A compliance checking and early warning method and device for high-risk operation processes

Abstract

The application discloses a kind of high-risk operation process compliance inspection and early warning method and equipment, it is related to industrial safety monitoring and intelligent early warning technical field, by constructing dynamic digital twin model for high-risk operation of thermal power plant, define operation step sequence, compliance rule set and dynamic safety boundary;Collect and fuse real-time multi-source data on site;Intelligent compliance reasoning and risk assessment module are used to carry out logical analysis and risk level judgment;Output graded early warning signal and simulate risk propagation boundary;Finally, start disposal process and continuously optimize system based on feedback. Real-time compliance inspection, dynamic risk assessment, accurate graded early warning and intelligent closed-loop optimization of high-risk operation whole process are realized, and the intelligent level and response efficiency of operation safety management are improved.

Description

Technical Field

[0001] This invention belongs to the field of industrial safety monitoring and intelligent early warning technology, specifically relating to a method and equipment for compliance inspection and early warning of high-risk operation processes. Background Technology

[0002] In high-risk industries such as thermal power generation, compliance supervision and risk control of high-risk work processes (such as working at heights, hot work, scaffolding work, and confined space work) are crucial for ensuring production safety. Traditional management models mainly rely on paper operation tickets / work permits, manual inspections, and post-event traceability, which suffers from problems such as regulatory lag, human negligence, and inconsistent rule enforcement. With the development of the Industrial Internet and sensor technology, although work monitoring systems based on video surveillance or single data sources have emerged, they typically lack the ability to dynamically map and deeply analyze the entire work process chain, making it difficult to achieve closed-loop management of pre-event warning, in-event intervention, and post-event optimization.

[0003] Existing technologies struggle to effectively integrate multi-source heterogeneous data (such as procedural texts, historical tickets, 3D models, and real-time sensor information), and are unable to construct digital regulatory models that are synchronized with physical operations in real time and whose rules can be dynamically adjusted. Furthermore, most systems can only achieve static rule judgments or simple threshold alarms, lacking the ability to dynamically simulate and visualize risk propagation paths, resulting in insufficient early warning accuracy and limited support for response decisions.

[0004] Therefore, there is an urgent need for a comprehensive system that can integrate digital twins, multimodal perception, intelligent reasoning and dynamic simulation to achieve full-process, real-time, adaptive compliance checks and risk warnings for high-risk operations, thereby improving the level of intelligence and proactive defense capabilities of safety management.

[0005] The following problems exist in the existing technology: Traditional high-risk work process safety supervision relies on human experience and static procedures, which makes it difficult to reflect the real risk status of the work site in real time and dynamically, resulting in delayed risk identification and untimely early warning; Existing safety monitoring systems mostly use isolated sensor data and lack the ability to fuse and collaboratively analyze multi-source heterogeneous data, resulting in one-sided risk assessment and a high false alarm rate; Traditional compliance checks are usually based on simple matching of fixed rules, lacking the ability to reason about complex and related risks, and unable to dynamically and progressively assess the overall risk of the work process; Traditional early warning systems typically only provide simple level alarms and lack the ability to visualize and simulate the propagation process of risk space and dynamically generate early warning boundaries, which is not conducive to precise intervention by command personnel. The existing system lacks a self-optimization mechanism based on actual handling feedback. Model and rule updates rely on manual intervention, making it difficult to adapt to dynamic changes in field conditions and the emergence of new risk patterns. Summary of the Invention

[0006] The purpose of this invention is to provide a method and device for compliance inspection and early warning of high-risk work processes, so as to solve the technical problems of poor accuracy and high false alarm rate of existing high-risk work process early warning methods.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A compliance check and early warning method for high-risk work processes includes the following steps: For each high-risk operation, a corresponding digital twin model is established. The digital twin model defines the sequence of operation steps for each high-risk operation, the set of compliance rules for each operation step, and the associated dynamic safety boundary. Based on the current work steps and associated dynamic safety boundaries defined by the digital twin model, real-time multi-source data from high-risk work sites are collected and fused in a targeted manner. Based on the compliance rule set defined by the digital twin model, the real-time multi-source data is subjected to correlation logic analysis and compliance judgment to generate a comprehensive risk assessment result for the current high-risk operation process; Based on the comprehensive risk assessment results, and by invoking the digital twin model, a tiered early warning signal is generated.

[0008] Furthermore, a risk management process is initiated based on the aforementioned tiered early warning signals, and information on the accuracy verification of early warnings and the effectiveness of management measures is collected during the risk management process to continuously optimize the digital twin model.

[0009] Furthermore, the optimization process of the digital twin model includes: Based on the graded early warning signals, a risk management process is initiated for high-risk operational processes. After the risk management process is completed, the comparison results between the graded early warning signals and the actual occurrence of the risk are collected as verification information for the accuracy of the early warning, and the management process records are collected as information on the effectiveness of the management measures. A Bayesian optimization method is used to probabilistically update the compliance rule set of the digital twin model and the physical parameters of the dynamic simulation using the collected early warning accuracy verification information and the effectiveness information of the response measures. A multi-agent reinforcement learning approach is adopted, in which compliance status and graded early warning signals are used as agent actions, and the effectiveness information of the disposal measures is used as environmental feedback. The dynamic Bayesian network, the mapping rules between early warning level and risk level, and the numerical model are updated through policy gradient algorithm. The optimized compliance rule set, dynamically simulated physical parameters, dynamic Bayesian network, mapping rules between warning levels and risk levels, and numerical model are deployed into the digital twin model. The complete process record and results of this risk handling process are used as historical data to complete the optimization loop.

[0010] Furthermore, the process of establishing the digital twin model includes: Obtain multi-source heterogeneous data related to high-risk operations from a pre-set local database; Based on multi-source heterogeneous data, the sequence of high-risk operation steps is parsed, and the safety entities, behavioral norms and constraints are identified and extracted in a structured manner. Based on the sequence of high-risk operation steps, the safety entities, behavioral norms and constraints are associated and integrated to construct an initial safety knowledge graph. For each high-risk operation, a digital twin model is generated by integrating the initial safety knowledge graph.

[0011] Furthermore, the real-time multi-source data acquisition and fusion process is as follows: Based on the compliance rule set and dynamic safety boundaries defined by the current work step in the execution state in the digital twin model, driven by high-risk work process data, generate awareness task instructions; Based on the spatial monitoring range and the required data types of sensing in the sensing task instructions, a subset of multi-source sensing devices with matching locations and functions is selected and activated from the sensor network deployed at high-risk work sites. Spatiotemporal alignment processing is performed on real-time multi-source data collected by a subset of multi-source sensing devices, and feature extraction and fusion are performed on the spatiotemporally aligned real-time multi-source data.

[0012] Furthermore, the process for generating the comprehensive risk assessment results of the high-risk operation process is as follows: Calculate the probability of violation for each rule in the set of compliance rules defined by the current job step that is in execution state; Based on the violation condition probability of each rule in the current operation step, the posterior probability distribution of the current operation step at different risk levels is calculated.

[0013] Furthermore, based on the risk type, level, and spatial location information identified by the comprehensive risk assessment results, and by calling the digital twin model, a graded early warning signal corresponding to the risk level is generated according to the preset mapping rules between early warning levels and risk levels.

[0014] Secondly, this invention provides a compliance check and early warning system for high-risk work processes, including a model definition module, a data acquisition module, a risk assessment module, and an early warning module, wherein: Model definition module: used to build a corresponding digital twin model for each high-risk operation. The digital twin model defines the sequence of operation steps for each high-risk operation, the set of compliance rules for each operation step, and the associated dynamic safety boundary. Data acquisition module: used to collect and fuse real-time multi-source data from high-risk work sites in a targeted manner, based on the current work steps and associated dynamic safety boundaries defined by the digital twin model; Risk assessment module: Based on the compliance rule set defined by the digital twin model, it performs correlation logic analysis and compliance judgment on the real-time multi-source data to generate a comprehensive risk assessment result for the current high-risk operation process; Early warning module: Used to generate graded early warning signals based on the comprehensive risk assessment results and by calling the digital twin model.

