Multi-modal collaborative ai agent management system and method for chemical production safety

Abstract

The application provides a multi-modal collaborative AI intelligent agent management system and method for chemical production safety, belonging to the field of chemical production safety management. The system includes an interactive perception module, a working condition fitting intelligent agent, a risk identification intelligent agent, an evolution identification intelligent agent, and an early warning module. The interactive perception module is used to establish a multi-modal spatio-temporal data field; the working condition fitting intelligent agent is used to construct a unified multi-modal generation model and form a high-dimensional hidden space distribution; the risk identification intelligent agent is used to generate a multi-dimensional risk disturbance vector; the evolution identification intelligent agent is used to establish a risk evolution trajectory set; and the early warning module is used to execute early warning and synchronously establish a safety intervention strategy sequence. Through unified representation and coupled modeling of multi-modal spatio-temporal data under mechanism and topological constraints, the risk disturbance is explained and decomposed, and the evolution trajectory is forward deduced, so that the timeliness, traceability and accuracy of chemical safety risk identification and early warning intervention are improved.

Description

Technical Field

[0001] This invention relates to the field of chemical safety production management, and in particular to a multimodal collaborative AI intelligent agent management system and method for chemical safety production. Background Technology

[0002] Chemical production processes often involve high temperature, high pressure, flammability, explosiveness, toxicity and hazard, and are generally characterized by long continuous processes, complex reaction mechanisms, and multi-media coupling. There is a strong coupling relationship between process parameters, equipment status, environmental conditions and personnel operation. Even a small disturbance in any link may be propagated and amplified along the process link, eventually evolving into a safety accident.

[0003] With the development of the Industrial Internet and intelligent sensing technology, chemical plants have widely deployed various data collection methods such as DCS, SIS, video surveillance, environmental monitoring, and personnel positioning, forming monitoring data covering multiple dimensions including processes, equipment, environment, and personnel. Existing chemical safety management solutions typically process and alarm independently for each subsystem when utilizing this data, with process data, equipment data, environmental data, and personnel data remaining independent of each other. In the risk identification stage, existing solutions often use threshold judgments or single-modal anomaly detection models, only issuing alarms when risks have become apparent, resulting in a lag in risk identification. In the risk analysis stage, the warning results of existing solutions often manifest as a single risk score or a general anomaly indication, making it difficult for operators to trace the specific source dimensions of the risk and the contribution degree of each dimension, and the warning results lack interpretability. In the risk prediction stage, existing solutions usually only determine the current state and cannot extrapolate the transmission and evolution of risks downstream along the process chain, making the risk evolution path untraceable, the warnings lack foresight, and fail to allow sufficient time for on-site intervention.

[0004] In summary, existing chemical safety production management schemes suffer from technical problems such as fragmented multimodal data, leading to delayed risk identification, untraceable evolution paths, and a lack of interpretability and foresight in early warning. Summary of the Invention

[0005] This invention addresses the technical problems in existing chemical safety production where multimodal data is fragmented, leading to delayed risk identification, untraceable evolution paths, and a lack of interpretability and foresight in early warning. It provides a multimodal collaborative AI intelligent agent management system and method for chemical safety production to solve these problems.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0007] In a first aspect, this invention provides a multimodal collaborative AI intelligent agent management system for safe chemical production, comprising: an interactive perception module, used to perform interactive perception of processes, equipment, environment, and personnel during the production and operation of a chemical plant, and to establish a multimodal spatiotemporal data field; a working condition fitting intelligent agent, used to construct a unified multimodal generation model based on historical calibration data of the multimodal spatiotemporal data field, introducing constraints of chemical reaction mechanisms and process topology, wherein the unified multimodal generation model is used to jointly model the coupling relationship between multimodal data, forming a high-dimensional latent space distribution characterizing the evolution law of normal working conditions; a risk identification intelligent agent, used to map the real-time collected multimodal spatiotemporal data field to the high-dimensional latent space distribution, and after calculating the deviation feature set, to perform interpretability decomposition of the deviation feature set, generating a multidimensional risk perturbation vector; an evolution identification intelligent agent, used to perform risk evolution recursive analysis within the unified multimodal generation model using the multidimensional risk perturbation vector, and to establish a risk evolution trajectory set; and an early warning module, used to execute early warnings based on the risk evolution trajectory set, and simultaneously establish a safety intervention strategy sequence.

[0008] Secondly, this invention provides a multimodal collaborative AI intelligent agent management method for chemical safety production, including: during the production and operation of a chemical plant, performing interactive perception of processes, equipment, environment, and personnel to establish a multimodal spatiotemporal data field; based on historical calibration data of the multimodal spatiotemporal data field, introducing constraints of chemical reaction mechanisms and process topology to construct a unified multimodal generation model, wherein the unified multimodal generation model is used to jointly model the coupling relationships between multimodal data, forming a high-dimensional latent space distribution characterizing the evolution law of normal operating conditions; mapping the real-time collected multimodal spatiotemporal data field to the high-dimensional latent space distribution, and after calculating the deviation feature set, performing interpretability decomposition of the deviation feature set to generate a multidimensional risk disturbance vector; using the multidimensional risk disturbance vector to perform risk evolution recursive analysis within the unified multimodal generation model to establish a risk evolution trajectory set; and executing early warning based on the risk evolution trajectory set, and simultaneously establishing a safety intervention strategy sequence.

[0009] The beneficial effects of this invention are:

[0010] First, through the interactive sensing module, interactive sensing of processes, equipment, environment, and personnel is performed during the production and operation of the chemical plant, establishing a multimodal spatiotemporal data field. This unifies the previously scattered process parameters, equipment status, environmental conditions, and personnel behavior data across various subsystems in both temporal and spatial dimensions, forming a foundational data base covering all elements of chemical production and providing a unified data source for subsequent cross-modal modeling. Based on this, a unified multimodal generation model is constructed by using a condition-fitting agent based on historical calibration data from the multimodal spatiotemporal data field, incorporating constraints from chemical reaction mechanisms and process topology. This unified multimodal generation model is then used to jointly model the coupling relationships between multimodal data, forming a high-dimensional latent space distribution representing the evolution of normal operating conditions. Chemical reaction mechanisms and process topology are integrated as prior constraints into data-driven modeling, ensuring that the generation model both follows objective physical and chemical laws and incorporates historical operating experience. This allows for a stable characterization of the coupling relationships of multimodal data under normal operating conditions within the high-dimensional latent space, providing a reference benchmark for subsequent risk identification. Subsequently, the risk identification agent maps the real-time collected multimodal spatiotemporal data field to a high-dimensional latent space distribution, calculates the deviation feature set, and then performs interpretability decomposition on the deviation feature set to generate a multidimensional risk disturbance vector. This structurally decomposes the difference between the real-time operating state and the normal operating condition baseline into different dimensions such as process, equipment, environment, and personnel, so that the deviation is no longer represented by a single, general score, but rather forms a disturbance vector with traceable contributions from each dimension, providing support for the interpretability of the early warning results. Furthermore, the evolutionary identification agent uses the multidimensional risk disturbance vector to perform risk evolution recursive analysis within a unified multimodal generation model, establishing a risk evolution trajectory set. Leveraging the operating condition evolution patterns already learned by the generation model, it uses the current risk disturbance as a starting point to prospectively extrapolate the subsequent transmission path of the risk along the process link, forming a traceable trajectory set. This extends risk management from current state judgment to the prediction of future evolution trends. Subsequently, the early warning module executes early warnings based on the risk evolution trajectory set and simultaneously establishes a sequence of safety intervention strategies. This unifies the risk identification results and evolution simulation results into on-site executable early warning signals and intervention actions, so that the early warning not only informs of the existence of the risk, but also provides a handling strategy that matches the evolution trajectory.

[0011] Through the above technical solution, a multimodal spatiotemporal data field established by the interactive sensing module serves as the data foundation. A working condition fitting agent constructs a unified multimodal generation model under mechanistic and topological constraints, enabling multimodal data to achieve unified representation and coupled modeling in a high-dimensional latent space. Then, a risk identification agent decomposes the deviation feature set into multidimensional risk disturbance vectors, achieving interpretable attribution of risk sources. Next, an evolutionary identification agent performs recursive analysis on the risk disturbance vectors within the generation model, obtaining a set of risk evolution trajectories, enabling forward-looking prediction and traceability of risk evolution paths. Finally, an early warning module translates the above results into early warning signals and safety intervention strategy sequences, thereby improving the timeliness, traceability, and accuracy of early warning and intervention in chemical safety risk identification. Attached Figure Description

[0012] Figure 1 A schematic diagram of the structure of the multimodal collaborative AI intelligent agent management system for chemical safety production provided by the present invention;

[0013] Figure 2 This is a flowchart illustrating the multimodal collaborative AI agent management method for chemical safety production provided by the present invention.

[0014] In the attached diagram, the components represented by each number are as follows:

[0015] Interactive perception module 11, working condition fitting agent 12, risk identification agent 13, evolution identification agent 14, and early warning module 15. Detailed Implementation

[0016] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0018] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0019] Example 1, as Figure 1 As shown, this embodiment of the invention provides a multimodal collaborative AI agent management system for chemical safety production, including:

[0020] Interactive sensing module 11 is used to perform interactive sensing of processes, equipment, environment and personnel during the production and operation of a chemical plant, and to establish a multimodal spatiotemporal data field.

[0021] Specifically, the interactive sensing module 11 is deployed between the edge-side data acquisition gateway and the central management platform in the chemical plant, and is responsible for providing data to subsequent intelligent agents. The objects of interactive sensing include four dimensions: process, equipment, environment, and personnel, and a multimodal spatiotemporal data field is established based on the interactive sensing results of the four dimensions.

[0022] For the process dimension, the objects of perception are the process parameters in chemical reactions and production flows. The data sources for this dimension are primarily the existing DCS (Distributed Control System) and SIS (Safety Instrumented System) of the chemical plant, accessed via OPC UA protocol, Modbus TCP protocol, or a dedicated plant communication bus. The data collected includes process variables such as temperature, pressure, liquid level, flow rate, pH value, conversion rate, feed ratio, catalyst addition, stirring speed, and reflux ratio in the reactor.

[0023] From an equipment perspective, the sensing objects are key electromechanical equipment in chemical plants that perform functions such as power generation, conveying, heat exchange, and agitation. Data sources for this dimension include vibration sensors, shaft temperature probes, smart meters, valve opening feedback loops, and equipment status word upload interfaces installed on the equipment itself. Data collected includes vibration spectra, bearing temperatures, current and voltage, and start / stop status for compressors and pumps; valve opening feedback and response times; temperature differences and pressure drops on the hot and cold sides of heat exchangers; and current and torque for agitators, used to reflect the health status and operational margin of the equipment.

[0024] For the environmental dimension, the sensing objects are the environmental conditions within the chemical plant area that may affect safe production. Data sources for this dimension include combustible gas detectors, toxic gas detectors, temperature and humidity sensors, anemometers, leak acoustic arrays, and infrared thermal imagers deployed in key areas. The data collected includes the concentrations of combustible components such as methane, hydrogen, and volatile organic compounds; the concentrations of toxic components such as hydrogen sulfide, carbon monoxide, and ammonia; ambient temperature and humidity, wind speed and direction, rain and snow conditions; and the acoustic signatures of pipeline leaks obtained through acoustic arrays and surface temperature distribution obtained through infrared imaging.

[0025] For the personnel dimension, the sensing objects are the location, behavior, and protective status of workers within the factory area. Data sources for this dimension include UWB positioning base stations, access control card readers, smart safety helmets and smart PPE, and video behavior recognition cameras. The data collected includes the real-time location coordinates of personnel, the time and duration of entry into restricted areas, the status of protective equipment wearing, whether they are in dangerous working postures, and whether there are any violations of operating procedures. The personnel dimension data includes both time-series coordinates from location tracking and structured events from video recognition.

[0026] After completing the acquisition of sensing data across the four dimensions of process, equipment, environment, and personnel, the interactive sensing module 11 further performs convergence processing on the interactive sensing results across these four dimensions. During this convergence processing, the interactive sensing module 11 uses the chemical plant's unified master clock as the time reference to time-align sensing data from different sources with different sampling periods. It also uses the chemical plant's digital plant area model and process flow topology as the spatial reference, assigning each sensing data point a corresponding spatial location identifier and process unit affiliation identifier. This establishes a correspondence between the sensing data across the four dimensions of process, equipment, environment, and personnel within the same spatiotemporal reference system, thereby forming a multimodal spatiotemporal data field with unified spatiotemporal reference characteristics. The multimodal spatiotemporal data field uses unified spatiotemporal coordinates as an index, supporting slice queries and joint calls of the sensing data across the four dimensions by process unit, spatial location, and time window.

[0027] By establishing a multimodal spatiotemporal data field, the interactive sensing module 11 organizes the sensing data that were originally scattered in various independent subsystems within the chemical plant into a data set with spatiotemporal correlation characteristics. This provides spatiotemporally consistent data input for subsequent joint modeling of the coupling relationship between multimodal data, high-dimensional latent space mapping, and risk identification processing, avoiding the loss of coupling relationship caused by the fragmentation of data in different dimensions.

[0028] Furthermore, in the interactive perception module 11, a multimodal spatiotemporal data field is established, including:

[0029] The interactive perception results of processes, equipment, environment, and personnel are managed for time synchronization and spatial registration, and a data set under a unified spatiotemporal reference system is constructed.

[0030] The structured organization and analysis of multimodal data relationships in the execution dataset forms a multimodal spatiotemporal data field with spatiotemporal correlation characteristics.

[0031] In one feasible implementation, when establishing a multimodal spatiotemporal data field, the first step is to perform time synchronization and spatial registration management on the interactive perception results of processes, equipment, environment, and personnel, and construct a data set under a unified spatiotemporal reference system.

[0032] Time synchronization management includes two aspects: time base unification and sampling rhythm unification. Regarding time base unification, the interactive sensing module 11 deploys a unified master clock in the chemical plant control network, using either the PTP precision time protocol or the NTP network time protocol to align the local clocks of various acquisition devices with the master clock. The alignment accuracy is set to milliseconds or sub-milliseconds depending on downstream processing requirements, ensuring that each sensing data source uses the master clock timestamp as its unique time index. Regarding sampling rhythm unification, considering that process parameters at the process dimension typically have sampling periods of sub-seconds to seconds, vibration spectrum sampling frequencies at the equipment dimension can reach kilohertz levels, gas concentration and meteorological parameter sampling periods at the environmental dimension are typically several seconds, personnel positioning data update periods are typically hundreds of milliseconds to seconds, and video behavior recognition events are non-sequential events, there are significant differences in sampling rhythms between different acquisition sources. The interactive sensing module 11 performs resampling processing on each sensing data source based on a pre-set unified sampling rhythm. For sensing data with a sampling frequency higher than the unified sampling rhythm, the mean, extreme values, or feature statistics are taken within the time window corresponding to the unified sampling rhythm, and the multi-point data within the window are aggregated into a single-point output. For sensing data with a sampling frequency lower than the unified sampling rhythm, the data points are supplemented between two adjacent sampling times by selecting either the previous value retention method or the linear interpolation method according to the physical meaning of the sensing data. Specifically, the previous value retention method is used for step sensing data such as valve opening degree and start / stop status, while the linear interpolation method is used for sensing data with continuous change characteristics such as temperature, pressure, and concentration. For non-uniform time-series sensing data such as video behavior recognition events, the interactive sensing module 11 uses the master clock timestamp of the event occurrence time as the alignment anchor point, and classifies the event into the corresponding time window according to its timestamp. Together with the periodic time-series sensing data within the time window, it constitutes the multimodal sensing data set of the time window. After the above time base unification and sampling rhythm unification processing, the sensing data of the four dimensions are all aligned according to the unified master clock timestamp and output according to the unified sampling rhythm, forming a time-series data set with the master clock timestamp as the time index.

