Digitally managed basement ventilation and smoke extraction system and method

Abstract

This invention discloses a basement ventilation and smoke extraction system and method based on digital management. The method includes: constructing a digital twin architecture comprising a terminal layer, an edge layer, and a cloud layer to perform real-time mapping and interaction between the physical system and the virtual model of basement ventilation and smoke extraction; using multiple sensors for integrated monitoring at the terminal layer to collect relevant data parameters of basement ventilation and smoke extraction; using a highly integrated edge computing gateway to preprocess the collected data; the area controller using a dynamic Bayesian network to fuse the preprocessed multi-source data to construct a local environmental state model; the edge layer using deep reinforcement learning algorithms and a digital twin model to simulate various scenarios for real-time state prediction and generate an emergency plan library; and the central server at the cloud layer running a dual-track decision engine, combining a deterministic rule base and an adaptive deep reinforcement learning model for adaptive adjustment. This invention achieves efficient and precise coordinated control of ventilation and smoke extraction.

Description

Technical Field

[0001] This invention relates to the field of ventilation and smoke extraction technology, specifically to a basement ventilation and smoke extraction system and method based on digital management. Background Technology

[0002] In existing complex basement environments, ventilation and smoke extraction systems mainly consist of smoke extraction terminals, manual control terminals, smoke prevention and extraction control terminals, and a central control terminal. These components are set up in multiple different areas and work together. The core components and their corresponding control terminals, such as regional manual terminals and centralized ventilation and smoke extraction terminals, are usually deployed in physically dispersed and logically isolated areas. Their operating status, fault alarms, start and stop signals, and other key data information are stored separately or enclosed in their respective independent subsystems, such as independent building automation subsystems, fire alarm subsystems, and dedicated equipment monitoring platforms.

[0003] Therefore, the data generated by the components of a distributed deployment are often scattered across their respective independent systems. This physical dispersion and logical isolation of data makes real-time global situational awareness difficult. Management terminals cannot seamlessly integrate and intuitively monitor dynamic information such as valve opening and closing status, fan operating parameters, terminal control commands, and module feedback signals on a single interface, making it difficult to quickly identify abnormal correlations across devices and regions. Simultaneously, system response is lagging and linkage efficiency is low. Data between different subsystems cannot achieve low-latency, high-fidelity interaction and fusion computing, resulting in low accuracy of cross-system collaborative analysis. In emergency situations, control commands must be transmitted layer by layer through multiple systems, and the manual intervention commonly used in existing systems mainly relies on distributed field terminals. This makes it impossible to adaptively respond to real-time conditions or dynamically optimize and adjust control based on real-time integrated global data. Summary of the Invention

[0004] The purpose of this invention is to provide a basement ventilation and smoke extraction system and method based on digital management to solve the problems mentioned in the background art.

[0005] The specific technical solution provided by this invention is as follows: a basement ventilation and smoke exhaust system based on digital management, comprising the following functional units: a global perception fusion unit, an edge intelligent gateway unit, a digital twin driving unit, a deep reinforcement learning decision-making unit, a dynamic execution verification unit, and an adaptive adjustment control unit.

[0006] A digitally managed basement ventilation and smoke extraction method includes the following steps:

[0007] S1: Construct a digital twin architecture including a terminal layer, an edge layer, and a cloud layer to perform real-time mapping and interaction between the physical system and the virtual model of basement ventilation and smoke extraction.

[0008] Preferably, the real-time mapping and interaction between the physical system and the virtual model for basement ventilation and smoke extraction specifically includes: terminal layer sensors in the digital twin architecture collect various parameters of the physical system according to a set sampling frequency; the edge layer preprocesses the received data and transmits it to the cloud layer, while simultaneously updating the corresponding parameters in the virtual model in real time based on this data; the virtual model uses the updated parameters to perform simulation calculations, simulating the operation of the physical system under different working conditions, and feeds back the simulation results to the edge layer and the cloud layer; the edge layer and the cloud layer then formulate control strategies based on the simulation results of the virtual model and actual needs, and send control commands to the actuators in the terminal layer. The actuators adjust the physical system according to the commands, realizing real-time interaction between the physical system and the virtual model.

[0009] S2: At the terminal layer, multiple sensors are used for integrated monitoring to collect data parameters related to ventilation and smoke extraction in the basement.

[0010] S3: The collected data is preprocessed using a highly integrated edge computing gateway.

[0011] S4: The regional controller uses a dynamic Bayesian network to fuse preprocessed multi-source data and construct a local environmental state model.

[0012] Preferably, constructing a local environment state model specifically includes:

[0013] S41: Determine the state variables and observation variables of the local environment;

[0014] S42: Construct a dynamic Bayesian network structure that includes the initial network structure and the transition network structure;

[0015] S43: Determine the parameters of the dynamic Bayesian network and obtain the conditional probability distribution among the variables;

[0016] S44: Using preprocessed multi-source data for dynamic Bayesian network inference;

[0017] S45: Construct a corresponding local environment state model as the basis for the edge layer to perform state prediction and generate emergency plans.

