Knowledge-based underground engineering intelligent ventilation control test system and method

Abstract

The invention discloses a knowledge-based underground engineering intelligent ventilation control test system and method, and the system constructs a knowledge-driven intelligent ventilation architecture, fuses physical adjustment and data knowledge management, and introduces a ventilation hypergraph network optimization regulation strategy through a bidirectional mapping mechanism of entity and knowledge data. The system adopts end-side cloud collaborative deployment, constructs a four-layer architecture of data, calculation, application and management, and realizes multi-source data acquisition, intelligent decision and closed-loop control. In combination with a deep reinforcement learning algorithm and various wind solving models, the system can automatically optimize ventilation parameters under a dynamic working condition, and on-demand air supply and energy consumption minimization are realized. The system is composed of a power subsystem, a pipeline subsystem, a physical model subsystem, a monitoring subsystem, a regulation subsystem and the like and an intelligent management and control platform, has high expansibility and high practicability, and is suitable for design, debugging and control research of an underground engineering ventilation system.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent ventilation control of underground engineering, and particularly relates to a knowledge-based intelligent ventilation control test system and method for underground engineering. BACKGROUND

[0002] Underground engineering includes exploration and development of mineral energy in the deep part of the earth, safe use of various types of underground space, and engineering facilities for disaster prevention and mitigation. It is gradually developing towards the mode of "unified planning, multi-functional integration, and large-scale construction". At the same time, the construction and maintenance of the project pay more attention to low energy consumption, compact space, high comfort and safety, and have characteristics such as multi-level, large-scale, green and low carbon, and intelligent construction. It faces challenges such as precise supply of ventilation, remote control and knowledge optimization, and urgently needs overall research on the test method and system of intelligent ventilation to provide a basis for the intelligent practice of ventilation engineering.

[0003] On-demand ventilation remote control can better meet the needs of human health and production process for air volume, while reducing ventilation cost and improving ventilation efficiency. It has good adaptability to variable working condition ventilation environment protection scenarios. At the same time, the ventilation system also needs to have the function of dynamically adjusting the operating parameters of the ventilation system according to the state and position information of equipment and personnel, such as dilution ventilation air volume control with pollutant concentration and fire smoke concentration deviation as signal input, and fan power pressure regulation with pressure deviation as signal input.

[0004] At present, with the continuous iteration of ventilation theory, methods and technology, and the development of current communication, digital, intelligent construction and smart operation and maintenance related disciplines, the intelligentization of underground engineering ventilation has made some progress, gradually changing the ventilation management mode relying on experience design and on-site manual control. For example, for traffic tunnel ventilation engineering, the ventilation control mode has gradually developed from direct control and indirect control to "active" ventilation control mode; for dust removal in cavern group construction, air environment monitoring and early warning, ventilation energy consumption optimization and local ventilation automatic variable frequency regulation technology research have been carried out; for mine ventilation safety, remote start-stop control of main ventilation power devices and local ventilation facilities. In summary, each engineering scene meets the specific ventilation protection needs in the field, and certain achievements have been made in ventilation equipment remote control technology and methods, but there is a lack of overall research on the intelligentization of ventilation systems based on knowledge, especially the intelligent ventilation optimization control method under different goals of multiple scenes, and the comprehensive test system platform of "perception-analysis-control-optimization" closed loop are still blank. Therefore, establishing an intelligent control system for simulating the ventilation environment of engineering site not only can carry out environmental monitoring design, intelligent ventilation parameter optimization, control model and other theoretical and application research under the demand of multi-working condition protection, but also can determine and test the technical parameters of the control execution equipment used on site, which is of great significance. SUMMARY

[0005] The technical problem to be solved by the present application is to provide an intelligent ventilation control test system and method based on knowledge for underground engineering, which can improve the intelligent level of ventilation control, realize on-demand ventilation and energy-saving operation, has strong expandability and high practicability, and is suitable for the design, debugging and control research of underground engineering ventilation systems.

[0006] The technical solution of the present application is: An intelligent ventilation control test system based on knowledge for underground engineering, the system comprises a knowledge-driven intelligent ventilation architecture, which fuses physical regulation and data knowledge management through a bidirectional mapping mechanism between ventilation system entities and knowledge data; the bidirectional mapping mechanism performs real-time information interaction in the process of ventilation regulation decision-making on one hand, including cyclic iterative optimization of data collection, transmission, analysis and feedback, and finally feeds back the optimized scheme to the ventilation equipment to provide regulation and control decision-making services; on the other hand, a ventilation hypergraph network reflecting the structure of the ventilation system and the air regulation process is established, and the ventilation system performs entity regulation and control according to the optimized regulation scheme of the hypergraph, so that the actual ventilation system and the generated data knowledge are consistent, and the data knowledge is checked and supplemented; the system further comprises four levels of data layer, calculation layer, application layer and management layer, wherein the data layer is responsible for data perception, integration and management; the calculation layer realizes the conversion of data to knowledge relying on cloud computing; the application layer constructs intelligent ventilation scene modules; the management layer provides user interaction and management services; the system adopts an end-edge-cloud deployment mode to realize the allocation and coordination of calculation and storage, and forms a closed-loop operation system; the hardware system of the system is divided into a power system, a pipeline system, a physical model system, a monitoring system and a regulation and control system, and is configured in the air sampling area, the physical model wind area, the primary air control area and the secondary air control area; the system further comprises an intelligent management and control software platform, including data management, model management and knowledge engineering modules.

[0007] Further, the bidirectional mapping mechanism integrates mechanism knowledge and experience knowledge in the ventilation regulation process, wherein the mechanism knowledge is expressed in mathematical form, including formulas or equations, and is used for the explicit characteristics of ventilation control; the experience knowledge represents the association between control actions and process environment, and is used for the implicit characteristics of ventilation control, and is embodied in the form of regulation data; the regulation and control decision-making service system adopts a deep reinforcement learning algorithm to mine data, provide decisions and instructions for the intelligent regulation and control device of air flow, realize iterative optimization of ventilation parameters, including a monitoring part for data collection and feature extraction processing, and a control part for receiving state information and mapping it into regulation actions to meet the air volume demand under different working conditions.

[0008] Further, the data layer adopts digital sensors to collect air environment, ventilation fan power, regulation facilities and branch network data, stores and manages them through relational databases, time series databases and graph databases, supports large-scale data retrieval and analysis; the computing layer includes data centers, information models and knowledge-based applications, wherein the data centers include data rooms, computing and storage, databases and disaster recovery backups, the information models have data exchange, ventilation scene management, three-dimensional spatial analysis and visualization expression functions through the establishment of Internet of Things and big data platforms, the knowledge-based applications realize knowledge representation, extraction, mining and reasoning based on the knowledge graph search engine; the application layer includes monitoring data, network calculation, dynamic feedback control, knowledge Q&A, data association, platform interface, early warning and permission authentication modules; the management layer provides monitoring data visualization, wind network dynamic calculation, intelligent decision support, remote monitoring, equipment regulation, abnormal diagnosis, disaster warning, authorized login and message processing services, displays data through three-dimensional modeling and parameter visualization, and provides system energy optimization, ventilation environment warning and ventilation equipment fault information.

[0009] Further, in the end-edge-cloud deployment mode, the end includes field sensor terminals and variable frequency fan regulation and damper regulation terminals, the integrated air environment monitoring frame is installed in the physical model air area, the air volume and air pressure sensors are installed inside the air pipe, the fan control cabinet and damper regulation device are installed near the variable frequency fan and adjustable damper; the edge is an edge computing server, which is deployed on the experimental field industrial computer, directly connected with the equipment and facilities of the end for local data processing and decision-making, and communicates with the cloud; the cloud is a cloud computing center, which performs big data analysis, storage and model training, and issues models and strategies; the field sensor terminal synchronously collects data through the wind speed sensor, electrochemical CO sensor, electrochemical CO2 sensor, electrochemical oxygen sensor and laser scattering dust sensor, converts the data into RS485 signal transmission of ModbusRTU protocol through the industrial signal conditioning module, controls the delay within 500 ms through time synchronization control technology, and controls the collection frequency to be 7.568 s; the cloud computing center receives the environment and ventilation parameters collected by the field sensor terminal, optimizes the ventilation parameters using the data model, and adjusts the working state of the variable frequency fan and the ventilation facility in real time according to the calculation result; the data model includes natural wind distribution, control type wind distribution and hybrid optimization, and contains network design, ventilation design and optimization analysis modules.

[0010] Further, the air sampling area is the area where the inlet and outlet of the model ventilation system are located, the power system is located in the primary air control area, including the forward and reverse frequency conversion fan unit and the intelligent fan control cabinet, supporting manual and automatic modes, as well as remote control, the intelligent fan control cabinet is arranged near the forward and reverse frequency conversion fan unit; the pipeline system is located in the primary air control area, and is formed by a ventilation network in series, parallel and angular connection form through soft wind belts with different diameters, equipped with air volume monitors and automatic adjusting air valves, the air volume monitors and automatic adjusting air valves are arranged at the fan outlet or the constant air volume air path straight pipe section; the secondary air control area includes the main lower branch area of the ventilation network, and is classified according to the pipeline architecture, equipped with corresponding intelligent electrically operated double-leaf multi-leaf regulating valves, intelligent air speed monitors and intelligent differential pressure monitors, through the combined opening degree adjustment of the electrically operated double-leaf multi-leaf regulating valves, the on-demand air volume distribution and pressure gradient adjustment of each ventilation network branch are realized; the physical model system is a scaled-down model of the underground engineering cavern group and the ventilation system, the model design follows the Navier-Stokes equation, Euler criterion and Reynolds criterion, the length scale ratio ranges from 1:30 to 1:150, and the model is made of modular assembly and acrylic plate materials; the monitoring system is located in the physical model air area, including an integrated air environment monitoring frame, intelligent smoke detectors, intelligent electrically operated louver air outlets, intelligent electrically operated double-leaf multi-leaf regulating valves, intelligent air speed monitors and intelligent differential pressure monitors, the monitoring system communicates with the control service by using the Modbus protocol, obtains the monitoring values of the on-site air environment monitoring, air volume and pressure monitoring, power equipment and air regulating facilities, the monitoring system monitors the deviation between the air environment in the air area and the target environment, and the deviation between the ventilation parameters and the predicted parameters, at the same time, receives the adjustment instructions issued by the cloud, and performs end control through the intelligent electrically operated louver air outlets and the intelligent electrically operated double-leaf multi-leaf regulating valves; the control system includes electrically operated regulating valves and intelligent detectors, the electrically operated regulating valves and the intelligent detectors communicate through the Modbus protocol, realizing opening degree adjustment and end control.

