Fire-fighting ventilation emergency intelligent regulation method and system based on reinforcement learning

Abstract

The present application relates to the technical field of reinforcement learning, in particular to a fire ventilation emergency intelligent regulation and control method and system based on reinforcement learning; the method of the present application firstly calculates a fire condition comprehensive intensity index after fusing multi-source data and standardizing preprocessing, realizing accurate quantification of fire danger degree; then, by coupling fire intensity and ventilation system condition adaptation degree, three fire grades are divided, the grading result is matched with the actual regulation and control ability of the equipment, and the disconnection between grading and execution is avoided; then, a grading differentiated strategy generation mechanism is adopted, the primary stage uses a quick lookup table method to ensure emergency response speed, the middle and high stages use a differentiated reward reinforcement learning to realize multi-objective optimization, and the real-time performance of low-risk scenarios and the regulation and control accuracy of high-risk scenarios are considered; finally, by ventilation hydraulic network decoupling and equipment instruction mapping, the global regulation and control target is converted into executable equipment parameters, ensuring that the strategy is compliant and controllable, and comprehensively improving the scientificity and effectiveness of fire ventilation emergency regulation and control.

Description

Technical Field

[0001] This invention relates to the field of reinforcement learning technology, specifically to a fire ventilation emergency intelligent control method and system based on reinforcement learning. Background Technology

[0002] In building fires, smoke is a major cause of casualties, and fire ventilation systems are core facilities for controlling smoke diffusion and ensuring personnel evacuation. Traditional fire ventilation systems mostly use fixed threshold on / off control, triggering fan start / stop and damper switching based solely on single-point smoke concentration or temperature. This results in low control precision, delayed response, and an inability to adapt to the dynamic development of a fire, easily leading to problems such as insufficient smoke extraction and smoke backflow.

[0003] Some existing technologies introduce intelligent algorithms to optimize ventilation control, but the classification of fire conditions and fire levels is based on a single dimension, relying solely on fire detection parameters. This leads to a disconnect between the hazard level classification and the actual capabilities of the equipment, resulting in poor adaptability of the control strategies. At the same time, emergency strategies lack differentiated design based on different levels. In minor fire scenarios, the use of complex intelligent algorithms results in wasted computing power, while in severe fire scenarios, the optimization objective is singular, failing to balance smoke extraction efficiency, smoke backflow prevention, personnel evacuation and protection, and energy consumption control. Summary of the Invention

[0004] The purpose of this invention is to provide a fire ventilation emergency intelligent control method and system based on reinforcement learning.

[0005] The technical solution of this invention is as follows: A fire ventilation emergency intelligent control method based on reinforcement learning includes the following operations: S1. Collect fire sensing data, ventilation system status data and building space basic data, preprocess them to obtain a standardized sensing dataset, and calculate the comprehensive intensity index of fire conditions. S2. Based on the comprehensive intensity index of fire conditions and the status data of the ventilation system, the adaptability index of the ventilation system is obtained. Combined with the comprehensive intensity index of fire conditions, the coupling risk value is calculated. According to the coupling risk value, the fire protection levels are divided into primary, intermediate and advanced levels. S3. For the basic fire protection level, a fast lookup table correction method using three-dimensional spatial grid matching is adopted to obtain the ventilation control strategy, which is mapped into equipment control commands and issued for execution. For the intermediate and advanced fire protection levels, a deep reinforcement learning model with differentiated optimization objectives is invoked. The strategy input state space is constructed using the comprehensive fire intensity index, ventilation system status data, and building space basic data. The strategy is iteratively trained through a hierarchical multi-objective reward function to generate multiple sets of candidate control strategies. The optimal control strategy is selected based on the reward score and then executed in S4. S4. Based on the preset ventilation hydraulic network model, the global objective of the optimal control strategy is decoupled into the parameters of each branch equipment, mapped into equipment control commands, and issued for execution.

[0006] In S3, the fast lookup table correction method for three-dimensional spatial grid matching is as follows: the interior space of the building is divided into three-dimensional grids of equal granularity. Based on the relationship between the grid position and the distribution of ventilation openings, multiple sets of preset benchmark control parameter combinations are matched for each grid. The grid coordinates of the fire source and the direction of smoke diffusion are obtained in real time. The ventilation control parameter combination of the corresponding grid is retrieved and matched. The ventilation control parameter combination is linearly corrected based on the deviation between the current smoke concentration and the flame temperature to obtain the ventilation control strategy.

[0007] The deep reinforcement learning model processing procedure corresponding to the intermediate fire response level is as follows: Based on the building space topology and ventilation hydraulic network topology, a causal mapping library of fire source spatial nodes and ventilation branches is constructed to determine the range of adjustable equipment and the causal correlation degree of branches corresponding to each fire source location; a causal mask is applied to the model action space in real time according to the spatial node where the fire source is located to shield the equipment action dimension of non-associated branches; a target operating range of smoke concentration in the fire area and total smoke exhaust volume of the system is preset; a strategy decision is triggered and the equipment parameter adjustment amount is output as a candidate control strategy only when the core smoke index deviates from the target range, the comprehensive intensity index of the fire condition fluctuates beyond the fluctuation threshold, or the risk of smoke backflow is detected; otherwise, the current equipment operating state is maintained; the hierarchical multi-objective reward function is obtained based on the basic reward for meeting the standard, the smoke exhaust efficiency approach reward, the causal differential backflow penalty, and the unit smoke exhaust energy consumption penalty.

[0008] In S3, the deep reinforcement learning model processing procedure corresponding to the advanced fire response level is as follows: A hierarchical reinforcement learning framework is constructed, comprising an upper-level option strategy layer, a lower-level action execution layer, and a global value commentator. Three types of high-rise control options are preset: smoke extraction, air supply balancing, and evacuation and traffic flow maintenance. The safety urgency of each evacuation route and the overall evacuation safety urgency are calculated in real time, serving as the core input state for the upper-level option strategy, dynamically matching the optimal control option corresponding to the current fire situation. The lower-level action execution layer outputs initial control parameters based on the selected control option. Before outputting the parameters, a Lyapunov safety barrier function is used for verification, correcting action parameters that exceed the evacuation safety threshold to within the safety boundary, and outputting candidate control strategies. The hierarchical multi-objective reward function is obtained by weighting the upper-level option matching reward, the lower-level core objective execution reward, and the evacuation safety margin reward.