[0018] S5: The edge layer uses deep reinforcement learning algorithms and digital twin models to simulate various scenarios for real-time state prediction and generate an emergency plan library.

[0019] Preferably, simulating multiple scenarios for real-time state prediction specifically includes:

[0020] S51: Determine the state space, action space, and reward function;

[0021] S52: Construct a DQN+LSTM model; the DQN+LSTM model uses an LSTM network and a deep neural network;

[0022] S53: Calculate the target value using the target network, and select actions and update parameters based on the evaluation network;

[0023] S54: Use a digital twin model to simulate multiple scenarios, and use the output of the digital twin model as environmental feedback to interact with the DQN+LSTM model;

[0024] S55: Obtain feedback based on the reward function, gradually optimize the action strategy, and based on the optimized strategy, make real-time predictions and outputs of the system state under different scenarios.

[0025] S6: The central server in the cloud layer runs a dual-track decision engine, which combines a deterministic rule base with an adaptive deep reinforcement learning model to perform overall adaptive and collaborative control.

[0026] Preferably, the operation process of the dual-track decision engine specifically includes:

[0027] S61: The dual-track decision engine simultaneously calls the deterministic rule base and the adaptive deep reinforcement learning model to analyze the global environment state;

[0028] S62: The deterministic rule base matches the current state according to the preset rules. If there is a rule that matches perfectly, the corresponding control command is generated directly. If there is no rule that matches perfectly or the execution effect of the rule is not good, the adaptive deep reinforcement learning model is started.

[0029] S63: The reinforcement learning model calculates the optimal control policy based on the global state, supplementing or adjusting the deterministic rules;

[0030] S64: Integrate and evaluate the control strategies generated by the two decision paths, and select the control strategy with the best overall performance as the final decision;

[0031] S65: The final control command is sent to each edge layer, which then forwards it to the actuator in the terminal layer to achieve coordinated control of the entire system.

[0032] S66: Simultaneously feed the decision-making process and execution results back to the reinforcement learning model for continuous model updates and optimization.

[0033] S7: Synchronously retain the hybrid decision-making mechanism. Under normal conditions, the system performs adaptive and collaborative control. Under emergency conditions, it utilizes human intervention terminals to transform human experience into system knowledge, forming a human-machine collaborative decision-making mechanism.

[0034] Compared to existing technologies, this invention effectively balances the needs of real-time performance and complex decision-making through a three-layer architecture. The edge intelligent gateway unit achieves local data preprocessing and primary policy caching through distributed edge computing, and adopts distributed ledger technology to promote the sharing of key states between adjacent edge nodes, enabling cross-regional policy collaboration. Simultaneously, by utilizing a dual-track decision engine that integrates a deterministic rule base and an adaptive reinforcement learning model, it ensures standardized responses in routine scenarios while optimizing dynamic strategies through global data. This overcomes the limitations of traditional systems that rely on fixed rules and cannot adaptively adjust, comprehensively improving the global perception capability, response speed, decision-making accuracy, and operational reliability of basement ventilation and smoke extraction systems, providing a digital and intelligent solution for the safety management of complex underground spaces. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the functional units of a basement ventilation and smoke extraction system based on digital management, provided in an embodiment of the present invention.

[0036] Figure 2 This is a flowchart of the operation method of a basement ventilation and smoke exhaust system based on digital management provided in an embodiment of the present invention;

[0037] Figure 3 This is a flowchart of the operation method for data fusion and model construction provided in an embodiment of the present invention;

[0038] Figure 4 This is a flowchart of the edge layer operation method provided in an embodiment of the present invention;

[0039] Figure 5 This is a flowchart of the operation method of the dual-track decision engine provided in the embodiment of the present invention;

[0040] Figure 6 This is a schematic diagram of the logic execution of the cloud layer provided in an embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0042] Example 1:

[0043] like Figure 1 As shown in the figure, the basement ventilation and smoke exhaust system based on digital management described in this embodiment includes: a global perception fusion unit, an edge intelligent gateway unit, a digital twin driving unit, a deep reinforcement learning decision-making unit, a dynamic execution verification unit, and an adaptive adjustment control unit.

[0044] In this embodiment, the global perception fusion unit is used to collect device status and environmental data in real time from all dimensions.