[0011] Further, the intelligent fan control cabinet is an independent box body, comprising: an upper box body, a lower box body, a control cabinet base, an upper cabinet door and a lower cabinet door, the upper box body and the lower box body are fixedly connected to form a main frame of the independent box body, the lower box body is fixedly connected with the control cabinet base, the upper cabinet door is connected with the upper box body, the lower cabinet door is connected with the lower box body, the control cabinet base is provided with a wire inlet hole for the entry of external cables, and forms a channel connection with the internal circuit of the box body; the upper box body comprises: a man-machine interaction display screen, a power module A, a communication module A, a wiring slot, a core control data processing unit, a wiring terminal block A and a battery pack monitoring and charging and discharging control module, the man-machine interaction display screen is fixed on the upper cabinet door through a display screen inlay door frame, the man-machine interaction display screen is connected with the core control data processing unit, the power module A is electrically connected with an external AC 220V power supply through the wiring terminal block A and the wire inlet hole, the power module A is electrically connected to the man-machine interaction display screen, the communication module A, the core control data processing unit and the battery pack monitoring and charging and discharging control module, the communication module A is connected with the core control data processing unit, the communication module A is connected with an external server through the wiring slot and the wiring terminal block A, the wiring slot is connected with the power module A, the communication module A, the core control data processing unit and the wiring terminal block A, the core control data processing unit is connected with the man-machine interaction display screen and the communication module A and is electrically connected with the wiring terminal block A, the wiring terminal block A comprises a leakage protection air switch and is electrically connected with the power module A, the core control data processing unit and the external cable; a battery is arranged in the lower box body, and the battery is electrically connected with the battery pack monitoring and charging and discharging control module; a transparent protection plate is arranged on the upper cabinet door, the transparent protection plate is fixed on the upper cabinet door and covers the man-machine interaction display screen, a transparent protection plate is also arranged on the lower cabinet door, and the transparent protection plate is fixed on the lower cabinet door; the intelligent fan control cabinet is connected with external equipment including a fan power distribution box frequency converter and an external server through the wire inlet hole and the wiring terminal block A.

[0012] Further, the integrated air environment monitoring frame comprises an integrated frame, sensors and an environment monitoring collection box, the integrated frame is a grid frame structure, the environment monitoring collection box and the sensors are fixed on the integrated frame, the sensors are fixed at the grid positions of the integrated frame to form an array arrangement; the sensors comprise a sensor display screen and a sensor probe, the sensor display screen is connected with the sensor probe; the environment monitoring collection box comprises a display screen, a leakage protection air switch, a power module B, a communication module B, a wiring terminal block B and a data acquisition control module, the display screen is connected with the data acquisition control module and is fixed on the box door of the environment monitoring collection box, the leakage protection air switch is electrically connected with the power module B and the wiring terminal block B, the power module B is electrically connected with an external AC 220V power supply through the wiring terminal block B, the power module B supplies power for the sensors, the display screen, the communication module B and the data acquisition control module, the communication module B is connected with the data acquisition control module and uploads the collected data to an external server, the wiring terminal block B is electrically connected with the data acquisition control module, the power module B and the leakage protection air switch, and the data acquisition control module is connected with the display screen and the wiring terminal block B; the data collected by the sensor probe is transmitted to the data acquisition control module for processing through the wiring terminal block B.

[0013] Further, the data management module of the intelligent management and control software platform provides computer end and mobile end display, the computer end provides model introduction, monitoring point information, monitoring index data curve change, early warning and prediction, monitoring data historical query and export function, the time period that needs to be inquired can be selected, and the original data, hourly average and daily average multiple data types are exported, the mobile end provides remote monitoring data query function, including environmental monitoring index, preset frequency, running frequency, current, local-remote state and fan running index such as power consumption, abnormal data is highlighted in red and the corresponding threshold is given;The model management module is divided into natural wind distribution, control type wind solution and hybrid type optimization model, including network design, ventilation design and optimization analysis module, the network design module completes the establishment process of the ventilation network structure through the extraction of the ventilation unit, the spatial correlation, and the definition of the number and number of the interface unit, the name and description, the coordinate and the directional connection parameter between units, the ventilation design module completes the process of assigning and calculating the characteristic parameters of each unit based on the graph structure, including unit physical structure parameters, ventilation characteristic parameters and constraint parameters, in the basic equipment performance data configuration interface, the performance curve parameters or the input corrected performance curve are configured, the input of the fan, the air duct and the special-shaped air duct structure performance data is provided, the optimization analysis module has the calculation functions of natural wind distribution, control type wind solution and hybrid type regulation air supply, and supports visualization, analysis and export of the ventilation network solution result, outputs node flow, branch flow and flow direction, branch impedance, fan total pressure demand and other parameter display, natural wind distribution is the minimum total resistance under the premise of given total air intake or total air exhaust;Control type regulation is the stability regulation branch combination optimization based on the sensitivity index under the premise of specified demand air supply branch requirement;Hybrid type regulation air supply is the fan valve combination optimization of the lowest fan operating power under the premise of specified demand air supply branch requirement;The knowledge engineering module constructs a knowledge graph through a physical layer, a data layer and an ontology layer, realizes knowledge storage, visualization, auxiliary design, knowledge retrieval, question and answer and report generation.

[0014] A knowledge-based underground engineering intelligent ventilation control method applied to the system, comprising the following steps: constructing a knowledge-driven intelligent ventilation architecture to realize bidirectional mapping and cyclic iteration optimization of entities and knowledge data;Carrying out data acquisition, processing and model training through end-edge-cloud deployment;Dividing hardware regions, including setting air intake and exhaust outlets in the air sampling area, configuring power and pipeline systems in the first-level wind control area, realizing wind valve joint debugging in the second-level wind control area, and monitoring deviations and executing regulation and control in the physical model area;Using a deep reinforcement learning algorithm for decision-making, including initializing state space and action space, collecting data as states, selecting actions, executing and observing rewards, and updating value functions;Applying optimization models, including natural wind distribution, control type wind solution and hybrid type regulation air supply.

[0015] The present application has the following advantages: 1. Achieve intelligent and closed-loop optimization management of ventilation control: The system has an intelligent closed-loop structure of perception-analysis-control-optimization. Through the collaborative operation of data acquisition, model reasoning and reinforcement learning algorithms, it can dynamically identify the operating status of the ventilation system and iteratively optimize the adjustment strategy to achieve efficient, stable and energy-saving operation of the ventilation system.

[0016] 2. Possesses a zoned and layered controllable structure: This invention constructs a layered closed-loop area, including the air collection area, the physical model air use area, the primary air control area, and the secondary air control area. It can realize the linkage control of variable frequency fans and multi-blade electric air valves, support the joint adjustment of air volume and pressure difference management between different ventilation levels, and improve the accuracy of regulation and system response speed.

[0017] 3. Highly reproduces on-site ventilation characteristics and has engineering simulation capabilities: The physical model system adopts similarity theory (such as Euler criterion and Reynolds criterion) for scaled-down design, which can accurately reproduce the flow field characteristics of complex ventilation environments in underground engineering, and provide high-fidelity physical basis for experimental verification and optimization simulation of ventilation systems.

[0018] 4. Support for multi-mode testing and research on complex ventilation conditions: This invention has the ability to simulate wind fields in large-span spaces and solve ventilation networks. It can study the migration mechanism of pollutants and the distribution law of airflow, and can also carry out ventilation strategy evaluation and intelligent optimization for air volume requirements under different operating conditions, providing experimental support for ventilation design and management in complex environments such as tunnel groups and cavern groups.

[0019] 5. Supports performance testing and durability assessment of ventilation system equipment: Through parameter calibration and long-term testing, the performance of core equipment such as fans, dampers, and sensors can be tested and the control accuracy can be assessed, verifying their response characteristics and stability under different operating conditions, and providing engineering adaptability verification functions.

[0020] 6. Integrated hardware and software design enhances platform integration capabilities and ease of operation: The integrated air environment monitoring rack and intelligent fan control cabinet integrate data acquisition, control execution and remote communication functions. Combined with wireless communication and cloud deployment, it can realize end-edge-cloud data linkage and remote control, with high integration and high automation level, and is suitable for training, research and system integration testing.

[0021] 7. Support knowledge graph construction and intelligent question answering services to enhance knowledge management and decision support capabilities: The system constructs a knowledge graph in the ventilation field through ontology modeling and knowledge extraction, realizing the fusion expression and reasoning support from data to knowledge, and can provide traceable knowledge support and expert assistance for the formulation of control strategies. Attached Figure Description

[0022] Figure 1This is a functional zoning diagram of a knowledge-based intelligent ventilation control test system for underground engineering, according to the present invention.

[0023] Figure 2 This is a schematic diagram of the system hierarchy of a knowledge-based intelligent ventilation control test system for underground engineering according to the present invention.

[0024] Figure 3 This is a schematic diagram of the edge-cloud system architecture of a knowledge-based intelligent ventilation control test system for underground engineering according to the present invention.

[0025] Figure 4 This is a schematic diagram of the external structure of the intelligent fan control cabinet of a knowledge-based intelligent ventilation control test system for underground engineering, according to the present invention.

[0026] Figure 5 This is a schematic diagram of the internal structure of the intelligent fan control cabinet of a knowledge-based intelligent ventilation control test system for underground engineering, according to the present invention.

[0027] Figure 6 This is a schematic diagram of the structure of an integrated air environment monitoring frame for a knowledge-based intelligent ventilation control test system for underground engineering, according to the present invention.

[0028] Figure 7 This is a schematic diagram of the sensor structure of an integrated air environment monitoring frame for a knowledge-based intelligent ventilation control test system for underground engineering, according to the present invention.

[0029] Figure 8 This is a schematic diagram of the environmental monitoring and data acquisition box of an integrated air environment monitoring frame for a knowledge-based intelligent ventilation control test system for underground engineering, according to the present invention.

[0030] Figure 9 This is a schematic diagram of the computer interface of the intelligent management and control software platform of the knowledge-based intelligent ventilation control test system for underground engineering according to the present invention.

[0031] Figure 10 This is a schematic diagram of the mobile terminal operation interface of the intelligent management and control software platform of the knowledge-based intelligent ventilation control test system for underground engineering according to the present invention.

[0032] In the diagram: 1-Integrated rack; 3-Sensor; 5-Environmental monitoring and acquisition box; 7-Independent cabinet; 301-Sensor display screen; 303-Sensor probe; 501-Display screen; 503-Leakage protection air switch; 505-Power module B; 507-Communication module B; 509-Terminal block B; 511-Data acquisition and control module; 701-Transparent protection board; 703-Upper cabinet door; 705-Lower cabinet door; 707-Cable inlet; 709-Control cabinet base; 711-Battery; 713-Human machine interaction display screen; 715-Display screen embedded in door frame; 717-Power module A; 719-Communication module A; 721-Wiring trough; 723-Core control data processing unit; 725-Terminal block A; 727-Battery pack monitoring and charging / discharging control module. Detailed Implementation

[0033] It should be noted that, in the description of this invention, unless otherwise specified and limited, the terms "installation", "connection", "linking", "interconnection", etc., should be interpreted broadly, that is, they can be fixed connections or detachable connections; they can be mechanical connections or electrical connections; they can be direct connections or indirect connections through an intermediate medium.

