Intelligent smoke-proof air curtain parameter optimization system and method based on proxy model and reinforcement learning

Abstract

The invention discloses an intelligent smoke-proof air curtain parameter optimization system and method based on a proxy model and reinforcement learning. The system comprises a high-fidelity simulation and order reduction module, a proxy model construction module, a reinforcement learning control strategy training module, a data acquisition module, a control execution module and an air curtain. The method comprises the following steps: establishing a multi-physics-field coupled fire smoke diffusion numerical model based on finite element simulation software; a low-order subspace model capable of being used for rapidly reconstructing system behaviors is constructed based on a reduced-order algorithm, and a performance expression data set under different control parameters is predicted; based on the low-order subspace model data, a deep learning method is driven to construct a reduced-order proxy model; based on reinforcement learning, training an optimal control strategy of the air curtain in the constructed reduced-order proxy model environment; the air speed and the spraying angle of the air curtain are optimized in real time through the control execution module, and effective control and dynamic adjustment of fire smoke are achieved. According to the invention, the smoke control efficiency, the energy efficiency level and the cross-scene adaptive capability of the system are improved.

Description

Technical Field

[0001] This invention relates to an intelligent smoke-proof air curtain parameter optimization system and method based on surrogate model and reinforcement learning, belonging to the field of intelligent air curtain control technology. Background Technology

[0002] With the widespread development of urban underground spaces, air curtains, as a fire and smoke control technology, have been widely used in narrow, enclosed spaces such as subways, tunnels, and urban utility tunnels. In traditional smoke control systems, once a fire occurs, fire doors are typically used to isolate the area, and the spread of smoke is controlled by delaying the activation of mechanical smoke extraction systems. Air curtain systems, on the other hand, are triggered by a linkage control system in the early stages of a fire. High-pressure fans spray air at a specific angle, forming a high-speed airflow barrier that effectively blocks the spread of smoke.

[0003] To enhance the smoke control capabilities of air curtains in fire scenarios, existing technologies focus on the airflow dynamics parameters of the air curtain, analyzing the impact of parameters such as jet velocity, nozzle thickness, and deflection angle on smoke control effectiveness. Current research mainly concentrates on: analyzing the influence of airflow parameters on smoke control performance through experiments and numerical simulations; exploring the application performance of specific structures (such as blow-suction types and modified structures) in different fire scenarios; and optimizing air curtain structural parameters to improve energy efficiency and smoke control efficiency. However, current research generally employs parameter scanning or empirical optimization methods, which suffer from long experimental cycles, high computational costs, and insufficient handling of coupling relationships between parameters. Especially when facing complex and variable building fire environments and operating conditions, air curtain systems lack an intelligent optimization mechanism with online learning and adaptive capabilities. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent smoke control air curtain parameter optimization system and method based on surrogate model and reinforcement learning. This method can combine surrogate model and reinforcement learning to construct an intelligent optimization system for air curtain parameters, realize intelligent optimization of air curtain parameters, improve the smoke control efficiency and energy efficiency of the system, and enhance its cross-scenario adaptive capability.

[0005] To achieve the above objectives, this invention provides an intelligent smoke-proof air curtain parameter optimization system based on surrogate model and reinforcement learning, including a high-fidelity simulation and order reduction module, a surrogate model construction module, a reinforcement learning control strategy training module, a data acquisition module, a control execution module, and an air curtain; The high-fidelity simulation and order reduction module is connected to the surrogate model construction module, providing simulation data input to the surrogate model construction module; the reinforcement learning control strategy training module uses the surrogate model construction module as its training environment; the data acquisition module is located on both sides of the air curtain for sampling flue gas features; and the control execution module is connected to the data acquisition module, the reinforcement learning control strategy training module, and the air curtain.

[0006] Furthermore, the high-fidelity simulation and order reduction module establishes a multi-physics coupled numerical model of fire smoke diffusion based on numerical simulation software (such as COMSOL Multiphysics) to simulate the influence of the air curtain on smoke diffusion behavior under different wind speeds, air supply angles, and fire source conditions. Through systematic parameter scanning experiments, high-fidelity simulation data of smoke flow, concentration distribution, and temperature evolution in key areas behind the air curtain are obtained. After forming a sample set, the high-dimensional flow field data is appropriately orthogonally decomposed, and combined with empirical interpolation or discrete singular value decomposition (SVD), the dominant modes and evolution characteristics of the system are extracted. A low-order subspace model that can quickly reconstruct the system behavior is then constructed as the order reduction model.

[0007] The surrogate model construction module employs a supervised learning method to construct a mapping model between the time series of states in front of the air curtain and the control parameters to the smoke response behind the air curtain, based on simulation samples. To fully exploit the temporal correlation in the smoke evolution process, the surrogate model adopts a temporal regression structure based on a long short-term memory network. Its input is the sensor state sequence over a past period, and its output is the key physical quantity of the response behind the air curtain, forming a hybrid surrogate model that combines physical priors and data-driven characteristics. This model has the advantages of capturing state change trends, handling lag relationships, and enhancing fitting ability, enabling it to more accurately predict smoke control performance under different control parameters. Through the construction of this module, the system can achieve rapid testing and optimization of multi-parameter control strategies for the air curtain without repeating high-fidelity simulations, meeting the needs of high-frequency interaction and cross-scenario generalization in reinforcement learning, and providing data support for intelligent optimization control of the air curtain. The reinforcement learning control strategy training module is used to train the optimal control strategy for air curtain wind speed and jet angle in a simulation environment constructed by the surrogate model, thereby achieving intelligent adjustment of smoke diffusion. To improve the responsiveness to dynamic changes in the fire environment, a time series modeling approach is introduced. The physical parameters of temperature and smoke concentration collected by the sensors in front of the air curtain in the past 5 seconds are used as the state input sequence to capture the time dependence and trend characteristics of the smoke evolution process. The time series state input is processed by a long short-term memory neural network to extract temporal features and input to the strategy network for decision output. The action space is a continuous variable, including the air curtain jet speed and angle. The system adopts a reinforcement learning algorithm such as SoftActor-Critic (SAC) to perform interactive learning in the simulation environment constructed by the surrogate model. A reward function is constructed based on the concentration, temperature and energy consumption surrogate output results behind the air curtain to guide strategy optimization. The data acquisition module is located on both sides of the air curtain, forming a spatial sensor array with multiple sensors. This array is used to collect temperature and smoke concentration parameters in the area in front of and behind the air curtain in real time during a fire, thus constructing system status information. The sensors are symmetrically or correspondingly distributed, covering typical smoke plume channels and forming a two-dimensional array to enhance the sampling density and coverage of spatial smoke characteristics. All sensors are connected to the control module via a standard interface, and the data is transmitted to the status construction module for integration according to a set sampling period. The control execution module is used to deploy the control strategy trained in the reinforcement learning control strategy training module to the actual air curtain control system, realizing real-time intelligent adjustment of air curtain wind speed and spray angle. The control execution module takes environmental parameters collected by the sensor array in front of the air curtain within the last 5 seconds as input, forming continuous time-series state data, and combines this with past control action sequences to form a dynamic control input window. This time-series input is updated at a fixed 1-second interval during system operation and is transmitted in real-time to the deployed control strategy network for forward inference. The control strategy network of the reinforcement learning control strategy training module is the policy function obtained during the reinforcement learning phase, including an encoding layer that extracts temporal features based on a long short-term memory structure, and an action prediction layer that outputs wind speed and angle commands based on a fully connected structure. The control execution module deploys this policy function to embedded control hardware or an edge computing platform, maintaining a data interface connection with the sensor acquisition module, continuously receiving the sensor state sequence within a 5-second sliding time window, and obtaining the current optimal air curtain control parameters through forward calculation.

