Industrial safety risk monitoring and early warning system and method of autonomous planning intelligent agent

Abstract

The invention provides an industrial safety risk monitoring and early warning system and method for an autonomous planning intelligent agent, relates to the technical field of industrial safety production, and solves various limitation problems still existing in an existing safety risk monitoring scheme. In the system, an environment sensing layer is used for collecting multi-source sensing data in an industrial production environment in real time; the edge calculation layer is used for carrying out primary processing on the multi-source sensing data and executing data compression and key information extraction so as to generate key analysis data; the cloud analysis layer is used for performing deep intelligent analysis on the key analysis data, executing security risk assessment and risk diffusion prediction, and dynamically generating and adjusting an early warning decision through an autonomous planning agent; and the action execution layer is used for executing a corresponding safety action instruction according to the early warning decision and feeding back an execution result to the cloud analysis layer to realize system optimization. According to the invention, the risk early warning accuracy and the emergency disposal timeliness can be improved finally.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial safety production, and in particular to an industrial safety risk monitoring and early warning system and method of autonomous planning agent. BACKGROUND

[0002] In the field of industrial safety monitoring and risk early warning, the existing technology mainly relies on traditional sensor networks, fixed rule models and centralized data analysis methods. These methods realize real-time perception of the environment state by deploying various physical sensors, upload the collected data to the central server or cloud for unified processing, and conduct risk assessment combined with artificial experience. Some systems also introduce edge computing nodes to assist in local feature extraction, while security, fire safety and production monitoring subsystems usually operate independently. Although such solutions have achieved certain results in basic information collection and preliminary analysis, their overall efficiency still faces significant challenges in complex industrial scenarios.

[0003] The existing technology has obvious deficiencies in the analysis ability in the time and space dimensions. The sensor network is often regarded as an isolated data point, without fully considering the dynamic correlation between spatial distribution, perception range and risk propagation path. For example, the static deployment strategy ignores the risk diffusion between regions, resulting in blind spots in monitoring coverage, making it difficult to accurately simulate the evolution trend of risk from the source to the surrounding. At the same time, traditional methods tend to separate the processing of time series and spatial information, lacking effective mechanisms for their fusion, which limits the accurate modeling of the dynamic evolution process of risk and affects the completeness and response timeliness of monitoring.

[0004] Human factors introduce an unavoidable bias in risk assessment. In the face of massive multi-dimensional data, operators rely too much on subjective experience for judgment, which is easy to miss low-frequency high-risk events or misjudge normal fluctuations as abnormal conditions. Due to the lack of systematic knowledge support and updating mechanism, the evaluation results vary from person to person and are difficult to maintain consistency. In addition, the existing system fails to integrate the latest regulations and policies, industry standards and accident cases in a timely manner, resulting in a lag in risk identification compared to actual needs, reducing the accuracy and reliability of early warning.

[0005] The lack of dynamic autonomous planning capability further restricts the adaptability of the system. Current solutions mostly use pre-set fixed analysis processes, which can only trigger simple alarms when an anomaly occurs, and cannot adjust the data collection frequency or analysis depth in real time according to the risk level. This static mechanism reacts slowly in the face of sudden risks, making it difficult to dig deep into the correlation between data, resulting in waste of information value. The system also lacks autonomous reasoning capability based on intelligent agents, which cannot complete the closed-loop process from perception to decision-making without human intervention, weakening the overall response efficiency.

[0006] The synergy efficiency between the edge and the cloud is generally low. The edge computing architecture often has problems such as ambiguous function division and unreasonable resource scheduling, for example, simple tasks are handled by the edge end, and complex analysis is all concentrated in the cloud, resulting in insufficient utilization of edge resources and excessive load on the cloud. In a high-concurrency scenario, data transmission delay increases significantly, seriously affecting real-time response capability. At the same time, there is a lack of unified model coordination mechanism between the edge and the cloud, making the same risk have inconsistent judgment standards at different levels, forming an information island and reducing the robustness of the system.

[0007] The lack of cross-system linkage capability is also a prominent defect of the existing technology. Industrial systems generally use a chimney architecture, with security, fire safety, and production monitoring modules isolated from each other, lacking a unified data sharing platform. For example, when a fire occurs, the alarm from the fire safety system cannot be effectively linked with video monitoring or personnel positioning systems, making it difficult for emergency command to grasp the situation on the ground in a timely manner. In addition, there is a lack of effective correlation between dynamic sensing data and static description data, limiting the comprehensive analysis capability of complex events and delaying critical disposal opportunities.

[0008] To address the above problems, existing research has attempted to introduce artificial intelligence and big data technology for improvement. Some solutions improve monitoring accuracy through multi-source data fusion, but still do not solve the bottleneck of edge and cloud synergy; others use machine learning models for risk prediction, but are difficult to adapt to dynamic scenarios due to single training data and rigid feature engineering; some explore the use of graph neural networks to enhance spatial modeling, but ignore the integration of the time dimension, resulting in incomplete risk evolution characterization; and attempts at natural language processing technology are limited by semantic understanding capabilities and cannot truly replace professional judgment. These efforts, while making some progress in specific areas, have not fundamentally innovated the system architecture and analysis paradigm, but may instead introduce new challenges such as decreased model explainability or computational resource overload.