[0045] For example, existing basement ventilation and smoke extraction systems involve control terminals equipped with various types of sensors, including those for current, vibration, temperature and humidity, and open / closed status. These sensors are uniformly connected to data from subsystems such as fire protection and building control via industrial protocol converters. Based on embedded RFID tags, physical devices are bound to digital objects, standardizing the heterogeneous data streams generated by each terminal. This unit, building upon RFID device binding, performs spatial topology registration. Specifically, by adding a spatial relationship modeling engine, it automatically constructs a three-dimensional topology map of "device-area-subsystem," such as "B1—Zone 1 fan—fire damper V1," simultaneously forming virtual sensing of blind spots. It integrates the CFD compensation capabilities of the digital twin drive unit, dynamically generating key parameters for uncovered areas, achieving compatibility with data from subsystems such as fire protection and building control, and automatically identifying conflicting data sources. For example, it triggers verification when the temperature difference between fire protection and building control exceeds a threshold. Building upon existing data binding methods that rely solely on devices, this unit addresses the initial data association problem caused by physical dispersion and difficulty in correlation analysis by adding the dual dimensions of spatial relationships and data reliability. This provides a highly correlated data foundation for subsequent digital management.

[0046] In this embodiment, the edge intelligent gateway unit is used to receive sensing data by configuring a separate edge computing unit in each region, perform multi-source data fusion, align the timestamps of multi-source heterogeneous data, generate a primary control policy cache, and realize local preprocessing of sensing data.

[0047] For example, this unit runs a lightweight DRL model, generating control policies based on local data when the network is down, such as according to... The concentration gradient autonomous adjustment valve enables cross-edge consensus coordination and utilizes distributed ledger technology to allow adjacent edge nodes to share key states. For example, by sharing the exhaust pressure value of area A, cross-area strategy coordination can be achieved, thus avoiding overload of the fan in area B. It can not only perform local preprocessing but also achieve autonomous coordination between edge nodes, eliminating cross-regional linkage lag.

[0048] In this embodiment, the digital twin drive unit is used to calculate the three-dimensional CFD flow field in real time, simulate the impact of control commands on the flow field, and generate virtual sensor data of the uncovered area for blind spot compensation.

[0049] For example, in this unit, the airflow / smoke diffusion effect is simulated based on control commands to pre-test the feasibility of strategies, such as displaying the backflow zone caused by closing a certain valve, verifying virtual and real interaction, comparing virtual sensor data with actual feedback, automatically marking the model error area when the deviation is greater than the set threshold, simulating the fire spread path, pre-calculating the smoke extraction efficiency of multiple strategies and ranking them.

[0050] In this embodiment, the deep reinforcement learning decision unit is used for real-time state change analysis of multi-scale anomaly detection based on graph neural networks, and realizes intelligent switching of system modes through a multi-threshold event triggering mechanism.

[0051] For example, this unit uses a multi-scale graph neural network to perform anomaly correlation analysis on the device layer, region layer, and system layer simultaneously, achieving cross-scale global optimization. It can also automatically adjust alarm thresholds based on historical event data, improve the sensitivity of electrical fire early warning when humidity rises in summer, and perform strategy conflict arbitration. When energy-saving mode conflicts with emergency smoke exhaust command, it will automatically make a decision according to a preset rule matrix.

[0052] In this embodiment, the dynamic execution verification unit is used to convert the strategy into a device control signal, compare the predicted value with the actual sensor feedback and dynamically correct the CFD parameter error, and roll back to the safety strategy when the deviation exceeds the threshold.

[0053] For example, this unit uses dual channels for execution verification. The prediction channel sends control commands to the three-dimensional situation unit for pre-simulation to verify effectiveness. The reality channel compares the device feedback with the expected state in real time. When the deviation exceeds the limit, a rollback is triggered. When the CFD model continues to make inaccurate predictions, a parameter calibration algorithm is automatically started, such as correcting the boundary turbulence coefficient. At the same time, a safety strategy container stores a pre-verified backup strategy library, which is directly invoked in emergency situations such as "anti-backflow safety valve opening sequence". Under the existing verification method that only corrects after the fact, this unit establishes a triple guarantee of pre-execution verification, real-time feedback and model self-healing to solve the problem of command transmission risk.

[0054] In this embodiment, an adaptive adjustment control unit is used to optimize the sensor network structure by combining dynamic topology adjustment. The controller, based on a deep reinforcement learning algorithm, performs condition prediction and self-learning, optimizing operating parameters based on historical data.

[0055] For example, this unit automatically updates the system topology based on equipment failures and additions. For instance, after fan 1 stops, it automatically assigns the associated valves to the control of the standby fan. At the same time, it compresses the large DRL model trained in the cloud into a lightweight version that can be deployed at the edge, updates it incrementally on a regular basis, and finally performs full-link tracing, recording the data versions of each stage of perception, decision-making, and execution to trace the root cause of the failure.

[0056] Example 2:

[0057] like Figure 2 As shown in the figure, the operation method of the basement ventilation and smoke exhaust system based on digital management described in this embodiment includes the following steps:

[0058] S1: Construct a digital twin architecture including a terminal layer, an edge layer, and a cloud layer to perform real-time mapping and interaction between the physical system and the virtual model of basement ventilation and smoke extraction.

