An artificial intelligence-based cross-cloud network traffic intelligent scheduling and optimization system

Abstract

The application relates to the technical field of cloud computing and network communication, in particular to a cross-cloud network flow intelligent scheduling and optimization system based on artificial intelligence, which comprises a data sensing and collecting module, a heterogeneous data fusion and feature engineering module, a cross-cloud network digital twin construction and updating module, a multi-target collaborative intelligent decision module, a strategy lossless translation and distributed execution module and a feedback and model self-evolution module for realizing closed-loop optimization, which are connected in sequence. The application generates a unified network view by constructing an intelligent closed loop and fusing multi-source data, performs strategy deduction based on light-weight digital twin, generates a global optimization scheduling strategy meeting multi-target constraints by deep reinforcement learning and multi-agent collaboration, and finally executes the strategy on a heterogeneous cloud platform through strategy translation and distributed coordination and drives model self-optimization.

Description

Technical Field

[0001] This invention relates to the fields of cloud computing and network communication technology, and in particular to an intelligent scheduling and optimization system for cross-cloud network traffic based on artificial intelligence. Background Technology

[0002] As enterprises deepen their digital transformation, adopting multiple public, private, or hybrid cloud architectures has become the norm. While this cross-cloud deployment brings flexibility and avoids vendor lock-in, it also makes network traffic management exceptionally complex. Cross-cloud network traffic scheduling faces several challenges: First, the network architectures, monitoring interfaces, and control APIs of different cloud platforms are heterogeneous, making it difficult to obtain a globally unified network view; second, network states change dynamically, and traffic patterns are complex, making traditional scheduling strategies based on static rules or thresholds difficult to adapt in real time; third, optimization objectives are diverse (such as low latency, low cost, high availability, and load balancing) and may conflict with each other, requiring coordinated trade-offs.

[0003] Existing technical solutions suffer from the following shortcomings: 1) Most solutions are limited to a single cloud platform or the product ecosystem of a specific cloud service provider, lacking true cross-cloud heterogeneous collaboration capabilities. 2) The optimization methods employed are mostly heuristic rules or traditional operations research models, which have limited ability to handle high-dimensional, nonlinear, and real-time changing cross-cloud network scheduling problems, and cannot autonomously learn and evolve from historical data. 3) The modules of the system are tightly coupled, lacking an end-to-end intelligent closed loop from perception and decision-making to execution, as well as the ability for continuous model self-optimization.

[0004] Therefore, there is an urgent need for a system that can deeply perceive the state of cross-cloud networks, make intelligent decisions and automatically execute optimized scheduling, and has the ability to continuously learn and evolve. Summary of the Invention

[0005] To overcome the aforementioned deficiencies in the prior art, this invention provides an artificial intelligence-based intelligent scheduling and optimization system for cross-cloud network traffic, thereby addressing the problems existing in the background art.

[0006] This invention provides the following technical solution: an intelligent scheduling and optimization system for cross-cloud network traffic based on artificial intelligence, comprising: The data sensing and acquisition module is used to collect multi-dimensional raw operational data from multiple heterogeneous cloud platforms in real time. The heterogeneous data fusion and feature engineering module is connected to the data perception and acquisition module. It is used to standardize, align and fuse the multidimensional raw operating data, and extract a unified spatiotemporal feature vector for traffic scheduling. The cross-cloud network digital twin construction and update module is connected to the heterogeneous data fusion and feature engineering module, and is used to dynamically construct and continuously update a lightweight digital twin model that reflects the global real-time state of the cross-cloud network based on the unified spatiotemporal feature vector. The multi-objective collaborative intelligent decision-making module is connected to the cross-cloud network digital twin construction and update module. It is used to use the lightweight digital twin model as a simulation environment and output a globally optimal or near-optimal traffic scheduling strategy that meets the constraints of latency, cost, load balancing and service level agreement (SLA) based on deep reinforcement learning and multi-agent collaborative algorithm. The strategy lossless translation and distributed execution module is connected to the multi-objective collaborative intelligent decision-making module. It is used to losslessly translate the globally optimal or near-optimal traffic scheduling strategy into a set of control instructions natively supported by each heterogeneous cloud platform, and coordinate the distributed execution of each cloud platform. The closed-loop feedback and model self-evolution module is connected to the output end of the policy lossless translation and distributed execution module, as well as the feedback input end of the cross-cloud network digital twin construction and update module and the multi-objective collaborative intelligent decision-making module. It is used to collect actual network state data after policy execution, calculate policy performance indicators, and use online learning mechanisms to drive the continuous iterative optimization of the lightweight digital twin model and the algorithm model in the multi-objective collaborative intelligent decision-making module.