[0034] like Figures 1-10 As shown, this invention proposes a knowledge-based intelligent ventilation control test system and method for underground engineering. The following describes in detail the system architecture, deep reinforcement learning control algorithm, and the specific implementation methods of three types of optimization models (natural air distribution, controlled air distribution, and hybrid regulated air supply) through multiple embodiments.

[0035] Example 1: The intelligent ventilation system in this example adopts a knowledge-driven architecture design. Through a two-way mapping mechanism between ventilation system entities and knowledge data, it deeply integrates the physical ventilation regulation process with data knowledge management to achieve intelligent control of ventilation in underground engineering projects. This two-way mapping mechanism includes two levels of meaning: Real-time interactive control: Real-time interactive information circulation is achieved in the ventilation regulation decision-making process. On the one hand, the system continuously collects environmental and equipment data, obtains ventilation parameters (such as wind speed, air volume, ambient gas and dust concentration, etc.) through sensors, and transmits them to the data center for analysis and processing via the communication network. On the other hand, based on the analysis results, optimized ventilation control schemes (such as adjusting fan frequency, valve opening, etc.) are generated, and the instructions are fed back to the ventilation equipment in real time through the control network for execution. This forms a cyclical iterative optimization of data collection, transmission, analysis, and feedback, and finally feeds the optimized scheme back to the ventilation equipment to provide refined ventilation regulation decision-making services, realizing closed-loop control.

[0036] Entity-Knowledge Consistency: The system establishes a ventilation hypergraph network model that can characterize the structure and air regulation process of the ventilation system. It abstracts the various elements of the ventilation system (tunnels, nodes, fans, air outlets / valves, etc.) and their relationships into a hypergraph structure. In this hypergraph, nodes represent the intersections of roadways in the ventilation network, and hyperedges can represent the connections between multiple air paths. Through full-element hypergraph modeling, the system can perform global optimization calculations on the ventilation network to form the optimal regulation scheme. Subsequently, this scheme is deployed to the ventilation system entities (such as regulating air valves, variable frequency fan operating frequency parameters, etc.) to ensure that the actual operating state of the ventilation system is consistent with the predictions of the knowledge model. The feedback data after execution is used to verify and enrich the knowledge base, gradually improving the hypergraph model and control strategy. In this way, the entity ventilation process and data knowledge form a two-way mapping and mutual calibration: entity operation generates data knowledge, and knowledge deduction guides entity regulation.

[0037] Knowledge fusion mechanism: In the above-mentioned two-way mapping intelligent ventilation control process, both mechanistic knowledge and empirical knowledge are incorporated to improve decision-making intelligence. Mechanism knowledge refers to the physical mechanisms and laws governing ventilation, explicitly expressed in mathematical formulas or equations, reflecting the explicit characteristics of ventilation control. Examples include pressure loss formulas, drag laws, and fan performance curves, all of which constitute mechanism knowledge and are used to constrain and guide control decisions. Mechanism knowledge ensures that control strategies conform to the physical laws of the ventilation system. For instance, using a scaled-down model of the Navier-Stokes equations and the flow continuity equation ensures that simulated airflow matches reality, or employing the wind resistance-squared relationship. (in, This indicates the pressure difference between the two ends of the ventilation branch. This represents the resistance coefficient of the branch, which is related to the tunnel cross-section, length, and surface roughness, among other factors. The regulation effect is predicted by the airflow through this branch; the mechanism formula provides explicit constraints for ventilation control, such as calculating and estimating the change in airflow after a damper is closed to a certain extent, providing a basis for algorithm decision-making.

[0038] Empirical knowledge refers to the control experience and patterns accumulated over long-term operation. It reflects the implicit characteristics between control actions and environmental responses and is stored in a knowledge base in the form of data. For example, based on historical data of CO2 concentration, wind speed, and corresponding fan operating frequency and valve opening values ​​at various measuring points under certain operating conditions, feature engineering can be constructed to analyze the relationship between input feature values ​​and target values. This empirical knowledge exists in the form of control data pairs or rules. After training with a deep neural network model, the implicit relationship between ventilation system regulation (fan frequency, valve opening) and air environment monitoring values ​​can be extracted from historical data. Empirical knowledge provides implicit guidance for ventilation control and can be used as supplementary feature inputs or a reference for the rule base when making algorithmic decisions. For example, during the regulation process, the valve opening is constrained to be no less than 70% and the fan operating pressure is less than 90% of the maximum pressure in the ventilation route.

[0039] In summary, the knowledge-driven intelligent ventilation architecture integrates the aforementioned mechanisms and empirical knowledge to achieve dual-driven data and knowledge: it leverages the rigor of physical models and algorithms while also drawing on expert experience and historical data patterns, making decision-making more accurate and efficient.

[0040] Layered Design: To achieve the above-mentioned architectural functions, this embodiment divides the system functions into four layers, forming a layered architecture of data layer, computing layer, application layer, and management layer, with each layer working collaboratively. Data Layer: Responsible for data sensing, acquisition, integration, and management. This layer acquires various environmental and equipment data related to ventilation through deployed digital sensors, including air environment parameters (wind speed, wind pressure, temperature, humidity, concentration of harmful gases such as CO / CO2 / oxygen content, dust concentration, etc.), ventilation power equipment parameters (fan speed, current, voltage, etc.), status of regulating facilities (valve opening), and ventilation network structure data. All sensor-collected data is aggregated through an industrial communication network and stored in multiple types of databases for unified management. For example, real-time high-frequency data (such as time-series data like wind speed and wind pressure) is stored in a time-series database, relational data (such as equipment ledgers and sensor information) is stored in a relational database, and complex topological relationships and knowledge graphs are stored in a graph database. The data layer provides large-scale data retrieval and analysis capabilities, providing a reliable data foundation for upper-layer intelligent algorithms.

[0041] The computing layer, relying on the cloud computing platform, realizes the transformation from data to knowledge and complex computing tasks. The computing layer consists of three parts: data center, information model, and knowledge-based applications. Data center: Deployed in the cloud or central computer room, it contains high-performance computing and storage resources, database clusters and disaster recovery backup facilities. It is the data processing core of the entire system and supports the storage, cleaning and analysis of massive amounts of data.

[0042] Information Model: Build an Internet of Things and big data platform, connect to the data center, and realize the information modeling and simulation functions of the ventilation system. The platform has functions such as data exchange (receiving data from the field edge and sending control commands), ventilation scene management (modeling the ventilation network, defining the air path topology and parameters), and 3D spatial analysis and visualization (presenting the ventilation system and underground space as a 3D model, allowing users to intuitively view the airflow path, sensor distribution, etc.).

[0043] Knowledge-based applications: Intelligent knowledge services are built based on knowledge graph technology, integrating structured knowledge (such as equipment information, regulations and standards) and unstructured knowledge (documents, papers) in the ventilation field to form a ventilation knowledge graph. On this basis, knowledge representation, extraction, mining and reasoning functions are realized. For example, through knowledge question answering engines (such as IBM Watson Discovery+ Knowledge Studio, Alibaba Industrial Brain, etc.), users can query "the percentage of time when the ventilation environment monitoring value meets the target value in a certain period", "the ventilation system adjustment design scheme under changing operating conditions", "the possible causes of a fan failure", and the system uses knowledge graph reasoning to provide answers. It can also generate explanations or reports of control schemes based on model calculation results and knowledge base information.

[0044] Application Layer: This layer is designed for specific business functions, building various intelligent ventilation application scenario modules to directly provide services for ventilation management and control. It mainly includes the following functional modules: Monitoring data module: Receives and displays environmental data and equipment operation data from each monitoring point in real time, such as wind speed and air volume curves, gas concentration changes, etc.

[0045] Network solution module: Performs calculations and solutions for the ventilation network, including calculations of air volume distribution and air pressure distribution. Based on the collected real-time data and network model, it can calculate the current air volume, resistance, pressure drop, etc. of each roadway branch. It also supports users to set different damper / valve states to simulate and calculate new air volume distribution results for analysis of the effect of adjustment measures.

[0046] Dynamic Feedback Control Module: This module works in conjunction with the monitoring module to achieve closed-loop control of key ventilation parameters. Based on real-time deviations in sensor data (such as actual airflow being lower than the target value), the dynamic feedback control module generates adjustment commands (such as opening a branch air valve) to correct the deviation and continuously monitors the effect to achieve dynamic airflow balance.

[0047] Knowledge Q&A Module: Provides knowledge-based query and expert guidance functions. Users can ask questions about ventilation systems in natural language (such as the basis of control schemes, the meaning of parameters, etc.), and the system will retrieve relevant knowledge from the knowledge graph and provide answers.

[0048] Data association module: This module links monitoring data, control records, design drawings, historical reports, etc., making it convenient for managers to view the corresponding data when ventilation is abnormal during a certain period, thus forming a comprehensive analysis of data and documents.

[0049] Platform interface module: Provides interfaces for interaction with other systems, such as integration with underground engineering integrated automation platforms or emergency response systems, receiving external commands or outputting ventilation data.

[0050] Early warning and forecast module: Based on thresholds and model predictions, it realizes ventilation safety early warning. For example, when the data of a certain sensor exceeds the threshold or the model predicts that harmful gas may accumulate in the future, it will promptly issue an alarm and trend forecast to the management.

[0051] Access control module: Ensures system security by setting viewing and control permissions for different user roles (operators, engineers, administrators) and login authentication.

[0052] Management layer: Provides the human-computer interaction interface and advanced management services. It serves as the window for users to interact with the system. Through a web-based monitoring center and mobile applications, the management layer presents the system's operating status and provides control functions to users. On the computer, the model and system can be comprehensively monitored and managed visually, including: system overview and model introduction, status of each monitoring point and sensor, real-time curves and historical trends of monitoring indicator data, abnormal alarm information and early warning forecast results; users can select time periods to query historical data and export it in multiple data granularities (such as raw data, hourly average, daily average) for offline analysis; the management terminal also provides an intelligent decision support interface to display the optimized ventilation plan (such as the air volume to be allocated to each branch, the valve and fan settings that need to be adjusted, etc.) for managers to review or execute with one click.

[0053] On the mobile app, lightweight remote monitoring and control functions are provided, allowing users to view key ventilation environment indicators (such as gas, temperature and humidity) and fan operating parameters (such as preset frequency, actual frequency, current, local / remote control status, power consumption, etc.) at any time. Abnormal data will be highlighted in red and the threshold will be noted, making it easier for maintenance personnel to identify problems in a timely manner. In addition, the mobile app supports receiving alarm notifications pushed by the system and issuing simple control commands remotely (subject to security authentication).