[0059] In this embodiment, the terminal layer serves as the data entry point for the entire system. It comprises hardware devices such as sensors and actuators deployed throughout the basement, responsible for real-time sensing of various physical environment parameters and executing control commands. This layer includes temperature sensors installed on the ceiling, smoke detectors distributed in different areas, and control terminals for smoke exhaust fans, all of which directly interact with the physical space. The edge layer primarily handles preliminary data processing and localized decision-making tasks. Composed of edge computing gateways, area controllers, and other devices, it enables rapid analysis and response before data transmission to the cloud, reducing data transmission volume and latency. For example, when the smoke concentration in a certain area suddenly increases, the edge layer can quickly mobilize the smoke exhaust equipment in that area for preliminary treatment, preventing the spread of the disaster. The cloud layer, acting as the system's "central brain," consists of high-performance servers and a cloud computing platform. It is responsible for the global management and optimization of the entire basement's ventilation and smoke exhaust system. It can store massive amounts of historical data and provide decision support to the edge layer through big data analysis and artificial intelligence algorithms. Simultaneously, it can monitor and remotely control the overall system's operational status in real time.

[0060] For example, the specific implementation process of real-time mapping and interaction between the physical system and the virtual model includes: the terminal layer sensors collect various parameters of the physical system, such as temperature, humidity, smoke concentration, and wind speed, according to a set sampling frequency, and transmit this data through Ethernet, LoRa, NB-IoT, etc. The data is transmitted to the edge layer via wired or wireless communication. After preprocessing the received data, the edge layer transmits the processed data to the cloud layer and updates the corresponding parameters in the virtual model in real time based on this data, enabling the virtual model to accurately reflect the current state of the physical system. The virtual model then uses the updated parameters to perform simulation calculations, simulating the operation of the physical system under different working conditions, and feeds the simulation results back to the edge layer and the cloud layer. The edge layer and the cloud layer then formulate control strategies based on the simulation results of the virtual model and actual needs, and send control commands to the actuators in the terminal layer. The actuators adjust the physical system according to the commands, thereby realizing real-time interaction between the physical system and the virtual model. When the virtual model simulates that the smoke extraction effect in a certain area is not good, the cloud layer will analyze the cause and formulate an optimization plan. The edge layer will then control the smoke extraction fan in that area to increase its speed or turn on the backup fan according to the optimization plan. At the same time, the virtual model will update the smoke extraction effect in that area in real time until the expected goal is achieved.

[0061] For example, in a large underground parking lot using a digital twin architecture, multiple temperature sensors, smoke detectors, and wind speed sensors are deployed at the terminal layer, distributed in various areas of the parking lot; corresponding edge computing gateways are set up at the edge layer, with each gateway responsible for managing the corresponding sensors and smoke extraction equipment in the corresponding area; and a cloud computing platform is built at the cloud layer, equipped with corresponding high-performance servers. When a vehicle spontaneously combusts in a parking lot area, temperature sensors and smoke detectors in that area quickly collect data on rising temperature and excessive smoke concentration, transmitting the data to the corresponding edge computing gateway. The edge gateway preprocesses the data, transmitting it to the cloud platform while simultaneously updating the temperature and smoke concentration parameters for that area in the virtual model. Based on these parameters, the virtual model simulates the fire's spread and smoke diffusion path, feeding the simulation results back to the edge gateway and cloud platform. The cloud platform, combining historical data and simulation results, formulates a global smoke extraction plan. The edge gateway, according to this plan, controls all smoke extraction fans in that area to start and adjusts their speed to maximum, while simultaneously closing the fireproof roller shutters connecting that area to other areas. The actuators at the terminal layer execute corresponding operations according to instructions, and changes in the physical system are reflected in the virtual model in real time. The virtual model continues to simulate and provide feedback until the fire is brought under control and the smoke is effectively extracted.

[0062] S2: At the terminal layer, multiple sensors are used for integrated monitoring to collect data parameters related to ventilation and smoke extraction in the basement.

[0063] In this embodiment, the present invention collects basement smoke exhaust and ventilation related data through a distributed sensor network. The terminal layer deploys three layers of sensor nodes via a hierarchical sensor network: a top-level environmental parameter sensor, a middle-level smoke and thermal imaging sensor, and a bottom-level airflow dynamic sensor. The top-level environmental parameter sensor is used to collect temperature, humidity, and other data. , Environmental parameters such as PM2.5 are collected by a mid-level smoke and thermal imaging sensor, which collects smoke data and thermal imaging data generated in the basement. A bottom-level airflow dynamic sensor collects airflow dynamic parameters. Each sensor collects target parameters according to a preset sampling period based on the changing characteristics of its parameters. The collected raw data undergoes signal conditioning circuitry within the sensor, including amplification, filtering, and A / D conversion, converting analog signals into digital signals. The digital signals are then transmitted to the aggregation node at the terminal layer or directly to the edge computing gateway at the edge layer via the data transmission module. The aggregation node summarizes and converts the received data from multiple sensors before transmitting it to the edge computing gateway, thus forming a high-resolution sensing network and improving spatial resolution.

