Edge network acceleration system and method based on AI intelligent scheduling

Abstract

The invention relates to the technical field of edge computing, and relates to an edge network acceleration system and method based on AI intelligent scheduling, and the system comprises an edge node sensing module, an AI intelligent scheduling decision module, a dynamic resource distribution module, a multipath data acceleration module, a network quality monitoring module, an edge and cloud collaboration module, and an AI model optimization module. Through AI dynamic scheduling and heterogeneous resource allocation, the CPU utilization rate of the edge nodes is improved, the bandwidth utilization rate is improved, and resource waste is reduced; lSTM prediction and QUIC protocol optimization are selected through multiple paths, real-time service delay is reduced, and the millisecond-level requirements of the Internet of Vehicles and industrial control are met; aI redundant data filtering and dynamic resource application, edge and cloud data transmission quantity reduction, cloud bandwidth occupation reduction, model pruning, quantification, incremental training, AI scheduling model volume reduction, reasoning computing power consumption reduction, low computing power hardware QoS real-time monitoring and abnormity alarm adaptive to edge nodes, network fault recovery time shortening and service interruption rate reduction are realized.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and in particular to an edge network acceleration system and method based on AI intelligent scheduling. Background Technology

[0002] With the popularization of 5G and IoT technologies, edge computing has become a core solution for reducing network latency due to its "nearby processing" characteristic. However, current edge networks face the following technical challenges: Traditional edge networks employ fixed resource allocation strategies, such as static bandwidth allocation and equal computing power distribution, which makes them unable to adapt to dynamic loads, resulting in low resource utilization. Single transmission paths are susceptible to link congestion and lack AI-driven path prediction and optimization, making it difficult to meet the millisecond-level latency requirements of applications such as vehicle networking and industrial control. Data interaction between edge nodes and the cloud lacks dynamic adaptation, with high-frequency redundant data consuming a large amount of bandwidth, and AI model deployment suffers inference latency due to edge computing power limitations. Traditional QoS monitoring relies solely on post-event data statistics and cannot predict network congestion in advance, resulting in long fault response times and impacting business continuity.

[0003] Therefore, there is an urgent need for an edge network acceleration solution that integrates AI technology to solve the above problems through intelligent scheduling, dynamic optimization, and collaborative management. Summary of the Invention

[0004] The purpose of this invention is to address the problems existing in the background technology by proposing an edge network acceleration system and method based on AI intelligent scheduling.

[0005] The technical solution of the present invention, in its first aspect, provides an edge network acceleration system based on AI intelligent scheduling, comprising: The edge node perception module is deployed at each edge node to acquire the resource status of the edge node hardware, network link parameters and business demand data in real time, and outputs the dimensionality-reduced key feature vector to the AI ​​intelligent scheduling decision module. The AI ​​intelligent scheduling and decision-making module receives the feature vectors output by the edge node perception module and outputs computing power resource allocation schemes, bandwidth allocation ratios and multi-path transmission strategies. The dynamic resource allocation module dynamically allocates CPU, GPU, and bandwidth resources based on the decisions output by the AI ​​scheduling decision module. The multi-path data acceleration module performs multi-path selection, data fragmentation and caching acceleration based on the decisions output by the AI ​​scheduling decision module; The network quality monitoring module monitors the QoS indicators of the dynamic resource allocation module and the multi-path data acceleration module in real time. If an anomaly is detected, an alarm is triggered, and the alarm result is fed back to the AI ​​intelligent scheduling decision module. The edge and cloud collaboration module, based on the decisions output by the AI ​​scheduling and decision-making module, requests GPU computing power from the cloud when overloaded; The AI ​​model optimization module uses model pruning, quantization, and incremental training to lightweight and iterate the performance of the AI ​​scheduling model, reducing edge computing power consumption.

[0006] Preferably, the edge node perception module specifically includes a multi-dimensional data acquisition unit, a data preprocessing unit, and a feature extraction unit; The multi-dimensional data acquisition unit collects edge node CPU utilization, memory usage, GPU load, link bandwidth, latency, packet loss rate, and business priority data, with a sampling frequency of 10Hz. The data preprocessing unit uses Kalman filtering to eliminate data noise; The feature extraction unit reduces the 128-dimensional original data to 32-dimensional based on the improved PCA algorithm and transmits it to the AI ​​scheduling module via bus.

