Internet of Things monitoring management method and system based on deep learning

Abstract

The invention discloses an Internet of Things monitoring management method and system based on deep learning, and relates to the related technical field of Internet of Things monitoring management, and the method comprises the steps: reading node multi-source data through an intelligent agent, constructing a deep causal structure field through cross-modal deep causal reasoning, and revealing the relationship between environmental inducements, state paths and risk propagation; constructing a space-time causal tensor representing dynamic risk distribution of a monitoring area, quantitatively evaluating monitoring income, energy consumption cost and risk exposure, executing game analysis under a multi-party game strategy, and establishing an intelligent agent behavior strategy; and updating the monitoring network topology, executing deep space-time-cause and effect joint reasoning, generating an exception explanation result, and performing exception report management. According to the invention, the technical problems of lack of causal interpretability, resource allocation rigidness and insufficient abnormal response capability of Internet of Things monitoring in the prior art are solved, and the technical effects of accurate abnormal management of Internet of Things nodes and improvement of resource utilization efficiency and risk decision interpretability are achieved.

Description

Technical Field

[0001] This invention relates to the field of IoT monitoring and management technology, specifically to IoT monitoring and management methods and systems based on deep learning. Background Technology

[0002] IoT monitoring nodes are widely used in environmental monitoring, industrial automation, and other fields. They are typically deployed in complex and dynamic environments to collect multi-source data such as temperature, humidity, pressure, vibration, and images in real time. However, they also face challenges such as data heterogeneity, resource constraints, and frequent anomalies. Traditional IoT monitoring cannot effectively handle cross-modal correlations of multi-source data and lacks a deep understanding of the causal mechanisms behind anomalies, making it difficult to automatically identify anomalies, explain their causes, and predict potential risks from IoT data. Furthermore, existing decision analysis methods lack interpretability and struggle to reveal the causal relationships between environmental triggers, state change paths, and risk propagation. In addition, IoT monitoring nodes have limited energy consumption and computing power, making it impossible to optimize resource allocation and reduce risk exposure while ensuring monitoring benefits.

[0003] Therefore, current technologies in the Internet of Things (IoT) suffer from technical problems such as a lack of causal explanatory power in monitoring, rigid resource allocation, and insufficient anomaly response capabilities. Summary of the Invention

[0004] This application provides a deep learning-based IoT monitoring and management method and system, which solves the technical problems of lack of causal interpretability, rigid resource allocation and insufficient anomaly response capability in existing IoT monitoring technologies. It achieves the technical effects of precise anomaly management of IoT nodes, improved resource utilization efficiency and interpretability of risk decisions.

[0005] This application provides a deep learning-based IoT monitoring and management method, comprising: configuring autonomous agents at IoT monitoring nodes; after using the autonomous agents to read multi-source monitoring data from the IoT monitoring nodes, establishing a deep causal structure field containing environmental causes, state change paths, and risk propagation relationships through a cross-modal deep causal reasoning model; constructing a spatiotemporal causal tensor based on the deep causal structure field to characterize the dynamic risk distribution of the monitoring area; using the autonomous agents to perform quantitative assessments of monitoring benefits, energy consumption costs, and risk exposure metrics based on the spatiotemporal causal tensor, performing game analysis under a multi-party game strategy, and establishing an agent behavior strategy; dynamically updating the IoT monitoring structure through topological evolution based on the agent behavior strategy, performing deep spatiotemporal-causal joint reasoning analysis under a self-organizing layout, and establishing anomaly explanation results, the anomaly explanation results including anomaly causes, propagation chains, and potential impact areas; and managing anomaly reporting based on the anomaly explanation results.

[0006] In a possible implementation, the deep learning-based IoT monitoring and management method further performs the following processing: after parsing the spatiotemporal causal tensor, an autonomous agent is used to evaluate the causal closed-loop stability of the corresponding monitoring node based on the parsing results, and a causal closed-loop stability evaluation result is established; after extracting the dynamic risk value of the monitoring area based on the spatiotemporal causal tensor, a neighborhood cooperation constraint is constructed based on the dynamic risk value of the monitoring area and the causal closed-loop stability evaluation result; the neighborhood cooperation constraint is used as a constraint condition for game analysis, and game constraint management of the agent's behavior strategy is performed.

[0007] In a possible implementation, the deep learning-based IoT monitoring and management method further performs the following processing: After establishing constraints, autonomous agents are used to perform joint utility modeling on monitoring benefits, energy consumption costs, and risk exposure, and the results are weighted and integrated according to preset risk preference and benefit preference rules to form node utility results describing the behavioral tendencies of monitoring nodes; the node utility results and the causal closed-loop stability evaluation results are used as the basic dataset for game analysis; based on the basic dataset, a local decision-making task with neighborhood cooperation constraints is constructed for each autonomous agent, and the local decision-making task is used to limit the adjustable range of behavioral parameters of the autonomous agent in terms of sampling frequency, monitoring direction, data reporting path, and neighborhood cooperation strength; the task-solving evolutionary game iteration of the local decision-making task is performed using a distributed policy update mechanism to output the agent's behavioral strategy.

[0008] In a possible implementation, the deep learning-based IoT monitoring and management method further performs the following processing: During the iterative process after acquiring the initial local decision-making task, the autonomous agent periodically receives the behavioral parameter adjustment intentions of neighboring agents based on the neighborhood cooperation constraints, forming policy interaction information to characterize the trend of neighborhood policy changes; before updating its own policy, the autonomous agent performs constraint consistency verification on candidate policies, which is used to detect the feasibility of candidate policies in terms of sampling frequency, monitoring direction, data reporting path, and neighborhood cooperation strength, and outputs the arbitrated and corrected candidate policies; based on the output results, the autonomous agent uses the node utility results and policy interaction information to perform policy response calculation, and selects a response policy from the output results to replace the current policy based on the policy response calculation results; performs multi-round iterative game evolution, and outputs the agent's behavioral policy.