[0054] Through the above four-layer architecture, the system realizes a closed loop of the entire process from bottom-level data collection, middle-level analysis and decision-making, upper-level application execution to management interaction.

[0055] This embodiment adopts a three-layer deployment of "end-edge-cloud" to make full use of on-site real-time performance and cloud computing power: Terminals: These refer to various sensor terminals and actuators on-site that directly interact with the physical environment. They mainly include environmental sensor terminals and variable frequency fan / valve actuators. In the experimental scenario, an integrated air environment monitoring rack is installed in the wind zone of the physical model to centrally collect air quality and airflow parameters for that area. Several wind speed and wind pressure sensors are installed at key locations in the ductwork to monitor airflow. Electrochemical sensors are used to monitor the concentration of harmful gases (CO, CO2, oxygen, etc.). Laser scattering dust sensors monitor dust concentration. These front-end sensors convert analog signals to RS485 (Modbus) signals through industrial-grade signal conditioning modules (such as RedLion signal conditioners, Dewesoft universal thermocouple signal conditioners, Advantech signal conditioning modules, etc.). RTU digital signals are used to achieve stable communication. Similarly, at the execution end, the intelligent fan control cabinet is connected to the main fan frequency converter, and the electric air valve adjustment device is installed on the adjustable air valve. These actuators also receive control commands through the Modbus network. The data acquisition frequency and communication delay of the end-side equipment have been optimized. For example, time synchronization control technology (such as NTP protocol) is used to control the transmission delay to within 500ms, and the sensor data acquisition frequency is set to once every 7 seconds, thereby ensuring the synchronization and real-time performance of environmental data.

[0056] Preferably, the intelligent fan control cabinet is an independent enclosure 7, including: an upper enclosure, a lower enclosure, a control cabinet base 709, an upper cabinet door 703, and a lower cabinet door 705. The upper and lower enclosures are fixedly connected to form the main frame of the independent enclosure 7. The bottom of the lower enclosure is fixedly connected to the control cabinet base 709. The upper cabinet door 703 is connected to the upper enclosure, and the lower cabinet door 705 is connected to the lower enclosure. The control cabinet base 709 is provided with a cable inlet hole 707 for the entry of external cables, forming a channel connection with the internal circuitry of the enclosure. The upper enclosure includes: a human-machine interface display screen 713, a power module A717, a communication module A719, and a wiring... The system includes a wiring trough 721, a core control data processing unit 723, a terminal block A725, and a battery pack monitoring and charging / discharging control module 727. The human-machine interface display screen 713 is fixed to the upper cabinet door 703 via a display screen inlay door frame 715. The human-machine interface display screen 713 is connected to the core control data processing unit 723. The power module A717 is electrically connected to an external AC220V power supply via the terminal block A725 and the inlet hole 707. The power module A717 is electrically connected to the human-machine interface display screen 713, the communication module A719, the core control data processing unit 723, and the battery pack monitoring and charging / discharging control module 727. The communication module A719 is connected to the core control data processing unit 723. The communication module A719 is connected to an external server via a wiring duct 721 and a terminal block A725. The wiring duct 721 is connected to the power module A717, the communication module A719, the core control data processing unit 723, and the terminal block A725. The core control data processing unit 723 is connected to the human-machine interface display screen 713 and the communication module A719, and is electrically connected to the terminal block A725. The terminal block A725 includes a leakage current protection circuit breaker and is connected to the power module A717 and the core control data processing unit 723. The data processing unit 723 is electrically connected to external cables; a battery 711 is installed inside the lower cabinet, and the battery 711 is electrically connected to the battery pack monitoring and charging / discharging control module 727; a transparent protective plate 701 is installed on the upper cabinet door 703, and the transparent protective plate 701 is fixed on the upper cabinet door 703 and covers the human-machine interaction display screen 713; a transparent protective plate 701 is also installed on the lower cabinet door 705, and the transparent protective plate 701 is fixed on the lower cabinet door 705; the intelligent fan control cabinet is connected to external equipment, including the fan distribution box inverter and external server, through the inlet hole 707 and the terminal block A725.

[0057] Preferably, the integrated air environment monitoring frame includes: an integrated frame 1, sensors 3, and an environmental monitoring and acquisition box 5. The integrated frame 1 is a grid frame structure. The environmental monitoring and acquisition box 5 and several sensors 3 are fixed on the integrated frame 1. The sensors 3 are distributed and fixed at the grid positions of the integrated frame 1, forming an array arrangement. Each sensor 3 includes: a sensor display screen 301 and a sensor probe 303. The sensor display screen 301 is connected to the sensor probe 303. The environmental monitoring and acquisition box 5 includes: a display screen 501, a leakage current protection circuit breaker 503, a power supply module B505, a communication module B507, a terminal block B509, and a data acquisition and control module 511. The display screen 501 is connected to the data acquisition and control module 511 and fixed on the door of the environmental monitoring and acquisition box 5. The air switch 503 is electrically connected to the power module B505 and the terminal block B509. The power module B505 is electrically connected to an external AC220V power supply through the terminal block B509. The power module B505 supplies power to the sensor 3, the display screen 501, the communication module B507, and the data acquisition and control module 511. The communication module B507 is connected to the data acquisition and control module 511 and uploads the collected data to an external server. The terminal block B509 is electrically connected to the data acquisition and control module 511, the power module B505, and the leakage protection air switch 503. The data acquisition and control module 511 is connected to the display screen 501 and the terminal block B509. The data collected by the sensor probe 303 is transmitted to the data acquisition and control module 511 for processing via the terminal block B509.

[0058] Edge: Deployed at the experimental site, the edge computing server or industrial control computer directly connects to a large number of end-side devices, responsible for local data preliminary processing, caching, and rapid response decision-making. Real-time control applications, such as PLC programs or embedded AI models, run on the edge server, enabling them to make decisions on end-side data at the millisecond level, such as locally controlling the start and stop of fans or adjusting the opening of air valves. In the event of network interruption or high cloud latency, edge computing ensures that basic safety control logic is executed locally. At the same time, the edge layer packages and uploads the aggregated data to the cloud and accepts advanced policy adjustment instructions issued by the cloud, realizing communication and linkage with the cloud.

[0059] The cloud refers to a remote cloud computing center or data center with powerful storage, computing, and model training capabilities. The cloud is responsible for global and in-depth analysis, such as mining historical big data, training deep reinforcement learning models, and performing complex ventilation network optimization calculations. The cloud center receives environmental and ventilation parameter data uploaded by field sensors, executes advanced algorithms (such as global wind network calculation, optimization solution, and AI decision-making), and distributes the results (such as updated control strategies and model parameters) to the edge and end devices. The cloud can continuously optimize the control model (including deep learning model training) based on long-term data and adjust the working mode of field fans and valves in real time when the strategy matures or changes in environmental conditions are detected. The three-layer division of labor between the end, edge, and cloud is clear: the end side is responsible for fine perception and execution, the edge side is responsible for real-time control and coordination, and the cloud side is responsible for deep computing and global optimization.

[0060] Hardware System Composition and Area Division: The hardware component of this embodiment comprises five main parts: a power system, a piping system, a physical model system, a monitoring system, and a control system. To simulate the ventilation environment of underground engineering projects, the experimental system is divided into four functional areas, with each hardware subsystem distributed within them: Air intake area: This is the area where the air inlet and outlet of the model ventilation system are located. The main function of the air intake area is to serve as a transitional connection between the test area and the atmospheric environment, ensuring that the model ventilation network is connected to the outside atmosphere. It is used for air intake and exhaust from the atmospheric environment. The air inlet and outlet are located at both ends of the laboratory site to reduce the impact of exhaust stale air on fresh air.

[0061] Level 1 Air Control Zone: This area is equipped with the main power system and main ductwork of the ventilation system. The power system consists of a reversible adjustable-speed main fan unit and its matching intelligent fan control cabinet. The main fan unit can operate in both directions (forward rotation for air supply, reverse rotation for air exhaust) and achieves precise airflow control through stepless speed regulation via a frequency converter. The intelligent fan control cabinet is installed near the main fan and supports manual / automatic and remote mode switching. Internally, it contains a human-machine interface display screen, a core control unit, a communication module, and a power module (its detailed structure is described above); the main fan outlet is connected to... The main air duct of the piping system is constructed in the primary air control zone by connecting flexible ducts of different diameters in series, parallel, and angles to form the main ventilation network, simulating the airflow of an underground tunnel. To monitor and regulate the airflow in the main air duct, airflow monitors and automatic regulating valves are installed at the fan outlets or important straight duct sections. The airflow monitors continuously measure the airflow passing through the cross-section, while the automatic regulating valves can change their opening degree according to control commands to control the amount of airflow passing through that section. Through these power and piping network components, the primary air control zone provides the total airflow for the entire system and performs primary distribution.

[0062] Secondary air control zone: This area includes the region where the ventilation network extends to the next level of branches, essentially the location of secondary branches of the ventilation network. Based on the ventilation duct architecture, this zone contains multiple branch nodes and local loops, and each major branch is equipped with an intelligent electric double-leaf regulating valve (which adjusts the ventilation resistance of the branch by adjusting the blade angle), an intelligent wind speed monitor, and an intelligent differential pressure monitor. These devices combine to form secondary local airflow control: through the coordinated opening adjustment of the regulating valves, on-demand air supply and automatic pressure gradient balancing are achieved for each ventilation branch. For example, when a branch (corresponding to a tunnel or working face) needs increased airflow, the system can increase the opening of the valve in that branch and correspondingly reduce the opening of valves in other branches to meet the branch's needs within a limited total airflow. In the secondary air control zone, each regulating valve is typically designed with a "double-leaf" structure (i.e., the valve consists of multiple blades that can open and close synchronously) to ensure smooth and wide-range resistance changes during adjustment. The wind speed monitor and differential pressure monitor monitor the wind speed of each branch and the pressure difference between the branch and a reference point in real time, providing feedback to determine whether the adjustment has achieved the expected results.