[0015] Thirdly, a terminal device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0016] Fourthly, a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0017] Compared with the prior art, the present invention has the following beneficial technical effects: This invention provides a compliance inspection and early warning method for high-risk work processes. It constructs a dynamic digital twin model for high-risk operations in thermal power plants, defining the sequence of work steps, a set of compliance rules, and dynamic safety boundaries. It collects and integrates real-time multi-source data from the site; utilizes intelligent compliance reasoning and risk assessment modules for logical analysis and risk level judgment; outputs tiered early warning signals and simulates risk propagation boundaries; and finally initiates the handling process and continuously optimizes the system based on feedback. This method achieves real-time compliance inspection, dynamic risk assessment, precise tiered early warning, and intelligent closed-loop optimization throughout the entire high-risk work process, improving the intelligence level and response efficiency of work safety management. It solves the problems of low efficiency in manual inspection, untimely early warning, reliance on experience for risk assessment, and lack of closed-loop optimization in risk handling in traditional high-risk work processes.

[0018] Preferably, by constructing a dynamic digital twin model for each high-risk operation, the operation process, compliance rules and safety boundaries are digitally and structurally expressed. The status of the twin model is dynamically updated based on real-time multi-source data, realizing a comprehensive and real-time mapping of the operation process, thereby significantly improving the real-time nature and accuracy of risk perception.

[0019] Preferably, based on the perception requirements defined by the digital twin model, multi-source perception device data is activated and fused in a targeted manner, and spatiotemporal alignment and feature fusion are performed to form a unified multimodal joint feature vector. This achieves precise scheduling of monitoring resources and high-quality data fusion, providing comprehensive and consistent input for subsequent intelligent analysis and improving the reliability and accuracy of compliance judgment. Rule-based conditional probability calculation and Bayesian network fusion technology are adopted to achieve hierarchical risk assessment from single steps to the overall process.

[0020] Preferably, the system models the entire workflow as a dynamic Bayesian network, enabling online reasoning about the evolution of risk states and generating a comprehensive assessment result that includes instantaneous risk and future risk exposure, thus achieving intelligent and dynamic risk assessment. While generating tiered early warning signals, the system dynamically simulates propagable risks using a physical field model and generates visualized dynamic early warning boundaries based on a digital twin environment. This function enables visualized deduction of the risk propagation process and precise definition of the early warning area, greatly improving the intuitiveness of the early warning and the targeted nature of decision support.

[0021] Preferably, after the early warning and handling loop is closed, the system automatically collects information on the accuracy of the early warning and the effectiveness of the handling, and uses methods such as Bayesian optimization and multi-agent reinforcement learning to automatically and probabilistically update the rules, parameters and inference network of the digital twin model. This mechanism enables the system to continuously learn and optimize itself, allowing the system to continuously accumulate experience, adapt to changes, and improve its long-term adaptability and early warning effectiveness. Attached Figure Description

[0022] Figure 1 This is a flowchart of a compliance check and early warning method for a high-risk work process according to an embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0025] It should be noted that the terms "first" and "second" in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] The present invention will now be described in further detail with reference to the accompanying drawings: like Figure 1 As shown, a compliance check and early warning method for high-risk work processes includes the following steps: Step 1: Establish a corresponding digital twin model for each high-risk operation. The digital twin model defines the sequence of operation steps for each high-risk operation, the set of compliance rules for each operation step, and the associated dynamic safety boundary. First, a dynamic and instantiable digital twin model is established for each high-risk operation in a thermal power plant. This digital twin model not only defines the standard sequence of steps for the operation, but also sets specific compliance rule sets and dynamic safety boundaries for each step, forming a virtual mapping and safety benchmark for operation execution.

[0027] Step one specifically includes: To establish and maintain a dynamic digital twin model for each high-risk operation, the following steps are involved: Obtain multi-source heterogeneous data related to high-risk operations in thermal power plants from a pre-set local database; Multi-source heterogeneous data includes at least the following: a library of operation tickets and work tickets instances containing standardized operating procedure documents, electrical operation ticket history records, and thermal machinery work ticket history records; a library of safety procedures and cases containing industry safety procedure documents, equipment maintenance procedure documents, and historical safety accident case records; and an asset database containing at least the 3D model of equipment and plant area spatial topology information of thermal power plants. From the database of operation tickets and work tickets, high-risk operation step sequences are parsed out; from the database of safety procedures and cases, safety entities, behavioral norms, and constraints are identified and extracted in a structured manner using natural language processing technology; based on the high-risk operation step sequences, safety entities, behavioral norms, and constraints are associated and integrated to construct an initial safety knowledge graph. For each high-risk operation, a digital twin model framework is generated by integrating key information from the initial safety knowledge graph with spatially relevant information from the asset database. This framework includes a virtual 3D environment model that maps to the physical power plant space and a unified spatiotemporal reference as the data fusion benchmark. During the execution of high-risk operations, the digital twin model is dynamically instantiated and its status is maintained using operation instructions and progress data received from the power plant's production management system as operation process data, and on-site perception data collected from the power plant site.

[0028] Specifically, this system pre-sets a local database for centralized storage and management of multi-source heterogeneous data related to high-risk operations in thermal power plants. The multi-source heterogeneous data mainly includes three categories: First, the operation ticket and work ticket instance library, which comes from the thermal power plant production management system and historical electronic archive system. The content includes at least standardized operation procedure documents (such as the "Standard Procedure for Electrical Switching Operation"), historical electrical operation ticket records, and historical thermal mechanical work ticket records. Each record contains information such as operation name, operation step sequence, operator, operation time, equipment number, and safety measures. Second, the safety procedure and case library, which comes from industry safety standard documents (such as the "Electric Power Safety Work Procedure"), internal equipment maintenance procedures, and historical safety accident report archives. The content includes at least industry safety procedures and equipment maintenance procedure documents, as well as historical safety accident case records (including accident description, cause analysis, responsible links, and rectification measures). Third, the asset database, which comes from the thermal power plant asset management system, 3D design drawings (such as BIM models), and plant area GIS system. The content includes at least the 3D geometric models of all equipment in the plant (such as STL or IFC format models of boilers and steam turbines), the physical attributes of the equipment (such as rated parameters), and the spatial topology information of the plant area (such as the connection relationship between equipment, workshop layout, and passage coordinates).

[0029] The system employs a text parsing method combining rule-based analysis and machine learning to parse the content of operation tickets and work tickets instance libraries. A named entity recognition model is used to identify key entities, such as equipment names (e.g., #1 main transformer). Sequence labeling models (e.g., BiLSTM-CRF) are used to segment operation tickets, identifying the step boundaries and sequence of high-risk operations, ultimately extracting a high-risk operation step sequence. For example, for the "110kV line power outage maintenance" operation, the parsed step sequence is: Step 1: Verify equipment name and number; Step 2: Disconnect the circuit breaker; Step 3: Verify that the equipment has no voltage, etc. Natural Language Processing (NLP) technology is used to process the content of safety regulations and case libraries. Term segmentation, part-of-speech tagging, and dependency parsing are performed on the regulations and case reports. A relation extraction model is used to identify and structurally extract safety entities (e.g., workers), behavioral norms (e.g., mandatory wearing of safety helmets), and constraints (e.g., safe distance greater than 1 meter). These are stored in the form of triples (entity, relation, entity / value), such as (worker, mandatory wearing, safety helmet).