[0009] In summary, existing industrial safety monitoring and risk warning technologies have systematic defects in spatio-temporal fusion, cognitive enhancement, dynamic planning, edge-cloud synergy, and cross-system linkage, making it difficult to meet the real-time, accuracy, and intelligence needs of complex industrial scenarios. Therefore, there is an urgent need to develop a new generation of monitoring and warning system that can deeply integrate spatio-temporal characteristics, enhance cognitive ability, support dynamic autonomous decision-making, and achieve efficient synergy, to comprehensively improve risk perception and response levels. SUMMARY

[0010] The present application aims to solve the various limitations of the existing security risk monitoring scheme, such as insufficient spatio-temporal fusion analysis, cognitive bias of personnel, insufficient dynamic autonomous planning analysis capability, low edge cloud cooperation efficiency, weak cross-system linkage capability, etc., and therefore proposes an industrial security risk monitoring and early warning system and method of autonomous planning agents. The present application improves the spatio-temporal coverage capability of security risk monitoring, reduces cognitive bias, enhances dynamic analysis efficiency, optimizes edge cloud cooperation, realizes cross-system intelligent linkage, and finally improves the risk early warning accuracy and emergency disposal timeliness by means of autonomous perception of the environment, autonomous cognition of security risks, autonomous planning of task execution, autonomous data collection, autonomous execution and evaluation of results.

[0011] The present application adopts the following technical solutions to achieve the purpose: An industrial security risk monitoring and early warning system of autonomous planning agents, comprising an environment perception layer, an edge computing layer, a cloud analysis layer and an action execution layer. The environment perception layer is used to collect multi-source perception data in the industrial production environment in real time, including environment parameters, personnel status and equipment running state information. The edge computing layer is used to preliminarily process the multi-source perception data, perform data compression and key information extraction, and generate key analysis data. The cloud analysis layer is used to perform deep intelligent analysis on the key analysis data, execute security risk assessment and risk diffusion prediction, and dynamically generate and adjust early warning decisions through autonomous planning agents. The action execution layer is used to execute corresponding safety action instructions according to the early warning decisions, and feedback the execution results to the cloud analysis layer to realize system optimization.

[0012] Specifically, the environment perception layer includes a sensor array for collecting multi-source perception data in the industrial production environment in real time, and the types of the multi-source perception data include video image data, environment temperature and humidity and harmful gas concentration data, field sound data and electrical equipment current and voltage parameter data.

[0013] Specifically, the edge computing layer is used to reduce transmission load by performing data dimensionality reduction compression processing, and apply a dynamic target detection model to identify abnormal behaviors in image data in real time, filter out key analysis data and optimize data transmission efficiency.

[0014] Specifically, the cloud analysis layer includes a multi-modal analysis unit, a space-time fusion engine, and an autonomous planning agent; the multi-modal analysis unit is used for integrating image, sound, and text data for comprehensive risk assessment; the space-time fusion engine is used for combining time and space dimensions to perform risk diffusion simulation, predict risk propagation path and impact range; the autonomous planning agent is used for dynamically generating early warning decision strategies based on a historical case library and a regulation standard library, and adjusting data collection and analysis processes according to real-time feedback.

[0015] Specifically, the action execution layer is used for converting the early warning decision into a safety action instruction and implementing execution, and simultaneously performing success or failure evaluation on the execution result, and feeding back error analysis information to the cloud analysis layer to optimize subsequent decisions when the evaluation fails.

[0016] The application also provides an industrial safety risk monitoring and early warning method of an autonomous planning agent, a hardware basis of the method being the industrial safety risk monitoring and early warning system; the method comprising: receiving a task target and a current environment state, and defining a task decomposition rule; collecting multi-source perception data of an industrial production environment in real time through an environment perception module pre-deployed in the environment perception layer; delivering the multi-source perception data to a cognitive engine module pre-deployed in the edge computing layer for processing and analysis, and the cognitive engine module judging whether new task planning is needed based on the analysis result; if it is judged that new task planning is needed, triggering a task planning module pre-deployed in the cloud analysis layer, which generates a task tree according to the task target, the current environment state, and the preset task decomposition rule; if it is judged that new task planning is not needed, directly triggering an action execution module pre-deployed in the action execution layer; after the task planning module generates the task tree, delivering it to the action execution module; the action execution module converts the preset or newly generated task tree into an action instruction and executes it, while monitoring the task progress in real time; evaluating the execution result, updating a state model of the industrial production environment if the evaluation is successful, and performing error analysis and delivering feedback information to the task planning module for adjustment and optimization if the evaluation fails; dynamically adjusting the task planning and execution strategies based on the evaluation result, forming a closed-loop decision and execution process.

[0017] Specifically, the task planning module realizes the process of generating a task tree according to a task target, a current environment state, and a preset task decomposition rule through a built-in task planning evaluation algorithm; the comprehensive evaluation function of the task planning evaluation algorithm adopts a weighted evaluation function for the final planning quality.

