Management optimization method for Internet of Things application platform

Abstract

The present invention discloses a management optimization method for an Internet of Things application platform, and relates to the technical field of the Internet of Things. The method adopts collaborative learning and management of an adaptive edge intelligent agent; introduces an adaptive edge intelligent agent AEIA model to give each edge node the ability to learn and make decisions autonomously; adopts a distributed collaborative learning mechanism; and adopts a self-organization and self-management strategy for edge resources based on context awareness: constructs a context-aware resource scheduling engine to monitor various environmental parameters including user behavior patterns, device energy consumption, and communication delays in the Internet of Things application platform in real time, and dynamically allocates and optimizes resources according to context changes; adopts a hierarchical self-management architecture to divide the self-management capabilities into device level, node level, and platform level; device-level self-management focuses on the monitoring and management of basic resources; intelligent energy efficiency management and optimization strategies based on prediction-driven; introduces an adaptive energy consumption prediction model; and adopts a dynamic adjustment of the device's operating mode and task scheduling strategy.

Description

Technical Field

[0001] The invention discloses a management optimization method for an Internet of Things application platform and relates to the technical field of the Internet of Things. Background Art

[0002] The supply chain hierarchical management optimization system based on the intelligent Internet of Things described in the Chinese invention patent CN118520995A mainly solves the problem that the supply chain management is too dependent on manual work, resulting in insufficient optimization and increased operating costs for upstream and downstream merchants. Through the synchronous collection of the bottom-end demand information of the supply chain and the upper-level supply source information by the Internet of Things technology, the control of the entire supply chain process information is realized, and hierarchical optimization management is carried out on this basis. The specific method includes a comprehensive assessment of the upstream supply capacity, downstream demand capacity and historical stability of the supply chain, and the implementation of full-process control to ensure the operational stability of the supply chain; in the case of not meeting the preset optimal selection conditions, intelligent selection analysis is carried out to promptly replace the supply source that meets the conditions. Although this method has achieved certain results in the field of supply chain management, it still has some drawbacks in the management and optimization of the Internet of Things application platform.

[0003] First, the main focus of existing technologies is on the information collection and control optimization of the entire process of supply chain management. Its optimization goal focuses on the dynamic matching and optimal selection of upstream and downstream resources in the supply chain, and focuses on how to improve the operating efficiency and stability of the supply chain. However, this method mainly focuses on specific application scenarios, especially in the field of supply chain management, and lacks sufficient consideration for the comprehensive management needs of edge computing, device management, data processing and resource optimization that are widely present in the IoT platform. Over-reliance on centralized information collection and analysis to optimize decision-making ignores the autonomous management and dynamic optimization capabilities of distributed nodes in the IoT platform. The number of edge nodes in the IoT environment is huge and widely distributed. The collaboration and adaptive adjustment between nodes are the key to ensuring the efficient operation of the entire system. However, this method fails to make full use of the distributed characteristics of edge computing, resulting in difficulty in achieving efficient resource scheduling and real-time optimization in the face of large-scale, multi-node complex IoT environments.

[0004] Secondly, the existing technology relies on a centralized supply chain management and control mode. Although it can collect and analyze upstream and downstream information of the supply chain in real time, this centralized management mode has significant bottleneck problems in large-scale Internet of Things platforms. In the Internet of Things platform, the number of node devices is huge and the types are diverse. Each node device may contain different sensors, actuators, data processing units and communication modules. In such a complex distributed environment, relying on a single centralized system for global management and optimization is prone to problems such as data transmission delay, network congestion and single point failure, and cannot guarantee the real-time nature of the data and the stability of the system. For example, when a node in the Internet of Things platform is abnormal or fails, relying on the central system for fault detection and processing will increase the response time of the system and may reduce the operating efficiency of the entire platform. In addition, this centralized model also has data privacy and security issues. The centralized storage and processing of a large amount of sensitive data can easily become the target of network attacks. At the same time, there are limitations in exception handling. Its exception handling mechanism mainly relies on the preset optimal selection conditions for optimal analysis and continuous adjustment, without combining real-time data and multiple anomaly detection models for dynamic optimization. In the IoT platform, abnormal situations are diverse and change rapidly, such as equipment failure, communication interruption, network attack and data leakage, which require more flexible and real-time abnormal detection and processing mechanisms. Based on the optimized autoencoder model and multi-layer anomaly detection network, real-time detection and response of abnormal behaviors can be better achieved, but this method lacks the corresponding intelligent anomaly detection and processing capabilities. In the face of complex and changing IoT environments, its ability to ensure system security is relatively weak. Summary of the invention

[0005] The purpose of the present invention is to provide a management optimization method for an Internet of Things application platform, thereby solving some of the drawbacks and shortcomings pointed out in the background technology.

[0006] The present invention solves the above-mentioned technical problems by adopting the following technical solutions: A management optimization method for an Internet of Things application platform, comprising: S1, collaborative learning and management using an adaptive edge intelligent agent:

[0007] S1.1. Introduce the Adaptive Edge Intelligent Agent AEIA model to give each edge node the ability to learn and make decisions autonomously; each AEIA dynamically perceives the device density, data flow, and task complexity information in the area, and autonomously selects the adaptive role including data collector, processor, or relay;

[0008] S1.2, adopting a distributed collaborative learning mechanism, so that AEIA can share model parameters or feature extraction methods with adjacent edge nodes when processing local data;

[0009] S2. Context-aware edge resource self-organization and self-management strategy:

[0010] S2.1. Build a context-aware resource scheduling engine to monitor various environmental parameters in the IoT application platform, including user behavior patterns, device energy consumption, and communication delays in real time, and dynamically allocate and optimize resources based on context changes;

[0011] S2.2. Introduce a self-organizing mechanism at the edge node level, so that each node can perceive the surrounding resources including computing power, storage space, and bandwidth and self-balance with other nodes;

[0012] S2.3, adopt a hierarchical self-management architecture, dividing the self-management capabilities into device level, node level and platform level; device level self-management focuses on the monitoring and management of basic resources; node level self-management focuses on the optimization of local resources and task scheduling; platform level self-management coordinates the global optimization and exception handling of the entire network;

[0013] S3. Intelligent energy efficiency management and optimization strategy based on prediction drive:

[0014] S3.1, introduce an adaptive energy consumption prediction model, combined with the autoregressive integrated moving average model ARIMA, to predict the future energy consumption requirements of IoT devices and edge nodes;

[0015] S3.2. Dynamically adjust the equipment's operating mode and task scheduling strategy, prioritize key tasks for efficient equipment, and extend the standby time of inefficient equipment.