[0007] Furthermore, the data sensing and acquisition module includes: The cloud platform native API adaptation submodule is used to adapt to the monitoring and management APIs of different cloud service providers; The active probing submodule is used to inject customized probing traffic packets into cross-cloud network paths to measure end-to-end network performance metrics. The passive traffic sniffing submodule is deployed in the virtual network devices provided by the cloud service provider or in the tenant's virtual machine to collect network metadata flowing through it. The multidimensional raw operational data includes, but is not limited to: real-time topology of virtual networks within each cloud platform, link bandwidth utilization, port packet loss rate, transmission latency and jitter, resource utilization of virtual machines / containers, and cost and billing data and SLA compliance status of cross-cloud peering connections.

[0008] Furthermore, the processing performed by the heterogeneous data fusion and feature engineering module includes: Spatiotemporal alignment: Establishing a unified timestamp sequence and spatial topology mapping relationship for data sources from different acquisition frequencies and clock offsets; Normalization and encoding: Normalizing and embedding encoding of heterogeneous numerical, categorical and textual data; High-dimensional feature extraction: The attention mechanism is applied to process the fused data to automatically identify and extract key spatiotemporal features that are strongly related to traffic scheduling, and to form the unified spatiotemporal feature vector.

[0009] Furthermore, the cross-cloud network digital twin construction and update module is specifically used for: The lightweight digital twin model is constructed based on a graph neural network, where nodes represent cloud regions, availability zones, or critical network devices, and edges represent network links or logical connections. The attribute states of nodes and edges in the graph are updated in real time using the unified spatiotemporal feature vector. The lightweight digital twin model can deduce and predict the results of candidate traffic scheduling strategies at a speed much faster than that of a real network.

[0010] Furthermore, the multi-objective collaborative intelligent decision-making module includes: The policy network employs an actor-critic framework in deep reinforcement learning, where the actor network generates traffic scheduling actions based on the current state of the lightweight digital twin model, and the critic network evaluates the long-term multi-objective value of the actions. The multi-agent coordinator models key scheduling points (such as cloud egress gateways and load balancers) in cross-cloud networks as cooperative agents, and they collaborate through a decentralized partially observable Markov decision process model and communication mechanism. The multi-objective optimizer integrates multiple optimization objectives, such as latency, cost, load balancing, and SLA compliance, into a composite reward function through an adaptive weighting algorithm, which is used to train the policy network.

[0011] Furthermore, the strategy lossless translation and distributed execution module includes: The strategy parsing submodule is used to parse the globally optimal or near-optimal traffic scheduling strategy and decompose it into a sequence of atomic operations for specific cloud platforms and network devices. The instruction translation submodule has a built-in template library of native control instructions for various cloud platforms, which maps the atomic operation sequence and generates executable native API calls or configuration commands. The distributed coordination and execution submodule adopts a two-phase commit protocol to coordinate the atomic execution of the control instructions across cloud platforms, ensuring the consistency of cross-cloud scheduling transactions.

[0012] Furthermore, the closed-loop feedback and model self-evolution module includes: The performance evaluation submodule is used to compare the actual network state after policy execution with the predicted state of the digital twin, and to calculate the prediction bias and actual benefits of the policy. An online learning engine, based on the prediction bias and actual gains, uses online gradient descent or evolutionary strategy algorithms to fine-tune the parameters of the lightweight digital twin model and the weights of the policy network in the multi-objective collaborative intelligent decision-making module. The model version management submodule seamlessly switches to a new model version when online learning leads to continuous improvement in model performance; and automatically rolls back to a stable version when performance deteriorates.

[0013] Furthermore, the system also includes a security and compliance barrier module. This module performs security audits and compliance checks on the strategy after the multi-objective collaborative intelligent decision-making module generates the strategy and before the strategy lossless translation and distributed execution module executes it, to ensure that the strategy does not violate preset security policies and compliance rules.