[0009] It also includes a strategy pre-training method for building fire dynamics simulation-driven approaches. This method generates a multi-scenario fire ventilation coupling sample set through a fire dynamics simulator and uses a meta-learning framework to perform offline pre-training of deep reinforcement learning models corresponding to intermediate and advanced fire protection levels, respectively.

[0010] The specific steps of the strategy pre-training method are as follows: Based on the fire dynamics simulator, traverse the combination scenarios of different fire source locations, different fire intensities, different initial ventilation states, and different building layouts to generate multiple sets of coupled time-series data of the entire fire development process and ventilation control response, forming a simulation sample library; use different scenarios in the simulation sample library as independent training tasks to perform multi-task pre-training on the deep reinforcement learning model to obtain initialization parameters; based on the digital twin model of the target building, generate simulation data of the real building scene, perform domain adaptation fine-tuning on the pre-trained model, reduce the distribution difference between the simulation domain and the real building scene to below the preset difference threshold, and obtain the pre-trained model.

[0011] The preprocessing of S1 includes unifying the spatiotemporal reference of fire perception data, ventilation system status data and building space basic data, removing triple constraint outliers, completing and resampling spatiotemporally coupled time series, normalizing node differences and encapsulating structured output.

[0012] A reinforcement learning-based intelligent emergency control system for fire ventilation, used to implement the aforementioned reinforcement learning-based intelligent emergency control method for fire ventilation, includes: The fire condition comprehensive intensity index generation module is used to collect fire sensing data, ventilation system status data and building space basic data, and after preprocessing, obtain a standardized sensing dataset to calculate the fire condition comprehensive intensity index. The fire protection rating module is used to obtain the ventilation system condition adaptability index based on the comprehensive fire intensity index and ventilation system status data, and to calculate the coupling risk value in combination with the comprehensive fire intensity index; based on the coupling risk value, it classifies the fire protection into three levels: primary, intermediate, and advanced. The first equipment instruction generation module is used to obtain ventilation control strategies for basic fire protection levels by using a fast lookup table correction method based on three-dimensional spatial grid matching, which is then mapped into equipment control instructions and issued for execution. For intermediate and advanced fire protection levels, a deep reinforcement learning model with differentiated optimization objectives is invoked. The strategy input state space is constructed using the comprehensive fire intensity index, ventilation system status data, and basic building space data. The strategy is iteratively trained through a hierarchical multi-objective reward function to generate multiple sets of candidate control strategies. The optimal control strategy is selected based on the reward score and then executed by the second equipment instruction generation module. The second equipment instruction generation module is used to decouple the global objective of the optimal control strategy into the parameters of each branch equipment based on the preset ventilation hydraulic network model, map them into equipment control instructions, and issue them for execution.

[0013] A fire ventilation emergency intelligent control device based on reinforcement learning includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the above-mentioned fire ventilation emergency intelligent control method based on reinforcement learning.

[0014] A computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described reinforcement learning-based intelligent emergency control method for fire ventilation.

[0015] The beneficial effects of this invention are as follows: This invention provides a reinforcement learning-based intelligent emergency control method for fire ventilation. First, it integrates multi-source data and calculates a comprehensive fire intensity index after standardized preprocessing, overcoming the shortcomings of traditional methods such as inconsistent data quality and a single dimension for determining operating conditions, thus achieving precise quantification of fire hazard levels. Second, it divides fire intensity into three fire levels by coupling fire intensity with the adaptability of ventilation system operating conditions, breaking away from the traditional model of determining levels solely based on disaster severity, ensuring that the classification results match the actual control capabilities of the equipment, and avoiding a disconnect between classification and execution. Third, it employs a differentiated strategy generation mechanism, using a fast lookup table method for the primary level to ensure emergency response speed, and reinforcement learning with differentiated rewards for intermediate and advanced levels to achieve multi-objective optimization, balancing real-time performance in low-risk scenarios with control precision in high-risk scenarios, and solving the problems of redundant computing power and misaligned optimization objectives in single algorithms. Finally, through decoupling the ventilation-hydraulic network and mapping equipment commands, it transforms global control objectives into executable equipment parameters, ensuring that the strategy implementation is compliant and controllable, comprehensively improving the scientific nature and effectiveness of fire ventilation emergency control. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the exemplary embodiments of this application clearer, the technical solutions in the exemplary embodiments of this application are described clearly and completely below. Obviously, the described exemplary embodiments are only some embodiments of this application, and not all embodiments.

[0017] This embodiment provides a fire ventilation emergency intelligent control method based on reinforcement learning, including the following operations: S1. Collect fire sensing data, ventilation system status data and building space basic data, preprocess them to obtain a standardized sensing dataset, and calculate the comprehensive intensity index of fire conditions. S2. Based on the comprehensive intensity index of fire conditions and the status data of the ventilation system, the adaptability index of the ventilation system is obtained. Combined with the comprehensive intensity index of fire conditions, the coupling risk value is calculated. According to the coupling risk value, the fire protection levels are divided into primary, intermediate and advanced levels. S3. For the basic fire protection level, a rapid lookup table correction method using three-dimensional spatial grid matching is adopted to obtain the ventilation control strategy, which is then mapped into equipment control commands and issued for execution. For intermediate and high fire protection levels, a deep reinforcement learning model with differentiated optimization objectives is invoked. The strategy input state space is constructed using the comprehensive fire intensity index, ventilation system status data, and building space basic data. The strategy is iteratively trained through a hierarchical multi-objective reward function to generate multiple sets of candidate control strategies. The optimal control strategy is selected based on the reward score and then executed in S4. S4. Based on the preset ventilation hydraulic network model, the global objective of the optimal control strategy is decoupled into the parameters of each branch equipment, mapped into equipment control commands and issued for execution. The specific steps are detailed below.

[0018] S1. Collect fire sensing data, ventilation system status data, and basic building space data. After preprocessing, obtain a standardized sensing dataset and calculate the comprehensive intensity index of the fire condition.

[0019] First, fire detection data, ventilation system status data, and basic building space data are collected. Fire detection data includes the location of the fire source, flame temperature, PM2.5 smoke concentration, CO concentration, and smoke diffusion velocity and direction, which can be obtained through fire detectors. Ventilation system status data includes real-time monitoring values ​​of fan speed, fan start / stop status, damper opening, and air velocity and pressure within the ventilation ducts, which can be obtained through a ventilation system monitoring platform. Basic building space data includes the building floor structure, ventilation duct layout, evacuation route locations, and distribution of digitally sealed areas, which can be obtained by extracting and analyzing the building BIM model.

