Artificial intelligence data management and distribution system based on Internet of Things

Abstract

The invention discloses an artificial intelligence data management and distribution system based on the Internet of Things, and the system comprises a state collection module, a dynamic graph construction module, a multi-agent module, a model reconstruction module, a reward construction module, a model training module, and a decision generation module. Constructing a dynamic graph state space containing a node and data interaction relationship, establishing a multi-agent reinforcement learning basic model based on a graph attention mechanism, and introducing a dynamic graph deep reinforcement learning framework to reconstruct a state modeling structure, a state transition relationship and a strategy updating path; and carrying out online training on the multi-agent decision model by combining a graph structure level reward feedback mechanism, and finally generating a data management and distribution decision strategy. According to the method, collaborative modeling of the data distribution decision and the equipment state evolution relationship in the Internet of Things environment is realized, and the adaptability to the complex topological relationship and the dynamic state change in the data management and distribution process is improved.

Description

Technical Field

[0001] This invention relates to the field of data management and intelligent scheduling technology that integrates the Internet of Things (IoT) and artificial intelligence (AI), and in particular to an AI-based data management and distribution system based on the IoT. Background Technology

[0002] Against the backdrop of the continuous development of IoT technology, various sensing devices, edge nodes, and data service nodes are being deployed on a large scale in application scenarios such as industrial manufacturing, smart cities, energy management, and public services, forming a highly distributed and dynamically changing data generation and interaction environment. During operation, IoT systems continuously generate a large amount of heterogeneous data, which needs to be managed and distributed among different device nodes and processing nodes to support status monitoring, business collaboration, and decision execution. Existing IoT data management and distribution technologies typically rely on preset data scheduling rules or static configuration strategies to forward and process data based on device type, network topology, or fixed priorities. Their core implementation methods are mostly based on rule matching, threshold judgment, or centralized control logic.

[0003] As IoT systems continue to expand in scale and their operating environments become increasingly complex, the constraints faced in the data distribution process exhibit highly dynamic characteristics, including frequent changes in node load states, significant fluctuations in network latency, and the continuous evolution of data interaction relationships. To address these changes, some existing technologies have introduced artificial intelligence methods to optimize the data management process, such as using machine learning models to predict node states or adjusting data scheduling strategies through reinforcement learning. However, these technologies often treat node states as independent inputs, lacking system modeling of the complex relationships between nodes, and thus failing to comprehensively depict the overall state of multi-node collaborative operation within an IoT system.

[0004] On the other hand, existing IoT data distribution schemes employing reinforcement learning typically rely on a single agent or a centralized decision-making structure. When faced with the need for parallel decision-making across multiple nodes, these schemes are prone to problems such as high decision coupling and an expansion of the state space dimension. Furthermore, some multi-agent methods fail to effectively incorporate the dynamic topological characteristics of IoT systems in practical applications, neglecting the impact of changes in inter-node interactions over time on data distribution decisions, resulting in insufficient model adaptability.

[0005] Therefore, how to provide an AI-based data management and distribution system based on the Internet of Things is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an AI-based data management and distribution system for the Internet of Things (IoT). Addressing the problems of existing IoT data management and distribution processes relying on preset rules, difficulty adapting to dynamic changes in operational states, and inability to form autonomous decisions, this invention proposes a data management and distribution technology solution that combines dynamic graph state modeling, graph attention multi-agent reinforcement learning, and dynamic graph deep reinforcement learning. By dynamically modeling the operational states and interaction relationships of device nodes, a multi-agent decision-making model is constructed, and a graph-based reward feedback and online training mechanism are introduced to achieve continuous updating and generation of data management and distribution decisions. This invention possesses the beneficial effect of adaptive modeling and decision updating of the data distribution process in a dynamic IoT environment.