[0063] The physical model wind zone houses the physical model system and a monitoring system covering the model, simulating the actual wind-exposed areas inside underground engineering projects (such as tunnels or caverns). The physical model is designed according to the geometric scale of underground spaces (such as tunnels and underground factories) (constraining the overall model size with a length scale of 1:30–1:150), employing a modular assembly method. The model internally constructs typical structures such as straight-walled circular arch tunnels, circular tunnels, large-section factories, shafts, tunnel bifurcation, and cavern connections to simulate the airflow distribution and gas accumulation characteristics in real space. The physical model design follows similarity theory, satisfying Euler's criterion and Reynolds' criterion to ensure that the airflow pattern within the model is similar to that at the actual scale. The monitoring system deploys dense sensing devices in the physical model area, including: an integrated air environment monitoring rack, intelligent smoke detectors (monitoring smoke and fire signs), and intelligent motorized louvered air vents (controlling the opening of the terminal air outlets). The system includes several components: a direction sensor for precisely allocating airflow into the model space; intelligent electric double-leaf regulating valves (similar to the secondary air control zone, used for regulating airflow in terminal branches); intelligent wind speed monitors; and intelligent differential pressure monitors (monitoring wind speed and differential pressure at key points within the model, such as the pressure difference between inside and outside a cavern, to determine ventilation effectiveness). The monitoring system communicates with the control service via the Modbus protocol, uploading environmental data and ventilation parameters collected from the model area to the host computer. Simultaneously, it receives adjustment commands from the cloud to execute terminal controls (such as automatically opening and closing louvers to change the airflow direction and volume according to commands). The model's airflow zone is crucial for verifying the effectiveness of ventilation control. The monitoring system focuses on the deviation between the actual environment and the target environment within the model, as well as the deviation between ventilation parameters and model-predicted parameters. Once a deviation exceeds a threshold, the system triggers a control algorithm for correction, such as further adjustment via terminal valves, to ensure that environmental indicators within the model reach the desired level.

[0064] Control System: This system spans all the aforementioned areas and comprises various actuators (electric regulating valves) and intelligent detectors, forming a closed-loop control network. The system connects via a unified industrial communication protocol (such as Modbus TCP / RTU): electric regulating valves receive control commands to adjust the airflow resistance opening, and intelligent detectors collect status feedback signals. Together, they achieve fully automatic control from the main fan to the terminal air outlets. For example, when management issues an airflow adjustment command, the control system can simultaneously adjust the main fan frequency and the opening of corresponding branch valves, while monitoring changes in wind speed and pressure difference at various locations, utilizing feedback to achieve on-demand joint adjustment. In the event of an anomaly (such as a sensor detecting smoke or a valve malfunction), the control system also has rapid response capabilities: triggering pre-planned commands (such as immediately closing relevant branch valves to prevent smoke and dust diffusion) and uploading the anomaly information to management for alarm purposes.

[0065] The above details the construction method of the knowledge-driven intelligent ventilation control system for underground engineering from aspects such as system architecture, deployment mode and hardware composition. Through knowledge and data fusion, layered design and edge-cloud collaboration, as well as a sound monitoring and control network, the system lays the foundation for the operation of subsequent intelligent algorithms. The following embodiments will further introduce the specific implementation of the intelligent control algorithms and optimization models adopted by the system.

[0066] Example 2: This example provides a ventilation control decision algorithm based on deep reinforcement learning to achieve intelligent airflow regulation. The method runs on the platform of the above system and utilizes a large amount of ventilation status data collected by the data layer. Through the learning algorithm, it automatically mines and optimizes the regulation strategy, provides decision instructions for airflow regulation devices such as fans and air valves, and realizes adaptive iterative optimization of ventilation parameters. The principle and implementation steps of this deep reinforcement learning control method are described in detail below.

[0067] First, the ventilation control problem in underground engineering needs to be modeled as a Markov decision process in order to apply reinforcement learning algorithms, with the following elements defined: State space S: The system state consists of various parameters of the ventilation environment and equipment operation, and can be represented by a multi-dimensional vector. Typical state variables include: environmental indicators at each monitoring point (such as CO concentration, CO2 concentration, oxygen concentration, temperature, humidity, dust concentration, etc.), wind speed and air volume at key nodes / branches of the ventilation network, operating parameters of the main fans (frequency, air pressure, power consumption), and the opening degree of each regulating valve, etc. The current state is denoted as S. ,but It can be represented as: , The formula above includes the air volume of each ventilation branch. Wind speed Gas concentration at each monitoring point , Environmental information, as well as fan pressure / power. Information such as equipment was used to construct a state space that can fully characterize the operating status of the ventilation system.

[0068] Action Space A: The actions that the intelligent agent (i.e., the control algorithm) can execute include various adjustment operations on the ventilation equipment. In this system, the main control methods are adjusting the operating parameters of the main fan (such as frequency or speed) and adjusting the opening degree of each branch damper. Therefore, the action vector... It can be defined as: , in, This indicates the amount of adjustment to increase or decrease the frequency (or speed) of the main fan. This indicates the adjustment amount for the opening of the nth adjustable damper (a positive value means opening the valve wider / reducing resistance, and a negative value means closing the valve narrower / increasing resistance); Depending on the system implementation, the action can be discrete (e.g., increasing the fan frequency by 1Hz or adjusting the valve by several fixed steps) or continuous (fine adjustment of any amplitude). If discrete action is used, the action space is a finite combination of discrete control commands. If continuous action is used (in this embodiment, since the variable frequency fan and electric valve can be continuously adjusted, a continuous action space is preferred), then the action space A is a multi-dimensional real number space, with each dimension taking values ​​within the allowable range (e.g., fan frequency ±5Hz range, valve opening 0-100% range).

[0069] State transition environment: The evolution of the ventilation system follows the combined effects of physical laws and control operations. When the agent is in a state... Execute the following action After that, the system state will transition to a new state. This process can be described using a transfer function. Represented as: , in, It implies the physical evolution of the ventilation network (such as how changing the fan frequency affects the air volume distribution, and how new monitoring quantities are obtained through wind resistance relationships and node balance). It indicates unmeasurable disturbances or uncertainties in the environment; although precise It is difficult to represent analytically, but reinforcement learning algorithms do not require explicit expression. Instead, it learns state transition patterns implicitly by interacting with environmental data.

[0070] Reward Function: The reward function defines the "good" or "bad" evaluation of each control action, guiding the agent to optimize its strategy. In ventilation control scenarios, the reward design must consider both ventilation effect and cost: on the one hand, it aims to meet the needs of on-demand air supply in each area and ensure safety standards are met; on the other hand, it aims to reduce energy consumption and equipment wear. Therefore, a continuous, real-time reward is designed. , in, This indicates a set of key monitoring items or branches with demand constraints (such as work surfaces that require guaranteed airflow, or points where low concentrations of harmful gases need to be controlled). This represents the target value of the i-th indicator (such as the required air volume for a certain area). This represents the current value of the i-th indicator. The larger the squared difference between the two, the greater the deviation from the target, resulting in negative utility. This represents the weighting coefficient, which is set manually or adjusted through experimentation. This represents the set of costs incurred, such as the power consumption of the main ventilator. The opening and closing costs of several control valves, etc. This represents the weighting coefficient, which is set manually or adjusted through experimentation. This represents a measure of the energy consumption of the j-th item, such as the power consumption of a fan or the cost of frequent valve operation; reward. The reward is a negative loss-type reward; the smaller the absolute value, the better the performance. When all indicators meet the target and energy consumption is low... It approaches 0; conversely, if the deviation is large and the energy consumption is high, then... With a relatively large negative value, this reward setting allows the agent to weigh ventilation quality (on-demand air supply, safety) and energy consumption costs when optimizing strategies, tending to find the control scheme with the best overall performance.

[0071] Based on the Markov decision process model described above, a deep reinforcement learning algorithm is used to solve for the optimal control strategy. This embodiment uses a value function-based method, represented by the Deep Q Network (DQN) algorithm, which utilizes a neural network to approximate the state-action value function. The algorithm approaches the optimal state through repeated trials and updates; the main steps are as follows: Step 1, Initialization: Initialize the parameters (random weights) of the deep Q-network. ), and initialize the experience replay pool to empty; set the discount factor. (Used to weigh the impact of future rewards), learning rate (Update step size), and initial parameters of the ε-greedy policy. (Used to balance exploration and exploitation).

[0072] Step 2, State Observation: Obtain the initial state from the environment. (The current ventilation status is obtained through system monitoring data).

[0073] Step 3, Action Decision: At each decision step, select an action based on the current Q-network. The action selection follows an ε-greedy strategy: based on probability. Randomly select an action (exploration), with probability... Select the action with the highest current value (exploitation); specifically, calculate... For all possible actions The estimated value, selection As the optimal action ;by Probabilistic execution ,by It performs a random action with probability; as training progresses... Gradual decay reduces random exploration.

[0074] Step 4: Perform the action and receive the reward: Select the action. The instructions are sent to the ventilation system for execution, such as adjusting the fan frequency and the opening of several valves. The environment then reports back the new status after a short period of time. (Acquired from a new round of sensor readings), and the immediate reward for this action is calculated simultaneously. (Based on the aforementioned reward function formula), record the triplet. .

[0075] Step 5: Store the experience gained from the above interactions. Store the experience in the experience replay pool. When the experience stored in the pool exceeds the preset capacity, discard the oldest experience.

[0076] Step 6, Training and Update: Randomly sample a small batch of replay samples from the experience pool to train the Q-network. For each sampled experience, construct the target value: , in, Indicates the target Q-network, This indicates the action to be chosen in the next state. This represents the expected value of Q calculated based on the Bellman optimal equation. Indicates the current instant reward. This represents the discount factor, ranging from 0 to 1, used to balance the importance of immediate rewards and future rewards; Indicates weight; Then calculate the current Q-network for state-action pairs. Estimate The loss function is constructed using mean squared error: , By analyzing the loss function with respect to Taking the derivative, the neural network weights are updated using methods such as stochastic gradient descent or Adam optimization. The formula is expressed as: , in, Indicates the optimizer step size (corresponding to the aforementioned) ), This represents the gradient of the loss with respect to the parameters. This indicates the current network loss; this update affects the network output. Value to Expectation To move closer.

[0077] Step 7, Status Progression: Update the environment status: If the current round has not ended, return to step 3 and continue to the next decision; repeat this process to allow the agent to continuously interact with the environment, accumulate experience and optimize its strategy.

[0078] Step 8, Iteration and Termination: Repeat steps 3-7, gradually decreasing the speed during training. To reduce random exploration and make greater use of the learned strategy, training ends when a predetermined number of training rounds is reached or the overall policy converges (e.g., the average reward improvement in recent rounds is not significant), ultimately yielding a near-optimal Q-network parameter. Thus, the optimal strategy was determined. During deployment and implementation, the intelligent agent will follow the optimal strategy. Select control actions for the ventilation system in real time.

[0079] To improve efficiency and security during training, this embodiment also incorporates the aforementioned mechanistic and empirical knowledge into the deep reinforcement learning algorithm: Mechanism knowledge integration: Physical models or simulators are used to provide safety boundaries and reference actions. For example, strong penalties for physical violations can be added to the reward function (such as giving a large negative reward when the air volume allocation does not meet the node continuity requirements) to ensure that the learning process does not attempt actions that obviously violate physical laws. In addition, ventilation simulation models (based on equations such as Navier-Stokes) can be used to generate some "ideal state-action-result" data pairs offline and pre-fill them into the experience replay pool to help with the initial training of the network. When the actual environmental state exceeds the safety range (such as fan overload, gas exceeding the limit), the agent should be constrained not to perform aggressive actions. This can be achieved by adding a physical check module before the policy output, that is, verifying whether the action meets the safety conditions of the mechanism model before it is issued. If it does not meet the conditions, it should be pruned or replaced with a safe action.