[0008] A parameter optimization method for intelligent smoke-proof air curtain based on surrogate model and reinforcement learning includes the following steps: S1. Establish a numerical model of fire smoke diffusion based on multi-physics field coupling using finite element simulation software; S2. Based on the order reduction algorithm, a low-order subspace model can be constructed to quickly reconstruct the system behavior and predict the performance dataset under different control parameters; S3. Based on low-order subspace model data, a temporal modeling method is used to extract the relationship between environmental state and control action, and a deep learning method is used to construct a reduced-order agent model. S4. Based on reinforcement learning, train the optimal control strategy for the air curtain in the constructed reduced-order agent model environment; S5. By controlling the execution module to optimize the air curtain's wind speed and spray angle in real time, effective control and dynamic adjustment of fire smoke can be achieved.

[0009] Furthermore, the construction process of the numerical model for fire smoke diffusion in S1 is as follows: COMSOL Multiphysics software was used as a high-fidelity simulation platform. Its built-in reduced-order model and surrogate modeling were used to process and reconstruct the flow field data. A three-dimensional computational domain containing the fire source, air curtain, and affected area was established in COMSOL Multiphysics. The preferred computational domain was 6 m long × 3 m wide × 3 m high. The air curtain nozzle was 0.03 m wide and located 1.0 m on the right side. The fire source area was 0.4 m × 0.2 m and placed 0.5 m away from the left wall. The computational domain included the fire source area, the air curtain area, and the area affected by the air curtain. To capture the smoke flow and temperature distribution, meshing was used. The mesh size was preferably set to 2 mm in the fire source area and the air curtain area to improve the calculation accuracy. In other areas, the mesh size was appropriately increased to reduce the amount of computation.

[0010] Furthermore, the specific process of S2 is as follows: S2.1 The mass equation is constructed as follows: ; Where ρ represents fluid density, t represents time, u represents fluid velocity vector, and ∇⋅(ρu) represents mass flux divergence per unit volume; The momentum equation is as follows: ; Where ρ represents fluid density, u represents fluid velocity vector, p represents pressure, μ represents fluid dynamic viscosity, ρ0 represents reference density, g represents gravitational acceleration vector, β represents thermal expansion coefficient, T represents local temperature, and T0 represents reference temperature; S2.2 The energy equation is constructed as follows: ; Where ρ represents the fluid density, c pLet represent the specific heat capacity at constant pressure, T represent the local temperature, u represent the fluid velocity vector, ∇T represent the gradient operator of the temperature field, and k represent the thermal conductivity. Indicates a heat source per unit volume; S2.3, The flue gas component equations are constructed as follows: ; Where C represents the flue gas mass fraction, t represents time, u represents the fluid velocity vector, ∇C represents the spatial gradient of the flue gas mass fraction, and D represents the diffusion coefficient of the flue gas in the air; S2.4 The state equations are constructed as follows: COMSOL uses incompressible + Boussinesq, where density changes only with temperature, and pressure contributes negligibly to density (Mach < 0.3, satisfying the fire smoke scenario): ; Where ρ represents the local fluid density, β represents the reference density, T represents the coefficient of thermal expansion, T represents the local temperature, and T0 represents the reference temperature. An air curtain was configured with an airflow source having a set wind speed and jet angle. The air curtain parameters were adjustable, with the jet angle ranging from 10° to 20° in 5° intervals, the outlet wind speed ranging from 4 m / s to 8 m / s in 0.5 m / s intervals, and the outlet width of 0.03 m. This configuration was used to simulate the smoke control effect under different combinations. To ensure the comprehensiveness and representativeness of the sample data, the experiment was conducted under different environmental conditions, including different initial smoke concentrations, different ambient temperatures, and different ventilation conditions. High-fidelity data on smoke flow, concentration distribution, and temperature evolution in key areas of the air curtain system were obtained by simulating the effects of the air curtain on smoke diffusion behavior under different wind speeds, air supply angles, and fire source conditions. These data formed a sample set covering various operating conditions.

[0011] After acquiring high-dimensional flow field snapshot data covering the parameter space, singular value decomposition is performed on the snapshot matrix in COMSOL's Proper Orthogonal Decomposition module to extract the dominant modes and evolutionary features of the system: Let s be the snapshots of the transient flow field collected by the system under different parameter conditions. i , where i = 1, 2, …, M, represents the i-th time step or simulation sample. The time mean field of all samples is calculated as follows: ; Construct the covariance matrix: ; in, This represents the average state vector of the system.

[0012] Perform singular value decomposition on the covariance matrix C, i.e.: ; Where V is the eigenvector matrix and Λ is the eigenvalue diagonal matrix, reflecting the energy occupied by each principal mode.

[0013] According to the energy principle Determine the truncation order r, and obtain the reduced-order basis by truncating the first r feature modules. ; Among them, 𝑉 𝑟 To form a matrix from the first r eigenvectors, Λ 𝑟 For the corresponding eigenvalue diagonal matrix, It is a reduced-order orthogonal basis matrix.

[0014] Depend on The online projection coefficients are obtained, and the reconstructed field of the original flow field is: ; in, This is the projection coefficient vector of the system in the reduced-order space, reflecting the excitation intensity of each principal mode; U is the average state vector of the system. r 'a' represents the reconstructed dynamic component.

[0015] For processing nonlinear operators, empirical interpolation methods are used to calculate the nonlinear terms at some sampling points, and the interpolation matrix is ​​then used to... accomplish A highly efficient approximation; Where N(s) is a nonlinear operator, Used to reconstruct global nonlinear behavior from a finite number of sampling points; This step significantly reduces computational complexity while ensuring model accuracy, allowing the system state equations to be solved only in a low-dimensional subspace, thus deriving a low-order subspace model R(·).

[0016] Furthermore, the specific process of S3 is as follows: S3.1 Constructing the Training Sample Set: The training samples come from the dataset generated by the low-order subspace model. The input part is a multi-dimensional time series sampled at a frequency of 1 Hz within the last 5 seconds, including the sensor status (temperature, smoke concentration) placed in front of the air curtain. The output part is the key physical response quantities behind the air curtain, including temperature and smoke concentration. After normalization, the input data is divided into a training set (70%), a validation set (15%), and a test set (15%) to ensure the generalization performance of the model. Verification shows that the surrogate model F can quickly predict the response behind the air curtain without high-fidelity simulation, supporting high-frequency interaction and cross-scene generalization. The data acquisition module is equipped with multiple sensors, forming a front and rear sensor array, to collect real-time temperature and flue gas concentration parameters in the areas before and after the air curtain. The front and rear sensors are arranged symmetrically or correspondingly, covering the typical plume channel range. The front array S... f Located on the windward side of the air curtain, it is used to capture the temperature T and concentration C of the smoke plume; the rear array S b Located on the wind side behind the air curtain, it is used to detect the smoke conditions behind the air curtain barrier.

[0017] The two arrays are symmetrically distributed in the plane, forming a two-dimensional sampling grid that covers the typical plume channel range. The data acquisition module synchronously collects sensor data at a set sampling period and aligns the timestamps to form a sliding time window sequence {(T,C)}{t-5s:t} of the most recent 5 seconds, providing input for subsequent agent modeling and policy inference; this spatial sensor array improves the sampling density and coverage of flue gas characteristics. These sensors are connected to the control module via a standard interface and transmit data to the state construction module for integration at a set period. S3.2, Build a descending-order proxy model using the Deep Learning Toolbox: 1) Set the sequenceInputLayer as the input layer to receive multidimensional time series data: The input is a multidimensional sequence of length T=50, with a sampling interval of 1s and a total duration of 5s. The expression is: ; Where, x t It is composed of historical sequences of temperature and flue gas concentration from 5 measuring points in front of the air curtain; 2) A two-layer LSTM is used, with each layer having a hidden state dimension h=128, to capture long-range dependencies and short-term fluctuations in the time series. The single-step update formula for the ℓth layer is:

[0018] Among them, f t i t o t These are the activation results of the forget gate, input gate, and output gate, respectively; c t For the state of the memory unit; h t Let be the hidden state at the current time step; σ(⋅) and tanh(⋅) are the Sigmoid and hyperbolic tangent activation functions, respectively.

