A laboratory safety risk intelligent early warning method, system, medium and product

Abstract

The application discloses a kind of laboratory safety risk intelligent early warning method, system, medium and product, it is related to risk early warning field.The method comprises the following steps: calibrating laboratory digital twin model based on real-time monitoring data, generating initial probability distribution based on laboratory digital twin model and fusing preset equipment failure probability model and personnel behavior uncertainty model;The probability density distribution change of initial probability distribution at a plurality of time points in the future is deduced by numerical algorithm, and the corresponding state prediction probability distribution is obtained;The probability value of each dangerous state developed at each time point in the future is obtained based on the state prediction probability distribution of each future time point;When the probability value of any dangerous state evaluated exceeds the preset threshold value, the optimal control operator is solved reversely by optimization algorithm, and the intervention action corresponding to the optimal control operator is executed.The above technical scheme can predict laboratory safety risk in advance.

Description

Technical Field

[0001] This application relates to the field of risk warning, and in particular to an intelligent early warning method, system, medium and product for laboratory safety risks. Background Technology

[0002] In existing technologies, methods commonly used to ensure laboratory safety include assigning dedicated personnel to conduct regular inspections of the laboratory to check the operational status of equipment and whether environmental parameters are normal; and installing sensors, such as smoke sensors and temperature sensors, to issue alarms when abnormalities are detected. These methods can monitor and manage the safety status of the laboratory to a certain extent and play a role in preventing laboratory safety accidents.

[0003] However, traditional laboratory safety management methods have significant shortcomings. Dedicated personnel patrols are time-sensitive, making real-time monitoring difficult and potentially missing unforeseen safety hazards. Traditional sensors can only monitor single environmental parameters, failing to comprehensively assess laboratory safety risks. Furthermore, these methods cannot predict future safety risks or implement effective interventions in advance, making it difficult to meet the increasingly complex safety management needs of modern laboratories. Summary of the Invention

[0004] This application provides a method, system, medium, and product for intelligent early warning of laboratory safety risks, which can predict laboratory safety risks in advance, display the risks in three-dimensional dynamic visualization, and take timely intervention measures to reduce the probability of danger.

[0005] Firstly, this application provides an intelligent early warning method for laboratory safety risks, the method comprising:

[0006] Acquire real-time monitoring data from multi-source heterogeneous sensors deployed in the laboratory, calibrate a preset laboratory digital twin model based on the real-time monitoring data, calculate and generate an initial probability distribution characterizing the current uncertainty of the system based on the state information of the calibrated laboratory digital twin model and by integrating a preset equipment failure probability model and a personnel behavior uncertainty model.

[0007] By using numerical algorithms to deduce the changes in the probability density distribution of the initial probability distribution at multiple future time points, the corresponding state prediction probability distribution is obtained.

[0008] The predicted probability distribution of the state at each future time point is projected onto multiple predefined dangerous state modes to calculate the probability value of developing into each dangerous state at each future time point and then output.

[0009] When the probability value of any assessed dangerous state exceeds a preset threshold, the optimal control operator is solved in reverse by an optimization algorithm, and the intervention action corresponding to the optimal control operator is executed.

[0010] By employing the above technical solution, real-time data collection of laboratory environment and equipment is achieved through multi-source heterogeneous sensors. Combined with the fusion analysis of digital twin models and probabilistic models, this enables high-precision characterization and uncertainty quantification of the current system state, improving the accuracy and comprehensiveness of safety monitoring. Numerical algorithms are used to perform probability density extrapolation of the system's future state, allowing for early prediction of the development trends of various hazardous scenarios. This enables a shift from passive response to proactive early warning, significantly extending the safety response window. By projecting the predicted state onto predefined hazardous modes, the probability of various risks is quantitatively assessed, and automated risk classification is achieved using preset thresholds, providing a scientific and intuitive basis for safety decision-making. When the risk probability exceeds the threshold, the system automatically solves for the optimal control strategy through optimization algorithms and executes corresponding intervention actions, achieving closed-loop management from risk warning to proactive handling. This effectively curbs accident development and enhances the laboratory's autonomous protection level.

[0011] In some embodiments, the calculation of an initial probability distribution characterizing the current uncertainty of the system based on the state information of the calibrated laboratory digital twin model, and by integrating a preset equipment failure probability model and a personnel behavior uncertainty model, specifically includes:

[0012] The current state parameters of all entities in the calibrated laboratory digital twin model are extracted and mapped into the system state space to form a state vector. Based on the state vector, a multivariate Gaussian probability distribution representing the state confidence range is generated as the base point of the initial probability distribution.

[0013] The system calls the equipment failure probability model in the laboratory digital twin model that matches the equipment model and current operating conditions of the target equipment. It inputs the current running time and load rate of the target equipment into the equipment failure probability model and receives the conditional failure probability density function output by the equipment failure probability model. The conditional failure probability density function defines the continuous probability of the current equipment state drifting in the neighborhood of the healthy state. The target equipment is any one of the equipment in the laboratory digital twin model.

[0014] The system calls the personnel behavior uncertainty model in the laboratory digital twin model that matches the role of the target personnel and the current operation task. It inputs the current task complexity and environmental interference into the personnel behavior uncertainty model and receives the discrete probability distribution list output by the personnel behavior uncertainty model. The discrete probability distribution list defines the possible discrete branches of the target personnel's behavior state. The target personnel is any one of the personnel in the laboratory digital twin model.

[0015] Based on the physical and logical relationships between entities defined in the laboratory digital twin model, the initial probability distribution is obtained by jointly calculating the base point, the conditional failure probability density function of all devices, and the discrete probability distribution list of all personnel in the system state space.

[0016] By employing the aforementioned technical solution, the state parameters of all entities in the calibrated digital twin model are extracted and mapped into state vectors in the state space, forming a confidence range characterized by a multivariate Gaussian distribution. This enables dynamic, structured, and probabilistic modeling of the laboratory physical system and its operational status, providing an accurate initial state benchmark for subsequent risk simulations. By separately invoking equipment failure probability models matched to equipment models and operating conditions, and personnel behavior uncertainty models matched to personnel roles and tasks, a simultaneous and refined quantitative expression of continuous equipment failure risks and discrete personnel behavior risks is achieved, overcoming the limitations of traditional early warning systems that only focus on a single dimension of equipment or environment. Based on the physical and logical relationships between entities defined in the digital twin model, various risk probability distributions are jointly calculated in the system state space, forming a unified overall uncertainty model that expresses the coupled effects of multiple factors such as equipment, personnel, and environment. This significantly improves the systematicness and completeness of the system risk representation.

[0017] In some embodiments, the initial probability distribution is obtained by jointly calculating the distribution of the base point, the conditional failure probability density function of all devices, and the discrete probability distribution list of all personnel in the system state space based on the physical and logical relationships between entities defined in the laboratory digital twin model, specifically including:

[0018] The base point, the conditional failure probability density function of all equipment, and the discrete probability distribution list of all personnel are each modeled as independent factor nodes in a probability tensor network.

[0019] Based on the physical connections between devices, the personnel-device operation relationships, and the environmental coupling relationships defined in the laboratory digital twin model, associated edges are defined between each of the factor nodes in the probability tensor network to construct the network topology.

[0020] Based on the network topology, tensor contraction is performed on the interconnected factor nodes along the shared associated edges. All intermediate variables are eliminated by summation to obtain the joint probability distribution tensor. The joint probability distribution tensor is then normalized to obtain the initial probability distribution.

[0021] By adopting the above technical solution, the equipment failure probability density function, the discrete distribution of personnel behavior, and the system state baseline are modeled as independent factor nodes in a probabilistic tensor network, respectively. This achieves structural decoupling and modular representation of continuous equipment risk, discrete personnel risk, and the overall system state, providing a clear framework for the mathematical expression of complex coupling relationships. Based on the entity relationships such as physical connections, human-machine operation, and environmental coupling defined in the digital twin model, dynamic association edges are established between factor nodes, forming a network topology that can adaptively adjust with changes in laboratory layout, equipment configuration, and work processes, giving the model good scalability and scenario adaptability. Tensor contraction operations are used to fuse information along the association edges of connected factor nodes. By summing and eliminating intermediate variables, the joint calculation of high-dimensional, heterogeneous probability distributions is efficiently completed, avoiding the problems of excessive computational complexity or large approximation errors in complex systems using traditional Bayesian networks or Monte Carlo methods.

