A high-efficiency network monitoring and scheduling system based on cloud-edge collaborative architecture

Abstract

The application discloses a kind of high efficiency network monitoring and scheduling systems based on cloud edge collaborative architecture, it is related to communication and video transmission technical field, and mutual authentication security communication channel based on equipment hardware security module and certificate system is established between the edge intelligent agent and the global arranger and carries out two-way communication;The edge intelligent agent is periodically or on demand according to the agreed data message format to the global arranger Report structured telemetry abstract, and the structured telemetry abstract includes timestamp, flow identification, source and destination IP and port, delay metric, queue depth and message count field.The application realizes structured telemetry abstract output by layered adaptive sampling and near-source abstract mechanism, solves the bandwidth storage pressure caused by the whole original data uplink, and the problem that key events are missed under low sampling rate, reduces the amount of uplink data, guarantees the data availability of key events and abnormal time, improves uplink efficiency and reduces cloud cost.

Description

Technical Field

[0001] This invention relates to the field of communication and video transmission technology, specifically to a high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture. Background Technology

[0002] With the rapid increase in the number of edge devices and video / IoT nodes, the amount of telemetry and media data generated by the network has grown significantly. Centralized uploading of all raw data to the cloud can lead to uplink bandwidth bottlenecks, increased cloud processing costs, and prolonged response latency. On the other hand, relying solely on isolated edge strategies lacks a holistic network perspective and makes it difficult to optimize scheduling in complex topologies and cross-domain failures. Furthermore, privacy compliance requirements for video and industrial data limit the large-scale uplink of raw data. Existing systems typically have shortcomings in sampling strategies, inference distribution, strategy generation, and audit credibility, making it impossible to simultaneously meet the multiple requirements of real-time performance, efficiency, global optimization, and compliance.

[0003] Patent CN111526061B discloses a monitoring traffic scheduling system and method for network range combat training scenarios. The patent enables more flexible monitoring traffic transmission, reduces monitoring traffic bandwidth load, and improves transmission efficiency.

[0004] The aforementioned patents reduce the bandwidth load of monitoring traffic and improve transmission efficiency, but privacy compliance requirements limit the large-scale uplink of raw data.

[0005] To this end, this application proposes a high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture that can achieve structured telemetry summary output. Summary of the Invention

[0006] The purpose of this invention is to provide a high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture, in order to solve the technical problems mentioned in the background art, such as the significant increase in the amount of telemetry and media data generated by the network, and the resulting uplink bandwidth bottleneck, increased cloud processing costs, and prolonged response latency caused by the centralized uploading of all raw data to the cloud.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture, the system comprising:

[0008] Multiple edge intelligent agents, each of which is connected to at least one data source device through an edge network interface, and each edge intelligent agent is equipped with a programmable data plane acquisition unit, a local preprocessing and compression unit, a lightweight local inference model module, a policy execution and reward recording unit, and a local immutable record area.

[0009] At least one cloud-based global orchestrator, which includes a telemetry aggregation database, a global model module, a policy generation module, an online validation module, and a federated learning coordination module;

[0010] The edge intelligent agent and the global orchestrator communicate bidirectionally through a secure communication channel established based on mutual authentication using the device hardware security module and certificate system.

[0011] The edge intelligent agent periodically or on demand reports a structured telemetry digest to the global orchestrator according to the agreed data packet format. The structured telemetry digest includes a timestamp, flow identifier, source and destination IP and port, latency metric, queue depth and packet count fields.

[0012] Preferably, the programmable data plane acquisition unit implements the inline insertion of telemetry tags by mounting a program on the data plane or a programmable switch program, and exports a structured telemetry message or summary containing the above fields in real time according to the data message format.

[0013] Preferably, the local preprocessing and compression unit includes: a traffic summarization module based on probability statistics, a distribution estimation module based on quantiles and a time series segmentation compression module, and a keyframe thumbnail module for video data. The keyframe thumbnail module outputs a thumbnail data structure containing region of interest hash and optical flow energy attributes. The thumbnail data structure is incorporated into the structured telemetry summary report with specified fields.