[0019] After the hidden state at the end of the last LSTM layer, add a fully connected layer and map it to the physical response: ; in, This is the final output of the LSTM hidden state. The hidden state is mapped to the target response behind the air curtain, and the two-dimensional output corresponds to the average temperature and smoke concentration 1 m behind the air curtain, respectively. 3) The mean squared error (MSE) is used as the loss function during training: ; Where, N batch The number of samples in a single training batch; by analyzing the predicted values Compared with the true value The average of the squared errors can effectively reflect the accuracy of the model's predictions.

[0020] The optimization method used was Adam, with an initial adaptive learning rate of 0.001, each mini-batch containing 64 samples, a maximum training epoch of 200, and an early stopping strategy on the validation set to avoid overfitting.

[0021] After network training, this LSTM model can rapidly predict the response behind the air curtain from the input state-action sequence within milliseconds. Compared to directly calling COMSOL full-order simulation, the inference speed of this model is improved by hundreds of times, which is sufficient to support high-frequency interaction and cross-scenario training of reinforcement learning control strategies. The resulting surrogate model has both physical prior and data-driven characteristics. On the one hand, it inherits the main modal distribution extracted by POD order reduction; on the other hand, it uses LSTM to capture complex nonlinear time evolution laws, thereby more accurately predicting the smoke control performance of the air curtain under different control parameters, providing reliable support for intelligent optimization control.

[0022] Furthermore, the specific process of S4 is as follows: S4.1. Each episode generates a completely new historical sequence of temperature and concentration ahead using random or adaptive perturbation as the initial state. The action policy network obtains the air curtain parameters and uses a proxy model to provide millisecond-level state transitions. ; Wherein, the state vector s t Forward temperature-concentration history sequence of 5 s × 1 Hz; action space These are continuous motion parameters, including wind speed of 4–8 m / s and angle of 10–20 degrees. S4.2 To capture time dependence and lag effects, both the policy network and the value network share a two-layer 128-dimensional LSTM mathematical skeleton. The tail output heads are connected to the regression layer MSE and the policy-value layer SAC loss, respectively, to achieve dual-task reuse in the same feature space. This extracts the main features from the original time series and completes action decision-making and value assessment based on these features. The policy network π ϕ It is a continuous action policy that outputs the mean μ of the actions. t and logarithmic standard deviation logσ t :

[0023] in, This is the final output of the LSTM hidden state, 𝜇 𝑡 Let be the mean vector of the action distribution, 𝜎 𝑡 Let be the standard deviation vector of the action distribution, 𝑊 𝜇 With 𝑊 𝜎 These are the corresponding weight matrices, 𝑏 𝜇 With 𝑏 𝜎 This is a bias term.

[0024] Action sampling: ; S4.3 In the design of the reward function, the system comprehensively incorporates the air curtain's smoke control performance, energy consumption, and control stability into the evaluation to ensure that the direction of strategy optimization meets the needs of actual application. The reward function consists of three parts: safety reward, energy consumption penalty, and action stability penalty, and its overall form is expressed as follows: ; in, w 1 , w 2 , w 3 The weight parameters are respectively for safety reward, energy consumption penalty, and motion smoothness penalty, satisfying... w 1 > w 2 , and w 1 > w 3 ; 1) Security Bonus Items: ; Where C(t) represents the smoke concentration behind the air curtain, T(t) represents the peak temperature; α1 and α2 are normalization coefficients; when C(t) or T(t) exceeds the safety threshold, the system applies additional negative rewards to enforce safety. 2) Energy consumption penalty item: ; Where P(t) is the power consumption of the air curtain at time t, and β is the energy consumption weighting coefficient; the energy consumption penalty term guides the strategy to minimize energy consumption while satisfying the smoke control effect; 3) Penalty for smooth movement: ; Among them, a t This represents the control action vector at the current moment, i.e., wind speed and injection angle; γ is the smoothness weighting coefficient; the action stability penalty term is used to suppress drastic fluctuations in control parameters, ensuring the stability of air curtain operation and the reliability of the equipment. Through the above reward function design, the system can automatically balance the three objectives of smoke control effect, energy efficiency and control stability during the training process, and finally obtain an intelligent air curtain control strategy that prioritizes safety, optimizes energy consumption and balances stability.

[0025] S4.4, Policy Gradient and Update adopt the Soft Actor-Critic framework, with the goal of maximizing cumulative reward: At the start of each episode, an initial state s0 is randomly generated, including historical sequences of temperature and concentration ahead. A strategy model for generating control actions, i.e., wind speed and injection angle, is then generated based on the temperature and concentration state sequence in front of the air curtain over the most recent 5 seconds. After extracting temporal features through LSTM encoding, the control command action parameters a are output. t ; Use the proxy model to quickly provide the next state s t+1 =F(s t ,a t The agent model performs interactive learning; in parallel, a value evaluation model is used to evaluate the long-term reward under the current state and action, so that the policy model receives correct feedback and calculates the reward value R. t ; the experience of each interaction (s t， a t， r t， s t+1 The data is stored in the replay cache D. A mini-batch is randomly drawn from the replay cache D to update the policy and value network, and then updated through the Adam optimizer. ; Among them, 𝑉 𝜓 For the value function network output, 𝜋 𝜙 is the output of the policy network, is the discount factor, and is the entropy adjustment coefficient.

[0026] To improve training efficiency and reduce the impact of temporal correlation between samples, empirical replay and mini-batch random sampling are introduced during training; simultaneously, a target value network is set, and a soft update method is adopted. The soft update coefficient τ is set to 0.01 to ensure smooth convergence of the value evaluation model during training. The rapid inference capability of the surrogate model enables the entire system to accumulate a large number of interaction samples in a short time, thereby training a robust air curtain control strategy that performs well in various scenarios and obtaining the optimal policy function. ; The system outputs control actions including the air curtain's wind speed setpoint and spray angle setpoint. The action range is normalized and subject to boundary constraints to ensure consistency with the actual controllable range of the physical device.

[0027] Among them, v min =4m / s, v max =8m / s, θ min =10°, θ max =20°.

[0028] Furthermore, the specific process of S5 is as follows: S5.1 The control execution module collects environmental parameters from the sensor array in front of the air curtain in real time, forming continuous time series state data and creating a dynamic control input window; the data acquisition module updates the 5-second sliding window state sequence at a 1-second cycle; the control execution module inputs this time series state into the strategy network for forward inference. The control strategy network includes an encoding layer that extracts time series features based on an LSTM structure, and an action prediction layer that outputs wind speed and angle based on a fully connected structure, to obtain the current optimal (v*, θ*); S5.2 Apply the optimal control strategy to the air curtain physical device to adjust the air curtain's wind speed and angle. The command is transmitted to the air curtain via the actuator to achieve coordinated adjustment of the jet wind speed and angle; front and rear arrays (S f S b Real-time data transmission forms a closed loop; Specifically, the control execution module sets the upper limits for the output wind speed and spray angle to be 8 m / s and 20°, respectively; if the output wind speed or angle exceeds its upper limit, the control execution module will automatically adjust it to the upper limit value. The control execution module sets the following limits on the slope of change of wind speed and jet angle: the slope of change of wind speed is limited to a maximum increase of 1 m / s per step; the slope of change of jet angle is limited to a maximum increase of 5° per step; if the slope of change of wind speed and jet angle output by the strategy exceeds these limits, the control execution module adjusts them to the maximum allowable range of change. The control execution module is equipped with a bypass mechanism to handle situations such as sensor malfunctions or insufficient policy confidence, namely: Sensor anomaly detection: By monitoring the deviation between the sensor's measured value and the historical average value, if the deviation exceeds a set threshold, the sensor is determined to be abnormal. Policy confidence assessment: The probability density of the policy output action is calculated. If the density is lower than a set threshold, the policy confidence is deemed insufficient.