[0030] Using the extracted high-risk work step sequence as the main framework, the extracted safety entities, behavioral norms, and constraints are linked and attached to the corresponding work step nodes through entity alignment and relational reasoning. For example, "Step 3: Verify that the equipment has no voltage" will be associated with the safety entity "voltage tester", the behavioral norm "voltage must be tested on each side of the equipment", and the constraint "voltage level of the voltage tester is matched and the test is valid". Finally, an initial safety knowledge graph centered on the high-risk work process is constructed. This graph is stored in a graph structure, with node types including work steps and safety entities, and edge types including associated constraints and dependent equipment.

[0031] For each specific high-risk operation (such as "#2 boiler superheater tube rupture repair"), an instantiable digital twin model framework is generated by integrating its corresponding initial safety knowledge graph and spatial information in the asset database (such as the 3D model of the ruptured boiler, the layout of surrounding equipment, etc.). This framework is a software template, and its core components include: a virtual 3D environment model that is constructed based on the equipment 3D model and the plant spatial topology information and mapped to the physical thermal power plant space; a unified spatiotemporal benchmark for all data fusion and calculation; and a predefined parameter interface for receiving external data to drive model state updates.

[0032] During the actual execution of high-risk operations, this digital twin model framework is dynamically activated. Through data interfaces, it receives real-time instructions and progress reports from the power plant's production management system (such as a work permit management system) as operational process data, and receives on-site sensing data (such as temperature and gas concentration) from sensors and other equipment deployed at the high-risk work site. The system uses the received operational process data to drive the virtual operational process within the digital twin model, synchronizing the operational progress in the virtual environment with the physical world. Simultaneously, it uses on-site sensing data to update the real-time status of entities and the environment in the virtual 3D environment model; for example, it updates the real-time location of personnel in the virtual environment based on UWB positioning data. Through continuous comparison and updates, it ensures a consistent dynamic mapping between the digital twin model and its physical entities, forming a digital twin model synchronized with the physical operational process.

[0033] The digital twin model defines the sequence of work steps for this high-risk operational process, the set of compliance rules for each work step, and the associated dynamic safety boundaries, including the following steps: The high-risk operation step sequence extracted from the initial safety knowledge graph is directly defined as the operation step sequence of the high-risk operation process mapped by the digital twin model, and a unique identifier and context information are configured for each operation step. The behavioral norms and constraints extracted from the initial safety knowledge graph are transformed into a set of condition judgment logics corresponding to each work step, forming a set of compliance rules for that work step; the multi-dimensional condition judgment logic integrates the status parameters of personnel, equipment, tools and environment related to the safety of high-risk operations; Define a dynamic safety boundary for each operation step. The spatial extent of this boundary is a dynamically changing three-dimensional region, and the core dimension of this region is the dynamic safety radius. The calculation formula is:

[0034] in, The baseline safety radius is determined based on the current work steps and equipment location. These are the risk factor values ​​for G environmental and equipment statuses that are sensed in real time. and These are the safety threshold and critical threshold for the corresponding risk factors. The normalized weighting coefficients for each risk factor. The sensitivity coefficient is dynamically adjustable; the spatial range of the dynamic safety boundary varies with... The changes are adjusted in real time.

[0035] Specifically, the high-risk operation step sequence extracted from the initial safety knowledge graph will be directly defined as the formal step sequence of the physical operation process mapped by the digital twin model. Each operation step in the sequence will be configured with a unique identifier (such as "Operation ID_Step Sequence Number", e.g., "OP2024001_Step03") and context information. The context information will include at least: operation step name, operation step type (e.g., preparation, operation), expected execution time range (based on historical data statistics, e.g., [5,15] minutes), a list of directly associated physical equipment, a list of necessary tools and equipment, the role or job allowed to perform the operation (e.g., certified electrical operator), and the predecessor and successor step identifiers IDs that clearly define the process logic.

[0036] The compliance rule set is the core logic set used by the digital twin model to automatically determine whether each work step is compliant. Its construction is a transformation process from natural language specifications to executable logic. It iterates through all behavioral norms and constraint triples associated with specific work steps in the initial safety knowledge graph. The rule engine parser transforms each natural language description triple into formalized conditional judgment logic. The rule transformation process follows multiple patterns. For constraints involving thresholds, they are directly converted into logical expressions. For example, "ambient temperature below 50℃" is converted into the logical expression: IF temperature_sensor_value<50 THEN Compliant ELSE Violation. For constraints involving the status of equipment, personnel, or tools, they are converted into state machine queries. For example, "circuit breaker must be in the open state" is converted into: IF circuit_breaker.status=="OPEN" THEN Compliant ELSE Violation. For combinations of multiple constraints, they are converted into logical operations. For constraints involving spatial or temporal relationships, location or timing judgments are introduced. The condition judgment logic for each rule is designed to receive and integrate multi-dimensional real-time status parameters. These parameters are organized into a status vector, derived from personnel status, equipment status, tool status, and environmental status. An independent compliance rule set object is generated for each work step, containing a rule list, rule priorities (some rules are graded by importance), and an evidence mapping table specifying the data source required to verify each rule (corresponding to a specific sensor ID or data stream).

[0037] Dynamic safety boundaries are used to define real-time hazardous areas or safety buffer zones for work procedures. Their spatial extent is a dynamically changing three-dimensional region, the core dimension of which is the dynamic safety radius. The calculation formula is dynamically adjusted based on real-time risk factors: ; in, The baseline safety radius is dynamically determined based on the current work step type and the location of associated equipment, taking into account industry safety regulations, the 3D model of the equipment in the asset database, the equipment body dimensions, and the operating space requirements. These are G risk factor values ​​related to environmental and equipment status, acquired in real-time at high-risk work sites using sensors and other equipment. Examples include oxygen concentration and combustible gas concentration in confined space hot work. Safety thresholds are also considered. These are safety limits specified in industry regulations, such as the lower safe limit for oxygen concentration in confined space operations being 19.5% (volume fraction); critical threshold. The threshold is taken from an immediate threat to life and health (IDLH) or an immediate cause of an accident; for example, the critical lower limit of oxygen concentration might be set at 18.0%. Normalized weighting coefficients. The value was determined using the Analytic Hierarchy Process (AHP) combined with statistical analysis of historical accident data. It ranges from 0 to 1 and satisfies the following conditions: ; The sensitivity coefficient is dynamically adjustable and its value is usually set between 0.5 and 2.0. The initial value is set according to the inherent risk level of the operation (such as level 1 hot work, level 2 hot work). For example, a larger value (1.5) is used for high-risk operations.

[0038] During the execution of the work steps, the dynamic safety radius is calculated cyclically at a fixed frequency (e.g., 1Hz). At the same time, it ensured And when any risk factor approaches the critical threshold, This will significantly increase and expand the scope of security alerts. In the virtual 3D environment model of the digital twin, the spatial coordinates of the main risk factors are used as the geometric center, and the real-time calculated... For a radius, generate or update a 3D geometry (such as a sphere or cylinder) as a visual representation of a dynamic safety boundary, which is typically rendered as a semi-transparent red (high risk) to yellow (low risk) gradient spherical shell, and updates accordingly. It scales dynamically according to changes. When any risk factor exceeds... In such cases, the highest level of warning will be triggered immediately, and the emergency response procedure will be initiated.

[0039] Step 2: Based on the current work steps and associated dynamic safety boundaries defined by the digital twin model, collect and integrate real-time multi-source data from high-risk work sites. Based on the current steps and safety boundaries indicated by the digital twin model, multi-source sensing devices (such as cameras, infrared sensors, gas sensors and positioning devices) deployed at the work site are activated and coordinated to collect real-time multi-source data, and spatiotemporal alignment and feature fusion are performed to generate a multimodal joint feature vector that can reflect the on-site status.