[0080] Incorporating experiential knowledge: Expert experience and historical data patterns are used as reward adjustments or policy priors. For example, the experiential rule "when gas exceeds the limit, the air volume in the area should be increased immediately" can be transformed into giving extra rewards for the action "increasing the frequency of the fan or opening the relevant valve" when the gas concentration is high. Conversely, if the action contradicts the experience, the reward is reduced. In addition, imitation learning can be used to pre-train the policy network using past control operation data, allowing the agent to imitate human or existing control logic behavior, and then switch to the reinforcement learning autonomous optimization stage to reduce exploration time.

[0081] Using the above methods, deep reinforcement learning algorithms can converge to high-performance ventilation control strategies more quickly. After training, the agent is deployed in the intelligent decision support module of the system application layer to read monitoring data in real time. And output optimization instructions Through actual operation verification, this deep reinforcement learning algorithm can automatically meet the air volume demand under different working conditions: for example, when the number of people working on a tunnel face increases and the oxygen demand rises, the algorithm detects that the wind speed or oxygen content in that branch decreases, and quickly increases the air volume in that branch (by increasing the frequency of the main fan and opening the valve in that branch), while avoiding excessive disturbance to other branches; when the overall ventilation load decreases (such as when some areas are shut down), the algorithm automatically reduces the fan speed and closes the valves in unoccupied areas to achieve energy-saving operation; this demonstrates that the agent has "learned" a balance strategy between on-demand air supply and energy saving through reinforcement learning.

[0082] Example 3: This example introduces the implementation and application logic of the natural air distribution optimization model in the system. Natural air distribution refers to a state in which the air volume of each branch of the ventilation network is naturally distributed according to the principle of minimum total resistance, given a total intake air volume or total exhaust air volume. In other words, without applying additional adjustments to each branch in the network (all air valves are open or there is no increase in resistance due to adjustment), the airflow will be distributed automatically based on the resistance characteristics of the pipeline itself. The natural air distribution state is usually regarded as a benchmark operating condition of the ventilation network and can be used as a reference for subsequent optimization and control. The mathematical model, optimization calculation method and application in this system are described in detail below.

[0083] Establishing a mathematical model of the ventilation network is fundamental to solving the problem of natural wind distribution. Using graph theory, the ventilation system can be represented as a network topology with loops, where nodes represent intersections of channels and edges represent ventilation channels (pipelines) or ventilation equipment (fans). For a network containing f main fans, we can abstract n branches, m nodes, and the number of independent loops. The steady state of the ventilation network satisfies the principles of mass conservation and energy conservation, which corresponds to: For each node i, the sum of the airflow flowing into that node equals the sum of the airflow flowing out of that node, which can be represented by a connection matrix as follows: , in, This represents the air volume of the j-th branch. This represents the node-branch correlation coefficient, i.e., when the airflow from branch j exits node i. When the flow enters node i When not passing through node i ; For each independent loop The sum of the pressure drops along the loop is zero (when there is no external force). If the loop includes a fan, the sum of the pressure drops equals the pressure provided by the fan. For example, for a closed loop without a fan: , , For a circuit containing one fan: But one of them: , in, This represents the pressure loss function of the j-th branch. Indicates branch resistance, This represents the loop-branch correlation coefficient, i.e., when branch j is in a loop. When the direction of the middle path is consistent with the direction of the loop traversal The opposite is true. If not in that circuit ; This represents the air pressure provided by the fan (denoted as a positive value). In the equation, the fan is treated as a special "branch" with its pressure drop. This indicates that energy is provided, and this set of equations ensures that in any loop, the pressure supplied by the wind turbine is balanced with the resistance loss along the way; The above-mentioned nodal balance and loop balance equations together form the basic equation set of the ventilation network. If the resistance of each branch is known... Given the fan operating conditions (given fan pressure or total air volume, etc.), solving this system of equations will yield the air volume of each branch under natural conditions. Distribution and nodal pressure; loop equations contain The term is solved using numerical iterative algorithms (such as the traditional Hardy-Cross loop-by-loop correction method or the Newton-Raphson iterative method based on nodal pressure); when the fan pressure limit is not considered, the solution is... This is the result of natural wind distribution.

[0084] Mathematically, natural wind distribution can be viewed as an optimization problem: minimizing the total energy consumption of the network while satisfying node balance constraints and total airflow constraints. Taking a single-fan system as an example, let the total airflow be... (Wind turbine provided), the optimization objective can be expressed as: , , , Wherein, objective function The total pressure loss power consumption of the entire network is represented by the constraints including constant total flow and node continuity. It can be proved by the Lagrange multiplier method that the solution of this optimization problem satisfies the aforementioned node / loop equation, which is the natural wind distribution solution. The above principle of minimum total resistance ensures optimal energy utilization. Without additional control, the system naturally tends to a certain airflow distribution, so that the overall resistance loss is minimized.

[0085] In the software platform of this system, the natural wind distribution model serves as a basic module, providing the following functions: The ideal air volume distribution is calculated based on the design parameters of the tunnel ventilation system (network topology, tunnel length, cross-section, facility resistance, etc.). This provides a benchmark for subsequent analysis. For example, by comparing the natural air distribution results with the actual monitored air volume, it is possible to find which branches are abnormal (if the actual value deviates too much from the theoretical value, there may be changes in resistance or air leakage).

[0086] During the ventilation design phase, the total air volume requirement is input, and the air volume of each working face / branch is calculated using the natural air distribution model. If the air volume at some key locations is insufficient, it indicates that the design scheme cannot meet the requirements without control, and it is necessary to increase the fan pressure or install local dampers for adjustment. This calculation guides the designers in setting the location and parameters of the initial control facilities.

[0087] When abnormal sensor airflow data occurs, assuming the current airflow provided by the fan is known, the theoretical distribution can be calculated using the natural airflow distribution algorithm. If the actual airflow of a certain branch is much lower than the natural airflow distribution value, there may be local blockage or a sudden increase in resistance; conversely, if it is much higher than the theoretical value, the airflow path may be blocked elsewhere, causing excessive airflow distribution. This helps to locate the fault point.

[0088] In energy consumption optimization, the total resistance loss of natural wind distribution is used as an ideal low-consumption reference. If the resistance loss generated by the control or hybrid scheme is close to this value, it indicates that the additional loss caused by regulation is small and the scheme is more economical.

[0089] In its implementation, the natural air distribution calculation module first reads the ventilation network structure and parameters from the database or user input, and then performs network calculation (a professional ventilation network solving algorithm can be selected). After the calculation is completed, the output will include: the air volume of each branch. Node pressure Total pressure required for each fan Users can view the natural air distribution results through a visual interface at the application layer: for example, the thickness of the arrows in the 3D model indicates the air volume of each air path, and output data reports such as the flow rate of each node, the impedance of each branch, and the pressure drop.

[0090] It should be noted that although the natural wind distribution state has the optimal energy, when the demand in certain areas differs significantly from that of the natural wind distribution, it is necessary to use the following controlled wind distribution or mixed-type regulated air supply model for adjustment and optimization.

[0091] Example 4: This example illustrates a control-type airflow optimization model, which is an optimization method for achieving stable airflow adjustment based on sensitivity analysis under specified on-demand air supply requirements. "Airflow adjustment" refers to adjusting the ventilation network to calculate the fresh air volume distribution that meets specific constraints. "Control-type" means that it includes constraints and optimal goals (such as certain branches must reach a specified air volume as a constraint, and the minimum number of facilities involved in adjustment actions as the optimal goal), which is achieved through a combination of control actions of the facilities. This model focuses on maintaining system stability as much as possible while meeting the demand and reducing unnecessary adjustment actions. The model principle, algorithm steps, and application logic are detailed below.

[0092] The task of a controlled airflow model is to solve for the adjustment values ​​of other adjustable branches, given that airflow targets are set in some branches, so that the entire network achieves a new airflow distribution; let... For the set of branches that need to meet the on-demand air supply requirements, for each branch Target air volume Then, control-type optimization can be expressed as: when satisfying... Under these conditions, select a set of branches that need to be adjusted. and their respective resistance changes , making as small as possible or The size should be as small as possible, while ensuring that the new steady state of the network is achievable without generating large pressure fluctuations.

[0093] To achieve the above optimization, the concept of a sensitivity coefficient for the ventilation network is introduced. Sensitivity Defined as: Under the current operating conditions, the unit adjustment action in branch j (such as increasing its resistance) The change in air volume of the target branch i caused by ) is: , The partial derivatives can be approximated by linearizing the ventilation network equations. The larger the value (positive or negative indicates direction), the more significant the impact of adjusting branch j on the airflow of branch i. A positive sensitivity means that increasing the resistance of branch j will increase the airflow of branch i, while a negative sensitivity will have the opposite effect. This is achieved by calculating the sensitivity vector of all candidate adjustment branches. Sort out the branches that have the greatest impact on the target branch i. These branches are the ones that may need to be adjusted to meet the requirements.

[0094] Based on the sensitivity analysis results, an iterative heuristic optimization process is adopted, including the following steps: Step 1, Initialization: Starting from natural wind distribution or current operating conditions, calculate each target branch. The current airflow is compared with the target airflow to obtain the difference. Define an empty adjustment set. .

[0095] Step 2: Select the adjustment target: For each target branch i that has not yet been satisfied, examine the branches with the largest absolute sensitivity values ​​(excluding those already satisfied). (in the middle), select one of them Branch j with the correct direction of influence is added. For example, if (If you need to increase the airflow in target branch i), then select... If branch j is positive and large, it means that reducing branch j will drive more airflow to the target branch i; if (If it is necessary to reduce the airflow of target branch i), then select... For a negative and large branch j (the sign depends on the defined direction), a set of candidate adjustment branches is obtained.

[0096] Step 3: Determine the adjustment range: For each selected... Determine an adjustment resistance increment Estimated using a linear approximation method: (When a branch j affects one target branch i), when a branch j affects multiple target branches i, establish a small-scale linear system of equations and solve multiple simultaneous equations. Solve For safety reasons, it is advisable to start with a smaller adjustment step (such as half of the expected amount) to avoid over-adjustment.

[0097] Step 4, Simulation Verification and Iteration: [The text abruptly ends here, likely due to an incomplete sentence or The adjustments in the model are applied to solve for the new airflow distribution when all If all targets are met (within the tolerance range), a wind removal solution has been successfully found; if there are still deviations, then update... Then continue iterating: Based on the new sensitivity or continuing to use the previously calculated sensitivity, further adjust the existing sensitivity. within The amplitude or other branch adjustments may be added.

