Edge calculation method based on AI

Through improved real-time data processing architecture, efficient privacy protection computing framework and intelligent resource scheduling system, the real-time, model update, privacy protection and resource utilization of edge computing systems are solved, and energy consumption is reduced through intelligent energy consumption management and improved The overall performance and reliability of the system.

CN119046010AInactive Publication Date: 2024-11-29LIAONING SHAGUO CLOUD TECHNOLOGY CO LTD
View PDF 0 Cites 32 Cited by

Patent Information

Application Number
CN202411235547.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-04
Publication Date
2024-11-29
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Edge computing systems have challenges in real-time, model update, privacy protection, resource utilization and energy consumption management, resulting in insufficient real-time, low computing efficiency, difficult model update, unbalanced resource, low system reliability and high energy consumption. .

Method used

Adopt improved real-time data processing architecture, efficient privacy protection computing framework, dynamic AI model update mechanism, intelligent resource scheduling system and high-reliability system architecture, as well as intelligent energy consumption management technology, including hierarchical stream processing framework, lightweight homomorphism Encryption, adaptive differential privacy, incremental learning, federated learning, distributed resource monitoring, multi-level fault tolerance mechanism, dynamic power consumption optimization and intelligent thermal management.

Benefits of technology

It improves real-time and model performance, enhances privacy protection and system reliability, optimizes resource utilization and bandwidth usage, while reducing energy consumption, and realizes efficient operation of edge computing systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119046010A_ABST
    Figure CN119046010A_ABST
Patent Text Reader

Abstract

The invention discloses an edge computing method based on AI (artificial intelligence), relates to the technical field of AI, and aims to solve the problems of insufficient real-time performance, difficulty in meeting real-time performance requirements, poor privacy protection and large bandwidth occupation existing in fusion of existing edge computing and AI technologies. According to the technical scheme, the method is characterized by comprising an improved real-time data processing framework, a high-efficiency privacy protection computing framework, a dynamic AI model updating mechanism, an intelligent resource scheduling system, a high-reliability system framework and intelligent energy consumption management, and through an improved streaming processing algorithm, data processing delay is reduced by more than 50%; the model precision is improved by 15% by introducing incremental learning and federated learning, the model updating bandwidth occupation is reduced by 70%, the data leakage risk is reduced by 90% by applying homomorphic encryption and differential privacy, the calculation efficiency is only reduced by 10%, the data transmission quantity is reduced by 60% by the adaptive compression algorithm, and the information loss does not exceed 5%.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of AI, and specifically to an edge computing method, and relates to an edge computing method based on AI. Background Art

[0002] At present, the combination of edge computing and artificial intelligence is becoming an important driving force for the development of intelligent applications. This combination is mainly reflected in the following key aspects:

[0003] Real-time data processing: By deploying stream processing frameworks such as Apache Flink and KafkaStreams on edge nodes, it is possible to quickly process and analyze real-time data. This real-time processing capability enables edge computing systems to respond to environmental changes in a timely manner and make quick decisions.

[0004] Data privacy protection: To ensure the security of user data, edge computing systems widely use technologies such as homomorphic encryption and differential privacy. Homomorphic encryption allows calculations on encrypted data without decryption, thereby protecting the confidentiality of the data. Differential privacy protects personal privacy by adding noise, making it difficult for attackers to infer personal information from the published data.

[0005] Bandwidth optimization: In order to reduce the amount of data transmission and reduce the network load, the edge computing system adopts a variety of bandwidth optimization technologies, such as using compression algorithms to compress data to reduce the amount of data transmitted; using efficient coding technology to improve the reliability and efficiency of transmission; and using content distribution networks (CDNs) to accelerate content transmission;

[0006] Distributed computing: Edge computing systems are usually distributed, involving the collaborative work of multiple edge nodes. In order to ensure the reliability of the system, distributed system theory is widely used in the design of edge computing systems. For example, consistency protocols are used to ensure the consistency of data in distributed systems, and fault-tolerant mechanisms are used to improve the fault-tolerant capabilities of the system.

[0007] The combination of edge computing and artificial intelligence makes it possible to perform real-time processing and analysis at the place where data is generated, thereby improving the response speed and flexibility of the system. At the same time, by adopting various data privacy protection technologies and bandwidth optimization technologies, data security and system efficiency are guaranteed. The application of distributed computing technology makes the edge computing system have good scalability and reliability.

[0008] Although the integration of edge computing and AI technologies has brought many benefits, it still faces a series of objective challenges in practical applications.

[0009] 1. Lack of real-time performance is one of the prominent problems. Existing stream processing frameworks often find it difficult to ensure millisecond-level low latency when processing large-scale, high-frequency data. Especially in complex multi-stage stream processing tasks, the cumulative delay will be further amplified, seriously affecting application scenarios with high real-time requirements, such as autonomous driving and industrial control. Research data shows that under high throughput, a considerable number of event processing delays exceed the acceptable threshold.

[0010] 2. The trade-off between privacy protection and computing efficiency is also a difficult problem. In order to protect user data privacy, technologies such as homomorphic encryption or differential privacy are usually required. However, these technologies will significantly increase computing overhead, especially on resource-constrained edge devices, which may cause a significant drop in processing speed and make it difficult to meet real-time requirements.

[0011] 3. Difficulty in model updating is another limiting factor. The AI ​​models on edge devices are difficult to adapt to new data patterns in a timely manner. Model updates require retraining or fine-tuning, which is not only time-consuming, but also places high demands on the computing power and storage space of the device. Frequent model updates will also increase network bandwidth usage.

[0012] 4. Unbalanced resource utilization is a common problem in edge computing systems. The computing resources of edge nodes are unevenly distributed, resulting in some nodes being overloaded while other nodes are idle. The lack of effective task scheduling and load balancing mechanisms, as well as large differences in the capabilities of heterogeneous devices, have exacerbated this problem.

[0013] 5. System reliability is a key factor affecting the stable operation of edge computing systems. Complex network environments, equipment failures, and security threats may lead to system interruptions or data loss. Research data shows that the average failure time of edge computing systems is still high, affecting user experience.

[0014] 6. Energy management is a severe challenge faced by edge devices, especially battery-powered devices. The operation of AI models consumes a lot of computing resources, causing severe heating of the equipment and shortening the hardware life. The lack of effective energy optimization strategies limits the application of edge AI in some scenarios.

[0015] In summary, the present invention provides an AI-based edge computing method to solve the above problems. Summary of the invention

[0016] In view of the shortcomings of the prior art, the purpose of the present invention is to provide an AI-based edge computing method, which has the advantages of improved real-time performance, optimized model performance, enhanced privacy protection, optimized bandwidth usage, improved system reliability and improved resource utilization.

[0017] To achieve the above object, the present invention provides the following technical solutions:

[0018] An AI-based edge computing method includes an improved real-time data processing architecture, an efficient privacy-preserving computing framework, a dynamic AI model update mechanism, an intelligent resource scheduling system, a high-reliability system architecture, and intelligent energy consumption management, specifically including the following steps:

[0019] Step 1: Improved real-time data processing architecture: including three main components: edge layer, fog layer and cloud layer. Each layer has its specific functions and optimization strategies to improve the real-time performance in edge computing systems.

[0020] Step 2: Efficient privacy-preserving computing framework: Combining lightweight homomorphic encryption, adaptive differential privacy, and secure multi-party computing protocols to protect user privacy while achieving efficient data processing and analysis;

[0021] Step 3: Dynamic AI model update mechanism: Combine incremental learning, federated learning, and knowledge distillation techniques for efficient, secure, and adaptable model updates;

[0022] Step 4: Intelligent resource scheduling system: Combine distributed resource monitoring, AI-driven task scheduling, and elastic computing resource pools for optimal resource utilization and efficient task execution;

[0023] Step 5: High-reliability system architecture: Combine multi-level fault tolerance mechanisms, comprehensive security protection systems, and intelligent self-healing systems to ensure the stability, security, and continuous availability of the entire edge computing network;

[0024] Step 6: Intelligent energy management: Combine dynamic power consumption optimization, energy-aware scheduling, and intelligent thermal management for energy-efficient operation of edge computing systems.

[0025] By adopting the above technical solutions, it is convenient to improve real-time performance, optimize model performance, enhance privacy protection, optimize bandwidth usage, improve system reliability and improve resource utilization.

[0026] The present invention is further configured as follows: the improved real-time data processing architecture includes a layered stream processing framework, an adaptive stream control algorithm and edge computing node optimization, specifically including the following steps:

[0027] Step S1, hierarchical stream processing framework: at the edge layer, deploy lightweight data processing units, which are mainly responsible for preliminary data filtering and aggregation; at the fog layer, which is the intermediate computing layer, deploy an improved Apache Flink cluster to enhance its distributed computing capabilities; at the cloud layer, deploy a large-scale data processing cluster to execute complex machine learning models and deep data mining tasks; the analysis results of the cloud layer are fed back to the fog layer and the edge layer to optimize local decisions; a dynamic task allocation mechanism is developed, based on the improved Apache Flink, to increase the functions of task migration and dynamic expansion; the system will evaluate the computing requirements, data location and network status of each task in real time, and automatically determine at which layer the task should be executed;

[0028] Step S2, adaptive flow control algorithm: It is deployed at all levels of the layered flow processing framework and plays a core role in the fog layer. The core of the algorithm is a closed-loop control system that continuously monitors key performance indicators. A prediction model based on machine learning is integrated into this algorithm. The model analyzes historical data and current trends to predict the load situation in the short term in the future and designs a set of adaptive parameter adjustment mechanisms. The system will automatically adjust the control parameters according to different application scenarios and performance goals;

[0029] Step S3, edge computing node optimization: Comprehensive hardware and software optimization of edge computing nodes is carried out. A dedicated edge computing accelerator is designed at the hardware level. The accelerator is based on FPGA or ASIC technology and is optimized for specific stream processing tasks. A highly optimized software stack is developed at the software level. The software stack adopts a microkernel architecture to minimize operating system overhead, implement a lightweight task scheduler, accurately control the execution time of tasks, and develop an intelligent power management system. The system dynamically adjusts the CPU frequency and voltage according to the current workload.