[0033] Spatial registration management includes establishing spatial reference benchmarks and binding spatial identifiers for sensing data. When establishing spatial reference benchmarks, the interactive sensing module 11 calls upon a pre-established digital plant area model of the chemical plant as the spatial reference benchmark. The digital plant area model includes, but is not limited to, a 3D plant area model based on BIM technology and a process flow topology model based on P&ID diagrams. Both types of models are labeled with the location coordinates and geometric extent of spatial entities such as reactors, heat exchangers, compressors, pumps, pipelines, valves, tank areas, loading and unloading areas, control rooms, and evacuation routes. They also label the division and boundaries of functional areas such as tank areas, reaction areas, loading and unloading areas, the area surrounding the control room, and evacuation routes, as well as the material flow direction, energy exchange relationships, and connection topology between each process unit. When binding spatial identifiers for sensing data, the interactive sensing module 11 binds three spatial identifiers to each sensing data stream: first, spatial location coordinates, corresponding to the 3D coordinates of the sensing device in the digital plant area model; second, spatial area identifier, corresponding to the functional area where the sensing device is located; and third, process unit affiliation identifier, corresponding to the process unit node to which the object measured by the sensing device belongs in the process flow topology. For sensing devices with fixed installation locations, the above three identifiers are statically bound during the device deployment phase. For sensing sources with dynamic location attributes, such as personnel positioning and mobile monitoring devices, the above three identifiers are dynamically generated by the interactive sensing module 11 based on the real-time coordinates reported by the device in the digital plant area model. After the above spatial reference benchmark establishment and sensing data spatial identifier binding processing, the sensing data in all four dimensions are spatially positioned according to the spatial location coordinates, spatial area identifiers, and process unit affiliation identifiers in the digital plant area model, forming a spatially aligned data set with the three spatial identifiers as spatial indices.

[0034] After the above time synchronization management and spatial registration management, the four-dimensional perception data are aligned with the master clock timestamp as the time index in the time dimension and with the three spatial identifiers in the digital factory area model as the spatial index in the spatial dimension. Together, they form a data set under a unified spatiotemporal reference system, which serves as the processing object for subsequent structured organization and analysis of multimodal data association relationships.

[0035] Next, the interactive sensing module 11 performs structured organization analysis of multimodal data association relationships on the dataset under the unified spatiotemporal reference system, forming a multimodal spatiotemporal data field with spatiotemporal association characteristics. The structured organization analysis includes intra-unit multimodal slice organization, inter-unit directional association organization, and cross-unit shared data association organization. First, in intra-unit multimodal slice organization, the interactive sensing module 11 organizes the dataset using the process unit affiliation identifier as the primary index and the time window as the secondary index. The time window uses the sampling period corresponding to the aforementioned unified sampling rhythm as the basic window length, and is set as an integer multiple of several basic windows according to downstream processing requirements. Overlapping intervals are set or not set between adjacent time windows as needed. For each process unit, the interactive sensing module 11 associates and organizes the sensing data of the four dimensions of process, equipment, environment, and personnel belonging to the process unit and falling within the same time window into a group of multimodal data slices; multiple groups of slices of the same process unit in adjacent time windows are arranged sequentially along the time axis to form a multimodal spatiotemporal slice sequence of the process unit. Then, during the organization of directional associations between units, the interactive sensing module 11 establishes a directional association index between the multimodal spatiotemporal slice sequences of adjacent process units based on the material flow direction and energy exchange relationship recorded in the process flow topology of the digital plant model. The directional association index stores the connection relationship from upstream process units to downstream process units in the form of an adjacency list. For connections with material transfer delays, such as pipeline transmission, the directional association index is supplemented with the transmission delay parameters estimated from the pipeline length and material flow rate in the digital plant model. This allows the sensing data of the upstream process unit at a given time to be mapped to the corresponding multimodal data slice of the downstream process unit according to the transmission delay parameters, thereby establishing a temporal correspondence between the sensing data of upstream and downstream process units. Subsequently, during the cross-unit shared data association organization, for shared sensing data spanning multiple process units, such as meteorological parameters, combustible gas concentration fields, and personnel trajectories at the plant scale, the interactive sensing module 11 determines the spatial coverage of the shared sensing data based on the nominal detection coverage of the sensing device or the pre-divided functional area boundaries in the digital plant model. It then associates the shared sensing data with the multimodal spatiotemporal slice sequence of all process units within the covered area. After the above structured organization and analysis, the dataset under the unified spatiotemporal reference system is organized into a data structure that possesses both spatiotemporal indexes and cross-modal and cross-unit association indexes—that is, a multimodal spatiotemporal data field with spatiotemporal association characteristics. The multimodal spatiotemporal data field supports slice querying, joint invocation, and cross-unit tracing of sensing data in four dimensions: process unit, spatial location, time window, and material energy flow direction. This provides structured data with spatiotemporal consistency and cross-modal association characteristics for subsequent joint modeling of coupling relationships between multimodal data, high-dimensional latent space mapping, and risk identification processing.

[0036] Furthermore, this embodiment also includes a resampling module. This resampling module is used to perform anomaly trigger discrimination based on structured organization analysis. When the anomaly trigger discrimination result is a trigger pass result, the corresponding data is subjected to linked resampling processing.

[0037] In a preferred embodiment, the resampling module performs preliminary anomaly triggering discrimination based on the state of each multimodal data slice in the multimodal spatiotemporal data field. When the discrimination result indicates that there is a potentially abnormal process unit within the current time window, it triggers the linkage supplementary sampling of the process unit and its associated units, so that the process unit and its associated units can obtain higher density sensing data in the early stage of the possible anomaly, providing a more sufficient data foundation for subsequent joint modeling and risk identification processing based on the multimodal spatiotemporal data field.

[0038] First, the resampling module performs anomaly trigger discrimination on the multimodal data slices output for each time window, based on the content of each multimodal data slice in the multimodal spatiotemporal data field formed by structured organization analysis. The criteria for anomaly trigger discrimination include single-sensor data deviation discrimination, cross-modal correlation consistency deviation discrimination, and time series change rate discrimination. For single-sensor data deviation discrimination, the resampling module calculates the normalized deviation of each sensing data within the multimodal data slice relative to its pre-set normal value range. The normalized deviation is obtained by dividing the difference between the current value of the sensing data and the median of the normal value range by half the width of the normal value range. For cross-modal correlation consistency deviation discrimination, the resampling module pre-establishes correlation reference relationships between cross-modal sensing data within the same process unit based on historical normal data. These correlation reference relationships include, but are not limited to, correlation coefficient matrices between process parameters and pairing change rules between process parameters and equipment status, such as an increase in reactor temperature accompanied by an increase in cooling medium flow rate, and an increase in compressor current accompanied by an increase in outlet pressure. The resampling module calculates the degree of deviation between the combination of cross-modal sensing data values ​​within the current time window and the correlation reference relationships, obtaining the cross-modal correlation consistency deviation. For time-series change rate discrimination, the resampling module uses multimodal data slices from the same process unit in adjacent time windows as objects, calculates the first-order change rate of each sensing data point between adjacent time windows, i.e., the ratio obtained by subtracting the value of the previous time window from the value of the later time window and dividing by the time window length. It then compares the first-order change rate with the change rate distribution of the sensing data under historical normal operating conditions to obtain the time-series change rate deviation. The resampling module compares the deviations obtained from the above three types of discrimination with the corresponding trigger thresholds. The trigger threshold is determined by performing quantile statistics on historical normal operating data and adjusting it based on the experience of process experts. Typically, it is taken as the 95th or 99th percentile of the corresponding deviation distribution under historical normal operating conditions. To avoid false triggers caused by instantaneous data jitter, the resampling module further sets a continuous discrimination condition. Only when the deviation of the same path or the same cross-modal combination exceeds the corresponding trigger threshold within a consecutive preset number of time windows (e.g., 2 to 5), is the multimodal data slice corresponding to the current time window determined as a trigger pass result, and the process unit affiliation identifier, time window identifier, and the sensing data category on which the trigger is based are recorded.

[0039] Then, the resampling module determines the scope of the linkage resampling based on the cross-modal association index and cross-unit association index in the multimodal spatiotemporal data field. For linkages within the same process unit, the resampling module includes all sensing data in the four dimensions of process, equipment, environment, and personnel within the corresponding multimodal data slice, based on the multimodal slice organization results within the unit. For linkages between units, the resampling module includes the upstream and downstream process units that trigger the linkage along the material flow direction and energy exchange path, based on the directional association index. The propagation hop count of the linkage scope is determined according to the pre-set propagation depth parameter, which is generally 1 to 3 hops, i.e., propagating 1 to 3 process unit nodes upstream and downstream along the material flow direction and energy exchange path, respectively. The time window offset corresponding to the upstream and downstream process units is determined according to the transmission delay parameter attached in the directional association index. For linkages with shared sensing data across units, the resampling module includes the shared sensing data that triggers the linkage across the corresponding process unit, based on the cross-unit shared data association organization results.

[0040] Subsequently, the resampling module sends a resampling command to the acquisition devices corresponding to the sensing data falling within the linkage range. This causes the acquisition devices to perform temporary supplementary sampling of the sensing data falling within the linkage range at a resampling rhythm higher than the pre-set uniform sampling rhythm. The resampling rhythm is set as an integer multiple of the uniform sampling rhythm based on the category of sensing data used to determine the trigger pass result and the degree of deviation from the abnormal trigger judgment. The acquisition devices report sensing data according to the resampling rhythm within the pre-set resampling duration. After the resampling duration ends, the system returns to the uniform sampling rhythm. During the linkage resampling execution, non-uniform temporal sensing sources for video behavior recognition temporarily increase the reporting accuracy and frequency of behavior recognition events after receiving the linkage resampling command.

[0041] Subsequently, the resampling module realigns the high-rhythm sensing data collected during the resampling period to the time index of the multimodal spatiotemporal data field according to the unified sampling rhythm processing method described in the aforementioned time synchronization management. For multi-point data within the time window corresponding to the unified sampling rhythm, the average, extreme values, or characteristic statistics are still aggregated to form the representative value for that time window. Simultaneously, the original high-rhythm sensing data from the resampling period is attached to the corresponding multimodal data slice as an additional data layer. The additional data layer is stored as an independent field in the data structure of the multimodal data slice. Within the additional data layer, the original sampling points of the high-rhythm sensing data are stored sequentially using the timestamp corresponding to the resampling rhythm as an index. It establishes a connection with the regular data layer through the process unit affiliation identifier and time window identifier shared with the multimodal data slice, enabling the multimodal spatiotemporal data field to simultaneously possess a regular data layer organized according to the unified sampling rhythm and a high-density additional data layer organized according to the resampling rhythm at that time window.

[0042] Through the above processing, the data density of the multimodal spatiotemporal data field on the process unit and its associated unit that may have anomalies is temporarily increased, so that the subsequent joint modeling of the coupling relationship between multimodal data, high-dimensional latent space mapping and risk identification processing can obtain more sufficient perception data in the early stage of suspected anomaly occurrence, and improve the response speed of anomaly identification and risk judgment.

[0043] The working condition fitting agent 12 is used to construct a unified multimodal generation model based on the historical calibration data of the multimodal spatiotemporal data field, by introducing constraints of chemical reaction mechanism and process topology. The unified multimodal generation model is used to jointly model the coupling relationship between multimodal data to form a high-dimensional latent space distribution that characterizes the evolution law of normal working conditions.

[0044] Specifically, the working condition fitting agent 12 receives the multimodal spatiotemporal data field output by the interactive perception module 11, constructs a unified multimodal generation model based on the multimodal spatiotemporal data field, and forms a high-dimensional latent space distribution.

[0045] First, historical calibration data of the multimodal spatiotemporal data field is acquired. This historical calibration data is selected from the chemical plant's historical operating records and represents the stable operating periods confirmed by process experts. To ensure the historical calibration data adequately covers the evolution of normal operating conditions of chemical production units, the data collection period typically covers at least one complete production cycle and includes operating records under different load conditions, different raw material ratios, and different environmental temperature and humidity conditions. Data corresponding to known maintenance periods, process changeover periods, and periods of recorded abnormal events are excluded to ensure that the data input for modeling are all samples of normal operating conditions.

[0046] Then, based on historical calibration data, joint modeling of the coupling relationships between multimodal data is performed. These coupling relationships refer to the mutual influence and constraints among changes in process parameters, equipment operating status, environmental conditions, and personnel operating behaviors during chemical production. Examples include the correspondence between reactor temperature changes and cooling medium flow rate changes, compressor vibration state changes and outlet pressure changes, and combustible gas concentration changes and personnel work area locations. In existing chemical safety management schemes, these coupling relationships cannot be explicitly expressed because the data from the four dimensions of process, equipment, environment, and personnel are fragmented. The operating condition fitting agent 12 performs cross-modal joint modeling on the four dimensions of perceived data in the multimodal spatiotemporal data field, enabling a unified expression of these coupling relationships in the modeling results. During the joint modeling process, the operating condition fitting agent 12 introduces two types of constraints—chemical reaction mechanism constraints and process topology constraints—to work together in the modeling process. Chemical reaction mechanism constraints, based on chemical reaction kinetic equations, material balance relationships, and energy conservation relationships, impose physical feasibility restrictions on the changes in state variables within the same process unit in the modeling results. This ensures that the multimodal data combinations within the same process unit in the modeling results always satisfy the mechanistic laws of the chemical production process, avoiding physically infeasible state combinations in the modeling results. Process topology constraints, based on the material flow direction, energy exchange relationships, and connection topology between process units in the chemical production unit, impose directional restrictions on the connection relationships and information transmission directions between state variables corresponding to different process units in the modeling results. This ensures that the information flow direction between adjacent process units in the modeling results is consistent with the actual material and energy flow direction of the chemical production unit, preventing the reverse transmission of state from downstream process units to upstream process units in the modeling results. These two types of constraints act on the same process unit and between adjacent process units, respectively, jointly ensuring that the joint modeling results possess physical interpretability in addition to data distribution fitting. Under the influence of these two types of constraints, the operating condition fitting agent 12 constructs a unified multimodal generative model as the carrier for joint modeling. The unified multimodal generative model is based on generative models. In its specific implementation, generative model architectures with controllable latent space structure, such as conditional variational autoencoders, diffusion generative models, or flow matching generative models, can be selected.For example, a unified multimodal generation model consists of a modal encoder, a coupled encoder, a latent space, a coupled decoder, and a modal decoder connected sequentially. The modal encoder sets up dedicated sub-encoders for each of the four dimensions of perceived data: process, equipment, environment, and personnel. Temporal-series perceived data uses a temporal convolutional network or Transformer encoder to extract features; video behavior recognition perceived data uses a convolutional neural network combined with a temporal aggregation module to extract features; and scalar-series perceived data uses a multilayer perceptron to extract features. The coupled encoder, based on the output features of the modal encoder, performs cross-modal coupling on the features of the four dimensions of process, equipment, environment, and personnel to obtain the coupled features corresponding to each process unit. The latent space refers to the space between the encoder and decoder in the unified multimodal generation model. The continuous vector space between encoders has a lower dimension than the space containing the original multimodal sensing data. After processing by the modal encoder and the coupled encoder, the original multimodal sensing data is compressed and mapped into vector points in the latent space. The dimensional values ​​of each vector point carry the essential state features extracted from the original multimodal sensing data. The distance and direction relationships between vector points in the latent space reflect the similarity and change relationships between the original multimodal sensing data. The latent space is based on the latent space sub-representation regions corresponding to each process unit as basic units, and the connection relationships and information transmission directions between the latent space sub-representation regions are structurally preset according to the process topology constraints. The coupled decoder and the modal decoder, as the inverse process of the encoder, reconstruct the sensing data of each modality from the latent space. The chemical reaction mechanism constraints are embedded in the latent space sub-representation regions corresponding to each process unit and the output layer of the modal decoder in the form of mechanism residual terms. The process topology constraints are embedded in the latent space in the form of the structured preset of the connection relationships and information transmission directions between the latent space sub-representation regions.