[0098] The application of the controlled air distribution model in the system is mainly reflected in on-demand air supply and localized air control scenarios: On-demand ventilation: When the construction intensity at a certain working face of the tunnel increases and a larger ventilation volume is required, a new target air volume is calculated based on the abrupt change mapping of environmental monitoring concentration values ​​for that branch target air volume. The system calls the control model and calculates a series of valve adjustment operations (such as closing some non-critical branch valves and increasing the operating frequency of related fans) to direct more air to the target area. After these operations are executed, the air volume in the target area increases to the required value, while the air volume in other areas remains within the preset acceptable range, and there is no multi-area air volume linkage change in the system.

[0099] Localized airflow control: In ventilation adjustments, selecting appropriate combinations avoids a domino effect. For example, with two adjacent work surfaces A and B, arbitrarily shutting off airflow to B while only increasing airflow to A might lead to severe insufficient airflow to B. Therefore, the model considers the needs of B and may even include B in the overall system. By working together, we can find a balance from a holistic perspective.

[0100] On the interface, when a user initiates an on-demand air supply solution, the system displays the selected target and its requirements, and then provides a list of suggested control schemes (such as specific instructions like "close damper X by 20%, damper Y to 30% opening, and increase the main fan frequency by 2Hz"). Users can view the scope of these instructions' impact and the predicted results (simulated in advance by the network calculation module). After confirmation, the system executes the command. During execution, the monitoring module verifies the achievement of each target air volume. If a target is not met, the system can automatically initiate the next round of iterative adjustments until the requirements are met.

[0101] Example 5: This example introduces the design and application of a hybrid regulating air supply optimization model. Hybrid regulating air supply refers to a system that, under the premise of specified on-demand air supply requirements, simultaneously considers the operational optimization of ventilation power equipment (fans) and achieves the minimum fan energy consumption through the joint regulation of fans and air valves. Compared with the aforementioned control model, the hybrid model further takes minimizing fan power as its objective, which falls under the category of energy-saving optimization. The mathematical description, solution approach, and application examples of the hybrid model are explained in detail below.

[0102] In hybrid optimization, the goal is to satisfy the airflow requirements of all critical branches while minimizing the operating power of the main ventilator, which constitutes a constrained optimization problem. Decision variables: Adjustable elements in the ventilation network, including the operating parameters of the main fans (such as fan frequency or speed, affecting fan pressure and air volume supply) and the opening degree of each adjustable damper (affecting the resistance of each branch), which can be represented by a vector. Represents all decision variables, such as ,in, Indicates the frequency of the fan. This indicates the opening degree of the k-th regulating valve (0-100% or the corresponding resistance value).

[0103] Optimization objective: Minimize wind turbine power For a given fan model, the power is related to the air volume and air pressure provided by the fan, and can be approximated as: , in, This indicates the total air supply volume of the fan. This indicates the total pressure of the fan, i.e., the output air pressure; This indicates the efficiency of the wind turbine.

[0104] Constraints include: On-demand air supply constraint: For each demand branch Its air volume ,in, This represents the actual air volume of branch i under the current adjustment state. This represents the desired airflow in target branch i; Network mechanical constraints: given the wind turbine frequency hour, and To fall on the fan characteristic curve, the valve opening... With additional resistance There are corresponding relationships, which can be embedded in the optimization model as implicit constraints, and the results are obtained by simulation programs.

[0105] The range of values ​​for decision variables includes the frequency range of the wind turbine (e.g., ...). (corresponding to the upper and lower limits of safe operation frequency of the fan) and the opening degree of each valve It may also include practical constraints such as the minimum number of valve adjustment steps.

[0106] Given the complexity of the optimization objectives and constraints, heuristic or intelligent optimization algorithms are often used to solve the problem, such as genetic algorithms, particle swarm optimization, or simulated annealing. These algorithms do not require the objective function to have an analytical gradient; performance can be evaluated by repeatedly trying new solutions. In the implementation of this system, the following strategy can be adopted: Step 1, Encoding and Initial Population: Encode the decision vector Encode individuals (such as gene strings encoded with real numbers), initially generating a set of random individuals (population).

[0107] Step 2: Set fitness To minimize the objective function, add penalties for constraint violations, for example: , in, Denotes the overall optimization objective function. This represents the total power of the fan under the current adjustment parameter x; This represents the penalty coefficient, used to measure the degree of punishment for not meeting the target airflow.

[0108] Step 3, Evolutionary Iteration: Perform selection, crossover, and mutation operations using a genetic algorithm, or update the velocity and position using particle swarm optimization, continuously generating new solutions, and calling the ventilation network simulation module for calculation. and each The fitness is evaluated, and after multiple iterations, the decision vector converges to an approximately optimal one. .

[0109] Step 4, Decoding and Result Verification: Decode the optimal individual back to the actual operating parameters to verify that all constraints are met and the wind turbine power is at its lowest.

[0110] In this system, the hybrid air supply regulation model is mainly used in ventilation energy-saving scenarios, such as night shifts or non-production periods in mines, where only basic ventilation is required. In this case, hybrid optimization can obtain an operating scheme with extremely low fan power, significantly saving energy. The application steps are illustrated below: Assume a mine has three main ventilation working faces, A, B, and C, with a minimum required ventilation volume of 100 m³ / s. 3 / min, 80 m 3 / min, 50 m 3 / min, during daytime production, the total air volume demand is high, and the fan frequency may need to operate at 50Hz to meet the total air volume; at night, production A stops and only ventilation maintenance is required, while activities in B and C are reduced; input the above data into the hybrid optimization model: demand set , etc.; Initial operating point: daytime The total pressure satisfies the total Q=300 m 3 At approximately / min, at night it is permissible to reduce it to the lowest possible level, but the total air volume required for A, B, and C must be 230 m³ / min. 3 / min is sufficient; the model searches using a genetic algorithm and obtains the following results: closing 80% of the irrelevant branch valves reduces the fan frequency to 40Hz, and the fan power decreases from 100 kW to 60 kW, resulting in A being 100 m 3 / min, B gets 80 m 3 / min, C gets 50 m 3 / min just meets the requirement, and this is the optimized running parameter.

[0111] At the system application level, it will be presented in the form of one-click optimization: when the administrator clicks "Energy-saving optimization operation", the system prompts "It is recommended to reduce the fan frequency to 40Hz, close the return air door of X zone to 20% opening, and stop the local ventilation fan of Y zone", etc. Then the user confirms the execution. After execution, the monitoring data will show that the total air volume has decreased, and each air demand surface has just met the standard and is not over-supplied. From this, the current power can be counted, and the power consumption can be seen to be reduced by about 40%.

[0112] By optimizing the air supply through hybrid regulation, the power of the fan under different operating conditions is dynamically optimized to the minimum required level, significantly reducing unnecessary air supply and thus reducing power consumption. At the same time, due to the coordinated regulation of valves, the rest of the network will not experience insufficient air supply or system disorder due to fan speed reduction. This method of joint optimization of fans and air paths is significantly more energy-efficient than the traditional fixed fan frequency plus manual air adjustment strategy.

[0113] Through the detailed description of Embodiments 1 to 5 above, the knowledge-based intelligent ventilation control test system and method for underground engineering provided by the present invention achieves comprehensive innovation from architecture design to algorithm control. Embodiment 1 illustrates how the system integrates the bidirectional mapping of physical entities and knowledge data. Embodiment 2 provides the algorithm steps and formula description for intelligent airflow regulation using deep reinforcement learning. Embodiments 3 to 5 respectively provide specific model implementation details and application scenarios for different optimization objectives (natural air distribution, controlled on-demand air distribution, and hybrid energy-saving air supply). The embodiments are interconnected and support each other, ultimately forming an intelligent ventilation control system that can self-perceive, self-determine, and self-optimize, and has broad application prospects in the field of underground engineering ventilation.

[0114] While specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that any non-creative modifications or equivalent substitutions to the technical solutions of the present invention, without departing from the scope of protection of the present invention, should be covered within the scope of the claims of the present invention.

Claims

1. A knowledge-based intelligent ventilation control test system for underground engineering, characterized in that, The system comprises a knowledge-driven intelligent ventilation architecture. This architecture integrates physical regulation with data knowledge management through a bidirectional mapping mechanism between ventilation system entities and knowledge data. This bidirectional mapping facilitates real-time information interaction during ventilation regulation decision-making, including iterative optimization of data acquisition, transmission, analysis, and feedback, ultimately feeding the optimized solution back to the ventilation equipment to provide regulation decision services. Furthermore, it establishes a ventilation hypergraph network that reflects the ventilation system structure and airflow adjustment process. The ventilation system performs entity regulation based on the optimized regulation scheme derived from all elements of the hypergraph, ensuring consistency between the actual ventilation system and the generation of data knowledge, and verifying and supplementing the data knowledge. The system further comprises four layers: a data layer, a computing layer, an application layer, and a management layer. The data layer is responsible for data perception, integration, and management; the computing layer relies on cloud computing to realize the transformation of data into knowledge. The application layer constructs intelligent ventilation scenario modules; the management layer provides user interaction and management services. The system adopts an edge-cloud deployment approach to achieve coordinated allocation of computing and storage, forming a closed-loop operation system. The hardware system of the system is divided into a power system, a pipeline system, a physical model system, a monitoring system, and a control system, and is configured in the wind collection area, the physical model wind use area, the primary wind control area, and the secondary wind control area. The system also includes an intelligent management and control software platform, including data management, model management, and knowledge engineering modules.

2. The knowledge-based intelligent ventilation control test system for underground engineering according to claim 1, characterized in that, The bidirectional mapping mechanism incorporates mechanistic and empirical knowledge into the ventilation regulation process. The mechanistic knowledge is expressed in mathematical form, including formulas or equations, and is used for the explicit characteristics of ventilation control. Experienced knowledge represents the correlation between control actions and the process environment, serving as the implicit characteristics of ventilation control and manifested in the form of regulation data. The regulation decision service system employs deep reinforcement learning algorithms to mine the data, providing decisions and instructions for the intelligent airflow regulation device, and realizing iterative optimization of ventilation parameters. It includes a monitoring part and a control part. The monitoring part is used for data acquisition and feature extraction processing, while the control part is used to receive status information and map it into adjustment actions to meet the airflow requirements under different operating conditions.

3. The knowledge-based intelligent ventilation control test system for underground engineering according to claim 1, characterized in that, The data layer uses digital sensors to collect data on air environment, ventilation fan power, regulation facilities, and branch networks. This data is stored and managed using relational databases, time-series databases, and graph databases, supporting large-scale data retrieval and analysis. The computing layer includes a data center, information model, and knowledge-based applications. The data center includes a data center, computing and storage, database, and disaster recovery backup. The information model, built on an IoT and big data platform, provides data exchange, ventilation scenario management, 3D spatial analysis and query, and visualization capabilities. The knowledge-based applications utilize a knowledge graph-based search engine to achieve knowledge fusion, representation, extraction, mining, and reasoning. The application layer includes modules for monitoring data, network calculation, dynamic feedback control, knowledge Q&A, data association, platform interfaces, early warning and forecasting, and access authentication. The management layer provides services such as monitoring data visualization, dynamic wind network calculation, intelligent decision support, remote monitoring, equipment regulation, anomaly diagnosis, disaster early warning, authorized login, and message processing. It displays data through 3D modeling and parameter visualization, providing system energy consumption optimization, ventilation environment early warning, and ventilation equipment fault information.