[0014] The technical effects and advantages of this invention are as follows: By fusing multi-source data, the problem of cross-cloud network state visibility is solved, providing high-quality and unified input for intelligent decision-making. Millisecond-level policy simulation using lightweight digital twins significantly reduces the cost and risk of trial and error in real networks. Combining deep reinforcement learning and multi-agent collaboration, it can adaptively handle dynamic optimization problems with multiple objectives and constraints, achieving decision quality far exceeding static rules. Through lossless policy translation and distributed coordination, automated and consistent implementation of high-level intelligent policies down to low-level heterogeneous control instructions is achieved. Closed-loop feedback and online learning enable the system to autonomously evolve with changes in network environment and business needs, maintaining long-term optimized performance. Mechanisms such as security and compliance verification, model version management, and execution transaction coordination ensure the reliability and compliance of system operation. Attached Figure Description

[0015] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0017] Please see Figure 1 As shown, an artificial intelligence-based cross-cloud network traffic intelligent scheduling and optimization system includes: The data sensing and acquisition module is used to collect multi-dimensional raw operational data from multiple heterogeneous cloud platforms in real time. The heterogeneous data fusion and feature engineering module is connected to the data perception and acquisition module. It is used to standardize, align and fuse the multidimensional raw operating data, and extract a unified spatiotemporal feature vector for traffic scheduling. The cross-cloud network digital twin construction and update module is connected to the heterogeneous data fusion and feature engineering module, and is used to dynamically construct and continuously update a lightweight digital twin model that reflects the global real-time state of the cross-cloud network based on the unified spatiotemporal feature vector. The multi-objective collaborative intelligent decision-making module is connected to the cross-cloud network digital twin construction and update module. It is used to use the lightweight digital twin model as a simulation environment and output a globally optimal or near-optimal traffic scheduling strategy that meets the constraints of latency, cost, load balancing and service level agreement (SLA) based on deep reinforcement learning and multi-agent collaborative algorithm. The strategy lossless translation and distributed execution module is connected to the multi-objective collaborative intelligent decision-making module. It is used to losslessly translate the globally optimal or near-optimal traffic scheduling strategy into a set of control instructions natively supported by each heterogeneous cloud platform, and coordinate the distributed execution of each cloud platform. The closed-loop feedback and model self-evolution module is connected to the output end of the policy lossless translation and distributed execution module, as well as the feedback input end of the cross-cloud network digital twin construction and update module and the multi-objective collaborative intelligent decision-making module. It is used to collect actual network state data after policy execution, calculate policy performance indicators, and use online learning mechanisms to drive the continuous iterative optimization of the lightweight digital twin model and the algorithm model in the multi-objective collaborative intelligent decision-making module. Detailed implementation method: The system is deployed in a software-defined manner within a separate control plane. Its workflow is as follows: Perception Phase: The data perception and acquisition module periodically (e.g., every 30 seconds) collects raw data such as virtual machine CPU / memory usage, virtual switch throughput, peer-to-peer connection latency and bandwidth utilization by agents deployed on cloud platforms such as Alibaba Cloud, AWS, and Tencent Cloud or by directly calling their cloud monitoring APIs.

[0019] Fusion and Representation Phase: The heterogeneous data fusion and feature engineering module unifies the above data into a time series based on UTC time and maps all resources onto a global network topology map. Through normalization and feature extraction, a unified feature vector is generated containing information such as "predicted link latency from cloud region A to B = 45ms, available bandwidth = 1Gbps, and unit traffic cost = 0.05 yuan / GB".

[0020] Twin Construction Phase: The cross-cloud network digital twin construction and update module receives the feature vector and updates a lightweight graph model maintained internally. This model can simulate the routing process of data packets from source to destination and predict its performance metrics.

[0021] Intelligent Decision-Making Phase: When scheduling a new batch of traffic (such as user requests to access a cross-cloud deployed service from Shanghai), the multi-objective collaborative intelligent decision-making module takes the current state of the digital twin as input. Its internal deep reinforcement learning agent rapidly simulates thousands of possible path choices within the twin (such as traversing Alibaba Cloud's intranet to Tencent Cloud, or via a public dedicated line), evaluating the latency, cost, and load of each path in the simulation. Ultimately, it outputs a scheduling strategy with the highest overall score, such as "allocating 80% of the traffic to path X and 20% to path Y."