[0009] In a possible implementation, the deep learning-based IoT monitoring and management method further performs the following processing: Based on the agent's behavior strategy, the autonomous agent dynamically evaluates the node connection relationships, data flow paths, and cooperation scope of the current IoT monitoring structure, identifying redundant, inefficient, and high-conflict links in the structure to form a structure adaptability result reflecting topology evolvability; using the structure adaptability result, the agent determines link maintenance, link reconnection, and link weight redistribution operations according to neighborhood cooperation constraints and its own behavior strategy, generating topology evolution instructions, which are used to gradually update the local monitoring structure in a distributed cooperative environment; based on the topology evolution instructions, the updated local monitoring structure undergoes self-organizing layout analysis, updating the causal correlation strength, information propagation direction, and risk coupling paths between monitoring nodes to form a structured layout expression; and based on the structured layout expression, deep spatiotemporal-causal joint reasoning analysis and management under the self-organizing layout are performed.

[0010] In a possible implementation, the deep learning-based IoT monitoring and management method further performs the following processing: taking the structured layout expression as input, it performs hierarchical aggregation of spatiotemporal observation data and causal relationships at three scales: node level, subgraph level, and region level, generating node status fragments, subgraph collaboration fragments, and region risk fragments respectively, and constructing a data semantic layer for joint inference; based on the data semantic layer, it uses a graph structure-aware temporal representation to synchronously encode the spatiotemporal features at each scale, the temporal representation being used to simultaneously capture the topological coupling relationship between nodes, the directionality of information propagation, and the temporal evolution trend, and outputs a spatiotemporal embedding vector for causal identification; it performs explicit scoring and screening of the spatiotemporal embedding vector under a causal attention mechanism to generate candidate causal chains ranked by confidence; it performs local data-driven forward simulation and counterfactual verification of the candidate causal chains, constructs anomaly explanation results based on the verification results, and outputs them.

[0011] In a possible implementation, the deep learning-based IoT monitoring and management method further performs the following processing: configuring an early warning response signal based on the anomaly interpretation result, and using the early warning response signal for anomaly reporting management.

[0012] This application also provides a deep learning-based IoT monitoring and management system, comprising: a causal structure field construction module, used to configure autonomous agents at IoT monitoring nodes, and after using the autonomous agents to read multi-source monitoring data from IoT monitoring nodes, to establish a deep causal structure field containing environmental causes, state change paths, and risk propagation relationships through cross-modal deep causal reasoning; a behavior strategy establishment module, used to construct a spatiotemporal causal tensor to characterize the dynamic risk distribution of the monitoring area based on the deep causal structure field, and to use the autonomous agents to perform quantitative assessments of monitoring benefits, energy consumption costs, and risk exposure based on the spatiotemporal causal tensor, perform game analysis under a multi-party game strategy, and establish agent behavior strategies; a joint reasoning analysis module, used to dynamically update the IoT monitoring structure through topological evolution based on the agent behavior strategies, perform deep spatiotemporal-causal joint reasoning analysis under a self-organizing layout, and establish anomaly explanation results, including anomaly causes, propagation chains, and potential impact areas; and an anomaly reporting management module, used to manage anomaly reporting based on the anomaly explanation results.

[0013] This application proposes a deep learning-based IoT monitoring and management method and system. This system utilizes intelligent agents to read multi-source data from nodes, constructs a deep causal structure field using cross-modal deep causal reasoning, and reveals the relationship between environmental triggers, state paths, and risk propagation. It constructs a spatiotemporal causal tensor representing the dynamic risk distribution of the monitoring area, quantitatively assesses monitoring benefits, energy costs, and risk exposure, performs game analysis under a multi-party game strategy, and establishes intelligent agent behavior policies. The system updates the monitoring network topology and performs deep spatiotemporal-causal joint reasoning to generate anomaly explanation results and manage anomaly reporting. This addresses the technical problems of existing IoT monitoring technologies, such as lack of causal interpretability, rigid resource allocation, and insufficient anomaly response capabilities. It achieves the technical effects of precise anomaly management of IoT nodes, improved resource utilization efficiency, and interpretable risk decisions. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings of the embodiments of this disclosure will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0015] Figure 1 This is a schematic diagram of the IoT monitoring and management method based on deep learning provided in an embodiment of this application.

[0016] Figure 2This is a schematic diagram of the structure of an IoT monitoring and management system based on deep learning, provided in an embodiment of this application.

[0017] Figure labeling: Causal structure field construction module 10, behavior strategy establishment module 20, joint reasoning analysis module 30, and anomaly reporting management module 40. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on the specific implementation methods, structures, features and effects of the present invention.

[0019] This application provides an IoT monitoring and management method based on deep learning, such as... Figure 1 As shown, the method includes: Step S100: Configure an autonomous intelligent agent on the IoT monitoring node. After reading the multi-source monitoring data of the IoT monitoring node using the autonomous intelligent agent, establish a deep causal structure field containing environmental causes, state change paths, and risk propagation relationships through a cross-modal deep causal reasoning mode.

[0020] Preferably, IoT monitoring nodes may include sensors, cameras, and dedicated data acquisition equipment. Autonomous agents with autonomous decision-making and computing capabilities are configured at each IoT monitoring node. These agents then read heterogeneous multi-source monitoring data from their respective IoT monitoring nodes, including physical environmental data such as temperature, humidity, pressure, and vibration; equipment status data such as power consumption, CPU load, communication signal strength, and fault codes; image / video data captured by visual sensors; and recorded operation and alarm event logs. The multi-source monitoring data is analyzed using a cross-modal deep causal reasoning model. This involves aligning, correlating, and jointly analyzing data from different modalities, such as temperature readings, equipment vibration signals, and on-site video clips. Deep learning models, such as deep neural networks, are used to automatically extract complex, high-level features from the high-dimensional raw data. Finally, causal reasoning is used to identify the causal mechanisms and effects of the multi-source monitoring data. Using this direction, a structured causal network model is established, encompassing environmental triggers, state change paths, and risk propagation relationships. This serves as a deep causal structural field, where nodes represent different variables or events, such as increased ambient temperature, equipment bearing wear, and fire risk, while edges represent the causal relationships between nodes. Specifically, environmental triggers are used to identify external environmental factors that cause changes in system state, such as consistently >80% ambient humidity. State change paths depict the intermediate chain of processes that leads from one state to another, for example, increased humidity → condensation on circuit boards → short circuit → abnormal equipment operation. Risk propagation relationships characterize how a risk event spreads and affects other monitoring nodes or areas through physical or logical dependencies between nodes after an event occurs. For example, a failure of monitoring node A → data loss in the area it is responsible for monitoring → failure of the global risk assessment model that relies on data from that area.