[0020] Preprocessing operations can be achieved through outlier removal, normalization with physical constraints, and time series completion.

[0021] This embodiment takes into account the complex topology of building spaces, the cross-layer spread of smoke, the multi-layered network of ventilation systems, and the strong heterogeneity of the acquired multi-source data. Conventional statistical preprocessing cannot guarantee the consistency between the data and spatial physical laws. Therefore, the preprocessing process designed in this embodiment is as follows: First, the fire sensing data, ventilation system status data, and basic building space data are unified in terms of spatiotemporal reference to complete the spatial mapping and timestamp reference alignment of the multi-source data; then, triple constraint outlier removal is performed, marking outlier data that does not meet the constraints as null values ​​and retaining the spatiotemporal index; subsequently, spatiotemporal coupling time series completion and resampling are performed to fill null values ​​and obtain a complete dataset aligned with the time dimension; then, node differential normalization is performed so that the normalized values ​​directly correspond to the node risk level and equipment load level; finally, structured encapsulation and output are performed.

[0022] The unified spatiotemporal benchmark operation is as follows: Based on the building space fundamental data, the entire building area is divided into spatial nodes including smoke control zones, evacuation route sections, and ventilation duct branches. A spatial topological adjacency matrix is ​​constructed, and the static spatial attributes of each node are pre-calculated, including node betweenness centrality, node path connectivity coefficient, node functional attributes, and node safety threshold parameters. Fire sensing data and ventilation system status data are bound to the corresponding spatial node IDs to complete the spatial mapping and timestamp benchmark alignment of multi-source data, resulting in a unified spatiotemporal benchmark dataset. This method establishes a unified spatiotemporal coordinate system from the bottom layer, deeply binding three types of heterogeneous data with the building space, solving the problem of the disconnect between sensing data and spatial risks in complex buildings, and providing a unified benchmark for subsequent spatial dimension verification, completion, and intensity calculation.

[0023] The triple constraint outlier removal operation is as follows: the unified spatiotemporal benchmark dataset is processed sequentially by the 3σ criterion to achieve statistical initial screening, spatial neighborhood consistency verification to achieve spatial secondary screening, and physical mechanism compliance verification to achieve mechanism-level final screening. Outliers that do not conform to statistical distribution, spatial gradient and physical mechanism are marked as null values ​​and their spatiotemporal indexes are retained to obtain the constraint verification dataset.

[0024] The spatial neighborhood consistency verification process is as follows: Based on the spatial topological adjacency matrix, consistency verification is performed on the same type of data of the same node and neighboring nodes. If the flue gas concentration or temperature data of a certain node deviates from the mean of neighboring nodes by more than a preset deviation threshold and does not conform to the preset spatial gradient law driven by the fire source location, for example, the concentration of nodes far away from the fire source is much higher than that around the fire source, then it is marked as a spatial outlier. If the wind speed or wind pressure data of the same duct upstream and downstream show a reverse sudden change and do not conform to the preset hydraulic resistance law, it is marked as ventilation data anomaly.

[0025] The physical mechanism compliance verification process is as follows: For the ventilation system status data, establish a parameter matching relationship library based on the ventilation hydraulic characteristic equation: verify the matching relationship between fan speed, air volume, and air pressure. If the parameter combination is not within the reasonable range of the fan performance curve, preferably deviating by 30%, it is judged as abnormal; verify the positive correlation between the opening of the air valve and the branch wind speed. If the air valve is fully closed but the branch wind speed is still higher than the wind speed threshold, it is judged as abnormal.

[0026] The spatiotemporal coupled time-series completion and resampling operation is as follows: Hierarchical completion is performed on the null value sequences in the constraint verification dataset: short missing sequences with consecutive missing values ​​≤ 3 time steps (based on a 5s time step) are completed using linear interpolation; long missing sequences with consecutive missing values ​​> 3 time steps are completed according to their data type, including completing long missing fire perception data using the neighborhood inverse distance weighting method, and completing long missing ventilation system data based on a ventilation hydraulic network model using mechanistic decomposition; after completion, resampling is performed at a fixed time step to obtain a spatiotemporally aligned dataset. Compared to pure time interpolation, this significantly improves accuracy, while the unified time step ensures the temporal synchronization of subsequent intensity calculations and strategy generation. When performing neighborhood inverse distance weighted interpolation on smoke sensing data, spatial topology constraints of building fire compartments are introduced. Only neighboring detector nodes that are spatially connected within the same fire compartment and are not completely separated by physical walls are selected to participate in weighted interpolation. Cross-enclosed fire compartments are prohibited from calling external measuring point data for interpolation to avoid illegal cross-compartment interpolation, which conforms to the physical diffusion law that smoke can only spread through channels and cannot penetrate enclosed walls.

[0027] The node differentiation normalization operation is as follows: based on the functional attributes and safety thresholds of different spatial nodes, differentiated normalization upper and lower bounds are set for the fire perception parameters and ventilation status parameters of each node in the spatiotemporal aligned dataset, and all parameters are uniformly mapped to the [0,1] interval, so that the normalized values ​​directly correspond to the node risk level and equipment load level, thus obtaining the normalized dataset.

[0028] The structured encapsulation output operation is as follows: the normalized dataset is structured and encapsulated according to the format of timestamp, spatial node ID, fire sensing parameter set, ventilation status parameter set, and spatial static attribute set, and the output is a standardized sensing dataset.

[0029] Then, based on the standardized sensing dataset, the comprehensive fire intensity index, which reflects the fire situation, is calculated using the following formula: , , , , The comprehensive intensity index of fire conditions. The heat intensity factor of the fire source. This refers to the real-time flame core temperature. Standard ambient temperature, The fire source spatial hazard level is derived from the betweenness centrality of nodes in the standardized perception dataset. It is obtained by normalizing the betweenness centrality of the complex network of nodes in the fire source space and represents the degree of influence of the fire source location on the overall building space. The concentration factor of flue gas toxicity. PM2.5 particulate matter concentration This refers to the carbon monoxide concentration. These are the first toxicity weighting coefficient and the second toxicity weighting coefficient, respectively. The threshold for the safe concentration of flue gas that is tolerable to personnel; As a risk factor for flue gas diffusion, The modulus of flue gas diffusion velocity. The angle between the main direction of flue gas diffusion and the axis of the nearest evacuation channel. The diffusion path connectivity coefficient is taken from the node path connectivity coefficient in the standardized sensing dataset and is obtained by normalizing the number of spatially connected nodes on the diffusion path. These are weights for heat intensity of the fire source, toxic concentration of flue gas, and risk of flue gas diffusion.