[0016] Furthermore, the construction of the adaptive edge intelligent agent AEIA model includes: on the edge node of the Internet of Things platform, the AEIA model optimizes the role selection in the network environment through reinforcement learning; the AEIA model dynamically perceives the network status and task requirements, selects the role and defines the adaptive learning optimization function F opt (s,a), the function combines the adaptability of the environment state s and the action strategy a:

[0017]

[0018] Where s represents the environment state vector, including the current load, data flow, and device density information of the edge node; a represents the action strategy vector of AEIA, covering multiple predictions of different role selections; σ(t, s, a) is the role adaptability function, which is used to describe the fitness of a specific environment state s and strategy a at time t; η is the balance factor, which is used to adjust the weights of different factors in role selection; ρ i is the weight adjustment coefficient of node i, which indicates the relative importance of different nodes in the role selection process; Δ(s,a) is the strategy difference function, which measures the adaptability difference between different strategies.

[0019] Furthermore, the construction of the adaptive edge intelligent agent AEIA model includes: multiple AEIA agents collaborate and compete in the edge computing environment of the Internet of Things platform to complete tasks together; each agent negotiates with other agents to determine its role allocation based on its own computing power and network load; the negotiation process is performed by the role negotiation function C collab (r1, r2) description, defined as follows:

[0020]

[0021] Among them, r1 and r2 represent the resource vectors of the two AEIA nodes, including the computing power, bandwidth and storage capacity of each node; P eff (·) represents the resource utility function, which evaluates the work efficiency and resource utilization of each AEIA node under different resource configurations; ω k is the negotiation weight coefficient, which indicates the weight ratio of each node in the overall decision during the negotiation process; K is the total number of nodes participating in the negotiation.

[0022] Furthermore, the construction of the adaptive edge intelligent agent AEIA model includes: the AEIA model dynamically allocates tasks and load balances by monitoring the network status and task requirements of edge nodes in real time, and the load balancing optimization function L balance (w,z) is defined as follows:

[0023]

[0024] Among them, w is the task load vector, which indicates the amount and type of tasks currently processed by each edge node; z is the node state vector, which reflects the real-time status of the node, such as computing power, remaining power, network delay, etc.; κ(t) is the load priority weight coefficient, which is used to dynamically adjust the priority of different tasks.

[0025] Furthermore, the hierarchical self-management architecture includes device-level self-management, node-level self-management and platform-level self-management. In the device-level self-management, each IoT device is equipped with a single lightweight local management agent, which is responsible for monitoring and managing basic resources. The local management agent dynamically adjusts the working mode of the device to optimize resource usage by deploying a local management module based on an adaptive rule engine, and defines a device resource status optimization function:

[0026]

[0027] Among them, R opt (t) represents the resource optimization objective function in the time period T, which indicates the comprehensive resource management effect of the equipment at different times; α(t) is the balance coefficient of the equipment state adjustment at time t, which is used to weigh the relationship between energy consumption and performance; Indicates that at time t, the energy consumption function E usage (t) relative to the device working mode S mode (t) represents the change rate of energy consumption under different working modes; β(t) is the data acquisition frequency adjustment coefficient at time t, which is used to adjust the data acquisition strategy of the device under different load conditions; M freq (t) is the data collection frequency function at time t, which represents the strategy of the device to adaptively adjust the data collection frequency.

[0028] Furthermore, in the node-level self-management, each node is equipped with an autonomous management agent, which is responsible for managing the data processing, caching and forwarding of multiple IoT devices connected locally, as well as the dynamic scheduling of tasks; a local resource optimization module based on graph neural network is used to control the dynamic allocation of node resources; the optimization function of node task scheduling is defined as:

[0029]

[0030] Among them, T opt (x) is the optimization function of node-level task scheduling, which represents the overall efficiency of task scheduling and resource allocation of all nodes under given conditions; x is the node state vector, including the load, bandwidth, and computing power information of each node; U task (x i ) is the task utility function, which represents the utility of executing a specific task on node i and evaluates the efficiency of the task in using node resources; R node (x i ) is the node resource allocation function, describing the resource usage of node i;

[0031] In addition, λ i is the task scheduling priority weight, indicating the priority of different tasks during scheduling; γ i is the resource allocation balance coefficient, which is used to weigh the fairness of resource allocation between nodes; It is the communication optimization item of Graph Neural Networks (GNN) based on topological structure learning, which describes the communication optimization effect between node i and node j.

[0032] Furthermore, in the platform-level self-management machine, the platform-level management agent is responsible for coordinating the global optimization and exception handling of the entire network, and adjusts the global optimization strategy in real time by summarizing and analyzing the status information from each node; platform-level self-management relies on the global optimization model based on federated learning and the monitoring system based on anomaly detection, and defines the global optimization function as:

[0033]

[0034] Among them, P opt(w) is the global optimization objective function, which is used to measure the comprehensive effectiveness of the platform-level management system in global optimization and anomaly detection; w is the model weight vector in federated learning, which represents the set of model parameters summarized from each node; F global (w) is the global utility function, which represents the utility of the overall optimization strategy across nodes and measures the effect of model update on the improvement of global performance; δ(t) is the global learning balance coefficient at time t, which is used to adjust the weight of global optimization and local node learning; ζ(t) is the anomaly detection priority coefficient, which determines the priority balance between anomaly detection and global optimization;

[0035] Among them, A detect (w, t) is an anomaly detection function, which detects abnormal network behaviors including sudden equipment failures, network attacks or data leaks in real time based on the optimized autoencoder model.