[0022] Strategy execution phase: The strategy lossless translation and distributed execution module translates the above percentage strategy into a series of specific instructions, such as API calls to modify weights in Alibaba Cloud Load Balancer CLB and API calls to add entries to Tencent Cloud routing tables, and coordinates the near-simultaneous issuance and execution of these instructions.

[0023] Feedback and Evolution Phase: After a period of time (e.g., 5 minutes) following policy execution, the closed-loop feedback and model self-evolution module collects actual network data again, calculating the actual average latency to be 48ms (slightly higher than the twin's prediction of 45ms). This deviation, along with benefits such as cost savings, constitutes the training signal. Online learning algorithms (e.g., stochastic gradient descent with momentum) are used to fine-tune the prediction model of the digital twin and the policy network parameters of the decision-making agent, making their subsequent predictions and decisions more accurate.

[0024] Through the closed-loop collaboration of the above six modules, the system achieves continuous adaptive optimization of cross-cloud traffic.

[0025] Furthermore, the data sensing and acquisition module includes: The cloud platform native API adaptation submodule is used to adapt to the monitoring and management APIs of different cloud service providers; The active probing submodule is used to inject customized probing traffic packets into cross-cloud network paths to measure end-to-end network performance metrics. The passive traffic sniffing submodule is deployed in the virtual network devices provided by the cloud service provider or in the tenant's virtual machine to collect network metadata flowing through it. The multidimensional raw operational data includes, but is not limited to: real-time topology of virtual networks within each cloud platform, link bandwidth utilization, port packet loss rate, transmission latency and jitter, resource utilization of virtual machines / containers, and cost and billing data and SLA compliance status of cross-cloud peering connections. Detailed implementation method: The cloud platform native API adaptation submodule encapsulates APIs from different cloud vendors. For example, for AWS, it calls the CloudWatch API's GetMetricData to retrieve NetworkIn and NetworkOut metrics; for Alibaba Cloud, it calls the Cms API's DescribeMetricList to retrieve network traffic data for ECS instances. This submodule unifies authentication (such as AK / SK), request frequency limits, and error retry logic for different APIs.

[0027] The proactive probing submodule periodically (e.g., every minute) sends ICMP or TCP probe packets from lightweight probe agents deployed in various cloud availability zones to peer agents in other cloud availability zones. By calculating round-trip time (RTT) and packet loss rate, the true performance of the inter-cloud link is directly measured. Probe packet size and frequency are configurable to avoid significant impact on production traffic.

[0028] The passive traffic sniffing submodule is deployed as a DaemonSet on each node of the Kubernetes cluster, or installed as an image in a virtual machine. It uses libpcap or eBPF technology to capture network packet header information (five-tuple, packet size, timestamp) flowing through the local machine, aggregates it (without involving payload to protect privacy), and reports it for analyzing micro-traffic matrices and identifying "elephant flows".

[0029] The data collected by these sub-modules is aggregated to form a structured document. Metrics (such as {metric:'cpu_util', value:65, provider:'aws', region:'us-east-1', resource_id:'i-123456', timestamp:1625097600}) and raw data pool of unstructured logs.

[0030] Furthermore, the processing performed by the heterogeneous data fusion and feature engineering module includes: Spatiotemporal alignment: Establishing a unified timestamp sequence and spatial topology mapping relationship for data sources from different acquisition frequencies and clock offsets; Normalization and encoding: Normalizing and embedding encoding of heterogeneous numerical, categorical and textual data; High-dimensional feature extraction: The attention mechanism is applied to process the fused data to automatically identify and extract key spatiotemporal features that are strongly related to traffic scheduling, and to form the unified spatiotemporal feature vector. Detailed implementation method: The specific operation of spatiotemporal alignment is as follows: The module maintains a global topology graph with network devices (such as virtual routers and load balancers) and links as nodes. All raw data with timestamps is interpolated or aggregated into a unified fixed time window (such as one point every 10 seconds). For example, AWS's 1-minute granular data and Alibaba Cloud's 30-second granular data are both generated into 10-second granular sequence data through linear interpolation and mounted onto the corresponding nodes or edges in the topology graph.