[0030] This embodiment integrates fire perception data, ventilation system status data, and basic building space data. After preprocessing, it effectively improves data reliability and spatiotemporal consistency. The comprehensive fire intensity index calculated based on this can accurately quantify the actual degree of fire danger, providing a basis for subsequent graded response and ventilation control.

[0031] S2. Based on the comprehensive intensity index of fire conditions and the status data of the ventilation system, the adaptability index of the ventilation system is obtained. Combined with the comprehensive intensity index of fire conditions, the coupling risk value is calculated. According to the coupling risk value, the fire protection levels are divided into primary, intermediate and advanced levels.

[0032] Traditional fire protection rating systems generally use single-point fire parameters, such as temperature and smoke concentration threshold triggering modes, to classify only the severity of the fire and completely ignore the actual load-bearing capacity of the ventilation system. This easily leads to a distorted classification where "under the same level of disaster, if ventilation is intact, the risk is controllable, but if ventilation fails, it quickly becomes uncontrollable." In order to solve the above technical problems, this embodiment first obtains the ventilation system condition adaptability index based on the comprehensive intensity index of the fire condition and the status data of the ventilation system.

[0033] The formula for calculating the ventilation system's operational suitability index is as follows: , The current actual total smoke exhaust volume of the ventilation system is calculated based on fan speed, damper opening, and duct velocity. The theoretical smoke exhaust volume required under fire conditions is calculated from the comprehensive fire intensity index F through a preset mapping function; The health status of the ventilation system is calculated by weighting the duct pressure deviation rate and the wind speed deviation rate, which characterizes the actual output efficiency of the ventilation system.

[0034] Then, based on the comprehensive fire intensity index and the ventilation system adaptability index, the coupling risk value is calculated to address the disconnect between fire classification and equipment performance capabilities, facilitating the matching of subsequent fire protection ratings with the adjustable range of the ventilation system. The formula for calculating the coupling risk value is as follows: , For coupling risk value, The comprehensive intensity index of fire conditions. This is the adaptability index for the ventilation system's operating conditions.

[0035] Finally, based on the coupled risk value, three fire response levels are divided into primary, intermediate, and advanced levels. The higher the coupled risk value, the higher the level. The classification results directly correspond to the actual fire safety risks of the building and can accurately identify hidden high-risk scenarios such as "small fires but ventilation failure".

[0036] S3. For the primary fire protection level, a fast lookup table correction method using three-dimensional spatial grid matching is adopted to obtain the ventilation control strategy, which is mapped into equipment control commands and issued for execution. A deep reinforcement learning model with differentiated optimization objectives is invoked to construct the strategy input state space using the comprehensive fire intensity index, ventilation system status data, and building space basic data. The strategy is iteratively trained through a hierarchical multi-objective reward function to generate multiple sets of candidate control strategies. The optimal control strategy is selected based on the reward score and executed in S4.

[0037] For basic fire safety levels, a rapid lookup table correction method based on three-dimensional spatial grid matching is employed to obtain ventilation control strategies, which are then mapped to equipment control commands and executed. Specifically, the rapid lookup table correction method based on three-dimensional spatial grid matching involves dividing the building's interior space into three-dimensional grids of equal granularity. Combining the grid location with the distribution of ventilation openings, multiple preset sets of baseline control parameter combinations are matched to each grid. The grid coordinates of the fire source and the smoke diffusion direction are acquired in real time. The corresponding ventilation control parameter combinations for each grid are retrieved and matched. Then, the deviation between the current smoke concentration and flame temperature is used to linearly correct the ventilation control parameter combinations, resulting in the ventilation control strategy. This method requires no complex iterative calculations, has low response latency, and is suitable for the rapid response needs of basic fires.

[0038] The specific steps for retrieving and matching the ventilation control parameter combination corresponding to the grid are as follows: The building's interior space is pre-divided into three-dimensional grids of equal granularity. Multiple sets of baseline control parameter combinations corresponding to different initial fire levels are pre-set for each grid, and a mapping table of adjacent diffusing grids covering the horizontal and vertical directions is constructed. Each set of baseline control parameter combinations includes the baseline speed multiplier of the smoke exhaust fan in the corresponding smoke control zone, the baseline opening value of the air valve around the ignition point, the baseline airflow matching ratio of the make-up air system, and the associated smoke exhaust port opening / closing status combination. The three-dimensional coordinates of the fire source are acquired in real time, and the central grid where the fire source is located is matched and determined. Based on the real-time main direction of smoke diffusion, the downstream diffusing grids are selected from the mapping table of adjacent diffusing grids of the central grid. If a vertical diffusion trend exists, the corresponding grids on the upper and lower floors are simultaneously included. The weight coefficients of the central grid and the influencing grids are allocated inversely according to the spatial distance. The baseline control parameter combinations corresponding to the same fire level for the central grid and the diffusing influence grids are retrieved, and the parameters are weighted and fused one by one to obtain the final matched ventilation control parameter combination. This embodiment of the method abandons the traditional crude mode of single-grid single-point table lookup. Combining the characteristics of the initial fire smoke already beginning to spread in a directional manner, it takes the fire source center grid as the core and the downstream diffusion grid as a supplement. By spatial distance weighted fusion of multi-grid reference parameters, it retains the low latency advantage of the table lookup method, while also taking into account the smoke diffusion trend. The matching accuracy is significantly improved, and no complex calculation is required, which is fully adapted to the rapid response requirements of the initial fire.