[0014] Preferably, the edge intelligent agent maintains an indicator vector locally, which consists of three values: queue depth, packet drop rate, and hotspot score. The edge intelligent agent switches between three sampling granularities based on a configurable threshold set and sampling window parameters. The three sampling granularities are: session-level digest, packet header-level digest, and payload key moment abbreviation. The threshold set and sampling window parameters are distributed by the global orchestrator in policy packets and can be adjusted locally by the edge intelligent agent according to rules.

[0015] Preferably, the lightweight local inference model module is a lightweight model obtained through distillation, and the model file size is limited to less than a predetermined size due to uplink bandwidth and local storage constraints. The global model module is jointly composed of a graph neural network of the basic network topology and a sequence model for temporal feature processing. The federated learning coordination module stipulates that each edge intelligent agent performs a limited number of training rounds locally and only reports the model summary or gradient difference after privacy protection processing. The global orchestrator performs secure aggregation on the reported information and then issues model update or distillation prompts.

[0016] Preferably, the policy generation module includes a model prediction module and a policy search-based scheduling module. The policy search-based scheduling module receives short-term traffic prediction vectors, queue depth vectors, edge-cloud resource availability vectors, and policy version identifiers with a defined input vector set and outputs a set of instructions defined as action vectors. The action vectors are defined in the action set as discrete action items, which include: rate limiting, path switching, transcoding level adjustment, stream fragment migration, and priority marking.

[0017] Preferably, when generating a policy package, the policy generation module parameterizes the reward function used as a set of scalar components and coefficients. The reward function is expressed as a linear combination of the scalar components. The scalar components include service level violation metrics, bandwidth consumption metrics, energy consumption metrics, and throughput metrics. The coefficient vector is distributed with the policy version and package and signed along with the policy package.

[0018] Preferably, the policy execution and reward recording unit implements a session-level flow granularity fragmentation mechanism. The fragmentation mechanism maps the flow to at least three types of sub-flow identifiers at the session layer: control class, important data class, and ordinary data class, and maintains a mapping table of sub-flows to transport bearers. The bearer protocol listed in the mapping table is at least one of multiplexing or multipath bearer protocols, and is issued as part of the policy packet.

[0019] Preferably, the global orchestrator further includes a security audit module. The security audit module stores policy packages, policy version records, and edge execution evidence digests in an append-only log structure. Before being distributed, the policy package is signed by the cloud and includes a policy version identifier, coefficient vector, mapping table, and rollback point identifier. When the edge intelligent agent executes the policy package, it records timestamped execution evidence and signs the execution evidence digest with the key of the local hardware security module before sending it back. The security audit module writes the execution evidence digest and policy digest into the configured verifiable ledger.

[0020] Preferably, the policy package includes an explicit rollback condition field for automatic rollback and a previous policy version summary field. After receiving the policy package, the edge intelligent agent stores the policy version, rollback point, and previous policy version summary in the local immutable record area. When the rollback condition in the form of a Boolean expression in the policy package is met, the agent automatically restores the previous configuration according to the rollback point. At the same time, the agent writes the recovery event and corresponding evidence into the immutable record area and sends it back to the global orchestrator for auditing.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. This invention achieves structured telemetry summary output through hierarchical adaptive sampling and near-source summarization mechanism, which solves the bandwidth and storage pressure caused by uploading all raw data, as well as the problem of missing key events at low sampling rates. It significantly reduces the amount of uploaded data, while ensuring the availability of data for key events and abnormal moments, improving upload efficiency and reducing cloud costs.

[0023] 2. This invention achieves secure collaborative updating of model weights between edge and cloud through edge-cloud hybrid inference and federated distillation update mechanism, which solves the contradiction between limited edge computing capabilities and lack of implementation response from the global perspective of the cloud, as well as the problem of centralized training under data privacy restrictions. Under the premise of protecting sensitive data, it achieves the coexistence of rapid response capability at the edge and global optimization capability in the cloud, improves detection accuracy and strategy effect, and reduces dependence on original data.