[0021] Step S200: Based on the deep causal structure field, a spatiotemporal causal tensor is constructed to characterize the dynamic risk distribution of the monitoring area. The autonomous agent is used to perform quantitative evaluation of monitoring benefits, energy consumption costs, and risk exposure based on the spatiotemporal causal tensor. Game analysis under a multi-party game strategy is performed to establish the agent's behavior strategy.

[0022] Preferably, the deep causal structure field defines the causal relationships between IoT monitoring nodes. Based on the deep causal structure field, a spatiotemporal causal tensor is constructed, which quantifies and spatiotemporally represents the causal network model, and determines a structured, multidimensional risk state table to characterize the dynamic risk distribution of the monitoring area. The spatiotemporal causal tensor may be a three-dimensional structure containing time, space, and risk variables. Here, time represents different time slices or time series, space represents different monitoring nodes or geographic grid locations, and risk variables represent various risk state variables defined in the causal structure field, such as overheat risk value, structural stress risk value, and communication interruption risk value. Furthermore, the risk value at each location in the spatiotemporal causal tensor is calculated and determined through the causal relationships defined in the causal structure field. For example, the risk value at location X may be determined by the risk value of upstream location Y at time T-1, propagated according to the weights on the causal edges.

[0023] Preferably, autonomous agents are used to quantitatively evaluate monitoring benefits, energy costs, and risk exposure based on the spatiotemporal causal tensor. This involves establishing a computable utility function for each action of the autonomous agent. Each agent analyzes the spatiotemporal causal tensor and calculates monitoring benefits, energy costs, and risk exposure. Specifically, monitoring benefits refer to the value gained by the agent in performing monitoring tasks. The level of benefit is directly related to the importance of the risks it can cover. For example, monitoring an area that is high-risk according to the spatiotemporal causal tensor and is upstream in the risk propagation chain yields very high benefits. Energy costs refer to the energy consumed by the agent in performing monitoring, computation, and communication tasks. Risk exposure refers to the possibility that risks may not be detected in time if the agent does not monitor a certain area or reduces its monitoring frequency, and the severity of its potential consequences. This is directly derived from the risk values ​​of the corresponding area and time in the spatiotemporal causal tensor.

[0024] Preferably, game analysis is performed under a multi-party game strategy, where each agent acts as an independent decision-maker, aiming to maximize its own utility by increasing the sampling rate, adjusting sensor orientation, and offloading tasks to neighboring nodes. This alters the surrounding risk situation and resource conditions, thereby affecting the utility of other agents. For example, if agent O focuses on monitoring region A, agent P can allocate more resources to region B, which is implicit cooperation. Conversely, if two agents compete to monitor the same high-yield region, it leads to resource conflict and waste. Game analysis refers to finding an equilibrium point through iterative calculation among all agents in their interdependent relationships. At this equilibrium point, no single agent can gain a better benefit by unilaterally changing its strategy. Ultimately, the agent behavior strategy is obtained, maximizing the overall monitoring efficiency.

[0025] Furthermore, step S200 also includes step S210, after parsing the spatiotemporal causal tensor, using an autonomous agent to evaluate the causal closed-loop stability of the corresponding monitoring node based on the parsing results, and establishing a causal closed-loop stability evaluation result; step S220, after extracting the dynamic risk value of the monitoring area based on the spatiotemporal causal tensor, constructing a neighborhood cooperation constraint based on the dynamic risk value of the monitoring area and the causal closed-loop stability evaluation result; step S230, using the neighborhood cooperation constraint as the constraint condition for game analysis, and executing game constraint management of the agent's behavior strategy.

[0026] Preferably, the analysis of the spatiotemporal causal tensor refers to the autonomous agent analyzing a multidimensional risk data table to identify its causal dynamics and obtain analytical results. For example, it is identified that the temperature rise of node M is the cause of the increased vibration of node N, and the vibration of node N further aggravates the temperature of node M through the feedback of the shared load-bearing structure. Then, the autonomous agent performs a causal closed-loop stability evaluation of the corresponding monitoring node based on the analytical results. That is, it performs a quantitative evaluation of the causal chain that forms the feedback loop, including using stability criteria in cybernetics or dynamic systems theory to calculate whether the feedback loop is stable. If it is stable, small disturbances are absorbed and the risk is not amplified. If it is unstable / divergent, small disturbances will be amplified step by step along the causal chain, eventually leading to system loss of control. The causal closed-loop stability evaluation result is then output to identify whether the local causal network in which the agent is located or is responsible is in a fragile / easily runaway state.

[0027] Preferably, the dynamic risk value of the current monitoring area is extracted from the spatiotemporal causal tensor, and then combined with the dynamic risk value and the causal closed-loop stability evaluation result to generate neighborhood cooperation constraints. If the dynamic risk value is high and the causal closed-loop stability is poor, it indicates that the node is a key risk point, and then mandatory cooperation constraints are generated. For example, it must maintain uninterrupted data synchronization with neighboring nodes, has the right to request other neighboring nodes to assist in monitoring when its own energy consumption is below a threshold, and cannot enter a low-power sleep mode without the consent of a majority of neighboring nodes. Conversely, if the dynamic risk value is low and the causal closed-loop stability is good, the neighborhood cooperation constraints are relatively relaxed, giving the agent more autonomous decision-making space.