[0007] Preferably, the AI ​​intelligent scheduling decision module includes a scheduling model training unit, a real-time decision reasoning unit, a dynamic strategy adjustment unit, and an AI model storage unit; The scheduling model training unit constructs the PPO model and trains it according to the preset training cycle and method to obtain a lightweight PPO model. The real-time decision-making inference unit deploys a lightweight PPO model and outputs computing power, bandwidth allocation, and path strategy. The dynamic policy adjustment unit triggers policy adjustment when the load threshold of a preset edge node is exceeded. The AI ​​model storage unit uses Redis to cache and store the model and decision logs.

[0008] Preferably, the dynamic resource allocation module includes a computing power resource allocation unit, a bandwidth resource allocation unit, and a resource conflict resolution unit; The computing resource allocation unit dynamically allocates heterogeneous resources from x86 / ARM CPUs, NVIDIA / AMD GPUs, and Xilinx FPGAs through a functional interface; The bandwidth resource allocation unit issues flow table rules, supporting DSCP QoS marking; The resource conflict resolution unit uses a priority arbitration algorithm to perform secondary resource allocation for real-time and non-real-time services.

[0009] Preferably, the multipath data acceleration module includes a path detection and selection unit, a transmission protocol optimization unit, a data fragmentation and reassembly unit, and a cache acceleration unit; The path detection and selection unit detects the path status, combines the LSTM long short-term memory network model to predict the latency within 100ms, and selects the path with a period of 500ms. The transmission protocol optimization unit establishes a connection with the target edge node, configures the congestion control algorithm, and sets the maximum number of concurrent streams. The data sharding and reassembly unit is used to shard the data source and add an index number and CRC32 check code to each shard; The cache acceleration unit is used to deploy distributed cache nodes to reduce redundant data transmission.

[0010] Preferably, the network quality monitoring module includes a QoS indicator monitoring unit, an anomaly detection and alarm unit, and a quality log storage unit; The QoS indicator monitoring unit is used to monitor latency, packet loss rate, and resource utilization during transmission. The anomaly detection and alarm unit detects abnormal states and triggers an MQTT alarm when an abnormal state persists for three consecutive cycles. The Quality Log Storage Unit is used to store QoS logs for at least 30 days.

[0011] Preferably, the edge and cloud collaboration module includes a data filtering and synchronization unit, a cloud resource scheduling unit, and a collaboration strategy update unit; The data filtering and synchronization unit is used to compress business data at edge nodes and synchronize critical data to the cloud. The cloud resource scheduling unit requests GPU computing power from the cloud in real time based on the edge GPU load; The collaborative strategy update unit dynamically adjusts the data synchronization cycle based on the current status of edge nodes and cloud resources.

[0012] Preferably, the AI ​​model optimization module includes a model lightweighting unit and an incremental training unit; Incremental training units update model parameters incrementally based on new data each day; The model lightweighting unit uses structured pruning, INT8 quantization, or knowledge distillation to lightweight the incrementally trained model.

[0013] A second aspect of the present invention provides an edge network acceleration method based on AI intelligent scheduling, comprising the following specific steps: S1. System initialization and parameter configuration; Start each module, configure data acquisition frequency, resource threshold, QoS indicator threshold, load pre-trained PPO model, and establish edge and cloud connection; S2. Multi-dimensional perception of edge nodes: Collect hardware resources, link parameters, and business data, and after preprocessing, output feature vectors by improving PCA dimensionality reduction. S3, AI intelligent scheduling decision generation; the lightweight PPO model receives feature vectors, outputs computing power, bandwidth allocation and path strategy, and stores decision logs; S4. Dynamic resource allocation execution; heterogeneous computing power is allocated through K3s, bandwidth flow tables are issued by SDN, and resource conflicts are arbitrated. S5: Accelerated multi-path data transmission; path status detection, LSTM latency prediction, QUIC protocol fragmentation transmission, and high-frequency data caching. S6. Real-time monitoring and adjustment of network quality; monitoring QoS indicators, triggering alarms when abnormalities occur, and adjusting strategies using the AI ​​scheduling module; S7, edge and cloud collaboration optimization; filtering redundant data and synchronizing it to the cloud; requesting cloud computing power and adjusting collaboration strategies when resources are insufficient. S8, AI model incremental optimization and system reset; incremental training of PPO model during off-peak hours, lightweight optimization, release of resources and storage of data after business ends.