[0007] According to an embodiment of the present invention, an artificial intelligence data management and distribution system based on the Internet of Things includes the following modules: Status acquisition module: Collects the operating status data of device nodes in the IoT system, performs time synchronization and normalization processing on the operating status data, and obtains the IoT operating status dataset; Dynamic graph construction module: Based on the IoT operation status dataset, a dynamic graph state space is constructed, which maps device nodes to nodes in the dynamic graph, maps the data interaction relationships between nodes to edges in the dynamic graph, and associates them with the corresponding operation status attributes to obtain the initial dynamic graph state representation. Multi-agent module: Taking the initial dynamic graph state representation as input, initialize the Graph-Attention multi-agent reinforcement learning basic model, configure the corresponding agent for each node, and perform weighted modeling of the nodes and their adjacency relationships in the dynamic graph through the graph attention mechanism to obtain the Graph-Attention multi-agent reinforcement learning basic model; Model Reconstruction Module: Introducing a dynamic graph deep reinforcement learning framework into the Graph-Attention multi-agent reinforcement learning base model, the dynamic graph state representations at multiple consecutive time steps are used as joint inputs to reconstruct the state modeling structure, state transition relationships, and policy update paths in the Graph-Attention multi-agent reinforcement learning base model, resulting in an improved multi-agent decision model. Reward Construction Module: Based on the improved multi-agent decision-making model, a graph-structured reward feedback mechanism is constructed to map the changes in the running state after the distribution and execution of IoT data into node-level reward information, and the reward is propagated along the node association relationship in the dynamic graph to obtain graph-structured reward data; Model training module: The improved multi-agent decision model is trained and its parameters are updated online using graph-structured reward data to obtain the trained multi-agent decision model; Decision generation module: Inputs the dynamic graph state representation of the current moment into the trained multi-agent decision model, generates and outputs data management and distribution decision strategies.

[0008] Optionally, the status acquisition module specifically includes: Raw status data reflecting the operating status of devices are collected from device nodes in the Internet of Things system. The raw status data includes the processing load information, communication latency information, data cache usage information, and data interaction records between device nodes, forming a set of raw operating status data of device nodes. Perform timestamp marking, time alignment, and data normalization on various types of status data in the original operating status data set of device nodes to generate an IoT operating status dataset.

[0009] Optionally, the dynamic graph construction module specifically includes: Based on the IoT operation status dataset, the operation status data is indexed and organized according to the device node identifier to form node status data records corresponding to the device nodes. Based on node status data records, a unique node identifier is assigned to each device node, and the device node is mapped to a node in a dynamic graph to generate a node set; Based on the data interaction relationship information reflected in the IoT operation status data, the data interaction connection relationship between device nodes is determined, and the data interaction connection relationship is mapped to the edge in the dynamic graph to generate an edge set; The node load status, communication latency status, and data cache occupancy status in the node status data record are associated with the corresponding dynamic graph nodes, and the interaction frequency and interaction direction in the data interaction connection relationship are associated with the corresponding dynamic graph edges, forming a dynamic graph structure data with attributes. Perform integrity checks on dynamic graph structure data with attributes and output the initial dynamic graph state representation.

[0010] Optionally, the multi-agent module specifically includes: Based on the initial dynamic graph state representation, the set of nodes, the set of edges, and the running state attributes associated with the nodes in the dynamic graph are obtained to form dynamic graph structure data containing node attributes and edge attributes. Based on the node set, configure a corresponding intelligent agent for each device node, and establish a one-to-one correspondence between device node identifiers and intelligent agent identifiers to generate a multi-agent set; Based on the edge set, the connection relationship information between device nodes is extracted, a graph structure information reflecting the topological relationship of device nodes is constructed, and the graph structure information is combined with the running state attributes associated with the nodes to form multi-agent input data; Introducing a graph attention computation process into a multi-agent ensemble, determining the adjacency relationship between nodes based on graph structure information, performing weighted computation of adjacency nodes on the multi-agent input data, and generating a node state representation containing adjacency node information; Based on the node state representation, the node state representation is collected and organized according to the agent identifier to form a multi-agent state representation corresponding to each agent; Output the multi-agent state representation to complete the construction of the Graph-Attention multi-agent reinforcement learning basic model and obtain the basic multi-agent state representation.