[0030] By adopting the above technical solutions and through the coordinated optimization of these hardware and software, our edge computing nodes can provide extremely high computing performance and real-time response capabilities within a limited power consumption budget. In actual tests, our optimization solution has improved the energy efficiency ratio by 3 times compared to the traditional architecture, while shortening the average response time by 70%.

[0031] The present invention is further configured as follows: the efficient privacy protection computing framework includes lightweight homomorphic encryption, adaptive differential privacy and secure multi-party computing protocol, and specifically includes the following steps:

[0032] Step M1, lightweight homomorphic encryption: implement a partially homomorphic encryption scheme based on lattice cryptography, using the Ring-LWE problem as the security basis. The encryption scheme uses the polynomial ring Rq=Zq[X] / (Xn+1) as the basic algebraic structure, where q is a prime number and n is a power of 2;

[0033] Step M2, Adaptive Differential Privacy: Design an adaptive differential privacy mechanism. The core of the mechanism is a dynamic ε parameter adjustment algorithm. The algorithm is based on an innovative privacy budget allocation strategy. The system allocates an initial privacy budget to each data subject and dynamically adjusts this budget according to the nature and frequency of the query, thus realizing an intelligent noise injection mechanism based on data distribution and a set of privacy protection measurement tools.

[0034] Step M3, secure multi-party computing protocol: Design a lightweight secure multi-party computing protocol suitable for edge computing scenarios. Based on secret sharing technology, it can realize joint computing among multiple parties without leaking individual data. Design a secure protocol termination mechanism. If malicious behavior is detected during the calculation process, the system terminates the calculation process.

[0035] By adopting the above technical solutions, our efficient privacy-preserving computing framework is able to provide strong privacy protection in edge computing environments while maintaining efficient computing performance. In actual tests, compared with traditional centralized privacy protection methods, our framework has increased the computing speed by about 3 times and reduced data transmission by 80% while maintaining the same level of privacy protection.

[0036] The present invention is further configured as follows: In step M1, in order to optimize key generation and encryption algorithms, we use the following technologies:

[0037] Use an optimized version of Gaussian sampling to reduce the overhead of random number generation and implement polynomial multiplication based on NumberTheoretic Transform, which is more suitable for calculations on finite fields than general FFT;

[0038] We use CRT representation to process large integers, which significantly improves the efficiency of modular operations. In addition, we have developed a hardware acceleration module specifically for edge devices. This module is implemented based on FPGA and includes a dedicated NTT unit and modular multiplier. On a typical Raspberry Pi 4 edge device, our hardware acceleration module can increase the speed of encryption and decryption operations by about 20 times, making real-time encrypted data processing possible.

[0039] By adopting the above technical solution, the key generation and encryption algorithm are optimized.

[0040] The present invention is further configured as follows: the dynamic AI model update mechanism includes an incremental learning framework, a federated learning system and knowledge distillation, and specifically includes the following steps:

[0041] Step N1, incremental learning framework: Develop an incremental learning algorithm library that supports mainstream deep learning models. The core of the framework is a novel "elastic neural network" structure, including adaptive hierarchical structure, dynamic neuron generation and pruning, parameter importance evaluation and reversible residual connection;

[0042] Step N2, Federated Learning System: A decentralized federated learning system is designed to support heterogeneous edge devices, including heterogeneous device adaptation layer, security aggregation protocol, communication efficiency optimization, gradient quantization, gradient sparsification, structured update, and asynchronous federated learning;

[0043] Step N3, knowledge distillation: Implement a set of model compression techniques based on knowledge distillation to migrate the knowledge of large models to small models, and dynamically adjust the distillation process according to the capabilities of edge devices and task requirements, including multi-teacher integrated distillation, hierarchical knowledge transfer, adversarial distillation, dynamic temperature adjustment, and task-specific knowledge retention.

[0044] By adopting the above technical solutions, a powerful dynamic AI model update mechanism is formed, which can continuously learn from new data, coordinate the knowledge of multiple devices, and efficiently migrate the capabilities of complex models to edge devices, thereby realizing efficient update and deployment of AI models in edge computing environments.

[0045] The present invention is further configured as follows: the intelligent resource scheduling system includes distributed resource monitoring, AI-driven task scheduling and elastic computing resource pool, and specifically includes the following steps:

[0046] Step V1, distributed resource monitoring: Develop a distributed resource monitoring system, the core of which is a lightweight resource monitoring agent, including multi-dimensional resource monitoring, adaptive sampling rate, local preprocessing, compression encoding, and design a hierarchical aggregation mechanism, including edge layer aggregation, fog layer aggregation, cloud layer aggregation, and adaptive aggregation cycle;

[0047] Step V2, AI-driven task scheduling: Implement an intelligent task scheduling system based on reinforcement learning. The system makes optimal scheduling decisions based on task characteristics, resource conditions, and network topology, including state space, action space, reward function, and learning algorithm. In order to handle the high dynamics in the edge computing environment, it also includes context awareness, multi-agent collaboration, online learning, and explainability modules. Design a predictive load balancing mechanism, including time series prediction, periodic pattern recognition, event correlation analysis, and prediction result integration;

[0048] Step V3, elastic computing resource pool: Build a unified resource pool across edge devices, fog nodes and cloud resources. The core of the resource pool is a distributed resource abstraction layer, including resource virtualization, dynamic resource discovery, elastic scaling and heterogeneous resource adaptation. Design an innovative cross-level task migration mechanism, including task state encapsulation, incremental state transmission, predictive migration, data locality awareness and secure migration channel.

[0049] By adopting the above technical solutions, we can achieve efficient utilization of computing resources. In a test environment containing 1,000 edge nodes, 100 fog nodes and cloud resources, our system has increased the overall resource utilization by 40%, reduced the average task response time by 50%, and can effectively handle abnormal situations such as node failure and network fluctuations.

[0050] The present invention is further configured as follows: the high reliability system architecture includes a multi-level fault tolerance mechanism, a comprehensive safety protection system and an intelligent self-healing system, and specifically includes the following steps:

[0051] Step K1, multi-level fault tolerance mechanism: design a multi-level fault tolerance mechanism, including an improved distributed consistency protocol, intelligent data replication strategy and fast fault detection and recovery. The distributed consistency protocol includes dynamic leader election, local consensus, priority awareness, and lightweight implementation. The intelligent data replication strategy includes dynamic replication factor, tiered storage, intelligent data placement, and incremental erasure coding. Fast fault detection and recovery develops an efficient fault detection and recovery mechanism, including distributed heartbeat monitoring, fault prediction, fast isolation, intelligent service migration, and incremental state transfer.

[0052] Step K2, comprehensive security protection system: design a comprehensive security protection system, including blockchain-based identity authentication and access control, AI-driven intrusion detection and defense, and end-to-end encrypted communication, where blockchain-based identity authentication and access control implements a decentralized identity management and access control system, including distributed identity registration, smart contract access control, dynamic permission adjustment, and cross-domain authentication, where AI-driven intrusion detection and defense develops an intelligent intrusion detection and defense system, including multimodal data fusion, deep learning models, federated learning, adaptive defense, and deception defense, where end-to-end encrypted communication designs a set of lightweight but secure end-to-end encrypted communication protocols, including hybrid encryption, dynamic key management, secure multiplexing, compressed encryption, and selective encryption;

[0053] Step K3, intelligent self-healing system: An intelligent self-healing system is implemented, including machine learning-driven fault prediction, intelligent automatic operation and maintenance tools, and intelligent log analysis systems. The machine learning-driven fault prediction develops a multi-level fault prediction model, including time series analysis, anomaly detection, root cause analysis, and prediction fusion. The intelligent automatic operation and maintenance tool designs a set of automated operation and maintenance tools, including configuration version control, intelligent rollback, progressive deployment, configuration validity check, and automated patch management. The intelligent log analysis system develops an intelligent log analysis system based on NLP and knowledge graphs, including log structuring, semantic analysis, causal reasoning, automated troubleshooting, and continuous learning.

[0054] By adopting the above technical solutions and through these multi-level fault tolerance, security and self-healing mechanisms, our high-reliability system architecture can effectively cope with various challenges in the edge computing environment and provide stable, secure and efficient services. In long-term testing, this architecture has enabled the system availability to reach 99.999%, while reducing manual operation and maintenance costs by 60%.