[0036] The management optimization method of the Internet of Things application platform of the present invention has the following beneficial effects: first, by introducing the adaptive edge intelligent agent (AEIA) model, each edge node in the Internet of Things platform can learn and make decisions autonomously, so that the system can dynamically select roles (such as data collectors, processors or relayers) according to real-time network status and task requirements, improve the resource utilization efficiency and task processing capabilities of edge nodes, reduce the load pressure of central servers, and improve the response speed and stability of the system. Secondly, based on the local resource optimization module of distributed collaborative learning and graph neural network, edge nodes can realize dynamic allocation of resources and adaptive scheduling of tasks locally, avoid the bottleneck of centralized control, and improve the resource utilization and data processing efficiency of the entire Internet of Things platform. Thirdly, through the device-level, node-level and platform-level self-management mechanisms of the hierarchical self-management architecture, each level realizes targeted resource optimization and task management, for example, device-level self-management realizes the monitoring and management of basic resources, node-level self-management focuses on the optimization and task scheduling of local resources, and platform-level self-management coordinates global optimization and exception handling to ensure that the system can operate efficiently and stably in a complex environment. In addition, platform-level self-management relies on the global optimization model of federated learning and the monitoring system based on anomaly detection, which can achieve global resource optimization and exception handling while protecting data privacy, thereby enhancing the security and robustness of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flow chart of the management optimization method of the Internet of Things application platform of the present invention.

[0038] Figure 2 A flow chart for constructing the adaptive edge intelligent agent AEIA model of the present invention.

[0039] Figure 3 This is a flow chart of the hierarchical self-management architecture of the present invention. DETAILED DESCRIPTION

[0040] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0041] Combined with Figure 1 Flowchart, the management optimization method of the Internet of Things application platform of the present invention comprises the following steps:

[0042] S1. Collaborative learning and management using adaptive edge intelligent agents:

[0043] S1.1. Introduce an adaptive edge intelligent agent model to give each edge node the ability to learn and make decisions autonomously. Each AEIA uses a built-in perception module to dynamically monitor and perceive information such as device density, data flow, and task complexity in its area in real time. Based on this perceived information, AEIA can autonomously select the appropriate role to perform specific tasks. For example, when the device density of an edge node is high and the data flow is large, AEIA can choose to act as a "data collector" to centrally process and cache data, thereby reducing the load pressure on the central server. When a node has strong computing power and storage resources, AEIA can choose the "data processor" role to pre-process, analyze or filter local data to reduce the delay in data transmission. When a node is in a key position for data transfer, AEIA can choose the "data relay" role to efficiently forward the processed data to the next node to optimize the data transmission path of the entire network. This adaptive role selection is based on a reinforcement learning algorithm. Through continuous environmental feedback and self-adjustment, AEIA can optimize its own role allocation and working status under different environmental conditions, thereby improving the resource utilization efficiency and task execution capability of edge nodes.

[0044] S1.2. A distributed collaborative learning mechanism is adopted to enable AEIA to share model parameters or feature extraction methods with adjacent edge nodes when processing local data. Each AEIA updates its own machine learning model parameters or feature extraction methods through a local training process, and sends these updated results as partial parameters to adjacent nodes for knowledge sharing and collaborative learning. This distributed collaborative learning mechanism can effectively utilize the computing power of each node to achieve rapid knowledge dissemination and efficient model updating, while avoiding the centralized transmission of large amounts of raw data, reducing network bandwidth usage and the risk of privacy data leakage. Through this collaborative approach, each AEIA model can acquire a large amount of learning experience in a short period of time, improve the accuracy and generalization ability of the overall model, and enable the entire IoT application platform to more intelligently manage and optimize itself in a complex and dynamic environment, ultimately achieving the purpose of improving the overall performance of the system, enhancing the ability to respond to dynamic changes, and saving resources and energy consumption.

[0045] S2. Context-aware edge resource self-organization and self-management strategy:

[0046] S2.1. Construct a context-aware resource scheduling engine. The context-aware resource scheduling engine monitors various environmental parameters within the IoT application platform in real time, including user behavior patterns, device energy consumption, communication delay, etc. Specifically, user behavior patterns include information such as the user's device usage frequency, data transmission requirements, and task processing requirements; device energy consumption refers to the current energy consumption status of each IoT device, involving the device's working mode, operating time, and remaining power; communication delay covers the data transmission time and network delay between different nodes; the scheduling engine uses multimodal data fusion technology to comprehensively analyze the changes in these environmental parameters, and dynamically models and perceives the context of the IoT platform in real time by building a model based on a graph convolutional network. After obtaining the context information, the scheduling engine dynamically allocates and optimizes resources according to the current environmental changes. For example, when the user behavior pattern shows a high-frequency data transmission requirement, the scheduling engine will prioritize allocating more bandwidth and computing resources to the corresponding node; when it is detected that some devices are in a high-energy consumption state, the scheduling engine can reduce their workload or transfer some tasks to nodes with lower energy consumption to achieve overall energy efficiency optimization;

[0047] S2.2. Introduce a self-organizing mechanism at the edge node level, so that each node can autonomously perceive the surrounding resource situation, including computing power, storage space, bandwidth, etc. Through autonomous perception, the node can judge its current resource usage status and the resource gap with other nodes, and perform self-balancing accordingly. The self-balancing mechanism is based on the negotiation algorithm of game theory. Each node dynamically adjusts resource configuration by exchanging information with neighboring nodes. For example, when a node detects that its own resources are insufficient and the neighboring nodes have surplus resources, the node can request the neighboring nodes to assist in completing some tasks, or forward data to the neighboring nodes for processing to reduce load pressure. This self-organizing mechanism enables each node to optimize resources locally, avoiding the bottleneck of centralized control and improving the response speed and resource utilization efficiency of the entire IoT platform;

[0048] S2.3. A hierarchical self-management architecture is adopted, and self-management capabilities are divided into device level, node level and platform level. Device-level self-management mainly focuses on the monitoring and management of basic resources, such as device battery status, memory utilization, sensor status, etc., and optimizes the device operation status through the adaptive adjustment function of the local management agent; node-level self-management focuses on the optimization of local resources and task scheduling. Each edge node is responsible for managing the data processing, caching and forwarding of multiple locally connected devices through an autonomous management agent, and dynamically adjusts the priority and resource allocation of tasks using reinforcement learning and graph neural network technology to ensure efficient data processing and communication; platform-level self-management coordinates the global optimization and exception handling of the entire network, and realizes global resource optimization and exception monitoring through global data analysis, cross-node collaboration and policy decision-making. The platform-level management agent summarizes and analyzes the status information from each node, uses the federated learning model to update the global optimization strategy, and adopts anomaly detection mechanism to detect and handle abnormal behaviors in the network in real time, such as device failures or network attacks, thereby ensuring the security and stability of the IoT application platform.