[0022] In some embodiments, the step of using numerical algorithms to deduce the change in probability density distribution of the initial probability distribution at multiple future time points to obtain the corresponding state prediction probability distribution specifically includes:

[0023] The state variables in the laboratory are discretized into a set of state basis vectors in the system state space. The state variables include environmental parameters, equipment status, personnel positions and actions.

[0024] Based on the standard operating procedures, physical laws, and equipment safety thresholds of the laboratory, a potential energy operator is constructed to characterize the inherent constraints of the system. Based on the physical connection and linkage logic between the laboratory equipment, a kinetic energy operator is constructed to characterize the trend of system state changes. The potential energy operator and the kinetic energy operator are linearly combined to form an equivalent Hamiltonian operator.

[0025] Using the initial probability distribution as the initial state vector and the equivalent Hamiltonian operator as the dynamic driver, the time evolution equation is solved by numerical integration to obtain the state vectors at multiple future time points.

[0026] Calculate the squared modulus of the state vectors at multiple future time points to obtain the corresponding state prediction probability distribution.

[0027] By employing the aforementioned technical solution, the laboratory state is discretized into state basis vectors, and potential and kinetic energy operators based on physical constraints and system correlations are constructed. This forms an equivalent Hamiltonian operator that integrates physical rules and operational logic, providing a high-dimensional, structured, and computable mathematical model for the dynamic evolution of complex coupled systems. Using the initial probability distribution as the initial state and the equivalent Hamiltonian operator as the evolution driver, the time evolution equation is solved through numerical integration, achieving efficient deduction of the overall uncertainty dynamic evolution of the system on a continuous time scale. This overcomes the limitations of traditional discrete-time Markov models in terms of time resolution and state continuity. The use of wave function evolution and modulus squared calculation methods within a quantum mechanical framework ensures probability conservation. The resulting state prediction probability distribution is not only mathematically self-consistent but also retains the physical interpretability of the system evolution, facilitating the analysis of risk propagation paths and key evolutionary nodes.

[0028] In some embodiments, the step of projecting the predicted probability distribution of the state at each future time point onto multiple predefined dangerous state patterns to obtain and output the probability value of developing into each dangerous state at each future time point specifically includes:

[0029] Regularly conduct correlation and retrospective analysis of historical monitoring data streams in the laboratory and recorded safety events to obtain abnormal evolution patterns of system states that are not predefined.

[0030] The abnormal evolution mode of the system state is transformed into a new dangerous eigenstate subspace in the state space, and the new dangerous eigenstate subspace is written into the dynamically updated dangerous state mode library;

[0031] During each risk assessment, the predicted probability distribution of the state at each future time point is projected onto all currently stored hazardous state patterns in the hazardous state pattern library in real time for calculation.

[0032] The results of the projection calculation are output as a set of probability values ​​for the system to develop into each type of dangerous state in the model library at each future time point, forming a full risk probability spectrum.

[0033] By employing the above technical solution, the system automatically identifies undefined abnormal evolution patterns by periodically analyzing the correlation between historical data and safety events, and transforms these patterns into new eigenstates of hazardous states. This enables dynamic updates to the hazardous state pattern library, allowing the system to continuously learn and adapt to unknown risks. During each risk assessment, the predicted probability distribution is projected in real-time onto all patterns in the dynamically updated hazardous pattern library, achieving comprehensive parallel assessment of both known hazards and potential unknown risks, significantly improving the system's ability to identify complex and sudden safety threats. By incorporating dynamically mined new hazardous patterns into the assessment process in real-time, the system achieves deep integration of emerging abnormal patterns with historical experience rules, enabling it to provide early warnings based on historical patterns and respond promptly to novel risk evolution paths that first appear during laboratory operations.

[0034] In some embodiments, obtaining and outputting the probability values ​​of development to each dangerous state at each future time point includes:

[0035] The probability value at each future time point is bound to the spatial location attribute and the scope of influence attribute of the corresponding hazard state in the laboratory digital twin model to generate a risk data object with spatial coordinates and risk intensity.

[0036] For each of the risk data objects, the corresponding visual representation parameters are dynamically calculated based on the probability value and the evolution stage of the corresponding hazard state. The visual representation parameters include color, transparency, basic geometric dimensions, and dynamic deformation parameters used to express the trend of risk diffusion or accumulation.

[0037] Based on the visual representation parameters, corresponding three-dimensional graphic elements are generated. All the generated three-dimensional graphic elements are used as independent rendering layers and superimposed on the basic three-dimensional scene of the laboratory digital twin model in real time to form a three-dimensional dynamic risk visualization screen.

[0038] By employing the aforementioned technical solution, risk data objects are generated by binding probability values ​​with the spatial location and impact range of hazardous states. This imbues abstract probability data with clear spatial coordinates and impact boundary attributes, transforming risk information from numerical description to spatially locatable information. This supports an intuitive understanding of risk distribution within a 3D twin scenario. Based on dynamically calculated visual representation parameters such as probability values ​​and evolution stages, a structured and hierarchical risk visual coding system is formed using various visual variables such as color, transparency, geometric dimensions, and dynamic deformation. This allows for the intuitive differentiation and understanding of risks of different levels, types, and evolutionary trends. By using the generated 3D graphic elements as independent rendering layers and overlaying them in real-time onto the digital twin model's base scene, dynamic and seamless integration and visualization of risk information and the physical environment are achieved. This enables safety managers to intuitively perceive the location, intensity, and evolution process of risks within a simulation environment.

[0039] In some embodiments, the reverse solution of the optimal control operator using an optimization algorithm includes:

[0040] The convergence objective function of the optimization algorithm is set as minimizing the sum of the projected integral values ​​of the predicted state probability distribution on the dangerous state mode within a preset future time range.

[0041] A control space is constructed using all controllable parameters of the laboratory digital twin model. A gradient-based optimization algorithm is used to perform a global iterative search within the control space to find a combination of control parameters that can reduce the value of the convergent objective function.

[0042] From the results of the global search, the control parameter adjustment amount that contributes the most to reducing the value of the convergence objective function is identified, and the mathematical expression corresponding to the control parameter adjustment amount is determined as the optimal control operator.

[0043] By adopting the above technical solution, a mathematical model oriented towards optimizing the overall cumulative risk of the system is established by setting the minimization of the projected integral of risk probability in the future time as the objective function. This transforms safety control from an empirical, rule-based response to a proactive optimization decision based on probability quantification. A gradient-based optimization algorithm is used to perform a global iterative search in the entire controllable parameter space, achieving joint optimization of complex multivariable control systems. This overcomes the inefficiency and local optima problems of traditional rule-based control or manual trial-and-error in scenarios with explosive parameter combinations. The identified control parameter adjustments are transformed into optimal control operators in mathematical expression, forming control commands that combine mathematical rigor and engineering interpretability. These commands can directly drive the digital twin model to perform simulation verification and are mapped to the actual physical control system to execute corresponding intervention actions.

[0044] In a second aspect, embodiments of this application provide a computer system including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the steps of the method described in any possible implementation of the first aspect.

[0045] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method described in any possible implementation of the first aspect.

[0046] Fourthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in any possible implementation of the first aspect.

[0047] It is understood that the computer system provided in the second aspect, the storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the method provided in this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0048] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0049] 1. Model and analyze the laboratory as a holistic system, tightly coupled with equipment, personnel, and environment, rather than monitoring individual risk points in isolation. This helps to identify complex systemic risks caused by the interaction and chain reactions of multiple factors;

[0050] 2. By using numerical algorithms to extrapolate the dynamic changes in probability distribution, it is possible to predict the probability distribution of system state at multiple future time points, thereby enabling the prediction of risk development trends and gaining a valuable time window for early intervention.

[0051] 3. Projecting the predicted future state onto predefined hazardous state patterns essentially maps a continuous probability distribution onto a specific, understandable hazardous scenario. This transforms abstract mathematical predictions into judgments about the probability of specific hazardous events occurring.