[0018] ​Preferably, the action execution module selects the optimal action matching the current environment state from the predefined action library through the state-action value mapping; executes the selected action and receives environment feedback; updates the state-action value mapping based on the feedback; and dynamically determines whether to continue or terminate the task execution process according to the updated environment state.

[0019] Preferably, the action execution module obtains the reward by receiving environment feedback The reward The corresponding algorithm formula is the reinforcement learning reward function formula.

[0020] Preferably, the task requirements and execution results are stored in the training data set; if the execution result is successful, the knowledge base is updated to accumulate experience data and strengthen subsequent task planning; if the execution result is unsuccessful, the failure reason is analyzed and the configuration of the task planning module, the action execution module or the environment perception module is dynamically adjusted according to the reason; and the system is self-adaptively optimized through knowledge base updating.

[0021] In summary, due to the adoption of the technical solutions, the present application has the following advantages: The present application significantly improves the precision, real-time performance and automation level of industrial safety management through the intelligent agent-driven dynamic risk monitoring and autonomous early warning system. The system uses multi-modal fusion analysis technology to integrate visual, acoustic and sensor data, and combines temporal and spatial diffusion methods to effectively improve the identification accuracy of high-risk scenarios such as gas leakage and equipment overload, significantly reducing the false negative rate, thereby enhancing the comprehensiveness and reliability of risk perception.

[0022] In terms of emergency early warning, the present application realizes efficient collaborative operation of the edge and cloud. The edge is responsible for real-time detection and preliminary analysis, while the cloud dynamically plans and generates targeted disposal schemes. The close cooperation between the two greatly shortens the overall response time, ensuring that risk events can be timely intervened and handled, and improving the timeliness and effectiveness of emergency disposal.

[0023] The intelligent agent-based linkage mechanism successfully breaks down the data barriers between security, fire safety and production monitoring systems, enabling seamless cross-system collaboration. Each subsystem shares information and responds in unison under a unified framework, providing complete situational awareness for emergency command and avoiding disposal delays caused by information silos, significantly enhancing the comprehensive response capability in complex scenarios.

[0024] The system also has self-adaptive learning capability, which considers safety, efficiency and resource balance through the reinforcement learning reward function, and combines with the retrieval-enhanced knowledge base to continuously learn and optimize decision-making strategies after deployment. This mechanism effectively reduces the false positive rate, improves the accuracy and adaptability of early warning, and enables the system to dynamically adapt to the changing needs of industrial scenarios.

[0025] This invention realizes a paradigm upgrade from passive response to autonomous prediction, planning and execution, providing an intelligent solution for industrial safety that covers the entire chain of risk perception, analysis, early warning and disposal, and comprehensively improving the intelligence level and overall efficiency of safety management. Attached Figure Description

[0026] The present invention further illustrates its embodiments and technical solutions in detail with reference to the following accompanying drawings, specifically including 5 figures as follows: Figure 1 This is a schematic diagram illustrating the composition, structure, and functional principle of the system of the present invention; Figure 2 This is a simplified schematic diagram illustrating the overall logical flow of the method of the present invention; Figure 3 Example diagram illustrating the specific process of the method of the present invention; Figure 4 This is a detailed flowchart illustrating the action execution process in the method of the present invention; Figure 5 This is a detailed flowchart illustrating the reflection and learning process in the method of this invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0028] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0029] Example 1 like Figure 1 As shown, an industrial safety risk monitoring and early warning system for autonomous planning intelligent agents includes an environmental perception layer, an edge computing layer, a cloud analysis layer, and an action execution layer. The environmental perception layer is used to collect multi-source sensing data in the industrial production environment in real time, including environmental parameters, personnel status, and equipment operating status information; The edge computing layer is used to perform preliminary processing on multi-source sensing data, perform data compression and key information extraction to generate key analysis data; The cloud analysis layer is used for in-depth intelligent analysis of key analysis data, performs security risk assessment and risk diffusion prediction, and dynamically generates and adjusts early warning decisions through autonomous planning intelligent agents; The action execution layer is used to execute corresponding security action instructions according to the early warning decisions, and feedback the execution results to the cloud analysis layer to realize system optimization.

[0030] In this embodiment, the environment perception layer includes a sensor array for real-time collection of multi-source perception data of the industrial production environment. The types of multi-source perception data include video image data, environmental temperature and humidity and harmful gas concentration data, on-site sound data, and electrical equipment current and voltage parameter data. The environment perception layer is the data source of the entire system, responsible for real-time collection of various environmental parameters, personnel status, and equipment status information. The static safety profile and dynamic real-time data of the production site safety risk, and various hardware collection devices for dynamic real-time data can specifically include: Video camera: used to capture on-site images and videos, providing visual basis for subsequent behavior recognition and anomaly detection; Temperature and humidity / gas sensor: monitors temperature, humidity, and harmful gas concentration in the environment to ensure the safety of the working environment; Sound pickup: records on-site sound, which can be used for voice recognition and acoustic anomaly detection; Current and voltage sensor: monitors the operating state of electrical equipment to prevent safety accidents caused by power failure.