[0014] Preferably, in step S3, the reward function of the AI ​​scheduling decision is used to evaluate the merits of the AI ​​scheduling strategy; The reward function for AI scheduling decisions is: ; in, The reward value is in the range [0, 1]. These are the weighting coefficients. ; This represents the CPU utilization of the edge node, ranging from [0, 1]. This represents the edge network bandwidth utilization rate, ranging from [0, 1]. The transmission delay deviation rate is in the range [0, 1]. .

[0015] Compared with the prior art, the present invention has the following beneficial technical effects: This invention improves edge node CPU utilization and bandwidth utilization while reducing resource waste through AI dynamic scheduling and heterogeneous resource allocation; optimizes multi-path selection LSTM prediction and QUIC protocol to reduce real-time service latency, meeting the millisecond-level requirements of vehicle networking and industrial control; reduces data transmission volume between edge and cloud and cloud bandwidth usage through AI redundant data filtering and dynamic resource application; reduces AI scheduling model size and inference computing power consumption through model pruning, quantization, and incremental training; adapts to low-computing-power hardware of edge nodes for real-time QoS monitoring and anomaly alarms; shortens network fault repair time; and reduces service interruption rate. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the edge network acceleration system in an embodiment of the present invention; Figure 2 This is a schematic diagram of the edge node sensing module in an embodiment of the present invention; Figure 3This is a schematic diagram of the AI ​​intelligent scheduling decision module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the dynamic resource allocation module in an embodiment of the present invention; Figure 5 This is a flowchart of the edge network acceleration method in an embodiment of the present invention; Figure 6 This is a schematic diagram of the AI ​​model optimization module in an embodiment of the present invention; Figure 7 This is a schematic diagram of the edge-cloud collaboration module in an embodiment of the present invention; Figure 8 This is a schematic diagram of the network quality monitoring module in an embodiment of the present invention; Figure 9 This is a schematic diagram of the multipath data acceleration module in an embodiment of the present invention. Detailed Implementation Example 1

[0017] like Figure 1-4 As shown in the figure, the edge network acceleration system based on AI intelligent scheduling proposed in this embodiment includes an edge node perception module, an AI intelligent scheduling decision module, a dynamic resource allocation module, a multi-path data acceleration module, a network quality monitoring module, an edge and cloud collaboration module, and an AI model optimization module. The edge node perception module is deployed at each edge node. Based on multi-dimensional data collection, preprocessing and feature extraction, it acquires the hardware resource status, network link parameters and business requirements of the edge nodes in real time, and outputs the key feature vector after dimensionality reduction.

[0018] The edge node perception module specifically includes a multi-dimensional data acquisition unit, a data preprocessing unit, and a feature extraction unit. In this embodiment, the multi-dimensional data acquisition unit can use the SNMPv3 protocol to collect edge node CPU utilization, memory usage, GPU load, link bandwidth, latency, packet loss rate, and service priority data, with a sampling frequency of 10Hz. It is equipped with temperature, humidity, and current sensors to monitor the equipment environment and avoid hardware failures. The data preprocessing unit can use Kalman filtering to eliminate data noise, use Z-Score normalization to map the data to the [0, 1] interval for convenient subsequent calculations, and use the DBSCAN algorithm to remove outliers. The delay of the preprocessing step is ≤10ms. The feature extraction unit can reduce the 128-dimensional original data to 32 dimensions based on the improved PCA algorithm, while retaining ≥95% of the feature information, and transmit it to the AI ​​scheduling module through the PCIe 4.0 bus.

[0019] The AI ​​intelligent scheduling and decision-making module uses the deep reinforcement learning PPO algorithm to build a scheduling model. It receives feature data from the edge node perception module and outputs a computing resource allocation scheme, bandwidth allocation ratio, and multi-path transmission strategy, supporting dynamic policy adjustment. The PPO algorithm is a deep reinforcement learning algorithm used to train an agent to make optimal decisions. Its full name is Proximal Policy Optimization. It is a constrained policy gradient method whose constraints are elegantly implemented through pruning terms in the objective function, thus achieving an excellent balance between training stability and ease of implementation.