[0028] Preferably, neighborhood cooperation constraints are used as hard constraints in game analysis and embedded in the game decision-making model of each autonomous agent. This executes game constraint management of agent behavior strategies, and the constraints cannot be violated during game analysis. For example, if an agent intends to hibernate to save energy, but its cooperation constraints require it to remain active to stabilize the fragile causal closed loop, then the hibernation option will be directly prohibited in the game. When selecting monitoring targets, agents must prioritize satisfying the cooperation tasks required by the constraints before pursuing additional monitoring benefits. This ensures that the behavior of individual agents in autonomous decision-making will not undermine the causal stability and risk control capabilities of the entire monitoring system, thereby achieving precise game management.

[0029] Furthermore, step S230 also includes step S231, after establishing constraints, using autonomous agents to perform joint utility modeling on monitoring benefits, energy consumption costs, and risk exposure, and integrating them according to preset risk preference and benefit preference rules to form node utility results used to describe the behavioral tendencies of monitoring nodes; step S232, using the node utility results and the causal closed-loop stability evaluation results as the basic dataset for game analysis; step S233, based on the basic dataset, constructing a local decision-making task with neighborhood cooperation constraints for each autonomous agent, the local decision-making task being used to limit the adjustable range of behavioral parameters of the autonomous agent in terms of sampling frequency, monitoring direction, data reporting path, and neighborhood cooperation strength; step S234, using a distributed policy update mechanism to execute the task-solving evolutionary game iteration of the local decision-making task, and outputting the agent's behavioral strategy.

[0030] Preferably, after using neighborhood cooperation constraints as constraints in game analysis, joint utility modeling is performed using autonomous agents. That is, a utility function is used to perform multi-objective joint modeling of monitoring benefits, energy consumption costs, and risk exposure. This is then transformed into single-objective optimization through weighted summation. The weights are the preset risk preference and benefit preference rules. For example, energy consumption costs have a higher weight when energy is scarce, and risk exposure has a higher weight when there are high security requirements. The current monitoring benefits, energy consumption costs, and risk exposure are substituted into the utility function as input variables, and the output is used to calculate the node utility result that describes the behavioral tendencies of the monitoring nodes. This represents the comprehensive satisfaction that the agent can obtain under a specific behavioral strategy.

[0031] Preferably, the node utility results and causal closed-loop stability evaluation results are integrated as the basic dataset for game analysis, ensuring that the game analysis considers both individual high utility and the global health and stability of the system. Then, based on the basic dataset, a local decision task with neighborhood cooperation constraints is constructed for each autonomous agent. That is, each autonomous agent defines an optimization objective and adjusts its own behavioral parameters to maximize its node utility results while satisfying all neighborhood cooperation constraints. The local decision task is used to limit the adjustable range of the autonomous agent's behavioral parameters such as sampling frequency, monitoring direction, data reporting path, and neighborhood cooperation strength. For example, the sampling frequency is selected between 1Hz and 10Hz, the monitoring direction is selected within 0° to 360°, the data reporting path is selected through the monitoring node, and the neighborhood cooperation strength determines how much resource to allocate to help neighboring nodes.

[0032] Preferably, the task-solving evolutionary game iteration of the local decision-making task is performed using a distributed policy update mechanism. Specifically, each agent randomly selects or initializes its behavior policy based on historical data, and then performs policy interaction and learning. That is, it periodically communicates with neighboring nodes to exchange its own policies and obtained utilities, and then compares them with the utilities of neighboring nodes. If it finds that the neighboring node's policy has higher utility and the behavior policy also meets its own constraints, it imitates the behavior policy and verifies the constraint consistency of the candidate policy to ensure its feasibility. After repeating this process multiple times, the behavior policy of the entire monitoring system gradually stabilizes, forming a Nash equilibrium or evolutionarily stable policy. At this point, no single agent can improve its own utility by unilaterally changing its behavior policy. Thus, the agent's behavior policy is output, which is the behavior parameters of the optimal sampling frequency, monitoring direction, data reporting path, and neighborhood cooperation strength finally determined by each autonomous agent, thereby maximizing local utility under global constraints.

[0033] Furthermore, step S241 also includes the following steps: during the iterative process after acquiring the initial local decision-making task, the autonomous agent periodically receives the behavioral parameter adjustment intentions of neighboring agents based on the neighborhood cooperation constraints, forming policy interaction information to characterize the trend of neighborhood policy changes; before updating its own policy, the autonomous agent performs constraint consistency verification on candidate policies, which is used to detect the feasibility of candidate policies in terms of sampling frequency, monitoring direction, data reporting path, and neighborhood cooperation strength, and outputs the arbitrated and corrected candidate policies; based on the output results, the autonomous agent uses the node utility results and policy interaction information to perform policy response calculation, selects a response policy from the output results to replace the current policy based on the policy response calculation results; performs multi-round iterative game evolution, and outputs the agent's behavioral policy.

[0034] Preferably, at the beginning of each iteration after acquiring the initial local decision-making task, the autonomous agent, based on its neighborhood cooperation constraints, receives behavioral parameter adjustment intentions from neighboring agents within its communication range. These intentions represent new behavioral strategies that neighboring agents are considering or about to adopt, and are used as policy interaction information to characterize the changing trends of other agents' strategies within its neighborhood, such as a general tendency towards higher sampling frequency or greater cooperation. Then, based on the received policy interaction information and its current state, the autonomous agent generates potential candidate strategies and performs constraint consistency verification on these strategies, including checking for consistency in sampling frequency, monitoring direction, and data. The feasibility of parameters for reporting paths and neighborhood cooperation strength is assessed to ensure that neighborhood cooperation constraints are not violated. For example, whether the sampling frequency is within the allowable upper and lower limits, whether the monitoring direction covers the monitoring sectors it is responsible for, whether the data reporting path is an available and reliable path or a path that avoids being prohibited, and whether the neighborhood cooperation strength reaches the promised minimum cooperation level. If the candidate strategy does not meet the constraints, arbitration correction is initiated. That is, if the candidate strategy proposes to reduce the cooperation strength below the constraint level, the correction program automatically raises it to the minimum allowable value, and then outputs the arbitrated corrected candidate strategy. Each strategy 100% satisfies all preset cooperation and system stability constraints.