[0052] 4. When the risk probability exceeds the threshold, the optimal control operator is solved in reverse through an optimization algorithm. This means that the system can automatically calculate a set of intervention strategies that theoretically can most effectively reduce the overall future risk probability based on the current specific risk situation, thereby improving the overall reliability of the system. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating an intelligent early warning method for laboratory safety risks in an embodiment of this application;

[0054] Figure 2 This is a flowchart illustrating the calculation of the initial probability distribution in an embodiment of this application;

[0055] Figure 3 This is a schematic diagram of an exemplary hardware structure of a computer system in an embodiment of this application. Detailed Implementation

[0056] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0057] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0058] The following is combined Figure 1 The method of the embodiments of this application will be described below.

[0059] Figure 1 This is a flowchart illustrating an intelligent early warning method for laboratory safety risks in an embodiment of this application, as shown below. Figure 1 As shown, the method includes the following steps:

[0060] S101. Obtain real-time monitoring data from multi-source heterogeneous sensors deployed in the laboratory, calibrate a preset laboratory digital twin model based on the real-time monitoring data, and calculate and generate an initial probability distribution characterizing the current uncertainty of the system based on the state information of the calibrated laboratory digital twin model and by integrating a preset equipment failure probability model and a personnel behavior uncertainty model.

[0061] S102. By using numerical algorithms, the probability density distribution of the initial probability distribution changes at multiple future time points to obtain the corresponding state prediction probability distribution;

[0062] S103. Project the predicted probability distribution of the state at each future time point onto multiple predefined dangerous state modes to obtain the probability value of developing into each dangerous state at each future time point and output it.

[0063] S104. When the probability value of any assessed dangerous state exceeds a preset threshold, the optimal control operator is solved in reverse by an optimization algorithm, and the intervention action corresponding to the optimal control operator is executed.

[0064] Multi-source heterogeneous sensors deployed within the laboratory include, but are not limited to, temperature sensors, humidity sensors, gas concentration sensors (such as VOC, O2, CO), smoke detectors, video surveillance equipment, access control systems, equipment operation status monitors (current, vibration, pressure), intelligent gas supply device monitoring units, and personnel positioning and behavior capture equipment. These sensors upload monitoring data to edge computing nodes or a central server in real time via wired or wireless communication protocols. The pre-set digital twin model is a virtualized, structured, and computable representation of the laboratory's physical entities (equipment, environment, personnel), business processes, and physical laws. The system compares and assimilates the real-time monitoring data with the state parameters of the corresponding entities in the twin model, using methods such as Kalman filtering, particle filtering, or deep learning-based regression to dynamically adjust model parameters (such as equipment efficiency coefficients, environmental thermal conductivity coefficients, etc.) to ensure that the twin model's state is consistent with the real laboratory. For each key piece of equipment in the digital twin model (such as reactors, fume hoods, and gas cylinder cabinets), the system pre-configures or learns its equipment failure probability model online. The model can be a survival analysis model trained based on historical equipment operating data (such as the Weibull distribution), a degradation model based on physical failure mechanisms, or a Bayesian network incorporating expert knowledge. Model inputs include the equipment's current operating time, load rate, number of work cycles, and number of recent abnormal events. The output is a conditional failure probability density function indicating a shift in the equipment's health status under the current conditions. For various personnel roles in the laboratory (such as lab technicians, visitors, and maintenance personnel), the system pre-configures or adaptively learns uncertainty models of their behaviors. This model can be constructed based on historical operation records, a behavioral rule base, and cognitive load theory, for example, using a Hidden Markov Model or a decision tree model. Model inputs include the current task complexity (such as the type of hazardous chemicals involved and the number of operation steps), environmental disturbances (such as noise and emergency alarms), and personnel fatigue index. The output is a series of possible behaviors that personnel may take and their corresponding discrete probability distributions. The system extracts and maps the current state parameters (such as temperature, pressure, valve opening, and personnel location) of all entities in the calibrated digital twin model to a high-dimensional system state space, forming a state vector. Based on this state vector, the conditional failure probability density function of all equipment and the discrete behavioral probability distribution of all personnel are integrated. A multivariate probability distribution characterizing the current uncertainty of the entire laboratory system is generated through joint probability calculation methods (such as probabilistic graphical model inference, Monte Carlo integration, or variational inference). This initial probability distribution quantifies the confidence that the system is in various possible states at the current moment.

[0065] Based on the physical and chemical laws, equipment operation logic, operating procedures, and human-computer interaction rules of the laboratory, a mathematical model describing the evolution of the system state is constructed. This model can be a differential equation, a stochastic differential equation, or a state transition probability matrix. The key is to abstract the physical connections between equipment, control logic, and personnel behavior strategies into coupling relationships between state variables. Using an initial probability distribution as initial conditions, numerical algorithms are used to extrapolate the uncertainty of the future system state. Specific methods may include solving the Fokker-Planck equation, Monte Carlo simulation, or operator-based quantum mechanical analogy methods. Through the above numerical extrapolation, estimates of the probability density function (or discrete probability mass function) of the system state at a series of discrete time points in the future (e.g., 1 minute, 5 minutes, 30 minutes), i.e., the state prediction probability distribution, are obtained. These distributions describe the probability that the system may be in various states at different future times.

[0066] Hazardous state models are mathematical descriptions of specific safety accident scenarios that may occur in a laboratory (such as fire, gas leak, equipment overheating, and mixed hazards caused by human error). Each model corresponds to a region or subspace (called a hazardous eigenstate) in the system state space. The model library can be predefined by safety experts based on historical accident analysis and safety procedures, and can be automatically mined and dynamically updated from historical anomaly data using machine learning methods. For the state prediction probability distribution at each future time point, for the target model (any hazardous state) in the hazardous state model library, the integral (or summation, for discrete states) of the state prediction probability distribution over the region corresponding to the target model is calculated. This integral value is the probability that the system is in the target model at time t. Performing the above projection calculation for each future time point and each hazardous state model yields a two-dimensional matrix (or a set of time series vectors), where the element Risk_Probability(t, i) represents the probability of the i-th type of hazard occurring at time t. These probability values ​​can be displayed in real time on the monitoring interface, stored in a database for trend analysis, or directly transmitted to the decision-making module.

[0067] A pre-defined probability alarm threshold θ_i is assigned to each hazardous state mode i. The system continuously monitors all Risk_Probability(t, i). Once a time point t and a certain mode i are found to have Risk_Probability(t, i) > θ_i, an optimized control process is triggered. The objective function is typically set to minimize the weighted integral sum of the probabilities of all hazardous states within a future look-ahead time window [t, t+T], or minimize the worst-case probability. Decision variables are the set of adjustable and controllable parameters in the digital twin model, such as the setpoint of the ventilation system, the heating power of a reactor, the on / off state of an emergency shut-off valve, and the content of warning instructions sent to specific personnel. These parameters constitute a control space. Constraints include equipment operating limits, safety procedures, and the acceptable level of personnel intervention. An optimization algorithm is used to search the control parameter trajectory in the control space that optimizes the objective function. The algorithm may include: (1) Model Predictive Control (MPC) framework: In each control cycle, solve a finite-time optimal control problem, execute only the first step of the control action, and then solve it again based on the new state estimate. (2) Gradient-based method: If the system model is differentiable, the gradient of the objective function with respect to the control parameters can be calculated using the adjoint method or automatic differentiation, and gradient descent algorithms (such as Adam, L-BFGS) can be used for efficient search. For complex nonlinear systems, a policy network (control operator) can be pre-trained. This network takes the current system state (or probability distribution characteristics) as input and directly outputs the control action. The optimal control policy obtained can be mathematically expressed as an operator (function) that maps the system state to the control action, i.e., the optimal control operator. Apply the optimal control operator to the actual control system: First, execute the control policy in the digital twin model, predict its effect, and verify its safety and effectiveness; after verification, send the control command to the corresponding actuator through the industrial control network, such as adjusting the valve opening, changing the equipment operation mode, activating the audible and visual alarm, issuing personnel evacuation guidance information, locking the access control of dangerous areas, etc. After the intervention is performed, the system continues to monitor changes in the laboratory status through sensors and returns to step S101 to start a new cycle.