[0020] The AI ​​intelligent scheduling and decision-making module includes a scheduling model training unit, a real-time decision inference unit, a dynamic policy adjustment unit, and an AI model storage unit. In this embodiment, the scheduling model training unit can build a PPO model based on the TensorFlow 2.15 framework, defining a state space S, an action space A, and setting a reward function during modeling, and training the built model at a training cycle of 24 hours / cycle. The real-time decision inference unit can deploy a lightweight PPO model optimized by TensorRT, outputting computing power, bandwidth allocation, and path strategies. The dynamic policy adjustment unit triggers policy adjustments when the edge node load is ≥80%. The AI ​​model storage unit uses Redis to cache and store the model and decision logs.

[0021] The dynamic resource allocation module is based on AI scheduling decisions and uses K3s and SDN technologies to achieve heterogeneous allocation and conflict arbitration of CPU, GPU and bandwidth resources, ensuring that resources are adapted to business needs.

[0022] The dynamic resource allocation module includes a computing power resource allocation unit, a bandwidth resource allocation unit, and a resource conflict resolution unit. In this embodiment, the computing power resource allocation unit can allocate heterogeneous resources such as x86 / ARM CPUs, NVIDIA / AMD GPUs, and Xilinx FPGAs through the K3s interface. The bandwidth resource allocation unit issues flow table rules based on the OpenDaylight SDN controller and supports DSCP QoS marking. Finally, the resource conflict resolution unit uses a priority arbitration algorithm to perform secondary resource allocation for real-time and non-real-time services, resolving resource contention issues. K3s technology is a simplified version of Kubernetes, a CNCF-certified, fully Kubernetes-compatible distribution. Its core design goals are: extreme lightweightness, simplicity, and ease of deployment, specifically designed for resource-constrained edge computing, IoT, CI / CD, and development and testing environments. Typical application scenarios include edge computing and IoT. In this embodiment, K3s plays the role of a unified abstraction layer and coordinator for heterogeneous resources: Hardware abstraction: Through device plugins, heterogeneous hardware such as x86 / ARM CPUs, NVIDIA / AMD GPUs, and Xilinx FPGAs are uniformly abstracted into Kubernetes-schedulable resources. Intelligent Scheduling: Combines standard schedulers and custom extensions to implement a priority-based arbitration algorithm. Network Coordination: Works with the SDN controller to ensure consistency between network and computing resource policies.

[0023] The multi-path data acceleration module uses the QUIC protocol and LSTM latency prediction to achieve multi-path selection, data fragmentation and transmission, and caching acceleration, thereby reducing data transmission latency.

[0024] In this embodiment, the multi-path data acceleration module specifically includes a path detection and selection unit, a transmission protocol optimization unit, a data fragmentation and reassembly unit, and a cache acceleration unit. As a preferred embodiment, the path detection and selection unit uses ICMPEcho to detect path status, combines it with an LSTM model to predict latency within 100ms, and selects a path with a period of 500ms based on the prediction results. The transmission protocol optimization unit establishes a connection with the target edge node, configures a congestion control algorithm, and sets the maximum number of concurrent streams. The data fragmentation and reassembly unit is used to fragment the data source and add an index number and CRC32 checksum to each fragment. The cache acceleration unit is used to deploy distributed cache nodes to reduce duplicate data transmission.

[0025] The network quality monitoring module monitors QoS indicators in real time, including latency, packet loss rate, and resource utilization. It triggers alarms through anomaly detection and feeds them back to the AI ​​scheduling module to adjust strategies.

[0026] The network quality monitoring module includes a QoS indicator monitoring unit, an anomaly detection and alarm unit, and a quality log storage unit; In this embodiment, the QoS indicator monitoring unit monitors latency status, setting latency indicators such as ≤50ms for real-time tasks and ≤200ms for non-real-time tasks; packet loss rate and resource utilization, such as CPU ≤80% and memory ≤85%, are also considered. The anomaly detection and alarm unit uses the isolated forest algorithm to detect anomalies, and triggers an MQTT alarm when an anomaly persists for three consecutive cycles. The quality log storage unit uses InfluxDB to store QoS logs and maintains them for at least 30 days.

[0027] The edge and cloud collaboration module uses AI redundancy filtering and dynamic resource allocation to achieve data synchronization and resource collaboration between the edge and the cloud, reducing bandwidth consumption.