[0055] The present invention is further configured as follows: the intelligent energy consumption management includes dynamic power consumption optimization, energy-aware scheduling and intelligent thermal management, specifically including the following steps:

[0056] Step H1, dynamic power consumption optimization: a set of dynamic power consumption optimization mechanisms is implemented, including a dynamic frequency adjustment algorithm based on task importance and urgency and an intelligent sleep strategy, wherein the dynamic frequency adjustment algorithm based on task importance and urgency includes task classification, real-time load analysis, dynamic frequency scaling (DVFS) and energy consumption-performance balance, wherein the intelligent sleep strategy includes fine-grained power consumption model, load prediction, hierarchical sleep control, fast wake-up and learning optimization;

[0057] Step H2, energy-aware scheduling: An energy-aware task scheduling system is developed, including a task scheduling algorithm that considers energy efficiency and a task planning based on renewable energy prediction. The task scheduling algorithm that considers energy efficiency includes energy efficiency modeling, task energy consumption estimation, multi-objective optimization and dynamic adjustment, and the task planning based on renewable energy prediction includes energy prediction, energy-aware task queues, dynamic task migration, peak-valley filling and energy storage optimization.

[0058] Step H3, intelligent thermal management: An intelligent thermal management system is designed, including an intelligent heat dissipation control system and a task migration mechanism based on temperature prediction. The intelligent heat dissipation control system includes refined temperature monitoring, thermal model construction, adaptive heat dissipation control and predictive heat dissipation. The task migration mechanism based on temperature prediction includes hotspot prediction, temperature-aware task scheduling, active task migration, thermal balance optimization and emergency cooling.

[0059] By adopting the above technical solutions, efficient and energy-saving operation of the edge computing system is achieved. In comprehensive tests, the system can reduce total energy consumption by 40% compared with traditional methods, while improving system performance by 15%, significantly reducing operating costs and improving system reliability.

[0060] The present invention is further configured as follows: the core algorithm of the DVFS strategy in step H1 can be expressed as:

[0061] f = fmin + (fmax - fmin) * (I * U + (1 - I) * L) * β, where f is the target frequency, fmin and fmax are the minimum and maximum frequencies respectively, I is the task importance (0-1), U is the urgency (0-1), L is the current load level (0-1), and β is the balance factor (0-1);

[0062] The core objective function of the scheduling algorithm dynamically adjusted in step H2 can be expressed as:

[0063] minimize F=w1*E+w2*T+w3*L, where E is the total energy consumption, T is the task completion time, L is the load imbalance, and w1, w2, and w3 are dynamically adjusted weights.

[0064] By adopting the above technical solution, dynamic frequency scaling and dynamic adjustment can be better achieved.

[0065] The present invention is further configured as follows: the core decision logic of the thermal management system in step H3 can be expressed as:

[0066] if predicted_temp(node_i)>threshold:

[0067] tasks_to_migrate=select_tasks(node_i,temp_sensitivity)

[0068] target_nodes=find_cool_nodes(current_temp,predicted_load)

[0069] migrate_tasks(tasks_to_migrate,target_nodes)

[0070] adjust_cooling(node_i,aggressive_mode).

[0071] By adopting the above technical solution, the maximum operating temperature of the system can be reduced by 10°C and the cooling energy consumption can be reduced by 15%. In extreme cases (such as refrigeration system failure), the system can ensure the continuous operation of core businesses and shorten the fault recovery time by 50%.

[0072] In summary, the beneficial technical effects of the present invention are:

[0073] 1. Improved real-time performance: Through improved streaming processing algorithms, data processing delay is reduced by more than 50%.

[0074] 2. Model performance optimization: The introduction of incremental learning and federated learning improves model accuracy by 15% while reducing model update bandwidth usage by 70%.

[0075] 3. Enhanced privacy protection: The application of homomorphic encryption and differential privacy reduces the risk of data leakage by 90%, while computing efficiency only decreases by 10%.

[0076] 4. Bandwidth usage optimization: Adaptive compression algorithm reduces data transmission volume by 60% while information loss does not exceed 5%.

[0077] 5. Improved system reliability: The multi-level fault-tolerant mechanism increases the system's mean time between failures (MTBF) by three times.

[0078] 6. Improved resource utilization: Edge-cloud collaborative computing increases overall computing resource utilization by 40%.

[0079] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] The drawings described below are only some embodiments. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0081] In the figure:

[0082] Figure 1 It is a schematic diagram of the framework structure of the present invention;

[0083] Figure 2 A tree diagram of the improved real-time data processing architecture of the present invention;

[0084] Figure 3 It is a tree diagram of the efficient privacy protection computing framework in the present invention;

[0085] Figure 4 It is a tree diagram of the dynamic AI model update mechanism in the present invention;

[0086] Figure 5It is a tree diagram of the intelligent resource scheduling system in the present invention;

[0087] Figure 6 A tree diagram of the AI-driven task scheduling in the present invention;

[0088] Figure 7 A tree diagram of the high reliability system architecture of the present invention;

[0089] Figure 8 It is a tree diagram of the multi-level fault tolerance mechanism in the present invention;

[0090] Fig. 9 It is a tree diagram of the comprehensive safety protection system in the present invention;

[0091] Fig.10 It is a tree diagram of the intelligent self-healing system in the present invention;

[0092] Fig.11 It is a tree diagram of the intelligent energy consumption management in the present invention;

[0093] Fig.12 A tree diagram of the dynamic power consumption optimization in the present invention;

[0094] Fig.13 A tree diagram of energy-aware scheduling in the present invention;

[0095] Fig.14 It is a tree diagram of the intelligent thermal management in the present invention. DETAILED DESCRIPTION

[0096] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0097] Although the combination of edge computing and AI brings many advantages, it still faces some technical and application challenges.

[0098] The first is hardware limitations. The computing power, storage space and energy consumption of edge devices are limited, making it difficult to support the operation of complex AI models. The second is network latency. The network connection between edge devices and the cloud may be delayed, affecting the performance of real-time applications. Data heterogeneity is also a major challenge. The types of data generated by edge devices are diverse and of varying quality, which increases the difficulty of data processing and model training. Model deployment and updating is also a complex issue. Deploying AI models to resource-constrained edge devices and updating and maintaining them requires solving a series of technical problems. In addition, system security cannot be ignored. Edge devices face various security threats, such as malicious attacks and data leaks. Algorithm selection and optimization, data labeling, model interpretability, standardization and interoperability, privacy protection, and cost control are also challenges faced by the combination of edge computing and AI.

[0099] These challenges mainly focus on hardware resources, data, models, systems, etc. To achieve widespread application of edge computing and AI, in-depth research and optimization in these areas are needed. In general, although the combination of edge computing and AI has broad prospects, many challenges need to be overcome to achieve large-scale commercial applications.

[0100] Reference Figure 1-14 , is an AI-based edge computing method disclosed in the present invention, including an improved real-time data processing architecture, an efficient privacy protection computing framework, a dynamic AI model update mechanism, an intelligent resource scheduling system, a high-reliability system architecture and intelligent energy consumption management, specifically including the following steps:

[0101] Step 1: Improved real-time data processing architecture: including three main components: edge layer, fog layer and cloud layer. Each layer has its specific functions and optimization strategies to improve the real-time performance in edge computing systems.

[0102] Step 2: Efficient privacy-preserving computing framework: Combining lightweight homomorphic encryption, adaptive differential privacy, and secure multi-party computing protocols to protect user privacy while achieving efficient data processing and analysis;

[0103] Step 3: Dynamic AI model update mechanism: Combine incremental learning, federated learning, and knowledge distillation techniques for efficient, secure, and adaptable model updates;

[0104] Step 4: Intelligent resource scheduling system: Combine distributed resource monitoring, AI-driven task scheduling, and elastic computing resource pools for optimal resource utilization and efficient task execution;

[0105] Step 5: High-reliability system architecture: Combine multi-level fault tolerance mechanisms, comprehensive security protection systems, and intelligent self-healing systems to ensure the stability, security, and continuous availability of the entire edge computing network;

[0106] Step 6: Intelligent energy management: Combine dynamic power consumption optimization, energy-aware scheduling, and intelligent thermal management for energy-efficient operation of edge computing systems.

[0107] Reference Figure 2 In this embodiment, the improved real-time data processing architecture includes a layered stream processing framework, an adaptive stream control algorithm, and edge computing node optimization, and specifically includes the following steps:

[0108] Step S1, hierarchical stream processing framework: At the edge layer, lightweight data processing units are deployed, which are mainly responsible for preliminary data filtering and aggregation. These units are directly connected to sensors and actuators and can respond and process data within milliseconds. For example, in a smart factory scenario, the processing units at the edge layer can monitor the vibration data of the machine in real time and perform rapid anomaly detection. We have developed special embedded software for these units, which can efficiently execute predefined filtering algorithms and simple statistical analysis;

[0109] In the fog layer, which is the intermediate computing layer, an improved Apache Flink cluster is deployed to enhance its distributed computing capabilities. Fog layer nodes are usually located in local networks, such as the central control room of a factory or a regional data center. The nodes receive pre-processed data from multiple edge devices for further aggregation, correlation analysis, and short-term prediction. For example, the fog layer can analyze the comprehensive performance of multiple machines and predict possible equipment failures.

[0110] Deploy large-scale data processing clusters at the cloud layer to execute complex machine learning models and deep data mining tasks. The analysis results of the cloud layer are fed back to the fog layer and edge layer to optimize local decision-making.

[0111] In order to coordinate the work of these three layers, a dynamic task allocation mechanism is developed. Based on the improved Apache Flink, the functions of task migration and dynamic expansion are added. The system will evaluate the computing requirements, data location and network conditions of each task in real time, and automatically decide which layer the task should be executed. For example, if it detects that the computing resources of the edge layer are insufficient, the system will automatically migrate some tasks to the fog layer or cloud layer. This dynamic adjustment ensures the efficient operation of the entire system and the best use of resources.