[0049] S3. Intelligent energy efficiency management and optimization strategy based on prediction drive:

[0050] S3.1. Introduce an adaptive energy consumption prediction model. The energy consumption prediction model is combined with an autoregressive integrated moving average model to analyze and predict the future energy consumption requirements of IoT devices and edge nodes. Specifically, the ARIMA model uses time series data to model the historical energy consumption of devices and nodes, identify trends and seasonal changes, and thus accurately predict future energy consumption. The autoregressive part (AR) of the ARIMA model predicts the current and future energy consumption status based on the past energy consumption values ​​of the device, and the integral part (I) handles the non-stationarity in the time series and converts the non-stationary energy consumption data into stationary data through differential operations. The moving average (MA) part uses historical prediction errors to correct the current energy consumption prediction value to ensure the accuracy and reliability of energy consumption prediction. In this prediction model, the dynamic energy consumption characteristics of IoT devices and edge nodes under different environmental conditions are considered, such as the operating frequency of the device, network status, data processing load and other factors. These factors are integrated into the ARIMA model as adaptive variables, so that the model can adjust the prediction results in real time according to environmental changes. Through accurate prediction of future energy consumption demand, the IoT platform can formulate corresponding resource scheduling and energy efficiency optimization strategies in advance to reduce the risk of excessive energy consumption and resource waste of equipment;

[0051] S3.2. Dynamically adjust the operation mode and task scheduling strategy of the equipment. Based on the results of the energy consumption prediction model, the system adjusts the operation mode and task scheduling plan of the equipment in real time, giving priority to assigning key tasks to efficient equipment, while extending the standby time of inefficient equipment. Specifically, during the task scheduling process, the system dynamically assigns the priority and execution order of tasks according to the energy efficiency and current working status of each device. Efficient equipment, that is, equipment with relatively low energy consumption and superior performance, will be given priority in assigning key tasks. These tasks usually require high computing power and low-latency response, such as data-intensive analysis, real-time processing tasks, etc.; while inefficient equipment, that is, equipment with relatively high energy consumption, will be assigned key tasks first. Devices with high or poor current status are scheduled to perform low-priority tasks or be in standby mode. By extending the standby time of these devices, their workload and energy consumption are reduced. In addition, the system adopts a reinforcement learning algorithm to continuously optimize the scheduling strategy, learn and adapt to different device performance and task requirements. In actual operation, the system will dynamically adjust the device's working mode according to the real-time status of the device and the predicted energy consumption changes. For example, when the device power is lower than a certain threshold, the system will switch it to low-power mode to reduce energy consumption, or when the device is idle for more than a certain period of time, the system will put it into deep sleep mode to further save energy.

[0052] Embodiment 1:

[0053] See attached Figure 2 In a large-scale smart city IoT application platform, multiple edge nodes are distributed in different urban areas. Each edge node is responsible for monitoring and managing various IoT devices in its area, such as cameras, temperature and humidity sensors, traffic light controllers, and environmental quality monitoring instruments. These edge nodes use the Adaptive Edge Intelligent Agent (AEIA) model to optimize their role selection in the network environment through reinforcement learning to more efficiently complete task management, data processing, and resource allocation. Specifically, the AEIA model dynamically selects the optimal role, such as data collector, data processor, or data relay, by sensing the network status (such as device density, data traffic, current load, etc.), thereby optimizing the overall system performance.

[0054] The AEIA model first collects the environmental state vector s of the edge node, including the following information: current load (such as CPU usage and memory occupancy), data traffic (such as the number of packets per second), device density (such as the number of IoT devices covered by each node). Assume that an edge node in a certain area is monitoring 500 IoT devices, with a data traffic of 200 packets per second, a CPU usage of 70%, and a memory usage of 65%. After receiving this state information, the AEIA model must decide its optimal action strategy vector a, for example: continue to act as a "data collector" to collect more data, or convert to a "data processor" to process existing data, or act as a "data relay" to pass data to a higher-level edge node or cloud server.

[0055] To make this decision, the AEIA model uses an adaptive learning optimization function F opt (s,a) is used to evaluate the adaptability of different strategies. The optimization function is defined as follows:

[0056]

[0057] Among them, σ(t,s,a) is the role fitness function, which is used to describe the fitness of a specific environment state s and strategy a at time t. In this scenario, σ(t,s,a)=0.8×(1-load factor)+0.2×data flow coefficient, where the load factor is the average of the node's current CPU usage (70%) and memory occupancy (65%), that is, 0.675; the data flow coefficient is related to the transmission volume of 200 data packets per second, standardized to between 0 and 1, and set to 0.6. Therefore, σ(t,s,a)=0.8×(1-0.675)+0.2×0.6=0.26+0.12=0.38.

[0058] On this basis, the AEIA model uses the balance factor η to adjust the weights of different factors in role selection. In this scenario, η = 0.5 is set to balance the influence between adaptability and strategy differences. At the same time, ρ i is the weight adjustment coefficient of node i, which indicates the relative importance of different nodes in the role selection process. In this scenario, there are 5 adjacent nodes, and the weight coefficient ρ i They are 0.1, 0.2, 0.25, 0.15, and 0.3 respectively, reflecting the importance of each node to task collaboration and data processing.

[0059] Δ(s,a) is a strategy difference function, which is used to measure the adaptability difference between different strategies. Assuming the current strategy is data collector, if the strategy difference coefficient of changing to data processor is 0.2, and the strategy difference coefficient of changing to data relay is 0.1, then Δ(s,a) can be expressed as the difference degree based on these strategy changes, and its value is set to 0.15.

[0060] Substituting these values ​​into the optimization function F opt (s,a), we can calculate:

[0061] F opt (s,a)=∫0 T 0.38dt+0.5×(0.1×0.15+0.2×0.15+0.25×0.15+0.15×0.15+0.3×0.15)

[0062] First calculate the integral part, setting the time T = 10 seconds:

[0063] ∫0 10 0.38dt=0.38×10=3.8

[0064] Then calculate the weighted part:

[0065] 0.5×(0.1×0.15+0.2×0.15+0.25×0.15+0.15×0.15+0.3×0.15)

[0066] =0.5×(0.015+0.03+0.0375+0.0225+0.045)=0.5×0.15=0.075

[0067] Therefore, the optimization function F opt The value of (s,a) is:

[0068] F opt (s,a)=3.8+0.075=3.875

[0069] Based on this result, the AEIA model will evaluate the effect of the role selection strategy. The adaptability score of the "data processor" role is set to 3.875, while the adaptability scores of other roles are 3.2 (data collector) and 3.6 (data relayer), so the AEIA model will select "data processor" as the current best role to maximize the resource utilization efficiency and task processing capacity of the node.