[0064] For example, the sensors are connected to the aggregation node of each fire compartment via an RS485 bus, and the aggregation node then transmits the data to the edge computing gateway via Ethernet. During daily operation, temperature sensors monitor temperature changes in the mall in real time, and upload data promptly when the temperature in a certain area exceeds a set threshold; smoke sensors constantly monitor for smoke generation, and immediately send alarm signals and smoke concentration data upon detection; wind speed sensors monitor the wind speed at the air supply and exhaust vents to ensure the normal operation of the ventilation system; CO concentration sensors monitor the CO content in the air to ensure personnel safety; current and vibration sensors monitor the operating status of the fans, and issue timely fault warnings when abnormal current or excessive vibration occurs. Through the coordinated work of these sensors, comprehensive and real-time acquisition of ventilation and smoke exhaust data parameters related to the underground mall is achieved.

[0065] S3: The collected data is preprocessed using a highly integrated edge computing gateway.

[0066] In this embodiment, the highly integrated edge computing gateway provides preprocessing operations for the raw data collected. It performs preprocessing operations, including data cleaning, data conversion, data fusion, and data compression, on the raw data collected by different types of connected sensors. Based on data cleaning, noise, outliers, and missing values ​​in the raw data are removed. Data conversion transforms data of different formats and units into a unified format and unit. Data fusion is used to comprehensively process data of the same or related types from multiple sensors. Finally, a data compression algorithm is used to compress the preprocessed data, thereby improving the efficiency of data transmission and storage while ensuring data accuracy.

[0067] S4: The regional controller uses a dynamic Bayesian network to fuse preprocessed multi-source data and construct a local environmental state model.

[0068] In this embodiment, as Figure 3 As shown, the regional controller utilizes a dynamic Bayesian network for data fusion and model building. The specific operation process is as follows:

[0069] S41: Determine the state variables and observation variables of the local environment;

[0070] For example, state variables describe key parameters of the local environmental state, such as normal, high, and excessively high temperature, normal, excessive, and severely excessive smoke concentration, and good, average, and poor ventilation. Observation variables are preprocessed data from various sensors, such as measurements from temperature sensors, smoke sensors, and wind speed sensors.

[0071] S42: Construct a dynamic Bayesian network structure that includes the initial network structure and the transition network structure;

[0072] For example, the initial network structure is used for the initial time. The probabilistic relationship between state variables and observed variables is described. Dependencies between variables are determined through expert knowledge or data learning; for example, temperature depends on the measurement value of a temperature sensor, and smoke concentration depends on the measurement value of a smoke sensor. A transfer network structure is used to describe the relationship between state variables and observed variables. and The probabilistic relationships between state variables at different times describe the evolution of environmental state over time. For example, the current temperature state will affect the temperature state at the next time, and the current ventilation state will affect the temperature and smoke concentration at the next time.

[0073] S43: Determine the parameters of the dynamic Bayesian network and obtain the conditional probability distribution among the variables.

[0074] For example, the parameters of the initial network are obtained by statistical analysis of historical data or estimation by expert experience. For example, the probability that the temperature is normal at the initial moment, the probability that the temperature sensor measurement value is within a certain range when the temperature is normal, etc. The parameters of the transition network are also learned from historical data. For example, the probability that the temperature becomes too high at the next moment when the current temperature is normal.

[0075] S44: Use preprocessed multi-source data to perform inference on dynamic Bayesian networks, that is, estimate the posterior probability distribution of state variables based on the values ​​of observed variables.

[0076] For example, the inference process employs a forward-backward algorithm, calculating the results at each time step based on the forward algorithm. , given before The posterior probability of the state variable at each observation is calculated, i.e., the forward probability; then, the backward algorithm is used to calculate the probability at each time step. Given from arrive The observed values ​​are the posterior probabilities of the state variables, i.e., the backward probabilities; then, by combining the forward and backward probabilities, we obtain the probability at each time step. The posterior probability distribution of the state variables is used to determine the state of the local environment. Finally, based on the posterior probability distribution of the state variables obtained through inference, a local environment state model is constructed. This model can reflect the dynamic changes of the local environment in real time and provide a basis for subsequent state prediction and control decisions.

[0077] S45: Construct a corresponding local environment state model as the basis for the edge layer to perform state prediction and generate emergency plans.

[0078] S5: The edge layer uses deep reinforcement learning algorithms and digital twin models to simulate various scenarios for real-time state prediction and generate an emergency plan library.