[0029] Bypass strategy: Once a sensor malfunction is detected or the strategy confidence is insufficient, the control execution module automatically switches to a conservative strategy or fixed parameter mode, setting the wind speed to 6 m / s and the injection angle to 15° to ensure that the system can still operate stably under abnormal conditions.

[0030] This invention sets up a high-fidelity simulation and order reduction module, a surrogate model construction module, a reinforcement learning control strategy training module, a data acquisition module, a control execution module, and an air curtain. It uses the surrogate model and order reduction model to quickly evaluate the smoke-blocking performance of the air curtain after parameter adjustment, and uses reinforcement learning algorithms to continuously optimize the control strategy in interaction. It realizes intelligent optimization of parameters such as air curtain wind speed and angle, improves the smoke control efficiency and energy efficiency of the system, and enhances its cross-scenario adaptive capability. It is suitable for fire prevention and control needs in different spatial forms such as subways, tunnels, and integrated pipe corridors. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a flowchart of the workflow of the method of the present invention. Detailed Implementation

[0032] The invention will now be further described with reference to the accompanying drawings.

[0033] like Figure 1 As shown, an intelligent smoke-proof air curtain parameter optimization system based on surrogate model and reinforcement learning includes a high-fidelity simulation and order reduction module, a surrogate model construction module, a reinforcement learning control strategy training module, a data acquisition module, a control execution module, and an air curtain. The high-fidelity simulation and order reduction module is connected to the surrogate model construction module, providing simulation data input to the surrogate model construction module; the reinforcement learning control strategy training module uses the surrogate model construction module as its training environment; the data acquisition module is located on both sides of the air curtain for sampling flue gas features; and the control execution module is connected to the data acquisition module, the reinforcement learning control strategy training module, and the air curtain.

[0034] As a preferred implementation, the high-fidelity simulation and order reduction module establishes a multi-physics coupled numerical model of fire smoke diffusion based on numerical simulation software (such as COMSOL Multiphysics) to simulate the influence of the air curtain on smoke diffusion behavior under different wind speeds, air supply angles, and fire source conditions. Through systematic parameter scanning experiments, high-fidelity simulation data of smoke flow, concentration distribution, and temperature evolution in key areas behind the air curtain are obtained. After forming a sample set, the high-dimensional flow field data is appropriately orthogonally decomposed, and combined with empirical interpolation or discrete singular value decomposition (SVD), the dominant modes and evolutionary characteristics of the system are extracted. A low-order subspace model that can quickly reconstruct the system behavior is then constructed as the order reduction model.

[0035] The surrogate model construction module employs a supervised learning method to construct a mapping model between the time series of states in front of the air curtain and the control parameters to the smoke response behind the air curtain, based on simulation samples. To fully exploit the temporal correlation in the smoke evolution process, the surrogate model adopts a temporal regression structure based on a long short-term memory network. Its input is the sensor state sequence over a past period, and its output is the key physical quantity of the response behind the air curtain, forming a hybrid surrogate model that combines physical priors and data-driven characteristics. This model has the advantages of capturing state change trends, handling lag relationships, and enhancing fitting ability, enabling it to more accurately predict smoke control performance under different control parameters. Through the construction of this module, the system can achieve rapid testing and optimization of multi-parameter control strategies for the air curtain without repeating high-fidelity simulations, meeting the needs of high-frequency interaction and cross-scenario generalization in reinforcement learning, and providing data support for intelligent optimization control of the air curtain. The reinforcement learning control strategy training module is used to train the optimal control strategy for air curtain wind speed and jet angle in a simulation environment constructed by the surrogate model, thereby achieving intelligent adjustment of smoke diffusion. To improve the responsiveness to dynamic changes in the fire environment, a time series modeling approach is introduced. The physical parameters of temperature and smoke concentration collected by the sensors in front of the air curtain in the past 5 seconds are used as the state input sequence to capture the time dependence and trend characteristics of the smoke evolution process. The time series state input is processed by a long short-term memory neural network to extract temporal features and input to the strategy network for decision output. The action space is a continuous variable, including the air curtain jet speed and angle. The system adopts a reinforcement learning algorithm such as SoftActor-Critic (SAC) to perform interactive learning in the simulation environment constructed by the surrogate model. A reward function is constructed based on the concentration, temperature and energy consumption surrogate output results behind the air curtain to guide strategy optimization. The data acquisition module is located on both sides of the air curtain, forming a spatial sensor array with multiple sensors. This array is used to collect temperature and smoke concentration parameters in the area in front of and behind the air curtain in real time during a fire, thus constructing system status information. The sensors are symmetrically or correspondingly distributed, covering typical smoke plume channels and forming a two-dimensional array to enhance the sampling density and coverage of spatial smoke characteristics. All sensors are connected to the control module via a standard interface, and the data is transmitted to the status construction module for integration according to a set sampling period. The control execution module is used to deploy the control strategy trained in the reinforcement learning control strategy training module to the actual air curtain control system, realizing real-time intelligent adjustment of air curtain wind speed and spray angle. The control execution module takes environmental parameters collected by the sensor array in front of the air curtain within the last 5 seconds as input, forming continuous time-series state data, and combines this with past control action sequences to form a dynamic control input window. This time-series input is updated at a fixed 1-second interval during system operation and is transmitted in real-time to the deployed control strategy network for forward inference. The control strategy network of the reinforcement learning control strategy training module is the policy function obtained during the reinforcement learning phase, including an encoding layer that extracts temporal features based on a long short-term memory structure, and an action prediction layer that outputs wind speed and angle commands based on a fully connected structure. The control execution module deploys this policy function to embedded control hardware or an edge computing platform, maintaining a data interface connection with the sensor acquisition module, continuously receiving the sensor state sequence within a 5-second sliding time window, and obtaining the current optimal air curtain control parameters through forward calculation.

[0036] like Figure 2 As shown, a method for optimizing the parameters of an intelligent smoke-proof air curtain based on an agent model and reinforcement learning includes the following steps: S1. Establish a numerical model of fire smoke diffusion based on multi-physics field coupling using finite element simulation software; S2. Based on the order reduction algorithm, a low-order subspace model can be constructed to quickly reconstruct the system behavior and predict the performance dataset under different control parameters; S3. Based on low-order subspace model data, a temporal modeling method is used to extract the relationship between environmental state and control action, and a deep learning method is used to construct a reduced-order agent model. S4. Based on reinforcement learning, train the optimal control strategy for the air curtain in the constructed reduced-order agent model environment; S5. By controlling the execution module to optimize the air curtain's wind speed and spray angle in real time, effective control and dynamic adjustment of fire smoke can be achieved.

[0037] As a preferred embodiment, the construction process of the numerical model for fire smoke diffusion in S1 is as follows: COMSOL Multiphysics software was used as a high-fidelity simulation platform. Its built-in reduced-order model and surrogate modeling were used to process and reconstruct the flow field data. A three-dimensional computational domain containing the fire source, air curtain, and affected area was established in COMSOL Multiphysics. The preferred computational domain was 6 m long × 3 m wide × 3 m high. The air curtain nozzle was 0.03 m wide and located 1.0 m on the right side. The fire source area was 0.4 m × 0.2 m and placed 0.5 m away from the left wall. The computational domain included the fire source area, the air curtain area, and the area affected by the air curtain. To capture the smoke flow and temperature distribution, meshing was used. The mesh size was preferably set to 2 mm in the fire source area and the air curtain area to improve the calculation accuracy. In other areas, the mesh size was appropriately increased to reduce the amount of computation.