[0039] The specific steps for linearly correcting the ventilation control parameter combination based on the deviation between the current flue gas concentration and flame temperature are as follows: First, pre-calibrate the reference flame temperature and reference flue gas concentration corresponding to each grid ventilation control parameter combination; calculate the temperature deviation between the real-time flame temperature and the reference flame temperature, and the concentration deviation between the real-time flue gas concentration and the reference flue gas concentration; based on a preset piecewise correction coefficient table, match the temperature correction coefficient corresponding to the temperature deviation, and the concentration correction coefficient corresponding to the concentration deviation; apply differentiated weighted two-dimensional corrections to the fan speed parameter and the damper opening parameter in the ventilation control parameter combination, with the fan speed parameter primarily corrected for temperature deviation and secondarily for concentration deviation, and the damper opening parameter primarily corrected for concentration deviation and secondarily for temperature deviation, to obtain preliminary correction parameters; perform equipment rated limit clamping verification on the preliminary correction parameters to obtain the corrected ventilation control parameters, which serve as the ventilation control strategy. Clamping verification includes performing equipment rated boundary verification on all corrected parameters, ensuring that the fan speed does not exceed the rated speed and is not lower than the minimum operating speed, and that the damper opening is limited to a preset opening range; parameters exceeding the limits are directly clamped to the boundary value.

[0040] The deep reinforcement learning model corresponding to the intermediate fire protection level is implemented using the soft Actor-Critic algorithm with event-triggered causal masking. The specific steps are detailed below.

[0041] Step 1: Based on the ventilation network topology and building space information in the ventilation system status data and building space basic data, construct a causal mapping library between fire source nodes and ventilation branches. Determine the range of adjustable equipment associated with each fire source location and the causal correlation degree of the branches, and clarify the range of smoke exhaust branches and make-up air branches that each fire source node can affect. The branch causal correlation degree is obtained by inversely normalizing the topological distance between the spatial node where the branch is located and the fire source node; the higher the correlation degree, the larger the value.

[0042] Step 2: A basic strategy network is built using a soft Actor-Critic algorithm. Causal masks are applied to the action space in real time based on the node where the fire source is located, masking the action dimensions of equipment in non-associated branches. This directly masks the action dimensions of equipment in branches outside the fire source's influence range, retaining only the adjustable parameters of associated branches, significantly compressing the decision space. The input states of the basic strategy network include the overall fire intensity, smoke concentration at each spatial node, and ventilation system operating parameters; the original action space includes continuous values ​​of all fan speed adjustments and all valve opening adjustments.

[0043] Step 3: Preset the target operating range for smoke concentration in the fire zone and system smoke exhaust volume. Only when the core indicators deviate from the corresponding target range, or the fire condition comprehensive intensity index fluctuates beyond the fluctuation threshold, or backflow risk is detected, a strategy decision is triggered, and the equipment parameter adjustment amount is output as a candidate control strategy; otherwise, the current operating state is maintained. The core indicators include smoke concentration in the fire zone, total system smoke exhaust volume, and single-step fluctuation range of the fire condition comprehensive intensity index.

[0044] In this process, the hierarchical multi-objective reward function adopts an interval-type hierarchical reward function for strategy iterative optimization. The total reward is composed of the basic reward for meeting the standard, the smoke emission efficiency reward, the causal differentiation penalty and the energy consumption penalty weighted. The candidate control strategy with the highest reward value is selected as the optimal control strategy according to the total reward value sorted from high to low.

[0045] The formula for calculating the bonus function for intermediate fire protection level is: , , , , These are, respectively, basic reward for meeting standards, smoke emission efficiency reward, causal difference penalty, and energy consumption penalty. , , , These are the basic reward weights for meeting standards, the reward weights for smoke emission efficiency, the penalty weights for causal differentiation, and the penalty weights for energy consumption.

[0046] Basic performance bonuses used to incentivize operating conditions to remain within the target range The calculation formula is as follows: , To achieve the target pullback reward, a one-time reward strategy is used to pull the indicator back to the normal range. This is the motion incentive coefficient. To control the amplitude of actions, such as changes in fan speed and adjustments in valve opening, For single-cycle steady-state maintenance rewards, the greater the magnitude of the action to eliminate risks, the higher the reward; simply maintaining the operating condition in place only yields a small fixed reward.

[0047] Smoke extraction efficiency rewards used to positively incentivize increased smoke extraction volume. The calculation formula is as follows: , Triggering time t The total smoke exhaust volume of the system. The median value of the target smoke emission range is used as the benchmark for optimal smoke emission, aiming to achieve a smoke emission value that closely matches the optimal value.

[0048] Causal Differentiation Penalty for Constraints Related to Backflow Risk The calculation formula is as follows: , This is a global baseline penalty item with a fixed empirical value. It applies a basic penalty whenever the risk of backflow is triggered, ensuring the baseline strength of safety constraints and preventing insufficient penalties or constraint failure due to low branch contribution. The base coefficient for backflow penalty, The branch road that caused the backflow i The degree of causal relationship of the branches.

[0049] Energy consumption penalty as an auxiliary optimization item The calculation formula is as follows: , To trigger the average power of the ventilation system within the cycle, This represents the system's real-time total power. The baseline unit exhaust energy consumption is the energy consumption required to exhaust a unit volume of air under rated operating conditions. As a normalization benchmark, this calculation measures the energy consumption level of a company's penalty unit exhaust volume, guiding strategies to achieve the same exhaust effect with lower energy consumption, rather than simply reducing total power.

[0050] The specific execution steps of the deep reinforcement learning model corresponding to the advanced fire protection level are as follows.

[0051] Step 1: Construct a hierarchical reinforcement learning framework that includes an upper-level option policy, a lower-level action policy, and a global value critic. Specifically, the Option-Critic hierarchical reinforcement learning framework is adopted, which is divided into three parts from top to bottom: an upper-level option policy layer, a lower-level action execution layer, and a global value critic. The framework is trained jointly in an end-to-end manner.

[0052] The system includes three preset high-rise control options: primary focus on smoke extraction, supplementary ventilation and balancing, and evacuation and ventilation maintenance, each corresponding to a different fire response objective at a different stage of the fire.

[0053] The core objective of the primary smoke extraction system is to maximize the smoke extraction rate in the fire area, adapting to stages where the fire is concentrated and evacuation pressure is relatively low. The core objective of the supplementary airflow balancing system is to balance the wind pressure in each smoke control zone and suppress backflow of smoke across zones, adapting to stages where the smoke diffusion range expands. The core objective of evacuation and traffic flow maintenance is to forcibly maintain the smoke concentration in all evacuation routes below the safe threshold, adapting to critical stages of densely populated evacuations.

