Smart learning method for thermal operation whole-process risk prevention and control of thermal power plant

Abstract

The application discloses a thermal state operation whole-process risk prevention and control execution method for intelligent learning of a thermal power plant, and particularly relates to the technical field of safety monitoring of a thermal power plant, and the method comprises the following steps: collecting multi-dimensional thermal state data of equipment, personnel and environment and performing space-time alignment, constructing a space-time synchronous thermal state tensor and a digital twin; calculating the thermal entropy increase rate of equipment and the thermal entropy increase rate of personnel based on the entropy increase principle of non-equilibrium thermodynamics, and constructing a joint entropy increase index to unify and quantify risks; inputting data into a deep reinforcement learning model to predict an operation track, and performing risk pre-rehearsal in the digital twin; automatically generating a three-level grading prevention and control strategy of prompting, intervention and termination according to the pre-rehearsal result and executing the three-level grading prevention and control strategy, and feeding back the execution result to optimize the model. The application realizes early identification and dynamic closed-loop prevention and control of thermal state operation risks, can advance risk identification by 15-25 minutes, has an identification accuracy rate of high-risk operations of 94.7%, and effectively reduces the accident rate by 73%.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring technology for thermal power plants, and more specifically, to a method for risk prevention and control throughout the entire process of hot operations in thermal power plants based on intelligent learning. Background Technology

[0002] Hot operations in thermal power plants involve high-temperature and high-pressure equipment such as boilers and steam turbines. The working environment is complex and the risk factors are highly coupled, which can easily lead to equipment damage and personnel casualties. Existing risk prevention and control methods are mainly divided into three categories: First, early warning methods based on equipment status monitoring, which monitor equipment parameters in real time by deploying temperature and pressure sensors and alarm when they exceed the threshold; second, personnel behavior analysis methods based on video images, which issue early warnings by identifying abnormal behaviors such as workers crossing boundaries and misoperation; and third, heat exposure assessment methods based on environmental parameters, which assess the risk of heatstroke by measuring the heat radiation intensity of the working area.

[0003] However, the above methods have the following shortcomings in practical applications: Equipment monitoring, personnel management, and environmental perception are all independent of each other, lacking a deep integration mechanism for multi-source heterogeneous data, making it difficult to characterize the risk evolution pattern under the coupling effect of human-machine-environment; Early warning mechanisms often use fixed thresholds or simple statistical methods, which cannot adapt to the time-varying and nonlinear dynamic characteristics of parameters in the hot operation of thermal power plants, resulting in high false alarm and false alarm rates. Existing technologies are mostly for real-time monitoring or post-event alarms, lacking the ability to predict the evolution of future risks and unable to effectively intervene before risks occur; The lack of unified quantitative indicators for equipment risk and personnel risk makes it difficult to assess the coupled amplification effect between the two.

[0004] To address the aforementioned issues, there is an urgent need for a risk prevention and control method for thermal power plant hot operations that can achieve deep integration of human-machine-environment multi-dimensional risks, possess risk prediction capabilities, and be dynamically self-optimizing. In view of this, the present invention provides a smart learning-based full-process risk prevention and control execution method for thermal power plant hot operations. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a method for risk prevention and control of the entire process of hot operation in thermal power plants based on intelligent learning, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a smart learning-based risk prevention and control method for the entire process of hot operation in thermal power plants, comprising: Multidimensional feature data are collected throughout the entire hot operation process. The multidimensional feature data includes time series data of equipment hot parameters, time series data of workers' physiological parameters, and time series data of thermal radiation parameters of the working environment. The multidimensional feature data is then spatiotemporally aligned to construct a spatiotemporally synchronized hot tensor. A digital twin of thermal operation in a thermal power plant is constructed. The digital twin includes a sub-model of equipment thermal stress evolution and a sub-model of personnel thermal balance, which are used to dynamically map the thermal stress state of key parts of the equipment and the thermal accumulation state of the workers, respectively. Based on the principle of non-equilibrium thermodynamic entropy increase, the thermal entropy increase rate of equipment and the thermal entropy increase rate of personnel are calculated respectively, and a joint entropy increase index is constructed based on the thermal entropy increase rate of equipment and the thermal entropy increase rate of personnel. The spatiotemporal synchronous thermal tensor is input into the operation behavior prediction model based on deep reinforcement learning, and the predicted operation trajectory of the operator and the equipment operation sequence are output within a preset time period in the future. Risk simulation is performed in a digital twin by combining the predicted operation trajectory with the equipment operation sequence, and the joint entropy increase exponential change curve corresponding to the predicted operation trajectory is calculated to identify potential risk points. Based on the risk simulation results, a tiered prevention and control strategy is automatically generated and executed, and the execution results are fed back to the operational behavior prediction model for reinforcement learning and iterative optimization.