[0070] Multiple edge nodes in the smart city IoT application platform use the Adaptive Edge Intelligent Agent (AEIA) model to collaborate and compete to complete tasks together. Each AEIA agent negotiates with other agents based on its own computing power, network load, and storage capacity to determine its best role configuration in the current environment. The negotiation process uses the role negotiation function C collab (r1, r2) to ensure that different nodes can collaborate efficiently.

[0071] It is assumed that there are three edge nodes in different areas of the city, namely nodes A, B and C. The resource vector r1 of node A includes 60% computing power, 80Mbps bandwidth, and 100GB storage capacity; the resource vector r2 of node B includes 70% computing power, 70Mbps bandwidth, and 120GB storage capacity; the resource vector r3 of node C includes 50% computing power, 90Mbps bandwidth, and 150GB storage capacity. Since node A has lower computing power but higher bandwidth, it is more suitable to act as a data relay; node B has higher computing power and storage capacity, and is more suitable as a data processor; and node C has obvious advantages in bandwidth and storage capacity, and can flexibly switch roles between data collection and data forwarding.

[0072] The role negotiation process is described using the following role negotiation function:

[0073]

[0074] Where r1 and r2 represent the resource vectors of two AEIA nodes A and B respectively; P eff (r) is the resource utility function, which evaluates the work efficiency and resource utilization of each AEIA node under different resource configurations; ω k is the negotiation weight coefficient, which indicates the weight ratio of each node in the overall decision; K is the total number of nodes participating in the negotiation.

[0075] In this scenario, the resource utility functions of nodes A and B are set as follows:

[0076] P eff (r1) = 0.6·computing power + 0.3·bandwidth + 0.1·storage capacity = 0.6×60+0.3×80+0.1×100

[0077] =36+24+10=70

[0078] P eff (r2) = 0.5·computing power + 0.4·bandwidth + 0.1·storage capacity = 0.5×70+0.4×70+0.1×120

[0079] =35+28+12=75

[0080] Next, the negotiation contribution value between the two nodes is calculated, that is, and Set these contributions to:

[0081]

[0082] Set the total number of nodes participating in the negotiation K = 3, and the negotiation weight coefficient of each node ω k are 0.3, 0.4 and 0.3. Substitute into the negotiation function:

[0083]

[0084] The negotiation function value C obtained by calculation collab (r1, r2) = 0.1, indicating the degree of cooperation between nodes A and B under the current resource configuration. When the negotiation results show that the resource utilization gap between the two nodes is small, they will tend to share tasks and data, thereby optimizing the overall resource usage.

[0085] In the actual smart city environment, node A is assigned to the role of data relay, node B is the data processor, and node C dynamically adjusts between data collection and forwarding. During a day's operation, node A can transmit a large amount of data to node B with low latency due to its high bandwidth; node B uses its high computing power and storage space to quickly process the received data and feed it back to the cloud server or other key decision-making systems; and node C flexibly switches between data collection and forwarding according to the density of sensor data to ensure the optimal data transmission path.

[0086] The effect of this strategy in actual operation shows that the average delay of node A is reduced by 20%, the data processing speed of node B is increased by 15%, and the network load of node C is effectively balanced, thereby improving the overall performance and stability of the entire IoT platform.

[0087] To further optimize resource utilization and system performance, the AEIA model dynamically allocates tasks and achieves load balancing by monitoring the network status and task requirements of edge nodes in real time. Each edge node is equipped with an AEIA model, which uses the load balancing optimization function L balance (w,z) to ensure the load balancing and efficiency maximization of each node. In this platform, the three main edge nodes A, B and C continue to participate in the cooperation of task processing and data transmission. They need to dynamically adjust the task allocation according to their respective resource conditions to cope with the changing network status.

[0088] The load balancing optimization function is defined as follows:

[0089]

[0090] Among them, w is the task load vector, which indicates the amount and type of tasks currently being processed by each edge node. For example, node A is processing 200 data analysis tasks, node B is responsible for 300 data forwarding tasks, and node C performs 150 real-time video streaming tasks; z is the node state vector, which reflects the real-time status of each node, such as computing power, remaining power, and network delay. The computing power of node A is 70%, the remaining power is 50%, and the network delay is 100ms; the computing power of node B is 80%, the remaining power is 60%, and the network delay is 120ms; the computing power of node C is 60%, the remaining power is 40%, and the network delay is 90ms. The load priority weight coefficient κ(t) is a time function, which is used to dynamically adjust the priority of different tasks. In this case, κ(t) is set to 1.2-0.01t, which means that as time goes by, the priority of the task gradually decreases, ranging from [0.8,1.2].

[0091] First calculate the key part of the load optimization function Set F opt (w) = 0.5 × task type coefficient + 0.3 × task quantity coefficient + 0.2 × node resource utilization. For node A, set its task type coefficient (data analysis task) to 0.7, task quantity coefficient to 200 / 500 = 0.4 (500 is the maximum number of tasks), and node resource utilization is the percentage of computing power, 70%, or 0.7, then:

[0092] F opt (w)=0.5×0.7+0.3×0.4+0.2×0.7=0.35+0.12+0.14=0.61

[0093] F opt (w) calculates the partial derivative with respect to the node state vector z, assuming that each state (computing power, remaining power, network delay) has the same impact on task allocation, and the derivative value is 0.1. Therefore:

[0094]

[0095] Substitute into the load balancing optimization function:

[0096] L balance (w,z)=∫0 10 [(1.2-0.01t)×0.1]dt

[0097] Calculate the integral part:

[0098]

[0099] The calculation results are:

[0100] L balance (w,z)=(0.12×10-0.0005×100)-(0.12×0-0.0005×0)=1.2-0.05=1.15

[0101] This value represents the optimization target value of node A under the current load state. Similarly, the optimization values ​​of nodes B and C can be calculated. The task type coefficient of node B is set to 0.6 (data forwarding task), the task type coefficient of node C is set to 0.8 (video streaming task), the task quantity coefficients are 300 / 500=0.6 and 150 / 500=0.3, and the resource utilization of nodes B and C are 0.8 and 0.6, respectively. Calculate the F of nodes B and C opt (w):

[0102] F opt (w B )=0.5×0.6+0.3×0.6+0.2×0.8=0.3+0.18+0.16=0.64

[0103] F opt (w C )=0.5×0.8+0.3×0.3+0.2×0.6=0.4+0.09+0.12=0.61

[0104] The corresponding derivative value is also 0.1, and the load balancing optimization function of nodes B and C is calculated:

[0105] L balance (w B ,z)=∫0 10 (0.12-0.001t)dt=1.15

[0106] L balance (w C ,z)=∫0 10 (0.12-0.001t)dt=1.15

[0107] According to the calculation results, the load balancing optimization function values ​​of nodes A, B, and C are all 1.15, indicating that the current task allocation and resource utilization status are relatively balanced. If the value of a node deviates from this balance value, the AEIA model will adjust the task allocation, such as reducing the task volume or lowering the priority, to return to a balanced state.