[0068] Figure 2 This is a flowchart illustrating the calculation of the initial probability distribution in an embodiment of this application, as shown below. Figure 2 As shown, the initial probability distribution representing the current uncertainty of the system is calculated and generated based on the state information of the calibrated laboratory digital twin model and by integrating a preset equipment failure probability model and a personnel behavior uncertainty model. Specifically, this includes:

[0069] S201. Extract and map the current state parameters of all entities in the calibrated laboratory digital twin model to the system state space to form a state vector. Generate a multivariate Gaussian probability distribution representing the state confidence range based on the state vector, and use it as the base point of the initial probability distribution.

[0070] S202. Call the equipment failure probability model in the laboratory digital twin model that matches the equipment model and current operating conditions of the target equipment. Input the current running time and load rate of the target equipment into the equipment failure probability model. Receive the conditional failure probability density function output by the equipment failure probability model. The conditional failure probability density function defines the continuous probability of the current equipment state drifting within the neighborhood of the healthy state. The target equipment is any one of the equipment in the laboratory digital twin model.

[0071] S203. Call the personnel behavior uncertainty model in the laboratory digital twin model that matches the role of the target personnel and the current operation task, input the current task complexity and environmental interference degree into the personnel behavior uncertainty model, and receive the discrete probability distribution list output by the personnel behavior uncertainty model. The discrete probability distribution list defines the possible discrete branches of the target personnel's behavior state. The target personnel is any one of the personnel in the laboratory digital twin model.

[0072] S204. Based on the physical and logical relationships between entities defined in the laboratory digital twin model, perform joint distribution calculations on the base point, the conditional failure probability density function of all devices, and the discrete probability distribution list of all personnel in the system state space to obtain the initial probability distribution.

[0073] The calibrated digital twin model includes all physical entities (such as device A, device B, ventilation duct P, regional temperature field T, and person Zhang San) and their dynamic attributes (such as the current temperature, rotation speed, and valve opening of device A; and the position coordinates and operating posture of person Zhang San). The system defines a unified, high-dimensional system state space, where each dimension corresponds to a key state variable (e.g., x1 represents the temperature of device A, x2 represents the pressure of device B, x3 represents the average VOC concentration in region T, and x4 represents the distance of person Zhang San relative to device A). The current state parameter values ​​of all entities are filled into the corresponding dimensions according to predefined mapping rules, forming a numerical vector X_current=[x1, x2, x3, ..., xN]. TThis vector represents the point estimate of the system's most probable current state. Considering sensor errors, model calibration residuals, and instantaneous fluctuations in the state, the current true state is not a definite point but rather exists within a confidence range. The system uses the state vector X_current as the mean vector m and sets a covariance matrix n based on historical data statistical analysis or a pre-defined model (e.g., considering sensor accuracy). This generates a multivariate Gaussian (normal) distribution: P_base(X) ~ N(m, n). This P_base(X) serves as the base point of the initial probability distribution. It describes the normal fluctuation range of the system state around the current best estimate, without considering sudden equipment failures or abnormal personnel operations.

[0074] The system maintains a library of equipment failure probability models. For any target device in the digital twin model (e.g., "centrifuge C123"), the system retrieves the corresponding failure probability model from the library based on its equipment model (e.g., "C123 high-speed centrifuge") and current operating condition (e.g., "high-speed operation," "low-temperature cooling"). This model can be trained based on a large amount of historical operating data and failure records of the same model of equipment, for example, using a survival analysis model (e.g., proportional hazards model) or a degradation process model (e.g., Wiener process, gamma process). Real-time parameters reflecting the current consumption or pressure of the equipment are input into the model. Core parameters typically include runtime (cumulative running time since the last major overhaul), load rate (current power / rated power), and may also include the number of recent abnormal vibrations, the degree of temperature deviation from the baseline, etc. The model outputs a conditional failure probability density function f_i(δ|p_i), where p_i represents the input conditions (runtime t, load rate l), and δ represents the drift of the equipment health status index relative to its health baseline (one or more dimensional vectors). For example, δ1 represents the vibration amplitude increment, δ2 represents the bearing temperature increment, and the function value of f_i(δ|p_i) represents the probability density of a specific drift δ in the health state of the equipment under given current operating conditions. It is a continuous distribution that describes the probability of the equipment evolving from its current state to a nearby fault state of varying degrees.

[0075] The system maintains a library of personnel behavior uncertainty models. For any target personnel in the digital twin model (e.g., "Experimenter Li Si"), the system retrieves the corresponding model from the library based on their role (e.g., "Senior Experimenter," "Intern") and the currently performed task (e.g., "Adding Catalyst A to the Reactor," "Transferring Flammable Solvent"). This model can be constructed based on behavioral analysis of the personnel's historical operation records when performing similar tasks. For example, a Hidden Markov Model can be used to describe standard operating procedures and their common deviations, or a Bayesian network can be used to characterize the relationship between task complexity, environmental stress, and the probability of error. Real-time factors affecting the reliability of personnel operations are input into the model, mainly including the current task complexity (e.g., number of steps, number of hazardous chemicals involved, operational precision requirements) and environmental interference (e.g., background noise level, whether other alarms are sounding, lighting conditions), and may also include the personnel's current physiological state (e.g., fatigue index provided by wearable devices connected to the system). The model outputs a list of discrete probability distributions {(a_j, q_j)}, where a_j represents a possible behavioral state branch, such as: "correctly complete the step", "miss step S5", "mistakenly add reagent B instead of reagent A", "discard the container under tension", and q_j represents the probability that the person will take action a_j under the current input conditions, and the sum of all q_j is one. This list defines the discrete probability distribution of the person's behavior at the current moment.

[0076] The digital twin model predefines the physical and logical relationships between entities. For example: "Person Zhang San is operating equipment A," "The cooling water for equipment B comes from pipe P," and "The airflow from fume hood V affects the diffusion of pollutants in the affected area." These relationships determine how the drift amount δ due to equipment failure and the personnel's behavioral choice a_j will affect the overall system state variable X. The system state node X is a high-dimensional random vector representing the complete state of the entire laboratory (e.g., temperature, pressure, concentration, location, etc.). The equipment failure drift node δ_i (i=1...M) represents the continuous drift of the health state of the i-th equipment relative to the baseline, and the personnel behavior node a_j (j=1...N) represents the discrete behavioral choice currently taken by the j-th person. The system state prior distribution P_base(X) is a multivariate Gaussian distribution calibrated from sensor data, representing the best estimate and confidence range of the system state without considering equipment failure and personnel anomalies. The equipment failure condition distribution f_i(δ_i|p_i) represents the failure probability density function of the i-th equipment. Given the current operating conditions p_i (running time, load rate, etc.), it outputs the probability density of the drift δ_i. The personnel behavior condition distribution q_j(a_j) represents the behavior probability distribution of the j-th person. Given the current task complexity, environmental interference, etc., it outputs the discrete probability of taking behavior a_j. The physical and logical relationships between entities are encoded by the indicator function I(X, {δ_i}, {a_j}). For example, if person a_j performs the "open valve" operation, the corresponding valve opening variable in system state X should be "open". If equipment δ_i experiences a "cooling efficiency decrease" drift, the temperature variable of the corresponding area in system state X should increase. If X is consistent with the state derived from {δ_i} and {a_j}, then I(X, {δ_i}, {a_j}) = 1; otherwise, I(X, {δ_i}, {a_j}) = 0. Calculating the marginal probability distribution P_initial(X) of system state X typically requires integrating (or summing) over all possible equipment drifts {δ_i} and personnel behaviors {a_j}. Since directly calculating this integral is extremely complex, approximate inference algorithms are needed. Monte Carlo methods: A large number of equipment drifts and personnel behaviors are sampled, and the resulting system state is calculated for each sample. Finally, the distribution of X is statistically analyzed. Variational inference: Assuming P_initial(X) belongs to a simple family of distributions (such as a Gaussian mixture model), optimization is used to find the member in this family that best approximates the true posterior distribution. Probabilistic tensor network methods: Each distribution is modeled as a tensor node, and the relationships are modeled as tensor edges. Tensor shrinkage and normalization operations are used to efficiently approximate the joint distribution. The P_initial(X) calculated through the above process is the initial probability distribution representing the current uncertainty of the system.