[0011] Optionally, the model reconstruction module specifically includes: Based on the basic multi-agent state representation, and by acquiring the basic multi-agent state representations corresponding to multiple consecutive time points in chronological order, time series state data is formed. Based on time-series state data, we construct time-series input data to describe the relationship between device node states and time changes, and then input the time-series input data into a dynamic graph deep reinforcement learning framework. In the dynamic graph deep reinforcement learning framework, time correlation modeling is performed on the temporal input data to generate a temporal state representation that reflects the evolution relationship of device node states; Based on temporal state representation, the state modeling structure in the Graph-Attention multi-agent reinforcement learning base model is updated so that the state modeling structure can receive and process temporal state representation. Based on temporal state representation, the state transition relationship in the Graph-Attention multi-agent reinforcement learning basic model is reconstructed to generate a state transition description corresponding to the time series state data. Based on the updated state modeling structure and the reconstructed state transition relationship, the policy update path in the Graph-Attention multi-agent reinforcement learning basic model is adjusted to obtain an improved multi-agent decision model.

[0012] Optionally, the reward construction module specifically includes: Based on the improved multi-agent decision-making model, and by acquiring the device node operating status data before and after the IoT data distribution execution; Based on the device node operating status data, the state changes of the device node before and after data distribution execution are compared and analyzed to form state change data corresponding to the device node. Based on state change data, calculate the corresponding node-level reward information for each device node, and generate node-level reward data that corresponds one-to-one with each device node. Based on dynamic graph structure data, node-level reward data is propagated according to the node connection relationships in the dynamic graph to form graph structure-level reward propagation results; Organize the results of graph-level reward propagation into graph-level reward data.

[0013] Optionally, the model training module specifically includes: Obtain the current parameter state of the improved multi-agent decision-making model; Based on graph-level reward data, according to the one-to-one correspondence between device nodes and agents, the graph-level reward data is allocated to the corresponding agents to form agent-level training input data; In the improved multi-agent decision-making model, gradient update calculations are performed on the policy parameters of each agent based on the agent-level training input data to generate updated agent policy parameters. Based on the updated agent policy parameters, the shared parameters in the multi-agent decision-making model are synchronously adjusted to form a consistent set of multi-agent model parameters. Write the synchronized and adjusted set of multi-agent model parameters into the improved multi-agent decision model to complete one model parameter update process. Output the multi-agent decision model with updated parameters.

[0014] Optionally, the decision generation module specifically includes: Based on a multi-agent decision-making model that has completed parameter updates, and by reading the dynamic graph state representation at the current moment; Based on the dynamic graph state representation at the current moment, the dynamic graph state representation is assigned to the corresponding intelligent agent according to the one-to-one correspondence between device nodes and intelligent agents, forming intelligent agent-level state input data; In the multi-agent decision-making model that has completed parameter updates, policy reasoning calculations are performed on the agent-level state input data to generate distribution decision candidate results corresponding to each agent. Based on the distribution decision candidate results, the outputs of each intelligent agent are integrated according to the device node identifier to form the device node-level data management distribution decision results; The device node-level data management and distribution decision results are mapped to IoT data distribution path information and distribution target information to generate data management and distribution decision strategies.

[0015] The beneficial effects of this invention are: This invention employs a technical solution combining dynamic graph state space construction and graph attention multi-agent reinforcement learning to achieve joint modeling of the operating states and interaction relationships of IoT device nodes, effectively improving the overall perception capability of node association information during data management and distribution. Simultaneously, by introducing a dynamic graph deep reinforcement learning framework into the multi-agent reinforcement learning base model and reconstructing the state modeling structure, state transition relationships, and policy update paths, it achieves continuous modeling of the time-varying characteristics of device node operating states, significantly enhancing the adaptability of data management and distribution decisions to dynamic operating environments. Furthermore, by constructing a training mechanism based on graph structure-level reward feedback, node-level state changes are propagated along the dynamic graph structure, effectively solving the problem in existing technologies where reward information cannot reflect the overall impact of node association relationships. This overcomes the limitations of single-node evaluation methods, enabling continuous training and updating of the multi-agent decision model, thereby effectively improving the coordination, stability, and reliability of system operation in IoT data management and distribution application scenarios. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0017] Figure 1 This is a schematic diagram of the process of an artificial intelligence data management and distribution system based on the Internet of Things proposed in this invention.