[0024] 3. This invention achieves rapid response and temporary handling by integrating telemetry within the programmable data plane with rapid micro-burst response, solving the problems of high statistical latency, inaccurate indicators, and difficulty in timely capture and handling of micro-bursts in the control plane and application layer, reducing the impact of link jitter on services, and improving system stability;

[0025] 4. This invention achieves automatic rollback and auditing through a policy engine based on predictive reinforcement learning and a decomposable and rollback-capable policy chain. It solves the problem that traditional rule-based or threshold policies are difficult to balance multiple objectives and lack an auditable rollback mechanism in complex cross-domain scenarios, improves scheduling intelligence, reduces deployment risks, and meets compliance and auditing requirements. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall system structure of the present invention;

[0027] Figure 2 This is a schematic diagram of the internal workflow of the edge intelligent agent module of the present invention;

[0028] Figure 3 This is a schematic diagram of the closed-loop process of cloud-based policy distribution and feedback in this invention;

[0029] Figure 4 This is a schematic diagram of the multipath migration and rollback timing of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figure 1 , Figure 2 , Figure 3 and Figure 4 The present invention provides an embodiment of a high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture, which is demonstrated by applying general deployment and data flow implementation.

[0032] Furthermore, an edge intelligent agent is deployed at each access point, and the edge agent is connected to the upstream switch via a gigabit or higher link. A global orchestrator is deployed in the cloud and connected to multiple edge agents.

[0033] Data plane acquisition: Deploy inline telemetry programs on the network stack or switch of the edge agent to insert and capture telemetry tags in the passing data packets and export a set of structured fields. The exported fields include at least timestamp, session identifier, source and destination addresses and ports, latency metric, queue depth and packet count. The acquisition program supports sampling triggered by probability or event.

[0034] Local preprocessing: The traffic is compressed using probabilistic summarization algorithms (e.g., segmented count minimization summary, quantile estimation), and segmented compression is used to preserve key inflection points in time series. Keyframes are extracted from video data, and region of interest hashes and optical flow energy are extracted from the keyframes as thumbnail information.

[0035] Lightweight Inference: Deploying distilled small sequence models for anomaly score calculation and short-term flow change forecasting within a sliding window;

[0036] Policy execution: After receiving the policy package from the cloud, you can implement rate limiting, scheduling to backup bearer, reducing transcoding level or adjusting priority for sessions or substreams locally, and record and sign execution evidence;

[0037] Telemetry aggregation: Periodically receive and add edge summaries to the database, merge summaries from multiple edges according to topological relationships, and fill in missing time-series information;

[0038] Global model: Extract the structural vectors of nodes and links based on the network topology, and input them along with temporal features into the joint model to generate policy parameters or distillation information;

[0039] Policy generation: Construct a policy package based on the model output, which includes fields such as policy version identifier, rollback point, mapping table (mapping of sub-streams to bearer protocols), reward function coefficient vector, and policy signature;

[0040] The edge reports a summary according to the configuration, the cloud generates and signs the policy package and distributes it, the edge executes and records the evidence and then sends back the execution summary, the cloud audit module writes the summary, policy package signature and execution evidence into the append log and writes them into the verifiable ledger as needed;

[0041] Example parameters: The sampling window is set to 10 seconds by default, which can be shortened to 2 seconds in high fluctuation mode. The threshold for determining high queue occupancy can be configured to 80%. The local inference latency of the lightweight model should be controlled within the range of 100 milliseconds to achieve rapid response.

[0042] Please see Figure 1 , Figure 2 , Figure 3 and Figure 4 The present invention provides an embodiment of a high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture, specifically implemented in a video surveillance scenario: in urban video surveillance, the video streams generated by a large number of cameras need to be monitored while ensuring the identification of key events and privacy compliance.

[0043] Furthermore, three-layer sampling is implemented for the video stream: under normal circumstances, only session-level summaries and keyframe summaries are uploaded. When a local abnormal score is detected to exceed the threshold, the edge uploads the thumbnail frames of key moments and the added message header summary to the cloud for detailed analysis. The keyframe thumbnail includes interpolation timestamps, ROI hashes, optical flow energy, and small visual feature vectors. All uploaded visual thumbnails are de-identified by the local privacy filtering module (e.g., blurring faces and deleting identifiable information) before being uploaded.

[0044] After receiving the abbreviated data, the cloud combines it with the time-series summary from the nearby edge and uses a topology-aware model to determine whether there is an anomaly in the link or upstream forwarding, and then issues the necessary stream migration or transcoding strategies. For areas with strict privacy or legal restrictions, the cloud only issues aggregation strategies and model distillation signals to avoid requiring the edge to upload the original video.