[0035] Preferably, the autonomous agent uses node utility results and policy interaction information as input to perform policy response calculation and evaluation for each arbitrated and corrected candidate policy. Specifically, it calculates the comprehensive utility predicted by the candidate policy and analyzes whether the candidate policy is in synergy or conflict with the trends of neighboring nodes. Policies that are in synergy with the neighborhood trend may bring additional synergistic benefits. The response value is obtained by combining the predicted utility and the degree of synergy scores. Then, based on the policy response calculation result, the policy with the highest response value is selected from the output to replace the current policy, thereby updating its current behavior policy. Then, multiple rounds of iterative game evolution are performed, that is, entering the next iteration cycle. All agents collect information, verify, calculate and update policies again, and finally reach Nash equilibrium, that is, the vast majority of agents no longer change their own behavior policies during the iteration. Finally, the agent behavior policy is output, which includes the optimal stable behavior policy of each autonomous agent relative to its neighbors under all constraints.

[0036] Step S300: Based on the agent behavior strategy, the IoT monitoring structure is dynamically updated through topological evolution, and deep spatiotemporal-causal joint reasoning analysis under self-organizing layout is performed to establish anomaly explanation results. The anomaly explanation results include the cause of the anomaly, the propagation chain, and the potential area of ​​influence.

[0037] Step S300 further includes step S310, which, based on the agent's behavior strategy, dynamically evaluates the node connection relationships, data flow paths, and cooperation scope of the current IoT monitoring structure using an autonomous agent, identifies redundant links, inefficient links, and high-conflict links in the structure, and forms a structure adaptability result that reflects the topology's evolvability; step S320, using the structure adaptability result, determines the operations of link maintenance, link reconnection, and link weight redistribution according to the neighborhood cooperation constraints and its own behavior strategy, and generates a topology evolution instruction, which is used to gradually update the local monitoring structure in a distributed cooperative environment; step S330, based on the topology evolution instruction, performs self-organizing layout analysis on the updated local monitoring structure, updates the causal correlation strength, information propagation direction, and risk coupling path between monitoring nodes, and forms a structured layout expression; step S340, performs deep spatiotemporal-causal joint reasoning analysis and management under the self-organizing layout according to the structured layout expression.

[0038] Preferably, each autonomous agent dynamically evaluates the node connection relationships, data flow paths, and cooperation scope of the current IoT monitoring structure based on its behavioral strategy. This includes assessing whether the current communication connection is necessary, whether there are better neighbor nodes, whether the data reporting path is efficient, whether there are bottlenecks or single points of failure, and whether the current combination of collaborating partners best supports its monitoring task. This identifies redundant links, inefficient links, and high-conflict links in the structure. Redundant links are those with multiple functionally identical connections that waste resources, inefficient links are those with poor connection quality that affect data transmission efficiency, and high-conflict links are those where multiple agents compete to use the same link or the same neighbor node, leading to conflict and performance degradation. This results in a structural adaptability score, which reflects the topology's evolvability, i.e., the degree to which the current topology supports the agent's behavioral strategy.

[0039] Preferably, the agent comprehensively considers the structural adaptation results, neighborhood cooperation constraints, and its own behavioral strategies to jointly decide on the operation of the links, including link preservation, link reconnection, and link weight reallocation. Link preservation involves retaining efficient and critical links, link reconnection involves disconnecting inefficient or highly conflicting links and establishing new connections with better neighbor nodes, and link weight reallocation involves adjusting the priority or resource allocation of links. This generates topology evolution instructions, which are used to gradually update the local monitoring structure in a distributed cooperative environment to avoid global oscillations. For example, this can be used to add, delete, or modify connections in the local monitoring structure.

[0040] Preferably, based on topology evolution instructions, the updated local monitoring structure is subjected to self-organizing layout analysis, including updating causal correlation strength, updating information propagation directionality, and updating risk coupling paths. Specifically, due to changes in connection relationships, the availability of data between nodes changes, thus recalculating the causal influence correlation strength; the new data flow path determines the new direction of risk and information propagation; based on the new connections and causal strengths, the potential path of risk propagation between nodes is re-depicted; thereby forming a structured layout expression, i.e., an updated causal network model that is completely synchronized with the current physical topology. Finally, deep spatiotemporal-causal joint inference analysis and management under the self-organizing layout is performed according to the structured layout expression, including anomaly detection, root cause localization, and risk prediction based on the updated structured expression, to ensure the accuracy and reliability of anomaly interpretation results.

[0041] Furthermore, step S340 also includes step S341, using the structured layout expression as input, performing hierarchical aggregation of spatiotemporal observation data and causal relationships at three scales: node level, subgraph level, and region level, respectively generating node situation fragments, subgraph collaboration fragments, and region risk fragments, and constructing a data semantic layer for joint inference; step S342, based on the data semantic layer, using a graph structure-aware temporal representation to synchronously encode the spatiotemporal features at each scale, the temporal representation being used to simultaneously capture the topological coupling relationship between nodes, the directionality of information propagation, and the temporal evolution trend, and outputting a spatiotemporal embedding vector for causal identification; step S343, performing explicit scoring and filtering of the spatiotemporal embedding vector under a causal attention mechanism to generate candidate causal chains ranked by confidence; step S344, performing local data-driven forward simulation and counterfactual verification on the candidate causal chains, constructing anomaly explanation results based on the verification results, and outputting them.