[0018] Figure 2 This is a schematic diagram of a multi-agent graph attention and dynamic graph reinforcement learning collaborative decision-making structure for an IoT-based artificial intelligence data management and distribution system proposed in this invention. Detailed Implementation

[0019] Combination Figures 1-2 The present invention will be described in further detail below. These accompanying drawings are simplified schematic diagrams, illustrating only the basic structure of the invention and showing the main components relevant to the invention. Figure 1 and Figure 2 The present invention provides an Internet of Things-based artificial intelligence data management and distribution system, which includes the following modules:

[0020] Status acquisition module: Collects the operating status data of device nodes in the IoT system, performs time synchronization and normalization processing on the operating status data, and obtains the IoT operating status dataset; Dynamic graph construction module: Based on the IoT operation status dataset, a dynamic graph state space is constructed, which maps device nodes to nodes in the dynamic graph, maps the data interaction relationships between nodes to edges in the dynamic graph, and associates them with the corresponding operation status attributes to obtain the initial dynamic graph state representation. Multi-agent module: Taking the initial dynamic graph state representation as input, initialize the Graph-Attention multi-agent reinforcement learning basic model, configure the corresponding agent for each node, and perform weighted modeling of the nodes and their adjacency relationships in the dynamic graph through the graph attention mechanism to obtain the Graph-Attention multi-agent reinforcement learning basic model; Model Reconstruction Module: Introducing a dynamic graph deep reinforcement learning framework into the Graph-Attention multi-agent reinforcement learning base model, the dynamic graph state representations at multiple consecutive time steps are used as joint inputs to reconstruct the state modeling structure, state transition relationships, and policy update paths in the Graph-Attention multi-agent reinforcement learning base model, resulting in an improved multi-agent decision model. Reward Construction Module: Based on the improved multi-agent decision-making model, a graph-structured reward feedback mechanism is constructed to map the changes in the running state after the distribution and execution of IoT data into node-level reward information, and the reward is propagated along the node association relationship in the dynamic graph to obtain graph-structured reward data; Model training module: The improved multi-agent decision model is trained and its parameters are updated online using graph-structured reward data to obtain the trained multi-agent decision model; Decision generation module: Inputs the dynamic graph state representation of the current moment into the trained multi-agent decision model, generates and outputs data management and distribution decision strategies.

[0021] In this embodiment, the status acquisition module specifically includes: Raw status data reflecting the operating status of devices are collected from device nodes in the Internet of Things system. The raw status data includes the processing load information, communication latency information, data cache usage information, and data interaction records between device nodes, forming a set of raw operating status data of device nodes. Perform timestamp marking, time alignment, and data normalization on various types of status data in the original operating status data set of device nodes to generate an IoT operating status dataset.

[0022] In this embodiment, the dynamic graph construction module specifically includes: Based on the IoT operation status dataset, the operation status data is indexed and organized according to the device node identifier to form node status data records corresponding to the device nodes. Based on node status data records, a unique node identifier is assigned to each device node, and the device node is mapped to a node in a dynamic graph to generate a node set; Based on the data interaction relationship information reflected in the IoT operation status data, the data interaction connection relationship between device nodes is determined, and the data interaction connection relationship is mapped to the edge in the dynamic graph to generate an edge set; The node load status, communication latency status, and data cache occupancy status in the node status data record are associated with the corresponding dynamic graph nodes, and the interaction frequency and interaction direction in the data interaction connection relationship are associated with the corresponding dynamic graph edges, forming a dynamic graph structure data with attributes. Perform integrity checks on dynamic graph structure data with attributes and output the initial dynamic graph state representation.

[0023] In this embodiment, the multi-agent module specifically includes: Based on the initial dynamic graph state representation, the set of nodes, the set of edges, and the running state attributes associated with the nodes in the dynamic graph are obtained to form dynamic graph structure data containing node attributes and edge attributes. Based on the node set, configure a corresponding intelligent agent for each device node, and establish a one-to-one correspondence between device node identifiers and intelligent agent identifiers to generate a multi-agent set; Based on the edge set, the connection relationship information between device nodes is extracted, a graph structure information reflecting the topological relationship of device nodes is constructed, and the graph structure information is combined with the running state attributes associated with the nodes to form multi-agent input data; Introducing a graph attention computation process into a multi-agent ensemble, determining the adjacency relationship between nodes based on graph structure information, performing weighted computation of adjacency nodes on the multi-agent input data, and generating a node state representation containing adjacency node information; Based on the node state representation, the node state representation is collected and organized according to the agent identifier to form a multi-agent state representation corresponding to each agent; Output the multi-agent state representation to complete the construction of the Graph-Attention multi-agent reinforcement learning basic model and obtain the basic multi-agent state representation.