[0047] Next, the working condition fitting agent 12, based on a unified multimodal generation model embedding two types of constraints, performs training processing on historical calibration data. The optimization objectives of the training process include two objectives: a data distribution fitting objective and a mechanism constraint objective. The data distribution fitting objective aims to minimize the reconstruction error between the reconstructed sensing data output by the modal decoder and the input sensing data, enabling the unified multimodal generation model to fit the distribution of multimodal data under normal working conditions. The mechanism constraint objective aims to minimize the mechanism residual term corresponding to the chemical reaction mechanism constraint, ensuring that the state combination output by the unified multimodal generation model satisfies the relationships of chemical reaction kinetics, material balance, and energy conservation. The working condition fitting agent 12 performs joint optimization on the above two objectives, using the weighted sum of the loss terms corresponding to the two objectives as the overall loss function. It iteratively updates the parameters of the unified multimodal generation model through gradient backpropagation until the overall loss function converges to a pre-set threshold, resulting in the trained unified multimodal generation model.

[0048] After training the unified multimodal generation model, the working condition fitting agent 12 maps historical calibration data to the latent space using the modal encoder and coupled encoder of the trained unified multimodal generation model, forming a high-dimensional latent space distribution representing the evolution law of normal working conditions. The high-dimensional latent space distribution is generally represented as a high-density distribution manifold formed by the aggregation of mapping points corresponding to normal working condition samples in the latent space. The direction of the high-density distribution manifold is consistent with the material flow direction and energy exchange path of the chemical production unit. Any point on the high-density distribution manifold corresponds to a set of normal working condition states that satisfy the constraints of the chemical reaction mechanism. The transition between adjacent points on the high-density distribution manifold corresponds to the state evolution path of the chemical production unit under normal working conditions. For mapping points that deviate from the high-density distribution manifold, they correspond to abnormal states that deviate from normal working conditions. The direction and magnitude of deviation can be further decomposed according to a predetermined mechanism dimension.

[0049] After the above processing, the working condition fitting agent 12 completes the construction of the unified multimodal generation model and the formation of the high-dimensional latent space distribution. The resulting unified multimodal generation model and high-dimensional latent space distribution serve as a reference benchmark for subsequent mapping processing and deviation feature set calculation of the real-time multimodal spatiotemporal data field. On the other hand, they serve as the state evolution space carrier for performing risk evolution recursive analysis within the latent space, ensuring that subsequent risk identification and risk evolution analysis are carried out under the same high-dimensional latent space distribution, thus guaranteeing the consistency of state representation throughout the entire process.

[0050] Furthermore, in the working condition fitting agent 12, a unified multimodal generation model is constructed, including:

[0051] According to the process unit division rules of chemical production plants, the historical calibration data is divided into subsets corresponding to reaction units, transmission units and heat exchange units respectively, and a directional process topology constraint structure is constructed based on the material flow direction and energy exchange relationship between each unit.

[0052] Based on the aforementioned process topology constraint structure, a structural framework for a unified multimodal generation model is defined, ensuring that the connection relationships and information transmission directions between different state variables in the unified multimodal generation model are consistent with the material flow direction and energy exchange path. Furthermore, a hidden space sub-representation region corresponding to each unit is preset within the structural framework.

[0053] The reaction unit, the transport unit and the heat exchange unit are respectively introduced with corresponding mechanistic constraints. The reaction unit is introduced with rate constraints based on the reaction kinetic equation, the transport unit is introduced with cumulative constraints based on the material balance relationship, and the heat exchange unit is introduced with heat exchange constraints based on the energy conservation relationship. The mechanistic constraints are embedded into the structural framework in the form of state constraint terms.

[0054] Based on the structural framework of embedded state constraints, the unified multimodal generation model is trained using the historical calibration data. By jointly optimizing the data distribution fitting target and the mechanism constraint target, the unified multimodal generation model forms a high-dimensional latent space distribution that represents the evolution law of normal operating conditions, under the condition of satisfying the local mechanism constraints of each process unit and the overall process topology constraints.

[0055] In a preferred embodiment, when constructing a unified multimodal generation model, the working condition fitting agent 12 first divides the historical calibration data into subsets corresponding to reaction units, transmission units, and heat exchange units according to the process unit division rules of chemical production plants. The process unit division rules are based on the process function attributes of the chemical production plants. For example, reactors, reaction towers, and catalytic reactors, which perform chemical reactions, are classified into reaction units; pipes, material pumps, buffer tanks, and storage tanks, which perform material conveying and buffering functions, are classified into transmission units; and heat exchangers, coolers, furnaces, and condensers, which perform heat exchange and temperature regulation functions, are classified into heat exchange units. For composite devices that perform multiple process functions simultaneously, such as jacketed reactors and reaction towers with internal heat exchange coils, they are classified into corresponding process units according to their main process function attributes. Auxiliary process function identifiers are added to the process topology constraint structure for supplementation. The process unit division rules are consistent with the process flow topology model established based on the P&ID diagram in the digital plant model, and the process function attribute annotations of each device in the digital plant model can be directly used for process unit division. For historical calibration data, the operating condition fitting agent 12 directly aggregates the multimodal data slices of the reaction units corresponding to the process unit affiliation identifiers in the multimodal spatiotemporal data field into a reaction unit subset, the multimodal data slices of the transmission units corresponding to the process unit affiliation identifiers into a transmission unit subset, and the multimodal data slices of the heat exchange units corresponding to the process unit affiliation identifiers into a heat exchange unit subset. After completing the unit division of the historical calibration data, the operating condition fitting agent 12 constructs a directional process topology constraint structure based on the material flow direction and energy exchange relationship between each process unit. The process topology constraint structure is expressed in the form of a directed graph, with each process unit as the node of the graph and the material flow direction connection and energy exchange connection between each process unit as the directed edge of the graph. For the material flow direction relationship between each process unit, a corresponding material directed edge is established in the process topology constraint structure. The direction of the material directed edge is based on the actual flow direction of the material, that is, from the material output side to the material receiving side; taking the case of the reaction unit transporting reaction products to the downstream transmission unit as an example, the material directed edge points from the reaction unit node to the transmission unit node. For the energy exchange relationship between each process unit, corresponding directed energy edges are established in the process topology constraint structure. The direction of the directed energy edges is based on the actual heat flow direction, that is, from the heat release side to the heat absorption side. Taking the energy exchange relationship between the heat exchange unit and the reaction unit as an example, when the heat exchange unit inputs heat to the reaction unit, the directed energy edge points from the heat exchange unit node to the reaction unit node; when the reaction unit releases reaction heat to the heat exchange unit, the directed energy edge points from the reaction unit node to the heat exchange unit node.Each directed edge is assigned a connection attribute parameter, which includes transmission medium type, transmission delay, transmission capacity limit, energy exchange method, etc. The transmission delay parameter is consistent with the transmission delay parameter attached to the directional association index in the multimodal spatiotemporal data field, so that the process topology constraint structure and the multimodal spatiotemporal data field are consistent in the expression of cross-unit directional relationships.

[0056] Then, the working condition fitting agent 12 defines the structural framework of the unified multimodal generation model based on the process topology constraint structure. The structural framework definition process ensures that the connection relationships and information transmission directions between different state variables in the unified multimodal generation model are consistent with the material flow direction and energy exchange path in the process topology constraint structure. Specifically, the working condition fitting agent 12 uses the process topology constraint structure as the structural template for information flow within the latent space of the unified multimodal generation model. For each directed edge from the upstream node to the downstream node in the process topology constraint structure, a corresponding state variable information transmission path is established in the latent space of the unified multimodal generation model. For cross-node connections that do not exist in the process topology constraint structure, no corresponding information transmission path is established in the latent space of the unified multimodal generation model. This approach ensures that the internal information flow direction of the unified multimodal generation model is limited by the process topology structure, preventing the unified multimodal generation model from establishing pseudo-connections opposite to the actual process flow direction during training. After completing the directional definition of the structural framework, the working condition fitting agent 12 pre-defines the latent space sub-representation regions corresponding to each process unit within the structural framework. The latent space sub-representation region refers to a local representation sub-space divided by process unit in the latent space of the unified multimodal generation model. Each process unit node corresponds to a latent space sub-representation region, which is specifically responsible for the latent space representation of the multimodal data corresponding to that process unit. The working condition fitting agent 12 reserves non-overlapping sub-spaces for the reaction unit, transport unit, and heat exchange unit in the latent space of the unified multimodal generation model. That is, the latent space sub-representation regions corresponding to each process unit occupy different dimensional index ranges in the overall latent space dimension, and there is no dimensional overlap between them. The output value range of each latent space sub-representation region is limited by selecting activation functions with bounded output amplitude characteristics, such as tanh activation function and sigmoid activation function. For the distribution shape of each latent space sub-representation region, the standard normal distribution is used as the prior distribution, and during training, the KL divergence term is used to constrain the actual distribution of each latent space sub-representation region to approach the standard normal distribution. The connection relationships between latent space sub-representation regions are established according to the directed edge relationships defined by the process topology constraint structure. The latent space sub-representation region of the upstream process unit transmits state information unidirectionally to the latent space sub-representation region of the downstream process unit. This is achieved by setting up a unidirectional information transmission layer between two adjacent sub-representation regions. The unidirectional information transmission layer is implemented through a unidirectional fully connected layer or an attention layer with an additional directional mask: the unidirectional fully connected layer only retains the weight connections from upstream to downstream, and the weights of the reverse connections are reset to zero and do not participate in gradient updates during training; the attention layer with an additional directional mask shields the reverse attention weights through a mask matrix during attention calculation, so that the attention signal only propagates along the upstream to downstream direction.

[0057] Next, the working condition fitting agent 12 introduces corresponding mechanistic constraints to the reaction unit, transmission unit and heat exchange unit respectively, and embeds the mechanistic constraints into the structural framework in the form of state constraint terms.

[0058] For the reaction unit, the operating condition fitting agent 12 introduces rate constraints based on the reaction kinetic equation. The reaction kinetic equation expresses the correspondence between the reaction rate and the reactant state in terms of reaction rate, reaction rate constant, reactant concentration, and reaction order. That is, the reaction rate is obtained by multiplying the reaction rate constant by the concentration of each reactant by a power of the reaction order. The reaction rate constant increases with increasing reaction temperature, and the correspondence between the reaction rate constant and the reaction temperature is given by the Arrhenius equation and is determined by the pre-exponential factor, activation energy, gas constant, and reaction temperature. The working condition fitting agent 12 treats reaction rate, reaction temperature, and concentrations of each reactant as state variables carried by the latent space sub-representation region of the reaction unit. It rewrites the reaction kinetic equation as state constraint terms for the reaction unit: specifically, by shifting terms on both sides of the reaction kinetic equation and subtracting to zero on the right-hand side, the resulting expression on the left-hand side is the construction form of the state constraint terms for the reaction unit. These state constraint terms are defined as the difference between the actual reaction rate value output by the unified multimodal generation model and the theoretical reaction rate value calculated based on the reaction rate constant and reactant concentrations according to the reaction kinetic equation. When the state variables of the reaction unit output by the unified multimodal generation model satisfy the reaction kinetic equation, the value of the state constraint terms is close to zero; when the state variables deviate from the reaction kinetic equation, the value of the state constraint terms increases significantly. The state constraint terms are embedded in the latent space sub-representation region corresponding to the reaction unit and the reaction unit output branch of the modal decoder in the form of residual terms, serving as the mechanistic constraint quantity of the reaction unit.

[0059] For the transmission unit, the working condition fitting agent 12 introduces cumulative constraints based on the material balance relationship. The material balance relationship is expressed as the conservation relationship between the inlet material flow rate, outlet material flow rate, and the accumulated material within the transmission unit. That is, the rate of change of the accumulated material within the transmission unit at any given time is equal to the difference between the inlet material flow rate and the outlet material flow rate at the same time. The working condition fitting agent 12 treats the accumulated material, inlet material flow rate, and outlet material flow rate as state variables carried by the latent space sub-representation region of the transmission unit, and rewrites the material balance relationship as a state constraint term for the transmission unit: Specifically, by shifting terms on both sides of the material balance relationship and subtracting them to make the right side of the equation zero, the expression on the left side is the construction form of the state constraint term for the transmission unit. The state constraint term for the transmission unit is defined as the difference between the actual value of the rate of change of the accumulated material output by the unified multimodal generation model and the theoretical value of the rate of change of the accumulated material calculated based on the material balance relationship according to the inlet and outlet material flow rates. When the state variables of the transmission unit output by the unified multimodal generation model satisfy the material balance relationship, the value of the state constraint term for the transmission unit is close to zero. The state constraint term of the transmission unit is embedded in the latent space sub-representation region corresponding to the transmission unit and the output branch of the transmission unit of the modal decoder in the form of residual term, as the mechanistic constraint quantity of the transmission unit.

[0060] For the heat exchange unit, the operating condition fitting agent 12 introduces heat exchange constraints based on the energy conservation relationship. The energy conservation relationship is expressed by the flow rate, inlet temperature, outlet temperature, and specific heat capacity of the medium on the hot and cold sides of the heat exchange unit. That is, the heat released by the hot side medium is equal to the heat absorbed by the cold side medium. The heat released or absorbed by the medium on either side is given by the product of the flow rate, specific heat capacity, and the difference between the inlet and outlet temperatures of that side medium. The working condition fitting agent 12 treats the flow rate, inlet temperature, and outlet temperature of the media on both the hot and cold sides as state variables carried by the latent space sub-representation region of the heat exchange unit. It rewrites the energy conservation relation as state constraint terms for the heat exchange unit: specifically, by shifting and subtracting terms from both sides of the energy conservation relation to make the right-hand side zero, the resulting expression on the left-hand side is the construction form of the state constraint terms for the heat exchange unit. These state constraint terms are defined as the difference between the heat released by the hot-side medium and the heat absorbed by the cold-side medium, as output by the unified multimodal generation model. When the state variables of the heat exchange unit output by the unified multimodal generation model satisfy the energy conservation relation, the value of the state constraint terms is close to zero. The state constraint terms are embedded in the latent space sub-representation region corresponding to the heat exchange unit and the heat exchange unit output branch of the modal decoder in the form of residual terms, serving as the mechanistic constraint quantity of the heat exchange unit.