[0032] Normalization and Encoding: For numerical data (such as bandwidth utilization), max-min normalization is used to scale to the [0,1] interval. For categorical data (such as cloud provider type: AWS, Alibaba), one-hot encoding or learned embedding vectors are used for representation. For textual SLA descriptions (such as "availability >= 99.95%), sentence vectors from a pre-trained natural language processing model (such as BERT) are used as encoding.

[0033] High-dimensional feature extraction: The spatiotemporally aligned and encoded data is organized into a three-dimensional tensor (node ​​× feature × time step) by time window. This tensor is input into an encoder based on an attention mechanism (e.g., a TransformerEncoder layer). The attention mechanism can automatically learn which nodes (e.g., a high-cost dedicated line) and which historical time points are more important under a specific scheduling objective (e.g., the current main challenge is cost). The final output of the encoder is flattened and dimensionality reduced to form a fixed-length unified feature vector rich in spatiotemporal correlation information, which is then used by downstream modules.

[0034] Furthermore, the cross-cloud network digital twin construction and update module is specifically used for: The lightweight digital twin model is constructed based on a graph neural network, where nodes represent cloud regions, availability zones, or critical network devices, and edges represent network links or logical connections. The attribute states of nodes and edges in the graph are updated in real time using the unified spatiotemporal feature vector. The lightweight digital twin model can deduce and predict the results of candidate traffic scheduling strategies at a speed much faster than that of a real network. Detailed implementation method: The digital twin model constructed by this module is essentially a parameterized graph neural network model. The node set V of the graph includes: each cloud region (e.g., ap-southeast-1), each availability zone (e.g., us-east-1a), and core virtual network devices (e.g., TGW). The edge set E includes: links between availability zones within the cloud, cross-cloud peering connections, and VPN tunnels.

[0036] Each node v has an attribute vector h_v, such as [CPU utilization, memory utilization, status (healthy / abnormal)]. Each edge e has an attribute vector h_e, such as [bandwidth utilization, current latency, packet loss rate, unit cost]. These attribute values ​​are derived from the unified spatiotemporal feature vector output by the feature engineering module.

[0037] At the heart of the model is a message-passing graph neural network. Its forward propagation process simulates the diffusion of network state effects: for example, an increase in the load of a node (virtual machine) will affect the queue latency of its connected edges (virtual switch ports), which in turn affects downstream nodes. Through several layers of GNN computation, the model can output predictions of network performance (such as end-to-end latency and available bandwidth) between any pair of nodes.

[0038] "Lightweight" is reflected in the model size (fewer layers, moderate number of parameters) and inference speed. When inferring a candidate scheduling strategy (such as "directing traffic from node A to node B"), only the strategy needs to be used as input (changing the "traffic allocation" attribute of the relevant edges in the graph), and a single GNN forward propagation (millisecond level) can be run to predict the steady-state performance index of the entire network under the new traffic distribution, without the need for minute-level real network simulation or waiting for real feedback.

[0039] Furthermore, the multi-objective collaborative intelligent decision-making module includes: The policy network employs an actor-critic framework in deep reinforcement learning, where the actor network generates traffic scheduling actions based on the current state of the lightweight digital twin model, and the critic network evaluates the long-term multi-objective value of the actions. The multi-agent coordinator models key scheduling points (such as cloud egress gateways and load balancers) in cross-cloud networks as cooperative agents, and they collaborate through a decentralized partially observable Markov decision process model and communication mechanism. The multi-objective optimizer integrates multiple optimization objectives, such as latency, cost, load balancing, and SLA compliance, into a composite reward function through an adaptive weighting algorithm, which is used to train the policy network. Detailed implementation method: The Actor-Critic Framework: The ActorNetwork is a neural network that takes the global state S_t (i.e., the aforementioned unified feature vector) as input to the digital twin representation and outputs a probability distribution π(a|S_t), representing the probability of choosing various scheduling actions (such as "choosing path 1 for service A" or "adjusting weights for service B") in the current state. The CriticNetwork is another neural network that takes the state S_t and action a_t as input and outputs a scalar value V(S_t, a_t), representing the long-term expected reward of that action.