[0024] In this embodiment, the model reconstruction module specifically includes: Based on the basic multi-agent state representation, and by acquiring the basic multi-agent state representations corresponding to multiple consecutive time points in chronological order, time series state data is formed. Based on time-series state data, we construct time-series input data to describe the relationship between device node states and time changes, and then input the time-series input data into a dynamic graph deep reinforcement learning framework. In the dynamic graph deep reinforcement learning framework, time correlation modeling is performed on the temporal input data to generate a temporal state representation that reflects the evolution relationship of device node states; Based on temporal state representation, the state modeling structure in the Graph-Attention multi-agent reinforcement learning base model is updated so that the state modeling structure can receive and process temporal state representation. Based on temporal state representation, the state transition relationship in the Graph-Attention multi-agent reinforcement learning basic model is reconstructed to generate a state transition description corresponding to the time series state data. Based on the updated state modeling structure and the reconstructed state transition relationship, the policy update path in the Graph-Attention multi-agent reinforcement learning basic model is adjusted to obtain an improved multi-agent decision model.

[0025] In this embodiment, the reward construction module specifically includes: Based on the improved multi-agent decision-making model, and by acquiring the device node operating status data before and after the IoT data distribution execution; Based on the device node operating status data, the state changes of the device node before and after data distribution execution are compared and analyzed to form state change data corresponding to the device node. Based on state change data, calculate the corresponding node-level reward information for each device node, and generate node-level reward data that corresponds one-to-one with each device node. Based on dynamic graph structure data, node-level reward data is propagated according to the node connection relationships in the dynamic graph to form graph structure-level reward propagation results; Organize the results of graph-level reward propagation into graph-level reward data.

[0026] In this embodiment, the model training module specifically includes: Obtain the current parameter state of the improved multi-agent decision-making model; Based on graph-level reward data, according to the one-to-one correspondence between device nodes and agents, the graph-level reward data is allocated to the corresponding agents to form agent-level training input data; In the improved multi-agent decision-making model, gradient update calculations are performed on the policy parameters of each agent based on the agent-level training input data to generate updated agent policy parameters. Based on the updated agent policy parameters, the shared parameters in the multi-agent decision-making model are synchronously adjusted to form a consistent set of multi-agent model parameters. Write the synchronized and adjusted set of multi-agent model parameters into the improved multi-agent decision model to complete one model parameter update process. Output the multi-agent decision model with updated parameters.

[0027] In this embodiment, the decision generation module specifically includes: Based on a multi-agent decision-making model that has completed parameter updates, and by reading the dynamic graph state representation at the current moment; Based on the dynamic graph state representation at the current moment, the dynamic graph state representation is assigned to the corresponding intelligent agent according to the one-to-one correspondence between device nodes and intelligent agents, forming intelligent agent-level state input data; In the multi-agent decision-making model that has completed parameter updates, policy reasoning calculations are performed on the agent-level state input data to generate distribution decision candidate results corresponding to each agent. Based on the distribution decision candidate results, the outputs of each intelligent agent are integrated according to the device node identifier to form the device node-level data management distribution decision results; The device node-level data management and distribution decision results are mapped to IoT data distribution path information and distribution target information to generate data management and distribution decision strategies.

[0028] Example 1: To verify the feasibility of this invention in practice, this example selects a typical large-scale IoT data management and distribution application scenario as the verification environment. In this scenario, the IoT system consists of a large number of heterogeneous device nodes distributed at different network layers, continuously generating status data and requiring dynamic data management and distribution decisions between the edge and the central side. In this application environment, the device nodes exhibit characteristics such as dynamic fluctuations in processing load, communication latency changing with network status, uneven data cache usage, and frequent changes in data interaction relationships between nodes. This places high demands on the real-time performance, adaptability, and global consistency of the data management and distribution strategy.