[0061] Subsequently, the operating condition fitting agent 12, based on the structural framework of embedded state constraints, trains the unified multimodal generative model using historical calibration data. The optimization objectives of the training process include two objectives: a data distribution fitting objective and a mechanism constraint objective. The data distribution fitting objective aims to minimize the reconstruction error between the reconstructed sensing data output by the modal decoder and the input sensing data. The mechanism constraint objective aims to minimize the weighted sum of the state constraints of the reaction unit, the transmission unit, and the heat exchange unit. The operating condition fitting agent 12 performs joint optimization on the two objectives. The overall loss function is obtained by adding the reconstruction loss term corresponding to the data distribution fitting objective to the weighted residual terms of the three mechanism constraint terms. Each of the three mechanism constraint terms corresponds to a weighting coefficient, which is set according to the importance of each process unit in the overall risk contribution of the chemical production plant. Typical values ​​range from 0.1 to 1.0, and can be specified by process experts based on the actual conditions of the plant. As an example, the weighting coefficient for the reaction unit is 0.6, for the transmission unit it is 0.3, and for the heat exchange unit it is 0.4. The working condition fitting agent 12 uses gradient backpropagation to iteratively optimize the overall loss function. To maintain the predefined directional connectivity in the aforementioned structural framework, the weight gradients corresponding to the reverse connections between the latent space sub-representation regions are forcibly set to zero or the corresponding weights are frozen and not updated during gradient backpropagation, ensuring that parameter updates only occur on the connection weights in the upstream-to-downstream direction. Simultaneously, the mechanistic constraint terms corresponding to each process unit participate in the gradient composition of the overall loss function, ensuring that parameter updates always move in the direction of minimizing the residuals of the three mechanistic constraint terms. These two constraints work together to ensure that the unified multimodal generation model satisfies both the local mechanistic constraints of each process unit and the overall process topology constraints. When the overall loss function converges to a pre-set threshold or the number of iterations reaches a pre-set upper limit, the training process ends, and a trained unified multimodal generative model is obtained. The historical calibration data is mapped to the latent space through the trained unified multimodal generative model to form a high-dimensional latent space distribution that represents the evolution law of normal working conditions.

[0062] Through the above processing, the working condition fitting agent 12 completes the definition of the unified multimodal generation model structure framework under the directional specification of the process topology constraint structure, completes the embedding of state constraint terms under the physical specification of the mechanism constraint of the sub-process unit, and completes the learning of model parameters under the data-driven joint optimization training, so that the obtained unified multimodal generation model has the ability to fit data distribution, the feasibility of local mechanism and the consistency of overall process topology.

[0063] Furthermore, in the operating condition fitting agent 12, a high-dimensional latent space distribution characterizing the evolution law of normal operating conditions is formed, including:

[0064] Using the latent space sub-representation region corresponding to each process unit as the basic unit, a set of state samples under normal operating conditions is constructed based on historical calibration data. Density clustering constraints are applied to the mapping result of the state sample set in the latent space to form a high-density distribution manifold of the stable operating state of the corresponding process unit within each latent space sub-representation region.

[0065] In the process of constructing the high-density distributed manifold, cross-unit consistency constraints based on process topology constraints are introduced to enable the latent space sub-representation regions of adjacent process units to form a continuous state transition relationship in the material flow direction and energy exchange direction, and to construct a directed state evolution manifold structure that conforms to the actual process flow in the overall latent space.

[0066] Based on mechanistic constraints, the gradient of each state variable in the directed state evolution manifold structure is constrained directionally, limiting the feasible direction of the state evolution path in the latent space, so that any state transition in the latent space satisfies the relationship of reaction kinetics, material balance and energy conservation, forming a set of latent space evolution trajectories with physical feasibility.

[0067] By combining the set of latent space evolution trajectories and applying sparse distribution constraints and directional decomposition constraints to samples that deviate from the high-density distribution manifold, the abnormal state is made to appear as a low-density offset distribution unfolding along a predetermined mechanism dimension in the latent space, thereby generating a structured high-dimensional latent space distribution that simultaneously includes the normal operating condition manifold and the direction of abnormal disturbance.

[0068] In a preferred embodiment, the process of forming a high-dimensional latent space distribution representing the evolution law of normal operating conditions in the working condition fitting agent 12 includes four processing steps: high-density distribution manifold construction, directed state evolution manifold structure construction, latent space evolution trajectory set formation, and structured high-dimensional latent space distribution generation. These four processing steps sequentially superimpose constraints on the previously trained unified multimodal generation model, gradually shaping the mapping results of historical calibration data in the latent space from a scattered distribution into a high-dimensional latent space distribution with structured characteristics.

[0069] In the aforementioned structural framework definition phase, the operating condition fitting agent 12 has pre-reserved a mechanistic residual direction dimension subspace in the latent space sub-representation region corresponding to each process unit. Specifically, the latent space sub-representation region corresponding to each process unit consists of two parts: a main dimension subspace and a mechanistic residual direction dimension subspace. The main dimension subspace is used to carry the state characteristics of the process unit under normal operating conditions, while the mechanistic residual direction dimension subspace is used to carry the deviation information of the mechanistic constraint residual direction of the process unit relative to normal operating conditions. The number of dimensions of the mechanistic residual direction dimension subspace is typically 3 to 5, corresponding to the mechanistic constraint residual directions involved in the process unit. Specifically, the mechanistic residual direction dimension subspace corresponding to the reaction unit carries the reaction kinetic residual direction, the mechanistic residual direction dimension subspace corresponding to the transport unit carries the material balance residual direction, and the mechanistic residual direction dimension subspace corresponding to the heat exchange unit carries the energy conservation residual direction. The mechanistic residual direction dimension subspace is distinguished from the main dimension subspace in the latent space sub-representation region in terms of dimension index, and there is no dimension overlap. The mechanism residual directional dimension subspace is the predetermined mechanism dimension in the subsequent description.

[0070] First, the operating condition fitting agent 12 uses the latent space sub-representation region corresponding to each process unit as the basic unit and constructs a set of state samples under normal operating conditions based on historical calibration data. Specifically, the operating condition fitting agent 12 inputs each multimodal data slice from the historical calibration data into the modal encoder and coupling encoder of the trained unified multimodal generation model according to its process unit affiliation identifier. After processing by the encoder, the mapping results of each multimodal data slice on the latent space sub-representation region of the corresponding process unit are obtained. All mapping results corresponding to the same process unit are aggregated to form the normal operating condition state sample set of that process unit. Each mapping result is a sample point on the corresponding latent space sub-representation region, and the coordinates of the sample point are given by the value of the dimension occupied by the latent space sub-representation region. After constructing the normal operating condition state sample set, the operating condition fitting agent 12 performs density clustering constraints on the mapping results of the state sample set in the latent space, so that a high-density distribution manifold of the stable operating state of the corresponding process unit is formed in each latent space sub-representation region. Specifically, the operating condition fitting agent 12 superimposes a density clustering loss term into the training loss function of the unified multimodal generative model. This density clustering loss term aims to reduce the average distance between sample points within the same process unit's state sample set; that is, the loss term increases as the average distance between sample points increases. Simultaneously, the operating condition fitting agent 12 typically uses Gaussian kernel density estimation to estimate the local density at each sample point in the latent space sub-representation region. The increase in local density at each sample point is used as the density clustering quantification index, and the negative values ​​of the local density at each sample point are included as an additional contribution to the density clustering loss term in the loss function. Under the constraint of density clustering, the state samples under normal operating conditions of the same process unit continuously converge towards the high-density center within the corresponding latent space sub-representation region, ultimately forming a compact, continuous, and high-density distribution structure—the high-density distribution manifold of the process unit. The high-density distribution manifold occupies a dimension far lower than the dimension of the latent space sub-representation region itself, exhibiting the characteristic of a low-dimensional manifold embedded in a high-dimensional space.

[0071] Then, based on the high-density distributed manifold construction, the working condition fitting agent 12 further introduces cross-unit consistency constraints based on process topology constraints, so that the hidden space sub-representation regions of adjacent process units form a continuous state transition relationship in the material flow direction and energy exchange direction. The implementation method of cross-unit consistency constraint is as follows: The working condition fitting agent 12 applies consistency constraints to the correspondence between the high-density distributed manifolds of adjacent process units based on the material flow direction and energy exchange path recorded by each directed edge in the aforementioned process topology constraint structure. Specifically, for each directed edge in the process topology constraint structure from the upstream process unit node to the downstream process unit node, the working condition fitting agent 12 takes the state sample point of the upstream process unit at a given time in the same production process and the state sample point of the downstream process unit at the time offset by the transmission delay parameter attached to the directed edge as a pair of paired samples. The upstream sample point is input into the aforementioned one-way information transmission layer to obtain the predicted downstream sample point position output by the one-way information transmission layer. The Euclidean distance between the predicted downstream sample point position and the actual paired downstream sample point position in the latent space sub-representation region of the downstream process unit is calculated. The sum of the squares of the obtained Euclidean distances is used as the value of the cross-unit consistency loss term, and this loss term is superimposed in the training loss function. The cross-unit consistency loss term drives the parameter update of the unidirectional information transfer layer during gradient backpropagation, aligning the predicted downstream sample point positions with the actual paired downstream sample point positions. Under the cross-unit consistency constraint, the high-density distribution manifolds of adjacent process units are no longer isolated from each other, but are continuously connected along the material flow and energy exchange directions, constructing a directed state evolution manifold structure that conforms to the actual process flow in the overall latent space. In the directed state evolution manifold structure, the high-density distribution manifolds of each process unit serve as local nodal manifolds, interconnected along the process topology direction; the overall directionality of the manifold structure is consistent with the material flow and energy exchange path of the chemical production unit.

[0072] Next, the working condition fitting agent 12 applies directional constraints to the gradient of each state variable in the directional state evolution manifold structure based on the three types of mechanistic constraints of the aforementioned embedded structural framework: the state constraint terms of the reaction unit corresponding to the reaction kinetic equation, the state constraint terms of the transport unit corresponding to the material balance relationship, and the state constraint terms of the heat transfer unit corresponding to the energy conservation relationship. Specifically, for any state transition process in a directed state evolution manifold structure, i.e., the transition path between any two adjacent sample points in the latent space, the working condition fitting agent 12 calculates the gradient of the change of each state variable along the time direction of the transition path, and substitutes the gradient into the state constraint term expression of the corresponding process unit to obtain the mechanism constraint residual corresponding to the state transition process. The excess between the mechanism constraint residual and the pre-set mechanism constraint residual threshold is used as the value of the directional gradient constraint loss term. That is, when the mechanism constraint residual value is within the pre-set threshold range, the loss term is zero, and the part exceeding the threshold is included in the loss term by square value. The working condition fitting agent 12 superimposes the directional gradient constraint loss term into the training loss function. During the training process, the parameter update is carried out in the direction of reducing the value of the directional gradient constraint loss term, i.e., in the direction of ensuring that the mechanism constraint residual corresponding to the state transition path output by the unified multimodal generation model does not exceed the pre-set threshold, so that the model after training no longer generates state transition paths that exceed the mechanism constraint residual threshold. The aforementioned directional constraint treatment ensures that any state transition in the latent space must satisfy reaction kinetics, material balance, and energy conservation relationships, forming a set of physically feasible latent space evolution trajectories. Each trajectory in the latent space evolution trajectory set corresponds to a state transition path that starts from the initial normal operating state, evolves along the process topology, and satisfies the three types of mechanistic constraints.

[0073] Subsequently, the working condition fitting agent 12, combined with the latent space evolution trajectory set, applies sparse distribution constraints and directional decomposition constraints to samples deviating from the high-density distribution manifold, making the abnormal state appear as a low-density offset distribution unfolding along a predetermined mechanism dimension in the latent space. During training, samples deviating from the high-density distribution manifold naturally arise in two situations: first, samples whose mapping points fall outside the high-density distribution manifold after mapping from historical calibration data via modal encoder and coupled encoder; second, samples generated during model training whose mechanism constraint residuals corresponding to their reconstruction results exceed a pre-set threshold. All of these deviation samples are uniformly treated as objects of sparse distribution constraints and directional decomposition constraints. The implementation method of sparse distribution constraint is as follows: the working condition fitting agent 12 superimposes a sparse distribution loss term into the training loss function. The sparse distribution loss term aims to make the distribution of deviation samples as sparse as possible in the predetermined mechanism dimension. That is, the deviation samples only show non-zero offsets in a few dimensions in the predetermined mechanism dimension, while maintaining values ​​close to zero in other dimensions. Specifically, L1 norm regularization is used to apply sparsity constraints to the offset vector of the deviation samples in the predetermined mechanism dimension, that is, the sum of the absolute values ​​of each component of the offset vector is used as the value of the sparse distribution loss term. The implementation method of directional decomposition constraints is as follows: For each deviation sample point that deviates from the high-density distribution manifold, the working condition fitting agent 12 performs component decomposition on its deviation vector along a predetermined mechanism dimension, decomposing the deviation vector into three components corresponding to the reaction kinetic residual, material balance residual, and energy conservation residual, respectively. At the same time, the working condition fitting agent 12 substitutes the multimodal state variable corresponding to the deviation sample point into the reaction unit state constraint term, the transport unit state constraint term, and the heat exchange unit state constraint term, and calculates the actual values ​​of the mechanism residuals of the deviation sample point in the three mechanism dimensions, respectively. The working condition fitting agent 12 superimposes the directional decomposition loss term into the training loss function. The directional decomposition loss term uses the difference between the three decomposed components and the actual values ​​of the corresponding mechanism residuals as the loss term value, so that the values ​​of each decomposed component in the corresponding mechanism residual direction match the actual values ​​of the mechanism residuals calculated by the corresponding state constraint term, and the values ​​in other mechanism residual directions are close to zero. Under the combined effect of sparse distribution constraints and directional decomposition constraints, samples that deviate from the high-density distribution manifold are manifested in the latent space as a sparse, low-density offset distribution that unfolds along the predetermined mechanism dimension and can be decomposed into components according to the mechanism dimension. Furthermore, each component in the offset direction maintains an interpretable correspondence with the specific mechanism violation type.

[0074] After the above four processing steps are superimposed with constraints, the working condition fitting agent 12 finally generates a structured high-dimensional latent space distribution that simultaneously includes the normal working condition manifold and the direction of abnormal disturbances. The structured high-dimensional latent space distribution has the following three structural features: First, the latent space sub-representation region corresponding to each process unit carries a high-density distribution manifold that characterizes the stable operating state of the process unit, which is used to characterize the normal working condition; Second, the high-density distribution manifolds of adjacent process units are connected to each other along the process topology direction, forming a directed state evolution manifold structure that runs through the entire latent space, and any state transition within this manifold structure satisfies the three types of mechanism constraints, constituting a set of latent space evolution trajectories with physical feasibility; Third, abnormal states that deviate from the high-density distribution manifold are sparse, low-density offset distributions that can be decomposed according to the mechanism type along the pre-reserved predetermined mechanism dimension, so that abnormal deviations can be quantitatively identified and can be explained and decomposed according to the mechanism dimension.

[0075] The structured high-dimensional latent space distribution, as the final output of the unified multimodal generation model, provides a structured reference benchmark for subsequent mapping processing and deviation feature set calculation of real-time multimodal spatiotemporal data fields. This enables subsequent processing to determine whether there is a risk by comparing the degree of deviation between the mapping point and the high-density distribution manifold, to obtain the contribution of the risk source in each mechanism dimension by decomposing the deviation vector along the predetermined mechanism dimension, and to predict the risk evolution path by extrapolating the state evolution of the directed state evolution manifold structure.

[0076] Risk identification agent 13 is used to map the real-time acquired multimodal spatiotemporal data field to a high-dimensional latent space distribution. After calculating the deviation feature set, it performs interpretability decomposition of the deviation feature set to generate a multidimensional risk perturbation vector.