[0042] Preferably, a structured layout representation is used as the input framework, injecting the original spatiotemporal observation data and causal relationships into it, and performing hierarchical aggregation at three scales: node level, subgraph level, and region level. Specifically, at the node level, the focus is on a single monitoring node, aggregating its own historical data and information from its direct causal neighbors to form a node status fragment, used to describe the local state of a single node and its direct causes and effects; at the subgraph level, the focus is on a community composed of multiple closely interacting nodes, aggregating the data and causal relationships of all nodes within the community to form a subgraph collaboration fragment, used to describe the collaborative working mechanism between nodes and the internal risk propagation pattern within a local area; at the region level, the focus is on the entire monitoring area or multiple subgraphs, performing higher-level abstraction and aggregation to form a regional risk fragment, used to describe the macro-risk situation and global patterns across subgraphs; and then a data semantic layer for joint inference is constructed, containing structured data with different granularities of information from micro to macro.

[0043] Preferably, based on the data semantic layer, a graph structure-aware temporal representation is used to synchronously encode the spatiotemporal features at each scale. The graph structure-aware temporal representation is usually a fusion model of graph neural network and temporal model, used to simultaneously capture the topological coupling relationship between nodes, the directionality of information propagation and the temporal evolution trend. Specifically, the graph neural network is used to understand the mutual influence of network connections between nodes, the flow of risk is determined according to the direction of the causal graph, the temporal model is used to identify the changes of each entity and its state over time, and finally it is transformed into a spatiotemporal embedding vector for causal identification. Then, a causal attention mechanism is used to explicitly score all possible node pairs, that is, to calculate the attention score of the influence of one node on another node, representing the confidence level of their causal relationship. All potential causal relationships are then ranked according to their confidence levels, retaining those with confidence levels exceeding a preset confidence threshold to form candidate causal chains ranked by confidence. Finally, local data-driven forward simulation and counterfactual verification are performed on the candidate causal chains. Forward simulation builds a simulation prediction model based on the candidate causal chains and uses real data from the cause nodes to simulate and predict the state of the result nodes. If the prediction results highly match the real data of the result nodes, the credibility of the causal chain is enhanced. Counterfactual verification is the gold standard of causal reasoning. It simulates or searches for scenarios where the cause nodes are not abnormal in the data, and then observes the state of the result nodes. For example, if the temperature of the cause node was not abnormal at that time, would the vibration of the result node be abnormal? Verification determines the anomaly explanation results containing the exact cause, reliable propagation chain, and accurate potential area of ​​influence, and outputs the results.

[0044] Step S400: Perform exception reporting management based on the exception explanation results.

[0045] Step S400 further includes configuring an early warning response signal based on the anomaly explanation result, and using the early warning response signal for anomaly reporting management.

[0046] Preferably, a structured early warning response signal is generated based on the anomaly explanation results, including the early warning level, response type, potential impact area, root cause location information, and expected evolution path. The early warning level is comprehensively assessed based on the fundamental nature of the cause, the length and speed of the propagation chain, and the scope and importance of the impact area. The potential impact area ensures that early warning information is only sent to management personnel in the relevant area. The root cause location information is the source of the anomaly. The expected evolution path is determined based on the propagation chain. Then, the early warning response signal is used for anomaly reporting management. Specifically, different levels and types of signals trigger different reporting channels. For example, the highest-level alarm simultaneously activates audible and visual alarms, SMS notifications, and work order creation; medium-level maintenance early warnings display a pop-up notification on the work interface. The early warning response signal is automatically routed to the corresponding area's maintenance personnel based on the potential impact area. Furthermore, it can identify multiple derivative anomalies caused by the same root cause and mark chain anomalies as "explained," thereby significantly reducing redundant alarms. This achieves accurate, targeted, operable, and traceable anomaly reporting management, improving resource utilization efficiency and the explainability of risk decisions.

[0047] In the above text, refer to Figure 1 A deep learning-based IoT monitoring and management method according to embodiments of the present invention is described in detail. Next, reference will be made to... Figure 2 This invention describes a deep learning-based Internet of Things (IoT) monitoring and management system according to an embodiment of the present invention.

[0048] The deep learning-based IoT monitoring and management system according to embodiments of the present invention addresses the technical problems in existing IoT monitoring, such as lack of causal interpretability, rigid resource allocation, and insufficient anomaly response capabilities. It achieves the technical effects of precise anomaly management of IoT nodes, improved resource utilization efficiency, and enhanced interpretability of risk decisions. Figure 2 As shown, the IoT monitoring and management system based on deep learning includes: a causal structure field construction module 10, a behavior strategy establishment module 20, a joint reasoning analysis module 30, and an anomaly reporting management module 40.

[0049] The causal structure field construction module 10 is used to configure autonomous agents at IoT monitoring nodes. After reading multi-source monitoring data from IoT monitoring nodes using the autonomous agents, a deep causal structure field containing environmental causes, state change paths, and risk propagation relationships is established through cross-modal deep causal reasoning. The behavior strategy establishment module 20 is used to construct a spatiotemporal causal tensor to characterize the dynamic risk distribution of the monitoring area based on the deep causal structure field. The autonomous agents use the spatiotemporal causal tensor to perform quantitative assessments of monitoring benefits, energy consumption costs, and risk exposure, and perform game analysis under a multi-party game strategy to establish agent behavior strategies. The joint reasoning analysis module 30 is used to dynamically update the IoT monitoring structure through topological evolution based on the agent behavior strategies, perform deep spatiotemporal-causal joint reasoning analysis under a self-organizing layout, and establish anomaly explanation results. The anomaly explanation results include anomaly causes, propagation chains, and potential impact areas. The anomaly reporting management module 40 is used to manage anomaly reporting based on the anomaly explanation results.