[0045] Typical process example: When a local frame packet loss increase is detected, the edge increases the sampling granularity and uploads the thumbnail frame. The cloud determines that the upstream link is congested and issues a strategy to migrate the ordinary sub-stream to the backup bearer and reduce the resolution of the ordinary sub-stream. The edge executes the strategy and reports the execution evidence. The cloud records the audit log and makes manual inspection suggestions when necessary.

[0046] Please see Figure 1 , Figure 2 , Figure 3 and Figure 4 The present invention provides an embodiment of a high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture, which is implemented in an industrial Internet of Things scenario and federated learning implementation: In industrial parks, distributed PLCs, sensors and high-definition inspection cameras require stable and reliable remote monitoring and scheduling, and production data is highly sensitive.

[0047] Furthermore, each edge intelligent agent performs several rounds of local training using locally collected telemetry summaries and limited label data to generate model update summaries. Before being reported, the update summaries are differentially privatized and signed by the local trusted execution environment. The cloud-based federated learning coordination module securely aggregates the update summaries from multiple edges to obtain new global models or distillation signals, which are then distributed to the corresponding edges.

[0048] Due to data sensitivity, the system specifies in the policy package which types of data must be processed locally and cannot be reported in the original payload. Based on this constraint, the cloud policy generation module generates a policy package and model distillation information that only contains reportable fields.

[0049] Example parameters: The number of local training rounds is usually limited to 1-3 rounds. The magnitude of differential privacy noise and the reporting frequency are issued by cloud policies and can be adjusted according to the compliance level.

[0050] Please see Figure 1 , Figure 2 , Figure 3 and Figure 4 The present invention provides an embodiment of a high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture, with an example of multi-tenant and cross-domain policy decomposition: when operators or cloud service providers provide edge cloud services to multiple tenants on the same physical infrastructure, it is necessary to isolate and prioritize resources based on tenant SLAs.

[0051] Furthermore, when generating global policies, the cloud first decomposes the global resource allocation target into tenant-level sub-targets, and then refines the tenant-level sub-targets into region-level and edge-level sub-policies and encapsulates them into policy packages for distribution.

[0052] Each policy package contains tenant identifier, policy priority, rollback conditions and mapping table. After receiving the policy, the edge performs priority conflict detection before executing it locally and adjusts it according to the conflict resolution rules defined in the policy package when a conflict occurs.

[0053] The resource mapping table explicitly specifies the mapping from sub-flows to bearers in the policy package, ensuring that high-priority tenants receive priority guarantees when resources are scarce, and mapping change operations must be recorded with audit evidence and returned after being signed.

[0054] Please see Figure 1 , Figure 2 , Figure 3 and Figure 4 The present invention provides an embodiment of a high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture. In terms of micro-burst detection and rapid migration implementation: short-term, high-intensity micro-burst traffic often occurs between the data center and the access network. It is necessary to detect the micro-burst traffic within a few hundred milliseconds and avoid link jitter in some local links to prevent service quality degradation.

[0055] Furthermore, the data plane acquisition program has micro-burst detection logic. When a sudden increase in queue depth and packet arrival rate is detected within the sliding window, the acquisition program immediately switches to high sampling mode and triggers abbreviated reporting at critical moments. At the same time, the lightweight inference model at the edge makes an online judgment on the cause of the burst and implements rate limiting or short-term priority adjustment locally. When local measures are insufficient, the edge reports the event and requests the cloud to issue a migration strategy.

[0056] After receiving the report, the cloud calculates alternative bearer paths based on the latest network topology information and issues sub-stream migration maps. The edge migrates sub-streams step by step according to the mapping table and sends execution evidence and migration progress back to the cloud during the migration process.

[0057] If the system metrics do not improve after migration or the rollback condition is triggered, the edge will restore the sub-stream to its previous bearer according to the rollback field in the policy package and record the recovery event.

[0058] Please see Figure 1 , Figure 2 , Figure 3 and Figure 4 The present invention provides an embodiment of a high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture, which combines cross-operator verifiable policies with a lightweight ledger. In scenarios involving multiple operators or regulators, it is necessary to achieve tamper-proof and multi-party verifiable evidence storage for policy issuance and execution.

[0059] Furthermore, after signing the policy package, the cloud writes the policy summary into a lightweight ledger maintained by multiple parties. After executing and signing the execution evidence summary, the edge sends the summary back to the cloud. The cloud then writes the execution evidence summary into the ledger or broadcasts it to multiple parties in the form of a ledger transaction. The ledger is only used to store summary and index information, and the actual amount of data that can be uploaded to the chain is limited to ensure latency and storage efficiency. The ledger adopts a lighter consensus mechanism to adapt to real-time requirements.