[0077] In some embodiments, the initial probability distribution is obtained by jointly calculating the distribution of the base point, the conditional failure probability density function of all devices, and the discrete probability distribution list of all personnel in the system state space based on the physical and logical relationships between entities defined in the laboratory digital twin model, specifically including:

[0078] The base point, the conditional failure probability density function of all equipment, and the discrete probability distribution list of all personnel are each modeled as independent factor nodes in a probability tensor network.

[0079] Based on the physical connections between devices, the personnel-device operation relationships, and the environmental coupling relationships defined in the laboratory digital twin model, associated edges are defined between each of the factor nodes in the probability tensor network to construct the network topology.

[0080] Based on the network topology, tensor contraction is performed on the interconnected factor nodes along the shared associated edges. All intermediate variables are eliminated by summation to obtain the joint probability distribution tensor. The joint probability distribution tensor is then normalized to obtain the initial probability distribution.

[0081] The system state X is represented by a multivariate Gaussian prior distribution P_base(X) ~ N(m, n). P_base(X) is represented as a d-order tensor T_base, where each dimension k of the tensor corresponds to a state variable x_k, and the dimension size is L_k. The element value at index (i1, i2, ..., id) of the tensor T_base is the probability mass of the system state being in a specific discrete state combination corresponding to these indices (approximately obtained by integrating the original continuous Gaussian distribution over this discrete unit). This tensor node encodes the prior knowledge of the system state in the network. Suppose the system state X consists of d variables. To perform tensor computation, each continuous state variable x_k needs to be discretized into L_k levels within its possible value range (e.g., temperature from 0°C to 100°C is discretized into 100 intervals of 1°C). In this way, the system state space is transformed from a continuous space into a d-dimensional discrete grid, with a total of L1×L2×...×Ld possible discrete state combinations. The equipment failure probability density function is modeled as a tensor node T_device_i: For the conditional failure probability density function f_i(δ_i|p_i) of the i-th device, its output drift δ_i is typically also a vector (e.g., [vibration drift, temperature drift]). Similarly, each component of δ_i is discretized into several levels, and f_i(δ_i|p_i) is represented as an M_i-order tensor T_device_i, where M_i is the dimension of the drift δ_i. Each element of this tensor represents the probability density value (normalized to probability mass) of the device health state drifting to a specific discrete level combination under a given condition p_i. Each device corresponds to an independent tensor node. The personnel behavior distribution list is modeled as a tensor node T_person_j: The discrete probability distribution list {(a_j, q_j)} of the j-th person is essentially a categorical distribution. Each possible behavior a_j is mapped to a unique index. This distribution is represented as a first-order tensor (i.e., a vector) T_person_j. The length of the vector is equal to the total number of possible actions K_j for that person. The value of the k-th element of the vector is the probability q_j(k) of taking the k-th action. Each person corresponds to an independent tensor node (vector). Based on the digital twin model, the physical and logical relationships between entities are analyzed. For example, physical connection: "The exhaust vent of device A is connected to the ventilation duct P." Operational relationship: "Person Zhang San is operating the control panel of device B." Environmental coupling: "The heat emitted by device C affects the temperature of its surrounding area R." These relationships mean that the state drift δ_i of a device or the action a_j of a person will directly affect one or more specific variables in the system state X.

[0082] All tensor nodes (T_base, T_device_1, ..., T_device_M, T_person_1, ..., T_person_N) are placed in the network. If, based on entity relationships, the drift δ_i of a device node T_device_i affects the k-th variable x_k in the system state X, then an edge is established between the tensor node T_device_i and T_base (representing X). This edge connects the index of a specific dimension of T_device_i corresponding to δ_i to the index of the k-th dimension of T_base. Similarly, if the behavior a_j of a person node T_person_j affects some variables in the system state X, then an edge is established between T_person_j and T_base. This ultimately forms a star or mesh topology, where the T_base node is typically located at the center and connected to various device and person nodes via multiple edges. Each edge represents a specific dependency or influence relationship.

[0083] Along all associated edges, the shared indices of the connected tensor nodes are summed to remove intermediate variables such as equipment drift and personnel behavior, ultimately resulting in a tensor that only relates to the system state X. This process is equivalent to computing a massive tensor network contraction. This operation is equivalent to exhaustively enumerating all possible combinations of equipment failure scenarios and personnel behaviors, weighting them according to their probabilities, and accumulating their impact on the system state X, ultimately obtaining the total probability distribution tensor T_initial of the system state X considering all sources of uncertainty. Due to discretization approximation and the conversion from probability density to probability mass, the sum of all elements of the contracted tensor T_initial may not be 1. The system performs normalization on T_initial, and the normalized tensor is the representation of the desired initial probability distribution P_initial(X) in the discrete state space. It is a d-order tensor, where each element explicitly gives the probability of the system being in a specific combination of discrete states.

[0084] In some embodiments, the step of using numerical algorithms to deduce the change in probability density distribution of the initial probability distribution at multiple future time points to obtain the corresponding state prediction probability distribution specifically includes:

[0085] The state variables in the laboratory are discretized into a set of state basis vectors in the system state space. The state variables include environmental parameters, equipment status, personnel positions and actions.

[0086] Based on the standard operating procedures, physical laws, and equipment safety thresholds of the laboratory, a potential energy operator is constructed to characterize the inherent constraints of the system. Based on the physical connection and linkage logic between the laboratory equipment, a kinetic energy operator is constructed to characterize the trend of system state changes. The potential energy operator and the kinetic energy operator are linearly combined to form an equivalent Hamiltonian operator.

[0087] Using the initial probability distribution as the initial state vector and the equivalent Hamiltonian operator as the dynamic driver, the time evolution equation is solved by numerical integration to obtain the state vectors at multiple future time points.

[0088] Calculate the squared modulus of the state vectors at multiple future time points to obtain the corresponding state prediction probability distribution.

[0089] The system extracts all key variables requiring dynamic prediction from the laboratory digital twin model. These variables are categorized into environmental parameters, equipment status, and personnel location and actions. Environmental parameters include temperature, humidity, and specific gas concentrations (such as O2 and VOCs) in different areas. Equipment status includes stirring speed, heating power, valve opening, and pump start / stop status in the reactor. Personnel location and actions include personnel coordinates and the currently performed action category (such as "walking," "sampling," and "recording"). The set of values ​​for all variables is constructed into a high-dimensional joint state space. To enable numerical computation, each continuous or discrete variable is hierarchically discretized. For example, temperature is uniformly divided into multiple intervals from T_min (minimum temperature) to T_max (maximum temperature), with the center value of each interval representing a discrete state. Valve opening is divided into multiple levels from 0% to 100%. Personnel action A is itself discrete, such as {idle, operating equipment A, handling hazardous materials, emergency evacuation}. Each possible combination of discrete states is defined as a state basis vector, denoted as . .For example, This can represent the complete scenario where "the temperature in region 1 is 25°C, the reactor rotation speed is 300 rpm, and person Zhang San is located at coordinates (10, 5, 0) and his action is 'adding reagent'...". All basis vectors The system spans a D-dimensional complex vector space (Hilbert space), where D is the product of the discrete order numbers of all variables. Any probability distribution of the system can be represented as a complex coefficient vector (wave function) over these basis vectors.

[0090] Standard Operating Procedures (SOPs): These are the prescribed operating procedures and safety protocols for personnel. A state that violates SOPs should have higher "potential energy" (i.e., be less stable). Physical Laws: Such as the laws of thermodynamics (heat flows from high to low temperature), fluid mechanics, and chemical reaction kinetics. States that conform to these natural laws have lower potential energy. Equipment Safety Thresholds: Such as upper limits for temperature alarms and pressure safety limits. States approaching or exceeding these thresholds are assigned extremely high potential energy, indicating that the system strongly "rejects" entering these dangerous states. The potential energy operator is a D×D diagonal matrix (in position basis). Its diagonal elements V_ss represent the basis vectors. The "potential energy" at that time. For example, in a state that fully complies with standard operating procedures and safety thresholds. Set V_safe, safe=0 (reference point).