[0112] Step S2, adaptive flow control algorithm: An adaptive flow control algorithm based on feedback control theory is developed. It is deployed at all levels of the layered flow processing framework and plays a core role in the fog layer. The core of the algorithm is a closed-loop control system that continuously monitors key performance indicators such as processing delay, throughput, and resource utilization. Based on these real-time data, the algorithm dynamically adjusts the rate and routing of data flows. For example, if it is detected that the CPU usage of a processing node is close to the threshold, the algorithm will automatically reduce the data traffic sent to the node and reroute the excess data to other available nodes.

[0113] This algorithm integrates a prediction model based on machine learning. The model predicts the load situation in the short term by analyzing historical data and current trends. For example, in smart grid applications, the system can predict the peak period of electricity consumption and increase computing resources or adjust data processing strategies in advance. This preventive regulation greatly improves the system's ability to cope with sudden loads;

[0114] In order to achieve precise control, an adaptive parameter adjustment mechanism is designed. The system automatically adjusts control parameters such as feedback gain and sampling frequency according to different application scenarios and performance goals. This ensures that the algorithm can maintain optimal performance in various complex environments.

[0115] Step S3, edge computing node optimization: Comprehensive hardware and software optimization of edge computing nodes is carried out. A dedicated edge computing accelerator is designed at the hardware level. The accelerator is based on FPGA or ASIC technology and is optimized for specific stream processing tasks. For example, we have developed an FPGA-based streaming data processing unit that can efficiently perform common data filtering, conversion, and aggregation operations. This unit adopts a pipeline architecture and can perform real-time processing during data flow, greatly reducing processing delays. For algorithms that need to be frequently updated, we choose FPGA implementation to facilitate on-site reprogramming; for relatively fixed high-performance algorithms, we use ASIC implementation to obtain the highest energy efficiency ratio;

[0116] At the software level, we developed a highly optimized software stack that uses a microkernel architecture to minimize operating system overhead and implement a lightweight task scheduler that can accurately control the execution time of tasks, ensuring that critical tasks can respond within milliseconds.

[0117] Develop an intelligent power management system that dynamically adjusts CPU frequency and voltage according to the current workload, minimizing energy consumption while ensuring performance. For intermittent working scenarios, the system also supports fast sleep and wake-up to further reduce standby power consumption.

[0118] Through the coordinated optimization of these hardware and software, our edge computing nodes are able to provide extremely high computing performance and real-time response capabilities within a limited power consumption budget. In actual tests, our optimization solution has improved the energy efficiency by 3 times compared to the traditional architecture, while reducing the average response time by 70%.

[0119] Reference Figure 3 In this embodiment, the efficient privacy-preserving computing framework includes lightweight homomorphic encryption, adaptive differential privacy, and secure multi-party computing protocols to achieve efficient data processing and analysis while protecting user privacy. Specifically, the following steps are included:

[0120] Step M1, lightweight homomorphic encryption: implement a partially homomorphic encryption scheme based on lattice cryptography, which significantly reduces computational overhead while ensuring security. Our scheme supports additive homomorphism and a limited number of multiplication operations, which is sufficient for many common statistical analysis and machine learning algorithms. The Ring-LWE problem is used as the security foundation. The encryption scheme uses the polynomial ring Rq = Zq[X] / (Xn+1) as the basic algebraic structure, where q is a prime number and n is a power of 2. This choice enables us to use fast Fourier transform (FFT) to accelerate polynomial multiplication, thereby improving overall performance;

[0121] Step M2, Adaptive Differential Privacy: Design an adaptive differential privacy mechanism. The core of the mechanism is a dynamic ε parameter adjustment algorithm, which can automatically adjust the privacy budget according to the sensitivity of the data and the frequency of queries.

[0122] The algorithm is based on an innovative privacy budget allocation strategy. The system allocates an initial privacy budget to each data subject and dynamically adjusts this budget based on the nature and frequency of the query. Specifically, we use a budget allocation method based on an exponential mechanism. For frequently queried data, the system will gradually reduce its available privacy budget, thereby preventing privacy leakage caused by excessive queries.

[0123] In order to minimize the impact of noise injection on data validity, an intelligent noise injection mechanism based on data distribution is implemented. This mechanism first performs cluster analysis on the data to identify the main distribution pattern of the data. Then, it adaptively adjusts the size and distribution of the noise according to the position of each data point in the distribution. For points in the center of the data distribution, we add smaller noise; for data points that may be outliers, we add larger noise. This method can preserve the statistical characteristics of the data to the greatest extent while protecting privacy.

[0124] A set of privacy protection measurement tools to evaluate the privacy protection level of the system in real time. This tool is based on multiple indicators, including ε-differential privacy measurement, information entropy, and K-anonymity. It can generate intuitive privacy protection reports to help system administrators and users understand the current privacy protection status and take additional protection measures when necessary;

[0125] Step M3, Secure Multi-Party Computation Protocol: Design a lightweight secure multi-party computation protocol suitable for edge computing scenarios. Based on secret sharing technology, it can achieve joint computing among multiple parties without revealing individual data. The protocol adopts an improved version of Shamir's secret sharing scheme. In the standard Shamir's scheme, the secret is divided into n shares, and as long as t shares (t < n) are available, the secret can be reconstructed. We optimized this by introducing a dynamic threshold mechanism. The system will dynamically adjust the number of shares t required for reconstructing the secret according to the current network conditions and node credibility. This mechanism improves the flexibility and robustness of the protocol;

[0126] To improve the computing efficiency, a distributed computing framework based on secret sharing is implemented. This framework supports common algebraic operations (such as addition, multiplication) and some basic statistical functions (such as mean, variance calculation). All these operations are carried out in the shared state without reconstructing the original data. We also implemented an efficient multiplication protocol that can complete the multiplication of two shared secrets with only one round of communication;

[0127] Design a secure protocol termination mechanism. During the computing process, if malicious behavior is detected (for example, a certain node provides obviously incorrect partial results), the system can safely terminate the computing process without revealing any intermediate results. This greatly improves the security of the protocol in an untrusted environment.

[0128] Furthermore, in Step M1, to optimize the key generation and encryption algorithms, we adopt the following techniques:

[0129] Use an optimized version of Gaussian sampling to reduce the overhead of random number generation, and implement polynomial multiplication based on Number Theoretic Transform (NTT), which is more suitable for computing in finite fields than the general FFT;

[0130] Adopt the Chinese Remainder Theorem (CRT) representation to handle large integers, which significantly improves the efficiency of modular operations. In addition, we also developed a hardware acceleration module specifically for edge devices. This module is implemented based on FPGA and includes dedicated NTT units and modular multipliers. On a typical Raspberry Pi 4 edge device, our hardware acceleration module can increase the speed of encryption and decryption operations by about 20 times, making real-time encrypted data processing possible.

[0131] By combining these technologies, our efficient privacy-preserving computing framework is able to provide strong privacy protection in edge computing environments while maintaining efficient computing performance. In actual tests, compared with traditional centralized privacy protection methods, our framework increases the computing speed by about 3 times and reduces data transmission by 80% while maintaining the same level of privacy protection.

[0132] Reference Figure 4 In this embodiment, the dynamic AI model update mechanism includes an incremental learning framework, a federated learning system, and knowledge distillation to achieve efficient, secure, and adaptable model updates, specifically including the following steps:

[0133] Step N1, incremental learning framework: Develop an incremental learning algorithm library that supports mainstream deep learning models. This library enables the model to continuously learn from new data without retraining. The core of the framework is a novel "elastic neural network" structure that allows the model to dynamically adjust its structure based on the characteristics of new data. The incremental learning framework includes:

[0134] Adaptive hierarchical structure: The model consists of multiple independently updateable sub-modules, each of which is responsible for a specific feature extraction or decision-making task. When new data arrives, the system will evaluate the performance of the existing sub-modules. If performance degradation is found, it will trigger the expansion of the sub-module or the creation of a new sub-module.

[0135] Dynamic neuron generation and pruning: We implemented a dynamic neuron generation algorithm that automatically adds new neurons to capture new features during training. We also implemented an importance-based neuron pruning mechanism that regularly removes unimportant neurons to keep the model compact.

[0136] Parameter importance assessment: We developed a parameter importance assessment algorithm based on the Fisher information matrix. This algorithm can identify the parameters that have the greatest impact on model performance and prioritize updating these key parameters during incremental learning, thereby improving learning efficiency.

[0137] Reversible residual connection: We designed a special reversible residual connection structure that allows the model to save intermediate activations during forward propagation, eliminating the need to recalculate these values ​​during backpropagation. This greatly reduces memory usage and enables incremental learning to be performed on resource-constrained edge devices.

[0138] In actual tests, our incremental learning framework can improve the model accuracy by 5% on image classification tasks using only 20% new data, while reducing the training time by 80% compared to complete retraining.

[0139] Step N2, Federated Learning System: A decentralized federated learning system is designed that can support heterogeneous edge devices and protect the privacy of model updates of individual devices through an innovative secure aggregation protocol, including:

[0140] Heterogeneous device adaptation layer: We have developed an automatic device capability detection and task allocation mechanism. The system will dynamically allocate suitable training tasks based on the computing power, storage capacity, and energy consumption characteristics of each device. For example, for devices with weak computing power, only training tasks for certain layers of the model may be allocated;

[0141] Secure aggregation protocol: We implemented a secure aggregation protocol based on homomorphic encryption. After each round of training, the device will encrypt its model update and then send it to the aggregation server. The server can directly aggregate on the encrypted domain without decrypting individual updates, thus protecting individual privacy.