[0054] Step 2: Calculate the safety urgency of each evacuation route and the overall evacuation safety urgency in real time. Use this as the core input of the upper-level option strategy to dynamically match the optimal control option for the current fire situation. The lower-level action strategy outputs the initial parameters of the control equipment based on the selected option. Before output, the parameters are verified by the Lyapunov safety barrier function. Actions that exceed the evacuation safety threshold are automatically corrected to within the safety boundary, and candidate control strategies are output.

[0055] The formula for calculating the safety urgency of a single evacuation route is as follows: , For the first Real-time flue gas concentration in each channel This represents the safe concentration threshold for the corresponding passage. The global evacuation safety urgency is the maximum value of the safety urgency of all passages. This value, along with the comprehensive fire intensity index, serves as the core input state for the upper-level option strategy, guiding the strategy to prioritize matching control options that ensure evacuation safety.

[0056] The formula for calculating the Lyapunov barrier function is: , This represents the current system state. The initial action output by the lower-level strategy. For action Predicted values ​​of flue gas concentration in evacuation routes after execution. The safe concentration threshold for evacuation routes.

[0057] The constraint rule is: if This indicates that the action will exceed the safety threshold, therefore the action... Make corrections, correction amount satisfy: , The gradient of the safety barrier function is the safety barrier function. Regarding action variables The gradient vector represents the rate and direction of change of the barrier function value with the control action. To correct the increment of the action, A positive convergence coefficient ensures that the smoke concentration in the evacuation passage falls back to within the safe threshold after correction, thus directly eliminating dangerous strategies from the decision-making stage.

[0058] In this process, a hierarchical multi-objective reward function is used to jointly train the upper and lower layer strategies. The total reward is composed of the upper layer option reward, the lower layer action reward and the safety margin positive reward, and the output is a dynamic optimal control strategy that satisfies hard safety constraints.

[0059] The formula for calculating the bonus function for advanced fire protection ratings is as follows: , , , These are respectively: upper-level option rewards, lower-level action rewards, and positive safety margin rewards. , , The reward weights are respectively assigned to upper-level options, lower-level actions, and positive safety margins.

[0060] Upper-level option rewards used to assess the rationality of high-level decisions The calculation formula is as follows: , The reward for matching fire situation options is based on the matching score between the current fire situation and the high-level control options, guiding the upper-level strategy to choose the correct response direction. To mitigate the impact of option switching, a fixed negative penalty is applied each time a high-level control option switch is triggered, thus preventing frequent strategy fluctuations.

[0061] Lower-level action rewards The calculation formula is as follows; , As for the core objective reward of the sub-options, when the high-level control option focuses on smoke extraction, the core objective is to bring the smoke extraction efficiency close to the optimal value. To guide efficient smoke extraction, the numerical range is [-1,1]. When the high-rise control option is makeup air leveling, the core is the uniformity of zoned pressure difference, correcting boundary risks. , , , These represent the maximum, minimum, and average values ​​of the total power difference in the system. This is a constant term; when the high-level control option is evacuation and traffic maintenance, , For the first Real-time flue gas concentration in each channel The maximum value of 1 is taken when all channels meet the safety concentration threshold, which directly corresponds to the personnel life safety target; As a differentiated general constraint, when the high-level control option focuses on smoke extraction or make-up air leveling, It is based on the minimum penalty for smoke emission volume and the penalty for energy consumption per unit of smoke emission; when the high-rise control option is evacuation and traffic maintenance. The unit emission energy consumption penalty is abolished, and the emission volume limit constraint is removed to avoid conflict with the traffic control target.

[0062] In conjunction with Lyapunov hard constraints, a positive reward margin of safety is provided to both ensure a bottom line and avoid overly conservative strategies. The calculation formula is as follows: , For the first Real-time flue gas concentration in each channel For the first The safety concentration threshold for each evacuation channel is determined by the difference between the actual concentration in the evacuation channel and the safety threshold. The greater the difference between the actual concentration in the evacuation channel and the safety threshold, the higher the safety margin and the higher the reward. The guidance strategy maintains safety redundancy within a controllable range, rather than simply preventing the threshold from being exceeded.

[0063] When the deep reinforcement learning model corresponding to the advanced fire protection level is executed, the highest priority is given to Lyapunov barrier function verification: after the deep reinforcement learning network initially generates candidate control actions, it uses the Lyapunov barrier function to predict the evolution trajectory of smoke spread conditions corresponding to the actions, directly correcting and eliminating extreme dangerous actions that would cause major safety risks such as large-scale smoke intrusion into evacuation routes, and only retaining feasible actions that meet basic safety constraints into the candidate action set; subsequently, the second priority is given to hierarchical multi-objective reward function evaluation, with the safety margin reward item only providing differentiated incentives for all feasible actions that have passed the barrier verification, guiding the agent to prioritize ventilation control schemes with greater safety redundancy and stronger ability to resist dynamic disturbances of fire smoke. The Lyapunov barrier function serves as a bottom-line safety safety net, used to intercept extreme control actions that exceed the safety red line; the safety margin reward completes strategy optimization within the set of legal feasible actions. The barrier function does not eliminate critical actions that are near the safety boundary, feasible but have weak anti-disturbance capabilities. Such actions are assigned to the safety margin reward to achieve differentiated scoring, avoiding the agent from converging to a critical control strategy that barely meets the lower safety limit. This adapts to the needs of scenarios with intense smoke development and dynamic and ever-changing spread under advanced fire conditions.

[0064] This embodiment adopts different protection measures for different fire levels. For the primary level, grid matching is used for fast response with low computing power consumption and low latency. For the intermediate and advanced levels, differentiated reinforcement learning models are used to adapt to the core requirements of different fire levels. The hierarchical reward function achieves multi-objective balance and the control strategy is more in line with the needs of the scenario.

[0065] S4. Based on the preset ventilation hydraulic network model, the global objective of the optimal control strategy is decoupled into the parameters of each branch equipment, mapped into equipment control commands, and issued for execution.