[0031] These sensors transmit raw data streams to the edge computing layer, laying the foundation for subsequent data processing. In this embodiment, the edge computing layer is used to reduce transmission load by performing data dimensionality reduction compression processing, and applies a dynamic target detection model for real-time anomaly behavior recognition of image data, filters out key analysis data, and optimizes data transmission efficiency. The edge computing layer undertakes the tasks of preliminary data processing and key information extraction, which can specifically include: Edge computing box: as a data aggregation point, it receives raw data from the environment perception layer and performs preliminary processing; Data dimensionality reduction and compression unit: reduces the dimensionality and compresses a large amount of raw data, reducing the amount of data and improving transmission efficiency; YOLO dynamic detection model: a dynamic detection model based on the YOLO algorithm, which can quickly identify abnormal behaviors or objects in images and filter out key analysis data.

[0032] In this embodiment, the cloud analysis layer includes a multi-modal analysis unit, a space-time fusion engine, and an autonomous planning agent. The multi-modal analysis unit is used to integrate image, sound, and text data for comprehensive risk assessment. The space-time fusion engine is used to combine time and space dimensions to perform risk diffusion simulation, predict risk propagation path and impact range. The autonomous planning agent is used to dynamically generate early warning decision strategies based on historical case library and regulatory standard library, and adjust data collection and analysis process according to real-time feedback.

[0033] In the cloud analysis layer, a cloud intelligence hub is configured as the control center of the entire cloud analysis layer, coordinating the work of each module to ensure efficient operation of the system. It enables the cloud analysis layer to conduct multi-dimensional and deep intelligent analysis on key data. In the example of this embodiment, it can be composed of the following modules or units: Multi-modal analysis unit: integrates multiple types of data, including images, sounds, texts, etc., and uses deep learning and natural language processing techniques for comprehensive analysis; Agent matrix: contains multiple agents, each responsible for different analysis tasks; Knowledge base RAG system: combines historical case library and regulatory standard library to provide reference for decision-making; Quality evaluation agent: evaluates data quality and analysis results to ensure accuracy and reliability of analysis; Autonomous planning agent: automatically adjusts data collection and analysis strategies according to current situation to improve system flexibility and adaptability; Cross-modal parser: unifies and fuses data of different modalities to generate comprehensive analysis results; Space-time fusion engine: considers time and space factors for multi-dimensional analysis of data to reveal potential risk trends; Risk diffusion simulation unit: based on the analysis results of the space-time fusion engine, simulates the diffusion process of risks in time and space, predicts possible impact range and consequences; Trajectory intersection space-time fusion safety risk prediction algorithm: performs time diffusion calculation and space diffusion calculation on unsafe behavior of people and unsafe state of objects; Safety risk assessment unit: assesses based on historical probability of unsafe behavior of people; Safety risk prediction unit: realizes static risk description and dynamic identification of system risk, dynamically identifies system risk occurrence, and dynamically calculates system risk release and diffusion.

[0034] In this embodiment, the action execution layer is used to convert early warning decisions into safety action instructions and implement execution, while evaluating success or failure of execution results, and feeding back error analysis information to the cloud analysis layer to optimize subsequent decisions when evaluation fails.

[0035] Through the above multi-level and multi-module design, the system of the embodiment realizes full-process automation from data collection to intelligent analysis, effectively improving the ability and efficiency of industrial safety monitoring and early warning.

[0036] When the above system is deployed to a certain chemical laboratory, in terms of hardware, the environment perception layer can specifically include H2S sensors, infrared thermal imagers and other devices; the edge computing box of the edge computing layer has a pre-built YOLO v8 model trained in advance; the cloud analysis layer has 4 sets of K100 GPU servers. Combined with the relevant rules of the construction guide for the safety emergency intelligent system of the chemical laboratory, the configuration of the system functions can include six parts of dangerous goods management, temperature and humidity management, gas leakage management, fire-fighting facility management, electrical equipment and line management, and administrative management; the personnel involved are professional scientific researchers and a small number of visitors, and the point positions to be monitored in each type of area are different, and the planning is according to the double-light monitoring point position planning of each area. The key monitoring matters include fire smoke detection, personnel quantity, illegal personnel intrusion, etc.; the above system functions are configured therein to realize industrial safety risk monitoring and early warning.

[0037] Embodiment 2 On the basis of embodiment 1, the embodiment provides an industrial safety risk monitoring and early warning method of an autonomous planning intelligent agent, and the hardware basis of the method is the industrial safety risk monitoring and early warning system in embodiment 1; the main content of the method includes receiving a task target and a current environment state, and defining a task decomposition rule; The multi-source perception data of the industrial production environment are collected in real time by the environment perception module pre-deployed in the environment perception layer; The multi-source perception data are transmitted to the cognitive engine module pre-deployed in the edge computing layer for processing and analysis, and the cognitive engine module judges whether new task planning is needed based on the analysis result; If it is judged that new task planning is needed, the task planning module pre-deployed in the cloud analysis layer is triggered, which generates a task tree according to the task target, the current environment state and the preset task decomposition rule; if it is judged that new task planning is not needed, the action execution module pre-deployed in the action execution layer is directly triggered; After the task planning module generates the task tree, the task tree is transmitted to the action execution module; The action execution module converts the preset or newly generated task tree into an action instruction and executes it, while monitoring the task progress in real time; The execution result is evaluated, if the evaluation is successful, the state model of the industrial production environment is updated, if the evaluation fails, error analysis is performed and feedback information is transmitted to the task planning module for adjustment and optimization; Based on the evaluation result, the task planning and execution strategy are dynamically adjusted to form a closed-loop decision and execution process.