[0040] The real-time multi-source data acquisition and fusion process is as follows: Based on the compliance rule set and dynamic safety boundary defined by the current work step in the execution state in the digital twin model driven by high-risk work process data, a perception task instruction is generated. The perception task instruction clarifies the specific target to be monitored, the type of sensor data required to verify the compliance of the work step, and the spatial monitoring range based on the dynamic safety boundary associated with the work step. Based on the spatial monitoring range and required sensor data types in the sensing task instructions, select and activate a subset of multi-source sensing devices with matching locations and functions from the sensor network deployed at high-risk work sites; the subset of multi-source sensing devices includes at least visible light camera devices, infrared thermal imaging devices, environmental gas sensing devices, and UWB positioning devices that cover the spatial monitoring range. Spatiotemporal alignment processing is performed on real-time multi-source data collected by a subset of multi-source sensing devices, and the timestamps and spatial coordinates of all data are uniformly synchronized to the virtual three-dimensional environment model and unified time reference associated with the digital twin model. Feature extraction and fusion are performed on the spatiotemporally aligned real-time multi-source data to generate a multimodal joint feature vector that represents the on-site compliance status of the current operation step.

[0041] Specifically, when a high-risk work process is underway, and the currently active step in the digital twin model is updated (e.g., switching from step 3 to step 4), a new perception task instruction is triggered. The compliance rule set of the current work step is analyzed, and the direct evidence and monitoring targets required for the verification of each rule are analyzed one by one. For example, the rule "oxygen concentration must be greater than 19.5%" directly corresponds to "ambient gas sensing equipment (oxygen sensor)". The dynamic safety boundary associated with this work step is determined through calculation. The values ​​and their geometric center coordinates (usually the location of the main hazardous equipment) are used to obtain a spatial monitoring range. Finally, a structured sensing task instruction is generated, which explicitly includes the specific target to be monitored (such as specific equipment or personnel), the data type of sensor required to verify compliance (such as gas concentration, location coordinates, etc.), and the spatial monitoring range based on dynamic safety boundaries. It also includes task priority (usually instructions for high-risk steps have the highest priority) and control information such as validity period (covering the expected execution time of the current step).

[0042] The system maintains a field sensor network registry covering high-risk work areas, recording the static information and real-time status of all sensing devices within these areas of the thermal power plant. Based on the generated sensing task instructions, a constraint-optimized device selection algorithm is run. This algorithm uses the spatial monitoring range and sensor data type requirements specified in the sensing task instructions as core constraints, with coverage integrity and resource consumption as optimization objectives. First, spatial and functional matching is performed, filtering from the registry all online devices whose detection range intersects with the spatial monitoring range in the instructions and whose device type matches the required sensor data type, forming an initial candidate device set. For example, online oxygen sensors that can cover the target area are selected. Subsequently, from the initial candidate device set, an optimal subset of multi-source sensing devices is selected based on principles such as minimizing coverage redundancy, optimizing data quality, and load balancing. This subset must contain at least several types of devices required by the sensing task instructions, including visible light cameras, infrared thermal imaging devices, environmental gas sensors, and UWB positioning devices. The system uses the industrial IoT network in the factory area to send specific control commands to each device in the selected subset of devices (for example, for sensor devices, the commands include setting the sampling frequency, range, trigger mode, etc.). After receiving the commands, the devices adjust their own parameters and collect data, and upload the raw data stream to the system in real time via the high-speed network.

[0043] Raw data from different devices with varying spatiotemporal references are transformed into a spatiotemporally unified, semantically correlated multimodal joint feature vector. First, spatiotemporal alignment is performed using a precise time protocol to synchronize the timestamps of all devices to a unified time reference. Spatially, using pre-calibrated installation positions and orientation parameters of each sensor, various data types are mapped onto a virtual 3D environment model within the digital twin model. Next, a parallel feature extractor is run for each data source to extract features. These feature extractors are typically based on pre-trained deep learning models or traditional image / signal processing algorithms. For example, object detection models are used to identify personnel, safety equipment, and equipment status in visible light video streams; statistical features of the temperature matrix are extracted from infrared thermal image data; filtered concentration values ​​are read from environmental gas data; and real-time coordinates and velocity are calculated from positioning data. Finally, to make compliance judgments, the system prepares corresponding data for each compliance rule in the current operation step and maintains a query list. This list clearly records "which rule requires which feature items". For example, the rule "whether the safety distance is met" corresponds to the two feature items "the location of personnel A" and "the location of equipment B". Based on this query list, all feature items used by the rule are collected and arranged in a pre-set fixed order (for example, first sort all location coordinates, then sort various sensor values, and finally sort the image recognition results) to form a one-dimensional array. This array is the multimodal joint feature vector.

[0044] Step 3: Based on the compliance rule set defined by the digital twin model, perform correlation logic analysis and compliance judgment on real-time multi-source data to generate a comprehensive risk assessment result for the current high-risk operation process; Based on the predefined set of compliance rules in the digital twin model, correlation analysis and compliance judgment are performed on the fused real-time data. The compliance status and risk level of each operation step are calculated through a probabilistic model (such as a Bayesian network). Furthermore, the entire process steps are integrated through a dynamic Bayesian network to form a comprehensive risk assessment result that covers instantaneous risks, future risk exposures, and a specific risk list.

[0045] Based on the compliance rule set defined by the digital twin model, correlation logic analysis and compliance judgment are performed on real-time multi-source data to generate the compliance status of each operation step, including the following steps: Based on the compliance rule set defined by the current work step in the execution state within the digital twin model driven by high-risk work process data, each rule in the compliance rule set is extracted; the compliance rule set contains J rules, and for the j-th rule... , Extract rule-related evidence from the multimodal joint feature vector to form an evidence vector. ; through this rule The corresponding likelihood function Calculate the probability of violation conditions ; Based on the violation conditional probabilities calculated from all J rules in the current work step, a Bayesian network is used to fuse them, and the posterior probability distribution of the current work step at different risk levels k is calculated. Risk level k is the internal evaluation level used to generate graded early warning signals. This posterior probability distribution represents the compliance status of the work step. : ,in, Let k be the prior probability of risk level k. Let be the likelihood probability under a risk level of k.

[0046] Specifically, the current execution steps are obtained from the digital twin model. The defined compliance rule set contains J formalized rules, denoted as... Each rule Essentially, it is a logical judgment function, for example: Temperature ≤ Simultaneously, the system receives multimodal joint feature vectors.

[0047] The system maintains a rule-evidence mapping table, which predefines each rule. The specific evidence items required for verification and their index positions in the multimodal joint feature vector; for example, rules. For region A, where the oxygen concentration is ≥19.5%, the corresponding evidence is the floating-point value named `oxygen_concentration_zone_A` in the multimodal joint feature vector. Based on the rule-evidence mapping table, extract the value corresponding to the rule from the multimodal joint feature vector. All relevant evidence values ​​are organized into an evidence vector. At the same time, the evidence values ​​are subjected to necessary normalization or standardization preprocessing to make them fall into a uniform numerical range (such as [0,1]).

[0048] For each rule Associate with a predefined likelihood function This function will This is mapped to a violation conditional probability between 0 and 1, i.e.: This value indicates that, given the rule and evidence vector Given the given conditions, the probability that the rule will be violated. Likelihood function. The form of the rule is designed based on the rule type and the nature of the evidence. For threshold-type rules involving numerical values ​​(single evidence), the Sigmoid function is usually used to make the probability transition smoothly around the threshold. Where T is the rule threshold. Steepness coefficient (sensitivity). The larger the value, the more drastic the probability change near the threshold. For Boolean rules (single evidence), where the evidence is a binary confidence level (such as the confidence level for object detection), the likelihood function can be designed as follows: For example, if the confidence level for detecting "wearing a safety helmet" is 0.95, then the probability of violation is 0.05. For complex logical rules (multiple evidence), first calculate the probability of violation for each sub-condition, and then fuse them according to logical relationships (AND, OR). For example, the rule "normal pressure AND valve closed"... ,in , These represent the probabilities of violating the two sub-conditions.