[0007] Preferably, the time-series data of the equipment thermal parameters include at least one of the following: boiler heating surface wall temperature distribution, main steam temperature change rate, main steam pressure change rate, metal thermal stress dynamics, and expansion displacement. The time-series data of the physiological parameters of the workers include at least one of the following: core body temperature, heart rate variability, skin surface heat flux density, and skeletal point data of the working posture; The time-series data of thermal radiation parameters of the working environment include at least one of the following: black sphere temperature, thermal radiation flux, and airflow organization distribution.

[0008] Preferably, the thermal entropy increase rate of the equipment is calculated according to the following formula:

[0009] in, The rate of increase of thermal entropy of the equipment. For the critical areas of the equipment in location At any time heat flux density, For the corresponding temperature.

[0010] Preferably, the rate of increase in thermal entropy of personnel is calculated according to the following formula:

[0011] in, For the rate of increase of thermal entropy of personnel, The rate of heat production from human metabolism. For external power output, To reduce heat loss from the skin For the core temperature, Skin temperature.

[0012] Preferably, the joint entropy increase index is constructed according to the following formula: in, The joint entropy increase index, The rate of increase of thermal entropy of the equipment. For the rate of increase of thermal entropy of personnel, , , These are the preset weighting coefficients.

[0013] Preferably, the operation behavior prediction model uses a Transformer encoder to extract long-term dependency features of the operation trajectory, uses an LSTM decoder to generate future operation sequences, and introduces an attention mechanism to perform heatmap saliency annotation on the device operation buttons.

[0014] Preferably, the risk simulation includes: simulating the equipment thermal stress change process and personnel thermal load change process corresponding to the operation trajectory in the digital twin, and calculating the time evolution curve of the joint entropy increase index during the simulation. When the evolution curve exceeds the preset dynamic threshold range, the corresponding operation trajectory point is marked as a potential risk point.

[0015] Preferably, the tiered prevention and control strategy includes: The first-level alert strategy uses augmented reality display devices to overlay a cloud map of the safe operating boundary and the distribution of thermal radiation intensity in the operator's field of vision; The secondary intervention strategy triggers a voice alarm or restricts device operation permissions when the distance between the predicted operation trajectory and the high-risk area is less than a preset threshold. The three-level termination strategy automatically executes a work stoppage command and generates a safe evacuation path when the joint entropy increase index exceeds the emergency threshold.

[0016] Preferably, the execution results are fed back to the operational behavior prediction model for reinforcement learning and iterative optimization, including: using the change data of the joint entropy increase index after the actual implementation of the hierarchical prevention and control strategy as a reward signal to update the network parameters of the operational behavior prediction model.

[0017] Preferably, the multidimensional feature data is subjected to spatiotemporal alignment processing, including: using a dynamic time warping algorithm to align the time series data of equipment thermal parameters, the time series data of workers' physiological parameters, and the time series data of thermal radiation parameters of the working environment at different sampling frequencies on the time axis, and mapping the aligned data to a unified digital twin space coordinate to form a spatiotemporally synchronized thermal tensor.