[0060] Regulators can verify the consistency between the ledger strategy and the execution summary; in the event of a dispute, they can further collect evidence based on the ledger records and the original logs in the cloud and at the edge.

[0061] Implementation details and interface specification examples:

[0062] Structured telemetry summary format: It is recommended to include the following fields: timestamp, stream / session identifier, source IP, destination IP, source port, destination port, short-term latency statistics (average, maximum, percentile), queue depth statistics, packet count, top-k hotspot stream summary, keyframe thumbnail pointer and model summary fingerprint;

[0063] Recommended policy package structure: A policy should include at least the following: policy version identifier, effective event, rollback point identifier, rollback conditions (expressed as Boolean expressions), reward function coefficient vector, sub-stream mapping table, signature field, and policy description metadata;

[0064] Communication and interface transmission examples: It is recommended to use secure bidirectional RPC or message bus to carry policy packets and telemetry summaries. The channel should use mutually authenticated transport layer security protocols and support batch and incremental reporting.

[0065] Security and key management; edge devices should load device keys and certificates into the hardware security module, and policy packages should use an asymmetric signature mechanism; execution evidence should be signed locally, sent back, and verified in the cloud.

[0066] Audit Logs and Evidence Storage: The audit module should persist the strategy package summary, execution evidence summary, and key transportation bureau in the form of append-only logs, and upload the summary to the blockchain or write it to a multi-party evidence storage system as needed.

[0067] This application is not limited to the above embodiments, and those skilled in the art can make substitutions or adjustments to the following aspects without departing from the spirit of this application:

[0068] The specific implementation of the data plane can be achieved using different technologies depending on the hardware capabilities (e.g., using only eBPF, or using P4 on nodes with programmable switches).

[0069] The specific network structure of the lightweight model and the global model can be adjusted according to the application scenario, and the parameters and number of training rounds can be adapted according to bandwidth and computing resource limitations.

[0070] The consensus mechanism, frequency of on-chain data uploads, and amount of on-chain data for auditing ledgers can be adjusted according to the needs and latency tolerance of the participants.

[0071] The policy rollback conditions, threshold set, and window length are all configurable parameters.

[0072] Working principle: High-precision telemetry is captured directly in the data plane of edge devices or edge agents, and the raw data is converted into structured telemetry summaries and visual thumbnails locally through hierarchical adaptive sampling and compression algorithms, so as to remove redundant and sensitive information as much as possible at the source;

[0073] The edge runs a lightweight inference model to perform short-term anomaly detection and traffic prediction on the local summary. When an anomaly or resource risk is detected, the edge executes the pre-defined sub-policy in real time to achieve a sub-second to hundreds of millisecond response and reports the execution result summary back to the cloud.

[0074] The cloud aggregates telemetry summaries from multiple edges and performs cross-domain root cause analysis and long-term policy training based on a topology-aware large model. The cloud generates policy packages with version and rollback information, which are then signed and distributed to the edge. After the policy is executed, the signed execution evidence is aggregated to form a verifiable audit chain, completing closed-loop collaboration.

[0075] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture, characterized in that: The system includes: Multiple edge intelligent agents are configured, each connected to at least one data source device via an edge network interface. Each edge intelligent agent includes a programmable data plane acquisition unit, a local preprocessing and compression unit, a lightweight local inference model module, a policy execution and reward recording unit, and a local immutable recording area. High-precision telemetry is directly captured in the data plane of the edge device or edge agent, and the raw data is converted into structured telemetry summaries and visual thumbnails locally through a hierarchical adaptive sampling and compression algorithm. The lightweight inference model runs at the edge to perform short-term anomaly detection and traffic prediction on the local summaries. The cloud aggregates telemetry summaries from multiple edges and performs cross-domain root cause analysis and long-term policy training based on a topology-aware large model, generating a policy package with version and rollback information in the cloud. At least one cloud-based global orchestrator, which includes a telemetry aggregation database, a global model module, a policy generation module, an online validation module, and a federated learning coordination module; The edge intelligent agent and the global orchestrator communicate bidirectionally through a secure communication channel established based on mutual authentication using the device hardware security module and certificate system. The edge intelligent agent periodically or on demand reports a structured telemetry summary to the global orchestrator according to the agreed data packet format. The structured telemetry summary includes timestamp, flow identifier, source and destination IP and port, latency metric, queue depth and packet count fields. The local preprocessing and compression unit includes: a traffic summarization module based on probability statistics, a distribution estimation module based on quantiles, and a time series segmentation compression module, as well as a keyframe thumbnail module for video data. The keyframe thumbnail module outputs a thumbnail data structure containing region of interest hash and optical flow energy attributes. The thumbnail data structure is incorporated into the structured telemetry summary report with specified fields.