[0049] The entire operation step is modeled as a Naive Bayes classifier, which integrates the violation probabilities of all J rules to calculate the posterior probability distribution of the current operation step at different internal risk levels. The posterior probability distribution represents the compliance status (CS) of the current operation step. The system predefines K discrete risk levels, denoted as... For example, k=1 (low risk / compliance), k=2 (medium risk / caution), k=3 (high risk / violation).

[0050] Risk level prior probability This represents the base probability of this step at each risk level when there is no evidence, which can be obtained statistically from historical operation databases (e.g., P(low risk) = 0.7, P(medium risk) = 0.2, P(high risk) = 0.1); likelihood probability. This represents the observed evidence vector under a risk level of k. (or its derived probability of violation) The probability distribution of the likelihood probability. Due to the evidence vector or the probability of violation derived therefrom Since most values ​​are continuous, a parameterized probability distribution (such as the Beta distribution) is used to model them. The Beta distribution has two shape parameters. and By analyzing historical data, when the actual risk level of a work step is k, the rule... The number of times a violation or compliance is determined: , ,in, For historical data, the step risk level is k and Violated (i.e.) The number of instances with higher values, such as greater than 0.5. For a step with risk level k and a rule To be obeyed (i.e.) The number of instances with lower values, such as less than or equal to 0.5; The probability density function (PDF) calculated as a Beta distribution is given in... Values ​​at: Using probability density to approximate probability is a common method in Bayesian inference with continuous variables.

[0051] For the current step, the evidence vectors for all J rules have been obtained. And calculate the corresponding probability of violation. For each risk level k, calculate its posterior probability. (i.e., compliance status) ): The posterior probability calculations for all K risk levels are normalized to ensure their sum is 1. Finally, the compliance status CS of the current work step is represented as an S-dimensional probability vector. For example, CS=[0.85,0.12,0.03] indicates that the step has an 85% probability of being low-risk (compliant), a 12% probability of being medium-risk (requires attention), and a 3% probability of being high-risk (non-compliant).

[0052] Integrate the compliance status of all operational steps to generate a comprehensive risk assessment result for the current high-risk operational process, including the following steps: The entire high-risk operation process is modeled as a dynamic Bayesian network, where each node in the network corresponds to a work step; for the i-th work step node... , , This represents the total number of steps in a high-risk work process, with each step corresponding to a risk level. Dynamic Bayesian networks determine the probability of risk state transitions between work steps by learning from relevant historical data in the database of operation tickets and work tickets, as well as the database of safety procedures and cases. ; Setting initial steps The prior probability distribution of risk level is As high-risk work processes are executed, an online forward inference algorithm is performed on the dynamic Bayesian network to recursively update the risk state confidence of each work step; as the high-risk work process progresses to work step... At time t, for the work step Any possible risk level Its risk status confidence level The calculation formula is: ;in, Indicates the steps of the operation arrive The observed sequence of compliance states, For the work steps The risk level is Confidence level of risk status at that time In the work steps The risk level is Observed compliance status under the conditions The likelihood probability; Based on the risk confidence level and compliance status of each work step, the instantaneous overall risk value of the entire high-risk work process is calculated. and future risk exposure It generates a structured risk list that includes the type, level, spatial location of the identified risks, and an index of the source operation steps associated with the risks; the instantaneous overall risk value, future risk exposure, and the structured risk list together constitute a comprehensive risk assessment result.

[0053] Specifically, it will include The high-risk workflow of each task step is modeled as a dynamic Bayesian network (DBN), where each node in the network corresponds to a task step. Each task step node Corresponding to a risk level , (For example, 1 represents low risk, and K represents high risk). Dynamic Bayesian networks reflect the propagation of risk along the workflow, i.e., the steps in the workflow. The risk status depends not only on itself, but also on the previous work steps. The impact of a risk state, its conditional probability dependency, simplifies to: .

[0054] The core parameter of a dynamic Bayesian network is the state transition probability matrix between steps. For adjacent operation steps... arrive There exists an H×H dimensional state transition probability matrix. Its elements Define the steps from the job Risk level m is transferred to the work step The probability of risk level n, .

[0055] The state transition probability matrix is ​​learned from historical data using machine learning methods. The data comes from a library of operation tickets and work tickets instances and a library of safety procedures and cases. The operation ticket and work ticket instance library provides a large number of historical operation step sequence records, used to learn typical, accident-free risk transfer patterns between operation steps. The safety procedures and case library (especially historical safety accident case records) provides a path description of the gradual evolution of risk from the initial step to an accident, used to learn abnormal, high-risk risk transfer patterns between steps. The expectation-maximization algorithm is used for parameter learning. The algorithm's input is a compliance state sequence obtained by "replaying" historical data. That is, for each historical operation instance, based on its archived perception data or reconstructed scenario, the compliance reasoning process is rerun to calculate a compliance state CS for each step in that instance, thus forming an observation sequence. When the true risk level (latent variable) of historical steps cannot be directly observed, the algorithm uses these observation sequences to refine the state transition probability matrix. Perform iterative estimation until convergence; for example, for a three-level risk system (K=3), a learned transition matrix might look like this: The matrix shows that if the previous operation step is low risk (m=1), the current operation step has an 85% probability of remaining low risk, a 10% probability of escalating to medium risk, and a 5% probability of escalating to high risk.

[0056] During the actual execution of high-risk operations, online real-time reasoning is performed as the operation progresses to the next step. At that time, the system had already observed the changes from the work steps. arrive Compliance status sequence The forward algorithm is used to recursively calculate the risk state confidence at each step. Forward variables are defined. This indicates that, given the observed compliance status sequence of the first i operation steps, the risk level of the i-th operation step is... The joint probability. Likelihood probability. An H×H dimensional confusion matrix is ​​obtained by learning from historical data. To represent, where elements This represents the probability of observing a compliance status of q when the true risk level is p. It can be initialized as an identity matrix. The calculation process is as follows: For the first step (when i=1)... ,in, This is the prior probability of the initial risk level. Its value is obtained from a historical statistical database based on the type of work, or it can be directly set by technical personnel according to the inherent risk level of this work. For example, an extremely high-risk work can be set to [0.1, 0.3, 0.6], indicating that the probability of the first step being low, medium, and high risk is 10%, 30%, and 60%, respectively. For subsequent work steps... ( ): ,in, State transition probability matrix The corresponding element in; through the Normalization will yield the current task steps. Posterior probability distribution of risk level The subsequent a priori probability distribution represents the confidence level of the risk state of that operation step.

[0057] Instantaneous overall risk value Quantify the execution of the task up to the current time t (high-risk task process execution up to the task step) When a high-risk operation process is in its most immediate stage, the formula for calculating its overall risk level is as follows: ,in, Risk level The corresponding risk weights, for example, are defined. =0.1 (low risk) =0.5 (medium risk) =1.0 (high risk), this weight reflects the severity of different risk levels and can be set according to the severity of the consequences; It is a continuous value between [0,1], with higher values ​​indicating greater overall risk. Future risk exposure. Predicting from the current work steps The cumulative amount of potential risks that may be encountered from the start to the end of the task; the confidence level of the risk status based on the current task step during calculation. and state transition probability matrix sequence The system recursively predicts the probability distribution of risk levels for all subsequent operational steps and then sums up the expected risk values ​​of these predicted distributions.