[0091] Physical connections between devices: such as pipes connecting devices, their pressure and flow states can influence each other. This defines the coupling between state variables, allowing a state change in one device to drive a state change in another device. Interlocking logic: such as interlocking control logic ("When a gas leak is detected, automatically close the main valve and start the exhaust"). This defines a deterministic or probabilistic transition that the system will definitely or is very likely to occur under a specific combination of states. The kinetic energy operator is a D×D sparse off-diagonal matrix. Matrix elements K_{s, s'} represent the transitions from state... transition to state The kinetic energy amplitude. Its construction rules, for example, are based on proximity transitions: only when two basis vectors... and In a physically sense, K_{s, s'} can only be non-zero if the values ​​of only one variable change by a minimum discrete order. The magnitude of the transition strength, K_{s, s'}, is determined by the rate constant or logical trigger probability of the relevant physical process. For example, according to the heat conduction equation, the strength of a temperature transition from T_i to T_{i+1} (higher) may be proportional to the temperature difference (T_{i+1} - T_i) and the thermal conductivity of the material. It is generally required that the kinetic energy operator be a Hermitian operator, which guarantees probability conservation. A linear combination of the potential energy operator and the kinetic energy operator constitutes an equivalent Hamiltonian operator. This equivalent Hamiltonian operator is not the Hamiltonian of a real quantum system, but rather an analogous mathematical model of the dynamics of a classical probabilistic system. It integrates the system's inherent constraints (potential energy operator) and dynamic transition capabilities (kinetic energy operator), fully encoding the evolution of the laboratory as a complex system.

[0092] The initial probability distribution P_initial(X) is transformed into its representation in the discrete state space to obtain the initial state vector. Specifically, for each basis vector, its corresponding coefficient ψ_s(0) is set to sqrt(P_initial(s)), where P_initial(s) is the probability that the system is in discrete state s. Therefore, It is a D-dimensional complex vector, and its squared magnitude is... This is exactly equal to the initial probability P_initial(s). Based on the equivalent Hamiltonian operator, the system constructs a time-dependent evolution operator, which describes the precise evolution of the system's state vector over any short time step. The mathematical definition of the evolution operator is as follows:

[0093] ;

[0094] in, This represents the evolution operator, where Δt represents the time step. This represents the equivalent Hamiltonian operator, where i represents the imaginary unit. It is an analog parameter used to adjust the evolution time scale (in actual calculations it is usually normalized or absorbed into the equivalent Hamiltonian operator).

[0095] Since directly calculating the value of the above formula can be very difficult in high-dimensional cases, the system employs an efficient numerical integration algorithm to approximate the continuous evolution of the state vector over time. The specific process is as follows: the total time range to be predicted [0, tF] is divided into many small time steps Δt; from the initial state vector... Initially, a numerical integration algorithm is applied at each time step to calculate the state vector for the next time step using the equivalent Hamiltonian operator. This process is iterated repeatedly, as if progressing step by step along the time axis. One of the most commonly used algorithms in the system is the Cronck-Nicholson method, which maintains the conservation of the magnitude of the state vector, thus ensuring that the total probability is always 1, which is crucial for probabilistic prediction. Its core iterative formula is as follows:

[0096] ;

[0097] Here, I is the identity matrix. At each time step, the system needs to solve a large system of linear equations to obtain the new state vector. Throughout the iteration process, the system records the complete state vector at each preset future time point {t1, t2, ..., tF}. These vectors completely capture all possible probability amplitude interference and evolution results up to that moment.

[0098] For the state vector of a future time t_f The s-th component ψ_s(tf) is a complex number. Calculating the squared modulus of this component yields the predicted probability that the system is in the discrete state s at time tf. Performing this calculation on all basis vectors yields a set of probability values, which constitutes the complete state prediction probability distribution at time tf. Repeating this process for all future time points {t1, t2, ..., tF} ultimately results in a series of probability distributions {P_pred(X, t1), P_pred(X, t2), ..., P_pred(X, tF)}, representing the state prediction probability distribution for multiple future time points.

[0099] In some embodiments, the step of projecting the predicted probability distribution of the state at each future time point onto multiple predefined dangerous state patterns to obtain and output the probability value of developing into each dangerous state at each future time point specifically includes:

[0100] Regularly conduct correlation and retrospective analysis of historical monitoring data streams in the laboratory and recorded safety events to obtain abnormal evolution patterns of system states that are not predefined.

[0101] The abnormal evolution mode of the system state is transformed into a new dangerous eigenstate subspace in the state space, and the new dangerous eigenstate subspace is written into the dynamically updated dangerous state mode library;

[0102] During each risk assessment, the predicted probability distribution of the state at each future time point is projected onto all currently stored hazardous state patterns in the hazardous state pattern library in real time for calculation.

[0103] The results of the projection calculation are output as a set of probability values ​​for the system to develop into each type of dangerous state in the model library at each future time point, forming a full risk probability spectrum.

[0104] The system continuously collects and stores historical monitoring data streams from all sensors in the laboratory, and aligns and correlates them with safety logs (such as manually reported anomalies, automatic alarm records, and accident reports) using timestamps. Through machine learning algorithms (such as unsupervised anomaly detection, sequence pattern mining, and causal inference models), the system proactively analyzes whether a specific, recurring system state evolution trajectory or characteristic combination exists in the period leading up to a safety incident, a pattern not explicitly listed in the initially preset hazard pattern library. It identifies potential "precursor patterns" or "novel accident evolution paths." For example, the analysis might reveal that "before the past three minor leaks, a specific combination of 'slowly rising temperature in area A' and 'abnormally decreasing wind speed in fume hood B' appeared, and this combination is rare in normal operation." The identified anomaly evolution patterns are then mathematically modeled. Typically, a hazard pattern corresponds to a continuous region or a region composed of multiple discrete basis vectors in the system state space (i.e., the high-dimensional space discretized in the previous steps), rather than a single point. This region is called the eigenstate subspace of the hazard pattern. For example, the aforementioned "leakage precursor pattern" might be defined as: in the state space, the set of all states that satisfy "temperature variable T_A > threshold T_th and wind speed variable V_B < threshold V_th". This set is a subspace. The system writes this newly defined subspace, along with its descriptive information (triggering conditions, associated events, severity level), as a new entry and updates it to the hazardous state pattern library. This pattern library thus becomes a knowledge base that is continuously enriched as system operating experience accumulates.

[0105] During each risk assessment cycle (e.g., every second or after receiving a new batch of data), the system acquires the predicted state probability distributions {P_pred(X, t1), P_pred(X, t2), ...} for multiple future time points {t1, t2, ..., tF} generated by previous steps. For each future time point tf, the system initiates an independent projection calculation based on the probability distribution P_pred(X, tf) and each currently stored hazardous state pattern Dk (k=1 to K, where K is the total number of patterns, including preset and dynamically learned ones). The projection calculation is mathematically equivalent to solving an integral (for continuous state variables) or a summation (for discrete state variables). The system calculates the sum of probabilities for all state points falling within the hazardous subspace Dk in the probability distribution P_pred(X, tf). The probability distribution of future states can be imagined as a probability "cloud" floating in the state space. Each hazardous pattern defines a specific "hazardous area." Projection calculations rapidly measure how much "mass" (i.e., total probability) of this probability cloud falls within each danger zone. Because calculations for different time points and different danger modes are independent of each other, the system employs a highly parallel computing architecture (such as utilizing GPUs or distributed computing clusters) to perform thousands of such projection calculations simultaneously, achieving millisecond-level real-time risk assessment.

[0106] The results of all parallel projection calculations are aggregated by the system. For each future time point tf, the system obtains a list containing K probability values: [Risk_Prob(tf, D1), Risk_Prob(tf, D2), ..., Risk_Prob(tf, DK)]. Here, Risk_Prob(tf, DK) represents the predicted probability value of the system evolving to the scenario defined by hazard mode DK at time tf. The results of all time points are organized along the time dimension to form a two-dimensional "risk-time" matrix, or it can be viewed as a three-dimensional risk surface (time, hazard mode category, probability value). This output is the "full risk probability spectrum." It provides a panoramic, quantitative, and time-lined view of risk. For example, it can clearly show: "After 1 minute, the probability of 'electrical short circuit fire' (mode D3) is 0.01%, and the probability of 'toxic gas accumulation' (mode D7) is 0.5%", and "After 5 minutes, the probability of 'toxic gas accumulation' rises to 2.1%, while a new risk of 'personnel evacuation obstruction' (dynamic learning mode Dnew) emerges with a probability of 0.3%".