[0038] Overall, the method of this embodiment is based on the perception of the industrial site environment by the environment perception layer and the decision of the cognitive engine module in the edge computing layer to implement the corresponding process execution. First, the system collects external information through the environment perception module, and then passes the data to the cognitive engine module for processing and analysis; the cognitive engine module determines whether a new plan is needed according to the current state, and if the result is "yes", a new task tree is generated in the task planning module; if the result is "no", it directly jumps to the action execution module.

[0039] In the task planning module, the system generates a detailed task tree according to the current environment and the target, and passes it to the action execution module. The action execution module is responsible for converting the task tree into specific action instructions and further executing. After the action execution is completed, the system enters the result evaluation stage to analyze and judge the execution result; if the evaluation result is "success", the state model of the industrial production environment is updated to reflect the latest environment state; if the evaluation result is "failure", error analysis is performed to find out the problem and feedback to the task planning module for adjustment and optimization.

[0040] The main process of the whole method can form a closed-loop diagram as shown in Figure 2 , which ensures that the system can dynamically adjust the strategy according to the changes in the environment and the task requirements, and realize efficient and accurate decision-making and execution.

[0041] In this embodiment, the environment perception module collects audio and video data, temperature and humidity, gas concentration, current and voltage, etc. Internet of Things data, and interprets the on-site environment in combination with the production site static safety archive history. The cognitive engine module needs to input the task target and the current environment state, and define the task decomposition rules; then, if task planning is needed, the task planning module formulates a task decomposition plan according to the target through a large model, which relies on pre-set basic rules, including step restrictions for task decomposition, requirements for atomic operations, and generation of alternative solutions, to ensure the feasibility and flexibility of the plan, finally constructs a prompt word and generates a task execution scheme, forming a task tree; after selecting the task execution scheme, it enters the execution task stage, and adopts the method of thinking while doing in the process, and monitors the task progress in real time; according to the task execution result, it is judged whether the plan needs to be adjusted, if it needs to be adjusted, it returns to the sub-task execution result feedback link for a new round of optimization and execution, until the task is completed, the process is ended. The method flow of this part can refer to the specific diagram of Figure 3 .

[0042] In this embodiment, the quality of task planning depends on three aspects: the causal consistency between the objective and the outcome, the efficiency of task execution (i.e., the executability of task decomposition), and the robustness of the task. The task planning module, through its built-in task planning evaluation algorithm, generates a task tree based on the task objective, the current environmental state, and preset task decomposition rules. The comprehensive evaluation function of the task planning evaluation algorithm assesses the quality of the final plan. The weighted evaluation function is as follows:

[0043] In the formula, , , These represent the corresponding weight values. This embodiment suggests... , , ; Represents the degree of goal achievement; Represents planning efficiency; Represents anti-interference capability; This represents causal consistency; This represents taking the minimum value.

[0044] Goal Achievement It is used to measure the degree of matching between planning results and expected goals, and its calculation formula is as follows:

[0045] In the formula, Representing the Individual goals; Represents the set of goals actually achieved; Represents the first The weight values ​​corresponding to each sub-objective; As an indicator function, it is 1 when the goal is achieved and 0 when it is not achieved; based on the degree of goal achievement... The calculation results are used to determine the success or failure of the task objective based on a preset threshold. The task objective was met and the result was excellent. Only qualified at that time. The mission objective failed.

[0046] Planning efficiency The formula used to quantify the cost-effectiveness of resource consumption versus time is as follows:

[0047] In the formula, , , All three are normalization coefficients, and their values ​​can be equal, all of which are set to 1. Represents the value of the task; It represents the sum of task execution time and task planning time, and the task planning time has a preset weight; Represents the weighted sum of system resources; for planning efficiency When the calculated result is less than 0.1, the task planning evaluation algorithm will re-plan the task.

[0048] Anti-interference capability Robustness entropy is used to assess the ability to cope with interference, and its calculation formula is as follows:

[0049] In the formula, Representing the Task success rate under various interference scenarios; Represents the total number of interference scenario types; based on anti-interference capability The calculation results distinguish the strength of the system's anti-interference ability. When the system is fully robust, it has strong anti-interference capabilities; when At that time, the system is relatively fragile and has weak anti-interference ability.

[0050] Causal consistency Used as a veto factor to verify the causal rationality of the planning logic, and when If the calculation result is less than a first preset threshold, the planning is directly deemed to have failed. This first preset threshold can be set to 0.5. The calculation formula is as follows:

[0051] In the formula, A cause-and-effect diagram derived from the planning process; A causal knowledge base representing the real world; This represents the number of causal edges in the causal knowledge base; when When the calculation result is greater than or equal to the second preset threshold, the corresponding plan is deemed qualified. The second preset threshold can be 0.9.