[0058] Generate a structured list, i.e., a structured risk list; each risk entry in the list should have a specific identified risk, including a unique risk identifier, risk type, risk level, spatial location of the risk occurrence (e.g., if it is equipment temperature exceeding the limit, the location is the three-dimensional coordinates of the equipment), and an index of the source operation step (determined by tracing back high-probability risk transfer paths in the DBN, for example, if the current operation step...). The high risk has a high confidence level, and the transition probability indicates that it largely stems from... The risk originates from the following steps: The system includes the associated compliance rule identifier (pointing to the specific rule being violated), risk confidence level, and timestamp. Based on the compliance rule content associated with the risk item (such as rules involving the environment, personnel, equipment, or processes), the system directly determines the risk type (e.g., environmental parameters exceeding limits, personnel violations, abnormal equipment status, or disordered process sequence), and further specifies the type by combining evidentiary data and a digital twin model.

[0059] Ultimately, the instantaneous overall risk value will be... Future risk exposure The structured risk list is encapsulated into a comprehensive data object, which serves as the output of the comprehensive risk assessment results for the current high-risk operational processes.

[0060] Step four: Based on the comprehensive risk assessment results and by calling the digital twin model, generate tiered early warning signals.

[0061] Based on the assessment results, graded early warning signals are generated according to risk type and level. Numerical simulation and visualization of the transmissible risks are performed to dynamically generate and display the early warning boundaries, thereby realizing real-time visual early warning and situational awareness of risks.

[0062] Based on the risk type, level, and spatial location information identified by the comprehensive risk assessment results, and by calling the digital twin model, a graded early warning signal corresponding to the risk level is generated according to the preset mapping rules between early warning level and risk level.

[0063] If the identified risk is a transmissible risk, then a corresponding numerical model is selected from a pre-set physical field model library for dynamic simulation based on the risk type. The spatial topology, obstacle information, and real-time multi-source data of the virtual three-dimensional environment model of the digital twin model are used as the initial and boundary conditions of the numerical model, and the control equations of the numerical model are numerically solved by an adaptive mesh refinement method. Based on the dynamic risk field obtained by numerically solving the governing equations, and according to predefined multi-level safety thresholds corresponding to different warning levels... Extracting isosurfaces Defined as dynamic early warning boundary: ,in, This is a simulated area defined to address the current risk transmission scenario. This represents the risk quantification index value at spatial location x and time t. Dynamic early warning boundaries It is spatiotemporally correlated with the graded early warning signals and superimposed onto the virtual three-dimensional environment model of the digital twin model for real-time visualization and situational simulation.

[0064] Specifically, a pre-defined early warning system with L levels is established. For example, L=4 levels are defined: l=1 (Normal / Blue), l=2 (Attention / Yellow), l=3 (Warning / Orange), and l=4 (Alarm / Red), corresponding to the response process from observation to emergency evacuation. An early warning mapping rule base is maintained, which contains mapping rules based on "if-then" logic. These rules map the type, level, and confidence level of risks identified by the comprehensive risk assessment results to specific early warning levels. , For example, a rule could be: if the risk type is flammable gas leak, the risk level is high, and the risk status confidence is greater than 0.6, then the warning level is l=3 (warning). The structured risk list in the comprehensive risk assessment results is traversed, and the warning mapping rule base is queried for each risk item. The warning level is determined based on the matching rule. It generates a corresponding graded early warning signal data object, which includes a unique early warning identifier, the associated risk list item ID, the early warning level, the early warning trigger timestamp, the risk spatial location (three-dimensional coordinates), a natural language early warning description (e.g., methane concentration exceeding the standard was detected at the flange of pipeline No. 3 in Area A, posing an explosion risk, please evacuate immediately!), and suggested disposal measures (e.g., start ventilation and evacuate personnel within a 50-meter radius).

[0065] Not all risks require dynamic simulation. The system has a built-in whitelist of transmissible risk types defined based on physicochemical principles. Typical transmissible risks include the spread of toxic / flammable gases (such as H2S and CH4), fire heat radiation and smoke spread, pressure wave propagation (such as steam or explosion shock waves), and liquid leakage spread (such as fuel oil). When the risk type corresponding to the generated warning signal exists in the whitelist, it is determined to be a "transmissible risk" and the dynamic simulation process is triggered.

[0066] Based on the risk type keywords of the warning signal, a corresponding numerical model is matched and selected from a pre-set physical field numerical model library for dynamic simulation. This numerical model contains the governing equations (usually a set of partial differential equations, such as convection-diffusion equations) describing the transport and diffusion processes of hazardous substances (such as gases). For example, for the risk type "methane gas leak", a "neutral / light gas diffusion model" (such as a Gaussian plume model) is selected. Subsequently, the virtual three-dimensional environment model of the digital twin model is invoked to obtain the spatial topology of the current work scene (the three-dimensional geometric model of the plant area and buildings, serving as the boundary and internal obstacles of the simulation area), obstacle information (such as the location, size, and material properties of equipment and walls), and real-time multi-source environmental data (such as wind speed and temperature). The location and intensity parameters of the risk source (such as the size of the leak, the leakage rate, and the physicochemical properties of the leaked substance (such as density and diffusion coefficient)) are obtained from the risk entries associated with the warning signal.

[0067] Centered on the location of the risk source, and based on the risk type, material properties (such as diffusion rate and density), environmental conditions, and a preset simulation time, a sufficiently large enclosing region Ω is defined in three-dimensional space using empirical engineering formulas (e.g., based on Pasquill-Gifford stability classification and maximum impact distance estimation) or conservative geometric extrapolation methods. This ensures that the impact of the risk is included within the dynamic simulation timeframe. An adaptive mesh refinement method is used to solve the governing equations. Initially, the simulation region Ω is divided into a basic structured or unstructured coarse mesh. The system defines mesh refinement criteria, such as automatically refining the mesh in areas with large risk quantification field gradients (e.g., near leak points or concentration fronts) and pre-refining the mesh in areas where critical equipment or personnel may reside. The mesh refinement area is dynamically adjusted based on equipment locations and real-time personnel positioning data from the digital twin model. The solution invokes a built-in or integrated computational fluid dynamics (CFD) solver kernel (such as a solver customized based on OpenFOAM). The solver uses the finite volume method or finite element method to spatially discretize the governing equations and employs explicit or implicit schemes for temporal discretization. It then performs iterative calculations on an adaptive mesh, updating the risk field values ​​on all mesh elements throughout the simulation region at each time step until a preset dynamic simulation end time is reached. The solution process uses an adaptive time step strategy to ensure numerical stability and computational efficiency. The solver outputs a spatiotemporally evolving risk quantization field C(x,t), where x is a three-dimensional spatial position vector (i.e., ), t is time; the physical meaning of C(x,t) depends on the type of risk, for example, for gas diffusion, C represents concentration (unit: kg / m³).

[0068] For each type of transmissible risk, a set of multi-level safety thresholds is predefined, corresponding one-to-one with the warning level. The safety threshold is set based on safety standards, and each This represents a specific level of danger; reaching or exceeding this level indicates that the risk has reached the warning level. The required level of initiation of the corresponding response; for example, for methane gas, (Attention) 10% of the lower explosive limit (LEL) can be taken. Based on the risk quantification field C(x,t), for each warning level... Extracting satisfaction The set of all spatial points, which constitutes an isosurface in three-dimensional space, is defined as the dynamic early warning boundary. : In terms of numerical implementation, isosurface extraction uses the classic moving cube algorithm or moving tetrahedron algorithm. This algorithm traverses all mesh cells in the simulation region and extracts isosurfaces based on the C-value at the cell vertex and... The comparison relationship generates a series of triangular patches to approximate the isosurface. Extracted isosurfaces This refers to the dynamic early warning boundary, which indicates that at time t, the area affected by the risk just reaches the warning level. The spatial profile corresponding to the danger level.

