network to the task scheduling module of the edge node, receive the real-time task load sequence as input, and output the task load for several time periods in the future.

4. The operation and maintenance method of a multi-cloud management platform based on cloud-edge collaboration according to claim 1 is characterized in that: The S3 specifically includes: S31. Obtain the load data output by the bidirectional long short-term memory network and define the resource status indicator function RS t Represents the resource load at time step t: Among them, RS t represents the resource status indicator function, w1, w2, w3 and w4 represent weight parameters, k represents the storage space adjustment parameter, U CPU Indicates CPU usage, U MEM Indicates memory usage, S FREE Indicates the remaining storage capacity, Q LEN Indicates the length of the task queue; S32. Define the task priority calculation function to dynamically adjust the task priority in the task queue: Among them, P represents the task priority calculation function, T RT Indicates the real-time requirements of the task, D SIZE Indicates the amount of data, T LIMIT Indicates the time limit for completion; S33. Based on the output of the task priority calculation function, the task strategy is classified, and the classification includes local execution, push to the cloud and delayed processing. If the priority calculation function value is greater than the high threshold of the priority, the task is executed preferentially at the local edge node. If the priority calculation function value is less than the low threshold of the priority, the task is delayed. If the priority calculation function value is between the high threshold and the low threshold of the priority, the task is pushed to the cloud for processing.

5. The operation and maintenance method of a multi-cloud management platform based on cloud-edge collaboration according to claim 1 is characterized in that: The S4 specifically includes: S41. Construct a task feature matrix and a cloud edge resource matrix. The task features include the real-time nature of the task, computational complexity, data volume, task arrival time, and priority. The cloud edge resources include computing resources, storage capacity, network bandwidth, and resource availability. T={(C i ,T RT,i ,D SIZE,i ,P i ,T ARR,i )∣i=1,2,,N}; R={(E CPU,j ,E MEM,j ,B j ,L j )∣j=1,2,,M}; Among them, T represents the task feature matrix, R represents the cloud-edge resource matrix, and C i represents the computational complexity, T RT,i Indicates the real-time nature of the task, D SIZE,i Indicates the amount of data, P i Indicates the task priority, T ARR,i represents the task arrival time, E CPU,j represents computing resources, E MEM,j Indicates storage capacity, B j Indicates network bandwidth, L j represents resource availability, N represents the total number of tasks, and M represents the total number of edge nodes; S42. Establish a task and resource matching model and calculate the matching degree between tasks and resources: Among them, S(i,j) represents the matching function, w1, w2, w3, w4, w5 and w6 represent the weight coefficients of the matching function, indicating the influence of different task characteristics on the matching degree; S43. According to the value of the matching degree model, a matching relationship matrix S between tasks and resources is constructed: S={S(i,j)|i=1,2,,N; j=1,2,,M}; Among them, S(i,j) represents the matching degree between task index i and resource node index j.

6. The operation and maintenance method of a multi-cloud management platform based on cloud-edge collaboration according to claim 1 is characterized in that: The S5 specifically includes: S51. Sort each row in the matching relationship matrix S to find the best match between the current task and the resource node, that is, S(i,j) with the largest matching value: j * =argmax j S(i,j),j=1,2,…,n; Among them, j * Indicates the best resource node index matching the i-th task, argmax j S(i,j) means finding the index of the maximum value of S(i,j); S52. Constructing a task resource allocation table: A={(i,j * )|i=1,2,…,m}; Among them, A represents the task resource allocation table, (i,j * ) indicates that task i is assigned to resource node j * ; S53, for each pair of allocation relations (i, j * ), re-evaluate the rationality of the allocation by optimizing the objective function: O(i,j * )=λ1·P(i)+λ2·A r (j * )+λ3·C r (j * )+λ4·B r (j * ); Among them, O(i,j * ) represents the optimized allocation score, λ1, λ2, λ3 and λ4 represent the weight coefficients of the matching function, and represent the optimization weight coefficients, which are used to balance the influence of priority and resource attributes. r (j * ) represents the jth * The computing power of resource nodes, B r (j * ) represents the jth * The network bandwidth of resource nodes, A r (j * ) represents the jth * The current availability of resource nodes; S54, if the assignment relationship of task i (i, j * ) score is low, then rematch the resource nodes and look for feasible alternative nodes to satisfy: O(i,j′)>O(i,j * ); Among them, O(i,j') represents the new assigned score.

7. The operation and maintenance method of a multi-cloud management platform based on cloud-edge collaboration according to claim 1 is characterized in that: The S6 specifically includes: S61, monitor the running status of the edge node in real time, collect multi-dimensional monitoring data, the multi-dimensional monitoring data includes CPU usage, storage occupancy, network delay and task execution status, and pre-process the multi-dimensional monitoring data; S62. Use a gated temporal convolutional network in combination with an autoencoder to construct a time series anomaly detection model, wherein the gated temporal convolutional network includes causal convolution, dilated convolution, and a gating mechanism, and the autoencoder consists of an encoder and a decoder; S63. Use the gated temporal convolutional network to extract multi-scale temporal features from the pre-processed multi-dimensional monitoring data and output the temporal embedding features: H t =σ(W f *S t +b f )⊙tanh(W g *S t +b g ); Among them, H t represents the temporal embedding feature at time step t, W f and W g Represent the weights of the gate and main convolution kernel, b f and b g represents the corresponding bias term, σ represents the Sigmoid activation function, ⊙ represents the point-by-point product, S t represents the input sequence; S64, inputting the time series embedding features into an autoencoder, compressing the high-dimensional time series embedding features into low-dimensional potential representations, and decoding the low-dimensional potential representations back into reconstructed data of the original input dimension through a decoder of the autoencoder; S65. Compare the input data sequence and the reconstructed data generated by the autoencoder, and calculate the reconstruction error: Where E represents the reconstruction error, t represents the time step, and x i represents the input data sequence, represents the reconstructed data generated by the autoencoder; S66, comparing the reconstruction error with a preset threshold, and determining it as abnormal when the reconstruction error is greater than the preset threshold; S67. When an abnormal edge node is detected, an autonomous recovery mechanism is triggered according to the task queue and resource allocation status. The autonomous recovery mechanism includes task migration, resource switching and task reallocation. The task migration means migrating high-priority tasks on the abnormal node to other normal nodes. The resource switching means enabling redundant resource nodes to take over the tasks of the abnormal node. The task reallocation means recalculating the priority of unfinished tasks and reallocating them to suitable resource nodes.

8. An operation and maintenance system of a multi-cloud management platform based on cloud-edge collaboration, executing the operation and maintenance method of a multi-cloud management platform based on cloud-edge collaboration according to any one of claims 1 to 7, characterized in that: include: The cloud management module is used to manage the configuration, distributed deployment and resource association of cloud resources with edge nodes; The edge management module is used to monitor the resource status, task allocation and operation status of edge nodes; The task scheduling module is used to generate the matching relationship between tasks and resources based on task characteristics and resource status, and optimize the task resource allocation strategy; Edge autonomy module, which is used to realize task priority classification and dynamic policy adjustment of edge nodes by optimizing network and load prediction; Anomaly detection module, which is used to monitor the status of edge nodes in real time, detect anomalies and make relevant adjustments; Resource distribution map generation module, used to generate and synchronize cloud-edge collaborative resource distribution maps and display resource status; Task management module, which is used to record task life cycle data, update task status and support scheduling and recovery operations; The communication interface module is used to realize task data transmission and status feedback between the cloud and edge nodes.

Images