[0029] In existing applications, common data management and distribution methods often employ static rules or local optimization strategies based on single-node states. These methods have limited ability to model the relationships between device nodes, making it difficult to depict the dynamic evolution of the entire IoT system. With the continuous growth in the number of nodes and the high complexity of data interaction relationships, existing methods show significant shortcomings in data distribution path selection, load balancing, and adaptation to temporal state changes. This can easily lead to problems such as excessive load on some nodes, increased communication congestion, and delayed data distribution decisions. This invention addresses these issues by introducing dynamic graph modeling, multi-agent reinforcement learning, and a collaborative design of graph attention mechanisms to achieve joint perception and decision-making regarding the global state of the IoT system.

[0030] In the specific implementation process, the status acquisition module first continuously collects operational status data from all device nodes in the IoT system. The collected data covers device node processing load, communication latency, data cache usage, and data interaction records between nodes. The collected raw status data undergoes unified timestamp marking, time alignment, and data normalization to form a consistent IoT operational status dataset, providing reliable input for subsequent dynamic graph modeling.

[0031] The dynamic graph construction module, based on the IoT operational status dataset, maps each device node to a node in the dynamic graph, maps the data interaction relationships between device nodes to edges in the dynamic graph, and associates corresponding operational status attributes with nodes and edges, thereby constructing an initial dynamic graph state representation. This dynamic graph state representation can simultaneously reflect the state of a node itself and the interaction relationships between nodes, providing a foundation for characterizing the overall operational status of the IoT system.

[0032] Building upon this foundation, the multi-agent module takes the initial dynamic graph state representation as input, configures a corresponding agent for each device node, and establishes a one-to-one correspondence between device node identifiers and agent identifiers. By introducing a graph attention calculation process, weighted modeling of nodes and their adjacency relationships in the dynamic graph is performed. This allows each agent to focus on the state information of neighboring nodes that are highly relevant to its own decisions during state modeling, thereby completing the construction of the Graph-Attention multi-agent reinforcement learning basic model and outputting a basic multi-agent state representation.

[0033] In the model reconstruction phase, a dynamic graph deep reinforcement learning framework is introduced to organize the basic multi-agent state representations at multiple consecutive time points into time-series state data. By performing temporal correlation modeling on the time-series state data, a temporal state representation reflecting the evolution of device node states over time is generated. Based on this temporal state representation, the state modeling structure, state transition relationships, and policy update paths in the Graph-Attention multi-agent reinforcement learning base model are reconstructed as a whole, ultimately resulting in an improved multi-agent decision model that can simultaneously perceive spatial structural relationships and temporal evolution characteristics.

[0034] In the reward construction phase, based on the improved multi-agent decision-making model, the operating status data of device nodes before and after IoT data distribution are collected. The changes in node status are compared and analyzed, and node-level reward information is calculated for each device node accordingly. Subsequently, the node-level reward information is propagated by utilizing the connection relationships between nodes in the dynamic graph structure, forming graph-level reward data, so that the reward feedback can reflect the impact of local state changes on the overall system.

[0035] During the model training phase, the improved multi-agent decision-making model is trained online using graph-structured reward data. During training, each agent performs gradient updates of policy parameters based on its assigned agent-level training input data, and a shared parameter synchronization mechanism ensures the consistency of the multi-agent decision-making model parameters. In this embodiment, a fixed learning rate of 0.0005 and a discount factor of 0.98 are used for model training. Each round of online training includes 120 policy update iterations, and the model parameters stabilize after 3000 consecutive time steps.

[0036] During the decision generation phase, the dynamic graph state representation at the current moment is input into the trained multi-agent decision model. Each agent outputs its corresponding distribution decision candidate results, which are then integrated according to the device node identifier to finally generate a device node-level data management and distribution decision strategy to guide the data distribution execution of the IoT system.

[0037] To verify the practical effectiveness of the technical solution of this invention, the performance of existing data distribution methods based on static rules and the method of this invention was compared and tested in the same experimental environment on several key indicators. During the test, five IoT operation samples of different sizes were selected, and each sample was continuously run for no less than 5000 time steps. Statistical analysis of the system operation status was performed. The relevant test results are shown in Table 1.