[0069] A dynamic simulation instance is created for each warning signal that triggers the dynamic simulation. This instance is an encapsulated data structure that includes: the complete definition of the physical field numerical model selected for this simulation (such as the governing equations), the key physical parameters of the model (such as the diffusion coefficient and reaction rate of the leaked substance), all geometric and topological data of the adaptive mesh generated for this simulation, the risk quantification field C(x,t), and the warning boundary. Each warning signal object is associated with a corresponding dynamic simulation instance reference. The dynamic warning boundary sequence is then... (in The warning trigger time is overlaid in real-time onto the virtual 3D environment model of the digital twin model for visualization. Dynamic warning boundaries for different warning levels are rendered using surfaces with different colors and transparency (e.g., semi-transparent yellow for l=2, semi-transparent red for l=4, possibly with a warning pulsating effect). Not only is the dynamic warning boundary displayed at the current moment, but based on dynamic simulation results, the dynamic changes (expansion, movement) of the warning boundary over a future period (e.g., the next 5 minutes) can be projected and displayed in the form of a time slider or animation, thereby predicting the risk diffusion path and impact range. Finally, the visualized scene of the warning boundary is rendered using a visualization engine (such as Unity3D) and output to monitoring terminals and mobile devices.

[0070] Optionally, based on the compliance rule set defined by the digital twin model, correlation logic analysis and compliance judgment are performed on real-time multi-source data to generate the compliance status of each operation step. The compliance status of all operation steps is then integrated to generate a comprehensive risk assessment result for the current high-risk operation process, including: Based on the graded early warning signals, a risk management process is initiated for high-risk operational processes, and the management process is tracked in real time. After the risk management process is completed, the comparison results between the graded early warning signals and the actual occurrence of the risk are collected as verification information for the accuracy of the early warning, and the management process records are collected as information on the effectiveness of the management measures. A Bayesian optimization method is employed, utilizing collected information on the accuracy verification of early warnings and the effectiveness of response measures to probabilistically update the compliance rule set of the digital twin model and the dynamically simulated physical parameters. For any parameter to be optimized... Its posterior distribution is determined by the prior distribution. and likelihood function According to Bayes' theorem The update yielded the following: Verify the accuracy of newly collected early warning information and the effectiveness of response measures; after updating, take the posterior expectation as the latest parameter value; A multi-agent reinforcement learning approach is adopted, in which compliance status and graded early warning signals are used as agent actions, and the effectiveness information of the disposal measures is used as environmental feedback. The dynamic Bayesian network, the mapping rules between early warning level and risk level, and the numerical model are updated through policy gradient algorithm. The optimized compliance rule set, dynamically simulated physical parameters, dynamic Bayesian network, mapping rules between warning level and risk level, and numerical model are deployed to the digital twin model and corresponding functional modules. The complete process record and results of this risk handling process are updated as historical data to the operation ticket and work ticket instance library and the safety procedure and case library to complete the optimization loop.

[0071] Specifically, the system has a built-in standard emergency response plan library. This library predefines corresponding response procedures and tasks for different warning levels, risk types, and spatial locations. For example, for a methane leak warning, the plan may include automatically activating the ventilation system near the leak point, sending evacuation instructions to relevant personnel, notifying the emergency response team to arrive, and isolating the hazardous area. Based on the warning level, risk type, and location information of the warning signal, the most suitable emergency response plan is matched from the plan library, and a structured list of response tasks is generated. Each task includes the execution actions, responsible roles, expected completion time, and dependencies between tasks. Through the power plant's operation scheduling system and personnel mobile terminals, the response tasks are automatically assigned to the corresponding execution entities (including automatic control systems and personnel). Simultaneously, a virtual response process is initiated in the virtual environment of the digital twin model for tracking and comparison.

[0072] The system tracks the execution status of disposal tasks in real time through multiple channels; for automatically executed tasks (such as starting ventilation), it receives confirmation signals of successful execution through the industrial control network; on-site personnel report the start, completion, or problems encountered in the task through mobile terminals (such as tablets); and it uses multimodal sensing equipment at high-risk work sites to monitor the effectiveness of disposal measures, such as whether the gas concentration has decreased, whether personnel have been evacuated, and whether isolation zones have been established.

[0073] After the virtual response process is completed (or the preset observation period is reached), information on the accuracy of the early warning is collected. The early warning signal is compared with the actual occurrence of the risk. For example, whether the risk actually occurred, whether it expanded, or whether it caused losses is extracted from historical perception data and safety incident reports, thereby forming verification labels such as true positive and false positive. Information on the effectiveness of the response measures is also collected. Each response measure taken is recorded, as well as the changes in risk indicators after the measures are implemented. For example, the rate at which methane concentration decreases after ventilation is started, and whether there are still people remaining in the danger zone after personnel evacuation. These are quantified into indicators such as response time and risk reduction rate.

[0074] By utilizing the collected information on the accuracy of early warnings and the effectiveness of response measures, key uncertain parameters in the digital twin model are updated probabilistically using a Bayesian optimization framework. These uncertain parameters mainly include parameters in the compliance rule set (such as the steepness coefficient of the likelihood function), parameters in the dynamic safety boundary calculation model (i.e., the weight coefficient and sensitivity coefficient of the dynamic safety radius calculation formula), and physical parameters of the physical field numerical model involved in the dynamic simulation module (such as the gas diffusion coefficient and turbulence model constants).

[0075] In the Bayesian optimization framework, for each parameter to be optimized (i.e., the uncertain parameter) It is treated as a random variable and a Gaussian process is used to build a surrogate model of it and the objective function (such as the early warning F1 score or the risk decay rate). Parameters prior distribution Based on historical data settings (such as setting it to a uniform or Gaussian distribution). When new observation data is obtained... After verifying the accuracy of early warnings or the effectiveness of response measures, the posterior distribution of the parameters is updated according to Bayes' theorem: , where the likelihood function This indicates that when the parameter is Observed at time The probability is calculated. The next evaluation point is selected using acquisition functions such as expectation improvement, and after iterative updates, the posterior expectation is taken as the latest value of the parameter. ,in It is all the accumulated observation data; Write the corresponding model parameters (such as compliance rule sets, dynamic security boundary calculation parameters, and physical field model parameters) into the configuration file or database to complete this parameter update.

[0076] For complex optimization problems involving multi-component collaboration and sequential decision-making, multi-agent reinforcement learning is employed to jointly optimize multiple core decision-making components. These components include a dynamic Bayesian network (DBN), a warning level mapping rule base, and the selection and parameters of a physical field numerical model. The compliance inspection and early warning system for the entire high-risk operation process is abstracted into a reinforcement learning environment, where the environment state... This includes the current state of the digital twin model, real-time multi-source sensing data, work process progress, and environmental actions. This includes generated compliance status, early warning signals, and handling recommendations, as well as environmental feedback. This information is derived from the effectiveness of the response measures (e.g., the faster the risk subsides, the higher the reward).

[0077] Multiple agents are employed, each responsible for optimizing different components. One agent optimizes the state transition probability matrix and observation likelihood matrix of the dynamic Bayesian network; another optimizes the mapping rule between warning levels and risk levels; and a third optimizes the selection of the physical field numerical model or its internal parameters. A centralized training and distributed execution framework is used, employing either Multi-Agent Deep Deterministic Policy Gradient (MADDPG) or Multi-Agent Proximal Policy Optimization (MAPPO) algorithms for training. Each agent designs a reward function based on its role; for example, the reward of a DBN agent is positively correlated with the degree of agreement between its predicted risk state confidence and the actual risk evolution. During training, each risk handling process is treated as a complete round, recording the state, action, and reward sequence and storing it in an experience replay buffer. Data is periodically sampled from the buffer, and the aforementioned algorithms are used to update the policy network and value network of each agent. After training, the policy network outputs of each agent are converted into specific, deployable model parameters and rules. For example, the policy output of the dynamic Bayesian network agent is converted into specific values ​​of the updated state transition probability matrix and observation likelihood matrix; the policy output of the early warning mapping rule agent is converted into updated mapping rule conditions (e.g., "risk state confidence greater than 0.6" is changed to "greater than 0.7"); and the policy output of the numerical model agent is converted into updated model selection priority or physical parameter values.