[0050] The specific configuration of the behavior strategy establishment module 20 will be described in detail below. The behavior strategy establishment module 20 further includes: after parsing the spatiotemporal causal tensor, using an autonomous agent to evaluate the causal closed-loop stability of the corresponding monitoring nodes based on the parsing results, and establishing a causal closed-loop stability evaluation result; after extracting the dynamic risk value of the monitoring area based on the spatiotemporal causal tensor, constructing neighborhood cooperation constraints based on the dynamic risk value of the monitoring area and the causal closed-loop stability evaluation result; and using the neighborhood cooperation constraints as constraints for game analysis to execute game constraint management of the agent's behavior strategy.

[0051] The specific configuration of the behavior strategy establishment module 20 will be described in detail below. The behavior strategy establishment module 20 further includes: after establishing constraints, using autonomous agents to perform joint utility modeling on monitoring benefits, energy consumption costs, and risk exposure, and integrating them according to preset risk preference and benefit preference rules to form node utility results describing the behavioral tendencies of monitoring nodes; using the node utility results and the causal closed-loop stability evaluation results as the basic dataset for game analysis; based on the basic dataset, constructing a local decision-making task with neighborhood cooperation constraints for each autonomous agent, the local decision-making task being used to limit the adjustable range of behavioral parameters of the autonomous agent in terms of sampling frequency, monitoring direction, data reporting path, and neighborhood cooperation strength; using a distributed policy update mechanism to execute the task-solving evolutionary game iteration of the local decision-making task, and outputting the agent's behavior strategy.

[0052] The specific configuration of the behavior strategy establishment module 20 will be described in detail below. The behavior strategy establishment module 20 further includes: during the iterative process after acquiring the initial local decision-making task, the autonomous agent periodically receives the behavioral parameter adjustment intentions of neighboring agents based on the cooperative constraints of its neighborhood, forming policy interaction information to characterize the trend of neighborhood policy changes; before updating its own policy, the autonomous agent performs constraint consistency verification on candidate policies, which is used to detect the feasibility of candidate policies in terms of sampling frequency, monitoring direction, data reporting path, and neighborhood cooperation strength, and outputs the arbitrated and corrected candidate policies; based on the output results, the autonomous agent uses the node utility results and policy interaction information to perform policy response calculation, selects a response policy from the output results to replace the current policy based on the policy response calculation results; performs multi-round iterative game evolution, and outputs the agent's behavior policy.

[0053] The specific configuration of the joint reasoning analysis module 30 will be described in detail below. The joint reasoning analysis module 30 further includes: dynamically evaluating the node connection relationships, data flow paths, and cooperation scope of the current IoT monitoring structure using the autonomous agent based on the agent's behavior strategy; identifying redundant, inefficient, and high-conflict links in the structure to form a structure adaptability result reflecting the topology's evolvability; using the structure adaptability result, determining link maintenance, link reconnection, and link weight redistribution operations according to neighborhood cooperation constraints and its own behavior strategy, generating topology evolution instructions, which are used to gradually update the local monitoring structure in a distributed cooperative environment; based on the topology evolution instructions, performing self-organizing layout analysis on the updated local monitoring structure, updating the causal correlation strength, information propagation direction, and risk coupling paths between monitoring nodes to form a structured layout expression; and performing deep spatiotemporal-causal joint reasoning analysis management under the self-organizing layout based on the structured layout expression.

[0054] The following will describe the specific configuration of the joint inference analysis module 30 in detail. The joint inference analysis module 30 further includes: using the structured layout expression as input, performing hierarchical aggregation of spatiotemporal observation data and causal relationships at three scales: node level, subgraph level, and region level, respectively generating node situation fragments, subgraph collaboration fragments, and region risk fragments, constructing a data semantic layer for joint inference; based on the data semantic layer, using a graph structure-aware temporal representation to synchronously encode the spatiotemporal features at each scale, the temporal representation being used to simultaneously capture the topological coupling relationships between nodes, the directionality of information propagation, and the temporal evolution trend, outputting a spatiotemporal embedding vector for causal identification; performing explicit scoring and filtering of the spatiotemporal embedding vector under a causal attention mechanism to generate candidate causal chains ranked by confidence; performing local data-driven forward simulation and counterfactual verification on the candidate causal chains, constructing and outputting anomaly explanation results based on the verification results.

[0055] The specific configuration of the anomaly reporting management module 40 will be described in detail below. The anomaly reporting management module 40 further includes: configuring an early warning response signal based on the anomaly interpretation result, and using the early warning response signal for anomaly reporting management.

[0056] The IoT monitoring and management system based on deep learning provided in this embodiment of the invention can execute the IoT monitoring and management method based on deep learning provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A deep learning-based IoT monitoring and management method, characterized in that, The method includes: Autonomous agents are configured on IoT monitoring nodes. After reading multi-source monitoring data from IoT monitoring nodes using the autonomous agents, a deep causal structure field containing environmental factors, state change paths, and risk propagation relationships is established through cross-modal deep causal reasoning. Based on the deep causal structure field, a spatiotemporal causal tensor is constructed to characterize the dynamic risk distribution of the monitoring area. The autonomous agent is used to perform quantitative evaluation of monitoring benefits, energy consumption costs, and risk exposure based on the spatiotemporal causal tensor, and to perform game analysis under a multi-party game strategy to establish the agent's behavior strategy. Based on the intelligent agent's behavior strategy, the IoT monitoring structure is dynamically updated through topological evolution, and deep spatiotemporal-causal joint reasoning analysis under self-organizing layout is performed to establish anomaly explanation results, which include anomaly causes, propagation chains, and potential impact areas. Anomaly reporting management is performed based on the anomaly explanation results.

2. The IoT monitoring and management method based on deep learning as described in claim 1, characterized in that, Game analysis under multi-party game strategies, establishing agent behavior strategies, including: After analyzing the spatiotemporal causal tensor, an autonomous intelligent agent is used to evaluate the causal closed-loop stability of the corresponding monitoring nodes based on the analysis results, and a causal closed-loop stability evaluation result is established. After extracting the dynamic risk value of the monitoring area based on the spatiotemporal causal tensor, a neighborhood cooperation constraint is constructed based on the dynamic risk value of the monitoring area and the causal closed-loop stability evaluation result. The neighborhood cooperation constraint is used as a constraint condition for game analysis to perform game constraint management of the agent's behavior strategy.