[0108] Embodiment 2:

[0109] like Figure 3As shown in the figure, in order to improve the energy efficiency and overall resource management capabilities of the devices in the smart city IoT application platform, the system adopts a hierarchical self-management architecture, including device-level self-management, node-level self-management, and platform-level self-management. In device-level self-management, each IoT device, such as a city air quality monitor, smart street light, or smart trash can, is equipped with a lightweight local management agent. This local management agent is responsible for monitoring and managing the basic resources of the device (such as battery power, CPU utilization, and memory usage), and uses an adaptive rule engine to dynamically adjust the device's working mode to optimize resource usage.

[0110] For example, in a smart street light network in a city, each street light is equipped with a light sensor, a communication module, and a control unit. The local management agent deploys a local management module based on an adaptive rule engine to monitor the device status in real time, including battery power, brightness requirements, and environmental changes. For example, when it is detected that the ambient brightness drops below the set threshold, the management agent of the smart street light will dynamically adjust the working mode according to the actual situation, such as increasing the brightness or enabling low power consumption mode.

[0111] In order to optimize resource usage, the device resource status optimization function is defined:

[0112]

[0113] Among them, R opt (t) represents the resource optimization objective function in the time period T, which is used to measure the comprehensive resource management effect of the equipment at different times; α(t) is the balance coefficient of the equipment state adjustment at time t, which is used to balance between energy consumption and performance; Indicates that at time t, the energy consumption function E usage (t) relative to the device working mode S mode The rate of change of (t) reflects the change of energy consumption under different working modes; β(t) is the data acquisition frequency adjustment coefficient at time t, which determines the data acquisition strategy of the device under different load conditions; M freq (t) is the data collection frequency function at time t, which represents the strategy of the device to adaptively adjust the data collection frequency according to environmental conditions.

[0114] The smart street light is set to operate at night in the time period T = 12 hours, and the working state needs to be adjusted according to different time periods and ambient brightness. Between t = 0 and t = 6 hours (midnight to early morning), due to the small number of pedestrians and vehicles, the smart street light enters low power consumption mode, and the value range of α(t) is between [0.4, 0.6], giving priority to reducing energy consumption. At this time, It means that in low power mode, energy consumption is reduced by 30%; and from t = 6 to t = 12 hours (early morning to daytime), as the environment gradually becomes brighter, the smart street light gradually dims until it turns off, and the value range of α(t) is between [0.2, 0.3], mainly to ensure energy efficiency and reduce brightness.

[0115] Data acquisition frequency adjustment coefficient β(t) and data acquisition frequency function M freq (t) reflects the adaptive response capability of the intelligent street light to the ambient brightness. For example, from time t = 0 to t = 6 hours, β(t) = 0.5, indicating that the data collection frequency needs to be moderate to maintain the necessary brightness monitoring; while from t = 6 to t = 12 hours, β(t) = 0.8, increasing the data collection frequency to respond to the brightness change at sunrise, M freq (t) = 1.2-0.1t, and its value range is [0.6, 1.2], which means that as the ambient brightness increases, the data collection frequency gradually decreases.

[0116] Substitute these actual data into the device resource status optimization function:

[0117]

[0118] In the time interval t=0 to t=6 hours:

[0119] R opt (t) = ∫0 6 (0.5·(-0.3)+0.5·(1.2-0.1t))dt=∫0 6 (-0.15+0.5×(1.2-0.1t))dt

[0120] =∫0 6 (-0.15+0.6-0.05t)dt=∫0 6 (0.45-0.05t)dt

[0121] Calculate the integral part:

[0122]

[0123] Similarly, for the time interval t=6 to t=12 hours:

[0124] R opt (t) = ∫6 12 (0.3·(-0.1)+0.8·(1.2-0.1t))dt=∫6 12 (-0.03+0.96-0.08t)dt

[0125] =∫6 12 (0.93-0.08t)dt

[0126] Calculate the integral part:

[0127]

[0128] Finally, the device resource status optimization target R opt (t) The total value in the entire time period T = 12 hours is:

[0129] R opt (t) = 1.8 + 1.26 = 3.06

[0130] This result shows that through the dynamic adjustment of the working mode and data collection frequency by the local management agent of the smart street lamp, effective resource management and energy-saving optimization can be achieved under a given balance between energy consumption and performance, making the entire IoT platform run more efficiently and better meet the application needs of smart cities.

[0131] To further optimize the utilization efficiency of IoT node resources, the system introduces autonomous management agents in node-level self-management. Each node is equipped with an autonomous management agent, which is responsible for managing the data processing, caching and forwarding of multiple IoT devices connected locally, as well as the dynamic scheduling of tasks. Each autonomous management agent uses a local resource optimization module based on Graph Neural Networks (GNN) to monitor and control the dynamic allocation of node resources in real time, thereby improving the resource utilization and task processing efficiency of the entire system.

[0132] Consider three key edge nodes A, B, and C distributed in different areas of the city, connected to multiple air quality monitors, smart street lights, and traffic signal controllers. These nodes need to be resource-scheduled by autonomous management agents to ensure efficient real-time data processing and transmission. Set the state vector x1 of node A to include a load of 70%, a bandwidth of 60Mbps, and a computing power of 80%; the state vector x2 of node B to include a load of 50%, a bandwidth of 70Mbps, and a computing power of 75%; the state vector x3 of node C to include a load of 65%, a bandwidth of 50Mbps, and a computing power of 70%.