[0066] The specific execution process is as follows: Based on the preset ventilation system hydraulic network model, the global objectives such as total smoke exhaust volume and target air pressure in the optimal control strategy are decoupled into sub-objectives of air volume allocation and air pressure allocation for each branch duct according to the branch resistance characteristics; based on the air volume and air pressure sub-objectives of each branch, the fan performance curve and valve resistance characteristic curve are called to map the target speed of each fan and the target opening degree of each valve; the target control parameters of each device are converted into control commands conforming to the BACnet or Modbus industrial standard communication protocol and sent to the corresponding ventilation execution components. The ventilation hydraulic network model is automatically constructed in advance by parsing the building BIM model to obtain the duct topology and ventilation equipment parameters; when collecting basic building space data in S1, the original BIM information is read synchronously to generate and initialize the ventilation hydraulic network model.

[0067] Furthermore, since existing reinforcement learning models mostly rely on online training, real-time computing power consumption is high and policy convergence speed is slow in fire emergency scenarios, which cannot meet the real-time requirements of emergency response. In order to solve this technical problem, this embodiment also constructs a policy pre-training method driven by fire dynamics simulation. A multi-scenario fire ventilation coupled sample set is generated through a fire dynamics simulator. The meta-learning framework is used to perform offline pre-training and domain adaptation adjustment on the corresponding deep reinforcement learning models at intermediate and advanced levels, respectively, without sharing initialization parameters. The specific execution steps of the strategy pre-training method are as follows: Based on a fire dynamics simulator, iterates through combinations of scenarios with different fire source locations, different fire intensities, different initial ventilation states, and different building layouts to generate multiple sets of coupled time-series data of the entire fire development process and ventilation control response, forming a simulation sample library; using a model-independent meta-learning framework, different scenarios in the simulation sample library are used as independent training tasks to perform multi-task pre-training on the deep reinforcement learning model to obtain initialization parameters that can quickly adapt to new scenarios; based on the digital twin model of the target building, a small batch of simulation data of real building scenarios is generated to perform domain adaptation fine-tuning on the pre-trained model, reducing the distribution difference between the simulation domain and the real building scenario to below a preset difference threshold, thus obtaining a pre-trained model that can be directly deployed.

[0068] This embodiment also provides a reinforcement learning-based intelligent emergency control system for fire ventilation, used to implement the above-mentioned reinforcement learning-based intelligent emergency control method for fire ventilation, including: The fire condition comprehensive intensity index generation module is used to collect fire sensing data, ventilation system status data and building space basic data, and after preprocessing, obtain a standardized sensing dataset to calculate the fire condition comprehensive intensity index. The fire protection rating module is used to obtain the ventilation system condition adaptability index based on the comprehensive fire intensity index and ventilation system status data, and to calculate the coupling risk value in combination with the comprehensive fire intensity index; based on the coupling risk value, it classifies the fire protection into three levels: primary, intermediate, and advanced. The first equipment instruction generation module is used to obtain ventilation control strategies for basic fire protection levels by using a fast lookup table correction method based on three-dimensional spatial grid matching, which is then mapped into equipment control instructions and issued for execution. For intermediate and advanced fire protection levels, a deep reinforcement learning model with differentiated optimization objectives is invoked. The strategy input state space is constructed using the comprehensive fire intensity index, ventilation system status data, and basic building space data. The strategy is iteratively trained through a hierarchical multi-objective reward function to generate multiple sets of candidate control strategies. The optimal control strategy is selected based on the reward score and then executed by the second equipment instruction generation module. The second equipment instruction generation module is used to decouple the global objective of the optimal control strategy into the parameters of each branch equipment based on the preset ventilation hydraulic network model, map them into equipment control instructions, and issue them for execution.

[0069] This embodiment also provides a fire ventilation emergency intelligent control device based on reinforcement learning, including a processor and a memory, wherein the processor executes the computer program stored in the memory to implement the above-mentioned fire ventilation emergency intelligent control method based on reinforcement learning.

[0070] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described reinforcement learning-based intelligent control method for fire ventilation emergency response.

[0071] This embodiment provides a reinforcement learning-based intelligent emergency control method for fire ventilation. First, it integrates multi-source data and calculates the comprehensive intensity index of fire conditions after standardized preprocessing, overcoming the shortcomings of traditional schemes such as inconsistent data quality and single dimension of condition judgment, and achieving accurate quantification of fire hazard level. Then, it divides fire intensity into three levels by coupling fire intensity with ventilation system condition adaptability, breaking the traditional mode of determining the level solely based on the disaster situation, and ensuring that the classification results match the actual control capability of the equipment, avoiding the disconnect between classification and execution. Next, it adopts a graded differentiated strategy generation mechanism. The primary level uses a fast lookup table method to ensure emergency response speed, while the intermediate and advanced levels use reinforcement learning with differentiated rewards to achieve multi-objective optimization, taking into account the real-time performance of low-risk scenarios and the control accuracy of high-risk scenarios, solving the problems of redundant computing power and misaligned optimization objectives of single algorithms. Finally, through ventilation-hydraulic network decoupling and equipment command mapping, the global control objective is transformed into executable equipment parameters, ensuring that the strategy is implemented in a compliant and controllable manner, and comprehensively improving the scientificity and effectiveness of fire ventilation emergency control.

[0072] While exemplary embodiments of the invention have been described herein, many other variations or modifications conforming to the principles of the invention can be directly determined or derived from the disclosure of this invention without departing from its spirit and scope. Therefore, the scope of the invention should be understood and recognized to cover all such other variations or modifications.

Claims

1. A fire ventilation emergency intelligent control method based on reinforcement learning, characterized in that, This includes the following operations: S1. Collect fire sensing data, ventilation system status data and building space basic data, preprocess them to obtain a standardized sensing dataset, and calculate the comprehensive intensity index of fire conditions. S2. Based on the comprehensive intensity index of fire conditions and the status data of the ventilation system, the adaptability index of the ventilation system conditions is obtained. Combined with the comprehensive intensity index of fire conditions, the coupling risk value is calculated. Based on the coupled risk value, fire protection levels are divided into three categories: primary, intermediate, and advanced. S3. For the basic fire protection level, a rapid lookup table correction method using three-dimensional spatial grid matching is adopted to obtain the ventilation control strategy, which is then mapped into equipment control commands and issued for execution. For intermediate and high fire protection levels, a deep reinforcement learning model with differentiated optimization objectives is invoked. The strategy input state space is constructed using the comprehensive fire intensity index, ventilation system status data, and building space basic data. The strategy is iteratively trained through a hierarchical multi-objective reward function to generate multiple sets of candidate control strategies. The optimal control strategy is selected based on the reward score and then executed in S4. S4. Based on the preset ventilation hydraulic network model, the global objective of the optimal control strategy is decoupled into the parameters of each branch equipment, mapped into equipment control commands, and issued for execution.