[0038] As a preferred embodiment, the specific process of S2 is as follows: S2.1 The mass equation is constructed as follows: ; Where ρ represents fluid density, t represents time, u represents fluid velocity vector, and ∇⋅(ρu) represents mass flux divergence per unit volume; The momentum equation is as follows: ; Where ρ represents fluid density, u represents fluid velocity vector, p represents pressure, μ represents fluid dynamic viscosity, ρ0 represents reference density, g represents gravitational acceleration vector, β represents thermal expansion coefficient, T represents local temperature, and T0 represents reference temperature; S2.2 The energy equation is constructed as follows: ; Where ρ represents the fluid density, c p Let represent the specific heat capacity at constant pressure, T represent the local temperature, u represent the fluid velocity vector, ∇T represent the gradient operator of the temperature field, and k represent the thermal conductivity. Indicates a heat source per unit volume; S2.3, The flue gas component equations are constructed as follows: ; Where C represents the mass fraction of flue gas, t represents time, u represents the fluid velocity vector, ∇C represents the spatial gradient of the mass fraction of flue gas, and D represents the diffusion coefficient of flue gas in air; S2.4 The state equations are constructed as follows: COMSOL uses incompressible + Boussinesq, where density changes only with temperature, and pressure contributes negligibly to density (Mach < 0.3, satisfying the fire smoke scenario): ; Where ρ represents the local fluid density, β represents the reference density, T represents the coefficient of thermal expansion, T represents the local temperature, and T0 represents the reference temperature. An air curtain was configured with an airflow source having a set wind speed and jet angle. The air curtain parameters were adjustable, with the jet angle ranging from 10° to 20° in 5° intervals, the outlet wind speed ranging from 4 m / s to 8 m / s in 0.5 m / s intervals, and the outlet width of 0.03 m. This configuration was used to simulate the smoke control effect under different combinations. To ensure the comprehensiveness and representativeness of the sample data, the experiment was conducted under different environmental conditions, including different initial smoke concentrations, different ambient temperatures, and different ventilation conditions. High-fidelity data on smoke flow, concentration distribution, and temperature evolution in key areas of the air curtain system were obtained by simulating the effects of the air curtain on smoke diffusion behavior under different wind speeds, air supply angles, and fire source conditions. These data formed a sample set covering various operating conditions.

[0039] After acquiring high-dimensional flow field snapshot data covering the parameter space, singular value decomposition is performed on the snapshot matrix in COMSOL's Proper Orthogonal Decomposition module to extract the dominant modes and evolutionary features of the system: Let s be the snapshots of the transient flow field collected by the system under different parameter conditions. i , where i = 1, 2, …, M, represents the i-th time step or simulation sample. The time mean field of all samples is calculated as follows: ; Construct the covariance matrix: ; in, This represents the average state vector of the system.

[0040] Perform singular value decomposition on the covariance matrix C, i.e.: ; Where V is the eigenvector matrix and Λ is the eigenvalue diagonal matrix, reflecting the energy occupied by each principal mode.

[0041] According to the energy principle Determine the truncation order r, and obtain the reduced-order basis by truncating the first r feature modules. ; Among them, 𝑉 𝑟 To form a matrix from the first r eigenvectors, Λ 𝑟 For the corresponding eigenvalue diagonal matrix, It is a reduced-order orthogonal basis matrix.

[0042] Depend on The online projection coefficients are obtained, and the reconstructed field of the original flow field is: ; in, This is the projection coefficient vector of the system in the reduced-order space, reflecting the excitation intensity of each principal mode; U is the average state vector of the system. r 'a' represents the reconstructed dynamic component.

[0043] For processing nonlinear operators, empirical interpolation methods are used to calculate the nonlinear terms at some sampling points, and the interpolation matrix is ​​then used to... accomplish A highly efficient approximation; Where N(s) is a nonlinear operator, Used to reconstruct global nonlinear behavior from a finite number of sampling points; This method significantly reduces computational complexity while ensuring model accuracy, allowing the system state equations to be solved only in a low-dimensional subspace, thus deriving a low-order subspace model R(·).

[0044] As a preferred embodiment, the specific process of S3 is as follows: S3.1 Constructing the Training Sample Set: The training samples come from the dataset generated by the low-order subspace model. The input part is a multi-dimensional time series sampled at a frequency of 1 Hz within the last 5 seconds, including the sensor status (temperature, smoke concentration) placed in front of the air curtain. The output part is the key physical response quantities behind the air curtain, including temperature and smoke concentration. After normalization, the input data is divided into a training set (70%), a validation set (15%), and a test set (15%) to ensure the generalization performance of the model. Verification shows that the surrogate model F can quickly predict the response behind the air curtain without high-fidelity simulation, supporting high-frequency interaction and cross-scene generalization. The data acquisition module is equipped with multiple sensors, forming a front and rear sensor array, to collect real-time temperature and flue gas concentration parameters in the areas before and after the air curtain. The front and rear sensors are arranged symmetrically or correspondingly, covering the typical plume channel range. The front array S... f Located on the windward side of the air curtain, it is used to capture the temperature T and concentration C of the smoke plume; the rear array S b Located on the wind side behind the air curtain, it is used to detect the smoke conditions behind the air curtain barrier.

[0045] The two arrays are symmetrically distributed in the plane, forming a two-dimensional sampling grid that covers the typical plume channel range. The data acquisition module synchronously collects sensor data at a set sampling period and aligns the timestamps to form a sliding time window sequence {(T,C)}{t-5s:t} of the most recent 5 seconds, providing input for subsequent agent modeling and policy inference; this spatial sensor array improves the sampling density and coverage of flue gas characteristics. These sensors are connected to the control module via a standard interface and transmit data to the state construction module for integration at a set period. S3.2, Build a descending-order proxy model using the Deep Learning Toolbox: 1) Set the sequenceInputLayer as the input layer to receive multidimensional time series data: The input is a multidimensional sequence of length T=50, with a sampling interval of 1s and a total duration of 5s. The expression is: ; Where, x t It is composed of historical sequences of temperature and flue gas concentration from 5 measuring points in front of the air curtain; 2) A two-layer LSTM is used, with each layer having a hidden state dimension h=128, to capture long-range dependencies and short-term fluctuations in the time series. The single-step update formula for the ℓth layer is:

[0046] Among them, f t i t o t These are the activation results of the forget gate, input gate, and output gate, respectively; c t For the state of the memory unit; h t Let be the hidden state at the current time step; σ(⋅) and tanh(⋅) are the Sigmoid and hyperbolic tangent activation functions, respectively.

[0047] After the hidden state at the end of the last LSTM layer, add a fully connected layer and map it to the physical response: ; in, This is the final output of the LSTM hidden state. The hidden state is mapped to the target response behind the air curtain, and the two-dimensional output corresponds to the average temperature and smoke concentration 1 m behind the air curtain, respectively. 4) The mean squared error (MSE) is used as the loss function during training: ; Where, N batchThe number of samples in a single training batch; by analyzing the predicted values Compared with the true value The average of the squared errors can effectively reflect the accuracy of the model's predictions.

[0048] The optimization method used was Adam, with an initial adaptive learning rate of 0.001, each mini-batch containing 64 samples, a maximum training epoch of 200, and an early stopping strategy on the validation set to avoid overfitting.

[0049] After network training, this LSTM model can rapidly predict the response behind the air curtain from the input state-action sequence within milliseconds. Compared to directly calling COMSOL full-order simulation, the inference speed of this model is improved by hundreds of times, which is sufficient to support high-frequency interaction and cross-scenario training of reinforcement learning control strategies. The resulting surrogate model has both physical prior and data-driven characteristics. On the one hand, it inherits the main modal distribution extracted by POD order reduction; on the other hand, it uses LSTM to capture complex nonlinear time evolution laws, thereby more accurately predicting the smoke control performance of the air curtain under different control parameters, providing reliable support for intelligent optimization control.