[0052] During the execution of the task planning and evaluation algorithm, metrics are monitored in real time. To perform quantitative evaluation of the execution process, and when When the deviation exceeds a preset threshold, a replanning process is triggered. The cumulative deviation during the planning process is used for analysis after the task planning is completed. Real-time monitoring metrics. The calculation formula is as follows:

[0053] In the formula, Represents time The actual state; Represents time the expected state; is a preset threshold for the normalized deviation. When the deviation between the actual state and the expected state exceeds the threshold, re-planning is triggered. The cumulative deviation can be used for post-analysis to understand the performance and deviation of the system over a period of time.

[0054] In this embodiment, the action execution module selects the optimal action matching the current environment state from the predefined action library through the state-action value mapping; executes the selected action and receives environment feedback; updates the state-action value mapping based on the feedback; and dynamically determines whether to continue or terminate the task execution process according to the updated environment state.

[0055] Specifically, reference can be made to Figure 4 For the detailed process of this part, the action execution module is an intelligent execution component based on reinforcement learning, which realizes efficient task execution and risk control through dynamic environment interaction and continuous optimization. Its execution process starts from the "start" phase, which first needs to build a state-action value mapping table (Q table) and a predefined action library.

[0056] In the process of building the Q table, the system will evaluate the value of taking various possible actions in different states based on historical data and experience, forming a comprehensive state-action value mapping. At the same time, the predefined action library contains a variety of pre-set action options, such as: collecting audio and video MCP, collecting MQTT IoT data MCP, file system access MCP, database access MCP, EHS system API access MCP, image understanding security risk, video understanding security risk, voiceprint understanding security risk, text understanding security risk, IoT data understanding security risk, device anomaly analysis MCP, encoding intelligent agent, sandbox running, etc.; These actions cover data collection, security risk analysis, device anomaly detection, dynamic coding, code execution, etc., providing the system with rich operation options.

[0057] Next, the system enters the "select the optimal action according to the plan" link, which relies on the initial state of the environment and the execution plan. The system will consider the current environmental conditions and task requirements to select the most suitable action from the predefined action library for execution. Subsequently, the system will execute the selected action and obtain rewards based on environmental feedback, which embodies the core idea of reinforcement learning - continuously optimizing decision-making strategies through trial and error and feedback.

[0058] After executing the action and obtaining the reward, the system will update the Q table, incorporating new experiences and knowledge into the state-action value mapping to guide subsequent decision-making. This process is dynamic, and as the environment changes and the task progresses, the Q table will be continuously adjusted and optimized to ensure that the system can adapt to complex and changing environmental conditions.

[0059] Finally, the system will determine whether to continue or terminate the process based on the current state of the environment. If the task is not yet completed or the environmental conditions change, the system will return to the "select the optimal action according to the plan" stage for a new round of decision-making and execution; if the task is completed or the predetermined goal is achieved, the process will enter the "end" stage, marking the successful completion of this task.

[0060] The action execution module emphasizes the importance of dynamic environment interaction, continuous optimization, and intelligent decision-making through this design. By constructing a Q-table and a predefined action library, combined with reinforcement learning algorithms, it achieves efficient task execution and risk control, and is suitable for automated management and decision support in a variety of complex scenarios.

[0061] In this embodiment, the action execution module obtains the reward by receiving environmental feedback. The reward The corresponding algorithm formula is the reinforcement learning reward function formula, as follows:

[0062] In the formula, , , , These represent the corresponding weights for each item, and the suggested weight values ​​are: , , , ; Represents a safety reward; Represents efficiency rewards; Represents resource rewards; This represents a penalty.

[0063] Safety Rewards The calculation formula is as follows:

[0064] In the formula, for ordinal numbers In this situation, Each corresponds to a different safety indicator. Its total number; Representing safety indicators The preset security threshold, Representative safety indicators Maximum allowed value; Representative safety indicators The corresponding weights; This represents the activation function used for a smooth transition. (In terms of total number) For example, then Representing the concentration of toxic gas, such as H2S≤10ppm; Representing the degree of deviation of the device state, such as the ratio of the real-time pressure value of the pressure tank to the safety pressure threshold value is less than or equal to 5%; Representing the number of personnel violations, which can be obtained through visual detection; the weights of the three safety indicators can be allocated in the manner of 0.4 / 0.3 / 0.3.

[0065] Efficiency reward The calculation formula is as follows:

[0066] In the formula, representing the standard disposal time; representing the actual disposal time; representing the task completion indication function.

[0067] Resource reward It is directly obtained by calculating the resource consumption value.

[0068] Penalty term The calculation formula is as follows:

[0069] In the formula, representing the corresponding weight coefficient of the th behavior or indicator, for example, the first penalty term is a missed report, the second is a false alarm, and the third is an unnecessary interruption of the experiment, and the respective corresponding weight coefficients are allocated in the manner of 10 / 5 / 3; representing the unit step indication function, which is used to determine whether the th behavior is in violation or the th indicator meets the requirements; if there is a violation or does not meet the requirements, then is 1, otherwise 0.