[0018] The technical effects and advantages of this invention are as follows: This invention constructs a human-machine-environment ternary heat exchange model, aligning and coupling the thermal stress state of equipment, the physiological thermal load of personnel, and the thermal radiation parameters of the working environment in the spatiotemporal dimension. This solves the problem that equipment monitoring, personnel management, and environmental perception are separate systems that cannot be integrated in the prior art. Furthermore, based on the principle of non-equilibrium thermodynamic entropy increase, this invention establishes unified quantitative indicators for the thermal entropy increase rate of equipment and personnel, and constructs a joint entropy increase index. From the perspective of the second law of thermodynamics, it realizes the physical unity of equipment aging and personnel fatigue, providing a quantifiable and comparable objective criterion for the risk of hot work. This invention combines deep reinforcement learning with digital twin technology. By using an operational behavior prediction model to predict and simulate the future trajectory of workers, it can identify potential hazards 15-25 minutes before the actual occurrence of a risk. Actual power plant tests have verified that the accuracy rate of high-risk operation identification using this method reaches 94.7%, which is about 35 percentage points higher than the traditional threshold method. The false alarm rate is reduced to 3.2%, which is about 80% lower than the traditional method. The average advance warning time for workers' risk of heatstroke is 18 minutes, leaving sufficient time for on-site intervention. This invention employs a closed-loop mechanism of pre-simulation-prevention-feedback-optimization, using the execution results of prevention and control strategies as reward signals to feed back to the operational behavior prediction model for reinforcement learning and iterative optimization. This achieves dynamic self-optimization of the prevention and control strategies. Furthermore, the three-tiered prevention and control strategy generated by this invention covers the entire process from risk perception to emergency response, effectively reducing the accident rate related to hot operations by approximately 73% and decreasing the number of unplanned equipment shutdowns due to thermal stress by approximately 45%. This significantly improves the overall safety level of hot operations in thermal power plants. This method is not only applicable to coal-fired power plants but, after parameter adjustments, can also be extended to high-temperature operation scenarios such as gas turbine power plants, nuclear power plant thermal maintenance, and steel smelting, demonstrating broad industry applicability. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Example

[0021] This embodiment provides a smart learning-based risk prevention and control method for the entire process of hot operations in thermal power plants. This method constructs a human-machine-environment ternary heat exchange model to realize the prediction and dynamic prevention and control of hot operation risks, specifically including the following steps.

[0022] S1. Collect multi-dimensional feature data throughout the entire hot operation process. Specifically, at the maintenance site of a 600MW coal-fired unit, temperature sensors, pressure sensors, and strain sensors are deployed at key parts of the boiler body, turbine, pipelines, and other equipment to collect time-series data of equipment hot parameters, including boiler heating surface wall temperature distribution, main steam temperature change rate, main steam pressure change rate, metal thermal stress dynamics, and expansion displacement. At the same time, wearable physiological monitoring devices are provided for the workers to collect core body temperature, heart rate variability, and skin surface heat flux density. The work posture skeletal point data is also collected through a visual recognition system. Black sphere thermometers, heat flux meters, and anemometers are deployed in the work area to collect time-series data of environmental thermal radiation parameters, including black sphere temperature, thermal radiation flux, and airflow organization distribution. Because the sampling frequencies of the three types of data mentioned above are different (equipment data is usually at the second level, physiological data is at the millisecond level, and environmental data is at the minute level), it is necessary to perform spatiotemporal alignment processing on the multidimensional feature data. In this embodiment, a dynamic time warping algorithm is used to align the time series data with different sampling frequencies on the time axis. Based on the real-time position coordinates of the operators in the digital twin space, the aligned data is mapped to a unified digital twin space coordinate to form a spatiotemporal synchronized hot state tensor. The dimension of this tensor is (time step × number of spatial positions × number of feature channels), which serves as the unified input format for subsequent models.

[0023] S2. Construct a digital twin of the thermal operation of a thermal power plant. This digital twin includes two core sub-models: an equipment thermal stress evolution sub-model and a personnel thermal balance sub-model. The equipment thermal stress evolution sub-model is based on the finite element heat transfer principle. According to the real-time collected wall temperature distribution and working fluid parameters, it dynamically maps the temperature field and stress field distribution of key parts such as boiler drum, superheater header, and main steam pipeline, and calculates the cumulative thermal fatigue damage index of key parts. For example, when the unit load changes rapidly, the model can calculate the thermal stress change caused by the temperature difference between the upper and lower walls of the drum in real time. The personnel thermal balance sub-model is built on the Gagge two-node model and is modified by combining real-time collected physiological parameters such as core body temperature and skin heat flux density. It calculates the human body heat accumulation rate and the remaining time of thermal stress tolerance. For example, when workers work continuously in high-temperature areas, the model can dynamically predict the time when their core body temperature reaches the safe threshold.