[0050] As a preferred embodiment, the specific process of S4 is as follows: S4.1. Each episode generates a completely new historical sequence of temperature and concentration ahead using random or adaptive perturbation as the initial state. The action policy network obtains the air curtain parameters and uses a proxy model to provide millisecond-level state transitions. ; Wherein, the state vector s t Forward temperature-concentration history sequence of 5 s × 1 Hz; action space These are continuous motion parameters, including wind speed of 4–8 m / s and angle of 10–20 degrees. S4.2 To capture time dependence and lag effects, both the policy network and the value network share a two-layer 128-dimensional LSTM mathematical skeleton. The tail output heads are connected to the regression layer MSE and the policy-value layer SAC loss, respectively, to achieve dual-task reuse in the same feature space. This extracts the main features from the original time series and completes action decision-making and value assessment based on these features. The policy network π ϕ It is a continuous action policy that outputs the mean μ of the actions. t and logarithmic standard deviation logσ t :

[0051] in, This is the final output of the LSTM hidden state, 𝜇 𝑡 Let be the mean vector of the action distribution, 𝜎 𝑡 Let be the standard deviation vector of the action distribution, 𝑊 𝜇 With 𝑊 𝜎 These are the corresponding weight matrices, 𝑏 𝜇 With 𝑏 𝜎 This is a bias term.

[0052] Action sampling: ; S4.3 In the design of the reward function, the system comprehensively incorporates the air curtain's smoke control performance, energy consumption, and control stability into the evaluation to ensure that the direction of strategy optimization meets the needs of actual application. The reward function consists of three parts: safety reward, energy consumption penalty, and action stability penalty, and its overall form is expressed as follows: ; in, w 1 , w 2 , w 3 The weight parameters are respectively for safety reward, energy consumption penalty, and motion smoothness penalty, satisfying... w 1 > w 2 , and w 1 > w 3 ; 1) Security Bonus Items: ; Where C(t) represents the smoke concentration behind the air curtain, T(t) represents the peak temperature; α1 and α2 are normalization coefficients; when C(t) or T(t) exceeds the safety threshold, the system applies additional negative rewards to enforce safety. 2) Energy consumption penalty item: ; Where P(t) is the power consumption of the air curtain at time t, and β is the energy consumption weighting coefficient; the energy consumption penalty term guides the strategy to minimize energy consumption while satisfying the smoke control effect; 3) Penalty for smooth movement: ; Among them, a t This represents the control action vector at the current moment, i.e., wind speed and injection angle; γ is the smoothness weighting coefficient; the action stability penalty term is used to suppress drastic fluctuations in control parameters, ensuring the stability of air curtain operation and the reliability of the equipment. Through the above reward function design, the system can automatically balance the three objectives of smoke control effect, energy efficiency and control stability during the training process, and finally obtain an intelligent air curtain control strategy that prioritizes safety, optimizes energy consumption and balances stability.

[0053] S4.4, Policy Gradient and Update adopt the Soft Actor-Critic framework, with the goal of maximizing cumulative reward: At the start of each episode, an initial state s0 is randomly generated, including historical sequences of temperature and concentration ahead. A strategy model for generating control actions, i.e., wind speed and injection angle, is then generated based on the temperature and concentration state sequence in front of the air curtain over the most recent 5 seconds. After extracting temporal features through LSTM encoding, the control command action parameters a are output. t ; Use the proxy model to quickly provide the next state s t+1 =F(s t ,a t The agent model performs interactive learning; in parallel, a value evaluation model is used to evaluate the long-term reward under the current state and action, so that the policy model receives correct feedback and calculates the reward value R. t ; the experience of each interaction (s t， a t， r t， s t+1 The data is stored in the replay cache D. A mini-batch is randomly drawn from the replay cache D to update the policy and value network, and then updated through the Adam optimizer. ; Among them, 𝑉 𝜓 For the value function network output, 𝜋 𝜙 is the output of the policy network, is the discount factor, and is the entropy adjustment coefficient.

[0054] To improve training efficiency and reduce the impact of temporal correlation between samples, empirical replay and mini-batch random sampling are introduced during training; simultaneously, a target value network is set, and a soft update method is adopted. The soft update coefficient τ is set to 0.01, ensuring smooth convergence of the value evaluation model during training. The rapid inference capability of the surrogate model enables the entire system to accumulate a large number of interaction samples in a short time, thereby training a robust air curtain control strategy that performs well in various scenarios and obtaining the optimal policy function π. θ*( s); The system outputs control actions including the air curtain's wind speed setpoint and spray angle setpoint. The action range is normalized and subject to boundary constraints to ensure consistency with the actual controllable range of the physical device.

[0055] Among them, v min =4m / s, v max =8m / s, θ min =10°, θ max =20°.

[0056] As a preferred embodiment, the specific process of S5 is as follows: S5.1 The control execution module collects environmental parameters from the sensor array in front of the air curtain in real time, forming continuous time series state data and creating a dynamic control input window; the data acquisition module updates the 5-second sliding window state sequence at a 1-second cycle; the control execution module inputs this time series state into the strategy network for forward inference. The control strategy network includes an encoding layer that extracts time series features based on an LSTM structure, and an action prediction layer that outputs wind speed and angle based on a fully connected structure, to obtain the current optimal (v*, θ*); S5.2 Apply the optimal control strategy to the air curtain physical device to adjust the air curtain's wind speed and angle. The command is transmitted to the air curtain via the actuator to achieve coordinated adjustment of the jet wind speed and angle; front and rear arrays (S f S b Real-time data transmission forms a closed loop; Specifically, the control execution module sets the upper limits for the output wind speed and spray angle to be 8 m / s and 20°, respectively; if the output wind speed or angle exceeds its upper limit, the control execution module will automatically adjust it to the upper limit value. The control execution module sets the following limits on the slope of change of wind speed and jet angle: the slope of change of wind speed is limited to a maximum increase of 1 m / s per step; the slope of change of jet angle is limited to a maximum increase of 5° per step; if the slope of change of wind speed and jet angle output by the strategy exceeds these limits, the control execution module adjusts them to the maximum allowable range of change. The control execution module is equipped with a bypass mechanism to handle situations such as sensor malfunctions or insufficient policy confidence, namely: Sensor anomaly detection: By monitoring the deviation between the sensor's measured value and the historical average value, if the deviation exceeds a set threshold (set to 0.1 in this embodiment), the sensor is determined to be abnormal. Policy confidence assessment: The probability density of the policy output action is calculated. If the density is lower than the set threshold (0.9 in this embodiment), the policy confidence is determined to be insufficient.

[0057] Bypass strategy: Once a sensor malfunction is detected or the strategy confidence is insufficient, the control execution module automatically switches to a conservative strategy or fixed parameter mode, setting the wind speed to 6 m / s and the injection angle to 15° to ensure that the system can still operate stably under abnormal conditions.

[0058] Experimental tests verified that the intelligent smoke-proof air curtain parameter optimization system proposed in this invention successfully adjusted the jet velocity and angle of the air curtain in multiple different fire environments, achieving effective control of smoke diffusion. Comparative experimental data shows that the system of this invention has significant advantages in control accuracy, energy consumption, and response speed compared to traditional fixed-parameter control systems.

[0059] In the experiment, the system was able to dynamically adapt to changing conditions in different fire scenarios, accurately predict and optimize the control strategy of the air curtain, thereby effectively improving the control effect of fire smoke diffusion. Even in more complex fire scenarios, the system maintained high-efficiency control performance, verifying the superiority of this method.

Claims

1. An intelligent smoke-preventing air curtain parameter optimization system based on a proxy model and reinforcement learning, characterized in that, The air curtain comprises a high-fidelity simulation and order reduction module, an agent model construction module, a reinforcement learning control strategy training module, a data acquisition module, a control execution module and an air curtain. The high-fidelity simulation and order reduction module is connected with the agent model construction module to provide input of simulation data for the agent model construction module; the reinforcement learning control strategy training module takes the agent model construction module as a training environment; and the data acquisition module is arranged on both sides of the air curtain to sample smoke features. The control execution module is connected with the data acquisition module, the reinforcement learning control strategy training module and the air curtain.