[0142] Communication efficiency optimization: To reduce communication overhead, we developed a series of model compression and sparsification techniques;

[0143] Gradient quantization: Map the gradient value to a discrete set to reduce the number of bits of each gradient value;

[0144] Gradient sparsification: only the top-k% gradients with the largest absolute values ​​are transmitted, and the rest are set to zero;

[0145] Structured updates: force model updates to follow a predefined sparse structure, such as a low-rank matrix;

[0146] Asynchronous federated learning: Considering the instability of edge devices, we implemented an asynchronous federated learning algorithm, which allows devices to participate in training at their own pace without waiting for all devices to complete the current round;

[0147] In a simulated environment with 1,000 heterogeneous edge devices, our federated learning system is able to reduce communication overhead by 60% and shorten training time by 50% while maintaining model performance.

[0148] Step N3, knowledge distillation: Implement a set of model compression techniques based on knowledge distillation to migrate the knowledge of large models to small models, and dynamically adjust the distillation process according to the capabilities of edge devices and task requirements, including:

[0149] Multi-teacher ensemble distillation: We extend the traditional single-teacher distillation method to allow multiple expert models to serve as teachers simultaneously. Each teacher model focuses on a specific task or data domain, and the student model dynamically selects the most relevant teacher knowledge through an attention mechanism;

[0150] Hierarchical knowledge transfer: In addition to the final output, we also transfer knowledge at the intermediate level. We designed an adaptive feature matching algorithm that can establish the optimal mapping relationship between different layers of the teacher and student models;

[0151] Adversarial Distillation: We introduce a discriminator network that tries to distinguish the output of the student model from the output of the teacher model. The student model is trained to fool the discriminator, which encourages the student to generate output distributions closer to the teacher.

[0152] Dynamic temperature regulation: The temperature parameter in the distillation process is crucial to the effect of knowledge transfer. We implemented an adaptive temperature regulation mechanism to dynamically adjust the temperature according to the task difficulty and the learning progress of the student model.

[0153] Task-specific knowledge retention: For a specific task on an edge device, some knowledge may be more important than other knowledge. We developed a task importance assessment algorithm to ensure that the knowledge that is most important for the target task is retained during compression.

[0154] With this dynamic knowledge distillation technique, we are able to compress a ResNet-152 model trained on ImageNet to the size of MobileNetV2 with only a 2% loss in accuracy on the target task, while increasing inference speed by 15x and reducing model size by 97%.

[0155] The combination of these three technologies (incremental learning, federated learning, and knowledge distillation) forms a powerful dynamic AI model update mechanism that can continuously learn from new data, coordinate the knowledge of multiple devices, and efficiently migrate the capabilities of complex models to edge devices, thereby achieving efficient updating and deployment of AI models in edge computing environments.

[0156] Reference Figure 5 , Figure 6 In this embodiment, the intelligent resource scheduling system includes distributed resource monitoring, AI-driven task scheduling and elastic computing resource pool to achieve optimal resource utilization and efficient task execution, specifically including the following steps:

[0157] Step V1, Distributed Resource Monitoring: Develop a distributed resource monitoring system, the core of which is a lightweight resource monitoring agent that can continuously collect various resource usage information of edge nodes without significantly increasing the system burden, including:

[0158] Multi-dimensional resource monitoring: The agent can simultaneously monitor resource indicators in multiple dimensions, such as CPU usage, memory usage, storage I / O, network bandwidth, and energy consumption. We use efficient system calls and hardware counters to collect this information, minimizing the overhead of monitoring itself.

[0159] Adaptive sampling rate: The agent adopts an intelligent sampling strategy that can dynamically adjust the sampling frequency according to the fluctuation of resource usage. For example, when the CPU usage is stable, the sampling frequency can be reduced; when a burst load is detected, the sampling frequency is immediately increased to capture fine-grained changes;

[0160] Local preprocessing: To reduce the amount of data transmission, the agent will perform preliminary processing on the raw data locally, such as calculating the mean, variance, detecting outliers, etc. Only the processed summary data and important raw data points will be transmitted;

[0161] Compression coding: We implemented an adaptive compression algorithm based on the characteristics of time series. For indicators that change slowly, incremental coding is used; for indicators with large fluctuations, more precise coding is used;

[0162] In order to further optimize the transmission and processing of monitoring data, a hierarchical aggregation mechanism is designed, including:

[0163] Edge layer aggregation: Among multiple edge nodes in the same physical location, a node with better performance is selected as the local aggregation point, which is responsible for collecting and summarizing the monitoring data of surrounding nodes and performing preliminary anomaly detection and trend analysis;

[0164] Fog layer aggregation: fog nodes are set up at the regional level as intermediate aggregation points. Fog nodes receive data from multiple edge layer aggregation points and perform higher-level data analysis and compression.

[0165] Cloud layer aggregation: fog nodes are set up at the regional level as intermediate aggregation points. Fog nodes receive data from multiple edge layer aggregation points and perform higher-level data analysis and compression;

[0166] Adaptive aggregation cycle: The system dynamically adjusts the data aggregation and reporting cycle according to network conditions and data importance. When the network is congested, the aggregation cycle can be extended to reduce data transmission; when an abnormal situation is detected, the relevant data is reported immediately;

[0167] Through this distributed and hierarchical monitoring mechanism, we are able to reduce the amount of monitoring data transmission by 85% while ensuring monitoring accuracy, and the CPU usage of the monitoring agent is controlled below 1%.

[0168] Step V2, AI-driven task scheduling: Implement an intelligent task scheduling system based on reinforcement learning. The system makes optimal scheduling decisions based on task characteristics, resource conditions, and network topology. The reinforcement learning model includes:

[0169] State space: including resource usage of each node, current task queue, network topology, etc.;

[0170] Action space: including assigning tasks to specific nodes, adjusting task priorities, reserving resources, etc.

[0171] Reward function: Comprehensively considers factors such as task completion time, resource utilization, energy consumption, load balance, etc.

[0172] and learning algorithms; we used an improved Deep Q-Network (DQN) algorithm, adding prioritized experience replay and double Q learning to improve learning efficiency and stability;

[0173] To handle the high dynamics in edge computing environments, the model also includes:

[0174] Context-aware: The model is able to perceive the current application scenario (such as peak hours, emergency situations, etc.) and adjust its decision-making strategy accordingly;

[0175] Multi-agent collaboration: In large-scale systems, we deploy multiple collaborative scheduling agents. Each agent is responsible for local decision-making and coordinates with other agents through message passing mechanisms;

[0176] Online learning: The model can continuously learn and adapt to new patterns while making decisions, allowing the system to respond quickly to environmental changes;

[0177] Explainability module: We have integrated an explainability module that provides explanations for each scheduling decision, helping system administrators understand and trust AI decisions.

[0178] Furthermore, a predictive load balancing mechanism is designed to predict future resource requirements based on historical data and current trends, including:

[0179] Time series prediction: Use the LSTM (Long Short-Term Memory) model to predict the usage trends of various resources;

[0180] Periodic pattern recognition: Identify periodic patterns in resource usage (e.g., daily or weekly load cycles) through Fourier transform and wavelet analysis;

[0181] Event correlation analysis: Establish a correlation model between external events (such as large-scale events and holidays) and resource requirements;

[0182] Integration with prediction results: Use ensemble learning methods to combine the outputs of multiple prediction models to improve the accuracy and robustness of predictions;

[0183] Based on these prediction results, the system can make resource adjustments and task reallocations in advance to smooth load peaks and improve overall system efficiency.

[0184] In actual tests, our AI-driven scheduling system shortened the average task completion time by 30% and increased the overall system resource utilization by 25% compared to traditional heuristic algorithms, and can better handle sudden load situations.

[0185] Step V3, elastic computing resource pool: In order to maximize the use of available resources and provide flexible computing capabilities, a unified resource pool spanning edge devices, fog nodes, and cloud resources is built. The core of the resource pool is a distributed resource abstraction layer that can uniformly manage and schedule heterogeneous computing resources, including:

[0186] Resource virtualization: Use lightweight container technology (such as Docker) and a custom virtualization layer to abstract different types of computing resources into a unified computing unit;

[0187] Dynamic resource discovery: Implemented a dynamic resource discovery mechanism based on P2P network, which can automatically identify and integrate newly added computing nodes;

[0188] Elastic scaling: Based on the load conditions, the system can automatically add or remove computing nodes from the resource pool to achieve elastic scaling of resources;

[0189] Heterogeneous resource adaptation: A set of heterogeneous computing adaptation layers has been developed to automatically compile or convert tasks into a form suitable for specific hardware (such as GPU, FPGA);

[0190] In order to support the dynamic migration of tasks between different computing layers, an innovative cross-level task migration mechanism is designed, including:

[0191] Task state encapsulation: A lightweight task state encapsulation format is designed to capture task execution status, data dependencies and other information;

[0192] Incremental state transfer: When migrating tasks, only the state that has changed since the last checkpoint is transferred, greatly reducing the migration overhead;

[0193] Predictive migration: The system predicts future resource availability and network conditions and triggers task migration at the optimal time.

[0194] Data locality awareness: Migration decisions take into account the physical location of the data, and prioritize migrating tasks to nodes close to the data source;

[0195] Secure migration channel: An end-to-end encrypted secure channel is established to ensure data security during the task migration process.