[0133] In order to achieve optimal scheduling of node tasks, the system defines the optimization function of node task scheduling:

[0134]

[0135] Among them, T opt (x) is the optimization function of node-level task scheduling, which represents the overall efficiency of task scheduling and resource allocation of all nodes under given conditions; x is the node state vector, including the load, bandwidth and computing power information of each node; U task (x i) is the task utility function, which represents the utility of executing a specific task on node i and evaluates the efficiency of the task in using node resources; R node (x i ) is the node resource allocation function, describing the resource usage of node i; λ i is the task scheduling priority weight, indicating the priority of different tasks during scheduling; γ i is the resource allocation balance coefficient, which is used to weigh the fairness of resource allocation between nodes; It is the communication optimization item of GNN based on topology structure learning, which describes the communication optimization effect between node i and node j.

[0136] Assume that nodes A, B, and C are performing data processing, data forwarding, and data collection tasks respectively. The task utility function U of node A is task (x1) = 0.7·computing power + 0.3·load, the task utility function of node B is U task (x2) = 0.6·computing power + 0.4·load, the task utility function of node C is U task (x3)=0.8·computing power+0.2·load.

[0137] Based on these utility function calculations, we set The change rates of are: 0.15 for node A, 0.12 for node B, and 0.1 for node C. Resource allocation function R node (x i ) represents the resource usage of the node, and the resource usage of node A is set to 70%, node B to 50%, and node C to 65%.

[0138] Introducing task scheduling priority weight λ i and resource allocation balance coefficient γ i . Under the current task requirements, λ i They are node A0.4, node B0.3, node C0.3, γ i They are node A0.3, node B0.4, and node C0.3 respectively. Communication optimization items based on topology learning Indicates the communication optimization effect between node i and node j in the network topology. Set

[0139] Substituting these values ​​into the optimization function:

[0140] T opt (x)=(0.4·0.15+0.3·(0.8·60+0.2·80))+(0.3·0.12+0.4·(0.8·70+0.2·75))

[0141] +(0.3 0.1+0.3(0.8 50+0.2 70))

[0142] Simplify the calculation of each part:

[0143] T opt (x)=(0.4·0.15+0.3·(48+16))+(0.3·0.12+0.4·(56+15))

[0144] +(0.3 0.1+0.3(40+14))

[0145] =(0.06+19.2)+(0.036+28.4)+(0.03+16.2)=19.26+28.436+16.23=63.926

[0146] The calculated T opt (x) = 63.926, which indicates the overall efficiency of the current node task scheduling and resource allocation. Through this optimization function, the autonomous management agent can identify the best task and resource allocation strategy for each node. In actual operation, node A will continue to focus on data processing tasks, node B will be responsible for data forwarding, and node C will flexibly perform data collection. At the same time, the communication optimization items between nodes The high efficiency ensures the optimal data transmission path, thereby further improving the operating efficiency of the system.

[0147] To further optimize the global performance of the IoT platform, the platform-level management agent is responsible for coordinating the global optimization and exception handling of the entire network. The platform-level management agent adjusts the global optimization strategy in real time by aggregating and analyzing the status information from each node (such as nodes A, B, and C). In this process, platform-level self-management relies on a global optimization model based on federated learning and a monitoring system based on anomaly detection to ensure that the entire IoT system can respond quickly to any abnormal situation while operating normally.

[0148] The global optimization function is defined as:

[0149]

[0150] Among them, P opt (w) is the global optimization objective function, which is used to measure the comprehensive effectiveness of the platform-level management system in global optimization and anomaly detection; w is the model weight vector in federated learning, which represents the set of model parameters summarized from each node; F global(w) is the global utility function, which represents the utility of the overall optimization strategy across nodes and measures the effect of model update on the improvement of global performance; δ(t) is the global learning balance coefficient at time t, which is used to adjust the weight of global optimization and local node learning; ζ(t) is the anomaly detection priority coefficient, which determines the priority balance between anomaly detection and global optimization; A detect (w, t) is an anomaly detection function that detects abnormal behaviors in the network in real time based on the optimized autoencoder model, including sudden device failures, network attacks, or data leaks.

[0151] Assume that within the current time period T = 24 hours, the platform-level management agent needs to optimize the resource utilization efficiency of the nodes and ensure the security of the system. The model weight vectors w of nodes A, B, and C are [0.6, 0.7, 0.8] respectively. These weights represent the optimal model parameters obtained by each node in the latest federated learning process. Global utility function F global (w) can be in the form of 0.4×node computing efficiency + 0.6×data transmission efficiency, assuming that the computing efficiency of node A is 0.75 and the data transmission efficiency is 0.8; the computing efficiency of node B is 0.7 and the data transmission efficiency is 0.85; the computing efficiency of node C is 0.65 and the data transmission efficiency is 0.9.

[0152] Compute the global utility function:

[0153] F global (w)=0.4×(0.75+0.7+0.65) / 3+0.6×(0.8+0.85+0.9) / 3=0.4×0.7+0.6×0.85

[0154] =0.28+0.51=0.79

[0155] Next, find the partial derivative of the global utility function with respect to the model weight w, and set its partial derivative to a constant of 0.05. The value range of the global learning balance coefficient δ(t) is [0.5, 1.0], which is used to adjust the global optimization weight at different times t, and its value is set to δ(t) = 0.8. The anomaly detection priority coefficient ζ(t) is used to adjust the balance between anomaly detection and global optimization, and its value range is set to [0.2, 0.7]. At the current time point t = 12 hours, the value ζ(t) = 0.5. Anomaly detection function A detect (w, t) Based on the autoencoder model, it detects sudden equipment failures and network attacks in real time, and sets the probability of detecting anomalies to 0.2 at the current moment.

[0156] Substitute into the global optimization function:

[0157] P opt (w) = ∫0 24(0.8·0.05+0.5·0.2)dt=∫0 24 (0.04+0.1)dt

[0158] =∫0 24 0.14dt=0.14×24=3.36

[0159] This calculation result P opt (w) = 3.36 represents the comprehensive performance score of the platform-level management agent in balancing global optimization and anomaly detection within 24 hours. Based on this optimization function, the platform-level management agent can effectively judge the overall operating status of the system. If the value is lower than the preset threshold (set to 3.0), the system needs to readjust the model weights or enhance the anomaly detection strength to ensure the security and efficiency of the IoT platform.

[0160] Through this global optimization strategy based on federated learning and anomaly detection, the platform-level management agent can dynamically adjust the model parameters and global learning strategies of each node. For example, when an abnormal behavior of a node is detected (such as an abnormal increase in data transmission at node C), the platform management agent can increase the priority coefficient ζ(t) of anomaly detection, improve the ability to respond to potential threats, and simultaneously adjust the resource utilization strategies of other nodes to ensure the security and stable operation of the entire IoT platform.