2. The high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture according to claim 1, characterized in that: The programmable data plane acquisition unit implements inline insertion of telemetry tags by mounting a program on the data plane or a programmable switch program, and exports structured telemetry messages or summaries containing the above fields in real time according to the data message format.

3. The high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture according to claim 1, characterized in that: The edge intelligent agent maintains an indicator vector locally, which consists of three values: queue depth, packet drop rate, and hotspot score. The edge intelligent agent switches between three sampling granularities based on a configurable threshold set and sampling window parameters: session-level digest, packet header-level digest, and payload key moment abbreviation. The threshold set and sampling window parameters are distributed by the global orchestrator in policy packets and adjusted locally by the edge intelligent agent according to rules.

4. The high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture according to claim 3, characterized in that: The lightweight local inference model module is a lightweight model obtained through distillation. The model file size is limited to a predetermined size due to uplink bandwidth and local storage constraints. The global model module is jointly composed of a graph neural network of the basic network topology and a sequence model for temporal feature processing. The federated learning coordination module stipulates that each edge intelligent agent performs a limited number of training rounds locally and only reports the model summary or gradient difference after privacy protection processing. The global orchestrator performs secure aggregation on the reported information and then issues model update or distillation prompts.

5. A high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture as described in claim 4, characterized in that: The policy generation module includes a model prediction module and a policy search-based scheduling module. The policy search-based scheduling module receives short-term traffic prediction vectors, queue depth vectors, edge-cloud resource availability vectors, and policy version identifiers with a defined input vector set and outputs a set of instructions defined as action vectors. The action vectors are defined in the action set as discrete action items, which include: rate limiting, path switching, transcoding level adjustment, stream fragment migration, and priority marking.

6. A high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture as described in claim 5, characterized in that: When generating a policy package, the policy generation module parameterizes the reward function used as a set of scalar components and coefficients. The reward function is expressed as a linear combination of the scalar components. The scalar components include service level violation metrics, bandwidth consumption metrics, energy consumption metrics, and throughput metrics. The coefficient vector is distributed with the policy version and package and signed along with the policy package.

7. A high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture as described in claim 6, characterized in that: The policy execution and reward recording unit implements a session-level flow granularity fragmentation mechanism. The fragmentation mechanism maps the flow to at least three types of sub-flow identifiers at the session layer: control class, important data class, and ordinary data class, and maintains a mapping table of sub-flows to transport bearers. The bearer protocol listed in the mapping table is at least one of multiplexing or multipath bearer protocols, and is issued as part of the policy packet.

8. A high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture as described in claim 7, characterized in that: The global orchestrator also includes a security audit module, which stores policy packages, policy version records, and edge execution evidence digests in an append-only log structure. Before being distributed, the policy package is signed by the cloud and includes a policy version identifier, coefficient vector, mapping table, and rollback point identifier. When the edge intelligent agent executes the policy package, it records timestamped execution evidence and signs the execution evidence digest with the key of the local hardware security module before sending it back. The security audit module writes the execution evidence digest and policy digest into the configured verifiable ledger.

9. A high-performance network monitoring and scheduling system based on a cloud-edge collaborative architecture as described in claim 8, characterized in that: The policy package includes an explicit rollback condition field for automatic rollback and a previous policy version summary field. After receiving the policy package, the edge intelligent agent stores the policy version, rollback point, and previous policy version summary in the local immutable record area. When the rollback condition in the form of a Boolean expression in the policy package is met, the agent automatically restores the previous configuration according to the rollback point. At the same time, the agent writes the recovery event and corresponding evidence into the immutable record area and sends it back to the global orchestrator for auditing.

Images