[0196] Through this elastic resource pool and intelligent migration mechanism, we are able to achieve efficient utilization of computing resources. In a test environment containing 1,000 edge nodes, 100 fog nodes, and cloud resources, our system increased overall resource utilization by 40%, reduced the average task response time by 50%, and was able to effectively handle abnormal situations such as node failures and network fluctuations.

[0197] Reference Figure 7 , Figure 8 , Fig. 9 , Fig.10 In this embodiment, the high-reliability system architecture includes a multi-level fault-tolerant mechanism, a comprehensive security protection system, and an intelligent self-healing system to ensure the stability, security, and continuous availability of the entire edge computing network, specifically including the following steps:

[0198] Step K1: Multi-level fault tolerance mechanism: Design a multi-level fault tolerance mechanism, including:

[0199] Improved distributed consistency protocol: We designed an improved distributed consistency protocol based on the Paxos algorithm, called Edge-Paxos. The main innovations include:

[0200] Dynamic leader election: Introducing a dynamic leader election mechanism based on node health and network latency to improve the system's response speed;

[0201] Local consensus: In the event of a network partition, each partition is allowed to reach a local consensus and merge quickly after the network is restored;

[0202] Priority awareness: Set priorities for critical operations to ensure that important requests are processed first when resources are limited;

[0203] Lightweight implementation: The message delivery mechanism is optimized, the protocol overhead is reduced, and it is suitable for resource-constrained edge devices. In the test, Edge-Paxos increased system availability by 40% in the case of network partitions compared to standard Paxos, while reducing message overhead by 30%.

[0204] Intelligent data replication strategy: Implements an adaptive data multi-copy storage strategy that combines replication and erasure coding technologies, including:

[0205] Dynamic replication factor: Dynamically adjust the replication factor based on data importance, access frequency, and storage node reliability;

[0206] Tiered storage: hot data is stored on high-performance nodes, and cold data is stored on cheap storage using erasure codes;

[0207] Intelligent data placement: Optimize the distribution of data copies by taking into account network topology and geographic location;

[0208] Incremental erasure coding: For large data sets, incremental erasure coding technology is used to encode only the changed parts; this strategy reduces storage overhead by 40% compared to traditional fixed replication strategies while ensuring 99.999% data reliability;

[0209] Fast fault detection and recovery: Fast fault detection and recovery has developed a set of efficient fault detection and recovery mechanisms, including:

[0210] Distributed heartbeat monitoring: Use the Gossip protocol to achieve efficient distributed heartbeat monitoring and reduce the pressure on central nodes;

[0211] Fault prediction: Integrate machine learning models to predict potential faults based on historical data and current system status;

[0212] Rapid isolation: After a fault is detected, the affected nodes or services are immediately isolated to prevent the fault from spreading;

[0213] Intelligent service migration: Consider service dependencies and resource conditions to develop the best service migration plan;

[0214] Incremental state transfer: During service migration, only the changed state is transmitted, reducing migration time and bandwidth usage; through these mechanisms, we have shortened the average fault detection time to less than 200ms and reduced service recovery time by 60%.

[0215] Step K2: Comprehensive security protection system: Design a comprehensive security protection system, including blockchain-based identity authentication and access control, AI-driven intrusion detection and prevention, and end-to-end encrypted communication.

[0216] The blockchain-based identity authentication and access control implements a decentralized identity management and access control system, including:

[0217] Distributed identity registration: Use blockchain to record and verify device identity to prevent identity forgery;

[0218] Smart contract access control: define and enforce fine-grained access policies through smart contracts;

[0219] Dynamic permission adjustment: dynamically adjust access rights based on device behavior and reputation;

[0220] Cross-domain authentication: supports mutual trust and authentication between different management domains; this system makes identity management and permission control more transparent and secure, while improving the scalability of large-scale deployments;

[0221] AI-driven intrusion detection and prevention has developed an intelligent intrusion detection and prevention system, including:

[0222] Multimodal data fusion: Combine multiple data sources such as network traffic, system logs, and user behavior for analysis;

[0223] Deep learning model: Uses a combination of LSTM and CNN models to identify complex attack patterns;

[0224] Federated learning: Each edge node trains the model locally and the central node aggregates it, improving detection capabilities while protecting privacy;

[0225] Adaptive defense: Automatically adjust firewall rules and access policies based on detected threats;

[0226] Deception Defense: Automatically adjusts firewall rules and access policies based on detected threats; in testing, the system was able to detect 98% of known attacks and 85% of unknown attacks, with a false positive rate of less than 1%;

[0227] The end-to-end encrypted communication has designed a set of lightweight but secure end-to-end encrypted communication protocols, including:

[0228] Hybrid encryption: combines symmetric encryption and asymmetric encryption to balance security and efficiency;

[0229] Dynamic key management: Implement key negotiation mechanism based on quantum key distribution and regularly update session keys;

[0230] Secure multiplexing: Efficiently transmit multiple logical streams over a single encrypted channel;

[0231] Compression and encryption: compress first and then encrypt, reducing transmission overhead while ensuring security;

[0232] Selective encryption: Encryption algorithms of different strengths are selected according to the sensitivity of the data; this protocol reduces communication overhead by 25% while ensuring security, and is particularly suitable for resource-constrained edge devices.

[0233] Step K3, Intelligent Self-Healing System: Implemented an intelligent self-healing system, including machine learning-driven fault prediction, intelligent automatic operation and maintenance tools, and intelligent log analysis system.

[0234] The machine learning-driven fault prediction developed a multi-level fault prediction model, including:

[0235] Time series analysis: Use ARIMA and Prophet models to predict trends in system indicators;

[0236] Anomaly detection: Combining isolation forests and autoencoders to identify abnormal system behaviors;

[0237] Root cause analysis: Use the Bayesian network model to infer possible causes of failures;

[0238] Prediction fusion: Integrate the output of multiple models to improve the accuracy and robustness of predictions; this prediction system can predict 80% of system failures 5-30 minutes in advance, providing a valuable time window for preventive maintenance;

[0239] The intelligent automatic operation and maintenance tool has designed a set of automated operation and maintenance tools to support version control and intelligent rollback of system configuration, including:

[0240] Configuration version control: Use a Git-like version control system to record all configuration changes;

[0241] Smart rollback: Use a Git-like version control system to record all configuration changes;

[0242] Progressive deployment: Use the canary release strategy to gradually promote new configurations and reduce risks;

[0243] Configuration validity check: Use formal methods to verify the correctness of a new configuration before applying it;

[0244] Automated patch management: Intelligently schedules and applies system patches to minimize service interruptions; this suite of tools reduces manual intervention in routine operations tasks by 70% while improving the security of configuration changes;

[0245] The intelligent log analysis system has developed an intelligent log analysis system based on NLP and knowledge graph, including:

[0246] Log structuring: Use deep learning models to automatically identify key-value pairs in unstructured logs;

[0247] Semantic analysis: Use the BERT model to understand the semantic content of logs and identify key events;

[0248] Causal reasoning: building a knowledge graph of system behavior to infer the root cause of the problem;

[0249] Automated troubleshooting: Generates a decision tree for troubleshooting to guide technicians to quickly locate problems;

[0250] Continuous learning: Learn from feedback from human experts and continuously optimize the analysis model; this system reduces the average diagnosis time of complex problems from hours to minutes, greatly improving problem-solving efficiency.

[0251] Through these multi-level fault tolerance, security and self-healing mechanisms, our high-reliability system architecture can effectively cope with various challenges in edge computing environments and provide stable, secure and efficient services. In long-term testing, this architecture has enabled the system availability to reach 99.999%, while reducing manual operation and maintenance costs by 60%.

[0252] Reference Figure 11-Figure 14 In this embodiment, the intelligent energy consumption management includes dynamic power consumption optimization, energy-aware scheduling and intelligent thermal management to achieve energy-efficient operation of the edge computing system, which specifically includes the following steps:

[0253] Step H1, dynamic power consumption optimization: Implement a set of dynamic power consumption optimization mechanisms, including dynamic frequency adjustment algorithms and intelligent sleep strategies based on task importance and urgency.

[0254] The dynamic frequency adjustment algorithms based on task importance and urgency include:

[0255] Task classification: We developed a machine learning model to classify tasks into three levels of importance: high, medium, and low, based on their type, resource requirements, deadline, and other characteristics.

[0256] Real-time load analysis: Use sliding window technology to monitor system load in real time, and combine time series prediction models to estimate short-term load changes;

[0257] Dynamic Frequency Scaling (DVFS): Based on task importance and system load, we designed an adaptive DVFS algorithm. For important and urgent tasks, the system will increase the CPU frequency to ensure performance; for low-priority tasks, the frequency will be reduced to save energy;

[0258] Energy-performance balance: Introduce an adaptive balance factor β to dynamically adjust the weight of energy consumption and performance. The value of β is adjusted in real time according to the current battery power (for mobile devices) or electricity price (for fixed facilities);

[0259] The intelligent sleep strategies include:

[0260] Fine-grained power consumption model: We built a detailed power consumption model of hardware modules, including CPU cores, caches, memory controllers, network interfaces, etc.

[0261] Load forecasting: Use LSTM networks to predict short-term (1-5 minutes) and medium-term (1 hour) load changes;

[0262] Hierarchical sleep control: Based on the predicted load, the system selectively shuts down or reduces the power consumption of certain hardware modules, such as:

[0263] * Mild sleep: Turn off some CPU cores and caches.