[0028] In this embodiment, the edge and cloud collaboration module includes a data filtering and synchronization unit, a cloud resource scheduling unit, and a collaboration strategy update unit. The data filtering and synchronization unit is used to compress the business data of the edge nodes and synchronize key data to the cloud. The cloud resource scheduling unit requests GPU computing power from the cloud in real time according to the edge GPU load. The collaboration strategy update unit dynamically adjusts the data synchronization cycle based on the current edge node and cloud resource status.

[0029] The AI ​​model optimization module achieves lightweighting and performance iteration of the AI ​​scheduling model through model pruning, quantization, and incremental training, reducing edge computing power consumption. The AI ​​model optimization module includes a model lightweighting unit and an incremental training unit; the incremental training unit updates the model parameters incrementally based on new data every day; the model lightweighting unit uses structured pruning, INT8 quantization, or knowledge distillation to lightweight the incrementally trained model. Example 2

[0030] like Figure 5 As shown, this embodiment provides an acceleration method for an edge network acceleration system based on AI intelligent scheduling. The system from Embodiment 1 is used for acceleration, and typical application scenarios such as cloud gaming and video surveillance are used as examples in this embodiment. Includes the following steps: Data parameters: Resolution: 4K; Color depth: 10-bit or 12-bit; Encoding format: Uncompressed RAW format; Step 1: System Initialization and Parameter Configuration Start the edge node perception module and configure the data collection parameters: set the sampling frequency to 10Hz, and sample hardware resource collection items, network link collection items, and service requirement collection items; hardware resource collection items include CPU utilization, memory usage, and GPU load; network link collection items include bandwidth, latency, and packet loss rate; service requirement collection items include priority, data volume, and latency threshold.

[0031] Load the pre-trained PPO model of the AI ​​intelligent scheduling decision module and initialize the model parameters, including reward function weights and state space dimensions. Configure resource thresholds for the dynamic resource allocation module: set CPU utilization cap to 80%, memory usage cap to 85%, and bandwidth utilization cap to 85%; configure the transmission protocol and fragment size for the multi-path data acceleration module. Start the network quality monitoring module and set the QoS indicator thresholds: real-time service latency ≤50ms, non-real-time service latency ≤200ms, packet loss rate ≤1%, and alarm triggering cycle 3 monitoring cycles. Establish edge-cloud collaborative connection: Connect edge nodes and cloud platform via MQTT-SN protocol, configure data synchronization period of 10 seconds, and set cloud resource request threshold to edge GPU load ≥90%; Step 2: Multi-dimensional perception of edge nodes The multi-dimensional data acquisition unit collects edge node hardware resource data, network link data, and service data via the SNMPv3 protocol, while also collecting device environment data. Current hardware resource data shows CPU utilization at 65%, memory usage at 70%, and GPU load at 50%. Network link data shows 5G link latency at 30ms, packet loss rate at 0.5%, and bandwidth usage at 40%. Service data shows video surveillance service priority at 2, data volume at 8MB / frame, and latency threshold at 40ms. Current ambient temperature is 25℃, and current is 1.2A. The data preprocessing unit performs noise reduction, normalization (mapping the 65% CPU utilization to 0.65), and anomaly detection on the collected data, outputting a preprocessed data matrix as a 32-dimensional feature vector. The feature extraction unit reduces the dimensionality of 32-dimensional data to 16-dimensional data by improving the PCA algorithm, and then transmits the dimensionality-reduced feature data to the AI ​​intelligent scheduling and decision-making module through the PCIe 4.0 bus. Step 3: AI-powered intelligent scheduling decision generation The real-time decision reasoning unit receives feature data and inputs it into the lightweight PPO model. Based on the current state, such as "CPU utilization 0.65 + bandwidth usage 0.4 + service priority 2", the model outputs the following decision results: ① Computing power allocation: 3-core CPU and 2GB GPU memory are allocated to video services; ② Bandwidth allocation: 50% of bandwidth is allocated to video services (50Mbps); ③ Path selection: 5G is the main path and Wi-Fi 6 is the backup path. The dynamic policy adjustment unit determines that the current edge node load CPU utilization is 65% < 80%, indicating no excessive load and no need to adjust the policy. The decision result is sent to the dynamic resource allocation module and the multi-path data acceleration module via the Ethernet 10Gbps bus. The AI ​​model storage unit stores the decision results—decision time, resource allocation scheme, and path strategy—into a Redis cache and records them in the decision log. The reward function for the AI ​​scheduling decision is:

[0032] in, The reward value is in the range [0, 1], and the larger the value, the better the scheduling strategy. These are the weighting coefficients. The default value is CPU utilization weight. Bandwidth utilization weight Delayed optimization weights ; This represents the CPU utilization of edge nodes, ranging from [0, 1]. For example, a CPU utilization of 65% corresponds to... ; This represents the edge network bandwidth utilization rate, ranging from [0, 1]. For example, 50% bandwidth utilization corresponds to... ; The transmission delay deviation rate is in the range [0, 1]. For example, an actual delay of 38ms corresponds to a threshold of 40ms. Used to evaluate the merits of AI scheduling strategies and guide PPO model training; Step 4: Dynamic resource allocation execution The computing resource allocation unit receives the decision results and creates a business container through the API interface of the K3s interface, configuring the container resources as follows: 3 CPU cores, 2GB GPU memory, and 4GB RAM. The bandwidth resource allocation unit converts "video service share of 50%" into SDN flow table rules, and simultaneously matches port 554, bandwidth limit of 50Mbps, DSCP tag EF, and distributes it to the edge switch through the OpenDaylight controller. The resource conflict resolution unit detects the current resource allocation status. For example, if the current status is that there are no multiple services competing for GPU resources, then no arbitration is required, and a "resource allocation completed" signal is sent to the network quality monitoring module. Step 5: Accelerate data transmission via multi-path methods The path detection and selection unit uses a decision-based path strategy to detect the 5G path latency of 30ms and packet loss rate of 0.5% through ICMPEcho, and the Wi-Fi 6 path latency of 43ms and packet loss rate of 0.8%. Combined with the LSTM model, it predicts the 5G path latency of 32ms in the next 100ms and confirms that 5G is the main path. The transmission protocol optimization unit enables the QUIC protocol, establishes a 0-RTT connection with the target edge node, configures the BBRv3 congestion control algorithm, and sets the maximum number of concurrent flows to 8. The data fragmentation and reassembly unit fragments the 8MB video frame into 1500-byte segments, totaling 5334 segments; each segment is indexed 1-5334 and a CRC32 checksum is added, and the segments are transmitted via the 5G main path, while the first 100 segments are backed up to the Wi-Fi 6 backup path. The caching acceleration unit detects that this video frame is the first transmission and does not trigger caching. The receiving end reassembles the fragments according to the index. After the checksum verification is successful, the complete video frame is output. The transmission delay is 38ms. The LSTM prediction delay formula is:

[0033] in, For prediction delay These are the weighting coefficients. Default path length weight Bandwidth usage weight LSTM prediction weights ; The physical length of the link, after normalization, ranges from [0, 1]. For example, a 5G path length of 1km corresponds to... ; This represents the link bandwidth utilization rate, ranging from [0, 1]. For example, a bandwidth utilization of 40% corresponds to... ; The base delay prediction value output by the LSTM model is obtained based on 100 historical delay samples. Step 6: Real-time monitoring and adjustment of network quality The QoS indicator monitoring unit monitors transmission delay, packet loss rate, bandwidth utilization, and CPU utilization at a frequency of 20Hz, and all indicators are within the threshold range. When the monitoring reached the 10th second, the 5G link latency suddenly increased to 65ms due to interference. The high latency remained for three consecutive monitoring cycles, triggering an alarm in the anomaly detection unit. The alarm unit sends a "5G path delay exceeded" signal to the AI ​​scheduling module and pushes alarm information to the operation and maintenance platform at the same time. After receiving the alarm, the AI ​​scheduling module outputs a new decision within 30ms: switch to the Wi-Fi 6 backup path. The multi-path module completes the path switching within 10ms, and the latency is reduced to 42ms after the switch. Step 7: Edge and Cloud Collaborative Optimization The data filtering and synchronization unit performs redundant filtering on the business data of edge nodes, including video frame metadata and device status data, eliminating 90% of normal status data and compressing 10% of key data, such as video frame anomaly markers and device temperature fluctuations, before synchronizing them to the cloud. The synchronization data volume is reduced from 1GB / h to 100MB / h. When the edge node GPU load rises to 88% due to the addition of new AI inference tasks, which is close to the 90% threshold, the cloud resource scheduling unit requests 1 GPU computing power from the cloud. The cloud responds within 3 seconds and migrates some inference tasks to the cloud via HTTPS protocol, reducing the edge GPU load to 65%. The collaborative strategy update unit adjusts the data synchronization cycle from 10 seconds to 15 seconds based on the current edge-cloud resource status, thereby reducing cloud bandwidth usage; The formula for calculating edge network load balancing is:

[0034] in, The load balancing degree ranges from [0, 1], with a value closer to 1 indicating a more balanced load. For the first The combined load of each edge node, after normalization, ranges from [0, 1]. ,in For the first CPU utilization of each node, For memory utilization, For bandwidth utilization; This represents the maximum combined load across all edge nodes. The minimum total load among all edge nodes; The average total load across all edge nodes; Step 8: Incremental optimization of AI model and system reset The AI ​​model optimization module starts incremental training at 2:00 AM every day. Based on the resource status, business data, and decision logs of the day, a total of 100,000 samples are used to incrementally update the PPO scheduling model. Only the parameters of the fully connected layer are updated, and the training time is 2 hours. The model lightweight unit performs INT8 quantization on the incrementally trained model, reducing the model size from 200MB to 60MB and the inference latency from 50ms to 35ms. When the business ends, the system releases all resources, such as shutting down the business container, deleting the SDN flow table, disconnecting the QUIC connection, and resetting each module to standby state. At the same time, the resource utilization, transmission latency, and decision accuracy of this navigation data are stored in the InfluxDB database for subsequent optimization.

[0035] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. An edge network acceleration system based on AI intelligent scheduling, characterized in that, include: The edge node perception module is deployed at each edge node to acquire the resource status of the edge node hardware, network link parameters and business demand data in real time, and outputs the dimensionality-reduced key feature vector to the AI ​​intelligent scheduling decision module. The AI ​​intelligent scheduling and decision-making module receives the feature vectors output by the edge node perception module and outputs computing power resource allocation schemes, bandwidth allocation ratios and multi-path transmission strategies. The dynamic resource allocation module dynamically allocates CPU, GPU, and bandwidth resources based on the decisions output by the AI ​​scheduling decision module. The multi-path data acceleration module performs multi-path selection, data fragmentation and caching acceleration based on the decisions output by the AI ​​scheduling decision module; The network quality monitoring module monitors the QoS indicators of the dynamic resource allocation module and the multi-path data acceleration module in real time. If an anomaly is detected, an alarm is triggered, and the alarm result is fed back to the AI ​​intelligent scheduling decision module. The edge and cloud collaboration module, based on the decisions output by the AI ​​scheduling and decision-making module, requests GPU computing power from the cloud when overloaded; The AI ​​model optimization module uses model pruning, quantization, and incremental training to lightweight and iterate the performance of the AI ​​scheduling model, reducing edge computing power consumption.

2. The edge network acceleration system based on AI intelligent scheduling according to claim 1, characterized in that, The edge node perception module specifically includes a multi-dimensional data acquisition unit, a data preprocessing unit, and a feature extraction unit; The multi-dimensional data acquisition unit collects edge node CPU utilization, memory usage, GPU load, link bandwidth, latency, packet loss rate, and business priority data, with a sampling frequency of 10Hz. The data preprocessing unit uses Kalman filtering to eliminate data noise; The feature extraction unit reduces the 128-dimensional original data to 32-dimensional based on the improved PCA algorithm and transmits it to the AI ​​scheduling module via bus.

3. The edge network acceleration system based on AI intelligent scheduling according to claim 1, characterized in that, The AI ​​intelligent scheduling decision module includes a scheduling model training unit, a real-time decision reasoning unit, a dynamic strategy adjustment unit, and an AI model storage unit; The scheduling model training unit constructs the PPO model and trains it according to the preset training cycle and method to obtain a lightweight PPO model. The real-time decision-making inference unit deploys a lightweight PPO model and outputs computing power, bandwidth allocation, and path strategy. The dynamic policy adjustment unit triggers policy adjustment when the load threshold of a preset edge node is exceeded. The AI ​​model storage unit uses Redis to cache and store the model and decision logs.

4. The edge network acceleration system based on AI intelligent scheduling according to claim 1, characterized in that, The dynamic resource allocation module includes a computing power resource allocation unit, a bandwidth resource allocation unit, and a resource conflict resolution unit. The computing resource allocation unit dynamically allocates heterogeneous resources from x86 / ARM CPUs, NVIDIA / AMD GPUs, and Xilinx FPGAs through a functional interface; The bandwidth resource allocation unit issues flow table rules, supporting DSCP QoS marking; The resource conflict resolution unit uses a priority arbitration algorithm to perform secondary resource allocation for real-time and non-real-time services.