[0024] S3. Based on the principle of entropy increase in non-equilibrium thermodynamics, calculate the entropy increase rate of equipment and the entropy increase rate of personnel, respectively. The thermal entropy increase rate of the equipment is calculated using the following formula:

[0025] in, The rate of increase of thermal entropy of the equipment. For the critical areas of the equipment in location At any time heat flux density, For the corresponding temperature, The volume of the critical area of ​​the equipment is a key indicator that represents the rate of energy dissipation caused by irreversible thermal processes inside the equipment. An abnormal increase in this indicator often indicates increased thermal fatigue or potential failure of the equipment. The rate of increase in thermal entropy of personnel is calculated using the following formula:

[0026] in, For the rate of increase of thermal entropy of personnel, The rate of heat production from human metabolism. For external power output, To reduce heat loss from the skin For the core temperature, Skin temperature is an indicator that represents the entropy generation rate of the human body's thermal regulation system. An abnormally high increase in this indicator suggests that the body's heat load is too high or that the thermal regulation function is unbalanced. Based on the two entropy increase rates mentioned above, a joint entropy increase index is constructed:

[0027] in, The joint entropy increase index, The rate of increase of thermal entropy of the equipment. For the rate of increase of thermal entropy of personnel, , , The preset weighting coefficients can be optimized and determined using machine learning methods based on historical data from different work scenarios. This joint index unifies the physical nature of equipment aging and personnel fatigue from the perspective of the second law of thermodynamics, providing a unified quantitative indicator for risk warning.

[0028] S4. Input the spatiotemporal synchronous thermal tensor into the operation behavior prediction model based on deep reinforcement learning, and output the predicted operation trajectory of the operator and the equipment operation sequence within a preset time period in the future. In this embodiment, the operation behavior prediction model uses a Transformer encoder to extract long-term dependency features of the operation trajectory, uses an LSTM decoder to generate future operation sequences, and introduces an attention mechanism to perform heatmap saliency annotation on the equipment operation buttons. The model is pre-trained based on historical operation data (including operation trajectory, equipment status, personnel status, and environmental parameters) and continuously optimized through reinforcement learning during actual operation. The model output is the predicted operation path (coordinate sequence in digital twin space) and the corresponding equipment operation sequence (such as valve opening adjustment, button operation, etc.) within the next 3-5 minutes.

[0029] S5. Conduct risk simulation of the predicted operation trajectory and equipment operation sequence in a digital twin. Specifically, this includes: simulating the equipment thermal stress change process and personnel thermal load change process corresponding to the predicted operation trajectory in the digital twin, and calculating the time evolution curve of the joint entropy increase index during the simulation. When the evolution curve exceeds the preset dynamic threshold range, the corresponding operation trajectory point is marked as a potential risk point. For example, if it is predicted that the operator will enter a high-temperature area in 3 minutes and the thermal entropy increase rate of the equipment in that area is on an upward trend, then the location is marked as a high-risk point. The dynamic threshold range is determined based on statistical analysis of safe operation samples from historical operational data and can be automatically updated as the model iterates.

[0030] S6. Automatically generate and execute a tiered prevention and control strategy based on the risk simulation results. In this embodiment, the tiered prevention and control strategy includes: Level 1 alert strategy: When the risk level is low, the safe operating boundary and the heat radiation intensity distribution cloud map are superimposed on the operator's field of vision through augmented reality display devices (such as AR glasses), so that the operator can intuitively perceive the distribution of heat risks in the surrounding area. Secondary intervention strategy: When the distance between the predicted operation trajectory and the high-risk area is less than a preset threshold (e.g., 1 meter), a voice alarm or device operation permission restriction is triggered. For example, the system automatically locks the device operation buttons in the high-risk area to prevent accidental operation. Three-level termination strategy: When the joint entropy increase index exceeds the emergency threshold, the operation is automatically suspended and a safe evacuation route is generated. For example, the system notifies all workers to evacuate immediately via broadcast and displays the safe evacuation route in the AR glasses.

[0031] S7. Feed the execution results back to the operational behavior prediction model for reinforcement learning and iterative optimization. Specifically, use the change in the joint entropy increase index after the actual implementation of the hierarchical prevention and control strategy as a reward signal to update the network parameters of the operational behavior prediction model. For example, if the prevention and control strategy successfully avoids the risk event (the joint entropy increase index does not exceed the emergency threshold), a positive reward is given; if the prevention and control strategy fails to intervene effectively (a risk event occurs), a negative reward is given. Through the closed-loop optimization mechanism, the model can continuously learn more accurate operational behavior predictions and more effective prevention and control strategies, achieving the effect of intelligent learning. Example