[0079] In this embodiment, combined with Figure 4 As shown, the edge layer combines deep reinforcement learning networks and long short-term memory networks using a deep reinforcement learning algorithm to effectively process time-series data and optimize decisions. Combined with the simulation capabilities of a digital twin model, it enables real-time prediction of the basement ventilation and smoke extraction system's status and simulation of various scenarios, ultimately generating an emergency response plan library. The specific process steps include:

[0080] S51: Determine the state space, action space, and reward function.

[0081] For example, the state space includes the state parameters of the current local environment and the operating state of the device; the action space includes the control actions that the edge layer can take; and the reward function is used to evaluate the effect of the action. For example, when a certain action is taken, a positive reward will be obtained if the smoke concentration decreases, the temperature decreases, or the device energy consumption decreases, and a negative reward will be obtained otherwise.

[0082] S52: Construct a DQN+LSTM model.

[0083] For example, LSTM networks are used to process historical state sequence data, capturing long-term dependencies in time series data, and their output serves as input to deep reinforcement learning algorithms. DQN, on the other hand, fits an action value function through a deep neural network to evaluate the expected reward of taking a certain action in a specific state. During model training, an experience replay mechanism is used to store the experience generated by the agent's interaction with the environment, including states, actions, rewards, and the next state, in an experience pool. Then, random sampled experience samples are used for training to improve the model's stability.

[0084] For example, the target value is calculated using the target network simultaneously, and actions are selected and parameters are updated based on the evaluation network. The parameters of the evaluation network are copied to the target network every certain number of steps to reduce fluctuations during training.

[0085] S53: Utilize digital twin models to simulate various scenarios.

[0086] For example, these scenarios include daily operation scenarios, equipment failure scenarios, and fire scenarios. During the simulation, the output of the digital twin model is used as environmental feedback and interacts with the DQN+LSTM model. The agent continuously tries different actions, receives feedback based on the reward function, and gradually optimizes its action strategy. Based on the optimized strategy, the system state under different scenarios is predicted in real time. The predicted content includes the temperature change trend, smoke concentration diffusion path, wind speed distribution, etc., over a period of time.

[0087] S54: The LSTM network uses historical state data and current state data to output a predicted value for the future state.

[0088] For example, DQN selects the optimal control action based on the predicted state value, ensuring that the system can prepare in advance. Finally, based on various simulated scenarios and corresponding optimal control actions, an emergency plan library is generated. Each emergency plan includes a scenario description, triggering conditions, control measures, and expected effects. When a scenario matching a plan occurs during actual operation, that plan is quickly invoked for processing.

[0089] S6: The central server in the cloud layer runs a dual-track decision engine, which combines a deterministic rule base with an adaptive deep reinforcement learning model to perform overall adaptive and collaborative control.

[0090] In this embodiment, the dual-track decision engine in the cloud layer combines the stability of deterministic rules with the adaptability of reinforcement learning to perform global optimization and collaborative control of the entire basement ventilation and smoke exhaust system. The deterministic rule base is a series of fixed control rules based on fire safety regulations, engineering experience, and expert knowledge. These rules specify standardized control measures to be taken under specific conditions, such as "when the smoke concentration in a certain area exceeds 0.1 mg / m³, immediately turn on all smoke exhaust fans in that area and close the fireproof roller shutters in adjacent areas," and "when the operating current of a smoke exhaust fan exceeds 120% of its rated current, automatically reduce its speed to 80% of its rated speed," etc. These rules are clear, fast, and reliable, and can handle some common and recurring situations. The adaptive deep reinforcement learning model employs a similar but larger-scale deep reinforcement learning algorithm to the edge layer, including a deep deterministic policy gradient algorithm. This model takes the environmental conditions of the entire basement, including temperature, smoke concentration, and equipment operating status in each area, as input, and outputs a global control strategy. Through interaction with the global environment, it continuously learns and optimizes the strategy to adapt to complex and changing situations.

[0091] For example, in combination Figure 5 and Figure 6 As shown, the implementation steps of the dual-track decision engine are as follows: The central server receives preprocessed data and local environmental state models uploaded from each edge layer, and integrates this information to construct a global environmental state model, reflecting the operating status of the ventilation and smoke extraction system in the entire basement. The dual-track decision engine simultaneously calls a deterministic rule base and an adaptive deep reinforcement learning model to analyze the global environmental state. The deterministic rule base matches the current state according to preset rules. If a completely matching rule exists, the corresponding control command is directly generated; if no completely matching rule exists or the rule execution effect is poor (e.g., the smoke concentration continues to rise after the rule is executed), the adaptive deep reinforcement learning model is activated. The reinforcement learning model calculates the optimal control strategy based on the global state. This strategy may supplement or adjust the deterministic rules, for example, by turning on a backup fan to enhance the smoke extraction effect, in addition to the rule specifying the activation of some smoke extraction fans. Then, the control strategies generated by the two decision paths are fused and evaluated. Evaluation indicators include smoke extraction efficiency, energy consumption, and safety. The control strategy with the best overall performance is selected as the final decision. Finally, the final control commands are sent to each edge layer, which then forwards them to the actuators in the terminal layer, achieving coordinated control of the entire system. Simultaneously, the decision-making process and execution results are fed back to the reinforcement learning model for continuous updating and optimization, ensuring the model can constantly adapt to new situations.