[0041] Multi-Agent Coordinator: Each cloud egress gateway or global load balancer is defined as an agent. Each agent can only observe a portion of the network's state (partial observability, POMDP). They share actor and critic network parameters. During decision-making, each agent derives its action from the actor network based on its local observations. To coordinate, agents exchange key information (such as "my side is congested") through a lightweight communication channel (e.g., sharing a context vector generated by an attention mechanism), thereby adjusting their actions to achieve global coordination.

[0042] Multi-objective optimizer: Define a composite reward function R_t = w1*R_latency + w2*R_cost + w3*R_balance + w4*R_sla. Here, R_latency is negatively correlated with prediction latency, R_cost is negatively correlated with traffic cost, R_balance is negatively correlated with the variance of load on each link, and R_sla is a large negative value when the SLA is violated. The weights w1-w4 are not fixed but dynamically adjusted based on business priorities and historical optimization results. For example, when the cost budget is tight at the end of the month, the system automatically increases the weight of w2. This reward function R_t is used to calculate the target value of the critic network and simultaneously update the actor and critic networks using a policy gradient algorithm, ultimately training a scheduling policy that balances multiple objectives.

[0043] Furthermore, the strategy lossless translation and distributed execution module includes: The strategy parsing submodule is used to parse the globally optimal or near-optimal traffic scheduling strategy and decompose it into a sequence of atomic operations for specific cloud platforms and network devices. The instruction translation submodule has a built-in template library of native control instructions for various cloud platforms, which maps the atomic operation sequence and generates executable native API calls or configuration commands. The distributed coordination and execution submodule adopts a two-phase commit protocol to coordinate the atomic execution of the control instructions across cloud platforms, ensuring the consistency of cross-cloud scheduling transactions. Detailed implementation method: The decision policy received by the policy parsing submodule may be an abstract instruction, such as: {"service":"web-frontend","traffic_distribution":{"path_A_via_AWS":0.7","path_B_via_Aliyun":0.3}}. The parser, based on a predefined knowledge graph (recording which VPCs the web-frontend's backend cluster resides in on AWS and Alibaba Cloud, and the traffic entry point), decomposes it into a sequence of atomic operations, for example: Op1: On AWSRoute53, set the weight of the endpoint pointing to the AWS cluster in the weighted routing policy for web-frontend.example.com to 70.

[0045] Op2: In the weighted routing policy of web-frontend.example.com on Alibaba Cloud DNS, set the weight of the endpoint pointing to the Alibaba Cloud cluster to 30.

[0046] The instruction translation submodule provides pre-defined API call templates for each atomic operation and cloud platform combination. For example, for Op1, it populates an AWSSDKChangeResourceRecordSets request template and generates a specific JSON request body. For Op2, it populates an Alibaba Cloud SDKUpdateDNSSLBWeight request template.

[0047] The distributed coordination and execution submodule acts as the "coordinator." It first sends a "ready" request, along with the specific API instructions, to all involved executors (clients responsible for calling AWS APIs and Alibaba Cloud APIs). Each executor checks the validity of the instructions and reserves resources (such as confirming sufficient permissions), then replies "ready." Once all executors have replied "ready," the coordinator sends a "commit" command, and all executors simultaneously execute the final API call. If any executor fails during the "ready" phase, the coordinator sends a "rollback" command to all executors, canceling the reservations and ensuring no partially effective policies remain, thus guaranteeing the atomicity of cross-cloud operations.

[0048] Furthermore, the closed-loop feedback and model self-evolution module includes: The performance evaluation submodule is used to compare the actual network state after policy execution with the predicted state of the digital twin, and to calculate the prediction bias and actual benefits of the policy. An online learning engine, based on the prediction bias and actual gains, uses online gradient descent or evolutionary strategy algorithms to fine-tune the parameters of the lightweight digital twin model and the weights of the policy network in the multi-objective collaborative intelligent decision-making module. The model version management submodule seamlessly switches to a new model version when online learning leads to continuous improvement in model performance; and automatically rolls back to a stable version when performance deteriorates. Detailed implementation method: At the end of an evaluation cycle (e.g., 10 minutes) after policy execution, the performance evaluation submodule obtains the actual measured average latency L_actual and actual cost C_actual from the perception module. Simultaneously, it obtains the predicted values ​​L_pred and C_pred for the same time period before policy execution from the digital twin module. The prediction bias is then calculated. (e.g., MSE=(L_actual-L_pred)^2) And actual benefits (e.g., R_actual=-(w1*L_actual+w2*C_actual)).