[0077] Specifically, the risk identification agent 13 receives the multimodal spatiotemporal data field output in real time by the interactive perception module 11 during the operation phase. Based on the unified multimodal generation model and the high-dimensional latent space distribution that have been trained by the aforementioned working condition fitting agent 12, it performs risk identification processing on the working condition state carried in the real-time multimodal spatiotemporal data field and outputs a multidimensional risk disturbance vector, which serves as the basis for the subsequent evolution identification agent 14 to perform risk evolution recursive analysis and the early warning and alarm management module 15 to perform early warning and alarm management.

[0078] First, the risk identification agent 13 maps the real-time acquired multimodal spatiotemporal data field to a high-dimensional latent space distribution. Specifically, the risk identification agent 13 uses the multimodal data slices continuously output by the interactive sensing module 11 during operation as the processing object. According to the process unit affiliation identifier, each multimodal data slice is input into the modal encoder and coupling encoder of the trained unified multimodal generation model. After processing by the encoder, the mapped state vector of each multimodal data slice on the corresponding process unit's latent space sub-representation region is obtained. The mapped state vector is a vector point on the corresponding latent space sub-representation region, and its coordinates are given by the value of the dimension occupied by the latent space sub-representation region. Following the sampling rhythm of the real-time multimodal spatiotemporal data field, the risk identification agent 13 continuously outputs the mapped state vectors of each process unit at the same sampling rhythm, forming a continuous real-time mapped state vector sequence along the time axis.

[0079] Then, the risk identification agent 13 calculates the deviation feature set of the real-time mapped state vector relative to the normal operating condition high-density distribution manifold in the high-dimensional latent space distribution. For each real-time mapped state vector, the risk identification agent 13 determines the nearest neighbor point of the mapped state vector to the normal operating condition high-density distribution manifold within the latent space sub-representation region of the corresponding process unit. There are two ways to determine the nearest neighbor point: one is the k-nearest neighbor search method, which uses the set of mapped sample points of historical calibration data on the normal operating condition high-density distribution manifold as the candidate set, and retrieves the sample point with the smallest Euclidean distance to the current mapped state vector as the nearest neighbor point; the other is the gradient descent method, which uses the distance of the current mapped state vector to the normal operating condition high-density distribution manifold as the objective function, and iteratively adjusts the position of the mapped state vector along the gradient descent direction of the objective function. The termination condition of the iteration adjustment is that the objective function converges to a pre-set convergence threshold, and the convergence position is taken as the nearest neighbor point. The difference vector between the mapped state vector and its nearest neighbor is used as the deviation vector corresponding to the mapped state vector. The deviation vector contains two types of information: deviation magnitude and deviation direction. The deviation magnitude is measured by the magnitude of the deviation vector, and the deviation direction is composed of the components of the deviation vector in each dimension of the latent space sub-representation region. The deviation vectors generated by each process unit at the same time are organized into a deviation feature set for that time time according to the process unit's attribution identifier.

[0080] The retention and discrimination of each deviation vector in the deviation feature set is determined by a pre-set deviation threshold. The deviation threshold is obtained by performing quantile statistics on historical calibration data. Specifically, after mapping the historical calibration data to a high-dimensional latent space distribution via the encoder, the deviation amplitude distribution of each sample point to the high-density distribution manifold under normal operating conditions is statistically analyzed. The 95th or 99th quantile of this distribution is taken as the deviation threshold, which can be further modified by process experts based on the actual situation of the equipment. For deviation vectors with a deviation amplitude less than the deviation threshold, it is determined that the process unit is within the normal operating range at that moment, and the corresponding deviation vector is represented as a zero vector in the deviation feature set. For deviation vectors with a deviation amplitude greater than or equal to the deviation threshold, it is determined that the process unit has an operating condition deviation at that moment, and the corresponding deviation vector is retained in the deviation feature set for subsequent processing.

[0081] Next, the risk identification agent 13 performs interpretability decomposition on the deviation feature set to generate a multidimensional risk disturbance vector. Each deviation vector in the deviation feature set is presented as a geometric vector in the latent space. The deviation vector itself cannot directly indicate the process unit affiliation or mechanism dimension affiliation of the operating condition deviation. Based on the structured division of the latent space sub-representation regions of each process unit in the aforementioned high-dimensional latent space distribution and the pre-reserved mechanism residual direction dimension sub-space, the risk identification agent 13 partitions each deviation vector in the deviation feature set according to the process unit boundary, decomposes it according to the mechanism residual direction, and reconstructs it according to the process topology direction. This transforms the deviation information, originally existing in the geometric form of the latent space, into component representations organized according to process units and mechanism dimensions, resulting in a multidimensional risk disturbance vector. Each component in the multidimensional risk disturbance vector corresponds to a specific process unit and a specific mechanism dimension. The component value represents the risk disturbance amplitude of that process unit in that mechanism dimension, and the component sign represents the disturbance direction.

[0082] Through the above processing, the risk identification agent 13 outputs a structured multidimensional risk disturbance vector at each sampling time. The multidimensional risk disturbance vector serves as the criterion for the early warning module 15 to perform early warning management, and as the initial disturbance input for the evolution identification agent 14 to perform risk evolution recursive analysis in the latent space. This enables subsequent risk evolution analysis and early warning processing to obtain the risk disturbance distribution at the current time according to the process unit and mechanism dimensions.

[0083] Furthermore, the risk identification agent 13 generates a multi-dimensional risk perturbation vector, including:

[0084] The real-time acquired multimodal spatiotemporal data field is input into the unified multimodal generation model to obtain the mapped state vector in the high-dimensional latent space. Based on the normal working condition high-density distribution manifold corresponding to the high-dimensional latent space distribution, the minimum deviation path of the mapped state vector relative to the high-density distribution manifold is calculated to form an initial deviation feature representation including the deviation magnitude and deviation direction.

[0085] The initial deviation feature representation is subjected to structured decomposition processing. The deviation direction is partitioned and projected according to the preset latent space sub-representation region in the unified multimodal generation model. Based on the mechanism constraints corresponding to each process unit, the initial deviation feature representation is decomposed into multiple disturbance components with physical meaning in the subspaces corresponding to the reaction unit, the transmission unit and the heat exchange unit, respectively, to obtain local risk disturbance sub-vectors associated with different process units.

[0086] Based on the process topology constraint structure, cross-unit propagation correlation analysis is performed on each local risk disturbance sub-vector. According to the material flow direction and energy exchange path, the transmission weight and coupling influence of each disturbance component between adjacent process units are calculated, and the local risk disturbance sub-vectors are weighted and reconstructed to form a coupled disturbance representation that characterizes the risk propagation relationship.

[0087] The coupled disturbance representation is subjected to direction correction and amplitude normalization, and the corrected disturbance components are combined and encoded according to a preset mechanism dimension to generate an interpretable multidimensional risk disturbance vector.

[0088] In a preferred embodiment, firstly, the risk identification agent 13 inputs the real-time collected multimodal spatiotemporal data field into the aforementioned trained unified multimodal generation model. It then obtains the mapped state vector and the deviation vector relative to the high-density distribution manifold under normal operating conditions using the aforementioned mapping and minimum deviation path calculation method. The deviation amplitude and deviation direction in the deviation vector are then organized together to form an initial deviation feature representation. This initial deviation feature representation is expressed in binary form. The first component of the binary tuple is the deviation amplitude, measured by the magnitude of the deviation vector; the second component is the deviation direction vector, composed of the components of the deviation vector in each dimension of the latent space sub-representation region of the corresponding process unit. The initial deviation feature representation is organized according to the attribution identifier of each process unit. Each process unit generates a corresponding initial deviation feature representation at each sampling time, which serves as the input object for subsequent structured decomposition processing.

[0089] Then, the risk identification agent 13 performs structured decomposition processing on the initial deviation feature representation. The structured decomposition processing includes partitioned projection and mechanism direction decomposition. In partitioned projection, the risk identification agent 13 projects the deviation direction component vector in the initial deviation feature representation according to the pre-defined latent space sub-representation regions in the unified multimodal generation model. In the aforementioned structural framework, latent space sub-representation regions with non-overlapping dimensional index ranges are reserved for the reaction unit, transmission unit, and heat exchange unit, respectively. Based on the dimensional index range occupied by the latent space sub-representation regions of each process unit, the risk identification agent 13 extracts the components of the deviation direction component vector that fall within the dimensional index range of the reaction unit as reaction unit deviation component vectors, the components that fall within the dimensional index range of the transmission unit as transmission unit deviation component vectors, and the components that fall within the dimensional index range of the heat exchange unit as heat exchange unit deviation component vectors. After partitioned projection processing, the deviation direction component vector of the original overall organization is separated into three deviation component vectors that fall into the latent space sub-representation regions of the three types of process units, respectively. In the mechanism direction decomposition, the risk identification agent 13, based on the mechanism constraints corresponding to each process unit, further decomposes the deviation component sub-vectors along predetermined mechanism dimensions within the latent space sub-representation region where each deviation component sub-vector resides. In the aforementioned structural framework, a mechanism residual direction dimension sub-space is pre-reserved within the latent space sub-representation region of each process unit. Each dimension in the mechanism residual direction dimension sub-space corresponds to the mechanism constraint residual direction involved in the process unit, where the reaction unit corresponds to the reaction kinetics residual direction, the transport unit corresponds to the material balance residual direction, and the heat exchange unit corresponds to the energy conservation residual direction. The risk identification agent 13 projects each deviation component sub-vector onto the mechanism residual direction dimension sub-space of its latent space sub-representation region, obtaining the projection components of the deviation component sub-vectors on each mechanism residual direction dimension. Each projection component is a physically meaningful perturbation component. The value of the perturbation component represents the perturbation amplitude of the process unit in the corresponding mechanism residual direction, and the sign of the perturbation component represents the perturbation direction. After mechanism-oriented decomposition, the deviation component sub-vectors corresponding to each process unit are decomposed into multiple disturbance components corresponding to different mechanism residual directions. All disturbance components of each process unit are collectively organized into the local risk disturbance sub-vector corresponding to that process unit. After the above-mentioned partitioned projection and mechanism-oriented decomposition, the risk identification agent 13 obtains three local risk disturbance sub-vectors associated with the reaction unit, the transmission unit, and the heat exchange unit, respectively. Each local risk disturbance sub-vector organizes its disturbance components according to the predetermined mechanism dimension of the corresponding process unit.

[0090] Next, the risk identification agent 13 performs cross-unit propagation correlation analysis on each local risk disturbance sub-vector based on the process topology constraint structure. The cross-unit propagation correlation analysis models the propagation relationship between the three local risk disturbance sub-vectors. In a chemical production unit, there is a coupling and transmission relationship between process units along the material flow direction and energy exchange path. A risk disturbance occurring in one process unit will be propagated along the process topology direction to the downstream process unit and have a coupling effect on it. Relying solely on the local risk disturbance sub-vectors within each process unit is insufficient to fully express the propagation relationship of risk disturbances in the overall process flow. The implementation method of cross-unit propagation correlation analysis is as follows: The risk identification agent 13 uses the material flow direction and energy exchange path recorded by each directed edge in the process topology constraint structure as a basis. For each directed edge pointing from the upstream process unit node to the downstream process unit node, it determines the transmission weight and coupling influence degree of that directed edge. The transmission weight is defined by the contribution intensity of the local risk disturbance sub-vector of the upstream process unit to the disturbance of the downstream process unit. The value of the transmission weight is directly extracted from the connection weight parameters learned by the unidirectional information transmission layer after training. Specifically, for a unidirectional information transmission layer implemented by a fully connected layer with unidirectional connections, the average absolute value of each element of the weight matrix in the upstream-to-downstream direction of the fully connected layer is taken as the transmission weight of the directed edge; for a unidirectional information transmission layer implemented by an attention layer with an additional directional mask, the average value of each element of the attention weight in the upstream-to-downstream direction of the attention layer is taken as the transmission weight of the directed edge. The degree of coupling influence is characterized by the proportion of disturbance retention during the transmission of disturbance from the upstream process unit to the downstream process unit along the directed edge. The value of the degree of coupling influence is determined by the connection attribute parameters recorded by the corresponding directed edge in the process topology constraint structure. Specifically, for a material directed edge, the degree of coupling influence is equal to the material mixing ratio or material transmission ratio recorded by the directed edge, reflecting the proportion of material from the upstream unit flowing into the downstream unit relative to the total material in the downstream unit; for an energy directed edge, the degree of coupling influence is equal to the normalized product of the heat exchange efficiency or heat transfer coefficient recorded by the directed edge and the heat transfer area, reflecting the proportion of heat released by the upstream unit being absorbed by the downstream unit. The degree of coupling influence ranges from 0 to 1. The closer the value is to 1, the higher the degree of retention of the disturbance during the propagation process along the directed edge.

[0091] After obtaining the transmission weights and coupling influence levels on each directed edge, the risk identification agent 13 performs weighted reconstruction processing on each local risk disturbance sub-vector. The weighted reconstruction processing method is as follows: for the local risk disturbance sub-vector of each process unit, the risk identification agent 13 takes the process unit as the convergence node, and according to each upstream directed edge pointing to the process unit in the process topology constraint structure, it performs weighted transmission of the local risk disturbance sub-vector of each upstream process unit according to the product of the transmission weight and coupling influence level on the corresponding directed edge, and superimposes the weighted transmission result with the local risk disturbance sub-vector of the process unit itself according to the fusion ratio [α, 1-α], where α is the weight of the local risk disturbance sub-vector of the process unit itself, 1-α is the weight of the upstream transmission weighted result, and the value of α is between 0.5 and 0.8, typically 0.7, that is, the process unit itself is the main disturbance, and the coupling transmission from the upstream is the auxiliary disturbance, to obtain the corresponding coupled disturbance sub-vector of the process unit. The coupled perturbation subvector represents the component associated with the process unit in the coupled perturbation representation. The coupled perturbation subvectors corresponding to each process unit together constitute the coupled perturbation representation characterizing the risk propagation relationship. The coupled perturbation representation retains the risk perturbation information of each process unit itself, and also integrates the coupled perturbation information propagated from upstream process units along the process topology in a structured manner.

[0092] Subsequently, the risk identification agent 13 performs direction correction and amplitude normalization on the coupled disturbance representation, and combines and encodes the corrected disturbance components according to a preset mechanism dimension to generate an interpretable multidimensional risk disturbance vector. During direction correction, the risk identification agent 13 uniformly corrects the signs of each disturbance component in the coupled disturbance representation according to a predefined mechanism residual sign convention, ensuring that the disturbance components in different process units and different mechanism dimensions of the multidimensional risk disturbance vector adopt a unified sign meaning. Specifically, the mechanism residual sign convention is as follows: for the reaction kinetics residual direction, a reaction rate higher than the theoretical value is positive, and a rate lower than the theoretical value is negative; for the material balance residual direction, a material accumulation rate change rate higher than the theoretical value is positive, and a rate lower than the theoretical value is negative; for the energy conservation residual direction, a heat release on the hot side is higher than the heat absorbed on the cold side is positive, and a rate lower than the heat absorbed on the cold side is negative. After direction correction, the signs of each disturbance component maintain a consistent physical meaning across different process units and at different times. During amplitude normalization, the risk identification agent 13 performs normalization processing on the amplitude of each disturbance component in the coupled disturbance representation. The normalization benchmark adopts the mechanism constraint residual threshold in the corresponding mechanism residual direction, that is, the ratio between the amplitude of each disturbance component and the corresponding mechanism constraint residual threshold is used as the normalized disturbance component amplitude. The normalization processing makes the amplitude of each disturbance component have uniform dimensions and comparability. A normalized amplitude greater than or equal to 1 indicates that the mechanism violation degree corresponding to the disturbance component has reached or exceeded the threshold level, and a normalized amplitude less than 1 indicates that the mechanism violation degree corresponding to the disturbance component is still within the threshold. Subsequently, the risk identification agent 13 combines and encodes each disturbance component after direction correction and amplitude normalization processing according to a preset mechanism dimension. Specifically, using the process unit as the primary index and the mechanism residual direction as the secondary index, each disturbance component is encoded as a quadruple of "process unit identifier + mechanism residual direction identifier + normalized amplitude + direction sign". All quadruples are arranged in the order of the primary and secondary indices to form the final output multidimensional risk disturbance vector. Each component in the multidimensional risk disturbance vector corresponds to a specific process unit and a specific mechanism dimension. The component value reflects the normalized amplitude and direction of the risk disturbance after cross-unit coupling in that mechanism dimension.