[0092] S7: Synchronously retain the hybrid decision-making mechanism. Under normal conditions, the system performs adaptive and collaborative control. Under emergency conditions, it utilizes human intervention terminals to transform human experience into system knowledge, forming a human-machine collaborative decision-making mechanism.

[0093] In this embodiment, the hybrid decision-making mechanism combines the advantages of automatic control and manual decision-making to ensure the efficient and reliable operation of the basement ventilation and smoke extraction system under various conditions. Adaptive collaborative control under normal conditions refers to the system automatically adjusting the ventilation and smoke extraction equipment based on real-time monitoring data, state prediction results, and optimized control strategies, achieving coordinated operation of equipment in various areas without manual intervention. Specifically, the edge layer and cloud layer continuously update the system's global state model through real-time data interaction and analysis. The cloud layer's dual-track decision engine generates globally optimal control commands based on the global state model, combined with a deterministic rule base and an adaptive deep reinforcement learning model. These control commands are transmitted through the edge layer to the actuators in the terminal layer, which adjust the equipment's operating state according to the commands, such as fan speed and valve opening. Simultaneously, the system monitors and evaluates the equipment's operating performance in real time. When a deviation is detected between the actual performance and the expected target, the control strategy is automatically adjusted, forming a closed-loop control. This adaptive collaborative control can adjust in real time according to environmental changes, improving system operating efficiency and energy-saving effects.

[0094] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0095] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A basement ventilation and smoke extraction method based on digital management, applied to a basement ventilation and smoke extraction system based on digital management, characterized in that: The system includes the following functional units: global perception fusion unit, edge intelligent gateway unit, digital twin driving unit, deep reinforcement learning decision unit, dynamic execution verification unit, and adaptive adjustment control unit; The full-domain perception fusion unit is used to collect device status and environmental data in real time from all dimensions; The edge intelligent gateway unit; This is used to receive sensing data by configuring a separate edge computing unit in each region, and to perform multi-source data fusion. The digital twin drive unit is used to solve the three-dimensional CFD flow field in real time, preview the impact of control commands on the flow field, and generate virtual sensor data of the uncovered area for blind spot compensation. The deep reinforcement learning decision unit is used to perform real-time state change analysis; The dynamic execution verification unit is used to convert the strategy into device control signals, compare the predicted values ​​with the actual sensor feedback, and dynamically correct the CFD parameter errors. The adaptive adjustment control unit is used to combine dynamic topology adjustment to optimize the sensor network structure, and the controller based on deep reinforcement learning algorithm performs operating condition prediction and self-learning to perform adaptive adjustment control. The operational steps of a digitally managed basement ventilation and smoke extraction method include: S1: Construct a digital twin architecture including a terminal layer, an edge layer, and a cloud layer to perform real-time mapping and interaction between the physical system and the virtual model of basement ventilation and smoke extraction; S2: At the terminal layer, multiple sensors are used for integrated monitoring to collect data parameters related to ventilation and smoke exhaust in the basement; S3: The collected data is preprocessed using a highly integrated edge computing gateway; S4: The regional controller uses a dynamic Bayesian network to fuse preprocessed multi-source data and construct a local environmental state model. S5: The edge layer uses deep reinforcement learning algorithms and digital twin models to simulate various scenarios for real-time state prediction and generate an emergency plan library; S6: The central server in the cloud layer runs a dual-track decision engine, which combines a deterministic rule base with an adaptive deep reinforcement learning model to perform overall adaptive and collaborative control. S7: Synchronously retain the hybrid decision-making mechanism. Under normal conditions, the system performs adaptive and collaborative control. Under emergency conditions, it utilizes human intervention terminals to transform human experience into system knowledge, forming a human-machine collaborative decision-making mechanism.

2. The basement ventilation and smoke extraction method based on digital management according to claim 1, characterized in that: The real-time mapping and interaction between the physical system and the virtual model for basement ventilation and smoke extraction specifically includes: The terminal layer sensors collect various parameters of the physical system according to the set sampling frequency; The edge layer preprocesses the received data and transmits it to the cloud layer, while simultaneously updating the corresponding parameters in the virtual model in real time based on this data. The virtual model uses the updated parameters to perform simulation calculations, simulates the operation of the physical system under different working conditions, and feeds back the simulation results to the edge layer and cloud layer. The edge layer and cloud layer then formulate control strategies based on the simulation results of the virtual model and actual needs, and send control commands to the actuators in the terminal layer. The actuators adjust the physical system according to the commands, realizing real-time interaction between the physical system and the virtual model.