4. The knowledge-based intelligent ventilation control test system for underground engineering according to claim 1, characterized in that, In the described edge-cloud deployment, the edge includes field sensor terminals and variable frequency fans and damper adjustment terminals. An integrated air environment monitoring rack is installed in the air zone used for the physical model. Air volume and air pressure sensors are installed inside the air ducts. The fan control cabinet and damper adjustment device are installed near the variable frequency fans and adjustable dampers. The edge is an edge computing server deployed on an industrial control computer at the experimental site. It directly connects to the equipment and facilities at the edge for local data processing and decision-making, and communicates with the cloud. The cloud is a cloud computing center that performs big data analysis, storage, and model training, and distributes models and strategies. The field sensor terminals collect multi-parameter data synchronously through wind speed sensors, electrochemical CO sensors, electrochemical CO2 sensors, electrochemical oxygen sensors, and laser scattering dust sensors. The data is converted to Modbus through an industrial-grade signal conditioning module. The RS485 signal transmission of the RTU protocol uses time synchronization control technology to keep the delay within 500ms and the acquisition frequency controlled at 7.568s. The cloud computing center receives environmental and ventilation parameters collected by field sensor terminals, optimizes ventilation parameters using data models, and adjusts the working status of variable frequency fans and air conditioning facilities in real time based on the calculation results. The data models are divided into three types: natural air distribution, controlled air distribution, and hybrid optimization, and include network design, ventilation design, and optimization analysis modules.

5. The knowledge-based intelligent ventilation control test system for underground engineering according to claim 1, characterized in that, The air collection area is the area where the air intake and exhaust outlets of the model ventilation system are located. The power system is located in the primary air control zone, including a variable frequency drive (VFD) fan unit supporting forward and reverse rotation and an intelligent fan control cabinet. The fan unit supports manual and automatic operation modes, as well as remote control. The intelligent fan control cabinet is located near the VFD fan unit. The piping system is located in the primary air control zone, forming a ventilation network through series, parallel, and angle connections of flexible air belts of different diameters. It is equipped with airflow monitors and automatic regulating valves, which are located at the fan outlet or the straight section of the constant airflow path. The secondary air control zone includes the main lower-level branches of the ventilation network, and is hierarchically arranged according to the piping structure. It is equipped with corresponding intelligent electric double-leaf regulating valves, wind speed monitors, and differential pressure monitors. By adjusting the opening of the electric double-leaf regulating valve group, the on-demand airflow distribution and pressure gradient adjustment of each ventilation network branch are achieved. The physical model system is a scaled-down model of the underground engineering cavern group and ventilation system. The model design follows... Based on the Navier-Stokes equations, Euler's criterion, and Reynolds' criterion, the length scale ranges from 1:30 to 1:

150. The model is constructed using modular assembly and acrylic sheet materials. The monitoring system, located in the air zone of the physical model, includes an integrated air environment monitoring frame, intelligent smoke detectors, intelligent electric louvered air vents, intelligent electric double-leaf regulating valves, intelligent wind speed monitors, and intelligent differential pressure monitors. The monitoring system communicates with the control service using the Modbus protocol to acquire on-site air environment monitoring, air volume and pressure monitoring, and monitoring values ​​of power equipment and air conditioning facilities. The monitoring system monitors the deviation between the air environment in the air zone and the target environment, and between ventilation parameters and predicted parameters. Simultaneously, it receives adjustment commands from the cloud and performs end-point control through the intelligent electric louvered air vents and intelligent electric double-leaf regulating valves. The control system includes electric regulating valves and intelligent detectors, which communicate via the Modbus protocol to achieve opening adjustment and end-point control.

6. The knowledge-based intelligent ventilation control test system for underground engineering according to claim 5, characterized in that, The intelligent fan control cabinet is an independent enclosure (7), including: an upper enclosure, a lower enclosure, a control cabinet base (709), an upper cabinet door (703), and a lower cabinet door (705). The upper enclosure and the lower enclosure are fixedly connected to form the main frame of the independent enclosure (7). The bottom of the lower enclosure is fixedly connected to the control cabinet base (709). The upper cabinet door (703) is connected to the upper enclosure, and the lower cabinet door (705) is connected to the lower enclosure. The control cabinet base (709) is provided with a cable inlet hole (707) for the entry of external cables to form a channel connection with the internal circuit of the enclosure. The upper enclosure includes: a human-machine interaction display screen (713), a power module A (717), a communication module A (719), and a wiring trough (721). The system includes a core control data processing unit (723), a terminal block A (725), and a battery pack monitoring and charging / discharging control module (727). The human-machine interaction display screen (713) is fixed to the upper cabinet door (703) via a display screen inlaid door frame (715). The human-machine interaction display screen (713) is connected to the core control data processing unit (723). The power module A (717) is electrically connected to an external AC220V power supply via a terminal block A (725) and an inlet hole (707). The power module A (717) is electrically connected to the human-machine interaction display screen (713), the communication module A (719), the core control data processing unit (723), and the battery pack monitoring and charging / discharging control module (727). 27), the communication module A (719) is connected to the core control data processing unit (723), the communication module A (719) is connected to the external server through the wiring trough (721) and the terminal block A (725), the wiring trough (721) is connected to the power module A (717), the communication module A (719), the core control data processing unit (723) and the terminal block A (725), the core control data processing unit (723) is connected to the human-machine interaction display screen (713) and the communication module A (719) and is electrically connected to the terminal block A (725), the terminal block A (725) includes a leakage current protection air switch, and is connected to the power module A (717), the core control data processing unit (723) and the core control data processing unit (725). The data processing unit (723) is electrically connected to the external cable; a battery (711) is installed in the lower cabinet, and the battery (711) is electrically connected to the battery pack monitoring and charging / discharging control module (727); a transparent protective plate (701) is installed on the upper cabinet door (703), and the transparent protective plate (701) is fixed on the upper cabinet door (703) and covers the human-machine interaction display screen (713); a transparent protective plate (701) is also installed on the lower cabinet door (705), and the transparent protective plate (701) is fixed on the lower cabinet door (705); the intelligent fan control cabinet is connected to external equipment through the inlet hole (707) and the terminal block A (725), including the fan distribution box frequency converter and the external server.

7. The knowledge-based intelligent ventilation control test system for underground engineering according to claim 5, characterized in that, The integrated air environment monitoring frame includes: an integrated frame (1), sensors (3), and an environmental monitoring acquisition box (5). The integrated frame (1) is a grid frame structure. The environmental monitoring acquisition box (5) and several sensors (3) are fixed on the integrated frame (1). The sensors (3) are distributed and fixed at the grid positions of the integrated frame (1) to form an array arrangement. The sensors (3) include: a sensor display screen (301) and a sensor probe (303). The sensor display screen (301) is connected to the sensor probe (303). The environmental monitoring acquisition box (5) includes: a display screen (501), a leakage current protection air switch (503), a power supply module B (505), a communication module B (507), a terminal block B (509), and a data acquisition control module (511). The display screen (501) is connected to the data acquisition control module (511) and fixed on the door of the environmental monitoring acquisition box (5). The gas switch (503) is electrically connected to the power module B (505) and the terminal block B (509). The power module B (505) is electrically connected to an external AC220V power supply through the terminal block B (509). The power module B (505) supplies power to the sensor (3), the display screen (501), the communication module B (507), and the data acquisition control module (511). The communication module B (507) is connected to the data acquisition control module (511) and uploads the collected data to an external server. The terminal block B (509) is electrically connected to the data acquisition control module (511), the power module B (505), and the leakage protection air switch (503). The data acquisition control module (511) is connected to the display screen (501) and the terminal block B (509). The data collected by the sensor probe (303) is transmitted to the data acquisition control module (511) for processing via the terminal block B (509).

8. The knowledge-based intelligent ventilation control test system for underground engineering according to claim 1, characterized in that, The data management module of the intelligent control software platform provides display on both computer and mobile devices. The computer device offers model introductions, monitoring point information, monitoring indicator data curve changes, early warning forecasts, historical data query and export functions, allowing selection of the query time period and export of data in multiple data types, including raw data, hourly averages, and daily averages. The mobile device provides remote monitoring data query functions, including environmental monitoring indicators, preset frequencies, operating frequencies, current, local-remote status, and power consumption, among other fan operating indicators. Abnormal data is highlighted in red and given corresponding thresholds. The model management module is divided into natural wind distribution, controlled wind distribution, and hybrid optimization models, including network design, ventilation design, and optimization analysis modules. The network design module completes the process of establishing the ventilation network graph structure by extracting ventilation units, spatially associating them, and defining the number, number, name, description, coordinates, and directed connection parameters between units on the interface. The ventilation design module assigns characteristic parameters to each unit based on the graph structure. The calculation process includes unit physical structure parameters, ventilation characteristic parameters, and constraint parameters. In the basic equipment performance data configuration interface, performance curve parameters are configured or corrected performance curves are input, providing input data for the structural performance of fans, ducts, and irregular ducts. The optimization analysis module has calculation functions for natural air distribution, controlled air distribution, and mixed-type regulated air supply. It supports visualization, analysis, and export of ventilation network calculation results, displaying parameters such as node flow rate, branch flow rate and direction, branch impedance, and fan total pressure requirement. Natural air distribution is the airflow allocation with the lowest total resistance given a total intake or exhaust air conditioner. Controlled air supply is the optimization of stable air supply branch combinations based on sensitivity indicators, given on-demand air supply branch requirements. Mixed-type regulated air supply is the optimization of fan and valve combinations with the lowest fan operating power, given on-demand air supply branch requirements. The knowledge engineering module constructs a knowledge graph through the physical layer, data layer, and ontology layer to achieve knowledge storage, visualization, assisted design, knowledge retrieval, question answering, and report generation.

9. A knowledge-based intelligent ventilation control method for underground engineering, applied to the system described in claim 1, characterized in that, Includes the following steps: A knowledge-driven intelligent ventilation architecture is constructed to achieve bidirectional mapping and iterative optimization between entities and knowledge data. Edge-cloud deployment is adopted for data acquisition, processing, and model training. Hardware zones are divided, including an intake and exhaust area for the intake zone, a primary control zone for power and piping systems, and a secondary control zone for valve coordination. A physical model monitors deviations in the air zones and executes controls. Deep reinforcement learning algorithms are used for decision-making, with steps including initializing the state and action spaces, collecting data as states, selecting actions, executing and observing rewards, and updating the value function. Optimization models are applied, including natural air distribution, controlled air distribution, and hybrid regulated air supply.

Images