[0032] This embodiment is basically the same as embodiment 1, except that a different implementation method is used for the spatiotemporal alignment processing of multidimensional feature data. In this embodiment, the time series data of equipment thermal parameters, the time series data of workers' physiological parameters, and the time series data of thermal radiation parameters of the working environment are first interpolated to the same time frequency (e.g., 1Hz). Then, the real-time position of the workers is tracked in the digital twin space, and the data of each sensor are weighted and interpolated according to the principle of spatial proximity to form a spatiotemporally synchronized thermal tensor. Specifically, for any spatial position, its corresponding feature value is the inverse weighted average of the distances of the sensor data of the neighboring sensors at that position. Example

[0033] This embodiment is basically the same as Embodiment 1, except that the weighting coefficients of the joint entropy increase index are determined differently. In this embodiment, the weighting coefficient , , Using a Bayesian optimization method, historical accident records are used as negative samples and safe operation records are used as positive samples. The objective function of minimizing false alarm rate and false negative rate is optimized to determine the weight coefficients. The optimized weight coefficients can better balance the contribution of equipment risk and personnel risk, and improve the accuracy of early warning. Example

[0034] This embodiment provides another way to implement an operational behavior prediction model; In this embodiment, the operation behavior prediction model adopts a graph neural network (GNN) combined with a Transformer structure. First, the layout of thermal power plant equipment and operation panels are constructed as graph structure data, where nodes represent equipment or operation buttons and edges represent spatial proximity or operational logic associations. Then, the operation trajectory of the operator is mapped as a path sequence on the graph. A graph convolutional network is used to extract spatial structure features, which are then input into a Transformer to extract temporal features. Finally, the future operation sequence is output. This structure can better utilize the spatial topological relationship between equipment and improve the accuracy of operation prediction.

[0035] It should be noted that all types of data involved in the above embodiments can be acquired using existing mature sensor technologies, such as: Boiler heating surface wall temperature distribution: can be measured using thermocouples or infrared temperature sensor arrays; Main steam temperature / pressure change rate: can be directly acquired by the DCS system; Dynamic thermal stress in metals: can be obtained by measuring with strain gauges and combining with finite element calculations; Expansion displacement: can be measured using a displacement sensor or a laser rangefinder; Core body temperature: can be measured using an ingestible temperature capsule or an infrared ear thermometer; Heart rate variability: can be measured by electrocardiogram electrodes or photoelectric heart rate sensors; Skin surface heat flux density: can be measured using a heat flux density sensor; Operational posture skeleton point data: can be obtained through depth cameras or inertial measurement units; Black ball temperature: can be measured using a black ball temperature meter; Thermal radiation flux: can be measured using a radiation heat flux meter; Airflow organization and distribution: can be measured using an anemometer and temperature sensor array.

[0036] For the formula for calculating the thermal entropy increase rate of equipment, the integration region For critical areas of equipment, such as boiler drums, superheater headers, and main steam pipes, which are prone to thermal fatigue, the heat flux density is... and temperature It can be obtained through sensor measurement or finite element simulation; For the formula of personnel thermal entropy increase rate, metabolic heat production rate The power output can be obtained by referring to a table based on the intensity of the work (such as walking, carrying, fine motor operations, etc.) or by estimating through heart rate. Under normal working conditions, the heat loss from the skin can be approximated to zero. It can be calculated based on the product of the heat flux density on the skin surface and the body surface area; For the joint entropy increase exponent, the cross term This reflects the coupling effect between equipment risk and personnel risk. That is, when the equipment condition changes drastically and the personnel heat load increases simultaneously, the risk will be amplified nonlinearly, which is consistent with the actual safety law of hot operation in thermal power plants.

[0037] In an actual test at a power plant, after applying the method of this invention: The accuracy rate for identifying high-risk operations reached 94.7%, an improvement of approximately 35 percentage points compared to the traditional threshold method; The false alarm rate was reduced to 3.2%, a decrease of approximately 80% compared to traditional methods; The average advance warning time for the risk of heatstroke among workers is 18 minutes, allowing sufficient time for intervention; The number of unplanned equipment outages due to thermal stress decreased by approximately 45%.