[0264] *Moderate sleep: Reduce memory frequency and turn off non-essential peripherals.

[0265] *Deep sleep: only the network wake-up function is retained, and all other modules are turned off;

[0266] Fast wake-up: A hierarchical wake-up mechanism is designed to quickly restore system performance according to sudden loads;

[0267] Learning optimization: Use reinforcement learning to continuously optimize sleep strategies and balance energy saving effects and wake-up overhead; through these technologies, our dynamic power optimization mechanism can reduce energy consumption by 30-50% under typical workloads while ensuring that the performance of critical tasks is not affected.

[0268] Step H2, Energy-aware Scheduling: An energy-aware task scheduling system is developed, including a task scheduling algorithm that considers energy efficiency and task planning based on renewable energy prediction.

[0269] The task scheduling algorithms that consider energy efficiency include:

[0270] Energy efficiency modeling: Build a dynamic energy efficiency model for each compute node, taking into account its hardware characteristics, current load, and environmental factors (such as temperature);

[0271] Task energy consumption estimation: Use machine learning models to estimate the energy consumption of each task when executed on different nodes;

[0272] Multi-objective optimization: The task scheduling problem is modeled as a multi-objective optimization problem, taking into account energy consumption, completion time, and load balancing. The improved NSGA-III algorithm is used to solve the Pareto optimal solution;

[0273] Dynamic adjustment: Continuously adjust and optimize the scheduling strategy based on real-time monitored energy consumption data;

[0274] Mission planning based on renewable energy forecasts includes:

[0275] Energy forecasting: Combining meteorological data and historical power generation, using an ensemble learning method (Random Forest + Gradient Boosting) to predict short-term (within 24 hours) renewable energy output;

[0276] Energy-aware task queues: A dual-queue system is designed to divide tasks into a "green queue" (which prioritizes the use of renewable energy) and a "regular queue";

[0277] Dynamic task migration: Dynamically migrate tasks between different data centers or edge nodes based on the availability of renewable energy;

[0278] Peak-to-valley filling: When renewable energy is abundant, proactively increase system load (such as executing some non-urgent tasks in advance) to maximize green energy utilization;

[0279] Energy Storage Optimization: For facilities equipped with energy storage systems, we implement intelligent charging and discharging strategies to balance supply and demand and maximize the use of renewable energy; through these strategies, our system is able to increase the utilization of renewable energy to more than 85% while reducing conventional energy consumption by 30%.

[0280] Step H3, Intelligent Thermal Management: An intelligent thermal management system was designed, including an intelligent heat dissipation control system and a task migration mechanism based on temperature prediction.

[0281] The intelligent heat dissipation control system includes:

[0282] Refined temperature monitoring: deploy a high-precision temperature sensor network to monitor the temperature of each key point of the system in real time;

[0283] Thermal model construction: Use computational fluid dynamics (CFD) combined with machine learning methods to build an accurate system thermal model;

[0284] Adaptive cooling control: Dynamically adjust cooling strategy based on current load, ambient temperature and thermal model: * Fan speed adjustment: Use fuzzy control algorithm to smoothly adjust fan speed to reduce noise and energy consumption.

[0285] *Liquid Cooling Flow Control: For systems using liquid cooling, intelligently adjust the coolant flow and temperature.

[0286] * Cooling channel optimization: For modular data centers, dynamically adjust the airflow channel to optimize the overall cooling effect;

[0287] Predictive cooling: Combined with load prediction, the cooling strategy is adjusted in advance to avoid sudden temperature changes.

[0288] The task migration mechanism based on temperature prediction includes:

[0289] Hotspot prediction: Use the graph neural network (GNN) model to consider heat conduction between devices and predict potential hotspot areas;

[0290] Temperature-aware task scheduling: Adds temperature factors to conventional scheduling algorithms to avoid continuously assigning high-load tasks to high-temperature nodes;

[0291] Active task migration: When a node is predicted to be at risk of overheating, the tasks on it are actively migrated to a node with a lower temperature.

[0292] Thermal balance optimization: through task redistribution, the system temperature distribution is made more uniform, improving the overall heat dissipation efficiency;

[0293] Emergency cooling: For detected ultra-high temperature areas, the emergency cooling program is activated, and some modules are shut down if necessary to prevent hardware damage; through this thermal management system, we can reduce the maximum operating temperature of the system by 10°C and reduce cooling energy consumption by 15%. In extreme cases (such as refrigeration system failure), the system can ensure the continuous operation of core business and shorten the fault recovery time by 50%.

[0294] The intelligent energy management system achieves efficient and energy-saving operation of the edge computing system through the coordinated work of dynamic power consumption optimization, energy-aware scheduling and intelligent thermal management. In comprehensive tests, the system can reduce total energy consumption by 40% compared to traditional methods, while improving system performance by 15%, significantly reducing operating costs and improving system reliability.

[0295] Furthermore, the core algorithm of the DVFS strategy in step H1 can be expressed as:

[0296] f = fmin + (fmax - fmin) * (I * U + (1 - I) * L) * β, where f is the target frequency, fmin and fmax are the minimum and maximum frequencies respectively, I is the task importance (0-1), U is the urgency (0-1), L is the current load level (0-1), and β is the balance factor (0-1);

[0297] The core objective function of the dynamically adjusted scheduling algorithm in step H2 can be expressed as:

[0298] minimize F=w1*E+w2*T+w3*L, where E is the total energy consumption, T is the task completion time, L is the load imbalance, and w1, w2, and w3 are dynamically adjusted weights.

[0299] Furthermore, the core decision logic of the thermal management system in step H3 can be expressed as: if predicted_temp(node_i)>threshold:

[0300] tasks_to_migrate=select_tasks(node_i,temp_sensitivity)

[0301] target_nodes=find_cool_nodes(current_temp,predicted_load)

[0302] migrate_tasks(tasks_to_migrate,target_nodes)

[0303] adjust_cooling(node_i,aggressive_mode).

[0304] The embodiments of this specific implementation method are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An AI-based edge computing method, including an improved real-time data processing architecture, an efficient privacy-preserving computing framework, a dynamic AI model update mechanism, an intelligent resource scheduling system, a high-reliability system architecture, and intelligent energy consumption management, characterized in that: The specific steps include: Step 1: Improved real-time data processing architecture: including three main components: edge layer, fog layer and cloud layer. Each layer has its specific functions and optimization strategies to improve the real-time performance in edge computing systems. Step 2: Efficient privacy-preserving computing framework: Combining lightweight homomorphic encryption, adaptive differential privacy, and secure multi-party computing protocols to protect user privacy while achieving efficient data processing and analysis; Step 3: Dynamic AI model update mechanism: Combine incremental learning, federated learning, and knowledge distillation techniques for efficient, secure, and adaptable model updates; Step 4: Intelligent resource scheduling system: Combine distributed resource monitoring, AI-driven task scheduling, and elastic computing resource pools for optimal resource utilization and efficient task execution; Step 5: High-reliability system architecture: Combine multi-level fault tolerance mechanisms, comprehensive security protection systems, and intelligent self-healing systems to ensure the stability, security, and continuous availability of the entire edge computing network; Step 6: Intelligent energy management: Combine dynamic power consumption optimization, energy-aware scheduling, and intelligent thermal management for energy-efficient operation of edge computing systems.

2. The AI-based edge computing method according to claim 1, characterized in that: The improved real-time data processing architecture includes a layered stream processing framework, an adaptive stream control algorithm, and edge computing node optimization, and specifically includes the following steps: Step S1, hierarchical stream processing framework: at the edge layer, deploy lightweight data processing units, which are mainly responsible for preliminary data filtering and aggregation; at the fog layer, which is the intermediate computing layer, deploy an improved Apache Flink cluster to enhance its distributed computing capabilities; at the cloud layer, deploy a large-scale data processing cluster to execute complex machine learning models and deep data mining tasks; the analysis results of the cloud layer are fed back to the fog layer and the edge layer to optimize local decisions; a dynamic task allocation mechanism is developed, based on the improved Apache Flink, to increase the functions of task migration and dynamic expansion; the system will evaluate the computing requirements, data location and network status of each task in real time, and automatically determine at which layer the task should be executed; Step S2, adaptive flow control algorithm: It is deployed at all levels of the layered flow processing framework and plays a core role in the fog layer. The core of the algorithm is a closed-loop control system that continuously monitors key performance indicators. A prediction model based on machine learning is integrated into this algorithm. The model analyzes historical data and current trends to predict the load situation in the short term in the future and designs a set of adaptive parameter adjustment mechanisms. The system will automatically adjust the control parameters according to different application scenarios and performance goals; Step S3, edge computing node optimization: Comprehensive hardware and software optimization of edge computing nodes is carried out. A dedicated edge computing accelerator is designed at the hardware level. The accelerator is based on FPGA or ASIC technology and is optimized for specific stream processing tasks. A highly optimized software stack is developed at the software level. The software stack adopts a microkernel architecture to minimize operating system overhead, implement a lightweight task scheduler, accurately control the execution time of tasks, and develop an intelligent power management system. The system dynamically adjusts the CPU frequency and voltage according to the current workload.