2. The agent model and reinforcement learning based intelligent smoke-prevention air curtain parameter optimization system according to claim 1, characterized in that, The high-fidelity simulation and order reduction module establishes a multi-physics coupling fire smoke diffusion numerical model based on a numerical simulation software to simulate the influence of the air curtain on smoke diffusion behavior under different wind speeds, air supply angles and fire source conditions; through systematic parameter scanning experiments, high-fidelity simulation data of smoke flow, concentration distribution and temperature evolution in a key area behind the air curtain are obtained, a sample set is formed, high-dimensional flow field data are properly orthogonally decomposed, and an experience interpolation method or discrete singular value decomposition is combined to extract system dominant modes and evolution characteristics, so that a low-order subspace model capable of quickly reconstructing system behavior is constructed as a reduction model; The agent model construction module adopts a supervised learning method to construct a mapping relationship model between state time series in front of the air curtain and control parameters and smoke response behind the air curtain based on simulation samples; the agent model adopts a time series regression structure based on a long short-term memory network, the input of which is a sensor state sequence in the past period of time, and the output of which is a key physical quantity of a response behind the air curtain; The reinforcement learning control strategy training module is used to train an optimal control strategy of air curtain wind speed and jet angle in the simulation environment constructed by the agent model to realize intelligent adjustment of smoke diffusion; by introducing a time series modeling method, temperature and smoke concentration physical parameters collected by sensors in front of the air curtain in the past 5 seconds are formed into a state input sequence to capture time dependence and trend characteristics in the smoke evolution process; The time series state input is extracted by a long short-term memory neural network to extract time sequence characteristics and is input into a policy network to output a decision, the action space being continuous variables including jet wind speed and angle of the air curtain; the system adopts a reinforcement learning algorithm to interactively learn in the simulation environment constructed by the agent model, constructs a reward function according to agent output results of concentration, temperature and energy consumption behind the air curtain to guide policy optimization; The data acquisition module is arranged on both sides of the air curtain, forms a spatial sensor array by arranging a plurality of sensors, and is used to collect temperature and smoke concentration parameters in front of and behind the air curtain in a fire environment to construct system state information. The control execution module is used for deploying the control policy trained in the reinforcement learning control policy training module to an actual air curtain control system, so as to realize real-time intelligent adjustment of the air curtain wind speed and the injection angle; the control execution module takes the environmental parameters collected by the sensor array in front of the air curtain in the last 5 seconds as input, forms a continuous time sequence state data, and combines a past control action sequence to form a dynamic control input window; the control policy network of the reinforcement learning control policy training module is a policy function obtained by training in a reinforcement learning stage, which includes an encoding layer for extracting time sequence features based on a long short-term memory structure and an action prediction layer for outputting wind speed and angle instructions based on a full connection structure; the control execution module deploys the policy function to an embedded control hardware or an edge computing platform, maintains a data interface connection with a sensor collection module, continuously receives a sensor state sequence in a 5-second sliding time window, and obtains the current optimal air curtain control parameters through forward calculation.

3. An intelligent smoke-preventing air curtain parameter optimization method based on a proxy model and reinforcement learning, characterized in that, The method comprises the following steps: S1, a multi-physics coupling fire smoke diffusion numerical model is established based on a finite element simulation software; S2, a low-order subspace model capable of quickly reconstructing system behavior is constructed based on a reduction algorithm, and performance data sets under different control parameters are predicted; S3, a relationship between environmental states and control actions is extracted based on the low-order subspace model data by using a time sequence modeling method, and a reduced-order proxy model is constructed by driving a deep learning method; S4, an optimal control policy of the air curtain is trained in the constructed reduced-order proxy model environment based on reinforcement learning; S5, the wind speed and the injection angle of the air curtain are optimized in real time through a control execution module, so as to realize effective control and dynamic adjustment of the fire smoke.

4. The agent model and reinforcement learning based intelligent smoke-prevention air curtain parameter optimization method according to claim 3, characterized in that, The construction process of the fire smoke diffusion numerical model of S1 is as follows: COMSOL Multiphysics software is used as a high-fidelity simulation platform, and a built-in reduction model and proxy modeling are used to process and reconstruct the flow field data; a three-dimensional calculation domain including a fire source, an air curtain and an affected area is established in COMSOL Multiphysics; The calculation region includes a fire source region, an air curtain region and a region affected by the air curtain, and grid division is used to capture the smoke flow and temperature distribution.

5. The agent model and reinforcement learning based intelligent smoke-prevention air curtain parameter optimization method according to claim 3, characterized in that, The specific process of S2 is as follows: S2.1, the mass equation is constructed as follows: ； Wherein, ρ represents the fluid density, t represents the time, u represents the fluid velocity vector, and ∇⋅(ρu) represents the mass flux divergence per unit volume; The momentum equation is as follows: ； Wherein, ρ represents the fluid density, u represents the fluid velocity vector, p represents the pressure, μ represents the fluid dynamic viscosity, ρ0 represents the reference density, g represents the gravity acceleration vector, β represents the thermal expansion coefficient, T represents the local temperature, and T0 represents the reference temperature; S2.2, the energy equation is constructed as follows: ； where p represents fluid density, c p represents the specific heat at constant pressure, T represents the local temperature, u represents the fluid velocity vector, V T represents the gradient operator of the temperature field, k represents the thermal conductivity, represents the heat source per unit volume; S2.3, the smoke component equation is constructed as follows: ； Wherein, C represents the smoke mass fraction, t represents the time, u represents the fluid velocity vector, ∇C represents the spatial gradient of the smoke mass fraction, and D represents the diffusion coefficient of the smoke in the air; S2.4, the state equation is constructed as follows: ； where p denotes the local fluid density, denotes the reference density, β denotes the thermal expansion coefficient, T denotes the local temperature, and To denotes the reference temperature. After collecting the high-dimensional snapshot data covering the parameter space, singular value decomposition is performed on the snapshot matrix in the Proper Orthogonal Decomposition module of COMSOL to extract the dominant modes and evolution characteristics of the system: Let the transient flow snapshots collected by the system under different parameter conditions be s i where i = 1, 2, …, M, represents the i-th time step or simulation sample; the time mean field of all samples is calculated as: ； Construct the covariance matrix: ; wherein, represents the average state vector of the system; A singular value decomposition is performed on the covariance matrix C, i.e.: ; where V is the eigenvector matrix, and A is the eigenvalue diagonal matrix, reflecting the energy occupied by each dominant mode. According to the energy criterion , determine the truncation order r, the first r characteristic modules before truncation get the reduced basis ; wherein, V 𝑟 is a matrix formed by taking the first r eigenvectors, Λ 𝑟 is a diagonal matrix of corresponding eigenvalues, is a reduced-order orthogonal basis matrix; obtained from obtained from ; wherein, is the projection coefficient vector of the system in the reduced order space, reflecting the excitation strength of each dominant mode; is the average state vector of the system, U r a is the reconstructed dynamic component; For the nonlinear operator, the nonlinear term is calculated on partial sampling points by combining the experience interpolation method, and the high efficient approximation of the nonlinear operator is realized by the interpolation matrix implementation of high efficient approximation; where N(s) is a nonlinear operator, for reconstructing the global nonlinear behavior from a finite number of samples; The low-order subspace model R(·) is derived, so that the system state equation only needs to be solved in the low-order subspace.