[0038] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart learning-based approach to risk prevention and control throughout the entire hot operation process of thermal power plants, characterized by: include: Multidimensional feature data are collected throughout the entire hot operation process. The multidimensional feature data includes time series data of equipment hot parameters, time series data of workers' physiological parameters, and time series data of thermal radiation parameters of the working environment. The multidimensional feature data is then spatiotemporally aligned to construct a spatiotemporally synchronized hot tensor. A digital twin of thermal operation in a thermal power plant is constructed. The digital twin includes a sub-model of equipment thermal stress evolution and a sub-model of personnel thermal balance, which are used to dynamically map the thermal stress state of key parts of the equipment and the thermal accumulation state of the workers, respectively. Based on the principle of non-equilibrium thermodynamic entropy increase, the thermal entropy increase rate of equipment and the thermal entropy increase rate of personnel are calculated respectively, and a joint entropy increase index is constructed based on the thermal entropy increase rate of equipment and the thermal entropy increase rate of personnel. The spatiotemporal synchronous thermal tensor is input into the operation behavior prediction model based on deep reinforcement learning, and the predicted operation trajectory of the operator and the equipment operation sequence are output within a preset time period in the future. Risk simulation is performed in a digital twin by combining the predicted operation trajectory with the equipment operation sequence, and the joint entropy increase exponential change curve corresponding to the predicted operation trajectory is calculated to identify potential risk points. Based on the risk simulation results, a tiered prevention and control strategy is automatically generated and executed, and the execution results are fed back to the operational behavior prediction model for reinforcement learning and iterative optimization.

2. The method according to claim 1, characterized in that: The time-series data of the equipment's thermal parameters include at least one of the following: boiler heating surface wall temperature distribution, main steam temperature change rate, main steam pressure change rate, metal thermal stress dynamics, and expansion displacement. The time-series data of the physiological parameters of the workers include at least one of the following: core body temperature, heart rate variability, skin surface heat flux density, and skeletal point data of the working posture; The time-series data of thermal radiation parameters of the working environment include at least one of the following: black sphere temperature, thermal radiation flux, and airflow organization distribution.

3. The method according to claim 1, characterized in that: The thermal entropy increase rate of the equipment is calculated using the following formula: ； in, The rate of increase of thermal entropy of the equipment. For the critical areas of the equipment in location At any time heat flux density, For the corresponding temperature.

4. The method according to claim 1, characterized in that: The rate of increase in thermal entropy of personnel is calculated using the following formula: ； in, For the rate of increase of thermal entropy of personnel, The rate of heat production from human metabolism. For external power output, To reduce heat loss from the skin For the core temperature, Skin temperature.

5. The method according to claim 1, characterized in that: The joint entropy increase index is constructed according to the following formula: ； in, The joint entropy increase index, The rate of increase of thermal entropy of the equipment. For the rate of increase of thermal entropy of personnel, , , These are the preset weighting coefficients.

6. The method according to claim 1, characterized in that: The operation behavior prediction model uses a Transformer encoder to extract long-term dependency features of the operation trajectory, uses an LSTM decoder to generate future operation sequences, and introduces an attention mechanism to perform heatmap saliency annotation on the device operation buttons.

7. The method according to claim 1, characterized in that: Risk simulation includes: simulating and predicting the changes in equipment thermal stress and personnel thermal load corresponding to the operation trajectory in a digital twin, and calculating the time-series evolution curve of the joint entropy increase index during the simulation. When the evolution curve exceeds the preset dynamic threshold range, the corresponding operation trajectory point is marked as a potential risk point.

8. The method according to claim 1, characterized in that: The tiered prevention and control strategy includes: The first-level alert strategy uses augmented reality display devices to overlay a cloud map of the safe operating boundary and the distribution of thermal radiation intensity in the operator's field of vision; The secondary intervention strategy triggers a voice alarm or restricts device operation permissions when the distance between the predicted operation trajectory and the high-risk area is less than a preset threshold. The three-level termination strategy automatically executes a work stoppage command and generates a safe evacuation path when the joint entropy increase index exceeds the emergency threshold.

9. The method according to claim 1, characterized in that: The execution results are fed back to the operational behavior prediction model for reinforcement learning and iterative optimization, including: using the change data of the joint entropy increase index after the actual implementation of the hierarchical prevention and control strategy as a reward signal to update the network parameters of the operational behavior prediction model.

10. The method according to claim 1, characterized in that: Spatiotemporal alignment processing of multidimensional feature data includes: using a dynamic time warping algorithm to align time-series data of equipment thermal parameters, physiological parameters of workers, and thermal radiation parameters of the work environment at different sampling frequencies on the time axis, and mapping the aligned data to a unified digital twin space coordinate to form a spatiotemporally synchronized thermal tensor.

Images