[0050] The online learning engine contains two parallel learning loops: The twin model update loop uses (state S_t, predicted action a_t, actual next state S_{t+1_actual}) as training samples, and uses the prediction bias as part of the loss function to update the parameters of the digital twin graph neural network through online gradient descent, making its predictions more accurate.

[0051] The decision model update loop: (state S_t, action a_t, actual reward R_actual, actual new state S_{t+1_actual}) is used as reinforcement learning experience samples and stored in the experience replay pool. Periodically, samples are taken from the pool, and the TD error is calculated using the actual reward R_actual (not the reward predicted by the twin) to update the weights of the actor and critic networks in the multi-objective collaborative intelligent decision module.

[0052] The model version management submodule monitors model performance metrics on the validation set (historical data) during online learning. When the comprehensive evaluation metrics of the new parameter model (such as improved prediction accuracy and increased decision benefits) exceed a certain threshold of the old model and stabilize for a period of time, a version switch is triggered: the new model parameter file is loaded into memory, replacing the old model currently in service. This process is unaffected by the scheduling service. If real-time monitoring detects an abnormal increase in key metrics (such as SLA violation rate) after the switch, the system automatically rolls back to the previous stable version and issues an alert.

[0053] Furthermore, the system also includes a security and compliance barrier module. This module performs security audits and compliance checks on the strategy after the multi-objective collaborative intelligent decision-making module generates the strategy and before the strategy lossless translation and distributed execution module executes it, to ensure that the strategy does not violate preset security policies and compliance rules. Detailed implementation method: The security and compliance barrier module serves as the final mandatory checkpoint before policy execution. It maintains a set of rule engines.

[0055] Security Audit: This checks whether policies would cause traffic to bypass audited paths or route traffic from sensitive data areas to cloud areas with lower security levels. For example, a rule might stipulate that "traffic from a financial database VPC must not be transmitted over the public Internet." If the policy generated by the decision module suggests such a path, this module will directly reject the policy and provide feedback to the decision module for recalculation.

[0056] Compliance Verification: This checks whether the policy complies with data sovereignty regulations. For example, a rule might stipulate that "data traffic from EU users (geo_ip=EU) must always terminate in the eu-central-1 or eu-west-1 region." This verifier analyzes the mapping between the user IP ranges that the policy might affect and the target cloud regions, and blocks execution if a violation is found.

[0057] This module is typically implemented as a "plugin," and the rules can be dynamically configured. It provides explicit outputs such as "pass," "reject and alert," and "reject and return suggestions," ensuring that the automated system, while pursuing optimal performance and cost, does not cross the line of safety and compliance.

[0058] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent scheduling and optimization system for cross-cloud network traffic based on artificial intelligence, characterized in that, include: The data sensing and acquisition module is used to collect multi-dimensional raw operational data from multiple heterogeneous cloud platforms in real time. The heterogeneous data fusion and feature engineering module is connected to the data perception and acquisition module. It is used to standardize, align and fuse the multidimensional raw operating data, and extract a unified spatiotemporal feature vector for traffic scheduling. The cross-cloud network digital twin construction and update module is connected to the heterogeneous data fusion and feature engineering module, and is used to dynamically construct and continuously update a lightweight digital twin model that reflects the global real-time state of the cross-cloud network based on the unified spatiotemporal feature vector. The multi-objective collaborative intelligent decision-making module is connected to the cross-cloud network digital twin construction and update module. It is used to use the lightweight digital twin model as a simulation environment and output a globally optimal or near-optimal traffic scheduling strategy that meets the constraints of latency, cost, load balancing and service level agreement (SLA) based on deep reinforcement learning and multi-agent collaborative algorithm. The strategy lossless translation and distributed execution module is connected to the multi-objective collaborative intelligent decision-making module. It is used to losslessly translate the globally optimal or near-optimal traffic scheduling strategy into a set of control instructions natively supported by each heterogeneous cloud platform, and coordinate the distributed execution of each cloud platform. The closed-loop feedback and model self-evolution module is connected to the output end of the policy lossless translation and distributed execution module, as well as the feedback input end of the cross-cloud network digital twin construction and update module and the multi-objective collaborative intelligent decision-making module. It is used to collect actual network state data after policy execution, calculate policy performance indicators, and use online learning mechanisms to drive the continuous iterative optimization of the lightweight digital twin model and the algorithm model in the multi-objective collaborative intelligent decision-making module.