[0107] In some embodiments, obtaining and outputting the probability values ​​of development to each dangerous state at each future time point includes:

[0108] The probability value at each future time point is bound to the spatial location attribute and the scope of influence attribute of the corresponding hazard state in the laboratory digital twin model to generate a risk data object with spatial coordinates and risk intensity.

[0109] For each of the risk data objects, the corresponding visual representation parameters are dynamically calculated based on the probability value and the evolution stage of the corresponding hazard state. The visual representation parameters include color, transparency, basic geometric dimensions, and dynamic deformation parameters used to express the trend of risk diffusion or accumulation.

[0110] Based on the visual representation parameters, corresponding three-dimensional graphic elements are generated. All the generated three-dimensional graphic elements are used as independent rendering layers and superimposed on the basic three-dimensional scene of the laboratory digital twin model in real time to form a three-dimensional dynamic risk visualization screen.

[0111] The system queries the laboratory digital twin model based on the hazard mode identifier to obtain the predefined "spatial location attribute" and "impact range attribute" of that mode. Spatial location: This is typically the core coordinates of the hazard source or hazardous area. For example, for the "reactor overheating" mode, its location attribute is the geometric center coordinates of the reactor; for the "corridor smoke accumulation" mode, its location attribute might be the centerline coordinates of the corridor. Impact range: This defines the spatial boundary of the hazard effect. It can be a three-dimensional bounding box (cube), a sphere radius, a polygonal region, or a path along equipment or pipelines. For example, the impact range of a gas leak mode might be a sphere with the leak point as its center and the diffusion radius as its radius. The system encapsulates this five-tuple information—[time point, hazard mode, probability value, spatial location, impact range]—into a structured risk data object. This object is the basic data unit for subsequent visualization processing.

[0112] The system pre-defines a coding dictionary that maps risk data to visual characteristics, with the following rules: Color Hue: Used to distinguish hazard types; for example, red represents fire / high temperature risk, yellow represents chemical pollution risk, blue represents electrical risk, and purple represents biological risk. Color Saturation: Linearly or non-linearly correlated with probability values. The higher the probability, the more saturated (or brighter) the color. Transparency: Usually related to the proximity of the risk's occurrence or its degree of certainty. For example, the further away a predicted risk is from the current time, the higher the transparency of its visualization elements, indicating greater uncertainty or lower urgency. Basic Geometric Dimensions: Directly bound to the hazard's impact range attribute; the size of visualization elements (such as spheres, cubes, icons) scales proportionally according to the size of the impact range, intuitively reflecting the potential spatial size of the hazard. Dynamic Deformation Parameters: Used to express the dynamic trend of the risk, including pulsation frequency / amplitude, diffusion ripples, and directional streamlines. Pulsation Frequency / Amplitude: Indicates that the risk is rapidly accumulating or approaching. A sharp increase in probability values ​​within a short period may cause the corresponding visualization elements to produce a high-frequency, large-amplitude pulsation effect. Diffusion Ripples: Indicate that a risk (such as a leak or contamination) is spreading spatially. Concentric ripples spreading outward will appear on or around the surface of the visualized element. Directional Streamlines: Indicate that a risk (such as flame spread or airflow) has a clear direction of propagation. Particle streams or arrows moving in the direction of propagation will be superimposed on the visualized element. For each risk data object, the system applies the above coding rules in real time to calculate a set of specific visual parameter values. For example, for a risk object of "hydrogen leak with a 15% probability in 5 minutes", the system may generate: Color = (cyan, saturation 75%), transparency = 40%, geometric size = (sphere with a radius of 2 meters), deformation parameter = (medium-speed diffusion ripples). Based on the calculated visual representation parameters, the graphics engine dynamically generates the corresponding 3D graphic elements. Element Type: It can be a simple geometry (sphere, cube, cone for representing point / volume risk), or a tubular body or patch generated along a path (for representing line / surface risk), or a more complex icon or warning symbol. Parameter Injection: Color, transparency, size, deformation parameters, etc., are assigned to the graphic element in real time. Deformation parameters are implemented through a shader program to achieve smooth dynamic effects. Rendering layer management: All generated 3D graphic elements are managed uniformly, forming an independent risk visualization layer (or overlay layer). Overlay rendering: This risk layer and the base 3D scene of the laboratory digital twin model (precise site, equipment, and piping models) are processed in two independent rendering passes, but ultimately composited using alpha blending within the same viewport. The entire process is completed in a high-frame-rate rendering loop. Whenever the risk assessment module outputs a new full-risk probability spectrum, the visualization module immediately updates the risk data object set, recalculates visual parameters, and refreshes the graphic elements of the risk layer, thus achieving real-time, dynamic updates to the risk image.

[0113] In some embodiments, the reverse solution of the optimal control operator using an optimization algorithm includes:

[0114] The convergence objective function of the optimization algorithm is set as minimizing the sum of the projected integral values ​​of the predicted state probability distribution on the dangerous state mode within a preset future time range.

[0115] A control space is constructed using all controllable parameters of the laboratory digital twin model. A gradient-based optimization algorithm is used to perform a global iterative search within the control space to find a combination of control parameters that can reduce the value of the convergent objective function.

[0116] From the results of the global search, the control parameter adjustment amount that contributes the most to reducing the value of the convergence objective function is identified, and the mathematical expression corresponding to the control parameter adjustment amount is determined as the optimal control operator.

[0117] The optimization algorithm uses a quantified objective function as a benchmark. The core calculation logic of this function is to sum the weighted probabilities of the system evolving into various predefined hazardous states within a predetermined timeframe (e.g., the next 5 minutes). This sum represents the total expected risk the system will face in the future. Different hazardous modes (such as fire and leaks) are assigned different weights based on their severity. The fundamental goal of the optimization algorithm is to find a control strategy that minimizes this calculated total expected risk. Based on a laboratory digital twin model, the system comprehensively identifies and lists all physical parameters and logical commands that can be directly or indirectly controlled by the software system, forming a complete set of controllable variables, i.e., the control space. Typical variables include: equipment control: ventilation system wind speed setpoint, air conditioning target temperature, reactor heating power, smart valve opening commands, lighting switches; security intervention: locking commands for electronic access control in specific areas, trigger signals for audible and visual alarms, and activation commands for emergency ventilation or sprinkler systems; information guidance: warning information sent to designated personnel terminals and evacuation route indications on public displays. This high-dimensional control space defines all possible intervention methods, within which the optimization algorithm explores the optimal combination. The system employs efficient numerical optimization algorithms (such as gradient-based optimizers) to conduct automated, iterative exploration within the control space. The process begins with an initial control setting and repeatedly executes the following loop: Assuming the currently explored set of control parameters is used, the system drives a digital twin model to perform closed-loop dynamic simulation, simulating the complete evolution of the laboratory under this control strategy over a future time period; based on the simulation results, the probability of the system evolving to each hazardous state in the future is recalculated, and the current expected total future risk (i.e., the objective function value) is calculated accordingly; the algorithm analyzes the effect of the current control strategy and intelligently determines how to adjust the control parameters (e.g., whether to increase or decrease the opening of a certain damper) to reduce the total risk in the next simulation. This iteration continues until the algorithm determines that the total risk can no longer be significantly reduced, or the preset number of iterations has been reached. At this point, the specific set of control parameters found by the algorithm that minimizes the total future risk is determined as the optimal solution for this optimization search. The system analyzes the optimal parameter combinations to identify the most critical and essential control actions that contribute significantly to risk reduction. This helps focus on the most crucial and effective interventions from numerous adjustments. These core control actions and their adjustment logic are encapsulated into a clear and structured decision rule, namely the optimal control operator. The operator not only contains specific control instructions (such as "increase the fan speed of fume hood No. 1 to 80%), but more importantly, it establishes a mapping relationship from a specific risk state mode to the corresponding optimal intervention action.For example, this operator essentially defines a rule like this: "When the system predicts that the probability of 'gas leakage risk of type A' exceeds the threshold, the most effective response is to 'close upstream valve X and activate the forced exhaust ventilation in area Y'." This operator is the final product of the optimization process, and it can be directly called by the system execution module to drive real-world equipment to perform corresponding actions, thereby achieving proactive and precise intervention in predicted risks.