[0070] As a preferred embodiment of the present embodiment, the task requirements and execution results can be stored in the training data set in the method; if the execution result is successful, the knowledge base is updated to accumulate experience data and strengthen subsequent task planning; if the execution result is a failure, the failure reason is analyzed and the configuration of the task planning module, the action execution module or the environment perception module is dynamically adjusted according to the reason; and the system is self-adapted and optimized through the knowledge base update.

[0071] Specifically, the method corresponding to this part can form an independent reflection and learning module in the system, which is based on a feedback-driven system evolution process, which can be referred to in the schematic of Figure 5 , which aims to improve the task success rate and system performance through continuous task execution result analysis and system optimization. First, the task execution result needs to be evaluated, and appropriate measures are taken according to different situations of success or failure.

[0072] The system stores the task requirements and execution results in the training dataset for subsequent model training and system optimization. When the task execution result is successful, the task requirements and execution results are updated to the knowledge base to facilitate subsequent reinforcement reference methods. This process helps accumulate valuable experience data and supports the long-term development of the system.

[0073] When the task execution result is a failure, the system will enter the "evolution system" section, analyze the reasons for the failure in depth, and propose improvement plans. Possible reasons include planning errors, execution deviations, and environmental changes. According to different failure reasons, the system will take corresponding optimization measures.

[0074] If the failure reason is a planning error, the system will update the task planning module, including updating the large language model, prompt words, and updating the retrieval enhancement method, to improve the accuracy and effectiveness of task planning. If the failure reason is an execution deviation, the system will update the action execution module, which also includes updating the corresponding large language model, prompt words, and model calling tools, to improve the stability and reliability of task execution. If the failure reason is environmental changes, the system will update the perception data, including prompting the user to increase data acquisition sources, to adapt to the changing external environment.

[0075] After completing the corresponding optimization measures, the system will update the knowledge base to ensure that the latest knowledge and experience can be effectively utilized. Finally, the process enters the "end" stage, marking the completion of this task processing and system optimization. This part emphasizes the importance of feedback-driven and continuous optimization, and through the analysis of task execution results and the dynamic adjustment of system components, it realizes the self-evolution and performance improvement of the system, which is suitable for complex task scenarios that require high reliability and adaptive ability.

Claims

1. An industrial safety risk monitoring and early warning system for autonomous planning intelligent agents, characterized in that: It includes an environment perception layer, an edge computing layer, a cloud analytics layer, and an action execution layer; The environmental perception layer is used to collect multi-source sensing data in the industrial production environment in real time, including environmental parameters, personnel status, and equipment operating status information; The edge computing layer is used to perform preliminary processing on the multi-source sensing data, including data compression and key information extraction, to generate key analysis data. The cloud-based analysis layer is used to perform in-depth intelligent analysis of the key analysis data, perform security risk assessment and risk spread prediction, and dynamically generate and adjust early warning decisions through autonomous planning intelligent agents; The action execution layer is used to execute corresponding safety action instructions based on the warning decision, and feed the execution results back to the cloud analysis layer to achieve system optimization.

2. The industrial safety risk monitoring and early warning system according to claim 1, characterized in that: The environmental perception layer includes a sensor array for real-time acquisition of multi-source sensing data of the industrial production environment. The types of multi-source sensing data include video image data, environmental temperature and humidity and harmful gas concentration data, on-site sound data, and electrical equipment current and voltage parameter data.

3. The industrial safety risk monitoring and early warning system according to claim 1, characterized in that: The edge computing layer is used to reduce transmission load by performing data dimensionality reduction and compression processing, and to apply a dynamic target detection model to identify abnormal behavior in image data in real time, filter out key analysis data, and optimize data transmission efficiency.

4. The industrial safety risk monitoring and early warning system according to claim 1, characterized in that: The cloud-based analytics layer includes a multimodal analysis unit, a spatiotemporal fusion engine, and an autonomous planning agent. The multimodal analysis unit integrates image, sound, and text data for comprehensive risk assessment. The spatiotemporal fusion engine combines time and space dimensions to perform risk diffusion simulations and predict risk propagation paths and impact ranges. The autonomous planning agent dynamically generates early warning decision-making strategies based on historical case libraries and regulatory standard libraries, and adjusts data collection and analysis processes according to real-time feedback.

5. The industrial safety risk monitoring and early warning system according to claim 1, characterized in that: The action execution layer is used to transform early warning decisions into safety action instructions and implement them. At the same time, it evaluates the success or failure of the execution results, and feeds back the error analysis information to the cloud analysis layer when the evaluation fails to optimize subsequent decisions.