2. The system according to claim 1, characterized in that, The data sensing and acquisition module includes: The cloud platform native API adaptation submodule is used to adapt to the monitoring and management APIs of different cloud service providers; The active probing submodule is used to inject customized probing traffic packets into cross-cloud network paths to measure end-to-end network performance metrics. The passive traffic sniffing submodule is deployed in the virtual network devices provided by the cloud service provider or in the tenant's virtual machine to collect network metadata flowing through it. The multidimensional raw operational data includes, but is not limited to: real-time topology of virtual networks within each cloud platform, link bandwidth utilization, port packet loss rate, transmission latency and jitter, resource utilization of virtual machines / containers, and cost and billing data and SLA compliance status of cross-cloud peering connections.

3. The system according to claim 1, characterized in that, The processing performed by the heterogeneous data fusion and feature engineering module includes: Spatiotemporal alignment: Establishing a unified timestamp sequence and spatial topology mapping relationship for data sources from different acquisition frequencies and clock offsets; Normalization and encoding: Normalizing and embedding encoding of heterogeneous numerical, categorical and textual data; High-dimensional feature extraction: The attention mechanism is applied to process the fused data to automatically identify and extract key spatiotemporal features that are strongly related to traffic scheduling, and to form the unified spatiotemporal feature vector.

4. The system according to claim 1, characterized in that, The cross-cloud network digital twin construction and update module is specifically used for: The lightweight digital twin model is constructed based on a graph neural network, where nodes represent cloud regions, availability zones, or critical network devices, and edges represent network links or logical connections. The attribute states of nodes and edges in the graph are updated in real time using the unified spatiotemporal feature vector. The lightweight digital twin model can deduce and predict the results of candidate traffic scheduling strategies at a speed much faster than that of a real network.

5. The system according to claim 1, characterized in that, The multi-objective collaborative intelligent decision-making module includes: The policy network employs an actor-critic framework in deep reinforcement learning, where the actor network generates traffic scheduling actions based on the current state of the lightweight digital twin model, and the critic network evaluates the long-term multi-objective value of the actions. The multi-agent coordinator models key scheduling points (such as cloud egress gateways and load balancers) in cross-cloud networks as cooperative agents, and they collaborate through a decentralized partially observable Markov decision process model and communication mechanism. The multi-objective optimizer integrates multiple optimization objectives, such as latency, cost, load balancing, and SLA compliance, into a composite reward function through an adaptive weighting algorithm, which is used to train the policy network.

6. The system according to claim 1, characterized in that, The strategy lossless translation and distributed execution module includes: The strategy parsing submodule is used to parse the globally optimal or near-optimal traffic scheduling strategy and decompose it into a sequence of atomic operations for specific cloud platforms and network devices. The instruction translation submodule has a built-in template library of native control instructions for various cloud platforms, which maps the atomic operation sequence and generates executable native API calls or configuration commands. The distributed coordination and execution submodule adopts a two-phase commit protocol to coordinate the atomic execution of the control instructions across cloud platforms, ensuring the consistency of cross-cloud scheduling transactions.

7. The system according to claim 1, characterized in that, The closed-loop feedback and model self-evolution module includes: The performance evaluation submodule is used to compare the actual network state after policy execution with the predicted state of the digital twin, and to calculate the prediction bias and actual benefits of the policy. An online learning engine, based on the prediction bias and actual gains, uses online gradient descent or evolutionary strategy algorithms to fine-tune the parameters of the lightweight digital twin model and the weights of the policy network in the multi-objective collaborative intelligent decision-making module. The model version management submodule seamlessly switches to a new model version when online learning leads to continuous improvement in model performance; and automatically rolls back to a stable version when performance deteriorates.

8. The system according to any one of claims 1-7, characterized in that, The system also includes a security and compliance barrier module. This module performs security audits and compliance checks on the strategy after the multi-objective collaborative intelligent decision-making module generates the strategy and before the strategy lossless translation and distributed execution module executes it, to ensure that the strategy does not violate preset security policies and compliance rules.

Images