2. The fire ventilation emergency intelligent control method based on reinforcement learning according to claim 1, characterized in that, In S3, the fast lookup table correction method for three-dimensional spatial grid matching is as follows: the interior space of the building is divided into three-dimensional grids of equal granularity. Based on the relationship between the grid position and the distribution of ventilation openings, multiple sets of preset benchmark control parameter combinations are matched for each grid. The grid coordinates of the fire source and the direction of smoke diffusion are obtained in real time. The ventilation control parameter combination of the corresponding grid is retrieved and matched. The ventilation control parameter combination is linearly corrected based on the deviation between the current smoke concentration and the flame temperature to obtain the ventilation control strategy.

3. The fire ventilation emergency intelligent control method based on reinforcement learning according to claim 1, characterized in that, In S3, the deep reinforcement learning model processing procedure corresponding to the intermediate fire response level is as follows: Based on the building space topology and ventilation hydraulic network topology, a causal mapping library of fire source space nodes and ventilation branches is constructed to determine the range of adjustable equipment and the causal correlation degree of the branches corresponding to each fire source location; a causal mask is applied to the model action space in real time according to the space node where the fire source is located to shield the equipment action dimension of non-associated branches; the target operating range of smoke concentration in the fire area and total smoke exhaust volume of the system is preset; only when the core smoke index deviates from the target range, the comprehensive intensity index of the fire condition fluctuates beyond the fluctuation threshold, or the risk of smoke backflow is detected, the strategy decision is triggered and the equipment parameter adjustment amount is output as a candidate control strategy; if not triggered, the current equipment operating state is maintained. The hierarchical multi-objective reward function is derived from the basic reward for meeting the standard, the reward for smoke exhaust efficiency, the penalty for causal differential backflow, and the penalty for unit smoke exhaust energy consumption.

4. The fire ventilation emergency intelligent control method based on reinforcement learning according to claim 1, characterized in that, In S3, the deep reinforcement learning model processing procedure corresponding to the advanced fire response level is as follows: A hierarchical reinforcement learning framework is constructed, which includes an upper-level option strategy layer, a lower-level action execution layer, and a global value commentator. It pre-sets three types of high-level control options: smoke extraction, air supply and balancing, and evacuation and ventilation. The system calculates the safety urgency of each evacuation route and the overall evacuation safety urgency in real time, which serves as the core input state for the upper-level option strategy, and dynamically matches the optimal control option corresponding to the current fire situation. The lower-level action execution layer outputs the initial parameters for device control based on the selected control options. Before the parameters are output, they are verified by the Lyapunov safety barrier function. The action parameters that exceed the evacuation safety threshold are corrected to the safety boundary, and the candidate control strategy is output. The hierarchical multi-objective reward function is obtained by weighting the upper-level option matching reward, the lower-level core objective execution reward, and the evacuation safety margin reward.

5. The fire ventilation emergency intelligent control method based on reinforcement learning according to claim 1, characterized in that, A strategy pre-training method driven by fire dynamics simulation is constructed. A multi-scenario fire ventilation coupling sample set is generated through a fire dynamics simulator. A meta-learning framework is used to pre-train the deep reinforcement learning models corresponding to intermediate and advanced fire protection levels offline.

6. The fire ventilation emergency intelligent control method based on reinforcement learning according to claim 5, characterized in that, The specific steps of the policy pre-training method are as follows: Based on a fire dynamics simulator, we traverse various scenarios with different fire source locations, fire intensities, initial ventilation states, and building layouts to generate multiple sets of coupled time-series data on the entire fire development process and ventilation control response, forming a simulation sample library. Different scenarios in the simulation sample library are used as independent training tasks to perform multi-task pre-training on the deep reinforcement learning model and obtain the initialization parameters. Based on the digital twin model of the target building, simulation data of the real building scene is generated. The pre-trained model is then fine-tuned to adapt to the domain, reducing the distribution difference between the simulation domain and the real building scene to below a preset difference threshold, thus obtaining the pre-trained model.

7. The fire ventilation emergency intelligent control method based on reinforcement learning according to claim 1, characterized in that, The preprocessing of S1 includes unifying the spatiotemporal reference of fire perception data, ventilation system status data and building space basic data, removing triple constraint outliers, completing and resampling spatiotemporally coupled time series, normalizing node differences and encapsulating structured output.

8. A fire ventilation emergency intelligent control system based on reinforcement learning, used to implement the fire ventilation emergency intelligent control method based on reinforcement learning as described in claim 1, characterized in that, include: The fire condition comprehensive intensity index generation module is used to collect fire sensing data, ventilation system status data and building space basic data, and after preprocessing, obtain a standardized sensing dataset to calculate the fire condition comprehensive intensity index. The fire protection rating module is used to obtain the ventilation system condition adaptability index based on the comprehensive fire intensity index and ventilation system status data, and to calculate the coupling risk value by combining the comprehensive fire intensity index. Based on the coupled risk value, fire protection levels are divided into three categories: primary, intermediate, and advanced. The first equipment instruction generation module is used to obtain ventilation control strategies for basic fire protection levels by using a fast lookup table correction method based on three-dimensional spatial grid matching, which is then mapped into equipment control instructions and issued for execution. For intermediate and advanced fire protection levels, a deep reinforcement learning model with differentiated optimization objectives is invoked. The strategy input state space is constructed using the comprehensive fire intensity index, ventilation system status data, and basic building space data. The strategy is iteratively trained through a hierarchical multi-objective reward function to generate multiple sets of candidate control strategies. The optimal control strategy is selected based on the reward score and then executed by the second equipment instruction generation module. The second equipment instruction generation module is used to decouple the global objective of the optimal control strategy into the parameters of each branch equipment based on the preset ventilation hydraulic network model, map them into equipment control instructions, and issue them for execution.

9. A fire ventilation emergency intelligent control device based on reinforcement learning, characterized in that, It includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the reinforcement learning-based intelligent control method for fire ventilation emergency as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the reinforcement learning-based intelligent control method for fire ventilation emergency response as described in any one of claims 1-7.