[0093] Through the above processing, the risk identification agent 13 progressively decomposes, propagates, correlates, corrects, normalizes, and encodes the deviation feature set corresponding to the real-time multimodal spatiotemporal data field, obtaining an interpretable multidimensional risk perturbation vector. This multidimensional risk perturbation vector serves both as the criterion for the early warning module 15 to perform early warning management and as the initial perturbation input for the evolutionary identification agent 14 to perform risk evolution recursive analysis in the latent space.

[0094] Evolutionary identification agent 14 is used to perform risk evolution recursive analysis within a unified multimodal generation model using the multidimensional risk perturbation vector, and to establish a set of risk evolution trajectories.

[0095] Specifically, the evolutionary identification agent 14 takes over the multidimensional risk perturbation vector output by the risk identification agent 13 at each sampling time, and performs risk evolution recursive analysis within the unified multimodal generation model and the high-dimensional latent space distribution based on the unified multimodal generation model trained by the working condition fitting agent 12, outputting a set of risk evolution trajectories as the basis for the subsequent early warning and alarm management by the early warning and alarm alarm module 15.

[0096] First, the evolutionary identification agent 14 determines the initial state for risk evolution recursive analysis. The initial state consists of two parts: a reference position and a disturbance superposition. The reference position is the mapped state vector obtained by mapping the real-time multimodal spatiotemporal data field at the current sampling time through the modal encoder and the coupled encoder, reflecting the actual position of the current operating condition in the latent space. The disturbance superposition is the superposition of the multidimensional risk disturbance vector at the current sampling time onto the corresponding coordinate components of the reference position after coordinate restoration, according to the process unit identifier and the mechanism residual direction identifier, reflecting the risk disturbance information identified at the current time. The specific processing method for perturbation superposition is as follows: Since the amplitude of each perturbation component in the multidimensional risk perturbation vector is the normalized amplitude after normalization based on the mechanism constraint residual threshold, which is inconsistent with the actual value of the latent space coordinates in terms of dimensions, the evolutionary recognition agent 14 first performs coordinate restoration processing on each perturbation component. It multiplies the normalized amplitude by the mechanism constraint residual threshold in the corresponding mechanism residual direction to obtain the actual coordinate value of the perturbation component in the corresponding mechanism residual direction dimension. Then, the evolutionary recognition agent 14 locates the latent space sub-representation region of the corresponding process unit on the mapped state vector based on the process unit identifier in the perturbation component. It then further locates the corresponding mechanism residual direction dimension in the latent space sub-representation region based on the mechanism residual direction identifier in the perturbation component. The actual coordinate value of the perturbation component after coordinate restoration is added to the coordinate component of the mapped state vector at that position, completing the superposition of one perturbation component. After performing the above coordinate restoration and position superposition processing on all perturbation components in the multidimensional risk perturbation vector in sequence, the initial state is obtained.

[0097] Then, the evolutionary recognition agent 14 performs state evolution recursion along the process topology direction, starting from the initial state, within the latent space of the unified multimodal generation model. The state evolution recursion uses a pre-set recursion time step Δt as the single-step unit. The recursion time step Δt is expressed in seconds or minutes, and its typical value is several times the sampling period corresponding to the unified sampling rhythm. It is specified by process experts based on the process reaction time scale of the chemical production plant. For example, for fast reaction devices such as gas phase reactors, the recursion time step Δt is smaller, typically 1 to 10 seconds; for slow reaction devices such as large storage tanks or heat exchanger systems, the recursion time step Δt is larger, typically 1 to 5 minutes.

[0098] In each single-step advancement, the evolutionary recognition agent 14 determines the propagation direction set for this single-step advancement based on the downstream directed edge distribution relationship of the current node in the process topology constraint structure. Specifically, when there are N downstream directed edges for the current node, this single-step advancement propagates along each of the N downstream directed edges, generating state advancements on N downstream nodes. For each downstream directed edge, the evolutionary recognition agent 14 inputs the latent space state corresponding to the current time node into the state transition module inside the unified multimodal generation model. The state transition module is composed of a unidirectional information transmission layer and the directed connection relationships between each latent space sub-representation region, reusing the parameters in the trained unified multimodal generation model. Based on the connection relationship corresponding to the directed edge and following the physically feasible directions defined by the established directed state evolution manifold structure and the set of latent space evolution trajectories, the state transition module performs single-step extrapolation on the latent space state of the current time node, outputting the latent space state of the downstream node corresponding to the next time node. After the above processing, each single-step advancement generates N next time node states downstream of the current node, corresponding to N downstream propagation paths. Since the directed state evolution manifold structure and the set of latent space evolution trajectories have been subject to mechanistic and topological constraints during the training phase, the next time node states output by the state transition module automatically satisfy the relationships of reaction kinetics, material balance, and energy conservation, eliminating the need to repeatedly perform mechanistic feasibility judgments during the operation phase. For process topologies with multiple levels of downstream nodes, the above single-step advancement continues to unfold at each level of downstream node according to the downstream directed edge distribution relationship of that level, allowing risk disturbances to propagate downstream layer by layer along the process topology. As the single-step advancement continues to iterate along the time axis, each propagation path at each time node continues to generate the next time node state, ultimately forming an evolutionary trajectory tree rooted at the initial disturbance node, branching downstream layer by layer along the process topology, and continuously extending into the future along the time axis.

[0099] The evolutionary identification agent 14 uses the state of the next time node obtained from each single step as the input for subsequent single steps, iteratively advancing along the time axis. This allows each evolutionary branch on the evolutionary trajectory tree to extend continuously towards future times from the initial perturbation node until the branch reaches a pre-set evolution termination condition. The evolution termination condition includes three scenarios: the calculation time reaches its upper limit, the latent space state converges, and the latent space state diverges. Specifically, when the cumulative calculation time of the evolutionary branch reaches the pre-set maximum evolution time Tmax, the calculation of that branch is terminated. The maximum evolution time Tmax is expressed in minutes or hours, typically ranging from 10 minutes to 6 hours from the current time, specified by process experts based on the risk response time window of the chemical production unit. When the latent space state converges into the high-density distribution manifold of normal operating conditions during the calculation of the evolutionary branch, i.e., the magnitude of the deviation vector corresponding to the calculated state decreases to the value set by the aforementioned risk identification agent 13... When the deviation is below a predetermined threshold, the deduction of that evolutionary branch is terminated, indicating that the risk disturbance on that evolutionary branch has naturally decayed and dissipated during the evolution process. When the deviation amplitude of the latent space state during the deduction of that evolutionary branch shows a monotonically increasing trend over several consecutive recursive time steps and exceeds the predetermined divergence threshold, the deduction of that evolutionary branch is terminated, indicating that the risk disturbance on that evolutionary branch has evolved into an uncontrollable divergence, and further deduction is no longer of analytical value. The divergence threshold is typically a multiple of the deviation threshold and is specified by process experts based on the risk tolerance of the equipment.

[0100] After the above-described state evolution recursive processing, the evolutionary recognition agent 14 obtains a risk evolution trajectory set composed of all evolutionary branches on the evolution trajectory tree. Each evolution trajectory in the risk evolution trajectory set consists of several time nodes arranged sequentially along the time axis starting from the initial perturbation node. Each time node records the latent space state corresponding to that time node, the process unit identifier involved in that state, the perturbation amplitude and direction corresponding to each mechanism residual direction on that process unit, and the path position information of that time node in the evolution trajectory tree. The path position information includes the process topology depth corresponding to that time node and the upstream node identifier of the evolutionary branch to which that node belongs, which is used to trace the propagation path of the risk perturbation along the process topology in subsequent processing.

[0101] Through the above processing, the evolutionary identification agent 14 produces a set of structured risk evolution trajectories at each sampling time. The risk evolution trajectory set serves as the basis for the early warning module 15 to perform early warning management and is used for subsequent processing based on risk evolution trend stage division, risk level determination and safety intervention strategy sequence establishment.

[0102] The early warning module 15 is used to issue early warnings based on the risk evolution trajectory set and simultaneously establish a sequence of safety intervention strategies.

[0103] Specifically, the early warning module 15 receives the risk evolution trajectory set output by the evolution recognition agent 14 at each sampling moment. Based on the future risk evolution information carried in the risk evolution trajectory set, it simultaneously performs early warning and safety intervention strategy sequence establishment processing at each sampling moment, and outputs early warning information and safety intervention instructions to the chemical plant operators, the scheduling control system and the on-site execution agency respectively.

[0104] For early warning processing, the early warning module 15 uses the risk evolution trajectory set as the basis for judgment, identifies the risk development trend evolving along the process topology after the current moment, and organizes the early warning information into levels according to the severity and time urgency of the risk development trend, sending early warning information to the corresponding recipients according to the early warning level. The information content of the early warning includes, but is not limited to, the distribution of risk disturbances identified at the current moment, the range of process units expected to be affected at future time nodes, the type and degree of mechanism violation expected to occur in each affected process unit, and the propagation path of the risk evolving from the current moment to each future time node. The output forms of the early warning include, but are not limited to, the visual early warning window on the control room display screen, the push message on the operator's handheld terminal, the digital early warning signal in the scheduling control system, and the alarm drive signal triggered to the on-site audible and visual early warning device.

[0105] For the establishment and processing of the safety intervention strategy sequence, the early warning module 15 constructs a set of intervention operation sequences organized according to the execution time based on the risk propagation path carried in the risk evolution trajectory set and the expected disturbance distribution at future time nodes. Each intervention operation in the safety intervention strategy sequence corresponds to a specific execution node, a specific execution time, and a specific intervention action. After the safety intervention strategy sequence is established, the early warning module 15 distributes each intervention operation in the sequence to the corresponding on-site execution agency, dispatch control system, or operator in sequence according to the execution time, so that the chemical production unit can receive proactive intervention before the risk disturbance develops to an uncontrollable level along the evolution trajectory.

[0106] The early warning and the establishment of a safety intervention strategy sequence are executed synchronously and in parallel at each sampling time. Both processes share the same set of risk evolution trajectories as their basis, but they target different output objects. Early warning focuses on conveying risk information to personnel and systems in a readable format, enabling relevant personnel and systems to understand the current risk status and future evolution trends. The establishment of a safety intervention strategy sequence focuses on distributing risk response measures to the site in an executable form, enabling chemical production units to proactively handle risks according to a pre-planned sequence.

[0107] Through the above processing, the early warning module 15 synchronously completes the hierarchical reporting of early warning information and the construction and distribution of safety intervention strategy sequences at each sampling moment, so that the chemical production unit can obtain corresponding early warning prompts and intervention measures at each stage of the development of risk disturbance along the evolution trajectory, forming a risk management closed loop from risk perception, risk identification, risk evolution analysis to risk early warning and intervention.

[0108] Furthermore, the early warning dispatch module 15 is also used for:

[0109] The risk development trend is divided into stages and the corresponding risk levels are determined based on the set of risk evolution trajectories.

[0110] Early warning and notification management is carried out using risk levels and phase division results.

[0111] In a preferred embodiment, when issuing an early warning, the early warning module 15 first divides the risk development trend into stages and determines the corresponding risk levels based on the risk evolution trajectory set. The stage division and risk level determination quantify the risk development trend from the time dimension and the severity dimension, respectively. Regarding stage division, the early warning module 15 divides the risk development trend into several stages along the time dimension based on the evolution of the disturbance amplitude of each evolution trajectory along the time axis in the risk evolution trajectory set. For example, the risk development trend can be divided sequentially from the current moment into four stages: the gestation stage, the development stage, the intensification stage, and the critical stage. Specifically, when the disturbance amplitude on the risk evolution trajectory exceeds the deviation threshold but has not yet continued to grow, the early warning module 15 determines that it is in the incubation stage, at which point the risk is still in an early recoverable state; when the disturbance amplitude on the risk evolution trajectory shows a continuous growth trend along the time axis but has not reached several times the mechanism constraint residual threshold, the early warning module 15 determines that it is in the development stage, at which point the risk has begun to propagate along the process topology; when the disturbance amplitude on the risk evolution trajectory exceeds several times the mechanism constraint residual threshold and continues to grow, the early warning module 15 determines that it is in the intensification stage, at which point the risk has significantly deviated from the controllable range and is strongly propagating to downstream process units; when the disturbance amplitude on the risk evolution trajectory exceeds the divergence threshold, or when the risk evolution trajectory reaches a sensitive process unit within a pre-set time window, the early warning module 15 determines that it is in the critical stage, at which point the risk is approaching an uncontrollable level. The specific criteria values ​​for each stage are specified by process experts based on the process safety margin of the chemical production unit. Regarding risk level determination, the early warning module 15 classifies risk development trends into corresponding risk levels based on three comprehensive indicators: the disturbance amplitude involved in the risk evolution trajectory set, the range of affected process units, and the degree of risk evolution divergence. For example, risks are divided into four levels from low to high: blue warning, yellow warning, orange warning, and red warning. Specifically, a blue warning is issued when the maximum disturbance amplitude in the risk evolution trajectory set is between the deviation threshold and the mechanism constraint residual threshold, the number of affected process units is 1, and there is no divergent evolution trajectory; a yellow warning is issued when the maximum disturbance amplitude exceeds the mechanism constraint residual threshold but does not reach the divergence threshold, or the number of affected process units is 2 to 3, or a single divergent evolution trajectory appears; an orange warning is issued when the maximum disturbance amplitude continues to increase to near the divergence threshold, or the number of affected process units is 4 or more, or several divergent evolution trajectories appear; and a red warning is issued when the maximum disturbance amplitude exceeds the divergence threshold, or the risk evolution trajectory reaches a sensitive process unit within a pre-set time window. The specific criteria for each level are determined by process experts based on the process characteristics and historical risk records of the chemical production plant.