[0038] Table 1 Comparison of Data Management and Distribution Performance

[0039] As shown in Table 1, the method of this invention can accurately predict the trend of system operating status changes under different samples, and the deviation between the predicted and measured values ​​remains within a small range. In key indicators such as communication latency, node load balancing, and distribution success rate, the method of this invention exhibits stable and consistent performance advantages, demonstrating that the comprehensive application of dynamic graph modeling, multi-agent collaborative decision-making, and graph-structured reward feedback mechanisms can effectively improve the overall quality and stability of IoT data management and distribution decisions, verifying the feasibility and practical value of this invention in real-world applications.

Claims

1. An artificial intelligence data management and distribution system based on the Internet of Things, characterized in that, Includes the following modules: Status acquisition module: Collects the operating status data of device nodes in the IoT system, performs time synchronization and normalization processing on the operating status data, and obtains the IoT operating status dataset; Dynamic graph construction module: Based on the IoT operation status dataset, a dynamic graph state space is constructed, which maps device nodes to nodes in the dynamic graph, maps the data interaction relationships between nodes to edges in the dynamic graph, and associates them with the corresponding operation status attributes to obtain the initial dynamic graph state representation. Multi-agent module: Taking the initial dynamic graph state representation as input, initialize the Graph-Attention multi-agent reinforcement learning basic model, configure the corresponding agent for each node, and perform weighted modeling of the nodes and their adjacency relationships in the dynamic graph through the graph attention mechanism to obtain the Graph-Attention multi-agent reinforcement learning basic model; Model Reconstruction Module: Introducing a dynamic graph deep reinforcement learning framework into the Graph-Attention multi-agent reinforcement learning base model, the dynamic graph state representations at multiple consecutive time steps are used as joint inputs to reconstruct the state modeling structure, state transition relationships, and policy update paths in the Graph-Attention multi-agent reinforcement learning base model, resulting in an improved multi-agent decision model. Reward Construction Module: Based on the improved multi-agent decision-making model, a graph-structured reward feedback mechanism is constructed to map the changes in the running state after the distribution and execution of IoT data into node-level reward information, and the reward is propagated along the node association relationship in the dynamic graph to obtain graph-structured reward data; Model training module: The improved multi-agent decision model is trained and its parameters are updated online using graph-structured reward data to obtain the trained multi-agent decision model; Decision generation module: Inputs the dynamic graph state representation of the current moment into the trained multi-agent decision model, generates and outputs data management and distribution decision strategies.

2. The IoT-based artificial intelligence data management and distribution system according to claim 1, characterized in that, The status acquisition module specifically includes: Raw status data reflecting the operating status of devices are collected from device nodes in the Internet of Things system. The raw status data includes the processing load information, communication latency information, data cache usage information, and data interaction records between device nodes, forming a set of raw operating status data of device nodes. Perform timestamp marking, time alignment, and data normalization on various types of status data in the original operating status data set of device nodes to generate an IoT operating status dataset.

3. The IoT-based artificial intelligence data management and distribution system according to claim 1, characterized in that, The dynamic graph construction module specifically includes: Based on the IoT operation status dataset, the operation status data is indexed and organized according to the device node identifier to form node status data records corresponding to the device nodes. Based on node status data records, a unique node identifier is assigned to each device node, and the device node is mapped to a node in a dynamic graph to generate a node set; Based on the data interaction relationship information reflected in the IoT operation status data, the data interaction connection relationship between device nodes is determined, and the data interaction connection relationship is mapped to the edge in the dynamic graph to generate an edge set; The node load status, communication latency status, and data cache occupancy status in the node status data record are associated with the corresponding dynamic graph nodes, and the interaction frequency and interaction direction in the data interaction connection relationship are associated with the corresponding dynamic graph edges, forming a dynamic graph structure data with attributes. Perform integrity checks on dynamic graph structure data with attributes and output the initial dynamic graph state representation.