5. The edge network acceleration system based on AI intelligent scheduling according to claim 1, characterized in that, The multipath data acceleration module includes a path detection and selection unit, a transmission protocol optimization unit, a data fragmentation and reassembly unit, and a cache acceleration unit. The path detection and selection unit detects the path status, combines the LSTM long short-term memory network model to predict the latency within 100ms, and selects the path with a period of 500ms. The transmission protocol optimization unit establishes a connection with the target edge node, configures the congestion control algorithm, and sets the maximum number of concurrent streams. The data sharding and reassembly unit is used to shard the data source and add an index number and CRC32 check code to each shard; The cache acceleration unit is used to deploy distributed cache nodes to reduce redundant data transmission.

6. The edge network acceleration system based on AI intelligent scheduling according to claim 1, characterized in that, The network quality monitoring module includes a QoS indicator monitoring unit, an anomaly detection and alarm unit, and a quality log storage unit; The QoS indicator monitoring unit is used to monitor latency, packet loss rate, and resource utilization during transmission. The anomaly detection and alarm unit detects abnormal states and triggers an MQTT alarm when an abnormal state persists for three consecutive cycles. The Quality Log Storage Unit is used to store QoS logs for at least 30 days.

7. The edge network acceleration system based on AI intelligent scheduling according to claim 1, characterized in that, The edge and cloud collaboration module includes a data filtering and synchronization unit, a cloud resource scheduling unit, and a collaboration strategy update unit. The data filtering and synchronization unit is used to compress business data at edge nodes and synchronize critical data to the cloud. The cloud resource scheduling unit requests GPU computing power from the cloud in real time based on the edge GPU load; The collaborative strategy update unit dynamically adjusts the data synchronization cycle based on the current status of edge nodes and cloud resources.

8. The edge network acceleration system based on AI intelligent scheduling according to claim 1, characterized in that, The AI ​​model optimization module includes a model lightweighting unit and an incremental training unit; Incremental training units update model parameters incrementally based on new data each day; The model lightweighting unit uses structured pruning, INT8 quantization, or knowledge distillation to lightweight the incrementally trained model.

9. An edge network acceleration method based on AI intelligent scheduling, using the system described in any one of claims 1-8 for acceleration, characterized in that, The specific steps include the following: S1. System initialization and parameter configuration; Start each module, configure data acquisition frequency, resource threshold, QoS indicator threshold, load pre-trained PPO model, and establish edge and cloud connection; S2. Multi-dimensional perception of edge nodes: Collect hardware resources, link parameters, and business data, and after preprocessing, output feature vectors by improving PCA dimensionality reduction. S3, AI intelligent scheduling decision generation; the lightweight PPO model receives feature vectors, outputs computing power, bandwidth allocation and path strategy, and stores decision logs; S4. Dynamic resource allocation execution; heterogeneous computing power is allocated through K3s, bandwidth flow tables are issued by SDN, and resource conflicts are arbitrated. S5: Accelerated multi-path data transmission; path status detection, LSTM latency prediction, QUIC protocol fragmentation transmission, and high-frequency data caching. S6. Real-time monitoring and adjustment of network quality; Monitor QoS metrics and trigger alarms when abnormalities occur; the AI ​​scheduling module then adjusts its strategy. S7, edge and cloud collaboration optimization; filtering redundant data and synchronizing it to the cloud; requesting cloud computing power and adjusting collaboration strategies when resources are insufficient. S8, AI model incremental optimization and system reset; incremental training of PPO model during off-peak hours, lightweight optimization, release of resources and storage of data after business ends.

10. The edge network acceleration method based on AI intelligent scheduling according to claim 9, characterized in that, In step S3, the reward function of AI scheduling decision is used to evaluate the merits of AI scheduling strategy. The reward function for AI scheduling decisions is: ； in, The reward value is in the range [0, 1]. These are the weighting coefficients. ; This represents the CPU utilization of the edge node, ranging from [0, 1]. This represents the edge network bandwidth utilization rate, ranging from [0, 1]. The transmission delay deviation rate is in the range [0, 1]. .

Images