[0161] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A management optimization method for an Internet of Things application platform, characterized in that The following steps are involved: S1. Collaborative learning and management using adaptive edge intelligent agents: S1.

1. Introduce the Adaptive Edge Intelligent Agent AEIA model to give each edge node the ability to learn and make decisions autonomously; each AEIA dynamically perceives the device density, data flow, and task complexity information in the area, and autonomously selects the adaptive role including data collector, processor, or relay; S1.2, adopting a distributed collaborative learning mechanism, so that AEIA can share model parameters or feature extraction methods with adjacent edge nodes when processing local data; S2. Context-aware edge resource self-organization and self-management strategy: S2.

1. Build a context-aware resource scheduling engine to monitor various environmental parameters in the IoT application platform, including user behavior patterns, device energy consumption, and communication delays in real time, and dynamically allocate and optimize resources based on context changes; S2.

2. Introduce a self-organizing mechanism at the edge node level, so that each node can perceive the surrounding resources including computing power, storage space, and bandwidth and self-balance with other nodes; S2.3, adopt a hierarchical self-management architecture, dividing the self-management capabilities into device level, node level and platform level; device level self-management focuses on the monitoring and management of basic resources; node level self-management focuses on the optimization of local resources and task scheduling; platform level self-management coordinates the global optimization and exception handling of the entire network; S3. Intelligent energy efficiency management and optimization strategy based on prediction drive: S3.1, introduce an adaptive energy consumption prediction model, combined with the autoregressive integrated moving average model ARIMA, to predict the future energy consumption requirements of IoT devices and edge nodes; S3.2, dynamically adjust the operation mode and task scheduling strategy of the equipment, give priority to assigning key tasks to efficient equipment, and extend the standby time of inefficient equipment; The construction of the adaptive edge intelligent agent AEIA model includes: on the edge node of the Internet of Things platform, the AEIA model optimizes the role selection in the network environment through reinforcement learning; the AEIA model dynamically perceives the network status and task requirements, selects the role and defines the adaptive learning optimization function F opt (s,a), the function combines the adaptability of the environment state s and the action strategy a: Where s represents the environment state vector, including the current load, data flow, and device density information of the edge node; a represents the action strategy vector of AEIA, covering multiple predictions of different role selections; σ(t, s, a) is the role adaptability function, which is used to describe the fitness of a specific environment state s and strategy a at time t; η is the balance factor, which is used to adjust the weights of different factors in role selection; ρ i is the weight adjustment coefficient of node i, which indicates the relative importance of different nodes in the role selection process; Δ(s,a) is the strategy difference function, which measures the adaptability difference between different strategies.

2. The management optimization method of the Internet of Things application platform according to claim 1 is characterized in that The construction of the adaptive edge intelligent agent AEIA model includes: multiple AEIA agents collaborate and compete in the edge computing environment of the Internet of Things platform to complete tasks together; each agent negotiates with other agents to determine its role allocation based on its own computing power and network load; the negotiation process is performed by the role negotiation function C collab (r1, r2) description, defined as follows: Among them, r1 and r2 represent the resource vectors of the two AEIA nodes, including the computing power, bandwidth and storage capacity of each node; P eff (·) represents the resource utility function, which evaluates the work efficiency and resource utilization of each AEIA node under different resource configurations; ω k is the negotiation weight coefficient, which indicates the weight ratio of each node in the overall decision during the negotiation process; K is the total number of nodes participating in the negotiation.

3. The management optimization method of the Internet of Things application platform according to claim 2 is characterized in that The construction of the adaptive edge intelligent agent AEIA model includes: the AEIA model monitors the network status and task requirements of edge nodes in real time, dynamically allocates tasks and balances the load.

4. The management optimization method of the Internet of Things application platform according to claim 1 is characterized in that The hierarchical self-management architecture includes device-level self-management, node-level self-management and platform-level self-management. In the device-level self-management, each IoT device is equipped with a single lightweight local management agent, which is responsible for monitoring and managing basic resources. The local management agent dynamically adjusts the device's working mode to optimize resource usage by deploying a local management module based on an adaptive rule engine, and defines a device resource status optimization function: Among them, R opt (t) represents the resource optimization objective function in the time period T, which indicates the comprehensive resource management effect of the equipment at different times; α(t) is the balance coefficient of the equipment state adjustment at time t, which is used to weigh the relationship between energy consumption and performance; Indicates that at time t, the energy consumption function E usage (t) relative to the device working mode S mode (t) represents the change rate of energy consumption under different working modes; β(t) is the data acquisition frequency adjustment coefficient at time t, which is used to adjust the data acquisition strategy of the device under different load conditions; M freq (t) is the data collection frequency function at time t, which represents the strategy of the device to adaptively adjust the data collection frequency.

5. The management optimization method of the Internet of Things application platform according to claim 4 is characterized in that In the node-level self-management, each node is equipped with an autonomous management agent, which is responsible for managing the data processing, caching and forwarding of multiple locally connected IoT devices, as well as the dynamic scheduling of tasks; a local resource optimization module based on graph neural networks is used to control the dynamic allocation of node resources.

6. The management optimization method of the Internet of Things application platform according to claim 5 is characterized in that In the platform-level self-management machine, the platform-level management agent is responsible for coordinating the global optimization and exception handling of the entire network, and adjusts the global optimization strategy in real time by summarizing and analyzing the status information from each node; platform-level self-management relies on the global optimization model based on federated learning and the monitoring system based on anomaly detection, and defines the global optimization function as: Among them, P opt (w) is the global optimization objective function, which is used to measure the comprehensive effectiveness of the platform-level management system in global optimization and anomaly detection; w is the model weight vector in federated learning, which represents the set of model parameters summarized from each node; F global (w) is the global utility function, which represents the utility of the overall optimization strategy across nodes and measures the effect of model update on the improvement of global performance; δ(t) is the global learning balance coefficient at time t, which is used to adjust the weight of global optimization and local node learning; ζ(t) is the anomaly detection priority coefficient, which determines the priority balance between anomaly detection and global optimization; Among them, A detect (w, t) is an anomaly detection function, which detects abnormal network behaviors including sudden equipment failures, network attacks or data leaks in real time based on the optimized autoencoder model.

Images