[0118] The intelligent early warning method for laboratory safety risks in the embodiments of this application has been described above. The computer system in the embodiments of this application will be described in detail below in conjunction with the above-described intelligent early warning method for laboratory safety risks.

[0119] Please see Figure 3 This is a schematic diagram of an exemplary hardware structure of a computer system in an embodiment of this application.

[0120] In some embodiments, the computer system 300 includes a computer device, which may be a terminal device. The computer device includes a processor 301, a memory 302, a sensor module 303, a communication module 304, an input device 305, and an output device 306 connected via a system bus. The processor 301 of the computer device provides computing and control capabilities. The memory 302 of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database is used to store data.

[0121] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0122] In some embodiments of this application, a computer-readable storage medium is provided, including instructions that, when executed on the computer system 300, cause the computer system 300 to perform the intelligent early warning method for laboratory safety risks as described in this application.

[0123] In some embodiments of this application, a computer program product is also provided, which, when run on a computer system 300, causes the computer system 300 to execute the intelligent early warning method for laboratory safety risks in the embodiments of this application.

[0124] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0125] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0126] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for intelligent early warning of laboratory safety risks, characterized in that, include: Acquire real-time monitoring data from multi-source heterogeneous sensors deployed in the laboratory, calibrate a preset laboratory digital twin model based on the real-time monitoring data, calculate and generate an initial probability distribution characterizing the current uncertainty of the system based on the state information of the calibrated laboratory digital twin model and by integrating a preset equipment failure probability model and a personnel behavior uncertainty model. By using numerical algorithms to deduce the changes in the probability density distribution of the initial probability distribution at multiple future time points, the corresponding state prediction probability distribution is obtained. The predicted probability distribution of the state at each future time point is projected onto multiple predefined dangerous state modes to calculate the probability value of developing into each dangerous state at each future time point and then output. When the probability value of any assessed dangerous state exceeds a preset threshold, the optimal control operator is solved in reverse using an optimization algorithm, and the intervention action corresponding to the optimal control operator is executed. The initial probability distribution characterizing the current uncertainty of the system is calculated and generated based on the state information of the calibrated laboratory digital twin model and by integrating a preset equipment failure probability model and a personnel behavior uncertainty model. Specifically, this includes: The current state parameters of all entities in the calibrated laboratory digital twin model are extracted and mapped into the system state space to form a state vector. Based on the state vector, a multivariate Gaussian probability distribution representing the state confidence range is generated as the base point of the initial probability distribution. The system calls the equipment failure probability model in the laboratory digital twin model that matches the equipment model and current operating conditions of the target equipment. It inputs the current running time and load rate of the target equipment into the equipment failure probability model and receives the conditional failure probability density function output by the equipment failure probability model. The conditional failure probability density function defines the continuous probability of the current equipment state drifting in the neighborhood of the healthy state. The target equipment is any one of the equipment in the laboratory digital twin model. The system calls the personnel behavior uncertainty model in the laboratory digital twin model that matches the role of the target personnel and the current operation task. It inputs the current task complexity and environmental interference into the personnel behavior uncertainty model and receives the discrete probability distribution list output by the personnel behavior uncertainty model. The discrete probability distribution list defines the possible discrete branches of the target personnel's behavior state. The target personnel is any one of the personnel in the laboratory digital twin model. Based on the physical and logical relationships between entities defined in the aforementioned laboratory digital twin model, the initial probability distribution is obtained by jointly calculating the distribution of the base point, the conditional failure probability density function of all equipment, and the discrete probability distribution list of all personnel in the system state space. Based on the physical and logical relationships between entities defined in the laboratory digital twin model, the initial probability distribution is obtained by jointly calculating the distribution of the base point, the conditional failure probability density function of all equipment, and the discrete probability distribution list of all personnel in the system state space, specifically including: The base point, the conditional failure probability density function of all equipment, and the discrete probability distribution list of all personnel are each modeled as independent factor nodes in a probability tensor network. Based on the physical connections between devices, the personnel-device operation relationships, and the environmental coupling relationships defined in the laboratory digital twin model, associated edges are defined between each of the factor nodes in the probability tensor network to construct the network topology. Based on the network topology, tensor contraction is performed on the interconnected factor nodes along the shared associated edges. All intermediate variables are eliminated by summation to obtain the joint probability distribution tensor. The joint probability distribution tensor is then normalized to obtain the initial probability distribution.

2. The method according to claim 1, characterized in that, The step of using numerical algorithms to deduce the change in probability density distribution of the initial probability distribution at multiple future time points to obtain the corresponding state prediction probability distribution specifically includes: The state variables in the laboratory are discretized into a set of state basis vectors in the system state space. The state variables include environmental parameters, equipment status, personnel positions and actions. Based on the standard operating procedures, physical laws, and equipment safety thresholds of the laboratory, a potential energy operator is constructed to characterize the inherent constraints of the system. Based on the physical connection and linkage logic between the laboratory equipment, a kinetic energy operator is constructed to characterize the trend of system state changes. The potential energy operator and the kinetic energy operator are linearly combined to form an equivalent Hamiltonian operator. Using the initial probability distribution as the initial state vector and the equivalent Hamiltonian operator as the dynamic driver, the time evolution equation is solved by numerical integration to obtain the state vectors at multiple future time points. Calculate the squared modulus of the state vectors at multiple future time points to obtain the corresponding state prediction probability distribution.

3. The method according to claim 1, characterized in that, The step of projecting the predicted probability distribution of the state at each future time point onto multiple predefined dangerous state modes to calculate and output the probability value of developing into each dangerous state at each future time point specifically includes: Regularly conduct correlation and retrospective analysis of historical monitoring data streams in the laboratory and recorded safety events to obtain abnormal evolution patterns of system states that are not predefined. The abnormal evolution mode of the system state is transformed into a new dangerous eigenstate subspace in the state space, and the new dangerous eigenstate subspace is written into the dynamically updated dangerous state mode library; During each risk assessment, the predicted probability distribution of the state at each future time point is projected onto all currently stored hazardous state patterns in the hazardous state pattern library in real time for calculation. The results of the projection calculation are output as a set of probability values ​​for the system to develop into each type of dangerous state in the model library at each future time point, forming a full risk probability spectrum.

4. The method according to claim 3, characterized in that, The probability values ​​for each dangerous state at each future time point are obtained and output, including: The probability value at each future time point is bound to the spatial location attribute and the scope of influence attribute of the corresponding hazard state in the laboratory digital twin model to generate a risk data object with spatial coordinates and risk intensity. For each of the risk data objects, the corresponding visual representation parameters are dynamically calculated based on the probability value and the evolution stage of the corresponding hazard state. The visual representation parameters include color, transparency, basic geometric dimensions, and dynamic deformation parameters used to express the trend of risk diffusion or accumulation. Based on the visual representation parameters, corresponding three-dimensional graphic elements are generated. All the generated three-dimensional graphic elements are used as independent rendering layers and superimposed on the basic three-dimensional scene of the laboratory digital twin model in real time to form a three-dimensional dynamic risk visualization screen.

5. The method according to claim 1, characterized in that, The process of finding the optimal control operator through reverse optimization algorithms includes: The convergence objective function of the optimization algorithm is set as minimizing the sum of the projected integral values ​​of the predicted state probability distribution on the dangerous state mode within a preset future time range. A control space is constructed using all controllable parameters of the laboratory digital twin model. A gradient-based optimization algorithm is used to perform a global iterative search within the control space to find a combination of control parameters that can reduce the value of the convergent objective function. From the results of the global search, the control parameter adjustment amount that contributes the most to reducing the value of the convergence objective function is identified, and the mathematical expression corresponding to the control parameter adjustment amount is determined as the optimal control operator.

6. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-5.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-5.

Images