[0112] The early warning notification module 15 uses both the phase division results and the risk level determination results as a structured description of the risk development trend. The phase division results reflect the evolutionary position of the risk in the time dimension, and the risk level determination results reflect the quantitative level of the risk in the severity dimension. Together, they constitute a two-dimensional characterization of the risk development trend. Then, the early warning notification module 15 uses the risk level and phase division results to perform early warning notification management. The core of early warning notification management lies in selecting the corresponding early warning notification strategy based on the combination of different risk levels and different phases, so that the frequency of early warning information issuance, the recipients, the output format, and the level of detail of the information match the current risk development trend. Specifically, the early warning notification module 15 adjusts the frequency of early warning information issuance and the scope of recipients from low to high risk levels. For example, under a blue alert level, alert information is issued at a lower frequency, primarily to the handheld terminals of on-duty operators; under a yellow alert level, alert information is issued at a medium frequency, simultaneously to on-duty operators, control room displays, and the dispatch and control system; under an orange alert level, alert information is issued at a higher frequency, further reported to the chemical plant's dispatch and command center and higher-level safety management departments in addition to the aforementioned recipients, and triggers on-site audible and visual alert devices; under a red alert level, alert information is continuously issued in real time, simultaneously triggering the plant-wide broadcast alert and emergency response system, notifying all relevant personnel and external emergency management departments. The alert reporting module 15 adjusts the level of detail and reporting focus of the alert information based on the stage division results. For example, during the incubation stage, early warning information is mainly suggestive, focusing on informing that the risk disturbance has been identified and the scope of the affected process units; during the development stage, early warning information is mainly trend-based, focusing on the direction of risk propagation along the process topology and the expected affected process units; during the escalation stage, early warning information is mainly urgent, focusing on informing about the current magnitude of the risk and the time remaining before the critical threshold; during the critical stage, early warning information is mainly instructional, focusing on informing about the emergency response measures that need to be implemented immediately.

[0113] The early warning module 15 performs a combined lookup on the risk level and stage division results at each sampling time. Based on the pre-established "risk level - stage - early warning strategy" correspondence table, it determines the early warning strategy for that time and selects the corresponding items from the early warning information content and output format according to the determined strategy to perform early warning, thus forming differentiated and hierarchical early warning management.

[0114] Through the above processing, the early warning module 15 enables the chemical production unit to receive tiered early warning information that matches the different stages and severity of the risk development trend. This avoids excessive early warnings in the early stages of a risk, which can cause operator fatigue, and also avoids insufficient early warnings in the escalation or critical stages of a risk, which could lead to missing the best opportunity for action.

[0115] Furthermore, the early warning dispatch module 15 is also used for:

[0116] Based on the aforementioned risk evolution trajectory set, identify risk sources and propagation paths, and establish retrospective time-series intervention strategies;

[0117] The aforementioned retrospective time-series intervention strategy is used as a sequence of security intervention strategies to perform intervention management.

[0118] In a preferred embodiment, firstly, the early warning module 15 identifies the risk source and propagation path based on the risk evolution trajectory set. During risk source identification, the early warning module 15 uses the initial disturbance node of each evolution trajectory in the risk evolution trajectory set as the starting point for backtracking. It traces back along the path location information recorded at each time node on each evolution trajectory, identifying the earliest disturbance in the process unit node in the time dimension, and determining that process unit node as the risk source node of the corresponding evolution trajectory. If multiple evolution trajectories in the risk evolution trajectory set originate from the same process unit node, that process unit node is determined as the primary risk source node of the current risk event. If multiple independent risk source nodes exist in the risk evolution trajectory set, the early warning module 15 sorts each risk source node according to the number of its corresponding evolution trajectories and the maximum disturbance amplitude, determining the top-ranked risk source node as the primary risk source node and the remaining nodes as secondary risk source nodes. During propagation path identification, the early warning module 15 uses the path location information recorded at each time node (including the process topology depth corresponding to that time node and the upstream node identifier of the evolution branch to which that node belongs) to trace the evolution trajectory tree from the risk source node downstream, reconstructing the complete propagation chain of the risk disturbance along the process topology. The complete propagation chain is expressed as a node sequence starting from the risk source node, passing through several intermediate transmission process unit nodes in sequence, and finally reaching the downstream affected process unit node. The connection relationship between adjacent nodes in the sequence is consistent with the corresponding directed edge in the process topology constraint structure. Each node is attached with the disturbance amplitude and mechanism residual direction component at the time node corresponding to that node. In the case where multiple evolution trajectories in the risk evolution trajectory set share some propagation nodes, the early warning module 15 merges multiple propagation chains into a propagation path graph. The propagation path graph is expanded along the process topology direction with the risk source node as the root node. Each branch node corresponds to a different propagation sub-path, so that the multi-path propagation relationship of the risk disturbance is uniformly expressed in the propagation path graph.

[0119] Then, the early warning module 15 establishes a retrospective time-series intervention strategy based on the identified risk source nodes and propagation path diagram. The retrospective time-series intervention strategy organizes each intervention operation in a "source first, propagation path second, downstream protection third" sequence, enabling the chemical production unit to first cut off the source of risk, then prevent the risk from continuing to propagate along the process topology, and finally implement protective measures for downstream process units that have been affected or are expected to be affected. Specifically, the intervention operations in the retrospective time-series intervention strategy are organized according to the following three time-series levels: the first time-series level is the source treatment operation. The early warning module 15 determines the corresponding source intervention action for the identified primary and secondary risk source nodes based on the type of mechanistic violation involved in each risk source node. Specifically, for risk sources in the direction of reaction kinetic residuals, typical source intervention actions include reducing reactant feed flow rate, reducing reaction temperature setpoint, increasing cooling medium flow rate, and adding reaction inhibitors; for risk sources in the direction of material balance residuals, typical source intervention actions include closing inlet material valves, initiating emergency pressure relief, and switching to backup storage tanks; for risk sources in the direction of energy conservation residuals, typical source intervention actions include increasing heat exchange medium flow rate, switching heat exchanger backup loops, and initiating emergency cooling systems. The execution time of source handling operations is set to the current time or the shortest response time after the current time. The second time sequence level is the propagation cutoff operation. The early warning module 15 determines the corresponding propagation cutoff intervention action for intermediate transmission process unit nodes located between the main risk source node and downstream affected nodes in the propagation path diagram, based on the directed edge type of each intermediate node in the process topology constraint structure. Specifically, for intermediate nodes on the directed edge of materials, propagation cutoff actions typically include closing the material conveying valve corresponding to the node, activating the material diversion loop corresponding to the node, and reducing the material conveying flow rate; for intermediate nodes on the directed edge of energy, propagation cutoff actions typically include switching the heat exchange loop and reducing energy transfer efficiency. The execution time of the propagation cutoff operation is set according to the transmission delay parameter of each intermediate node from the main risk source node in the propagation path diagram, that is, the propagation cutoff operation at each intermediate node is executed within the reserved response time before the expected risk disturbance reaches the node. The third timing level is the downstream protection operation. The early warning module 15 determines the corresponding downstream protection intervention action for the downstream affected process unit nodes in the propagation path diagram based on the process function attributes and current operating status of each downstream node, such as activating the emergency shutdown procedure of the downstream process unit, increasing the safety interlock priority of the downstream process unit, performing preventive load reduction operation on the downstream process unit, and activating the emergency pressure relief system of the downstream process unit. The timing of downstream protection operations is set according to the arrival time of the corresponding time node in the risk evolution trajectory set, that is, protection measures are initiated in advance before the expected risk disturbance reaches the downstream node.

[0120] After the retrospective time-series intervention strategy is established, the data structure of each intervention operation in the strategy includes three elements: execution node, execution time, and intervention action. The execution node corresponds to the specific identifier of the on-site execution agency, scheduling control system, or operator; the execution time is expressed as a delay time from the current time; and the intervention action is expressed as a specific process operation instruction, including the operation object, operation parameters, and operation duration. This data structure is consistent with the data structure of each intervention operation in the safety intervention strategy sequence; that is, the retrospective time-series intervention strategy constitutes the specific expression form of the safety intervention strategy sequence.

[0121] Subsequently, the early warning module 15 utilizes a backtracking time-series intervention strategy as the sequence of security intervention strategies for intervention management. The early warning module 15 sequentially distributes each intervention operation in the strategy to the corresponding execution nodes according to their execution time. When the execution time for each intervention operation arrives, the early warning module 15 sends the instruction signal for that intervention operation to the corresponding execution node and waits for the execution result feedback from the execution node. For intervention operations that receive successful execution feedback, the early warning module 15 marks the operation as executed in the strategy record. For intervention operations that receive failed execution feedback, the early warning module 15, based on pre-set failure handling rules, either retryes the intervention operation, issues a backup intervention operation, or upgrades the warning level.

[0122] During the intervention management process, the early warning module 15 continuously receives the latest risk disturbance vector and risk evolution trajectory set output by the risk identification agent 13 and the evolution identification agent 14. If the risk development trend reflected by the latest risk evolution trajectory set changes significantly compared with the previous analysis results (such as the migration of the main risk source node, the reconstruction of the propagation path, or the upgrading of the risk level), the early warning module 15 updates the retrospective time-series intervention strategy, re-identifies the risk source and propagation path, reconstructs the retrospective time-series intervention strategy, and replaces the unexecuted part of the original strategy that is being executed with the updated strategy, so that the intervention management always matches the latest risk development trend.

[0123] Through the above processing, the early warning module 15 enables the chemical production unit to receive proactive intervention and handling in the "source-propagation-downstream" sequence as the risk disturbance develops along the process topology. It can also dynamically adjust the intervention strategy based on the latest risk development trend, avoiding the failure to effectively contain the risk due to misalignment of the intervention sequence or omission of the intervention target during the risk evolution process.

[0124] Furthermore, this embodiment also includes a verification module, which is used to verify the consistency of the risk evolution trajectory set based on the time-series update results of the multidimensional risk perturbation vector, generate verification feedback, and perform supervised update management of the unified multimodal generation model based on the verification feedback.

[0125] Specifically, firstly, the verification module receives the time-series update results of the multidimensional risk perturbation vector. Specifically, the verification module continuously receives the multidimensional risk perturbation vector output by the risk identification agent 13 at each sampling time, and organizes the multidimensional risk perturbation vectors at each sampling time into a time-series sequence of multidimensional risk perturbation vectors according to the sampling time order. Simultaneously, the verification module receives the risk evolution trajectory set output by the evolution identification agent 14 at each sampling time, and organizes the risk evolution trajectory set at each sampling time into a time-series sequence of risk evolution trajectory sets according to the sampling time order.

[0126] Then, the verification module performs consistency verification on the risk evolution trajectory set based on the time series sequence of the multidimensional risk perturbation vector and the time series sequence of the risk evolution trajectory set. The core of the consistency verification lies in comparing the degree of agreement between "the predicted perturbation for future times in the risk evolution trajectory set produced at a certain sampling time t" and "the multidimensional risk perturbation vector actually observed at subsequent sampling times t+Δt, t+2Δt, ...". The specific processing method for consistency verification is as follows: For the risk evolution trajectory set at a certain sampling time t in the time series of the risk evolution trajectory set, the verification module extracts the predicted perturbation distribution of each evolution trajectory at a future time node t+nΔt (n is a positive integer) as the predicted value; For the subsequent actual sampling time t+nΔt in the time series of the multidimensional risk perturbation vector, the verification module extracts the multidimensional risk perturbation vector actually output by the risk identification agent 13 at that time as the observed value; The verification module aligns the predicted value and the observed value item by item according to the process unit identifier and the mechanism residual direction identifier, and calculates the difference in perturbation amplitude and perturbation direction between the two in the corresponding process unit and the corresponding mechanism residual direction. The quantitative indicators for consistency verification include two categories: perturbation amplitude consistency indicators and perturbation direction consistency indicators. The perturbation amplitude consistency indicator is defined as the absolute value of the difference between the normalized perturbation amplitude of the predicted value and the observed value at corresponding positions; a smaller difference indicates a better indicator. The perturbation direction consistency indicator is defined as whether the signs of the perturbation directions of the predicted value and the observed value are the same at corresponding positions; the same sign indicates consistency, and opposite signs indicate inconsistency. The verification module compares these two types of quantitative indicators with a pre-set consistency threshold to obtain the consistency verification result. The consistency threshold is obtained by performing statistical analysis on the consistency distribution of predicted and observed data during historical operation, typically taking the median or 75th percentile of the historical consistency indicator distribution. When both types of indicators are within the consistency range defined by the consistency threshold, the verification module determines that the risk evolution trajectory set at sampling time t is consistent with the subsequent actual observation results; when either type of indicator exceeds the consistency range defined by the consistency threshold, the verification module determines that the risk evolution trajectory set is inconsistent with the actual observation results.

[0127] Next, the verification module generates verification feedback based on the consistency verification results. The verification feedback is expressed in structured data form, containing four elements: verification conclusion, inconsistency location, inconsistency degree, and inconsistency samples. Specifically, the verification conclusion identifies whether the result of this consistency verification is consistent or inconsistent; the inconsistency location, when the verification conclusion is inconsistent, records the specific process unit identifier and the direction of the mechanistic residual where the inconsistency occurs between the predicted and observed values; the inconsistency degree, when the verification conclusion is inconsistent, records the magnitude and direction of the perturbation amplitude difference between the predicted and observed values ​​at the inconsistency location; and the inconsistency samples record the predicted and observed data pairs involved in this verification for subsequent monitoring and updates.

[0128] Subsequently, the validation module performs supervised update management on the unified multimodal generative model based on the validation feedback. The implementation of supervised update management varies depending on the validation conclusions and the degree of inconsistency in the validation feedback. For sampling moments with consistent validation conclusions, the validation module does not update the parameters of the unified multimodal generative model; it only records the validation result in the model's health record for subsequent overall model performance evaluation. For sampling moments with inconsistent validation conclusions, the validation module performs targeted model parameter updates. The specific processing method for parameter updates is as follows: the validation module uses the inconsistent samples in the validation feedback as incremental training data and the difference between predicted and observed values ​​as a supervision signal. It performs incremental training on the parameters of the latent space sub-representation regions corresponding to the inconsistent positions in the unified multimodal generative model and their upstream and downstream unidirectional information transfer layers. The training loss function adopts the established overall loss function form. Iterative training is performed to update the model parameters in the direction of reducing the difference between predicted and observed values. After the incremental training is completed, the parameter updates overwrite the parameters at the corresponding positions in the original model; the remaining parameters not involving inconsistent positions remain unchanged.

[0129] When a large number of inconsistent verification feedbacks accumulate within a certain time window, the verification module triggers the global update management of the unified multimodal generation model. That is, all inconsistent samples accumulated within a certain time window are used as global incremental training data to perform a complete retraining process on the unified multimodal generation model. The retraining process still uses the established overall loss function and joint optimization method to make the unified multimodal generation model adapt to the long-term evolution of the process characteristics of chemical production equipment.

[0130] Through the above processing, the verification module enables the unified multimodal generation model to be continuously corrected and updated based on actual observation data during the actual operation of the chemical production plant. This avoids the decline in prediction accuracy caused by factors such as long-term evolution of process characteristics, equipment aging, and changes in raw material composition, and keeps the risk identification and risk evolution analysis capabilities in line with the current actual state of the chemical production plant.

[0131] Example 2, as Figure 2 As shown, based on the same inventive concept as the multimodal collaborative AI agent management system for chemical safety production provided in Embodiment 1, this embodiment of the invention also provides a multimodal collaborative AI agent management method for chemical safety production, including:

[0132] During the production and operation of a chemical plant, the interaction and perception of processes, equipment, environment, and personnel are implemented to establish a multimodal spatiotemporal data field;

[0133] Based on the historical calibration data of the multimodal spatiotemporal data field, chemical reaction mechanism constraints and process topology constraints are introduced to construct a unified multimodal generation model. The unified multimodal generation model is used to jointly model the coupling relationship between multimodal data to form a high-dimensional latent space distribution that characterizes the evolution law of normal operating conditions.

[0134] The real-time acquired multimodal spatiotemporal data field is mapped to a high-dimensional latent space distribution. After calculating the deviation feature set, the interpretability decomposition of the deviation feature set is performed to generate a multidimensional risk perturbation vector.