4. The IoT-based artificial intelligence data management and distribution system according to claim 1, characterized in that, The multi-agent module specifically includes: Based on the initial dynamic graph state representation, the set of nodes, the set of edges, and the running state attributes associated with the nodes in the dynamic graph are obtained to form dynamic graph structure data containing node attributes and edge attributes. Based on the node set, configure a corresponding intelligent agent for each device node, and establish a one-to-one correspondence between device node identifiers and intelligent agent identifiers to generate a multi-agent set; Based on the edge set, the connection relationship information between device nodes is extracted, a graph structure information reflecting the topological relationship of device nodes is constructed, and the graph structure information is combined with the running state attributes associated with the nodes to form multi-agent input data; Introducing a graph attention computation process into a multi-agent ensemble, determining the adjacency relationship between nodes based on graph structure information, performing weighted computation of adjacency nodes on the multi-agent input data, and generating a node state representation containing adjacency node information; Based on the node state representation, the node state representation is collected and organized according to the agent identifier to form a multi-agent state representation corresponding to each agent; Output the multi-agent state representation to complete the construction of the Graph-Attention multi-agent reinforcement learning basic model and obtain the basic multi-agent state representation.

5. The IoT-based artificial intelligence data management and distribution system according to claim 1, characterized in that, The model reconstruction module specifically includes: Based on the basic multi-agent state representation, and by acquiring the basic multi-agent state representations corresponding to multiple consecutive time points in chronological order, time series state data is formed. Based on time-series state data, we construct time-series input data to describe the relationship between device node states and time changes, and then input the time-series input data into a dynamic graph deep reinforcement learning framework. In the dynamic graph deep reinforcement learning framework, time correlation modeling is performed on the temporal input data to generate a temporal state representation that reflects the evolution relationship of device node states; Based on temporal state representation, the state modeling structure in the Graph-Attention multi-agent reinforcement learning base model is updated so that the state modeling structure can receive and process temporal state representation. Based on temporal state representation, the state transition relationship in the Graph-Attention multi-agent reinforcement learning basic model is reconstructed to generate a state transition description corresponding to the time series state data. Based on the updated state modeling structure and the reconstructed state transition relationship, the policy update path in the Graph-Attention multi-agent reinforcement learning basic model is adjusted to obtain an improved multi-agent decision model.

6. The IoT-based artificial intelligence data management and distribution system according to claim 1, characterized in that, The reward construction module specifically includes: Based on the improved multi-agent decision-making model, and by acquiring the device node operating status data before and after the IoT data distribution execution; Based on the device node operating status data, the state changes of the device node before and after data distribution execution are compared and analyzed to form state change data corresponding to the device node. Based on state change data, calculate the corresponding node-level reward information for each device node, and generate node-level reward data that corresponds one-to-one with each device node. Based on dynamic graph structure data, node-level reward data is propagated according to the node connection relationships in the dynamic graph to form graph structure-level reward propagation results; Organize the results of graph-level reward propagation into graph-level reward data.

7. The IoT-based artificial intelligence data management and distribution system according to claim 1, characterized in that, The model training module specifically includes: Obtain the current parameter state of the improved multi-agent decision-making model; Based on graph-level reward data, according to the one-to-one correspondence between device nodes and agents, the graph-level reward data is allocated to the corresponding agents to form agent-level training input data; In the improved multi-agent decision-making model, gradient update calculations are performed on the policy parameters of each agent based on the agent-level training input data to generate updated agent policy parameters. Based on the updated agent policy parameters, the shared parameters in the multi-agent decision-making model are synchronously adjusted to form a consistent set of multi-agent model parameters. Write the synchronized and adjusted set of multi-agent model parameters into the improved multi-agent decision model to complete one model parameter update process. Output the multi-agent decision model with updated parameters.

8. The IoT-based artificial intelligence data management and distribution system according to claim 1, characterized in that, The decision generation module specifically includes: Based on a multi-agent decision-making model that has completed parameter updates, and by reading the dynamic graph state representation at the current moment; Based on the dynamic graph state representation at the current moment, the dynamic graph state representation is assigned to the corresponding intelligent agent according to the one-to-one correspondence between device nodes and intelligent agents, forming intelligent agent-level state input data; In the multi-agent decision-making model that has completed parameter updates, policy reasoning calculations are performed on the agent-level state input data to generate distribution decision candidate results corresponding to each agent. Based on the distribution decision candidate results, the outputs of each intelligent agent are integrated according to the device node identifier to form the device node-level data management distribution decision results; The device node-level data management and distribution decision results are mapped to IoT data distribution path information and distribution target information to generate data management and distribution decision strategies.

Images