3. The IoT monitoring and management method based on deep learning as described in claim 2, characterized in that, Using the neighborhood cooperation constraint as a constraint condition for game analysis, the game constraint management of the agent's behavioral strategy is executed, including: After establishing the constraints, the autonomous intelligent agent is used to perform joint utility modeling on the monitoring benefits, energy consumption costs and risk exposure. The results are then weighted and integrated according to the preset risk preference and benefit preference rules to form the node utility results used to describe the behavioral tendencies of the monitoring nodes. The node utility results and the causal closed-loop stability evaluation results are used as the basic dataset for game analysis. Based on the aforementioned basic dataset, a local decision-making task with neighborhood cooperation constraints is constructed for each autonomous agent. The local decision-making task is used to limit the adjustable range of behavioral parameters of the autonomous agent in terms of sampling frequency, monitoring direction, data reporting path, and neighborhood cooperation strength. The task-solving evolutionary game iterative process utilizes a distributed policy update mechanism to perform local decision-making tasks and outputs the agent's behavioral policy.

4. The IoT monitoring and management method based on deep learning as described in claim 3, characterized in that, The task-solving evolutionary game iterative method utilizes a distributed policy update mechanism to perform local decision-making tasks, outputting agent behavior policies, including: During the iterative process after acquiring the initial local decision-making task, the autonomous agent periodically receives the behavioral parameter adjustment intentions of the neighboring agents based on the neighborhood cooperation constraints, forming policy interaction information to characterize the trend of neighborhood policy changes. Before updating its own strategy, the autonomous agent performs a constraint consistency verification on the candidate strategy. The constraint consistency verification is used to detect the feasibility of the candidate strategy in terms of parameters such as sampling frequency, monitoring direction, data reporting path and neighborhood cooperation strength, and outputs the candidate strategy after arbitration correction. Based on the output results, the autonomous agent uses the node utility results and policy interaction information to perform policy response calculations, and selects a response policy from the output results to replace the current policy based on the policy response calculation results. Perform multiple rounds of iterative game evolution and output the agent's behavioral strategy.

5. The IoT monitoring and management method based on deep learning as described in claim 1, characterized in that, Based on the aforementioned agent behavior strategy, the IoT monitoring structure is dynamically updated through topology evolution, including: Based on the aforementioned agent behavior strategy, the autonomous agent is used to dynamically evaluate the node connection relationship, data flow path and cooperation scope of the current Internet of Things monitoring structure, and to identify redundant links, inefficient links and high-conflict links in the structure, forming a structural adaptability result that reflects the topology's evolvability. Using the structure fit results, the operations of link maintenance, link reconnection, and link weight reallocation are jointly determined according to the neighborhood cooperation constraints and its own behavior strategy, and a topology evolution instruction is generated. The topology evolution instruction is used to gradually update the local monitoring structure in a distributed cooperative environment. Based on the aforementioned topology evolution instructions, the updated local monitoring structure is executed to perform self-organizing layout analysis, update the causal correlation strength, information propagation direction, and risk coupling path between monitoring nodes, and form a structured layout expression. According to the structured layout expression, deep spatiotemporal-causal joint reasoning analysis management is performed under a self-organizing layout.

6. The IoT monitoring and management method based on deep learning as described in claim 5, characterized in that, Based on the structured layout expression, deep spatiotemporal-causal joint reasoning analysis and management under a self-organizing layout are performed, including: Using the structured layout expression as input, spatiotemporal observation data and causal relationships are hierarchically aggregated at three scales: node level, subgraph level, and region level, generating node situation fragments, subgraph collaboration fragments, and region risk fragments respectively, and constructing a data semantic layer for joint reasoning; Based on the data semantic layer, a graph structure-aware temporal representation is used to synchronously encode the spatiotemporal features at each scale. The temporal representation is used to simultaneously capture the topological coupling relationship between nodes, the directionality of information propagation, and the temporal evolution trend, and outputs a spatiotemporal embedding vector for causal identification. The spatiotemporal embedding vectors are subjected to explicit scoring and filtering under a causal attention mechanism to generate candidate causal chains ranked by confidence. The candidate causal chains are subjected to local data-driven forward simulation and counterfactual verification. Anomaly explanations are constructed and output based on the verification results.

7. The IoT monitoring and management method based on deep learning as described in claim 1, characterized in that, Configure an early warning response signal based on the anomaly explanation result, and use the early warning response signal for anomaly reporting management.

8. A deep learning-based IoT monitoring and management system, characterized in that, The system is used to implement the deep learning-based IoT monitoring and management method according to any one of claims 1 to 7, the system comprising: The causal structure field construction module is used to configure autonomous intelligent agents in IoT monitoring nodes. After reading multi-source monitoring data of IoT monitoring nodes using the autonomous intelligent agents, a deep causal structure field containing environmental causes, state change paths, and risk propagation relationships is established through cross-modal deep causal reasoning mode. The behavior strategy establishment module is used to construct a spatiotemporal causal tensor based on the deep causal structure field to characterize the dynamic risk distribution of the monitoring area, and to use the autonomous agent to perform quantitative evaluation of monitoring benefits, energy consumption costs, and risk exposure based on the spatiotemporal causal tensor, perform game analysis under multi-party game strategy, and establish the agent's behavior strategy. The joint reasoning analysis module is used to dynamically update the IoT monitoring structure through topological evolution based on the agent's behavior strategy, perform deep spatiotemporal-causal joint reasoning analysis under self-organizing layout, and establish anomaly explanation results, including anomaly causes, propagation chains, and potential impact areas. The exception reporting management module is used to manage exception reporting based on the exception explanation results.

Images