6. The agent model and reinforcement learning based intelligent smoke-prevention air curtain parameter optimization method according to claim 3, characterized in that, The specific process of S3 is as follows: S3.1, Constructing a training sample set: the training samples come from the data set generated by the low-order subspace model, the input part is the multi-dimensional time series sampled at a frequency of 1 Hz in the last 5 seconds, including the state of the sensor arranged in front of the air curtain, and the output part is the key physical response quantity behind the air curtain, including temperature and smoke concentration; After normalization, the input data is divided into training set, validation set and test set; S3.2, build the following reduced-order surrogate model through Deep Learning Toolbox: 1) Set sequenceInputLayer as the input layer to receive the multi-dimensional time series: The input is a multi-dimensional sequence with a length of T=50, a sampling interval of 1s, and a total time of 5s, expressed as: ； wherein x t The temperature and smoke concentration history sequence of the 5 measuring points in front of the air curtain are spliced. 2) Use two layers of LSTM, each with a hidden state dimension of h=128, to capture long-range dependencies and short-term fluctuations in time series, and the ℓ-th layer single-step update formula is: ； wherein f t , i t , o t are the activation results of the forget gate, input gate and output gate, respectively; c t is the memory cell state; h t is the hidden state at the current time; and σ(·) and tanh(·) are the Sigmoid and hyperbolic tangent activation functions, respectively. Add a fully connected layer after the last hidden state of the LSTM layer and map it to the physical response: ； wherein, is the final output of the LSTM hidden state, ; maps the hidden state to the target response quantity behind the air curtain, outputting two dimensions corresponding to the average temperature and smoke concentration 1 m behind the air curtain, respectively; During training, the mean square error MSE is used as the loss function: ； wherein N batch is the number of samples in a single training batch; the prediction accuracy of the model is effectively reflected by averaging the squared error between the predicted value and the true value . The optimization method uses Adam, the adaptive learning rate initial value is set to 0.001, each mini-batch contains 64 samples, the maximum training period is set to 200, and the early stopping strategy is used on the validation set to avoid overfitting.

7. The agent model and reinforcement learning based intelligent smoke-prevention air curtain parameter optimization method according to claim 3, characterized in that, The specific process of S4 is as follows: S4.1, Each episode generates a new front temperature and concentration history sequence as the initial state with random or adaptive disturbance, the action policy network obtains the air curtain parameters, and the surrogate model provides millisecond-level state transition: ； where the state vector s t is a 5 s x 1 Hz forward temperature-concentration history sequence; action space is a continuous action parameter, including wind speed 4-8 m / s, angle 10-20 degrees; S4.2、For capturing time dependence and hysteresis effects, both policy and value networks share a two-layer 128-dimensional LSTM backbone, whose tail outputs are connected to a regression layer MSE and a policy-value layer SAC loss, respectively, to realize dual-task multiplexing in the same feature space, extract main features from the original time series, and complete action decision and value evaluation on this basis; the policy network π ϕ is a continuous action policy, outputting the mean μ t and log standard deviation logσ t of the action ； where, is the final output of the LSTM hidden state, 𝑡 is the mean vector of the action distribution, 𝑡 is the standard deviation vector of the action distribution, 𝜇 and 𝜎 are the corresponding weight matrices, 𝜇 and 𝜎 are the bias terms; Action sampling: ; S4.3, In the design of the reward function, the system integrates the air curtain smoke control performance, energy consumption, and control stability into the evaluation to ensure that the direction of policy optimization meets the actual application requirements; The reward function is composed of three parts: safety reward, energy consumption penalty, and action stability penalty, and its overall form is expressed as: ； wherein, w 1 , w 2 , w 3 are weight parameters of the safety reward, the energy consumption penalty, and the action smoothness penalty, respectively, satisfying w 1 > w 2 , and w 1 > w 3 ; 1) Safety reward term: ； Wherein, C(t) represents the smoke concentration behind the air curtain, T(t) represents the temperature peak value; is a normalization coefficient; when C(t) or T(t) exceeds the safety threshold, the system applies an additional negative reward to force the safety guarantee; 2) Energy consumption penalty term: ； Where P(t) is the power consumption of the air curtain at time t, and β is the energy consumption weight coefficient; 3) Action stability penalty term: ； wherein a t represents the control action vector at the current time, i.e. the wind speed and the jet angle; γ is a smoothing weight coefficient; the action smoothness penalty term is used to suppress the drastic fluctuation of the control parameters, ensuring the stability of the air curtain operation and the reliability of the equipment; S4.4, The strategy gradient and update use the Soft Actor-Critic framework, which aims to maximize the cumulative reward: At the beginning of each episode, an initial state s0 is randomly generated, including the temperature and concentration history sequence in front; the strategy model of the generated control action, i.e., the wind speed and the spray angle, is based on the temperature and concentration state sequence in the last 5 seconds in front of the air curtain, and the control instruction action parameters a are output after the time sequence features are extracted by LSTM encoding t ; the next state s t+1 is quickly provided by using the agent model t =F(s t ,a t ), and the interactive learning of the agent model is performed; in parallel with this, the value evaluation model is used to evaluate the long-term return under the current state and action, so that the strategy model can obtain correct feedback, and the reward value R t， is calculated t， ; the experience (s t， a t+1 ) of each interaction is stored in the replay buffer D, a mini-batch is randomly extracted from the replay buffer D, which is used to update the strategy and value network, and the strategy and value network are updated by the Adam optimizer: ； where V 𝜓 is the output of the value function network, π 𝜙 is the output of the policy network, γ is the discount factor, and β is the entropy regulation coefficient. In order to improve the training efficiency and reduce the influence of time correlation between samples, experience replay and small batch random sampling are introduced in the training; at the same time, the target value network is set, and the soft update method is adopted: Wherein, the soft update coefficient τ is 0.01, which makes the value evaluation model converge smoothly in the training; The fast reasoning ability of the surrogate model enables the whole system to accumulate a large number of interaction samples in a short time, so as to train the air curtain control strategy which performs stably in various scenes and obtain the optimal strategy function . The control actions output by the system include the air curtain's wind speed set value and jet angle set value, and the action range is normalized and boundary-constrained to ensure consistency with the actual controllable interval of the physical device: ； where v min = 4 m / s, v max = 8 m / s, θ min = 10°, and θ max = 20°.

8. The agent model and reinforcement learning based intelligent smoke-prevention air curtain parameter optimization method according to claim 3, characterized in that, The specific process of S5 is as follows: S5.1, the control execution module collects the environmental parameters of the sensor array in front of the air curtain in real time, forms a continuous time sequence state data, and forms a dynamic control input window; the data acquisition module updates the 5s sliding window state sequence at a period of 1s; the control execution module forwards the time sequence state to the strategy network for inference, the control strategy network includes an encoding layer based on LSTM structure to extract time sequence features, and an action prediction layer based on full connection structure to output wind speed and angle, to obtain the current optimal (v*, θ*); S5.2, apply the optimal control strategy to the air curtain physical device, adjust the air curtain wind speed and angle, instruct the air curtain to act on the execution mechanism, realize the linkage adjustment of the jet wind speed and angle; front and rear sensor arrays S f , S b Real-time data transmission, forming a closed loop; Specifically, the upper limit of the wind speed and jet angle set by the control execution module is 8 m / s and 20° respectively; if the wind speed or angle output by the strategy exceeds the upper limit, the control execution module automatically adjusts it to the upper limit value; The control execution module limits the change slope of the wind speed and jet angle respectively: the wind speed change slope is limited to an increase of at most 1 m / s per step; the jet angle change slope is limited to an increase of at most 5° per step; if the change slope of the wind speed and jet angle output by the strategy exceeds the limit, the control execution module adjusts it to the maximum allowable change range; The control execution module sets a bypass mechanism to deal with sensor abnormalities or insufficient strategy confidence, which includes: Sensor anomaly detection: by monitoring the deviation of the sensor measurement value from the historical average value, if the deviation exceeds the set threshold, the sensor is determined to be abnormal; Strategy confidence evaluation: by calculating the probability density of the strategy output action, if the density is lower than the set threshold, the strategy confidence is determined to be insufficient; Bypass strategy: once the sensor anomaly or insufficient strategy confidence is detected, the control execution module automatically switches to a conservative strategy or a fixed parameter mode, sets the wind speed to 6 m / s and the jet angle to 15°, to ensure that the system can still operate stably under abnormal conditions.

Images