[0135] Using the aforementioned multidimensional risk perturbation vector, risk evolution recursive analysis is performed within a unified multimodal generation model to establish a set of risk evolution trajectories;

[0136] Early warnings are issued based on the risk evolution trajectory set, and a sequence of safety intervention strategies is established simultaneously.

[0137] Furthermore, a unified multimodal generation model is constructed, including:

[0138] According to the process unit division rules of chemical production plants, the historical calibration data is divided into subsets corresponding to reaction units, transmission units and heat exchange units respectively, and a directional process topology constraint structure is constructed based on the material flow direction and energy exchange relationship between each unit.

[0139] Based on the aforementioned process topology constraint structure, a structural framework for a unified multimodal generation model is defined, ensuring that the connection relationships and information transmission directions between different state variables in the unified multimodal generation model are consistent with the material flow direction and energy exchange path. Furthermore, a hidden space sub-representation region corresponding to each unit is preset within the structural framework.

[0140] The reaction unit, the transport unit and the heat exchange unit are respectively introduced with corresponding mechanistic constraints. The reaction unit is introduced with rate constraints based on the reaction kinetic equation, the transport unit is introduced with cumulative constraints based on the material balance relationship, and the heat exchange unit is introduced with heat exchange constraints based on the energy conservation relationship. The mechanistic constraints are embedded into the structural framework in the form of state constraint terms.

[0141] Based on the structural framework of embedded state constraints, the unified multimodal generation model is trained using the historical calibration data. By jointly optimizing the data distribution fitting target and the mechanism constraint target, the unified multimodal generation model forms a high-dimensional latent space distribution that represents the evolution law of normal operating conditions, under the condition of satisfying the local mechanism constraints of each process unit and the overall process topology constraints.

[0142] Furthermore, a high-dimensional hidden space distribution representing the evolution of normal operating conditions is formed, including:

[0143] Using the latent space sub-representation region corresponding to each process unit as the basic unit, a set of state samples under normal operating conditions is constructed based on historical calibration data. Density clustering constraints are applied to the mapping result of the state sample set in the latent space to form a high-density distribution manifold of the stable operating state of the corresponding process unit within each latent space sub-representation region.

[0144] In the process of constructing the high-density distributed manifold, cross-unit consistency constraints based on process topology constraints are introduced to enable the latent space sub-representation regions of adjacent process units to form a continuous state transition relationship in the material flow direction and energy exchange direction, and to construct a directed state evolution manifold structure that conforms to the actual process flow in the overall latent space.

[0145] Based on mechanistic constraints, the gradient of each state variable in the directed state evolution manifold structure is constrained directionally, limiting the feasible direction of the state evolution path in the latent space, so that any state transition in the latent space satisfies the relationship of reaction kinetics, material balance and energy conservation, forming a set of latent space evolution trajectories with physical feasibility.

[0146] By combining the set of latent space evolution trajectories and applying sparse distribution constraints and directional decomposition constraints to samples that deviate from the high-density distribution manifold, the abnormal state is made to appear as a low-density offset distribution unfolding along a predetermined mechanism dimension in the latent space, thereby generating a structured high-dimensional latent space distribution that simultaneously includes the normal operating condition manifold and the direction of abnormal disturbance.

[0147] Furthermore, a multidimensional risk perturbation vector is generated, including:

[0148] The real-time acquired multimodal spatiotemporal data field is input into the unified multimodal generation model to obtain the mapped state vector in the high-dimensional latent space. Based on the normal working condition high-density distribution manifold corresponding to the high-dimensional latent space distribution, the minimum deviation path of the mapped state vector relative to the high-density distribution manifold is calculated to form an initial deviation feature representation including the deviation magnitude and deviation direction.

[0149] The initial deviation feature representation is subjected to structured decomposition processing. The deviation direction is partitioned and projected according to the preset latent space sub-representation region in the unified multimodal generation model. Based on the mechanism constraints corresponding to each process unit, the initial deviation feature representation is decomposed into multiple disturbance components with physical meaning in the subspaces corresponding to the reaction unit, the transmission unit and the heat exchange unit, respectively, to obtain local risk disturbance sub-vectors associated with different process units.

[0150] Based on the process topology constraint structure, cross-unit propagation correlation analysis is performed on each local risk disturbance sub-vector. According to the material flow direction and energy exchange path, the transmission weight and coupling influence of each disturbance component between adjacent process units are calculated, and the local risk disturbance sub-vectors are weighted and reconstructed to form a coupled disturbance representation that characterizes the risk propagation relationship.

[0151] The coupled disturbance representation is subjected to direction correction and amplitude normalization, and the corrected disturbance components are combined and encoded according to a preset mechanism dimension to generate an interpretable multidimensional risk disturbance vector.

[0152] Furthermore, a multimodal spatiotemporal data field is established, including:

[0153] The interactive perception results of processes, equipment, environment, and personnel are managed for time synchronization and spatial registration, and a data set under a unified spatiotemporal reference system is constructed.

[0154] The structured organization and analysis of multimodal data relationships in the execution dataset forms a multimodal spatiotemporal data field with spatiotemporal correlation characteristics.

[0155] Furthermore, based on the structured organization analysis, anomaly trigger discrimination is performed. If the anomaly trigger discrimination result is a trigger pass result, then the corresponding data linkage resampling processing is performed.

[0156] Furthermore, the risk development trend is divided into stages based on the risk evolution trajectory set, and the corresponding risk level is determined; the risk level and stage division results are used for early warning and notification management.

[0157] Furthermore, risk sources and propagation paths are identified based on the risk evolution trajectory set, and a retrospective time-series intervention strategy is established; the retrospective time-series intervention strategy is used as a sequence of security intervention strategies to implement intervention management.

[0158] Furthermore, consistency verification of the risk evolution trajectory set is performed based on the temporal update results of the multidimensional risk disturbance vector, verification feedback is generated, and supervised update management of the unified multimodal generation model is executed based on the verification feedback.

[0159] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0160] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A multimodal collaborative AI intelligent agent management system for chemical safety production, characterized in that: The system includes: The interactive sensing module is used to perform interactive sensing of processes, equipment, environment, and personnel during the production and operation of a chemical plant, and to establish a multimodal spatiotemporal data field. The working condition fitting agent is used to construct a unified multimodal generation model based on the historical calibration data of the multimodal spatiotemporal data field, by introducing constraints of chemical reaction mechanism and process topology. The unified multimodal generation model is used to jointly model the coupling relationship between multimodal data to form a high-dimensional latent space distribution that characterizes the evolution law of normal working conditions. A risk identification agent is used to map real-time acquired multimodal spatiotemporal data fields to a high-dimensional latent space distribution. After calculating the deviation feature set, it performs interpretable decomposition of the deviation feature set to generate a multidimensional risk perturbation vector, including: The real-time acquired multimodal spatiotemporal data field is input into the unified multimodal generation model to obtain the mapped state vector in the high-dimensional latent space. Based on the normal working condition high-density distribution manifold corresponding to the high-dimensional latent space distribution, the minimum deviation path of the mapped state vector relative to the high-density distribution manifold is calculated to form an initial deviation feature representation including the deviation magnitude and deviation direction. The initial deviation feature representation is subjected to structured decomposition processing. The deviation direction is partitioned and projected according to the preset latent space sub-representation region in the unified multimodal generation model. Based on the mechanism constraints corresponding to each process unit, the initial deviation feature representation is decomposed into multiple disturbance components with physical meaning in the subspaces corresponding to the reaction unit, the transmission unit and the heat exchange unit, respectively, to obtain local risk disturbance sub-vectors associated with different process units. Based on the process topology constraint structure, cross-unit propagation correlation analysis is performed on each local risk disturbance sub-vector. According to the material flow direction and energy exchange path, the transmission weight and coupling influence of each disturbance component between adjacent process units are calculated, and the local risk disturbance sub-vectors are weighted and reconstructed to form a coupled disturbance representation that characterizes the risk propagation relationship. The coupled disturbance representation is subjected to direction correction and amplitude normalization, and the corrected disturbance components are combined and encoded according to a preset mechanism dimension to generate an interpretable multidimensional risk disturbance vector. An evolutionary identification agent is used to perform risk evolution recursive analysis within a unified multimodal generation model using the multidimensional risk perturbation vector, and to establish a set of risk evolution trajectories. The early warning module is used to issue early warnings based on the risk evolution trajectory set and simultaneously establish a sequence of safety intervention strategies.

2. The multimodal collaborative AI intelligent agent management system for chemical safety production as described in claim 1, characterized in that, In the aforementioned working condition fitting agent, a unified multimodal generation model is constructed, including: According to the process unit division rules of chemical production plants, the historical calibration data is divided into subsets corresponding to reaction units, transmission units and heat exchange units respectively, and a directional process topology constraint structure is constructed based on the material flow direction and energy exchange relationship between each unit. Based on the aforementioned process topology constraint structure, a structural framework for a unified multimodal generation model is defined, ensuring that the connection relationships and information transmission directions between different state variables in the unified multimodal generation model are consistent with the material flow direction and energy exchange path. Furthermore, a hidden space sub-representation region corresponding to each unit is preset within the structural framework. The reaction unit, the transport unit and the heat exchange unit are respectively introduced with corresponding mechanistic constraints. The reaction unit is introduced with rate constraints based on the reaction kinetic equation, the transport unit is introduced with cumulative constraints based on the material balance relationship, and the heat exchange unit is introduced with heat exchange constraints based on the energy conservation relationship. The mechanistic constraints are embedded into the structural framework in the form of state constraint terms. Based on the structural framework of embedded state constraints, the unified multimodal generation model is trained using the historical calibration data. By jointly optimizing the data distribution fitting target and the mechanism constraint target, the unified multimodal generation model forms a high-dimensional latent space distribution that represents the evolution law of normal operating conditions, under the condition of satisfying the local mechanism constraints of each process unit and the overall process topology constraints.

3. The multimodal collaborative AI intelligent agent management system for chemical safety production as described in claim 1, characterized in that, In the aforementioned working condition fitting agent, a high-dimensional latent space distribution representing the evolution law of normal working conditions is formed, including: Using the latent space sub-representation region corresponding to each process unit as the basic unit, a set of state samples under normal operating conditions is constructed based on historical calibration data. Density clustering constraints are applied to the mapping result of the state sample set in the latent space to form a high-density distribution manifold of the stable operating state of the corresponding process unit within each latent space sub-representation region. In the process of constructing the high-density distributed manifold, cross-unit consistency constraints based on process topology constraints are introduced to enable the latent space sub-representation regions of adjacent process units to form a continuous state transition relationship in the material flow direction and energy exchange direction, and to construct a directed state evolution manifold structure that conforms to the actual process flow in the overall latent space. Based on mechanistic constraints, the gradient of each state variable in the directed state evolution manifold structure is constrained directionally, limiting the feasible direction of the state evolution path in the latent space, so that any state transition in the latent space satisfies the relationship of reaction kinetics, material balance and energy conservation, forming a set of latent space evolution trajectories with physical feasibility. By combining the set of latent space evolution trajectories and applying sparse distribution constraints and directional decomposition constraints to samples that deviate from the high-density distribution manifold, the abnormal state is made to appear as a low-density offset distribution unfolding along a predetermined mechanism dimension in the latent space, thereby generating a structured high-dimensional latent space distribution that simultaneously includes the normal operating condition manifold and the direction of abnormal disturbance.

4. The multimodal collaborative AI intelligent agent management system for chemical safety production as described in claim 1, characterized in that, In the interactive perception module, a multimodal spatiotemporal data field is established, including: The interactive perception results of processes, equipment, environment, and personnel are managed for time synchronization and spatial registration, and a data set under a unified spatiotemporal reference system is constructed. The structured organization and analysis of multimodal data relationships in the execution dataset forms a multimodal spatiotemporal data field with spatiotemporal correlation characteristics.

5. The multimodal collaborative AI intelligent agent management system for chemical safety production as described in claim 4, characterized in that, The system also includes: The resampling module is used to perform anomaly trigger discrimination based on structured organization analysis. When the anomaly trigger discrimination result is a trigger pass result, the corresponding data is subjected to linked resampling processing.

6. The multimodal collaborative AI intelligent agent management system for chemical safety production as described in claim 1, characterized in that, The warning dispatch module is also used for: The risk development trend is divided into stages and the corresponding risk levels are determined based on the set of risk evolution trajectories. Early warning and notification management is carried out using risk levels and phase division results.

7. The multimodal collaborative AI intelligent agent management system for chemical safety production as described in claim 6, characterized in that, The warning dispatch module is also used for: Based on the aforementioned risk evolution trajectory set, identify risk sources and propagation paths, and establish retrospective time-series intervention strategies; The aforementioned retrospective time-series intervention strategy is used as a sequence of security intervention strategies to perform intervention management.

8. The multimodal collaborative AI intelligent agent management system for chemical safety production as described in claim 1, characterized in that, The system also includes: The verification module is used to verify the consistency of the risk evolution trajectory set based on the time-series update results of the multidimensional risk disturbance vector, generate verification feedback, and perform supervised update management of the unified multimodal generation model based on the verification feedback.

9. A multimodal collaborative AI agent management method for chemical safety production, characterized in that: The method includes: During the production and operation of a chemical plant, the interaction and perception of processes, equipment, environment, and personnel are implemented to establish a multimodal spatiotemporal data field; Based on the historical calibration data of the multimodal spatiotemporal data field, chemical reaction mechanism constraints and process topology constraints are introduced to construct a unified multimodal generation model. The unified multimodal generation model is used to jointly model the coupling relationship between multimodal data to form a high-dimensional latent space distribution that characterizes the evolution law of normal operating conditions. The real-time acquired multimodal spatiotemporal data field is mapped to a high-dimensional latent space distribution. After calculating the deviation feature set, an interpretability decomposition of the deviation feature set is performed to generate a multidimensional risk perturbation vector, including: The real-time acquired multimodal spatiotemporal data field is input into the unified multimodal generation model to obtain the mapped state vector in the high-dimensional latent space. Based on the normal working condition high-density distribution manifold corresponding to the high-dimensional latent space distribution, the minimum deviation path of the mapped state vector relative to the high-density distribution manifold is calculated to form an initial deviation feature representation including the deviation magnitude and deviation direction. The initial deviation feature representation is subjected to structured decomposition processing. The deviation direction is partitioned and projected according to the preset latent space sub-representation region in the unified multimodal generation model. Based on the mechanism constraints corresponding to each process unit, the initial deviation feature representation is decomposed into multiple disturbance components with physical meaning in the subspaces corresponding to the reaction unit, the transmission unit and the heat exchange unit, respectively, to obtain local risk disturbance sub-vectors associated with different process units. Based on the process topology constraint structure, cross-unit propagation correlation analysis is performed on each local risk disturbance sub-vector. According to the material flow direction and energy exchange path, the transmission weight and coupling influence of each disturbance component between adjacent process units are calculated, and the local risk disturbance sub-vectors are weighted and reconstructed to form a coupled disturbance representation that characterizes the risk propagation relationship. The coupled disturbance representation is subjected to direction correction and amplitude normalization, and the corrected disturbance components are combined and encoded according to a preset mechanism dimension to generate an interpretable multidimensional risk disturbance vector. Using the aforementioned multidimensional risk perturbation vector, risk evolution recursive analysis is performed within a unified multimodal generation model to establish a set of risk evolution trajectories; Early warnings are issued based on the risk evolution trajectory set, and a sequence of safety intervention strategies is established simultaneously.

Images