6. A method for monitoring and early warning of industrial safety risks of autonomous planning intelligent agents, characterized in that, The hardware basis of the method is the industrial safety risk monitoring and early warning system as described in claim 1; the method includes: Receive the task objective and the current environment state, and define the task decomposition rules; The environmental sensing module pre-deployed in the environmental sensing layer collects multi-source sensing data of the industrial production environment in real time. The multi-source sensing data is transmitted to a pre-deployed cognitive engine module in the edge computing layer for processing and analysis. The cognitive engine module determines whether new task planning is needed based on the analysis results. If it is determined that new task planning is needed, the pre-deployed task planning module in the cloud analysis layer is triggered, which generates a task tree based on the task objective, the current environment status, and preset task decomposition rules; if it is determined that no new task planning is needed, the pre-deployed action execution module in the action execution layer is triggered directly. After generating the task tree, the task planning module passes it to the action execution module. The action execution module converts the preset or regenerated task tree into action instructions and executes them, while monitoring the task progress in real time. The execution results are evaluated. If the evaluation is successful, the state model of the industrial production environment is updated. If the evaluation fails, error analysis is performed and the feedback information is passed to the task planning module for adjustment and optimization. Based on the evaluation results, the task planning and execution strategies are dynamically adjusted to form a closed-loop decision-making and execution process.

7. The industrial safety risk monitoring and early warning method according to claim 6, characterized in that: The task planning module, through its built-in task planning evaluation algorithm, generates a task tree based on the task objective, current environmental state, and preset task decomposition rules. The comprehensive evaluation function of the task planning evaluation algorithm assesses the quality of the final plan. The weighted evaluation function is as follows: In the formula, , , These represent the corresponding weight values; Represents the degree of goal achievement; Represents planning efficiency; Represents anti-interference capability; This represents causal consistency; This represents taking the minimum value; where, the degree of goal achievement It is used to measure the degree of matching between planning results and expected goals, and its calculation formula is as follows: In the formula, Representing the Individual goals; Represents the set of goals actually achieved; Represents the first The weight values ​​corresponding to each sub-objective; As an indicator function, it is 1 when the goal is achieved and 0 when it is not achieved; based on the degree of goal achievement... The calculation results are used to determine the success or failure of the task objective based on a preset threshold. Planning efficiency The formula used to quantify the cost-effectiveness of resource consumption versus time is as follows: In the formula, , , All are normalized coefficients; Represents the value of the task; It represents the sum of task execution time and task planning time, and the task planning time has a preset weight; Represents the weighted sum of system resources; for planning efficiency If the calculation result is less than a preset threshold, the task planning and evaluation algorithm will re-plan the task. Anti-interference capability Robustness entropy is used to assess the ability to cope with interference, and its calculation formula is as follows: In the formula, Representing the Task success rate under various interference scenarios; Represents the total number of interference scenario types; based on anti-interference capability The calculation results distinguish the strength of the system's anti-interference ability. When the system is fully robust, it has strong anti-interference capabilities; when At that time, the system was relatively fragile and had weak anti-interference capabilities; Causal consistency Used as a veto factor to verify the causal rationality of the planning logic, and when If the calculation result is less than the first preset threshold, the planning is directly deemed to have failed; the calculation formula is as follows: In the formula, A cause-and-effect diagram derived from the planning process; A causal knowledge base representing the real world; This represents the number of causal edges in the causal knowledge base; when When the calculation result is greater than or equal to the second preset threshold, the corresponding plan is judged to be qualified. During the execution of the task planning and evaluation algorithm, metrics are monitored in real time. To perform quantitative evaluation of the execution process, and when When the deviation exceeds a preset threshold, a replanning process is triggered. The cumulative deviation during the planning process is used for analysis after the task planning is completed.

8. The industrial safety risk monitoring and early warning method according to claim 6, characterized in that: The action execution module selects the optimal action that matches the current environment state from a predefined action library through state-action value mapping; executes the selected action and receives environmental feedback; updates the state-action value mapping based on the feedback; and dynamically determines whether the task execution process should continue or terminate based on the updated environment state.

9. The industrial safety risk monitoring and early warning method according to claim 8, characterized in that: The action execution module obtains rewards by receiving environmental feedback. The reward The corresponding algorithm formula is the reinforcement learning reward function formula, as follows: In the formula, , , , These represent the corresponding weights of each item; Represents a safety reward; Represents efficiency rewards; Represents resource rewards; Represents a penalty item; Safety Rewards The calculation formula is as follows: In the formula, for ordinal numbers In this situation, Each corresponds to a different safety indicator. Its total number; Representative safety indicators The preset security threshold, Representative safety indicators Maximum allowed value; Representative safety indicators The corresponding weights; This represents the activation function used for smooth transitions; Efficiency Rewards The calculation formula is as follows: In the formula, Standard processing time; Represents the actual processing time; This represents a function indicating task completion. Resource rewards Obtained directly by calculating resource consumption values; Penalty items The calculation formula is as follows: In the formula, Representing the The corresponding weight coefficient for each behavior or indicator; The unit step indicator function is used to determine the first step. Is the behavior in violation of regulations or the first Does each indicator meet the requirements? If there are violations or the requirements are not met, then... It is 1 if it is true, otherwise it is 0.

10. The industrial safety risk monitoring and early warning method according to claim 6, characterized in that: The task requirements and execution results are stored in the training dataset; if the execution result is successful, the knowledge base is updated to accumulate experience data and strengthen subsequent task planning; if the execution result is unsuccessful, the reasons for the failure are analyzed and the configuration of the task planning module, action execution module or environment perception module is dynamically adjusted according to the reasons; and the system adaptive optimization is achieved through knowledge base updates.

Images