[0078] Before applying the optimized compliance rule set, dynamic Bayesian network parameters, early warning mapping rules, and numerical model parameters to a high-risk operational environment, a shadow mode test was conducted. This involved running the old and new models in parallel, with the old model's output still used for actual control, while the new model's output was only used for recording and comparison. This process continued for a period to confirm that the new model was no less accurate or stable than the old model. For changes to key rules, A / B testing was employed, dividing the field into two groups: one using the old rules and the other using the new rules. The differences in early warning accuracy and response effectiveness between the two groups were compared. After successful testing, a hot deployment mechanism was used to deploy the updated compliance rule set, dynamic simulation physical parameters, dynamic Bayesian network, early warning mapping rules, and numerical model to the corresponding functional modules without service interruption. A model version management system was maintained to record the content, time, optimization goals, and performance improvements of each update for review and auditing purposes.

[0079] Finally, the complete record of this risk handling process (including warning signals, handling tasks, execution feedback, changes in perceived data, and parameter and model versions before and after optimization) is structured and stored in historical data. Specifically, it is updated to two core libraries: the execution record of this high-risk operation is added to the operation ticket and work ticket instance library as a new historical instance; in the safety procedures and case library, if this handling process discovers a new risk pattern or verifies effective handling measures, a new case record can be automatically generated or added to the case library after confirmation by a safety engineer, thereby completing the learning loop from handling to optimization.

[0080] In another embodiment of the present invention, a compliance inspection and early warning system for high-risk work processes is provided, comprising a model definition module, a data acquisition module, a risk assessment module, and an early warning module, wherein: Model definition module: Used to build a corresponding digital twin model for each high-risk operation. The digital twin model defines the sequence of operation steps for each high-risk operation, the set of compliance rules for each operation step, and the associated dynamic safety boundary. Data acquisition module: Used to collect and integrate real-time multi-source data from high-risk work sites based on the current work steps and associated dynamic safety boundaries defined by the digital twin model; Risk assessment module: Based on the compliance rule set defined by the digital twin model, it performs correlation logic analysis and compliance judgment on real-time multi-source data, and generates a comprehensive risk assessment result for the current high-risk operation process; Early warning module: Used to generate tiered early warning signals based on comprehensive risk assessment results and by calling the digital twin model.

[0081] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can be implemented in one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs) containing computer-usable program code. The form of a computer program product implemented on ROM (optical memory, etc.).

[0082] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0083] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or substitutions to the specific implementation of the invention, but these changes, modifications or substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. A method for compliance inspection and early warning of high-risk work processes, characterized in that, Includes the following steps: For each high-risk operation, a corresponding digital twin model is established. The digital twin model defines the sequence of operation steps for each high-risk operation, the set of compliance rules for each operation step, and the associated dynamic safety boundary. Based on the current work steps and associated dynamic safety boundaries defined by the digital twin model, real-time multi-source data from high-risk work sites are collected and fused in a targeted manner. Based on the compliance rule set defined by the digital twin model, the real-time multi-source data is subjected to correlation logic analysis and compliance judgment to generate a comprehensive risk assessment result for the current high-risk operation process; Based on the comprehensive risk assessment results, and by invoking the digital twin model, a tiered early warning signal is generated.

2. The compliance inspection and early warning method for high-risk work processes according to claim 1, characterized in that, Based on the aforementioned graded early warning signals, a risk management process is initiated, and information on the accuracy verification of early warnings and the effectiveness of management measures is collected during the risk management process to continuously optimize the digital twin model.

3. The compliance inspection and early warning method for high-risk operation processes according to claim 2, characterized in that, The optimization process of the digital twin model includes: Based on the graded early warning signals, a risk management process is initiated for high-risk operational processes. After the risk management process is completed, the comparison results between the graded early warning signals and the actual occurrence of the risk are collected as verification information for the accuracy of the early warning, and the management process records are collected as information on the effectiveness of the management measures. A Bayesian optimization method is used to probabilistically update the compliance rule set of the digital twin model and the physical parameters of the dynamic simulation using the collected early warning accuracy verification information and the effectiveness information of the response measures. A multi-agent reinforcement learning approach is adopted, in which compliance status and graded early warning signals are used as agent actions, and the effectiveness information of the disposal measures is used as environmental feedback. The dynamic Bayesian network, the mapping rules between early warning level and risk level, and the numerical model are updated through policy gradient algorithm. The optimized compliance rule set, dynamically simulated physical parameters, dynamic Bayesian network, mapping rules between warning levels and risk levels, and numerical model are deployed into the digital twin model. The complete process record and results of this risk handling process are used as historical data to complete the optimization loop.

4. The compliance inspection and early warning method for high-risk work processes according to claim 1, characterized in that, The process of establishing the digital twin model includes: Obtain multi-source heterogeneous data related to high-risk operations from a pre-set local database; Based on multi-source heterogeneous data, the sequence of high-risk operation steps is parsed, and the safety entities, behavioral norms and constraints are identified and extracted in a structured manner. Based on the sequence of high-risk operation steps, the safety entities, behavioral norms and constraints are associated and integrated to construct an initial safety knowledge graph. For each high-risk operation, a digital twin model is generated by integrating the initial safety knowledge graph.

5. The compliance inspection and early warning method for high-risk work processes according to claim 1, characterized in that, The real-time multi-source data acquisition and fusion process is as follows: Based on the compliance rule set and dynamic safety boundaries defined by the current work step in the execution state in the digital twin model, driven by high-risk work process data, generate awareness task instructions; Based on the spatial monitoring range and the required data types of sensing in the sensing task instructions, a subset of multi-source sensing devices with matching locations and functions is selected and activated from the sensor network deployed at high-risk work sites. Spatiotemporal alignment processing is performed on real-time multi-source data collected by a subset of multi-source sensing devices, and feature extraction and fusion are performed on the spatiotemporally aligned real-time multi-source data.

6. The compliance inspection and early warning method for high-risk work processes according to claim 1, characterized in that, The process for generating the comprehensive risk assessment results of the high-risk operation process is as follows: Calculate the probability of violation for each rule in the set of compliance rules defined by the current job step that is in execution state; Based on the violation condition probability of each rule in the current operation step, the posterior probability distribution of the current operation step at different risk levels is calculated.

7. The compliance inspection and early warning method for high-risk work processes according to claim 1, characterized in that, Based on the risk type, level, and spatial location information identified by the comprehensive risk assessment results, and by calling the digital twin model, a graded early warning signal corresponding to the risk level is generated according to the preset mapping rules between early warning levels and risk levels.

8. A compliance inspection and early warning system for high-risk work processes, characterized in that, It includes a model definition module, a data acquisition module, a risk assessment module, and an early warning module, among which: Model definition module: used to build a corresponding digital twin model for each high-risk operation. The digital twin model defines the sequence of operation steps for each high-risk operation, the set of compliance rules for each operation step, and the associated dynamic safety boundary. Data acquisition module: used to collect and fuse real-time multi-source data from high-risk work sites in a targeted manner, based on the current work steps and associated dynamic safety boundaries defined by the digital twin model; Risk assessment module: Based on the compliance rule set defined by the digital twin model, it performs correlation logic analysis and compliance judgment on the real-time multi-source data to generate a comprehensive risk assessment result for the current high-risk operation process; Early warning module: Used to generate graded early warning signals based on the comprehensive risk assessment results and by calling the digital twin model.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements as claimed in claim 1.

7. The steps of any of the methods described.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements as described in claim 1.

7. The steps of any of the methods described.

Images