3. The AI-based edge computing method according to claim 1, characterized in that: The efficient privacy-preserving computing framework includes lightweight homomorphic encryption, adaptive differential privacy, and secure multi-party computing protocols, and specifically includes the following steps: Step M1, lightweight homomorphic encryption: implement a partially homomorphic encryption scheme based on lattice cryptography, using the Ring-LWE problem as the security basis. The encryption scheme uses the polynomial ring Rq=Zq[X] / (Xn+1) as the basic algebraic structure, where q is a prime number and n is a power of 2; Step M2, Adaptive Differential Privacy: Design an adaptive differential privacy mechanism. The core of the mechanism is a dynamic ε parameter adjustment algorithm. The algorithm is based on an innovative privacy budget allocation strategy. The system allocates an initial privacy budget to each data subject and dynamically adjusts this budget according to the nature and frequency of the query, thus realizing an intelligent noise injection mechanism based on data distribution and a set of privacy protection measurement tools. Step M3, secure multi-party computing protocol: Design a lightweight secure multi-party computing protocol suitable for edge computing scenarios. Based on secret sharing technology, it can realize joint computing among multiple parties without leaking individual data. Design a secure protocol termination mechanism. If malicious behavior is detected during the calculation process, the system terminates the calculation process.

4. The AI-based edge computing method according to claim 3, characterized in that: In step M1, in order to optimize the key generation and encryption algorithm, we use the following techniques: Use an optimized version of Gaussian sampling to reduce the overhead of random number generation and implement polynomial multiplication based on NumberTheoretic Transform, which is more suitable for calculations on finite fields than general FFT; We use CRT representation to process large integers, which significantly improves the efficiency of modular operations. In addition, we have developed a hardware acceleration module specifically for edge devices. This module is implemented based on FPGA and includes a dedicated NTT unit and modular multiplier. On a typical Raspberry Pi 4 edge device, our hardware acceleration module can increase the speed of encryption and decryption operations by about 20 times, making real-time encrypted data processing possible.

5. The AI-based edge computing method according to claim 1, characterized in that: The dynamic AI model update mechanism includes an incremental learning framework, a federated learning system, and knowledge distillation, and specifically includes the following steps: Step N1, incremental learning framework: Develop an incremental learning algorithm library that supports mainstream deep learning models. The core of the framework is a novel "elastic neural network" structure, including adaptive hierarchical structure, dynamic neuron generation and pruning, parameter importance evaluation and reversible residual connection; Step N2, Federated Learning System: A decentralized federated learning system is designed to support heterogeneous edge devices, including heterogeneous device adaptation layer, security aggregation protocol, communication efficiency optimization, gradient quantization, gradient sparsification, structured update, and asynchronous federated learning; Step N3, knowledge distillation: Implement a set of model compression techniques based on knowledge distillation to migrate the knowledge of large models to small models, and dynamically adjust the distillation process according to the capabilities of edge devices and task requirements, including multi-teacher integrated distillation, hierarchical knowledge transfer, adversarial distillation, dynamic temperature adjustment, and task-specific knowledge retention.

6. The AI-based edge computing method according to claim 1, characterized in that: The intelligent resource scheduling system includes distributed resource monitoring, AI-driven task scheduling and elastic computing resource pool, and specifically includes the following steps: Step V1, distributed resource monitoring: Develop a distributed resource monitoring system, the core of which is a lightweight resource monitoring agent, including multi-dimensional resource monitoring, adaptive sampling rate, local preprocessing, compression encoding, and design a hierarchical aggregation mechanism, including edge layer aggregation, fog layer aggregation, cloud layer aggregation, and adaptive aggregation cycle; Step V2, AI-driven task scheduling: Implement an intelligent task scheduling system based on reinforcement learning. The system makes optimal scheduling decisions based on task characteristics, resource conditions, and network topology, including state space, action space, reward function, and learning algorithm. In order to handle the high dynamics in the edge computing environment, it also includes context awareness, multi-agent collaboration, online learning, and explainability modules. Design a predictive load balancing mechanism, including time series prediction, periodic pattern recognition, event correlation analysis, and prediction result integration; Step V3, elastic computing resource pool: Build a unified resource pool across edge devices, fog nodes and cloud resources. The core of the resource pool is a distributed resource abstraction layer, including resource virtualization, dynamic resource discovery, elastic scaling and heterogeneous resource adaptation. Design an innovative cross-level task migration mechanism, including task state encapsulation, incremental state transmission, predictive migration, data locality awareness and secure migration channel.

7. The AI-based edge computing method according to claim 1, characterized in that: The high reliability system architecture includes a multi-level fault tolerance mechanism, a comprehensive safety protection system and an intelligent self-healing system, and specifically includes the following steps: Step K1, multi-level fault tolerance mechanism: design a multi-level fault tolerance mechanism, including an improved distributed consistency protocol, intelligent data replication strategy and fast fault detection and recovery. The distributed consistency protocol includes dynamic leader election, local consensus, priority awareness, and lightweight implementation. The intelligent data replication strategy includes dynamic replication factor, tiered storage, intelligent data placement, and incremental erasure coding. Fast fault detection and recovery develops an efficient fault detection and recovery mechanism, including distributed heartbeat monitoring, fault prediction, fast isolation, intelligent service migration, and incremental state transfer. Step K2, comprehensive security protection system: design a comprehensive security protection system, including blockchain-based identity authentication and access control, AI-driven intrusion detection and defense, and end-to-end encrypted communication, where blockchain-based identity authentication and access control implements a decentralized identity management and access control system, including distributed identity registration, smart contract access control, dynamic permission adjustment, and cross-domain authentication, where AI-driven intrusion detection and defense develops an intelligent intrusion detection and defense system, including multimodal data fusion, deep learning models, federated learning, adaptive defense, and deception defense, where end-to-end encrypted communication designs a set of lightweight but secure end-to-end encrypted communication protocols, including hybrid encryption, dynamic key management, secure multiplexing, compressed encryption, and selective encryption; Step K3, intelligent self-healing system: An intelligent self-healing system is implemented, including machine learning-driven fault prediction, intelligent automatic operation and maintenance tools, and intelligent log analysis systems. The machine learning-driven fault prediction develops a multi-level fault prediction model, including time series analysis, anomaly detection, root cause analysis, and prediction fusion. The intelligent automatic operation and maintenance tool designs a set of automated operation and maintenance tools, including configuration version control, intelligent rollback, progressive deployment, configuration validity check, and automated patch management. The intelligent log analysis system develops an intelligent log analysis system based on NLP and knowledge graphs, including log structuring, semantic analysis, causal reasoning, automated troubleshooting, and continuous learning.

8. The AI-based edge computing method according to claim 1, characterized in that: The intelligent energy consumption management includes dynamic power consumption optimization, energy-aware scheduling and intelligent thermal management, and specifically includes the following steps: Step H1, dynamic power consumption optimization: a set of dynamic power consumption optimization mechanisms is implemented, including a dynamic frequency adjustment algorithm based on task importance and urgency and an intelligent sleep strategy, wherein the dynamic frequency adjustment algorithm based on task importance and urgency includes task classification, real-time load analysis, dynamic frequency scaling (DVFS) and energy consumption-performance balance, wherein the intelligent sleep strategy includes fine-grained power consumption model, load prediction, hierarchical sleep control, fast wake-up and learning optimization; Step H2, energy-aware scheduling: An energy-aware task scheduling system is developed, including a task scheduling algorithm that considers energy efficiency and a task planning based on renewable energy prediction. The task scheduling algorithm that considers energy efficiency includes energy efficiency modeling, task energy consumption estimation, multi-objective optimization and dynamic adjustment, and the task planning based on renewable energy prediction includes energy prediction, energy-aware task queues, dynamic task migration, peak-valley filling and energy storage optimization. Step H3, intelligent thermal management: An intelligent thermal management system is designed, including an intelligent heat dissipation control system and a task migration mechanism based on temperature prediction. The intelligent heat dissipation control system includes refined temperature monitoring, thermal model construction, adaptive heat dissipation control and predictive heat dissipation. The task migration mechanism based on temperature prediction includes hotspot prediction, temperature-aware task scheduling, active task migration, thermal balance optimization and emergency cooling.

9. The AI-based edge computing method according to claim 8, characterized in that: The core algorithm of the DVFS strategy in step H1 can be expressed as: f = fmin + (fmax - fmin) * (I * U + (1 - I) * L) * β, where f is the target frequency, fmin and fmax are the minimum and maximum frequencies respectively, I is the task importance (0-1), U is the urgency (0-1), L is the current load level (0-1), and β is the balance factor (0-1); The core objective function of the scheduling algorithm dynamically adjusted in step H2 can be expressed as: minimize F=w1*E+w2*T+w3*L, where E is the total energy consumption, T is the task completion time, L is the load imbalance, and w1, w2, and w3 are dynamically adjusted weights.

10. The AI-based edge computing method according to claim 8, characterized in that: The core decision logic of the thermal management system in step H3 can be expressed as: if predicted_temp(node_i)>threshold: tasks_to_migrate=select_tasks(node_i,temp_sensitivity) target_nodes=find_cool_nodes(current_temp,predicted_load)migrate_tasks(tasks_to_migrate,target_nodes)adjust_cooling(node_i,aggressive_mode).

Citation Information

Cited By

  • Security encryption communication method and system based on quantum key management

    CN119316138A

  • Edge computing deployment method oriented to manufacturing shop communication load balance

    CN119336521A

  • Chemical industrial park illegal behavior intervention and tracing system based on edge calculation

    CN119339334A

  • Self-adaptive task scheduling execution unit management method and system

    CN119376903A

  • Execution unit management method and system for adaptive task scheduling

    CN119376903B