3. The basement ventilation and smoke extraction method based on digital management according to claim 2, characterized in that: The terminal layer uses a hierarchical sensor network to deploy three layers of sensor nodes, including a top layer, a middle layer, and a bottom layer, to collect parameters in real time.

4. The basement ventilation and smoke extraction method based on digital management according to claim 3, characterized in that: S4 specifically includes: S41: Determine the state variables and observation variables of the local environment; The state variables are key parameters used to describe the local environmental state, and the observed variables are preprocessed data values ​​from each sensor. S42: Construct a dynamic Bayesian network structure that includes the initial network structure and the transition network structure; The initial network structure is used to determine the dependencies between variables through expert knowledge or data learning, and to obtain the probabilistic relationship between the initial state variables and the observed variables; the transition network structure is used to describe the probabilistic relationship between the state variables at different times, and to obtain the evolution law of the environmental state over time. S43: Determine the parameters of the dynamic Bayesian network and obtain the conditional probability distribution among the variables; The parameters of the initial network and the transition network are obtained through statistical analysis of historical data or expert experience estimation. S44: Using preprocessed multi-source data for dynamic Bayesian network inference; S45: Construct a corresponding local environment state model as the basis for the edge layer to perform state prediction and generate emergency plans.

5. The basement ventilation and smoke extraction method based on digital management according to claim 4, characterized in that: The inference process of the dynamic Bayesian network specifically includes: Estimate the posterior probability distribution of the state variables based on the values ​​of the observed variables; The forward algorithm is used to calculate the results at each time step. , given before The posterior probability of the state variable for each observation is used to obtain the forward probability. The backward algorithm is used to calculate at each time step. Given from To the specified period The observed values ​​are the posterior probabilities of the state variables, from which the backward probabilities are obtained; Then, by combining the forward and backward probabilities, we obtain the probability at each time step. The posterior probability distribution of state variables determines the state of the local environment; Finally, based on the posterior probability distribution of the state variables obtained through reasoning, a local environment state model is constructed to reflect the dynamic changes of the local environment in real time.

6. The basement ventilation and smoke extraction method based on digital management according to claim 5, characterized in that: S5 specifically includes: S51: Determine the state space, action space, and reward function; S52: Construct a DQN+LSTM model; the DQN+LSTM model uses an LSTM network and a deep neural network; S53: Calculate the target value using the target network, and select actions and update parameters based on the evaluation network; S54: Use a digital twin model to simulate multiple scenarios, and use the output of the digital twin model as environmental feedback to interact with the DQN+LSTM model; S55: Obtain feedback based on the reward function, gradually optimize the action strategy, and based on the optimized strategy, make real-time predictions and outputs of the system state under different scenarios.

7. The basement ventilation and smoke extraction method based on digital management according to claim 6, characterized in that: Simulating multiple scenarios for real-time state prediction also includes: LSTM network using historical and current state data to output predicted values ​​of future states; DQN selecting the optimal control action based on the predicted state value; and generating an emergency plan library based on multiple simulated scenarios and corresponding optimal control actions. Each emergency plan includes a scenario description, triggering conditions, control measures, and expected effects. When a scenario matching a plan occurs in actual operation, the plan is invoked for processing.

8. The basement ventilation and smoke extraction method based on digital management according to claim 7, characterized in that: The dual-track decision engine includes a deterministic rule base and an adaptive deep reinforcement learning model. The deterministic rule base formulates fixed control rules based on fire protection codes, engineering experience, and expert knowledge. The adaptive deep reinforcement learning model uses a deep reinforcement learning algorithm, taking the environmental state data of the entire basement as input, to optimize the output of the global control strategy.

9. The basement ventilation and smoke extraction method based on digital management according to claim 8, characterized in that: The operation process of the dual-track decision engine includes: S61: The dual-track decision engine simultaneously calls the deterministic rule base and the adaptive deep reinforcement learning model to analyze the global environment state; S62: The deterministic rule base matches the current state according to the preset rules. If there is a rule that matches perfectly, the corresponding control command is generated directly. If there is no rule that matches perfectly, the adaptive deep reinforcement learning model is started. S63: Deep reinforcement learning models calculate the optimal control policy based on the global state, supplementing or adjusting deterministic rules; S64: Integrate and evaluate the control strategies generated by the two decision paths, and select the control strategy with the best overall performance as the final decision; S65: The final control command is sent to each edge layer, which then forwards it to the actuator in the terminal layer to achieve coordinated control of the entire system. S66: Simultaneously feed the decision-making process and execution results back to the deep reinforcement learning model for continuous model updates and optimization.

Images